This paper was converted on www.awesomepapers.org from LaTeX by an anonymous user.
Want to know more? Visit the Converter page.

Connecting real and hyperarithmetical analysis

Sam Sanders Department of Philosophy II, RUB Bochum, Germany sasander@me.com
Abstract.

Going back to Kreisel in the Sixties, hyperarithmetical analysis is a cluster of logical systems just beyond arithmetical comprehension. Only recently natural examples of theorems from the mathematical mainstream were identified that fit this category. In this paper, we provide many examples of theorems of real analysis that sit within the range of hyperarithmetical analysis, namely between the higher-order version of Σ11\Sigma_{1}^{1}-AC0\textup{{AC}}_{0} and weak-Σ11\Sigma_{1}^{1}-AC0\textup{{AC}}_{0}, working in Kohlenbach’s higher-order framework. Our example theorems are based on the Jordan decomposition theorem, unordered sums, metric spaces, and semi-continuous functions. Along the way, we identify a couple of new systems of hyperarithmetical analysis.

Key words and phrases:
Higher-order arithmetic, hyperarithmetical analysis
2010 Mathematics Subject Classification:
03B30, 03F35

1. Introduction

1.1. Motivation and overview

The aim of this paper is to exhibit many natural examples of theorems from real analysis that exist in the range of hyperarithmetical analysis. The exact meaning of ‘hyperarithmetical analysis’ and the previous boldface text is discussed in Section 1.3, but intuitively speaking the latter amounts to being sandwiched between known systems of hyperarithmetical analysis or their higher-order extensions. We shall work in Kohlenbach’s framework from [kohlenbach2], with which we assume basic familiarity.

We introduce some necessary definitions and axioms in Section 1.2. We shall establish that the following inhabit the range of hyperarithmetical analysis.

  • Basic properties of (Lipschitz) continuous functions on compact metric spaces without second-order representation/separability conditions, including the generalised intermediate value theorem (Section 2).

  • Properties of functions of bounded variation, including the Jordan decomposition theorem, where the total variation is given (Section 3).

  • Properties of semi-continuous functions and closed sets (Section 4.1).

  • Convergence properties of unordered sums (Section 4.2).

These results still go through if we restrict to arithmetically defined objects by Theorem 2.8. To pinpoint the exact location of the aforementioned principles, we introduce a new ‘finite choice’ principle based on finite-Σ11\Sigma_{1}^{1}-AC0 from [gohzeg] (see Section 1.2), using Borel’s notion of height function ([opborrelen4, opborrelen5]).

Finally, as to conceptual motivation, the historical examples of systems of hyperarithmetical analysis are rather logical in nature and natural examples from the mathematical mainstream are a relatively recent discovery, as discussed in Section 1.3. Our motivation is to show that third-order arithmetic exhibits many robust examples of theorems in the range of hyperarithmetical analysis, similar perhaps to how so-called splittings and disjunctions are much more plentiful in third-order arithmetic, as explored in [samsplit]. In this paper, we merely develop certain examples and indicate the many possible variations.

1.2. Preliminaries

We introduce some basic definitions and axioms necessary for this paper. We note that subsets of {\mathbb{R}} are given by their characteristic functions as in Definition 1.2, well-known from measure and probability theory. We shall generally work over ACA0ω\textup{{ACA}}_{0}^{\omega} -defined right below- as some definitions make little sense over the base theory RCA0ω\textup{{RCA}}_{0}^{\omega}. We refer to [kohlenbach2] for the latter.

First of all, full second-order arithmetic Z2{\textsf{{Z}}}_{2} is the ‘upper limit’ of second-order RM. The systems Z2ω{\textsf{{Z}}}_{2}^{\omega} and Z2Ω{\textsf{{Z}}}_{2}^{\Omega} are conservative extensions of Z2{\textsf{{Z}}}_{2} by [hunterphd]*Cor. 2.6. The system Z2Ω{\textsf{{Z}}}_{2}^{\Omega} is RCA0ω\textup{{RCA}}_{0}^{\omega} plus Kleene’s quantifier (3)(\exists^{3}) (see e.g. [dagsamXIV] or [hunterphd]), while Z2ω{\textsf{{Z}}}_{2}^{\omega} is RCA0ω\textup{{RCA}}_{0}^{\omega} plus (SSk2)(\SS_{k}^{2}) for every k1k\geq 1; the latter axiom states the existence of a functional SSk2\SS_{k}^{2} deciding Πk1\Pi_{k}^{1}-formulas in Kleene normal form. The system Π11-CA0ωRCA0ω+(SS12)\Pi_{1}^{1}\text{-{{CA}}}_{0}^{\omega}\equiv\textup{{RCA}}_{0}^{\omega}+(\SS_{1}^{2}) is a Π31\Pi_{3}^{1}-conservative extension of Π11-CA0\Pi_{1}^{1}\text{-{{CA}}}_{0} ([yamayamaharehare]), where SS12\SS_{1}^{2} is also called the Suslin functional. We also write ACA0ω\textup{{ACA}}_{0}^{\omega} for RCA0ω+(2)\textup{{RCA}}_{0}^{\omega}+(\exists^{2}) where the latter is as follows

(E:{0,1})(f)[(n)(f(n)=0)E(f)=0].(\exists E:{\mathbb{N}}^{{\mathbb{N}}}\rightarrow\{0,1\})(\forall f\in{\mathbb{N}}^{{\mathbb{N}}})\big{[}(\exists n\in{\mathbb{N}})(f(n)=0)\leftrightarrow E(f)=0\big{]}. (2\exists^{2})

The system ACA0ω\textup{{ACA}}_{0}^{\omega} is a conservative extension of ACA0\textup{{ACA}}_{0} by [hunterphd]*Theorem 2.5. Over RCA0ω\textup{{RCA}}_{0}^{\omega}, (2)(\exists^{2}) is equivalent to (μ2)(\mu^{2}), where the latter expresses the existence of Feferman’s μ\mu (see [kohlenbach2]*Prop. 3.9), defined as follows for all ff\in{\mathbb{N}}^{{\mathbb{N}}}:

μ(f):={nif n is the least natural number such that f(n)=0 0if f(n)>0 for all n.\mu(f):=\begin{cases}n&\textup{if $n$ is the least natural number such that $f(n)=0$ }\\ 0&\textup{if $f(n)>0$ for all $n\in{\mathbb{N}}$}\end{cases}.

The following schema is essential to our enterprise, as discussed in Section 1.3.

Principle 1.1 (QF-AC0,1\textup{{QF-AC}}^{0,1}).

For any Y:Y:{\mathbb{N}}^{{\mathbb{N}}}\rightarrow{\mathbb{N}}, if (n)(f)(Y(f,n)=0)(\forall n\in{\mathbb{N}})(\exists f\in{\mathbb{N}}^{{\mathbb{N}}})(Y(f,n)=0), then there exists a sequence (fn)n(f_{n})_{n\in{\mathbb{N}}} in {\mathbb{N}}^{{\mathbb{N}}} with (n)(Y(fn,n)=0)(\forall n\in{\mathbb{N}})(Y(f_{n},n)=0).

The local equivalence between sequential and ‘epsilon-delta’ continuity cannot be proved in ZF, but can be established in RCA0ω+QF-AC0,1\textup{{RCA}}_{0}^{\omega}+\textup{{QF-AC}}^{0,1} (see [kohlenbach2]). Thus, it should not be a surprise that the latter system is often used as a base theory too.

Secondly, we make use the following standard definitions concerning sets.

Definition 1.2 (Sets).
  • A subset AA\subset{\mathbb{R}} is given by its characteristic function FA:{0,1}F_{A}:{\mathbb{R}}\rightarrow\{0,1\}, i.e. we write xAx\in A for FA(x)=1F_{A}(x)=1, for any xx\in{\mathbb{R}}.

  • A set AA\subset{\mathbb{R}} is enumerable if there is a sequence of reals that includes all elements of AA.

  • A set AA\subset{\mathbb{R}} is countable if there is Y:Y:{\mathbb{R}}\rightarrow{\mathbb{N}} that is injective on AA, i.e.

    (x,yA)(Y(x)=0Y(y)x=y).(\forall x,y\in A)(Y(x)=_{0}Y(y)\rightarrow x=_{{\mathbb{R}}}y).
  • A set AA\subset{\mathbb{R}} is strongly countable if there is Y:Y:{\mathbb{R}}\rightarrow{\mathbb{N}} that is injective and surjective on AA; the latter means that (n)(xA)(Y(x)=n)(\forall n\in{\mathbb{N}})(\exists x\in A)(Y(x)=n).

  • A set AA\subset{\mathbb{R}} is finite in case there is NN\in{\mathbb{N}} such that for any finite sequence (x0,,xN)(x_{0},\dots,x_{N}), there is iNi\leq N with xiAx_{i}\not\in A. We sometimes write ‘|A|N|A|\leq N’.

Thirdly, we list the following second-order system needed below.

Principle 1.3 (finite-Σ11\Sigma_{1}^{1}-AC0, [gohzeg]).

The system RCA0\textup{{RCA}}_{0} plus for any arithmetical φ\varphi:

(n)( nonzero finitely many X)φ(n,X)((Xn)n)(n)φ(n,Xn),(\forall n\in{\mathbb{N}})(\exists\textup{ nonzero finitely many }X\subset{\mathbb{N}})\varphi(n,X)\rightarrow(\exists(X_{n})_{n\in{\mathbb{N}}})(\forall n\in{\mathbb{N}})\varphi(n,X_{n}),

where ‘( nonzero finitely many X)φ(n,X)(\exists\textup{ nonzero finitely many }X\subset{\mathbb{N}})\varphi(n,X)’ means that there is a non-empty sequence (X0,,Xk)(X_{0},\dots,X_{k}) such that for any XX\subset{\mathbb{N}}, φ(n,X)(ik)(Xi=X)\varphi(n,X)\leftrightarrow(\exists i\leq k)(X_{i}=X).

We let height-Σ11\Sigma_{1}^{1}-AC0 be finite-Σ11\Sigma_{1}^{1}-AC0 where we additionally assume gg\in{\mathbb{N}}^{{\mathbb{N}}} to be given such that for all nn, g(n)k+1g(n)\geq k+1 where k+1k+1 is the length of the sequence (X0,,Xk)(X_{0},\dots,X_{k}) in the formula ‘( nonzero finitely many X)φ(n,X)(\exists\textup{ nonzero finitely many }X\subset{\mathbb{N}})\varphi(n,X)’. We have the following straightforward connections:

Σ11-AC0finite-Σ11-AC0height-Σ11-AC0weak-Σ11-AC0,\textup{{$\Sigma_{1}^{1}$-AC${}_{0}$}}\rightarrow\textup{{finite-$\Sigma_{1}^{1}$-AC${}_{0}$}}\rightarrow\textup{{height-$\Sigma_{1}^{1}$-AC${}_{0}$}}\rightarrow\textup{{weak-$\Sigma_{1}^{1}$-AC${}_{0}$}},

i.e. height-Σ11\Sigma_{1}^{1}-AC0 is also a system of hyperarithmetical analysis by Section 1.3. In the grand scheme of things, gg is a height function, a notion that goes back to Borel ([opborrelen3, opborrelen5]) and is studied in RM in [samBIG, samBIG3].

1.3. On hyperarithmetical analysis

Going back to Kreisel ([kreide]), the notion of hyperarithmetical set (see e.g. [simpson2]*VIII.3) gives rise to the second-order definition of theory/theorem of hyperarithmetical analyis (THA for brevity, see e.g. [skore3]). In this section, we recall known results regarding THAs, including the exact (rather technical) definition, for completeness.

First of all, well-known THAs are Σ11\Sigma_{1}^{1}-CA0\textsf{CA}_{0} and weak-Σ11\Sigma_{1}^{1}-CA0\textsf{CA}_{0} (see [simpson2]*VII.6.1 and VIII.4.12), where the latter is the former with the antecedent restricted to unique existence. Any system between two THAs is also a THA, which is a convenient way of establishing that a given system is a THA.

Secondly, at the higher-order level, ACA0ω+QF-AC0,1\textup{{ACA}}_{0}^{\omega}+\textup{{QF-AC}}^{0,1} from Section 1.2 is a conservative extension of Σ11\Sigma_{1}^{1}-CA0\textsf{CA}_{0} by [hunterphd]*Cor. 2.7. This is established by extending any model \mathcal{M} of Σ11\Sigma_{1}^{1}-AC0 to a model 𝒩\mathcal{N} of ACA0ω+QF-AC0,1\textup{{ACA}}_{0}^{\omega}+\textup{{QF-AC}}^{0,1}, where the second-order part of 𝒩\mathcal{N} is isomorphic to \mathcal{M}. In this paper, we study (higher-order) systems that imply weak-Σ11\Sigma_{1}^{1}-CA0\textsf{CA}_{0} and are implied by ACA0ω+QF-AC0,1\textup{{ACA}}_{0}^{\omega}+\textup{{QF-AC}}^{0,1}. In light of the aforementioned conservation result, it is reasonable to refer to such intermediate third-order systems as existing in the range of hyperarithmetical analysis.

Thirdly, finding a natural THA, i.e. hailing from the mathematical mainstream, is surprisingly hard. Montalbán’s INDEC from [monta2], a special case of Jullien’s [juleke]*IV.3.3, is generally considered to be the first such statement. The latter theorem by Jullien can be found in [aardbei]*6.3.4.(3) and [roosje]*Lemma 10.3. The monographs [roosje, aardbei, juleke] are all ‘rather logical’ in nature and INDEC is the restriction of a higher-order statement to countable linear orders in the sense of RM ([simpson2]*V.1.1), i.e. such orders are given by sequences. In [dagsamXI]*Remark 2.8 and [samcount]*Remark 7 and §3.4, a number of third-order statements are identified, including the Bolzano-Weierstrass theorem and König’s infinity lemma, that are in the range of hyperarithmetical analysis. Shore and others have studied a considerable number of THAs from graph theory [skore1, skore2, gohzeg]. A related concept is that of almost theorem/theory of hyperarithmetical analysis (ATHA for brevity, [skore3]), which is weaker than ACA0\textup{{ACA}}_{0} but becomes a THA when combined with the latter.

Finally, we consider the official definition of THA from [monta2] based on ω\omega-models.

Definition 1.4.

A system TT of axioms of second-order arithmetic is a theory/theorem of hyperarithmetical analysis in case

  • TT holds in HYP(Y)\textup{{HYP}}(Y) for every YωY\subset\omega, where HYP(Y)\textup{{HYP}}(Y) is the ω\omega-model consisting of all sets hyperarithmetic in YY,

  • all ω\omega-models of TT are hyperarithmetically closed.

Here, an ω\omega-model is hyperarithmetically closed if it is closed under disjoint union and for every set X,YωX,Y\subset\omega, if XX is hyperarithmetically reducible to YY and YY is in the model, then XX is in the model too. In turn, this notion of reducibility means that nXn\in X can be expressed by a Δ11\Delta_{1}^{1}-formula with YY as a parameter; we refer to [monta2]*Theorem 1.14 for equivalent formulations.

2. Metric spaces

We introduce the well-known definition of metric space (M,d)(M,d) to be used in this paper (Section 2.1), where we always assume MM to be a subset of {\mathbb{R}}, up to coding of finite sequences. We show that basic properties of (Lipschitz) continuous functions on such metric spaces exist in the range of hyperarithmetical analysis (Section 2.2), even if we restrict to arithmetically defined objects (Theorem 2.8). We have previously studied metric spaces in [sammetric]; to our own surprise, some of these results have nice generalisations relevant to the study of hyperarithmetical analysis.

2.1. Basic definitions

We shall study metric spaces (M,d)(M,d) as in Definition 2.1. We assume that MM comes with its own equivalence relation ‘=M=_{M}’ and that the metric dd satisfies the axiom of extensionality relative to ‘=M=_{M}’ as follows:

(x,y,v,wM)([x=Myv=Mw]d(x,v)=d(y,w)).(\forall x,y,v,w\in M)\big{(}[x=_{M}y\wedge v=_{M}w]\rightarrow d(x,v)=_{{\mathbb{R}}}d(y,w)\big{)}.

Similarly to functions on the reals, ‘F:MF:M\rightarrow{\mathbb{R}}’ denotes a function from MM to the reals that satisfies the following instance of the axiom of function extensionality:

(x,yM)(x=MyF(x)=F(y)).(\forall x,y\in M)(x=_{M}y\rightarrow F(x)=_{{\mathbb{R}}}F(y)). (EM)

We recall that the study of metric space in second-order RM is at its core based on equivalence relations, as discussed explicitly in e.g. [simpson2]*I.4 or [damurm]*§10.1.

Definition 2.1.

A functional d:M2d:M^{2}\rightarrow{\mathbb{R}} is a metric on MM if it satisfies the following properties for x,y,zMx,y,z\in M:

  1. (a)

    d(x,y)=0x=Myd(x,y)=_{{\mathbb{R}}}0\leftrightarrow x=_{M}y,

  2. (b)

    0d(x,y)=d(y,x),0\leq_{{\mathbb{R}}}d(x,y)=_{{\mathbb{R}}}d(y,x),

  3. (c)

    d(x,y)d(x,z)+d(z,y)d(x,y)\leq_{{\mathbb{R}}}d(x,z)+d(z,y).

We shall only study metric spaces (M,d)(M,d) with MM\subset{\mathbb{N}}^{{\mathbb{N}}} or MM\subset{\mathbb{R}}. To be absolutely clear, quantifying over MM amounts to quantifying over {\mathbb{N}}^{{\mathbb{N}}} or {\mathbb{R}}, perhaps modulo coding of finite sequences, i.e. the previous definition can be made in third-order arithmetic for the intents and purposes of this paper. Since we shall study compact metric spaces, this restriction is minimal in light of [buko]*Theorem 3.13.

Sub-sets of MM are defined via characteristic functions, like for the reals in Definition 1.2, keeping in mind (EM). In particular, we use standard notation like BdM(x,r)B_{d}^{M}(x,r) to denote the open ball {yM:d(x,y)<r}\{y\in M:d(x,y)<_{{\mathbb{R}}}r\}.

Secondly, the following definitions are now standard, where we note that a different nomenclature is sometimes used in second-order RM. A sequence (wn)n(w_{n})_{n\in{\mathbb{N}}} in (M,d)(M,d) is Cauchy if (k)(N)(m,nN)(d(wn,wm)<12k)(\forall k\in{\mathbb{N}})(\exists N\in{\mathbb{N}})(\forall m,n\geq N)(d(w_{n},w_{m})<\frac{1}{2^{k}})

Definition 2.2 (Compactness and around).

For a metric space (M,d)(M,d), we say that

  • (M,d)(M,d) is weakly countably-compact if for any (an)n(a_{n})_{n\in{\mathbb{N}}} in MM and sequence of rationals (rn)n(r_{n})_{n\in{\mathbb{N}}} such that we have MnBdM(an,rn)M\subset\cup_{n\in{\mathbb{N}}}B^{M}_{d}(a_{n},r_{n}), there is mm\in{\mathbb{N}} such that MnmBdM(an,rn)M\subset\cup_{n\leq m}B^{M}_{d}(a_{n},r_{n}),

  • (M,d)(M,d) is countably-compact if for any sequence (On)n(O_{n})_{n\in{\mathbb{N}}} of open sets in MM such that MnOnM\subset\cup_{n\in{\mathbb{N}}}O_{n}, there is mm\in{\mathbb{N}} such that MnmOnM\subset\cup_{n\leq m}O_{n},

  • (M,d)(M,d) is compact in case for any Ψ:M+\Psi:M\rightarrow{\mathbb{R}}^{+}, there are x0,,xkMx_{0},\dots,x_{k}\in M such that ikBdM(xi,Ψ(xi))\cup_{i\leq k}B_{d}^{M}(x_{i},\Psi(x_{i})) covers MM,

  • (M,d)(M,d) is sequentially compact if any sequence has a convergent sub-sequence,

  • (M,d)(M,d) is limit point compact if any infinite set in MM has a limit point,

  • (M,d)(M,d) is complete in case every Cauchy sequence converges,

  • (M,d)(M,d) is totally bounded if for all kk\in{\mathbb{N}}, there are w0,,wmMw_{0},\dots,w_{m}\in M such that imBdM(wi,12k)\cup_{i\leq m}B_{d}^{M}(w_{i},\frac{1}{2^{k}}) covers MM.

  • a function f:Mf:M\rightarrow{\mathbb{R}} is topologically continuous if for any open VV\subset{\mathbb{R}}, the set f1(V)={xM:f(x)V}f^{-1}(V)=\{x\in M:f(x)\in V\} is also open.

  • a function f:Mf:M\rightarrow{\mathbb{R}} is closed if for any closed CMC\subset M, we have that f(C)f(C) is also closed. ([munkies, ooskelly, leelee, searinghot]).

Regarding the final item, the set f(C)f(C) does not necessarily exist in ACA0ω\textup{{ACA}}_{0}^{\omega}, but ‘f(C)f(C) is closed’ makes sense111In particular, ‘yf(C)y\in f(C)’ means ‘(xC)(f(x)=y)(\exists x\in C)(f(x)=y)’ and ‘f(C)f(C) is closed’ means ‘(yf(C))(N)(zB(z,12N))(zf(C))(\forall y\not\in f(C))(\exists N\in{\mathbb{N}})(\forall z\in B(z,\frac{1}{2^{N}}))(z\not\in f(C))’, as expected. as shorthand for the associated well-known definition. We could study other notions, e.g. the Lindelöf property or compactness based on nets, but have opted to stick to basic constructs already studied in second-order RM.

Finally, fragments of the induction axiom are sometimes used, even in an essential way, in second-order RM (see e.g. [neeman, skore3]). The equivalence between induction and bounded comprehension is also well-known in second-order RM ([simpson2]*X.4.4). We shall need a little bit of the induction axiom as follows.

Principle 2.3 (IND2\textup{{IND}}_{2}).

Let Y2Y^{2} satisfy (n)(f2)[Y(n,f)=0](\forall n\in{\mathbb{N}})(\exists f\in 2^{{\mathbb{N}}})[Y(n,f)=0]. Then (n)(w1)[|w|=n(i<n)(Y(i,w(i))=0)](\forall n\in{\mathbb{N}})(\exists w^{1^{*}})\big{[}|w|=n\wedge(\forall i<n)(Y(i,w(i))=0)\big{]}.

We let IND0\textup{{IND}}_{0} and IND1\textup{{IND}}_{1} be IND2\textup{{IND}}_{2} with ‘(f2)(\exists f\in 2^{{\mathbb{N}}})’ restricted to respectively ‘( at most one f2)(\exists\textup{ at most one }f\in 2^{{\mathbb{N}}})’ and ‘(!f2)(\exists!f\in 2^{{\mathbb{N}}})’. We have previously used INDi\textup{{IND}}_{i} for i=0,1,2i=0,1,2 in the RM of the Jordan decomposition theorem ([dagsamXI]). By the proof of [dagsamXI]*Theorem 2.16, Z2ω+IND2{\textsf{{Z}}}_{2}^{\omega}+\textup{{IND}}_{2} cannot prove the uncountability of the reals formulated as: the unit interval is not strongly countable.

2.2. Metric spaces and hyperarithmetical analysis

2.2.1. Introduction

In this section, we identify a number of the basic properties of metric spaces in the range of hyperarithmetical analysis, as listed on the next page. The Axiom of Choice for finite sets as in Principle 2.4 naturally comes to the fore. Clearly, the principle Finite Choice implies finite-Σ11\Sigma_{1}^{1}-AC0 over ACA0ω\textup{{ACA}}_{0}^{\omega}.

Principle 2.4 (Finite Choice).

Let (Xn)n(X_{n})_{n\in{\mathbb{N}}} be a sequence of non-empty finite sets in [0,1][0,1]. Then there is (xn)n(x_{n})_{n\in{\mathbb{N}}} such that xnXnx_{n}\in X_{n} for all nn\in{\mathbb{N}}.

In more detail, we will establish that the following theorems are intermediate between ACA0ω+QF-AC0,1\textup{{ACA}}_{0}^{\omega}+\textup{{QF-AC}}^{0,1} and ACA0ω+Finite Choice\textup{{ACA}}_{0}^{\omega}+\textup{{Finite Choice}}.

  • Basic properties of continuous functions on sequentially compact metric spaces (Section 2.2.2).

  • Basic properties of sequentially continuous functions on (countably) compact metric spaces (Section 2.2.3).

  • Restrictions of the previous results to arithmetically defined or Lipschitz continuous functions (Section 2.2.4).

  • Basic properties of connected metric spaces, including the generalisation of the intermediate value theorem (Section 2.2.5).

We sometimes obtain elegant equivalences, like for the intermediate value theorem (Theorem 2.13). We believe there is no ‘universal’ approach to the previous results: each section is based on a very particular kind of metric space.

2.2.2. Sequentially compact spaces

In this section, we establish that basic properties of sequentially compact spaces inhabit the range of hyperarithmetical analysis. The following theorem is our first result, to be refined below.

Theorem 2.5 (ACA0ω+IND2\textup{{ACA}}_{0}^{\omega}+\textup{{IND}}_{2}).

The principle Finite Choice follows from any of the items (a)-(j) where (M,d)(M,d) is any metric space with MM\subset{\mathbb{R}}; the principle QF-AC0,1\textup{{QF-AC}}^{0,1} implies items (a)-(i).

  1. (a)

    For sequentially compact (M,d)(M,d), any continuous f:Mf:M\rightarrow{\mathbb{R}} is bounded.

  2. (b)

    The previous item with ‘is bounded’ replaced by ‘is uniformly continuous’.

  3. (c)

    For sequentially compact (M,d)(M,d) and continuous f:Mf:M\rightarrow{\mathbb{R}} with infxMf(x)=y\inf_{x\in M}f(x)=y\in{\mathbb{R}} given, there is xMx\in M with f(x)=yf(x)=y.

  4. (d)

    (Dini). Let (M,d)(M,d) be sequentially compact and let fn:(M×)f_{n}:(M\times{\mathbb{N}})\rightarrow{\mathbb{R}} be a monotone sequence of continuous functions converging to continuous f:Mf:M\rightarrow{\mathbb{R}}. Then the convergence is uniform.

  5. (e)

    For a sequentially compact metric space (M,d)(M,d), equicontinuity implies uniform equicontinuity ([magnus]*Prop. 4.25).

  6. (f)

    For a sequentially compact metric space (M,d)(M,d) with M[0,1]M\subset[0,1] infinite, there is a discontinuous function f:Mf:M\rightarrow{\mathbb{R}}.

  7. (g)

    (Closed map lemma, [munkies, leelee, kura, mannetti]) For a sequentially compact metric space (M,d)(M,d) any continuous function f:Mf:M\rightarrow{\mathbb{R}} is closed.

  8. (h)

    For sequentially compact (M,d)(M,d) and disjoint closed C,DMC,D\subset M, d(C,D)>0d(C,D)>0.

  9. (i)

    (weak Cantor intersection theorem) For a sequentially compact metric space (M,d)(M,d) and a sequence of closed sets with MCnCn+1M\supseteq C_{n}\supseteq C_{n+1}\neq\emptyset, such that limndiam(Cn)=0\lim_{n\rightarrow\infty}\textsf{{diam}}(C_{n})=0, there is a unique wnCnw\in\cap_{n\in{\mathbb{N}}}C_{n}.

  10. (j)

    (Ascoli-Arzelà) For sequentially compact (M,d)(M,d), a uniformly bounded equi-continuous sequence of functions on MM has a uniformly convergent sub-sequence.

The theorem still goes through if we require a modulus of continuity in item (a) or if we replace ‘continuity’ by ‘topological continuity’ in items (a)-(f).

Proof.

We first derive Finite Choice from item (a) via a proof-by-contradiction. To this end, fix a sequence of non-empty finite sets of reals (Xn)n(X_{n})_{n\in{\mathbb{N}}}. Suppose there is no sequence (xn)n(x_{n})_{n\in{\mathbb{N}}} of reals such that xnXnx_{n}\in X_{n} for all nn\in{\mathbb{N}}. We now define

M0:={w1:(i<|w|)(w(i)Xi)},M_{0}:=\{w^{1^{*}}:(\forall i<|w|)(w(i)\in X_{i})\}, (2.1)

where w1w^{1^{*}} is a finite sequence of reals of length |w||w|, readily coded using (2)(\exists^{2}). We define the equivalence relation ‘=M0=_{M_{0}}’ as follows: the relation w=M0vw=_{M_{0}}v holds if |w|=0|v||w|=_{0}|v|, where w,vM0w,v\in M_{0}. The metric d0:M02d_{0}:M_{0}^{2}\rightarrow{\mathbb{R}} is defined as d0(w,v)=|12|v|12|w||d_{0}(w,v)=|\frac{1}{2^{|v|}}-\frac{1}{2^{|w|}}| for any w,vM0w,v\in M_{0}. We then have d0(v,w)=0|v|=0|w|v=M0wd_{0}(v,w)=_{{\mathbb{R}}}0\leftrightarrow|v|=_{0}|w|\leftrightarrow v=_{M_{0}}w as required. We also have 0d0(v,w)=d0(w,v)0\leq d_{0}(v,w)=_{{\mathbb{R}}}d_{0}(w,v) for any v,wM0v,w\in M_{0}, while for any zM0z\in M_{0} we observe:

d0(v,w)=|12|v|12|w||=|12|v|12|z|+12|z|12|w|||12|v|12|z||+|12|z|12|w||=d0(v,z)+d0(z,w)\textstyle d_{0}(v,w)=|\frac{1}{2^{|v|}}-\frac{1}{2^{|w|}}|=|\frac{1}{2^{|v|}}-\frac{1}{2^{|z|}}+\frac{1}{2^{|z|}}-\frac{1}{2^{|w|}}|\leq|\frac{1}{2^{|v|}}-\frac{1}{2^{|z|}}|+|\frac{1}{2^{|z|}}-\frac{1}{2^{|w|}}|=d_{0}(v,z)+d_{0}(z,w)

by the triangle equality of the absolute value on the reals. Hence, (M0,d0)(M_{0},d_{0}) is a metric space as in Definition 2.1.

To show that (M0,d0)(M_{0},d_{0}) is sequentially compact, let (wn)n(w_{n})_{n\in{\mathbb{N}}} be a sequence in M0M_{0} and consider the following case distinction. In case (n)(|wn|<m0)(\forall n\in{\mathbb{N}})(|w_{n}|<m_{0}) for some fixed m0m_{0}\in{\mathbb{N}}, then (wn)n(w_{n})_{n\in{\mathbb{N}}} contains at most (m0+1)!(m_{0}+1)! different elements. The pigeon hole principle now implies that at least one wn0w_{n_{0}} occurs infinitely often in (wn)n(w_{n})_{n\in{\mathbb{N}}}, i.e. (wn0)n(w_{n_{0}})_{n\in{\mathbb{N}}} is a convergent sub-sequence. In case (m)(n)(|wn|m)(\forall m\in{\mathbb{N}})(\exists n\in{\mathbb{N}})(|w_{n}|\geq m), the sequence (wn)n(w_{n})_{n\in{\mathbb{N}}} yields a sequence (xn)n(x_{n})_{n\in{\mathbb{N}}} such that xnXnx_{n}\in X_{n} for all nn\in{\mathbb{N}}, which is impossible by assumption. Hence, (M0,d0)(M_{0},d_{0}) is a sequentially compact metric space.

Next, define f:M0f:M_{0}\rightarrow{\mathbb{R}} as f(w):=|w|f(w):=|w|, which is clearly not bounded on M0M_{0}, which one shows using IND2\textup{{IND}}_{2}. To show that ff is continuous at w0M0w_{0}\in M_{0}, consider the formula |12|w0|12|v||=d0(v,w0)<12N|\frac{1}{2^{|w_{0}|}}-\frac{1}{2^{|v|}}|=d_{0}(v,w_{0})<\frac{1}{2^{N}}; the latter is false for N|w0|+2N\geq|w_{0}|+2 and any vM0w0v\neq_{M_{0}}w_{0}. Hence, the following formula is vacuously true:

(k)(N)(vBd0M0(w0,12N))(|f(w0)f(v)|<12k).\textstyle(\forall k\in{\mathbb{N}})(\exists N\in{\mathbb{N}})(\forall v\in B_{d_{0}}^{M_{0}}(w_{0},\frac{1}{2^{N}}))(|f(w_{0})-f(v)|<_{{\mathbb{R}}}\frac{1}{2^{k}}). (2.2)

i.e. ff is continuous at w0M0w_{0}\in M_{0}, with a modulus of continuity given by h(w,k):=12|w|+k+2h(w,k):=\frac{1}{2^{|w|+k+2}}. To see that ff is also topologically continuous, fix an open set VV\subset{\mathbb{R}} and fix w0f1(V)w_{0}\in f^{-1}(V). Then for N0:=|w0|+2N_{0}:=|w_{0}|+2, one verifies that BdM0(w0,12N0)f1(V)B_{d}^{M_{0}}(w_{0},\frac{1}{2^{N_{0}}})\subset f^{-1}(V), i.e. f1(V)f^{-1}(V) is open. Thus, f:M0f:M_{0}\rightarrow{\mathbb{R}} is a continuous but unbounded function on a sequentially compact metric space (M0,d0)(M_{0},d_{0}), contradicting item (a). Item (b) also implies Finite Choice as ff is not uniformly continuous. For item (c), g:M0g:M_{0}\rightarrow{\mathbb{R}} defined as g(w):=12|w|g(w):=\frac{1}{2^{|w|}} is continuous in the same way as for ff. However, using IND2\textup{{IND}}_{2}, the infimum of gg on M0M_{0} is 0, but there is no wM0w\in M_{0} with g(w)=0g(w)=_{{\mathbb{R}}}0, by definition. Hence, item (c) also implies Finite Choice.

Now assume item (d) and suppose Finite Choice is again false; letting (Xn)n(X_{n})_{n\in{\mathbb{N}}} and (M0,d0)(M_{0},d_{0}) be as in the previous paragraph, we define fn:(×M0)f_{n}:({\mathbb{N}}\times M_{0})\rightarrow{\mathbb{R}} as:

fn(w):={|w| if |w|n0 otherwise.f_{n}(w):=\begin{cases}|w|&\textup{ if $|w|\leq n$}\\ 0&\textup{ otherwise}\end{cases}. (2.3)

Clearly, limnfn(w)=f(w)\lim_{n\rightarrow\infty}f_{n}(w)=f(w) and fn(w)fn+1(w)f_{n}(w)\leq f_{n+1}(w) for wM0w\in M_{0}; fnf_{n} is continuous in the same way as for ff. Item (d) implies that the convergence is uniform, i.e.

(k)(N)(wM0)(nN)(|fn(w)f(w)|<12k),\textstyle(\forall k\in{\mathbb{N}})(\exists N\in{\mathbb{N}})(\forall w\in M_{0})(\forall n\geq N)(|f_{n}(w)-f(w)|<\frac{1}{2^{k}}), (2.4)

which yields a contradiction by letting N1N_{1}\in{\mathbb{N}} be as in (2.4) for k=1k=1 and choosing w1Mw_{1}\in M of length N1+1N_{1}+1 using IND2\textup{{IND}}_{2}. One derives Finite Choice from item (e) in the same way.

Next, regarding item (g), suppose Finite Choice is false and consider again (M0,d0)(M_{0},d_{0}). Define the continuous function f:M0f:M_{0}\rightarrow{\mathbb{R}} by f(w)=q|w|f(w)=q_{|w|} where (qn)n(q_{n})_{n\in{\mathbb{N}}} is an enumeration of the rationals without repetitions. Using IND2\textup{{IND}}_{2}, we have f(M0)=f(M_{0})={\mathbb{Q}} and the latter is not closed while M0M_{0} is, contradicting item (g), and Finite Choice must hold. To obtain the latter from item (f), note that (M0,d0)(M_{0},d_{0}) is infinite (using IND2\textup{{IND}}_{2}) while all functions f:M0f:M_{0}\rightarrow{\mathbb{R}} are continuous as (2.2) is vacuously true. Regarding item (j), assuming again that Finite Choice is false, the sequence (fn)n(f_{n})_{n\in{\mathbb{N}}} as in (2.3) is equicontinuous:

(k,wM0)(N)(vBd0M0(w,12N))(n)(|fn(w)fn(v)|<12k).\textstyle(\forall k\in{\mathbb{N}},w\in M_{0})(\exists N\in{\mathbb{N}})(\forall v\in B_{d_{0}}^{M_{0}}(w,\frac{1}{2^{N}}))(\forall n\in{\mathbb{N}})(|f_{n}(w)-f_{n}(v)|<_{{\mathbb{R}}}\frac{1}{2^{k}}).

which (vacuously) holds in the same way as for (2.2). However, as for item (d), uniform convergence (of a sub-sequence) is false, i.e. item (j) also implies Finite Choice. For item (i), suppose Finite Choice is false, define Cn:={wM0:|w|>n+1}C_{n}:=\{w\in M_{0}:|w|>n+1\}, and verify that this closed and non-empty set has diameter at most 12n\frac{1}{2^{n}}, using IND2\textup{{IND}}_{2}. Since nCn=\cap_{n\in{\mathbb{N}}}C_{n}=\emptyset, we obtain Finite Choice from item (i). For item (h), suppose Finite Choice is false, and define C={wM0:|w| is odd}C=\{w\in M_{0}:|w|\textup{ is odd}\} and D={wM0:|w| is even}D=\{w\in M_{0}:|w|\textup{ is even}\}. One readily verifies that CD=C\cap D=\emptyset, C,DC,D are closed, and d(C,D)=0d(C,D)=0.

To establish the items in the theorem in ACA0ω+QF-AC0,1\textup{{ACA}}_{0}^{\omega}+\textup{{QF-AC}}^{0,1}, the usual proof-by-contradiction goes through. A proof sketch of item (a) as follows: let (M,d)(M,d) be as in the latter and suppose f:Mf:M\rightarrow{\mathbb{R}} is continuous and unbounded, i.e. (n)(xM)(|f(x)|>n)(\forall n\in{\mathbb{N}})(\exists x\in M)(|f(x)|>n). Since MM\subset{\mathbb{R}} and real numbers are represented by elements of Baire space, we may apply QF-AC0,1\textup{{QF-AC}}^{0,1} to obtain (xn)n(x_{n})_{n\in{\mathbb{N}}} in MM such that |f(xn)|>n|f(x_{n})|>n for all nn\in{\mathbb{N}}. Since MM is sequentially compact, (xn)n(x_{n})_{n\in{\mathbb{N}}} has a convergent sub-sequence, say with limit yMy\in M. Clearly, ff is not continuous at yMy\in M, a contradiction. To obtain (f), apply QF-AC0,1\textup{{QF-AC}}^{0,1} to the statement that M[0,1]M\subset[0,1] is infinite, yielding a sequence (wn)n(w_{n})_{n\in{\mathbb{N}}} in MM. Now define f:Mf:M\rightarrow{\mathbb{R}} as f(xn)=nf(x_{n})=n and f(y)=0f(y)=0 for yxmy\neq x_{m} for all mm\in{\mathbb{N}}. Since ff is unbounded on MM, it is discontinuous by item (a). Most other items are established using QF-AC0,1\textup{{QF-AC}}^{0,1} in the same way.

We also sketch how QF-AC0,1\textup{{QF-AC}}^{0,1} implies item (g). To this end, let f,Mf,M be as in the closed map lemma and suppose f(C)f(C) is not closed for closed CMC\subset M. Hence, there is y0f(C)y_{0}\not\in f(C) such that (k)(yB(y0,12k))(yf(C))(\forall k\in{\mathbb{N}})(\exists y\in B(y_{0},\frac{1}{2^{k}}))(y\in f(C)). By definition, the latter formula means (k)(xC)(|f(x)y0|<12k)(\forall k\in{\mathbb{N}})(\exists x\in C)(|f(x)-y_{0}|<\frac{1}{2^{k}}). Apply QF-AC0,1\textup{{QF-AC}}^{0,1} to obtain a sequence (xn)n(x_{n})_{n\in{\mathbb{N}}} in CC with (k)(|f(xk)y0|<12k)(\forall k\in{\mathbb{N}})(|f(x_{k})-y_{0}|<\frac{1}{2^{k}}). By sequential compactness, there is a convergent sub-sequence (zn)n(z_{n})_{n\in{\mathbb{N}}}, say with limit zz. Since CC is closed, we have zCz\in C and since ff is continuous (and hence sequentially continuous) also f(z)=y0f(z)=y_{0}. This contradicts y0f(C)y_{0}\not\in f(C) and the closed map lemma therefore follows from QF-AC0,1\textup{{QF-AC}}^{0,1}. ∎

The final part of the proof also goes through if ff is only usco (see Def. (4.1)). As to other generalisations of Theorem 2.5, the latter still goes through for ‘continuity’ replaced by ‘absolute differentiability’ from [incell] formulated222The correct formulation based on [incell] is that ‘f:Mf:M\rightarrow{\mathbb{R}} is (absolutely) differentiable on the metric space (M,d)(M,d)’ in case we have (k,pM)(N)(x,yM)(0<d(x,p),d(y,p)<12N||f(x)f(p)|d(x,p)|f(y)f(p)|d(y,p)|<12k),\textstyle(\forall k\in{\mathbb{N}},p\in M)(\exists N\in{\mathbb{N}})(\forall x,y\in M)(0<d(x,p),d(y,p)<\frac{1}{2^{N}}\rightarrow\big{|}\frac{|f(x)-f(p)|}{d(x,p)}-\frac{|f(y)-f(p)|}{d(y,p)}\big{|}<\frac{1}{2^{k}}), which is the ‘epsilon-delta’ definition formulated to avoid the existence of the derivative. appropriately.

Finally, we observe that (M0,d0)(M_{0},d_{0}) from (2.1) is not (countably) compact, i.e. we need a slightly different approach for the latter, to be found in the next section.

2.2.3. Compact spaces

In this section, we establish that basic properties of (countably) compact spaces inhabit the range of hyperarithmetical analysis.

First of all, the following theorem is a version of Theorem 2.5 for (countably) compact spaces and sequential continuity. We seem to (only) need sequential compactness to guarantee that everything remains provable in ACA0ω+QF-AC0,1\textup{{ACA}}_{0}^{\omega}+\textup{{QF-AC}}^{0,1}.

Theorem 2.6 (ACA0ω+IND2\textup{{ACA}}_{0}^{\omega}+\textup{{IND}}_{2}).

The principle Finite Choice follows from any of the items (a)-(d) where (M,d)(M,d) is any metric space with MM\subset{\mathbb{R}}; the principle QF-AC0,1\textup{{QF-AC}}^{0,1} implies all these items.

  1. (a)

    For (weakly) countably-compact and sequentially compact (M,d)(M,d), any sequentially continuous f:Mf:M\rightarrow{\mathbb{R}} is bounded.

  2. (b)

    The previous item with ‘is bounded’ replaced by ‘is (uniformly) continuous’.

  3. (c)

    For a (weakly) countably-compact (M,d)(M,d) that is infinite, there is f:Mf:M\rightarrow{\mathbb{R}} that is not sequentially continuous.

  4. (d)

    The first item with ‘(weakly) countably-compact’ replaced by ‘compact’ or ‘complete and totally bounded’.

Proof.

We first derive Finite Choice from item (a) via a proof-by-contradiction. To this end, fix a sequence of non-empty finite sets of reals (Xn)n(X_{n})_{n\in{\mathbb{N}}}. Suppose there is no sequence (xn)n(x_{n})_{n\in{\mathbb{N}}} of reals such that xnXnx_{n}\in X_{n} for all nn\in{\mathbb{N}} and recall M0M_{0} from (2.1). Now define M1=M0{0M1}M_{1}=M_{0}\cup\{0_{M_{1}}\} where 0M10_{M_{1}} is a new symbol such that wM10w\neq_{M_{1}}0 for wM0w\in M_{0} and ‘=M1=_{M_{1}}’ is ‘=M0=_{M_{0}}’ otherwise. Define d1:M12d_{1}:M_{1}^{2}\rightarrow{\mathbb{R}} as d0d_{0} on M0M_{0}, as d1(w,0M1):=d(0M1,w)=12|w|d_{1}(w,0_{M_{1}}):=d(0_{M_{1}},w)=\frac{1}{2^{|w|}} for wM0w\in M_{0}, and d1(0M1,0M1)=0d_{1}(0_{M_{1}},0_{M_{1}})=0. Then (M1,d1)(M_{1},d_{1}) is a metric, which is shown in the same way as for (M0,d0)(M_{0},d_{0}).

To show that (M1,d1)(M_{1},d_{1}) is countably-compact, let (On)n(O_{n})_{n\in{\mathbb{N}}} be an open cover of M1M_{1} and suppose n1n_{1}\in{\mathbb{N}} is sucht hat 0M1On10_{M_{1}}\in O_{n_{1}}. By definition, there is N1N_{1}\in{\mathbb{N}} such that Bd1M1(0M1,12N1)On0B_{d_{1}}^{M_{1}}(0_{M_{1}},\frac{1}{2^{N_{1}}})\subset O_{n_{0}}, i.e. d(0M1,w)=12|w|<12N1d(0_{M_{1}},w)=\frac{1}{2^{|w|}}<\frac{1}{2^{N_{1}}} implies wOn1w\in O_{n_{1}} for wM0w\in M_{0}. Now use IND2\textup{{IND}}_{2} to enumerate the finitely many vM0v\in M_{0} such that |v|N1|v|\leq N_{1}. This finite sequence is covered by some nn2On\cup_{n\leq n_{2}}O_{n}, i.e. we have obtained a finite sub-covering of M1M_{1}, namely nmax(n1,n2)On\cup_{n\leq\max(n_{1},n_{2})}O_{n}. Moreover, (M1,d1)(M_{1},d_{1}) is sequentially compact, which can be proved via the same case distinction as for (M0,d0)(M_{0},d_{0}) in the proof of Theorem 2.5.

Next, define g:M1g:M_{1}\rightarrow{\mathbb{R}} as g(w):=|w|g(w):=|w| for wM0w\in M_{0} and g(0M1)=0g(0_{M_{1}})=_{{\mathbb{R}}}0, which is clearly not bounded on M1M_{1}; this follows again via IND2\textup{{IND}}_{2}. Then gg is continuous at w0M0w_{0}\in M_{0} in the same way as ff from the proof of Theorem 2.5, namely since (2.2) is vacuously true. To show that gg is sequentially continuous at 0M10_{M_{1}}, let (wn)n(w_{n})_{n\in{\mathbb{N}}} be a sequence converging to 0M10_{M_{1}}. In case this sequence is eventually constant 0M10_{M_{1}}, clearly g(0M)=limng(wn)g(0_{M})=\lim_{n\rightarrow\infty}g(w_{n}) as required. In case (wn)n(w_{n})_{n\in{\mathbb{N}}} is not eventually constant 0M10_{M_{1}}, the convergence to 0M10_{M_{1}} in the d1d_{1}-metric implies that for any nn\in{\mathbb{N}}, there is mnm\geq n with |wm|>n|w_{m}|>n. Thus, (wn)n(w_{n})_{n\in{\mathbb{N}}} yields a sequence (xn)n(x_{n})_{n\in{\mathbb{N}}} such that xnXnx_{n}\in X_{n} for all nn\in{\mathbb{N}}, which contradicts our assumptions, i.e. this case cannot occur. As a result, g:M1g:M_{1}\rightarrow{\mathbb{R}} is sequentially continuous. Since, it is also unbounded (thanks to IND2\textup{{IND}}_{2}), we obtain a contradiction with item (a). Thus, (a) implies Finite Choice, and the same for item (b). To obtain Finite Choice from item (c), note that (M1,d1)(M_{1},d_{1}) is infinite (using IND2\textup{{IND}}_{2}) while all functions f:M1f:M_{1}\rightarrow{\mathbb{R}} are sequentially continuous by the previous.

To show that (M1,d1)(M_{1},d_{1}) satisfies the properties in item (d), note that for Ψ:M1+\Psi:M_{1}\rightarrow{\mathbb{R}}^{+}, the ball Bd1M1(0M1,Ψ(0M1))B_{d_{1}}^{M_{1}}(0_{M_{1}},\Psi(0_{M_{1}})) covers all but finitely many points of M1M_{1} (in the same way as On0O_{n_{0}} from the second paragraph of the proof). Hence, (M1,d1)(M_{1},d_{1}) is compact, and totally boundedness follows in exactly the same way. For completeness, let (wn)n(w_{n})_{n\in{\mathbb{N}}} be a Cauchy sequence in M1M_{1}, i.e. we have

(k)(N)(n,mN)(d1(wn,wm)<12k).\textstyle(\forall k\in{\mathbb{N}})(\exists N\in{\mathbb{N}})(\forall n,m\geq N)(d_{1}(w_{n},w_{m})<\frac{1}{2^{k}}).

As above, (wn)n(w_{n})_{n\in{\mathbb{N}}} is either eventually constant or provides a sequence (xn)n(x_{n})_{n\in{\mathbb{N}}} such that xnXnx_{n}\in X_{n} for all nn\in{\mathbb{N}}. The latter case is impossible by assumption and the former case is trivial.

To establish the items in the theorem in ACA0ω+QF-AC0,1\textup{{ACA}}_{0}^{\omega}+\textup{{QF-AC}}^{0,1}, the usual proof-by-contradiction goes through as in the proof of Theorem 2.5. ∎

We believe that we cannot use epsilon-delta or topological continuity in the previous theorem. Nonetheless, we have the following corollary that makes use of the sequential333A function f:Mf:M\rightarrow{\mathbb{R}} is called sequentially uniformly continuous if for any sequences (wn)n(w_{n})_{n\in{\mathbb{N}}}, (vn)m(v_{n})_{m\in{\mathbb{N}}} in MM such that limnd(wn,vn)=0\lim_{n\rightarrow\infty}d(w_{n},v_{n})=0, we have limn|f(wn)f(vn)|=0\lim_{n\rightarrow\infty}|f(w_{n})-f(v_{n})|=0. definition of uniform continuity.

Corollary 2.7.

Over ACA0ω+IND2\textup{{ACA}}_{0}^{\omega}+\textup{{IND}}_{2}, items (a)-(f) in Theorem 2.5 and items (a)-(d) in Theorem 2.6 are intermediate between QF-AC0,1\textup{{QF-AC}}^{0,1} and Finite Choice if we replace ‘continuity’ by ‘sequential uniform continuity’.

Proof.

The usual proof-by-contradiction using QF-AC0,1\textup{{QF-AC}}^{0,1} (and (2)(\exists^{2})) shows that sequential uniform continuity implies uniform continuity. For the remaining implications, consider g:M1g:M_{1}\rightarrow{\mathbb{R}} from the the proof of Theorem 2.6. This function is sequentially continuous at 0M10_{M_{1}} since any sequence converging to 0M10_{M_{1}} must be eventually constant 0M10_{M_{1}}. Similarly, for sequences (wn)n(w_{n})_{n\in{\mathbb{N}}}, (vn)n(v_{n})_{n\in{\mathbb{N}}} in M1M_{1} limnd1(wn,vn)=0\lim_{n\rightarrow\infty}d_{1}(w_{n},v_{n})=0 implies that the sequences are eventually equal. Hence, gg is also sequentially uniformly continuous. A similar proof goes through for (M0,d0)(M_{0},d_{0}) and ff from Theorem 2.5. ∎

We have identified a number of basic properties of continuous functions on compact metric spaces that exist in the range of hyperarithmetical analysis. A number of restrictions and variations are possible, which is the topic of the next section.

2.2.4. Restrictions

We show that some the above principles still inhabit the range of hyperarithmetical analysis if we restrict to arithmetically defined objects or Lipschitz continuity.

First of all, the following theorem establishes that Theorem 2.5 holds if we restrict to arithmetically defined objects.

Theorem 2.8.

Item (a) from Theorem 2.5 still implies weak-Σ11\Sigma_{1}^{1}-AC0 over ACA0ω+IND1\textup{{ACA}}_{0}^{\omega}+\textup{{IND}}_{1} if all objects in the former item have an arithmetical definition.

Proof.

In a nutshell, we can modify the above proofs to obtain (only) weak-Σ11\Sigma_{1}^{1}-AC0 while all relevant objects can be defined using (2)(\exists^{2}). To this end, let φ\varphi be arithmetical and such that (n)(!X)φ(n,X)(\forall n\in{\mathbb{N}})(\exists!X\subset{\mathbb{N}})\varphi(n,X), but there is no sequence (Xn)n(X_{n})_{n\in{\mathbb{N}}} with (n)φ(n,Xn)(\forall n\in{\mathbb{N}})\varphi(n,X_{n}). Use 2\exists^{2} to define η:[0,1](2×2)\eta:[0,1]\rightarrow(2^{{\mathbb{N}}}\times 2^{{\mathbb{N}}}) such that η(x)=(f,g)\eta(x)=(f,g) outputs the binary expansions of xx, taking f=gf=g if there is only one. Define the following set using 2\exists^{2}:

A:={x[0,1]:(m)(φ(m,η(x)(1))φ(m,η(x)(2)))}A:=\{x\in[0,1]:(\exists m\in{\mathbb{N}})(\varphi(m,\eta(x)(1))\vee\varphi(m,\eta(x)(2)))\}

and Y(x):=(μm)[φ(m,η(x)(1)φ(m,η(x)(2)))]Y(x):=(\mu m)[\varphi(m,\eta(x)(1)\vee\varphi(m,\eta(x)(2)))]. Then YY is injective and surjective on AA. In particular Am:={x[0,1]:φ(m,η(x)(1))φ(m,η(x)(2))}A_{m}:=\{x\in[0,1]:\varphi(m,\eta(x)(1))\vee\varphi(m,\eta(x)(2))\} contains exactly one element by definition. Using Xn=AnX_{n}=A_{n}, the metric space (M0,d0)(M_{0},d_{0}) as in (2.1) in the proof of Theorem 2.5 now has an arithmetical definition. The same holds for the function F:M0F:M_{0}\rightarrow{\mathbb{R}} where F(w):=|w|F(w):=|w|. The rest of the proof of item (a) of Theorem 2.5 now goes through, using IND1\textup{{IND}}_{1} instead of IND2\textup{{IND}}_{2} where relevant, yielding in particular a contradiction. Hence, there must be a sequence (Xn)n(X_{n})_{n\in{\mathbb{N}}} with (n)φ(n,Xn)(\forall n\in{\mathbb{N}})\varphi(n,X_{n}), i.e. weak-Σ11\Sigma_{1}^{1}-AC0 follows as required. ∎

Secondly, we show that we may replace ‘continuity’ by ‘Lipschitz continuity’ in some of the above principles.

Definition 2.9.

A function f:Mf:M\rightarrow{\mathbb{R}} is α\alpha-Hölder-continuous in case there exist M,α>0M,\alpha>0 such that for any v,wMv,w\in M:

|f(v)f(w)|Md(v,w)α,|f(v)-f(w)|\leq Md(v,w)^{\alpha},

A function is Lipschitz (continuous) if is is 11-Hölder-continuous.

Theorem 2.10 (ACA0ω+IND2\textup{{ACA}}_{0}^{\omega}+\textup{{IND}}_{2}).

The principle Finite Choice follows from any of the items (a)-(e) where (M,d)(M,d) is any metric space with MM\subset{\mathbb{R}}; the principle QF-AC0,1\textup{{QF-AC}}^{0,1} implies items (a)-(e).

  1. (a)

    For a metric space (M,d)(M,d), any sequentially compact CMC\subseteq M is bounded, i.e. there are wM,mw\in M,m\in{\mathbb{N}} with (vC)(d(v,w)m)(\forall v\in C)(d(v,w)\leq m) (see [bartle2]*p. 333).

  2. (b)

    For sequentially compact (M,d)(M,d), any uniformly continuous f:Mf:M\rightarrow{\mathbb{R}} is bounded.

  3. (c)

    The previous item with ‘uniformly’ replaced by ‘α\alpha-Hölder’ or ‘Lipschitz’.

  4. (d)

    For sequentially compact (M,d)(M,d) that is infinite, there exists f:Mf:M\rightarrow{\mathbb{R}} that is bounded but not Lipschitz continuous.

  5. (e)

    For sequentially compact and bounded (M,d)(M,d) and Lipschitz f:Mf:M\rightarrow{\mathbb{R}} with infxMf(x)=y\inf_{x\in M}f(x)=y\in{\mathbb{R}} given, there is xMx\in M with f(x)=yf(x)=y.

Proof.

First of all, to derive item (a) from QF-AC0,1\textup{{QF-AC}}^{0,1}, fix a metric space (M,d)(M,d) and let CMC\subseteq M be sequentially compact. Suppose CC is not bounded, i.e. for some fixed w0Mw_{0}\in M, we have (m)(vC)(d(w0,v)>m)(\forall m\in{\mathbb{N}})(\exists v\in C)(d(w_{0},v)>m). Apply QF-AC0,1\textup{{QF-AC}}^{0,1} to obtain a sequence (vn)n(v_{n})_{n\in{\mathbb{N}}} such that d(w0,vn)>nd(w_{0},v_{n})>n for all nn\in{\mathbb{N}}. Clearly, this sequence cannot have a convergent sub-sequence, a contradiction, and CC must be bounded. To derive item (d) from QF-AC0,1\textup{{QF-AC}}^{0,1}, apply QF-AC0,1\textup{{QF-AC}}^{0,1} to the statement that MM is infinite. The resulting sequence (wn)n(w_{n})_{n\in{\mathbb{N}}} has a convergent sub-sequence, say (vn)n(v_{n})_{n\in{\mathbb{N}}} with limit vv. Define f(w)=1f(w)=1 (resp. f(w)=1f(w)=-1) if vn=wv_{n}=w and nn is even (resp. odd), and f(w)=0f(w)=0 otherwise. Clearly, f:Mf:M\rightarrow{\mathbb{R}} is bounded but not (Lipschitz) continuous. By, Theorem 2.5, the other items follow from QF-AC0,1\textup{{QF-AC}}^{0,1}.

Secondly, to derive Finite Choice from item (a), suppose (Xn)n(X_{n})_{n\in{\mathbb{N}}} is a sequence of finite sets such that there is no sequence (xn)n(x_{n})_{n\in{\mathbb{N}}} with xnXnx_{n}\in X_{n} for all nn\in{\mathbb{N}}. Recall the set M0M_{0} from (2.1) and define d2:M02d_{2}:M_{0}^{2}\rightarrow{\mathbb{R}} as d2(v,w)=||v||w||d_{2}(v,w)=\big{|}|v|-|w|\big{|} for v,wM0v,w\in M_{0}. That d2d_{2} is a metric is readily verified: the first and third item of Definition 2.1 hold by definition and the triangle equality of the absolute value; the second item in this definition holds since d2(v,w)=0|u|=|w|u=M0wd_{2}(v,w)=0\leftrightarrow|u|=|w|\leftrightarrow u=_{M_{0}}w. Now, the set C={wM0:|w| is even}C=\{w\in M_{0}:|w|\textup{ is even}\} is sequentially compact, as every sequence in CC either has at most finitely many different members, or yields a sequence (xn)n(x_{n})_{n\in{\mathbb{N}}} such that xnXnx_{n}\in X_{n} for all nn\in{\mathbb{N}}. We have excluded the latter by assumption, while the former trivially yields a convergent sub-sequence. Using IND2\textup{{IND}}_{2}, CC is however not bounded in (M0,d2)(M_{0},d_{2}), a contradiction, and item (a) implies Finite Choice.

Thirdly, to derive Finite Choice from the remaining items, let (M0,d2)(M_{0},d_{2}) be as above and note that the latter is sequentially compact as in the previous paragraph. Now define f:M0f:M_{0}\rightarrow{\mathbb{R}} as f(u):=|u|2f(u):=\frac{|u|}{2} and observe that |f(u)f(v)|=12||u||v||12d2(u,v)|f(u)-f(v)|=\frac{1}{2}\big{|}|u|-|v|\big{|}\leq\frac{1}{2}d_{2}(u,v), i.e. ff is Lifschitz (and uniformly) continuous. However, IND2\textup{{IND}}_{2} shows that ff is not bounded, a contradiction, and items (b)-(c) imply Finite Choice. Similarly, item (d) implies Finite Choice as (M0,d2)(M_{0},d_{2}) is such that every bounded function f:M0f:M_{0}\rightarrow{\mathbb{R}} is automatically Lipschitz. Indeed, if |f(w)|M0|f(w)|\leq M_{0} for all wM0w\in M_{0}, then the Lipschitz constant for ff can be taken to be 2M02M_{0}.

Finally, to derive Finite Choice from item (e), suppose the former is false and consider again (M0,d0)(M_{0},d_{0}), which is trivially bounded due to the definition of d0d_{0}. Now define g:M0g:M_{0}\rightarrow{\mathbb{R}} as g(u):=12|u|+1g(u):=\frac{1}{2^{|u|+1}}. This function is Lipschitz on (M0,d0)(M_{0},d_{0}) as

|g(u)g(v)|=|12|u|+112|v|+1|=12|12|u|12|v||=12d(u,v).\textstyle|g(u)-g(v)|=|\frac{1}{2^{|u|+1}}-\frac{1}{2^{|v|+1}}|=\frac{1}{2}|\frac{1}{2^{|u|}}-\frac{1}{2^{|v|}}|=\frac{1}{2}d(u,v).

However, gg has infimum equal to zero (using IND2\textup{{IND}}_{2}) but is strictly positive on M0M_{0}, contradicting item (e), which establishes the theorem. ∎

In conclusion, many implications between the notions in Definition 2.2 exist in the range of hyperarithmetical analysis, as well as the associated Lebesgue number lemma for countable coverings of open sets. These are left to the reader.

2.2.5. Connectedness

We show that basic properties of connected metric spaces exist in the range of hyperarithmetical analysis, including the intermediate value theorem. We also obtain some elegant equivalences in Theorem 2.13.

First of all, Cantor and Jordan were the first to study connectedness ([wilders]), namely as in the first item in Definition 2.11. The connectedness notions from the latter are equivalent for compact metric spaces in light of [mannetti]*§4.39 or [pugh]*p. 123.

Definition 2.11 (Connectedness).
  • A metric space (M,d)(M,d) is chain connected in case for any w,vMw,v\in M and ε>0\varepsilon>0, there is a sequence w=x0,x1,,xn1,xn=vMw=x_{0},x_{1},\dots,x_{n-1},{x_{n}}=v\in M such that for all i<ni<n we have d(xi,xi+1)<εd(x_{i},x_{i+1})<\varepsilon.

  • A metric space (M,d)(M,d) is connected in case MM is not the disjoint union of two non-empty open sets.

We shall study the following generalisation of the intermediate value theorem.

Principle 2.12 (Intermediate Value Theorem).

Let (M,d)(M,d) be a sequentially compact and chain connected metric space and let f:Mf:M\rightarrow{\mathbb{R}} be continuous. If f(w)<c<f(v)f(w)<c<f(v) for some v,wMv,w\in M and cc\in{\mathbb{R}}, then there is uMu\in M with f(u)=cf(u)=c.

The approximate intermediate value theorem is the previous principle with the conclusion weakened to ‘then for any ε>0\varepsilon>0 there is uMu\in M with |f(u)c|<ε|f(u)-c|<\varepsilon.’ The latter theorem is well-known from constructive mathematics (see e.g. [bridge1]*p. 40).

Theorem 2.13 (ACA0ω+IND2\textup{{ACA}}_{0}^{\omega}+\textup{{IND}}_{2}).

The principle Finite Choice follows from any of the items (a)-(g) where (M,d)(M,d) is any metric space with MM\subset{\mathbb{R}}; the principle QF-AC0,1\textup{{QF-AC}}^{0,1} implies items (a)-(g).

  1. (a)

    The intermediate value theorem as in Principle 2.12.

  2. (b)

    Principle 2.12 for Lipschitz continuous functions.

  3. (c)

    The approximate intermediate value theorem.

  4. (d)

    Let (M,d)(M,d) be a sequentially compact and chain connected metric space and let f:M{0,1}f:M\rightarrow\{0,1\} be continuous. Then ff is constant on MM.

  5. (e)

    Let (M,d)(M,d) be a sequentially compact and chain connected metric space and let f:Mf:M\rightarrow{\mathbb{R}} be locally constant. Then ff is constant on MM.

  6. (f)

    Let (M,d)(M,d) be sequentially compact and chain connected and let f:Mf:M\rightarrow{\mathbb{R}} be locally constant and continuous. Then ff is constant on MM.

  7. (g)

    For a sequentially compact metric space (M,d)(M,d), chain connectedness implies connectedness

  8. (h)

    Let (M,d)(M,d) be a sequentially compact and chain connected metric space and let f:Mf:M\rightarrow{\mathbb{R}} be (Lipschitz) continuous. Then ff is bounded on MM.

  9. (i)

    Item (h) with ‘ff is bounded’ replaced by ‘f(M)f(M) is not dense in {\mathbb{R}}’.

  10. (j)

    Item (h) with ‘ff is bounded’ replaced by ‘f(M)f(M) is closed’.

Moreover, items (a), (c), and (d)-(g) are equivalent.

Proof.

First of all, we show that item (a) implies Finite Choice. To this end, suppose the latter is false and consider M0M_{0} as in (2.1). Let (qn)n(q_{n})_{n\in{\mathbb{N}}} be an enumeration of the rationals (without repetitions) and define d3:M02d_{3}:M_{0}^{2}\rightarrow{\mathbb{R}} as follows: d3(w,v):=|q|w|q|v||d_{3}(w,v):=|q_{|w|}-q_{|v|}| for w,vM0w,v\in M_{0}. Then (M0,d3)(M_{0},d_{3}) is a sequentially compact metric space, which is proved in the same way as for the previous metrics d0,d1,d2d_{0},d_{1},d_{2}, namely that any sequence in M0M_{0} can have at most finitely many different elements. That (M0,d3)(M_{0},d_{3}) is chain connected is proved using IND2\textup{{IND}}_{2}. Indeed, fix u,wM0,ε>0u,w\in M_{0},\varepsilon>0 and consider d3(w,v)=|q|w|q|v||d_{3}(w,v)=|q_{|w|}-q_{|v|}|. Let q|w|=r0,r1,,rk1,rk=q|v|q_{|w|}=r_{0},r_{1},\dots,r_{k-1},r_{k}=q_{|v|}\in{\mathbb{Q}} be a finite sequence such that |riri+1|<ε|r_{i}-r_{i+1}|<\varepsilon for i<ki<k. Using IND2\textup{{IND}}_{2}, there are wiM0w_{i}\in M_{0} such that q|wi|=riq_{|w_{i}|}=r_{i} for i<ki<k, and chain connectedness of M0M_{0} follows.

Now define f:M0f:M_{0}\rightarrow{\mathbb{R}} by f(w)=12q|w|f(w)=\frac{1}{2}q_{|w|}, which is (Lipschitz) continuous, essentially by the definition of d3d_{3}, as we have:

|f(w)f(v)|=|12q|w|12q|v||=12|q|w|q|v||12d3(w,v).\textstyle|f(w)-f(v)|=|\frac{1}{2}q_{|w|}-\frac{1}{2}q_{|v|}|=\frac{1}{2}|q_{|w|}-q_{|v|}|\leq\frac{1}{2}d_{3}(w,v).

However, the range of ff consists of rationals, i.e. it does not have the intermediate value property. This contradiction yields Finite Choice. The same proof goes through for items (h)-(j).

Secondly, assume QF-AC0,1\textup{{QF-AC}}^{0,1} and let (M,d),f:M,w,vM,(M,d),f:M\rightarrow{\mathbb{R}},w,v\in M, and cc\in{\mathbb{R}} be as in Principle 2.12. Since MM is chain connected, we have

(k)(z1)(z(0)=wz(|w|1)=v(i<|z|1)d(z(i),z(i+1))<12k).\textstyle(\forall k\in{\mathbb{N}})(\exists z^{1^{*}})(z(0)=w\wedge z(|w|-1)=v\wedge(\forall i<|z|-1)d(z(i),z(i+1))<\frac{1}{2^{k}}). (2.5)

Apply QF-AC0,1\textup{{QF-AC}}^{0,1} to obtain a sequence (zk)k(z_{k})_{k\in{\mathbb{N}}} of finite sequences. Define a sequence (tk)k(t_{k})_{k\in{\mathbb{N}}} in MM where tkt_{k} is the first element tt in zkz_{k} such that f(t)cf(t)\geq c. By sequential completeness, there is a convergent sub-sequence (sk)k(s_{k})_{k\in{\mathbb{N}}} with limit sMs\in M. Since ff is continuous, we have limkf(sk)=f(s)\lim_{k\rightarrow\infty}f(s_{k})=f(s) and hence f(s)=cf(s)=c.

Thirdly, to show that item (g) implies Finite Choice, again suppose the latter is false and consider (M0,d3)(M_{0},d_{3}). By the above, the latter is sequentially compact and chain connected. To show that it is not connected, define O1:={wM0:q|w|>π}O_{1}:=\{w\in M_{0}:q_{|w|}>\pi\} and O2:={wM0:q|w|<π}O_{2}:=\{w\in M_{0}:q_{|w|}<\pi\}, verify that they are open and disjoint, and observe that M0=O1O2M_{0}=O_{1}\cup O_{2}, i.e. item (g) is false. Note also that f:M0f:M_{0}\rightarrow{\mathbb{R}} defined as f(w)=1f(w)=1 if wO1w\in O_{1} and 0 otherwise, is continuous but not constant, i.e. item (d) also implies Finite Choice.

To derive item (g) from QF-AC0,1\textup{{QF-AC}}^{0,1}, let (M,d)(M,d) be as in the former, i.e. sequentially compact and chain connected. Suppose MM is not connected, i.e. M=O1O2M=O_{1}\cup O_{2} where the latter are open, disjoint, and non-empty. Now fix vO1v\in O_{1} and wO2w\in O_{2} and consider (2.5). Apply QF-AC0,1\textup{{QF-AC}}^{0,1} to obtain a sequence (zk)k(z_{k})_{k\in{\mathbb{N}}} of finite sequences. Define sequences (sk)k(s_{k})_{k\in{\mathbb{N}}} and (tk)k(t_{k})_{k\in{\mathbb{N}}} in MM where tkt_{k} is the first element tt in zkz_{k} such that tO2t\in O_{2} and sks_{k} is the predecessor of tt in zkz_{k}. By sequential completeness, (sk)k(s_{k})_{k\in{\mathbb{N}}} and (tk)k(t_{k})_{k\in{\mathbb{N}}} have convergent sub-sequences, with the same limit by construction. However, if this limit is in O1O_{1}, then so is (tk)k(t_{k})_{k\in{\mathbb{N}}} eventually, a contradiction. Similarly, if this limit is in O2O_{2}, then so is (sk)k(s_{k})_{k\in{\mathbb{N}}} eventually, a contradiction. In each case we obtain a contradiction, i.e. MM must be connected, and item (g) follows. The same proof goes through for item (d).

Next, item (g) implies item (e) as in case the latter fails for f:Mf:M\rightarrow{\mathbb{R}}, say with f(w)<f(v)f(w)<_{{\mathbb{R}}}f(v), then O1={zM:f(z)f(w)}O_{1}=\{z\in M:f(z)\leq f(w)\} and O2={zM:f(z)>f(w)}O_{2}=\{z\in M:f(z)>f(w)\} are open, disjoint, and non-empty sets such that M=O1O2M=O_{1}\cup O_{2}, i.e. item (g) fails too. To show that item (e) implies Finite Choice, suppose the latter is false and let (M0,d3)(M_{0},d_{3}) be as above. Define g:M0g:M_{0}\rightarrow{\mathbb{R}} as g(w)=ng(w)=n in case |q|w||[nπ,(n+1)π]|q_{|w|}|\in[n\pi,(n+1)\pi]. Clearly, ff is locally constant but not constant, i.e. item (e) is false. To derive item (g) from item (e) (and item (f)), suppose the former is false, i.e. (M,d)(M,d) is a sequentially compact and chain connected metric space that is not connected. Let M=O1O2M=O_{1}\cup O_{2} be the associated decomposition and note that f:Mf:M\rightarrow{\mathbb{R}} defined by f(w)=1f(w)=1 if wO1w\in O_{1} and 0 otherwise, is locally constant (and continuous) but not constant, i.e. item (e) (and (f)) also fails. The equivalence for item (d) follows in the same way.

To show that item (g) implies item (a), suppose the latter is false for f:Mf:M\rightarrow{\mathbb{R}} and cc\in{\mathbb{R}}, i.e. f(w)cf(w)\neq c for all wMw\in M. By assumption, O1:={wM:f(w)<c}O_{1}:=\{w\in M:f(w)<c\} and O2:={wM:f(w)>c}O_{2}:=\{w\in M:f(w)>c\} are open, disjoint, and non-empty, i.e. item (g) also fails. To show that item (a) (and item (c)) implies item (g), suppose the latter fails for M=O1O2M=O_{1}\cup O_{2}, i.e. the latter are open, non-empty, and disjoint. Then f:Mf:M\rightarrow{\mathbb{R}} defined by f(w)=1f(w)=1 if wO1w\in O_{1} and 0 otherwise, is continuous but does not have the (approximate) intermediate value property. ∎

Regarding item (i), we could not find a way of replacing ‘f(M)f(M) is not dense in {\mathbb{R}}’ by ‘f(M)f(M) has finite measure’. We could study local connectedness and obtain similar results, but feel this section is long enough as is.

In conclusion, we have identified many basic properties of metric spaces that exist in the range of hyperarithmetical analysis. We believe there to be many more such principles in e.g. topology.

3. Functions of Bounded variation and around

We introduce functions of bounded variation (Section 3.1) and show that their basic properties exist in the range of hyperarithmetical analysis (Section 3.2). Similar to Theorem 2.8, we could restrict to arithmetically defined functions.

3.1. Bounded variation and variations

The notion of bounded variation (often abbreviated BVBV) was first explicitly444Lakatos in [laktose]*p. 148 claims that Jordan did not invent or introduce the notion of bounded variation in [jordel], but rather discovered it in Dirichlet’s 1829 paper [didi3]. introduced by Jordan around 1881 ([jordel]) yielding a generalisation of Dirichlet’s convergence theorems for Fourier series. Indeed, Dirichlet’s convergence results are restricted to functions that are continuous except at a finite number of points, while BVBV-functions can have infinitely many points of discontinuity, as already studied by Jordan, namely in [jordel]*p. 230. In this context, the total variation Vab(f)V_{a}^{b}(f) of f:[a,b]f:[a,b]\rightarrow{\mathbb{R}} is defined as:

supax0<<xnbi=0n|f(xi)f(xi+1)|.\textstyle\sup_{a\leq x_{0}<\dots<x_{n}\leq b}\sum_{i=0}^{n}|f(x_{i})-f(x_{i+1})|. (3.1)

The following definition provides two ways of defining ‘BVBV-function’. We have mostly studied the first one ([dagsamXI, samBIG, samBIG3]) but will use the second one in this paper.

Definition 3.1 (Variations on variation).
  1. (a)

    The function f:[a,b]f:[a,b]\rightarrow{\mathbb{R}} has bounded variation on [a,b][a,b] if there is k0k_{0}\in{\mathbb{N}} such that k0i=0n|f(xi)f(xi+1)|k_{0}\geq\sum_{i=0}^{n}|f(x_{i})-f(x_{i+1})| for any partition x0=a<x1<<xn1<xn=bx_{0}=a<x_{1}<\dots<x_{n-1}<x_{n}=b.

  2. (b)

    The function f:[a,b]f:[a,b]\rightarrow{\mathbb{R}} has total variation zz\in{\mathbb{R}} on [a,b][a,b] if Vab(f)=zV_{a}^{b}(f)=z.

We recall the ‘virtual’ or ‘comparative’ meaning of suprema in RM from e.g. [simpson2]*X.1. In particular, a formula ‘supAb\sup A\leq b’ is merely shorthand for (essentially) the well-known definition of the supremum.

Secondly, the fundamental theorem about BVBV-functions is formulated as follows.

Theorem 3.2 (Jordan decomposition theorem, [jordel]*p. 229).

A BVBV-function f:[0,1]f:[0,1]\rightarrow{\mathbb{R}} is the difference of two non-decreasing functions g,h:[0,1]g,h:[0,1]\rightarrow{\mathbb{R}}.

Theorem 3.2 has been studied via second-order representations in [groeneberg, kreupel, nieyo, verzengend]. The same holds for constructive analysis by [briva, varijo, brima, baathetniet], involving different (but related) constructive enrichments. We have obtained many equivalences for the Jordan decomposition theorem, formulated using item (a) from Definition 3.1 in [dagsamXI, samBIG3], involving the following principle.

Principle 3.3 (cocode0\textup{{cocode}}_{0}).

A countable set A[0,1]A\subset[0,1] can be enumerated.

This principle is ‘explosive’ in that ACA0ω+cocode0\textup{{ACA}}_{0}^{\omega}+\textup{{cocode}}_{0} proves ATR0\textup{{ATR}}_{0} while Π11-CA0ω+cocode0\Pi_{1}^{1}\text{-{{CA}}}_{0}^{\omega}+\textup{{cocode}}_{0} proves Π21-CA0\Pi_{2}^{1}\text{-{{CA}}}_{0} (see [dagsamX]*§4).

Thirdly, f:f:{\mathbb{R}}\rightarrow{\mathbb{R}} is regulated if for every x0x_{0} in the domain, the ‘left’ and ‘right’ limit f(x0)=limxx0f(x)f(x_{0}-)=\lim_{x\rightarrow x_{0}-}f(x) and f(x0+)=limxx0+f(x)f(x_{0}+)=\lim_{x\rightarrow x_{0}+}f(x) exist. Feferman’s μ\mu readily provides the limit of (f(x+12n))n(f(x+\frac{1}{2^{n}}))_{n\in{\mathbb{N}}} if it exists, i.e. the notation λx.f(x+)\lambda x.f(x+) for regulated ff makes sense in ACA0ω\textup{{ACA}}_{0}^{\omega}. On a historical note, Scheeffer and Darboux study discontinuous regulated functions in [scheeffer, darb] without using the term ‘regulated’, while Bourbaki develops Riemann integration based on regulated functions in [boerbakies]. Finally, BVBV-functions are regulated while Weierstrass’ ‘monster’ function is a natural example of a regulated function not in BVBV.

3.2. Bounded variation and hyperarithmetical analysis

We identify a number of statements about BVBV-functions that exist within the range of hyperarithmetical analysis, assuming ACA0ω\textup{{ACA}}_{0}^{\omega}. We even obtain some elegant equivalences and discus the (plentiful) variations of these results in Section 4.3.

First of all, the following principle appears to be important, which is just cocode0\textup{{cocode}}_{0} from the previous section restricted to strongly countable sets.

Principle 3.4 (cocode1\textup{{cocode}}_{1}).

A strongly countable set A[0,1]A\subset[0,1] can be enumerated.

Some RM-results for cocode1\textup{{cocode}}_{1} may be found in [dagsamXI]*§2.2.1; many variations are possible and these systems all exist in the range of hyperarithmetical analysis. The cited results are not that satisfying as they mostly deal with properties of strongly countable sets, in contrast to the below.

Secondly, we have the following theorem, establishing that items (ii)-(v) exist in the range of hyperarithmetical analysis.

Theorem 3.5 (ACA0ω+IND1\textup{{ACA}}_{0}^{\omega}+\textup{{IND}}_{1}).

The higher items imply the lower ones.

  1. (i)

    The principle QF-AC0,1\textup{{QF-AC}}^{0,1}.

  2. (ii)

    (Jordan) For fBVf\in BV with V01(f)=1V_{0}^{1}(f)=1, there are non-decreasing g,h:[0,1]g,h:[0,1]\rightarrow{\mathbb{R}} such that f=ghf=g-h.

  3. (iii)

    For fBVf\in BV with V01(f)=1V_{0}^{1}(f)=1, there is a sequence that includes all points of discontinuity of ff.

  4. (iv)

    For fBVf\in BV with V01(f)=1V_{0}^{1}(f)=1, the supremum555To be absolutely clear, we assume, for the existence of a functional Φ:2\Phi:{\mathbb{Q}}^{2}\rightarrow{\mathbb{R}} such that (p,q[0,1])(Φ(p,q)=supx[p,q]f(x))(\forall p,q\in{\mathbb{Q}}\cap[0,1])(\Phi(p,q)=\sup_{x\in[p,q]}f(x))). supx[p,q]f(x)\sup_{x\in[p,q]}f(x) exists for p,q[0,1]p,q\in[0,1]\cap{\mathbb{Q}}.

  5. (v)

    cocode1\textup{{cocode}}_{1}.

  6. (vi)

    weak-Σ11\Sigma_{1}^{1}-AC0.

Items (ii)-(iii) are equivalent; we only use IND1\textup{{IND}}_{1} to derive cocode1\textup{{cocode}}_{1} from item (5).

Proof.

Assume QF-AC0,1\textup{{QF-AC}}^{0,1} and let fBVf\in BV be such that V01(f)=1V_{0}^{1}(f)=1. By [dagsamXIV]*Theorem 2.16, ACA0ω\textup{{ACA}}_{0}^{\omega} suffices to enumerate all jump discontinuities of a regulated function, while ff is regulated by [dagsamXI]*Theorem 3.33. Then V01(f)=1V_{0}^{1}(f)=1 implies that

(k)(x0,,xmI)[(i<m)(xi<xi+1)112k<j=0m|f(xj)f(xj+1)|].\textstyle(\forall k\in{\mathbb{N}})(\exists x_{0},\dots,x_{m}\in I)\big{[}(\forall i<m)(x_{i}<x_{i+1})\wedge 1-\frac{1}{2^{k}}<\sum_{j=0}^{m}|f(x_{j})-f(x_{j+1})|\big{]}.

The formula in square brackets is arithmetical, i.e. since (2)(\exists^{2}) is available we may apply QF-AC0,1\textup{{QF-AC}}^{0,1} to obtain a sequence of finite sequences (wn)n(w_{n})_{n\in{\mathbb{N}}} witnessing the previous centred formula. This sequence includes all removable discontinuities of ff. Indeed, suppose y0[0,1]y_{0}\in[0,1] is such that f(y0)=f(y0+)f(y0)f(y_{0}-)=f(y_{0}+)\neq f(y_{0}) is not among the reals in (wn)n(w_{n})_{n\in{\mathbb{N}}}. Let k0k_{0}\in{\mathbb{N}} be such that |f(y0+)f(y0)|>12k0|f(y_{0}+)-f(y_{0})|>\frac{1}{2^{k_{0}}} and note that 112k0+1<j=0mk0+1|f(xj)f(xj+1)|1-\frac{1}{2^{k_{0}+1}}<\sum_{j=0}^{m_{k_{0}+1}}|f(x_{j})-f(x_{j+1})| for wk0+1=(x0,,xmk0+1)w_{k_{0}+1}=(x_{0},\dots,x_{m_{k_{0}+1}}) by assumption. Extending wk0+1w_{k_{0}+1} with y0y_{0} and points z0<y0<u0z_{0}<y_{0}<u_{0} close enough to y0y_{0}, we obtain a partition of [0,1][0,1] that witnesses that V01(f)>1V_{0}^{1}(f)>1, contradicting our assumptions. Since ff is regulated, it only has removable and jump discontinuities, i.e. item (iii) follows from QF-AC0,1\textup{{QF-AC}}^{0,1} as required.

By [dagsamXI]*Theorem 3.33, ACA0ω\textup{{ACA}}_{0}^{\omega} suffices to enumerate the points of discontinuity of any monotone g:[0,1]g:[0,1]\rightarrow{\mathbb{R}}, i.e. item (ii) implies item (iii). To obtain item (ii) from item (iii), note that the supremum over {\mathbb{R}} in (3.1) can be replaced by a supremum over {\mathbb{Q}} and any sequence that includes all points of discontinuity of ff. Hence, we may use (2)(\exists^{2}) to define the weakly increasing function g(x):=λx.V0x(f)g(x):=\lambda x.V_{0}^{x}(f). One readily verifies that h(x):=g(x)f(x)h(x):=g(x)-f(x) is also weakly increasing, i.e. f=ghf=g-h as in item (ii) follows. To obtain item (5) from item (iii), note that -similar to the previous- the supremum over {\mathbb{R}} in supx[p,q]f(x)\sup_{x\in[p,q]}f(x) can be replaced by a supremum over {\mathbb{Q}} and any sequence that includes all points of discontinuity of ff.

To derive cocode1\textup{{cocode}}_{1} from item (5), let A[0,1]A\subset[0,1] and Y:[0,1]Y:[0,1]\rightarrow{\mathbb{R}} such that the latter is injective and surjective on the former. Now define f:[0,1]f:[0,1]\rightarrow{\mathbb{R}} as follows: f(x):=12Y(x)f(x):=\frac{1}{2^{Y(x)}} if xAx\in A, and 0 otherwise. Using IND1\textup{{IND}}_{1}, ff is in BVBV and V01(f)=1V_{0}^{1}(f)=1. Now use (2)(\exists^{2}) to decide whether supx[0,12]f(x)<1\sup_{x\in[0,\frac{1}{2}]}f(x)<1; if the latter holds, ‘11’ is the first bit of the binary expansion of x0[0,1]x_{0}\in[0,1] such that Y(x0)=0Y(x_{0})=0. Using the supremum functional and (2)(\exists^{2}), the usual interval-halving technique then allows us to enumerate AA, as required for cocode1\textup{{cocode}}_{1}. For the final part, let φ\varphi be arithmetical and such that (n)(!X)φ(n,X)(\forall n\in{\mathbb{N}})(\exists!X\subset{\mathbb{N}})\varphi(n,X). Use 2\exists^{2} to define η:[0,1](2×2)\eta:[0,1]\rightarrow(2^{{\mathbb{N}}}\times 2^{{\mathbb{N}}}) such that η(x)=(f,g)\eta(x)=(f,g) outputs the binary expansions of xx, taking f=gf=g if there is only one. Then En={x[0,1]:φ(n,η(x)(1))φ(n,η(x)(2))}E_{n}=\{x\in[0,1]:\varphi(n,\eta(x)(1))\vee\varphi(n,\eta(x)(2))\} is a singleton and Y(x):=(μn)(xEn)Y(x):=(\mu n)(x\in E_{n}) is injective and surjective on A=nEnA=\cup_{n\in{\mathbb{N}}}E_{n}. The enumeration of AA provided by cocode1\textup{{cocode}}_{1} yields the consequent of weak-Σ11\Sigma_{1}^{1}-AC0. ∎

As to the role of the Axiom of Choice in Theorem 3.5, we note that the items (ii)-(v) can also be proved without QF-AC0,1\textup{{QF-AC}}^{0,1}. Indeed, λx.V0x(f)\lambda x.V_{0}^{x}(f) as in (3.1) involves a supremum over {\mathbb{R}}, which can be defined in Z2Ω{\textsf{{Z}}}_{2}^{\Omega} using the well-known interval-halving technique, i.e. the usual textbook proof (see e.g. [voordedorst]) goes through in Z2Ω{\textsf{{Z}}}_{2}^{\Omega}.

Thirdly, we have the following corollary using slightly more induction.

Corollary 3.6.

Over ACA0ω+IND2\textup{{ACA}}_{0}^{\omega}+\textup{{IND}}_{2}, item (iii) from Theorem 3.5 is equivalent to:

for fBVf\in BV with V01(f)=1V_{0}^{1}(f)=1 and with Fourier coefficients given, there is a sequence (xn)n(x_{n})_{n\in{\mathbb{N}}} outside of which the Fourier series converges to f(x)f(x).

Proof.

We note that IND2\textup{{IND}}_{2} suffices to guarantee that BVBV-functions are regulated by [dagsamXI]*Theorem 3.33. Now, the Fourier series of a BVBV-function always converges to f(x+)+f(x)2\frac{f(x+)+f(x-)}{2} and this fact is provable in ACA0ω\textup{{ACA}}_{0}^{\omega} if the Fourier coefficients are given, as discussed in (a lot of detail in) [samBIG]*§3.4.4. Hence, item (iii) of Theorem 3.5 immediately implies the centred statement in item (a), while for the reversal, the centred statement provides a sequence that includes all removable discontinuities, i.e. where f(x)f(x+)f(x)\neq f(x+) but f(x+)=f(x)f(x+)=f(x-). By [dagsamXIV]*Theorem 2.16, ACA0ω\textup{{ACA}}_{0}^{\omega} suffices to enumerate all jump discontinuities of a regulated function. Since there are no other discontinuities for ff, the corollary follows. ∎

We could obtain similar results for e.g. Bernstein or Hermit-Fejer polynomials as analogous results hold for BVBV-functions (see [samBIG3]). Other variations are discussed in Remark 4.3 below.

Fifth, as noted in Section 3.1, enumerating the points of discontinuity of a regulated function implies cocode0\textup{{cocode}}_{0}; the latter yields ATR0\textup{{ATR}}_{0} when combined with ACA0ω\textup{{ACA}}_{0}^{\omega}. By contrast, item (ii) in the following theorem is much weaker.

Theorem 3.7 (ACA0ω+IND0\textup{{ACA}}_{0}^{\omega}+\textup{{IND}}_{0}).

The higher items imply the lower ones.

  1. (i)

    The principle QF-AC0,1\textup{{QF-AC}}^{0,1}.

  2. (ii)

    For regulated f:[0,1]f:[0,1]\rightarrow{\mathbb{R}} with infinite DfD_{f}, there is a sequence of distinct points of discontinuity of ff.

  3. (iii)

    The principle Finite Choice.

  4. (iv)

    The principle finite-Σ11\Sigma_{1}^{1}-AC0.

Proof.

The first downward implication is immediate by applying QF-AC0,1\textup{{QF-AC}}^{0,1} -modulo (2)(\exists^{2})- to ‘DfD_{f} is not finite’. The final implication is straightforward. For the second downward implication, let (Xn)n(X_{n})_{n\in{\mathbb{N}}} be a sequence of non-empty finite sets and let η:[0,1]\eta:[0,1]\rightarrow{\mathbb{R}} be such that η(x)\eta(x) is the binary expansion of xx, choosing a tail of zeros if necessary. Define h:[0,1]h:[0,1]\rightarrow{\mathbb{R}} as:

h(x):={12n if η(x)=1111k+1-times0g0gk and (ik)(giXi)0otherwise.h(x):=\begin{cases}\frac{1}{2^{n}}&\textup{ if $\eta(x)=\underbrace{11\dots 11}_{\textup{$k+1$-times}}*\langle 0\rangle*g_{0}\oplus\dots\oplus g_{k}$ and $(\forall i\leq k)(g_{i}\in X_{i})$}\\ 0&\textup{otherwise}\end{cases}.

Using IND2\textup{{IND}}_{2}, one readily shows that hh is regulated (with left and right limits equal to zero) and that DhD_{h} is infinite if nXn\cup_{n\in{\mathbb{N}}}X_{n} is. Any sequence in DhD_{h} then yields a sequence as in the consequent of Finite Choice. ∎

An interesting variation is provided by the following corollary. We conjecture that Finite Choice cannot be obtained from the second item.

Corollary 3.8 (ACA0ω\textup{{ACA}}_{0}^{\omega}).

The higher items imply the lower ones.

  1. (i)

    The principle QF-AC0,1\textup{{QF-AC}}^{0,1}.

  2. (ii)

    For f:[0,1]f:[0,1]\rightarrow{\mathbb{R}} in BVBV with infinite DfD_{f}, there is a sequence of distinct points of discontinuity of ff.

  3. (iii)

    (Finite Choice\textup{{Finite Choice}}^{\prime}) Let (Xn)n(X_{n})_{n\in{\mathbb{N}}} be a sequence of non-empty finite sets in [0,1][0,1] and let gg\in{\mathbb{N}}^{{\mathbb{N}}} be such that |Xn|g(n)|X_{n}|\leq g(n). Then there is a sequence (xn)n(x_{n})_{n\in{\mathbb{N}}} such that xnXnx_{n}\in X_{n} for all nn\in{\mathbb{N}}.

  4. (iv)

    The principle height-Σ11\Sigma_{1}^{1}-AC0.

Proof.

The final implication is straightforward while the first one follows as in the proof of the theorem. For the second downward implication, let (Xn)n(X_{n})_{n\in{\mathbb{N}}} be a sequence of non-empty finite sets with |Xn|g(n)|X_{n}|\leq g(n). Define h:[0,1]h:[0,1]\rightarrow{\mathbb{R}} as in the proof of the theorem but replacing ‘12n\frac{1}{2^{n}}’ in the first case by 12n1g(n)+1\frac{1}{2^{n}}\frac{1}{g(n)+1}. By construction, hh is in BVBV with V01(f)1V_{0}^{1}(f)\leq 1 and the set DhD_{h} is infinite if nXn\cup_{n\in{\mathbb{N}}}X_{n} is. Any sequence in DhD_{h} then yields the sequence as in the consequent of Finite Choice\textup{{Finite Choice}}^{\prime}. ∎

Finally, we discuss numerous possible variations of the above results in Section 4.3, including Riemann integration and rectifiability.

4. Other topics in hyperarithmetical analysis

4.1. Semi-continuity and closed sets

We show that basic properties of semi-continuous functions, like the extreme value theorem, exist in the range of hyperarithmetical analysis. Since upper semi-continuous functions are closely related to closed sets, the latter also feature prominently.

First of all, we need Baire’s notion of semi-continuity first introduced in [beren].

Definition 4.1.

For f:[0,1]f:[0,1]\rightarrow{\mathbb{R}}, we have the following definitions:

  • ff is upper semi-continuous at x0[0,1]x_{0}\in[0,1] if for any kk\in{\mathbb{N}}, there is NN\in{\mathbb{N}} such that (yB(x0,12N))(f(y)<f(x0)+12k)(\forall y\in B(x_{0},\frac{1}{2^{N}}))(f(y)<f(x_{0})+\frac{1}{2^{k}}).

  • ff is lower semi-continuous at x0[0,1]x_{0}\in[0,1] if for any kk\in{\mathbb{N}}, there is NN\in{\mathbb{N}} such that (yB(x0,12N))(f(y)>f(x0)12k)(\forall y\in B(x_{0},\frac{1}{2^{N}}))(f(y)>f(x_{0})-\frac{1}{2^{k}}).

We use the common abbreviations ‘usco’ and ‘lsco’ for the previous notions. We say that ‘f:[0,1]f:[0,1]\rightarrow{\mathbb{R}} is usco’ if ff is usco at every x[0,1]x\in[0,1]. Following [martino], the extreme value theorem does not really generalise beyond semi-continuous functions.

Secondly, we have the following theorem, a weaker version of which is in [dagsamXVI]. We repeat that since the characteristic function of a closed set is usco, the connection between items (ii) and ClC is not that surprising.

Theorem 4.2 (ACA0ω+IND2\textup{{ACA}}_{0}^{\omega}+\textup{{IND}}_{2}).

The higher items imply the lower ones.

  1. (i)

    The principle QF-AC0,1\textup{{QF-AC}}^{0,1}.

  2. (ii)

    (Extreme value theorem) For usco f:f:{\mathbb{R}}\rightarrow{\mathbb{R}} with y=supx[n,n+1]f(x)y=\sup_{x\in[n,n+1]}f(x) for all nn\in{\mathbb{N}}, there is (xn)n(x_{n})_{n\in{\mathbb{N}}} such that (n)(xn[n,n+1]f(xn)=y)(\forall n\in{\mathbb{N}})(x_{n}\in[n,n+1]\wedge f(x_{n})=y).

  3. (iii)

    (ClC, [dagsamXVI]) Let (Cn)n(C_{n})_{n\in{\mathbb{N}}} be a sequence of non-empty closed sets in [0,1][0,1]. Then there is (xn)n(x_{n})_{n\in{\mathbb{N}}} such that xnCnx_{n}\in C_{n} for all nn\in{\mathbb{N}}.

  4. (iv)

    For usco and regulated f:f:{\mathbb{R}}\rightarrow{\mathbb{R}} with y=supx[n,n+1]f(x)y=\sup_{x\in[n,n+1]}f(x) for all nn\in{\mathbb{N}}, there is (xn)n(x_{n})_{n\in{\mathbb{N}}} such that (n)(xn[n,n+1]f(xn)=y)(\forall n\in{\mathbb{N}})(x_{n}\in[n,n+1]\wedge f(x_{n})=y).

  5. (v)

    (Finite Choice) Let (Xn)n(X_{n})_{n\in{\mathbb{N}}} be a sequence of non-empty finite sets in [0,1][0,1]. Then there is (xn)n(x_{n})_{n\in{\mathbb{N}}} such that xnXnx_{n}\in X_{n} for all nn\in{\mathbb{N}}.

  6. (vi)

    The principle finite-Σ11\Sigma_{1}^{1}-AC0.

Proof.

For the first downward implication, if the supremum yy is given, we have (n,k)(x[n,n+1])(f(x)>y12k)(\forall n,k\in{\mathbb{N}})(\exists x\in[n,n+1])(f(x)>y-\frac{1}{2^{k}}), and applying QF-AC0,1\textup{{QF-AC}}^{0,1} yields a sequence (xn,k)n,k(x_{n,k})_{n,k\in{\mathbb{N}}}. Since (2)ACA0(\exists^{2})\rightarrow\textup{{ACA}}_{0}, the latter has a convergent sub-sequence (for fixed nn\in{\mathbb{N}}), with limit say yn[n,n+1]y_{n}\in[n,n+1] by sequential completeness. One readily verifies that f(yn)=yf(y_{n})=y for any nn\in{\mathbb{N}} as ff is usco. For the second implication, fix a sequence (Cn)n(C_{n})_{n\in{\mathbb{N}}} of closed sets and define h:[0,1]h:[0,1]\rightarrow{\mathbb{R}} as follows using Feferman’s μ\mu:

h(x):={1xnCnn>00 otherwise.h(x):=\begin{cases}1&x-n\in C_{n}\wedge n>0\\ 0&\textup{ otherwise}\end{cases}. (4.1)

Since hh is essentially the characteristic function of closed sets, hh is usco on [n,n+1][n,n+1] by definition, for each nn\in{\mathbb{N}}. The sequence provided by item (ii) then clearly satisfies xnCnx_{n}\in C_{n}. To show that ClC implies item (ii), let f:[0,1]f:[0,1]\rightarrow{\mathbb{R}} and yy\in{\mathbb{R}} be as in the latter and define Cn,k={x[n,n+1]:f(x)y12k}C_{n,k}=\{x\in[n,n+1]:f(x)\geq y-\frac{1}{2^{k}}\} which is non-empty by definition and closed as ff is usco. The sequence provided by ClC yields xn[n,n+1]x_{n}\in[n,n+1] such that f(xn)=yf(x_{n})=y. The function hh from (4.1) is also regulated in case each CnC_{n} is finite, i.e. the fourth implication also follows. ∎

We note that item (ii) is equivalent to e.g. the sequential version of the Cantor intersection theorem ([dagsamXVI]).

Thirdly, ClC from Theorem 4.2 is provable in WKL0\textup{{WKL}}_{0} if assume that the closed sets are given by a sequence of RM-codes (see [simpson2]*IV.1.8). We next study ClC for an alternative representation of closed sets from [browner, brownphd, browner2] as follows.

Definition 4.3.

A (code for a) separably closed set is a sequence S=(xn)nS=(x_{n})_{n\in{\mathbb{N}}} of reals. We write ‘xS¯x\in\overline{S}’ in case (k)(n)(|xxn|<12k(\forall k\in{\mathbb{N}})(\exists n\in{\mathbb{N}})(|x-x_{n}|<\frac{1}{2^{k}}. A (code for a) separably open set is a code for the (separably closed) complement.

Next, item (i) in Theorem 4.4 is a weakening of [simpson2]*V.4.10, which in turn is a second-order version of the countable union theorem. In each case, the antecedent only expresses that for every nn, there exists an enumeration of AnA_{n}; abusing notation666In particular, the formula ‘XAn¯X\in\overline{A_{n}}’ in Theorem 4.4 is short-hand for ((Ym)m)[(Y)(YAn(m)(Ym=Y))(k)(l)(X¯k=Yl¯kYlAn)],(\exists(Y_{m})_{m\in{\mathbb{N}}})\big{[}(\forall Y\subset{\mathbb{N}})(Y\in A_{n}\rightarrow(\exists m\in{\mathbb{N}})(Y_{m}=Y))\wedge(\forall k\in{\mathbb{N}})(\exists l\in{\mathbb{N}})(\overline{X}k=\overline{Y_{l}}k\wedge Y_{l}\in A_{n})\big{]}, which is slightly more unwieldy. slightly, we still write ‘XAn¯X\in\overline{A_{n}}’ as in Definition 4.3, leaving the enumeration of AnA_{n} implicit. We sometimes identify subsets XX\subset{\mathbb{N}} and elements f2f\in 2^{{\mathbb{N}}}.

Theorem 4.4 (ACA0\textup{{ACA}}_{0}).

The following items are intermediate between Σ11\Sigma_{1}^{1}-AC0 and weak-Σ11\Sigma_{1}^{1}-AC0.

  1. (i)

    Let (An)n(A_{n})_{n\in{\mathbb{N}}} be a sequence of analytic codes such that each AnA_{n} is enumerable and non-empty. There is a sequence (Xn)n(X_{n})_{n\in{\mathbb{N}}} with (n)(XnAn¯)(\forall n\in{\mathbb{N}})(X_{n}\in\overline{A_{n}}).

  2. (ii)

    Let (An)n(A_{n})_{n\in{\mathbb{N}}} be a sequence of analytic codes such that AnA_{n} is enumerable and An¯\overline{A_{n}} has positive measure. There exists (Xn)n(X_{n})_{n\in{\mathbb{N}}} with (n)(XnAn¯)(\forall n\in{\mathbb{N}})(X_{n}\in\overline{A_{n}}).

  3. (iii)

    Let (An)n(A_{n})_{n\in{\mathbb{N}}} be a sequence of analytic codes such that AnA_{n} is enumerable and An¯\overline{A_{n}} is not enumerable. There exists (Xn)n(X_{n})_{n\in{\mathbb{N}}} with (n)(XnAn¯)(\forall n\in{\mathbb{N}})(X_{n}\in\overline{A_{n}}).

  4. (iv)

    Let (An)n(A_{n})_{n\in{\mathbb{N}}} be a sequence of analytic codes such that for all nn\in{\mathbb{N}}, AnA_{n} is RM-open. There exists (Xn)n(X_{n})_{n\in{\mathbb{N}}} with (n)(XnAn)(\forall n\in{\mathbb{N}})(X_{n}\in{A_{n}}).

Proof.

To prove the items in Σ11\Sigma_{1}^{1}-AC0, apply the latter to (n)(X)[XAn](\forall n\in{\mathbb{N}})(\exists X\subset{\mathbb{N}})[X\in A_{n}], noting that the formula in square brackets is Σ11\Sigma_{1}^{1} if AnA_{n} is an analytic code. To derive weak-Σ11\Sigma_{1}^{1}-AC0 from item (i), let φ\varphi be arithmetical and such that (n)(!X)φ(X,n)(\forall n\in{\mathbb{N}})(\exists!X\subset{\mathbb{N}})\varphi(X,n) and define ‘XAnX\in A_{n}’ as φ(X,n)\varphi(X,n) using [simpson2]*V.1.7. Clearly, XAn¯X\in\overline{A_{n}} then implies φ(X,n)\varphi(X,n) as AnA_{n} codes a singleton, i.e. item (i) implies weak-Σ11\Sigma_{1}^{1}-AC0. To obtain weak-Σ11\Sigma_{1}^{1}-AC0 from item (iv), let φ\varphi be as in the antecedent of the former and consider Ψ(X,n,k)\Psi(X,n,k) defined as

(Y)[φ(Y,n)Y¯k=X¯k],(\exists Y\subset{\mathbb{N}})[\varphi(Y,n)\wedge\overline{Y}k=\overline{X}k], (4.2)

which yields a sequence of Σ11\Sigma_{1}^{1}-formulas, yielding in turn a sequence of analytic codes (An,k)n,k(A_{n,k})_{n,k\in{\mathbb{N}}} by [simpson2]*V.1.7. In light of (4.2), An,kA_{n,k} is a basic open ball in 22^{{\mathbb{N}}}. In case Xn,kAn,kX_{n,k}\in A_{n,k} for all n,kn,k\in{\mathbb{N}}, define Yn:=λk.Xn,k¯kY_{n}:=\lambda k.\overline{X_{n,k}}k and note that φ(Yn,n)\varphi(Y_{n},n) for all nn\in{\mathbb{N}}. To obtain weak-Σ11\Sigma_{1}^{1}-AC0 from item (ii), let φ\varphi be as in the antecedent of the former and define Ψ(X,n,k)\Psi(X,n,k) as

(Y)[φ(X¯kY,n)(σ2<)(X=σ00)],(\exists Y\subset{\mathbb{N}})[\varphi(\overline{X}k*Y,n)\wedge(\exists\sigma\in 2^{<{\mathbb{N}}})(X=\sigma*00\dots)],

which yields a sequence of Σ11\Sigma_{1}^{1}-formulas, yielding in turn a sequence of analytic codes (An,k)n,k(A_{n,k})_{n,k\in{\mathbb{N}}} by [simpson2]*V.1.7. For fixed n0n_{0}\in{\mathbb{N}}, there is a unique X0X_{0}\subset{\mathbb{N}} such that φ(X0,n0)\varphi(X_{0},n_{0}), immediately yielding an enumeration of An0,kA_{n_{0},k} for any kk\in{\mathbb{N}}. Essentially by definition, An,k¯\overline{A_{n,k}} has measure 1/2k1/2^{k}. In case Xn,kAn,k¯X_{n,k}\in\overline{A_{n,k}} for all n,kn,k\in{\mathbb{N}}, define Yn:=λk.Xn,k¯kY_{n}:=\lambda k.\overline{X_{n,k}}k and note that φ(Yn,n)\varphi(Y_{n},n) for all nn\in{\mathbb{N}}. Item (iii) also follows as enumerable sets have measure zero. ∎

We would like to formulate item (i) using Borel codes from [simpson2]*V.3, but the latter seem to need ATR0\textup{{ATR}}_{0} to express basic aspects. The items from the theorem also imply finite-Σ11\Sigma_{1}^{1}-AC0, which is left as an exercise.

Finally, we formulate a higher-order result for comparison; we continue the abuse of notation involving Sn¯\overline{S_{n}} as in Theorem 4.2.

Theorem 4.5 (ACA0ω\textup{{ACA}}_{0}^{\omega}).

The higher items imply the lower ones.

  1. (i)

    The principle QF-AC0,1\textup{{QF-AC}}^{0,1}.

  2. (ii)

    Let (Sn)n(S_{n})_{n\in{\mathbb{N}}} be a sequence of sets in [0,1][0,1] such that for all nn\in{\mathbb{N}}, SnS_{n} is enumerable and non-empty. There is (xn)n(x_{n})_{n\in{\mathbb{N}}} with (n)(xnSn¯)(\forall n\in{\mathbb{N}})(x_{n}\in\overline{S_{n}}).

  3. (iii)

    Let (Sn)n(S_{n})_{n\in{\mathbb{N}}} be a sequence of sets in [0,1][0,1] such that for all nn\in{\mathbb{N}}, SnS_{n} is enumerable and Sn¯\overline{S_{n}} has positive measure. There is (xn)n(x_{n})_{n\in{\mathbb{N}}} with (n)(xnSn¯)(\forall n\in{\mathbb{N}})(x_{n}\in\overline{S_{n}}).

  4. (iv)

    cocode1\textup{{cocode}}_{1}

Proof.

The first downward implication follows by applying QF-AC0,1\textup{{QF-AC}}^{0,1} to ‘SnS_{n} is non-empty for all nn\in{\mathbb{N}}’. For the third downward implication, let Y:[0,1]Y:[0,1]\rightarrow{\mathbb{R}} and A[0,1]A\subset[0,1] be such that (n)(!xA)(Y(x)=n)(\forall n\in{\mathbb{N}})(\exists!x\in A)(Y(x)=n). Define the set

En,k:={x[n,n+1]:(q)(Y(xn+q)=nxn+qA|q|12k+1)}\textstyle E_{n,k}:=\{x\in[n,n+1]:(\exists q\in{\mathbb{Q}})(Y(x-n+q)=n\wedge x-n+q\in A\wedge|q|\leq\frac{1}{2^{k+1}})\}

and note that this sequence has a straightforward enumeration while the associated separably closed set has measure 12k\frac{1}{2^{k}}. Let (xn,k)n,k(x_{n,k})_{n,k\in{\mathbb{N}}} be the sequence provided by item (iii). Using sequential compactness, yn=limkxn,ky_{n}=\lim_{k\rightarrow\infty}x_{n,k} is a real in [n,n+1][n,n+1] satisfying Y(yn)=nY(y_{n})=n, for any nn\in{\mathbb{N}} as required. ∎

Variations of the previous theorem are possible, e.g. replacing ‘enumerable’ by ‘(strongly) countable’. Nonetheless, we are not able to derive e.g. cocode1\textup{{cocode}}_{1} from ClC restricted to closed sets of positive measure, i.e. the previous two theorems may well be due to the coding of closed sets as in Definition 4.3.

4.2. Unordered sums

The notion of unordered sum is a device for bestowing meaning upon sums involving uncountable index sets. We introduce the relevant definitions and then prove that basic properties of unordered sums exist in the range of hyperarithmetical analysis.

First of all, unordered sums are essentially ‘uncountable sums’ xIf(x)\sum_{x\in I}f(x) for any index set II and f:If:I\rightarrow{\mathbb{R}}. A central result is that if xIf(x)\sum_{x\in I}f(x) somehow exists, it must be a ‘normal’ series of the form if(yi)\sum_{i\in{\mathbb{N}}}f(y_{i}), i.e. f(x)=0f(x)=0 for all but countably many x[0,1]x\in[0,1]; Tao mentions this theorem in [taomes]*p. xii.

By way of motivation, there is considerable historical and conceptual interest in this topic: Kelley notes in [ooskelly]*p. 64 that E.H. Moore’s study of unordered sums in [moorelimit2] led to the concept of net with his student H.L. Smith ([moorsmidje]). Unordered sums can be found in (self-proclaimed) basic or applied textbooks ([hunterapp, sohrab]) and can be used to develop measure theory ([ooskelly]*p. 79). Moreover, Tukey shows in [tukey1] that topology can be developed using phalanxes, which are nets with the same index sets as unordered sums.

Secondly, as to notations, unordered sums are just a special kind of net and a:[0,1]a:[0,1]\rightarrow{\mathbb{R}} is therefore written (ax)x[0,1](a_{x})_{x\in[0,1]} in this context to suggest the connection to nets. The associated notation x[0,1]ax\sum_{x\in[0,1]}a_{x} is purely symbolic. We only need the following notions in the below. Let fin()\textup{{fin}}({\mathbb{R}}) be the set of all finite sequences of reals without repetitions.

Definition 4.6.

Let a:[0,1]a:[0,1]\rightarrow{\mathbb{R}} be any mapping, also denoted (ax)x[0,1](a_{x})_{x\in[0,1]}.

  • We say that (ax)x[0,1](a_{x})_{x\in[0,1]} is convergent to aa\in{\mathbb{R}} if for all kk\in{\mathbb{N}}, there is Ifin()I\in\textup{{fin}}({{\mathbb{R}}}) such that for Jfin()J\in\textup{{fin}}({{\mathbb{R}}}) with IJI\subseteq J, we have |axJax|<12k|a-\sum_{x\in J}a_{x}|<\frac{1}{2^{k}}.

  • A modulus of convergence is any sequence Φ01\Phi^{0\rightarrow 1^{*}} such that Φ(k)=I\Phi(k)=I for all kk\in{\mathbb{N}} in the previous item.

For simplicity, we focus on positive unordered sums, i.e. (ax)x[0,1](a_{x})_{x\in[0,1]} such that ax0a_{x}\geq 0 for x[0,1]x\in[0,1].

Thirdly, we establish that basic properties of unordered sums exist in the range of hyperarithmetical analysis.

Theorem 4.7 (ACA0ω+IND1\textup{{ACA}}_{0}^{\omega}+\textup{{IND}}_{1}).

The higher items imply the lower ones.

  1. (i)

    QF-AC0,1\textup{{QF-AC}}^{0,1}.

  2. (ii)

    For a positive and convergent unordered sum x[0,1]ax\sum_{x\in[0,1]}a_{x}, there is a sequence (yn)n(y_{n})_{n\in{\mathbb{N}}} of reals such that ay=0a_{y}=0 for all yy not in this sequence.

  3. (iii)

    For a positive and convergent unordered sum x[0,1]ax\sum_{x\in[0,1]}a_{x}, there is a modulus of convergence.

  4. (iv)

    cocode1\textup{{cocode}}_{1}.

Proof.

Assume QF-AC0,1\textup{{QF-AC}}^{0,1} and note that the convergence of an unordered sum to some aa\in{\mathbb{R}} implies

(k)(Ifin())(|axIax|<12k).\textstyle(\forall k\in{\mathbb{N}}\textup{)}(\exists I\in\textup{{fin}}({\mathbb{R}}))\big{(}|a-\sum_{x\in I}a_{x}|<\frac{1}{2^{k}}\big{)}. (4.3)

Apply QF-AC0,1\textup{{QF-AC}}^{0,1} to (4.3) to obtain a sequence (In)n(I_{n})_{n\in{\mathbb{N}}} of finite sequences of reals. This sequence must contain all yy\in{\mathbb{R}} such that ay0a_{y}\neq 0. Indeed, suppose y0y_{0}\in{\mathbb{R}} satisfies ay0>12k0a_{y_{0}}>_{{\mathbb{R}}}\frac{1}{2^{k_{0}}} for fixed k0k_{0}\in{\mathbb{N}} and y0y_{0} is not included in (In)n(I_{n})_{n\in{\mathbb{N}}}. By definition, Ik0+2I_{k_{0}+2} satisfies |axIk0+2ax|<12k0+2|a-\sum_{x\in I_{k_{0}+2}}a_{x}|<\frac{1}{2^{k_{0}+2}}. However, for J=Ik0+2{y0}J=I_{k_{0}+2}\cup\{y_{0}\}, we have aJ>aa_{J}>a, a contradiction. Hence, QF-AC0,1\textup{{QF-AC}}^{0,1} implies item (ii). The second and third items are readily seen to be equivalent.

For the final downward application, let A[0,1]A\subset[0,1] and Y:[0,1]Y:[0,1]\rightarrow{\mathbb{R}} be such that the latter is injective and surjective on the former. Define ax:=12Y(x)+1a_{x}:=\frac{1}{2^{Y(x)+1}} if xAx\in A, and 0 otherwise. One readily proves that x[0,1]ax\sum_{x\in[0,1]}a_{x} is convergent to 11, for which IND1\textup{{IND}}_{1} is needed. The sequence from the second item now yields the enumeration of the set AA required by cocode1\textup{{cocode}}_{1}. ∎

We note that height-Σ11\Sigma_{1}^{1}-AC0 can be obtained from item (ii) in Theorem 4.7; we conjecture that finite-Σ11\Sigma_{1}^{1}-AC0 cannot be obtained. Since unordered sums are just nets, one could study statements like

a convergent net has a convergent sub-sequence,

which for index sets defined over Baire space is equivalent to QF-AC0,1\textup{{QF-AC}}^{0,1} ([samhabil]).

4.3. Variations and generalisations

We discuss variations and generalisations of the above results.

First of all, many variations of the results in Section 3.2 exist for rectifiable functions. Now, Jordan proves in [jordel3]*§105 that BVBV-functions are exactly those for which the notion of ‘length of the graph of the function’ makes sense. In particular, fBVf\in BV if and only if the ‘length of the graph of ff’, defined as follows:

L(f,[0,1]):=sup0=t0<t1<<tm=1i=0m1(titi+1)2+(f(ti)f(ti+1))2\textstyle L(f,[0,1]):=\sup_{0=t_{0}<t_{1}<\dots<t_{m}=1}\sum_{i=0}^{m-1}\sqrt{(t_{i}-t_{i+1})^{2}+(f(t_{i})-f(t_{i+1}))^{2}} (4.4)

exists and is finite by [voordedorst]*Thm. 3.28.(c). In case the supremum in (4.4) exists (and is finite), ff is also called rectifiable. Rectifiable curves predate BVBV-functions: in [scheeffer]*§1-2, it is claimed that (4.4) is essentially equivalent to Duhamel’s 1866 approach from [duhamel]*Ch. VI. Around 1833, Dirksen, the PhD supervisor of Jacobi and Heine, already provides a definition of arc length that is (very) similar to (4.4) (see [dirksen]*§2, p. 128), but with some conceptual problems as discussed in [coolitman]*§3.

Secondly, regulated functions are not necessarily BVBV but have bounded Waterman variation W01(f)W_{0}^{1}(f) (see [voordedorst]), which is a generalisation of BVBV where the sum in (3.1) is weighted by a Waterman sequence, which is a sequence of positive reals that converges to zero and with a divergent series. Some of the above results generalise to regulated function for which the Waterman variation is known, say W01(f)=1W_{0}^{1}(f)=1.

Thirdly, one can replace the consequent of item (iii) in Theorem 3.5 by a number of similar conditions, like the existence of a Baire 1 representation (which can be defined in ACA0ω\textup{{ACA}}_{0}^{\omega} for monotone functions), the fundamental theorem of calculus at all reals but a given sequence, or the condition that if the Riemann integral of f:[0,1][0,1]f:[0,1]\rightarrow[0,1] in BVBV is zero, f(x)=0f(x)=0 for all x[0,1]x\in[0,1] but a given sequence. Many similar conditions may be found in [samBIG, samBIG3, samBIG4].

Fourth, Theorem 3.7 is readily generalised to (almost) arbitrary functions on the reals. To make sure the resulting theorem is provable in ACA0ω+QF-AC0,1\textup{{ACA}}_{0}^{\omega}+\textup{{QF-AC}}^{0,1}, it seems we need oscillation functions777For any f:f:{\mathbb{R}}\rightarrow{\mathbb{R}}, the associated oscillation functions are defined as follows: oscf([a,b]):=supx[a,b]f(x)infx[a,b]f(x)\textup{{osc}}_{f}([a,b]):=\sup_{{x\in[a,b]}}f(x)-\inf_{{x\in[a,b]}}f(x) and oscf(x):=limkoscf(B(x,12k)).\textup{{osc}}_{f}(x):=\lim_{k\rightarrow\infty}\textup{{osc}}_{f}(B(x,\frac{1}{2^{k}})).. Riemann, Ascoli, and Hankel already considered the notion of oscillation in the study of Riemann integration ([hankelwoot, rieal, ascoli1]), i.e. there is ample historical precedent. In the same way as for Theorem 3.7, one proves that the higher items imply the lower ones over ACA0ω\textup{{ACA}}_{0}^{\omega}.

  • The principle QF-AC0,1\textup{{QF-AC}}^{0,1}.

  • Any infinite set X[0,1]X\subset[0,1] has a limit point.

  • For any f:[0,1]f:[0,1]\rightarrow{\mathbb{R}} with oscillation function oscf:[0,1]\textup{{osc}}_{f}:[0,1]\rightarrow{\mathbb{R}}, the set Df={x[0,1]:oscf(x)>0}D_{f}=\{x\in[0,1]:\textup{{osc}}_{f}(x)>0\} is either finite or has a limit point.

  • For a non-piecewise continuous f:[0,1]f:[0,1]\rightarrow{\mathbb{R}} with oscillation function oscf:[0,1]\textup{{osc}}_{f}:[0,1]\rightarrow{\mathbb{R}}, the set Df={x[0,1]:oscf(x)>0}D_{f}=\{x\in[0,1]:\textup{{osc}}_{f}(x)>0\} has a limit point.

  • The arithmetical Bolzano-Weierstrass theorem ABW0\textup{{ABW}}_{0} ([coniving]).

We note that oscf:[0,1]\textup{{osc}}_{f}:[0,1]\rightarrow{\mathbb{R}} is necessary to make ‘xDfx\in D_{f}’ into an arithmetical formula while ‘xx is a limit point of DfD_{f}’ is a meaningful (non-arithmetical) formula even if DfD_{f} does not exist as a set.

Acknowledgement 4.8.

Our research was supported by the Klaus Tschira Boost Fund via the grant Projekt KT43. The initial ideas for this paper, esp. Section 2, were developed in my 2022 Habilitation thesis at TU Darmstadt ([samhabil]) under the guidance of Ulrich Kohlenbach. The main ideas of this paper came to the fore during the Trends in Proof Theory workshop in February 2024 at TU Vienna. We express our gratitude towards all above persons and institutions.

References

  • \bibselectallkeida