151 (1996)
The Zahorski theorem is valid in Gevrey classes
by
Jean S c h m e t s (Li`ege) and Manuel V a l d i v i a (Valencia)
Abstract. Let {Ω, F, G} be a partition of R
nsuch that Ω is open, F is F
σand of the first category, and G is G
δ. We prove that, for every γ ∈ ]1, ∞[, there is an element of the Gevrey class Γ
γwhich is analytic on Ω, has F as its set of defect points and has G as its set of divergence points.
1. Introduction. Let f be a real C
∞-function on R
n. The set where f is analytic is of course an open subset of R
n; denote it by Ω
f. It is clear that x belongs to Ω
fif and only if the radius of convergence %
f(x) of the Taylor series of f at x is strictly positive and this series represents f on some neighbourhood of x.
As the set T
fof x ∈ R
nsuch that %
f(x) > 0 is easily seen to be an F
σ-set, the set F
f= T
f\ Ω
fis also F
σand its elements x are characterized by the fact that %
f(x) > 0 and that the Taylor series of f at x represents f on no neighbourhood of x. By use of a lemma of R. P. Boas ([1], p. 234), one easily sees that F
fis also a first category set (cf. [3] or [4]).
Finally, one may consider the set G
f= R
n\ T
f, a G
δ-set given by
%
f(x) = 0.
It is clear that {Ω
f, G
f, F
f} is a partition of R
n.
The Zahorski theorem (cf. [6]) asserts conversely that for every partition {Ω, F, G} of [0, 1], where Ω is an open subset of [0, 1], F is a first category F
σ-subset of [0, 1] and G is a G
δ-subset of [0, 1], there is a real C
∞-function f on [0, 1] such that Ω = Ω
f, F = F
fand G = G
f. In [3], H. Salzmann and K. Zeller have provided a shorter proof of the Zahorski theorem and in [4], J. Siciak has extended this result to R
n.
1991 Mathematics Subject Classification: Primary 26E10; Secondary 26E05.
Key words and phrases: Gevrey classes, defect point, divergence point.
Research of the first author partially supported by the Belgian C.G.R.I.
Research of the second author partially supported by DGICYT PB91–0326 and the Convenio Hispano-Belga.
[149]
The purpose of this paper is to prove that, for every γ ∈ ]1, ∞[, the Zahorski theorem has a solution f belonging to the Gevrey class Γ
γ.
Let us recall that, for an open subset Ω of R
nand for γ ∈ ]1, ∞[, the Gevrey class Γ
γ(Ω) is the set of f ∈ C
∞(Ω) for which there are constants a, b > 0 such that
kD
αf k
Ω≤ ab
|α|(|α|!)
γ, ∀α ∈ N
n0. If Ω = R
n, we simply write Γ
γinstead of Γ
γ(R
n).
It is known that
(a) a function f ∈ C
∞(Ω) belongs to Γ
γ(Ω) if and only if there are constants c, d > 0 such that
kD
αf k
Ω≤ cd
|α||α|
γ|α|, ∀α ∈ N
n0.
(b) the Denjoy–Carleman–Mandelbrojt result (cf. [2]) states that, for every closed ball b of R
nand every γ ∈ ]1, ∞[, there is a nonzero function f ∈ Γ
γwith support contained in b.
In order to get an efficient way to state the results, for a real C
∞-function f on R
n, let us call the elements of Ω
f(resp. F
f; G
f) the analytic points (resp. the defect points; the divergence points) of f .
The purpose of this article is to prove the following result.
Theorem 1.1. For every partition {Ω, F, G} of R
n, where Ω (resp. F ; G) is an open set (resp. a first category F
σ-set; a G
δ-set) and every γ ∈ ]1, ∞[, there is an element of Γ
γhaving Ω (resp. F ; G) as its set of analytic points (resp. defect points; divergence points).
R e m a r k. It is a direct matter to check that the Zahorski theorem ex- tends to the case when R
nis replaced by an open or a closed subset of R
n. 2. An auxiliary result. We begin with the following easy result, where, as usual, D
r(Ω) denotes the space of C
r-functions on the open subset Ω of R
nwhich have a compact support contained in Ω.
Proposition 2.1. Let Ω be a nonvoid open subset of R
n. For every f ∈ C
∞(Ω), g ∈ D
∞(Ω) and λ > 0, it is well known that
h(x) = π
−n/2λ
n\
Ω
f (y)g(y)e
−λ2|x−y|2dy
belongs to C
∞(R
n). If moreover a
1, a
2, b
1, b
2> 0 and ζ > 1 are such that kD
αf k
Ω≤ a
1b
|α|1(|α|!)
ζand kD
αgk
Ω≤ a
2b
|α|2(|α|!)
ζ, ∀α ∈ N
n0, then
kD
αhk
Rn≤ a
1a
2(b
1+ b
2)
|α|(|α|!)
ζ, ∀α ∈ N
n0.
P r o o f. As g has a compact support contained in Ω, up to extension by 0 on R
n\ Ω, we may suppose that the product f g is a D
∞-function on R
nwith compact support contained in Ω. So in the definition of h(x), we may consider that we integrate on R
n. Therefore integrating by parts |α| times gives
D
αh(x) = π
−n/2λ
n\
Rn
D
α(f g)(y)e
−λ2|x−y|2dy, hence
|D
αh(x)| ≤ π
−n/2λ
n\
Rn
X
β≤α
α β
|D
βf (y)D
α−βg(y)|e
−λ2|x−y|2dy
≤ X
β≤α
α β
a
1a
2b
|β|1b
|α−β|2(|β|!)
ζ(|α − β|!)
ζ≤ a
1a
2(b
1+ b
2)
|α|(|α|!)
ζ.
3. Special compact covers of open subsets of R
n. In the following results, we are going to use systematically the following construction and notations. Let Ω be a nonvoid open subset of R
n. Then we set
Ω
m= (R
n\ Ω) ∪ {x ∈ R
n: |x| ≥ m √
n}, ∀m ∈ N,
and denote by µ the first positive integer m for which there is at least one cube Q of the type
Y
n j=1[2
−ma
j, 2
−m(a
j+ 1)] with a ∈ Z
ncontained in Ω and such that d(Q, Ω
m) > 2
−m√
n. Let Q
1,1, . . . , Q
1,p1be these cubes (of course, we have p
1∈ N) and set
H
1=
p1
[
h=1
Q
1,h.
Now we proceed by recursion. If the sets H
1, . . . , H
rare obtained, we let Q
r+1,1, . . . , Q
r+1,pr+1denote all the cubes of the type
Y
n j=1[2
−µ−ra
j, 2
−µ−r(a
j+ 1)] with a ∈ Z
ncontained in Ω, disjoint from the interior of H
1∪ . . . ∪ H
rand such that d(Q, Ω
µ+r) > 2
−µ−r√
n. Then we set
H
r+1=
p
[
r+1h=1
Q
r+1,h.
At this point let us remark that
d(H
r, H
r+2) ≥ 2
−µ−r, ∀r ∈ N.
Finally, we set
K
r= H
1∪ . . . ∪ H
r, ∀r ∈ N.
It is clear that {K
r: r ∈ N} is a compact cover of Ω such that K
r⊂ K
r+1ofor every r ∈ N.
4. The auxiliary functions v
r. With this construction in mind and the notations therein, we now prove the following result.
Proposition 4.1. Let Ω be a nonvoid open subset of R
nand ζ ∈ ]1, ∞[.
Then there are integers c, d ∈ N and functions v
r−2∈ C
∞(R
n) for r ∈ {3, 4, 5, . . .} such that
(a) supp(v
r−2) ⊂ K
r+1\ K
r−2o, (b) v
r−2(R
n) ⊂ [0, 1],
(c) v
r−2(H
r) = {1},
(d) kD
αv
r−2k
Rn≤ c(2
r−2d)
|α|(|α|!)
ζ, ∀α ∈ N
n0, for every integer r > 2.
P r o o f. Let ϕ be an element of C
∞(R) for which there are l, d > 0 such that
ϕ(t) > 0 if |t| < n
−1/22
−µ−4, ϕ(t) = 0 if |t| ≥ n
−1/22
−µ−4, kϕ
(s)k
R≤ ld
s(s!)
ζ, ∀s ∈ N
0, \
R
ϕ(t) dt = 1.
The existence of such a ϕ is provided by the Denjoy–Carleman–Mandelbrojt theorem (cf. [2]). Then we define
ψ(x) = ϕ(x
1) . . . ϕ(x
n), ∀x ∈ R
n.
Clearly ψ belongs to C
∞(R
n), has compact support equal to [−n
−1/22
−µ−4, n
−1/22
−µ−4]
nand satisfies
kD
αψk
Rn≤ l
nd
|α|(|α|!)
ζ, ∀α ∈ N
n0, \
Rn
ψ(x) dx = 1.
Now for every integer r ∈ N, we set
L
r= {x ∈ R
n: d(x, H
r) ≤ 2
−µ−r−1},
ψ
r(x) = 2
(r−2)nψ(2
r−2x), ∀x ∈ R
n,
and note that ψ
rbelongs to C
∞(R
n) and has a compact support of diameter 2
−µ−r−1. Then we set
v
r−2(x) = ψ
r∗ χ
Lr(x) = \
Rn
ψ
r(y)χ
Lr(x − y) dy, ∀x ∈ R
n.
It is well known or easy to check that the function v
r−2belongs to C
∞(R
n) and has the properties (a)–(c) as well as
kD
αv
r−2k
Rn≤ \
Rn
|D
αψ
r| dx ≤ `(supp(ψ
r))kD
αψ
rk
Rn≤ c(2
r−2d)
|α|(|α|!)
ζfor every α ∈ N
n0if we set c = l
n(n
−1/22
−µ−3)
n, a constant which does not depend on r > 2 nor on α ∈ N
n0.
5. Approximation in Gevrey classes and consequences. In the proof of Theorem 5.2, we shall make use of the following property which results immediately from the proof of Lemma 5 of [5].
Proposition 5.1 Let r ∈ N and g ∈ D
r(R
n). Then, for every ε > 0, there is λ
0> 0 such that, for every λ ≥ λ
0, the function
h(x) = π
−n/2λ
n\
Rn
g(y)e
−λ2|x−y|2dy
belongs to C
∞(R
n) (in fact, it is analytic on R
n) and satisfies kD
αh − D
αgk
Rn≤ ε if |α| ≤ r.
Theorem 5.2. Let Ω be a nonvoid open subset of R
nand let ζ, γ be real numbers such that 1 < ζ < γ. Then, for every f ∈ Γ
ζ(Ω), there is g ∈ Γ
γ(Ω) which is analytic on Ω and such that
kD
αf − D
αgk
Ω\Ks+1≤ 1
s if |α| ≤ s and s ≥ 2, with K
s+1defined as in the special compact cover of Ω.
P r o o f. Of course there are numbers a, b > 1 such that
kD
αf k
Ω≤ ab
|α|(|α|!)
ζand kD
αv
r−2k
Ω≤ a(2
r−2b)
|α|(|α|!)
ζfor every integer r ≥ 3 and α ∈ N
n0.
Now we introduce by recursion a sequence (g
s)
s∈Nin C
∞(R
n) such that kD
αg
sk
Rn≤ 2
s−1a
s+12
(s+1)|α|b
|α|(|α|!)
ζ, ∀α ∈ N
n0, ∀s ∈ N.
At this point, to get the functions g
s, we just need to consider a strictly increasing sequence (λ
s)
s∈Nof ]0, ∞[ but later on we shall make more strin- gent restrictions on these positive numbers.
We start with
g
1(x) = π
−n/2λ
n1\
Rn
v
1(y)f (y)e
−λ21|x−y|2dy, ∀x ∈ R
n,
where of course v
1f has been extended by 0 on R
n\ Ω. By Proposition 2.1, g
1belongs to C
∞(R
n) and satisfies
kD
αg
1k
Rn≤ a
23
|α|b
|α|(|α|!)
ζ≤ 2
0a
22
2|α|b
|α|(|α|!)
ζfor every α ∈ N
n0. Now if g
1, . . . , g
sare obtained, we first remark that we certainly have
D
αf −
X
s j=1g
jΩ
≤ kD
αf k
Ω+ X
s j=1kD
αg
jk
Ω≤ ab
|α|(|α|!)
ζ+ X
s j=12
j−1a
j+12
(j+1)|α|b
|α|(|α|!)
ζ≤ 2
sa
s+12
(s+1)|α|b
|α|(|α|!)
ζfor every α ∈ N
n0and then check by direct use of Proposition 2.1 that the function g
s+1defined by
g
s+1(x) = π
−n/2λ
ns+1\
Rn
v
s+1(y)
f (y) −
X
s j=1g
j(y)
e
−λ2s+1|x−y|2dy
suits our purpose.
Of course we have D
αX
s j=1g
jRn
≤
X
s j=12
j−1a
j+12
(j+1)|α|b
|α|(|α|!)
ζ≤ 2
sa
s+12
(s+1)|α|b
|α|(|α|!)
ζfor every s ∈ N and α ∈ N
n0.
With this majorant at our disposal, we are in a position to make a more precise (but not yet final) choice of the numbers λ
s(we are free to take them larger but strictly increasing). As we have
|α|→∞
lim
2
sa
s+12
(s+1)|α|b
|α|(|α|!)
ζ(|α|!)
γ= 0, ∀s ∈ N,
there is a strictly increasing sequence (A
s)
s∈Nin N such that, for every s ∈ N,
2
sa
s+12
(s+1)|α|b
|α|(|α|!)
ζ≤ (|α|!)
γif |α| ≥ A
s.
Then we can also fix a strictly increasing sequence (B
s)
s∈Nin N such that sup
|α|≤As
a(2
sb)
|α|(|α|!)
ζ≤ B
s, ∀s ∈ N.
Next we introduce the following elements of D
∞(R
n):
h
1(x) =
v
1(x)f (x), ∀x ∈ Ω,
0, ∀x ∈ R
n\ Ω,
and, for every integer s ≥ 2, h
s(x) =
v
s(x)(f (x) − P
s−1j=1
g
j(x)), ∀x ∈ Ω,
0, ∀x ∈ R
n\ Ω.
Finally, by recursive use of Proposition 5.1, we may require that the numbers λ
sare such that, for every s ∈ N,
kD
αh
s− D
αg
sk
Rn≤ (2
s+2+As+1B
s+1)
−1if |α| ≤ A
s+1. Now we consider the series
g(x) = X
∞ s=1g
s(x), ∀x ∈ Ω.
We first prove that g is defined and belongs to C
∞(Ω). Indeed, for every s ∈ N and every α ∈ N
n0such that |α| ≤ A
s+1, we get
kD
αh
s+1k
Hs+2≤ X
β≤α
α β
kD
βv
s+1k
Hs+2D
α−βf −
X
s j=1g
jHs+2
≤
(∗)
X
β≤α
α β
a(2
s+1b)
|β|(|β|!)
ζ(2
s+2+As+1B
s+1)
−1≤ X
β≤α
α β
B
s+1(2
s+2+As+1B
s+1)
−1≤ 2
−s−2−As+12
|α|≤ 2
−s−2(at (∗), we have used the fact that f − P
sj=1
g
j= h
s− g
son H
s+2). As v
s+1(x) = 0 for every x ∈ K
s+1, we get
kD
αh
s+1k
Ks+2≤ 2
−s−2if s ∈ N and |α| ≤ A
s+1, hence
kD
αg
s+1k
Ks+2≤ kD
αg
s+1− D
αh
s+1k
Ω+ kD
αh
s+1k
Ks+2≤ (2
s+3+As+2B
s+2)
−1+ 2
−s−2≤ 2
−s−1for every s ∈ N and α ∈ N
n0such that |α| ≤ A
s+1. Now it is clear that g ∈ C
∞(Ω).
We establish next that g satisfies the inequalities announced in the state- ment of the theorem—in fact, we are going to prove more. Consider an inte- ger s ≥ 2 and a multi-index α ∈ N
n0such that |α| ≤ A
s—this is certainly the case if |α| ≤ s. For every x
0∈ Ω \ K
s+1, there is a first integer p ≥ 2 such that x
0∈ K
s+p. According to the last established inequality, we of course have
|D
αg
s+r(x
0)| ≤ 2
−s−r, ∀r ∈ {p − 1, p, p + 1, . . .}.
As v
s+p−2(H
s+p) = {1}, we also get
D
αf (x
0) − D
αs+p−2
X
j=1
g
j(x
0)
= |D
αh
s+p−2(x
0) − D
αg
s+p−2(x
0)| ≤ 2
−s−p.
Thus
|D
αf (x
0) − D
αg(x
0)| ≤
D
αf (x
0) − D
αs+p−2
X
j=1
g
j(x
0) +
X
∞ r=p−1|D
αg
s+r(x
0)|
≤ 2
−s−p+ X
∞ r=p−12
−s−r≤ 2
−s−p+ 2
−s−p+2< 1 s , and hence
kD
αf − D
αgk
Ω\Ks+1≤ 1
s if s ≥ 2 and |α| ≤ A
s.
At this step, if we proceed as in the proof of Lemma 6 of [5], we see that it is possible to select successively the numbers λ
sin such a way that g is an analytic function on Ω. (This is our last refinement on the choice of λ
s.)
We still have to prove that g ∈ Γ
γ(Ω).
Set
sup
|α|≤A2
kD
αgk
K3/(|α|!)
γ= a
1and consider α ∈ N
n0and x
0∈ Ω.
On the one hand, if |α| ≤ A
2and (i) if x
0∈ K
3, we trivially have
|D
αg(x
0)| ≤ kD
αgk
K3≤ a
1(|α|!)
γ, (ii) if x
0∈ Ω \ K
3, we get
|D
αg(x
0)| ≤ |D
αf (x
0)| + |D
αf (x
0) − D
αg(x
0)|
≤ ab
|α|(|α|!)
ζ+ 1
2 ≤ 2ab
|α|(|α|!)
γ. Hence there are constants a
2, b
2> 0 such that
kD
αgk
Ω≤ a
2b
|α|2(|α|!)
γif |α| ≤ A
2.
On the other hand, if |α| > A
2, we first let s be the integer such that A
s< |α| ≤ A
s+1(of course s ≥ 2) and then consider the following two possibilities:
(i) if x
0∈ Ω \ K
s+2, then we have at once
|D
αg(x
0)| ≤ |D
αf (x
0)| + |D
αf (x
0) − D
αg(x
0)|
≤ ab
|α|(|α|!)
ζ+ 1
s + 1 ≤ 2ab
|α|(|α|!)
γ,
(ii) if x
0∈ K
s+2, we have |D
αg
s+r(x
0)| ≤ 2
−s−rfor every r ∈ N, hence X
∞r=1
|D
αg
s+r(x
0)| ≤ X
∞ r=12
−s−r= 2
−s. But we also have
D
βX
s j=1g
j(x
0)
≤ 2
sa
s+12
(s+1)|β|b
|β|(|β|!)
ζfor every β ∈ N
n0and s ∈ N, hence
D
αX
s j=1g
j(x
0)
≤ (|α|!)
γsince |α| ≥ A
s. Therefore
|D
αg(x
0)| ≤ D
αX
s j=1g
j(x
0) +
X
∞ r=1|D
αg
s+r(x
0)| ≤ 2
−s+ (|α|!)
γ≤ 2(|α|!)
γ. Consequently, there are constants a
3, b
3> 0 such that
kD
αgk
Ω≤ a
3b
|α|3(|α|!)
γif |α| > A
2.
Corollary 5.3. For every open and nonvoid subset Ω of R
nand γ ∈ ]1, ∞[, there is a function g ∈ Γ
γwhich is
(a) analytic on Ω,
(b) identically 0 on no connected component of Ω,
(c) flat on R
n\ Ω (i.e. identically 0 together with all its derivatives on R
n\ Ω).
P r o o f. If Ω is connected, we choose ζ ∈ ]1, γ[ and f ∈ Γ
ζ(Ω) with compact support contained in K
4o\ K
3and such that kf k
Ω> 1/2. Then Theorem 5.2 provides g ∈ Γ
γ(Ω) which is analytic on Ω and such that
kD
αf − D
αgk
Ω\K2+1≤ 1
2 if |α| ≤ 2 (which implies that g is not identically 0 on Ω) as well as
kD
αgk
Ω\Ks+1= kD
αg − D
αf k
Ω\Ks+1≤ 1
s if |α| ≤ s and s ≥ 3.
It is then well known that extending g by 0 on R
n\ Ω provides a solution.
If Ω has a finite number of connected components—say Ω
1, . . . , Ω
m— then, for every k ∈ {1, . . . , m}, there is g
k∈ Γ
γwhich is analytic on Ω
k, not identically 0 on Ω
kand flat on R
n\ Ω
k. It is then clear that g = P
mk=1
g
kis a solution.
As Ω always has countably many connected components, to conclude,
we just have to settle the case when {Ω
m: m ∈ N} is the set of connected
components of Ω. For this purpose, fix ζ ∈ ]1, γ[. By the first part of the proof, for every m ∈ N, there is g
m∈ Γ
ζwhich is analytic on Ω
m, not identically 0 on Ω
m, flat on R
n\ Ω
mand such that
kD
αg
mk
Ωm\Km,s+1≤ 1
s if |α| ≤ s and s ≥ 2,
where of course K
m,ris the rth element of the special cover of Ω
m. So, for every m ∈ N, there are constants a
m, b
m> 1 such that
kD
αg
mk
Rn≤ a
mb
|α|m(|α|!)
ζ, ∀α ∈ N
n0, hence there is an integer k
m≥ m such that
a
mb
|α|m(|α|!)
ζ≤ (|α|!)
γif |α| ≥ k
m. Then we set
c
m= ( sup
|α|≤km
a
mb
|α|m(|α|!)
ζ)
−1and g = X
∞ m=12
−mc
mg
m.
It is clear that g is a function defined on R
nwhich is analytic on Ω, identi- cally 0 on no connected component of Ω and identically 0 on R
n\Ω. Now we prove that g belongs to C
∞(R
n) and is flat on R
n\Ω. Let x ∈ R
n\Ω. For ev- ery integer k ≥ 3, there is r > 0 such that the ball b = {y ∈ R
n: |x − y| ≤ r}
is disjoint from the compact sets K
1,k+1, . . . , K
k,k+1. For every α ∈ N
n0such that |α| ≤ k, this leads to
sup
x∈b
X
∞ m=12
−mc
m|D
αg
m|
≤ sup
sup
m≤k
2
−mc
mk , sup
m>k
2
−mc
mkD
αg
mk
Rn≤ sup{1/k, 2
−k},
hence g belongs to C
∞(R
n) and is flat on R
n\ Ω. We still have to prove that g ∈ Γ
γ. This is immediate: for every α ∈ N
n0, we have
kD
αgk
Rn= sup
m∈N
2
−mc
mkD
αg
mk
Rn≤ sup
m∈N
sup{2
−m, 2
−m(|α|!)
γ} ≤ (|α|!)
γby consideration of the cases |α| ≤ k
mand |α| > k
m.
Corollary 5.4. For every γ ∈ ]1, ∞[ and nondegenerate compact in- tervals I, J of R such that J ⊂ I
o, there are f ∈ C
∞(R) and c > 0 such that
(a) f has no divergence point, (b) f (R) ⊂ [0, 1],
(c) f (R \ I) = {0}, f (I
o) ⊂ ]0, 1] and f (J) = {1},
(d) kf
(k)k
R≤ ck
γk, ∀k ∈ N.
P r o o f. Let I = [a
1, b
1] and J = [a
2, b
2]. We choose ζ ∈ ]1, γ[ and apply Corollary 5.3 with Ω = ]a
1, a
2[: there is a nonzero g ∈ Γ
ζwhich is analytic on ]a
1, a
2[ and such that supp(g) = [a
1, a
2]. Now we choose k > 0 such that g
1= kg
2satisfies T
R
g
1(x) dx = 1 and define the function f
1on R by f
1(x) =
x
\
−∞
g
1(t) dt, ∀x ∈ R.
It is clear that f
1belongs to Γ
ζ, is analytic on ]a
1, a
2[ and satisfies f
1(]−∞, a
1]) = {0}, f
1(]a
1, a
2[) ⊂ ]0, 1] and f
1([a
2, ∞[) = {1}.
So it is clear that f
1has no divergence point.
Similarly there is f
2∈ Γ
ζwhich is analytic on ]b
2, b
1[, has no divergence point and satisfies
f
2(]−∞, b
2]) = {0}, f
2(]b
2, b
1[) ⊂ ]0, 1] and f
2([b
1, ∞[) = {1}.
Finally, we set
f (x) = f
1(x)f
2(b
1+ b
2− x), ∀x ∈ R.
Of course f belongs to Γ
ζand satisfies (a)–(c). Let us establish that f also satisfies (d). As f ∈ Γ
ζ, there are a, b > 0 such that
kf
(k)k
R≤ ab
kk
ζk, ∀k ∈ N
0. Since
k→∞
lim
ab
kk
ζkk
γk= 0,
there is k
0∈ N such that ab
kk
ζk≤ k
γkfor every k ≥ k
0; therefore there is c > 0 such that
kf
(k)k
R≤ ck
γk, ∀k ∈ N.
Proposition 5.5. Let γ ∈ ]1, ∞[, let p ∈ N and let I, J be nondegenerate compact intervals of R such that J ⊂ I
o. Then there is m
0∈ N such that, for every integer m ≥ m
0, there is a function u ∈ C
∞(R) satisfying the following conditions:
(a) u has no divergence point, (b) supp(u) ⊂ I,
(c) ku
(k)k
R≤ 2
−m, ∀k ∈ {0, 1, . . . , m}, (d) ku
(k)k
R≤ 2
kk
γk, ∀k ∈ N,
(e) for every x ∈ J, one has either
|u
(pm)(x)| ≥ 5
−mm
γ(p−1)mor
|u
(pm+1)(x)| ≥ 5
−(m+1/p)pm + 1 p + 1
γ(p−1)(m+1/p).
P r o o f. Corollary 5.4 provides a function f ∈ C
∞(R) and a constant c > 0 such that f has no divergence point and satisfies
f (R) ⊂ [0, c], f (R \ I) = {0}, f (J) = {1}
as well as
kf
(k)k
R≤ ck
γk, ∀k ∈ N.
For any m ∈ N, we can introduce
a = m
γ, b = (2
2mcm
γm)
−1, u(x) = bf (x) sin(ax), ∀x ∈ R.
It is then clear that u is a C
∞-function on R satisfying the conditions (a) and (b) as well as
kuk
R≤ bc = (2
2mm
γm)
−1≤ 2
−m. Moreover, for every x ∈ R and k ∈ N, we have
|u
(k)(x)| ≤ b X
k h=0k h
|f
(h)(x)|a
k−h≤ bc X
k h=0k h
h
γha
k−h≤ (2
2mm
γm)
−1(k
γ+ a)
k, hence
|u
(k)(x)| ≤ (2
2mm
γm)
−1(m
γ+ m
γ)
m= 2
−mif 1 ≤ k ≤ m as well as
|u
(k)(x)| ≤ (2
2mm
γm)
−1(k
γ+ k
γ)
k≤ 2
kk
γkif k > m;
i.e. u also satisfies the conditions (c) and (d).
Now we investigate (e). Let x ∈ J. Of course we have u
(k)(x) = a
kb sin(kπ/2 + ax), ∀k ∈ N.
Now, for every m ∈ N, we certainly have
sup{|sin(pmπ/2 + ax)|, |sin((pm + 1)π/2 + ax)|} ≥ 2
−1/2. So on the one hand, if |sin(pmπ/2 + ax)| ≥ 2
−1/2, we get
|u
(pm)(x)| ≥ 2
−1/2a
pmb = 4
−m( √
2c)
−1m
γ(p−1)m, and on the other hand, if |sin((pm + 1)π/2 + ax)| ≥ 2
−1/2, then
|u
(pm+1)(x)| ≥ 2
−1/2a
pm+1b = 4
−m( √
2c)
−1m
γ((p−1)m+1)≥ 4
−(m+1/p)( √
2c)
−1m
γ(p−1)(m+1/p)≥ 4
−(m+1/p)( √ 2c)
−1pm + 1 p + 1
γ(p−1)(m+1/p).
To conclude, it is enough to take as m
0any positive integer m
0such that sup{4( √
2c)
1/m0, 4( √
2c)
1/(m0+1/p)} ≤ 5.
6. Characterizing the sets of divergence points. In this section, we establish the following result.
Theorem 6.1. For every γ ∈ ]1, ∞[ and every G
δ-subset G of R
n, there is an element of Γ
γhaving G as its set of divergence points.
P r o o f. We proceed in several steps.
S t e p 1: the numbers γ
jand p
j. We fix a strictly increasing sequence (γ
j)
j∈N0⊂ ]1, γ[ and, for every j ∈ N, denote by p
ja positive integer such that p
j(γ
j− γ
j−1) > γ
j.
S t e p 2: some auxiliary inequalities and the numbers q
r. For every r ∈ N, we certainly have
p
r− 1
p
rγ
r> γ
r−1> . . . > γ
0> 1.
Therefore, for every j ∈ {0, . . . , r − 1}, it is a straightforward matter to check the following limits:
(1) lim
m→∞
5
m2
2prm(p
rm)
γjprm1 m
γr(pr−1)m= lim
m→∞
4.5
1/prp
γrjm
γr(pr−1)/pr−γj prm= 0,
(2) lim
m→∞
5
m+1/pr2
2(prm+1)(p
rm+1)
γj(prm+1)p
r+ 1 p
rm + 1
γr(pr−1)(m+1/pr)= lim
m→∞
4.5
1/pr(p
r+ 1)
γr(pr−1)/pr(p
rm + 1)
γr(pr−1)/pr−γj prm+1= 0, (3) lim
m→∞
5
m(nr)
prm(p
rm + 1)
prm1 m
γr(pr−1)m= lim
m→∞
5
1/prnrp
γrr(pr−1)/pr(p
rm)
γr(pr−1)/pr−1 prmp
rm + 1 p
rm
prm= 0,
(4) lim
m→∞
5
m+1/pr(nr)
prm+1(p
rm + 2)
prm+1p
r+ 1 p
rm + 1
γr(pr−1)(m+1/pr)= lim
m→∞
5
1/prnr(p
r+ 1)
γr(pr−1)/pr(p
rm + 1)
γr(pr−1)/pr−1 prm+1p
rm + 2 p
rm + 1
prm+1= 0.
With these limits at our disposal, we find that, for every r ∈ N, there is
q
r∈ N such that, for every integer m ≥ q
rand j ∈ {1, . . . , r − 1}, we have
the following auxiliary inequalities:
(I) 4 · 2
2prm(p
rm)
γjprm< 5
−mm
γr(pr−1)m, (II) 4 · 2
2(prm+1)(p
rm + 1)
γj(prm+1)< 5
−(m+1/pr)p
rm + 1 p
r+ 1
γr(pr−1)(m+1/pr), (III) (nr)
prm(p
rm + 1)
prm< 5
−mm
γr(pr−1)m,
(IV) (nr)
prm+1(p
rm+2)
prm+1< 5
−(m+1/pr)p
rm + 1 p
r+ 1
γr(pr−1)(m+1/pr). S t e p 3: the sets G
l, Q
r, P
r, I
r,jand J
r,j. Being a G
δ-subset of R
n, G is equal to the intersection of a sequence (G
l)
l∈Nof open subsets of R
nthat we may suppose decreasing.
Proceeding as in the construction of the special compact cover of an open set, we find that each G
lis the union of countably many compact cubes Q
l,m,hthat we may renumber as a sequence, say (Q
l,k)
k∈N. Then for every l, k ∈ N, we denote by P
l,kthe compact cube in R
nhaving the same center as Q
l,kand
32diam(Q
l,k) as diameter. Now we arrange N
2into a sequence ((l
r, k
r))
r∈N, set
Q
r= Q
lr,krand P
r= P
lr,kr,
and let I
r,jand J
r,jfor j ∈ {1, . . . , n} be the compact intervals in R such that
Q
r= Y
n j=1J
r,jand P
r= Y
n j=1I
r,j.
Of course this construction leads to J
r,j⊂ I
r,jofor every r ∈ N and j ∈ {1, . . . , n}.
S t e p 4: the functions u
r,jand the numbers m
r. At this point, every- thing is set up to introduce the functions u
r,jfor r ∈ N and j ∈ {1, . . . , n}, as well as the sequence (m
r)
r∈Nof N by the following recursion.
An application of Proposition 5.5 to γ = γ
1and p = p
1leads to an integer m
1> q
1and to functions u
1,1, . . . , u
1,n∈ C
∞(R) such that, for every j ∈ {1, . . . , n},
(a) u
1,jhas no divergence point, (b) supp(u
1,j) ⊂ I
1,j,
(c) ku
(k)1,jk
R≤ 2
−m1, ∀k ∈ {0, 1, . . . , m
1}, (d) ku
(k)1,jk
R≤ 2
kk
γ1k, ∀k ∈ N,
(e) for every t ∈ J
1,j, one has either
|u
(p1,j1m1)(t)| ≥ 5
−m1m
γ11(p1−1)m1or
|u
(p1,j1m1+1)(t)| ≥ 5
−(m1+1/p1)p
1m
1+ 1 p
1+ 1
γ1(p1−1)(m1+1/p1).
Now, for an integer r ≥ 2, if the functions u
t,jfor t ∈ {1, . . . , r − 1}
and j ∈ {1, . . . , n} and the integers m
1, . . . , m
r−1are obtained, we ap- ply Proposition 5.5 to γ = γ
rand p = p
rand obtain an integer m
r>
sup{p
r−1m
r−1, q
r} and functions u
r,1, . . . , u
r,n∈ C
∞(R) such that, for ev- ery j ∈ {1, . . . , n},
(a) u
r,jhas no divergence point, (b) supp(u
r,j) ⊂ I
r,j,
(c) ku
(k)r,jk
R≤ 2
−mr, ∀k ∈ {0, 1, . . . , m
r}, (d) ku
(k)r,jk
R≤ 2
kk
γrk, ∀k ∈ N,
(e) for every t ∈ J
r,j, one has either
|u
(pr,jrmr)(t)| ≥ 5
−mrm
γrr(pr−1)mror
|u
(pr,jrmr+1)(t)| ≥ 5
−(mr+1/pr)p
rm
r+ 1 p
r+ 1
γr(pr−1)(mr+1/pr). S t e p 5: the functions u
rand u. Finally, for every r ∈ N, we define,
u
r(x) = u
r,1(x
1) . . . u
r,n(x
n), ∀x ∈ R
n, and consider the series u = P
∞r=1
u
r. For every k ∈ N, we certainly have k ≤ m
k. Therefore, for every α ∈ N
n0,
X
∞ r=sup{1,|α|}kD
αu
rk
Rn≤
X
∞ r=sup{1,|α|}2
−nmr≤ 1;
this implies that u is a bounded C
∞-function on R
n. Moreover, for every α ∈ N
n0such that |α| ≥ 1, we have
kD
αuk
Rn≤
|α|−1
X
r=1
kD
αu
rk
Rn+ X
∞ r=|α|kD
αu
rk
Rn≤
|α|−1
X
r=1