Professional Documents
Culture Documents
0
6
r
e
v
i
s
i
o
n
:
1
9
9
4
-
1
1
-
2
4
m
o
d
i
f
i
e
d
:
1
9
9
4
-
1
1
-
2
4
amsppt.styFSp90a.tex
Pointwise compact and stable sets
of measurable functions
S.Shelah & D.H.Fremlin
Hebrew University, Jerusalem
University of Essex, Colchester, England
[University of Essex Mathematics Department Research Report 91-3]
Introduction In a series of papers culminating in [Ta84], M.Talagrand, the second author and others
investigated at length the properties and structure of pointwise compact sets of measurable functions. A
number of problems, interesting in themselves and important for the theory of Pettis integration, were solved
subject to various special axioms. It was left unclear just how far the special axioms were necessary. In
particular, several results depended on the fact that it is consistent to suppose that every countable relatively
pointwise compact set of Lebesgue measurable functions is stable in Talagrands sense; the point being that
stable sets are known to have a variety of properties not shared by all pointwise compact sets. In the present
paper we present a model of set theory in which there is a countable relatively pointwise compact set of
Lebesgue measurable functions which is not stable, and discuss the signicance of this model in relation to
the original questions. A feature of our model which may be of independent interest is the following: in it,
there is a closed negligible set Q [0, 1]
2
such that whenever D [0, 1] has outer measure 1 then
Q
1
[D] = x : y D, (x, y) Q
has inner measure 1 (see 2G below).
1. The model We embark immediately on the central ideas of this paper, setting out a construction of a
partially ordered set which forces a fairly technical proposition in measure theory (1S below); the relevance
of this proposition to pointwise compact sets will be discussed in 2. The construction is complex, and
rather than give it in a single stretch we develop it cumulatively in 1E, 1I, 1Q below; it is to be understood
that each notation introduced in these paragraphs, as well as those in the denitions 1A, 1K, 1L, is to stand
for the remainder of the section. After each part of the construction we give lemmas which can be dealt
with in terms of the construction so far, even if their motivation is unlikely to be immediately clear.
When we come to results involving Forcing, we will try to follow the methods of [Ku80]; in particular, in
a p.o.set, p q will always mean that p is a stronger condition than q.
1A Denition If / is any family of sets not containing , set
dp(/) = min#(I) : I A ,= A /.
Observe that dp(/) = 0 i / = and that dp(/B) is at most the cardinal sum of dp(/) and dp(B). (Of
course much more can be said.)
1B Lemma Suppose that n, l, k N, with n, l not less than 2, and that is such that 0 <
1
2
and
l
k
(k + 2) ln n. Then there is a set W n n (we identify n with the set of its predecessors) such that
#(W) n
2
and whenever I [n]
l
and J
0
, . . . , J
l1
[n]
k
are disjoint, there are i I, j < l such that
i J
j
W.
proof If k = 0 this is trivial; suppose that k > 0. Set = T(n n). Give a probability for which the
events (i, j) W, as (i, j) runs over n n, are independent with probability . If W is a random set,
then
Pr(#(W) n
2
) >
1
4
because
1
2
and #(W) has the binomial distribution B(n
2
, ). On the other hand, if J [n]
k
and i < n,
Pr(i J W)
k
. So if I [n]
l
and J
0
, . . . , J
l1
are disjoint members of [n]
k
,
1
4
0
6
r
e
v
i
s
i
o
n
:
1
9
9
4
-
1
1
-
2
4
m
o
d
i
f
i
e
d
:
1
9
9
4
-
1
1
-
2
4
2
Pr(i J
j
, W i I, j < l) (1
k
)
l
2
exp(l
2
k
).
Accordingly
Pr( I [n]
l
, disjoint J
0
, . . . , J
l1
[n]
k
such that i J
j
, W i I, j < l)
#([n]
l
)#([n]
k
)
l
exp(l
2
k
)
n
l
n
kl
exp(l
2
k
)
= exp((k + 1)l ln n l
2
k
)
1
4
because
l
2
k
(k + 1)l ln n l ln n 2 ln2.
There must therefore be some W of the type required.
Remark Compare the discussion of cliques in random graphs in [Sp87], pp. 18-20.
1C Lemma Let m and l be strictly positive integers and / a non-empty family of non-empty sets. Let
T be the family of non-empty sets T /
m
. For T T write T
= tj : t T , j m
jm
/
j
. For
T , T
0
T say that T T
0
if T T
0
and
dp(u : t
u T
) dp(u : t
u T
0
)/2l
for every t T
T . Fix T
0
T and a cover o
i
)
i<2l
of T
0
. Then there is a T T
0
such that T o
i
for
some i < 2l.
[Notation: In this context we use ordinary italics, u, for members of /, and bold letters, t, for nite
sequences of members of /.]
proof For t T
0
T
0
set
t
= dp(u : t
u T
0
)/2l > 0.
For i < 2l dene o
(j)
i
)
jm
by setting o
(m)
i
= o
i
T
0
,
o
(j)
i
= t : t /
j
T
0
, dp(u : t
u o
(j+1)
i
)
t
for j < m. An easy downwards induction (using the fact that dp is subadditive) shows that T
0
/
j
=
i<2l
o
(j)
i
for every j m. In particular, there is some i < 2l such that o
(0)
i
. Now dene T by
T = t : t /
m
, tj o
(j)
i
j m T
0
o
i
,
and see that T T
0
, as required.
1D Corollary Let n, l, k and W be as in Lemma 1B. Take r k, let Z be the cartesian product n
r
and
set
j
= t : (H(t), z
j
)
W for j < l.
4
0
6
r
e
v
i
s
i
o
n
:
1
9
9
4
-
1
1
-
2
4
m
o
d
i
f
i
e
d
:
1
9
9
4
-
1
1
-
2
4
3
By Lemma 1C, there is a T T
0
such that either T o
i
for some i I or T o
j
for some j < l. But we
cannot have T o
j
, because z
j
/ A; so T o
i
for some i, and we have the rst alternative.
Remark 1C-1D are of course elementary, but their signicance is bound to be obscure; they will be used
in 1R below. An essential feature of 1C is the fact that the denominator 2l in the denition of is
independent of the size of /.
1E Construction: part 1 (a) Take a sequence n
k
)
kN
of integers increasing so fast that
(i) n
0
4;
(ii) n
k
> 2
k+1
;
(iii) writing c
l
=
i<l
2
ni
, then
ln(2
k1
n
k
) 2
l
(k + 1)( c
l+1
l
ln 2 + c
l1
l
ln n
l1
) for 1 l k;
(iv) ln n
k
(2
k+1
)
k
(k + 2);
(v) writing | for the least integer greater than or equal to ,
(k + 1) ln(2(ln n
k
)
2
|) 2
k
ln(2
k1
n
k
)|;
(vi) 2
k
k(ln n
k
)
2
|(
i<k
2
ni
)
k+1
n
k
;
(vii) ln(2
k1
n
k
) (k + 1) ln(2 c
k+1
k
+ 2k)
for every k N.
For each k N, let V
k
be the cartesian product
i<k
n
i
.
(b) For each k N, let T
k
be the set of those subsets t of V
k
expressible as t =
i<k
C
i
(t) where C
i
(t) n
i
and #(C
i
(t)) (1 2
i1
)n
i
for each i < k. Set T =
kN
T
k
, and for t T say that rank(t) = k if t T
k
.
For t, t
T say that t t
if rank(t) rank(t
) and C
i
(t) = C
i
(t
: t
t or t t
, suc(t) for t
: t t
, rank(t
) = rank(t) + 1.
(c) For k N, set
k
= (k + 1)/ ln(2
k1
n
k
|);
2
k1
n
k
> 1 by (a)(ii) above. For t T dene d
t
: TT R by writing
d
t
(S) =
rank(t)
ln(dp(C : t C S))
for every S T, allowing d
t
(S) = if S suc(t) = . Observe that d
t
(T) k + 1 whenever rank(t) = k
(because
dp(C : C n
k
, #(C) (1 2
k1
)n
k
) 2
k1
n
k
|.)
(d) Let Q be the set of subtrees q T such that
q ,= ;
if t t
q then t q;
if t q then q suc(t) ,= ;
writing
k
(q) = mind
t
(q) : t q T
k
for k N, lim
k
k
(q) = .
Observe that
k
(T) k + 1 for each k N, so that T Q and Q ,= .
(e) For q, q
Q say that q q
if q q
if q q
and q T
k
= q
T
k
and
d
t
(q) min(k, d
t
(q
)) 2
k
for every t q. Note that
k
is not transitive unless k = 0.
Remarks Of course the point of the sequence n
k
)
kN
on which the rest of this construction will depend
is that it increases as fast as we need it to. The exact list given in (a) above is of no signicance and will
be used only as a list of clues to the (elementary) arguments below which depend on the rapidly-increasing
nature of the sequence. This is why we have made no attempt to make the list as elegant or as short as
possible.
Three elements may be distinguished within the construction of Q. First, it is a p.o.set of rapidly
branching trees; that is, if t q Q, q suc(t) is large compared with T
rank(t)
, except for t of small rank.
This is the basis of most of the (laborious but routine) work down to 1P below. Second, there is a natural
Q-name for a subset of X =
kN
n
k
of large measure; a generic lter in Q leads to a branch of T and
4
0
6
r
e
v
i
s
i
o
n
:
1
9
9
4
-
1
1
-
2
4
m
o
d
i
f
i
e
d
:
1
9
9
4
-
1
1
-
2
4
4
hence to the of 1Q(d). Third, the use of dp in the denition of rapidly branching ((c)-(d) above) is what
makes possible the side-step in the last part of the proof of 1R.
1F Lemma Q is proper.
proof This is a special case of Proposition 1.18 in [Sh326]. (In fact, the arguments of 1G-1H below show
that Q satises Axiom A, and is therefore proper; see [Ba84], 2.4.)
1G Lemma Let k N and let be an ordinal. Suppose that A is a set with #(A) exp(2
k
/
j
) 1
for every j k, and that is a Q-name for a member of A. Let be a Q-name for a countable subset of
. Then for every q Q there are a q
k
q, a function H : T
k
A and a countable (ground-model) set
D such that
q
T
(t)
Q
= H(t) t q
T
k
,
q
Q
D.
proof (a) Set m = #(A). The point is that if j k and t T
j
and S
i
)
im
is a family of subsets of T,
then
d
t
(
_
im
S
i
) =
j
ln(dp(
_
im
C : t C S
i
))
j
ln(
im
dp(C : t C S
i
))
j
ln((m + 1) max
im
dp(C : t C S
i
))
=
j
ln(m + 1) + max
im
j
ln(dp(C : t C S
i
))
2
k
+ max
im
d
t
(S
i
).
(b) For each a A, let S
a
be the set
t : t q, rank(t) k, p Q, D []
,
p
k
q T
(t)
, p
Q
= a & D.
If t q S
a
and rank(t) k, then d
t
(S
a
) < min(k, d
t
(q)) 2
k
. For if S
a
suc(t) = , d
t
(S
a
) = . While
if S
a
suc(t) ,= , then for each s suc(t) S
a
we can nd p
s
Q and D
s
[]
such that p
s
k
q T
(s)
,
p
s
Q
= a and p
s
Q
D
s
. If we now set
p =
ssuc(t)Sa
p
s
, D =
ssuc(t)Sa
D
s
,
then p q T
(t)
and p
Q
= a and p
Q
D. Because t / S
a
, p ,
k
q T
(t)
and there must be an
s p such that d
s
(p) < min(k, d
s
(q T
(t)
)) 2
k
; evidently s = t and
d
t
(S
a
) < min(k, d
t
(q)) 2
k
,
as claimed.
(c) Suppose, if possible, that there is a t
0
q T
k
aA
S
a
. Set
p = t : t q T
(t0)
, t
aA
S
a
t
t.
Then p is a subtree of T. For every t p with t t
0
,
d
t
(q) max(d
t
(p) d
t
(S
a
) : a A) + 2
k
because #(A) = m. But d
t
(S
a
) < d
t
(q) 2
k
for every a A, by (b) above, so d
t
(p) d
t
(q) 2
k
(and
p suc(t) ,= ). This shows both that p has no maximal elements and that
i
(p)
i
(q) 2
k
for every
i k, so that p Q. Because Q is proper, we can nd a p
D
([Sh82], p. 81, III.1.16). Next, there are p
, a A such that p
Q
= a. Let j N be such that
i
(p
T
(t)
witnesses that
t S
a
; which is impossible.
(d) Accordingly we have for every t q T
k
an H(t) A, a countable set D
t
and a p
t
Q such that
p
t
Q
= H(t) & D
t
,
p
t
k
q T
(t)
.
4
0
6
r
e
v
i
s
i
o
n
:
1
9
9
4
-
1
1
-
2
4
m
o
d
i
f
i
e
d
:
1
9
9
4
-
1
1
-
2
4
5
Set q
tqT
k
p
t
, D =
tqT
k
D
t
; then q
k
q, D is countable, q
Q
D and q
T
(t)
Q
= H(t)
for every t q T
k
.
1H Lemma Let q
k
)
kN
be a sequence in Q such that q
2k+2
k+1
q
2k+1
k
q
2k
for every k N. Then
q =
kN
q
k
belongs to Q and is accordingly a lower bound for q
k
: k N in Q; also qT
k+1
= q
2k+1
T
k+1
for each k N.
proof Because each q
k
is a nitely-branching subtree of T with no maximal elements, so is q, and d
t
( q) =
lim
k
d
t
(q
k
) for every t q. Moreover, if t q and k l N,
d
t
(q
2l
) min(k, d
t
(q
2k
)) 3.2
k
+ 3.2
l
,
(induce on l, using the denition of
l
), so we have
i
( q) = lim
l
i
(q
2l
) min(k,
i
(q
2k
)) 3.2
k
for every i, k N; consequently lim
i
i
( q) = and q Q. Now if k N and 2k + 1 l, q
l+1
T
k+1
=
q
l
T
k+1
, so q T
k+1
= q
2k+1
T
k+1
.
amsppt.styFSp90b.tex Version of 16.9.92
1I Construction: part 2 Let be the cardinal c
+
(evaluated in the ground model).
(a) Let (P
, Q
)
<
) be a countable-support iteration of p.o.sets, as in [Ku80], chap. 8, such that
each Q
is a P
.
(b) If , K []
<
, k N and p, p
belong to P
, say that p
K,k
p
if p p
and
p
P
p()
k
p
() K,
taking
k
here to be a P
, where P
is
a P
, as the denition of P
in V .
(c) For every < ,
11
P
< .
(d) If has uncountable conality, A is a (ground-model) set,
f is a P
p and a P
f = g.
(e) If A is a (ground-model) set and
f is a P-name for a sequence in A, then we can nd a < and a
P
) where p P
and p
is a P
) P
, we have a P
-name
J for the support of p
which in V
P
is a
countable subset of . But because P
). Thus P
+
is dense in P
, as claimed.
(c) [Sh82], p. 96, III.4.1.
(d) [Sh82], p. 171, V.4.4.
(e) By [Sh82], p. 96, III.4.1, P satises the -c.c.; because is regular, (d) gices the result.
1K Denition Let , p P
.
4
0
6
r
e
v
i
s
i
o
n
:
1
9
9
4
-
1
1
-
2
4
m
o
d
i
f
i
e
d
:
1
9
9
4
-
1
1
-
2
4
6
(a) Dene U(p), p
(u)
)
uU(p)
as follows. A nite function u T belongs to U(p) if
either u = , in which case p
(u)
= p,
or u = v (, t) where v U(p), dom(v) < , and
p
(v)
P
t p
(v)
(),
in which case p
(u)
is dened by writing
p
(u)
() = p
(v)
() ,
p
(u)
() = p
(v)
() T
(t)
.
(b) Observe that if u U(p) then p
(u)
() = p() for dom(u), p
(u)
() = p()T
(u())
if dom(u);
U(p) is just the set of nite functions u for which these formulae dene such a p
(u)
P
. Of course p
(u)
p
for every u U(p).
(c) Note that if , p P
, u U(p) and v T is a nite function such that dom(u) dom(v) and u() v() in
T for every dom(u), then v U(p) i v U(p
(u)
), and in this case p
(v)
= (p
(u)
)
(v)
(induce on #(v)).
(e) We shall mostly be using not the whole of U(p) but the sets U(p; K, k) = U(p) T
K
k
for K []
<
,
k N, writing T
K
k
for the set of functions from K to T
k
.
1L Denition For , K []
<
, k N and p P
there are p
1
K,k
p, a function H : T
K
k
A
and a countable (ground-model) set D such that
p
1
is (K, k)-xed,
p
(u)
1
P
= H(u) u U(p
1
; K, k),
p
1
P
D.
proof Induce on m = #(K). If m = 0 we may take any a A, p
1
p such that p
1
= a, and (again
using [Sh82], III.1.16, this time based on 1Ja) a countable D and a p
1
p
1
such that p
1
P
D; now
set H() = a.
For the inductive step to #(K) = m 1, let be max K. As explained in 1Jb, P
may be regarded as a
dense subset of P
+1
P
; arguing momentarily in V
P
+1
we can nd a P
+1
-name r
0
for a member of P
, a
P
+1
-name
,
(p + 1, r
0
)
P
+1
P
= ,
(p + 1, r
0
)
P
+1
P
.
Now let
0
be a P
+1
-name for a countable subset of ( + 1) such that
11
P
+1
P
+1
supp( r
0
) =
0
.
Because #(A) = a < 2
c
m
a
c
m1
exp(2
k
/
i
) for every i k, we can use Lemma 1G in V
P
to nd
H, q,
such that
H is a P
is a P
,
p
P
q
k
p(),
p
P
_
q T
(t)
=
H(t) t q T
k
_
,
p
P
_
q
Q
_
.
Now consider the pair (
H, q T
k
). This can be regarded as a P
: T
K
k
A
T
k
, F
: T
K
k
TT
k
and a countable D such that
p
1
is (K , k)-xed,
p
(u)
1
P
H = H
(u) & q T
k
= F
(u)
for every u U( p
1
; K , k), and
p
1
P
D.
At this point we observe that
p
1
P
_
q
Q
supp( r
0
) D
_
.
Now the only dierence between P
+1
P
and P
dened by
saying that p
1
= p
1
, p
1
P
p
1
() = q, and p
1
P
p
1
() = r
0
() for < < ; then supp(p
1
)
supp( p
1
) (D ) is countable.
Consequently p
1
P
t) = H
(u)(t) for u T
K
k
, t T
k
,
p
1
K,k
p,
p
1
P
D,
U(p
1
; K, k) = u
t : u U( p
1
; K , k), t F
(u),
p
(v)
1
P
= H(v) v U(p
1
; K, k)
and nally
p
(u)
1
P
p
1
() T
k
= F
(u) u U(p
1
; K , k),
so that p
1
is (K, k)-xed, and the induction proceeds.
1N Lemma Suppose , K
k
)
kN
is an increasing sequence of nite subsets of , p
k
)
kN
is a sequence
in P
; suppose that
p
2k+2
K
k
,k+1
p
2k+1
K
k
,k
p
2k
for every k N and that
kN
supp(p
k
)
kN
K
k
. Then there is a p P
such that p p
k
for every
k N, supp( p)
kN
K
k
and
p
P
p() T
k
= p
2k+1
T
k
k N, K
k
,
p
P
p() T
k+1
= p
2k+2
T
k+1
k N, K
k+1
,
so that
U( p; K
k
, k) U(p
2k+1
; K
k
, k) and U( p; K
k
, k + 1) U( p
k+2
; K
k
, k + 1)
for every k N.
proof For each < choose p() such that
11
P
p() =
kN
p
k
().
An easy induction on shows that p P
kN
K
k
then
11
P
p() = T = 11
Q
,
while if k N and K
k
then
p
P
p
2l+2
()
l+1
p
2l+1
()
l
p
2l
() l k,
so that by Lemma 1H,
p
P
p() Q
& p() T
k+1
= p
2k+1
T
k+1
= p
2k+2
T
k+1
& p() T
k
= p
2k+1
T
k
.
It follows at once that U( p; K
k
, k) U(p
2k+1
; K
k
, k), U( p; K
k
, k+1) U(p
2k+2
; K
k
, k+1) for every k N.
1O Lemma Suppose that 0 < , is a P
. Then we
can nd a p and sequences K
k
)
kN
, H
k
)
kN
such that
p P
, p p;
K
k
)
kN
is an increasing sequence of subsets of , #(K
k
) k + 1 for every k, K
0
= 0;
supp( p)
kN
K
k
;
p is (K
k
, k)-xed and (K
k
, k + 1)-xed for every k;
H
k
is a function from T
K
k
k+1
to n
k
for every k;
p
(u)
(k) = H
k
(u) whenever k N and u U( p; K
k
, k + 1).
proof Using Lemma 1M, we can nd sequences p
k
)
kN
, K
k
)
kN
and H
k
)
kN
such that
4
0
6
r
e
v
i
s
i
o
n
:
1
9
9
4
-
1
1
-
2
4
m
o
d
i
f
i
e
d
:
1
9
9
4
-
1
1
-
2
4
8
p = p
0
, K
0
= 0;
#(K
k+1
) k + 2, K
k+1
K
k
;
p
2k+1
K
k
,k
p
2k
, p
2k+1
is (K
k
, k)-xed;
H
k
: T
K
k
k+1
n
k
is a function;
p
2k+2
K
k
,k+1
p
2k+1
, p
2k+2
is (K
k
, k + 1)-xed,
p
(u)
2k+2
P
(k) = H
k
(u) u U(p
2k+2
; K
k
, k + 1)
for every k N. Furthermore, we may do this in such a way that
kN
K
k
includes
kN
supp(p
k
). We
need of course to know that the n
k
are rapidly increasing; specically, that
2
c
k+1
k
exp(2
k
/
i
) i k
(when choosing p
2k+1
) and that
2
c
k+1
k+1
n
c
k
k+1
k
exp(2
k1
/
i
) i k + 1
(when choosing p
2k+2
), where we write c
k
= #(T
k
). But as c
k
i<k
2
ni
, this is a consequence of 1E(a)(i)
and (iii).
Armed with the sequences p
k
)
kN
, K
k
)
kN
we may now use Lemma 1N to nd a p as described there.
Because p
2k+1
is (K
k
, k)-xed and U( p; K
k
, k) U(p
2k+1
; K
k
, k) we must have equality here and p is
(K
k
, k)-xed for every k N. Similarly, p is (K
k
, k+1)-xed for every k. Moreover, if u U( p; K
k
, k+1) =
U(p
2k+2
; K
k
, k + 1) we have p
(u)
p
(u)
2k+2
, so
p
(u)
(k) = H
k
(u)
as required.
1P Lemma Suppose that , p P
, k N, K []
<
and V is a non-empty subset of U(p; K, k).
Then we have a p
1
=
_
vV
p
(v)
dened (up to -equivalence in P
) by saying
if K then p
1
() = p();
if K then
(p
1
)
(u)
p
1
() =
p() T
(t)
: v V such that v + 1 = u
t
for u v : v V.
Now p
1
p and if < , t p
1
(), rank(t) k we shall have
p
1
P
suc(t) p
1
() = suc(t) p();
so if < , i k we have
p
1
P
i
(p
1
())
i
(p()).
If p
2
p
1
there is some v V such that p
2
is compatible with p
(v)
1
= p
(v)
. If k l N, K L []
<
then
U(p
1
; L, l) = w : w U(p; L, l), v V such that v() w() K,
and p
(w)
1
= p
(w)
for every w U(p
1
; L, l); consequently, p
1
is (L, l)-xed if p is.
proof Requires only a careful reading of the denitions.
Remark Note that 1G-1P are based just on the fact that Q is a p.o.set of rapidly branching trees; the exact
denition of rapidly branching in 1E(c) is relevant only to some of the detailed calculations. Similar ideas
may be found in [BJSp89]and [Sh326].
1Q Construction: part 3 (a) Set X =
kN
n
k
. Then X, with its product topology, is a compact
metric space. Let be the natural Radon probability on X, the product of the uniform probabilities on the
factors.
(b) For each k N set l
k
= (ln n
k
)
2
|. Take W
k
n
k
n
k
such that #(W
k
) 2
k1
n
2
k
and whenever
I [n
k
]
l
k
and J
0
, . . . , J
l
k
1
are disjoint members of [n
k
]
k
, there are i I and j < l
k
such that i J
j
W
k
. (This is possible by Lemma 1B and 1E(a)(iv).) Set W
k
= W
k
(i, i) : i < n
k
.
Write R for
(x, y) : x, y X, (x(k), y(k)) W
k
k N, k : x(k) = y(k) is nite;
then R is negligible for the product measure of X X. For r N write R
r
for the set
(x, y
i
)
i<r
) : x X, (x, y
i
) R i < r X X
r
.
We shall frequently wish to interpret the formulae for the sets X, R
r
in V
P
; when doing so we will write
X, R
r
.
(c) Write
4
0
6
r
e
v
i
s
i
o
n
:
1
9
9
4
-
1
1
-
2
4
m
o
d
i
f
i
e
d
:
1
9
9
4
-
1
1
-
2
4
9
L = L
k
)
kN
: L
k
n
k
k N,
kN
#(L
k
)/n
k
> 0.
Again, we shall wish to distinguish between the ground-model set L and a corresponding P-name L.
(d) For each k N let
k
be the P-name for a subset of n
k
dened (up to equivalence) by saying that
p
P
k
= C
k
(t)
whenever rank(t) > k and p(0) T
(t)
. (Here C
k
(t) is the kth factor of t, as described in 1E(b).) Let
k
,
be P-names for the subsets of X given by
11
P
P
k
= : X, (i)
i
i k, =
kN
k
.
Then we have
11
P
P
#(
k
) (1 2
k1
)n
k
k N,
so that
11
P
P
() = 1.
1R Main Lemma If r N and D X
r
is a (ground-model) set such that D(
kN
L
k
)
r
,= for every
(ground-model) sequence L
k
)
kN
L, then for every (ground-model) sequence L
k
)
kN
L
11
P
P
kN
L
k
R
r
1
[D (
kN
L
k
)
r
].
proof (a) Let L
k
)
kN
L, let be a P-name such that
11
P
P
kN
L
k
,
and let p P. Write D
for D (
kN
L
k
)
r
. Let k
0
r, p
1
p be such that p
1
P
k0
. By Lemma
1O, we have a p
2
p
1
, an increasing sequence K
k
)
kN
of nite subsets of , and a sequence H
k
)
kN
of
functions such that
p
2
is (K
k
, k)-xed and (K
k
, k + 1)-xed for every k N,
p
(u)
2
P
(k) = H
k
(u) whenever u U(p
2
; K
k
, k + 1), k N;
kN
K
k
supp(p
2
);
0 K
0
, #(K
k
) k + 1 for every k N.
(b) For k k
0
, let Z
k
be the cartesian product set n
r
k
and take
W
k
to be
(i, z) : i < n
k
, z Z
k
, (i, z(j)) W
k
j < r.
Set /
k
= Tn
k
and T
k
= T(/
K
k
k
) ; dene
k
on T
k
as in Lemma 1C, taking l
k
and K
k
(with the
order induced by that of ) in place of l and m there.
For each u U(p
2
; K
k
, k) set
T
u
= c : u
c U(p
2
; K
k
, k + 1) T
k
,
where for u T
K
k
, c (Tn
k
)
K
we write
u
c = u() c())
K
.
By Corollary 1D, we may nd for each such u a w
u
< n
k
and a set J
u
n
k
such that #(J
u
) rl
k
and
either there is a T
k
T
u
such that H
k
(u
c) = w
u
for every c T
or for every z (n
k
J
u
)
r
there is a T
k
T
u
such that (H
k
(u
c), z)
W
k
, that is,
(H
k
(u
c), z(j)) W
k
j < r,
for every c T .
Set
I
k
= w
u
: u U(p
2
; K
k
, k),
J
k
=
J
u
: u U(p
2
; K
k
, k),
so that
#(
I
k
) #(U(p
2
; K
k
, k)) #(T
K
k
k
) (
i<k
2
ni
)
k+1
,
#(
J
k
) kl
k
(
i<k
2
ni
)
k+1
2
k
n
k
by 1E(a)(vi).
(c) Let k
1
max(k
0
, 1) be such that
k
(p
2
(0)) 2,
J
k
, L
k
for every k k
1
. Take any v
U(p
2
; K
k11
, k
1
) and set p
3
= p
(v
)
2
. Then p
3
is (K
k
, k)-xed and
(K
k
, k + 1)-xed for every k k
1
, and
T
u
= c : u
c U(p
3
; K
k
, k + 1)
whenever k k
1
, u U(p
3
; K
k
, k).
(d) We have
4
0
6
r
e
v
i
s
i
o
n
:
1
9
9
4
-
1
1
-
2
4
m
o
d
i
f
i
e
d
:
1
9
9
4
-
1
1
-
2
4
10
p
3
P
(i) = H
i
(v
i
) i < k
1
,
where v
i
is that member of T
Ki
i+1
such that v
i
() v
() for every K
i
. Set L
k
= H
k
(v
k
) for k < k
1
,
L
k
= L
k
J
k
for k k
1
; then
kN
#(L
k
)/n
k
> 0, because
kN
#(L
k
)/n
k
> 0 and
kN
#(
J
k
)/n
k
< .
So there is a z D (
kN
L
k
)
r
D
. Writing z
k
= z(j)(k))
j<r
for k N, we have z
k
(n
k
J
k
)
r
for
k k
1
, and
p
3
P
((k), z
k
)
W
k
for k < k
1
, because (i, i) W
k
for i < n
k
.
(e) For each k k
1
, u U(p
3
; K
k
, k) choose T
u
k
T
u
T
k
such that
either H
k
(u
c) = w
u
I
k
for every c T
u
or (H
k
(u
c), z
k
)
W
k
for every c T
u
.
Dene o
k
)
kk1
, p
k
)
kk1
by
p
k1
= p
3
,
o
k
= u
c : u U( p
k
; K
k
, k), c T
u
,
p
k+1
=
_
p
(v)
k
: v o
k
for every k k
1
, as in Lemma 1P. An easy induction on k shows that
T
u
= c : u
c U( p
k+1
; K
k
, k + 1)
whenever u U( p
k
; K
k
, k), k k
1
, that p
k
is (K
l
, l)-xed and (K
l
, l + 1)-xed whenever k
1
k l, and
that p
(v)
k
= p
(v)
3
whenever k
1
k l and v U( p
k
; K
l
, l) U( p
k
; K
l
, l + 1). Also supp( p
k
)
lN
K
l
for
every k k
1
.
(f ) It is likewise easy to see that, for k k
1
,
p
k+1
p
k
,
p
k+1
P
p
k+1
() T
k
= p
k
() T
k
< ,
p
k+1
P
p
k+1
() suc(t) = p
k
() suc(t) t p
k+1
() T
i
unless i = k and K
k
,
p
k+1
P
(k)
I
k
or ((k), z
k
)
W
k
.
(g) On the other hand, if k k
1
and K
k
,
p
k+1
P
dp(C : t C p
k+1
()) dp(C : t C p
k
())/2l
k
t p
k+1
() T
k
.
To see this, take any q p
k+1
and t such that
q
P
t p
k+1
() T
k
= p
k
() T
k
.
We may suppose that v
0
o
k
is such that q p
(v0)
k
= p
(v0)
k+1
= q
1
. Now p
k+1
is (K
k
, k +1)-xed so there
must be a t
t such that
q
1
P
T
k+1
p
k+1
().
There is accordingly a v
1
o
k
such that v
0
= v
1
and v
1
() = t
. Express v
1
as u
c
1
where u
U( p
k
; K
k
, k) and c
1
T
u
. Of course u() = t.
Now
q
1
P
C : t C p
k+1
() c() : c T
u
, c = c
1
,
q
1
P
C : t C p
k
() = c() : c T
u
, c = c
1
because p
k
and p
k+1
are both (K
k
, k + 1)-xed, while v
0
= (u)
(c
1
). But because T
u
k
T
u
,
dp(c() : c T
u
, c = c
1
) dp(c() : c T
u
, c = c
1
)/2l
k
.
So we get
q q
1
P
dp(C : t C p
k+1
()) dp(C : t C p
k
())/2l
k
.
As q and t are arbitrary, we have the result.
(h) Because
k
ln(2l
k
) 2
k
(by 1E(a)(v)),
p
k+1
K
k
,k
p
k
for every k k
1
. Also, supp( p
k
)
lN
K
l
for every k k
1
. By Lemma 1N, there is a p
4
P such that
p
4
p
k
for every k k
1
. Moreover, we may take it that
p
4
(0) =
kk1
p
k
(0)
(as in Lemma 1H), so that
4
0
6
r
e
v
i
s
i
o
n
:
1
9
9
4
-
1
1
-
2
4
m
o
d
i
f
i
e
d
:
1
9
9
4
-
1
1
-
2
4
11
i
(p
4
(0))
i
(p
3
(0))
i
ln(2l
i
) 1
whenever i k
1
. Note that for k k
1
,
p
4
P
(k)
I
k
or ((k), z
k
)
W
k
,
while for k < k
1
,
p
4
P
((k), z
k
)
W
k
.
(i) Now dene p
5
P by setting
i
=
I
i
z(j)(i) : j < r i N,
p
5
(0) = t : t p
4
(0), C
i
(t)
I
i
= whenever k
1
i < rank(t),
p
5
() = p
4
() if 0 < < .
Of course we must check that p
5
(0), as so dened, belongs to Q
0
= Q; but because
k
(p
4
(0)) 1 for k k
1
,
we have
dp(C : t C p
4
(0)) exp(1/
k
) 2(
i<k
2
ni
)
k+1
+ 2k 2#(
k
)
for every k k
1
, t p
4
(0) T
k
, using 1E(a)(vii). Of course p
5
(0) T
k11
= p
4
(0) T
k11
= v
(0), so
every element of p
5
(0) has a successor in p
5
(0), and also
i
(p
5
(0))
i
(p
4
(0))
i
ln 2
for i k
1
. (Here at last is the key step which depends on using dp in our measure of rapidly branching
given in 1E(d).) Thus p
5
(0) Q and p
5
P. But also
p
5
P
k
I
k
= k k
1
,
so that
p
5
P
(k) /
I
k
, (k) ,= z(j)(k) j < r, ((k), z
k
)
W
k
for k k
1
; nally
p
5
P
(, z) R
r
, R
r
1
[D
];
as p
5
p and p, are arbitrary,
11
P
P
kN
L
k
R
r
1
[D
],
as claimed.
1S Theorem For each r N,
11
P
P
if D
i
X and D
i
kN
L
k
,= L
k
)
kN
L, i < r,
then L
k
)
kN
L x
i
)
i<r
i<r
(D
i
kN
L
k
)
such that (x
j
, x
i
) R i < j < r.
proof Induce on r. If r = 0 the result is trivial. For the inductive step to r +1, take P-names
i
for subsets
of X such that
11
P
P
i
kN
L
k
,= L
k
)
kN
L, i r.
Take a P-name L for a member of L. Because members of L can be coded by simple sequences, we may
suppose that L is a P
-name for some < (1Je). The inductive hypothesis tells us that
11
P
P
L
k
)
kN
L x
i
)
i<r
i<r
(
i
kN
L
k
)
such that (x
j
, x
i
) R if i < j < r.
Next, using 1Jc-e, we can nd a and a P
i<r
i
,
11
P
(x
j
, x
i
) R
()
whenever x
l
)
l<r
, i < j < r,
(
kN
L
k
)
r
,= L
k
)
kN
L
()
,
where we write . . .
()
to indicate that we are interpreting some formula in V
P
.
4
0
6
r
e
v
i
s
i
o
n
:
1
9
9
4
-
1
1
-
2
4
m
o
d
i
f
i
e
d
:
1
9
9
4
-
1
1
-
2
4
12
Now we remark that by 1Jb P can be regarded, for forcing purposes, as an iteration P
, where P
is
a P
to say that
11
P
_
11
P
P
()
L R
r
1
[ (
L)
r
]
_
,
using the notation
()
to indicate which version of the P-name we are trying to use. Moving to V
P
for
a moment, we have
()
= 1 and (
L) > 0, so
11
P
P
l N, (
()
l
L) > 0.
Also, of course, every
()
l
L ,= .
We can therefore nd a P-name
r
for a member of
r
()
-names
i
, for
i < r, such that
11
P
P
i
)
i<r
(
L)
r
, (
r
,
i
)
i<r
) R
r
.
But of course we now have
11
P
P
i
)
ir
ir
(
i
L), (
j
,
i
) R i < j r.
As
i
)
ir
, L are arbitrary, this shows that the induction proceeds.
Version of 6.8.91
2. Pointwise compact sets of measurable functions We turn now to the questions in analysis which
the construction in 1 is designed to solve. We begin with some denitions and results taken from [Ta84].
2A Denitions (a) Let (X, , ) be a probability space. Write L
0
= L
0
() R
X
for the set of -
measurable real-valued functions on X. Let T
p
be the topology of pointwise convergence, the usual product
topology, on R
X
. Let T
m
be the (non-Hausdor, non-locally-convex) topology of convergence in measure
on L
0
, dened by the pseudometric
(f, g) =
_
min([f(x) g(x)[, 1)(dx)
for f, g L
0
.
(b) A set A R
X
is stable if whenever < in R, E and E > 0 there are k, l 1 such that
k+l
(x, y) : x E
k
, y E
l
, f A, f(x(i)) &f(y(j)) i < k, j < l
< (E)
k+l
,
writing
k+l
for the usual product outer measure on X
k
X
l
. (See [Ta84], 9-1-1.)
2B Stable sets Suppose that (X, , ) is a probability space and that A R
X
is a stable set.
(a) If (X, , ) is complete, then A L
0
(). ([Ta84], 9.1.)
(b) The T
p
-closure of A in R
X
is stable.
(c) If A is bounded above and below by members of L
0
, its convex hull is stable ([Ta84], 11-2-1).
(d) If A L
0
(as in (a)), then T
m
A, the subspace topology on A induced by T
m
, is coarser than T
p
A.
([Ta84], 9-5-2.)
For more about stable sets, see [Ta84]and [Ta87].
2C Pettis integration Let (X, , ) be a probability space and B a (real) Banach space.
(a) A function : X B is scalarly measurable if g : X R is -measurable for every g B
, the
continuous dual of B.
(b) In this case, is Pettis integrable if there is a function : B such that
_
E
gd exists = g(E) E , g B
.
(c) If : X B is bounded and scalarly measurable, then
A = g : g B
, |g| 1 L
0
is T
p
-compact. In this case is Pettis integrable i
f
_
E
f : A R
4
0
6
r
e
v
i
s
i
o
n
:
1
9
9
4
-
1
1
-
2
4
m
o
d
i
f
i
e
d
:
1
9
9
4
-
1
1
-
2
4
13
is T
p
A-continuous for every E ([Ta84], 4-2-3). In particular (by 2B(d)) is Pettis integrable if A is
stable.
2D The rivals Write
L
for Lebesgue measure on [0, 1], and
L
for its domain. Consider the following
two propositions:
(*) [0, 1] is not the union of fewer than c closed negligible sets;
() there are sequences n
k
)
kN
, W
k
)
kN
such that
n
k
2
k
, W
k
n
k
n
k
, #(W
k
) 2
k
n
2
k
k N;
taking X =
kN
n
k
, the usual Radon probability on X,
R = (x, y) : x, y X, (x(k), y(k)) W
k
k N, k : x(k) = y(k) is nite,
then whenever D X,
L
.
(e) If E
n
)
nN
is a stochastically independent sequence of measurable subsets of [0, 1], with lim
n
E
n
=
0 but
nN
(E
n
)
k
= for every k N, then there is an ultralter T on N such that
lim
nF
E
n
= x : n : x E
n
T
is non-measurable.
proof (a) See [Ta84], 9-3-1(b). (b) Use (a) and 2B(c). (c) Use (a) and 2B(d). (d) Use (a) and [Ta84],
10-2-1. (e) Observe that, writing E
n
for the characteristic function of E
n
, the set E
n
: n N is not
stable, and use (a).
2F Theorem Assume ().
(a) There is a bounded Pettis integrable function : [0, 1]
such that g : g (
, |g| 1 is
not stable in L
0
(
L
).
(b) There is a separable convex T
p
-compact subset of L
0
(
L
) which is not stable.
proof (We write
, :
by setting
(x)(m) = h
m
(x) m N, x X,
(E)(m) =
_
E
h
m
(x)(dx) m N, E .
We aim to show
4
0
6
r
e
v
i
s
i
o
n
:
1
9
9
4
-
1
1
-
2
4
m
o
d
i
f
i
e
d
:
1
9
9
4
-
1
1
-
2
4
14
(i) that A is not stable;
(ii) that if is a Radon probability on A
= Z A then
_
u(x) (du) = 0 for -almost every x;
(iii) f
E
: Z R is continuous for every E ;
(iv) is the indenite Pettis integral of , so that is Pettis integrable;
(v) K = g : g C(Z)
k
x : x X
k
, x(i)(m) / I i < k (1 n
1
m
l)
k
.
So we conclude that
k+l
(x, y) : x X
k
, y X
l
, f A, f(x(i)) = 0 i < k, f(y(j)) = 1 j < l
(1 2
m
)
l
2
(1 n
1
m
l)
k
(by Fubinis theorem). Because k, l and m are arbitrary, A cannot be stable.
ad (ii) Because each 1
m
is nite, any member of A
X, k : x(k) ,= x
(k) is nite x
E.
Note also that if x, y E then (x, y) / R; because either x(k) = y(k) for innitely many k, or there are k,
I such that x(k) ,= y(k), I 1
k
and x, y both belong to H
kI
, in which case (x(k), y(k)) / W
k
.
Now let be a Radon probability on A
in R
X
. If x, x
) for every u A
ir
u(x
i
) 1 for every u A
, while
ir
w(x
i
) (r+1) > 1,
and w cannot belong to the closed convex hull of A
.
Accordingly
, E .
It is enough to consider positive linear functionals g of norm 1. For any such g we have a Radon probability
on Z such that
g(f(h
m
))
mN
) =
_
Z
f(u) (du) for every f C(Z),
using the Riesz representation of positive linear functionals on C(Z). Set
m
= h
m
, = 1
mN
m
=
A
on A
such that
g(f(h
m
))
mN
) =
mN
m
f(h
m
) +
_
A
f(u)
(du)
for every f C(Z). Now an easy calculation (using (ii)) shows that
g((E)) = g(f
E
(h
m
))
mN
) =
mN
m
_
E
h
m
(x) (dx) =
_
E
g((x)) (dx)
for every E .
ad (v) If m N then h
m
= e
m
K where e
m
(
is dened by setting e
m
(z) = z(m) for every
z
is w
L
D = 1 then
L
Q
1
[D] = 1;
()
[0, 1]
2
such that whenever C, D [0, 1] and (C D) Q
= then
one of C, D is negligible.
4
0
6
r
e
v
i
s
i
o
n
:
1
9
9
4
-
1
1
-
2
4
m
o
d
i
f
i
e
d
:
1
9
9
4
-
1
1
-
2
4
15
Then 11
P
P
(). For start by taking Q
1
to be
(x, y) : x, y X, (x(k), y(k)) W
k
k N,
the closure of R in X X. Then the argument for 1S shows that
11
P
P
if D X and D
kN
L
k
,= L
k
)
kN
L
then < such that
()
Q
1
1
[D].
Consequently
11
P
P
if D X and
D = 1 then (Q
1
1
[D]) = 1.
Accordingly we have in V
P
the version of () in which ([0, 1],
L
) is replaced by (X, ). However there is
now a continuous inverse-measure-preserving function f : X [0, 1], and taking
Q = (f(x), f(y)) : (x, y) Q
1
, taking Q
to be
(x +q, y +q
) : (x, y) Q, q, q
), greatly simplied,
demolishes () also). The weaker form ()
), both ()
and
(