nLab fan theorem

Introduction

The fan theorem is one of the basic principles of intuitionism that make it more specific (even in mathematical practice, independent of any philosophical issues) than garden-variety constructive mathematics. Its main use is to justify pointwise analysis; without it, one really needs locale theory for point-free topology instead. In classical mathematics, the fan theorem is true.

Statement

Consider the finite and infinite sequences of binary digits. Given an infinite sequence $\alpha$ and a natural number $n$, let $\bar \alpha n$ be the finite sequence consisting of the first $n$ elements of $\alpha$.

Let $B$ be a collection of finite sequences of bits (or bitlists), that is a subset of the free monoid on the boolean domain. Given an infinite sequence $\alpha$ and a natural number $n$, we say that $\alpha$ $n$-bars $B$ if $\bar \alpha n \in B$; given only $\alpha$, we say that $\alpha$ bars $B$ if $\alpha$ $n$-bars $B$ for some $n$.

We are interested in these three properties of $B$:

• $B$ is decidable: For every finite sequence $u$, either $u \in B$ or $u \notin B$. (This is trivial in classical logic but will hold constructively only for some subsets $B$.)
• $B$ is barred: For every infinite sequence $\alpha$, $\alpha$ bars $B$.
• $B$ is uniform: For some natural number $M$, for every infinite sequence $\alpha$, if $\alpha$ bars $B$ at all, then $\alpha$ $n$-bars $B$ for some $n \leq M$.

A bar is a barred subset $B$.

Decidable Fan Theorem

Every decidable bar is uniform. (In other words, if a collection of bitlists is decidable and barred, then it is also uniform.)

(Full) Fan Theorem

Every bar is uniform. (In other words, if a collection of bitlists is barred, then it is also uniform.)

Although the fan theorem is about bars, it is different from the bar theorem, also known as “bar induction”, which is related but stronger. The fan theorem discusses bars on lists of elements of $\mathbb{B}$; the bar theorem discusses bars on lists of elements of $\mathbb{N}$.

Bar induction has three forms; decidable bar induction, monotone bar induction, and full bar induction (each stronger than the previous). The decidable fan theorem follows from decidable bar induction; the full fan theorem follows from monotone bar induction. Full bar induction, the strongest of the three, is a classical theorem; however, full bar induction implies Bishop’s limited principle of omniscience. Kleene's second realizability model provides a model of monotone bar induction plus Brouwer’s continuity principle.

Obfuscation

Let $\mathbb{B}$ be the set $\{0,1\}$ of binary digits (bits) and $\mathbb{N}$ the set $\{0,1,2,\ldots\}$ of natural numbers (numbers). Given a set $A$, let $A^*$ be the set of finite sequences of elements of $A$, let $A^{\mathbb{N}}$ be the set of infinite sequences of elements of $A$, and let $\mathcal{P}_{\Delta}A$ be the set of decidable subsets of $A$. Then the fan theorem is about (elements of) $\mathbb{B}^*$, $\mathbb{B}^{\mathbb{N}}$, and $\mathcal{P}_{\Delta}\mathbb{B}^*$.

However, the sets $\mathbb{N}$, $\mathbb{B}^*$, and $\mathbb{N}^*$ are all isomorphic. Similarly, the sets $\mathbb{B}^{\mathbb{N}}$, $\mathcal{P}_{\Delta}\mathbb{N}$, $\mathcal{P}_{\Delta}\mathbb{B}^*$, and $\mathcal{P}_{\Delta}\mathbb{N}^*$ are all isomorphic. In much of the literature on bars, one tacitly uses all of these isomorphisms, taking $\mathbb{N}$ and $\mathbb{B}^{\mathbb{N}}$ as chosen representatives of their isomorphism classes. Thus, everything in sight is either a natural number or an infinite sequence of bits.

The fan theorem is hard enough to understand when $\alpha$ is an infinite sequence of bits and $\bar \alpha n$ is a finite sequence of bits; it is even harder to understand when $\bar \alpha n$ is a natural number that bears no immediate relationship to the digits in the sequence $\alpha$.

Use in analysis

In classical mathematics, the fan theorem is simply true.

In constructive mathematics, the full fan theorem is equivalent to any and all of the following statements:

• As a locale, Cantor space has enough points (is topological).
• As a topological space, Cantor space is compact.

Assuming some amount of countable/dependent choice (probably pretty weak: haven’t checked myself though), the full fan theorem is equivalent to each of the following statements:

• As a topological space, the (located Dedekind) unit interval is compact (the Heine–Borel theorem).
• As a topological space, the (located Dedekind) real line is locally compact.
• There exists a class of “kontinuous” partial functions from the set $\mathbb{R}$ of (located Dedekind) real numbers to itself (see Waaldijk) such that
  • the restriction of a kontinuous function to any smaller domain is kontinuous;
  • the identity function on $\mathbb{R}$ is kontinuous;
  • the composite of two kontinuous functions is kontinuous;
  • a function whose domain is the unit interval is kontinuous if and only if it is uniformly continuous (in the usual metric-space sense); and
  • the function $(x \mapsto 1/x)$ defined on $\dot{\mathbb{R}}^+$ is kontinuous.

However, some of these equivalences fail in the absence of countable choice. See

• Moerdijk, Ieke. “Heine-Borel Does Not Imply the Fan Theorem.” Journal of Symbolic Logic, vol. 49, no. 2, 1984, pp. 514–519., doi:10.2307/2274182

which shows that the compactness of $[0, 1]$ does not necessarily imply the fan theorem constructively if one does not have some amount of countable choice available.

The decidable fan theorem (assuming some amount of countable/dependent choice) is equivalent to any of the following statements:

• Every uniformly continuous function from Cantor space to the metric space $\dot{\mathbb{R}}^+$ of positive real numbers has a positive lower bound.
• Every uniformly continuous function from the unit interval to $\dot{\mathbb{R}}^+$ has a positive lower bound.

For the above equivalences with the decidable fan theorem, see

• Julian, W., & Richman, F. (1984). A uniformly continuous function on [0,1] that is everywhere different from its infimum. Pacific Journal of Mathematics, 111(2), 333–340. https://doi.org/10.2140/pjm.1984.111.333

The above paper unfortunately uses different terminology than this article; for example, using “branch-bounded” and “bounded” in place of “barred” and “uniform”. The above paper also proves that, assuming all functions $\mathbb{N} \to \mathbb{N}$ are computable, there exists a decidable bar which is not uniform (and, as the title suggests, a uniformly continuous function $[0, 1] \to (0, \infty)$ whose range has an infimum of zero).

Some form of the fan theorem follows from any of these statements (some work needs to be done to verify these and get citations):

• The bar theorem holds (see above)
• As a locale, Baire space has enough points.
• Every pointwise-continuous function on Cantor space is uniformly continuous.
• Every pointwise-continuous function on the unit interval is uniformly continuous.
• Every uniformity (or indeed any metric) compatible with the usual topology of Cantor space is totally bounded.
• The property of being “complete and totally bounded”, i.e. Bishop-compact, is invariant under topological homemorphism between uniform (or metric) spaces.

I need to figure out how it relates to the various versions of König's Lemma?, as well as these statements (which are mutually equivalent):

• As a locale, the unit interval has enough points.
• As a locale, the real line (the locale of real numbers) has enough points.

Some of the results above may use countable choice, but probably no more than $AC_{0,0}$ (which is choice for relations between $\mathbb{N}$ and itself).

Analysis without the fan theorem

Point-wise real analysis without the fan theorem is very difficult, as shown by the example above from Waaldijk regarding “kontinuous” functions: without the fan theorem there isn’t really even a good notion of continuity! This was Brouwer's motivation for introducing the fan theorem.

However, the fan theorem (and bar theorem) can be avoided by instead using locales or another point-free approach to analysis.

Sheaf models

Fourman and Hyland prove that the Fan theorem is true for all toposes of sheaves on a topological space, while also providing a sheaf model not satisfying the fan theorem. In other words, for Localic toposes, the underlying locale being spatial implies that the internal locale of real numbers is spatial, and that the internal cantor locale is spatial.

The reference also proves that sheaf toposes over locally countably compact topological spaces satisfy the stronger bar theorem.

Proofs

I should write down the classical proof (which uses excluded middle), as well as Brouwer's argument.

Ideally, the page on bar induction would contain the classical proof of full bar induction, as well as Brouwer’s reasoning for why bar induction should hold. We would then provide constructive proofs here that monotone bar induction implies the full fan theorem, and that decidable bar induction implies the decidable fan theorem.

References

• Thanks to Giovanni Curi on constructive news.

• Frank Waaldijk pointed out exactly why point-wise analysis needs the fan theorem.

I need to read the relevant parts here:

• Mike Fourman, R. Grayson, Formal Spaces. In: The L.E.J. Brouwer Centenary Symposium, A.S. Troelstra and D. van Dalen, eds. North Holland (1982), pp. 107–122.

More links that I need to keep in mind:

• http://golem.ph.utexas.edu/category/2008/12/the_status_of_coalgebra.html
• http://www.jaist.ac.jp/is/labs/ishihara-lab/wcalm2010/berger.pdf
• http://www.cairn.info/revue-internationale-de-philosophie-2004-4-page-483.htm

Also:

• Mike Fourman, Martin Hyland, Sheaf models for analysis. PDF