THE BANACH-TARSKI PARADOX

TOM WESTON

1. Introduction
The Banach-Tarski paradox is one of the most celebrated paradoxes in mathematics. It states that given any
two subsets A and B of R3 , which are bounded and have non-empty interior, it is possible to cut A into a finite
number of pieces which can be moved by rigid motions (translations and rotations) to form exactly B. This has
many amusing consequences; it is most commonly stated as the assertion that a pea can be cut into a finite number
of pieces which could then be reassembled to form the sun.
Intuitively, of course, this is absurd. Rigid motions are supposed to preserve volume. The Banach-Tarski paradox
seems to completely contradict this. This is why it is called a paradox and not a theorem. This is also why many
people have tried to argue that it must be false. Since the logic of the proof is unquestioned, they argue that one of
the underlying assumptions must be flawed.
The assumption that they pick is called the axiom of choice. The axiom of choice is the seemingly innocent
statement that given any collection of sets it is possible to pick one element from each of the sets. (Actually, the
collection of sets must itself form a set; however, we do not want to delve too deeply into axiomatic set theory here.)
Here, however, our intuition seems to cry out that this statement is obviously true. How then are we to reconcile
these two conflicting statements?
I have two main goals in writing this paper. The first is to show that the proof of the Banach-Tarski paradox is
not difficult. The argument is fairly long and intricate, but never requires anything very deep. In fact, the key step
is a simple observation about free groups which is stated and proved in Section 2. The rest of the proof is really just
fixing up the little things that do not quite fit. My proof basically follows that given in [3, Chapters 1-3], although
I have dispensed with the terminology he introduces. (Until Section 7, when a bit of it is indispensible.) I have also
tried to keep things as explicit as possible, to the point where we are eventually writing out decompositions of the
unit ball into 52 pieces.
My second goal is to try to show that the Banach-Tarski paradox is not really a paradox. I am aware of several
fairly convincing arguments for this. I will save most of them for the final section, but since in Sections 5 and 6 we
will need a variant on one, I will present it here. It is an example of a similar paradox in R2 , which exists without
any use of the axiom of choice.
So, let S1 be the unit circle in R2 . Let ` be the segment (0, 1) along the x-axis; ` is simply one of the radii of
1
1
S , without the center. Next, let be a counterclockwise rotation by, say, 10
radians around the origin. So, (`) is
1
n
another radius of S , and since 2 is irrational we see that (`) never comes back to coincide with `. So, define a
set C by

]
C=
= rhon (`).
n=0

(Here, and throughout the paper, the symbol ] stands for disjoint union. We are really using it is as a convenient
notation device; A ] B means that we are considering the set A B and that A and B are disjoint.)
Now, consider the set S1 ] C; it looks like some sort of wheel with an infinite number of spokes. (See Figure 2.)
1
Now, consider what happens when we rotate C by 1 ; that is, clockwise by 10
radians. Since n (`) 6= ` for any n,
we see that
n1 (`) 6= 1 (`)
for any n. Thus, 1 (`) does not intersect the set S1 ] C. That is,
S1 ] 1 (C) = S1 ] C ] 1 (`).

Figure 1. The Sphere S2

1

TOM WESTON

Figure 2. The Circle Trick

So we have cut S1 ] C into two pieces, moved one of them, and obtained the original set plus an extra line.
This is certainly not nearly as striking as the Banach-Tarski paradox, but it does illustrate that geometric paradoxes
can happen even in simple situations. The Banach-Tarski paradox obtains its additional power from the extra
freedom that we get by working in 3 dimensions.
The plan of the paper is as follows: We will begin in Section 2 with a quick review of free groups, and we will
then proceed to give the fundamental paradoxical decomposition of the free group of rank 2. In Section 3 we
will construct a free group of rank 2 in the group SO3 of rotations of the sphere S2 . (Note that S2 is simply the
hollow sphere, not the solid ball B3 .) In Section 4 we will use the natural group action of SO3 on S2 to extend this
paradox to most of the sphere; this intermediate result is called the Hausdorff Paradox. In Section 5 we will use a
3-dimensional version of the circle trick above to extend our paradox to the entire sphere S2 . Then in Section 6 we
will extend our paradox to the solid ball B3 , which will again require a variant on the circle trick. Finally, in Section
7 we will use a trick due to Banach to extend our paradox to arbitrary bounded subsets of R3 with interior points.
In Section 8 we will return to the underlying philosophical issues behind the Banach-Tarski paradox.
I have tried to keep the prerequisites to a minimum. However, rather than attempt a half-assed introduction to
certain subjects myself, I felt that I would refer the reader to Artins excellent book [1]. In particular, the reader will
need to be familiar with group theory and linear transformations at the level of Chapters 1-3 and 4.1, 4.2. Familiarity
with Chapters 5 and 6 is helpful, but not necessary, except for Section 6.7, which is essential. Lastly, some knowledge
of the cardinality of sets (countable and uncountable sets) will be needed. The contents of Chapter 2 of [2] are more
than sufficient.
2. Free Groups
Let F(x, y) be the free group on generators. Recall that F(x, y) consists of all possible strings in x, y, x1 , y 1
(called words) subject only to the relations
(1)

xx1 = x1 x = yy 1 = y 1 y = 1,

with group operation given simply by concatenation of words. Further recall that every word is equivalent (under
the relations (1)) to a unique reduced word, in which all possible pairs of inverses have been canceled. We will tend
to identify elements of F(x, y) with reduced words, rather than worry about equivalence classes. Despite its simple
definition, the group F(x, y) is extremely complicated, and the main step in the proof of the Banach-Tarski paradox
will be an application of an easy observation about F(x, y), which we will now explain.
Define the set W(x) to consist of all of the elements of F(x, y) which begin with x. (Remember that we are
considering only reduced words here.) We can define W(x1 ), W(y) and W(y 1 ) in the same way. Together with
{1} this gives a partition of F(x, y); that is,
(2)

F(x, y) = {1} ] W(x) ] W(x1 ) ] W(y) ] W(y 1 ).

Now consider what happens when we multiply W(x) by x1 on the left; that is, we set
x1 W(x) = {x1 u F(x, y) | u W(x)}.
Since u W(x), it begins with an x, which we can then cancel with the x1 . We claim that x1 W(x) will consist
of 1 and all words beginning with x, y or y 1 .
Proposition 2.1.
x1 W(x) = {1} ] W(x) ] W(y) ] W(y 1 ).
Proof. Take v W(x). Then v is a reduced word beginning with x, so it could not possibly begin xx1 . Therefore
x1 v can not begin with x1 , so x1 W(x) is disjoint from W(x1 ). By (2) this shows that
x1 W(x) {1} ] W(x) ] W(y) ] W(y 1 ).
Now take v {1} ] W(x) ] W(y) ] W(y 1 ); for example, suppose v W(y). Then v begins with y, so xv W(x),
as the x has nothing to cancel with. Therefore v = x1 xv x1 W(x). So,
x1 W(x) W(y).

THE BANACH-TARSKI PARADOX

Figure 3. The Rotations and

In exactly the same way we get that {1} x1 W(x), W(x) x1 W(x) and W(y 1 ) x1 W(x). Thus,
x1 W(x) {1} ] W(x) ] W(y) ] W(y 1 ),
which completes the proof.

Note that an analogous proof also establishes a similar result with x replaced by x1 , y or y 1 . In particular we
will need to use
y 1 W(y) = {1} ] W(x) ] W(x1 ) ] W(y).
We can now state the result that will be the main step in the Banach-Tarski paradox.
Corollary 2.2.
F(x, y) = x1 W(x) ] W(x1 )
and
F(x, y) = y 1 W(y) ] W(y 1 ).
Proof. This is immediate from (2) and Proposition 2.1.

Essentially, we have divided F(x, y) into five pieces, and then moved two of these pieces to reassemble two copies
of F(x, y). The connection with the Banach-Tarski theorem should be clear; we would like to be able to use this
construction on the group of rigid motions of a sphere to cut the sphere into five pieces, which we can then shift and
reassemble into two copies of the sphere. This is the main step in the Banach-Tarski theorem, although it will not
quite work and there will be some cleaning up to do.
We will need one last piece of information about F(x, y).
Proposition 2.3. F(x, y) is countable.
Proof. Let Fn be the subset of F(x, y) of (reduced) words of length n, where the length of a word is the number of
symbols in it. Then

]
F(x, y) =
Fn .
n=0

Each Fn is certainly finite however; in fact, it can have at most 4n elements. (It will actually have fewer elements than
that for n 2, because of cancellation.) Thus F(x, y) is a countable union of finite sets, and thus is countable. 
3. Independent Rotations
In order to apply the preceding construction to the sphere S2 , we first must find a free group of rank 2 in SO3 , the
group of rotations of S2 . Recall that SO3 can be realized as the group of orthogonal 3 3 matrices of determinant
1; that is,
SO3 = {A GL3 (R) | At A = I, det A = 1}
where GL3 (R) is the general linear group of invertible 3 3 matrices, I is the 3 3 identity matrix and At denotes
the transpose of A.
Also recall that SO3 has a natural group action on S2 : for any SO3 and p S2 , (p) is the image of p after
rotation by . More concretely, if is represented by a matrix A (with respect to the standard basis of R3 ) and
p = (x, y, z), then

x
(p) = A y .
z
We will now give two rotations which satisfy no relations; that is, they generate a free group of rank 2 in SO3 . Let
be a counterclockwise rotation by arccos 31 around the x-axis and let be a counterclockwise rotation by arccos 31

TOM WESTON

around the z-axis. (See Figure 3.) It is easy to see that 1 and 1 are represented by the following matrices (with
respect to the usual basis of R3 , of course) :

1
0
0
3
0
0
1

1
232 = 0
1
2 2
1 = 0
3
3
2 2
1
0
2
2
1
0 3
3

1
232 0
1
2 2 0
3
1

1
1 = 2 2
1
0 .
0 = 3 2 2
3
3
0
0
3
0
0
1
Proposition 3.1. and are independent as elements of SO3 ; that is, they generate a free group of rank 2 inside
of SO3 .
Proof. We first must explain what exactly we are proving. When we speak of a word in and , we will mean
some word in the free group F(x, y), with x replaced by and y replaced by . We have to show that the only one
of these reduced word which evaluates to 1 in the group SO3 is the empty word. We will do this by showing that
(0, 1, 0) 6= (0, 1, 0) for any word in and .
First, we observe that if has length n, then (0, 1, 0) has the form

1 
2,
b,
c
2
a
3n
for integers a, b, c. We prove this by induction on n. If n = 0, then = 1 and (0, 1, 0) = (0, 1, 0), which has the
correct form. So suppose that n > 1. Let be a reduced word of length n. Then we can write in one of the forms
0 , 1 0 , 0 , 1 0
for some reduced word 0 of length n 1. Then by the induction hypothesis we can write

1 
0 (0, 1, 0) = n1 a 2, b, c 2
3
for some integers a, b, c. Therefore (0, 1, 0) is one of



1 
0
(0, 1, 0) =
a 2, b, c 2
3n1

3
0
0
a 2
1
b
1
2 2
= n 0
3
0 2 2
1
c 2

1 
= n 3a 2, b 4c, (2b + c) 2
3 


1 
1 0
(0, 1, 0) = 1
a
2,
b,
c
2
3n1

30
0
a 2

1
b
0
1
2 2
= n
3
0 2 2
1
c 2

1 
= n 3a 2, b + 4c, (2b + c) 2
3


1 
0
(0, 1, 0) =
a
2,
b,
c
2
3n1

1
2 2 0
a 2
1
b
= n
2 2
1
0
3
0
0
3
c 2


1 
= n (a 2b) 2, 4a + b, 3c 2
3

THE BANACH-TARSKI PARADOX

1 0

(0, 1, 0) =




a 2, b, c 2
3n1

1 2 2 0
a 2
2 2
b
1
0
0
0
3
c 2



(a + 2b) 2, 4a + b, 3c 2

1
3n

1
3n
which all have the required form. This completes the induction.
Now, suppose that we have a non-identity reduced word in and of length n with (0, 1, 0) = (0, 1, 0). Writing
(0, 1, 0) in the form above, we see that we must have a = c = 0 and b = 3n . In particular, since 6= 1, we must
have n 1, so that
a b c 0 (mod 3).
=

We will show that this is not possible.

To do this, let us define, for any F(, ), N () to be the triple (a, b, c) above, considered modulo 3. This
makes sense since a, b, c are all integers. We must show that for any non-identity reduced word in and we
always have N () 6= (0, 0, 0).
Now, if N () = (a, b, c), then our above computations show that
N ()
= (3a, b 4c, 2b + c)
N (1 ) = (3a, b + 4c, 2b + c)
N ()
= (a 2b, 4a + b, 3c)
N ( 1 ) = (a + 2b, b 4a, 3c)

=
=
=
=

(0, b c, c b)
(0, b + c, b + c)
(a + b, a + b, 0)
(a b, b a, 0).

Iterating these, it is easy to see that for n > 0,

(
(0, b c, c b), n odd;
N ( ) =
(0, c b, b c), n even;
(
(0, b + c, b + c)
n odd;
N (n ) =
(0, b c, b c), n even;
(
(a + b, a + b, 0)
n odd;
n
N ( ) =
(a b, a b, 0) n even;
(
(a b, b a, 0) n odd;
N ( n ) =
(b a, a b, 0) n even.
n

(For example, N (2 ) = N (()), where N () = (0, b c, c b). Thus

N (2 ) = (0, (b c) (c b), (c b) (b c)) = (0, 2b 2c, 2c 2b) = (0, c b, b c).
Similarly, N (3 ) = (0, b c, c b), and it alternates from there. The other checks are done in the same way.)
So, let be a non-identity reduced word in and . Then alternates non-zero powers of and . Suppose first
that ends (on the right) with a power of , so that it has the form
n 3 n 2 n1
where the ni are non-zero. Then N (n1 ) = N (n1 1), where N (1) = (0, 1, 0). Thus, plugging into our above
formulas,
N (n1 ) {(0, 1, 2), (0, 2, 1), (0, 1, 1), (0, 2, 2)}.

(3)
Again using our above formulas,

N ( n2 n1 ) {(1, 1, 0), (2, 2, 0), (2, 1, 0), (1, 2, 0)}.

(We just have to plug the 4 values in (3) into the equations for N ( n ). Although there are 16 different possibilities
to check, we only get these 4 triples.) Continuing, we find that
N (n3 n2 n1 ) {(0, 1, 2), (0, 2, 1), (0, 1, 1), (0, 2, 2)}

TOM WESTON

again. Pictorially,
n1

n2

{(0, 1, 0)} {(0, 1, 2), (0, 2, 1), (0, 1, 1), (0, 2, 2)} {(1, 1, 0), (2, 2, 0), (2, 1, 0), (1, 2, 0)}
n3

{(0, 1, 2), (0, 2, 1), (0, 1, 1), (0, 2, 2)}

We thus see that we could not have N () = (0, 0, 0). Thus (0, 1, 0) 6= (0, 1, 0), so is not the identity rotation.
Similarly, if starts with a power of , we get
n1

n2

{(0, 1, 0)} {(1, 1, 0), (2, 2, 0), (2, 1, 0), (1, 2, 0)} {(0, 1, 2), (0, 2, 1), (0, 1, 1), (0, 2, 2)}
n3

{(1, 1, 0), (2, 2, 0), (2, 1, 0), (1, 2, 0)}

So again N () 6= (0, 0, 0), so (0, 1, 0) 6= (0, 1, 0). This completes the proof.

4. The Hausdorff Paradox

Now that we have realized the free group on 2 generators as a subgroup of SO3 , namely h, i = F(, ), we can
attempt to pass our paradoxical construction of Section 2 to S2 . This is accomplished through the natural action
of SO3 on S2 discussed in Section 3. This SO3 -action induces an equivalence relation on S2 , where p, q S2 are
equivalent if there is a SO3 with (p) = q. The equivalence classes are called the SO3 -orbits; if p S2 , its orbit
is simply the set
{(p) | SO3 }.
Of course, the orbits of the SO3 -action are not very interesting; for any points p, q S2 there are lots of rotations in
SO3 sending p to q. Therefore all points in SO3 are equivalent, and S2 itself is the only orbit.
The orbits get considerably more interesting if we look only at the F(, )-action. In fact, since F(, ) is countable
and the orbit of a point p S2 is just
{(p) | F(, )},
the orbits must also be countable. Since S2 is uncountable, there must be uncountably many of these F(, )-orbits.
(This is because a countable union of countable sets is still countable.)
We now need to invoke the axiom of choice. We want to pick exactly one representative of each F(, )-orbit.
This is precisely what the axiom of choice says that we can do. So let M0 be the set of all of these. We will now try
to use this set M0 to pass our paradox from F(, ) to S2 . This is not quite going to work, for reasons that we will
see in a moment, but with a slight modification we will obtain the key step in the Hausdorff paradox.
Our plan is as follows: We have our partition
F(, ) = 1 ] W() ] W(1 ) ] W() ] W( 1 )

(4)

of F(, ). Now, for any F(, ), we can form the set

(M0 ) = {(p) | p M0 }.
Then,
F(, )M0 = {(p) | F(, ), p M0 }
will be all of S , since M0 contains an element of each F(, )-orbit. Therefore, applying (4) to M0 , we obtain
2

S2 = F(, )M0 = M0 W()M0 W(1 )M0 W()M0 W( 1 )M0 .

Unfortunately, these unions need no longer be disjoint. For example, if by some chance (1, 0, 0) M0 then, since
(1, 0, 0) = (1, 0, 0), (1, 0, 0) will be both in M0 and W()M0 . More generally, if p is a fixed point of any F(, ),
then p could lie in two of the sets of the decomposition.
We will remedy this in the most naive possible way: if fixed points of rotations in F(, ) are causing problems,
we will simply get rid of them. So let D be the set of all points of S2 fixed by any non-identity F(, ). That is,
D = {p S2 | There is a F(, ), 6= 1, with (p) = p.}.
Since a non-identity rotation fixes exactly its axis of rotation, D will consist exactly of the points of S2 which lie
on axes of rotations in F(, ). In particular, since each non-identity element of F(, ) will contribute only 2 such
points, D is countable.
Now, we would like to consider our F(, )-action on S2 D. We first must check that we even have a well-defined
action here. That is, we must check that if p S2 D, then (p) S2 D for all F(, ). We will prove the
contrapositive. So suppose we have a point p S2 and a rotation F(, ) with (p) D. Then by the definition

THE BANACH-TARSKI PARADOX

of D there is a non-identity F(, ) with ((p)) = (p). But then 1 is also a non-identity rotation, and
1 (p) = p. So p D.
Now, as before, by the axiom of choice we can choose a representative of each F(, )-orbit on S2 D. Let M be
the set of these representatives. This time our partition of F(, ) will give us a partition of S2 D. In fact, suppose
1 , 2 F(, ) and 1 (M) 2 (M) 6= . Thus there is a point p with p 1 (M) and p 2 (M). Then there are
points p1 , p2 M with p = 1 (p1 ) and p = 2 (p2 ). But then
1
1
1 2 (p2 ) = 1 (p)
= p1

so p1 and p2 lie in the same orbit. Since M contains exactly one representative of this equivalence class, we must
have p1 = p2 . But then p1 is a fixed point of 1
/ D, we must have 1
1 2 . Since p1
1 2 = 1, so 1 = 2 . So we have
shown that if 1 6= 2 , then 1 (M) and 2 (M) are disjoint.
From this it follows that our partition (4) gives rise to a partition of S2 D:
(5)

S2 D = F(, )M = M ] W()M ] W(1 )M ] W()M ] W( 1 )M.

We are now in a position to apply Corollary 2.2. By our disjointness above, we also get disjoint unions
S2 D = 1 (W()M) ] W(1 )M

(6)
and

S2 D = 1 (W()M) ] W( 1 )M.

(7)

Let us write this more suggestively. Define four subsets of S2 as follows :

W
W1
W
W1

=
=
=
=

W()M
W(1 )M
W()M
W( 1 )M.

Then (5) tells us that

S2 D = M ] W ] W1 ] W ] W1
and (6) and (7) tell us that we can rotate W and W to obtain
S2 D = 1 W ] W1
and
S2 D = 1 W ] W1 .
This is known as the Hausdorff paradox.
Theorem 4.1 (Hausdorff Paradox). There is a countable set D S2 such that S2 D can be divided into 5 pieces
which can be rotated to form 2 copies of S2 D.
Let us pause to consider the sets we have constructed. D is an explicitly given countable set: it is simply the points
on the axes of a known set of rotations F(, ). The other 5 sets, M, W , W1 , W , W1 , are rather mysterious.
They are certain uncountable sets which only exist due to the axiom of choice, and there is very little more that we
can say about them.
5. The Banach-Tarski Paradox on S2
We must now use our paradoxical decomposition of S2 D to construct one on all of S2 . First we make three
simple observations about disjoint unions.
(1) If a set A is a disjoint union of subsets Ai ,
]
(8)
A=
Ai ,
i

and if each Ai in turn is a disjoint union of subsets Aij ,

]
Ai =
Aij ,
j

TOM WESTON

Figure 4. The rotations `

then A is the disjoint union of the sets Aij :

A=

Aij .

i,j

(2) If a set A S2 is a disjoint union of subsets Ai , as in (8), and if SO3 is any rotation, then (A) is the
disjoint union of the sets (Ai ):
]
(A) =
(Ai ).
i

(3) If a set A S2 is a disjoint union of subsets Ai , as in (8), and if B S2 is some other set, then A B is the
disjoint union of the (possibly empty) sets Ai B:
]
A B = (Ai B).
i

These three facts will be enough to insure that all of the operations we will perform below preserve disjoint unions.
Since there will be enough else going on to keep us busy, we will not comment on these issues again.
We would now like to find a three-dimensional analogue of the circle construction in the introduction. We must
find a rotation SO3 which we can somehow use to erase the set D. To begin, select a line ` through the origin
which does not intersect D; this is certainly possible since D is countable and there are uncountably many lines
through the origin in R3 . Define ` SO3 to be the counterclockwise rotation by radians around the line `. (See
Figure 4.) We would like to be one of these rotations, with the additional property that no matter how many times
we apply it to a point in D we never again land in D. So, define a set T by
T = { [0, 2) | There is a p D and a positive integer n with `n (p) D}.
Now, for each of the countably many pairs of points (p, q) D D we get at most countably many angles to put in
T , so T will be countable. So, since [0, 2) is uncountable we may choose some 0
/ T . Then set
= `0 .
By its construction we have that
n (D) D =
for all positive integers n. In fact, applying m to this, we find that
m+n (D) m (D) = ,
and then replacing n by n m we have
n (D) m (D) =
for all integers m, n with m 6= n. Now, define
E=

n (D) S2 .

n=0

Then D ( E.
Next, consider the disjoint union
S2 = (S2 E) ] E.

THE BANACH-TARSKI PARADOX

Since
E =

n (D)

n=0

=
=
=

n=0

]
n=0

n (D)
n+1 (D)
n (D)

n=1

= E D,
we have
S2 D = (S2 E) ] (E D) = (S2 E) ] (E).
That is, rotating E by and leaving the rest of S2 fixed we have erased D and obtained S2 D. Looking at this
the other way, starting with a decomposition of S2 D we will now be able to obtain one of S2 . We will now carry
this out with our decompositions (5), (6) and (7) of S2 D.
First we make two simple observations that will make our manipulations a lot easier to follow. Since D ( E, we
have S2 E ( S2 D, so
S2 E = (S2 D) (S2 E).
Also,
E = 1 (E)
= 1 (E D)


= 1 (S2 D) E .
Thus, we have
S2 = (S2 E) ] E




= (S2 D) (S2 E) ] 1 (S2 D) E .
We now want to plug in our three decompositions of S2 D. So,




S2 = (M ] W ] W1 ] W ] W1 ) (S2 E) ] 1 (M ] W ] W1 ] W ] W1 ) E

 

= (1 W ] W1 ) (S2 E) ] (1 W ] W1 ) E

 

= ( 1 W ] W1 ) (S2 E) ] ( 1 W ] W1 ) E .
This suggests that we should define
M0
W0
W0 1

= M (S2 E)
= W (S2 E)
= W1 (S2 E)

M1
W1
W1 1

= ME
= W E
= W1 E

W0
W0 1

= W (S2 E)
= W1 (S2 E)

W1
W1 1

= W E
= W1 E.

Then,
S2 = (M0 ] W0 ] W0 1 ] W0 ] W0 1 ) ] 1 (M1 ] W1 ] W1 1 ] W1 ] W1 1 )
= M0 ] 1 M1 ] W0 ] 1 W1 ] W0 1 ] 1 W1 1 ] W0 ] 1 W1 ] W0 1 ] 1 W1 1 .

10

TOM WESTON

In fact, in order to also use the decompositions (6) and (7), we further break up W and W as follows:
W00
W01
W10
W11
W00
W01
W10
W11

W0 (S2 E)
W0 (E)
W1 (S2 E)
W1 (E)
W0 (S2 E)
W0 (E)
W1 (S2 E)
W1 (E)

=
=
=
=
=
=
=
=

=
=
=
=
=
=
=
=

W (S2 E) (S2 E)
W (S2 E) (E)
W E (S2 E)
W E (E)
W (S2 E) (S2 E)
W (S2 E) (E)
W E (S2 E)
W E (E)

Now,
S2 = M0 ] 1 M1 ] W00 ] W01 ] 1 W10 ] 1 W11 ] W0 1 ] 1 W1 1 ]
W00 ] W01 ] 1 W10 ] 1 W11 ] W0 1 ] 1 W1 1 .
Also,




S2 = (1 W ] W1 ) (S2 E) ] 1 (1 W ] W1 ) E







= 1 W (S2 E) ] W1 (S2 E) ] 1 (1 W E) ] (W1 E)








= 1 W (S2 E) ] W1 (S2 E) ] 1 1 (W (E)) ] W1 E
h
i
h
i
= 1 (W00 ] W10 ) ] W0 1 ] 1 1 (W01 ] W11 ) ] W1 1
= 1 W00 ] 1 W10 ] W0 1 ] 1 1 W01 ] 1 1 W11 ] 1 W1 1 .
Similarly,
S2 = 1 W00 ] 1 W10 ] W0 1 ] 1 1 W01 ] 1 1 W11 ] 1 W1 1 .
We will rewrite this one last time. Define
M0
W00
W01
W0 1

A1
A3
A5
A7

=
=
=
=

A9
A11
A13

= W00
= W01
= W0 1

1 M1
1 W10
1 W11
1 W1 1

A2
A4
A6
A8

=
=
=
=

A10
A12
A14

= 1 W10
= 1 W11
= 1 W1 1 .

Then,
S2 = A1 ] A2 ] ] A14
S2 = 1 A3 ] 1 A4 ] 1 1 A5 ] 1 1 A6 ] A7 ] A8
S2 = 1 A9 ] 1 A10 ] 1 1 A11 ] 1 1 A12 ] A13 ] A14 .
This, finally, proves the Banach-Tarski paradox for S2 .
Theorem 5.1 (The Banach-Tarski Paradox, Version 1). The sphere S2 can be partitioned into a finite number of
pieces which can be rotated to form two copies of S2 .
6. The Banach Tarski Paradox on B3
The next step, then, is to extend this construction to the solid ball B3 . Note that this means that we no longer
must work exclusively with rotations in SO3 ; we can also work with some rotations around axes not containing the

THE BANACH-TARSKI PARADOX

11

Figure 5. The Rotation

origin. To avoid getting completely lost in notation, we define

3
5
7
9
11
13

=
=
=
=
=
=

1
1 1
1
1
1 1
1

4
6
8
10
12
14

1
1 1
1
1
1 1
1.

=
=
=
=
=
=

Thus, we can now write

S2 =

14
]

Ai =

i=1

8
]

i A i =

i=3

14
]

i A i .

i=9

We begin our extension in the obvious way: Define the sets

Bi B3 , 1 i 14,
by extending each Ai radially inward. That is, using spherical coordinates,
Bi = {(r, , ) | (1, , ) Ai , 0 < r 1}.
This gives us partitions of all of B3 except for the origin 0:
B3 0 =

14
]

Bi =

i=1

8
]

i Bi =

i=3

14
]

i Bi .

i=9

We must once again find a way to deal with this little complication. Luckily, essentially the same trick that we used
to deal with D will work.
There are many ways to do this, and all of our choices here are pretty much arbitrary. First, choose a point near
1
, 0, 0). Next take the line parallel to the y-axis through this point. Let be a rotation around this line
0, say ( 10
such that n (0) 6= 0 for all n > 0. (If we are willing to use the fact that is irrational, then an angle of 1 radian will

do. Otherwise, we can still use some angle like 2 , since no multiple of 2 can give 360.) Note that n (0) B3
for all n.
Now, since n (0) 6= 0 for all n, as before we find that
n (0) 6= m (0)
for all m 6= n. So define
F=

n (0).

n=0

Then (F) = F 0. So, since

B3 = (B3 F) ] F
we have
B3 0 = (B3 F) ] (F).
Again, as before we find that




B3 = (B3 0) (B3 F) ] 1 (B3 0) F .

12

TOM WESTON

Thus,
3

B =

" 14
]

Bi (B F) ]

8
]

i Bi (B F) ]

"

i=3

" 14
]

Bi F

i=1

i=1

"

" 14
]

8
]

i Bi F

i=3

i Bi (B F) ]

i=9

" 14
]

i Bi F .

i=9

So, finally, as before we define

Bi0 = Ai (B3 F)
Bi1 = Ai F
for i = 1, 2, and
3
Bi00 = Bi (B3 F) 1
i (B F)

Bi01 = Bi (B3 F) 1
i (F)
3
Bi10 = Bi F 1
i (B F)

Bi11 = Bi F 1
i (F)
for i = 3, . . . , 14. (Many of these may be empty, but it does not matter.)
Thus,
S2 = B10 ] 1 B11 ] B20 ] 1 B21 ]

14
]

Bi00 ]

i=3

8
]

i Bi00 ]

8
]

i Bi10 ]

8
]

1 i Bi01 ]
1 i Bi01 ]

i=3

i=3

i=3

14
]

14
]

14
]

i=9

i Bi00 ]

i=9

i Bi10 ]

14
]

Bi01 ]

14
]

i=3

i=3

8
]

1 i Bi11

14
]

1 i Bi11 .

1 Bi10 ]

14
]

1 Bi11

i=3

i=3

i=9

i=9
3

This, then, is our final paradoxical decomposition of B . Of course, there is nothing special about the ball of radius
1 centered at the origin; using rotations around axes not containing the origin we can obtain a similar decomposition
of any solid ball in R3 .
Theorem 6.1 (The Banach-Tarski Paradox, Version 2). Any solid ball B in R3 can be cut into finitely many pieces
which can be rotated to form 2 copies of B.
7. The Generalized Banach-Tarski Paradox
To get the final version of the Banach-Tarski paradox we will have to consider subsets of R3 much more general
than B3 . It will thus no longer be at all practical to carry around the sort of explicit decompositions that we have
been. We will therefore have to formalize our notions somewhat. First, since we will need translations as well as
rotations to do our decompositions, let M3 be the group of translations and rotations in R3 . (For an analysis of this
group, see [1, Chapter 5]. We will not need any deep facts about it.)
Definition 1. Let A and B be two subsets of R3 . We will say that A and B are equidecomposable, written A B,
if there exist partitions of A and B into the same number of pieces,
(9)

A=

n
]

Ai ,

B=

i=1

i=1

and gi M3 such that

gi Ai = Bi
for all i.

n
]

Bi

THE BANACH-TARSKI PARADOX

13

Roughly, A and B are equidecomposable if we can cut A into finitely many pieces which we can then rearrange
by rigid motions to form exactly B. Let us restate the Banach-Tarski paradox in this language.
Corollary 7.1. Let B be a ball in R3 of radius r. Then there is a subset B0 of B such that BB0 is equidecomposable
with a disjoint union of two balls of radius r.

Proof. Simply take B0 = B10 ] 1 B11 ] B20 ] 1 B21 . This is then precisely what we proved in Section 6.

We now check that the relation of equidecomposability is an equivalence relation.

(1) Reflexivity: Take A1 = A, g1 = 1. Then g1 A1 = A1 , so A A.
(2) Symmetry: Suppose A B. Then we can write A and B in the form (9) with gi M3 such that gi Ai = Bi
for all i. Then Ai = gi1 Bi , so B A.
(3) Transitivity: Suppose A B and B C. Then we can write

A=

n
]

Ai ,

B=

i=1

n
]

Bi =

i=1

n
]

Bj0 ,

C=

j=1

j=1

with gi , gj0 M3 satisfying gi Ai = Bi , gj0 Bj0 = Cj for all i and j.

Set Aij = gi1 (Bi Bj0 ). Then

Aij =

i,j

n
n ]
]

Aij

i=1 j=1
0

n ]
n
]

gi1 (Bi Bj0 )

i=1 j=1

n
]

i=1

=
=
=

n
]

i=1
n
]
i=1
n
]

i=1

=A

gi1 Bi

n
]

j=1

gi1 (Bi B)
gi1 (Bi )
Ai

n
]

Bj

Cj

14

TOM WESTON

so the Aij partition A. Further,

]

gj0 gi Aij =

i,j

gj0 gi gi1 (Bi Bj0 )

i,j

n ]
n
]

gj0 (Bi Bj0 )

j=1 i=1
0

n
]

gj0

Bj0

n
]

j=1

Bi

i=1

n
]

gj0 (Bj0 B)

j=1
0

n
]

gj0 (Bj )

j=1
0

n
]

Cj

j=1

= C.
So A C.
In fact, the relation has two other important properties.
(4) If A B, then there is a bijection : A B such that C (C) for any C A.
Proof. Since A B we have decompositions as in (9) together with gi M3 satisfying gi Ai = Bi . Now, for
any p A, let Aip be the subset in which p lies. Then define (p) = gip (p). will be a bijection since the
gi are all bijections. Further, if C A, write
n
]
C=
(Ai C).
i=1

Then by the definition of we have

n
]

(C) =

gi (Ai C),

i=1

so C (C).

(5) If A B = A0 B 0 = , and if A A0 and B B 0 , then A ] B A0 ] B 0 .

Proof. Write
A=

n
]

Ai ,

A =

i=1

n
]

A0i ,

B=

i=1

n
]

Bj ,

j=1

B =

n
]

Bj0 ,

j=1

with gi , gi0 M3 satisfying gi Ai = A0i , gj0 Bj = Bj0 . Then,

A]B =

n
]

i=1

Ai ]

A ]B =

n
]

i=1

Bj

j=1

and
0

n
]

gi Ai ]

n
]

gj0 Bj

j=1

so A ] B A0 ] B 0 .
We can also use to define another relation on subsets of R3 .
Definition 2. Let A and B be subsets of R3 . We write A 4 B if there is a B 0 B with A B0 .

THE BANACH-TARSKI PARADOX

15

That is, A 4 B if A is equidecomposable with a subset of B. It is easy to see that 4 is reflexive and transitive.
Also, if A B and B 4 C, then A 4 C. Similarly, if A 4 B and B C, then A 4 C.
We are now ready to begin the proof of the final version of the Banach-Tarski paradox. First let us consider what
properties of subsets of R3 are preserved by equidecomposability. The Banach-Tarski paradox shows that volume,
whatever precisely that means, is not so preserved. However, there are at least two properties which are. First,
if a set A has non-empty interior (that is, if the A contains any solid ball of any non-zero radius), then any set
equidecomposable with A must also have non-empty interior. (This is because we are only cutting A into finitely
many pieces; one of these will still have to contain some, possibly smaller, ball.) Second, if A is bounded (that is, if
A is contained in some ball of finite radius), then any set equidecomposable with A will also be bounded. (This is
because the finitely many gi M3 can all only move A by a finite distance.) The generalized Banach-Tarski paradox
states that any two sets having these properties are equidecomposable. We will now prove this using properties of
4.
Proposition 7.2. Let A and B be bounded subsets of R3 with non-empty interior. Then A 4 B.
Proof. By our hypotheses, we can choose solid balls BA and BB with A BA and BB B. Now, choose n large
enough so that BA can be covered by n (overlapping) copies of BB . Let S R3 be a disjoint union of n copies of
BB . Then, by the definition of n we can move pieces of S to cover BA . Taking appropriate subsets of these pieces,
we can then exactly cover BA . Thus, BA 4 S.
Next, by the Banach-Tarski paradox we can repeatedly duplicate a subset of BB to form as many copies of BB as
we want. In particular, we can form n copies of BB , so S 4 BB . Thus,
A BA 4 S 4 BB B,
so A 4 B.

Of course, we can swap A and B above, so B 4 A. Thus, the last step is to show that A 4 B and B 4 A implies
that A B.
Proposition 7.3 (The Banach-Schr
oder-Bernstein Theorem). Suppose A 4 B and B 4 A. Then A B.
Proof. By our hypotheses, we can choose A0 A and B 0 B such that A B 0 and A0 B. Let : A B 0 and
: A0 B be the bijections of Property 4 of above. Now, set C0 = A A0 and for n 1 set
Cn = 1 (Cn1 ) A.
Let C be the (not necessarily disjoint) union of the Cn :
C=

Cn A.

n=0

Now, since C0 C and C0 A0 = , we have

A C = A0 (C A0 ) = A0

Cn .

n=1

Thus, since and are bijections,

(A C) =

Cn

n=1
0

= (A )

n=1

=B
=B

n=1

[
n=0

= B (C).

Cn1
Cn

(Cn1 )

16

TOM WESTON

So by Property 4,
A C B (C).
Also by Property 4,
C (C).
Thus, by Property 5,
(A C) ] C (B (C)) ] (C)
A B.

Corollary 7.4 (The Banach-Tarski Paradox, Version 3). Any two bounded subsets of R3 with interior points are
equidecomposable.
8. Final Remarks
I will conclude with a few remarks on the alleged paradoxical nature of the Banach-Tarski paradox. I would argue
that this result is in no way a paradox. It does go against our intuition, but very often our intuition is flawed; that
is why mathematics requires rigor.
The reason that the Banach-Tarski paradox is so disturbing to people, of course, is that it assaults our intuition
about 3-dimensional geometry, a subject in which we have a great deal of first-hand experience. The corresponding
paradox for free groups, while somewhat surprising, is certainly not something that anyone would argue with.
However, as I hope the reader was able to determine from the above proof before it completely degenerated into
a mess of strange sets, the result on free groups is the key step in the proof of the Banach-Tarski paradox. From
this point of view, the Banach-Tarski paradox is not a statement about R3 so much as it is a statement about the
complexity of the group M3 .
What, then, of our intuition about the preservation of volume by rigid motions? Well, as any student of measure
theory is aware, when one attempts to rigorously define a notion of volume, complications immediately arise. Any
hope of being able to give a consistent definition of volume to every subset of R3 must quickly be abandoned; indeed,
when one looks at the definitions and theorems of measure theory it is immediately apparent that there is absolutely
no reason why we should be able to assign a volume to every set. This is precisely what happens in the Banach-Tarski
paradox; most of the sets we have constructed have no volume to preserve, so when put together in strange ways
they are able to change their collective volume.
I would like to make one final argument about the Banach-Tarski paradox. Since subsets of R3 with interior points
all have the same cardinality, the following statement is an immediate consequence of the definition of bijection:
If A and B are bounded subsets of R3 with interior points, then A can be cut into uncountably many
pieces which can be moved by rigid motions to form exactly B.
The Banach-Tarski paradox, then, is simply this statement with uncountably replaced by finitely. That this replacement can be made, then, is really a statement about the size and complexity of the group M3 ; it is large enough that
we can group the uncountably many points together in such a way that only finitely many distinct rigid motions are
necessary. For all of these reasons, it is my opinion that the Banach-Tarski paradox is in no way a convincing argument against the axiom of choice; the axiom of choice may or may not be a reasonable assumption in mathematics,
but the Banach-Tarski paradox is no reason to reject it.
References
[1] Michael Artin, Algebra. Prentice Hall, Englewood Cliffs, New Jersey, 1991.
[2] Walter Rudin, Principles of Mathematical Analysis, 3rd edition. McGraw Hill Book Company, New York, 1976.
[3] Stanley Wagon, The Banach-Tarski Paradox. Cambridge University Press, Cambridge, 1985.