Fundamenta Informaticae XX (2006) 125 1

IOS Press
Recursive Analysis Characterized as a Class of Real
Recursive Functions
Olivier Bournez
INRIA
LORIA (UMR 7503 CNRS-INPL-INRIA-Nancy2-UHP)
Campus scientique, BP 239, 54506 Vanduvre-L`es-Nancy, FRANCE
Olivier.Bournez@loria.fr
Emmanuel Hainry
INPL
LORIA (UMR 7503 CNRS-INPL-INRIA-Nancy2-UHP)
Campus scientique, BP 239, 54506 Vanduvre-L`es-Nancy, FRANCE
Emmanuel.Hainry@loria.fr
Abstract. Recently, using a limit schema, we presented an analog and machine independent
algebraic characterization of elementary functions over the real numbers in the sense of
recursive analysis.
In a dierent and orthogonal work, we proposed a minimalization schema that allows to
provide a class of real recursive functions that corresponds to extensions of computable
functions over the integers.
Mixing the two approaches we prove that computable functions over the real numbers in
the sense of recursive analysis can be characterized as the smallest class of functions that
contains some basic functions, and closed by composition, linear integration, minimalization
and limit schema.
1. Introduction
Recursive analysis, also called computable analysis, has been introduced by Turing [35], Grze-
gorczyk [16], Lacombe [19]. It has shown to provide a very robust concept of computability,
that enables to discuss most arguments of mathematical analysis from the computability point
of view: see e.g. monograph [36].
2 O. Bournez, E. Hainry / Recursive Analysis Characterized as a Class of Real Recursive Functions
In this framework, a function f : R R over the reals is considered as computable, if there
is some computable functional, or Type 2 machine, that maps any sequence quickly converging
to some x to a sequence quickly converging to f(x), for all x. That means that this notion of
computability requests a priori to deal with functionals, or higher order Turing machines.
In a recent work [4, 6], extending some classes proposed by Campagnolo, Moore and Costa
[10, 11, 12] by a suitable limit schema, we proved that a particular subclass of computable
functions over the reals can be characterized algebraically in a machine independent way. Indeed,
elementary functions in the sense of recursive analysis were characterized as the smallest class
of functions that contains some basic functions, and closed by composition, linear integration,
and a simple limit schema.
This results was obtained by using deeply a result from Campagnolo, Moore, and Costa
[10, 11, 12] characterizing algebraically functions over the reals that extend elementar functions
over the integers.
However, an algebraic and machine characterization of the whole class of computable func-
tions over the real numbers was missing. If we were to follow the steps of the arguments of [4, 6],
the point was rst to be able to nd a minimalization schema that could provide a result similar
to the previous one for computable functions over the integers, and second to understand how
and whether it could be arranged with the arguments of [4, 6], to provide such a characterization.
The rst step was solved recently in paper [5]. This journal paper presents detailed proofs
of the claims of [5], with several extensions. In particular, it characterizes also non-total func-
tions. More importantly, it proves that this is indeed possible to do the second step: mix these
constructions with the ones of [4, 6] to get a characterization of the whole class of computable
functions over the reals. This is done by extending the constructions of [4], and in particular
provides extensions of [4, 6] that allow to talk about functions dened on non-compact domains.
Indeed, computable functions over the reals are characterized in an algebraic and machine
independent way as the smallest class of functions that contains some basic functions, and closed
by composition, linear integration, minimalization and limit schema.
This result has several consequences. First, that proves that it is possible to dene com-
putability in the sense of recursive analysis in a machine independent way, avoiding to talk
about higher order Turing machines, or functionals, nor less natural characterization such as [8].
Second, that proves that the study of mathematical concepts through recursive analysis
can be investigated by talking in terms of these algebraic classes and operators, providing a
rather natural continuous setting to deal with continuous problems, instead of needing to discuss
continuous problems with discrete models.
Third, it provides strong connections with several analog models. Indeed, the classes from
Campagnolo, Costa and Moore, are inspired from a class of functions over the reals, called real
recursive functions, introduced by Moore in [21]. Real recursive functions have been shown
(see [21] with corrections from Graca and Costa in [15]) to be strongly connected to functions
computable by the General Purpose Analog Computer (GPAC) of Shannon [33]. GPAC is in
turn an abstraction of some systems that really existed [34, 9, 7], or is an abstraction of easy
realizable systems using todays electronic. Extensions of the GPAC have been discussed in [32]
and [20].
Fourth, these results show that the provided class of functions does not exhibit super-Turing
phenomena such as [21, 3, 2, 17, 14], and benets from all the robustness results that have been
O. Bournez, E. Hainry / Recursive Analysis Characterized as a Class of Real Recursive Functions 3
established for computable functions in recursive analysis.
The paper is organized as follows. In Section 2, we recall some basic mathematical properties
that we will use, as well as basic denitions from classical recursion theory. Section 3 recalls
recursive analysis. Section 4 recalls some results established by Campagnolo, Costa and Moore.
Section 5 introduced our proposed minimalization schema, and shows that adding this schema
provides a class that corresponds to extensions of computable functions over the integers. Section
6 discuss some alternative to our minimalization schema. Section 7 presents our limit schema,
and proves our main result.
2. Preliminaries
2.1. Mathematical preliminaries
Let Z, N, Q, R, R
0
denote the set of integers, natural integers, the set of rational numbers, the
set of real numbers, and the set of non-negative real numbers respectively. Given x R
n
, we
write

x to emphasize that x is a vector.
Lemma 2.1. (Bounding Lemma for Linear Dierential Equations (see e.g. [1]))
For linear dierential equation

x

= A(t)

x , if A is dened and continuous on interval I = [a, b],

where a 0 b, then, for all

x
0
, the solution of

x

= A(t)

x with initial condition

x (0) =

x
0
is dened and unique on I. Furthermore, we know that the solution satises
|

x (t)| |

x
0
| exp( sup
[0,t]
|A()|t).
Lemma 2.2. (Implicit Functions Theorem (see e.g. [30]))
Let f : T 1 R
k+1
R where T 1 is a product of closed intervals be a function of class
1
(
k
, for k 1. Assume that for all

x T, the equation f(

x , y) = 0 has exactly one solution y

0
and that this y
0
belongs to the interior of 1. Assume for all

x that
f
y
(

x , y
0
) ,= 0
in the corresponding root y
0
. Then function g : R
k
R that maps

x to the corresponding root
y
0
is dened over T and also of class (
k
.
Lemma 2.3. (Basic fact (see e.g. [6]))
Let F : R 1 R
k+1
R
l
be a function of class (
1
, and (

x ) : 1 R, K(

x ) : 1 R be
some continuous functions. Assume that for all t and

x ,
|
F
t
(t,

x )| K(

x ) exp(t(

x )).
Let T be the subset of the

x 1 with (

x ) > 0.
Then,
1
Recall that function f : D R
k
R
l
, k, l N, is said to be of class C
r
if it is r-times continuously dierentiable
on D. It is said to be of class C

if it is of class C
r
for all r.
4 O. Bournez, E. Hainry / Recursive Analysis Characterized as a Class of Real Recursive Functions
for all

x T, F(t,

x ) has a limit L(

x ) in t = +.
Function L(

x ) is a continuous function.
Furthermore
|F(t,

x ) L(

x )|
K(

x ) exp(t(

x ))
(

x )
.
2.2. Classical Recursion Theory
Classical recursion theory deals with functions over integers. Most classes of classical recursion
theory can be characterized as closures of a set of basic functions by a nite number of basic
rules to build new functions [13, 31, 27, 28]: given a set T of functions and a set O of operators
on functions (an operator is an operation that maps one or more functions to a new function),
[T; O] will denote the closure of T by O [13].
Proposition 2.1. (Classical settings: see e.g. [31, 27, 28])
Let f be a function from N
k
to N for k N. Function f is
elementar i it belongs to c = [0, S, U, +, ; COMP, BSUM, BPROD];
primitive recursive i it belongs to T = [0, S, U; COMP, REC];
recursive
2
i it belongs to ec = [0, S, U; COMP, REC, MU].
A function f : N
k
N
l
is elementar (resp: primitive recursive, recursive) i its projections
are elementar (resp: primitive recursive, recursive).
The basic functions 0, (U
m
i
)
i,mN
, S, +, and the operators BSUM, BPROD, COMP, REC,
MU are given by
1. 0 is the constant 0; U
m
i
: N
m
N, U
m
i
: (n
1
, . . . , n
m
) n
i
; S : N N, S : n n + 1;
+ : N
2
N, + : (n
1
, n
2
) n
1
+ n
2
; : N
2
N, : (n
1
, n
2
) max(0, n
1
n
2
);
2. BSUM : bounded sum. Given f, h = BSUM(f) is dened by h : (

x , y)

z<y
f(

x , z);
BPROD : bounded product. Given a function f, the bounded product h = BPROD(f) is
dened by h : (

x , y)

z<y
f(

x , z);
3. COMP : composition. Given f
1
, . . . , f
p
and g, h = COMP(f
1
, . . . , f
p
, g) is dened as the
function verifying h(

x ) = g(f
1
(

x ), . . . , f
p
(

x ));
4. REC : primitive recursion . Given f and g, h = REC(f, g) is dened as the function
verifying h(

x , 0) = f(

x ) and h(

x , n + 1) = g(

x , n, h(

x , n));
5. MU : minimalization. Given a function f, function f is dened on all

x for which there is

a y such that z y, f(

x , z) is dened and f(

x , y) = 0. For such

x , the minimalization
of f is f :

x infy; f(

x , y) = 0.
2
This class is often called partial recursive since it contains partial functions as opposed to the class of total
recursive functions.
O. Bournez, E. Hainry / Recursive Analysis Characterized as a Class of Real Recursive Functions 5
Observe that minimalization operator can actually be reinforced into a unique minimalization
operator as follows:
Proposition 2.2. A function f from N
k
to N
l
, for k, l N, is recursive i its projections belong
to [0, U, S; COMP, REC, UMU] where operator UMU is dened as follows:
1. UMU: unique minimalization. Given f, that satises that for all

x , there is at most one y
with f(

x , y) dened and equal to 0, the unique minimalization of f, denoted by !(f)(

x ),
is dened on all

x for which there is a (unique) y with f(

x , y) = 0. For such

x , !(f)(

x )
is dened as that unique y.
Proof:
The inclusion [0, U, S; COMP, REC, UMU] ec is easy: given f with !(f) dened, given

x ,
!(f)(

x ) can be computed by testing in parallel for all y whether f(

x , y) = 0 until one nds

the correct y, if it exists.
Conversely, let be a function from ec. It is well known [18, 31] that can be written as
= () with and in c and such that for all

x on which is dened, there is at least a
y with (

x , y) = 0. Let be the elementar function dened by (m, n) =

z<n
(m, z). Given
m, let us note n
0
= ()(m). We have n n
0
, (m, n) ,= 0 and n > n
0
, (m, n) = 0.Let
(m, n) = sgn(max(1 (m, n), (m, n + 1))).
We have clearly n < n
0
, (m, n) = 1, (m, n
0
) = 0 and n > n
0
, (m, n) = 1, hence
() =!() = (). is an elementar function and we have = !(), hence belongs to
[0, U, S; COMP, REC, UMU]. .
We have c T ec, and the inclusions are known to be strict [31, 27, 28]. If TIME(t)
and SPACE(t) denote the classes of functions that are computable with time and space t, then,
c = TIME(c), and T = TIME(T) = SPACE(T) [31, 27, 28]. Class T corresponds to
functions computable using For-Next programs. Class c corresponds to computable functions
bounded by some iterate of the exponential function [31, 27, 28]. At most two nested For-Next
loops are required for a function of class c, whereas general functions from class T may require
an arbitrary high number of such nested loops.
In classical computability, more general objects than functions over the integers can be
considered, in particular functionals, i.e. functions : (N
m
)
N
N
k
N
l
. A functional will be
said to be elementarily (or primitive recursively, recursively) computable when it belongs to the
corresponding
3
class.
3
Formally, a function f over the integers can be considered as functional f : (V,

n ) f(

n ). Similarly, an
operator Op on functions f1, . . . , fm over the integers can be extended to an operator over functionals by xing
rst argument Op(F1, . . . , Fm) : (V,

n ) Op(f1(V, .), . . . , fm(V, .))(

n ).
In that spirit, given some set F of basic functions N
k
N
l
and a set O of operators on functions over the integers,
we will still (abusively) denote by [f1, . . . , fp; O1, . . . , Oq] for the smallest class of functionals that contains basic
functions f1, . . . , fp, plus the functional Map : (V, n) Vn, the nth element of sequence V , and which is closed
by the operators O1, . . . , Oq. For example, a functional will be said elementarily computable i it belongs to
E = [Map, 0, S, U, +, ; COMP, BSUM, BPROD].
6 O. Bournez, E. Hainry / Recursive Analysis Characterized as a Class of Real Recursive Functions
3. Computable Analysis
The idea sustaining computable analysis, also called recursive analysis, is to dene computable
functions over real numbers by considering functionals over fast-converging sequences of rationals
[35, 19, 16, 36].
Let
Q
: N Q be the following representation
4
of rational numbers by integers:

Q
(p, r, q))
p r
q + 1
,
where ., ., .) : N
3
N is a computable bijection.
A sequence of integers (x
i
) N
N
represents a real number x if (
Q
(x
i
)) converges quickly
toward x (denoted by (x
i
) x) in the following sense : i, [
Q
(x
i
) x[ < exp(i). For a
sequence of k-tuples (

x
i
) (N
k
)
N
, we write (

x
i
)

x when it holds componentwise.

Note that many sequences can represent the same real number and also that the chosen
bound is arbitrary and could be replaced by another function converging fast toward 0. In
particular, our notion of computability is equivalent to the one of [36], or [12].
Denition 3.1. (Recursive analysis [36])
A function f : R
k
R is said computable (or real computable) if there exists a recursive
functional : (N
k
)
N
N N such that for all

x R
k
, for all sequence X = (

x
n
) (N
k
)
N
,
we have ((X, j))
j
f(

x ) whenever X

x . A function f : R
k
R
l
, with l > 1, is said
computable if all its projections are.
A function f will be said elementary whenever the corresponding functional is elementarily
computable. The class of computable (respectively elementary ) functions over the reals will be
denoted by ec(R) (resp. c(R)).
4. Real sub-recursive and sub-recursive functions
Campagnolo, Moore and Costa proposed in [10, 11, 12] to consider the following class, built in
analogy with elementar functions over the integers.
Denition 4.1. ([12, 11])
Let us dene L as being the class of functions f : R
k
R
l
, for some k, l N, dened by
L = [0, 1, 1, , U,
3
; COMP, LI] where the basic functions 0, 1, 1, , (U
m
i
)
i,mN
,
3
and the
schemata COMP and LI are the following:
1. 0, 1, 1, are the corresponding constants; U
m
i
: R
m
R are, as in the classical settings,
projections: U
m
i
: (x
1
, . . . , x
m
) x
i
;
2.
3
: R R is dened as
3
: x x
3
if x 0, 0 otherwise;
3. COMP: composition is dened as in the classical settings: Given f
1
, f
2
, . . . , f
p
and g,
h = COMP(f
1
, . . . , f
p
; g) is dened by h(

x ) = g(f
1
(

x ), . . . , f
p
(

x ));
4
Many other natural representations of rational numbers can be chosen and provide the same class of computable
functions: see [36].
O. Bournez, E. Hainry / Recursive Analysis Characterized as a Class of Real Recursive Functions 7
4. LI: linear integration. From g and h, LI(g, h) is the maximal solution of the linear dier-
ential equation
f
y
(

x , y) = h(

x , y)f(

x , y) with f(

x , 0) = g(

x ).
In this schema, if g goes to R
n
, f = LI(g, h) goes to R
n+1
and h(

x , y) is a (n+1) (n+1)
matrix with elements in L.
Class L includes common functions like +,sin,cos,,,exp, or constants r for all r Q (see
[12, 11]), but contains only total functions [11]:
Proposition 4.1. ([11])
All functions from L are continuous, dened everywhere, and of class (
2
.
Actually, observing the proofs from [12, 11], schema LI can be strengthened as follows:
Proposition 4.2. Class L is also the class of functions f : R
k
R
l
, for some k, l N, dened
by L = [0, 1, 1, , U,
3
; COMP, CLI] where CLI is the following schema:
1. CLI: controlled linear integration. From g and h, and c, with h dierentiable and norm
5
of rst partial derivatives of h bounded by c, CLI(g, h, c) is the maximal solution of the
linear dierential equation
f
y
(

x , y) = h(

x , y)f(

x , y) with f(

x , 0) = g(

x ).
In this schema, if g goes to R
n
, f = CLI(g, h, c) goes to R
n+1
and h(

x , y) is a (n+1)(n+1)
matrix with elements in L.
Class L can be related to the class c of elementar functions over the integers. A real extension

f of a function f : N
k
N
l
over the integers is a function

f from R
k
to R
l
whose restriction to
N
k
is f. Observe that a function

f : R
k
R
l
over the reals is an extension of a function over
the integers i its preserves integers:

f(N
k
) N
l
.
Denition 4.2. (Discrete Part)
Given a class ( of real functions, we denote by DP(() the class of functions over the integers
that have a real extension in (.
Proposition 4.3. ([12, 11])
c = DP(L). I.e.:
If a function from L preserves integers, then its restriction to integers is elementar.
Any elementar function over the integers, has a real extension that belongs to L.
Actually, class L can also be partially related to the class c(R) of functions over the real
numbers elementary in the sense of recursive analysis: any function from L is in c(R) [12, 11].
We proved in [4] that the inclusion is actually strict, but that adding a limit schema to class L,
allows us to capture the whole class c(R) for functions dened over a compact domain.
5
Say sup norm. All norms would provide equivalent results.
8 O. Bournez, E. Hainry / Recursive Analysis Characterized as a Class of Real Recursive Functions
5. Real recursive and recursive functions
We are now going to extend the class L with a minimalization schema in order to get a class
whose discrete part corresponds to recursive functions over the integers.
To do so, we need to introduce a zero-nding operator that permits to simulate the classical
discrete minimalization schema over the integers. However, this operator needs to be stricter
than a simple return the smallest root since this idea, investigated in [21], has shown to be the
source of numerous problems, including ill-dened problems and super-Turing Zeno phenomena.
These problems are discussed, and pointed in [12, 11, 23, 22, 21]. Papers [23, 25] do provide
well-dened alternatives, replacing minimalizations by limit-takings. We propose here to keep
to a minimalization schema, not as general as the one from [21]. Compared to the approach
from Costa and Mycka, our schemata are also more restricted than theirs.
Our idea is to use the alternative UMU schema which is equivalent to schema MU for classical
computability, but has real counterparts which turn out to preserve real computability. It means
that in a discrete context, the search of a unique zero is sucient to capture the whole class of
discrete recursive functions, and moreover in this continuous context, computing a unique zero
does not demonstrate the over-power of the standard minimalization operator.
Indeed, motivated by Proposition 2.2, by Lemma 2.2, and by results from recursive analysis
about the computability of zeros (see e.g. [36] where theorems 6.3.5 and 6.3.8 state that the
search of a unique zero is computable), we dene our unique-zero-nding operator UMU as
follows:
Denition 5.1. Given a dierentiable function f from (T 1) R
k+1
to R where T 1 is
a product of closed intervals, if for all

x T, y f(

x , y) is a non-decreasing function with

a unique root y
0
on 1 such that y
0
is in the interior of 1 and
f
y
(

x , y
0
) > 0, then UMU(f) is
dened on T as follows:
UMU(f) :

T R

x y
0
such that f(

x , y
0
) = 0.
We also slightly modify CLI schema, by allowing not-necessarily maximal solutions of linear
dierential equations to be considered. By abuse of notation, CLI will denote this schema in
what follows.
Denition 5.2. (CLI schema)
From g and h, and c, with h dierentiable and the norm
6
of rst partial derivatives of h bounded
by c, CLI(g, h, c) is any solution dened on a product of closed intervals of the the linear dier-
ential equation
f
y
(

x , y) = h(

x , y)f(

x , y) with f(

x , 0) = g(

x ).
In this schema, if g goes to R
n
, f = CLI(g, h, c) goes to R
n+1
and h(

x , y) is a (n+1)(n+1)
matrix with elements in L.
Denition 5.3. (Class L+!)
Let L+! be the set of functions dened by
L+! = [0, 1, U,
3
; COMP, CLI, UMU].
6
Say sup norm. All norms would provide equivalent results.
O. Bournez, E. Hainry / Recursive Analysis Characterized as a Class of Real Recursive Functions 9
Remark 5.1. The previous schema CLI yields a function in the class from g,h and c in the
class. The reason why the rst partial derivative of h is required to be bounded by c in the
schema is to ensure computability of f from computability of g, h and c. Notice that the bound
is on the derivative of h, and not on h, and hence functions like exponential can still be dened.
Notice, that when UMU operator is not present, it follows from Campagnolo, Moore and Costa
[10, 11, 12], that there is always some such function c in the class, and hence that there is no
need to require this bound.
Lemma 5.1. L L+!.
Proof:
(sketch) We only need to prove that constants 1 and are in L+!. Indeed, 1 is the unique
root of x x + 1, and = 4 arctan(1), where arctan(x) is the solution of linear dierential
equation arctan(0) = 0 and arctan

(x) =
1
1+x
2
, and x
1
1+x
2
can be obtained by applying UMU
on x, y (1 + x
2
)y 1. .
Lemma 5.2. All functions fromL+! are of class (
2
and dened on a product of closed intervals.
Proof:
By structural induction. Basic functions U,
3
are dened on R
k
and of class (
2
. Now, the
properties on the domain are preserved by the denition of composition, linear integration, and
schema UMU. The (
2
property is also preserved by Lemma 2.2 for schema UMU, and classical
results about dierential equations (see e.g. [1]) for schema CLI. .
It follows in particular, that there is no way to obtain functions such as x 1/x dened
on (0, +). It can be shown that any restriction to a closed interval of this function is in class
L+!.
Now, observe that operator UMU preserves real computability.
Lemma 5.3. ([36])
Given f : T 1 R real computable, if UMU(f) is dened, then UMU(f) is also real
computable.
Proof:
We reprint here mostly a restatement of a (slight generalization of) Corollary 6.3.9 from [36],
that we think helpful in order to understand next remark.
Write 1 = [a, b] with a and b possibly innite, where 1 is as in the denition of schema UMU.
Given

x T, let y
0
1 be the unique y
0
with f(

x , y
0
) = 0. Since f(

x , .) is continuous,
non-decreasing, and with a unique root, we have f(

x , y) < 0 for y < y

0
, and f(

x , y) > 0 for
y > y
0
.
There exists m N, such that f(

x , max(a, m)) < 0 and f(

x , min(m, b)) > 0: one just

need to take any integer m with m < y
0
< m. Actually, such an m can be computed as follows.
m = 1
Repeat
Compute f
1
= f(

x , min(b, m)) and f

2
= f(

x , max(a, m))) at pr e c i s i o n 2
m
10 O. Bournez, E. Hainry / Recursive Analysis Characterized as a Class of Real Recursive Functions
m = m + 1
Until (f
1
> 2
m
and f
2
< 2
m
)
Return m
Indeed, given any integer m
0
N with m
0
< y
0
< m
0
, (for example [y
0
[| +1), we have for
all m m
0
, f(

x , min(m, b)) f(

x , m
0
) > 0 and f(

x , max(a, m)) f(

x , m
0
) < 0. Now,
for m big enough (i.e. m m
0
, 2
m
[f(

x , max(a, m
0
))[, and 2
m
[f(

x , min(b, m
0
))[) we
have f
1
> 2
m
and f
2
< 2
m
and the process halts with an m such thatf(

x , max(a, m)) < 0

and f(

x , min(m, b)) > 0.

Computing y
0
then reduces to compute the unique root of function f(

x , .) over a compact
[m, m] 1. The fact that this is indeed computable can be seen as a consequence of the results
in [36]. See [5] for a direct proof. .
Remark 5.2. The proof is non constructive in the following sense: it follows from constructions
from [26] that there is no way to determine eectively from the code of f whether UMU(f) is
dened. Now, when it is, UMU(f) is real computable from previous arguments.
Lemma 5.4. Given h, g and c real computable, then f = CLI(g, h, c) is also real computable.
Proof:
Observing carefully [12, 11], if given

x R
k
and some y Q one can bound eectively the norms
of h(

x , y), f(

x , y),

2
f
y
2
(

x , y) for [y[ y, then f will be real computable: use the constructions

and bounds based on Eulers method to prove preservation of elementarity by linear integration
in [12, 11], but replacing elementar bounds by computable bounds.
Now, from [36], it is known that one can bound eectively the norm of any real computable
function on a closed domain, and so we only need to care about f(

x , y) and

2
f
y
2
(

x , y). But the

norm of f(

x , y) can be bounded eectively by Lemma 2.1 from bounds on the norms of g(

x )
and h(

x , y) on the corresponding domain, which are computable by previous argument. Now,

|

2
f
y
2
(

x , y)| = |(h
2
(

x , y) +
h
y
(

x , y))f(

x , y)|,
hence is bounded by (|h
2
(

x , y)| + |c(

x , y))|) |f(

x , y)|. First factor can still be bounded

eectively since h
2
(

x , y) and c(

x , y) are particular real computable functions, and we just saw

that second factor can be bounded eectively. .
From previous two Lemmas, the fact that basic functions are real computable and observ-
ing that composition is known to preserve real computability for functions dened over closed
intervals (see [36]), we obtain:
Theorem 5.1. Every function belonging to L+! is real computable.
Remark 5.3. As observed above, the proof is non-constructive. There is no way to obtain
eectively a functional that would compute a function f from the code of f. However, when f
belongs to L+!, there is one functional.
O. Bournez, E. Hainry / Recursive Analysis Characterized as a Class of Real Recursive Functions 11
0
0.5
1
1.5
2
2.5
3
3.5
4
4.5
5
-1 0 1 2 3 4
omega(x)
Figure 1. Graphical representation of
We now prove the converse direction. The following lemma is a weaker form of Lemma 7.3.
It provides a canonical extension of all functions in L. This terminology comes from [24] where
a similar result is proved. The idea is to exhibit a function

f that matches f on each integer
points, but is kept controlled with respect to x and y on each square.
Lemma 5.5. Given f : R
2
R in L, there exists

f : R
2
R in L such that (m, n) Z
2
,
(x, y) R
2
,

f(m, n) = f(m, n)

f(m, y) [f(m, y|), f(m, y + 1|)] (or [f(m, y + 1|), f(m, y|)]).

f(x, n) [f(x|, n), f(x + 1|, n)] (or [f(x + 1|, n), f(x|, n)]).
Proof:
Let =
3
2
. Let : x
3
(sin(2x)). i,

i+1
i
= 1 and is equal to 0 on [i +
1
2
, i + 1] for
i Z. Let its primitive equal to 0 at 0, and int : x (x
1
2
). Function int is a function
similar to the integer part: i Z, x [i, i +
1
2
], int(x) = i = x|. Figures 1 and 2 show
graphical representations of and int.
Let (i, y) = f(i, y + 1) f(i, y). Then for all i Z, y R, we have
(y)(i, int(y)) =

0 whenever y y| 1/2
(y)(i, y|) otherwise.
Let G be the solution of the linear dierential equation

G(x, 0) = f(x, 0)
G
y
(x, y) = (y)(x, int(y)).
12 O. Bournez, E. Hainry / Recursive Analysis Characterized as a Class of Real Recursive Functions
-1
0
1
2
3
4
-1 0 1 2 3 4
int(x)
Figure 2. Graphical representation of int
An easy induction on j then shows that G(i, j) = f(i, j) for all integer j. Furthermore, by
construction, i Z, G(i, y) belongs to the interval delimited by G(i, y|) = f(i, y|) and
G(i, y + 1|) = f(i, y + 1|).
Now, let

f be the solution of the linear dierential equation

f(0, j) = G(0, j)

f
x
(x, y) = (x)(G(int(x + 1), y) G(int(x), y)).
We have (i, j) Z
2
,

f(i, j) = f(i, j). And i Z,

f(i, y) belongs to the interval delimited by

f(i, y|) = f(i, y|) and

f(i, y + 1|) = f(i, y + 1|). And also, j Z,

f(x, j) belongs to the
interval delimited by

f(x|, j) = f(x|, j) and

f(x + 1|, j) = f(x + 1|, j).
.
Theorem 5.2. Every total recursive function over the integers has a real extension in L+!.
This also holds more generally for recursive functions dened on [a, b] N for some a, b
possibly innite.
Proof:
Let be a function from ec. Let T = [a, b] be its domain, taking a and b innite if is total.
We have = !() as in the proof of Proposition 2.2. Let
(m, n) = 2 (1 (m, n)) + (1 (m, n))
where is the same as in the proof of Proposition 2.2. m N T, for n = n
0
=!()(m, n),
we have (m, n
0
) = 1, and before this n
0
, (m, n) is equal to 0 and after this n
0
, (m, n) is equal
O. Bournez, E. Hainry / Recursive Analysis Characterized as a Class of Real Recursive Functions 13
to 2. Let i be a real extension of in L given by Proposition 4.3. Let

i be the function from L
obtained by Lemma 5.5 on f(m, x) : m, x i(m, x) 1.
m TN, there exists exactly one y R (given by y
0
=!()(m, n)) such that

i(m, y) = 0.
But, we can not directly apply schema UMU, since we have no assurance
7
that it also holds for
non integer values m. However, from the constructions in the proof of Lemma 5.5, given m N,
we have

i(m, y) equal to 1 for y y
0
1, and equal to (y) for y [y
0
1, y
0
+ 1], where
is dened in that proof.
Consider /(x) =
3
(x + 1). We have /(x) = 0 if x 1 and /(x) 1 if x 0. Let us
dene g as the solution on [a, b] R of the dierential equation

g(x, 0) = 1
g
y
(x, y) = /(

i(x, y)).
Let us choose (maple says =
1024
2609
) such that

0
1
/((x))dx = 1. We have m N,
g(m, y) = 0 y =!()(m, n).
Then dene g as the solution on [a, b] R of the linear dierential equation g(x, 0) = 1,
g
y
(x, y) = /( g(x, y)). If we choose adequately
8
(maple says =
a
4
b
4
c
2
+d
for some integers
a, b, c, d) , we will still have m N, g(m, y) = 0 y =!()(m, n).
The point is that, since / is always non-negative, we know that x R, y g(x, y) is
non-decreasing, and, because of Lemma 5.5, and from the denition of function /(x), it must
go to innity when y goes to innity. Actually, it must be equal to 1 up to a certain value y

,
then be strictly increasing, and since it goes to innity, it must have a root y
0
strictly greater
than y

. Now the derivative in this root y

0
cannot be 0 since /(x) is zero only when x 1.
This g is such that x, !y
0
such that g(x, y
0
) = 0 and
g
y
(x, y
0
) ,= 0 and for all x, y g(x, y)
is non-decreasing. We can thus apply UMU to this g. Now if we extend in a real function h
belonging to L using Proposition 4.3, we have hUMU(g) extending = () and belonging
to L+!. .
From previous two theorems, we obtain:
Theorem 5.3. For total functions ec = DP(L+!). I.e:
If a function from L+! extends some total function over the integers, this latter function
is total recursive.
Any total recursive function over the integers, has a real extension that belongs to L+!.
Proof:
The second item is Theorem 5.2. The rst item is immediate from Theorem 5.1: if a function f
belonging to L+! preserves integers, then a recursive function that equals f on N
k
can easily
be obtained from the functional computing f. .
Corollary 5.1. L is strictly included in L+!.
7
Actually, another problem is that the derivative relative to the second variable in the root point is 0.
8
This is in L since it can be obtained as a
4
UMU(x (b
4
c
2
+ d)x 1).
14 O. Bournez, E. Hainry / Recursive Analysis Characterized as a Class of Real Recursive Functions
6. Alternative schemata
6.1. An alternative: searching a unique root on some domain
Slight modications of the proofs in this paper can easily prove that schema UMU could be
actually be replaced by the following schema.
Proposition 6.1. We have also
ec = DP(L+!
[m,M]
).
where L+!
[m,M]
= [0, 1, U,
3
; COMP, CLI, !
[m,M]
)] where !
[m,M]
is the following schema:
1. Given a function f from R
k+1
to R, and two functions m, M : R
k
R, if for all

x ,
y f(

x , y) has a unique root y on [m(

x ), M(

x )], on which
f
y
(

x , y) ,= 0, then !
[m,M]
f
is dened as the function that maps

x to that root for all

x .
This schema is obviously more dicult to apply than UMU since it requires to give bounds
on the researched zero. However, it is more straightforward considering known theorems of
zero-searching in recursive analysis.
6.2. An other alternative: searching the minimum of a convex function
Observing that the (always unique) minimum of non-monotone convex function is real com-
putable, and that the zero of a non-decreasing function is the minimum of its primitive, and
that this latter primitive is a non-monotone convex function, schema UMU can actually be also
replaced by the following schema.
Proposition 6.2. We have also
ec = DP(L + min convex).
where L+min convex = [0, 1, U,
3
; COMP, CLI, min convex)] and min convex is the following
schema:
1. Given a function f from R
k+1
to R, such that for all

x , y f(

x , y) is a convex non-
monotone function, whose second derivative exists and is non-null on its minimum, then
min convex(f) is dened as the function that maps

x to the minimum of y f(

x , y).
Note that the idea of considering minimum of convex functions comes partly from discus-
sions with several people, including Manuel Campagnolo, Felix Costa and Cris Moore. The
question whether this precise schema would be equivalent to previous one was raised by Manuel
Campagnolo (private discussion).
O. Bournez, E. Hainry / Recursive Analysis Characterized as a Class of Real Recursive Functions 15
7. Link with computable analysis
In [4, 6], we proved that by adding a well chosen limit operator to the class L, it was possible
to capture not only discrete elementar functions but also elementary functions in the sense of
computable analysis.
If we carefully add this limit operator to the here-dened L+! class, we now prove that one
similarly obtains a class that captures not only discrete recursive functions but real recursive
functions (in the sense of recursive analysis).
To do so, we consider the following schema, already considered in [6]: a polynomial over
x R is a function of the form : R R, : x

n
i=0
a
i
x
i
for some a
0
, . . . , a
n
R.
A polynomial over

x = (x
1
, . . . , x
k+1
) R
k+1
is a function of the form : R
k+1
R,
:

x

n
i=0
a
i
x
i
k+1
for some a
0
, . . . , a
n
polynomial over (x
1
, . . . , x
k
) R
k
.
Denition 7.1. (LIM
w
schema)
Let f : R T R
k+1
R
l
, K : T R and : T R a polynomial with the following
hypothesis: for all

x , for all t |

x |,
|
f
t
(t,

x )| K(

x ) exp(t(

x )).
Then, on every product of closed intervals I R
k
on which (

x ) > 0, lim
t+
f(t,

x )
exists by Lemma 2.3. We dene F by F(

x ) = lim
t
f(t,

x ). If F is of class (
2
, then we
dene LIM
w
(f, K, ) as this function F : I R.
Let us now dene a new class L

!
.
Denition 7.2.
L

!
= [0, 1, U,
3
; COMP, CLI, UMU, LIM
w
]
We have the following theorem:
Theorem 7.1. For functions of class (
2
dened on a compact domain,
L

!
= ec(R).
The proof of this theorem will be done for one direction by structural induction, the other
by applying a more general property linked to the LIM
w
schema in the rest of this section.
7.1. Proof of Theorem 7.1: Upper bounds
We will now prove the rst direction of the theorem, namely L

!
ec(R).
Proposition 7.1. Every function belonging to L

!
is real computable in the sense of computable
analysis.
16 O. Bournez, E. Hainry / Recursive Analysis Characterized as a Class of Real Recursive Functions
Proof:
We have already proved that 0, 1, U,
3
belong to ec(R) and that COMP, CLI and UMU
preserve ec(R). Hence, we only need to show that LIM
w
also preserves ec(R).
Let g = LIM
w
(f, K, ), with f computed by recursive functional . We give the proof for f
dened on R ( to R where ( is a closed interval of R. The general case is easy to obtain.
Let x R, with (x) > 0. Since (x) is a polynomial, 1/(x) can be bounded by some
integer N computable from x. Similarly, K(x) can be bounded by some integer K computable
from x. In a same way, the norm of x can be bounded by some integer X computable from x.
Let (x
n
) x. For all i, j N, if we write abusively i for the constant sequence k i, we
have [
Q
((((i, x
n
), j)) f(i, x)[ < exp(j).
By Lemma 2.3, if i is big enough (i > |x|), we have
[f(i, x) g(x)[
K exp((x)i)
(x)
KN exp((x)i).
Hence,
[
Q
(((i, x
n
), j)) g(x)[ < exp(j) + KN exp((x)i).
If we take j

j + 1, i

N(j + 1 + ln(KN)|), we have exp(j

)
1
2
exp(j), and
KN exp((x)i

)
1
2
exp(j). Hence g is computed by the functional
: ((x
n
), j) ((max(X, N(j + 1 +ln(KN)|), x
n
)), j + 1).
since for all j,
|
Q
(((x
n
), j)) g(x)|
exp(j)
2
+
exp(j)
2
exp(j).
.
7.2. Proof of Theorem 7.1: Lower bounds
We will now prove the converse direction of the theorem: We are going to prove that for functions
of class (
2
dened on a compact domain, ec(R) L

!
.
In fact, we are going to prove a more general proposition that more or less states that given
a discrete class C c that has some basic properties of closure and a class of real functions
( such that C DP((), then the class of recursive analysis dened with functions from C,
denoted by C(R) is included in ( +LIM
w
(dened as being the class ( with LIM
w
as additional
operator.
The researched inclusion ec(R) L

!
for functions of class (
2
over compact domains will
follow considering C = ec and ( = L+!.
Formally,
Denition 7.3. Let C be a class of functions over the integers, with c C. A function
f : R
k
R is said to belong to C(R) as in Denition 3.1: There exists a functional
9
C such
that for all

x , for all sequence X, X

x = ((X, j))
j
f(

x ).
A function f : R
k
R
l
, with l > 1, is said to belong to C(R) if all its projections are.
9
In the sense of Footnote 3.
O. Bournez, E. Hainry / Recursive Analysis Characterized as a Class of Real Recursive Functions 17
Denition 7.4. Given a class of real function ( = [T; O], we will denote ( +LIM
w
for the class
[T; O, LIM
w
].
Denition 7.5. A modulus of continuity of a function f : T R
k
R
l
dened over a compact
domain is a function M : N N such that for all i N, for all x, y,
|x y| < exp(M(i)) |f(x) f(y)| < exp(i).
More generally (the modulus of continuity of a function dened over a compact domain gives
clearly a uniform modulus of continuity).
Denition 7.6. A uniform modulus of continuity of a function f : T R
k
R
l
dened over
a closed domain is a function M : N N N such that for all integer K, i N, for all
x, y [K, K]
k
,
|x y| < exp(M(K, i)) |f(x) f(y)| < exp(i).
Proposition 7.2. Let C be a class of functions over the integers, closed by composition, that
contains c, and ( be a class of real functions that contains L, that is closed under composition,
and by taking primitives, such that C DP(().
Then for functions of class (
2
dened on a compact domain, whose derivatives have a modulus
of continuity in C, C(R) ( + LIM
w
.
The researched lower bound ec(R) L

!
for functions of class (
2
over compact domains
indeed follows: indeed, consider C = ec, ( = L+!, and the following two well-known results,
that can be seen as slight generalizations of Corollary 6.4.8 and Theorem 6.2.7 of [36], or of
Theorem 2 of Chapter 1 of [29].
Lemma 7.1. Let f : T R
k
R
l
be a function of class (
2
dened over compact domain T.
If f is in C(R), then its partial derivatives also are.
Proof:
We give the proof for a function f dened on interval [0, 1] to R. The general case is easy to
obtain.
Since f

is continuous on a compact set, f

is bounded by some constant M. By mean value

theorem, we have [f

(x) f

(y)[ M[x y[ for all x, y.

Given x [0, 1], and i N, an approximation z of f

(x) at precision exp(i) can be

computed as follows: compute n with M exp(n) exp(i)/2. Compute y
1
a rational at most
exp(i n2) far from f(x), and y
2
a rational at most exp(i n2) far from f(x+exp(n)).
Take z = (y
1
y
2
)/ exp(n).
18 O. Bournez, E. Hainry / Recursive Analysis Characterized as a Class of Real Recursive Functions
This is indeed a value at most exp(i) far from f

(x) since by mean value theorem there

exists [x, x + exp(n)] such that f

(
j
) =
f(x+exp(n))f(x)
exp(n)
. Now
[z f

(x)[
|y
1
f(x)|
exp(n)
+
|y
2
f(x+exp(n))|
exp(n)
+[
f(x+exp(n))f(x)
exp(n)
f

(x)[
exp(i n 2) exp(n) + exp(i n 2) exp(n)
+[f

(
j
) f

(x)[
2 exp(i 2) + M exp(n)
exp(i)/2 + exp(i)/2
exp(i).
.
The following lemma is easy.
Lemma 7.2. If f : R
k
R
l
is dened over a product of compact intervals, f is of class (
1
, and
f C(R), then f has a modulus of continuity in C.
Proof:
The norm of any derivative of f, as a continuous function over a compact is bounded by some
integer m. By mean value theorem, function M(i) = m + i is easily shown to be a modulus of
continuity of f. As it is a linear function, it belongs to c C. .
Remark 7.1. The proof is non-constructive: m can not be obtained from the code of f in the
general case, and hence the modulus of continuity can not be obtained from the code of f.
But, for all function satisfying our hypotheses, there is a modulus of continuity in C.
7.3. Proof of Proposition 7.2
To prove Proposition 7.2, we use arguments similar to [4] and some properties from [36].
Indeed, the following lemma is used to get functions for which we know the behavior ev-
erywhere given only their values over a discrete set of points. It is a rened version of lemma
5.5.
Lemma 7.3. Let : R R be some decreasing function of (, with (x) > 0 for all x and going
to 0 when x goes to +, and 1/(x) (. Write
i
for (i|), Z
i
for j
i
; j Z, and x|

i
for
maxy Z

i
; y < x.
Given f : R
2
R
l
in (, there exists F : R
2
R
l
in ( with the following properties:
For all i N, x Z
i
, F(i, x) = f(i, x)
For all i N, x R, |F(i, x) f(i, x|

i
)| |f(i, x|

i
+
i
) f(i, x|

i
)|
For all i R
0
, x R,
|
F
i
(i, x)| 5|f(i + 1|, x|

i
) f(i|, x|

i
)|
+25|f(i|, x|

i
+
i
) f(i|, x|

i
)|
+25|f(i + 1|, x|

i+1
+
i+1
) f(i + 1|, x|

i+1
)|.
O. Bournez, E. Hainry / Recursive Analysis Characterized as a Class of Real Recursive Functions 19
Proof:
We are going to reuse the functions and int dened in the proof of lemma 5.5 whose graphical
representations are shown in Figures 1 and 2.
Let (i, x) = f(i, x + (i)) f(i, x). For all i,x, we have
(x/(i))
(i)
(i, (i) int(x/(i))) = 0 whenever x x|
(i)
(i)/2
=
(x/(i))
(i)
(i, x|
(i)
) otherwise.
Let G be the solution of the linear dierential equation

G(i, 0) = f(0)
G
x
(i, x) =
(x/(i))
(i)
(i, (i)int(x/(i)))
An easy induction on j then shows that G(i, j(i)) = f(i, j(i)) for all j Z.
On [j(i), (j + 1)(i)),
G(i, x) f(i, x|
(i)
) =

x
j(i)
(t/(i))
(i)
(i, t|
(i)
)dt,
hence, for all i N,
|G(i, x) f(i, x|

i
)| |(i, x|

i
)| = |f(i, x|

i
+
i
) f(i, x|

i
)|.
Now, let

(i, x) = G(i + 1, x) G(i, x). For all i,x we have

(i)

(int(i), x) = 0 whenever i i| 1/2

= (i)

(i|, x) otherwise
Let F be the solution of linear dierential equation

F(0, x) = G(0, x)
F
i
= (i)

(int(i), x)
An easy induction on i shows that F(i, x) = G(i, x) for all integer i, and all x R. Hence
F(i, x) = f(i, x) for all i N, x Z
i
and
|F(i, x) f(i, x|

i
)| |f(i, x|

i
+
i
) f(i, x|

i
)|
for all i N, x R.
Now,
F
i
is either 0 or (i)

(i|, x) = (i)(G(i + 1|, x) G(i|, x)). In any case, it is

derivable in x, and hence

2
F
xi
is either 0 or (i)(
G
x
(i + 1|, x)
G
x
(i|, x)).
When x Z
i
, bounding by 5 ( 5),
|
F
i
| 5|f(i + 1|, x) f(i|, x)|.
When x R,
|

2
F
xi
| |
G
x
(i + 1|, x| +|
G
x
(i|, x)|.
20 O. Bournez, E. Hainry / Recursive Analysis Characterized as a Class of Real Recursive Functions
The term |
G
x
(i|, x)| can be either 0 or
5|
(x/
i
)

i
(i|, x|

i
)|
25

i
|(i|, x|

i
)|

25

i
|f(i|, x|

i
+
i
) f(i|, x|

i
)|.
A similar bound holds for the other term, replacing i by i + 1.
Using mean value theorem,
|
F
i
(i, x)| |
F
i
(i, x|

i
)| +|

2
F
xi
(i, x)|(x x|

i
)
|
F
i
(i, x|

i
)| + (i)|

2
F
xi
(i, x)|
,
which yields the expected bound. .
Lemma 7.4. If f : T R R is dened over a closed interval containing 0, with bounds
either rational or innite, belongs to C(R), of class (
1
, with a uniform modulus of continuity in
C, then the primitive

(f) that is equal to 0 at 0 is in ( + LIM
w
.
Proof:
Let M
N
: N
2
N be the uniform modulus of continuity in C of function f: given some integer,
K, M
N
(K, ) is a modulus of continuity of function f on [K, K].
For all i, j N, let x
j
= j exp(M
N
(i + 1, i)), so that for all x, y [x
j
, x
j+1
] [i 1, i +1],
we have
[f(x) f(y)[ exp(i).
For all j, let p
j
and q
j
two integers such that p
j
exp(q
j
) is at most exp(i) far from
f(x
j
). The functions p
N
: N
2
N, and q
N
: N
2
N that map (i, j) to corresponding p
j
and q
j
are in C.
Since C DP(() these functions as well as function M
N
can be extended to function
p : R
2
R, q : R
2
R, M : R
2
R L. Consider function g : R ( R dened on all
(i, x) R ( by g(i, x) = p(i, exp(M(i + 1, i))x)e
q(i,exp(M(i+1,i))x)
. By construction, for i, j
integer, we have
g(i, x
j
) = p
j
exp(q
j
).
Consider the function F given by Lemma 7.3 for function g and : i exp(M(i + 1, i)).
We have
F(i, x
j
) = g(i, x
j
)
and
|g(i, x
j
) f(x
j
)| exp(i)
for all i, j.
O. Bournez, E. Hainry / Recursive Analysis Characterized as a Class of Real Recursive Functions 21
For all x (, and all integer i |x| we have (observe that if x
j
denotes x|

, we have
x
j
, x
j+1
[i 1, i + 1])
|F(i, x) f(x)| |F(i, x) F(i, x|

)| +|F(i, x|

) g(i, x|

)|
+|g(i, x|

) f(x|

)| +|f(x|

) f(x)|
|F(i, x|

+ ) F(i, x|

)| + 0 + exp(i) + exp(i)
|g(i, x
j+1
) g(i, x
j
)| + 2 exp(i)
|g(i, x
j+1
) f(x
j+1
)| +|g(i, x
j
) f(x
j
)|
+|f(x
j+1
) f(x
j
)| + 2 exp(i)
5 exp(i).
Consider the function G : R
2
R dened for all i, x R by the linear dierential equation

G(i, 0) = 0
G
x
(i, x) = F(i, x)
Hence
G(i, x) =

x
0
F(i, u)du.
We get
|
G
x
(i, x) f(x)| = |F(i, x) f(x)| 5 exp(i)
and by mean value theorem on function G(i, x) f(x), we get
|G(i, x)

x
0
(f)(x)| (5 exp(i))[x[,
when i |x|.
Hence,

(f)(x) is the limit of G(i, x) when i goes to + with integer values. We just need
to check that schema LIM
w
can be applied to function G of L

to conclude: indeed, the limit of

G(i, x) when i goes to + will exist and coincide with this value, i.e.

(f)(x).
Since
G
i
=

x
0
F
i
(i, u)du implies
|
G
i
|

x
0
|
F
i
|du [x[ |
F
i
| (x
2
+ 1) |
F
i
|,
we only need to prove that we can bound |
F
i
| by K(x) exp(i) for some function K L

,
and i |x|.
But from Lemma 5.5, we know that for all i, x,
|
F
i
(i, x)| 5|g(i + 1|, x|

i
) g(i|, x|

i
)|
+25|g(i|, x|

i
+
i
) g(i|, x|

i
)|
+25|g(i + 1|, x|

i+1
+
i+1
) g(i + 1|, x|

i+1
)|.
First term can be bounded by 5 exp(i) + 5 exp(i) = 10 exp(i).
22 O. Bournez, E. Hainry / Recursive Analysis Characterized as a Class of Real Recursive Functions
Second term can be bounded by
25(|g(i|, x|

i
+
i
) f(x|

i
+
i
)| +|f(x|

i
+
i
) f(x|

i
)| +|g(i|, x|

i
) f(x|

i
)|).
Hence by 25 exp(i) + 25 exp(i) + 25 exp(i) = 75 exp(i), for i |x|.
Similarly for third term, replacing i by i + 1.
Hence, when i |x|,
|
F
i
(i, x)| 160 exp(i),
and
|
G
i
(i, x)| 160 (x
2
+ 1) exp(i),
and so schema LIM
w
can be applied on function G of L

to get function

(f). This ends the
proof.
.
Actually, the previous lemma can easily be extended a little bit to get any primitive (clearly
this implies Proposition 7.2 for functions from a closed subset of R to R, considering a function
as the primitive of one of its derivative. The case R
k
to R
l
can be obtained by adapting our
arguments to functions of several variables.).
Lemma 7.5. Let h be a function of C(R) dened at 0.
If f : T R R is dened over a closed interval containing 0, with bounds either rational
or innite, belongs to C(R), of class (
1
, with a uniform modulus of continuity in C, then the
primitive of f equal to h(0) at 0 is in ( + LIM
w
.
Proof:
Replace in previous proof the initial condition G(i, 0) = 0 of the dierential equation dening
function G, by G(i, 0) = g(i) where g : R R is a function converging to h(0), obtained by
extending a suitably chosen function g : N N. .
This ends the proof that ec(R) = L

!
for functions dened over product of compact in-
tervals, which is an analog characterization of recursive functions in the sense of computable
analysis. In fact, we obtained a more general result concerning the power of our LIM
w
opera-
tor that can indeed be seen as a description of the missing link between discrete computability
theory and computable analysis.
7.4. Other results
From those results and those proofs, we can derive some other results of interest.
We can apply Proposition 7.2 to other classes than L+!. For example, is presented in [21]
a class

T = [0, 1, 1, U; COMP,

I]
that contains extensions of all recursive primitive functions. In other words,
T DP(

T).
O. Bournez, E. Hainry / Recursive Analysis Characterized as a Class of Real Recursive Functions 23
Where

I is dened as an integration operator: given functions f
1
, ..., f
m
of arity n and g
1
, ..., g
m+1
of arity n + 1 + m, if there is a unique set of functions h
1
, ..., h
m
such that
h
i
(x, 0) = f
i
(x)
h
i
y
(x, y) = g
i
(x, y,

h (x, y)) for all y I S

on an interval I containing 0 where S I is a countable set of isolated points and h and
h
y
are
both continuous, then h = h
1
is dened.
Since

T+
3
contains L, is stable under composition and integration, we have by Proposition
7.2 the following result:
Proposition 7.3. For functions of class (
2
dened on a compact domain,
T(R)

T +
3
+ LIM
w
Remark 7.2. Observe that
3
was not considered as a basic function in [21], since this integra-
tion operator allows to get non-analytic functions that can play the role of function
3
. For the
same reasons, we believe that actually T(R)

T + LIM
w
.
We have also some results for closed intervals non necessarily compact:
Proposition 7.4. Let f : T R
k
R
l
be some function over the reals of class (
2
, with T
product of closed intervals.
If f and the derivatives of f are in ec(R) then f L

!
.
Proof:
For k = l = 1, this follows from Lemma 7.5, considering function f as the primitive of one of
its derivative, and observing that it is known that any computable function over the reals has a
computable uniform modulus of continuity [36]. The case R
k
to R
l
can be obtained by adapting
arguments to functions of several variables. .
Remark 7.3. The previous arguments holds for any classes C, ( for which it is known that
any function over the reals in C(R) has a uniform modulus of continuity in C. In particular,
for elementar functions, all levels of the Grzegorczyk hierarchy [6] or for primitive recursive
functions.
In particular:
Proposition 7.5. Let f : T R
k
R
l
be some function over the reals of class (
2
, with T
product of closed intervals.
If f and the derivatives of f are in T then f

T +
3
+ LIM
w
.
8. Thanks
The authors would like to thank deeply Felix Costa for all his helpful comments and suggestions
about this manuscript.
24 O. Bournez, E. Hainry / Recursive Analysis Characterized as a Class of Real Recursive Functions
References
[1] Arnold, V. I.: Ordinary dierential equations, Springer-Verlag, Berlin, 1992.
[2] Asarin, E., Maler, O.: Achilles and the Tortoise Climbing Up the Arithmetical Hierarchy, Journal
of Computer and System Sciences, 57(3), December 1998, 389398.
[3] Bournez, O.: Complexite Algorithmique des Syst`emes Dynamiques Continus et Hybrides, Ph.D.
Thesis, Ecole Normale Superieure de Lyon, Janvier 1999.
[4] Bournez, O., Hainry, E.: An Analog Characterization of Elementarily Computable Functions Over
the Real Numbers, 31th International Colloquium on Automata Languages and Programming
(ICALP04) (J. Daz, J. Karhumaki, A. Lepisto, D. T. Sannella, Eds.), 3142, Springer, Turku,
Finland, 2004.
[5] Bournez, O., Hainry, E.: Real Recursive Functions and Real Extentions of Recursive Functions, Ma-
chines, Computations and Universality (MCU2004) (M. Margenstern, Ed.), 3354, Springer-Verlag,
Saint-Petersburg, Russia, 2004.
[6] Bournez, O., Hainry, E.: Elementarily Computable Functions Over the Real Numbers and R-Sub-
Recursive Functions, Theoretical Computer Science, 348(2-3), December 2005, 130147.
[7] Bowles, M. D.: U.S. technological enthusiasm and British technological skepticism in the age of the
analog brain, IEEE Annals of the History of Computing, 18(4), OctoberDecember 1996, 515,
ISSN 1058-6180.
[8] Brattka, V.: Computability over Topological Structures, in: Computability and Models (S. B.
Cooper, S. S. Goncharov, Eds.), Kluwer Academic Publishers, New York, 2003, 93136.
[9] Bush, V.: The dierential analyser, Journal of the Franklin Institute, 1931, 447488.
[10] Campagnolo, M., Moore, C., Costa, J. F.: An Analog Characterization of the Subrecursive Functions,
Proc. 4th Conference on Real Numbers and Computers (P. Kornerup, Ed.), Odense University Press,
2000.
[11] Campagnolo, M., Moore, C., Costa, J. F.: An analog characterization of the Grzegorczyk hierarchy,
Journal of Complexity, 18(4), 2002, 9771000.
[12] Campagnolo, M. L.: Computational complexity of real valued recursive functions and analog circuits,
Ph.D. Thesis, IST, Universidade Tecnica de Lisboa, 2001.
[13] Clote, P.: Computational Models and Function Algebras, in: Handbook of Computability Theory
(E. R. Grior, Ed.), North-Holland, Amsterdam, 1998, ISBN 0-444-89882-4, 589681.
[14] Etesi, G., Nemeti, I.: Non-Turing computations via Malament-Hogarth space-times, International
Journal Theoretical Physics, 41, 2002, 341370.
[15] Graca, D. S., Costa, J. F.: Analog computers and recursive functions over the reals, Journal of
Complexity, 19(5), 2003, 644664.
[16] Grzegorczyk, A.: Computable functionals, Fundamenta Mathematicae, 42, 1955, 168202.
[17] Hogarth, M. L.: Does General Relativity Allow an Observer to View an Eternity in a Finite Time?,
Foundations of Physics Letters, 5, 1992, 173181.
[18] Kalmar, L.: Egyszer u pelda eldonthetetlen aritmetikai problemara, Mate es zikai lapok, 50, 1943,
123.
O. Bournez, E. Hainry / Recursive Analysis Characterized as a Class of Real Recursive Functions 25
[19] Lacombe, D.: Extension de la notion de fonction recursive aux fonctions dune ou plusieurs variables
reelles III, Comptes Rendus de lAcademie des Sciences Paris, 241, 1955, 151153.
[20] Mills, J.: Programmable VLSI Extended Analog Computer for Cyclotron Beam Control, Technical
Report 441, Indiana University Computer Science, September 1995.
[21] Moore, C.: Recursion theory on the reals and continuous-time computation, Theoretical Computer
Science, 162(1), 5 August 1996, 2344.
[22] Mycka, J.: Innite limits and R-recursive functions, Acta Cybernetica, 16, 2003, 8391.
[23] Mycka, J.: -Recursion and innite limits, Theoretical Computer Science, 302, 2003, 123133.
[24] Mycka, J., Costa, J. F.: The P ,= NP conjecture, Submitted.
[25] Mycka, J., Costa, J. F.: Real recursive functions and their hierarchy, Journal of Complexity, 20(6),
2004, 835857.
[26] Mycka, J., Costa, J. F.: Undecidability over continuous-time, 2005, Submitted to the Logic Journal
of the IGPL, Oxford University Press.
[27] Odifreddi, P.: Classical Recursion Theory, Volume I, vol. 125 of Studies in Logic and the Foundations
of Mathematics, North-Holland, April 1992.
[28] Odifreddi, P.: Classical Recursion Theory, Volume II, vol. 143 of Studies in Logic and the Founda-
tions of Mathematics, North-Holland, 1999.
[29] Pour-El, M. B., Richards, J. I.: Computability in Analysis and Physics, Perspectives in Mathematical
Logic, Springer, Berlin, 1989.
[30] Ramis, E., Deschamp, C., Odoux, J.: Cours de Mathematiques Speciales, Tome 3, Topologie et
elements danalyse, Masson, February 1995.
[31] Rose, H.: Subrecursion: Functions and Hierarchies, Clarendon Press, 1984.
[32] Rubel, L. A.: The Extended Analog Computer, Advances in Applied Mathematics, 14, 1993, 3950.
[33] Shannon, C. E.: Mathematical Theory of the Dierential Analyser, Journal of Mathematics and
Physics MIT, 20, 1941, 337354.
[34] Thomson (Lord Kelvin), W.: On an instrument for calculating the integral of the product of two
given functions, Proceedings of the Royal Society of London, 24, 1876.
[35] Turing, A.: On computable numbers, with an application to the Entscheidungsproblem, Proceed-
ings of the London Mathematical Society, 42(2), 1936, 230265.
[36] Weihrauch, K.: Computable Analysis, Springer, 2000.