Table of Contents

Between Turing and Kleene . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

arXiv:2111.05052v1 [math.LO] 9 Nov 2021

Sam Sanders
1 Between Turing and Kleene computability . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 Short summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Extending the scope of Turing computability . . . . . . . . . . . . . . . . . . 2
1.2.1 A new notion of reduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2.2 Continuous and discontinuous functionals . . . . . . . . . . . . . . . 5
1.3 The need for an extension of Turing computation . . . . . . . . . . . . . . 7
1.3.1 Computing with second-order representations . . . . . . . . . . . . 8
1.3.2 On higher-order computation . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2 Some results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.1 Nets and computability theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.1.1 Nets, a very short introduction . . . . . . . . . . . . . . . . . . . . . . . . 10
2.1.2 Nets and convergence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.1.3 Nets and compactness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.2 On the uncountability of R . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.3 Discontinuous functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.3.1 Grilliot’s trick . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.3.2 Discontinuous functions and N -reduction . . . . . . . . . . . . . . . 17
 Between Turing and Kleene

Sam Sanders

Department of Philosophy II, RUB Bochum,

Universitätsstrasse 150, 44780 Bochum, Germany
[email protected]

Abstract. Turing’s famous ‘machine’ model constitutes the first intu-

itively convincing framework for computing with real numbers. Kleene’s
computation schemes S1-S9 extend Turing’s approach to computing with
objects of any finite type. Both frameworks have their pros and cons and
it is a natural question if there is an approach that marries the best
of both the Turing and Kleene worlds. In answer to this question, we
propose a considerable extension of the scope of Turing’s approach. Cen-
tral is a fragment of the Axiom of Choice involving continuous choice
functions, going back to Kreisel-Troelstra and intuitionistic analysis. Put
another way, we formulate a relation ‘is computationally stronger than’
involving third-order objects that overcomes (many of) the pitfalls of
the Turing and Kleene frameworks.

Keywords: Computability theory, Kleene S1-S9, Turing machines

1 Between Turing and Kleene computability

1.1 Short summary
In a nutshell, we propose a sizable extension of the scope of Turing’s ‘machine’
model of computation ([50]), motivated by a fragment of the Axiom of Choice
involving continuous choice functions, going back to Kreisel-Troelstra and in-
tuitionistic analysis ([22]). In particular, we formulate a relation ‘is computa-
tionally stronger than’ involving third-order objects but still based on Turing
computability by and large.
The interested reader will find the aforementioned extension discussed in
more detail in Section 1.2, along with a critical discussion of the scope of our
extension. The critical reader will learn about the pressing need for the aforemen-
tioned extension in Section 1.3. In particular, the latter section seeks to alleviate
worries that existing frameworks are somehow sufficient for our (foundational)
needs. The (problems involving the) representation of third-order objects via
second-order ones is a particularly important ‘case in point’.
Next, some elegant results in our proposed extension are listed in Section 2
pertaining to the following topics:
– convergence theorems for nets in the unit interval (Section 2.1.2),
2 S. Sanders

– covering theorems for the unit interval R (Section 2.1.3),

– the uncountability of the real numbers R (Section 2.2),
– discontinuous functions on the real numbers R (Section 2.3).
We note that all our results are part of classical mathematics, while we have
found constructive mathematics highly inspiring on our journey towards this
paper. We will assume familiarity with Turing-style computability theory ([44])
and higher-order primitive recursion like in Gödel’s system T ([24, p. 74]); knowl-
edge of Kleene’s higher-order computability theory, in particular the computa-
tion schemes S1-S9 (see [18, 24]), is useful but not essential.
Finally, we will discuss a number of theorems of real analysis and the following
remark discusses how the representations of real numbers can be done in a
straightforward and non-intrusive way.
Remark 1 (Representation of real numbers) Kohlenbach’s ‘hat function’
from [20, p. 289] guarantees that every element of NN defines a real number
via the well-known representation of reals as fast-converging Cauchy sequences.
Despite the definition of the latter being Π10 , a quantifier ‘(∀x ∈ R)’ amounts to
a quantifier over NN .
Moreover, Kohlenbach’s ‘tilde’ function from [21, Def. 4.24] guarantees that
‘(∀x ∈ [0, 1])’ also just amounts to a quantifier over NN , despite 0 ≤R x ≤R 1
being Π10 (in addition). These functions ensure a smooth treatment of R, [0, 1],
and 2N and functions between such spaces. We will always assume that real
numbers and R → R-functions are given in this way, i.e. as in the aforementioned
references [20, 21], so as to ensure a smooth treatment.

1.2 Extending the scope of Turing computability

In this section, we discuss the extension of Turing computability mentioned in
Section 1.1. In particular, we introduce this new concept in Section 1.2.1 and
discuss its scope in Section 1.2.2. The reader will have a basic understanding
of Turing computability theory ([44]) and higher-order primitive recursion like
Gödel’s system T ([24, p. 74]).

1.2.1 A new notion of reduction In this section, we formulate (4), which is

a relation formalising ‘is computationally stronger than’ involving third-order
objects but still based on Turing computability. We first need some preliminaries,
starting with (1).
First of all, many theorems in e.g. analysis can be given the form

(∀Y : NN → N)(∃x ∈ NN )A(Y, x), (1)

where NN is the Baire space and N is the set of natural numbers. Indeed, as
discussed in Remark 1, some basic primitive recursive operations relegate the
coding of real numbers (via elements of NN ) to the background. Moreover, a list
 Between Turing and Kleene 3

of theorems that can be brought in the form (1) can be found in Example 2
below, while we discuss the scope of theorems that can be brought in this form
at the end of this section and in Section 1.2.2.
Secondly, to improve readability, one often uses type theoretic notation in
(1), i.e. n0 for type 0 objects n ∈ N, x1 for type 1 objects x ∈ NN , and Y 2
for type 2 objects Y : NN → N. We will only occasionally need type 3 objects,
which map type 2 objects to natural numbers. We generally use Greek capitals
Θ3 , Λ3 , . . . for such objects.
Thirdly, to compare the logical strength of theorems of the form (1), one
establishes results of the following form over weak systems:

(∀Y 2 )(∃x1 )A(Y, x) → (∀Z 2 )(∃y 1 )B(Z, x), (2)

as part of Kohlenbach’s higher-order Reverse Mathematics (see [20] for an in-

troduction). The computational properties of (1) and (2) following S1-S9 can
then be studied as follows: let Θ3 and Λ3 be realisers for the antecedent and
consequent of (2) i.e. (∀Y 2 )A(Y, Θ(Y )) and (∀Z 2 )B(Z, Λ(Z)).
A central computability theoretic question concerning (2) is whether a re-
aliser Θ3 for the antecedent of (2) computes, in the sense of S1-S9, a realiser Λ3
for the consequent of (2), i.e. whether there is a Kleene algorithm with index
e ∈ N satisfying the following:

(∀Θ3 ) (∀Y 2 )A(Y, Θ(Y )) → (∀Z 2 )B(Z, {e}(Θ, Z)) .

 
(3)

Next, we list some theorems that have been studied via the above paradigm
based on (3) and S1-S9.
Example 2 (Some representative theorems)
– The Lindelöf, Heine-Borel, and Vitali covering theorems involving uncount-
able coverings ([29, 30, 34]),
– The Lebesgue number lemma ([33, 34]),
– The Baire category theorem ([33]),
– Convergence theorems for nets ([39, 40, 42]),
– Local-global principles like Pincherle’s theorem ([34]),
– The uncountability of R and the Bolzano-Weierstrass theorem for countable
sets in Cantor space ([35, 37]),
– Weak fragments of the Axiom of (countable) Choice ([36]).
– Basic properties of functions of bounded variation, like the Jordan decom-
position theorem ([38]).
Many more theorems are equivalent -in the sense of higher-order RM as in [20]-
to the theorems in the above list, as can be found in the associated references.
Fourth, for all the reasons discussed in Section 1.3, we formulate a version of (3)
based on Turing computability as follows:

(∀Z 2 , x1 ) A(t(Z), x) → [{e}s(Z,x) ↓ ∧ B(Z, {e}s(Z,x) )] ,

 
(4)
4 S. Sanders

where s2→1 , t2→2 are terms of Gödel’s T and ‘{e}X ’ is the e-th Turing machine
with oracle X ⊂ N. We note that (4) readily1 implies (3); we discuss the gener-
ality of (4) at the end of this section.
In line with the nomenclature of computability theory, we call the antecedent
and consequent of (2) ‘problems’ and say that
solving the problem (∀Z 2 )(∃y 1 )B(Z, x) N -reduces to solving the problem
(∀Y 2 )(∃x1 )A(Y, x)
in case (4) holds for the parameters mentioned. We view the N -reduction relation
as ‘neutral’ between the Turing and Kleene framework and the reader readily
verifies that N -reduction is transitive. In case the term s(Z, x) can be replaced
by a term u(x), i.e. the latter has no access to Z, we refer to (4) as strong
N -reduction.
Finally, the critical reader may wonder about the generality of (4). The latter
is quite general, for the following two reasons.
– It is an empirical observation based on [29–38] that positive results in S1-
S9 computability theory can be witnessed by terms of Gödel’s T of low
complexity. In this light, there is no real loss of generality if we use terms of
Gödel’s T as in (4).
– A theorem of (third-order) ordinary mathematics generally has the form
(1), unless the former implies the existence of a discontinuous function on
R. In the latter case, an ‘indirect’ treatment is still possible via the so-called
Grilliot’s trick, which we sketch in Section 2.3.1.
Like the reader, we feel that the second item deserves a more detailed expla-
nation, which is in Section 1.2.2. Regarding the first item, intellectual honesty
compels us to admit that many of our S1-S9 results are witnessed by terms of
Gödel’s T additionally involving Feferman’s search operator (already found in
Hilbert-Bernays [16]) defined for any f 1 as:
(
the least n0 such that f (n) = 0 (∃m0 )(f (m) = 0)
µ(f ) := . (5)
0 otherwise

While not strictly necessary always, it is convenient to have access to µ2 as we

then do not have to worry how spaces like [0, 1] or 2N are represented. Based on
this observation, we introduce the following:
solving the problem (∀Z 2 )(∃y 1 )B(Z, x) µN -reduces to solving the problem
(∀Y 2 )(∃x1 )A(Y, x)
in case (4) holds for the parameters mentioned except that t(Z) is replaced by
t(Z, µ2 ). Then ‘strong’ µN -reduction is defined similarly.
Finally, one could study (4) for other extensions of Gödel’s T , e.g. involving
‘minimization’ (see [24, §5.1.5]), but (4) seems more salient.
1
For e ∈ N and s2→1 , t2→2 as in (4), define e0 ∈ N as the Kleene algorithm such that
{e0 }(Θ, Z) := {e}s(Z,Θ(t(Z))) , which is total by assumption.
 Between Turing and Kleene 5

1.2.2 Continuous and discontinuous functionals We discuss the motiva-

tion behind our notion of N -reduction and establish its scope. To this end, we
have to make the following classical case distinction.
– If a given third-order theorem is consistent with Brouwer’s continuity the-
orem that all functions on R are continuous ([8]), then we can directly
analyse it via N -reduction.
– If a given third-order theorem implies the existence of a discontinuous func-
tion on R, we can indirectly analyse it via N -reduction based on Grilliot’s
trick, where the latter is sketched in Section 2.3.1.
To make sense of the above, we first sketch the ‘standard’ higher-order gener-
alisation of (second-order) comprehension, exemplified by Kleene’s ∃2 as in (6).
We then discuss another (less famous) formulation of comprehension, called the
neighbourhood function principle as in Definition 4, a fragment of the Axiom of
Choice involving continuous choice functions, going back to intuitionistic anal-
ysis ([22, 47]).
First of all, the commonplace one cannot fit a round peg in a square hole has
an obvious counterpart in computability theory: a type 2 functional cannot be
the oracle of a Turing machine. Nonetheless, a continuous type 2 functional can
be represented by a type 1 Kleene associate as in Definition 3, where we employ
the same2 notations as in [19]. Associates do ‘fit’ as oracles of Turing machines.

Definition 3 (Kleene associate from [19])

– A function α1 is a neighbourhood function if
• (∀β 1 )(∃n0 )(α(βn) > 0) and
∗ ∗
• (∀σ 0 , τ 0 )(α(σ) > 0 → α(σ ∗ τ ) = α(σ)).
– A function α1 is a (Kleene) associate for Y 2 if
• (∀β 1 )(∃n0 )(α(βn) > 0) and
• (∀β 1 , n0 )(n is least s.t. α(βn) > 0 → α(βn) = Y (β) + 1).
As in [19, §4], we additionally assume that an associate is a neighbourhood func-
tion, as the former can readily be converted to the latter.

Hence, we should specify that a discontinuous type two functional cannot be

the oracle of a Turing machine. Now, the archetypal example of a discontinuous
function is Kleene’s quantifier ∃2 defined as:

(∀f 1 ) (∃n0 )(f (n) = 0) ↔ ∃2 (f ) = 0 .

 
(6)

Clearly, (6) is the higher-order version of arithmetical comprehension (see e.g.

[43, III]) stating that {n ∈ N : A(n)} exists for arithmetical formulas A, which
includes the underlined formula in (6). We point out Grilliot’s trick, a method
∗
2
In particular, σ 0 is a finite sequence in N with length |σ| and we assume the well-
known coding of such finite sequences by natural numbers. Moreover, f n is the finite
∗
sequence (f (0), . . . , f (n − 1)) for any f 1 and n0 , and any f 0 in case |f | ≤ n.
6 S. Sanders

for (effectively) obtaining ∃2 from discontinuous functionals on e.g. R or 2N , as

also discussed in Section 2.3.1. To our own surprise, this kind of effective result
is essentially the prototype of (4), as discussed in Section 2.3.2.
Now, as noted above, Kleene’s ∃2 can decide the truth of arithmetical for-
mulas. In general, for a formula class Γ , one can study higher-order functionals
that decide the truth of formulas γ ∈ Γ . Examples are Kleene’s quantifiers ∃n
([24, Def. 5.4.3]) and the Feferman-Sieg functionals νn from [9, p. 129], which
we shall however not need.
Secondly, we consider the neighbourhood function principle NFP from [48],
studied in [22, 47] under a different name.

Definition 4 [NFP] For any formula A(n0 ), we have

(∀f 1 )(∃n0 )A(f n) → (∃γ ∈ K0 )(∀f 1 )A(f γ(f )), (7)

where ‘γ ∈ K0 ’ means that γ 1 is a (total) Kleene associate.

Clearly, (7) is a fragment of the Axiom of Choice involving continuous choice

functions. Not as obvious is that NFP is a ‘more constructive’ formulation of the
comprehension axiom (see Remark 5 below). We also note that NFP involving
third-order parameters has the form (1), namely for a formula A(n0 , Y 2 ) with
all parameters shown, (7) yields

(∀Y 2 )(∃γ 1 ) (∀f 1 )(∃n0 )A(f n, Y ) → [γ ∈ K0 ∧ (∀f 1 )A(f γ(f ))] .

 
(8)

Now, NFP proves the Lindelöf lemma ([48]), inspiring Theorem 12.
Finally, we can combine the above as follows: the case distinction from the
beginning of this section distinguishes between whether a given theorem T in
the language of third-order arithmetic implies the existence of a discontinuous
function (on R or 2N ), or not. It is then an empirical observation based on [29–38]
that for this theorem T, either the theorem T implies the existence of ∃2 via the
aforementioned Grilliot’s trick, or T is provable from a fragment of NFP where
A may include third-order parameters.
In the former case, the theorem T can be analysed ‘indirectly’ using N -
reduction, namely via Grilliot’s trick, as discussed in Section 2.3. In case the the-
orem T is provable from a fragment NFP (with third-order parameters), we can
generally bring T in the form (1) and hence analyse it directly via N -reduction.
In light of (8), fragments of NFP can always be analysed via N -reduction.
In conclusion, second-order comprehension has been generalised to higher
types in two (more-or-less-known ways) ways, namely as follows.
– Formulate ‘characteristic functionals’ like ∃2 from (6) that decide the truth
of certain formulas.
– Formulate NFP as in Definition 4 for formulas involving higher-order param-
eters and variables.
 Between Turing and Kleene 7

If a given theorem implies the existence of ∃2 , we can analyse it ‘indirectly’ via

N -reduction, namely via Grilliot’s trick. If a given theorem is provable from NFP
involving third-order fragments, we can (readily) analyse it via N -reduction. In
other words, if a third-order theorem is consistent with Brouwer’s continuity
theorem that all functions on R are continuous ([8]), then we can analyse it
directly via N -reduction.

Finally, we show that NFP classically follows from comprehension and vice
versa, assuming a fragment of the induction axiom.

Remark 5 (NFP and comprehension) To obtain NFP from comprehension

∗
modulo coding of finite sequences, let X be such that σ ∈ X ↔ A(σ 0 ) for
∗
any finite sequence σ 0 in N. Then define γ(σ) := |σ| + 1 in case σ ∈ X, and 0
otherwise. Assuming the antecedent of (7), this yields a (total) Kleene associate.
By definition, γ also satisfies the consequent of (7).

To obtain comprehension from NFP, suppose towards a contradiction that

comprehension is false, i.e. there is some formula A(n) such that
 
(∀X ⊂ N)(∃n ∈ N) [n ∈ X ∧ ¬A(n)] ∨ [A(n) ∧ n 6∈ X] . (9)

Now apply NFP to (9) (coding X ⊂ N as elements of 2N ) to obtain γ ∈ K0 . The

latter has an upper bound k0 ∈ N on 2N , i.e. n ∈ N in (9) is bounded by k0 .
However, the induction axiom readily proves ‘finite comprehension’ as follows:
 
(∀k ∈ N)(∃X ⊂ N)(∀n ≤ k) n ∈ X ↔ A(n) . (10)

Hence, for k = k0 + 1, (10) yields a contradiction.

1.3 The need for an extension of Turing computation

We argue why the extension of Turing computation sketched in Section 1.2 is

necessary and even most welcome, as follows.

– Higher-order objects are ‘coded’ as reals so as to accommodate their study

via Turing machines. It has recently been established that this ‘coding prac-
tise’ yields very different results compared to Kleene’s approach, even for
basic objects like functions of bounded variation (Section 1.3.1)
– The conceptual complexity of Kleene’s extension of Turing computability is
considerable, while the extension to ‘infinite time’ Turing machines is too
general for our purposes (Section 1.3.2).

Put another way, N -reduction is an attempt at formulating a relation ‘is com-

putationally stronger than’ for third-order statements that overcomes the above
pitfalls, namely the conceptual complexity of Kleene’s S1-S9 and the problems
associated with second-order representations.
8 S. Sanders

1.3.1 Computing with second-order representations We show that there

are huge differences between ‘computing with higher-order objects’ and ‘com-
puting with representations of higher-order objects’, even for basic objects like
functions of bounded variation on [0, 1].
Now, various3 research programs have been proposed in which higher-order
objects are represented/coded as real numbers or similar representations, so as
to make them amenable to the Turing framework. It is then a natural question
whether there is any significant difference4 between the Kleene S1-S9 approach
or the Turing-approach-via-codes.
Continuous functions being well-studied4 in this context, Dag Normann and
the author have investigated functions of bounded variation, which have at most
countably many points of discontinuity ([38]). A central result is the Jordan
decomposition theorem which implies that f : [0, 1] → R of bounded variation
on [0, 1] satisfies f = g − h on [0, 1] for monotone g, h : [0, 1] → R. We have the
following results.

– In case f : [0, 1] → R of bounded variation is given via a second-order

representation, then the monotone g, h : [0, 1] → R such that f = g − h, can
be computed from finite iterations of the Turing jump with f as a parameter
by [23, Cor. 10].
– A Jordan realiser J takes as input f : [0, 1] → R of bounded variation and
outputs J (f ) = (g, h), i.e. monotone g, h : [0, 1] → R with f = g − h on
[0, 1]. No Jordan realiser is computable (S1-S9) in any type 2 functional by
[37, Theorem 3.9].

Regarding the second item, a Jordan realiser is therefore not computable from
(finite iterations of) ∃2 , the higher-order counterpart of the Turing jump. The
same holds for S2k , which is a type two functional that can decide Πk1 -formulas
(involving first- and second-order parameters). The usual proof of the Jordan
decomposition theorem implies that Kleene’s ∃3 computes a Jordan realiser. But
∃3 implies full second-order arithmetic, and the same holds for the combination
of all S2k .
In conclusion, there is a huge difference in the computational hardness of the
Jordan decomposition theorem depending on whether we use representations or
not. However, this theorem deals with functions of bounded variation, a class
‘very close’ to the class of continuous functions. Hence, (Turing) computing with
3
Examples of such frameworks are: reverse mathematics ([43, 46]), constructive analy-
sis ([3, I.13], [5]), predicative analysis ([13]), and computable analysis ([52]). Bishop’s
constructive analysis is not based on Turing computability directly, but one of its
‘intended models’ is (constructive) recursive mathematics, as discussed in [7]. One
aim of Feferman’s predicative analysis is to capture Bishop’s approach.
4
The fan functional constitutes an early natural example of this difference: it has a
computable code but is not S1-S9 computable (but S1-S9 computable in Kleene’s ∃2
from Section 1.2.2). The fan functional computes a modulus of uniform continuity
for continuous functions on Cantor space; details may be found in [24].
 Between Turing and Kleene 9

representations, interesting as it may be, is completely different from (Kleene)

computing with actual higher-order objects. In this light, there is a clear need
for a notion like N -reduction that allows us to compute with actual higher-order
objects while staying close to Turing computability.

1.3.2 On higher-order computation We argue that the conceptual com-

plexity of Kleene’s S1-S9 is considerable, while the extension to ‘infinite time’
Turing machines is too general for our purposes (Section 1.3.2).
First of all, as noted above, Turing’s famous ‘machine’ model constitutes
the first intuitively convincing framework for computing with real numbers ([50])
while Kleene’s S1-S9 extend Turing’s approach to computing with objects of any
finite type ([18, 24]).
We have studied or made extensive use of Kleene’s S1-S9 computability the-
ory in [29–38]. In our opinion, while vastly more general in scope, Kleene’s S1-S9
has the following conceptual drawbacks.
– Turing computability boasts the elementary ‘Kleene T -predicate’ (see e.g.
[44, p. 15]) where T (e, x, y) intuitively expresses that y codes the compu-
tation steps of the e-th Turing machine program with input x. There is no
such construct for S1-S9.
– Kleene’s recursion theorem is one of the most elegant and important results
in Turing computability ([44, p. 36]) and is derived from first principles. By
contrast, Kleene’s schemes S1-S8 formalise higher-order primitive recursion
(only), while S9 essentially hard-codes the recursion theorem for S1-S9.
– Natural space and time constraints can be formulated for Turing machines,
yielding a canonical complexity theory ([1]); to the best of knowledge, no
such canonical theory exists for higher-order computation in general or S1-
S9 in particular.
– Even basic questions concerning S1-S9 computability theory can be chal-
lenging. We have formulated a most basic example in Section 2.2 concerning
the uncountability of R, arguably one of the most basic properties of the
real numbers, which nonetheless yields very hard problems regarding S1-S9
computability.
In conclusion, the previous items suggest that the much greater scope of S1-S9
comes at the cost of conceptual clarity and causes technical difficulties. It is then
a natural question whether we can find a ‘sweet spot’ between the conceptual
clarity of Turing computability on one hand, and the generality of S1-S9, leading
us to N -reduction.
Secondly, an infinite time Turing machine (ITTM) ([15]) is a generalisa-
tion of Turing computability involving infinite time or space. Welsh provides an
overview in [53] and Dag Normann studies non-montone inductive definitions
and the connection to ITTMs in [28].
In particular, Normann shows that ITTMs can outright compute many of the
functionals introduced in [29, 31, 34], including realisers for the covering lemmas
10 S. Sanders

due to Vitali, Heine-Borel, and Lindelöf. However, all these functionals are not
S1-S9 computable in any type two functional, i.e. the former are ‘hard to com-
pute’ (see [29,31,34]). As a result, ITTMs yield ‘too strong’ a baseline framework
for our purposes.

2 Some results

We establish some results based on our freshly minted notion of N -reduction

from Section 1.2, namely concerning the following topics.

– Convergence theorems for nets (Sections 2.1.2 and 2.1.3).

– Covering theorems (Sections 2.1.3).
– The uncountability of R (Section 2.2).
– Discontinuous functions on R and Grilliot’s trick (Section 2.3).

The below just constitutes an illustrative first collection of examples: we do not

claim our results to be particularly deep or ground-breaking. We do point out
that the above items yield functionals that are, like the Jordan realisers from
Section 1.3.1, hard to compute in that no type 2 functional can (S1-S9) compute
them, while ∃3 can.
Finally, the curious reader of course wonders what the counterpart of the
Turing jump is for N -reduction. We believe this to be the ‘J’ operation discussed
in Section 2.1.2.

2.1 Nets and computability theory

We study basic properties of nets via N -reduction. Nets are a generalisation of

sequences, and the latter hark back to the early days of computability theory
([45]). Filters provide an alternative to nets, but will not be discussed here for
reasons discussed in Remark 9.

2.1.1 Nets, a very short introduction Nets are the generalisation of the
concept of sequence to possibly uncountable index sets, nowadays called nets
or Moore-Smith sequences. These were first described in [25] and then formally
introduced by Moore and Smith in [26] and by Vietoris in [51]. These authors also
established the generalisation to nets of various basic theorems due to Bolzano-
Weierstrass, Dini, and Arzelà ([26, §8-9] and [51, §4]).
One well-know application is the formulation of fundamental topological no-
tions like compactness in terms of nets, as pioneered in [4], while Kelley’s text-
book [17] is standard. Tukey’s monograph [49] builds a similar framework, based
on very specific nets, called phalanxes, where the index sets consist of finite
subsets ordered by inclusion. We now list some basic definitions.

Definition 6 A set D 6= ∅ with a binary relation ‘’ is directed if

 Between Turing and Kleene 11

a.  is transitive, i.e. (∀x, y, z ∈ D)([x  y ∧ y  z] → x  z),

b. for x, y ∈ D, there is z ∈ D such that x  z ∧ y  z,
c.  is reflexive, i.e. (∀x ∈ D)(x  x).

For a directed set (D, ) and a topological space X, any mapping x : D → X

is a net in X. We denote λd.x(d) as ‘(xd )d∈D ’ or ‘xd : D → X’ to suggest
the connection to sequences. The directed set (D, ) is not always explicitly
mentioned together with a net xd : D → X.

The following definitions readily generalise from the sequence notion.

Definition 7 [Convergence of nets] If xd : D → X is a net, we say that it

converges to the limit limd xd = y ∈ X if for every neighbourhood U of y, there
is d0 ∈ D such that for all e  d0 , xe ∈ U .

Definition 8 [Increasing nets] A net xd : D → R is increasing if a  b implies

xa ≤R xb for all a, b ∈ D.

Now, we shall mostly use nets where the index set consists of finite sets of real
numbers ordered by inclusion, i.e. Tukey’s ‘phalanxes’ from [49]. As noted in
Remark 1, real numbers can readily be represented via elements of Baire space
using primitive recursive operations. Thus, such phalanxes are essentially nets
indexed by NN . The notion of ‘sub-sequence’ of course generalises to ‘sub-net’
(see e.g. [42]), but we do not need this (slightly technical) notion here.
Finally, we discuss an alternative to nets and why it is not suitable here.

Remark 9 (Nets and filters) For completeness, we discuss the intimate con-
nection between filters and nets. Now, a topological space X is compact if and
only if every filter base has a refinement that converges to some point of X,
which follows by [2, Prop. 3.4].
Whatever the meaning of the previous italicised notions, the similarity to the
Bolzano-Weierstrass theorem for nets is obvious, and not a coincidence: for every
net r, there is an associated filter base B(r) such that if the erstwhile converges,
so does the latter to the same point; one similarly associates a net r(B) to a
given filter base B with the same convergence properties (see [2, §2]).
Hence, filters provide an alternative to nets, but we have chosen to work with
nets for the following reasons, where the second one is the most pressing.

– Nets have a greater intuitive clarity compared to filters, in our opinion, due
to the similarity between nets and sequences.
– Nets are ‘more economical’ in terms of ontology: consider the aforementioned
filter base B(r) associated to the net r. By [2, Prop. 2.1], the base has strictly
higher type than the net. The same holds for r(B) versus B.
– The notion of refinement mirrors the notion of sub-net ([2, §2]). The former
is studied in [41] in the context of paracompactness; the associated results
suggest that the notion of sub-net works better in weak systems.
12 S. Sanders

On a conceptual note, the well-known notion of ultrafilter corresponds to the

equivalent notion of universal net ([2, §3]). On a historical note, Vietoris intro-
duces the notion of oriented set in [51, p. 184], which is exactly the notion of
‘directed set’. He proceeds to prove (among others) a version of the Bolzano-
Weierstrass theorem for nets. Vietoris also explains that these results are part
of his dissertation, written in the period 1913-1919, i.e. during his army service
for the Great War.

2.1.2 Nets and convergence We obtain a first result concerning N -reduction

and convergence theorems for nets. In particular, as promised above, we connect
the latter to the following operation, which is central and seems to play the role
of the Turing jump: for given Y 2 , define

J(Y ) := {n ∈ N : (∃f 1 )(Y (f, n) = 0)}.

We now have Theorem 10 where C is Cantor space ordered via the lexicographic
ordering ≤lex , i.e. the notion of ‘increasing net in C’ is obvious following Defi-
nition 8. We note that subsets of NN or R are given by characteristic functions,
well-known from measure and probability theory and going back one hundred
plus of years ([12]).

Theorem 10 The following strongly N -reduce to one and other:

– for all Y 2 , there is X ⊂ N such that X = J(Y ),

– a monotone net in C indexed by Baire space, has a limit.

Proof. To show that the second item strongly N -reduces to the first one, let
fd : D → C be an increasing net in C indexed by Baire space and consider the
∗
formula (∃d ∈ D)(fd ≥lex σ ∗ 00 . . . ), where σ 0 is a finite binary sequence. The
latter formula is equivalent to a formula of the form (∃g 1 )(Y (g, n) = 0) where
Y has the form t(λd.fd , n) for a term t of Gödel’s T . Now use J(Y ) to define
the limit f = limd fd , as follows: f (0) is 1 if (∃d ∈ D)(fd ≥lex 100 . . . ) and zero
otherwise. One then defines f (n + 1) in terms of f n in the same way. Note that
we only used J(Y ) to define f , i.e. we have a strong N -reduction.
∗
For the remaining case, fix some Y 2 and let w1 be a sequence of ele-
ments in NN . Define fw : D → C as fw := λk.F (w, k) where F (w, k) is 1 if
∗
(∃i < |w|)(Y (w(i), k) = 0), and zero otherwise. Then λw1 .fw is a monotone net
(phalanx) in C indexed by Baire space (modulo coding). In case limw fw = f ,
then it is readily verified that:

(∀n0 ) (∃g 1 )(Y (g, n) = 0) ↔ f (n) = 1 .

 
(11)
∗
In the notation of (4), the net λw1 fw has the form t(Y )(w) while s does not
depend on Y , i.e. we have a strong N -reduction. ⊓
⊔

The reader is warned that not all N -reduction results are as elegant.
 Between Turing and Kleene 13

2.1.3 Nets and compactness We connect the Heine-Borel theorem and

convergence theorems for nets via N -reduction.
First of all, the Heine-Borel theorem, aka Cousin’s lemma, ([6, 11]) pertains
to open-cover compactness, which we study for the unit interval. Clearly, each
Ψ : [0, 1] → R+ yields a ‘canonical’ covering ∪x∈[0,1] B(x, Ψ (x)), which must have
a finite sub-covering. This yields the principle HBU, which has the form (1).

(∀Ψ : [0, 1] → R+ )(∃x0 , . . . , xk ∈ [0, 1]) [0, 1] ⊂ ∪i≤k B(xi , Ψ (xi )) .

(HBU)

The reals in HBU are hard to compute (S1-S9) in terms of Ψ , as shown in

[29, 30], as no type two functional can perform this task. Computing a Lebesgue
number 5 is similarly hard as shown in [34]. Nonetheless, HBU seems stronger
than the Lebesgue number lemma expressing that a Lebesgue number exists for
any Ψ : [0, 1] → R+ . We believe that Theorem 11 expresses this fundamental
difference.
Theorem 11
– HBU µN -reduces to: for a monotone convergent net in [0, 1] indexed by Baire
space, there is a modulus6 of convergence.
– The Lebesgue number lemma strongly µN -reduces to: a monotone net in
[0, 1] indexed by Baire space, has a limit.
∗
Proof. For the first part, fix Ψ : [0, 1] → R+ and define the following where w1
is a finite sequence of reals:
(
1 (∀q ∈ Q ∩ [0, 1])(q ∈ ∪i<|w| B(w(i), Ψ (w(i))))
xw := .
B(w)/2 otherwise

Here, B(w) is the left-most end-point in [0, 1] of the intervals of the form
B(w(i), Ψ (B(w(i)))) for i ≤ k that is not covered by the union. Note that B(w)
∗
and xw are readily defined using µ2 . Modulo coding of reals, λw1 .xw can be
viewed as a monotone net (phalanx) indexed by Baire space and we must have
limw xw = 1. If (wk )k∈N is a modulus of convergence, then |xw2 − 1| < 41 by
definition, implying xw2 = 1. Hence, ∪i<|w2 | B(w(i), Ψ (w(i))) covers [0, 1] ∩ Q.
Now adjoin to w2 all the points w2 (i) ± Ψ (w2 (i)) for i < |w2 |, to obtain a cover-
ing of [0, 1]. This ‘adjoining’ takes the form of s(Ψ, w2 ) while xw takes the form
t(Ψ, µ2 )(w) for terms s, t of Gödel’s T , using the notation from (4).
For the second part, replace the output 1 by 34 + 2N1+3 in the first case of xw ,
where N is as follows: adjoin to w all the points w(i) ± Ψ (w(i)) for i < |w|, to
5
The notion of Lebesgue number is familiar from topology (see e.g. [27, p. 175]) and
amounts to the following: for a metric space (X, d) and an open covering O of X,
the real number δ > 0 is a Lebesgue number for O if every subset Y of X with
diam(Y ) := supx,y∈Y d(x, y) < δ is contained in some member of the covering.
6
A modulus of convergence for a net xd : D → R with limd xd = x is a sequence
(dk )k∈N with (∀k ∈ N)(∀d  dn )(|xd − x| < 21k ).
14 S. Sanders

obtain a covering of [0, 1]. Now use µ2 to find N ∈ N such that 21N is a Lebesgue
number for the latter covering. Note that the modified net is still monotone
as extending w can only increase the associated Lebesgue number. Clearly, any
cluster point of the modfied net is found in ( 34 , 1). A straightforward unbounded
search can now recover a Lebesgue number from the cluster point of the net
without access to Ψ , i.e. we have a strong µN -reduction. ⊓
⊔

In light of the first part of the previous proof, the ‘post-processing’ term s in (4)
seems necessary as a Turing machine cannot evaluate a third-order functional
at a given point due to type restrictions.

As shown in [42], the existence of a modulus of convergence as in the first

item of the theorem requires a fragment of the Axiom of Choice (AC) beyond
ZF. In fact, one readily shows that the former existence statement N -reduces
(and vice versa) to the following fragment of AC:

(∀Y 2 ) (∀n0 )(∃f 1 )(Y (f, n) = 0) → (∃Z 0→1 )(∀n0 )(Y (Z(n), n) = 0) ,
 

where we exclude the trivial case (∃f 1 )(∀n0 )(Y (f, n) = 0).

Finally, we connect the Lebesgue number lemma and NFP as follows.

Theorem 12 The Lebesgue number lemma strongly N -reduces to NFP for A(n) ≡
(∃f 1 )(Y (f, n) = 0) for any Y 2 .

Proof. By Remark 1, quantifying over 2N or [0, 1] amounts to nothing more

than quantifying over Baire space. To see this, define b : NN → 2N as follows:
b(f )(n) := 0 if f (n) = 0, and 1 otherwise. Also, define r(f ) := ∞ b(f )(n)
P
n=0 2n as
N +
the real in [0, 1] coded by f ∈ N . For Ψ : R → R , the following formula is
trivial (take g = f and large n):
1
(∀f ∈ NN )(∃n ∈ N) (∃g ∈ NN )[B(r(f ),
 
2n ) ⊂ B(r(g), Ψ (r(g)))] ,

which merely expresses that for every x ∈ [0, 1], there is n ∈ N and y ∈ [0, 1]
such that B(x, 21n ) ⊂ B(y, Ψ (y)). Applying NFP with parameter Ψ , we obtain
γ ∈ K0 such that
1
(∀f ∈ NN ) (∃g ∈ NN )[B(r(f ),
 
2γ(f )
) ⊂ B(r(g), Ψ (r(g)))] .

Now compute an upper bound for γ on 2N , using the Kleene associate for the fan
functional ([24, §8.3.2]). This upper bound yields the required Lebesgue number,
which only depends on γ 1 , not on Ψ , i.e. we have obtained a strong N -reduction.
⊓
⊔

We conjecture that HBU does not strongly µN -reduce to the fragment of NFP
from Theorem 12.
 Between Turing and Kleene 15

2.2 On the uncountability of R

We study one of the most (in)famous properties of R, namely its uncountability,

established by Cantor in 1874 as part of his/the first set theory paper ([10]).
The following two principles were first studied in [35, 37].

– NIN: there is no injection from [0, 1] to N.

– Cantor’s theorem: for a set A ⊂ [0, 1] and Y : [0, 1] → N injective on A,
there is x ∈ [0, 1] \ A).

A trivial manipulation of definitions shows that NIN and Cantor’s theorem are
logicially equivalent. We however have the following theorem and associated
Conjecture 14.

Theorem 13

– The problem NIN N -reduces to the Heine-Borel theorem HBU.

– Cantor’s theorem N -reduces to the Heine-Borel theorem HBU restricted to
Baire class 2 functions.
1
Proof. For the first part, fix Z : [0, 1] → N and define t(Z)(x) := 2Z(x)+1 moti-
vated by the notation in (4). In case x0 , . . . , xk ∈ [0, 1] is a finite sub-covering of
∪x∈[0,1] B(x, t(Z)(x)), there are i, j ≤ k with

Z(xi ) = Z(xj ) ∧ xi 6= xj . (12)

Indeed, in case there are no i, j ≤ k as in (12), then the measure of ∪i≤k B(xi , t(Z)(xi ))
Pk 1
is at most n=0 2i+1 < 1, contradicting the fact that ∪i≤k B(xi , t(Z)(xi )) cov-
ers [0, 1]. In light of (12), given the finite sequence s(Z, x0 , . . . , xk ) defined as
x0 , t(Z)(x0 ), . . . , xk , t(Z)(xk ), we can perform an unbounded search (on a Tur-
ing machine) to find i, j ≤ k and k ∈ N such that t(Z)(xi ) =Q t(Z)(xj ) and
[|xi − xj |](k) >Q 21k , where [z](m) is the approximation of z ∈ R up to 2m+1 1
.
Hence, we also obtain the consequent of (4) for the case at hand.
For the second part, fix A ⊂ [0, 1] and Y : [0, 1] → N such that Y is injective
on A. Now consider the following:
(
1
Y (x)+5 x∈A
t(Y, A)(x) := 12 .
8 x 6∈ A

One readily shows that t(Y, A) : R → R is Baire class 2, as it only has countably
many points of discontinuity by definition. For a finite sub-covering x0 , . . . , xk ∈
[0, 1] of ∪x∈[0,1] B(x, t(Y, A)(x)), there must be j ≤ k, with xj 6∈ A. Indeed, as
in the previous paragraph, the measure of ∪i≤k B(xi , t(Y, A)(xi )) is otherwise
Pk 1
at most n=0 2i+5 < 1, a contradiction. One can effectively decide whether
1 1
t(Y, A)(xi ) < 8 or t(Y, A)(xi ) > 16 for i ≤ k, i.e. one readily finds a j ≤ k with
xj 6∈ A. ⊓
⊔
16 S. Sanders

In light of the previous proof, the ‘post-processing’ term s in (4) again seems
necessary as a Turing machine cannot evaluate a third-order functional at a
point due to type restrictions.
Based on the previous proof, we conjecture the following.
Conjecture 14 The problem NIN does not N -reduce to the Heine-Borel theo-
rem HBU restricted to Baire class 2 functions, nor to the (full ) Lebesgue number
lemma.

2.3 Discontinuous functions

We show that a representative equivalence from the Reverse Mathematics lit-
erature involving (∃2 ) gives rise to N -reductions between the members of the
equivalence. That N -reduction applies here was surprising to us, as the exis-
tence of a discontinuous function like ∃2 does not have the syntactic form (1).
A central role is played by Grilliot’s trick, a method for (effectively) obtaining
∃2 from a discontinuous function ([14]). We discuss this trick in some detail
in Section 2.3.1, while the connection between this trick and N -reduction is
discussed in Section 2.3.2.

2.3.1 Grilliot’s trick In a nutshell, Grilliot’s trick is a method for effec-

tively obtaining ∃2 from a discontinuous function, say on NN or R. Clearly, ∃2
is discontinuous at 11 . . . , making the former functional a kind of ‘canonical’
discontinuous function.
First of all, Grilliot’s paper [14] pioneers the aforementioned method, nowa-
days called Grilliot’s trick; we refer to [24, Remark 5.3.9] for a discussion of the
general background and history. We note that Kohlenbach formalises Grilliot’s
trick in a weak logical system (namely his ‘base theory’ RCAω 0 ) in [20, §3].

Secondly, Kohlenbach’s rendition of Grilliot’s trick ([20, §3]) is quite easy to

understand conceptually. Indeed, assume we have a function F : R → R and a
sequence (xn )n∈N with limn→∞ xn = x such that limn→∞ F (xn ) 6= F (x), i.e. F
is not sequentially continuous at x. Then there is a term t3 of Gödel’s T of low
complexity such that E(f ) := λf 1 .t(F, λn.xn , x, f ) is Kleene’s ∃2 as in (6). All
technical details, including the exact definition of t, are found in [20, §3].
Thirdly, Kohlenbach uses Grilliot’s trick in [20, §3] to show that e.g. the
following sentence implies the existence of ∃2 :

(∃ε)(∀g ∈ L([0, 1]))[ε(g) ∈ [0, 1] ∧ (∀y ∈ [0, 1])(g(y) ≤ g(ε(g)))]. (13)

Here, ε(g) is a real in [0, 1] where the Lipschitz-continuous7 function g : [0, 1] → R

with constant 1 attains its maximum. We note that the underlined quantifiers can
7
A function g : [0, 1] → R is Lifschitz-continuous with constant 1 on [0, 1] if (∀x, y ∈
[0, 1])(|g(x)−g(y)| < |x −y|). Hence, to (effectively) recover the graph of g, it suffices
to have access to the sequence (g(q))q∈Q∩[0,1].
 Between Turing and Kleene 17

be brought in the form7 ‘(∃Y 2 )(∀α1 )’, which can also be obtained by representing
continuous functions via second-order codes.

2.3.2 Discontinuous functions and N -reduction In this section, we dis-

cuss the connection between Grilliot’s trick from Section 2.3.1 and N -reduction.
In particular, we show that the proof of [20, Prop. 3.14], establishing the equiv-
alence (13) ↔ (∃2 ), gives rise to N -reductions involving (13) and (∃2 ).
First of all, consider (13) from Section 2.3.1. The functional ε from (13) yields
a discontinuous function on R, which yields ∃2 in turn, following the proof of
[20, Prop. 3.14]. If we make all steps in the latter proof explicit8 , we obtain a
term t of Gödel’s T of low complexity such that
(∀ε) (∀g)A(g, ε(g)) → (∀f 1 )B(t(ε), f ) ,
 
(14)
where A(g, x) expresses that x ∈ [0, 1] is a real where the Lipschitz-continuous
function g : [0, 1] → R with Lipschitz constant 1 attains its maximum; the
formula (∀f 1 )B(∃2 , f ) is (6), i.e. the specification of ∃2 . Clearly, (14) implies by
contraposition that:
(∀ε, f 1 ) ¬B(t(ε), f ) → (∃g)¬A(g, ε(g)) ,
 
(15)
which is ‘almost’ the definition of N -reduction as in (4). Indeed, ‘(∃g)’ in (15)
is essentially a quantifier over R by Footnote 7, whence (∀ε) can be viewed as
a quantifier (∀Z 2 ). Furthermore, a detailed inspection of the proof that (13)
implies the existence of ∃2 in [20, Prop. 3.14], reveals the following: this proof
still goes through if we restrict (13) to a sentence of the form:
(∃ε)(∀n0 ))[ε(gn ) ∈ [0, 1] ∧ (∀q ∈ [0, 1] ∩ Q)(gn (q) ≤ gn (ε(gn )))], (16)
9
for some effective sequence of functions (gn )n∈N all in L([0, 1]). In case the
formula in square brackets in (16) is false for some n ∈ N, an unbounded search
will yield this number. Hence, we can replace ‘(∃g)¬A(g, ε(g)’ in (15) by
{e}s(ε,f ) ↓ ∧ ¬A({e}s(ε,f ) , ε({e}s(ε,f ) )

for some index e ∈ N and term s of Gödel’s T , which is exactly (4). The details
are somewhat tedious, but we nonetheless can say that the negation of (13)
N -reduces to the negation of (∃2 ).
Finally, the usual ‘interval-halving’ proof of the existence of a maximum of a
continuous function on [0, 1], can be done using ∃2 , yielding a term t of Gödel’s
T such that:
(∀E 2 ) (∀f 1 )B(E, f ) → (∀g)A(g, t(E)(g)) .
 
(17)
8
The construction of a discontinuous function on R in the proof of [20, Prop. 3.14]
depends on whether ε(g0 ) ∈ [0, 21 ] or ε(g0 ) ∈ [ 12 , 1], where g0 is the constant 0
function and ε as in (13). This non-effective case distinction can be replaced by an
effective case distinction whether ε(g0 ) < 34 or ε(g0 ) > 14 . The proof in the first case
goes through unmodified, while one replaces yx in the second case by −yx.
9
The join of the sequences (qx − q)q∈Q∩[0,1] and (−qx)q∈Q∩[0,1] suffices.
18 S. Sanders

The contraposition of (17) then has the same form as (15). One readily obtains
an index e ∈ N and term s of Gödel’s T with

(∀E 2 , g) ¬A(g, t(E)(g)) → [{e}s(E,g) ↓ ∧ ¬B(E, {e}s(E,g) ) ,

 
(18)

as one only needs to decide g(r) ≥ g(q) for r, q ∈ [0, 1] ∩ Q to find a maximum
of a (Lipschitz) continuous function g : [0, 1] → R. Hence, an unbounded search
on a Turing machine will find f 1 with ¬B(E, f ). We note that (18) is a case of
N -reduction of the negation of (∃2 ) to the negation of (13).
In conclusion, we observe that the negation of (∃2 ) will N -reduce to the
negation of (13), and vice versa. Thus, it perhaps makes sense to drop the
‘negation of’ here and distinguish between (1) and its negation in the definition
of N -reduction.

Acknowledgement 15 I thank Anil Nerode for his most helpful advise. My

research was kindly supported by the Deutsche Forschungsgemeinschaft via the
DFG grant SA3418/1-1. I thank the anonymous referees for their suggestions,
which have greatly improved this paper.

Bibliography
[1] Sanjeev Arora and Boaz Barak, Computational complexity. A modern approach.,
Cambridge University Press, 2009.
[2] Robert G. Bartle, Nets and filters in topology, Amer. Math. Monthly 62 (1955),
551–557.
[3] Michael J. Beeson, Foundations of constructive mathematics, Ergebnisse der Math-
ematik und ihrer Grenzgebiete, vol. 6, Springer, 1985.
[4] Garrett Birkhoff, Moore-Smith convergence in general topology, Ann. of Math. (2)
38 (1937), no. 1, 39–56.
[5] Errett Bishop, Foundations of constructive analysis, McGraw-Hill, 1967.
[6] E. Borel, Leçons sur la théorie des fonctions, Gauthier-Villars, Paris, 1898.
[7] Douglas Bridges and Fred Richman, Varieties of constructive mathematics, Lon-
don Mathematical Society Lecture Note Series, vol. 97, Cambridge University
Press, Cambridge, 1987.
[8] L. E. J. Brouwer, Collected works. Vol. 1, North-Holland Publishing Co., Amster-
dam, 1975. Philosophy and foundations of mathematics; Edited by A. Heyting.
[9] Wilfried Buchholz, Solomon Feferman, Wolfram Pohlers, and Wilfried Sieg, Iter-
ated inductive definitions and subsystems of analysis, LNM 897, Springer, 1981.
[10] Georg Cantor, Ueber eine Eigenschaft des Inbegriffs aller reellen algebraischen
Zahlen, J. Reine Angew. Math. 77 (1874), 258–262.
[11] Pierre Cousin, Sur les fonctions de n variables complexes, Acta Math. 19 (1895),
1–61.
[12] Lejeune P. G. Dirichlet, Über die Darstellung ganz willkürlicher Funktionen durch
Sinus- und Cosinusreihen, Repertorium der physik, von H.W. Dove und L. Moser,
bd. 1, 1837.
[13] Solomon Feferman, How a Little Bit goes a Long Way: Predicative Foundations
of Analysis, 2013. unpublished notes from 1977-1981 with updated introduction,
https://math.stanford.edu/∼feferman/papers/pfa(1).pdf .
 Between Turing and Kleene 19

[14] Thomas J. Grilliot, On effectively discontinuous type-2 objects, J. Symbolic Logic

36 (1971), 245–248.
[15] Joel D. Hamkins and Andy Lewis, Infinite time Turing machines, Journal of Sym-
bolic Logic 65 (1998), 567-604.
[16] David Hilbert and Paul Bernays, Grundlagen der Mathematik. II, Zweite Auflage.
Die Grundlehren der mathematischen Wissenschaften, Band 50, Springer, 1970.
[17] John L. Kelley, General topology, Springer-Verlag, 1975. Reprint of the 1955 edi-
tion; Graduate Texts in Mathematics, No. 27.
[18] Stephen C. Kleene, Recursive functionals and quantifiers of finite types. I, Trans.
Amer. Math. Soc. 91 (1959), 1–52.
[19] Ulrich Kohlenbach, Foundational and mathematical uses of higher types, Reflec-
tions on the foundations of mathematics, Lect. Notes Log., vol. 15, ASL, 2002,
pp. 92–116.
[20] , Higher order reverse mathematics, Reverse mathematics 2001, Lect. Notes
Log., vol. 21, ASL, 2005, pp. 281–295.
[21] , Applied proof theory: proof interpretations and their use in mathematics,
Springer Monographs in Mathematics, Springer-Verlag, Berlin, 2008.
[22] G. Kreisel and A. S. Troelstra, Formal systems for some branches of intuitionistic
analysis, Ann. Math. Logic 1 (1970), 229–387.
[23] Alexander P. Kreuzer, Bounded variation and the strength of Helly’s selection
theorem, Log. Methods Comput. Sci. 10 (2014), no. 4, 4:16, 15.
[24] John Longley and Dag Normann, Higher-order Computability, Theory and Appli-
cations of Computability, Springer, 2015.
[25] E. H. Moore, Definition of Limit in General Integral Analysis, PNAS 1 (1915),
no. 12, 628–632.
[26] E. H. Moore and H. Smith, A General Theory of Limits, Amer. J. Math. 44 (1922),
102–121.
[27] James R. Munkres, Topology, Prentice-Hall, 2000, 2nd edition.
[28] Dag Normann, Computability And Non-Monotone Induction, Submitted, arxiv:
https://arxiv.org/abs/2006.03389 (2020), pp. 41.
[29] Dag Normann and Sam Sanders, Nonstandard Analysis, Computability Theory,
and their connections, Journal of Symbolic Logic 84 (2019), no. 4, 1422–1465.
[30] , The strength of compactness in Computability Theory and Nonstandard
Analysis, Annals of Pure and Applied Logic 170 (2019), no. 11.
[31] , On the mathematical and foundational significance of the uncount-
able, Journal of Mathematical Logic, https://doi.org/10.1142/S0219061319500016
(2019).
[32] , Representations in measure theory, Submitted, arXiv:
https://arxiv.org/abs/1902.02756 (2019).
[33] , Open sets in Reverse Mathematics and Computability Theory, Journal of
Logic and Computation 30 (2020), no. 8, pp. 40.
[34] , Pincherle’s theorem in reverse mathematics and computability theory,
Ann. Pure Appl. Logic 171 (2020), no. 5, 102788, 41.
[35] , On the uncountability of R, Submitted, arxiv:
https://arxiv.org/abs/2007.07560 (2020), pp. 37.
[36] , The Axiom of Choice in Computability Theory and Reverse Mathematics,
Journal of Logic and Computation 31 (2021), no. 1, 297-325.
[37] , On robust theorems due to Bolzano, Weierstrass, and Cantor in Reverse
Mathematics, See https://arxiv.org/abs/2102.04787 (2021), pp. 30.
[38] , Betwixt Turing and Kleene, Submitted, arxiv:
https://arxiv.org/abs/2109.01352 (2021), pp. 15.
20 S. Sanders

[39] Sam Sanders, Nets and Reverse Mathematics: initial results, Lecture notes in Com-
puter Science 11558, Proceedings of CiE19, Springer (2019), 253-264.
[40] , Reverse Mathematics and computability theory of domain theory, Lecture
notes in Computer Science 11541, Proceedings of WoLLIC19, Springer (2019),
550-568.
[41] , Reverse Mathematics of topology: dimension, paracompactness, and split-
tings, Notre Dame Journal for Formal Logic 61 (2020), no. 4, 537-559.
[42] , Nets and Reverse Mathematics: a pilot study, Computability 10 (2021),
no. 1, 31-62.
[43] Stephen G. Simpson, Subsystems of second order arithmetic, 2nd ed., Perspectives
in Logic, CUP, 2009.
[44] Robert I. Soare, Recursively enumerable sets and degrees, Perspectives in Mathe-
matical Logic, Springer, 1987.
[45] Ernst Specker, Nicht konstruktiv beweisbare Sätze der Analysis, J. Symbolic Logic
14 (1949), 145–158 (German).
[46] J. Stillwell, Reverse mathematics, proofs from the inside out, Princeton Univ.
Press, 2018.
[47] A. S. Troelstra, Choice sequences, Clarendon Press, Oxford, 1977. A chapter of
intuitionistic mathematics; Oxford Logic Guides.
[48] Anne Sjerp Troelstra and Dirk van Dalen, Constructivism in mathematics. Vol.
I, Studies in Logic and the Foundations of Mathematics, vol. 121, North-Holland,
1988.
[49] John W. Tukey, Convergence and Uniformity in Topology, Annals of Mathematics
Studies, no. 2, Princeton University Press, Princeton, N. J., 1940.
[50] Alan Turing, On computable numbers, with an application to the Entscheidungs-
problem, Proceedings of the London Mathematical Society 42 (1936), 230-265.
[51] Leopold Vietoris, Stetige Mengen, Monatsh. Math. Phys. 31 (1921), no. 1, 173–204
(German).
[52] Klaus Weihrauch, Computable analysis, Springer-Verlag, Berlin, 2000.
[53] P. D. Welch, Transfinite machine models, Turing’s legacy: developments from Tur-
ing’s ideas in logic, Lect. Notes Log., vol. 42, Assoc. Symbol. Logic, La Jolla, CA,
2014, pp. 493–529.