You are on page 1of 22

The Kakeya Problem

1407493

1
Contents
1 Introduction and History 3

2 Besicovitchs Original Construction by Perron Trees 4

3 Basic Definitions and Results 9


3.1 Hausdorff Measure and Dimension . . . . . . . . . . . . . . . . . 10
3.2 Irregularity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

4 Projection Of Fractal Sets 13

5 Duality of points and lines 15

6 Projective proof of Kakeya 17


6.1 Cantor Dust . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

7 Further Problems 21
7.1 Kakeya Set Conjecture . . . . . . . . . . . . . . . . . . . . . . . . 21

8 Bibliography 21

2
Abstract
In 1917 Japanese mathematician S.Kakeya posed the problem of find-
ing the smallest planar region in which a unit line segment could be rotated
through 180 without leaving the region. It was later shown in 1928 by
A.S.Besicovitch that such a region could have arbitrarily small measure.
Here I will demonstrate the original construction of such sets by Besi-
covitch using a technique called the Perron tree, I will then spend the
remainder of the text building up to a second, more surprising, construc-
tion by Besicovitch using the notion of duality.

1 Introduction and History


In 1917 the Japanese mathematician S.Kakeya posed the problem:
What is the minimal plane area in which a given figure can complete a full
revolution?
In 1920 this was included alongside some of Kakeyas other geometrical ques-
tions in the American Mathematical Monthly. [1] The question of minimal area
here was included with some special cases, the most famous of which become
known as the Kakeya problem.
What is the minimal plane area in which a needle can complete a full
revolution?
Definition 1 (Needle). A needle is a unit line segment

Then the problem involves investigating sets in which we have the ability to ro-
tate a needle, clearly the disk of unit diameter is such a set with area 4 0.785.
When first studied an additional condition of convexity was raised to which
M.Fujiwara and S.Kakeya conjectured [2] the solution to be an equilateral tri-
angle of height 1 which has area 13 0.577, this was later proved by J.Pal
in [3].
For the problem without convexity it seemed that Kakeya believed the mini-
mum simply connected solution would be that of the three-pointed deltoid with
area 8 0.393, yet it was shown in 1965 by both Melvin Bloom and I.J Schoen-
berg independently that it was
possible to create simply connected Kakeya sets
(52 2)
with area tending to 24 0.284, which is now known as the Bloom-
Schoenberg number. This was then conjectured to be the lower bound for a
simply connected Kakeya set, until 1971 when F.Cunningham demonstrated [4]
that for some > 0 we can find a simply connected kakeya set E such that

(E) < 108 + 0.03.

Figure 1: The three regions described above

3
In a similar time to when Kakeya first posed the problem Russian mathe-
matician A.S.Besicovitch was working on the seemingly unrelated problem in
analysis [5];

If f is a Riemann integrable function defined on the plane, is it always


possible to find orthogonal
R coordinate axis such that with respect to these
coordinates the integral f (x, y)dx exists as a Riemann integral for all y
where the resulting function of y is also Riemann integrable?

Besicovitch realised that by constructing a compact set in the plane of Lebesgue


measure zero which contained a unit line segment in every direction he would
have a counterexample to the statement of his question. He succeeded in con-
structing such a set, a construction which was then simplified over time to give
the construction of the Perron Tree that well see later. This result was pub-
lished in 1919 [5] during the Russian civil war, a period when there was little
communication available with the rest of the world, due to this Besicovitch
hadnt heard of Kakeyas problem and so it wasnt until he left Russia for Cam-
bridge that he realised a small modification of his result could give the existence
of a Kakeya set with arbitarily small measure.

2 Besicovitchs Original Construction by Perron


Trees
Before we get started with the construction we need to make the important
distinction between two very similar definitions that will appear in this text.
Definition 2 (Besicovitch Set). A Borel set containing a line segment in every
direction which has measure 0

Definition 3 (Kakeya Set). A set in which a needle can be moved through a


full rotation is a Kakeya Set
We should note that by the definitions a needle cannot be rotated at all in
a Besicovitch set as a rotation by any angle, as small as we like, requires some
area to take place. A proof of this is given below as taken from [6, 1.22].

Theorem 1. There doesnt exist a Kakeya set of measure zero.


Proof. Assume for the sake of contradiction that there does exist such a Kakeya
set E such that (E) = 0. Then there exists a continuous function

l : [0, 1] E
l : t 7 l(t)

taking times t to unit line segments in E which we can parameterise by

l(t) = {(x(t) + s cos (t), y(t) + s sin (t)) : 0.5 s 0.5}

for some continuous functions x, y, : [0, 1] R. Then a continuous function


on a compact set is uniformly continuous so we have that for some fixed > 0

|x(t) x(t0 )|, |y(t) y(t0 )|, |(t) (t0 )| 0.001

4
for t, t0 [0, 1] whenever |t t0 | < Then we can find t0 < t1 [0, 1] with
|t0 t1 | < such that (t0 ) 6= (t1 ) as (t) must not be constant (else the
needle wouldnt rotate).
We can assume that x(t0 ) = y(t0 ) = (t0 ) = 0 and then choose any a
[0.4, 0.4]. By the above inequality it is clear that for t (t0 , t1 ) the line
l(t) intersects the vertical line x = a at some point (a, ya (t)). Then as we
vary t (t0 , t1 ) the point ya (t) varies continuously. So by the fact that each
(a, ya (t)) E the intermediate value theorem implies that the interval between
(a, ya (t0 )) and (a, ya (t1 )) must lie in E.
If we take the union of this interval with all those corresponding to each a
[0.4, 0.4] then we find a non-trivial sector contained in E, and so E cannot
have zero measure.

As mentioned earlier the original construction for a Besicovitch set [7] has
since been simplified by a number of mathematicians including; Perron [8],
Schoenberg [9], Besicovitch himself [10], and more. The construction I demon-
strate will follow very closely both that given by Falconer in [11] and also Bar-
rowclough in [12].
Lemma 1. For a triangle T with base on line L of length 2b and height h we
can form two subtriangles T1 and T2 each with base length b and height h, then
by translating T2 towards T1 along the base line by a distance of 2(1 )b for
some ( 12 , 1) we form a new figure S such that;
S is the union of a triangle U with two auxiliary triangles
U is homothetic to T (similar and similarly situated) with (U ) = 2 (T )
(T ) (S) = (T )(1 )(3 1)
Where is the planar area or plane Lebesgue measure
Proof. This proof is a simple calculation so has been omitted for sake of brevity.

Theorem 2. Let T be a triangle based on the line L. By dividing the base of


T into 2k equal segments and joining each point to the top vertex of T we form
2k subtriangles T1 , T2 , . . . , T2k . Then by choosing k to be large enough we can

5
translate the subtriangles in such a way along L that the resulting figure S has
arbitrarily small area.

Proof. First consider consecutive triangles T2i , T2i1 for 1 i 2k1 then move
T2i along L towards T2i1 as described in the above lemma to obtain 2k1 new
figures S11 , S21 , . . . , S21k1 where each Si1 consists of a triangle Ti1 homothetic to
T2i1 T2i with (Ti1 ) = 2 (T2i1 T2i ) and two auxiliary triangles. Then by
the lemma the new figures have area

(Si1 ) = (T2i T2i1 ) (T2i T2i1 )(1 )(3 1)



k1
k1
2[ 2[
1
= Si = (T2i T2i1 ) (T2i T2i1 )(1 )(3 1)
i=1 i=1

= (T ) (T )(1 )(3 1)

So we see an overall reduction in area by (T )(1 )(3 1)


We now continue the construction by applying the same technique to con-
secutive Si1 for 1 i 2k1 since the triangles T2i 1 1
and T2i1 in each pair of
figures share a common side we can apply the lemma again. Forming 2k2 fig-
ures S12 , S22 , . . . , S22k1 each consisting of a triangle Ti2 homothetic to T2i1
1 1
T2i
2 2 1 1 2 1 1
with (Ti ) = (T2i1 T2i ) = ((T2i1 ) + (T2i )) and auxiliary triangles.
1 1
Then as before the reduction in area by overlapping T2i1 and T2i is at least

(Si2 ) (Si1 ) = (1 )(3 1)((T2i1


1 1
) + (T2i ))
= (1 )(3 1)2 (T4i3 T4i2 T4i1 T4i ))

Then we repeat this procedure in an inductive way, where at the (r + 1)th


stage in the construction we obtain figures Sir+1 by sliding previous figures S2i
r
r
and S2i1 in such a S way to reduce the area through overlapping by at least
(1 )(3 1)2r ( j2i Tj ) where 2i is the set of indices relating to the
r r
elementary triangles which form S2i and S2i1 . Then
S after k iterations we have
k
a final figure S1 =: S. If we compare the areas for i Sir for increasing r we get
that

(S) (T ) (1 )(3 1)(1 + 2 + + 2(k1) )(T )


(3 1)(1 2k )
 
= 1 (T )
(1 + )

So by examining this formula it doesnt take much to realise that by making


close enough to 1 and k as large as necessary we can make (S) as small as we
require.

6
From this construction we are left with a set known as a Perron-Schoenberg
tree, the point of this was to construct a planar set with line segments at every
angle between 60 and 120 to the horizontal, this was achieved by starting with
the equilateral triangle, a set which clearly satisfies this angle condition then
through cutting and sliding this triangle we have preserved these line segments
whilst making its area arbitrarily small. These sets can then be duplicated
and rotated to form a set of arbitrarily small area with a line segment in every
direction.
By constructing this set we are much closer to a solution to the Kakeya prob-
lem, the problem we are now faced with is the fact that although we have line
segments representing each angle as required we could not continuously rotate
between them, by inspecting the figure above you can convince yourself of this.
We require the ability to jump between parallel branches of the tree we have
made. The solution to this problem came from the Hungarian mathematician
J.Pal (who also proved the equilateral triangle gave the solution to the Kakeya
problem in convex sets) using a method now known as the P al join.
Lemma 2 (P al join). For lines L1 and L2 parallel in the plane and a given
> 0 there exists a set E with L1 , L2 E and (E) < such that a unit
segment may be moved continuously from L1 to L2 remaining in E.
[4]

7
Proof. Consider two points x1 , x2 on the lines L1 and L2 respectively and let M
be the line segment joining x1 and x2 . Then denote by E the set of L1 , L2 , M
and the unit sectors Si between Li and M (shaded grey in the figure above).
We want to show that we can make the area of E as small as we desire. This
can be done by taking x1 and x2 to be as far apart as necessary to ensure that
< ( the area of the shaded sectors) then

(E) = (S1 ) + (S2 )


= 2(S1 )
12
=2
2
=<

Then through this method we can see that by sliding a unit line segment up
and down L1 rotating at point x1 sliding through M and rotating back at x2
we can travel between parallel lines in arbitrary area.
We are now ready to give our solution to the Kakeya problem.

Theorem 3 (The Kakeya Problem). For > 0 there exists a set E such that a
unit segment may be moved continuously inside E to its original position having
gone through a rotation of 180 and with (E) <
Proof. E will be formed of three sets E0 , E1 , E2 which each will have the prop-
erty of being able to rotate through 60 from its original position, and such that
each for each Ei we have (Ei ) < 3 .
Starting with E0 we first consider a triangle T with unit height positioned on
a baseline L, then we divide T into 2k elementary triangles T1 . . . T2k and slide
S2k
them along L in such a way that ( 1 Ti ) < 6 as we did in the construction
k
of the Perron-Schoenberg trees. Denote E 0 = S2 Ti , in our construction of
1
these we have the property that for each i = 1, . . . , 2k 1 one side of Ti is
parallel to the opposite side of Ti+1 and so we may use a Pal join to find a set
Pi with (Pi ) < 6 21k such that a unit segment may be moved from Ti to Ti+1
continuously without leaving E 0 Pi . Then we define E0 = E
0 P1 P2k 1

8
hence
k
2X 1
0 ) +
(E0 ) = (E (Pi )
i=1

< + (2k 1) k
6 2 6

<
3
and we have that a unit line segment can be moved continuously to its original
position having gone through a rotation of 60 inside E0 . Then we defined E1
and E2 as copies of E0 rotated by 60 and 120 respectively. Then the set
E := E0 E1 E2 satisfies our requirements as a Kakeya set with (E) <
(It may be necessary to use some more Pal joins to connect the components Ei
but as weve seen this adds arbitrarily small area.)
And so finally we have shown the existence of an arbitrarily small solution
to the Kakeya problem, whilst this proof doesnt use any difficult ideas and
is geometrically quite pleasant, it is very cumbersome. Throughout the rest
of this paper we will give the necessary information to give a proof using the
method of duality which we will see gives the result much quicker if we have
some knowledge of projections.

3 Basic Definitions and Results


The method of duality which we plan to use requires some key results in ge-
ometric measure theory which this chapter will give the underlying knowledge
for. This section will heavily follow that of Falconer [13] throughout
The most basic reason for the study of measure theory is to generalise our ideas
of length, area, volume. These ideas become very important in the study of
integration or probability theory. In our discussion we will only be considering
measures on Borel subsets of Rn .
Definition 4. A function : P(Rn ) [0, ] (Where P(Rn ) is the power set
of Rn ) is called a measure if
() = 0
(A) (B) if A B
if A1 , A2 , . . . is a countable collection of subsets of Rn then

!
[ X
Ai (Ai )
i=1 i=1

This definition is actually that of an outer measure where we have that Borel
sets are measurable but for our purposes as we will only be discussing Borel sets
we give this definition to save a discussion of measurability.
Example. The most intuitive measure if probably the counting measure, where
we define (A) to be the number of points in A if A is finite and (A) =
otherwise.

9
Example. When considering subsets of R the Lebesgue measure follows our
idea of length where for normal open and closed intervals we have ((a, b)) =
([a, b]) = |b a|, then for arbitrary sets A we define
(
)
X [
(A) = inf (bi ai ) : A [ai , bi ]
i=1 i=1

This Lebesgue measure can be generalised to higher dimension which is


exactly what we used in our earlier presentation of the Perron-Schoenberg Trees
to give a notion of area.

3.1 Hausdorff Measure and Dimension


Definition 5. For a subset A Rn and given s > 0. Define for any > 0
(
)
X [
s s
H (A) = inf diam(Ui ) : A Ui and diam(Ui ) <
i=1 i=1

Where diam(U ) = sup{|x y| : x, y U }


Then the s-dimensional Hausdorff measure of A is

Hs (A) = lim Hs (A)


Its easy to see this limit exists by considering the fact that as we take
smaller the number of suitable covers of our set increases, (i.e. if 0 < then
every 0 cover is also a cover). So Hs is monotonically decreasing in .
Proposition 1. Hs satisfies our definition of a measure
Proof. First note it is clear that the image of Hs (A) [0, ]. To prove this is a
measure we then we need to show the three properties from our definition hold
for Hs .

Hs () = 0 : diam() = 0 so by taking A = we get Hs (A) = 0 and


hence Hs (A) = 0

A B = Hs (A) Hs (B) : Every -cover of B is a -cover of A,


therefore Hs (A) Hs (B) for each , and in taking limits the inequality
holds.

Hs ( Hs (Aj ) : For > 0 choose a -cover of Aj , {Uj,i :


S P
jJ Aj ) jJ
i Ij }, where Ij is countable, such that
X
diam(uj,i )s Hs (Aj ) +
2j
iIj

10
S
Then we have that j=1 {Uj,i : i Ij } is a -cover of A = Aj
XX
So Hs (A) (diam(Uj,i )s )
j=1 iIj
X 
Hs (Aj ) +
2j
jN

X
= 2 + Hs (Aj )
j=0
P
= Hs (A) j=0 Hs (Aj )
Hence by taking the supremum in (equivalent to taking limit due to
monotone increasing function) we get that

X
Hs (A) sup Hs (Aj )
j=0

X
sup Hs (Aj )
j=0

X
= Hs (Aj )
j=0

The idea of dimension is to indicate how much space a set occupies near to
each of its points. The Hausdorff dimension is the most widely used definition
in fractal geometry to give a dimension to any set.
To come up with our definition of the Hausdorff dimension we need to again
consider the definition we used in the measure. Note that for A Rn and
< 1 we have that Hs (A) is non-increasing with s and hence the same holds
for Hs (A). If we have t > s we can actually say more about the relation when
changing s. Suppose also that {Ui : i I} is a -cover of A then we have
X X X
diam(Ui )t diam(Ui )ts diam(Ui )s ts diam(Ui )s
iI iI iI

Then if we take infimum over -covers we get Ht (A) ts Hs (A) and so by


taking limits in 0 its clear that if Hs (A) < then Ht (A) = 0 t > s

11
This point at which the s-dimensional Hausdorff measure jumps from to
0 is the Hausdorff Dimension of A, dimH (A).
Definition 6. For a set A Rn
dimH (A) = inf{s 0 : Hs (A) = 0} = sup{s : Hs (A) = }
Where by convention we define the supremum of to be 0.

It is for our purposes this idea of dimension which we will use to categorise
a fractal set, giving the same definition as Mandelbrot in his original essay
where he coined the term fractal. Here as for all of our purposes we restrict our
discussion to Euclidean space.
Definition 7. A fractal set is some set A Rn for which the Hausdorff dimen-
sion dimH (A) is strictly greater than its topological dimension

Unfortunately this definition isnt a strict definition for a fractal as it misses


out some sets which ought to be seen as fractals. Its interesting that this def-
inition makes no reference to self-similarity which is often one of the defining
characteristics when a lay-person thinks of a fractal.
When we come to the proof of Kakeya by duality we will consider lines of
the form Lc : x = c, it will be useful then to have some results on the Hausdorff
dimension of sets of the form E Lc .
Proposition 2. For a subset E R2 if 1 s 2 then
Z
Hs1 (E Lx )dx Hs (E) (3.1)

Proof. For some > 0 consider a -cover {Ui } of E such that


X
diam(Ui )s Hs (E) +
i

Then for each Ui we can find a square Si of side length diam(Ui ) with sides
parallel to the coordinate axes such that Ui Si . Then for each x the collection
{Si Lx } forms a -cover of E Lx and so
X
Hs1 (E Lx ) diam(Si Lx )s1
i
X
= diam(Ui )s2 diam(Si Lx )
i
X Z
= diam(Ui )s2 1Si (x, y)dy
i

Then
Z X Z Z
Hs1 (E Lx )dx diam(Ui )s2 1Si (x, y)dxdy
i
X
= diam(Ui )s
i
Hs (E) +

12
By taking , 0 we get
Z
Hs1 (E Lx )dx Hs (E)

Corollary 1. Let E R2 . Then for -almost all x (, ) dimH (ELx )


max{0, dimH (E) 1}
Proof. By taking s > dimH (E) we get that Hs (E) = 0. If s > 1, then (3.1)
implies we have that
Z
Hs1 (E Lx )dx Hs (E) = 0

and so Hs1 (E Lx ) = 0 then dimH (E Lx ) s 1 for almost all x.

3.2 Irregularity
One of the key features of fractal sets is that of a fine structure i.e. detail on
arbitrary small scale. This leads to difficulty in analysing the structure of such
sets using classical calculus where we find approximations to curves and surfaces
via tangent lines and planes.
One approach we can take with fractals is to study the densities of points in
local neighbourhoods, in this section we restrict our study to s-sets that is a set
with finite Hausdorff dimension s with a positive finite s-dimensional Hausdorff
measure.
Definition 8. For an s-set A Rn we define the upper and lower densities of
A at a point x Rn to be

Hs (A B(x, r))
Ds (A, x) = lim
r0 (2r)s
s
H (A B(x, r))
Ds (A, x) = lim
r0 (2r)s

where B(x, r) is the closed ball centered at x of radius r. If both limits are equal
i.e. Ds (A, x) = Ds (A, x) then we say the density of A at x exists and write
Ds (A, x)
Definition 9. We call a point x A a regular point of A if Ds (A, x) =
Ds (A, x) = 1 and otherwise x is an irregular point. We say A is a regu-
lar s-set if Hs -almost all of its points are regular points i.e. Hs ({x A :
x is not a regular point of A}) = 0 and an irregular s-set if almost all of its
points are irregular .

4 Projection Of Fractal Sets


This section again follows Falconer [13] chapter 6. Its possible to have some
intuitive sense of orthogonal projections in lower dimensions as shadows. For
example when we think of classical sets in R3 i.e. a 1-dimensional curve in

13
R3 we know will have a 1-dimensional shadow, a 2-dimensional surface we can
imagine its shadow on the plane to be 2-dimensional but we also know that a
3-dimensional solid will have a 2-dimensional shadow in the plane.

Figure 2: a) shows projections of regular sets in R3


b) shows the same projections for fractal sets. 1

The same is in general true for fractal sets that is, for a fractal set E R3
the projection will have dimension 2 if dimH (E) > 2 and dimension dimH (E)
if dimH (E) < 2 and in this section we shall prove some of the results that will
be useful when we come to consider duality in the next sections.
Definition 10. For a set E R2 and a line L , that is the line in R2 through the
origin which makes an angle of with the horizontal axis. We define proj (E)
to be the orthogonal projection of E onto L to be the points (x, y) L such
that line intersecting (x, y) perpendicular to L also intersects the set E at some
point.
We use the following projection theorem to formalize the discussion given
before the definition.
Theorem 4 (Projection Theorem). Let E R2 then
1. If dimH (E) 1 then dimH (proj (E)) = dimH (E) for almost all [0, ]
2. If dimH (E) > 1 then dimH (proj (E)) = 1 for almost all [0, ]
The proof of this is quite involved, using ideas of mass distributions and
Fourier series so I omit it from here but it can be found in

14
In Falconer this theorem has the following corollaries which will become
useful to us when we start to investigate duality.

Corollary 2. Let E be a 1-set in R2 . If E is irregular then proj (E) has length


(proj (E)) = 0 for almost all [0, ). If E has a regular part with positive
measure then proj (E) has positive length for all but at most one value of .
Corollary 3. A 1-set in R2 is irregular it has projections of zero length
in at least two directions.

5 Duality of points and lines


A concept that students can often go their whole university career without
experiencing is that of duality in points and lines. This is the idea of mapping
points to lines and lines to points. It often allows us to take a more easy
approach to what can originally be a difficult problem and the Kakeya problem
is a fantastic example of this. It was many years after his original proof that
Besicovitch realised that duality could give a much shorter proof. [10]
In this section I will give the most popular and one of the most natural
dualities on the Euclidean plane following [14], then in the next when we discuss
the way this technique can be used in application to the Kakeya problem I will
give the definition used by Falconer to present the argument by Besicovitch.

Consider two planes, which we will denote the primal plane and the dual
plane, for the primal plane we will use coordinate axes x, y and for the dual
plane we use u, v to save confusion. Then the duality is a mapping

: primal plane dual plane

where a point p and a line l are mapped to the line p and the point l respec-
tively

p = (px , py ) 7 p : v = px u py
l : y = ax + b 7 l = (a, b)

15
Remark. This definition isnt able to handle vertical lines With this definition
we have a couple of interesting properties: For a point p = (px , py ) and a line
l : y = ax + b
Self Inverse : (p ) = p
proof: p : v = px u py then (p ) = (px , (py ))
Swapped Vertical Order : p lies above (respectively below) the line
l l lies above (respectively below) the line p
proof: p is above l py > apx + b and similarly l = (a, b) is
above p : v = px u py b > px a py rearrangment shows these
inequalities are the same.
Preserved Vertical Distance : the vertical distance between p and l is
the same as that between p and l
proof: The vertical distance between p and l is given by |py l(px )| =
|py (apx + b)| then the vertical distance between p : v = px u py and
l = (a, b) is |(b) p (a)| = | b (px a py )| = |py (apx + b)|
Incidence : p l l p
proof: Follows from the previous property, p l vertical dis-
tance from p to l is 0. Then the vertical distanc between duals is also 0.

We can also consider the dual of a line segment, if we consider the segment
pq on a line l then we define the dual of this segment to be the double wedge
not containing the vertical line through point l , this definition can be seen by
considering a point r moving along the segment and noticing how its dual r
rotates between the lines p and q about the point l .

Armed with this knowledge of duality we can consider an example problem


[15] where this approach will great simplify our solution.
Problem. In a set of n points determine whether there exist 3 which are collinear
A brute force attempt at solving this would see us checking each set of 3
points so we would only be able to solve the problem for every case in O(n3 )
time.

16
If we dualize this problem however we find the much simpler problem of
finding three lines intersecting at a point. Having our problem in this dualized
for we can use the idea of an arrangement to find a faster solution. An arrange-
ment A(L) for a set of lines is a subdivision induced by the line set. It consists
of faces, edges and vertices.
The problem of creating an arrangement has algorithms with runtime O(n2 )
where you incrementally add lines, then at each incrementation it is a constant
time check to see if the addition of the new line will cause 3 lines to intersect,
hence by considering the dual we find a solution to our original problem which
is O(n3 ).

6 Projective proof of Kakeya


Throughout this section I follow [1113] When Besicovitch first gave his proof
using the method of duality in 1964 [16] he used a form of duality by a technique
called polar reciprocity:
For a point x we define L(x) = x to be the line distance |x| 1
from the
origin and perpendicular to the radius vector of x. This then has the same
property of incidence as the duality we described in the previous section where
x L(y) y L(x). Here we will use the form instead given by Falconer
[13] as it provides a much simpler proof than that of Besicovitch.

So using the same notation now as in Falconer [13] for a point (a, b) (R)2
we let L(a, b) = (x, y)S: y = a + bx, then if E is a subset in the plane we define
the line set L(E) = (a,b)E L(a, b). For a constant c we call the vertical line
x = c as Lc then

L(a, b) Lc = (c, a + bc) = (c, (a, b) (1, c))

So for a subset E (R)2

L(E) Lc = {(c, (a, b) (1, c)) : (a, b) E}

When we take a product (a, b) (1, c) this has the effect of projecting our
point (a, b) onto the line L where = arctan(c). The resulting projection is

17
p
also scaled by a factor (1 + c2 ) but as this has no effect on the Hausdorff
dimension of the set it can be ignored. So we obtain the facts that
dimH (L(E) Lc ) = dimH (proj (E)) (6.1)
(L(E) Lc )) = 0 (proj (E)) (6.2)

Then we have developed a relation using duality between projections of a


point set, which we have studied, and the intersection of L(E) with vertical
lines.
This knowledge of projections gives the following result
Theorem 5. For a Borel set E R2 consider the line set L(E). Then
a) dimH L(E) min{2, 1 + dimH E}
b) if E is a 1-set then (L(E)) = 0 E is irregular
Proof. For part a) we consider Theorem 4 we studied earlier which tells us that
dimH (proj E) = min{1, dimH E} for almost all [0, ) and hence by (6.1) in
the remark made above we have that almost everywhere
dimH (L(E) Lc ) = dimH (proj E) = min{1, dimH E}
Then by Corollary 1 from our study of the Hausdorff dimension we have that
for almost all c (, )
dimH (L(E) Lc ) max{0, dimH L(E) 1}
= min{1, dimH E} max{0, dimH L(E) 1}
= min{1, dimH E} dimH L(E) 1
= min{2, dimH E + 1} dimH L(E)
proving part a) of the theorem.
For part b) we let E be a 1-set (i.e. dimH E = 1 with 0 < H(E) < ) then
by Corollary 2 we know that if E is irregular then (proj E) = 0 for almost all
or (proj E) > 0 for almost all if E not irregular. So by (6.2) we get that
if E is irregular then

(L(E) Lc ) = 0 for almost all c


and if E isnt irregular then
(L(E) Lc ) > 0 for almost all c
Then using the simple fact that
Z
(L(E)) = (L(E) Lc )dc

the results follows immediately, clearly


(L(E) Lc ) = 0 = (L(E)) = 0
and
(L(E) Lc ) > 0 = (L(E)) > 0

18
Remark. We can also consider in particular projections onto the y-axis, for a
point (a, b) (R)2 then projecting to the y-axis we have proj 2 (a, b) = b which
is under our duality just the gradient of the line L(a, b). So for a set E with
b proj 2 E there must exist a line in L(E) with gradient b.
Theorem 6. There exists Besicovitch sets in the plane, i.e a set of 0 area which
contains a line in every direction.
Proof. Let E R2 be an irregular 1-set such that [0, 1] proj/2 . Then by the
previous theorem as E is irregular we know that L(E) has area 0. Then since
[0, 1] proj/2 and our remark above we know that L(E) must contain lines of
the form y = a + bx for each gradient b [0, 1] and so we have lines intersecting
the x-axis at each angle in [0, 4 ].
Then 3 copies of L(E) rotated by 4 , 2 , 3
4 along with the original L(E) form
the desired set of area 0 which contains a line in every direction.

Theorem 7. There exists Kakeya sets in the plane.


Sketch of Proof. Let F be the Besicovitch set shown to exist by the previous
theorem then consider an -cover {Ui } of F B(x, r) where x and r are chosen
such that F B(x, r) still contains a line segment in every direction. Then by
the fact that F B(x, r) is compact we can find a finite subcover {Ui }ki=1 . So
for a needle starting in one of these Ui we can move between each other Uj via
some combination of Pal joins, and so can travel reach any direction via a finite
number of Pal joins. Then the union of our cover with these Joins can have area
as small as we desire.

6.1 Cantor Dust


To see an example [17] of this proof in action we can consider quite a famous
fractal as constructed by Cantor, the Cantor Dust D. To construct this we start
with the unit square D0 centered at the origin, and at each iteration k replace
each remaining square of side length d with four squares side length 41 d arranged
as in the figure below.

19
Then at each iteration we have the property that both projections to the x
and y axis contain the whole unit interval [0, 1] and so by our study of duality
when we consider the line set L(D) (Where D is the limit of this iterating
procedure) we find that it contains all lines with slopes in [0, 1] so by taking 3
copies of L(D) rotated by 4 , 2 , 3
4 we find a set containing line segments in each
direction.
Theorem 8. [13] D is a 1-set.
Proof. At the k th stage of the constructionpDk is formed of 4k squares each of
side length 4k and sophave diameter 4k (2). Then by taking these as a -
cover of D for = 4kp (2) we can find an upper bound for H1 (D) by realising
that H1 (D) k
p 4 4k (2) and so by limiting k i.e. taking 0 we find
1
H (D) (2).
Then to find a lower bound we use the fact that projections dont increase
distances (Projections are a Lipschitz mapping i.e. |proj(x) proj(y)| |x y)
then although we havent proved it here this implies that dimH (proj(E))
dimH (E). So in our example as the projection of the cantor dust to either axis
is the interval [0, 1] we have that

1 = ([0, 1]) = H1 ([0, 1]) H1 (D)

Theorem 9. [13] D is irregular.


Proof. By considering the lines L1 = x2 and L2 = 2x then the projection of D
onto each of these has length 0. (By considering the first few iterations in the
above diagram it is quite easy to convince yourself of this.)
Then by Corollary 3 the set D is irregular.

20
So having shown that the Cantor dust is an irregular 1-set we can apply
the statement of theorem 5 to find that L(D) has 0 area and so satisfies the
definition of a Besicovitch set.

7 Further Problems
7.1 Kakeya Set Conjecture
An open problem in the study of Kakeya and Besicovitch sets is the following
Theorem 10 (Kakeya Set Conjecture). A Besicovitch set in Rn has Hausdorff
and Minkowski dimension n.
The Minkowski dimension is a similar concept to the Hausdorff dimension
sometimes referred to as the Box-Counting Dimension.
Definition 11. For a set E Rn suppose for > 0, E can be covered by N ()
n-dimensional hypercubes then we define

N ()
dimbox (E) = lim
0 log(1/)

So far the conjecture has been proved for n = 1 and n = 2 only.


The result for n = 1 is trivial and below I give the proof for n = 2 found in
Falconer.
Theorem 11. Any Besicovitch set in R2 has Hausdorff dimension 2.
Proof. [13] Let E be a set containing a line in every direction then define

F = {(a, b) : L(a, b) E}

Then proj 2 is the entire y-axis and as seen earlier projection cant increase
dimension and so we have that dimH (F ) 1. So by Theorem 5 (b) we have
that dimH (L(F )) = 2, then since L(F ) E we see that dimH (E) = 2

8 Bibliography
References
[1] W. A. Hurwitz. New questions. The American Mathematical Monthly,
1920.
[2] M. Fujiwara and S. Kakeya. On some problems of maxima and minima for
the curve of constant breadth and the in - revolvable curve of the equilateral
triangle. Thoku Mathematical Journal.
[3] J. Pl. Ein minimumproblem frovale. Mathematische Annalen, 1921.

[4] F Cunningham. The kakeya problem for simply connected and for star-
shaped sets. American Mathematical Monthly, 1971.

21
[5] A.S. Besicovitch. Sur deux questions dint egrabilit e des fonctions. J. Soc.
Phys.-Math, pages 105123, 1919.

[6] T. Tao. Poincars legacies: pages from year two of a mathematical blog.
Department of Mathematics, UCLA, page 1.22, 2009.
[7] A.S. Besicovitch. On kakeyas problem and a similar one. Math. Zeitschrift,
pages 312320, 1928.
[8] O. Perron. ber einen satz von besicovitch. Math. Zeitschrift, pages 383386,
1928.
[9] I.J. Schoenberg. On the besicovitch-perron solution of the kakeya problem.
In Studies in Mathematical Analysis-Essays in Honor of G. Polya, pages
359363, 1962.

[10] A.S. Besicovitch. The kakeya problem. The American Mathematical


Monthly, 1963.
[11] K Falconer. The Geometry Of Fractal Sets. Cambridge University Press,
1985.
[12] O J. D. Barrowclough. The Kakeya Problem. PhD thesis, Sintef, 2008.

[13] K falconer. Fractal Geometry. John Wiley And Sons Ltd, second edition,
2003.
[14] S. Har-Peled. Duality, 2008.
[15] M van Kreveld. Lectures in geometric algorithms, 2016.

[16] A.S. Besicovitch. On fundamental geometric properties of plane line-sets.


Journal of the London Mathematical Society, 1964.
[17] M Furtner. The Kakeya Problem. PhD thesis, Ludwig-Maximilians-
Universit at Mu nchen, 2008.

22

You might also like