4

4
5


r
e
v
i
s
i
o
n
:
1
9
9
4
-
0
6
-
2
2







m
o
d
i
f
i
e
d
:
1
9
9
4
-
0
6
-
2
2


Every null-additive set is meager-additive
Saharon Shelah
1. The basic denitions and the main theorem.
1. Denition. (1) We dene addition on

2 as addition modulo 2 on each component,
i.e., if x, y, z

2 and x + y = z then for every n we have z(n) = x(n) + y(n) (mod 2).
(2) For A, B

2 and x

2 we set x + A =
df
x + y : y A, and we dene A + B
similarly.
(3) We denote the Lebesgue measure on

2 with . We say that X

2 is null-additive
if for every A

2 which is null, i.e. (A) = 0, X + A is null too.
(4) We say that X

2 is meager-additive if for every A

2 which is meager also X +A
is meager.
2. Theorem. Every null-additive set is meager-additive.
3. Outline and discussion. Theorem 2 answers a question of Palikowsi. It will be
proved in 2. In 3 we shall present direct characterizations of the null-additive sets, and
in 4 we shall do the same for the meager-additive sets.
It is obvious that every countable set is both null-additive and meager-additive. Are
there uncountable null-additive sets, and even null-additive sets of cardinality 2

0
? It will
be shown in 5 that if the continuum hypothesis holds then there is such a set. Haim Judah
has shown that there is a model of ZFC in which all the null-additive sets are countable,
but there are in it uncountable meager-additive sets. This is the model obtained by adding
to L more than
1
Cohen reals. In this model the Borel conjecture holds, and therefore
every null-additive set is strongly meager and hence countable. On the other hand, in this
model the uncountable set of all constructible reals is meager-additive.
2 The proof of Theorem 2.
4. Notation. (1) we shall use variables as follow: i, j, k, l, m, n for natural numbers,
f, g, h for functions from to , , , , , for nite sequences of 0s and 1s, x, y, z for
members of

2, A, B, X, Y for subsets of

2, and S, T for trees.
(2)
>
2 =

n<
n
2. We shall denote subsets of
>
2 with U, V . For
>
2, U
>
2
and x

2 we shall write + x for + x length(), and U + x for + x : U.
(3) For ,
>
2 we write if is an extension of .
(4) A tree is a nonempty subset of
>
2 such that
(a) If and T then also T, and
(b) If T and n > length () then there is a of length n such that and
T.
(5) For a tree T LimT = x

2 : for every n < x n T.
(6) A tree T is said to be nowhere dense if for every T there is a
>
2 such that
Publ. No. 445. First version written in April 91. Partially supported by the Basic Research Fundations of
the Israel Academy of Sciences and the Edmund Landau Center for research in Mathematical Analysis, supported
by the Minerva Foundation (Germany)
We thank Azriel Levy for rewriting the paper. The reader should thank him for including more proofs and
making the paper self contained.
1
4
4
5


r
e
v
i
s
i
o
n
:
1
9
9
4
-
0
6
-
2
2







m
o
d
i
f
i
e
d
:
1
9
9
4
-
0
6
-
2
2


and / T. A set B

2 is said to be nowhere dense if B LimT for some nowhere
dense tree T.
(7) For every x, y

2 we write x y if x(n) = y(n) for all but nitely many n < . For
A

2 A
n
=
df
y

2 : y x for some x A.
(8) U
[]
=
df
U : or (read: U through ).
(9) U

=
df

>
2 :

U (read: U above ), and for

>
2

k
=
df
(k + i) : i < length k).
(10) For ,
>
2

2 we write
n
if length () = length () and (i) = (i) for
every n i < length (). For S
>
2

2 we dene S
n
= :
n
for some S.
5. Outline of the proof. Let X

2 be null-additive. It clearly suces to prove that
for every A

2 which is nowhere dense X + A is meager. Given a nowhere dense tree
S we shall prove in Lemma 6 a condition which is sucient for a tree T to be such that
T +S is nowhere dense. Then we shall split X to a union X =

i=1
X
i
such that for each i
X
i
LimT
i
where T
i
is a tree which satises that condition. Thus for a nowhere dense S
each set X
i
+LimS Lim(T
i
+S) is nowhere dense, hence X+LimS

i=1
Lim(T
i
+S)
is meager.
6. Lemma. Let T be a tree such that
(a) T is nowhere dense.
(b) f = f
T
is the function from to given by
f(n) = minm : for every
n
2 there is a
m
2 such that and / T.
Thus for every sequence of length n there is a witness of length f(n) that T is nowhere
dense. Obviously for every n < f(n) > n, and if n < m then f(n) f(m).
Let g be a function from to and n = n
i
: i < ), n

= n

i
: i < ) increasing
sequences of natural numbers such that
(c) f
g(i)
(n
i
) n

i
< n
i+1
for every i < , where f
m
denotes the m-th iteration of f.
Then for every tree S which satises
(d) S is of width (n

, g), i.e., for every i < [

n

i
2 S[ g(i),
T + S is nowhere dense.
Proof. Let
n
i
2. We shall show the existence of an

i
2 such that

and

/ T + S.
By (c) there is a sequence m
0
, . . . , m
g(i)
such that m
0
= n
i
, f(m
k
) m
k+1
for
0 k g(i) and m
g(i)
= n

i
. Let
k
: k < k
i
) enumerate the set
n

i
2 S. k
i
g(i) by (d).
We dene
k

m
k
2 for 0 k k
i
by recursion as follows.
0
= . Given
k

m
k
2,
for k < k
i
, we shall dene
k+1

m
k+1
2 so that for no extension

i
2 of
k+1
we
shall have

+
k
T. We have
k
+
k
m
k

m
k
2 and by the denition of f and by
the choice of the m
k
s
k
+
k
m
k
has an extension
m
k+1
2 such that / T. If we
take
k+1
= +
k
m
k+1
then
k
+
k
m
k
implies
k

k+1
,
k+1

m
k+1
2 and

k+1
+
k
m
k+1
= / T, and therefore for every

i
2 such that
k

we have

+
k
/ T. Let

=
k
i
, and assume that

T + S. Then, for some k < k

i
g(i)

+
k
T, contradicting our choice of
k+1
=

m
k+1
. Thus

/ T + S.
7. Lemma. If S, T
i
, i are trees and LimS

i
LimT
i
then for some S and
j S
[]
T
j
.
Proof. Suppose that this is not the case, i.e., for every S and i < there is a
2
4
4
5


r
e
v
i
s
i
o
n
:
1
9
9
4
-
0
6
-
2
2







m
o
d
i
f
i
e
d
:
1
9
9
4
-
0
6
-
2
2


such that S
[]
and / T
i
. Once there is such a we can assume that and
length > length. We dene now, by induction on i,
i
and k
i
so that k
i
= length
i
,
k
0
= 0,
0
= ),
i

i+1
, k
i
< k
i+1
,
i+1
S and
i+1
/ T
i
. Let y =

i

i
.
Since
i
S for every i y LimS

i
LimT
i
, hence for some j y LimT
j
.
However, y k
j+1
=
j+1
/ T
j
, contradicting y LimT
j
.
8. Lemma. Let S and T be trees such that LimS (LimT)
n
. Then there are k < ,
,
k
2, S such that S

.
Proof. For n < ,
1
,
2

n
2 and
2
T we dene
T

1
,
2
=
df
:
1

1

:
2

T (This is the tree T

[
2
]
with
2
replaced
by
1
). Clearly
(1) (LimT)
n
=
_
n<,
1
,
2

n
2,
2
T
LimT

1
,
2
Since there are only countably many T

1
,
2
s in (1) there are by Lemma 7 a S and j <
such that S
[]
T

1
,
2
. Clearly there is an with and a with length = length
such that S

. (If
1
then =
1
and =
2
, else
1
and then = and
=
2

[length, length
2
) ).
9. Lemma. Let X be a null-additive set. Let T be a tree such that (LimT) > 0.
There is a tree T

such that (LimT

) > 0, for every T

(Lim(T
[]
)) > 0, and
((

2 (LimT)
n
) + X) LimT

= , and then X =

T

, length=length
Y
,
where
Y
,
= x X :

x
length
+ T
[]
T.
Proof. Since (LimT) > 0 then, as easily seen, ((LimT)
n
) = 1, hence
(

2 (LimT)
n
) = 0. Since X is null-additive also (X + (

2 (LimT)
n
)) = 0. Hence
there is a tree T

such that (LimT

) > 0 and (X + (

2 (LimT)
n
)) LimT

= .
Without loss of generality we can assume that T

has been pruned so that for T

(LimT
[]
) > 0.
Let x X then

2 (x + (LimT)
n
) = x + (

2 (LimT)
n
) X + (

2 (LimT)
n
).
Hence (

2 (x + (LimT)
n
)) LimT

(X + (

2 (LimT)
n
) LimT

= , i.e.,
LimT

x + (LimT)
n
, and therefore Lim(x + T

) = x + LimT

(LimT)
n
. By
Lemma 8 there are T

and
length
2 such that x
length
+ T

. Let
= + , then + = and therefore

x
length
+ T
[]
T
[]
T, hence x Y
,
.
10. Lemma. Let X be null-additive, and let n = n
i
: i < ), n

= n

i
: i < ) be
such that for every i < n
i
< n

i
and n

i
+ i 2
n

i
n
i+1
, then we can represent X as

m<
X
m
such that for each m, for some real a
m
(0, 1) and S
m
of width (n

, g
a
m
) we
have X
m
Lim(S
m
), where for every real a (0, 1) g
a
is the function on given by
g
a
(0) = 1, g
a
(i) = max(1, int(log
2
(a)/ log
2
(1 2
i
))), where for a real d int(d) is the
integral part of d.
Proof. Since n

i
+i 2
n

i
< n
i+1
we can x for each 0 < i < a sequence u
i,
:
n

i
2) of
pairwise disjoint subsets of the interval [n

i
, n
i+1
) having i members each. Let B

2 be
given by
B = y

2 : (j > 0)(k u
j,yn

j
) y(k) = 1.
B is clearly a closed subset of

2 hence for T = y n : y B n B = Lim(T).
3
4
4
5


r
e
v
i
s
i
o
n
:
1
9
9
4
-
0
6
-
2
2







m
o
d
i
f
i
e
d
:
1
9
9
4
-
0
6
-
2
2


The properties of T in which we are interested are
(B0) T
n
1
2.
(B1) For each T
n

i
2

T
[]

n
i+1
2

= 2
(n
i+1
n

i
)
(1 2
i
).
(B2) If ,
0
, . . . ,
k1

n

i
2,
+
0
, . . . ,
+
k1

n
i+1
2, +
l
T,
l

+
l
for l < k and

0
, . . . ,
k1
is with no repetitions then

+
:
+

n
i+1
2, (l < k)(
+
+
+
l
T)

2
n
i+1
n

i
_
1 2
i
_
k
.
(B3) For every
n

i
2 we have : n
i
T implies T.
These properties can be established by an obvious counting argument.
By (B0), (B1) and (B3) we have
(LimT) =
_

i=1
x

2 : x n
i
T
_
= (x

2 : x n
1
T)

i=1
(x

2 : x n
i+1
T)
(x

2 : x n
i
T)
= 1

i=1
[T
n
i+1
2[/2
n
i+1
[T
n
i+1
2[/2
n
i
=

i=1
(1 2
i
) > 0
For the T which we constructed let T

and Y
,
be as in Lemma 9. For
length
2
let Y
,,
= y Y
,
: y length = . Clearly
(2) X =
_
T

, length=length =length
Y
,,
Since there are only countably many Y
,,
s they can be taken to be the X
m
s we are
looking for, provided we show that every such Y
,,
is a subset of Lim(S) for some tree
S of width n

, g
a
) for some real 0 < a < 1. We shall see that this is indeed the case if we
take S = y m : y Y
,,
, m < and a = (T
[]
). a > 0 by what we assumed about
T

. As, obviously, Y
,,
Lim(S) all we have to do is to show that S is of width n

, g
a
).
We can choose a set W S
n
j+1
2 such that the function mapping W to n

j
is one to one and onto S
n

j
2
We x now , , and denote Y
,,
by Y and the length of , , by n. Let z

2
be such that z n = + and z(i) = 0 for i n. Then for every y such that y n = we
have y + z =

y
n
. Therefore, by the denition of Y we have
(3) Y = y

2 : yn = , (

y
n
)+T
[]
T = y

2 : yn = , y+z+T
[]
T
for every y Y there is a unique W such that n

j
= y n

j
( may be y n
j+1
).
Clearly [W[ = [S
n

j
2[ and we denote [W[ with s, so it suces to prove s g
a
(j). If n

j
n
then the only member of S
n

j
2 is n

j
hence s = 1, so s g
a
(j). We shall now deal with
the case where n

j
> n. Let
0
, . . .
s1
be the members of W. For m < s
m
= y n
j+1
for
some y Y , hence, by (3),
m
+z +T
[]
T and therefore (z +T
[]
)
n
j+1
2
m
+T.
Since this holds for every W we have
(4) z + T
[]

n
j+1
2

m<s

m
+ T
4
4
4
5


r
e
v
i
s
i
o
n
:
1
9
9
4
-
0
6
-
2
2







m
o
d
i
f
i
e
d
:
1
9
9
4
-
0
6
-
2
2


Let us nd out the size of

m<s
(
m
+ T). Let
n

j
2, and we shall ask how many
members of

m<s
(
m
+ T) extend . Now
m
+ T for each m < s i +
m
T
for each m < s. If for some m < s +
m
n

j
/ T then also +
m
/ T, hence has
no extension in

m<s
(
m
+ T). If for every m < s +
m
n

j
T then by (B2) (where
= ,
m
=
m
n

j
and
+
m
=
m
), since
m
n

j
,=
l
n

j
for m ,= l, the number of s
such that
n
j+1
2 and +
m
T for every m < s is 2
n
j+1
n

j
(1 2
j
)
s
. Since there
are 2
n

j
dierent s in
n

j
2 we have
(5)

m<s
(
m
+ T)

2
n
j+1
(1 2
j
)
s
.
On the other hand, since (T
[]
) = a T
[]

n
j+1
2 has at least a 2
n
j+1
members,
and so has z + T
[]

n
j+1
2. Comparing (4) with (5) we get a 2
n
j+1
2
n
j+1
(1 2
j
)
s
,
i.e., a (1 2
j
)
s
, log
2
(a) s log
2
(1 2
j
), s log
2
(a)/ log
2
(1 2
j
).
11. Proof of Theorem 2. Let X be null-additive. As mentioned in 5 it suces to show
that for every nowhere dense tree T X + Lim(T) is meager. Let f = f
T
as in Lemma 6.
Dene by recursion n
0
= 0, n

i
= f
g
1/(i+1)
(i)
(n
i
)+1 and n
i+1
= n

i
+i2
n

i
+1. By Lemma 10
X

m<
Lim(S
m
), where for some a
m
(0, 1) S
m
is of width n

, g
a
m
), hence it suces
to show that if S is of width n

, g
a
) for some a (0, 1) then Lim(S)+Lim(T) = Lim(S+T)
is meager. Let j be such that
1
j+1
a and let
1
, . . . ,
k
be all the members of S of
length n

j
. Then S =

k
l=1
S
[
l
]
and LimS =

k
l=1
Lim(S
[
l
]
). Therefore it suces to
prove that for 1 l k Lim(S
l
) + Lim(T) is meager and this follows once we show
that S
l
+ T is nowhere dense. To prove this we show that the requirements of Lemma
6 hold here for S
l
, T. (a) and (b) hold by our choice of T and f. Let g be dened by
g(i) = 1 for i < j and g(i) = g
a
(i) for i j. Now we shall see that (c) holds. For
i < j we have n

i
= f
g
1/(i+1)
(i)
(n
i
) + 1 f(n
i
) + 1 = f
g(i)
(n
i
) + 1, since f(n) n for
every n, and for i j we have n

i
= f
g
1/(i+1)
(i)
(n
i
) + 1 f
g
a
(i)
(n
i
) + 1 = f
g(i)
(n
i
) + 1,
since a
1
j+1

1
i+1
and g
a
(i) is a decreasing function of a. Thus for every i <
f
g(i)
(n
i
) f
g
1/(i+1)
(i)
(n
i
) n

i
. (d) of Lemma 6 holds since for i < j [
n

i
2 S
l
[ = 1 = g(i)
and for i j [
n

i
2 S
l
[ [
n

i
2 S[ g
a
(i) = g(i).
3 Characterization of the null-additive sets
12. Denition. By a corset we mean a non decreasing function f from to 0
which converges to innity (i.e., for every n < f(m) > n for all suciently large m).
For a corset f, we say that a tree T is of width f if for every n < [T
n
2[ f(n); and
we say that T is almost of width f if [T
n
2[ f(n) for all suciently large n.
13. Theorem. For every X

2 the following conditions are equivalent:
a. X is null-additive.
b. For every corset f there is a tree S of width f such that X Lim(S)
n
.
c. For every corset f there are trees S
m
, m < , which are almost of width f such that
X

m<
Lim(S
m
)
n
.
d. For every corset f there are trees S
m
, m < , of width f such that X

m<
Lim(S
m
).
Proof. (b)(c) is obvious.
5
4
4
5


r
e
v
i
s
i
o
n
:
1
9
9
4
-
0
6
-
2
2







m
o
d
i
f
i
e
d
:
1
9
9
4
-
0
6
-
2
2


(c)(d). Let S be a tree almost of width f. Then for some k we have [T
n
2[ f(n) for
all n k. By (1) of Lemma 8 Lim(S)
n
=

1
,
2

k
2,
2
S
Lim(S

1
,
2
). Each S

1
,
2
is of
width f since for n k we have [S

1
,
2

n
2[ = 1 and for n > k we have [S

1
,
2

n
2[
[S
n
2[ f(n). Therefore, if X

m<
Lim(S
n
)
n
as in (c) then each S
m
can be replaced
by countably many S

1
,
2
s and (d) holds.
(d)(b). Let f be a corset. We can easily dene by recursion a sequence 0 = n
0
< n
1
< . . .
of natural numbers and a corset f

such that for all j < and m n

j+1
we have
(j + 1) 2
n
j
f

(m) f(m).
For a given corset f, if X satises (d) let S

m
, m < , be as in (d) for the corset f

.
We construct now a set S
>
2 by dening S
m
2 by recursion on m. S
0
2 = ).
For n
i
< m n
i+1
let
S
m
2 =
m
2 : n
i
S
n
i
2 and S
n
j
j
for some j < i j = 0.
S can be easily seen to be a tree, and clearly Lim(S)
n

n<
Lim(S

n
) X. For m n
1
easily , for n
i
m < n
i+1
, i 1 we have
[S
m
2[

ji

S
n
j
j

m
2

ji
2
n
j
[S

j

m
2[ (i + 1) 2
n
j
f

(m) f(m),
thus S is of width f.
(d)(a). Assume now that (d) holds for X, and let A

2, (A) = 0; we shall prove
that (X + A) = 0. First we shall mention two lemmas of measure theory the proof of
which is left to the reader.
Lemma A. For every tree T with (Lim(T)) = a > 0 and > 0 there is an N such
that for every n N there is a t
n
2 T such that [t[ 2
n
(a ) and for each t
(Lim(T
[]
) > 2
n
(1 ).
Using Lemma A one can prove
Lemma B. For every tree T with (Lim(T)) > 0, every > 0 and every sequence

i
: 0 < i < ) of positive reals there is a subtree T

of T and an increasing sequence

n
i
: i < ) of natural numbers such that n
0
= 0, (Lim(T

)) > (Lim(T)) and

(6) for i > 0 and every
n
i
2 T

(Lim(T
[]
)) > 2
n
i
(1
i
)
By basic mesure theory (A
n
) = 0 so there is a tree T such that (Lim(T)) > 0 and
Lim(T) A
n
= hence Lim(T)
n
A = . Given < (Lim(T)) and
i
: i < ) as in
Lemma B we obtain a subtree T

of T as in that lemma with (Lim(T

)) > 0. The union

of suciently many nite translates of T

, i.e., trees T

1
,
2
as in (1) of Lemma 8 is a
tree T

satisfying (6) with (Lim(T

))
1
2
. Lim(T

)
n
= Lim(T

)
n
Lim(T)
n
and
hence Lim(T

) A Lim(T)
n
A = . We take now
i
=
1
4(i+1)
3
and take T to be T

and we get (Lim(T))

1
2
and
(7) for i > 0 and every
n
i
2 T (Lim(T
[]
)) > 2
n
i
(1
1
4(i + 1)
3
)
Let f be the corset given by f(n) = i +1 for n
i
n < n
i+1
. By (d) there are trees S
m
of width f such that X

m<
Lim(S
m
). To show that (X + A) = 0 it clearly suces
to show that for every tree S of width f (Lim(S) + A) = 0.
6
4
4
5


r
e
v
i
s
i
o
n
:
1
9
9
4
-
0
6
-
2
2







m
o
d
i
f
i
e
d
:
1
9
9
4
-
0
6
-
2
2


We dene
T

=
>
2 : + T for every S of the same length as
We do not show that T

is a tree but obviously if T

then T

, thus
Lim(T

) is dened. If (Lim(T

)) > 0 then, by a well-known property of the mea-

sure, (Lim(T

)
n
) = 1, hence in order to prove (Lim(S) + A) = 0 it suces to
prove (Lim(S) + A) Lim(T

)
n
= . Assume y (Lim(S) + A) Lim(T

)
n
. Since
y Lim(T

)
n
there is a y

2 such that y

(n) = y(n) for all suciently big ns and

y

Lim(T

). Since y Lim(S) + A there is an x Lim(S) such that y + x A, hence

y +x / Lim(T)
n
, hence y

+x / Lim(T). Therefore, for some n y

n+x n / T, hence,
by the denition of T

, y

n / T

contradicting y

Lim(T

).
We still have to prove that (Lim(T

)) > 0. We shall prove, by induction on i, that

(8) n
i
n n
i+1
[(T T

)
n
2[ 2
n

j<i
1
4(j + 1)
2
.
Once we establish (8) we notice that since
Lim(T) Lim(T

) =

n<
Lim(T) x

2 : x n T

, and the set

Lim(T) x

2 : x n T

is increasing with n hence (Lim(T) Lim(T

))
= lim
n
(Lim(T) x

2 : x n T

) lim
n
2
n
[(T T

)
n
2[
lim
n

n
j=0
1
4(j+1)
2
=

j=0
1
4(j+1)
2
=

2
24
<
1
2
and since (Lim(T))
1
2
(Lim(T

)) > 0.
To prove (8), assume now n
i
n n
i+1
. By the denition of T

(T T

)
n
2 = T
n
2 : ( S
n
2) + / T
= T
n
2 : ( S
n
2)( n
i
+ n
i
/ T)

S
n
2
T
n
2 : n
i
+ n
i
T + / T

n
2 : n
i
T T

S
n
2

n
2 : +
n
2 : n
i
T / T.
Therefore [(T T

)
n
2[ 2
nn
i
[(T T

)
n
i
2[ +[S
n
2[[
n
2 : n
i
T / T[.
For i > 0 we have, by the induction hypothesis [T T

n
i
2[ 2
n
i

j<i
1
4(j+1)
2
. For
i = 0 we have (T T

)
n
i
2 = since n
0
= 0 and T

. [S
n
2[ f(n) = i and
[
n
2 : n
i
T / T[
2
n
4(i+1)
3
, by (7). Thus
[(T T

)
n
2[ 2
nn
i
2
n
i

j<i
1
4(j+1)
2
+ (i + 1)
2
n
4(i+1)
3
2
n

j<i+1
1
4(j+1)
2
which is
what we had to show.
(a)(c). Most of the proof follows that of Lemma 10. We need also the following Lemma
14, which will be proved later. Let f be a corset.
14. Lemma. There is an innite sequence 0 = n
0
< n
1
< n
2
< . . . and a tree T such
that for every i f(n
i+1
) > (i + 1) 2
i+1
+ 1 and
(B1) For each T
n
i
2 we have [T
[]

n
i+1
2[ = 2
(n
i+1
n
i
)
(1 2
(i+1)
).
(B2) If ,
0
, . . . ,
k1

n
i
2,
+
0
, . . . ,
+
k1

n
i+1
2,
+
j
,=
+
l
for j < l < k, +
l
T,

+
l
for l < k then

+
:
+

n
i+1
2, (l < k)(
+
+
+
l
T)

2
n
i+1
n
i
_
1 2
(i+1)
_
k1
.
7
4
4
5


r
e
v
i
s
i
o
n
:
1
9
9
4
-
0
6
-
2
2







m
o
d
i
f
i
e
d
:
1
9
9
4
-
0
6
-
2
2


Let n
i
: i ) and T be as in Lemma 14. As in the proof of Lemma 10 we get
(LimT) > 0. Let T

and Y
,
be as in Lemma 9 and let Y
,,
, S and z be as in the
proof of Lemma 10. All we have to do is to show that S is almost of width f. Let us x
, and . We shall now see that
(9)
If

T
[]

n
i
2 then
[
+
:

+
T

n
i+1
2[/2
(n
i+1
n
i
)
(1 2
(i+1)
)
|S
n
i
2|1
Let
+
T
[]

n
i+1
2, then, by the denition of S (see (3)), if
+
S
n
i+1
2 then

+
+
+
+ z T. Thus

+
:

+
T

n
i+1
2
+
:

n
i+1
2, (
+
S)
+
+
+
+ z T.
Let us take in (B2) =

, k = [S
n
i
2[,
l
: l < k = S
n
i
2,
+
l
: l < k S
n
i+1
2,
and for l < k
+
l
n
i
=
l
,
l
=
l
+ z,
+
l
=
+
l
+ z, hence
l
=
+
l
n
i
for l < k. Since
for l < k
+
l
+ z =
+
l
S
n
i+1
2 we have

+
:

n
i+1
2, (
+
S)(
+
+
+
+ z T

+
:

n
i+1
2, (l < k)(
+
l
+
+
T),
therefore by (B2)
[
+
:

n
i+1
2, (
+
S)(
+
+
+
+ z T[ 2
n
i+1
n
i
(1 2
(i+1)
)
|S
n
i
2|1
,
which establishes (9).
(9) tells us how T

grows from the level n

i
to the level n
i+1
and therefore
[T

n
i
2[ 2
n
i

j<i
(1 2
(j+1)
)
|S
n
j
2|1
.
Let c
0
= (LimT

). We know that c
0
> 0 and we can assume c
0
< 1. Then
< log c
0
log([T

n
i
2[ 2
n
i
)

j<i
_
log(1 2
(j+1)
) ([S
n
j
2[ 1)
_
. Since
log(1 x)
1
2
x we get

j<i
2
(j+2)
([S
n
i+1
2[ 1) log
1
c
0
. We shall denote 4 log
1
c
0
by c, so

j<i
2
j
([S
n
j
2[ 1) c, and for every j 2
j
([S
n
j
2[ 1) c, hence
[S
n
j
2[ c 2
j
+ 1. For j > c we have, by our choice of the n
i
s,
f(n
j
) > j 2
j
+ 1 > c 2
j
+ 1 [S
n
j
2[, hence S is almost of width f.
Lemma 14 follows immediately from the following Lemma.
15. Lemma. For every n and 0 < p < 1 there is an N > n such that for every
n

N and t
n
2 there is a t

2 which satises the following (i)(iii).

(i) For each t

n t.
(ii) For each t [t
[]
[ 2
n

n
p.
(iii) If 0 < k 2
n
,
0
, . . . ,
k1

n
2,
+
0
, . . . ,
+
k1

n

2,
+
j
,=
+
l
for j < l < k,
+
l
t,
l
=
+
l
n for l < k then [
+
:
+

2, (l < k)
+
+
+
l
t

[ 2
n

n
p
k1
.
Proof. We shall prove the lemma by the probabilistic method. Let n

> n and let

A =
+

2 :
+
n t. We construct a subset A

of A as follows. We take a coin

which yields heads with probability p. For each
+
A we toss this coin and we put
+
in A

i the coin shows heads. We shall see that if we take t

= A

then, for suciently

large n

, the probability that (ii) holds has a positive lower bound which does not depend
on n

while the probability that (iii) holds is arbitrarily close to 1. Hence there is an N
and a t

as claimed by the lemma. We prove rst two lemmas.

Lemma 16. For k, ,
0
, . . . ,
k1
,
+
0
, . . . ,
+
k1
as im Lemma 15 there are reals c
1
, c
2
> 0
8
4
4
5


r
e
v
i
s
i
o
n
:
1
9
9
4
-
0
6
-
2
2







m
o
d
i
f
i
e
d
:
1
9
9
4
-
0
6
-
2
2


which depend only on p, n and k such that
Pr
_
[
+
:
+

2,

l<k

+
+
+
l
A[ p
k1
2
n

n
_
< c
1
e
c
2
2
n

Proof. We denote 2
n

n
with m. We set (
n

2)
[]
=
+
j
: j < m. Let G be the graph on
m given by
iGj i
+
i
+
+
l
: l < k
+
j
+
+
l
: l < k ,=
Obviously each i < m has at most k
2
neighbors in G hence, by a well known theorem,
m can be decomposed into k
2
+ 1 pairwise disjoint sets B
0
, . . . , B
k
2 such that for every
i k
2
if j, l B
i
and j ,= l then jGl does not hold. Let d <
1
2
minp
l1
p
l
: l 2
n
=
1
2
p
2
n
1
(1 p) > 0.
(10)
Pr
_
[j < m :

l<k

+
j
+
+
l
A

[ m p
k1
_
Pr
_
[j < m :

l<k

+
j
+
+
l
A

[ > m(p
k
+ d)
_
since p
k
+ d < p
k1
Assume that
(11)
for every i k
2
such that [B
i
[
dm
2k
2
+ 2
we have [j B
i
:

l<k

+
j
+
+
l
A

[B
i
[(p
k
+
d
2
)
then
j < m :

l<k

+
j
+
+
l
A

_
ik
2
, |B
i
|
dm
2k
2
+2
j B
i
:

l<k

+
j
+
+
l
A

_
ik
2
, |B
i
|<
dm
2k
2
+2
B
i
hence
[j < m :

l<k

+
j
+
+
l
A

ik
2
, |B
i
|
dm
2k
2
+2
[j B
i
:

l<k

+
j
+
+
l
A

[ +

ik
2
, |B
i
|<
dm
2k
2
+2
[B
i
[

ik
2
, |B
i
|
dm
2k
2
+2
[B
i
[(p
k
+
d
2
) +

ik
2
, |B
i
|<
dm
2k
2
+2
[B
i
[, by (11)
m(p
k
+
d
2
) + (k
2
+ 1)
dm
2k
2
+ 2
= m(p
k
+ d)
9
4
4
5


r
e
v
i
s
i
o
n
:
1
9
9
4
-
0
6
-
2
2







m
o
d
i
f
i
e
d
:
1
9
9
4
-
0
6
-
2
2


Therefore the event [j < m :
_
l<k

+
j
+
+
l
A

[ > m(p
k
+d) is incompatible with (11),
so we continue the inequality (10) by
(12)
Pr
_

ik
2
, |B
i
|
dm
2k
2
+2
_
[j B
i
:

l<k

+
j
+
+
l
A

[ > [B
i
[(p
k
+
d
2
)
_
_

ik
2
, |B
i
|
dm
2k
2
+2
Pr
_
[j B
i
:

l<k

+
j
+
+
l
A

[ > [B
i
[(p
k
+
d
2
)
_
For a xed j < m the events
+
j
+
+
l
A

for dierent ls are independent hence

Pr
_
_
l<k

+
j
+
+
l
A

_
= p
k
. For a xed i the events
_
l<k

+
j
+
+
l
A

for dif-
ferent js in B
i
are independent since, by the denition of the B
i
s, if j
1
, j
2
B
i
and
j
1
,= j
2
then
+
j
1
+
+
l
1
,=
+
j
2
+
+
l
2
. We have here [B
i
[ independent events, each with proba-
bility p
k
. By a formula of probability theory (see, e. g., the formula Pr [X > a] < e
2a
2
/n
in Spencer [2], p. 29)
Pr
_
j B
i
:

k<l

+
j
+
+
l
A

[ > [B
i
[p
k
+
_
< e

2
2
|B
i
|
and taking =
1
2
[B
i
[d we get
Pr
_
j B
i
:

k<l

+
j
+
+
l
A

[ > [B
i
[(p
k
+
d
2
)
_
< e

d
2
|B
i
|
2
Continuing (12) we get

ik
2
, |B
i
|
dm
2k
2
+2
e

d
2
|B
i
|
2

ik
2
, |B
i
|
dm
2k
2
+2
e

d
2
2
dm
2k
2
+2
(k
2
+ 1)e

d
3
2
n

n
4k
2
+4
Combining this with the inequalities (10) and (12) we get
Pr
_
[
+
:
+

2,

l<k

+
+
+
l
A

[ p
k1
2
n

n
_
< (k
2
+ 1)e

d
3
2
n

n
4k
2
+4
= (k
2
+ 1)e

d
3
2
n
2
n

4k
2
+4
Since d =
1
2
p
2
n
1
(1 p) this proves Lemma 16.
17. Lemma. There are c
3
, c
4
which depend only on p and n such that
(13)
Pr
_

k,,
0
,...,
k1
,
+
0
,...,
+
k1
[
+
:
+

2, (l < k)
+
+
+
l
A

[ 2
n

n
p
k1
_
c
3
(2
n

)
2
n
e
c
4
2
n

10
4
4
5


r
e
v
i
s
i
o
n
:
1
9
9
4
-
0
6
-
2
2







m
o
d
i
f
i
e
d
:
1
9
9
4
-
0
6
-
2
2


where k, ,
0
, . . . ,
k1
,
+
0
, . . . ,
+
k1
are as in (iii) of Lemma 15.
Proof. By our requirements on k, ,
0
, . . . ,
k1
,
+
0
, . . . ,
+
k1
there are at most 2
n
possi-
ble ks and s and (2
n

)
2
n
sequences
+
0
, . . . ,
+
k1
), while
0
, . . . ,
k1
are determined by

+
0
, . . . ,
+
k1
and n. Therefore we get, by Lemma 16,
Pr
_

k,,
0
,...,
k1
,
+
0
,...,
+
k1
_
[
+
:
+

2, (l < k)
+
+
+
l
A

[ 2
n

n
p
k1
_
_

k,,
0
,...,
k1
,
+
0
,...,
+
k1
Pr
_
[
+
:
+

2, (l < k)
+
+
+
l
A

[ 2
n

n
p
k1
_
2
n
2
n
(2
n

)
(2
n
)
c
1
e
c
2
2
n

Proof of Lemma 15 (continued). For each

+

2 such that
+

+
A

if the
coin shows heads and dierent tosses are independent [A
[]
[ is a binomial random variable
with expectation 2
n

n
p. By the central limit theorem of probability theory (see, e. g.,
Feller [1, Ch. 7]) the limit, as n

, of Pr
_
[A
[]
[ 2
n

n
p
_
is
_

0
1

2
e

x
2
2
dx =
1
2
,
hence there is an N such that for every n

N Pr
_
[A
[]
[ 2
n

n
p
_

1
3
. For dierent
t the random variables [A
[]
[ are idependent, hence
(14) Pr
_

t
([A
[]
[ 2
n

n
p)
_

1
3
|t|

1
3
2
n
The right-hand side of (13) clearly vanishes as n , let us take N to be such that for
n

N the right-hand side of (13) is < 3

2
n
. Therefore we have, by (13) and (14),
(15)
Pr
_

k,,
0
,...,
k1
,
+
0
,...,
+
k1
([
+
:
+

2, (l < k)
+
+
+
l
A

[ < 2
n

n
p
k1
)

t
([A
[]
[ 2
n

n
p)
_
> 0
By (15) there is a t

as required by the lemma.

4 Characterization of the meager-additive sets
18. Theorem. For every X

2 the following conditions are equivalent:
a. X is meager additive.
b. For every sequence n
0
< n
1
< n
2
< . . . of natural numbers there is a sequence
i
0
< i
1
< . . . of natural numbers and a y

2 such that for every x X and for every
suciently big k < there is an l [i
k
, i
k+1
) such that x [n
l
, n
l+1
) = y [n
l
, n
l+1
).
Proof. Throughout this proof, if x

2
>
2, k, l and k < l then x [k, l) will
denote the sequence
lk
2 such that (i) = x(k + i) for all i < l k.
11
4
4
5


r
e
v
i
s
i
o
n
:
1
9
9
4
-
0
6
-
2
2







m
o
d
i
f
i
e
d
:
1
9
9
4
-
0
6
-
2
2


(b)(a). In order to prove (a) it clearly suces to show that X + LimT is meager for
every nowhere dense tree T.
For a nowhere dense tree T let n
i
: i < ) be an ascending sequence of natural
numbers such that n
0
= 0 and for every i there is a sequence
i

n
i+1
n
i
2 such
that for every
n
i
2
i
/ T. Let i
j
: j < ) and y be as in (b), then, by (b),
X =

k
X
k
where X
k
= x X : (m k)(l [i
m
, i
m+1
))x[n
l
, n
l+1
) = y(n
l
, n
l+1
).
It clearly suces to prove that X
k
+ LimT is nowhere dense.
Let
n
i
m
2 for some m k; we shall show that has an extension which is not in
X
k
+ LimT. Let =
i
m

i
m
+1
. . .
i
m+1
1
and let = y [n
i
m
, n
i
m+1
) + . We
show that no extension z of is in X
k
+LimT. Suppose z X
k
+LimT then
z = x+w, x X
k
w LimT. Therefore =
1
+
2
and =
1
+
2
such that
1

1
x
and
2

2
w, hence
2

2
T. Let
n
i
m
2 be such that (j) = 0 for every j < n
i
m
,
and let

= ,

1
=
1
,

2
=
2
. Clearly

1
+

2
. Since x X
k
there is,
by (b), an l [i
m
, i
m+1
) such that x [n
l
, n
l+1
) = y [n
l
, n
l+1
). Since
1

1
x we have

1
[n
i
m
, n
i
m
+1
) = x [n
i
m
, n
i
m
+1
) and hence
1
[n
l
, n
l+1
) = x [n
l
, n
l+1
) = y [n
l
, n
l+1
)
Therefore, by the denition of and
y [n
l
, n
l+1
) +

2
[n
l
, n
l+1
) =

1
[n
l
, n
l+1
) +

2
[n
l
, n
l+1
) =

[n
l
, n
l+1
)
= y [n
l
, n
l+1
) +
l
hence

2
[n
l
, n
l+1
) =
l
. By the denition of
l

2

2
/ T, contradicting
2

2
T.
(a)(b). Let X be meager-additive. Let n
i
: i < ) be an ascending sequence of natural
numbers. Let B = x

2 : j(k [n
j
, n
j+1
)) x(k) ,= 0 and T = x n : x B, n .
Clearly B = LimT is nowhere dense, so X + LimT is meager, hence there are nowhere
dense trees S
n
, n such that for every n S
n
S
n+1
and X + LimT

n
S
n
. We
dene now the sequences i
l
: l < ), which is an ascending sequence of natural numbers,
and
l
: l < ) by recursion as follows. i
0
= 0. Given i
l
let
l
and i
l+1
be such that

l

n
i
l+1
n
i
l
2 and for every
n
i
l
2
l
/ S
l
; there are such
l
and i
l+1
since S
l
is
nowhere dense. Let y

2 be given by y [n
i
l
, n
i
l+1
) =
l
for every l < . We shall now
prove that i
l
: l < ) and y are as required by (b).
Let x X, so Lim(x + T) = x + LimT X + LimT

n
S
n
. Therefore,
by Lemma 7 (where we take x + T for S) there is an T and n such that
x + T
[]
S
n
. Let k be such that k n and i
k
length . By x + T
[]
S
n
we have
x n
i
k+1
+ (T
[]

n
i
k+1
2) S
n
S
k
. Thus for every T
[]

n
i
k+1
2 x n
i
k+1
+ S
k
,
hence, by the denition of
k
and y, x[n
i
k
, n
i
k+1
)+[n
i
k
, n
i
k+1
) ,=
k
= y[n
i
k
, n
i
k+1
) and
therefore x[n
i
k
, n
i
k+1
)y[n
i
k
, n
i
k+1
) ,= [n
i
k
, n
i
k+1
), i.e., x[n
i
k
, n
i
k+1
)y[n
i
k
, n
i
k+1
) /
[n
i
k
, n
i
k+1
) : T
[]
. Since i
k
> length this can happen, by the denition of T,
only if for some i
k
j < i
k+1
x [n
j
, n
j+1
) y [n
j
, n
j+1
) is identically zero, and this is
what we had to prove.
5. An uncountable null-additive set.
19. Theorem. If the continuum hypothesis holds then there is an uncountable null-
additive set.
12
4
4
5


r
e
v
i
s
i
o
n
:
1
9
9
4
-
0
6
-
2
2







m
o
d
i
f
i
e
d
:
1
9
9
4
-
0
6
-
2
2


Proof. Let f

: <
1
) be a sequence containing all corsets and let T

: <
1
) be
a sequence containing all perfect trees. Let E be the set of all limit ordinals <
1
such
that for every , < and n < there is a < such that
T

, T

n
2 = T

n
2 and for all m [T

m
2[ max ([T

m
2[, f

(m))
Clearly E is closed. For every , <
1
there is a perfect tree T such that T T

,
T
n
2 = T

n
2 and for all m < [T
m
2[ max ([T

n
2[, f

(m)). This tree T is T

for some <

1
. By a simple closure argument this implies that E is unbounded.
We need now the following lemma which will be proved later.
20. Lemma. There is an increasing and continuous sequence

: <
1
) of ordinals
in E such that for every <
1
, k < and <

there is an ordinal which is good for

(, , k), where by is good for (, , k) we mean that
(16)
(i) <
+1
(ii) T

, T

k
2 = T

k
2
(iii) for all such that

> and for every <

, there is a <

such that T

and T

is almost of width f

For <
1
let

be the which is good for (, 0, 0). We choose

LimT

: < ,
and let X =

: <
1
. X is clearly uncountable. We shall prove that X is null-additive
by proving that X satises condition (c) of Theorem 13. For a given corset f f = f

for
some <
1
. Let <
1
be such that

> . Let Z = <

+1
: T

is almost of width f

.
We shall see that
X

Z
LimT

. Since Z and are countable condition (c) of Theorem 13

holds.
Let > , it suces to prove that

LimT

for some Z. <

and since

is good for = k = 0 hence there is a <

such that T

, and T

is of width f

.
Thus Z and

LimT

LimT

.
Proof of Lemma 20. We dene

: <
1
) as follows.
0
is the least member of E.
For a limit ordinal

=

<

. Since

E for < also

E. We shall now
dene
+1
. We shall assume, as an induction hypothesis, that for each < the lemma
holds. For each <

and k < we shall nd a (, k) which is good for (, , k) and

we shall choose
+1
to be the least member of E greater than all these (, k)s.
First we shall show that what the lemma claims holds for the case where is a successor
or 0. Whenever we shall write 1 we shall assume that is a successor. Let <

and
k < be given, and let
n
: n < = : <

. We dene sequences
n
: n < ) and
k
n
: n < ) so that
(a) k
0
= k. If = 0 or <
1
then
0
= . If
1
then
0
is an ordinal which is
good for ( 1, , k). In any case
0
<

, T

0
T

and T

0

k
2 = T

k
2.
(b)
n+1
<

.
(c) T

n+1
T

n
.
(d) T

n+1

k
n
2 = T

n

k
n
2.
(e) T

n+1
is almost of width f

n
.
13
4
4
5


r
e
v
i
s
i
o
n
:
1
9
9
4
-
0
6
-
2
2







m
o
d
i
f
i
e
d
:
1
9
9
4
-
0
6
-
2
2


(f) k
n+1
> k
n
and every T

n+1

k
n
2 has at least two extensions in T

n+1

k
n+1
2.
There are indeed such sequences
n
: n < ) and k
n
: n < ). (a) determines k
0
and
0
;
if <
1
then there is an
0
as in (a) by the induction hypothesis.

in E and let us
take
n
,
n
, k
n
,
n+1
for , , n, in the denition of E, then

E says that there is an

n+1
which satises (b)(e). Since T

n+1
is perfect there is a k
n+1
as in (f).
Let T =

n
T

n
. By (c),(d),(f) T is a perfect tree, hence it is T

for some <

1
.
Since T, and therefore also , depend on and k we denote with (, k). As is easily
seen T
(,k)
T

, T
(,k)

k
2 = T

k
2, and for every <

T
(,k)
T

l+1
T

,
where l is such that =
l
. This means that (iii) of (16) holds for = . We shall have to
show that (iii) holds for < and to deal with the case where is a limit ordinal.
If is a limit ordinal let
n
: n < ) be an increasing sequence such that

0
> and

n<

n
= . We construct the sequences
n
: n < ) and k
n
: n < ) as in the case
where is a successor, except that (a),(b),(e) are replaced by
(a

) k
0
= k,
0
= .
(b

)
n
<

n
.
(e

)
n+1
is good for (
n
, , k).
By the induction hypothesis that the lemma holds for the
n
s there are indeed such
sequences
n
: n < ) and k
n
: n < ). Let T =

n<
T

n
. As above, T is a perfect
tree and T = T
(,k)
, T
(,k)
T

and T
(,k)

k
2 = T

k
2.
We shall now see that for both cases of with which we are dealing (iii) holds for
< . If is a successor then 1 and since
o
is, by (a), good for ( 1, , k)
there is a <

such that T

0
T

and T

is almost of width f

. Note that if
<
1
then, by the induction hypothesis, we have a <

which is good for (, , k),

and if = 0 then (iii) holds vacuously, hence we may assume that > 0 and [
1
,

).
Since T
(,k)
T

0
is as required by (iii). If is a limit ordinal then
n
for some
n < . Since
n+1
is good for
n
then there is a <

such that T

n+1
T

and
T

is almost of width f

. Since T
(,k)
T

n+1
is as required by (iii).
The only case left is that where is a limit ordinal and = in (iii). Since , <
also , <
n
for some n < .
n+1
is good for
n
hence there is a <

n
such that
T

n+1
T

and T

is almost of width f

. Since T
(,k)
T

n+1
and
n
< is as
required by (iii).
Bibliography
[1] William Feller, An Introduction to Probability Theory and its Applications. Wiley,
New York & London, 1950.
[2] Joel Spencer, Ten Lectures on the Probabilistic Method. CBMS-NSF Conference Series
in Applied Mathematics. SIAM 1987.
14