Discriptive Set Theory PDF
Discriptive Set Theory PDF
Discriptive Set Theory PDF
David Marker
Fall 2002
Contents
2 Borel Sets 14
4 Analytic Sets 34
5 Coanalytic Sets 43
6 Determinacy 54
7 Hyperarithmetic Sets 62
1
These are informal notes for a course in Descriptive Set Theory given at
the University of Illinois at Chicago in Fall 2002. While I hope to give a fairly
broad survey of the subject we will be concentrating on problems about group
actions, particularly those motivated by Vaught’s conjecture. Kechris’ Classical
Descriptive Set Theory is the main reference for these notes.
Notation: If A is a set, A<ω is the set of all finite sequences from A. Suppose
σ = (a0 , . . . , am ) ∈ A<ω and b ∈ A. Then σbb is the sequence (a0 , . . . , am , b).
We let ∅ denote the empty sequence. If σ ∈ A<ω , then |σ| is the length of σ. If
f : N → A, then f |n is the sequence (f (0), . . . , f (n − 1)).
If X is any set, P(X), the power set of X is the set of all subsets X.
If X is a metric space, x ∈ X and ² > 0, then B² (x) = {y ∈ X : d(x, y) < ²}
is the open ball of radius ² around x.
Part I
Classical Descriptive Set Theory
1 Polish Spaces
Definition 1.1 Let X be a topological space. We say that X is metrizable if
there is a metric d such that the topology is induced by the metric. We say that
X is separable if there is a countable dense subset.
A Polish space is a separable topological space that is metrizable by a com-
plete metric.
There are many classical examples of Polish spaces. Simple examples include
Rn , Cn , I = [0, 1], the unit circle T, and Qnp , where Qp is the p-adic field.
b y) = d(x, y)
d(x,
1 + d(x, y)
2
Suppose
Q dn is a complete metric on Xn , with dn < 1, for n = 0, 1, . . .. Define
db on Xn by
X∞
b g) = 1
d(f, n+1
dn (f (n), g(n)).
n=0
2
If f0 , f1 , . . . is a Cauchy-sequence, then f1 (i), f2 (i), . . . is a Cauchy-sequence
in Xi for each i. Let g(n) = lim fi (n). Then g is the limit of f0 , f1 , . . ..
i→∞
Suppose xi0 , xi1 , . . . is a dense subset of Xi . For σ ∈ N<ω let
½ n
xσ(n) if i < |σ|
fσ (n) = .
xn0 otherwise
Q
The {fσ : σ ∈ N<ω } is dense in X.
In particular, the Hilbert cube H = IN is Polish. Indeed, it is a universal
Polish space.
3
Lemma 1.6 If X is a Polish space, US⊆ X isSopen and ² > 0, then there are
open sets U0 , U1 , U2 , . . . such that U = Un = U n and diam (Un ) < ² for all
n.
Proof Let D be a countable dense set. Let U0 , U1 , . . . list all sets B n1 (d) such
that d ∈ D, n1 < ²/2 and B n1 (d) ⊆ U . Let x ∈ U . There is n > 0 such
that n1 < ², B n1 (x) ⊂ U . There is d ∈ D ∩ B 3n 1 (x). Then x ∈ B 1 (d) and
S 3n
. . . 0} |11 {z
φ(f ) = |00 {z . . . 1} |00 {z
. . . 0} 1 . . . .
f (0) f (1)+1 f (2)+1
4
Show that φ is a continuous and one-to-one. What is the image of φ?
Exercise 1.12 We say that x ∈ [0, 1] is a dyadic-rational is x = 2mn for some
m, n ∈ N. Otherwise, we say x is a dyadic-irrational. Show that N is homeo-
morphic to the dyadic-irrationals (with the subspace topology). [Hint: let φ be
as in Exercise 1.11 and map f to the dyadic-irrational with binary expansion
φ(f ).]
†
Exercise 1.13 Show that
1
f 7→ 1
1 + f (0) + 1+f (1)+ 1
1+f (2)+ 1
1+...
[T ] = {f ∈ N : f is a path through T }.
5
there is an m, such that the first n values of f (x) are determined by the first m
values of x. In §4 we will show how this brings in ideas from recursion theory.
Another key feature of the Baire space is that powers of the Baire space are
homeomorphic to the Baire space. Thus there is no natural notion of dimension.
Lemma 1.16 i) If k > 0, then N is homeomorphic to Nd × N k .
ii) N is homeomorphic to N N .
Proof If α = (n1 , . . . , nd , f1 , . . . , fk ) ∈ Nd × N k , let
φ(f ) = (n1 , . . . , nd , f1 (0), f2 (0), . . . , fk (0), f1 (1), . . . , fk (1), . . . , f1 (n), . . . , fk (n), . . .).
If β = (f0 , f1 , . . .) ∈ N N , let
ψ(β) = (f0 (0), f0 (1), f1 (0), . . .).
It is easy to see that φ and ψ are homeomorphisms.
A third important feature of the Baire space is that every Polish space is a
continuous image of the Baire space. We first prove that every closed subset of
N is a continuous image of N .
Theorem 1.17 If X is a Polish space, then there is a continuous surjective
φ : N → X.
Proof Using Lemma 1.6 build a tree of sets (Uσ : σ ∈ N<ω ) such that:
i) U∅ = X;
ii) Uσ is an open subset of X;
1
iii) diam (Uσ ) < |σ| ;
iv) Uτ ⊆ Uσ for σ ⊂ τ ;
[∞
v) Uσ = Uσ i .
i=0
If f ∈ N , then by 1.5 there is φ(f ) such that
∞
\ ∞
\
φ(f ) = Uf |n = Uf |n = {φ(f )}.
n=0 n=0
6
Lemma 1.18 Suppose X is a Polish space and Y ⊆ X is an Fσ -set and S ² > 0.
There are disjoint Fσ -sets Y0 , Y1 , . . . with diam (Yi ) < ², Y i ⊆ Y and Yi = Y .
S
Proof Let Y = Cn where Cn is closed. Replacing Cn by C0 ∪ . . . ∪ Cn
we may assume that C0 ⊆ C1 ⊆ . . .. Thus Y is the disjoint union of the sets
C0 , C1 \ C0 , C2 \ C1 , . . .. Since Ci \ Ci+1 ⊆ Ci ⊆ Y , it suffices to show that each
Ci \Ci−1 is a disjoint union of Fσ -sets of diameter less than ². Suppose Y = F ∩O
where F is closed and O is open. S By Lemma
S 1.6, we can find O0 , O1 , . . . open sets
with diam (On ) < ² and O = On = On . Let YnS= F ∩(On \(O0 ∩. . . On−1 )).
The Yi are disjoint, Yi ⊆ Oi ⊂ O, so Yi ⊆ Y , and Y i = Y .
Proof Using the previous lemma, we build a tree (Xσ : σ ∈ N<ω ) of Fσ -sets
such that
i) X∅ = X; S∞
ii) Xσ = i=0 Xσ i ;
iii) Xτ ⊆ Xσ if τ ⊂ σ;
1
iv) diam (Xσ ) < |σ| ;
v) if i 6= j, thenTXσ i ∩ Xσ j = ∅.
If f ∈ N , then Xf |n contains at most one point. Let
n ∞
\ o
F = f ∈ N : ∃x ∈ X x ∈ Xf |n .
n=0
T
Let φ : F → X such that φ(f ) = Xf |n . As above φ is continuous. By v) φ
is one-to-one.
T For any x ∈ X we can build a sequence σ0 ⊂ σ1 ⊂ . . . such that
x ∈ Xσn . We need only show that F is closed.
Suppose (fn ) is a Cauchy sequence in F . Suppose fn → f ∈ N . We must
show f ∈ F . For any n there is an m such that fi |n = fm |n for i > m. But then
d(φ(fi ), φ(fm )) < n1 . Thus φ(fn ) is a Cauchy sequence. Suppose φ(fn ) → x.
T T
Then x ∈ Xf |n = Xfn , so φ(f ) = x and f ∈ F .
Cantor–Bendixson analysis
We next show that the Continuum Hypothesis is true for Polish spaces, and
closed subsets of Polish spaces.
Definition 1.21 Let X be a Polish space. We say that P ⊆ X is perfect if X
is a closed set with no isolated points.
Note that ∅ is perfect. Nonempty perfect sets have size 2ℵ0 .
7
Lemma 1.22 If P ⊆ X is a nonempty perfect set, then there is a continuous
injection f : C → P . Indeed, there is a perfect F ⊆ P , homeomorphic to C. In
particular |P | = 2ℵ0 .
is closed.
Definition 1.24 If F ⊆ X is closed, the Cantor–Bendixson derivative is
8
Lemma 1.25 Suppose X is a Polish space and F ⊆ X is closed.
i) Γα (F ) is closed for all α < ω1 ;
ii) |Γα+1 (F ) \ Γα (F )| ≤ ℵ0 ;
iii) if Γ(F ) = F , then F is perfect, and Γα (F ) = F for all α < ω1 .
iv) there is an ordinal α < ω1 such that Γα (F ) = Γα+1 (F )
Proof i)–iii) are clear. For iv), let U0 , U1 , . . . be a countable basis for X. If
Γα+1 \ Γα 6= ∅, we can find nα ∈ N such that Unα isolates a point of Γα (F ).
By construction Unα does not isolate a point of Γβ (F ) for any β < α. Thus
nα 6= nβ for any β < α.
If there is no ordinal α with Γα (F ) = Γα+1 (F ), then α 7→ nα is a one-to-one
function from ω1 into N, a contradiction.
The Cantor–Bendixson rank of F , is the least ordinal α such that Γα (F ) =
α+1
Γ (F ).
Exercise 1.26 † Show that for all α < ω1 , there is a closed F ⊆ R with
Cantor–Bendixson rank α.
Polish subspaces
Suppose X is a Polish space and F ⊆ X is closed. If (xn ) is a Cauchy sequence
with each xn ∈ F , then lim xn ∈ F . Thus F is also a Polish space.
If U ⊂ X is open, then Cauchy sequences in U , may not converge to elements
of U . For example, (0, 1) ⊂ R and n1 → 0 6∈ (0, 1). The next lemma shows that
when U is open we are able to define a new complete metric on U compatible
with the topology.
9
Proof Let d be a complete metric on X compatible with the topology, we may
assume d < 1.
Let ¯ ¯
¯ 1 1 ¯
b ¯
d(x, y) = d(x, y) + ¯ − ¯.
d(x, X \ U ) d(y, X \ U ) ¯
b x) is a metric. Since d(x,
It is easy to see that d(x, b y) ≥ d(x, y), every d-open,
b
set is d-open. Suppose x ∈ U , d(x, X \ U ) = r > 0 and ² > 0. Choose δ > 0 such
η
that if 0 < η ≤ δ, then η + r(r−η) < ². If d(x, y) < δ, then d(y, X \ U ) > r − δ.
Hence ¯ ¯ ¯ ¯
¯1 1 ¯¯ ¯ −δ ¯
b ¯
d(x, y) ≤ δ + ¯ − ≤δ+¯ ¯ ¯ < ².
r r + δ¯ r(r − δ) ¯
b
Thus the d-ball of radius ² around x, contains the d-ball of radius δ. Hence
b
every d-open subset is open. Thus db is compatible with the subspace topology
on U . We need only show db is complete.
b
Suppose (xn ) is a d-Cauchy sequence. Then (xn ) is also a d-Cauchy sequence,
so there is x ∈ X such that xn → x. In addition for each n
¯ ¯
¯ 1 1 ¯
¯
lim ¯ − ¯ = 0.
i,j→∞ d(xi , X \ U ) d(xj , X \ U ) ¯
b
If (xi ) is a d-Cauchy sequence, then (xi ) is dn -Cauchy for each n. Thus there
is x T∈ X such that each xi → x in each On . Since each On is complete
x ∈ On = Y . Hence db is complete.
10
Can we generalize Corollary 1.31 further? We already saw that Q ⊆ R is not
a Polish subspace. Since Q is countable it is Fσ . Thus we can not generalize
this to Fσ -sets. Indeed the converse to the corollary is true.
Spaces of L-structures
We conclude this section with another important example of a Polish space.
Let L be a countable first-order language. Let Mod(L) be the set of all
L-structures with universe N. We will define two topologies on Mod(L). Let
{c0 , c1 , . . .} be a set of countably many distinct new constant symbols and let
L∗ = L ∪ {c0 , c1 , . . .}. If M ∈ Mod(L), then we can naturally view M as an
L∗ -structure by interpreting the constant symbol ci as i.
If φ is an L∗ -sentence, let Bφ = {M ∈ Mod(φ) : M |= φ}. Let τ0 be the
topology with basic open sets {Bφ : φ a quantifier-free L∗ -formula} and let τ1 be
the topology with basic open sets {Bφ : φ an L∗ -formula}. Clearly the topology
τ1 -refines τ0 .
11
the interpretation of the symbols of L in M and let χRM : N2 → 2 be the
characteristic function of RM . The function M 7→ (χRM , f M , cM ) is a bijection
between Mod(L) and X.
We will prove (Mod(L), τ0 ) is Polish by showing that this map is a homeo-
morphism. Let Y0 = {(g, h, n) ∈ X : g(i, j) = 1}, Y1 = {(g, h, n) : h(i, j) = k},
Y3 = {(g, h, n) ∈ X : n = m}. The inverse images of these sets are the ba-
sic clopen sets BR(ci ,cj ) , Bf (ci ,cj )=ck and Bcm =c , respectively. It follows that
this map is continuous. We need to show that if φ is quantifier-free, then the
image of Bφ is clopen. This is an easy induction once we show it for atomic
formulas. For formulas of the form R(ci , cj ) or f (ci , cj ) = ck , this is obvious.
A little more care is needed to deal with formulas built up from terms. For
example, let φ be the formula f (c0 , f (c1 , c2 )) = c3 . Then the image of Bφ is
Y = {(g, h, n) : h(0, h(1, 2)) = 3}. Then
[
Y = {(g, h, n) : h(1, 2) = i ∧ h(0, i) = 3}
i∈N
is open and
[ [
¬Y = {(g, h, n) : h(1, 2) = i ∧ h(0, i) = j}
i∈N j6=3
and X1 is closed.
Also \
X2 = {f : f (φ) = 0 ↔ f (¬φ) = 1}
φ∈S
and \
X3 = {f : f (ci = cj ) = 0}
i6=j
12
are closed.
Let F be the set of L∗ -formulas with one free-variable. Then
\³ [ ´
X4 = {f : f (∃vφ(v)) = 0} ∪ {f : f (φ(cn )) = 1}
φ∈F n∈N
is Gδ . Since X = X1 ∩ . . . ∩ X4 , X is Gδ .
Thus X is a Polish subspace of 2S .
If M ∈ Mod(L), let fM (φ) = 1 if M |= φ and fM (φ) = 0 if M |= ¬φ. It is
easy to see that fM ∈ X. If f ∈ X, then Henkin’s proof of Gödel’s Completeness
Theorem shows that there is an L-structure M with universe N such that:
i) if R is an n-ary relation symbol, then RM = {(n1 , . . . , nm ) : f (R(cn1 , . . . , cnm ) =
1};
ii) if g is an m-ary function symbol, then g M : Nm → N is the function
where g M (n1 , . . . , nm ) = k if and only if f (g(cn1 , . . . , cnm ) = cnk ) = 1;
iii) if c is a constant symbol, then cM = n if and only if f (c = cn ) = 1.
Thus M 7→ fM is a bijection between Mod(L) and X. The image of Bφ
is {f ∈ X : f (φ) = 1}. Thus this map is a homeomorphism and Mod(L) is a
Polish space.
Exercise 1.41 Show that the Hausdorff metric on K(X) is compatible with
the Vietrois topology.
13
2 Borel Sets
Definition 2.1 If X is any set, a σ-algebra on X is a collection of subsets of X
that is closed under complement and countable union. A measure space (X, Ω)
is a set X equipped with a σ-algebra Ω.
If (X, ΩX ) and (Y, ΩY ) are measure spaces, we say f : X → Y is a measurable
function if f −1 (A) ∈ ΩX for all A ∈ ΩY . We say that (X, ΩX ) and (Y, ΩY )
are isomorphic if and only if there is a measurable bijection with measurable
inverse.
Definition 2.2 If X is a topological space, the class of Borel sets B(X) is the
smallest σ-algebra containing the open sets.
If X and Y are topological spaces, we say that f : X → Y is Borel measurable
if it is a measurable map between the measure spaces (X, B(X)) and (Y, B(Y )).
We say that a measure space (X, Ω) is a standard Borel space if there is a
Polish space Y such that (X, Ω) is isomorphic to (Y, B(Y )).
S ii) Suppose−1O is open. S There are basic open sets U0 , U1 , . . . such that O =
Ui . Then f (O) = f −1 (Ui ) is a countable union of Borel sets and hence
Borel.
iii) Let U0 , U1 , . . . be a basis for the topology of Y . Then the graph of f is
∞
\
({(x, y) : y 6∈ Un } ∩ {(x, y) : x ∈ f −1 (Un )}.
n=0
−1
Since each f (Un ) is Borel so is the graph of f .
By ii) any continuous f : X → Y is Borel measurable. We will see later that
the converse of iii) is also true.
Since \ [
Ai = X \ (X \ Ai ),
any σ-algebra is also closed under countable intersections. Thus B(X) contains
all of the open, closed, Fσ , and Gδ sets. We could generalize this further by
14
taking Fσδ , intersections of Fσ -sets, Gδσ , unions of Gδ -sets, Fσδσ , Gδσδ . . ..
There is a more useful way of describing these classes.
Definition 2.4 Let X be a metrizable space. For each α < ω1 we define Σ0α (X)
and Π0α (X) ⊂ P(X) as follows:
Σ01 (X) is the collection of all open subsets of X;
Π0α (X) is the collection of all sets X \ A where A ∈ Σ 0
Sα (X);
0
For α > 1, Σα (X) is the collection of all sets X = Ai where each Ai ∈
Π0βi (X) for some βi < α.
We say that A ∈ ∆0α (X) if A ∈ Σ0α (X) and A ∈ Π0α (X).
When we are working in a single space we omit the X and write Σ0α and Π0α
instead of Σ0α (X) and Π0α (X).
Closed sets are Π01 , Fσ -sets are Σ02 , Gδ -sets are Π02 ,. . . .
Lemma 2.5 Suppose X is metrizable.
i) Σ0α ∪ Π0α S⊆ ∆0α+1 for all α < ω1 .
ii) B(X) = α<ω1 Σ0α .
iii) If X is infinite, then |B(X)| = 2ℵ0 .
Proof In any metric space every open set is both Fσ and Gδ , thus Σ01 ∪ Π01 ⊆
∆10 . i) then follows easily by induction. An easy induction shows that any
σ-algebra containing the open sets must contain Σ0α for each α < ω1 .
iii)SIf U0 , U1 , . . . is a basis for the topology, then every open set is of the
form n∈S Un for some S ⊆ N, thus |Σ0α | ≤ 2ℵ0 . Clearly |Π0α | = |Σ0α |. Suppose
S
α < ω1 and |Π0β | ≤ 2ℵ0 for all β < α. Then | β<α Π0β | < α and if F is the set
S
of f : N → β<α Π0β , then |F| ≤ (2ℵ0 )ℵ0 = 2ℵ0 and for any A ∈ Σ0α , there is
S
f ∈ F such that A = f (n). Thus |Σ0α | ≤ 2ℵ0 . Thus
¯ [ ¯
¯ ¯
|B(X)| = ¯ Σ0α ¯ ≤ ℵ1 × 2ℵ0 = 2ℵ0 .
α<ω1
15
Suppose we have proved ii) for all β < α. Then
∞ [
[ ∞
A0 ∩ A 1 = (B0,i ∩ B0,j )
i=0 j=0
and each B0,i ∩ B0,j is Π0β for some β < α. Thus A0 ∩ A1 is Σ0α .
iii) is immediate from i) and ii).
iv) Suppose f : X → Y is continuous. We prove that if A ⊆ YS is Σ0α
(respectively
S −1 Π0α ), then so is f −1 (A). If α = 0, this is clear. Since f −1 ( Ai ) =
f (A) and f −1 (Y \ A) = X \ f −1 (A), this follows easily by induction.
Examples
We give several examples.
Point are closed, so every countable set is a countable union of closed sets.
is Σ02 .
16
is closed. Let A1 = {x : ∀n∃m x(m) = n}. Then
∞ [
\ ∞
A1 = {x : x(m) = n}
n=0 m=0
∀n∃σ ∈ 2n x(σ) = 1
At first this looks Π02 , but the existential quantifier is only over a finite set.
Indeed \ [
W F2 = {x : x(σ) = 1}
n∈N σ∈2n
S
and σ∈2n {x : x(σ) = 1} is a clopen set.
17
Let N = {x ∈ C : x is normal}.
x is normal if and only if
à ¯ ¯ !
¯1 1 X
n ¯ 1
¯ ¯
∀k > 0∃m∀n n>m→¯ − x(i)¯ < .
¯2 n + 1 ¯ k
i=0
If ( ¯ ¯ )
¯1 1 X
n ¯ 1
¯ ¯
An,k = x∈C:¯ − x(i)¯ < ,
¯2 n + 1 ¯ k
i=0
and ∞
[
Mod(∃v φ(v)) = Mod(φ(cn )).
n=0
is Borel.
Exercise 2.16 Show that if φ is a first order L∗ -sentence, then Mod(φ) is Σ0n
for some n. (hint: prove that n depends only on the quantifier rank of φ.)
Conclude that if T is a first order theory, then Mod(T ) is Π0ω .
In the topology τ1 , Mod(φ) is clopen for all first order φ. Thus Mod(T ) is
closed.
18
Suppose M ∈ Mod(L). There is φM ∈ Lω1 ,ω , the Scott sentence of M (see
[11] 2.4.15) such that if M1 is a countable L-structure, then M ∼
= M1 if and
only if M1 |= φM . Thus
{M1 ∈ M : M1 ∼
= M} = Mod(φM )
is a Borel set.
∀n (Un ∩ B = ∅ → Un ∩ A = ∅).
A is finite if and only there are basic open set U1 , . . . , Un such that A ⊆
U1 ∪ . . . ∪ Un such that if Vi and Vj are disjoint basic open subsets of Ui , then
A ∩ V0 = ∅ or A ∩ V1 = ∅.
Fix U0 , U1 , . . . a basis for the open sets. If F ⊆ N is finite, then
[
BF = {A : A ⊆ Vi }
i∈F
is Σ02 .
Exercise 2.21 Show that {A ∈ K(X) : A is perfect} is Π02 .
19
Changing the Topology
Suppose X is a Polish space, let τ be the topology of X. We will often prove
interesting results about Borel sets A ⊆ X, by refining τ to a new topology τ1
with the same Borel sets such that A is clopen in the new topology.
We start with one preparatory lemma.
Lemma 2.22 Suppose X and Y are disjoint Polish spaces. The disjoint union
X ] Y is the space X ∪ Y where U ⊆ X ∪ Y is open if and only if U ∩ X and
U ∩ Y are both open. Then X ] Y is a Polish space.
Proof We know that X \ F has a Polish topology and F has a Polish topology.
Let τ1 be the Polish topology on the disjoint union of X \ F and F . Then F
is open. The open subsets of τ1 are either open in τ or intersections of τ open
sets with F . In particular they are all Borel in τ . Thus the Borel sets of τ1 are
the Borel sets of τ .
20
T
Since Ai is τi -clopen, it is also clopen in τ ∗ . Thus Ai is τ ∗ -closed. One
∗ ∗∗
further application of the previous T lemma allows us to refine τ to τ keeping
the same Borel sets but making Ai clopen.
Thus Ω is a σ-algebra, so Ω = B(X).
We can use this observation to deduce several important results.
Theorem 2.25 (Perfect Set Theorem for Borel Sets) If X is a Polish space
and B ⊆ X is an uncountable Borel set, then B contains a perfect set.
Proof Let τ be the topology on X. We can refine the topology to τ1 such that B
is closed. Since B is uncountable, by 1.27 there is a nonempty τ1 -perfect P ⊆ B
and f : C → P a homeomorphism. Since τ1 refines τ0 , f is also continuous in the
topology τ . Since C is compact, P is τ -closed. Since P has no isolated points
in τ1 , this is still true in τ , so P is a perfect subsets of B.
Borel Isomorphisms
Example 2.30 If A ⊆ X and B ⊆ Y are countable and |A| = |B|, then any
bijection f : A → B is a Borel isomorphism.
21
In this case the inverse image of any open set is countable, and hence, Fσ .
Example 2.31 The Cantor space C is Borel isomorphic to the closed unit in-
terval I.
22
Corollary 2.34 i) If X is a Polish space and A ⊆ X is an uncountable Borel
set, then A is Borel isomorphic to C.
ii) Any two uncountable Polish spaces are Borel isomorphic.
iii) Any two uncountable standard Borel spaces are isomorphic.
Proof
i) If (X, τ ) is a Polish space and A ⊆ X is Borel, we can refine the topology
of X making A clopen but not changing the Borel sets. Then A is a Polish space
with the new subspace topology. We have shown that A is Borel isomorphic to
a Borel subset of C and, by the Perfect Set Theorem, there is a Borel subset of
A homeomorphic to C. Thus there is a Borel isomorphism f : A → C. Since
the new topology has the same Borel sets as the original topology, this is also a
Borel isomorphism in the original topology.
ii) and iii) are clear from i).
Exercise 2.35 † Prove that if X and Y are Polish spaces, A ⊆ X is Borel and
f : X → Y is continuous, and f |A is one-to-one, then f (A) is Borel. Conclude
that f |A : A → B is a Borel isomorphism. [This can be proved by the methods
at hand, but we will give a very different proof later.]
Lemma 2.37 If X is a separable metric space, then for all 1 < α < ω1 there
is a Σ0α -universal set Uα ⊆ C × X and a Π0α -universal set Vα ⊆ C × X.
23
cα = {(n, f, x) ∈ N × C × X : (f, x) ∈ Vβn }. Then U cα = S
Let U n∈N {n} × Vβn
is Σ0α .
Using the natural homeomorphism between C and C N we can identify every
f ∈ C with (f0 , f1 , f2 , . . .) ∈ C N . Then Uα = {(f, x) : ∃n (n, fn , x) ∈ U
cα } is Σ0 .
S α
0 0
If A ⊆ X is Σα , then there are B0 , B1 , . . . such that Bn is Πβn and A = Bn .
Choose fn such that x ∈ Bn if and only if (fn , x) ∈ Vβn and choose f coding
(f0 , f1 , f2 , . . .). Then (f, x) ∈ Uα if and only if x ∈ A.
We can now prove that in an uncountable Polish space the Borel hierarchy
is a strict hierarchy of ω1 -levels.
Proof
i) Let U ⊆ C × C be Σ0α -universal. Let Y = {x : (x, x) 6∈ Uα }. Clearly
Y ∈ Π0α . If Y ∈ Σ0α , then there is y ∈ X such that x ∈ Y if and only if
(y, x) ∈ Uα . Then
y ∈ Y ⇔ (y, y) ∈ Uα ⇔ y 6∈ Y,
a contradiction.
ii) Suppose X is an uncountable Polish space. Then X contains a perfect
set P homeomorphic to C. If Σ0α (X) = Π0α (X), then, by 2.8, Σ0α (P ) = Σ0α (P ),
contradicting i).
Since Σ0α (X) 6= Π0α (X), there is A ∈ Π0α (X) \ Σ0α (X).
This gives us the following picture of the Borel Hierarchy.
..
.
HH ©
0 H ©©
Σα H ©© Π0α
H
©©H
© © ∆0α HH
© .. H
H
.
H ©
H ©©
Σ02 HH ©© Π02
H
©©H
© © ∆02 HH
© H
H
H ©
©
H ©
Σ01 HH ©© Π01
H
©©H
© © ∆01 HH
© H
H
24
We write A ≤w B if A is Wadge-reducible to B.
Note that if A ≤w B, then X \ A ≤w Y \ B.
Example 2.40 If A ⊆ N is open, then A ≤w {x ∈ N : ∃n x(n) = 1}.
S
Let A = n Un where each Un is a basic clopen set. Let f : N → N such
that n
1 if x ∈ Un
(f (x))(n) = .
0 otherwise.
If σ ∈ 2<ω , \ \
f −1 (Nσ ) = Un ∩ X \ Un
σ(n)=1 σ(n)=0
25
The Baire Property
We begin by recalling some basic ideas from analysis. Let X be a Polish space.
Definition 2.46 We say that A ⊆ X is nowhere dense set , if whenever U ⊆ X
is open and nonempty, there is a nonempty open V ⊆ U such that A ∩ U = ∅.
We say that B ⊆ X is meager if X is a countable union of nowhere dense
sets.
Exercise 2.47 Show that Cantor’s “middle third” set is nowhere dense.
Exercise 2.48 We say that I ⊆ P(X) is an ideal if i) ∅ ∈ I, ii) if A ∈ I and
B ⊆ A, then B ∈ S I, and iii) if A, B ∈ I, then A ∪ B ∈ I. We say that an ideal
I is a σ-ideal if n An ∈ I, whenever A0 , A1 , . . . ∈ I.
a) Show that the nowhere dense sets form an ideal.
b) Show that the meager sets form a σ-ideal.
Exercise 2.49 a) Show that if A is nowhere dense, then A is nowhere dense.
b) Show that every meager set is contained in a meager Fσ -set.
Exercise 2.50 Show that if U is open, then U \ U is nowhere dense.
26
Lemma 2.56 If A has the Baire property, then X \ A has the Baire property.
X \ A =∗ X \ U =∗ intr(X \ U ).
Exercise 2.59 Show that if A has the property of Baire, then either A is
meager or there is σ such that Nσ \ A is meager.
Do all sets have the Baire property?
Exercise 2.60 Use the axiom of choice to construct a subset of R without the
Baire Property.
27
We can code computer programs by integers so that each integer codes a
program. Let Pe be the machine coded by e. Let φe be the partial recursive
function computed by e. We write φe (n) ↓s if Pe halts on input n by stage s
and φe (n) ↓ if Pe halts on input n at some stage. Our enumeration has the
following features.
Fact 3.2 i) [universal function] The function (e, n) 7→ φe (n) is partial recursive.
ii) The set {(e, n, s) : φe (n) ↓s } is recursive.
iii) [halting problem] The set {(e, n) : φe (n) ↓} is not recursive.
iv) [parameterization lemma] If F : N2 → N is partial recursive, there is a
total recursive d : N → N such that
for all x, y.
Exercise 3.6 If you haven’t seen them before prove the statements in the last
Fact.
A program with oracle x ∈ N is a computer program which, in addition to
the usual steps, is allowed at any stage to ask the value of x(n).
We say that f is partial recursive in x if there is a program with oracle x
computing f and say that A ⊆ N is recursive in x if the characteristic function of
A is recursive in x. The facts above relativize to oracle computations. We write
φxe (n) for the value of the partial recursive function in x with oracle program
Pe on input n. One additional fact is useful.
Fact 3.7 (Use Principle) If φxe (n) ↓, then there is m such that if x|m = y|m,
then φye (n) = φxe (n).
28
Proof The computation of Pe with oracle x on input n makes only finitely
many queries about x. Choose m greater than all of the queries made.
We may also consider programs with finitely many oracles.
For x, y ∈ N we say x is Turing-reducible to y and write x ≤T y if x is
recursive in y. There is another useful reducibility that is the analog of Wadge-
reducibility for N.
Definition 3.8 We say A is many-one reducible to B if there is a total recursive
f such that n ∈ A if and only if f (n) ∈ B for all n ∈ N. We write A ≤m B if A
is many-one reducible to B.
Clearly if A ≤m B, then A ≤T B.
There is one subtle fact that will eventually play a key role.
Proof Let ½
φφx (x) (y) if φx (x) ↓
F (x, y) = .
↑ if φx (x) ↑
By the Parameterization Lemma, there is a total recursive d such that
For those of you who haven’t seen this before here is a sample of the many
applications of the Recursion Theorem. Let
(
1 if e = n
g(e, n) = 0 if e 6= n
.
By the Parameterization Lemma, there is a total recursive f such that φf (e) (n) =
g(e, n). By the Recursion Theorem there is an e such that φe (n) = 1 if n = e and
φe (n) = 0 if n 6= e. So this function “recognizes” its own code. The Recursion
Theorem will be very useful in §7.
Computable Functions on N
There is also a notion of computable function f : N → N .
Definition 3.10 We say that f : N → N is computable if there is an oracle
program P such that if x ∈ N and P is run with oracle x on input n, then P
halts and outputs (f (x))(n).
29
We say that f : N → N is computable from z if there is a two oracle program
P such that if x ∈ N and P is run with oracles z and x on input n, then M
halts and outputs (f (x))(n).
Proof
(⇐) Let P be the oracle program computing f from z. Suppose f (x) = y.
By the Use Principle, for any m there is an n such that if x|n = z|n, then
f (z)|m = y|m. Thus Nx|n ⊆ f −1 Ny|m and f is continuous.
(⇒) Let X = {(τ, σ) : f −1 (Nσ ) ⊆ Nτ }. Since f is continuous, for all x ∈ N
if f (x) = y, then for all n there is an m such that (x|m, y|n) ∈ X.
We claim that f is computable from X. Suppose we given oracles X and x
and input n. We start searching X until we find (τ, σ) ∈ X such that τ ⊂ x
and |σ| > n. Then (f (x))(n) = σ(n).
30
We will tend to prove things only for Σ01 sets. The relativization to Σ01 (x)
sets is usually straightforward.
For the two interesting examples X = N and X = N we get slightly more
informative characterizations.
Proof
i) This is clear since the recursively enumerable sets are exactly the images
of partial recursive functions.
ii) In this case SX = N<ω . Clearly if SS ⊆ N<ω is recursive, there is f : N →
<ω 0
N partial recursive
S with image S and σ∈N<ω Nσ is Σ1 .
Suppose A = n Nf (n) where f is partial recursive let S = {σ : there is
n ≤ |σ|, the computation of f (n) halts by stage |σ| and f (n) ⊆ σ. It is easy to
see that S is recursive. If σ ∈ S, then there is an n such that f (n) ⊆ σ, then
Nσ ⊆ Nf (n) ⊆ A. On the hand if f halts on input n, there is m ≥ n, |f (n)| such
that f halts by stage m. If τ ⊃ σ and |τ | ≥ m, then τ ∈ S. Thus
[ [
Nσ ⊃ Nτ = N σ .
σ∈S τ ⊃f (n),|τ |=m
S
It follows that A = σ∈S Nσ .
We have natural analogs of the finite levels of the Borel hierarchy.
Definition 3.16 Let X = Nk × N l . We say that A ⊆ X is Π0n if and only if
X \ A is Σ0n .
We say that A ⊆ X is Σ0n+1 if and only if there is B ⊆ N × X in Π0n such
that
x ∈ A if and only if ∃n (n, x) ∈ B.
We say that A is ∆0n if it is both Σ0n and Π0n .
We say that A ⊆ X is arithmetic if A ∈ ∆0n for some n.
31
Exercise 3.18 a) Show that if T is a recursive pruned tree, then the left-most
path through T is recursive.
b) Let T = {σ ∈ N<ω : if e < |σ| and φe (e) halts by stage |σ|, then φe (e)
halts by stage σ(e)}. Show that T is a recursive tree. Suppose f ∈ [T ]. Show
that φe (e) halts if and only if it halts by stage f (e). Conclude that there are
no recursive paths through T and, using a), that there is no recursive pruned
subtree of T .
We show that Σ0n and Π0n have closure properties analogous to those proved
in 2.6. The definition of computable function made in 3.10 makes sense for maps
f : X → Y where both X and Y are of the form Nk × N l
Lemma 3.19 i) Σ0n is closed under finite unions, finite intersections, and com-
putable inverse images.
ii) If A ⊆ N × X ∈ Σ0n , then {x ∈ X : ∃n (n, x) ∈ A} ∈ Σ0n .
iii) If f : X → N is computable and A ⊆ N × X is Σ0n then {x ∈ X : ∀m <
f (x) (m, x) ∈ A} ∈ Σ0n .
iv) Similarly Π0n is closed under union, intersection, computable inverse im-
ages, ∀n and ∃n < f (x).
v) Σ0n ⊆ ∆0n+1 .
Proof We prove this for Σ01 and leave the induction as an exercise.
S i) Suppose W0 and W1 are recursively enumerable subsets of SX and Ai =
η∈Wi Nη . Replacing Wi by the recursively enumerable set {ν : ∃η ∈ Wi η ⊆ ν}
if necessary we may assume that if ν ∈ Wi and η ⊃ ν, then η ∈ Wi . Then
[
A0 ∪ A 1 = Nη
η∈W0 ∪W1
and [
A0 ∩ A 1 = Nη
η∈W0 ∩W1
and W0 ∪W1 and W0 ∩W1 are recursively enumerable. Thus A0 ∪A1 and A0 ∩A1
are Σ01 .
If f : X → Y is computable with program Pe , let
G = {(η, ν) ∈ SX × SY : x ∈ Nη ⇒ f (x) ∈ Nν }.
Then (η, ν) ∈ G if and only if for all m < |ν| the program Pe using oracle
2
η halts on input
S m and outputs ν(m). Thus G is recursively enumerable.
Suppose A = ν∈W Nν where W is recursively enumerable, let V = S {η : ∃ν (ν ∈
W ∧ (η, ν) ∈ G}. Then V is recursively enumerable and f −1 (A) = η∈V Nη .
ii) Suppose AS⊆ N × X is Σ01 . There is a recursively enumerable W ⊆ SN×X
such that A = η∈W Nη . Let V = {ν ∈ SX : ∃n (n, ν) ∈ W }. Then V is
2 We assume that if the computation makes any queries about numbers i ≥ |η|, then the
32
recursively enumerable and
[
{x : ∃n (n, x) ∈ A} = Nν .
ν∈V
Exercise 3.20 Give the inductive steps to complete the proof of 3.19
We can make two interesting observations about universal sets. We state
these results for Σ0n , but the analogous results hold for Π0n .
Proposition 3.21 i) There is U ⊆ N × X a Σ0n -set that is Σ0n -universal.
ii) There is V ⊆ N × X a Σ0n -set that is Σ0n -universal.
Proof
i) Indeed the universal sets produced in 2.37 are Σ0n . Fix f : N → SX a
recursive bijection. The set U1 = {(x, y) : ∃n (x(n) = 1 ∧ y ∈ Nf (n) )} is Σ01 and
Σ0n -universal.
If Un∗ ⊆ N × N × X is Σ0n and Σ0n -universal for N × X, then
Un+1 = {(x, y) : ∃n (x, n, y) 6∈ Un∗ }
is Σ0n+1 and Σ0n -universal.
ii) Let V1 = {(n, x) : ∃m (φn (m) ↓ ∧x ∈ Nφn (m) )}. Let g : N × N → SN×X
be partial recursive such that g(n, m) = (n, φn (m)), then
[
V1 = Ng(n,m)
n,m
33
Fact 3.23 i) {e : dom (φe ) 6= ∅} is Σ01 -complete.
ii) {e : φe is total} is Π02 -complete.
iii) {e : dom (φe ) is infinite} is Π02 -complete.
iv) If U ⊆ N × N is Γ-universal, then U is Γ-complete.
4 Analytic Sets
When studying the Borel sets, it is often useful to consider a larger class of sets.
Definition 4.1 Let X be a Polish space. We say that A ⊆ X is analytic if there
is a Polish space Y , f : Y → X continuous and B ∈ B(Y ) such that A = f (B)
the image of B. We let Σ11 (X) denote the collection of all analytic subsets of
X.
If no confusion arises we write Σ11 rather than Σ11 (X). The following lemma
gives several alternative characterizations of analytic sets. In general if X × Y
is a product space, we let πX and πY denote the projections onto X and Y .
Proof
i)⇒ ii) Since A ∈ Σ11 , there is a Polish space Y , f : Y → X continuous, and
B ⊆ Y Borel such that f (B) = A. By 2.26 there is a continuous g : N → Y ,
such that g(N ) = B. Then f ◦ g is a continuous map from N to X whose image
is A.
ii)⇒ iii) Suppose f : N → X is continuous and f (N ) = A. Let G(f ) ⊆
N × X be the graph of f . Then G(f ) is closed and πX (G(f )) = A.
iii)⇒ iv) and iv) ⇒ i) are clear.
Exercise 4.3 a) Show that if A ∈ Σ11 (X), then there is B ∈ Π02 (C × X) such
that π(B) = A.
b) Show that this cannot be improved to Σ02 (C × X).
34
Lemma 4.5 i) Σ11 and Π11 are closed under countable unions and intersections.
ii) If f : X → Y is Borel measurable, and A ∈ Σ11 (X), then f (A) ∈ Σ11 (Y ).
iii) Σ11 and Π11 are closed under Borel measurable inverse images.
Proof
1 0
i) Suppose
S AS0 , A1 , . . . ∈ Σ1 (X). Let Ci ∈ Π1 (N × X) such that π(Ci ) = Ai .
1
Then Ai = π( Ci ) ∈ Σ1 .
Using the homeomorphism between N and N N we can view f ∈ N as coding
(f0 , f1 , . . .) ∈ N N . Let
C = {(f, x) ∈ N × X : ∀n (fn , x) ∈ Ci }.
T
Then C is a closed subset of N × X and Ai = π(C) ∈ Σ11 .
Since Σ11 is closed under countable unions and intersections, so is Π11 .
ii) If f : X → Y is Borel measurable, then, by 2.3 iii), G(f ), the graph of f ,
is a Borel subset of X × Y . Suppose A ∈ Σ11 (X), there is a closed C ⊆ N × X
such that A = πX (C). Let
Examples
Example 4.6 Let LO be as in 2.12 and let W O = {x ∈ LO : x is a well order}.
Then W O is Π11 .
A linear order is a well order if and only if there are no infinite descending
chains. Thus
Example 4.7 Let T r be T rN as in 2.13, codes for subsets of N<ω that are trees.
We say that T ∈ T r is well-founded if [T ] = ∅. Let WF = {x ∈ T r : x is well
founded}. Then WF is Π11 .
35
Let s : (N<ω )2 → 2 be such that s(σ, τ ) = 1 if and only if σ ⊂ τ . Then
For notational simplicity, we consider only the case L = {R} where R is a binary
relation symbol, the general case is similar. Then
M0 ∼
= M1 ⇔ ∃f : N → N ∀i, j ∈ N RM0 (i, j) ↔ RM1 (f (i), f (j)).
An uncountable closed set contains a perfect set. In 2.19 and 2.21 we saw
that {(A, B) : A ⊆ B} and {P ∈ K(X) : P is perfect} are Borel. But A is
uncountable if and only if
36
Proof We first prove this for X = C. Let A = {a ∈ N : (a, a) 6∈ U }. If U ∈ Π11 ,
then A ∈ Σ11 and A = Ua for some a ∈ C. But then
a ∈ A ⇔ a ∈ Ua ⇔ (a, a) ∈ U ⇔ a 6∈ A,
a contradiction.
If X is an uncountable Polish space, there is f : C → X be a Borel isomor-
phism. Let A ⊆ C be Σ11 but not Π11 . By 4.5 ii) f (A) ∈ Σ11 (X). By 4.5 iii), if
f (A) ∈ Π11 (X), then A = f −1 (f (A)) ∈ Π11 , a contradiction.
The last proof illustrates an important point. Since any two uncountable
Polish spaces are Borel isomorphic, if we are trying to prove something about
Borel and analytic sets, it is often enough to prove it for one particular Polish
space (like N or C) and then deduce it for all Polish spaces.
Corollary 4.14 If A ∈ ∆11 (X), then A is Borel. Thus B(X) = ∆11 (X).
Proof Since A and X \ A are disjoint Σ11 -sets. They are separated by a Borel
set. The only set separating A and X \ A is A.
We can now prove the converse to 2.3 iii).
37
Proof i) ⇒ ii) is 2.3 iii), and ii) ⇒ iii) is obvious.
Suppose the graph of f is analytic and A ∈ B(Y ), then
x ∈ f −1 (A) ⇔ ∃y (y ∈ A ∧ f (x) = y)
⇔ ∀y (f (x) = y → y ∈ A).
Proof Suppose not. For each n > 1 let Un,0 , Un,1 , . . . be an open cover of X
by open balls of radius n1 . Choose x(n) such that Un,x(n) ∩ A is uncountable.
Let An = A \ U n,x(n) . If An is uncountable we can find an open set V disjoint
from Un,x(n) such that V ∩ A is uncountable, thus An is uncountable. But
[ \
A\( An ) = Un ,
n n
S
since diam (Un ) → 0, there is at most one element in A \ ( n An ) and hence A
is countable, a contradiction.
38
A little extra care would allow us to conclude that f ◦ g(C) is perfect. We
will give a second proof of this theorem in §6.
The Perfect Set Theorem for Σ11 -sets is the best result of this kind that we can
prove in ZFC. The next natural question is whether the Perfect Set Theorem is
true for Π11 -sets. Unfortunately, this depends on set theoretic assumptions. Let
L be Gödel’s constructible universe. If x ⊆ N, let L(x) be the sets constructible
from x.
Baire Property
We will show that analytic sets have the Baire Property.
We begin by giving another normal form for Σ11 -sets. Let X be a Polish
space.
Definition 4.19 Suppose Bσ ⊆ X for all σ ∈ N<ω . We define
[ \
A({Bσ }) = Bσ .
f ∈N n∈N
39
Then [
A \ A1 = {A ∩ Ui : A ∩ Ui is meager},
a meager set. Let B = A ∪ A1 . Since B = A1 ∪ (A \ A1 ) is the union of a closed
set and a meager set, B has the Baire Property.
Suppose B 0 ⊇ A has Baire Property, then C = B \B 0 has the Baire Property.
We must show that C is meager. If not then Ui \ C is meager for some i. Then
Ui ∩ A is meager. Since Ui ∩ C 6= ∅ and C ⊆ A1 , there is x ∈ Ui such that
x ∈ A1 . Thus Ui ∩ A is not meager, a contradiction.
Theorem 4.22 Suppose Aσ has the Baire Property for all σ ∈ N<ω . Then
A = A({Aσ }) has the Baire Property.
Proof
T We may assume that Aσ ⊆ Aτ for τ ⊆ σ; otherwise replace Aσ by
<ω
τ ⊆σ Aτ . For σ ∈ N let
[ \
Aσ = Ax|n ⊆ Aσ .
x⊃σ n∈N
By 4.21 there is B σ ⊇ Aσ with the Baire Property such that if B ⊇ Aσ has the
Baire Property, then B σ \ B is meager.T We may assume that B σ ⊆ Aσ and that
B σ ⊆ B τ for τ ⊆ σ, replacing B σ by τ ⊆σ B τ if necessary.
S S
Let Cσ = B σ \ n B σ n .SSince Aσ ⊆ n B σ n , our choice of B σ s insures
that Cσ is meager. Let C = σ∈N<ω Cσ . Clearly C is meager.
Claim B ∅ \ C ⊆ A.
Let b ∈ B ∅ \ C. Since b 6∈ C∅ , there is x(0), such that b ∈ B x(0) . Suppose we
have x(0), . . . , x(n) such that b ∈ B x(0),...,x(n) . Since b 6∈ Cx(0),...,x(n) , there is
x(n + 1) such that b ∈ B x(0),...,x(n+1) . Continuing this way we construct x ∈ N
such that \ \
b∈ B x|n ⊆ Ax|n ⊆ A.
n n
Thus b ∈ A.
Then B ∅ \ A ⊆ C. Hence B ∅ \ A is meager. In particular B σ \ A, and hence
A, have the Baire Property.
Corollary 4.23 If X is a Polish space, then every Σ11 -set has the Baire Prop-
erty.
In §8 it will be useful to notice that our proof that the collection of sets with
the Baire Property is closed under the Souslin operator works in any topological
space with a countable basis (not just Polish spaces).
Exercise 4.24 a) Prove that if A ⊆ Rn , then there is B ⊇ A such that if B 0 ⊆ A
is Lebesgue measurable, then B \ B 0 has measure zero. [Hint: If µ∗ (A) < ∞,
choose B ⊇ A measurable with µ(B) = µ∗ (A), where µ∗ is Lebesgue outer
measure. Otherwise write A as a union of sets with finite outer measure.]
b) Modify the proof of 4.22, using a), to prove the following theorem.
40
Theorem 4.25 The collection of Lebesgue measurable subsets of R n is closed
under the Souslin operator. In particular every Σ11 -set is Lebesgue measurable.
We give another restatement.
Definition 4.26 Let C be the smallest σ-algebra containing the Borel sets and
closed under the Souslin operator A.
We have proved that every C-measurable set is Lebesgue measurable.
©H
©© ∆12 HH
©© H
© HH
H ©©
HH ©
Σ11 HH © Π1
H ©© 1
©©H 1 H
©© ∆1 HH
© © HH
41
Example 4.30 The set {x ∈ N : x ∈ L} is Σ12 .
The idea of the proof is that there is a sentence Θ such that ZF ` Θ and L
is absolute for transitive models of Θ. Using the Mostowski collapse
x ∈ L if and only if there is M a countable well-founded model of Θ + V = L
with x ∈ M.
This is a Σ12 definition of N ∩ L.
Example 4.31 If V = L, then there is a ∆12 well-order of N of order type ω1 .
Indeed the canonical well-ordering of L is ∆12 .
These example can be used to show that projective sets need not have nice
regularity properties. We will use Fubini’s Theorem, that a measurable A ⊆ R 2
has positive measure if and only if {a : {b : (a, b) ∈ A} has positive measure}
has positive measure.
Lemma 4.32 If R is a well-ordering of R of order type ω1 , then R is not
Lebesgue measurable.
Proof Suppose R is Lebesgue measurable. We consider For each x ∈ [0, 1],
Rx = {y : yRx} is countable and hence measure zero. By Fubini’s Theorem,
R has measure zero. We now exchange the order of integration. For each x,
Rx = {y : xRy} has a measure zero complement. Thus, by Fubini’s Theorem,
R has a measure zero complement, a contradiction.
Corollary 4.33 If V = L, then there is a nonmeasurable ∆12 -set.
Fubini’s Theorem has a category analog.
Let X be a Polish space and suppose A ⊆ X × X. For x ∈ X let Ax = {y ∈
X : (x, y) ∈ A}.
Theorem 4.34 (Kuratowski-Ulam Theorem) If A has the Baire property,
then A is nonmeager if and only if {a ∈ X : Aa is nonmeager} is nonmeager.
For a proof see [6] 8.41.
Exercise 4.35 Show that if V = L, then there is a ∆12 -set that does not have
the Baire property.
On the other hand we (probably) can’t prove in ZFC that there is a projective
set where any of the regularity properties above fail.
Theorem 4.36 (Solovay) If ZFC + ∃κ κ inaccessible is consistent then so
is ZFC + every uncountable projective set contains a perfect subset + every
projective set is Lebesgue measurable and has the property of Baire.
See [9] §42 for Solovay’s proof. The same arguments also show that if ZFC
+ ∃κ κ inaccessible is consistent, then so is ZF+every set of reals is Lebesgue
measurable and has the Baire property.
By 4.18 the consistency of an inaccessible is needed to prove the consistency
of every uncountable Π11 -set containing a perfect subset. Shelah has shown that
it is also needed to prove the consistency of all projective sets being measurable,
but not to prove the consistency of all projective sets having the Baire property.
42
The Effective Projective Hierarchy
We also have effective analogs of the projective point classes. Let X = Nk × N l
for some k, l ∈ N.
Definition 4.37 We say that A ⊆ X is Σ11 if there is a B ⊆ N × X such that
B ∈ Π01 and A = {x : ∃y (y, x) ∈ B}.
We say A ⊆ X is Π1n if X \ A is Σ1n and we say that A ⊆ X is Σ1n+1 if there
is a B ⊆ N × X with B ∈ Π1n such that A = {x : ∃y (y, x) ∈ B}.
We say A is ∆1n if it is both Σ1n and Π1n .
The next theorem summarizes a number of important properties of the
classes Σ1n and Π1n . We leave the proofs as exercises.
Theorem 4.38 i) The classes Σ1n and Π1n are closed under union, intersection,
∃n ∈ N, ∀n ∈ N and computable inverse images.
ii) If A ⊆ X × N is arithmetic, then {x : ∃y(x, y) ∈ A} is Σ1n .
iii) There is U ⊆ N × X a Σ1n -set that is Σ1n -universal.
iv) There is V ⊆ N × X a Σ1n -set that is Σ1n -universal.
v) Σ1n ⊂ ∆1n+1 , but Σ1n 6= ∆1n .
vi) The set WF of wellfounded trees is Π11 .
5 Coanalytic Sets
In this section we will study the structure of Π11 -sets. Because any two un-
countable Polish spaces are Borel isomorphic, it will be no loss of generality to
restrict our attention to the Baire space.
We begin by giving a normal form for Σ11 and Π11 -sets.
Definition 5.1 We say that T ⊆ N<ω × N<ω is a tree if:
i) |σ| = |τ | for all (σ, τ ) ∈ T ;
ii) if (σ, τ ) ∈ T and n ≤ |σ|, then (σ|n, τ |n) ∈ T .
If f, g ∈ N we say that (f, g) is a path through T if (f |n, g|n) ∈ T for all
n ∈ N. We let [T ] be the set of all paths through T .
Exercise 5.2 Show that F ⊆ N × N is closed if and only if there is a tree
T ⊆ N<ω × N<ω such that F = [T ].
If A ⊆ N is Σ11 , then there is C ⊆ N × N closed such that A = {x :
∃y (x, y) ∈ A}. Let T be a tree such that [T ] = C. For each x ∈ N , let
T (x) = {σ ∈ N<ω : (x|n, σ) ∈ T for some n}.
Then T (x) is a tree and x ∈ A if and only if there is f ∈ [T (x)]. Let T r be as
in 2.13.3 Then x 7→ T (x) is a continuous map from N to T r and x ∈ A if and
only if T (x) is ill-founded (i.e., not well-founded).
<ω
3 Although T r was defined to be a subset of 2 , we will (by suitable coding) view it as
a subset of C or N .
43
Let IF be the set of ill-founded trees. We saw in §3 that WF the set of well-
founded trees is Π11 , thus IF is Σ11 . We have just argued that IF is Σ11 -complete.
In particular, IF is Σ11 but not Borel.
If A is Π11 , then (N \ A) ≤w IF, so A ≤w WF. We summarize these results
in the following theorem.
Theorem 5.3 (Normal form for Π11 ) If A ⊆ N is Π11 , then there is a tree
T ⊆ N<ω × N<ω such that x ∈ A if and only if T (x) ∈ WF.
Ranks of Trees
Our analysis of Π11 -sets starts with an analysis of trees.
Definition 5.5 If T is a tree, let T 0 = {σ ∈ T : ∃τ ∈ T σ ⊂ τ } be the subtree
of nonterminal nodes of T . For α < ω1 define T α as follows:
i) T 0 = T ;
ii) T α+1 =T(T α )0 ;
iii) T α = β<α T β for α a limit.
Lemma 5.6 For any tree T there is an α < ω1 , such that T α = T β for all
β > α.
44
Proof
i) If σ ⊂ τ and τ ∈ T α , then σ ∈ T α+1 .
ii) By i) ρ(σ) ≥ sup{ρ(σbi) + 1 : σbi ∈ T }. On the other hand if α =
sup{ρ(σbi) + 1 : σbi ∈ T }, then σ has no extensions in T α so σ 6∈ T α+1 . Thus
ρ(σ) = sup{ρ(σbi) + 1 : σbi ∈ T }.
iii) If ρ(σ) = ∞, then by ii) there is σbi ∈ T with ρ(σbi) = ∞. This allows
us to inductively build f ∈ [T ] with f ⊃ σ.
iv) Clear from i)–iii).
Exercise 5.9 a) Show that if T 6= ∅, then ρ(T ) = ρT (∅).
b) Show that for all α < ω1 there is a tree T with ρ(T ) = α.
c) If T is a tree and σ ∈ N<ω , let Tσ = {τ ∈ N<ω : σbτ ∈ T }. Show that Tσ
is a tree and if T 6= ∅, then ρ(T ) = supn∈N (ρ(Thni ) + 1).
Lemma 5.11 i) If S, T ⊆ N<ω are trees, then ρ(S) ≤ ρ(T ) if and only if there
is an order preserving f : S → T .
ii) If T is a well-founded tree, then ρ(S) < ρ(T ) if and only if S = ∅ and
T 6= ∅ or there is n ∈ N and f : S → Thni order preserving.
Proof
i) If f : S → T is order preserving, then an easy induction on rank shows
that ρS (σ) ≤ ρT (f (σ)) for all σ ∈ S. Thus ρ(S) ≤ ρ(T ). For the converse, we
build f by induction such that ρS (σ) ≤ ρT (f (σ)) for all σ ∈ S. Let f (∅) = ∅.
Suppose we have defined f (σ) with ρS (σ) ≤ ρT (f (σ)) and σbi ∈ T . By 5.8 ii)
and iii) there is j ∈ N such that f (σ)bj ∈ T and ρT (f (σ)bj) ≥ ρS (σbi). Let
f (σbi) = f (σ)bj.
ii) If f : S → Thni is order preserving. Then
Conversely, if ρ(S) < ρ(T ) and S 6= ∅, then there is n ∈ N such that ρ(S) ≤
ρ(Thni ) and by i) there is an order-preserving f : S → Thni .
If α < ω1 , let WFα = {T ∈ WF : ρ(T ) < α}. We will show that WFα is
Borel.
Since T 7→ Thni
S is continuous, by induction, WFα is Borel. If α is a limit ordinal,
then WFα = β<α WFβ . Thus WFα is Borel.
45
Ranks of Π11 sets
We can now see how Π11 -sets are built up from Borel sets.
Corollary 5.14 If A ∈ Π11 and |A| > ℵ1 , then A contains a perfect set. In
particular, |A| ≤ ℵ1 or |A| = 2ℵ0 .
S
Proof Let A = α<ω1 Aα where Aα is Borel. If any Aα is uncountable, then
A contains a perfect set. Otherwise |A| ≤ ℵ1 .
It is consistent with ZFC that there is a Π11 -set that has cardinality ℵ1 < 2ℵ0 .
For example, this is true in any model of ZFC where ℵL1 = ℵV 1 <2 .
ℵ0
Proof
i)
ρ(S) ≤ ρ(T ) if and only ∃f : S → T order-preserving.
ii) For T ∈ WF,
T ∈ WF ⇔ ∃S (S ∈ A ∧ ρ(T ) ≤ ρ(S)}
46
We will recast the work we just did in more modern language introduced by
Moschovakis. This point of view is useful when one attempts to extend these
ideas to higher levels of the projective hierarchy.
Definition 5.17 A norm on a set A is a function φ : A → On where On is the
class of ordinals.
Suppose A ∈ Π11 . We say that φ : A → On is a Π11 -norm if there are
Π1 Σ1
relations ≤φ 1 ∈ Π11 (N × N ) and ≤φ 1 ∈ Σ11 (N × N ) such that if y ∈ A, then
Π1
x ∈ A ∧ φ(x) ≤ φ(y) ⇔ x ≤φ 1 y
Σ1
⇔ x ≤φ 1 y.
and
x ∈ B0 if and only if ¬(ρ(f (x)) ≤ ρ(g(x))).
By 5.15 A0 and B0 are Π11 .
Definition 5.21 We say that Γ has the separation property if whenever A, B ∈
Γ and A ∩ B = ∅ there is C ∈ Γ ∩ Γ̌ such that A ⊆ C and C ∩ B = ∅.
Lemma 5.22 If Γ̌ has the reduction property, then Γ has the separation prop-
erty.
47
Proof Suppose A, B ∈ Γ and A ∩ B = ∅. Then X \ A, X \ B ∈ Γ̌ and
X \ A ∪ X \ B = X. Let C ⊆ X \ B, D ⊆ X \ A such that C, D ∈ Γ̌, C ∩ D = ∅
and C ∪D = X. Then C = X \D so C ∈ Γ. If x ∈ A, then x ∈ (X \B)\(X \A).
Thus x ∈ C. Similarly if x ∈ B, then x ∈ D = X \ C. Thus C separates A and
B.
This gives a different proof that Σ11 has the separation property. We next
show that it is harder to separate Π11 -sets.
x 6∈ C ⇒ (b, x) ∈ U ⇒ x ∈ P ⇒ x ∈ Q ⇒ (a, x) ∈ U ⇒ x ∈ C
a contradiction.
Uniformization
48
Proof By 5.23 there are Π11 -sets A0 , A1 ⊆ N such that A0 ∩ A1 = ∅ but there
is no Borel set B with A0 ⊆ B and A1 ∩ B = ∅.
There are closed sets C0 , C1 ⊆ N such that
X \ Ai = {x : ∃y (x, y) ∈ Ci }.
49
iii) if (x, z) ∈ An and z(n) = y(n), then ρ(T (x, y)σn ) ≤ ρ(T (x, z)σn ).
In other words, we first find mx,n minimal such that there is a z with (x, z) ∈
An and z(n) = mx,n , we then find αx,n minimal such that there is (x, z) ∈ An
with z(n) = mx,n and ρ(T (x, z)σn ) = αx,n . Then (x, y) ∈ An+1 if and only if
(x, y) ∈ An , y(n)
T = mx,n and ρ(T (x, y)σn ) = αx,n .
Let B = n An . We will show that B is a Π11 -uniformization of A. Let π
be the projection (x, y) 7→ x. If x ∈ π(A), then, by induction, x ∈ π(An ) for all
n. Define yx ∈ N by yx (n) = mx,n .
Claim 1 If (x, y) ∈ B, then y = yx .
If (x, y) ∈ An , y(n) = mx,n , thus y = yx .
We need to show that (x, yx ) ∈ B for all x ∈ π(A). Fix x ∈ π(A).
Claim 2 Suppose σi , σj ∈ T (x, yx ) and σi ⊂ σj . Then αx,j < αx,i .
Choose (x, z) ∈ Aj+1 . Since i < j, (x, z) ∈ Ai+1 . Since z|j + 1 = y|j + 1 and
|σj | ≤ j, σj ∈ T (x, z). Thus
αx,i = ρ(T (x, z)σi ) > ρ(T (x, z)σj ) = αx,j ,
as desired.
Claim 3 (x, yx ) ∈ A.
If σi0 ⊂ σi1 ⊂ . . . is an infinite path through T (x, yx ), then, by claim 2,
αx,1 > αx,2 > . . . a contradiction. Thus T (x, yx ) is well founded and (x, yx ) ∈ A.
Claim 4 (x, yx ) ∈ B.
An induction on T (x, yx ) shows that if σn ∈ T (x, yx ), then ρ(T (x, yx )σn ) ≤
αx,n . By choice of αx,n , we have ρ(T (x, yx )σn ) = αx,n and (x, yx ) ∈ An for all
n.
Thus B is the graph of a function uniformizing A.
Claim 5 B is Π11 .
Define R(x, y, n) by
∃z[(∀k < n (y(k) = z(k) ∧ ρ(T (x, y)σk ) = ρ(T (x, z)σk ))∧
(z(n) < y(n) ∨ (z(n) = y(n) ∧ ρ(T (x, z)σn ) < ρ(T (x, y)σn ).
By 5.15 R is Σ11 . Suppose (x, y) ∈ An . If (x, y) 6∈ An+1 , then either y(n) 6= mx,n
or ρ(T (x, y)σn ) 6= αx,n . In either case R(x, y, n) holds. On the other hand,
suppose R(x, y, n) and z witnesses the existential quantifier. Since (x, y) ∈ An ,
(x, z) ∈ An . It is easy to see that either y(n) 6= mx,n or ρ(T (x, y)σn ) 6= αx,n .
Thus (x, y) 6∈ An+1 . Thus
(x, y) ∈ B ↔ (x, y) ∈ A ∧ ∀n ¬R(x, y, n)
and B is Π11 .
Corollary 5.30 Σ12 has the uniformization property.
Proof Suppose A ⊆ N × N is Σ12 . There is a Π11 -set B ⊆ N 3 such that
b⊆B
A = {(x, y) : ∃z (x, y, z) ∈ B}. By Kondo’s Theorem, there is a Π11 -set B
such that
b
π(A) = {x : ∃y∃z (x, y, z) ∈ B} = {x : ∃y∃z (x, y, z) ∈ B}
50
b Let
and for all x ∈ π(A) there is a unique pair (y, z) such that (x, y, z) ∈ B.
b = {(x, y) : ∃z (x, y, z) ∈ B}.
A b Then π(A)
b = π(A) and for all x ∈ π(A) there is
a unique y such that (x, y) ∈ B.b
Π11 -sets
Many of the proofs in this section work just as well for Π11 -sets. Here are
statements of the effective versions.
Recursive Ordinals
The set WF is Π11 . If A ⊆ N is Π11 , we know that A ≤w WF. We will show that
the reduction f can be chosen computable. There is a recursive tree T , such
that
x ∈ A ⇔ ∀y (x, y) 6∈ [T ] ⇔ T (x) ∈ WF.
The function x 7→ T (x) is computable.
51
A similar construction gives rise to a Π11 -complete (for ≤m ) subset of N.
Let O = {e ∈ N : φe is the characteristic function of a Well-founded tree
Te ⊆ N<ω }. Then e ∈ O if and only if
i) ∀σ φe (σ) ↓
ii) ∀σ ∈ N<ω ∀τ ∈ N<ω ((σ ⊆ τ ∧ φe (τ ) = 1) → φe (σ) = 1).
iii) ∀f : N → N<ω ∃n (φe (f (n)) = 0 ∨ φe (f (n + 1)) = 0 ∨ f (n) 6⊂ f (n + 1)).
Conditions i) and ii) are Π02 while iii) is Π11 . Thus O is Π11 .
Proposition 5.35 O is Π11 -complete.
Proof We will argue that N \ O is Σ11 -complete. Suppose A ∈ Σ11 . There is
B ⊆ N × N in Π01 such that n ∈ A if and only if ∃x (n, x) ∈ B. There is a
recursive tree T ⊆ N × N<ω such that (n, x) ∈ A if and only if (n, x|m) ∈ T for
all m ∈ N. There is a recursive f : N → N such that φf (n) is the characteristic
function of {σ : (n, σ) ∈ T }. Then φf (n) is the characteristic function of a tree
Tn and
n ∈ A ⇔ Tn 6∈ WF ⇔ f (n) 6∈ O.
52
Lemma 5.40 a) If α is a recursive ordinal and β < α, then β is a recursive
ordinal.
b) If α is a recursive ordinal, then α + 1 is a recursive ordinal.
c) Suppose f : N → N, g : N → N are recursive functions such that Pf (n) is
a program to compute the characteristic function of An , Pg(n) is a program that
computes the characterisitic function of ≺n a well-order of An and (An , ≺n ) has
order-type αn . Then sup αn is a recursive ordinal.
P
Proof a) and b) are routine. For c) we show that An is a recursive well-
order. Let A = {(n, m) : Mf (n) (m) = 1} and let (n, m) ≺ (n0 , m0 ) if and only
if n < n0 or n = n0 and m ≺n m0 . Then (A, ≺) is a recursive well-order. Let α
be the order type of A. Then αn ≤ α for all n. Since sup αn ≤ α, sup αn is a
recursive ordinal.
There are only countably many recursive well-orders. Thus there are only
countably many recursive ordinals.
Definition 5.41 Let ω1ck be the least non-recursive ordinal. We call this ordinal
the Church–Kleene ordinal.
More generally for any x we let ω1x be the least ordinal not recursive in x.
We need to be able to compare ordinals with trees.
Definition 5.42 For σ, τ ∈ N<ω we say σ / τ if τ ⊂ σ or there is an n such that
σ(n) 6= τ (n), but σ(m) = τ (m) for all m < n. We call / the Kleene–Brower
order.
Exercise 5.43 a) / is a linear order of N<ω .
b) If T ⊆ N<ω is a tree, then T is well-founded if and only if (T, /) is a
well-order.[Hint: If σ0 , σ1 , . . . is an infinite descending sequence in (T, /), define
x inductively by x(n) = least m such that (x(0), . . . , x(n − 1), m) / σi for some
i. Prove that x ∈ [T ].]
c) Prove that ω1ck = sup{ρ(T ) : T ⊆ N<ω a recursive well founded tree}.
The proof of 5.15 actually shows the following.
53
6 Determinacy
In this section we will introduce another logical tool that sheds new light on
Borel, analytic, and coanalytic sets, and is indispensable in the study of higher
levels of the projective hierarchy.
Let X be any nonempty set and let A ⊆ X N . We define an infinite two
player game G(A). Players I and II alternate playing elements of X. Player I
plays x0 , Player II replies with x1 , Player I then plays x3 . . . . A full play of the
game looks like this.
Player I Player II
x0
x1
x2
x3
x4
x5
.. ..
. .
Player I Player II
τ (∅)
x0
τ (x0 )
x1
τ (x0 , x1 )
x2
τ (x0 , x1 , x2 )
.. ..
. .
54
Definition 6.3 We say that the game G(A) is determined if either Player I or
Player II has a winning strategy.
We first show that if A is not too complicated, then G(A) is determined. We
consider X with the discrete topology and X N with the product topology.
55
Theorem 6.9 (Borel Determinacy) If A ⊆ N is Borel, then G(A) is deter-
mined.
x = σ0bj0bσ1bj1 σ2bj2 . . . ∈ C.
f (x) = τ (∅)bx(0)bτ (x0 )bx1bτ (x(0), x(1))bx2bτ (x(0), x(1), x(2), x(3)) . . . .
Then f (x) ⊃ µbx(n) and f (y) ⊃ µby(n). Thus f (x) 6= f (y). Thus f is continu-
ous and one-to-one. Hence f (C) is an uncountable closed subset of A.
56
Proposition 6.13 If Player II has a winning strategy in G∗ (A), then A is
countable.
Corollary 6.16 If PD holds, the any uncountable projective set contains a per-
fect subset.
57
There is a technique of “unfolding” games, that allows us to show that if
Det(Σ1n ) holds, then every uncountable Σ1n+1 set contains a perfect subset. We
will illustrate this idea by giving another proof of the perfect set theorem for
Σ11 -sets using only the determinacy of closed games.
Suppose A ⊆ C is Σ11 . Let B ⊆ C × N such that A = {x : ∃y (x, y) ∈ B}.
Consider the game G∗u (A) where at stage i Player I plays σi ∈ 2<ω and y(i) ∈ N
and Player II responds with ji ∈ 2. Together they play
x = σ0bj0bσ1bj1b. . .
and
y = (y(0), y(1), . . .).
Player I wins if (x, y) ∈ A. By closed determinacy (or more correctly by 6.7),
G∗u (A) is determined.
Lemma 6.17 If Player I has a winning strategy in G∗u (A), then A contains a
perfect subset.
Lemma 6.18 If Player II has a winning strategy in G∗u (A), then A is countable.
x(m) = 1 − τ (σ0 , y(0), . . . , σn−1 , y(n − 1), hx(0), . . . , x(m − 1)i, y(n)i.
Indeed for each possible value of y(n), there is at most one x rejected at p. Thus
the set of x rejected at p is countable and A is countable.
Lemmas 6.17 and 6.18 together with the determinacy of closed games gives
a second proof of the Perfect Set Theorem for Σ11 .
In §7 we will examine this game again. At that time it will be useful to note
that if x is rejected at p, then x is recursive in τ .
Banach-Mazur Games
We will show that, assuming Projective Determinacy, all projective sets have
the Baire property. Unfolding this argument will prove in ZFC that all Σ11 sets
have the Baire property (and hence all Π11 -sets have the Baire property).
Let A ⊆ N . Consider the Banach–Mazur game G∗∗ (A) where at stage i,
Player I plays σ2i ∈ N<ω and Player II playsSσ2i+1 ∈ N<ω such that σ0 ⊂ σ1 ⊂
σ2 ⊂ . . .. The final play of the game is x = σn and Player I wins if x ∈ A.
58
Lemma 6.19 Player II has a winning strategy in G∗∗ (A) if and only if A is
meager.
Proof S
(⇐) Suppose A = n An where each Ai is nowhere dense. We informally
describe a winning strategy for Player II. If, at stage i, Player I plays σ2i ∈ N<ω ,
then Player II plays σ2i+1 ⊃ σ2i such that NSσ2i+1 ∩ Ai = ∅. Since each Ai is
nowhere dense this is always possible. If x = σn is the final play of the game,
then, for each i,
x ∈ Nσ2i+1 ⊆ N \ A.
Thus this is a winning strategy for Player II.
(⇒) Suppose τ is a winning strategy for Player II. Suppose x ∈ A. Let
p = (σ0 , . . . , σ2m−1 ) be a position in the game where Player II has used τ . We
say that x is rejected at p if and only if x ⊃ σ2m−1 but for all σ2m ⊃ σ2m−1 if
x ⊃ σ2m−1 , then x 6⊃ τ (σ0 , σ2 , . . . , σ2m ).
Claim If x ∈ A, then there is a position p = (σ0 , . . . , σ2m−1 ) such that x ⊃
σ2m−1 and x is rejected at p.
Suppose not. Because x is not rejected at ∅, there is σ0 such that x ⊃ τ (σ0 ).
Inductively we build σ0 , σ2 , . . . such that
x ⊃ τ (σ0 , σ2 , . . . , σ2m )
for all m. But then if Player I plays σ0 , σ2 , . . . and Player II uses τ , then the
eventually play x ∈ A contradicting the claim that τ is a winning strategy for
Player II.
Let Rp = {x ∈ A : x is rejected at position p}.
Claim Rp is nowhere dense.
Note that Rp ⊆ Nσ2m−1 . For all σ ⊃ σ2m−1 let ησ = τ (σ0 , . . . , σ2m−1 , σ).
Then ησ ⊃ σ and Rp ∩ Nησ = ∅. Thus Rp is nowhere dense.
S
Thus A = p Rp is meager.
Lemma 6.20 If Player I has a winning strategy in G∗∗ (A), then there is η ∈
N<ω such that Nη \ A is meager.
Proof Let τ be Player I’s winning strategy for G∗∗ (A). Suppose Player I’s
first move in G∗∗ (A) is η. We will show that Nη \ A is meager, by showing that
Player II has a winning strategy τb in G∗∗ (Nη \ A).
Let
τb(σ0 , σ2 , . . . , σ2m ) = τ (σ0 , σ2 , . . . , σ2m ).
In other words Player II plays G∗∗ (Nη \ A), by pretending to be Player I using
τ in a game of G∗∗ (A).
If Player’s I first move in G∗∗ (Nη \ A) is σ0 , Player II checks to see how
Player I would reply if Player II played σ0 in G∗∗ (A). The following picture
describes the play of the games.
59
G∗∗ (Nη \ A) G∗∗ (A)
Player I Player II Player I Player II
σ0 η
τ (σ0 ) σ0
σ1 τ (σ0 )
τ (σ0 , σ1 ) σ1
σ2 τ (σ0 , σ1 )
τ (σ0 , σ1 , σ2 ) σ2
.. .. .. ..
. . . .
If x is a play of G∗∗ (Nη \ A) where Player II uses the strategy τb, then x is
also a play of G∗∗ (A) where Player I uses τ . Thus x ∈ Nη ∩ A and Player II
wins the play of G∗∗ (Nη \ A). Thus τb is a winning strategy for Player II and
Nσ \ A is meager.
Theorem 6.21 Assuming Projective Determinacy all projective sets have the
Baire property.
Proof Let A ⊆ N . If Player II has a winning strategy in G∗∗ (A), then by 6.20
A is meager.
<ω
S let S = {σ ∈ N
Suppose Player I has a winning strategy, : Nσ \ A is
meager}. By 6.19, S is nonempty. Let U = σ∈S Nσ \ A. Then
[
U \A = Nσ
σ∈S
Further Results
Projective Determinacy can also be used to prove that all projective sets are
Lebesgue measurable (see [6] §21 or [9] §43).
Can every projective set be uniformized by a projective set? If V = L, then
we can use the ∆12 well-ordering of N to show that they can. Moschovakis
showed that Projective Determinacy also leads to an interesting answer.
60
One key idea is to use determinacy to build Π12n+1 prewellorderings. For
proofs see [6] §39.
Another interesting class of games are the Wadge games. Suppose A, B ⊆ N .
Consider the game Gw (A, B) where Player I plays x(0), x(1), . . ., and Player II
plays y(0), y(1) with x(i), y(i) ∈ N. Player II wins if x ∈ A if and only if x ∈ B.
Proof Suppose B ∈ Σ0α and B 6≤w A. Then Player II does not have a winning
strategy in Gw (B, A). By Borel Determinacy, Player I has a winning strategy.
Thus A ≤w (N \ B) and A ∈ Π0α , a contradiction.
Exercise 6.26 Show that under Projective Determinacy and non-Borel Σ11 -set
is Σ11 -complete.
We write A <w B if A ≤w B but B 6≤w A.
61
x ∈ A. Show that if τ is a winning strategy for Player I, then Cτ ⊆ A, while if
τb is a winning strategy for Player II, then Cτ ∩ A = ∅.]
c) Let Ω be the collection of Turing-invariant Borel subsets of N . Show that
Ω is σ-algebra.
d) Let µ : Ω → 2 be µ(A) = 1 if and only if A contains a cone. Show that µ
is a σ-additive measure on Ω. µ is called the Martin-measure.
Assuming Projective Determinacy we can consider projective sets instead of
Borel sets.
The axiom of choice tells us there are undetermined games. It is interesting
to abandon the axiom of choice and consider ZF with the Axiom of Determinacy
AD which asserts that all games are determined. While AD is refutable from
ZFC, it is consistent with large cardinals that ZFC + L(R) |= AD. ZF+AD has
wild consequences. For example:
7 Hyperarithmetic Sets
Our first goal is to try to characterize the ∆11 -sets. In particular we will try to
formulate the “light-faced” version of
∆11 = Borel.
Borel Codes
Let X = Nk × N l . Let SX be as in §3.
Definition 7.1 A Borel code for a subset of X is a pair hT, li where T ⊆ N <ω
is a well-founded tree and l : T → ({0} × {0, 1}) ∪ ({1} × S X ) such that:
i) if l(∅) = h0, 0i, then σb0 ∈ T and σbn 6∈ T for all n ≥ 1;
ii) if l(∅) = h1, ηi, then σbn 6∈ T for all n ∈ N.
Let BC be the set of all Borel codes. It is easy to see that BC is Π11 .
If x = hT, li is a Borel code, we can define B(x) the Borel set coded by x. If
σ ∈ T , recall that Tσ = {τ : σbτ ∈ T }. We let lσ : Tσ → {0} × 2 ∪ {1} × SX by
lσ (τ ) = l(σbτ ). It is easy to see that hTσ , lσ i is also a Borel code.
Definition 7.2 We define B(hT, li) inductively on the height of T .
i) B(h∅, ∅i) = ∅.
ii) If l(∅) = h1, ηi, then B(hT, li) = Nη .
iii) If l(∅) = h0, 0i, then B(hT, li) = X \ B(hTh0i , lh0i i).
62
iv) If l(∅) = h0, 1i, then
[
B(hT, li) = B(hThni , lhni i).
hni∈T
Lemma 7.4 There are R ∈ Σ11 and S ∈ Π11 such that if x ∈ BC then
Proof We define a set A such that (x, y, f ) ∈ A if and only if x is a pair hT, li
where T ⊆ N<ω is a tree, l : T → {0} × 2 ∪ {1} × SX and f : T → 2 such that
for all σ ∈ T :
i) if l(∅) = h1, ηi, then f (σ) = 1 if and only if y ∈ Nη ;
ii) if l(∅) = h0, 0i, then f (σ) = 1 if and only if f (σb0) = 0;
iii) if l(∅) = h0, 1i, then f (σ) = 1 if and only if f (σbn) = 1 for some n.
An easy induction shows that if x = hT, li is a Borel code then (x, y, f ) ∈ A
if and only if f is the function
63
Proposition 7.7 If α < ω1ck , then WFα is a recursively coded Borel set.
Proof We have already shown that every recursively coded Borel set is ∆11 .
Suppose A is ∆11 . Since A is Π11 , there is a computable f : X → T r such that
x ∈ A if and only if f (x) ∈ WF. The set
f (A) = {y : ∃x x ∈ A ∧ f (x) = y}
is a Σ11 -subset of WF. By Σ11 -Bounding, there is α < ω1ck such that f (A) ⊆ WFα .
By 7.7 WFα is recusively coded, and by 7.6 A = f −1 (WFα ) is recursively coded.
For notational simplicity we will assume X = N , but all our arguments
generalize easily.
Let BCrec = {(e, x) : φxe is a total function and φxe ∈ BC}. Then
We say e ∈ BCrec and x ∈ Brec (e) if (e, ∅) ∈ BCrec and x ∈ Brec (e, ∅).
The proofs of both 7.6 and 7.7 will use the Recursion Theorem to do a
transfinite induction.
We begin with the base case of the induction
64
Proof i) φe is a code for a pair hT, li. Let
T 0 = {∅} ∪ {0bη : η ∈ T }
and l0 (∅) = h0, 0i, l0 (0bη) = l(η). It is easy to find Hc such that Hc (e) codes
hT 0 , l0 i and that if e ∈ BCrec , then Hc (e) is a code for the complement.
ii) Suppose φe (n) code a pair hTn , ln i. Let
T = {∅} ∪ {nbσ : σ ∈ Tn }
and let l(∅) = h0, 1i and l(nbσ) = ln (σ). It is easy to find Hu such that Hu (e)
codes hT, li. If each hTn , ln i is a Borel code, then hT, li codes their union.
Theorem 7.6 follows from the next lemma.
Proof
We define a recursive function g : N × N × T → N as follows:
i) If l(σ) = h1, ηi, then g(i, e, σ) = F (e, η);
ii) If l(σ) = h0, 0i, then g(i, e, σ) = Hc (φi (e, σb0));
iii) Suppose l(σ) = h0, 1i. Choose j such that φj (n) = φi (e, σbn). Then
g(i, e, σ) = Hu (j).
By the Recursion Theorem, there is bi such that φi (e, σ) = g(bi, e, σ) for all
e, σ. Let G(e, σ) = φi (e, σ).
We prove by induction on T , that G(e, σ) is a code for f −1 (B(hTσ , lσ i). By
i) this is clear if l(σ) = h1, η). We assume the claim is true for all τ ⊃ σ.
If l(σ) = h0, 0i, then
If l(σ) = h0, 1i, then G(e, σbn) is a Borel code for An = f −1 (B(hTσbn, lσ n i).
We choose j such that φj (n) is a code for An and G(e, σ) = Hu (j) is a code for
S
An .
Theorem 7.7 follows from the next lemma.
65
Note that ρ(S) ≤ ρ(T ) if and only if for all n ∈ N there is m ∈ N such that
ρ(Shni ) ≤ ρ(Thmi ). Thus
\ [
−1
{S ∈ T r : ρ(S) ≤ ρ(Tσ )} = fhni ({S ∈ T r : ρ(S) ≤ ρ(Tσ m )}).
n∈N m∈N
We can do this using the functions F, Hu and Hc above. Of course for some i,
this may well be undefined.
By the Recursion Theorem there is bi such that φi (σ) = g(bi, σ) for all σ.
An easy induction shows that G = φi is the desired function.
Hyperarithmetic Sets
66
Proof Clearly O is Π11 (x). There is T recursive in x such that T ∈ WF and
ρ(T ) > ω1ck . Then
O = {e : e codes a recursive tree S and ρ(S) < ρ(T )}
is Σ11 (x). Thus O ≤hyp x.
We delay the proof to the end of the section and look at some important
corallaries.
67
Corollary 7.21 Suppose A ⊆ N × N is a Borel set such that every section is
countable. Then X the projection of X is Borel and X can be uniformized by a
Borel set.
Theorem 7.23 If Player II has a winning strategy in G(A), then Player II has
a hyperarithmetic winning strategy.
68
that if Player II has a winning strategy, then there is a hyperarithmetic winning
strategy τ . Then A is countable. The proof of 6.18 shows that every x ∈ A is
rejected at some position and x ≤T τ . Thus x is hyperarithmetic and A ⊆ HYP.
Exercise 7.25 Show that if A is Σ11 and uncountable, then there is a continuous
injection f : C → A with f computable in x for some x ≤hyp O.
Proof Let
C is ∆11 . Let ½
0 if i 6∈ C
x(i, n) = .
m if i ∈ D ∧ (n, m) ∈ Brec (i)
Then A ⊆ {x0 , x1 , . . .}.
Exercise 7.28 Show that the same is true if A is Σ11 . [Hint: First show that
any Σ11 subset of HYP is contained in an ∆11 subset of HYP.]
Proof of Theorem 7.26 By relativizing, we assume that A is ∆11 . Suppose
A ⊆ N × N has countable sections. By the Effective Perfect Set Theorem,
for any x the set Ax = {y : (x, y) ∈ X} is a ∆11 (x) subset of {y : y ≤hyp x}.
By relativising the lemma, there is y ∈ NN such that y ≤hyp x and Ax ⊆
2
{y0 , y1 , . . .}.
Let B = {(x, j) ∈ N × N : j ∈ BCrec (x) ∧ ∀z ((x, z) ∈ A →
∃n∀m∀k (z(m) = k) ↔ (n, m, k) ∈ Brec (x, j)}. Then B is Π11 and π(B) = π(A).
By 7.21 π(A) is ∆11 . Thus, by selection, there is a ∆11 function t : π(A) → N
uniformizing B.
69
Let gn : π(A) → N be such that gn (x) = y if and only if
Since gn has a ∆11 -graph, it is Borel measurable and A is contained in the union
of the graphs of the gn .
Let ( )
n−1
^
Cn = x ∈ π(A) : (x, gn (x)) ∈ A ∧ gn (x) 6= gi (x)
i=0
and let fn = gn |Cn . Each gn is Borel measurable and A is the disjoint union of
the graphs of the gn .
Lemma 7.29 Suppose AS⊆ N is ∆11 and open. Then there is a hyperarithmetic
S ⊆ N<ω such that A = σ∈S Nσ .
S
Proof Let S = {σ : ∀x x ⊃ σ → x ∈ A}. Then S is Π11 and A = σ∈S Nσ .
There is a computable f : N<ω → T r such that
Let
B = {τ : ∀x ∈ A∃σ ρ(f (τ )) 6≤ ρ(f (σ)) ∧ σ ⊂ x}.
Note that B is Π11 and N<ω \ S ⊆ B. If B = N<ω \ S, then S is ∆11 , as desired.
If not, there is τ ∈ S ∩ B. Then S0 = {σ : f (σ) < f (τ )} is ∆11 and
[ [
A= Nσ ⊆ Nσ = A.
σ∈S0 σ∈S
S
If A is ∆11 and closed, then there is a ∆11 -set S ⊆ N<ω with N \A = σ∈S Nσ .
Let T = {σ ∈ N<ω : ∀τ ⊆ σ τ 6∈ S}. Then T is a hyperarithmetic tree and
A = [T ]. If A is compact, we can go a bit further.
Lemma 7.30 If A is ∆11 and compact, then there is a finite branching hyper-
arithmetic tree T such that A = [T ]. More generally, if A is a compact Σ 11 -set, F
is a closed ∆11 -set, and A ⊆ F , then there is a finite branching hyperarithmetic
tree T such that A ⊆ [T ] ⊆ F .
70
Proof By the remarks above, there is a hyperarithmetic tree T such that
F = [T ].
Let BV= {(σ, C) : σ ∈WN<ω ∧ C ⊂ N finite ∧ ∀x ((x ∈ A ∧ σ ⊂ x) →
( i∈C σbi ∈ T ∧ i∈C x ⊃ σbi).
Then B is Π11 . If Nσ ∩A = ∅, then (σ, C) ∈ B for all finite C ⊂ N. If Nσ ∩A 6= ∅,
then, by compactness, there is a finite set C ⊂ N such that
[
A ∩ Nσ = A ∩ Nσ i .
i∈C
Clearly (σ, C) ∈ B. By 5.28 there is a ∆11 -function f such that if σ ∈ N<ω , then
(σ, f (σ)) ∈ B. Let
^
T1 = {σ ∈ T : σ(i) ∈ f (σ|i)}.
i<|σ|
Proof Clearly π(A) is Σ11 . By relativizing the previous corollary, we see that
71
ExerciseS 7.33 Suppose A ⊆ N is closed, f : N → N is continuous and
f (A) ⊆ Fn where each Fn is closed. There is σ ∈ N<ω and n ∈ N such that
A ∩ Nσ 6= ∅ and f (A ∩ Nσ ) ⊆ Fn . [Hint: Suppose not. BuildS σ0 ⊂ σ1 ⊂ . . .
such that A ∩ Nσi 6= ∅ and f (Nσi+1 ∩ Fi ) = ∅. Consider x = σi to obtain a
contradiction.]
Definition 7.34 We say that A is a Kσ set if it is a countable union of compact
sets.
Since Rn is locally compact, in Rn every Fσ -set is a Kσ -set.
Lemma 7.35 Suppose A is ∆11 and a Kσ -set. Then there is x ∈ A ∩ HYP.
Proof There is a Π01 -set B ⊆ N × N such that A is the projection of B. By
the previous exercise, there is a basic open set N such
A1 = {x : ∃y (x, y) ∈ N × B}
is Σ11 and contained in a closed subset of F of A. Then
A1 = {x : ∀σ (σ ⊂ x → ∃y (y ∈ A1 ∧ σ ⊂ y))}
the closure of A1 is also Σ11 and contained in A.
Let B = {(x, σ) : x 6∈ A ∧ σ ⊂ x ∧ ∀y (y ⊃ σ → y 6∈ A1 }. Then B is Π11 and
for all x 6∈ A there is a σ such that (x, σ) ∈ B. By 5.28 there is a ∆11 function
f such that (x, f (x)) ∈ B for all x 6∈ A.
Let W0 = {f (x) : x 6∈ A}, let W1 = {σ : ∀y ⊃ σ y 6∈ A1 }. Then W0 is
Σ11 , W1 is Π11 and W0 ⊆ W1 . By Σ11 -separation, there is a ∆11 -set W such that
W0 ⊆ W ⊆ W1 . Let
T = {σ : ∀τ ⊆ σ : τ 6∈ W }.
Then T is a ∆11 tree. Since W ⊆ W1 , A1 ⊆ [T ]. Since W0 ⊆ W , [T ] ⊆ A.
By 7.30 there is a finite branching tree T1 ∈ ∆11 such that T1 ⊂ T and
A1 ⊆ [T1 ] ⊆ [T ] ⊆ A.
As in 7.31 there is x ∈ [T1 ] ∩ HYP.
Corollary 7.36 (Aresenin, Kunugui) If A ⊆ N × N is Borel and every
section if Kσ , then π(A) is Borel and A has a Borel uniformization.
These proofs are a little unsatisfactory as we have only proved the uniform-
results for N × N or more generally recursively presented Polish spaces (like
Rn ). Since “compact” and “Fσ ” are not preserced by Borel isomorphisms we
can not immediately transfer these results to arbitrary Polish spaces. In fact
these results are true in general (see [6] 35.46).
Exercise 7.37 a) Modify the proof of the Effective Perfect Set Theorem, using
the Banach–Mazur game, to prove that if A ⊆ N is a nonmeager ∆11 -set, then
there is a hyperarithmetic x ∈ A.
b) Prove that any Borel set with nonmeager sections can be uniformized by
a Borel set.
72
Part II
Borel Equivalence Relations
The second half of these notes will be concerned with the descriptive set theory
of equivalence relations.
Part of our interest in Descriptive Set Theory is motivated by Vaught’s
Conjecture. Suppose L is a countable language and T is an L-theory. Let
I(T, ℵ0 ) be the number of isomorphism classes of countable models of T .
Vaught’s Conjecture If I(T, ℵ0 ) > ℵ0 , then I(T, ℵ0 ) = 2ℵ0 .
Of course if the Continuum Hypothesis is true, Vaught’s Conjecture is true.
But perhaps it is provable in ZFC (though at the moment there is a manuscript
with a plausible counterexample due to Robin Knight).
We have seen before that Mod(T ) is a Polish space and ∼= is a Σ11 -equivalence
relation on Mod(T ). A first hope would be to deduce Vaught’s Conjecture from a
perfect set theorem for Σ11 -equivalence relations. This won’t work. For example,
consider the following equivalence relation on T r.
T ∼ S ⇔ ρ(S) = ρ(T )
Then ∼ is Σ11 . There is one equivalence class for all the ill-founded trees
and then one for each possible value of ρ. Thus ∼ has exactly ℵ1 -equivalence
classes.
We will see in §8, that while there is a perfect set theorem for Π11 (and hence
Borel) equivalence relations and a weaker perfect set theorem forΣ11 -equivalence
relations.
The rest of the notes will be concerned with two special cases of Σ11 -equivalence
relations:
i) Borel Equivalence relations
ii) Orbit Equivalence realations, suppose G is a Polish group, X is a Borel
set in a Polish space and µ : G × X → X is a continuous action of G on A.
Let EG be the equivalence relation xEG y if and only if there is g ∈ G such that
gx = y.
It is easy to see that the orbit equivalence relations EG are Σ11 . Of par-
ticular interest is the case where S∞ acts on Mod(T ). In this case EG is the
isomorphism equivalence relation on Mod(T ).
The study of these equivalence relations is also tied up with the study of the
dynamics of group actions and these ideas will also play a key role.
73
that Π11 -equivalence relations, and in particular Borel equivalence relations, are
better behaved.
Proof S
Let E ∩ U × U = An where each An is nowhere dense.
We build (Uσ : σ ∈ 2<ω ) nonempty basic clopen sets such that:
i) U∅ ⊆ U ;
ii) Uσ ⊆ Uτ for σ ⊆ τ ;
1
iii) diam Uσ < |σ| ;
iv) if |σ| = |τ | = n and σ 6= τ , then E ∩ (Uσ ∩ Uτ ) ∩ (A0 ∪ . . . ∪ An−1 ) = ∅.
T
For f ∈ C, let xf = Uf |n . By contstruction if f 6= g, then xf E 6 xg . Thus
there is a perfect set of E-inequivalent elements.
We choose U∅ an nonempty basic clopen subset of U .
Suppose we have constructed Uσ for all σ with |σ| = n satisfying i)–iv). Let
{(σi , τi ) : i = 1, . . . k} list all pairs of distinct sequences of length n + 1. If
|σ| = n + 1 we inductively define Uσi for i = 0, . . . k.
Let Uσ0 = Uσ|n .
If σ 6= σi and σ 6= τi , then Uσi+1 = Uσi . Otherwise, since A0 ∪ . . . ∪ An + 1 is
nowhere dense in U × U . We can find basic clopen Uσi+1 i
⊆ Uσi i and Uτi+1
i
⊆ Uτii
such that ¡ ¢
E ∩ Uσi+1
i
× Uτi+1
i
= ∅.
1
Choose Uσ a basic closed subset of Uσk of diameter less that |σ| .
While the argument above will be the model for our proof of Silver’s theorem,
there are some significant obstacles. First and foremost, if E is a Π11 -equivalence
relation, there is no reason to believe that there is an open set U such that
E ∩ (U × U ) is meager. Harrington’s insight was to change the topology so that
that this is true.
74
Gandy Topology
Proof
Suppose A0 , A1 , . . . are τG -nowhere dense subsets of U . It is easy to construct
U = U0 ⊇ U1 ⊇ U2 ⊇ . . . a sequence of nonempty Σ11 -sets such that T Un ∩An = ∅.
To prove the Lemma we need only do this in such a way that Un 6= ∅. This
will require a bit more work.
At stage s of the construction we will have:
i) nonempty Σ11 -sets U = U0 ⊇ U1 ⊇ U2 . . . ⊇ Us such that Ui ∩ Ai = ∅ for
all i ≤ s;
ii) recursive trees T0 , T1 , . . . Ts such that Ui = {x : ∃y (x, y) ∈ [Ti ]};
iii) sequences σ0 ⊂ σ1 . . . ⊂ σs such that Ui ⊂ Nσi for all i;
iv) sequences ηji for i ≤ j ≤ s such that ηii ⊂ ηi+1 i
⊂ ηsi and there is (x, y)
such that σs ⊂ x, ηsi ⊂ y and (x, y) ∈ [Ti ] for all i ≤ s .
S S i
Suppose we have done this. T Let x = n∈N Sσn and yi = n≥i ηn . Then
(x, yi ) ∈ [Ti ] for all i. Hence x ∈ Un and x 6∈ An .
At stage 0 we let U0 = U and σ0 = η00 = ∅.
At stage s + 1 let
s
^ ^
W = {x ∈ Us : x ⊃ σs ∧ ∃y0 . . . ∃ys (x, yi ) ∈ [Ti ] ∧ yi ⊃ ηsi }.
i=0 i≤s
75
k,l
Exercise 8.5 Modify the proof of 8.4 to show each topology τG also satisfies
the Baire category theorem.
We will need the following technical lemma. Let A ⊆ N 2 and let A∗ =
{(x, y, z) ∈ N 3 : (x, z) ∈ A}.
1,1 2,1
Lemma 8.6 If A is τG -nowhere dense, then A∗ is τG -nowhere dense.
Harrington’s Proof
We will prove Silver’s Theorem for Π11 -equivalence relations. The proof will
easily relativize to Π11 -equivalence relations.
Suppose E is a Π11 -equivalence relation on N with uncountably many equiv-
alence classes. We say that A ⊆ N is E-small if xEy whenever x, y ∈ A. Let
yEx ⇔ ∀z (z ∈ B → zEy).
Thus the E-class of x is Π11 . By Σ11 -separation there is a ∆11 set A such that
B ⊆ A ⊆ {y : yEx}.
76
Proof
1,1
We first argue that E has the Baire property in the τG -topology.
1
Claim If A ⊆ N is Σ1 , then there are basic open sets (Bσ : σ ∈ N<ω ) such
that A = A(Bσ ).
Let T ⊆ N<ω × N<ω be a tree such that A = {x : ∃y (x, y) ∈ [T ]}. Let
Bσ = {x : (x||σ|, σ) ∈ T }. Then
[ \
A= By|n = A(Bσ ).
y∈N n∈N
1,1
The basic open sets of N × N are open in the topology τG , and hence
have the Baire Property, in this topology. Since the Souslin operator preserves
the Baire Property, every Σ11 subset of N × N has the Baire Property in the
1,1
τG -topology.
1,1
Suppose for purposes of contradiction that E ∩ (U × U ) is τG -nonmeager.
Since E has the Baire property there are nonempty Σ11 -sets A, B ⊆ U , such that
1,1
E is τG -comeager in A × B.
Let A1 = {(x0 , x1 ) ∈ A × A : x0 E 6 x1 }. Since A ⊆ U is a nonempty Σ11 -set,
A is not E-small. Thus A1 is a nonempty Σ11 -set.
Let Ci = {(x0 , x1 , y) : (x0 , x1 ) ∈ A1 , y ∈ B, xi E
6 y}, for i = 0, 1.
2,1
Claim Ci is τG -meager.
1,1 1,1
Since E is τG -comeager in A × B, there are D0 , D1 , . . . τG -nowhere dense,
such that [
Dn = {(x, y) ∈ A × B : x E 6 y}.
2,1
By Lemma 8.6, Dn0 = {(x0 , x1 , y) : (x0 , y) ∈ Dn } is τG -nowhere dense, and
S 0 2,1
Ci ⊆ Dn is τG -meager.
2,1
Since τG satisfies the Baire Category Theorem, There is
6 x1 , a contradiction.
But then x0 Ey, x1 Ey and x0 E
77
iv) each T σ ⊂ N<ω ×N<ω is a recursive tree such Uσ = {x : ∃y (x, y) ∈ [T σ ]}
and T σ ⊇ T τ for σ ⊆ τ ;
v) µσ ∈ N<ω , µσ ⊂ µτ for σ ⊂ τ , and Uσ ⊆ Nµσ ;
vi) ητσ ∈ N<ω , ητσ0 ⊂ ητσ1 if σ ⊆ τ0 ⊂ τ1 , and
78
Lemma 8.13 A is a closed unbounded subset of ω1 .
Proof Let Eα0 = Eφ(α) . Then the sequence Eα0 has the desired properties.
79
Lemma 8.15 Suppose A ⊆ N contains at least ℵ2 E-classes. Then there is
α < ω1 and a ∈ A such that {x ∈ A : xEα a} and {x ∈ A : x 6 Eα a} each
contain at least ℵ2 elements. In particular there is b ∈ A such that b 6 E a and
{x ∈ A : x E
6 α b} also contains at least ℵ2 E-classes
Proof Since each Eα has only countably many equivalence classes. For each
α < ω1 we can find aα ∈ A such that {x ∈ A : xEα aα } represents at least
ℵ2 E-inequivalent elements. If the lemma is false, then {x ∈ A : x 6 Eα aα }
represents at most ℵ1 -inequivalent elements. Thus
[
B= {x ∈ A : x E
6 α aα }
α<ω1
To finish the proof of Burgess’ Theorem we use 8.15 to build (Uσ : σ ∈ 2<ω )
such that:
i) Uσ is a nonempty Σ11 -set with Uσ ⊆ Uτ for σ ⊆ τ such that Uσ represents
at least ℵ2 E-classes;
ii) if x ∈ Uσ 0 and y ∈ Uσ 1 , then x E 6 y;
T T
If f, g ∈ C, x ∈ Uf |n and y ∈ Ug|n , and m is least such thatTf (m) 6= g(m)
then x E 6 αm y and x E6 y. This would suffice if we knew that the Uf |n 6= ∅ for
f ∈ C. We can insure this as in the proof of Silver’s Theorem.
We also build (T σ : σ ∈ 2<ω ), (µσ : σ ∈ 2<ω ), and (ητσ : σ ⊆ τ ∈ 2<ω ) such
that:
iii) each T σ ⊂ N<ω × N<ω is a tree such Uσ = {x : ∃y (x, y) ∈ [T σ ]} and
T ⊇ T τ for σ ⊆ τ ;
σ
80
If j ∈ N and ξ = (kσ : σ ⊆ τ ) where each kσ ∈ N let
^
Wij,ξ = {x ∈ Wi : x ⊇ µτbj ∧ ∃y ⊇ ητσbkσ (x, y) ∈ [T σ ]}.
σ⊆τ
Since [
Wi = Wij,ξ
j,ξ
Morley’s original proof uses the Perfect Set Theorem for Σ11 -sets, but does
not use Silver’s Theorem. His proof is given in [11] §4.4.
Using Scott’s analysis of countable models (see [11] §2.4) it is easy to see
that isomorphism is an intersection of ℵ1 Borel equivalence relations.
If M and N are countable L-structures a ∈ M n and b ∈ N n we define
(M, a) ∼α (N , b) as follows:
(M, a) ∼0 (N , b) if and only if M |= φ(a) if and only if N |= φ(b) for all
quantifier free formulas;
(M, a) ∼α+1 (N , b) if and only if for all c ∈ M there is d ∈ N such that
(M, a, c) ∼α (N , b, d) and for all d ∈ N there is c ∈ M such that (M, a, c) ∼α
(N , b, d);
if α is a limit ordinal, then (M, a) ∼α (N , b) if and only if (M, a) ∼β (N , b)
for all β < α.
Exercise 8.17 Prove that ∼α is a Borel equivalence relation on Mod(L).
Proposition 8.18 If M a countable L-structure, there is α < ω1 such that if
N is countable and M ∼α N , then M and N are isomorphic.
T
For a proof see [11] 2.4.15. It follows that α<ω1 ∼α is the isomorphism
equivalence relation. This proposition makes isomophism easier to analyze than
general Σ11 -equivalence relations. In particular for any M there is an α such
that the ∼α -class of M is the isomorphism class. This makes the counting
argument much easier.
1
Exercise 8.19 Give an example of a ΣT 1 -equivalance relation E on a Polish
space X and x ∈ E such that if E = α<ω1 Eα where each Eα is a Borel
equivalence relation, then for all α < ω1 there is y ∈ X such that xEα y and
xE6 y.
81
9 Tame Borel Equivalence Relations
In this section we will look at some general results about Borel equivalence
relations. Let X be a Polish space and let E be a Borel equivalence relation on
X. If x ∈ X we let [x] denote the equivalence class of X.
We start by looking at some of the simpler Borel equivalence relations.
Definition 9.1 T ⊆ X is a transversal for E if
|T ∩ [x]| = 1
for all x ∈ X.
We say that s : X → X is a selector for E if s(x)Ex for all x ∈ X and
s(x) = s(y) if xEy.
For example, let X be the set of all f : N → R such that f is Cauchy and
let E be the equivalence relation
1
f Eg ⇔ ∀n∃m∀k > m |f (k) − g(k)| < .
n
Then the set T of constant sequences is a Borel transversal.
Lemma 9.2 Let E be a Borel equivalence relation on a Polish space X. Then
E has a Borel transversal if and only if E has a Borel-measurable selector.
Proof
(⇒) If T is a Borel transversal, let
s(x) = y ⇔ y ∈ T and xEy.
Since the graph of s is Borel, s is Borel measurable by Lemma 2.3.
(⇐) If s is a Borel measurable selector, then
T = {x : s(x) = x}
is a Borel selector.
Exercise 9.3 Suppose E is a Borel equivalence relation on X and Ω is a σ-
algebra on X containing the Borel sets. Show that more generally E has a
transversal in Ω if and only if E has an Ω-measurable selector.
82
Proposition 9.5 If E is a Borel equivalence relation on X, then E is tame if
and only if there is a Borel measurable f : X → C such that xEy if and only if
f (x) = f (y).
Proof
(⇒) If (An : n ∈ N) is a Borel separating family, let (f (x))(n) = 1 if and
only if x ∈ An . Then xEy if and only if f (x) = f (y).
(⇐) Let An = {x : (f (x))(n) = 1}. Then An is Borel and (An : n ∈ N) is a
separating family.
Proposition 9.5 leads us to the following key idea for comparing the com-
plexity of Borel equivalence relations.
Definition 9.6 Suppose E is a Borel equivalence relation on X and E ∗ is a
Borel equivalence relation on Y . We say that E is Borel reducible to E ∗ if there
is Borel measurable f : X → Y such that
∆(1) <B ∆(2) <B . . . <B ∆(n) <B . . . <B ∆(N) <B ∆(C).
E ≤B ∆(X) ≤B ∆(C).
Thus E is tame.
In general tame equivalence relations need not have Borel transversals. Sup-
pose C ⊆ N × N is a closed set such that π(C) is not Borel. Let E be the
equivalence relation (x, y)E(u, v) if and only if x = u on C. Clearly π shows
83
E is tame. If T is a transversal for E is a Borel uniformization of C. If T is
Borel, then, since π|T is one-to-one, by Corollary 7.22, π(C) = π(T ) is Borel, a
contradiction.
In some important cases these notions are equivalent.
Ax = {y : f (y) = x}
84
Definition 9.12 We say that A ⊆ X is E-invariant if whenever x ∈ A and
yEx, then y ∈ A. If µ is a Borel probability measure, we say that µ is E-ergodic,
if µ(A) = 0 or µ(A) = 1 whenever A is µ-measurable and E-invariant.
Definition 9.13 We say that A ⊆ X is E-atomic if there is x ∈ X with
µ([x]) > 0.
If the equivalence relation is clear from the context we will refer to E as
“atomic” rather than E-atomic.
Lemma 9.15 (Zero-one law for tail events) Let µ be the usual Lebesgue
measure on C. If A ⊆ C is E0 -invariant, then µ(A) = 0 or µ(A) = 1.
Proof Since A is Lebesgue measurable, for any ² > 0, there is an open set U
such that U ⊇ A and µ(U \ A) < ².
If U ⊆ C is open, there is a tree T on 2<ω such that C \ U = [T ]. Let
S = {σ 6∈ T : ∀τ ⊆ σ τ ∈ T }.
S
Note that U = σ∈S Nσ and Nσ ∩ Nτ = ∅ for σ, τ distinct elements of S.
Thus X X 1
µ(U ) = µ(Nσ ) = .
σ∈S σ∈S
2|σ|
But Nσ and A are independent events. Thus µ(Nσ ∩ A) = µ(Nσ )µ(A) and
X X
µ(A) = µ(Nσ ∩ A) = µ(Nσ )µ(A) = µ(U )µ(A).
σ∈S σ∈S
85
Proof Let µ be Lebesgue measure on C. By the zero-one law for tail events µ
is E0 -ergodic. If x ∈ C, then [x] is countable and hence measure zero. Thus µ
is nonatomic. Thus E0 is not tame.
Indeed if Ω is the σ-algebra of Lebesgue measurable subsets of C, our proof
shows that E0 is not Ω-tame. In particular E0 is not C-tame.
The first major result on Borel equivalence relations is the next theorem of
Harrington, Kechris and Louveau.4 It says that E0 is the simplest nontame
Borel equivalence relation.
The proof of Theorem 9.17 heavily uses effective descriptive set theory. We
postpone the proof. For now we will be content giving the following corollary.
Proof We have already shown i)⇒ ii), i)⇒ iii) and are assuming i) ⇔ iv).
ii) ⇒ i) Suppose E in not tame. Then E0 ≤B E. Let f : C → X be a Borel
reduction of E0 to E. If T is a C-measurable transversal for E, then f −1 (T ) is
a C-measurable transversal for E0 . But then E0 is C-tame, a contradiction.
iii)⇒ i) If E is not tame, there is f : C → X a Borel reduction of E0 to E.
Let µ be Lebesgue measure on C. We define a measure ν on X by
as desired.
If A ⊆ X is E-invariant, then f −1 (A) is E0 -invariant. Thus
compact group. Effros extended this to the case where E is an Fσ orbit equivalence relation
for a Polish group. The general case is due to Harrington, Kechris and Louveau [3].
86
so ν.. For any x ∈ X, either f −1 ([x]E ) = ∅ and ν([x]E = 0, or there is y ∈ C
with f (y) = x. Then f −1 ([x])E = [y]E0 and ν([x]E ) = 0. Thus ν is an E-ergodic
nonatomic probability measure on X.
For arbitrary Borel actions, the orbit equivalence relation is Σ11 , but if G is
countable _
xEy ⇔ gx = y.
g∈G
87
For each i, j let Ei,j Sbe the equivalence relation generated by Ri,j . Then
Ei,j is Borel and E = i,j Ei,j . We claim that there is a Borel measurable
gi,j : X → X such that Ei,j -classes are the orbits of gi,j .
For each x there is at most one y such that xRi,j y and at most one z such
that zRi,j x. Thus every Ei,j -class is of one of the following forms:
1) {xi : i ∈ Z};
2) {xi : i ∈ N};
3) {x−i : i ∈ N}; or
4) {xi : i = 0, . . . , n} for some n ∈ N, where xk Ri,j xk+1 .
Let Bi = {x : [x] is of type i)}. Then Bi is a Borel set.
We define gi,j as follows.
1) On classes of type 1) gi,j (xk ) = xk+1 .
2) On classes of type 2)
(x if k = 0
1
gi,j (xk ) = xk−2 if k > 0 is even .
xk+2 if k is odd.
3) On classes of type 3)
(x if k = 0
−1
gi,j (xk ) = x−k+2 if k > 0 is even .
x−k−2 if k is odd.
4) On classes of type 4)
½
xk+1 if k < n
gi,j (xk ) = .
x0 if k = n
88
Our main goal is to show that if F2 is the free group on two generators, then
the orbit equivalence relation for the natural action of F2 on 2F2 is universal.
If G is a countable group and X is a standard Borel set, then (g, f ) 7→ gf is
a Borel action of G on X G . We let E(G, X) denote this action.
We first show that there is a universal G-action.
Clearly, φ is one-to-one.
If h ∈ H and ρ(h∗ ) = h, then
and
h∗ φ(f )(g) = φ(f )(h−1 −1
∗ g) = f (ρ(h∗ g)).
89
Proof By the Feldman-Moore Theorem there is a countable group G and a
Borel action of E on X such that E is the orbit equivalence relation for the
action. By Lemma 10.6 E ≤B E(G, C). There is a surjective homomorphism
ρ : Fℵ0 → G. Thus by Lemma 10.7
We will simplify this example a bit more after a couple of simple lemmas.
If g 6∈ H, then h−1 g 6∈ H so
Since φ is one-to-one, we may argue as above that f1 E(H, X)f2 if and only if
φ(f1 )E(G, X)φ(f2 ).
Suppose a, b are free generators of F2 . Then {an ban : n = 1, 2, . . .} freely
generate a subgroup of F2 isomorphic to Fℵ0 . Thus E(F2 , C) is also a universal
countable Borel equivalence relation.
90
Since φ is one-to-one, f1 = hf . Thus
Proof
E ≤B E(Fℵ0 , C)
≤B E(Fℵ0 × Z, 2) by 10.10
≤B E(Fℵ0 , 2) by 10.7
≤B E(F2 , 2) by 10.9 since we can embed Fℵ0 into F2
91
11 Hyperfinite Equivalence Relations
Definition 11.1 We say that a Borel equivalence relation E is finite if every
equivalence class is finite. We say that E is hyperfinite
S if there are finite Borel
equivalence relations E0 ⊆ E1 ⊆ . . . such that E = En .
The equivalence relation E0 is hyperfinite. Let Fn be the equivalence relation
on C
xFn y ⇔ ∀m > n x(m) = y(m).
S
Then E0 = Fn and each Fn is finite.
The main goal of this section will be to give the following characterizations
of hyperfinite equivalence relations.
We first show that, for countable Borel equivalence relations, finite ⇒ tame
⇒ hyperfinite.
T = {x ∈ X : ∀g ∈ G x ≤ gx}
92
Theorem 11.5 Let E be a countable Borel equivalence relation, then E is hy-
perfinite if and onlyS if there are tame Borel equivalence relations E 0 ⊆ E1 ⊆
E2 ⊆ . . . with E = n En .
Z-actions
Suppose E is a Borel equivalence relation on X and <[x] is a linear order of [x].
We say that [x] 7→<[x] is Borel if there is a Borel R ⊆ X × X × X such that
i) R(x, y, z) ⇒ (xEy ∧ xEz);
ii) R(x, y, z) ⇒ y <[x] z;
iii) if xEx1 then R(x, y, z) ⇔ R(x1 , y, z).
93
iii) There is a Borel [x] 7→<[x] such that each infinite E-class has order type
Z.
iv) There is a Borel action of Z on X such that E is the orbit equivalence
relation.
v) There is a Borel automorphism T : X → X such that E-equivalence
classes are T -orbits.
Proof
It is clear that iv) ⇔ v)
i) ⇒ ii) S Let E0 ⊆ E1 ⊆ E2 ⊆ . . . be finite Borel equivalence relations such
that E = En . We may assume that E0 is equality. We may also assume that
there is an ordering < of X.
We inductively define <[x]En as follows.
1) <[x]E0 is trivial, since [x]E0 = {x}.
2) Suppose y, zEn x and yEn−1 z, then y <[x]En z if and only if y <[y]En−1 z.
3) Suppose y, zEn x and y 6 En−1 z. Let yb be the <[y]En−1 -least element of
[y]n−1 and zb be the <[z]En−1 -least element of [z]En−1 . If yb < zb, then y <[x]En z.
Otherwise z <[x]En y.
In other words: we order [x]En be breaking it into finitely many En−1 classes
C1 , . . . , Cm . We then order the classes Ci by letting yi be the <[yi ]En−1 -least
element and saying S that Ci < Cj if yi < yj .
Let <[x]E = <[x]En . If xEn y and x <[x] z <[x] y, then xEn x. It follows
that <[x] is a discrete union of finite orders. Thus <[x] is either a finite order or
has order type ω, ω ∗ or Z.
We need only argue that the assignments [x] 7→<[x]En is Borel. The only
difficulty is picking x b the <[x]En -least element of [x]En . There is a countable
groups G and a Borel actions of G on X such that En is the orbit equivalence
relation of Gn . Then
and
{x : ∃g ∈ G∀h ∈ G gx ≤[x] hx}
are Borel we can determine the order type of each class. If a class has order
type ω of ω ∗ we can reorder it so that it has order type Z. For example if [x] is
x0 <[x] x1 <[x] < . . . we define a new order <∗ so that
94
iii) ⇒ iv) We define a Borel automorphism g : X → X such that E is the
orbits of g. If x is the <[x] -maximal element of [x], then [x] is the <[x] -least
element of [x]. Otherwise let g(x) be the <[x] -successor of x. Arguing as above
g is Borel. We let Z act on X by nx = g (n) x. Clearly E is the orbit equivalence
relation.
iv) ⇒ iii) Let g : X → X be a Borel automorphism such that E-classes are
g-orbits. Then X0 = {x : ∃n 6= 0 : g (n) x = x} is Borel. On X0 we can define
<[x] using < is a fixed linear order of X. Thus, without loss of generality, we
may assume that every E-class is infinite. But then we can define x <[x] y if
and only if there is an n > 0 such that g (n) x = y. Clearly this is a Z-ordering
of [x].
iii) ⇒ i) We may assume that E is an equivalence relation on C. For each
equivalence class C we define a tree TC ⊆ 2<ω by
TC = {σ ∈ 2<ω : ∃x ∈ C x ⊃ σ}.
95
Clearly each En class is finite and if xEy, then xEn y for all sufficiently large
n.
The i)⇔ iii) is due to Slaman and Steel. The direction iv) ⇒ i) is due to
Weiss.
It follows immediately that there is a universal hyperfinite Borel equivalence.
Reducibility to E0
Theorem 11.11 (Doughrety-Jackson-Kechris) If E is a hyperfinite Borel
equivalence relation, then E ≤B E0 .
96
If x ∈ X and y 6∈ X, then f (x) 6 E0 f (y), since the even part of f (x) is 0 and
the even part of f (y) is 1.
If x, y ∈ X, then
b
f(x)E b
0 f(y) ⇔ p(g(x))E0 p(g(y)) ⇔ g(x) = g(y) ⇔ xEy.
If x, y 6∈ X, then
b
f(x)E b
0 f(y) ⇔ f (x)E0 f (y) ⇔ xEy.
Proof Since each Xi is invariant we mayS assume that the Xi are disjoint. If fi :
Xi → C is a Borel reduction and f : Xi → C is the function f (x) = 0i 1bfi (x)
for x ∈ Xi , then f is a Borel reducition of E|X to ∆(C).
The two lemmas allow us to work “modulo tame sets”, i.e. if X is tame we
may ignore it and assume we are just working with E|(C Z \ X).
Proof of Theorem 11.11 We will view each x ∈ C Z as a Z × N array of
zeros and ones. The columns are . . . , x−2 , x−1 , x0 , x1 , x2 , . . . where xi ∈ C. If
σ ∈ (2n )n , we view σ as (σ0 , . . . , σn−1 ) where each σi ∈ 2n . We say that σ
occurs in x at k if σi = xk+i |n for i = 0, . . . , n.
Fix σ ∈ (2n )n . Let Y be the set of all x ∈ C Z such that there is a largest k
such that σ occurs in x at k. We will argue that Y is tame. Suppose x ∈ X and
k is maximal such that σ occurs in x at k. If n ∈ Z, then (nx)i (j) = xi−n (j).
Thus k − n is the largest i such that σ occurs in nx at i. Thus Y is Z-invariant.
Let s : X → X be the function s(x) = kx where k is maximal such that σ
occurs in x at k. Then s(x) is the unique element of [x] where 0 is the largest i
such that σ occurs at i. Thus s is E-invariant and Y is tame.
Similarly, the set of x such that there is a least k such that σ occurs in x at
k is tame. By throwing out these tame Borel sets, we may restrict attention to
a Borel set X that for all σ and x ∈ X, the set of k such that σ occurs in x at
k is unbounded in both directions.
If σ ∈ (2n )n and m < n we let σ|m = (σ0 |m, . . . , σm−1 |m). Fix <n a linear
order of (2n )n such that if σ, τ ∈ (2n )n and σ|m <m τ |m for some m < n, then
σ <n τ .
For x ∈ X let fn (x) be the <n -least element of (2n )n occuring in x. Our
assumptions on <n , insure that fn (x)|m = fm (x) for m < n. Define f : X → C N
by
f (x) = (y0 , y1 , . . .)
where fn (x) = (y0 |n, y1 |n, . . . , yn−1 |n) for all n. Note that each fn and f are
E-invariant.
We say that g ∈ C N occurs in x at k if xk+i = g(i) for all i ∈ N. Let Y be
the set of x ∈ X such that f (x) occurs in x and there is a least k such that
97
f (x) occurs in x at k. Then Y is Borel, E-invariant and the function s(x) = kx
where k is least f (x) occurs at k is a Borel selector. Thus Y is tame. Let W be
the set of all x ∈ X such that f (x) occurs in x at k for arbitrarily small k. If
x ∈ W , then the action of Z on [x] is periodic. Thus [x] is finite and W is tame.
Throwing out Y and W we may assume that f (x) does not occur in x for all
x ∈ X.
For x ∈ X and n ∈ N define
k0x = 0
x x
k2n+1 = the least k such that k > k2n and f2n+1 (x) occurs in x at k.
x x
k2n+2 = the largest k such that k < k2n and f2n+2 (x) occurs in x at k.
Then
. . . ≤ k4x ≤ k2x ≤ k0x < k1x ≤ k3x ≤ . . . .
x x
Since f (x) does not occur in x, k2n → ∞ and k2n+2 → −∞.
We make the usual identification between C and P(N) by identifying sets
with their characteristic functions. Under this identification
Fix a bijection
p : N × (2<ω )<ω → N.
For x ∈ X and n ∈ N let
txn = |kn+1
x
− knx | + 1
x
and let rnx ∈ (2<ω )tn be (σ0 , . . . , σtxn −1 ) where
This looks more confusing then it is. Suppose n is even. Then knx < kn+1 x
x
and rn is just the block of the matrix x where we look take rows 0, . . . , n − 1
and columns knx to kn+1
x
.
Let G(x) = {p(n, rnx ) :∈ N}. From G(x), and knowing k0x = 0, we can
reconstruct the sequence (kix : i ∈ N) and x. Thus G is one-to-one.
Suppose G(x)4G(y) is finite. Then there is an m such that rnx = rny for all
n > m. It follows that y is obtained by shifting x. Thus xEy.
Suppose xEy. There is m ∈ Z such that xm+i = yi for all i ∈ N. Without
x
loss of generality assume m > 0. Let n0 be least such that k2n 0 +1
> m. Since
f (x) = f (y),
x y
k2n 0 +1
= m + k2n 0 +1
.
Thus knx = m + kny for all n > n0 . It follows that G(x)4G(y) is finite.
Thus G reduces E to E0 .
98
Growth Properties
Our next goal is to show that there are countable groups G 6= Z such that every
G-action is hyperfinite.
Definition 11.15 Suppose G is a finitely generated group. We say that G
has polynomial
S growth if there is a finite X ⊆ G closed under inverse such that
G = X n and there are C, d ∈ Z such that
|X n | ∈ O(nd )
We will prove that all Borel actions of finitely generated groups of polynomial
growth induce hyperfinite orbit equivalence relations. In fact we will work in
a more general context which will also allow us to understand actions of some
nonfinitely generated groups like Qd .
Definition 11.17 Let G be a countable group. We say that G has the mild
growth property of order c, if there is a sequence of finite sets K0 ⊆ K1 ⊆ K2 . . .
such that:
S
i) Ki = G;
ii) 1 ∈ K0 ;
iii)Ki = Ki−1 for all i;
iv) Ki2 ⊆ Ki+1 for all i;
v) |Ki+4 | < c|Ki | infinitely many i.
99
Proof Let X be a set of generators closed under inverse such that |X n | ≤ Cnd
n n n n+1
for all n > 0. Let Kn = X 2 . Clearly i)—iii) hold. Since X 2 X 2 = X 2 iv)
holds. We need only argue v). Suppose not. Then there is n0 such that
and
(16d + 1)k |Kn0 | ≤ C2n0 d (16d )k
for all k. But this is clearly impossible.
Proof Let Ki,0 ⊆ Ki,1 ⊆ . . . witness that Gi has the mild growth property of
order c. Let σ : N → N × N be a bijection such that if σ(i) = (j, k), then j ≤ i.
We will build K0 ⊆ K1 ⊆ . . . ⊂ G. For notational convenience let K−1 =
{1}.
Suppose we have Ki ⊆ Gi for i < 5k. We will show how to define K5k+i for
i = 0, . . . , 4. Let
2
K = K5k−1 ∪ Kσ(k) ⊆ G5k .
We can find an n such that
Let K5k+i = K5k,n+i for i = 0, . . . , 4. It is easy to see that i)–iv) hold and
S∞ 1 d
Since Qd = n=1 n! Z , Qd has the mild growth property.
In particular the orbit equivalence relation for any Borel action of a finitely
generated Abelian group is hyperfinite and the orbit equivalence relation for any
Borel action of Qd is hyperfinite. It is still an open question if any action of a
countable Abelian group is hyperfinite.
100
The theorem will follow from several lemmas.
Definition 11.21 Let F be a symmetric, reflexive Borel binary relation on a
Borel space X. We say that F is locally finite if {y : yF x} is finite for all x. We
say that Y ⊆ X is F -discrete if ¬(xF y) for all distinct x, y ∈ Y and we say that
Y is maximal F -discrete if it is discrete and for all x ∈ X there is y ∈ Y with
xF y.
Lemma 11.22 Let F be a locally finite, symmetric, reflexive Borel binary re-
lation on X. Then there is a maximal F -discrete Y ⊆ X.
Lemma 11.24 If G is a group with the mild growth property and E is the
orbit equivalence relation for a Borel action of G, then there are locally finite,
symmetric, reflexive BorelS binary relations F0 ⊆ F1 ⊆ . . . satisfying the Weiss
condition such that E = Fi .
Thus
xi x−1 −1
j = g i gj ∈ Kn+1
and xi Fn+1 xj , a contradiction.
101
If h ∈ Kn , then hgi ∈ Kn+4 . Thus
N |Kn | ≤ |Kn+4 | ≤ c|Kn |
and N ≤ c. Hence there are infinitely many n such that any Fn+1 -discrete
subset of {y : yFn+3 x} has cardinality at most c and (Fn : n ∈ N) has the Weiss
condition.
Lemma 11.25 Suppose E ⊆ E ∗ are countable Borel equivalence relations. If
E is hyperfinite and every E ∗ -class contains finitely many E-classes, then E ∗
is hyperfinite.
Proof Since E ∗ is the orbit equivlance relation for a Borel action of some
countable group G. Then |[x]E ∗ /E| = k if and only if there are g1 , . . . , gk ∈ G
such that gi x 6E gj x for i 6= j and for all g ∈ G gxEgi x for some i = 1, . . . , k.
Thus {x : |[x]E ∗ /E| = k} is Borel and, without loss of generality we may assume
that there is a fixed k such that each E ∗ class contains exactly k E-classes.
Let G = {g0 , g1 , . . .}. We inductively define functions f1 (x), . . . , fk (x) by
f1 (x) = x. Let Ni+1 (x) be the least n such that gn x 6E fj (x) for all j ≤ i and
fi+1 (x) = gNi+1 (x) x. Then
k
[
[x]E ∗ = [fi (k)]E .
i=1
S
Suppose E = En where E0 ⊆ E1 ⊆ . . . are finite Borel equivalence rela-
tions. Let xEn∗ y if and only if there is σ a permutaion on {1, . . . , k} such that
fi (x)En fσ(i) (y) for all i.
Clearly En∗ is an equivalence relation and En∗ ⊆ En+1 ∗
. If xEn∗ y, then
∗ ∗
xEn fi (y) for some i. Thus each En class is finite. If xE y, then there is a
permutation σ such that fi (x)Efσ(i) (y) for all i. SThere is an m such that
fi (x)En fσ(i) (y) for all i and all m > n. Thus E ∗ = En∗ is hyperfinite.
Proof of Theorem 11.20 By Lemma 11.24 we can find F0 ⊆ F1 ⊆ . . . a
sequence of locally finite, symmetric,
S reflexive Borel binary relations with the
Weiss condition such that Fn = E. Let Yn be a Borel maximal Fn -disjoint
set. Let sn : X → Yn be a Borel measurable function such that sn (x)Fn x. Let
πn : X → X be sn ◦ sn−1 ◦ . . . ◦ s0 and let xEn y if and only if πn (x) = πn (y).
Clearly En is an equivalence relation and En ⊆ En+1 . Since each sn is
finite-to-one, πn is finite-to-one and En is a finite equivalence
S relation. An easy
induction shows that if xEn y, then xEy. Thus E ∗ = En is a hyperfinite
equivalence relation and E ⊆ E ∗ . By Lemma 11.25 it suffices to show that
every E-class contains at most finitely many E ∗ -classes.
Suppose (Fn : n ∈ N) satisifies the Weiss condition with constant c. Sup-
pose x1 , . . . , xN are E-equivalent but E ∗ -inequivalent. We can find arbitrarily
large n such that x1 , . . . , xN Fn x1 and any Fn -discrete subset of {y : yFn+2 x1 }
has cardinality at most c. Then πn (x1 ), . . . , πn (xN ) are distinct elements of
Yn and hence are Fn -discrete. Since Fi2 ⊆ Fi+1 , we see, by induction, that
πn (xi )Fn+1 xi . Since xi Fn x1 , πn (xi )Fn+2 x1 . Thus N < c. Thus every E-class
contains at most c, E ∗ -classes and E is hyperfinite.
102
Ammenability
Throughout this section Γ will be a countable group.
Theorem 11.26 Suppose Γ acts freely on a standard Borel space X and µ is
a Γ-invariant probability measure on X. If the orbit equivalence relation E is
hyperfinite, then Γ is ammenable.
The proof we give was pointed out to me by Greg Hjorth. It uses one of the
many useful characterizations of ammenability.
Suppose K is a compact metric space. Let P (K) be the space of all Borel
probability measures on K. We topologize P (K) with the weakest topology
such the maps Z
µ 7→ f dµ
are continuous for all bounded continuous f : K → R. The space P (K) is also
a compact metric space and if K is a Polish space so is P (K) (see [6] 17.E).
A continuous action of Γ on K induces an action of Γ on P (K) by
gµ(A) = µ(g −1 A).
We say that µ is Γ-invariant if gµ = µ for all g ∈ Γ.
Theorem 11.27 A countable group Γ is ammenable if and only if for every
compact metric space K and every continuous action of Γ on K, there is a
Γ-invariant measure in µ(K).
103
Theorem 11.29 (Zimmer) There is an α-invariant, µ-measurable x 7→ νx
from X to P (K).
Assuming Zimmer’s result we will complete the proof. We claim that there
is a Γ-invariant Borel probability measure on K.
For A ⊆ K Borel let Z
ν(A) = νx (A) dµ.
X
Since µ and each νx are probability measures, ν is a probability measure, we
need only show that it is Γ-invariant.
Z
ν(gA) = νx (gA) dµ
ZX
= g −1 νx (A) dµ
X
Z
= νβ(g−1 ,x)x (A) dµ
X
Z
= νg−1 x (A) dµ.
X
But Z Z Z
F (x) dµ = gF (x) dµ = F (g −1 x) dµ
X X X
for any µ-measurable F : X → R and g ∈ Γ. Thus
Z Z Z
νx (gA) dµ = νg−1 x (A) dµ = νx (A) dµ
X X X
104
References
[1] H. Becker and A. Kechris, The Descriptive Set Theory of Polish Group
Actions, Cambridge Univ, Press, Cambridge UK, 1996.
[2] R. Dougherty, S. Jackson and A. Kechris, The structure of hyperfinite Borel
equivalence relations, Trans. Amer. Math. Soc., 341, 193–225, 1994.
[3] L. Harrington, A. Kechris and A. Louveau, A Glimm-Effros dichotomy for
Borel equivalence relations, Jour. Amer. Math. Soc., (3) 1990, 903–928.
[4] L. Harrington, D. Marker and S. Shelah, Borel orderings, Trans. Amer.
Math. Soc. (310) 1998, 293–302.
[5] G. Hjorth and A. Kechris, Rigidity theorems for actions of product groups
and countable Borel equivalence relations, preprint.
[6] A. Kechris, Classical Descriptive Set Theory, Springer-Verlag, New York,
1995.
[7] A. Kechris, Lectures on definable group actions and equivalence relations,
unpublished notes.
[8] S. Jackson, A. Kechris and A. Louveau, Countable Borel equivalence rela-
tions,
[9] T. Jech, Set Theory, Academic Press, New York, 1978.
[10] R. Mansfield and G. Weitkamp, Recursive Aspects of Descriptive Set The-
ory, Oxford Science Pub., Oxford UK, 1985.
[11] D. Marker, Model Theory: An Introduction, Springer, New York, 2002.
[12] D. M. Martin and A. S. Kechris, Infinite games and effective descriptive set
theory, Analytic Sets, C. A. Rogers et. al. ed., Academic Press, London,
1980.
[13] D. Martin and J. Steel, A proof of projective determinacy, J. Amer. Math.
Soc, 2 (1989) 71-125.
[14] Y. Moschovakis, Descriptive Set Theory, North Holland, Amsterdam, 1980.
[15] S. M. Srivastava, A Course on Borel Sets. Springer, New York, 1998.
[16] J. Steel, On Vaught’s Conjecture, in Cabal Seminar 76–77, A. Kechris and
Y. Moschovakis, eds., Springer-Verlag, New York, 1978.
105