INROADS Real Analysis Exchange

Vol. 27(2), 2001/2002, pp. 783–794

Krzysztof Ciesielski∗, Department of Mathematics, West Virginia University,

Morgantown, WV 26506-6310, USA. e-mail: K Cies@math.wvu.edu
web page: http://www.math.wvu.edu/~kcies
Hajrudin Fejzić, Department of Mathematics, California State University,
San Bernardino, CA 92407, USA. e-mail: hfejzic@csusb.edu
Chris Freiling, Department of Mathematics, California State University, San
Bernardino, CA 92407, USA. e-mail: cfreilin@csusb.edu

MEASURE ZERO SETS WITH

NON-MEASURABLE SUM

Abstract

For any C ⊆ R there is a subset A ⊆ C such that A + A has inner

measure zero and outer measure the same as C + C. Also, there is a
subset A of the Cantor middle third set such that A + A is Bernstein in
[0, 2]. On the other hand there is a perfect set C such that C + C is an
interval I and there is no subset A ⊆ C with A + A Bernstein in I.

1 Introduction.
It is not at all surprising that there should be measure zero sets, A, whose sum
A+A = {x+y : x ∈ A, y ∈ A} is non-measurable. Ask a typical mathematician
why this should be so and you are likely to get the following response:

The Cantor middle-third set, when added to itself gives an entire

interval, [0, 2]. So certainly there exists a measure zero set that
when added to itself gives a non-measurable set.

Key Words: sets, algebraic sum, measurability

Mathematical Reviews subject classification: 26A21, 28A05
Received by the editors November 9, 2001
∗ The work of the first author was partially supported by NATO Grant PST.CLG.977652.

Papers Authored or Co-Authored by a Contributing Editor are managed by a Managing

Editor or one of the other Contributing Editors.

783
784 K. Ciesielski, H. Fejzić, C. Freiling

The intuition being that an interval has much more content than is needed for
a non-measurable set.
Indeed such sets do exist (in ZFC). Sierpiński (1920) seems to be the first
to address this issue. Actually, he shows the existence of measure zero sets
X, Y such that X +Y is non-measurable (see [7]). The paper by Rubel (see [6])
in 1963 contains the first proof that we could find for the case X = Y (see also
[5]). Ciesielski [3] extends these results to much greater generality, showing
that A can be a measure zero Hamel basis, or it can be a (non-measurable)
Bernstein set and that A+A can also be Bernstein. He also establishes similar
results for multiple sums, A + A + A etc.
This paper is mainly about the statement above and the intuition behind
it. Below we list four conjectures, each of which seems justified by extending
this line of reasoning.
1. Not only does such a set exist, but it can be taken to be a subset of the
Cantor middle-third set, C 13 . (This does not seem to immediately follow
from any of the above proofs. Thomson [9, p. 136] claims this to be
true, but without proof.)
2. The intuition really has nothing to do with the precise structure of the
Cantor set, which might lead one to conjecture the following. Suppose
C is any set with the property that C + C contains a set of positive
measure. Then there must exist a subset A ⊆ C such that A + A is
non-measurable.
3. The intuition relies on the fact that non-measurable sets can have far
less content than an entire interval. Therefore, the claim should also
hold when non-measurable is replaced by other similar qualities. Recall
that if I is a set, then a set S is called Bernstein in I if and only if
both S and its complement intersect every non-empty perfect subset
of I. Constructing a set that is Bernstein in an interval is one of the
standard ways of establishing non-measurability. Certainly, any set that
is Bernstein in an interval has far less content than the interval itself.
Therefore, we might conjecture that there is a subset A ⊆ C 31 with A+A
Bernstein in [0,2].
4. Combining the reasoning behind the Conjectures 2 and 3, let C be any
set with the property that C + C contains an interval, I. We might
conjecture that there must exist a subset A ⊆ C such that A + A is
Bernstein in I.
We will settle these four conjectures in the next four sections. In partic-
ular, in Section 2 we will give a proof of the first conjecture using transfinite
Measurable A with Non-Measurable A + A 785

induction. This provides what is possibly the simplest proof of the original
assertion. However, the proof depends on a particular property of symmetric
perfect sets. In Section 3 we give a slightly more complicated argument to
prove the second conjecture. We show that any set C contains a subset, A,
such that the inner measure of A+A is zero and the outer measure is the same
as C + C. In Section 4 we will settle the third conjecture, finding a subset A
of the Cantor set such that A + A is Bernstein in [0,2]. Finally, in Section 5 we
will give a counterexample to the fourth conjecture, showing that the above
lines of reasoning are indeed limited. In all of these proofs, we will be assuming
the axioms of ZFC. No additional set theoretical assumptions will be used.
The existence of these sets is interesting historically. Suppose E is a mea-
surable set and f is a measurable function. Several researchers in the area of
generalized derivatives have taken for granted the measurability of sets such
as:
(   )
 f (x + h) − 2f (x) + f (x − h)  1
Ej = x ∈ E :  < j for all 0 < |h| < .
 
 h2  j

It wasn’t until 1960 that Stein and Zygmund [8] pointed out that the measur-
ability of these sets is not automatic, and not until 1993 that Fejzić and Weil
re-proved these results without this assumption. In their paper [4] they also
show that this measurability assumption can be reduced to the (false) claim
that A + A is measurable whenever A is measurable (see also [5]).

2 Conjecture 1 is True.
It is well known that every real x in [0, 2] can be expressed as x = a + b where
a, b are in the Cantor set C 13 . It is also well known (a proof occurs below
in Lemma 4) that for almost every such x there are continuum many such
representations. The following theorem therefore gives a positive answer to
Conjecture 1.

Theorem 1. Let C be any subset of R and let

R = {x ∈ R : x has c many representations x = a + b with a, b ∈ C} .

Then there is a subset A ⊆ C such that A + A is Bernstein in R.

Proof. If R has no non-empty perfect subsets then we are done. Otherwise it

has continuum many such sets. Let {Pξ : ξ < c} be the family of all non-empty
perfect subsets of R. We will find an A ⊆ C such that each Pξ intersects both
786 K. Ciesielski, H. Fejzić, C. Freiling

A + A and its complement. Construct a sequence,

hhaξ , bξ , cξ , dξ i ∈ C × C × Pξ × Pξ : ξ < ci

such that for each ξ < c,

(∗) cξ = aξ + bξ and {dη : η ≤ ξ} ∩ (Aξ+1 + Aξ+1 ) = ∅,

S
where Aξ = η<ξ {aη , bη }. This will ensure that A = Ac has the desired
properties, since then {cξ : ξ < c} ⊆ A + A ⊆ R \ {dξ : ξ < c}.
To make an inductive step, assume that for some α < c we have already
constructed a partial sequence satisfying (∗) for all ξ < α. First choose a
cα ∈ Pα \ {dξ : ξ < α}. Then choose aα , bα in C such that aα + bα = cα and
neither aα nor bα is in {dξ : ξ < α} − Aα . Construction is finished by choosing
dα ∈ Pα − (Aα+1 + Aα+1 ).

3 Conjecture 2 is True.
Conjecture 2 is settled by the following theorem. Note that it proves more than
what is needed for Conjecture 2. However, as is the case with Theorem 1, it
falls short of establishing that A + A is Bernstein in C + C.
Theorem 2. For every C ⊂ R there exists an A ⊂ C such that A + A has
inner measure zero and outer measure the same as C + C.
Proof. We can assume that the inner measure of C + C is positive, since
otherwise A = C is as desired.
Let G be a Gδ -set containing C +C such that G\(C +C) has inner measure
zero. We will construct a set A ⊂ C such that every perfect set P ⊂ G of
positive measure intersects both G \ (A + A) and A + A. This implies that
A + A has inner measure zero and the same outer measure as C + C.
Let Ē be the family of all perfect subsets P of G of positive measure such
that |P ∩(C +C)| < c.1 Let E be a maximalS subfamily of Ē of pairwise disjoint
sets. Then E is at most countable, so E = E is an Fσ -set and E ∩ (C + C)
has cardinality less than c. SFor every e ∈ E ∩ (C + C) fix ce , de ∈ C such
that e = ce + de . Let Z = {{ce , de } : e ∈ E ∩ (C + C)}. Then Z ⊂ C also
has cardinality less than c and E ∩ (C + C) ⊂ Z + Z. Notice that by the
maximality of E,
1 Since P ∩ (C + C) must have positive outer measure, this set must be uncountable.

Also Ē is empty if either C + C is measurable or Martin’s axiom holds. However, there are
models of ZFC containing sets C of cardinality less than c with full outer measure. In that
case, C + C also has these properties and so Ē contains all perfect subsets of G of positive
measure.
Measurable A with Non-Measurable A + A 787

(a) |P ∩ (C + C)| = c for every perfect set P ⊂ G \ E of positive measure.

Next, let K̄ be the family of all perfect subsets F of G \ E of positive

measure for which there exists an XF ⊂ C such that

|XF | < c and F ∩ (C + C) ⊂ XF + C.2

Let K be a maximalSsubfamily of K̄ of pairwise disjoint

S sets. Then K is at most
countable, so K = K is an Fσ -set and X = F ∈K XF ⊂ C has cardinality
less than c. Clearly K ∩ (C + C) ⊂ X + C. Moreover, by the maximality of K,

(b) for every P ⊂ G \ (E ∪ K) of positive measure and any set X ⊆ C of

cardinality less than c, (P ∩ (C + C)) * (X + C).

Let P be the family of all non-empty perfect subsets P of G \ E of positive

measure such that either P ⊂ K or P ∩ K = ∅. Let {Pξ : ξ < c} be an
enumeration of P. We will find an A ⊂ C containing Z such that each Pξ
intersects both A + A and G \ (A + A). First notice that such a set will be as
desired.
Indeed, take a perfect set P ⊂ G of positive measure. We have to show
that P intersects both G \ (A + A) and A + A. If P ∩ E has positive measure
then ∅ 6= P ∩ E ∩ (C + C) ⊂ Z + Z ⊂ A + A, while P ∩ E 6⊂ A + A since
E ∩ (A + A) ⊂ E ∩ (C + C) has cardinality less than c. On the other hand, if
P ∩ E has measure zero, then P \ E contains one of sets Pξ , since in this case
either (P \ E) ∩ K or (P \ E) \ K has to have positive measure.
To define A construct hhaξ , bξ , cξ , dξ i ∈ C × C × Pξ × Pξ : ξ < ci such that
for every ξ < c

• cξ = aξ + bξ and {dη : η ≤ ξ} ∩ (Aξ+1 + Aξ+1 ) = ∅,

S
where Aβ = Z∪X∪ η<β {aη , bη }. This will ensure that A = Ac has the desired
properties since then {cξ : ξ < c} ⊂ A + A and {dξ : ξ < c} ∩ (A + A) = ∅.
To make an inductive step assume that for some ξ < c we have already
constructed a partial sequence
for which {dη : η < ξ} ∩ (Aξ + Aξ ) = ∅. Let
Y = LINQ Aξ ∪ {dη : η < ξ} , where LINQ (S) denotes the set of finite linear
combinations of elements in S with rational coefficients. Notice that |Y | < c
since |S| < c implies |LINQ (S)| < c. Consider two cases.
Case 1: Pξ ⊂ K. Choose cξ ∈ Pξ ∩ (C + C) \ Y . The choice can be made by
(a), since |Pξ ∩ (C + C)| = c as Pξ ⊂ G \ E has positive measure.
2 The family K̄ may be non-empty even if C has measure zero. Although this cannot

happen under Martin’s axiom, it happens in any model of ZFC in which there is a non-
meager measure zero set of cardinality less than c. (See e.g. [2, thms. 2.7.3 and 2.1.7].)
788 K. Ciesielski, H. Fejzić, C. Freiling

As cξ ∈ Pξ ∩ (C + C) ⊂ K ∩ (C + C) ⊂ X + C there are aξ ∈ X ⊂ Aξ and

bξ ∈ C such that cξ = aξ + bξ . We now show that dη ∈ / Aξ+1 + Aξ+1 for each
η < ξ. Assume otherwise. Since Aξ+1 = Aξ ∪ {bξ } and dη ∈ / Aξ + Aξ , it must

be that dη ∈ {bξ } + (Aξ ∪ {bξ }). But then bξ ∈ Y = LINQ Aξ ∪ {dη : η < ξ}
and so also is cξ = aξ + bξ , contradicting the choice of cξ .
Case 2: Pξ ⊂ G \ (E ∪ K). We will first show that

(∗) there exist a, b ∈ C \ Y such that a + b ∈ Pξ \ Y .

To see this, for each p ∈ Pξ ∩ (C + C) choose ap , bp ∈ C with p = ap + bp . If

(∗) is false then for every p ∈ Pξ ∩ (C + C) \ Y we would have {ap , bp } ∩ Y 6= ∅.
But then Pξ ∩ (C + C) \ Y ⊂ (Y ∩ C) + C. So, for X0 = (Y ∩ C) ∪ {ap : p ∈ Y }
we have Pξ ∩ (C + C) ⊂ X0 + C contradicting (b).
Now, take a and b as in (∗) and put aξ = a, bξ = b, and
cξ = a + b. Then
aξ , bξ , cξ ∈
/ Y . In particular, dη ∈
/ {aξ , bξ , cξ } + LINQ Aξ for every η < ξ. So,
{dη : η < ξ} ∩ (Aξ+1 + Aξ+1 ) = ∅.
We finish the inductive steps by choosing a dξ ∈ Pξ \ LINQ (Aξ+1 ).

4 Conjecture 3 is True.
In this section we embellish the argument in Section 1 to settle the third
conjecture.
Definition 3. Two real numbers x, y will be called equivalent and we write
x ∼ y if and only if there is a ternary expansion of x and a ternary expansion
of y such that the two expansions disagree on only finitely many digits.
Note that if x is a ternary rational, then it will have two possible ternary
expansions. According to the definition, all such ternary rationals are equiv-
alent. Every other real x has a unique ternary expansion. The following
theorem fulfills the promise made in Section 2, showing that almost every
x ∈ [0, 2] has c many representations as the sum of elements in the Cantor set.
Lemma 4. Let x ∈ [0, 2]. If x/2 has infinitely many ones in its ternary expan-
sion, then there are c many representations of x as the sum of two Cantor-set
elements. Otherwise, x has only finitely many such representations and all of
the elements of C 13 used to represent x are equivalent to x/2.
Proof. Every element of C 13 has a ternary expansion consisting of only even
digits. Fix x ∈ [0, 2] and let c = x/2. If x = a + b with a, b ∈ C 13 , then c is
the average of a and b. If ci , ai , bi are the ith digits of c, a, and b, respectively,
then either
Measurable A with Non-Measurable A + A 789

• ci = 0 and ai = bi = 0,
• ci = 1 and ai = 0, bi = 2,
• ci = 1 and ai = 2, bi = 0, or
• ci = 2 and ai = bi = 2.
Suppose first that c is a ternary rational. Then the digits of c must end in either
a sequence of zeros or a sequence of two’s. In either case, the digits of a and b
must do likewise and so they are also ternary rationals. Therefore, a ∼ b ∼ c.
Now consider the case when c is not a ternary rational, so there is a unique
ternary expansion of c. Let us construct the numbers a and b using only even
digits. For each ci that is zero or two, we must have ai = bi = ci . But for each
ci that is one, we have a choice, either ai = 0, bi = 2 or ai = 2, bi = 0. Thus
if k ∈ {0, 1, 2, . . . , ω} is the number of digits in c that have the value 1, then
there are 2k possible choices for the pair a, b. In particular, if c has infinitely
many ones in its expansion, then there are 2ω = c many representations for x.
If there are only finitely many ones, then the digits of a, b, c will all agree on
a tail end.
Theorem 5. There is a set A ⊆ C 13 such that A + A is Bernstein in [0, 2].

Proof. Let R0 be the set of elements of [0, 2] that can be expressed in c many
ways as the sum of elements of C 31 , and let R1 be the elements that can be
expressed in only finitely many ways. Let {Pξ : ξ < c} be the family of all
non-empty perfect subsets of [0, 2]. We will find an A ⊆ C 13 such that each Pξ
intersects both A + A and its complement. Construct

hhaξ , bξ , cξ , dξ i ∈ C × C × Pξ × Pξ : ξ < ci

such that for each ξ < c,

(∗) cξ = aξ + bξ and Dξ ∩ (Aξ+1 + Aξ+1 ) = ∅,
S
where Aξ = η<ξ {aη , bη } and Dξ = {dη : η ≤ ξ}. This will ensure that A = Ac
has the desired properties, since then {cξ : ξ < c} ⊆ A + A ⊆ R \ Dξ . To make
the inductive step, assume that for some α < c we have already constructed
a partial sequence satisfying (∗) for all ξ < α. We first choose aα , bα , cα such
that aα +bα = cα and such that neither aα nor bα is in Dα −Aα . We distinguish
two cases.
Case 1: Pα intersects R0 in a set of cardinality c. First choose cα ∈
Pα ∩ R0 \ Dα . Then choose aα , bα in C 31 such that aα + bα = cα and neither
aα nor bα is in Dα − Aα .
790 K. Ciesielski, H. Fejzić, C. Freiling

Case 2: Pα intersects R1 in a set of cardinality c. First choose cα ∈

Pα ∩ R1 \ Dα such that cα /2 is not equivalent to any element of Dα − Aα .
Then choose aα , bα ∈ C 31 such that aα + bα = cα . Since, by Lemma 4, both
aα and bα are equivalent to cα /2, neither of them is in Dα − Aα .
Construction is finished by choosing dα ∈ Pα \ (Aα+1 + Aα+1 ).

5 Conjecture 4 is False.
In this section we will construct a set A such that A + A contains an interval,
I, yet there is no subset B ⊆ A with B + B Bernstein in I. Let C 12 be the
Cantor middle-half set, which will be the basis of our construction. That is,
C 12 is the set of points, x, in the unit interval such that there is a base four
expansion of x that uses only zeros and threes. Note that if the expansion
of x ends in a sequence of zeros, then there will be an equivalent expansion
ending in a sequence of threes. In this case we will say that x is a quaternary
rational. In all other cases, the base 4 decimal expansion is unique.
When we dealt with sums from C 13 + C 13 , it was easier to think in terms of
averages, so that the work could be carried out digit-wise. Similarly, when we
work with C 12 it will be easier to think in terms of the following auxiliary sets.
Let U be the set of elements of [0, 1] that use only zeros and twos in one of
its base four expansions, and let V be the set of elements that use only zeros
and ones. Since threes are not allowed in either set, each element of U and V
has only one valid expansion. Furthermore, the sums in U + V and the sums
in V + V can be carried out digit-wise since there will be no carries.
Our construction will be based on the following three lemmas. The proofs
of the first two of them can be seen geometrically by examining Figure 1.
Lemma 6. C 21 + C 12 has measure zero.

Proof. Let x = a + b with a, b ∈ C 12 . Let c = 13 x = 13 a + 31 b. Using a base 4

expansion of a and b where all digits are divisible by three, the computation
of 31 a + 13 b can be carried out digit-wise and results in an expansion where the
digit 3 is not used. Therefore, unless c is a quaternary rational, its expansion
will never use the digit three. Thus 13 (C 12 + C 12 ) has measure zero and so does
C 21 + C 12 .
Now consider a sum c = u + v where u ∈ U and v ∈ V . There are four
possibilities for the ith digits of this calculation.
• ci = 0, ui = vi = 0
• ci = 1, ui = 0, vi = 1
Measurable A with Non-Measurable A + A 791

6 6

A A
A A
@ A A A A
@ A A A A
@ A A A A
@ A A A A
A A
@ A A A A
@ @ A A
@ hole
R @@
@ R - AAU A
AU AAU A
UA -
0 1 0 1 1.5

Figure 1: Consider the pairs of reals from the unit interval with the middle
half removed. The projection along the direction with a slope −1 has holes,
while the projection along the direction with a slope −2 fills the interval [0, 1.5]
without any overlaps.

• ci = 2, ui = 2, vi = 0

• ci = 3, ui = 2, vi = 1

Unlike the sums of C 31 + C 13 this time there is no ambiguity. Every choice of

the digits of c gives a unique choice for the digits of u and v. The next lemma
relates this to the sum C 21 + 21 C 12 .

Lemma 7. C 12 + 12 C 12 = [0, 1.5]. Furthermore, each element in [0, 1.5] can be

expressed as such a sum in at most two ways, and except for a countable set,
each element in [0, 1.5] can be expressed in a unique way.

Proof. Fix an x ∈ [0, 1.5], and suppose x = a + b with a ∈ C 21 and b ∈ 21 C 12 .

Let c = 32 x = 23 a + 23 b. Using the fact that all of the digits of a are divisible
by three, the computation of 32 a can be carried out digit-wise and results in
an element of U . Similarly, 23 b ∈ V . But each such c has at most two such
representations and except when c is a quaternary rational, each such c has a
unique representation.
The equality is justified by 23 (C 12 + 12 C 12 ) = U + V = [0, 1] = 32 [0, 1.5].

We are now ready to define the set A that will serve as our counterexample.
Let A = C 12 ∪ 12 C 12 .
792 K. Ciesielski, H. Fejzić, C. Freiling

Lemma 8. There are two non-empty perfect subsets P and Q of A such that
every element of P + Q can be expressed uniquely as the sum of two elements
in A.

Proof. A + A is the union of three closed sets: C 12 + C 12 , 12 C 12 + 12 C 12 , and

C 12 + 12 C 12 . By Lemma 6, the first and second sets are both measure zero.
Choose an open interval I ⊆ [0, 1.5] that is disjoint from the first two sets. By
Lemma 7, the third set is the interval [0, 1.5]. Furthermore, this set can be
partitioned into two sets X and Y such that X is countable and every element
in Y has a unique representation as a sum of two elements, one in C 12 and
the other in 12 C 21 . Choose an x in Y ∩ I and let x = a + b with a ∈ C 12
and b ∈ 12 C 12 . Now choose a neighborhood J of a and a neighborhood K of
b small enough that the closure cl(J + K) of J + K is a subset of I. Let R
be the intersection of A with cl(J) and let S be the intersection of A with
cl(K). Then both R and S are non-empty perfect subsets of A. Let D be the
countable set consisting of all numbers used in the representations of elements
of X, and let P and Q be non-empty perfect subsets of R \ D and S \ D
respectively.
Fix x ∈ P + Q. We must show that x has a unique representation as a
sum of elements of A. Since x ∈ (R \ D) + (S \ D) then there exist a ∈ R \ D
and b ∈ S \ D with x = a + b. Since R ⊆ cl(J) and S ⊆ cl(K), this gives us
that x ∈ I. But then x is not in the first two pieces of A + A so it must be
that one of the elements a, b is in C 21 and the other is in 12 C 12 . Since a and b
are not in D, x cannot be in X. Therefore, x ∈ Y , and we are done.

Theorem 9. There is no subset B ⊆ A such that B + B is Bernstein in

[0, 1.5].

Proof. Suppose that such a set B exists. Let P and Q be as in the previous
lemma. B can’t contain a non-empty perfect subset, since that would imply
B + B also contains a non-empty perfect subset of [0, 1.5]. Therefore, there is
some element x in P \ B. Then x + Q is a perfect subset of P + Q and so each
element of x + Q has a unique representation as a sum of elements in A. But
then since x ∈ / B no element of x + Q is in B + B. So, B + B is not Bernstein
in [0, 1.5], which is a contradiction.

References
[1] J. M. Ash, S. Catoiu, and R. Rı́os-Collantes-de-Terán, On the n-th quantum
derivative, preprint.
Measurable A with Non-Measurable A + A 793

[2] T. Bartoszyński, H. Judah, Set Theory, A K Peters, 1995.

[3] K. Ciesielski, Measure zero sets whose complex sum is non-measurable,
Real Anal. Exchange, 26(2) (2000–2001), 919–922.
[4] H. Fejzić and C. E. Weil, Repairing the proof of a classical differentiation
result, Real Anal. Exchange, 19(2) (1993–94), 639–643.
[5] S. N. Mukhopadhyay and S. Mitra, An extension of a theorem of Ash on
generalized differentiability, Real Anal. Exchange, 24(1) (1998–99), 351–
371.
[6] L. A. Rubel, A pathological Lebesgue measurable function, J. London Math.
Soc., 38 (1963), 1–4.
[7] W. Sierpiński, Sur la question de la measurabilité de la base de M. Hamel,
Fund. Math., 1 (1920), 105–111.
[8] E. M. Stein and A. Zygmund, Smoothness and differentiability of functions,
Annals of the University of Sciences, Budapest, III-IV (1960–61), 295–307.
[9] B. S. Thomson, Symmetric Properties of Real Functions, Marcel Dekker,
1994.
794 K. Ciesielski, H. Fejzić, C. Freiling