On the foundations of constructive mathematics

— especially in relation to the theory of continuous functions

∗
Frank Waaldijk

July 6, 2003

Abstract
We discuss the foundations of constructive mathematics, including recursive mathematics and
intuitionism, in relation to classical mathematics. There are connections with the foundations
of physics, due to the way in which the different branches of mathematics reflect reality. Many
different axioms and their interrelationship are discussed. We show that there is a fundamental
problem in bish (Bishop’s school of constructive mathematics) with regard to its current definition
of ‘continuous function’. This problem is closely related to the definition in bish of ‘locally compact’.

Possible approaches to this problem are discussed. Topology seems to be a key to understanding
many issues. We offer several new simplifying axioms, which can form bridges between the various
branches of constructive mathematics and classical mathematics (‘reuniting the antipodes’). We
give a simplification of basic intuitionistic theory, especially with regard to so-called ‘bar induction’.
We then plead for a limited number of axiomatic systems, which differentiate between the various
branches of mathematics. Finally, in the appendix we offer bish an elegant topological definition
of ‘locally compact’, which unlike the current definition is equivalent to the usual classical and/or
intuitionistic definition in classical and intuitionistic mathematics respectively.

∗ currently employed at Royal Haskoning, Nijmegen, The Netherlands. Email at home: [email protected]

1
1 Introduction

1.1 This paper is concerned with the foundations of constructive mathematics, as the title suggests.
There are connections with the foundations of physics, due to the way in which the different branches of
mathematics reflect reality. As an example, our investigations lead to a physical experiment which might
give statistical evidence whether the physical world is ‘deterministic’ or not. This old debate is put in a
clear-cut mathematical perspective which we feel deserves attention, even though the experiment most
likely will fail to settle the debate.

The main focus of the paper is however of a mathematical nature. The paper discusses many different
axioms and their interrelationship. In the introduction we will already use their abbreviations such as
FT, MP, AC11 , etc. No worry however: all these abbreviations will be treated in detail in the rest of
the paper. Finally some recommendations are made for a constructive framework, which should form
the real benefit of this paper. We start with a brief historical outline.

1.2 In the beginning of the twentieth century Brouwer (in [Brouwer1907]) put forward a sharp foun-
dational critique of classical mathematics (class). Having made a name for himself in the new field
of topology through a number of impressive results, Brouwer turned to building a new mathematics:
intuitionism (int). Though at this time heard by many, his views were not largely accepted, not in
the least because his critique seemingly left most of classical mathematics in ashes. The outcome of
Brouwer’s dispute with Hilbert illustrates that even mathematics is no objective, but a subjective and
also very social affair. Most mathematicians continue to follow Hilbert’s classical views, but without
a proper understanding of Brouwer’s critique. Classical mathematics is mainstream, and taught for
instance in high school. Intuitionism is an oddity, seen mostly as useful in logic and informatics, and
hardly anyone encounters it.

Yet a number of well-known classical mathematicians supported Brouwer throughout the twentieth
century. For this paper we especially name Kleene, who became famous for his work on recursion theory,
and Bishop, a well-known analyst. Kleene wrote a book together with Vesley ([Kleene&Vesley1965])
called ‘The Foundations of Intuitionistic Mathematics — especially in relation to the theory of recursive
functions’. In this book Kleene presents Brouwer’s insights in a clear and straightforward way, and
axiomatizes them. Then he shows that Brouwer’s framework is consistent. For this Kleene uses the
theory of recursive functions and a notion called ‘realizability’. In passing, Kleene also shows that
Brouwer’s fan theorem FT is inconsistent with the assumption that all sequences of natural numbers
are given by a recursive rule. This assumption is called ‘Church’s thesis’ (CT). CT is adopted together
with Brouwer’s intuitionistic logic by a group of mainly Russian mathematicians (notably Markov), to
form a branch of constructive mathematics known as russ.

In logic and later on informatics the impact of Brouwer’s intuitionism was profoundly felt. But notwith-
standing Kleene’s and other people’s work, Brouwer’s impact on mainstream mathematics seemed to
dwindle. Then in [Bishop1967] Bishop introduced what we will call bish: a variety of constructive math-
ematics which adopts much of Brouwer’s intuitionistic views, but at the same time rejects Brouwer’s
most striking insights such as the fan theorem FT and the axiom of continuous choice AC11 .

Bishop wanted to use largely the same terminology as classical mathematics, with the same primarily
mathematical focus. In this way more mathematicians could understand constructive mathematics and

2
be attracted to it. Bishop viewed the axioms and terminology of intuitionism as mystic (notwithstanding
the clarifications offered in [Kleene&Vesley1965], which although phrased in ‘logical language’ form a
solid axiomatic foundation for int)1 . He preferred to first develop in a constructive way as much of
classical analysis as possible, before thinking things through foundationally. Out of his work, Richman
and Bridges developed a view of bish as a mathematics, the methods and results of which are acceptable
in class, int and russ alike. Hereby they disregard the philosophical differences, which give very
different interpretations of the objects and the universe involved.

Bishop’s approach was quite successful, and attracted a respectable number of mathematicians. This
gave new life to constructive mathematics. We will show in this paper however that bish, contrary
to its claim, does not take up a neutral position with respect to the antipodes class, int and russ.
Also we are led to investigate several mathematical and meta-mathematical issues which impact on the
foundations of int, russ, class, and bish. One conclusion that we like to draw right here is that
topology (Brouwer’s specialty) provides a fundamental key for understanding these issues. Topology
has not received much attention yet in constructive mathematics, although powerful topological tools
from classical mathematics can also be developed in bish and int.2

1.3 Bishop maintained that in the context of analysis one could do without FT and axioms of continu-
ity by restricting one’s attention to continuous functions known to be uniformly continuous on compact
subspaces of their domain. Only these functions deserved his predicate ‘continuous’, other functions if
existent did not merit much attention. It therefore seems that intuitively at least Bishop agreed with
Brouwer’s analysis which led Brouwer to put forward FT. One can show that the role of this axiom is
nothing more than to bring about the same restriction on continuous functions.

For in this paper we show that the existence of any class of ‘continuous functions’ satisfying the above
property ‘uniformly continuous on compact subspaces’ and a rather minimal list of other conditions is
equivalent within bish to the fan theorem FT. This throws a fundamental light on both bish and FT,
which has not received the attention it deserves in the constructive literature. The main theorem in
this respect is:

Theorem: Within bish the following statements are equivalent:

(1) The fan theorem FT.

(2) There exists a class of real-valued functions called ‘kontinuous’ functions such that:

(a) If f is a uniformly continuous real-valued function defined on [0, 1] , then f is kontinuous.

(b) If f and g are kontinuous functions such that Ran(f ) ⊆Dom(g) , then the composition g ◦f
is kontinuous.
(c) If f is a kontinuous function defined on [0, 1] , then f is uniformly continuous.
1
(d) The function x −→ x , defined on R+ , is kontinuous.

From this theorem we derive a result concerning the two different definitions of ‘continuous function’
which are currently being used in bish, namely the definition in [Bishop1967] (repeated in [Bishop&Bridges1985])
1 Bishop did mention [Kleene&Vesley1965] in his paper ‘Schizophrenia in Contemporary Mathematics’, [Bishop1985],
as well as in its list of references.
2 See [Waaldijk1996].

3
and the definition in [Bridges1979]. The latter is an attempt to repair the [Bishop1967] definition.3 To
formulate this result let us call the first definition ‘continuousBIS ’ and the second definition ‘continuousBRI ’.

corollary: Within bish the following three statements are equivalent:

(1) The fan theorem FT.

1
(2) If f is a continuousBIS function from [0, 1] to R+ , then the composition x −→ f (x) of f with
the continuousBIS function x −→ x1 is continuousBIS .
1
(3) The function x −→ x , defined on (R+ , dR ) , is continuousBRI .

remark: The following statement (more general than (2) above) therefore implies FT:

(2’) If f and g are real-valued continuousBIS functions such that Ran(f ) ⊆Dom(g) , then the compo-
sition g ◦f is continuousBIS .

However, we do not see how to prove (2’) using only FT. There is a proof of (2’) using FT and a weak,
classically true, version of AC11 (see 8.8).

1.4 Further on we study theorem 1.3 and its relation with the foundations of bish. We look at the
consequences of omitting one or more conditions on ‘continuous function’ from the list (2)(a) through
(2)(d) in the above theorem.

1.5 As a side product of our study, we obtain within the framework of russ a counterexample to FT
which is in essence the same as the counterexample in [Kleene&Vesley1965], but with a different flavour:
we present a recursive element of the Hilbert cube which is recursively apart from every recursive binary
element of the Hilbert cube (a binary element has a binary expansion in each coordinate). Recursive
mathematics (russ) plays an important role in our discussion. russ takes the viewpoint that any
infinite sequence must be given by a finite algorithm (recursive rule) which produces the sequence step
by step4 . If applied to the physical world, the viewpoint of russ in the author’s eyes corresponds to
the philosophical standpoint of our world being deterministic. We will elucidate this in section 7.

Early in his intuitionistic career Brouwer maintained similar ideas, but later on he dropped the algorith-
mic requirement and stated that infinite sequences can be produced step by step, without any rule for
and any prior knowledge of what the next steps might bring. Brouwer based his step-by-step creative
sequences on the human intuition of infinite time. Of course one can restrict oneself beforehand to
follow a recursive rule, but according to Brouwer this is no necessity. Therefore in Brouwer’s universe
one has in the strongest sense possible what we will call ‘incomplete information’5 . If applied to the
3 These definitions are not equivalent, yet so far in constructive papers no explicit reference is made to this duality (can

results from [Bridges1979] be used when the definitions come from [Bishop&Bridges1985]?).
4 Thinking things through, the author does not see how to convince himself that a certain sequence has infinite potential,

unless we produce precisely such a finite algorithm and show that it will produce a result for every n ∈N . What guarantee
do we have that other sequences, for instance from reality, are potentially infinite? What if this world comes to an end?
Of course then any algorithm would also stop, but at least one has the feeling that the algorithms as developed by Church
and Turing will be the same in any possible other world. Like an encapsulated virus an algorithmic infinite sequence can
be turned on at any time, in any situation, to produce the same outcome.
5 However, one also has incomplete information in russ, since it is impossible to algorithmically determine for an

arbitrary algorithm whether it will produce in fact an infinite sequence, or whether it fails on a given input.

4
physical world, the viewpoint of both int and class in the author’s eyes implies that our world is not
deterministic, see section 7.

The failure of FT in russ can be ascribed to the extra information that one has about elements of the
mathematical universe N N , namely that every such element is produced by a finite algorithm. In russ
each sequence is even produced by a known finite algorithm. However, reality comes to us in a different
way. Many natural phenomena change with time, and therefore present us with sequences of numbers
for which we have a priori no algorithm. This does not prove that such an algorithm ultimately doesn’t
exist. Who knows what the physicists or the theologians may come up with? But so far in most cases
we cannot pinpoint any such algorithm.

So we have basically two possible ways to model our real world. The first way is to assume that it is
non-deterministic, and then both int and class are remaining. The difference between these then is
the difference between a world of incomplete information (int) and a world of complete information
(class). Interpreted in this way it is clear that class leads to an unrealistic model, since in real life we
deal with many situations where we do not have complete information. So the principle of the excluded
middle (PEM, ‘For all A: A is true or A is false’) often doesn’t apply since we have no method to
tell which of the extremes is true. Brouwer very sharply demonstrated this mathematically, and also
identified the cause of the problems: the intuition behind PEM comes from finite decidable situations,
but this intuition cannot be transferred outside this context.

The second way to model our real world is to assume that it is deterministic. Then both russ and
classical recursion theory remain applicable. Classical recursion theory is again a world of complete
information. But russ is in between ‘complete information’ and ‘incomplete information’. It would
be worthwile to explore the consequences of a deterministic world with incomplete information (since
under the assumption of determinacy in the author’s eyes this comes closest to real life). That is a world
in which each infinite sequence is given by an algorithm, which in most cases is completely unknown.
We can model such a world by introducing two players, where player I picks algorithms and hands
out the computed values of these algorithms to player II, one at a time. Sometimes player I discloses
(partial) information about the algorithms themselves. Player II can of course construct her or his own
algorithms, but still is confronted with recursive elements of player I about which she/he has incomplete
information. One is tempted to think that in such a world intuitionistic principles might hold. We
believe CP will be consistent. But it is not difficult to show that FT remains invalid in such a world.

There is a surprising conclusion to this line of thought. It is possible to conduct a physical experiment
which might give statistical evidence whether the real world is deterministic or not (in the interpretation
of determinism as preferred by the author, see section 7). We believe that the actual carrying out of
this experiment can be of relevance to both mathematical foundations as to physics and philosophy in
general. The experiment is based on the way in which FT either fails or holds, given the determinacy
or non-determinacy assumption.

1.6 Our investigations lead us to consider the possibilities for ‘reuniting the antipodes’6 . The antipodes
here being classical mathematics (class) and intuitionism (int). From a purely axiomatic point of
view, they have far more in common than the current scope of bish, and they are much closer to each
other than to russ. As stated, they also share the non-determinacy of the universe as a philosophical
insight. It therefore seems worthwile to explore the ‘formal’ common ground of classical and intuitionistic
6 We gladly adapt the terminology of Peter Schuster to this setting.

5
mathematics. If systematically developed, many intuitionistic results would be seen to hold classically
as well, and thus would offer a way to develop a strong constructive theory which is still consistent
with the rest of classical mathematics. Such a constructive theory can form a conceptual framework for
applied mathematics, and information technology. These sciences now use an ad-hoc approach to reality,
since the classical framework is inadequate. This ad-hoc approach leads to inefficiency in methods used
and in research conducted. Bishop’s motivation to develop bish was similar. But int is stronger, and
comes closer to class.

The intuitionist can on the other hand more easily use the richness of ideas already present in classical
mathematics, if classical mathematics were to be systematically developed along the common grounds
before the unconstructive elements are brought in.7

One can also investigate the ‘formal’ common ground of all three varieties class, int and russ. This
common ground is much less than that of class and int, but still is larger than what is offered in bish.
We will indicate several nice theorems belonging to the common ground of all three varieties. 8

One of the conclusions is that basic topology is too often neglected in constructivism. Many approaches
and results in the constructive literature can be obtained in a simpler way with the help of basic topol-
ogy9 . Especially from the standpoint of topology, the current bish definition of ‘continuous function’
merits revision, and the same holds for the current bish-definition of ‘locally compact’.

Finally the paper offers some simplification of basic intuitionistic theory regarding bars and bar in-
duction. A further strengthening of intuitionism can be obtained by adopting Markov’s Principle MP
(true in class and russ). The author believes that there is very strong intuitive support for MP, and
intuitionism will benefit from the simplifications that MP brings about.

2 Basic Axioms and Definitions

2.1 In this section we present some basic axioms described in the literature, pertaining to constructive
mathematics in particular (int and russ). To define the relevant intuitionistic axioms of (continuous)
choice we need a number of straightforward definitions.

definition: Let σω denote the universal spread of all infinite sequences of natural numbers ( σω =N N ).
On σω the natural metric dω is defined by putting, for α and β in σω : dω (α, β) = sup({2−n | α(n) =
β(n) | n ∈N } . Write σ ω for the set of finite sequences of natural numbers (often written like this:
7 However, much of classical mathematics will remain an exercise in charting the consequences of the unrealistic principle

of the excluded middle. An intuitionist will be hard put to understand that this is a worthwile pursuit.
8 Also intuitionism and recursive mathematics share a noteworthy topological insight: in a complete metric space the

so-called ‘apartness topology’ coincides with the metric topology; which explains the metric continuity of everywhere-
defined functions in a more fundamental way than the usual approach. For russ, this insight relies partly on Markov’s
Principle MP which is classically true. We will plead for MP also in the context of int.
9 For many results on complete metric spaces (X, d) it suffices to prove the result for (σ , d ) , and then consider
ω ω
a continuous surjection from (σω , dω ) to (X, d) , ‘pulling back’ the topology. The same holds, mutatis mutandis, for
compact spaces and (σ2 , dω ) . Also consider so-called ‘partitions of unity’ (see [Waaldijk1996]). With the help of this
tool one proves the Dugundji extension theorem, rather than the weaker Tietze extension theorem. Also see the previous
footnote. There is a large list of similar simplifications.

6
 σ ω =N∗ ). For α in σω we write α(n) for the finite sequence α(0), . . . , α(n −1) formed by the first
n values of α . Then α(n) ∈σ ω , and vice versa σ ω ={α(n) |α ∈σω , n ∈N} . A subset B of σ ω is
called decidable iff ∀a ∈σ ω [ a ∈B ∨ a ∈ B ] . A subset B of σ ω is a bar on a subset A of σω iff
∀α ∈A ∃n ∈N [ α(n) ∈B ] . Notice that a bar on A is the same as an open cover of A consisting of
basic open sets (in the product topology).

Now let a be in σ ω , in other words a is a finite sequence of natural numbers. Then we write lg(a) for
the length of this finite sequence. So if a =a0 , . . . , an−1 then lg(a) =n . There is a sequence of length
0 , namely the empty sequence denoted by < > . For i < lg(a) we then write ai for the ith element of
this finite sequence. If a =a0 , a1 , . . . , alg(a)−1 and b =b0 , b1 , . . . , blg(b)−1 are in σ ω then we write a b
for the concatenation a0 , a1 , . . . , alg(a)−1 , b0 , b1 , . . . , blg(b)−1 of a and b . We write a  b iff there is a
c in σ ω such that b =a c , and we write a < b iff in addition lg(b) > lg(a) .

Likewise for α in σω and a in σ ω we write a α for the concatenation of a and α , which is the
element β of σω given by β(lg(a)) =a and for all n ∈N : β(lg(a) +n) =α(n) . For each a in σ ω we
define a canonical element αa of σω by putting: αa =a 0 where 0 =0, 0, 0, . . . . We write σω ∩a for
the set {α ∈σω |α(lg(a)) =a} of sequences ‘running through a ’.

A function f from σω to N is called a spread-function iff there is a function g from σ ω to N such that

for each α in σω : ∃!n ∈N [g(α(n)) > 0] and moreover for all n ∈N : g(α(n)) > 0 → f (α) =g(α(n)) −1 10 .
More generally a function f from σω to σω is called a spread-function iff there is a function g from
σ ω to σ ω such that for each α in σω and n ∈N there is an m ∈N such that: f (α)(n) =g(α(m)) ,
and moreover g(α(n))  g(α(n +1)) . In both definitions we say that g represents f . (Notice that a
spread-function is continuous with respect to dω .).

The fundamental intuitionistic axiom of continuous choice AC11 can now be formulated as follows:

AC11 Let A be a subset of σω ×σω such that:

(
) ∀α ∈σω ∃β ∈σω [ (α, β) ∈A ]

Then there is a spread-function γ from σω to σω such that for each α in σω : (α, γ(α)) is in A . We
say that γ fulfills (
).

We formulate four weaker versions of this axiom: AC10 , CP, AC01 , and AC00 . The last two are
simple axioms of countable choice, whereas AC10 is still an axiom of continuous choice, also known as
‘Brouwer’s principle for numbers’. AC10 implies the so-called continuity principle CP. In section 8 we
will indicate two, both classically and intuitionistically valid axioms called BDD and AC11weak , such
that AC11 is equivalent to the combination of CP, BDD and AC11weak . Therefore, CP may be held
solely responsible for intuitionistic results which are not acceptable when interpreted classically1112 .
Also we ask attention for axioms of dependent choice. We do not explain or defend the axioms in this
section since they are broadly discussed in the literature (they can all be found in [Kleene&Vesley1965],
[Gielen et al. 1981], [Veldman1981] and [Troelstra&vanDalen1988]). We begin with the weaker axioms
dealing with continuous choice:
10 Notice that {a ∈ σ |g(a) > 0} is a decidable thin bar in the sense of the second definition in 8.8. This shows that the
ω
concept of spread-function is inherently the same as the concept of a decidable (thin) bar.
11 We disregard the so-called Brouwer-Kripke axiom, feeling it is not part of mainstream intuitionism.
12 In a follow-up paper we intend to give a simple classical game-theoretic setting in which all of the intuitionistic axioms

hold true. This should provide classical mathematicians with an intuitive understanding of what intuitionism is about.

7
 AC10 Let A be a subset of σω ×N such that:

(

) ∀α ∈σω ∃n ∈N [ (α, n) ∈A ]

Then there is a spread-function γ from σω to N such that for each α in σω : (α, γ(α)) is in A . We
say that γ fulfills (

).

CP Let A be a subset of σω ×N such that:

∀α ∈σω ∃n ∈N [ (α, n) ∈A ]

Then: ∀α ∈σω ∃n ∈N ∃m ∈N ∀β ∈σω [ α(m) =β(m) → (β, n) ∈A ] .

We present two simple axioms of countable choice in decreasing order of strength:

AC01 Let A be a subset of N ×σω such that:

(∗) ∀n ∈N ∃α ∈σω [ (n, α) ∈A ]

Then there is a function h from N to σω such that for each n ∈N : (n, h(n)) is in A . We say that h
fulfills (∗).

AC00 Let A be a subset of N ×N such that:

(∗∗) ∀n ∈N ∃m ∈N [ (n, m) ∈A ]

Then there is a function h from N to N such that for each n ∈N : (n, h(n)) is in A . We say that h
fulfills (∗∗).

2.2 Finally we present two axioms of dependent choice in decreasing order of strength. For an
intuitionistic justification of these axioms we refer the reader to [Waaldijk1996].

DC1 Let δ be in σω , and let A be a subset of σω . Suppose R is a subset of A ×A such that:

δ ∈A ∧ ∀α ∈A ∃β ∈A [ (α, β) ∈R ]

Then there is a sequence (γn )n∈N of elements of σω such that γ0 =δ and for each n ∈N : (γn , γn+1 )
is in R .

DC0 Let s ∈N , and let A be a subset of N . Suppose R is a subset of A ×A such that:

s ∈A ∧ ∀n ∈A ∃m ∈A [ (n, m) ∈R ]

Then there is an α in σω such that α(0) =s and for each n ∈N : (α(n), α(n +1)) is in R .

2.3 It is now but a small step to formulate the principle of Bar Induction for Decidable bars: BID .
A simple definition:

definition: A subset A of σ ω is called inductive iff for all a in σ ω : ∀n ∈N [a n ∈A] → a ∈A .

BID Let B be a decidable bar on σω . Suppose A is an inductive subset of σ ω such that B ⊆A .

Then the empty sequence < > (of length 0 ) is in A .

8
remark: In classical mathematics BID can be derived from the principle of the excluded middle. The
above version of the bar theorem is therefore classically true. In section 8 we derive BID from an axiom
called BT (‘Brouwer’s Thesis’) which also holds both in class and int.

One of the results following from BID is the axiom known as the fan theorem FT. We need a preliminary
definition.

definition: Let σ2 denote the binary fan ( σ2 ={0, 1}N ). Write σ2 for the set of finite sequences
of elements of {0, 1} (often written like this: σ2 ={0, 1}∗ ). Then σ2 ={α(n) |α ∈σ2 , n ∈N} . A sub-
set B of σ2 is called decidable iff ∀a ∈σ2 [ a ∈B ∨ a ∈ B ] . A subset B of σ2 is a bar on σ2 iff
∀α ∈σ2 ∃n ∈N [ α(n) ∈B ] .

FT If B is a decidable bar on σ2 , then B contains a finite bar on σ2 (in other words: then
∃n ∈N ∀α ∈σ2 ∃m < n [ α(m) ∈B ] ).

2.4 The basic axiom in russ is of course Church’s Thesis: ‘every sequence of natural numbers is given
by a recursive rule’ (many results in russ already follow from the weaker statement: ‘the set of partial
functions from N to N is countable’). A partial recursive function α from N to N will usually be
denoted by something like ‘ φe ’ where the natural number e is the recursive index of α . This recursive
index is nothing but the encoding of the finite algorithm which for each n ∈N tries to compute α(n) .
There is a decidable subset I(N, N) of N such that each e in I(N, N) is a properly formed recursive
index of a partial recursive function from N to N , and vice versa for each partial recursive function α
from N to N there is an e in I(N, N) such that ∀n ∈N [ α(n) = • ’ stands for: ‘equal
• φ (n) ] , where ‘ =
e
if one of the algorithms terminates, given the input’.

It turns out one can canonically encode each finite recursive computation as a natural number, see
[Kleene1952]. This is the basis of Kleene’s decidable T -predicate on triples of natural numbers (e, n, k) ,
given by:

T (e, n, k) ⇐⇒ e is a recursive index and k is the canonical encoding of the computation of φe (n) .

In particular if T (e, n, k) then the algorithm φe terminates on the input n . But we are mostly interested
in the result of the computation k , and in its length (the number of canonical subcomputations leading
to the result). Both can be canonically derived from k of course, using recursive funtions Outc and
Lgth . So if T (e, n, k) , then φe (n) =Outc(k) and the length of the computation k equals Lgth(k) .

In this terminology we formulate Church’s thesis thus:

CT ∀α ∈σω ∃e ∈I(N, N) ∀n ∈N [ α(n) =φe (n) ]

If ∀n ∈N [ α(n) =φe (n) ] for α ∈σω and e ∈I(N, N) , then in particular we have: ∀n ∈N ∃m ∈N [ T (e, n, k) ] .
Therefore the set TOT = {e ∈I(N, N) |∀n ∈N ∃k ∈N [ T (e, n, k) ]} plays an important role in russ.

The combination of CT with AC00 is equivalent to an axiom known as CT0 , which illustrates the
connection between CT and choice axioms (cf. 8.13):

CT0 Let A be a subset of N ×N such that:

∀n ∈N ∃m ∈N [ (n, m) ∈A ]

9
 Then there is a recursive function h from N to N such that for each n ∈N : (n, h(n)) is in A .

From CT0 we can derive a more complex choice axiom CT01 , which plays a part in our discussion later
on:

CT01 Let A be any subset of N ×N , and B a decidable subset of N ×N such that:

∀n ∈N [∃y ∈N [(n, y) ∈B] → ∃m ∈N [(n, m) ∈A] ]

Then there is a partial recursive function h from N to N such that for each n ∈N : if ∃y ∈N [(n, y) ∈B]
then h(n) is defined and (n, h(n)) is in A .

lemma: CT01 follows from CT0 .

proof: We have by the assumption in CT01 :

(
) ∀(n, p) ∈N ×N ∃(r, s) ∈N ×N [ (r =0 ∧(n, p) ∈B ∧(n, s) ∈A) ∨ (r =1 ∧(n, p) ∈ B) ]

Now identify N ×N with N and apply CT0 to find a recursive function f from N ×N to N ×N fulfilling
(
). The desired partial recursive h can be defined by describing its working on n : let h(n) be the
second coordinate of f (n, p) where p is the least m ∈N for which the first coordinate of f (n, m) equals
0 •

CT01 is the first step to an even broader choice axiom known as ‘Extended Church’s Thesis’ (ECT0 ),
which is widely accepted in russ. But the formulation of ECT0 and its defense are in logical terms and
do not appeal to the author. We will present a simpler version with an intuitive defense in 8.13, but
first we present ECT0 . A subset A of N is called almost negative iff for each n ∈N the membership
of A is equivalent to the truth of a logical formula A(n) , where A contains the existential quantifier
∃ only directly in front of prime formulas, and furthermore is built only from the logic connectives →
and ∧ and the universal quantifier ∀ . Notice that in particular the set TOT is almost negative.

ECT0 Let A be any subset of N ×N , and B an almost negative subset of N such that:

∀n ∈N [n ∈B → ∃m ∈N [(n, m) ∈A] ]

Then there is a partial recursive function h from N to N such that for each n ∈B we have: h(n) is
defined and (n, h(n)) is in A .

2.5 The second important axiom in russ is called Markov’s Principle: ‘if it is impossible that a total
recursive function α does not achieve the value 1 for some n in N , then there is an n in N with
α(n) =1 ’. Formally:

MP Let α be in σω such that ¬ ¬∃n ∈N [ α(n) =1 ] . Then ∃n ∈N [ α(n) =1 ] .

The motivation in russ is as follows. Existence of an object (‘there is’) in constructive mathematics
always denotes a method to come by this object. Having an algorithm for computing α(n) for each
n ∈N , and knowing it cannot for all n ∈N avoid returning the value 1, we can simply wait until this
algorithm comes up with the desired n . In other words the method here is ‘wait’. This reasoning
seems valid outside the recursive context also, nevertheless Markov’s Principle is not generally accepted

10
in int. The usual reason given for not accepting MP is that MP does not enable one to fix a prior
bound on the number of computations needed. However, this reason can be questioned. Suppose that
to deserve the predicate ‘there is’ one must give a prior bound to the number of computations necessary.
For a recursive α , the complexity of giving a prior bound to the number of computations needed for
computing α(n) is not less than the complexity of α itself. So to show ‘there is’ a prior bound, one
must give a prior bound for the number of computations needed to give a prior bound, and this task
is at least as hard as simply computing α . And then one must give a prior bound for giving the prior
bound for giving the prior bound etc. Therefore giving a prior bound is not less complex than letting
α run by itself.

Also the complexity of natural induction (see Ind in 2.6) is such that in general one cannot indicate
beforehand a bound on the number of computations needed to complete the induction. One can only
show that the complexity is reduced by one step, with each step. So really one in general has no other
method to evaluate an inductive definition or computation than to wait. But when thinking it through,
this is precisely the same as what happens in the waiting process when we let α run. So the intuition
behind Ind is not different from the intuition behind MP.

This extends to the elements of the non-recursive intuitionistic universe just as well. Even though an
element α of σω need not be recursive, the reasoning above still applies. So in the author’s eyes it is
worthwile to consider adopting MP in intuitionism as well.

definition: Since recursion theory also is a part of class and int, we can prove most theorems
concerning russ in a constructive recursion theoretic setting. We therefore use terminology such as
σω REC , σ2 REC , etcetera to indicate the subset of recursive sequences in σω , σ2 , etcetera.

2.6 Finally we present the principle of induction as an axiom. This completes our presentation of
basic axioms. More axioms will be discussed in section 8.

Ind Let A be a subset of N such that 0 ∈A and for all n ∈N : n ∈A implies n +1 ∈A . Then A =N ,
that is: n ∈A for all n ∈N .

3 Basic definitions regarding continuous functions

3.1 We assume the reader to be familiar with the (constructive and/or intuitionistic and/or even
classical) definition of the real numbers, ‘compact metric space’, etcetera. Only those definitions which
can be considered special or specifically relevant will be repeated here. We turn immediately to the
definition of ‘(uniformly) continuous function’. We believe the following to be the most fruitful definition
in the context of metric spaces.

definition: Let f be a function from a metric space (X, d) to another metric space (Y, dY ) . Then f
is continuous iff for all x ∈X and all n ∈N there is an m ∈N such that for all z ∈B(x, 2−m ) we have:
f (z) ∈B(f (x), 2−n ) . We say that f is uniformly continuous iff for all n ∈N there is an m ∈N such
that for all x ∈X and all z ∈B(x, 2−m ) we have: f (z) ∈B(f (x), 2−n ) .

11
Furthermore we will be sloppy with our notation of the well-known metric spaces involved, for instance
writing R+ instead of (R+ , dR ) whenever confusion is unlikely. We do not advocate this in general,
but within the simple scope of this paper mistakes will not occur.

3.2 Now let us give the conditions (2)(a)–(d) in theorem 1.3 a more precise name and place.

definition: We define a number of ‘continuity properties’ which a given class of continuous functions
K may or may not possess:

Cont I If f is a uniformly continuous function, then f is in K

Cont II If f and g are functions in K such that Ran(f ) ⊆Dom(g) , then the composition g ◦f is
in K

Cont III If f is a function in K defined on a compact space, then f is uniformly continuous

1
Cont IV The function x −→ x , defined on R+ , is in K

We will study these properties in section 5. To strengthen our results we will also study slightly weaker
variants of these properties, such as restricting Cont I to uniformly continuous functions defined on
[0, 1] . In theorem 1.3 we also restrict Cont III to functions defined on [0, 1] . Disregarding such subtle
strengthenings theorem 1.3 can be rephrased thus: the existence of a class of functions satisfying Cont
I through Cont IV implies FT, and vice versa.

3.3 Bishop’s 1967 definition of ‘continuous function’ will be called ‘continuousBIS function’ in this
paper. The definition is limited to functions f such that Dom(f ) and Ran(f ) are subspaces of the
real numbers R , and is given directly below. The class of continuousBIS functions is easily seen to satisfy
Cont I, Cont III and Cont IV.

definition: Let f be a real-valued function defined on a subspace X of R . Then f is continuousBIS

iff f is continuous, and uniformly continuous on each compact subspace of X .

Already at its birth it was seen that this definition has a possibly serious drawback: how to prove within
bish that, when defined, the composition of two arbitrary continuousBIS functions is again continuousBIS ?
In other words, does the class of continuousBIS functions satisfy Cont II? Bishop proposed a lemma of
convenience, rather than answering the question with theorem 1.3. It follows from this theorem however
that the lemma doesn’t resolve the fundamental problem. We repeat this lemma for completeness’ sake,
leaving the simple proof to the reader.

lemma: (Bishop 1967) Let f be a real-valued continuousBIS function defined on a subspace X of R .

Let g be a real-valued continuousBIS function defined on a subspace Y of R such that Ran(f ) ⊆Y .
Suppose f maps each compact subspace of X into a compact subspace of Y . Then the composition
g ◦f is again continuousBIS .

Still a second drawback was discovered later: how to extend the [Bishop1967] definition ‘continuousBIS ’
to metric spaces in general, not just (R, dR ) ? For locally compact spaces there is no problem, simply

12
 drop the conditions ‘real-valued’ and ‘subspace of R ’. But if the resulting definition (which we also call
continuityBIS for simplicity) is also used for non-locally-compact spaces, one loses a desirable property in
the process. For given a continuousBIS function f from a metric space (X, d) to another metric space
(Y, dY ) , and given a compact subspace (Z, d) of (X, d) , we do not see how to conclude in general from
the given definition that:

(
) ∀n ∈N ∃m ∈N ∀z ∈Z ∀x ∈X [ d(x, z) < 2−m → dY (f (x), f (z)) < 2−n ]

If one wishes for continuous functions which are uniformly continuous on compact subspaces, then (
)
seems even better, and is true both in classical and intuitionistic mathematics. Actually the pragmatic
solution would be to simply add (
) to the definition.

3.4 An attempt to repair the two drawbacks mentioned above is found in [Bridges1979], where an
alternative definition of ‘continuous function’ (which we call ‘continuousBRI function’ ) is given. When
defined, the composition of two continuousBRI functions is always a continuousBRI function. Every
continuousBRI function is continuousBIS . To understand the [Bridges1979] definition we need a prelimi-
nary.

definition: Let (X, d) be a metric space. Then (X, d) is called a compact image iff (X, d) is the
uniformly-continuous image of the binary fan (σ2 , dω ) (for dω see def. 2.1).

remark: This is easily seen to be equivalent to the more complicated definition in [Bridges1979], if we
keep in mind that every compact space is the uniformly-continuous image of the binary fan (σ2 , dω ) 13 .
The finer constructive view finds a distinction between uniformly-continuous images of (σ2 , dω ) which
are metrically complete (and then called compact) and uniformly-continuous images of (σ2 , dω ) which
are not necessarily complete. Classically all such spaces are compact themselves, and the above definition
is superfluous. Intuitionistically one can prove many a compact image to be metrically incomplete, yet
it is possible to develop an attractive general topological theory, with a non-metrical (but metrizable)
concept of ‘compact’ such that the continuous image of a compact space is again compact. This theory
closely parallels the classical approach, and depends on FT (see [Waaldijk1996]).

3.5 The [Bridges1979] definition now runs like this:

definition: Let f be a function from a metric space (X, d) to another metric space (Y, dY ) . Let
(Z, d) be a subspace of (X, d) . We say that f is uniformly continuous near (Z, d) iff

∀n ∈N ∃m ∈N ∀z ∈Z ∀x ∈X [ d(x, z) < 2−m → dY (f (x), f (z)) < 2−n ]

We say that f is continuousBRI iff f is uniformly continuous near each compact image contained in
(X, d) .

remark: Notice that trivially every continuousBRI function is continuousBIS . So far no general choice
has been made between these two definitions. In [Bishop&Bridges1985] the [Bishop1967] definition is
repeated, likewise in [Bridges&Richman1987]. It would seem that the [Bishop1967] definition is favoured,
but there is no written motivation for this that the author knows of. Meanwhile, the two definitions
13 This topological fact was already proven by Brouwer (see the collected works [Brouwer1975]).

13
 are not equivalent, and a theorem proved using one definition should be checked to see whether it still
holds under the other definition.

One easily sees that the class of continuousBRI functions satisfies Cont I through Cont III. Therefore,
as corollary 1.3 points out, the assertion of Cont IV for the class of continuousBRI functions is equivalent
to FT14 .

4 Proof of theorem 1.3

4.1 The following lemma will play a nice role in this paper.

lemma: Let f be a uniformly continuous function from (σ2 , dω ) to (R, dR ) , then there is a decidable
subset B of σ2 such that:

(i) ∀α ∈σ2 ∀n ∈N [ α(n) ∈B → f (α) > 2−n ]

(ii) ∀α ∈σ2 [ f (α) > 0 → ∃m ∈N [α(m) ∈B] ] .

proof: Since f is uniformly continuous we have:

(
) ∀n ∈N ∃m ∈N ∀α, β ∈σ2 [ dω (α, β) < 2−m → dR (f (α), f (β)) < 2−n ]

By AC00 there is a function u from N to N fulfilling (
). Without loss of generality this function
is such that for all n ∈N : u(n +1) > u(n) ≥ n . For each n ∈N and each α in σ2 at least one of the
two following statements is true: f (α) > 2−n+1 or f (α) < 2−n . Therefore we find (with αa =a 0 see
definition 2.1):

(

) ∀n ∈N ∀a ∈σ2 ∃t ∈{0, 1} [ (t =0 ∧f (αa ) > 2−n ) ∨(t =1 ∧f (αa ) < 2−n+1 ) ]

By AC00 there is a function v from N ×σ2 to {0, 1} fulfilling (

). We are ready to define the desired
decidable B which satisfies the lemma, as follows. Put:

B = {a ∈σ2 |∃s < lg(a) [ v(s, a) =0 ∧ u(s +1) < lg(a) ] }

Clearly B is decidable, it remains to show that ∀α ∈σ2 ∀n ∈N [ α(n) ∈B → f (α) > 2−n−1 ] and
∀α ∈σ2 [ f (α) > 0 → ∃m ∈N [α(m) ∈B] ] . So first let α in σ2 and n ∈N and suppose a =α(n) ∈B . Then
there is s < lg(a) =n such that v(s, a) =0 . Therefore f (αa ) > 2−s and also u(s +1) < lg(a) =n , there-
fore dω (α, αa ) < 2−u(s+1) and so dR (f (α), f (αa )) < 2−s−1 and so f (α) > 2−s−1 ≥ 2−n . Now finally
let α in σ2 be such that f (α) > 0 . Determine s ∈N such that f (α) > 2−s+2 . Determine t =u(s +1)
and put a =α(t +1) , then dω (α, αa ) < 2−u(s+1) and so f (αa ) > 2−s+1 . Therefore v(s, a) =0 . Since
u(s +1) =t < lg(a) we see that a is in B •

As a direct consequence we obtain the following proposition, already proven in [Beeson1985], which
plays a central part in our proof of theorem 1.3:

proposition: Within bish the following statements are equivalent:

14 This is due to the clause ‘each compact image’ in the definition of continuityBRI . Replace this with ‘each compact

subspace’, and one gets an improvement of continuityBIS , losing Cont II in the process.

14
 (1) The fan theorem FT.

(2) If f is a uniformly continuous function from (σ2 , dω ) to (R+ , dR ) , then there is an N ∈N such
that for all α in σ2 : f (α) > 2−N .

proof: To prove (1) ⇒ (2) assume FT and let f be a uniformly continuous function from σ2 to
R+ . By the above lemma there is a decidable bar (!) B on σ2 such that ∀α ∈σ2 ∀n ∈N [ α(n) ∈B →
f (α) > 2−n−1 ] . By FT we have that B contains a finite bar on σ2 , in other words there is an N ∈N
such that ∀α ∈σ2 ∃m < N [ α(m) ∈B ] . Therefore for all α in σ2 we have: f (α) > 2−N .
To prove (2) ⇒ (1) assume (2) and let B be a decidable bar on σ2 . Define a uniformly contin-
uous function f from σ2 to R+ by putting, for α in σ2 : f (α) =2−s where s =µn ∈N [α(n) ∈B]
(‘ µn ∈N ’ means: ‘the smallest n ∈N such that’). Then by (2) there is an N ∈N such that all α in σ2 :
f (α) > 2−N . Clearly this implies that ∀α ∈σ2 ∃m < N [ α(m) ∈B ] , in other words B contains a finite
bar •

4.2 The second observation that we will use is the one-to-one correspondence of σ2 and the Cantor
space C , which is the well-known compact subspace of [0, 1] given by the formula:


C={ 2α(n) ·3−n−1 |α ∈σ2 }
n∈N

The one-to-one correspondence is of course given by the uniformly continuous function jC given by

jC (α) = n∈N 2α(n) ·3−n−1 for α in σ2 , with the uniformly continuous inverse jC−1 .

proposition: Within bish the following statements are equivalent:

(1) The fan theorem FT.

(2) If f is a uniformly continuous function from [0, 1] to R+ , then there is an N ∈N such that for
all x in [0, 1] : f (x) > 2−N .

proof: The implication (1) ⇒ (2) is shown as follows. Assume FT, and let f be a uniformly continuous
function from [0, 1] to R+ . Let g be a uniformly continuous surjection from σ2 to [0, 1] . Then h =f ◦g
is a uniformly continuous function from σ2 to R+ , so by proposition 4.1 there is an N ∈N such that
for all α in σ2 : h(α) > 2−N . We conclude that for all x in [0, 1] : f (x) > 2−N . To prove (2) ⇒ (1)
assume (2) and let B be a decidable bar on σ2 . Define a uniformly continuous function h from σ2 to
R+ by putting, for α in σ2 : h(α) =2−s where s =µn ∈N [α(n) ∈B] . Clearly by the above one-to-one
correspondence between σ2 and C we can define a corresponding uniformly continuous function g
from C to R+ by putting g =h ◦jC−1 . By the Tietze extension theorem (see [Bishop1967]) or simply
by linear interpolation g can be extended to a uniformly continuous function f from [0, 1] to R+ .
Then by (2) there is an N ∈N such that for all x in [0, 1] : f (x) > 2−N . So for all x in C we
have: g(x) > 2−N . Clearly this implies that for all α in σ2 : h(α) > 2−N , which in turn implies that
∀α ∈σ2 ∃m < N [ α(m) ∈B ] , in other words B contains a finite bar •

4.3 We are ready to prove the main theorem 1.3, which we repeat here for the reader’s convenience:

Theorem: (repeated from 1.3) Within bish the following statements are equivalent:
(1) The fan theorem FT.

15
 (2) There exists a class of real-valued functions called ‘kontinuous’ functions such that:

(a) If f is a uniformly continuous real-valued function defined on [0, 1] , then f is kontinuous.

(b) If f and g are kontinuous functions such that Ran(f ) ⊆Dom(g) , then the composition g ◦f
is kontinuous.
(c) If f is a kontinuous function defined on [0, 1] , then f is uniformly continuous.
1
(d) The function x −→ x , defined on R+ , is kontinuous.

proof: The implication (1) ⇒ (2) follows by taking for ‘kontinuous’ the class of continuousBRI functions.
Assume FT, then for this class all but (2)(d) follow immediately. To prove (2)(d) we use proposition 4.1,
which says that FT implies that every compact image in R+ has a positive distance from 0 . Therefore
the function x −→ x1 is uniformly continuous near every compact image, and thus a continuousBRI
function.

Now for the implication (2) ⇒ (1). Suppose we have a class of real-valued functions called ‘kontinuous’
functions such that (2)(a)–(d) hold. We need to derive from this the fan theorem FT. By proposition 4.2
it suffices to show:

(
) If f is a uniformly continuous function from [0, 1] to R+ , then there is an N ∈N such that for all x
in [0, 1] : f (x) > 2−N .

Therefore let f be a uniformly continuous function from [0, 1] to R+ . By (2)(a) f is kontinuous.

h 1
By (2)(d) and (2)(b) the composition function h defined by x −→ f (x) is kontinuous. Therefore by
(2)(c) h is a uniformly continuous real-valued function defined on [0, 1] . This in turn means that there
1
is an N ∈N such that for all x in [0, 1] : h(x) = f (x) < 2N . But then of course for all x in [0, 1] :
f (x) > 2−N . Thus we have shown (
), and by proposition 4.2 this implies FT •

4.4 Since at least corollary 1.3 should elicit some surprise, we will prove it explicitly. This will also
clarify the applicability of the main theorem itself. For the reader’s sake we repeat the corollary here:

corollary: (repeated from 1.3) Within bish the following three statements are equivalent:

(1) The fan theorem FT.

1
(2) If f is a continuousBIS function from [0, 1] to R+ , then the composition x −→ f (x) of f with
the continuousBIS function x −→ x1 is continuousBIS .
1
(3) The function x −→ x , defined on (R+ , dR ) , is continuousBRI .

proof: The implication (3) ⇒ (1) is shown as follows. Suppose (3) holds, then the class of continuousBRI
functions satisfies (2)(a)–(d) of theorem 1.3. The easy verification of this is left to the reader. So taking
‘kontinuous’ to mean ‘continuousBRI ’ we obtain (2) of theorem 1.3, which by the theorem implies FT.
For the implication (1) ⇒ (3) we use proposition 4.1, which says that FT implies that every compact
image in R+ has a positive distance from 0 . Therefore the function x −→ x1 is uniformly continuous
near every compact image, and thus a continuousBRI function.

Now for the implication (2) ⇒ (1). Suppose (2) holds, then we must deduce FT. However, using (2)
we can pinpoint a subclass K of the class of continuousBIS functions, such that this subclass satisfies

16
(2)(a)–(d) of theorem 1.3. To this end let K be the class of functions which contains precisely all
uniformly continuous real-valued functions with domain [0, 1] and also the functions x −→ x1 and the
identity x −→ x both with domain R+ . Trivially the class K satisfies (2)(a), (2)(c) and (2)(d). To
show that (2)(b) holds as well, let f and g be in K such that Ran(f ) ⊆Dom(g) . It remains to show
that the composition g ◦f is also in C . If both f and g are equal to one of the functions x −→ x1
or x −→ x (with domain R+ ) then we are done, the composition is again one of these functions. If
both f and g have as domain [0, 1] , then by definition of K they are both uniformly continuous,
therefore their composition g ◦f is again a uniformly continuous function with domain [0, 1] , and so in
C . The only other possible case is that f is a uniformly continuous function with domain [0, 1] , and
g is either x −→ x1 or the identity x −→ x (with domain R+ ). If g is the identity then we are done,
the composition trivially is in C . If finally g is x −→ x1 , then the statement that the composition
1
x −→ f (x) of f with the function x −→ x1 is continuousBIS and therefore uniformly continuous is
precisely the assumption (2). So the class K satisfies (2)(a)–(d) of theorem 1.3, and by this theorem
this implies the fan theorem FT.

The implication (1) ⇒ (2) is shown as follows. Assume FT, and let f be a uniformly continuous function
from [0, 1] to R+ . By proposition 4.2 there is an N ∈N such that for all x in [0, 1] : f (x) > 2−N . But
of course the function x −→ x1 is uniformly continuous on {x ∈R+ |x > 2−N } , therefore the composition
1
x −→ f (x) of f with the function x −→ x1 is uniformly continuous and therefore continuousBIS •

5 Possible alternative definitions?

5.1 In the author’s opinion, theorem 1.3 and its corollary show that in the absence of the axiom FT
the possibilities for an elegant definition of ‘continuous function’ which allows one to constructively
reproduce a general part of classical mathematics are limited. To clarify this opinion, we briefly investi-
gate the properties that a class of ‘continuous functions’ can be stated to have without (by this stating)
implying the fan theorem. Theorem 1.3 says this is only possible by (partly) dropping or amending the
properties Cont I through Cont IV for the then-defined class of continuous functions. Let’s review
some alternatives one at a time.

The first and very obvious alternative is to drop Cont III altogether, and possibly allow for continuous
functions on [0, 1] which are not uniformly continuous. This amounts to the normal definition of
‘continuous function’ (which we also adopt, see 3.1). However one of Bishop’s ends was to constructively
reproduce a certain general part of classical mathematics, namely continuous functions being uniformly
continuous on compact (sub)spaces. So in our opinion amending Cont III means changing the spirit of
bish. Still, this could be quite acceptable. And the connection with classical, intuitionistic and recursive
mathematics becomes simplified, since the definitions can be the same.

remark: Notice the current role of bish with respect to recursive mathematics. There are many
interesting continuous recursive functions on the recursive continuum ( [0, 1]REC ) which are not uniformly
continuous and therefore not continuousBIS or continuousBRI . This in turn means that these functions are
currently not studied as continuous functions in bish, whereas bish claims to be a neutral playground
for recursive, intuitionistic and classical mathematics. This leads to the question whether the current

17
focus of bish is any different than the focus determined by FT.

Another alternative is to amend or drop Cont IV. This alternative corresponds in fact to the solution
obtained with the definition of continuityBRI , see corollary 1.3. The author of this article does not believe
this to be a viable alternative. More to the point: the author doesn’t believe that a perfectly beautiful
constructive function such as x −→ x1 can seriously be doubted to be continuous. Such a doubt, if
fundamental, reveals a lack of elegance in definitions or axioms which must avenge itself in all sorts
of unforeseen ways, when more of mathematics will be constructively charted. The function x −→ x1
for example plays a crucial part in complex analysis and the theory of Riemannian manifolds, where
topology is all-pervasive.

Next we consider partly dropping or amending Cont II. This alternative corresponds to the solution
obtained with the definition of continuityBIS , see corollary 1.3 (however the definition of continuityBIS
should be improved on, see 3.3). If one wishes to retain the spirit of bish, then this alternative seems
the least problematic. In the field of analysis the author cannot straightaway pinpoint troublesome
consequences arising from the uncertainty whether the composition of two continuous functions is again
continuous.15 He believes this to be partly because not enough of classical mathematics has been charted
constructively. Once this is done, the author believes that the uncertainty about Cont II will be seen
to create many problems for elegance and generality, especially in the field of topology. The pragmatist
however need not necessarily care.

Finally there is the possibility of dropping or amending condition Cont I. One must then find a subclass
‘kontinuous’ of the class of continuous functions for which Cont II through Cont IV hold. Then Cont
I for this class becomes equivalent to the fan theorem. So then without FT one cannot prove that all
uniformly continuous functions defined on [0, 1] are ‘kontinuous’. We will investigate this possibility
in some detail in order to explain why the author thinks it doesn’t work. Some interesting results are
obtained in the process.

5.2 Roughly speaking, amending Cont I comes down to separating the uniformly continuous functions
on (σ2 , dω ) (or equivalently on C , or even on [0, 1] ) in two classes: the ‘kontinuous’ class and the ‘not-
kontinuous’ class. However, in order to do nice mathematics the kontinuous class should contain enough
functions. Therefore ‘nice’ uniformly continuous functions should certainly be kontinuous. Let us for a
moment extend our view to include R2 and R2 -valued functions as well. Then we feel that given any
element z of R2 the distance function dz given by dz (w) =dR2 (z, w) (for all w in R2 , where dR2
is the Euclidean distance function) should be kontinuous. Given this and Cont II through Cont IV
we will show that there is a single uniformly continuous function fbar from σ2 to R2 such that the
kontinuity of fbar implies the fan theorem.

5.3 The definition of this fbar hinges on the possibility to encode an arbitrary decidable subset B
of σ2 with an element αB of σ2 , as follows. Fix (for this section) a bijection h from N to σ2 , with
inverse h−1 . Let B be a decidable subset of σ2 , then αB is given by:

0 if h(n) ∈ B
αB (n) = (for all n ∈N ).
D
1 if h(n) ∈B

Vice versa each α in σ2 corresponds via h to exactly one decidable subset Bα ={ h(n) |α(n) =1} of
15 In functional analysis one expects trouble the soonest.

18
σ2 .

Let jC be the one-to-one correspondence between σ2 and C given in 4.2, where jC (α) = n∈N 2α(n) ·3−n−1
for α in σ2 . Let α be in σ2 , write αe for the element of σ2 given by αe (n) =α(2n) (for all n ∈N ), and
αo for the element of σ2 given by αo (n) =α(2n + 1) (for all n ∈N ). Consider the uniformly continuous
function fbar from σ2 to R2 given by:

fbar (α) = (jC (αe ), sup({0} ∪ {2−n | αo (n) ∈Bαe })).

A more graphical description: for each α in σ2 we look at jC (αe ) on the x-coordinate axis, and we
put a dot at (jC (αe ), 2−m ) as soon as we encounter the first m such that αo (m) is in Bαe . To ensure
that we put a dot even if this never occurs we actually put the dot at the supremum sup({0} ∪ {2−n |
αo (n) ∈Bαe }). The collection of all such ‘dots’ is the image of σ2 under fbar .

5.4 We can now prove another theorem very much like theorem 1.3. We use σ2 instead of [0, 1] in
this theorem in (2)(c), coming a little closer to the full property Cont III (in fact it would be equivalent
to Cont III in the presence of a fifth property stating that the restriction of a ‘kontinuous’ function is
again ‘kontinuous’).

Theorem: Within bish the following statements are equivalent:

(1) The fan theorem FT.

(2) There exists a class of real-valued and R2 -valued functions called ‘kontinuous’ functions such that:

(aa) For all z ∈R2 the function dz is kontinuous.

(ab) The function fbar is kontinuous.
(b) If f and g are kontinuous functions such that Ran(f ) ⊆Dom(g) , then the composition g ◦f
is kontinuous.
(c) If f is a kontinuous function defined on σ2 , then f is uniformly continuous.
1
(d) The function x −→ x , defined on R+ , is kontinuous.

proof: The implication (1) ⇒ (2) follows by taking for ‘kontinuous’ the class of all continuousBRI
real-valued and R2 -valued functions. The only nontrivial issue is to show that (2)(d) holds, and this
was done in subsection 4.3 in the proof of theorem 1.3. The implication (2) ⇒ (1) is trickier but not
difficult. Assume (2) and let B be a decidable bar on σ2 . Determine γ =αB in σ2 (see subsection 5.3).

claim ∀α ∈σ2 [ fbar (α) #(jC (γ), 0) ].

proof Let α be in σ2 . Since B is a decidable bar on σ2 , we can determine the least N ∈N such
that αo (N ) is in B . Determine the natural number M =h−1 (αo (N )) (see subsection 5.3). We now
check on αe (M ) . If αe (M ) =0 then αe #γ (because obviously γ(M ) = 1 ) and so jC (αe ) #jC (γ) and
so fbar (α) #(jC (γ), 0) since jC (αe ) is the x-coordinate of fbar (α) . If on the other hand αe (M ) =1
then fbar (α) =(jC (αe ), 2−M ) and so also fbar (α) #(jC (γ), 0) ◦

Now if we take z =(jC (γ), 0) , then dz is a kontinuous function (by (2)(aa)). Also fbar is kontinuous by
(2)(ab). Therefore the composition dz ◦fbar is a kontinuous function by (2)(b) and by the above claim
it goes from σ2 to R+ . The function x −→ x1 is kontinuous by (2)(d), therefore the composition g

19
 g
given by α −→ dz ◦f1bar (α) is kontinuous by (2)(b). By (2)(c) g is uniformly continuous and therefore
we can determine N ∈N such that for all α in σ2 we have g(α) < 2N . So for all α in σ2 we have
dz ◦fbar (α) > 2−N .

claim ∀β ∈σ2 ∃n ∈N [ n < N ∧ β(n) ∈B ].

proof Let β be in σ2 . Determine α in σ2 such that αe =γ and αo =β . Then the x-coordinate

of fbar (α) is jC (γ) . We had already discovered that dz ◦fbar (α) > 2−N , or in other words that
dR2 (fbar (α), (jC (γ), 0)) > 2−N . Since the x-coordinate of fbar (α) is jC (γ) we find that the y-coordinate
of fbar (α) is greater than 2−N . By the definition of fbar this implies: ∃n ∈N [ n < N ∧ β(n) ∈B ] ◦

So assuming (2) we have shown that an arbitrary decidable bar B on σ2 contains a finite bar, which
amounts to FT •

5.5 With the above theorem 5.4 we already hope to convince the reader that amending Cont I for
finding a workable definition of ‘continuous function’ which reproduces enough of classical mathematics
and yet does not imply the fan theorem is not feasible.

Some however might state that fbar is constructed only for this article, and therefore is not of much
interest and might be excluded from the class of ‘kontinuous’ functions. We have pondered on the
possibility of such a reaction. It probably wouldn’t help if we argued that fbar is a very beautiful
uniformly continuous function, nor if we argued that other even nicer functions which play a same
role as fbar are likely to show up. We did however find an ‘extremely nice’ function which already
causes enough problems for amending Cont I, although we have been unable to show that ‘kontinuity’
of this function implies FT. The function yields a counterexample to FT in the context of recursive
mathematics. It is in essence the same as the counterexample given in [Kleene&Vesley1965], but with
an added flavour which is useful to us later on. Since the role of recursive mathematics is important in
the discussion in section 8, we feel that this ‘extremely nice’ function merits presentation in the next
section.

5.6 This section’s aim was to give some background on possible alternative definitions of ‘continuous
function’. The most obvious alternative is to drop Cont III as a required property. To maintain the
spirit of bish one then still has to look especially into continuous functions being uniformly continuous
near compact subspaces (see subsection 5.1 and section 8). This will eventually bring about the same
problems, so it is a postponement of trouble. Nonetheless it merits serious consideration, since it provides
an equal footing for classical, intuitionistic and recursive mathematics, with transparent definitions. It
is a pragmatic approach which can serve as a basis for different general approaches.

Perhaps the so far next best alternative, dropping Cont II as in continuityBIS , can be done without
too many problems for elegance and generality. This deserves to be carefully looked at, investigating
consequences also in other fields than (linear) analysis.

20
6 A counterexample to FT in recursive mathematics

6.1 The search for an ‘extremely nice’ uniformly continuous function defined on σ2 which causes
similar problems for finding a suitable class of ‘kontinuous’ functions as fbar (see 5.3), leads to a
counterexample to FT in recursive mathematics which is presented in this section. In essence it is the
same as the example given in [Kleene&Vesley1965], but its specific form here should be of interest.

For this we will indicate a recursive element of RN which is recursively apart from all elements of RN
which have a recursive binary expansion in each coordinate (definitions follow). This revindicates an
old question posed by Brouwer in [Brouwer1922], the question whether each real number has a decimal
expansion. Again we need some preliminary definitions.

definition: A real number x is said to be a binary real iff there is an m ∈Z and an α in σ2 such that

x ≡ R m + n∈N α(n) ·2−n−1 . For x in [0, 1] we can take m =0 , so that x in [0, 1] is a binary real iff

there is an α in σ2 such that x ≡ R n∈N α(n) ·2−n−1 and this α is then called a binary expansion
of x . An x in R is called a recursive real iff it is equivalent to a recursively defined Cauchy-sequence
of rational numbers. An x in [0, 1] is called a recursive binary real iff it has a recursively defined
binary expansion (this can be easily generalized to R ). We write Rbin for the set of all binary reals,
and [0, 1]bin for the set of all binary reals in [0, 1] . Likewise we write RN
bin
(resp. [0, 1]N
bin
) for the set
of all infinite sequences of binary reals (resp. binary reals in [0, 1] ). We define the canonical uniformly
continuous surjection kbin from σ2 to [0, 1]bin as follows:


kbin (α) =
D
α(n) ·2−n−1 (for all α in σ2 ).
n∈N

Next let h be the unique bijection from N2 to N such that h(n, m) < h(s, t) ⇐⇒ (n, m) < lexsum (s, t)
where (n, m) < lexsum (s, t) iff n +m < s +t ∨(n +m =s +t ∧n < s) . We can use h to view each α in σ2
as an infinite sequence of elements of σ2 . For this we define, for each n ∈N and α in σ2 the element
α[n] of σ2 as follows:

α[n] (m) =
D
α(h(n, m)) (for all m ∈N ).
N
We define a uniformly continuous surjection kbin from σ2 to [0, 1]N
bin
thus:
N
kbin (α)(n) =
D
kbin (α[n] ) (all α in σ2 and all n ∈N ).

We say that an element x of RN is a recursive infinite-real iff x is a recursive sequence of recursive

reals. Finally we say that an element x of [0, 1]N
bin
is a recursive binary infinite-real iff x is a recursive
sequence of recursive binary reals.

remark: An element x of [0, 1]N bin

is a recursive binary infinite-real iff there is a recursive α in σ2
N
such that x ≡ N kbin (α) . This may be taken as the formal definition if necessary.
R

N
6.2 The uniformly continuous function kbin is the ‘extremely nice’ function promised at the end of
N
the previous section. The following theorem will eventually show that kbin causes problems similar to
the problems caused by fbar in the previous section.

21
Theorem: There is a recursive infinite-real β#bin in [0, 1]N which is recursively apart from each
recursive binary infinite-real in [0, 1]N
bin
.

The proof of this theorem involves some technicalities and is postponed until subsection 10.1 so that
we can directly focus on the interesting consequences of the theorem. Mathematicians experienced in
constructive / intuitionistic mathematics will have no trouble to deduce from theorem 6.2 the following
corollary:

corollary: (to theorem 6.2) There is a decidable subset B of σ2 such that B is a bar on σ2 REC
and yet B contains no finite bar on σ2 REC . In other words: the fan theorem FT fails in recursive
mathematics.

N
proof: We use lemma 4.1. Let β#bin be as in the theorem, let f be given by: f (x) = d (kbin
D RN
(x), β#bin )
which is a uniformly continuous function from (σ2 , dω ) to (R, dR ) . By lemma 4.1 there is a decidable
subset B of σ2 such that:

(i) ∀α ∈σ2 ∀n ∈N [ α(n) ∈B → f (α) > 2−n ]

(ii) ∀α ∈σ2 [ f (α) > 0 → ∃m ∈N [α(m) ∈B] ]

Clearly this B is the desired decidable bar on σ2 REC containing no finite bar on σ2 REC (since the image
N
of σ2 REC under kbin is dense in [0, 1]N
bin
)•

6.3 But the main reason for presenting theorem 6.2 is because it reflects on our analysis concerning
the definition of ‘continuous function’ in bish. This is best spelled out with another theorem quite like
theorem 5.4:

Theorem: It is not possible to prove within bish that there exists a class of real-valued and infinite-
real-valued functions called ‘kontinuous’ functions such that:
d
(aa) For all x ∈RN the function dx given by y −→
x
d N (x, y) is kontinuous.
R
N
(ab) The function kbin is kontinuous.

(b) If f and g are kontinuous functions such that Ran(f ) ⊆Dom(g) , then the composition g ◦f is
kontinuous.

(c) If f is a kontinuous function defined on σ2 , then f is uniformly continuous.

1
(d) The function x −→ x , defined on R+ , is kontinuous.

proof: The idea of the proof is of course the same as in the proof of theorem 5.4. Suppose we can prove
within bish the existence of a class of real-valued and infinite-real-valued functions called ‘kontinuous’
functions which satisfies (aa)–(d) above. Then this class exists also in recursive mathematics. Now focus
on recursive mathematics, and consider the recursive infinite-real β#bin in [0, 1]N given by theorem 6.2,
such that β#bin is recursively apart from each recursive binary infinite-real in [0, 1]N bin
. Since [0, 1]N
bin
is dense in [0, 1]N , we have that inf({dβ (x) |x ∈[0, 1]N
bin
}) = 0 , but at the same time the function
#bin
dβ restricted to [0, 1]N
bin
has for its range a subset of R+ . So finally we consider the composition
#bin
N 1
of the three kontinuous functions kbin , dβ and x −→ x (kontinuous by (ab), (aa) and (d)). This
#bin

22
 1
is the function k given by k(α) = dβ (α) . By (b) k is kontinuous, and by (c) k is uniformly
#bin
continuous on σ2 . However this is impossible, since for each N ∈N we can find an α in σ2 REC such
that k(α) > N . Contradiction, therefore the theorem holds •

We hope that in the light of section 5 this theorem will convince the remaining few that the amending
of Cont I is not a feasible option for coming to a bish-definition of ‘continuous function’. The proof of
theorem 6.2 omitted in this section will be given in the appendix which is the final section.

7 Reality, Determinism and Church’s Thesis

7.1 The fact that FT does not hold in russ is well-known (in the next section some mathemat-
ical foundational discussion about this is offered). In this section we look at a theorem proved in
[Bridges&Richman1987], which shows a little more than the failure of FT in russ. The theorem carries
in it the possibility for an experiment in the real world, which could give statistical evidence on whether
the real world is ‘deterministic’ or not. This is an old debate, unlikely to be easily resolved. This
section aims at providing a mathematical perspective, which might help clarify the debate. Even with
this mathematical perspective, the statistical setting of the experiment is complicated, and touches on
the foundations of probability theory. It remains the question, therefore, whether the old debate can
be resolved at all. The main relevance of this section is for our foundational discussion: it clarifies the
different perspectives on reality of russ on one side, and class and int on the other.

The section is built around a constructive-mathematician’s analysis of what we mean with ‘the real
world is deterministic’. Determinism is a much studied ‘doctrine’ in philosophy, and there are more
than one philosophical ‘definitions’ (read: interpretations) of the word ‘determinism’. Since whole
books have been written on the subject (see e.g. [Earman1986]), we cannot possibly reproduce all
the possible philosophical positions here, nor do we wish to do so. Our main interest devolves from
the interpretations of determinism that modern physics tries to prove or disprove. These efforts have
received quite some attention, both in philosophy and in physics.

The most common interpretation of determinism is an ontological one, called Laplacian determinism.
The ‘definition’ of Laplacian determinism is commonly phrased thus: ‘the instantaneous state of the
world at any time uniquely determines the state of the world at any later time’. Basically the idea is that
there is no freedom at all in the temporal evolution of the world: from a given beginstate, everything
follows ‘mechanically’. (This does not necessarily mean that the world is then predictable, since it may
then be and most likely is! fundamentally impossible to obtain the necessary ‘data’ and the necessary
‘machine’ or ‘mechanical computations’ to actually predict in advance what is going to happen.).

Unfortunately, there is no widely accepted sharper definition of Laplacian determinism. In the author’s
eyes this has led to a number of mathematically vague attempts to prove or disprove the statement ‘the
real world is deterministic’ in modern physics. One sees indeed that various attempts are not disputed
for their physical outcome, but for whether from this outcome Laplacian determinism can be conclusively
proven or dismissed. Now fortunately, the frameworks developed in constructive mathematics offer a
sharp mathematical interpretation of Laplacian determinism. In fact it is a very consequent carrying
through of Church’s Thesis. Church’s Thesis has been related to determinism by many authors, even

23
as radically as we propose (see [FoM2001]). We will see in this section that this leads to an interesting
physical experiment, to (dis)prove determinacy of the real world.

However, we also note in advance that our preferred interpretation is more or less dismissed in [Ear-
man1986] (chapter VI: Determinism, mechanism and effective computability) as being ‘simply not true’.
It should be noted that in [Earman1986] it is taken for granted, without explanation, that classical math-
ematics is the only truth available when it comes to describing the real world. This tacit and unexplained
assumption underlies many discussions in modern physics, and especially the one on determinism. We
cannot therefore emphasize enough that the foundational issues of modern mathematics are of vital
importance to the foundational issues of modern physics (and vice versa).

7.2 Like some other authors, we propose that Laplacian determinism be seen in the light of constructive
mathematics and Church’s Thesis. This means amongst other things that infinite sequences (of natural
numbers; a real number is then given by such an infinite sequence) are never ‘finished’, instead we
see them developing in the course of time. Now a very consequent, therefore elegant interpretation
of Laplacian determinism runs as follows. Suppose that there is in the real world a developing-infinite
sequence of natural numbers, say α . Then how to interpret the statement that this sequence is ‘uniquely
determined’ by the state of the world at time zero? At time zero we can have at most finite information
since according to our constructive viewpoint, infinity is never attained. So this finite information about
α supposedly enables us to ‘uniquely determine’ α in its course of time. It is now hard to see another
interpretation of this last statement, than the one given by Church’s Thesis, namely that this finite
information must be a (Turing-)algorithm that we can use to compute α(n) for any n ∈N .

With classical logic and omniscience, the previous can be stated thus: ‘for every (potentially infinite)
sequence of numbers (an )n∈N taken from reality there is a recursive algorithm α such that α(n) =an for
each n ∈N ’. This statement is sometimes denoted as ‘CTphys ’, see the discussion on the internet forum
‘Foundations of Mathematics’ ([FoM2001]). As stated in the introduction, this classical omniscient
interpretation is easily seen to fail in real life. Therefore we adopt the constructive viewpoint. The
statement ‘the real world is deterministic’ can then best be interpreted as: ‘a (potentially infinite)
sequence of numbers (an )n∈N taken from reality cannot be apart from every recursive algorithm α (in
symbols: ¬∀α ∈σω REC ∃n ∈N [ α(n) =  an ] )’.

We adopt this interpretation in the rest of the paper. We do not have the time and space (...) here to
extensively defend this interpretation, since we are primarily concerned with the foundations of mathe-
matics. We will however point out the connection between these foundations and our interpretation of
determinism. Also there is a direct connection with physics.

For under the above interpretation of Laplacian determinism, we hold that there is a physical experiment
which can give statistical evidence whether the real world is deterministic or not. The experiment
consists ‘simply’ of testing whether an arbitrary sequence taken from the physical reality is given by a
recursive rule or not (so even if one disagrees with the linking of determinism to Church’s Thesis, one
at least has an experiment which might decide whether the physical world is capable of producing a
non-recursive infinite sequence). For this experiment, we first introduce some well-known mathematical
notions in a simplified form. Measure theory is a mathematical discipline which introduces the concept
of ‘size’ also to infinite sets, in such a way that the measure of the union of two disjunct measurable
sets is the sum of the measures of the two sets involved . One basic motivation for this is the theory
of integration. A restricted class of functions which are defined ‘almost everywhere’ on e.g. [0, 1] can

24
be integrated. ‘Almost everywhere’ means that the subset on which such a function is not defined, has
measure 0. There are infinite subsets of [0, 1] which have measure 0 , such as Q ∩[0, 1] , but also the
uncountable Cantor set C . On the other hand, the measure of any real interval [α, β] is simply its
length β −α (it doesn’t matter whether the endpoints of the interval are included or not, since they
have measure 0).

However, even in class not every subset of [0, 1] is measurable. This is one of the issues that sparked
foundational disagreements in the beginning of the twentieth century. The construction in class of a
non-measurable subset of [0, 1] by Vitali in 1905 made essential use of the newfound classical axiom of
choice, which arose out of Cantor’s work on set theory. In constructive mathematics already a countable
union of disjoint intervals need not be measurable, since we cannot always determine the supremum of
the partial sums of the lengths of the intervals involved.

7.3 Linking measure theory to probability theory, the measure of a subset of [0, 1] can be seen as the
likelihood that an arbitrary element of [0, 1] is in that subset. Now in [Bridges&Richman1987] (Ch. 3,
thm. 4.1) it is shown how to construct, for every n ∈N , a bar Rn on [0, 1]REC (in fact a countable
sequence of intervals such that the union contains [0, 1]REC ) such that this bar in class has measure
less than 2−n . Constructively we cannot arrive at the same conclusion, since we cannot assert that
these bars are measurable16 . This is due as indicated to the constructive difficulty in determining the
supremum of the partial sums of the lengths of the intervals involved. Nonetheless it is easily seen that
for each n ∈N , the partial sums of the lengths of the intervals involved in Rn can never exceed 2−n .

Therefore we can still say that classically and intuitionistically the likelihood of an arbitrary element of
[0, 1] being in Rn is less than 2−n . But clearly, in russ this probability is 1 . This is the basis for our
real-world experiment. We conduct this experiment in the usual scientific way, with two hypotheses H0
and H1 , and an uncertainty parameter α0 . The aim is to disprove H0 , by arriving at an experimental
event which under the assumption of H0 will take place with probability less than α0 . As a result we
then accept H1 . This procedure is not symmetrical and it can be worthwile to switch H0 and H1 (which
is done most easily by introducing a reciprocal uncertainty parameter α1 ). For most experiments α0
is taken somewhere in the interval [10−3 , 10−2 ] , but of course we can afford to be liberal, and there is
much at stake, so we suggest for this experiment α0 =2−40 ≈10−12 .

We start with the easiest experimental setting by defining our hypotheses as follows:

H0 The world is infinite in time and non-deterministic, therefore the recursive sequences form a
negligible subset of the set of all infinite sequences.

H1 The world is infinite in time yet deterministic, therefore there is no sequence which is not given
by a recursive algorithm.

We concentrate on H0 and H1 , and disregard the finitistic point of view.17 Now for the corresponding
experiment. We take some natural fluctuating phenomenon (atmospheric pressure, oceanic waterflow,
temperature?) which gives us fluctuating data which are measured with high precision18 , say every
millisecond. It is easy to convert this growing stream of data to a fast-growing ‘infinite’ sequence of
16 In fact there is a constructive theorem saying that every measurable bar on [0, 1] has measure 1 .
17 If the world is bounded in time in such a way that the experiment comes to an end before giving a result, then
everything ends and there is no problem. If not, then there is no problem either.
18 The high-precision is needed to have the natural fluctuation numerically expressed.

25
decimals, to obtain a real-time decimal expansion (more precise: approximation) of an element β of
[0, 1] . At the same time we start constructing the recursive bar R40 , which can be enumerated as
a series of open intervals (Im )m∈N . The experiment finally consists of nothing more than checking
for each m ∈N , whether β is in Im . H0 states that the probability of this occurring for any m ∈N
is less than 10−12 . Therefore if for some m ∈N we do indeed arrive at the conclusion that β is in
Im , then the world is deterministic with overwhelming probability. On the other hand, if the world is
deterministic then this will eventually be discovered by this experiment, if we wait long enough19 !

7.4 As said, the experiment is not symmetrical w.r.t. the hypotheses above. It would be worthwile
to consider α1 =2−40 and see if we can define a reciprocal event in the above experiment which under
assumption of H1 will take place with probability less than α1 . For this we would have to be able for
each n ∈N , to tell something about the probability that for an arbitrary recursive element γ of [0, 1] ,
we have to wait longer than n before discovering an m ∈N for which γ is in Im . We frankly don’t
know whether it is possible to pin down these probabilities. It is too complicated for us, and for the
context of this paper, but hopefully somebody better-versed in probability and recursion theory will
take up this challenge.20

As it stands described above, if the world is deterministic the experiment gives a possibility for russ
to prove that class and int are unrealistic. But we don’t know how long we must wait for such a
result, since even under H1 the algorithm which computes our natural phenomenon is still unknown to
us. Now if the definition of a reciprocal event is possible, then one has a reciprocal statistical way of
proving russ to be unrealistic (depending on the outcome of course). It seems to the author that this is
interesting enough to look into. One must not be surprised of course, in the event that such a reciprocal
event can be defined, when it turns out that the numerical complexity of such an experiment will elude
our limited computational powers. Therefore the finitistic standpoint cannot be completely ignored.

Since in any way this experiment has not yet been conducted, the author is free to believe in his own
bias, namely that the world is non-deterministic. This bias will be seen to influence the discussion in the
next section. Notice that the next section is written in a more familiar mathematical setting, almost as
if this section did not exist. Yet an outcome of an experiment as described above would impact heavily,
if one is concerned with how accurately mathematics reflects reality.21
19 This is some form of Markov’s Principle of course.
20 The author sees a connection with Benford’s law and the fundamental problem of ‘picking a natural number at random’
(see [Hill1996], [DeFinetti1972] and [DeFinetti1974]). For suppose H1 holds, than (with classical logic) every sequence
(finite or potentially infinite) is given by a partial recursive algorithm, which can be seen as an element of I(N, N) . Consider
the probability that for a given e in I(N, N) the sequence β in our experiment runs through an Im which is derived by
letting e compute its first so many values. Many mathematicians would state that this probability cannot be equal for all
e , since by the hypothesis the sum of all these probabilities equals 1. Therefore as e goes to infinity, the probability of
e being the first number to yield an Im containing β goes to zero, in such a way that the sum of all these probabilities
equals 1. But the distribution of these probabilities in the real physical world is unknown. This is a phenomenon which
holds for the set N as well: what is the probability of an arbitrary natural number turning up in reality? In a follow-up
paper we intend to give an alternative explanation of Benford’s law, which opens up a new direction how to tackle these
questions. In [DeFinetti1974] a fundamental different position is stated: countable additivity of probabilities is rejected,
equal (zero) probability of natural numbers then becomes possible with the sum of all probabilities still equal to 1.
21 Even without outcome, the experiment has the merit of putting the debate ‘deterministic or non-deterministic’ in a

blunt and therefore clear perspective.

26
8 Reuniting the antipodes?

8.1 In this section we at times adopt a slightly provocative point of view, which we hope will not
antagonize. Our aim is to stimulate discussion in the realm of constructive mathematics, especially
foundational matters. We certainly do not claim to have final wisdom. And our standpoint will be
recognizably biased.

We have seen that the spirit of bish — constructive mathematics following classical mathematics as
much as possible — is intimately connected to the fan theorem FT. From theorem 1.3 the question
arises whether the scope and/or focus of bish as it stands today is any different from the scope and/or
focus determined by bish + FT. In other words: is the intuition behind bish any different from the
intuition behind FT?

We think not. We are so bold to suggest that a historic problem of communication is the main cause of
the separation between bish and FT. Bishop clearly thought the fan theorem to be a mystic element of
intuitionism. His understandable reaction was not to accept it, however we doubt that he did this with
a clear view of the relationship between the fan theorem and the theory of continuous functions, as put
forward in theorem 1.3. We deduce this from writings of Bishop himself.

8.2 In an unfinished never-published preprint called ‘The neat category of stratified spaces’ Bishop
admits the problem that one cannot prove within bish that the composition of two continuousBIS func-
tions is again continuousBIS . In this preprint he puts forward the solution of continuityBRI (adopted
in [Bridges1979]) as the definitive and final constructive definition of continuity. Theorem and corol-
lary 1.3 show that this approach yields an equivalence between the fan theorem and the continuityBRI
of the beautiful function x −→ x1 . The unpublished preprint in our eyes therefore reflects Bishop’s
unawareness of the relationship between the fan theorem and the theory of continuous functions, as in
theorem 1.3.

Brouwer, on the other hand, formulated the fan theorem rather late in his intuitionistic career. At that
time he had accumulated a vast knowledge and intuition of topology. His treatment of the fan theorem is
exceptional: he proves it using a stronger axiom of bar induction22 . In hindsight we think that Brouwer
at least suspected relationships like in theorem 1.3. We deem it extremely likely that Brouwer knew or
felt that an elegant intuitionistic theory was not possible without the fan theorem. For a topologist like
Brouwer, ‘elegant’ would almost surely have to imply that the composition of two continuous functions
is again continuous, and that a continuous function is indeed uniformly continuous on a compact space.

However, simply adopting the fan theorem in itself as an axiom might strike one as being rather op-
portunistic (this reflects some of Bishop’s resentment of the fan theorem). So if our little ‘psychological
analysis’ is correct, which it probably isn’t!, Brouwer was faced with the problem of motivating the
fan theorem in a fundamental way, transcending possible opportunistic motives. Personally we think
that his solution to this challenge: bar induction (or rather ‘Brouwer’s Thesis’ as presented later in this
section), is a stroke of pure genius.

But quite understandably, Brouwer never motivated the fan theorem in opportunistic terms. The
drawback of this has been that the vast topological backing of the fan theorem remained largely in
22 To be more precise he formulated what [Veldman1981] calls ‘Brouwer’s Thesis’. We investigate this further on.

27
Brouwer’s head23 . Therefore to the analyst Bishop the theoretical edifice leading to the fan theorem
probably lacked the conviction of practice, which the topologist Brouwer had in abundance.

Brouwer’s style of writing probably wasn’t much help either. Still Bishop could have consulted [Kleene&-
Vesley1965] to see that the fan theorem isn’t mystic at all (and neither is CP). Once again an expla-
nation for Bishop’s non-adoption of [Kleene&Vesley1965] might lie in the difference of language and
style. Bishop wanted to do mathematics, practical, down-to-earth. And although he was convinced of
Brouwer’s view that classical mathematics lives in a dream world (‘lacks numerical meaning’ according
to Bishop), he was strongly against the formal and foundational setting of intuitionism at that time,
which had attracted mainly logicians and philosophers.

8.3 So Bishop and more so his followers, notably Bridges and Richman (see [Bridges&Richman1987]
and many other publications), set out on another mission: to reunite the antipodes24 . In our humble
opinion they both succeeded and failed. Successful they have been with respect to changing the focus
of constructive mathematics back to analysis (from logic and philosophy). The author is impressed
with Bishop’s achievement in this sense, since in our view any true mathematics should primarily offer
mathematical challenges, and not focus mainly on the meta-mathematical part. By not adopting CP,
results of bish can easily be used by both intuitionistic and classical mathematics, thus bringing closer
the antipodes.

Also, results in bish hold in recursive mathematics. Therefore some see bish as a generalization of
classical, intuitionistic and recursive mathematics, disregarding philosophical issues. But bish in our
eyes does not take up a neutral position with respect to russ, especially since there are many continu-
ous functions in recursive mathematics which are not continuousBIS or continuousBRI . Therefore these
functions are not studied as continuous functions in bish. The only way to become a common theory for
intuitionism, classical and recursive mathematics is to change the spirit of bish, redirecting bish more
towards recursive mathematics by giving this branch more serious attention. In terms of the definition
of ‘continuous function’ this means dropping Cont III from the list of properties that a continuous
function should possess. In other words, this means adopting definition 3.1 and live with the possible
existence of continuous functions on [0, 1] which are not uniformly continuous.

It then seems wise to develop a general theory of continuous functions, as well as a general theory of
uniformly continuous functions. In this setting, the functions for which we know that they are uniformly
continuous near each compact subspace (let’s call these functions continuousBIS+ ) do not automatically
form a nice subclass (again, this would imply FT). But one can formulate conditions under which the
composition of two continuousBIS+ functions is again continuousBIS+ , as in lemma 3.3. Then all three
antipodes might benefit from this approach in an equal manner. In the light of the previous section,
the obvious advantage is that such a mathematics will reflect reality whether reality is deterministic or
not. As long as we cannot settle that question, it won’t be harmful to keep the options open. So we
plead for some adaptation of bish, see the discussion later on.

In the same vein, one should then develop russ to the point where russ takes into consideration that
even if the world is deterministic, in reality we still will not be supplied with much knowledge about
the algorithms producing our ‘infinite’ sequences. As stated earlier, we think it possible to model this
23 Itmust be said that a large part of Brouwer’s unpublished writings was lost, quite possibly his notes would have shed
a different light altogether on this discussion.
24 Thanks to Peter Schuster for this terminology.

28
mathematically by introducing two players, player I and player II. Player I hands out infinite sequences
to player II, but in principle does not reveal what algorithm is behind the sequences, although sometimes
in her benevolence she gives crumbs of knowledge to player II... We think that in such a model CP is
still consistent. But FT still fails, and of course player II can do the same experiment as described in
the previous section to find this out.

From the previous section there is one question which in this context begs asking. We have seen that
there are bars on σ2 REC with arbitrary small classical measure. But these bars are not measurable in
a constructive sense. We saw that constructively measurable bars on [0, 1] all have measure 1. Now
could it be that an adaptation of FT which limits itself to measurable bars, would also hold in russ?
The following proposition shows that this is not the case.

proposition: There is a measurable bar on σ2 REC which does not contain a finite bar on σ2 REC .

proof: It suffices to consider the bar B given in the proof of corollary 6.2. The standard measure on
σ2 is determined by the measure of the basic open sets, given as follows: µ({α ∈σ2 |α(n) =a}) = 2−n ,
for a in σ2 , lg(a) =n . It is not difficult to see that B (more precise: {β ∈σ2 |∃n ∈N [β(n) ∈B]} ) is
N
measurable and has measure 1 , since {β ∈σ2 |∃n ∈N [β(n) ∈B]} is the inverse image under kbin of the
N
open set C = {α ∈[0, 1] |α #β#bin } . By the proof of corollary 6.2 we have that B does not contain a
finite bar on σ2 REC •

8.4 Back to our main theme: the position of bish. It is safe to assume that Bishop didn’t really believe
that russ is the most realistic mathematics. The fact that bish has consequently sought to maintain
Cont III shows that the heart of bish lies with class and int. In bringing these antipodes together
we feel bish has fallen short. For we think that if one wishes to reduplicate enough classical results,
which was at least part of the thrust of bish, then dropping the fan theorem is too radical a move. It
is present both in intuitionistic and classical mathematics, whereas its formal absence causes all sorts
of problems for elegance and generality, testified to by theorem 1.3. Therefore in bish one comes across
constructions and make-do definitions to avoid a fan theorem to which both the antipodes agree25 . As a
consequence, it is at times not very attractive for intuitionism or classical mathematics to really adopt
the bish approach.

8.5 Asked by one of our critics for an illustration of this unattractiveness, we look at the classical and
intuitionistic homeomorphism h from R to R+ given by:

1
2−x for x ≤ 1
h(x) =D
x for x ≥ 1

It should be noted that under current definitions in bish h is not a homeomorphism (and the
continuityBRI of h is again equivalent to FT). Related to this is the fact that in bish the space R+
is not locally compact. In bish a metric space (X, d) is ‘locally compact’ iff every d-bounded sub-
set of (X, d) is contained in a compact subspace of (X, d) (this implies metrical completeness). It is
not a topological notion, by which we mean invariant under homeomorphisms (see the appendix). It
does not correspond to the classical and intuitionistic notion, which amounts to the usual approach to
25 Moreover the fan theorem in our opinion underlies the intuition behind bish as well, as explained before.

29
‘local’ properties in topology26 , namely: a topological space (X, T ) is locally compact iff for every x
in X and every open U x there is a compact neighborhood W ⊆U of x .27 The current bish def-
inition of ‘locally compact’ might lead to confusion, especially considering the claim that every bish
result is a classical result. Take for instance the following bish variant of the Tietze extension theorem
([Bishop&Bridges1985]):

theorem: (only in bish!) Let (A, d) be a locally compact subspace of a metric space (X, d) . Let
f be a continuous function from (A, d) to ([0, 1], dR ) . Then there is a continuous extension of f to
(X, d) .

Notice that without translation to the not-so-well-known bish definition of ‘locally compact’ the theorem
is both classically and intuitionistically false. The bish definition of ‘locally compact’ and ‘homeomor-
phism’ is partly tied in with the problems arising out of theorem 1.3. We think a better definition is
obtained by defining a metric space to be locally compact iff it has a one-point compact extension. We
will show in the appendix that this amounts to the usual definition in all three antipodes (see 10.2).
However this does mean trouble for bish in relation to Cont II, or Cont III, for continuityBIS and
continuityBRI .

Interestingly, with the usual definition of ‘continuous function’ (the one we adopt), we can prove in bish
the Dugundji Extension Theorem, which is stronger than the Tietze extension theorem. We formulate
a convenient version of the constructive Dugundji theorem proved in [Waaldijk1996]28 :

Theorem: Let (A, d) be a complete and located subspace of a metric space (X, d) . Let f be a
continuous function from (A, d) to (L, dL ) , a complete and locally convex linear space. Then there is
a continuous extension f˜ of f to (X, d) , such that for all x in X the image f˜(x) is in the closure of
the convex hull of f (A) .

The problem with this theorem for bish is that (so far) we have not discovered a general way of proving
within bish that the extension is continuousBIS or continuousBRI if the original function is continuousBIS
respectively continuousBRI . But this problem is non-existent in int, in class and even in russ29 ,
so we think the extension theorem above will be preferred by intuitionistic, classical and recursive
mathematicians alike. This ends our illustration.

8.6 Recapitulating, a drawback in Bishop’s approach is that the common ground of intuitionism and
classical mathematics is greater than one would suspect from the results of bish. So if one would care to
reunite these antipodes, it can be done in a much more powerful way. class and int are much closer to
each other, axiomatically speaking, than to russ. There is a philosophical motivation for this, namely
the earlier mentioned argument that class and int describe a non-deterministic world, whereas russ
describes a deterministic world. For the rest of the discussion in this section we focus on the axiomatical
differences and similarities. The eccentric position of russ as compared to class and int then comes
to the fore naturally.

26 See [Waaldijk1996] for a treatment of local properties in intuitionistic topology.

27 Inint it seems wise to add the condition that (X, T ) be spreadlike.
28 A stronger version is available, but its formulation needs several definitions which would be too troublesome here.
29 Provided one isn’t interested in uniform continuity of the resulting function near compact subspaces.

30
8.7 Therefore we work to reunite the antipodes class and int at their formal intersection which
‘contains’ BID . The intersection also contains the Lindelöf property for complete metric spaces and a
contrapositive form of Baire’s category theorem (intuitionistically derived from CP). We now go into
this in greater detail, and apologize beforehand for what will seem to some a myriad of axioms. The
result will be a plea for four ‘axiomatic systems’, three of which are currently not basking in attention.
The first axiomatic system will in essence be good old bish. The second will be a formal intersection of
all three antipodes, with the addition of MP. The third will be a formal intersection of class and int,
with the addition of MP. The last axiomatic system that we will plead for is traditional intuitionistism
int30 , also with the addition of MP. Other possible systems are left to the reader.

8.8 Remember that we took the trouble of repeating the intuitionistic axioms of continuous choice
AC11 and the weaker AC10 . These axioms conflict with classical mathematics, but there is a very
important consequence of AC10 which also holds, mutatis mutandis, in class and russ. We formulate
this consequence as an axiom called BDD (bar decidable descent). But we first need a definition:

definition: Let B and C be two bars on σω , then B descends from C iff for all c in C there is a
b in B such that b  c .

BDD Every bar on σω descends from a decidable bar on σω .

remark: Notice that AC10 is equivalent to the combination of CP, BDD and AC00 (proof in sec-
tion 10, where we also prove that BDD holds in russ). For elegance we introduce the concept of a
‘thin’ bar, meaning a bar without redundancy, in the next definition and lemma.

definition: A bar B on σω is called thin iff for all b in B and all a in σ ω we have: a < b → a ∈ B .

lemma: Every decidable bar on σω contains a decidable thin bar.

proof: Let B be a decidable bar on σω . Consider the bar C given by: C = {b ∈B |∀a ∈ σ ω [a < c →
a ∈ B] }. It is easy to check that C is a thin bar and a decidable subset of B •

corollary: BDD implies that every bar on σω descends from a decidable thin bar.

Call a metric space spreadlike iff it is the continuous image of (σω , dω ) , then it is a well-known
observation that every complete metric space is spreadlike. From BDD it follows that every spreadlike
metric space is Lindelöf (every open cover admits an enumerable refinement consisting of basic open
sets). From this Lindelöf property it follows that every open cover of a spreadlike metric space has a
subordinate partition of unity. This in turn gives rise to a number of important theorems concerning
continuous functions, such as the Michael Selection Theorem (see [Waaldijk1996]). Of course, combining
BID with BDD one gets the so called principle of monotone bar induction, usually denoted BIM . Also
with BDD one can strengthen FT to the statement that every bar on σ2 contains a finite bar, dropping
the precondition of decidability. This leads to the following important proposition:

proposition: From BDD and FT it follows that every open cover of a compact space has a finite
30 Without the Brouwer-Kripke axiom of course.

31
subcover.

proof: Let (X, d) be a compact space, let U be a collection of open sets such that for every x in X
belongs to a U in U . Since (X, d) is compact, there is a uniformly continuous surjection f from σ2
to X . Then the collection:

{ f −1 (U ) | U ∈U }

is an open cover of σ2 . Therefore the following subset B of σ2 is a bar on σ2 :

B = { α(n) | ∃U ∈U ∀β ∈σ2 [β(n) = α(n) → β ∈f −1 (U )] }

By BDD we know that B descends from a decidable bar C , which by FT contains a finite bar. But
then B also contains a finite bar, namely the descent of the finite bar contained in C . Therefore U
contains a finite subcover •

corollary: From BDD and FT it follows that every continuous function is continuousBRI and therefore
continuousBIS , or in other words: the three definitions coincide. So added to FT, the introduction of
BDD finally resolves the continuity problem which sparked this article. Proof of the other statements
in this subsection is given in the appendix, section 10.

8.9 In [Veldman1985] a very readable description of Brouwer’s justification of BID is given. In fact,
Brouwer more or less postulates an even more general axiom, from which one can derive BID . Veldman
calls (part of) this even more general axiom ‘Brouwer’s Thesis’. We will present it below using our own
terminology, and adopt the abbreviation BT (BT is classically true also, in fact it is equivalent to the
combination of BID and BDD). For the presentation we need to introduce so-called genetic bars. (We
will show later that the concept of a genetic bar coincides with the concept of a decidable thin bar). As
the name suggests, the definition is a genetic one, and runs thus:

definition: A genetic bar is a bar obtained by application of the following two formation rules:

I. The trivial bar {< >} is a genetic bar.

II. If B0 , B1 , . . . is a sequence of genetic bars, then so is the bar B given by: B = {n a |a ∈Bn }.

These rules supply all the genetic bars.

The definition of genetic bars justifies the principle of genetic induction PGI, which runs as follows:

PGI The definition of genetic bars is valid. Moreover, if P is a property of bars, such that:

I. The trivial bar {< >} has the property P;

II. If B0 , B1 , . . . is a sequence of genetic bars with the property P, then so is the bar B given by:
B = {n a |a ∈Bn };

then all genetic bars have the property P.

Brouwer’s Thesis can now be formulated like this:

BT Every bar descends from a genetic bar; and PGI holds.

32
An intuitionistic plea for BT can be give in the following way. The intuitionistic universe σω is
inhabited by choice sequences arising step by step in the course of time. This means that in general
only the minimum of information about an element is known. For the author, the axiom expresses that
— given such a universe — the only one way to convince ourselves that a subset B of σ ω is indeed a
bar, is to show that it descends from something that we can intuitively grasp as a bar, namely a genetic
bar.31

Kleene calls this aptly ‘reversing the arrows’. The genetic definition in fact mirrors what our intuition
tries to do when visualising an arbitrary bar. The author has no trouble accepting this definition, and
in this acceptance lies the immediate intuitive justification of PGI. Brouwer’s justification looks rather
more complex even when explained by Veldman, but we believe it to be essentially the same as our
presentation above. Some nice examples (due to Kleene) of bars descending fom genetic bars are given
in [Kleene&Vesley1965], and more readable in [Veldman1985]. Kleene uses the terminology ‘reversing
the arrows’ in connection with BID . We agree with Veldman, that the justification of BT is more
intuitive than the justification of BID given by Kleene, especially since the necessity of the decidability
condition in BID is left undiscussed in [Kleene&Vesley1965]. Above we hope to have simplified the
treatment in [Veldman1985].

Notice that BT is false in recursive mathematics. In the recursive universe, where far more information
about each element is available than in the constructive and intuitionistic universe, it is possible to
construct bars which do not descend from a genetic bar, see section 7 (these ‘sparse recursive bars’
capture every recursive element of σω , but do not form a bar in intuitionistic or classical mathematics;
whereas any recursive genetic bar in recursive mathematics is also a bar in the constructive, classical and
intuitionistic universe) . This explains the failure of BT and therefore BID in recursive mathematics
(see the next theorem). In classical mathematics BT simply holds.

Theorem: BT is equivalent to the combination of BID and BDD.

proof: First we show that BT implies BDD. For this we prove slightly more, using PGI, namely that
every genetic bar is a decidable thin bar. The proof is easy. Clearly the trivial bar is a decidable thin
bar. Now let B0 , B1 , . . . be a sequence of genetic bars which are decidable and thin. We must show
that the genetic bar B = {n a |a ∈Bn } is likewise decidable and thin. Obviously the empty sequence
< > is not in B . So let a be in σ ω , with length greater than zero. Then there is a b in σ ω and an
n ∈N with a =n b . Now a is in B iff b is in Bn . Therefore B is decidable. To show that B is thin,
suppose that a as above is in B , and that c is in σ ω with c < a . We must show that c is not in B .
For c equal to the empty sequence clearly c is not in B . Else there is a d in σ ω with c =n d , and
trivially d < b (remember a =n b ). Since Bn is thin we find that d is not in Bn , therefore c is not
in B . This shows that B is thin. By PGI we conclude that every genetic bar is a decidable thin bar.
And so BT implies that every bar descends from a decidable thin bar, which implies BDD.

Then we must show that BT implies BID . For this we turn to PGI, with a suitable property P. Let’s
say that a bar C is inductable iff for every inductive A such that C ⊆A , we have that the empty
sequence < > is in A . Clearly BID is equivalent to the statement ‘every decidable bar on σω is
inductable’. By BT this is equivalent to the statement ‘every decidable bar on σω which descends from
31 In fact the more precise analysis is that there are two basic methods that can be employed in ascertaining the ‘bar’
status of a given subset B of σ ω . The direct method is to have B given as a genetic bar. The other method is to see
that B descends from a previously ascertained bar. But these two methods put together yield the method of checking
whether B descends from a genetic bar.

33
a genetic bar is inductable’. This last statement we will prove using PGI. Let’s say that a genetic bar
B is inducing iff every decidable bar which descends from B is inductable.

We now prepare the two basic ingredients of PGI. Firstly, clearly the trivial bar {< >} itself is inducing.
Secondly let B0 , B1 , . . . be a sequence of genetic inducing bars. To apply PGI we must show that then
the bar B given by: B = {n b |b ∈Bn } is likewise inducing. This is not so hard. Let C be a decidable
bar descending from B . Let A be an inductive subset of σ ω such that C ⊆A . Our task is to show
that the empty sequence < > is in A . If < > ∈C then we are done. Else put Cn ={c |n c ∈C} , then
Cn is a decidable bar on σω for each n ∈N . For each n ∈N we have that An = {a |n a ∈A} is an
inductive subset of σ ω . Since C ⊆A , we find that Cn ⊆An and Cn descends from Bn for all n ∈N .
By induction assumption this implies that < > is in An for each n ∈N . Consequently n is in A for
each n ∈N , and since A is inductive we find that < > is in A . Therefore B is also inducing. We
apply PGI to conclude that every genetic bar is inducing, and therefore by BT we have BID .

The other way around is more work. Assume BDD and BID . Now we must derive BT. For this we first
prove that every decidable thin bar is in fact a genetic bar. Therefore the notion ‘decidable thin bar’
and ‘genetic bar’ will be seen to coincide, making the definition of genetic bars valid (the first statement
in PGI). So let B be a decidable thin bar on σω . For each a in σ ω , put Ba = {b ∈σ ω |a b ∈B}.
Now consider the set A = {a ∈σ ω |Ba is a genetic bar }. We hold that A is an inductive set, which
can be seen as follows. Suppose a is in σ ω such that a n is in A for each n ∈N . By the definition
of A this implies that Ba n is a genetic bar for each n ∈N . But then Ba is a genetic bar also, since
Ba = {n b |n ∈N, b ∈Ba  n }. Therefore a is in A , showing that A is inductive. It is easily seen that
B is a subset of A , since Bb is the trivial bar {< >} (and therefore genetic) for every b in B , because
B is thin. So to round it up, we use BID to conclude that the empty sequence < > is in A . By
definition of A this implies that B is genetic. In combination with BDD and corollary 8.8 we now
have that every bar descends from a genetic bar, which is the first half of BT.

For the second half of BT we must derive PGI from BDD and BID . The first step concerns the
validity of the definition of genetic bars, but this has been dealt with above. Next, suppose P is a
property of bars such that the trivial bar has property P. Suppose moreover that if B0 , B1 , . . . is a
sequence of genetic bars having property P, then the bar B given by: B = {n a |a ∈Bn } likewise
has property P. We must now show, using BDD and BID , that every genetic bar has the property P.
For this we refine our reasoning above in proving the first half of BT. Let B be an arbitrary genetic
bar. Then by BDD (and corollary 8.8) B descends from a decidable thin bar B  . Consider the
set A = {a ∈σ ω | every genetic bar which descends from Ba has property P }. We hold that A is an
inductive set, which can be seen as follows. Suppose a is in σ ω such that a n is in A for each n ∈N .
By the definition of A this implies that for each n ∈N , every genetic bar descending from Ba  n has
property P.

Now let C be an arbitrary genetic bar descending from Ba . To show that a is in A we must show
that C has property P. We distinguish two cases, following the definition of genetic bars. The first
case is that C is the trivial bar. Then trivially C has property P. The second case is that there is
a sequence of genetic bars C0 , C1 , . . . such that C = {n c |n ∈N, c ∈Cn } . Now it is easy to see that
Cn descends from Ba  n for each n ∈N , and therefore by the assumption on a has property P. Then
C has property P also, by the assumption on P. Since C was arbitrary, this shows that a is in A ,
showing that A is inductive. We hold that B  is a subset of A . To see this, let b be in B  . Since
Bb is the trivial bar {< >} (remember B  is thin), every genetic bar descending from Bb has to be
the trivial bar also, and so has property P. So to sum it up, we use BID to conclude that the empty

34
sequence < > is in A . By definition of A this implies that every genetic bar descending from B  has
property P, and so B has property P. This proves PGI, the second half of BT, and so we have derived
BT from BDD and BID as promised •

8.10 With the focus on a reunion of the antipodes, we reask the question posed in 2.4: why not
accept Markov’s Principle in intuitionism as well? The author thinks this will be a fruitful change to
intuitionism32 . We present a first consequence of MP, which is in fact an equivalence.

proposition: The following statements are equivalent:

(i) Markov’s Principle MP.

(ii) ‘Every function f from a metric space X to another metric space Y is strongly extensional, that
is: for all x, y in X : if f (x) #f (y) then x #y ’.

We think one should have started by taking the strong extensionality in (ii) as the definition of a
function, and this only illustrates the force of intuition behind MP; just as theorem 1.3 illustrates the
force of intuition behind FT. Notice that for continuous functions strong extensionality is a triviality.
We advocate that strong extensionality become the standard definition for constructive functions, even
if one doesn’t accept MP. Reserving the term ‘weak function’ for any function which is only known to
‘send equivalents to equivalents’, we believe we will never come across a weak function for which we
cannot conclude it is a (strongly extensional) function33 . In this way static in constructivism is reduced.

8.11 Banach’s ‘open mapping theorem’, which has received a lot of constructive attention, is an
interesting case for the antipodes. It holds, mutatis mutandis, in class, int and russ. In [Ishihara1994]
a nice weaker version is proved in bish, this version concerns what we will call the sequential topology (see
subsection 8.17), rather than the metric topology. In class and int, one usually proves the theorem
using a contrapositive form of Baire’s category theorem34 . Classically this form of Baire’s category
theorem is derived from PEM, intuitionistically it follows from CP. It is a powerful topological insight,
which certainly deserves a place in a reunion of the two antipodes.35 We present it below as an axiom
derived from CP.

CPBaire Let (An )n∈N be an enumerable collection of closed subsets of a complete metric space X ,
such that each x in X is in some An . Then there is an An which contains an open ball in the metric
given.

remark: The attentive reader will see our strategy. Historically int and class have been presented
as incompatible, and intuitionism has been called ‘the revolution’. Intuitionism has also been called
‘mystic’. But when we analyze, we find that almost all intuitionistic axioms are valid classically as well.
It is ‘only’ CP which is essentially responsible for intuitionistic results which do not hold in class. And
with CPBaire we isolate from CP yet another powerful topological tool which is valid in class, thus
32 It is true that Markov’s Principle conflicts with the Brouwer-Kripke axiom, but it so gives just another motivation for

the rejection of this Brouwer-Kripke axiom (the full BK axiom also conflicts with AC11 ).
33 The formalization of this belief is exactly MP, by the above proposition.
34 Ishihara’s version gives an interesting though slightly weaker alternative.
35 The contrapositive form is false in russ, even though Baire’s category theorem holds in bish. We have not found a

‘strong enough in our eyes’ adaptation which holds in all three antipodes. The reader may consult [Bridges&Richman1987]
for various adaptations which do hold in all three antipodes.

35
further reducing the distance of int to class. The distance the other way round is however still quite
large, due to the use in class of PEM outside of decidable situations.

proposition: CPBaire fails in russ.

proof: Consider for n ∈N the closed subset An = {α ∈σω |∀m ∈N [α(m) =φn (m) ]} of σω REC •

8.12 One other matter needs to be settled to finish our proposal for reuniting (some of) the antipodes,
as far as this simple paper goes. This concerns axioms of choice. In bish (cf. [Bishop1967]) some very
liberal axioms of choice are formulated, which place little to no restriction on the ‘domain’ and ‘range’ of
the situation36 . Contrary to the popular claim that bish is a common theory for intuitionism, classical
and recursive mathematics, we will conclude below that these liberal axioms of choice are unacceptable
from both a recursive and an intuitionistic point of view. Let us put forward Bas Spitters’ argument
why even a weak, non-continuous axiom of choice will fail in recursive mathematics (many authors have
commented on similar situations).

proposition: (Bas Spitters) There is a subset A of σω REC ×N , such that ∀ α ∈σω REC ∃n ∈N [ (α, n) ∈A ]
and yet there is no recursive function f from σω REC to N which for all α in σω REC gives us that
(α, f (α)) ∈A .

proof: Take A = {(α, e) ∈σω REC ×I(N, N) |e is a recursive index of α }. Clearly ∀ α ∈σω REC ∃n ∈N [ (α, n) ∈A ] .
But if there were a choice function f from σω REC to N which for all α in σω REC gives us that
(α, f (α)) ∈A , then for instance equivalence of two elements α , β of σω REC (meaning that α(n) =β(n)
for each n ∈N ) would be decidable, since one would only have to check whether f (α) equals f (β) or
not. This decidability of equivalence contradicts well-known basic results in recursive-function theory •

This proposition goes to show that one cannot put forward any ‘conventional’ axiom of choice to hold
in bish. Before we meditate on possible solutions, we first investigate the intuitionistic situation. In
intuitionism, the intuition behind AC11 (and AC10 ) depends vitally on the fact that the assertion
∃y [A(x, y)] holds for all x in all of σω , that σω is a non-deterministic universe with incomplete
information, and that A is a subset of σω ×σω . An intuitionist is definitely not convinced of the
validity of choice axioms for ‘arbitrary domains’ and ‘arbitrary ranges’, and this is vindicated by the
proposition above (notice that σω REC is a subset of σω ). The axiom DC1 is therefore also phrased
in the context of elements and subsets of σω 37 . DC1 is weaker than the axiom of dependent choice
currently accepted in bish, so we feel that DC1 can be safely adopted for any reunion of antipodes.

8.13 A way out of some of the problems above can be found in the presentation axiom described
in [Troelstra&vanDalen1988] (sect. 4.2). If we rephrase and restrict a little, then basically this axiom
states the following (where the notion of ‘set’ is left vague):
36 Many authors have commented on this, but there has been no formal change of phrasing of these axioms. In practice
only the weaker forms that we propose further on seem to be really used.
37 For ‘higher-order’ settings even the combination of AC
11 and DC1 does not always suffice, even if these settings seem
intuitively clear. It probably is worthwile to investigate if intuition allows for higher-order choice axioms than AC11 and
DC1 .

36
 PA Let A be a set. Then we can find a subset C of σω , and an equivalence relation ≡ A on C ,
such that
(1) A coincides with C equipped with ≡ A , under a mapping h from C to A ,

(2) If B is a subset of A ×σω such that ∀α ∈A ∃β ∈σω [ (α, β) ∈B ], then there is a function f from
C equipped with ≡ ω to σω such that ∀α ∈A ∃γ ∈C [ h(γ) =α ∧(α, f (γ)) ∈B ].
Informally speaking, the presentation axiom states that for every set A , there is an underlying ‘pre-
sentation’ set C for which an axiom of choice holds, even if it fails for A . By C being an ‘underlying
presentation’ we mean that C is a subset of σω such that A can be derived from C by imposing
an equivalence relation ≡ A on C . The reader may check that (only) for the example in the proof
of proposition 8.12, taking TOT as presentation of σω REC would resolve the problem (consider N as
a subset of σω , then TOT is also a subset of σω and take s ≡ A t = ∀n ∈N [φs (n) =φt (n)] ). But we
have little faith in the usefulness of the presentation axiom, since it does not offer enough to resolve
the choice problem for russ. This because in russ choice functions must be partial recursive functions,
otherwise they are of little use. And if one exacts this in the conclusion of PA, then PA is false in russ
(see [Troelstra&vanDalen1988] (sect. 4.11)).

Another way to solve many problems is by taking the desired result of a possible choice axiom (that
there is a partial recursive choice function) as a premise in statements one wants to prove. For example,
instead of saying ‘every open cover of a complete metric space has an enumerable subcover’ one can
also state: ‘every effective open cover of a complete metric space has an enumerable subcover’, where
‘effective’ means ‘given by a suitable algorithm’. So once again, the discussion is whether to adopt
axioms, or to restrict one’s attention beforehand to objects already having the desired properties. In
the case of choice axioms for russ, the latter seems an elegant enough way. Nonetheless we formulate
an intuitively motivated axiom of choice for russ, called CT11 . With this axiom one can prove BDD
(see the appendix, CT11 is not needed for the very similar statement ‘every effective bar descends from
a decidable bar’). We now go into this matter in greater detail.

The different nature of recursive mathematics is revealed by the fact that in russ there is indeed a non-
trivial underlying domain for σω , namely the subset TOT of N of indices of total recursive functions.
This is not the case in classical and intuitionistic mathematics. The set TOT however is a truly difficult
mathematical entity from a constructive point of view. It is not even an enumerable subset of N , let
alone decidable. So it cannot be constructed on its own. In order to work with total recursive functions
one has to consider the partial recursive functions as well. Now the choice problem comes to the fore,
but it can be resolved with the axiom ECT0 (see 2.4) which is broadly accepted in russ. From ECT0
it follows that if for all n in TOT a certain natural number exists, then there is already a partial
recursive function from N to N which for n in TOT yields precisely such a number. However, the
axiom ECT0 is motivated in [Troelstra&vanDalen1988] (sect. 4.4) only in the formal logical setting
of being consistent with Heyting Arithmetic (HA). The author finds the phrasing and the motivation
unappealing. We believe that an easier and more intuitive motivation can be given for a straightforward
choice axiom. Hence we phrase an ‘axiom of choice for the underlying domain N ’ for russ.

CT11 Let A be a subset of N ×N such that

∀n ∈TOT ∃m ∈N [ (n, m) ∈A ] .
Then there is a partial recursive function h from N to N such that for all n ∈TOT we have:
(n, h(n)) ∈A .

37
We motivate this axiom as follows. Given ∀n ∈TOT ∃m ∈N [ (n, m) ∈A ] , we analyze the way in which
the information n ∈TOT can lead to the pinpointing of an m ∈N for which (n, m) ∈A . Since such an
m must be produced at a finite point in time, we cannot have used more information than n itself and
a finite number k of computations of φn say on the input numbers 0, 1 . . . , M . So the method which
produces the desired m ∈N does not depend on the fact that n is in TOT, but on a finite number of
succesful computations derived from n 38 .

If we stipulate a canonical way to compute (e.g. we first take one computational step in computing
φn (0) , then we compute the first step in φn (1) , then another step in φn (0) , then another in φn (1)
and a step in φn (2) , then back to φn (0) etc.), then we arrive at the following conclusion. Given n and
a finite number k of computations derived from n we know whether this finite number k is sufficient
to produce m ∈N or not. Our method either says: ‘ok, enough information’ or it says ‘too bad, need
some more information’. Therefore there is a decidable subset B of N ×N such that for all (n, k) ∈B
we have: the number k of computations derived from n is sufficient to produce an m ∈N such that
n ∈TOT implies (n, m) ∈A .

Now put A = {(n, m) ∈N |n ∈TOT → (n, m) ∈A} . By our reasoning above we find:

∀n ∈N [∃k ∈N [(n, k) ∈B] → ∃m ∈N [(n, m) ∈A ] ]

Then we can apply CT01 , which we formally derived from CT0 in paragraph 2.4, to produce a partial
recursive h from N to N such that for all n ∈N : if ∃k ∈N [(n, k) ∈B] then h(n) is defined and
(n, h(n)) ∈A . Notice that h then is as required in the conclusion of CT11 , by the definition of A .

remark: In our reasoning we have reduced CT11 to the core of russ, namely CT. Notice that CT11 is
not an instance of the presentation axiom. We do not see a way to phrase a common axiom of choice for
all three antipodes which for russ amounts to CT11 . That our terminology is more or less consistent
is the content of the following proposition.

proposition: CT0 follows from CT11 .

proof: Let A be a subset of N ×N such that ∀n ∈N ∃m ∈N [ (n, m) ∈A ]. Then in particular we find:

∀n ∈TOT ∃m ∈N [ (φn (0), m) ∈A ] . Put A = {(n, m) ∈TOT ×N |(φn (0), m) ∈A} , then trivially:

∀n ∈TOT ∃m ∈N [ (n, m) ∈A ]

So by CT11 there is a recursive function f from N to N such that for each n ∈TOT : (n, f (n)) is in
A . It is standard to construct a recursive function g from N to N such that for each n ∈N we have:
∀m ∈N [φg(n) (m) =n] , and therefore g(n) ∈TOT . Now the function h given by h =f ◦g is a recursive
function from N to N such that for all n ∈N we have: (n, h(n)) ∈A •

8.14 Next we turn to class and int. Since their basis is more conventional than russ, there is a
conventional way to phrase an axiom of general choice for the reunion of classical and intuitionistic
mathematics. Such an axiom of general choice must be restricted in a way comparable to AC11 : the
‘domain’ of choice should be limited to a spread ( σω is sufficient). However the intuitionistic insight
that the resulting choice function is continuous, must of course be dropped for the reunion. So we first
38 But the conclusion that (n, m) ∈A can depend on the fact that n is in TOT! For instance if A is specified as a

subset of TOT ×N . . .

38
 propose a weak version of AC11 , which in the conclusion drops the continuity of the choice function (so
it leaves open whether the choice function is a spread-function).

AC11weak Let A be a subset of σω ×σω such that:

(
) ∀α ∈σω ∃β ∈σω [ (α, β) ∈A ]

Then there is a function γ from σω to σω such that for each α in σω : (α, γ(α)) is in A . We say
that γ fulfills (
).

remark: Because it is not asserted that the choice function is a spread-function, the choice function
cannot be seen to be representable by an element of σω . We have not investigated the consequences of
this, which might be further-reaching than expected at first glance (but see some propositions below).
We have that AC11 is equivalent to the combination of AC11weak , AC10 and AC01 . AC10 in turn is
equivalent to the combination of CP, BDD and AC00 .

Using BDD we can prove some interesting propositions, which for all three antipodes can serve as
alternative ‘axioms’ of continuous choice. The first proposition is a surrogate for AC10 . Notice that its
proof is also a proof that AC10 follows from the combination of CP, BDD and AC00 .

proposition: (using BDD) Let A be a subset of σω ×N such that:

(
) ∀α ∈σω ∃n ∈N ∃m ∈N ∀β ∈σω [ α(m) =β(m) → (β, n) ∈A ] .

Then there is a spread-function γ from σω to N such that for each α in σω : (α, γ(α)) is in A .

proof: From (
) we deduce that the set B = {a ∈σ ω |∃n ∈N ∀β ∈σω ∩a [(β, n) ∈A] } is a bar on σω .
By lemma 8.8 the bar B descends from a decidable thin bar C . By definition of B , for each c in C
there is an n ∈N such that ∀β ∈σω ∩c [(β, n) ∈A] } . Now by AC00 the desired spread-function γ can
easily be defined on C •

corollary: Every continuous function from a spreadlike metric space to another metric space can be
represented by a spread-function.

8.15 The second proposition, which for all three antipodes can serve as a surrogate for AC11 , closely
resembles the Michael Selection Theorem. In [Waaldijk1996] an intuitionistic version of the Michael
Selection Theorem is proved, using AC10 and DC1 (but by corollary 8.14 one sees that in fact only
BDD is used, so the theorem holds in all three antipodes). This theorem is concerned with conditions
under which a continuous choice function may be constructed, given a subset A from a product space
X ×Y such that for all x in X there is a y in Y with (x, y) ∈A . The Michael theorem provides
a continuous choice function39 under the following conditions. Firstly X must be a spreadlike metric
space and Y must be a Banach space. Moreover the set A is ‘Lower Semi Continuous’, and for each x in
X the set {y ∈Y |(x, y) ∈A} is convex and closed. The proposition below indicates that for X =σω =Y ,
lower semicontinuity and the closedness condition suffice for the existence of a suitable continuous choice
function. To make this precise, we use a simplified definition of lower semicontinuity40 .
39 Such a function is called a continuous selection.
40 For the general metric definition, see [Waaldijk1996].

39
definition: Let A be a subset of σω ×σω such that ∀α ∈σω ∃β ∈σω [ (α, β) ∈A ]. Then A is called
lower semicontinuous iff for all (α, β) ∈A and for all n ∈N , there is an m ∈N such that: ∀γ ∈σω [γ(m) =α(m) →
∃δ ∈σω [δ(n) =β(n) ∧(γ, δ) ∈A] ] .

proposition: (Michael Selection Theorem for σω , using BDD and DC1 ) Let A be a lower semicon-
tinuous subset of σω ×σω , such that for all α in σω the set {β ∈σω |(α, β) ∈A} is closed. Then there
is a continuous (spread-)function γ from σω to σω , such that for all α in σω we have (α, γ(α)) ∈A .

proof: The proof is a simplified transposition of the constructive proof of the Michael theorem in
[Waaldijk1996]. By definition we have:

∀α ∈σω ∀n ∈N ∃m ∈N ∃β ∈σω ∀γ ∈σω [γ(m) =α(m) → ∃δ ∈σω [δ(n) =β(n) ∧(γ, δ) ∈A] ] .

Therefore for n ∈N the set An = {(α, a) |lg(a) =n ∧∃β ∈σω ∩a [(α, β) ∈A ]} satisfies the requirements
of proposition 8.14 above. So the set Rn = {γ ∈σω | γ is a spread-function ∧∀α ∈σω [(α, γ(α) ∈An ]} is
inhabited for each n ∈N .

This shows that for each n ∈N we can find a spread-function δ in σω which for each α in σω give
the first n values of a β such that (α, β) is in A . Now the lower semicontinuity of A guarantees
us that given such a δ , we can always ‘extend’ δ to a spread-function δ  which for each α gives us
the first n +1 values of a β with (α, β) in A , and such that δ(α) < δ  (α) . Therefore we can apply
DC1 to find a sequence of spread-functions (γn )n∈N such that for all α in σω and all n ∈N we have
both ∃β ∈σω [β(n) =γn (α) ∧(α, β) ∈A] and γn (α) < γn+1 (α) . It is trivial to construct from (γn )n∈N
a spread-function γ from σω to σω such that for all n ∈N and α in σω one has γ(α)(n) =γn (α) .
So then we finally apply the fact that for all α in σω the set {β ∈σω |(α, β) ∈A} is closed to conclude
that γ is the required spread-function •

remark: The proof of this proposition illustrates the necessity of surrogates for AC10 . The axiom DC1
is strict in the sense that it requires the objects which one would chain together to be representable by
elements of σω . For continuous functions from a spreadlike metric space to another metric space, this
representability is the case, by corollary 8.14.

8.16 All the axiomatic ingredients having been presented, we now ask attention for four axiomatic
systems. First of all we shortly plead for bish , which is bish with AC01 and DC1 as axioms of
choice rather than the ones currently in use. The advantage of this axiomatic system is that practically
everyone will agree to its axioms. The leaps of faith required are relatively small. The drawback is
that intuition gets little space, therefore certain intuitive truths cannot be deduced. As stated, we
recommend a different approach towards continuous functions for this system.

8.17 Next we shortly plead for a formal system which incorporates as much common intuition of the
three antipodes class, russ and int as possible, with the addition of MP. Let us call this system
bish+ . The axioms of bish+ are therefore: the axioms of bish , MP, and BDD. 41 The advantage of
bish+ is that results derived ‘hold in reality’ independently of whether reality is deterministic or non-
deterministic (see section 7). Another advantage is that a common framework for all three antipodes can
be developed in such a way that work done in one antipode becomes transferable to another antipode
41 Intuitionists which do not believe in MP have the option to leave this out of course.

40
to the extent possible. This system can be extended with ‘common insights’, to give an oversight of
what theorems hold in all three antipodes.

One such ‘common insight’ of the three antipodes is that in a complete metric space the so-called se-
quential topology coincides with the metric topology42 . Therefore every sequentially continuous function
from a complete metric space to another metric space is a continuous function. Let us explain shortly
what is meant by ‘sequential topology’, since in the literature this definition does not occur. A subset A
of a metric space (X, d) is sequentially open iff for every a in A , and every Cauchy-sequence (an )n∈N
in (X, d) converging towards a , there is an N ∈N such that for all m ∈N , m > N , we have am is in
A.

We do not see how to turn this common insight into (or prove it from) an axiom which has a simple
defense. The same goes for Banach’s open mapping theorem, which also holds in all three antipodes.

remark: In [Waaldijk1996] the so-called apartness topology is introduced as follows. Let (X, #) be
an apartness space. Let A be a subset of X . Then A is open in (X, #) iff for all x in A we have:
∀y ∈X [ y ∈A ∨y #x ] . It is easy to prove in bish that every (strongly extensional) function from a
topological apartness space to another topological space is continuous. Now in both int and russ one
can show that for a complete metric space the metric topology coincides with the apartness topology.
From this one arrives at the conclusion that every function from a complete metric space to another
topological space is continuous.

8.18 Of course we plead for a formal system which is a reunion of class and int with the addition
of MP.43 Let us call this system clint+ . Then the axioms of clint+ are the axioms of bish+ , plus
AC11weak , BT and CPBaire . This system develops a substantial part of classical mathematics in an
intuitionistic way, such that both classical and intuitionistic mathematicians can benefit. Therefore this
system comes closest to what we believe was Bishop’s primary aim.

8.19 Finally, we plead for traditional intuitionism int with the addition of MP. In the author’s
personal opinion int is the branch of constructive mathematics which is closest to reality. In a world of
incomplete information the axiom CP is an unmistakeable constructive truth, that merits serious study.
The addition of MP is a simple consequence of the same attitude: if an axiom has a clear intuitive
defense, then this axiom is worth adopting to study its consequences.

9 Conclusions and Acknowledgements

9.1 To really open a discussion, we sum up what we feel to be the most important conclusions to
be drawn from this paper. They are presented directly below as numbered statements, meant to draw
attention.
42 From this insight it follows that every surjective linear function with located kernel from a Banach space to another

Banach space, is open. In other words: a ‘located kernel’ version of Banach’s open mapping theorem, see [Ishihara1994].
43 Intuitionists which do not believe in MP have the option to leave this out of course.

41
 1. The intuition that with regard to continuous functions one can restrict oneself to continuous
functions which are uniformly continuous on compact (sub)spaces is in fact equivalent to the
intuition behind the fan theorem FT (within bish). This intuition has so far been the very
cornerstone of bish. Therefore bish implicitly implies the fan theorem FT. It is not yet clear
whether a ‘satisfactory’ theory of continuous functions on compact spaces can be built without
explicitly relying on the fan theorem. To understand some of the problems see theorem 1.3. But
also see section 7, since ‘satisfactory’ obviously depends on our view of reality.

2. This foundational issue should be studied with priority. Several possibilities need investigation.

3. One important possibility is that of ‘reuniting the antipodes’. Hereby we mean the adoption of a
simple set of axioms which, when suitably interpreted, are acceptable both intuitionistically and
classically. Sweeping aside historical misunderstandings, such a constructive mathematics can be
much more powerful than bish currently is, both mathematically and communicatively speaking.

4. For the purpose of such a reunion we propose the system clint+ , with axioms BT, CPBaire , MP,
AC11weak , DC1 , and Ind, at least for the time being.

5. CP is the only axiomatic insight of intuitionism which is not classically acceptable44 . This simply
means that CP is a truly original intuition. Adding CP to the above list we get full-blown int
with MP added. This combination, called int+ , is worth serious study with regard to all fields
of mathematics.

6. The link between reality and foundations of mathematics can be studied experimentally (see sec-
tion 7). On the other hand, the link between reality and foundations of physics can be studied
mathematically. This can have a profound impact on physics and mathematics. Physicists there-
fore should be aware of foundational issues in mathematics, and mathematicians should be aware
of foundational issues in physics.

9.2 I would like to thank the following people for their support in the epigenesis of this paper. Firstly
Henk Taale, fellow mathematician at the Transport Research Centre, for convincing me to write this
paper at all. Then of course Wim Veldman, who was my thesis supervisor and much more, for his
continuing interest in this topic. In the last three years Bas Spitters joined the discussions, and like
Wim gave many valuable comments. Their friendly help and criticism have greatly improved the quality
of the presentation. Peter Schuster contributed his enthusiasm, which has been an important stimulant
in the whole thing, and his apt terminology ‘reuniting the antipodes’ which enabled me to find a good
focus for the discussion. I am happy that Wim Couwenberg also became interested, his enthusiastic
support means a lot to me. Also, I thank the anonymous referee, who made kind and very helpful
remarks, and thoughtfully included some extra references on history and on determinism.

Since 1993 my spiritual master Parthasarathi Rajagopalachari has helped me find a direction in math-
ematics which is close to my heart and my intuition. Of course this is only a byproduct of his help in
more important aspects of life than mathematics.

I also wish to thank my friends Thijs and Reinier, who were more than casually interested both in the
contents and the proceedings of this paper. Finally I wish to thank my children Nora and Femke and
my wife Suzan for their patience with this neverending project, and for their love.
44 We intend to show in a follow-up paper that there is a simple classical game-theoretic setting, in which all intuitionistic

axioms hold true classically (including CP).

42
10 Appendix: remaining proofs and remarks

10.1 We still owe the reader the proof of theorem 6.2. For the proof we will draw freely from the
theory expounded in [Kleene1952]. There is a decidable subset I(N, {0, 1}) of N such that each e in
I(N, {0, 1}) is a properly formed recursive index of a partial recursive function from N to {0, 1} , and
vice versa for each partial recursive function α from N to {0, 1} there is an e in I(N, {0, 1}) such
that ∀n ∈N [ α(n) = • ’ stands for: ‘equal if one of the algorithms terminates, given
• φ (n) ] , where ‘ =
e
the input’.

proof: (of theorem 6.2) For each n ∈N we define a recursive real βn in [0, 1] by defining for each
m ∈N the rational number βn (m) (a recursive real is given by a Cauchy-sequence of rational numbers
‘converging with fixed speed’). For this remember that we fixed a recursive bijection h from N2 to N
in subsection 6.1, and let n ∈N and m ∈N . Then (writing ‘ µk ≤ m [A(k)] ’ to mean ‘the least natural
number k less than m such that A(k) ’):
 1

 2 if not n ∈I(N, {0, 1})

 1
2 if n ∈I(N, {0, 1}) but not ∃k ≤ m [ T (n, h(n, 0), k) ]
βn (m) =
D  1 −s

 −2 if n ∈I(N, {0, 1}) and where s =µk ≤ m [ T (n, h(n, 0), k) ] and Outc(s) =1
 21 −s
2 +2 if n ∈I(N, {0, 1}) and where s =µk ≤ m [ T (n, h(n, 0), k) ] and Outc(s) =0
This definition probably needs some explanation. Informally speaking we are checking whether n
is the index of a partial recursive function α from N to {0, 1} . If so, then α possibly is a total
recursive function, in other words α might be a recursive element of σ2 . In that case using h we
can see α as a recursive sequence of recursive elements (α[n] )n∈N of σ2 , and apply the function kbin
N

to α to find a recursive infinite-real in [0, 1]N

bin
. For our diagonal argument we are now interested in
[n]
αn (0) =α(h(n, 0)) . If at a certain moment s in time the algorithm α finishes computing and yields
α(h(n, 0)) , then we can check whether α(h(n, 0)) =1 or α(h(n, 0)) =0 . In the first case we are now
sure that if α is a recursive element of σ2 , then the nth coordinate of kbin
N
(α) is a real number greater
1
than or equal to 2 (check!). Likewise in the second case we are sure that if α is a recursive element of
σ2 , then the nth coordinate of kbin
N
(α) is a real number less than or equal to 12 .

claim Let n ∈N . Suppose n is a recursive index of a recursive element α of σ2 . Then the nth
N
coordinate of kbin (α) is a real number apart from the real number βn defined above.

proof Since n is a recursive index of α , we have for all t ∈{0, 1} that α(h(n, 0)) =t iff there is an
s ∈N such that T (n, h(n, 0), s) and Outc(s) =t .

case 1 t =1
1
Then the nth coordinate of kbin N
(α) is a real number greater than or equal to 2 . On the other hand
1 −s
βn ≡ R 2 −2 therefore the claim holds.

case 2 t =0
1
Then the nth coordinate of kbin N
(α) is a real number less than or equal to 2 . On the other hand
1 −s
βn ≡ R 2 +2 therefore the claim holds ◦

Now let β#bin be the recursive infinite-real given by β#bin =β0 , β1 , . . . (or equivalently β#bin =(βn )n∈N

43
 [n]
or equivalently ∀n ∈N [β#bin =βn ] , whichever notation the reader finds most convenient).

claim The recursive infinite-real β#bin in [0, 1]N is recursively apart from each recursive binary
infinite-real in [0, 1]N
bin
.

proof Let γ =γ0 , γ1 , . . . be a recursive element of [0, 1]N bin

. Determine a recursive α in σ2 such
N
that γ ≡ N kbin (α) (see remark 6.1). Determine a recursive index e of α . Determine s ∈N such that
R
T (e, h(e, 0), s) . By the claim and its proof above dR (γe , βe ) ≥ 2−s , showing that γ is apart from β#bin .
Since every function involved is recursive, the claim holds and theorem 6.2 is proved ◦ •

10.2 Now we turn to our remark in 8.5. Following our analysis that the dropping of Cont III is
probably wisest, we suggest the following adaptation of the bish definition of ‘locally compact’:

definition: Let (X, d) and (Y, dY ) be metric spaces. (X, d) coincides with (Y, dY ) iff there is a
continuous surjection h from (X, d) to (Y, dY ) with a continuous inverse h−1 . Such an h is called
a homeomorphism . (Y, dY ) is called a one-point compact extension of (X, d) iff (Y, dY ) is compact
and there is a y in Y such that (X, d) coincides with ({z ∈Y |z #y}, dY ) . A metric space (X, d) is
locally compact iff (X, d) has a one-point compact extension. (X, d) is called topologically complete
iff (X, d) coincides with a complete metric space (then there is a d-equivalent d on X such that
(X, d ) is complete).

Theorem: Let (X, d) be a locally compact space. Then:

1. For every x in X and every open U x there is a compact neighborhood W ⊆U of x .

2. There is an enumerable neighborhood cover of (X, d) consisting of compact subspaces.

3. There is a d-equivalent metric d on X such that in (X, d ) every bounded subset is contained
in a compact subspace (therefore (X, d) is topologically complete).

We hold that in all three antipodes the above definition amounts to the desired usual definition of
‘locally compact’.

Theorem: (using BDD) For a metric space (X, d) the following statements are equivalent:

1. (X, d) is locally compact

2. For every x in X and every open U x there is a compact neighborhood W ⊆U of x .

3. There is an enumerable neighborhood cover of (X, d) consisting of compact subspaces.

4. There is a d-equivalent metric d on X such that in (X, d ) every bounded subset is contained
in a compact subspace (therefore (X, d) is topologically complete).

We leave the proof of these theorems as an exercise to the reader, who may consult [Waaldijk1996] and
[Bishop&Bridges1985] (thm. 4.4.9).

10.3 Then we need, for the sake of completeness, to prove a number of statements made in subsec-
tion 8.7 and further on. First we show:

44
 proposition: AC10 is equivalent to the combination of CP, BDD and AC00 .

proof: For the implication from right to left see the proof of proposition 8.14. For the implication from
left to right the only non-trivial thing to show is that BDD follows from AC10 . Let B be a bar on
σω , let A =σω ×B . Then we have:

(
) ∀α ∈σω ∃n ∈N [ (α, α(n)) ∈A ]

By AC10 there is a spread-function γ from σω to N such that for each α in σω : (α, γ(α)) is in A .
Therefore γ(α) is in B for each α in σω . Now consider α in σω . There is a unique n ∈N such that
γ(α(n)) > 0 and we know that γ(α) then equals γ(α(n)) −1 . We hold that γ(α(n))  α(n) . Otherwise
we would have α(n +1)  γ(α(n)) , and there would be a β in σω with β(n) =α(n) yet β(n) =  α(n) ,
whereas by definition of spread-function γ(β) =γ(α) and β(n +1)  γ(α(n)), contradiction. So we can
take the desired decidable bar to be the following one: {a ∈σ ω |γ(a) > 0} •

10.4 lemma: BDD follows from CT11 in russ.

proof: The proof is a combination of simple building blocks from the literature, which only need little
adaptation, and the reasoning in the proof of the lemma above. First one shows, using CT11 : ‘every
bar on σω REC contains an effective bar on σω REC ’, where ‘effective’ means ‘given by a partial recursive
function’. Then one proves, using CT0 : every effective bar on σω REC contains a recursively enumerable
bar on σω REC . This proof can be found in [Ishihara1993]. Finally we must show:

claim Every recursively enumerable bar on σω REC descends from a decidable bar.

proof Let A = {φe (n) |n ∈N} be a bar on σω REC , enumerated by the total recursive function
φe from N to σ ω (for simplicity we view σ ω as a subset of N ). We remind the reader that
if T (e, n, k) holds for e, n, k ∈N , then n < k and Outc(k) < k . Now consider the set B given by:
B = {a ∈σ ω |∃n < lg(a) ∃b < a [T (e, n, lg(a)) ∧Outc(lg(a)) =b]} . Since A is a bar, B is a decidable bar
such that A descends from B ◦ •

10.5 We also need to derive CPBaire from CP. For this we repeat a theorem from [Waaldijk1996],
which is proved therein using only CP.

Theorem: (CPcm ) let (X, d) be a complete metric space. Let A be a subset of X ×N such that:

(i) ∀x ∈X ∃n ∈N [ (x, n) ∈A ]

(ii) ∀x, y ∈X ∀n ∈N [ (x ≡y ∧ (x, n) ∈A) → (y, n) ∈A ]

Then: ∀x ∈X ∃n, m ∈N ∀y ∈X [ d(x, y) < 2−m → (y, n) ∈A ] .

From this theorem it is easy to derive CPBaire , we leave this as an exercise.

10.6 remark: As a last remark we offer a weak version of FT which is true in bish if one accepts the
definition of genetic bars and PGI. The proof is a simple application of PGI. Notice that FT follows
trivially from FT0 and Brouwer’s Thesis.

45
proposition: (FT0 ) Every genetic bar on σ2 contains a finite bar.

Bibliography

[Bishop1967] E. Bishop, Foundations of Constructive Analysis, McGraw-Hill, New York,

1967.
[Bishop&Bridges1985] E. Bishop and D.S. Bridges, Constructive Analysis, Springer, Berlin, 1985.

[Bishop1985] E. Bishop, Schizophrenia in Contemporary Mathematics, Erret Bishop:

Reflections on Him and His Research (ed. Rosenblatt), pp. 1-32; AMS,
Providence, Rhode Island, 1985.
[Bridges1979] D.S. Bridges, Constructive Functional Analysis, Pitman, London, 1979.

[Bridges&Richman1987] D.S. Bridges and F.Richman, Varieties of constructive mathematics, Lon-

don Math. Soc. Lecture Notes no. 93, Cambridge University Press, 1987.
[Brouwer1907] L.E.J. Brouwer, Over de Grondslagen der Wiskunde, PhD thesis, Univer-
siteit van Amsterdam, 1907.
[Brouwer1922] L.E.J. Brouwer, Besitzt jede reelle Zahl eine Dezimalbruch-Entwickelung?,
Math. Annalen 83, 201-210, 1922.
[Brouwer1975] L.E.J. Brouwer, Collected works, (vol. I, II) North-Holland, Amsterdam,
1975.
[vanDalen1999] D. van Dalen, Mystic, Geometer and Intuitionist — The Life of L.E.J.
Brouwer, vol. 1 The Dawning Revolution, Clarendon Press, Oxford, 1999.
[DeFinetti1972] B. De Finetti, Probability, Induction and Statistics, Wiley, New York, 1972.

[DeFinetti1974] B. De Finetti, Theory of Probability — A critical introductory treatment,

(vol. I, II) Wiley, New York, 1974.
[Dugundji1951] J. Dugundji, An extension of Tietze’s theorem, Pacific Journal of Mathe-
matics 1, 353-367, 1951.
[Earman1986] J. Earman, A Primer on Determinism, D. Reidel Publ., Dordrecht, 1986.

[FoM2001] S. Simpson (ed.), Foundations of Mathematics forum,

http://www.cs.nyu.edu/pipermail/fom/1998-November and /2001-
January, 1998-2001.
[Gielen et al. 1981] W. Gielen, H. de Swart and W.H.M. Veldman, The Continuum Hypothesis
in Intuitionism, Journal of Symbolic Logic 46 (1), 121–136, 1981.
[Heyting1956] A. Heyting, Intuitionism — an Introduction, North-Holland, Amsterdam,
1956 (3rd rev. ed. 1971).
[Hill1996] T. Hill, A statistical derivation of the significant-digit law, Statistical Sci-
ence 10, 354–363, 1996.
[Ishihara1993] H. Ishihara, Markov’s Principle, Church’s thesis and Lindelöf ’s theorem,
Indag. Mathem., N.S., 4 (3), 321–325, 1993.

46
[Ishihara1994] H. Ishihara, A constructive version of Banach’s inverse mapping theorem,
New Zealand Journal of Mathematics, vol. 23, 71–75, 1994.
[Johnson1987] D.M. Johnson, L.E.J. Brouwer’s coming of age as a topologist, MAA Stud-
ies in Mathematics, vol. 26, Studies in the History of Mathematics (ed.
E.R. Phillips), 61–97, 1987.
[Kleene1952] S.C. Kleene, Introduction to metamathematics, North-Holland, Amster-
dam, 1952.
[Kleene&Vesley1965] S.C. Kleene and R.E. Vesley, The Foundations of Intuitionistic Mathemat-
ics — especially in relation to the theory of recursive functions, North-
Holland, Amsterdam, 1965.
[Kushner1985] B.A. Kushner, Lectures on Constructive Mathematical Analysis, American
Mathematical Society, Providence, R.I., 1985.
[Michael1956] E. Michael, Continuous selections (I), Annals of Mathematics 63, 361-382,
1956.
[Troelstra&vanDalen1988] A.S. Troelstra and D. van Dalen, Constructivism in Mathematics, (vol. I,
II) North-Holland, Amsterdam, 1988.
[Veldman1981] W.H.M. Veldman, Investigations in intuitionistic hierarchy theory, PhD
thesis, University of Nijmegen, 1981.
[Veldman1985] W.H.M. Veldman, Intuı̈tionistische wiskunde, (lecture notes in Dutch) Uni-
versity of Nijmegen, 1985.
[Waaldijk1996] F.A. Waaldijk, modern intuitionistic topology, PhD thesis, University of
Nijmegen, 1996.

47