9

5
8


r
e
v
i
s
i
o
n
:
2
0
1
1
-
0
5
-
2
3







m
o
d
i
f
i
e
d
:
2
0
1
1
-
0
5
-
2
3


A Hanf number for saturation and omission
John T. Baldwin

University of Illinois at Chicago

Saharon Shelah
Hebrew University of Jerusalem
Rutgers University
May 23, 2011
Abstract
Suppose t = (T, T1, p) is a triple of two countable theories T T1 in vocab-
ularies 1 and a 1-type p over the empty set. We show the Hanf number for
the property: There is a model M1 of T1 which omits p, but M1 is saturated
is essentially equal to the L owenheim number of second order logic. In Section 4
we make exact computations of these Hanf numbers and note some distinctions
between rst order and second order quantication. In particular, we show that
if is uncountable, h
3
(L,(Q), ) = h
3
(L
1
,, ), where h
3
is the normal
notion of Hanf function (Denition 4.12.)
Newelski asked in [New] whether it is possible to calculate the Hanf number of the
following property P
N
. In a sense made precise in Theorem 0.2, we show the answer
is no. In accordance with the original question, we focus on countable vocabularies for
the rst three sections. We deal with extensions to larger vocabularies in Section 4.
Denition 0.1 We say M
1
[= t where t = (T, T
1
, p) is a triple of two theories in
vocabularies
1
, respectively, T T
1
and p is a
1
-type over the empty set if M
1
is a model of T
1
which omits p, but M
1
is saturated. Let K
t
denote the class of
models M
1
which satisfy t.
For K = K
t
, with t in a countable vocabulary, let P
c
N
(K
t
, ) hold if [
1
[
0
and for some M
1
with [M
1
[ = , M
1
[= t. P
f
N
(K
t
, ) is the same property restricted
to triples where T
1
and T are nitely axiomatizable in nite vocabularies and p is
denable in second order logic.
Recall Hanfs observation [Han60] that for any such property P(K, ), where K
ranges over a set of classes of models, there is a cardinal = H(P) such that is the
least cardinal satisfying: if P(K, ) holds for some then P(K, ) holds for

We give special thanks to the Mittag-Lefer Institute where this research was conducted. This is pa-
per 958 in Shelahs bibliography. Baldwin was partially supported by NSF-0500841. Shelah thanks the
Binational Science Foundation for partial support of this research.
1
9
5
8


r
e
v
i
s
i
o
n
:
2
0
1
1
-
0
5
-
2
3







m
o
d
i
f
i
e
d
:
2
0
1
1
-
0
5
-
2
3


arbitrarily large . H(P) is called the Hanf number of P. E.g. P(K, ) might be the
property that K has a model of power . Similarly the L owenheim number (P) of a
set P of classes is the least cardinal such that any class K P that has a model has
one of cardinality .
Theorem 0.2 Assume the collection of with
<
= is a proper class. H(P
f
N
) =
(L
II
) where L
II
denotes the collection of sentences of second order logic.
Since H(P
c
N
) H(P
f
N
), this shows that the Hanf number in the abstract is at least
(L
II
), as asserted. In Section 1 we introduce a variant
2
(L
II
) on the L owenheim
number of second order logic which is essentially equal to (L
II
) (i.e. equal modulo
a mild set theoretic hypothesis: Assumption 0.3). It is fairly easy to show (Claim 2.9)

2
(L
II
) H(P
f
N
) giving the essentially equal of the abstract. We will replace this
essential equality with an exact computation and deal with uncountable languages in
Section 4.
Jouko Vaananen provided the following summary of the effect of this result by
indicating the size of (L
II
). (L
II
) is bigger than the rst (second, third, etc) xed
point of any normal function on cardinals that itself can be described in second order
logic. For example it is bigger than the rst such that =

, bigger than the

rst such that there are cardinals below such that =

, etc. It is easy
to see that if there are measurable (inaccessible, Mahlo, weakly compact, Ramsey,
huge) cardinals, then the Lowenheim number of second order logic exceeds the rst of
them (respectively, the rst inaccessible, Mahlo, weakly compact, Ramsey, huge) (and
second, third, etc). So even under V = L, the L owenheim number is bigger than any
large cardinal that is second order denable and consistent with V = L. Such results
are discussed in Vaananens paper Hanf numbers of unbounded logics[Vaa79]. A
result of Magidor [Mag71] shows the Lowenheim number of second order logic is
always below the rst supercompact. Vaananens paper Abstract logic and set theory
II: Large cardinals gives lower bounds for the Lowenheim number of equicardinality
quantiers and thus a fortiori for second order logic [Vaa82]. In simple terms, if E()
is the statement that 2

++
then the rst cardinals (if any) such that E() holds
is less than the Lowenheim number of second order logic. This shows that by forcing
we can push the Lowenheim number up at will.
We make the following assumption throughout.
Assumption 0.3 Assume the collection of with
<
= is a proper class.
This assumption follows from GCH, but if GCH fails badly the only such cardinals
are strongly inaccessible. The key point for our use of the condition is that
<
=
> [
T
[ +
0
is a sufcient condition for the existence of a saturated model in of
a complete theory T; if T is unstable
<
= is also necessary. We will explore this
issue for stable theories, in the absence of Assumption 0.3, elsewhere. In Section 1 we
reviewsome properties of second order logic and showthe equality of two L owenheim
numbers in our context. In Section 2, we state two technical results, prove one, and
deduce Theorem 0.2 from them. In Section 3, we prove the more difcult technical
result. In Section 4, we code syntax more carefully and obtain a uniform equivalence
for vocabularies of all cardinalities.
2
9
5
8


r
e
v
i
s
i
o
n
:
2
0
1
1
-
0
5
-
2
3







m
o
d
i
f
i
e
d
:
2
0
1
1
-
0
5
-
2
3


Newelskis question arose in the study of the model theory of groups and the ex-
istence of groups with bounded orbits. The authors acknowledge very fruitful discus-
sions with Jouko V a an anen and Tapani Hyttinen concerning the material of this paper.
We also thank the referee for an unusually detailed and helpful series of reports.
1 Some Second Order Logic
By (pure) second order logic, L
II
, we mean the logic with individual variables and
variables for relations of all arities but no non-logical constants. The atomic formulas
are equalities between variables and expressions X(x) where X is an n-ary relation
variable and x is an n-tuple of individual variables. Note that a structure A for this
logic is simply a set so is determined entirely by its cardinality. But we use the full
semantics: the n-ary relation variables range over all n-ary relations on A.
We explain the connection of our restriction to with =
<
to the computation
of some variants on the Hanf and L owenheim numbers.. In general for any class K
of models write spec(K) for the collection of such that there is a model in K with
cardinality .
Denition 1.1 Let be a sentence of second order logic.
1. spec
1
() = : [= .
2. spec
2
() = : =
<
[= .
Note that there is a sentence in second order logic which has a model of size
if and only if
<
= . Namely, let assert there is an extensional relation R on sets
such that each element denotes, via R, a set of smaller cardinality than the universe and
each such set is coded by R. We will generally write
<
= to denote this sentence.
Denition 1.2 Dene H
2
and
2
to be the Hanf and Lowenheim numbers with respect
to spec
2
.
Well write
1
for and H
1
for H where it is convenient for comparison. Note the
following easy transformations in second order logic.
Fact 1.3 Fix L
II
.
1. spec
1
():
(a) There is a
1
L
II
with min(spec()) < min(spec(
1
)).
(b) If spec() is bounded and nonempty there is a
2
L
II
with spec(
2
)
bounded and nonempty and sup(spec()) < sup(spec(
2
)).
2. spec
2
():
(a) There is a
3
L
II
with min(spec()) < min(spec(
3
)) and if [=
3
,

<
= .
3
9
5
8


r
e
v
i
s
i
o
n
:
2
0
1
1
-
0
5
-
2
3







m
o
d
i
f
i
e
d
:
2
0
1
1
-
0
5
-
2
3


(b) If spec
2
() is bounded and nonempty there is a
4
L
II
with spec
2
(
4
)
bounded and nonempty and sup(spec
2
(
4
)) < sup(spec
2
(
4
)).
These transformations imply:
Fact 1.4 1. H
1
(L
II
), H
2
(L
II
),
1
(L
II
),
2
(L
II
) are strong limit cardinals.
2. There is no sentence attaining any of these values exactly. (E.g., there is no
L
II
with sup(spec()) = H
1
(L
II
).)
3. For either spectrum,
i
(L
II
) = supminspec
i
() : L
II
has a model
and similarly H
i
(L
II
) = supsupspec
i
(): L
II
is bounded.
Note that any logic satisfying Fact 1.3 will also satisfy Fact 1.4. We use this obser-
vation without comment in studying innitary second order logics in Section 4.
Using Assumption 0.3 we can show:
Lemma 1.5 H(L
II
) = H
2
(L
II
), (L
II
) =
2
(L
II
)
Proof. One direction is easy. For every sentence of second order logic, there is a
sentence

such that:
spec
2
() = spec
1
(

).

just expresses the conjunction of with

<
= . Recall Fact 1.4.3 Since every
2-spectrum is a 1-spectrum
2
(L
II
)
1
(L
II
) and H
2
(L
II
) H
1
(L
II
).
But the opposite inequality also holds. Let be a sentence with a non-empty 2-
spectrum. Let f() denote the least > with
<
= . It is easy to construct for
each second order sentence a sentence

such that
spec(

) = spec
2
(

) = f() : spec().
Clearly the map

shows
2
(L
II
)
1
(L
II
) and H
2
(L
II
) H
1
(L
II
).

1.5
2 The main result
Recall our notation from Denition 0.1.
Notation 2.1 We will write t (possibly with subscripts) for a triple (T, T
1
, p). The
expression t has a model in means there is a model of T
1
with cardinality that
omits p and whose reduct to L(T) = is saturated.
Convention 2.2 When
1
is nite we consider it to be a subset of . Thus the set of
rst order
1
-sentences is recursive and we can code them as natural numbers.
We concentrate rst on P
f
N
(K
t
, ) from Denition 0.1. We need some additional
coding to handle non-nitely axiomatizable theories and consider this generalization
in Section 4. We begin by clarifying a notion from Denition 0.1.
4
9
5
8


r
e
v
i
s
i
o
n
:
2
0
1
1
-
0
5
-
2
3







m
o
d
i
f
i
e
d
:
2
0
1
1
-
0
5
-
2
3


Denition 2.3 A type p in a vocabulary satisfying Convention 2.2 is denable in sec-
ond order logic if we can code the type as a subset A
p
of so that in the vocabulary
with constant symbol 0 and relation symbol S there is a second order sentence and
second order formula (x) satisfying the following condition. For the rst cardinal
which satises if M is (, 0, S), with 0 interpreted as 0 and S as successor on the
natural numbers, then A
p
= n: M [= (n).
Now for convenience we restrict our triples to those satisfying the convention. For-
mally:
Notation 2.4 T
f
denotes the set of triples t such that T is nitely axiomatizable and
p is second order denable.
Theorem 2.5 For every second order sentence , there is a triple t

T
f
such that if

<
= , then the following are equivalent:
1. t

has a model in .
2. has a model in every cardinal strictly less than .
We prove Theorem 2.5 in Section 3.
Lemma 2.6 For every t T
f
there is a second order
t
, such that
t
has a model in
if and only if t has a model in .
Proof. Recalling the restrictions involved in T
f
, it is easy to write a second order
sentence such that M [= if and only if M [= T
1
, M omits p and M is saturated.

2.6
We could strengthen Lemma 2.6 by restricting the second order quantication to
sets of size strictly less than the size of the model, but that is not important here. We
now deduce Theorem 0.2 from Theorem 2.5 and Lemma 2.6. We use the following
notation.
Notation 2.7 Spec(t) is the collection of cardinals such that there is an M
1
satisfy-
ing t with [M
1
[ = .
We have not established that the Hanf and Lowenheim numbers for the P
N
satisfy
Fact 1.4. This complicates the argument for the following two results.
Claim 2.8 H(P
f
N
)
2
(L
II
) where L
II
denotes second order logic.
Proof. Lemma 2.6 shows that for any t T
f
, there is a
t
L
II
with spec(t) =
spec(
t
). Suppose for contradiction that H(P
f
N
) >
2
(L
II
). Then there is a triple
t T
f
such that sup(spec(t))
2
(L
II
) .
Let C = : =
<
. Choose
t
L
II
so that [=
t
iff for every innite
cardinal C there is [, ) spec(t).
Let
t
be the minimal element of Cspec(t)
c
. Then
t

2
(L
II
). For any t, the
denition of P
f
N
guarantees that if
<
= , > and some M [= t has cardinality
5
9
5
8


r
e
v
i
s
i
o
n
:
2
0
1
1
-
0
5
-
2
3







m
o
d
i
f
i
e
d
:
2
0
1
1
-
0
5
-
2
3


then some N [= t has cardinality (take a saturated elementary submodel). Thus,
[=
t
if and only if every C belongs to spec(t).
Now spec(
t
) is exactly :
t
, whence spec(
t
) is : >
t
. So the
L owenheim number of
t
is (
t
)
+
>
2
(L
II
) and this contradiction completes the
proof.

2.8
Lemma 2.9 H(P
f
N
)
2
(L
II
) where L
II
denotes second order logic.
Proof. Suppose for contradiction that there is a second order sentence such that

0
= min(spec
2
()) H(P
f
N
). By the denition of spec
2
,
<0
0
=
0
. Let

express (U)(
U
[U[
<|U|
= [U[). We apply Theorem 2.5 to (

). Note that

is
true on all cardinals
0
and false on all <
0
. By Theorem 2.5,
0
[= t
(

)
and

0
H(P
f
N
). So t
(

)
and therefore (

) has arbitrarily large models. But (

) has
no models larger than
0
. This contradiction yields the theorem.
2.9
In the next section we prove the crucial Theorem 2.5. In the last section we remove
the restrictions to nitely axiomatizable theories and countable languages.
3 Essential Lemmas
Now we prove Theorem 2.5. For convenience, we list here the two vocabularies. We
describe the axioms of T and T
1
below.
Notation 3.1 1. contains unary predicates Q
1
, Q
2
, a binary relation R and par-
tial binary functions F and F
2
. It contains two constant symbols c
0
, c

and a
unary function symbol g.
2.
1
adds a unary predicate Q
0
and a binary relation <
1
.
Remark 3.2 (Proof Sketch) For each second order , we construct a triple t

. But
most of the construction is independent of the particular and so we rst construct a
theory T
1
which does not depend on . The vocabulary will contain unary predicates
Q
1
, Q
2
. The axioms will assert that Q
1
, Q
2
partition the universe. Q
0
is in
1
. Omis-
sion of the type p will guarantee that Q
0
Q
1
is countable. Omission of the type in a
model M of T
1
whose -reduct is
1
-saturated and some coding involving the partial
order <
0
in will guarantee that Q
1
(M) is well-ordered by a relation symbol <
1
in
1
.
A relation symbol R in will code subsets of Q
1
by elements of Q
2
. Thus rst order
quantication on Q
2
will encode second order quantication on Q
1
. In particular, we
can code a given second order sentence and thus extend T
1
to T

. But the encod-

ing guarantees correctness only on subsets whose every subset is coded in Q
2
. The
construction will guarantee that if < and M is -saturated, then is a <
1
-initial
segment Q
1
. Since < each subset of is coded by a type of size so the encoded
semantics is correct and is a model of .
Beginning the Proof of Theorem 2.5. We gradually introduce the vocabulary and
theory explaining the use of various predicates as they are introduced; we repeat a bit
6
9
5
8


r
e
v
i
s
i
o
n
:
2
0
1
1
-
0
5
-
2
3







m
o
d
i
f
i
e
d
:
2
0
1
1
-
0
5
-
2
3


of the proof sketch. Below we say certain conditions hold to mean they hold in any
model of T. We rst describe and T. In particular, contains unary predicates
Q
1
, Q
2
that partition the universe.
There is a binary relation <
0
, which is a partial order of Q
1
. There is a partial
function F mapping Q
1
Q
1
into Q
1
. We write F
a
for the partial function from Q
1
into Q
1
indexed by a. Any model of T satises: a
0
b implies F
a
F
b
.
We have two further properties of F. F
c0
is the empty function. For every a Q
1
and every e Q
1
, if e , domF
a
, then there are b, d Q
1
with a <
0
b and F
b
=
F
a
e, d).
Further there is a pairing function F
2
on Q
1
and an extensional relation R between
Q
1
and Q
2
so that each element of Q
2
codes a subset of Q
1
via R. We write U
b
for
a: R(b, a) (for a Q
1
and b Q
2
).
T asserts that Q
1
is preserved by g, that g is a permutation, and Q
1
(c
0
).
The set of U
a
: a Q
2
is closed under Boolean operations and if U
b
is such
a set so is F
a
(U
b
) for any a Q
1
. For each a Q
1
, there is b Q
2
such that
U
b
= c: c <
1
a.
Secondly, we turn to the description of
1
and T
1
. In
1
, there is a binary
1
-
relation <
1
, which is a linear order of Q
1
and such that Q
1
(x) implies x <
1
g(x) and
no element of Q
1
lies between x and g(x), and x <
1
c

implies g(x) <

1
c

. T
1
further
asserts (Q
1
, <
1
) is internally well-ordered in the following sense. For every a Q
2
,
if U
a
is non-empty, it has a <
1
-least element. Finally, there is a unary relation Q
0
such
that Q
0
Q
1
and T
1
asserts c
0

1
x <
1
c

if and only if Q
0
(x). (Q
0
is just an
abbreviation and <
1
is the crucial symbol added to create
1
.) Thus, each g
i
(c
0
) Q
0
and g
i
(c
0
) : i < name countably many elements of Q
0
which are <
1
- ordered in
order type .
The type p asserts Q
0
(x) and x is not a g
i
(c
0
) for any i < . Thus if p is omitted
in a model M, Q
0
(M) = g
i
(c
0
): i < .
Claim 3.3 If a model M of T
1
is such that its reduct to is an
1
-saturated model of
T but M omits p, (Q
1
, <
1
) is a well-ordering in M.
Proof. Suppose there is a countable <
1
-descending chain B = b
i
: i < in
(Q
1
, <
1
). Using the properties of F, we can dene a <
0
-increasing chain of a
n
in
Q
1
such that F
an
= c
0
, b
0
), . . . , g
n
(c
0
), b
n
), where the g
i
(c
0
) are images of c
0
by iterating g. Since the model is
1
-saturated there is an a

Q
1
such that each
F
an
F
a
. But then B = F
a
(g
i
(c
0
) : i < ). Note that while the choice of b
i
involved the
1
-symbol <
1
, the existence of a

is by the consistency of a -type so the

use of saturation is legitimate.
Since M omits p, g
i
(c
0
) : i < = Q
0
(M) = a: c
0
a <
1
c

and therefore
is coded by an element of Q
2
. By the closure properties of the coded sets, B = U
d
for
some d Q
2
. This contradicts the internal well-ordering of Q
1
.
3.3
Now translate to the rst order formula

(v) by translating each bound second

order variable Xto a rst order formula in x and v. Replace each occurrence of X(z) by
R(z, v) R(z, x). This translation has the following consequence. (This is immediate
for monadic second order but we included a pairing function F
2
on Q
1
so it extends to
arbitrary sentences.)
7
9
5
8


r
e
v
i
s
i
o
n
:
2
0
1
1
-
0
5
-
2
3







m
o
d
i
f
i
e
d
:
2
0
1
1
-
0
5
-
2
3


Fact 3.4 If M [= T, a Q
2
(M) and each subset of U
a
is coded by an element of
Q
2
(M), then M [=

(a) if and only U

a
(M) [= .
Add the following axiom to T
1
to obtain the theory T

:
(u)(w)[((z)R(z, w) z <
1
u)

(w)].
This completes t

as T, T

, p).
Claim 3.5 If < =
<
and M is a model of T

with cardinality that omits p but

whose reduct to is saturated then [= .
Conversely, if is true on all < =
<
, there is a model M
1
of T

with
cardinality that omits p but whose reduct to is saturated.
Proof. Since < , is an initial segment of Q
1
so = a Q
1
: R(y, d) for
some d Q
2
. But then each subset Y of gives rise to a type:
q
Y
(x) = R(y, d) R(y, x) : y Y R(y, x) : y , Y .
For each Y the -type q
Y
(x) has cardinality less than and so is realized by saturation.
We nish by Fact 3.4.
For the converse, well-order Q
1
by <
1
in order type . Add in Q
2
a code for
each subset of cardinality < . Let the F
a
list the partial functions of cardinality less
than from Q
1
to Q
1
and let <
0
denote the natural partial ordering on Q
1
induced by
inclusion of the named functions. Since is true below , each innite initial segment
in denes a model of and the denition of T

shows that we have a saturated model

of T when we take the reduct to . Finally, let Q
0
include exactly the rst elements
of Q
1
.

3.5
Concluding the Proof of Theorem 2.5.
Letting t

be the triple (T, T

, p) we have a triple satisfying Theorem 2.5.

2.5
4 The exact strength
In this section we remove the restrictions to nitely axiomatizable theories and count-
able languages. In Theorem 4.11 we prove actual equality of the Hanf number studied
here (for any triple of theories and types of any cardinality) with a L owenheim number
of second order logic; the cost is that we must move into innitary second order logic
and (in the proof) allow relation constants (i.e. predicate symbols other than equality)
in the vocabulary of the second order sentence.
Instead of Theorem 2.5 we could slightly more easily prove
H(P
f
N
)
2
(L
II
) H(P
c
N
),
which gives our answer to Newelskis question but is not quite as sharp. That is, if we
had just required t

in Theorem 2.5 to be in a countable language rather than nitely

axiomatizable, this would have no effect on the proof of Lemma 2.9 and it would have
8
9
5
8


r
e
v
i
s
i
o
n
:
2
0
1
1
-
0
5
-
2
3







m
o
d
i
f
i
e
d
:
2
0
1
1
-
0
5
-
2
3


simplied the proof of Theorem 2.5 since we could have worked with countably many
constants and omitted the function g. Similarly the arguments of Sections 2 and 3 ex-
tend from nitely axiomatizable to arithmetic by coding a model of arithmetic in the
second order sentence. And it is easy to see that the theory constructed in Theorem 2.5
is recursive. This observation is generalized in Theorem 4.11 to remove the restric-
tions on axiomatizability. The key idea is to see that we can use the same ideas as in
Section 3 to code the syntax of innitary second order logic by a triple t.
We extend our notion of second order logic in two ways. First we allow innite
conjunctions and strings of quantiers. Secondly we nowallowsome relation constants
instead of dealing with pure second order logic.
Denition 4.1 1. L

denotes the stage in the construction of the inner model L.

2. Let L

+
,
(II) denote second order logic allowing strings of second order quan-
tiers of cardinality <
+
and conjunctions and disjunctions of cardinality
.
Remark 4.2 Again using Assumption 0.3, note that as in Fact 1.3 the L owenheimnum-
ber of L

+
,
(II) is a strong limit cardinal of conality > and is an accumulation
point of : =
<
.
Notation 4.3 We denote by L(II,

) second order logic in the vocabulary

con-
sisting of constant symbols c

, c

, c

, c

, a unary predicate Q, and a binary relation

R
1
.
Notation 4.4 Let K = K
t
, where the set of triples t is as in Notation 2.1 but requiring
that
1
L

and omitting the requirement that p is second order denable. Then,

P

N
(K
t
, ) holds if [
1
[ and for some M
1
with [M
1
[ = , M
1
[= t.
In the following H() denotes the set of all sets whose transitive closure has car-
dinality less than . We use implicitly that
<
= implies that is regular and so
(H(), ) satises all axioms of ZFC except power set.
We now construct a sentence L(II,

) and for every L

and every in
L

+
,
(II)() a set A
,
of ordinals so that and A
,
codes .
Denition 4.5 Let X be a transitive subset of H(
+
) with cardinality . We say that
X is coded by A if there is an injection f from X into such that:
A = pr(f(a), f(b)): a b X.
Here pr is the standard pairing function on ordinals.
Now we show that the coding of X by A does not depend on the choice of f.
Lemma 4.6 If A codes X
1
by f
1
and X
2
by f
2
then X
1
= X
2
.
Proof. We rst note that for any X, A if A codes X by f then f is a bijection
between X and B = : pr(, ) A pr(, ) A. Applying this remark to
9
9
5
8


r
e
v
i
s
i
o
n
:
2
0
1
1
-
0
5
-
2
3







m
o
d
i
f
i
e
d
:
2
0
1
1
-
0
5
-
2
3


f
1
, f
2
, we see g = f
1
2
f
1
is a bijection fromX
1
onto X
2
. But then g is an isomorphism
with respect to as it is easy to check (from the denition of coding) that for any
y
1
, z
1
X
1
with y
2
= g(y
1
) and z
2
= g(z
1
), y
1
z
1
if and only if y
2
z
2
. That is,
y
1
z
1
if and only pr(f
1
(y
1
), f
1
(z
1
)) A if and only pr(f
2
(y
2
), f
2
(z
2
)) A if and
only y
2
z
2
. But then since X
1
X
2
, -induction yields that X
1
= X
2
.
4.6
We rst dene certain set of ordinals A
,
in V that codes and
L

+
,
(II)(). Here tc(X) denotes the transitive closure of X.
Denition 4.7 For every vocabulary L

and every sentence L

+
,
(II)()
we dene A
,
to be a set which codes tc(, ) , ) in the sense of Deni-
tion 4.5.
Such a code exists since the standard construction of L

+
,
(II)() yields
that each formula is in H(
+
), each subformula of tc(), and has at most
subformulas.
Now we dene the sentence L(II,

), which does not depend on or , so

that the interpretation of a predicate Qas A
,
in a model M of will identify as the
sentence under consideration. The function G in Denition 4.8.2c simply formalizes
the normal denition of truth.
Denition 4.8 We dene a sentence L(II,

) M [= if and only if M satises:

1. (M, R
M
1
) (H(), ) for some with
<
= and
2. the following

-axioms.
(a) c

and c

are cardinals in the sense of M.

(b) c

is a vocabulary of cardinality c

contained in L

.
(c) The set of formulas of L
c
+

,c
(II)(c

) are given their usual inductive de-

nition in M by a formula (x, c

, c

, c

) L(II)(

).
(d) c

L
c
+

,c
(II)(c

), i.e. M [= (c

, c

, c

, c

).
(e) Q(x) xR
1
c

.
(f) Q is a set of codes (in the sense of M) of tc(c

, c

) c

, c

), where tc
is with respect to R
1
.
3. There is a denable function Gwhich denes truth of sentences of L

+
,
(II, c

)
on c

-structures b

M: for every sentence , M [= G(b

) = 1 if and only
if
b [= .
(Here is an actual sentence, b is a c

-structure in H(), and

(b

) is the
member of M mapped to (b) under the isomorphism in Denition 4.8.1. )
The goal of the following lemma is to compute the L owenheim number of
L

+
,
(II). Since it is certainly greater than , we may assume
<
> .
10
9
5
8


r
e
v
i
s
i
o
n
:
2
0
1
1
-
0
5
-
2
3







m
o
d
i
f
i
e
d
:
2
0
1
1
-
0
5
-
2
3


Lemma 4.9 Fix
+
, a vocabulary L

and L

+
,
(II)(). Choose A
,
and L(II,

) satisfying Denitions 4.7 and 4.8, respectively. For any cardinal

=
<
> , the following are equivalent.
1. has no model of cardinality < .
2. There is a model (M, R
M
1
, Q
M
, c
M

, c
M

, c
M

, c
M

) with cardinality of the sen-

tence dened in Denition 4.8 such that letting P
M
denote
b: M [= b is an ordinal bR
M
1
c

and P
M
1
denote
b: M [= b is an ordinal bR
M
1
c

;
(a) (P
M
, R
M
1
) has order type ;
(b) (P
M
1
, R
M
1
) has order type ;
(c)
A
,
= < : for some a Q
M
, = otp(bR
1
c

: bR
M
1
a, R
M
1
).
Proof. Suppose 2). Without loss of generality, we identify (M, R
M
1
) with H(, ).
H() since L

. Then [M[ = , (P
M
, R
M
1
) has order type , and A
,
is the
image of Q
M
under an isomorphism from (P
M
, R
M
1
) to . By the choice of A
,
, the
model M correctly recognizes the vocabulary and the formula . The function G
M
correctly represents truth in M by Denition 4.8.3 and Denition 4.8.4. So fails on
all subsets of M with cardinality < by Denition 4.8 2d. Thus 2) implies 1). Clearly
if 1) holds we can construct a model M satisfying 2).

4.9
We continue to use the conventions regarding P
M
, P
M
1
from the proof of
Lemma 4.9.
Denition 4.10 For as in Denition 4.9, spec(, , , A
,
) is the set of the cardi-
nalities of models M of with (P
M
, P
M
1
, Q, R
M
1
) (, , A
,
, <).
Theorem 4.11 For any cardinals , , the following four cardinals are equal.
1.
1
is the Hanf number of P

N
.
2.
2
is the L owenheim number of L

+
,
(II) =
2
(L

+
,
(II)).
3.
3
is the L owenheim number of L

+
,
+(II) =
2
(L

+
,
+(II)).
4.
4
= supspec(, , , A
,
) : L(II,

), L

+
,
+(II), and A
,

such that spec(, , , A
,
) is bounded.
Proof. We chose the logic L

+
,
precisely so
1

2
(by a proof like that of
Lemma 2.6 but now we have conjunctions of cardinality ). So in fact the demand
that a model omits the type p becomes the model satises the sentence in L

+
,
,
11
9
5
8


r
e
v
i
s
i
o
n
:
2
0
1
1
-
0
5
-
2
3







m
o
d
i
f
i
e
d
:
2
0
1
1
-
0
5
-
2
3


(x)

p so the worries in the rst sections about the second order denability of p
disappear. Clearly
2

3
.
With from Lemma 4.9 and applying that lemma with as = and with =
+
)
yields:
min(spec
2
()) : L

+
,
+ sup(spec
2
(, , A
,
)) : L

+
,
+ is bounded.
(We can replace by a

whose only model is the model of with minimum cardi-

nality to guarantee the containment.) Thus,
3

4
.
The proof that
4

1
is obtained by modifying the proof of Theorem 2.5.
Add to the vocabulary in the T

from the proof in section 3 of Theorem 2.5, sym-

bols P, P
1
, Q, R
1
and symbols c

for each A
,
and use the same coding ideas
to guarantee that P
1
, Q are contained in P and all three are well-ordered by R
1
.
Thus, for L

+
,
+(II) we can construct t

, encoding the second order sen-

tence L(II,

) dened in Denition 4.8 and where the type p also codes that
Q
M
A
,
so that the two spectra are related as in Theorem 2.5. The type p is just
Q(x) x ,= c

: A
,
. This yields
4

1
by slightly modifying the
argument for Lemma 2.9.

4.11
Our discussion of the Hanf and L owenheim numbers of second order logic fo-
cused on two vocabularies: = and

. In contrast, in many studies of the Hanf and

L owenheim numbers of logics the number is taken as the supremum for a given logic
over all vocabularies of a bounded cardinality. That is, a L owenheim or Hanf function
(with argument the cardinality of the vocabulary) is dened:
Denition 4.12 For any logic L and any cardinal .
1. Let the L owenheim function
3
(L, ) be the least cardinal such that for any
vocabulary of cardinality and any L(), if has a model it has one
of cardinality less than or equal .
2. Let the Hanf function h
3
(L, ) be the least cardinal such that for any vocabu-
lary of cardinality and any L(), if has a model of cardinality , it
has arbitrarily large models.
The cardinalities of the vocabularies play a signicant role. A trivial example is
that both the L owenheim function and Hanf function in the sense of Denition 4.12
of rst order logic map the cardinality of the vocabulary to itself. A more interesting
example is that for the logic L(Q) (with Q interpreted as there exists uncountably
many), the L owenheim number in = = is
1
, while for an arbitrary vocabulary
it is
1
+ [[. The Hanf number of L(Q) for countable vocabularies is

(See e.g.
3.3.12 of [Sch85]). Thus,

= h
3
(L
,
(Q),
0
) < h
3
(L
1,
,
0
) =
1
.
Computing the Hanf function of L(Q) for vocabularies of cardinality
1
is
considerably more complicated. The key point is the following observation which
does not depend on Assumption 0.3. Since we have not been able to nd it in print we
describe the key innovation here.
12
9
5
8


r
e
v
i
s
i
o
n
:
2
0
1
1
-
0
5
-
2
3







m
o
d
i
f
i
e
d
:
2
0
1
1
-
0
5
-
2
3


Theorem 4.13 If is uncountable, h
3
(L
,
(Q), ) = h
3
(L
1,
, ).
Proof. Lopez-Escobar and Chang (e.g. [Cha68] showed how to code sentences of
L
1,
as rst order theories omitting types. Since each sentence will have only count-
ably many subformulas, each sentence regardless of the cardinality of the vocabulary
can be coded by a rst order sentence omitting countably many types. And it is known
(e.g. by the proof of Theorem 5.1.4 of [She78]) that this omitting types problem can be
reduced to omitting one type p of the form P(x) x ,= c
n
: n < . More precisely,
we can nd for any pair (T, ) in a vocabulary (where is a countable collection of
types) a pair (T
1
, p) in an expanded vocabulary
1
such that each -structure omitting
can be expanded to a
1
-structure omitting p and each model of T
1
omitting p also
omits each type in . This translation can be done for of any cardinality. The crux
of the current extension is that with an uncountable language one can further reduce
the omission of the type p to the assertion (Qx)P(x). For this, add to the language
unary function symbols F

for <
1
and a binary relation symbol <. Fix a family

: <
1
) of distinct functions from to 2. Now add axioms asserting
1. < linear orders P;
2. c
n+1
is the <-successor of c
n
and c
0
is <-minimal;
3. Each F

maps P onto P;
4. For < <
1
, if (

n ,=

n) then (x)(c
n
< x F

(x) ,= F

(x)).
Now in any model of these axioms if the type p is realized by some c, the F

(c) for
<
1
are uncountably many elements of P. This establishes the reduction.
4.13
Remark 4.14 This result can extended substantially; we could have replaced the study
of sentences of L
1,
by theories of cardinality
1
without loss. Basic facts con-
cerning the Hanf number of L

+
,0
for vocabularies of cardinality
1
appear
in VII.5 of [She78] and Chapter 4 of [Bal09]. General sufcient conditions to show
h
3
(L
1,
, ) = h
3
(L

+
,
, ) are studied in [GS05].
However, the dependence on the size of the vocabulary disappears for the
L owenheim number of innitary second order logic.Without loss of generality we re-
strict to relational languages. There is no loss in our restriction to vocabularies of
cardinality at most since no sentence of L

+
,
(II) can contain more than relation
symbols.
Note that just by existentially quantifying out the relation symbols, for any
+
:

3
(L

+
,
(II), )
2
(L

+
,
+(II)).
Combining this observation with
2
=
3
in Theorem 4.11, we have:
Corollary 4.15
3
(L

+
,
(II), ) =
2
(L

+
,
(II)).
13
9
5
8


r
e
v
i
s
i
o
n
:
2
0
1
1
-
0
5
-
2
3







m
o
d
i
f
i
e
d
:
2
0
1
1
-
0
5
-
2
3


We arrived at this result using our analysis of the Hanf number of P

N
. But a variant
of Corollary 4.15 can be obtained, using the same ideas of coding syntax, without
the detour through P

N
. It shows that with only a nite vocabulary one can code any
reasonable similarity type.
Notation 4.16

is the vocabulary containing one unary predicate P, one binary

relation symbol < and a ternary predicate R.
Theorem 4.17
3
(L

+
,
(II), ) =
2
(L

+
,
(II)(

)).
Proof. Instead of considering an arbitrary vocabulary of size we can consider
a universal vocabulary of binary relations. (It is easy to code the rst in the
second.) So the claim is that for any sentence L

+
,
(II)() there is a sentence

of L

+
,
(II)(

) such that spec() = spec(

). (The converse is obvious.)

Let

assert
1. (P, <) is a well-order.
2. the formula obtained by replacing each occurence of R
i
(
1
,
2
) in by the
formula (where z does not occur in ):
(z)[P(z) otp(y : P(y) y < z) = i R(z,
1
,
2
)].
(Here
1
,
2
are arbitrary terms from L

+
,
(II)(), i.e. individual vari-
ables.)
Now there is a model N with cardinality of

where P
N
has order type ,
such that N [=

if and only if there is a -structure M with cardinality satisfying

.
4.17
.
References
[Bal09] John T. Baldwin. Categoricity. Number 51 in University Lecture Notes.
American Mathematical Society, 2009. www.math.uic.edu/ jbaldwin.
[Cha68] C. C. Chang. Some remarks on the model theory of innitary languages. In
J. Barwise, editor, The syntax and semantics of innitary languages, pages
3664. Springer-Verlag, 1968. LNM 72.
[GS05] R. Goebel and S. Shelah. How rigid are reduced products? J Pure and
Applied Algebra, 202:230258, 2005.
[Han60] William Hanf. Models of languages with innitely long expressions. In
Abstracts of Contributed papers from the First Logic, Methodology and
Philosopy of Science Congress, Vol.1, page 24. Stanford University, 1960.
[Mag71] M. Magidor. On the role of supercompact and extendible cardinals in logic.
Israel Journal of Mathematics, 10:147157, 1971.
14
9
5
8


r
e
v
i
s
i
o
n
:
2
0
1
1
-
0
5
-
2
3







m
o
d
i
f
i
e
d
:
2
0
1
1
-
0
5
-
2
3


[New] Ludomir Newelski. Bounded orbits and measures on a group. preprint.
[Sch85] J. Schmerl. Transfer theorems and their applications to logics. In J. Barwise
and S. Feferman, editors, Model-Theoretic Logics, pages 177209. Springer-
Verlag, 1985.
[She78] S. Shelah. Classication Theory and the Number of Nonisomorphic Models.
North-Holland, 1978.
[Vaa79] Jouko Vaananen. On the Hanf numbers of unbounded logics. In B.Mayoh
F.Jensen and K.Moller, editors, Proceedings from 5th Scandinavian Logic
Symposium, pages 309328. Aalborg University Press, 1979.
[Vaa82] Jouko Vaananen. Abstract logic and set theory II: Large cardinals. Journal
of Symbolic Logic, 47:335346, 1982.
15