Ark. Mat.

, 36 (1998), 307316
c 1998 by Institut Mittag-Leer. All rights reserved

On a theorem of Kaufman: Cantor-type

construction of linear fractal Salem sets
Christian Bluhm(1 )

Abstract. In this paper we present a deterministic Cantor-type construction of linear fractal

Salem sets with prescribed dimension. The construction rests on a paper of Kaufman [10] where
he investigated the Fourier dimension of the set of -well approximable numbers in R.

1. Introduction
In metric Diophantine approximation, sets with fractal Hausdor dimension
occur often. As an example we recall the well-known theorem of Jarnk [7] and
Besicovitch [1], which states that for >0 the set E() of -well approximable
numbers in R has Hausdor dimension dimH (E())=2/(2+). In later years various authors generalized this theorem in many directions (see Dodson [4]). In 1981
Kaufman [10] proved that E() carries a probability measure with compact
support whose Fourier transform is of order

(x) = o(log |x|)|x|1/(2+)

(|x| ).

By a well-known theorem of Frostman (cf. Mattila [11]) this implies the lower bound
in the theorem of Jarnk and Besicovitch. Furthermore, it shows that the Hausdor
dimension of the support of equals its Fourier dimension, where the Fourier
dimension of a compact set K Rd is dened by
dimF (K) := sup{ [0, d ] | there is M1+ (K) with
(x) = O(|x|/2 ) (|x| )}.
Here M1+ (K) denotes the set of all probability measures with support in K, and

+
means the Fourier transform of a measure M1 (K) dened by

(x) := exp(2ix, y) (dy) (x Rd ).

(1 ) Supported by Deutsche Forschungsgemeinschaft DFG, Grant BL446/1-1.

bluhm.tex, typeset 2003-08-12 10:05 with arkiv.sty version 1998-11-04

308

Christian Bluhm

A compact set K Rd is called a Salem set if dimF (K)=dimH (K). The dimension of a Salem set means its Fourier resp. Hausdor dimension. The theorem of
Frostman already mentioned implies that the Fourier dimension of compact sets is
majorized by their Hausdor dimension. In certain random constructions the occurrence of fractal Salem sets seems to be natural (cf. Kahane and Mandelbrot [9],
Kahane [8, Chapters 1718], Salem [12], and Bluhm [3]), but Kaufmans work mentioned above is the only deterministic construction of a Salem set of prescribed
dimension known to the author. However, his account is by no means easy to follow. In this paper we have tried by modifying his construction and casting it in a
more geometric form to produce an easier deterministic construction of linear Salem
sets with prescribed dimension. This work is an extended version of Chapter 2 in
the authors dissertation [2].

2. Cantor-type constructions
First of all we need some notation. For xR
x := min |xm|
mZ

describes the distance from x to the nearest integer. The set of prime numbers will
be denoted by P, and we set
PM := P[M, 2M ]
for a positive integer M . Now we explain the Cantor-type construction considered
in this paper. Fix >0 and choose a sequence of positive integers (Mk )kN with
M1 < 2M1 < M2 < 2M2 < M3 < 2M3 < ... .
Later we are going to determine recursively a sequence (Mk )kN for which the set
S :=

{x [0, 1] | px p1 }

k=1 pPMk

is a Salem set of dimension 2/(2+) (Theorem 3.3). Let us for a moment explain
the structure of S . For abbreviation we set

q () := {x [0, 1] | qx q1 }
E

bluhm.tex, typeset 2003-08-12 10:05 with arkiv.sty version 1998-11-04

On a theorem of Kaufman: Cantor-type construction of linear fractal Salem sets

309

q () can be written as a union of closed intervals:

for every qN. Obviously, E
q1 

(1)

q () = [0, q2 ] 
E

m=1


m 2 m 2
q
, +q
[1q 2 , 1].
q
q

Therefore, the set S is compact.

We assume the following Condition 2.1 on (Mk )kN to be fullled throughout
the paper.
Before formulating it we should recall the prime number theorem in the following form (Hardy and Wright [6, (22.19.3)])
(2)

lim

M +

#PM
= 1,
M/ log M

where #A denotes the number of elements of a nite set A. Therefore, if M1 is

large enough we are able to nd a sequence (Mk )kN which fullls the following
condition.
Condition 2.1. Let M1 N be large enough so that
PMk =

and

#PMk

Mk
2 log Mk

for every kN.

We are now in a position to state the following proposition.
Proposition 2.2. The set S is a nonempty compact set in [0, 1] and has
nite Hausdor measure for the measure function h(r)=r 2/(2+) log(e+r1 ).

p () for every
Proof. For proving S = it is sucient to observe that 0, 1 E
pPMk and that PMk = for all kN (Condition 2.1).
q () can be covered by q+1 intervals of length 2q2 .
For qN the set E
Once more applying the prime number theorem (2) it is straightforward to show
that the set S has nite Hausdor measure for the measure function h(r)=
r2/(2+) log(e+r1 ).

As an immediate consequence we obtain dimF (S )dimH (S )2/(2+).
Remark 2.3. Closely related to S is the set
E() :=

{x [0, 1] | qx < q 1 }

k=1 q=k

of -well-approximable numbers. As mentioned in the introduction, the Hausdor

dimension of E() is 2/(2+), which was proved by Jarnk [7] and Besicovitch [1].
However, the set E() is dense in [0, 1] and therefore quite dierent from S .

bluhm.tex, typeset 2003-08-12 10:05 with arkiv.sty version 1998-11-04

310

Christian Bluhm

3. Construction of
In [10] Kaufman constructed a positive measure with support in E() whose
Fourier transform is of order

(x) = o(log |x|)|x|1/(2+)

(|x| ).

In this section we construct a measure with a similar decay and support in S

based on a certain sequence (Mk )kN , which will be constructed recursively according to Lemma 3.2 below. The construction rests on a modication of Kaufmans
construction. Here Cc2 (R) denotes the space of all twice continuously dierentiable
functions with compact support.
Before stating Lemma 3.2 we need to introduce some functions. Fix M N
with
R := (4M )1 < 12 ,

(3)

5
and dene a function FM on [ 12 , 12 ] by FM (x)= 15
(R2 x2 )2 when |x|R,
16 R
1
FM (x)=0 when R<|x| 2 . In the following we assume FM to be dened on the
whole real line with period 1. Because FM C 2 (R) its Fourier series FM (x)=

(M ) 2imx
(M )
converges uniformly to FM , where the Fourier coecients am
mZ am e
are given by

1/2
(M )
am =
FM (t)e2imt dt.
1/2

By (partial) integration we obtain

(4)

(M )

a0

= 1,

)
|a(M
m | 1,

and

)
2 2
|a(M
R
m |m

(m N)

for the Fourier coecients of FM . Now set


 
) 2impx
qM (x) :=
FM (px) =
a(M
.
m e
pPM

mZ pPM

Therefore, qM C 2 (R) is a 1-periodic function. We intend to normalize qM by

1
M (0)=1. Because of
multiplication with a constant c1
M in order to obtain cM q

)
a(M
(5)
qM (k) =
m
mZ
pPM
k=mp

it is clear that one has to choose cM :=#PM . For abbreviation we set

gM := c1
M qM .

bluhm.tex, typeset 2003-08-12 10:05 with arkiv.sty version 1998-11-04

On a theorem of Kaufman: Cantor-type construction of linear fractal Salem sets

311

Proposition 3.1. If gM (x)>0, then there exist pPM with pxp1 .

Proof. The function FM is 1-periodic, which yields to
gM (x) > 0 = there are p PM and m Z with |pxm| R = (4M )1 < 12 .
This implies the assertion of the proposition.

Roughly spoken, we are going to construct a measure with support in S by

repeated multiplication of densities gMk where (Mk )kN will be chosen recursively
according to the following lemma. We introduce the function
(x) := (1+|x|)1/(2+) log(e+|x|) log log(e+|x|)
for the sake of a clearer presentation.
Lemma 3.2. For every Cc2 (R) and >0 there exists a positive integer
M0 =M0 (, ) such that

|[gM ] (x) (x)|

(x)

for x R,

for all M M0 .
Before proving the lemma we use it for the construction of an appropriate
sequence (Mk )kN and a corresponding measure carried by S .
We start with a function 0 : R+ R+ with the properties


2
0 (x) dx = 1, 0 |]0,1[ > 0, and 0 |R\[0,1] 0.
(6)
0 Cc (R),
Now we choose 0< < 12 . According to Lemma 3.2 we nd
M1 = M1 (0 , 21 ),
M2 = M2 (0 gM1 , 22 ),
..
.
Mk = Mk (0 gM1 gM2 ... gMk1 , 2k ) (k N).
We assume S to be constructed according to (Mk )kN . Now we build products
G0 := 1,

and

Gk :=

k


gMj

j=1

bluhm.tex, typeset 2003-08-12 10:05 with arkiv.sty version 1998-11-04

(k N).

312

Christian Bluhm

Using Lemma 3.2 we obtain for all kN0 and all xR

|[0 Gk+1 ] (x)[0 Gk ] (x)| 2k1 (x).

(7)

Denote by 1 the Lebesgue measure on the Borel sets in R. Dene a sequence of

measures by
k := 0 Gk 1 (k N0 )
with Fourier transforms
k (x)=[0 Gk ] (x) (xR, kN0 ). Because of (7) the sequence (
k )kN0 is a Cauchy sequence with respect to the supremum norm. This
implies that
there exists M1+ ([0, 1])

such that

ck

where denotes weak convergence and c=c( ) is a positive constant which normalizes to mass 1. Now (under assumption of Lemma 3.2) we are able to prove
the main theorem of this paper.
Theorem 3.3. The measure obeys

(x) = O((x))

(|x| ).

Therefore, S is a Salem set of dimension 2/(2+).

Proof. The claimed Fourier asymptotic of follows from (7) and a simple geometric series argument, also taking into account that
p (x)=O(|x|2 ) for
xed p. The second assertion follows from Proposition 2.2, dimF (S )dimH (S ),
and Proposition 3.1 (which implies that the support of is contained in S ).

4. Proof of Lemma 3.2

Our task in this section is to prove Lemma 3.2. To begin with, x M N for a
(M )
moment. Because of (5) and |am |1 we have |
qM (k)|#{(m, p)ZPM |k=mp}.
Because |k| has a unique factorization by prime numbers, we easily obtain
|
qM (k)|

(8)

log |k|
log M

for all kZ\{0}. Additionally, by (4) and (5) we have the implication
mp = k
(9)

|k|
|k|
)
2 2

= |a(M
R 4k 2 M 2 R2 ,
m |m
p
2M
4k 2 M 2 R2 log |k|
= |
qM (k)|
log M

= |m| =

for all kZ\{0}. We prove Lemma 3.2 in three steps.

bluhm.tex, typeset 2003-08-12 10:05 with arkiv.sty version 1998-11-04

On a theorem of Kaufman: Cantor-type construction of linear fractal Salem sets

313

Step 1. There exists M1 >0 and A=A()>0 such that for all M M1 ,
|
gM (k)| AM 1 log M

for all k Z\{0},

|
gM (k)| A|k|1/(2+) log |k|

for all k Z with |k| > 4M R1 .

Proof of Step 1. We consider two cases.

Case 1: 1|k|4M R1 . Using (8) with Condition 2.1 we have
c1
2M 1 log M (log(4M )log R)
M log |k|

log M
log M
1
= 2M (log 4+log M +(1+)(log 4+log M )) 4(2+)M 1 log M.

qM (k)|
|
gM (k)| = c1
M |

Case 2: |k|>4M R1 =(4M )2+ . By (9) we obtain the following estimation,

qM (k)|
|
gM (k)| = c1
M |

2
c1
M 2 R2 log |k| 2M 1 (log M )4k 2 M 2 R2 log |k|
M 4k

log M
log M

= 8k 2 M R2 log |k| = 8k 2 41 (4M )3+2 log |k| 2|k|1/(2+) log |k|

It remains to show that |
gM (k)|AM 1 log M for |k|>4M R1 for all M M1 with
large M1 and some constant A=A(). But this is easily veried by combining
elementary properties of the logarithm with the estimations in Case 2 (e.g. A=
2(2+)).

From now on let always M M1 , and let Cc2 (R) be given.
Step 2. There exists B =B(, )>0 such that

BM 1 log M
|[gM ] (x) (x)|

for x R.

Proof of Step 2. Writing gM as a Fourier series we obtain



[gM ] (x) =
gM (k)(xk).
kZ

Then, Cc2 (R) and gM (0)=1 imply




|
gM (k)||(xk)|
B1
|
gM (k)|(1+|xk|)2
|[gM ] (x) (x)|
k=0

B1

k=0

(1+|xk|)2 sup |
gM (k)| BM 1 log M,

k=0

k=0


with constant B =B(, ):=2AB1 k=1 k 2 , where A=A() is the constant from
Step 1. Therefore, Step 2 is proven.

Now let >0 be arbitrarily small.

bluhm.tex, typeset 2003-08-12 10:05 with arkiv.sty version 1998-11-04

314

Christian Bluhm

Step 3. There exists M2 >0 such that for all M M2 ,

|[gM ] (x) (x)|

(x)

for x R.

Proof of Step 3. We consider two cases.

Case 1: |x|<8M R1 =2(4M )2+ . In this case the assertion follows for large
M from Step 2 and some tedious manipulation.
Case 2: |x|8M R1 . We divide the sum


B1 |
gM (k)|(1+|xk|)2

k=0

arising in the proof of Step 2 in two parts by summing rst over k with |xk| 12 |x|
and second over k with |xk|< 12 |x|. It is easy to see that for large M the rst
sum is majorized by C1 |x|1 with a constant C1 independent of M and x. For
estimating the second sum we apply Step 1 and use 12 |x|4M R1 to obtain


B1 |
gM (k)|(1+|xk|)

2B1

|xk|<|x|/2

k=1

B
for all M which are large enough.

sup
|k|>|x|/2

sup
|k|>|x|/24M R1

|
gM (k)|

(|k|1/(2+) log |k|) (x)

The assertion of Lemma 3.2 follows by choosing M0 (, )=M2 .

5. Conclusions
The proof of Lemma 3.2 shows that the construction of M0 (, ) is explicit.
Therefore, Lemma 3.2 provides a recursive explicit construction of the sequence
(Mk )kN .
By choosing an appropriate >0 the method of this paper results in an explicit method for constructing linear (fractal) Salem sets with prescribed dimension
in ]0, 1[.
It is possible to generalize the results of this paper in two directions.
The rst consists in considering a decreasing function :NR+ instead of
the function q q 1 . This leads to sets S (closely related to the set E() of
-well approximable numbers) instead of S . Dodson [4] calculated the Hausdor
dimension of E(), and in [2] we proved that S is a Salem set. The second

bluhm.tex, typeset 2003-08-12 10:05 with arkiv.sty version 1998-11-04

On a theorem of Kaufman: Cantor-type construction of linear fractal Salem sets

315

generalization consists in considering all xRd with |x|S . Then a paper of

Gatesoupe [5] shows that this leads to Salem sets in Rd (invariant under rotations)
with dimensions in ]d1, d[ (see [2]).
Acknowledgements. The author thanks his supervisor Professor D. K
olzow for
steady support. The author is extremely indebted to Professor R. S. Strichartz for
his hospitality at Cornell. He also would like to thank Professor T. W. K
orner for
some helpful comments on the manuscript. The author is grateful to an anonymous referee for suggesting various improvements which led to a shorter and clearer
presentation of Section 3.

bluhm.tex, typeset 2003-08-12 10:05 with arkiv.sty version 1998-11-04

316

Christian Bluhm:
On a theorem of Kaufman: Cantor-type construction of linear fractal Salem sets

References
1. Besicovitch, A. S., Sets of fractional dimensions (IV): on rational approximation
to real numbers, J. London Math. Soc. 9 (1934), 126131.
2. Bluhm, C., Zur Konstruktion von Salem-Mengen, Ph. D. Dissertation, Erlangen,
1996.
3. Bluhm, C., Random recursive construction of Salem sets, Ark. Mat. 34 (1996), 5163.
4. Dodson, M. M., Geometric and probabilistic ideas in metric Diophantine approximation, Uspekhi Mat. Nauk 48:5 (1995), 77-106 (Russian). English transl.:
Russian Math. Surveys 48:5 (1993), 73102..
5. Gatesoupe, M., Sur un theor`eme de R. Salem, Bull. Sci. Math. 91 (1967), 125127.
6. Hardy, G. H. and Wright, E. M., An Introduction to the Theory of Numbers, 5th
ed., Oxford Univ. Press, Oxford, 1979.
7. Jarnk, V., Zur metrischen Theorie der diophantischen Approximation, Prace Mat.Fiz. 36 (1928/29), 91106.
8. Kahane, J.-P., Some Random Series of Functions, 2nd ed., Cambridge Univ. Press,
Cambridge, 1985.
9. Kahane, J.-P. and Mandelbrot, B., Ensembles de multiplicite aleatoires, C. R.
Acad. Sci. Paris Ser. I Math. 261 (1965), 39313933.
10. Kaufman, R., On the theorem of Jarnk and Besicovitch, Acta Arith. 39 (1981),
265267.
11. Mattila, P., Geometry of Sets and Measures in Euclidean Spaces. Fractals and Rectiability, Cambridge Univ. Press, Cambridge, 1995.
12. Salem, R., On singular monotonic functions whose spectrum has a given Hausdor
dimension, Ark. Mat. 1 (1950), 353365.
Received May 21, 1997
in revised form February 12, 1998

bluhm.tex, typeset 2003-08-12 10:05 with arkiv.sty version 1998-11-04

Christian Bluhm
Institut f
ur Mathematik und Informatik
Universit
at Greifswald
Jahnstrae 15a
D-17487 Greifswald
Germany
email: bluhm@rz.uni-greifswald.de