8

2
5


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







m
o
d
i
f
i
e
d
:
2
0
0
3
-
1
0
-
1
4


A denable nonstandard model of the reals
Vladimir Kanovei

Saharon Shelah

August 2003
Abstract
We prove, in ZFC, the existence of a denable, countably saturated elemen-
tary extension of the reals.
Introduction
It seems that it has been taken for granted that there is no distinguished, denable
nonstandard model of the reals. (This means a countably saturated elementary exten-
sion of the reals.) Of course if V = L then there is such an extension (just take the
rst one in the sense of the canonical well-ordering of L), but we mean the existence
provably in ZFC. There were good reasons for this: without Choice we cannot prove
the existence of any elementary extension of the reals containing an innitely large
integer.
1 2
Still there is one.
Theorem 1 ( ZFC). There exists a denable, countably saturated extension

R of the
reals R, elementary in the sense of the language containing a symbol for every nitary
relation on R.
The problem of the existence of a denable proper elementary extension of R was
communicated to one of the authors (Kanovei) by V. A. Uspensky.
A somewhat dierent, but related problem of unique existence of a nonstandard
real line

R has been widely discussed by specialists in nonstandard analysis.
3
Keisler
notes in [3, 11] that, for any cardinal , either inaccesible or satisfying 2

=
+
, there
exists unique, up to isomorphism, -saturated nonstandard real line

R of cardinality
, which means that a reasonable level of uniqueness modulo isomorphism can be

Partial support of RFFI grant 03-01-00757 and DFG grant acknowledged.

Supported by The Israel Science Foundation. Publication 825.

1
In fact, from any nonstandard integer we can dene a non-principal ultralter on N, even a
Lebesgue non-measurable set of reals [4], yet it is consistent with ZF (even plus Dependent Choices)
that there are no such ultralters as well as non-measurable subsets of R [5].
2
It is worth to be mentioned that denable nonstandard elementary extensions of N do exist in
ZF. For instance, such a model can be obtained in the form of the ultrapower F/U, where F is the
set of all arithmetically denable functions f : N N while U is a non-principal ultralter in the
algebra A of all arithmetically denable sets X N.
3
What is needed is an underlying set theory which proves the unique existence of the hyperreal
number system [. . . ] (Keisler [3, p. 229]).
1
8
2
5


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







m
o
d
i
f
i
e
d
:
2
0
0
3
-
1
0
-
1
4


achieved, say, under GCH. Theorem 1 provides a countably saturated nonstandard
real line

R, unique in absolute sense by virtue of a concrete denable construction in
ZFC. A certain modication of this example also admits a reasonable model-theoretic
characterization up to isomorphism (see Section 4).
The proof of Theorem 1 is a combination of several known arguments. First of
all (and this is the key idea), arrange all non-principal ultralters over N in a linear
order A, where each ultralter appears repetitiously as D
a
, a A. Although A is
not a well-ordering, we can apply the iterated ultrapower construction in the sense of
[1, 6.5] (which is a nite support iteration in the forcing nomenclature), to obtain
an ultralter D in the algebra of all sets X N
A
concentrated on a nite number
of axes N. To dene a D-ultrapower of R, the set F of all functions f : N
A
R,
also concentrated on a nite number of axes N, is considered. The ultrapower F/D is
OD, thar is, ordinal-denable, actually, denable by an explicit construction in ZFC,
hence, we obtain an OD proper elementary extension of R. Iterating the D-ultrapower
construction
1
times in a more ordinary manner, i. e., with direct limits at limit steps,
we obtain a denable countably saturated extension.
To make the exposition self-contained and available for a reader with only fragmen-
tary knowledge of ultrapowers, we reproduce several well-known arguments instead of
giving references to manuals.
1 The ultralter
As usual, c is the cardinality of the continuum.
Ultralters on N hardly admit any denable linear ordering, but maps a : c
P(N), whose ranges are ultralters, readily do. Let A consist of all maps a : c
P(N) such that the set D
a
= ran a = a() : < c is an ultralter on N. The
set A is ordered lexicographically: a <
lex
b means that there exists < c such that
a = b and a() < b() in the sense of the lexicographical linear order < on
P(N) (in the sense of the identication of any u N with its characterictic function).
For any set u, N
u
denotes the set of all maps f : u N.
Suppose that u v A.
If X N
v
then put X u = x u : x X.
If Y N
u
then put Y v = x N
v
: x u Y .
We say that a set X N
A
is concentrated on u A, if X = (X u) A; in other
words, this means the following:
x, y N
A

x u = y u = (x X y X)

. ()
We say that X is a set of nite support, if it is concentrated on a nite set u A. The
collection X of all sets X N
A
of nite support is closed under unions, intersections,
complements, and dierences, i. e., it is an algebra of subsets of N
A
. Note that if ()
holds for nite sets u, v A then it also holds for u v. (If x (u v) = y (u v)
then consider z N
A
such that z u = x u and z v = y v .) It follows that for any
X X there is a least nite u = [[X[[ A satisfying ().
In the remainder, if U is any subset of P(I), where I is a given set, then Ui (i)
(generalized quantier) means that the set i I : (i) belongs to U.
2
8
2
5


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







m
o
d
i
f
i
e
d
:
2
0
0
3
-
1
0
-
1
4


The following denition realizes the idea of a nite iteration of ultralters. Suppose
that u = a
1
< < a
n
A is a nite set. We put
D
u
= X N
u
: D
an
k
n
. . . D
a
2
k
2
D
a
1
k
1
(k
1
, k
2
, ..., k
n
) X) ;
D = X X : X [[X[[ D
||X||
.
The following is quite clear.
Proposition 2. (i) D
u
is an ultralter on N
u
;
(ii) if u v A, v nite, X N
u
, then X D
u
i X v D
v
;
(iii) D X is an ultralter in the algebra X ;
(iv) if X X, u A nite, and [[X[[ u, then X D X u D
u
.
2 The ultrapower
To match the nature of the algebra X of sets X N
A
of nite support, we consider
the family F of all f : N
A
R, concentrated on some nite set u A, in the sense
that
x, y N
A

x u = y u = f(x) = f(y)

. ()
As above, for any f F there exists a least nite u = [[f[[ A satisfying ().
Let R be the set of all nitary relations on R. For any n-ary relation E R and
any f
1
, ..., f
n
F, dene
E
D
(f
1
, ..., f
n
) D x N
A
E(f
1
(x), ..., f
n
(x)) .
The set X = x N
A
: E(f
1
(x), ..., f
n
(x)) is obviously concentrated on u = [[f
1
[[
[[f
n
[[, hence, it belongs to X, and [[X[[ u = [[f
1
[[ [[f
n
[[.
In particular, f =
D
g means that D x N
A
(f(x) = g(x)). The following is clear:
Proposition 3. =
D
is an equivalence relation on F, and any relation on F of the
form E
D
is =
D
-invariant.
Put [f]
D
= g F : f =
D
g, and

R = F/D = [f]
D
: f F. For any n-ary
( n 1) relation E R, let

E be the relation on

R dened as follows:

E([f
1
]
D
, ..., [f
n
]
D
) i E
D
(f
1
, ..., f
n
) i D x N
A
E(f
1
(x), ..., f
n
(x)).
The independence on the choice of representatives in the classes [f
i
]
D
follows from
Proposition 3. Put

R =

E : E R. Finally, for any r R we put

r = [c
r
]
D
,
where c
r
F satises c
r
(x) = r, x.
Let L be the rst-order language containing a symbol E for any relation E R.
Then R; R) and

R;

R) are L-structures.
Theorem 4. The map r

r is an elementary embedding (in the sense of the
language L ) of the structure R; R) into

R;

R) .
3
8
2
5


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







m
o
d
i
f
i
e
d
:
2
0
0
3
-
1
0
-
1
4


Proof. This is a routine modication of the ordinary argument. By L[F] we denote
the extension of L by functions f F used as parameters. It does not have a direct
semantics, but if is a formula of L[F] and x N
A
then [x] will denote the
formula obtained by the substitution of f(x) for any f F which occurs in . Thus,
[x] is an L-formula with parameters in R.
Lemma 5 ( Los). For any closed L[F]-formula (f
1
, ..., f
n
) (all parameters f
i
F
indicated), we have:

R;

R) [= ([f
1
]
D
, ..., [f
n
]
D
) D x (R; R) [= (f
1
, ..., f
n
)[x]).
Proof. We argue by induction on the logic complexity of . For an atomic relation
E(f
1
, ..., f
n
), the result follows by the denition of

E. The only notable induction step
is in the direction = . Suppose that is y (y, f
1
, ..., f
n
), and
D x (R; R) [= (f
1
, ..., f
n
)[x]), that is, D x (R; R) [= y (y, f
1
, ..., f
n
)[x]) .
Obviously there exists a function f F, concentrated on u = [[f
1
[[ [[f
n
[[, such
that, for any x N
A
, if there exists a real y satisfying R; R) [= (y, f
1
, ..., f
n
)[x],
then y = f(x) also satises this formula, i. e., R; R) [= (f, f
1
, ..., f
n
)[x]. Formally,
x N
A

y R (R; R) [= (y, f
1
, ..., f
n
)[x]) = R; R) [= (f, f
1
, ..., f
n
)[x]

.
This implies D x (R; R) [= (f, f
1
, ..., f
n
)[x]). Then, by the inductive assumption,

R;

R) [= ([f]
D
, [f
1
]
D
, ..., [f
n
]
D
), hence

R;

R) [= ([f
1
]
D
, ..., [f
n
]
D
), as required.
(Lemma)
To accomplish the proof of Theorem 4, consider a closed L-formula (r
1
, ..., r
n
)
with parameters r
1
, ..., r
n
R. We have to prove the equivalence
R; R) [= (r
1
, ..., r
n
)

R;

R) [= (

r
1
, ...,

r
n
) .
Let f
i
= c
r
i
, thus, f
i
F and f
i
(x) = r
i
, x. Obviously (f
1
, ..., f
n
)[x] coincides with
(r
1
, ..., r
n
) for any x N
A
, hence (r
1
, ..., r
n
) is equivalent to D x (f
1
, ..., f
n
)[x].
On the other hand, by denition,

r
i
= [f
i
]
D
. Now the result follows by Lemma 5.
3 The iteration
Theorem 4 yields a denable proper elementary extension

R;

R) of the structure
R; R). Yet this extension is not countably saturated due to the fact that the ultra-
power

R was dened with maps concentrated on nite sets u A only. To x this
problem, we iterate the extension used above
1
-many times.
Suppose that M ; M) is an L-structure, so that M consists of nitary relations
on a set M, and for any E R there is a relation E
M
M of the same arity,
associated with E. Let F
M
be the set of all maps f : N
A
M concentrated on
nite sets u A. The structure F
M
/D =

M ;

M), dened as in Section 2, but with
the modied F, will be called the D-ultrapower of M ; M). Theorem 4 remains true
in this general setting: the map x

x (x M) is an elementary embedding of
M ; M) in

M ;

M) .
We dene a sequence of L-structures M

; M

),
1
, together with a system
of elementary embeddings e

: M

; M

) M

; M

), <
1
, so that
4
8
2
5


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







m
o
d
i
f
i
e
d
:
2
0
0
3
-
1
0
-
1
4


(i) M
0
; M
0
) = R; R) ;
(ii) M
+1
; M
+1
) is the D- ultrapower of M

; M

), that is, M
+1
; M
+1
) =
F

/D, where F

= F
M
consists of all functions f : N
A
M

concentrated on
nite sets u A. In addition, e
,+1
is the associated

-embedding M

; M

)
M
+1
; M
+1
), while e
,+1
= e
,+1
e

for any < (in other words,

e
,+1
(x) = e
,+1
(e

(x)) for all x M

);
(iii) if
1
is a limit ordinal then M

; M

) is the direct limit of the structures

M

; M

), < . This can be achieved by the following steps:

(a) M

is dened as the set of all pairs , x) such that x M

and x , rane

for all < .

(b) If E R is an n-ary relation symbol then we dene an n-ary relation E

on M

as follows. Suppose that x

i
=
i
, x
i
) M

for i = 1, ..., n. Let

= sup
1
, ...,
n
and z
i
= e

i
,
(x
i
) for every i, so that
i
<
and z
i
M

. (Note that if
i
= then e

i
,
is the identity.) Dene
E

(x
1
, ..., x
n
) i M

; M

) [= E(z
1
, ..., z
n
).
(c) Put M

= E

: E R then M

; M

) is an L-structure.
(d) Dene an embedding e

: M

( < ) as follows. Consider any

x M

. If there is a least < such that there exists an element y M

with x = e

(y) then let e

(x) = , y). Otherwise put e

(x) = , x) .
A routine verication of the following is left to the reader.
Proposition 6. If <
1
then e

is an elementary embedding of M

; M

)
to M

; M

) .
Note that the construction of the sequence of models M

; M

) is denable, hence,
so is the last member M

1
; M

1
) of the sequence. It remains to prove that the L-
structure M

1
; M

1
) is countably saturated.
This is also a simple argument. Suppose that, for any k,
k
(p
k
, x) is an L-formula
with a single parameter p
k
M

1
(the case of many parameters does not essentially
dier from the case of one parameter), and there exists an element x
k
M

1
such
that

ik

i
(p
i
, x
k
) is true in M

1
; M

1
) in other words, we have M

1
; M

1
) [=

i
(p
i
, x
k
) whenever k i. Fix an ordinal <
1
such that for any k, i there exist
(then obviously unique) y
k
, q
i
M

with x
k
= e

1
(y
k
) and p
i
= e

1
(q
i
). Then

i
(q
i
, y
k
) is true in M

; M

) whenever k i .
Fix a A such that D
a
is a non-principal ultralter, that is, all conite sub-
sets of N belong to D
a
. Consider the structure M
+1
; M
+1
) as the D-ultrapower
of M

; M

). The corresponding set F

consists of all functions f : N

A
M

concentrated on nite sets u A. In particular, the map f(x) = y

k
whenewer
x(a) = k belongs to F

. As any set of the form k : k i belongs to D

a
, we
have D
a
k (M

; M

) [=
i
(q
i
, y
k
)), that is, D x N
A
(M

; M

) [=
i
(q
i
, f)[x]), for
any i N. It follows, by Lemma 5, that
i
(

q
i
, y) holds in M
+1
; M
+1
) for any i,
where

q
i
= e
,+1
(q
i
) M
+1
while y = [f]
D
M
+1
is the D-equivalence class of
f in F

. Put x = e
+1,
1
(y); then
i
(p
i
, x) is true in M

1
; M

1
) for any i because
obviously p
i
= e
+1,
1
(

q
i
), i .
(Theorem 1)
5
8
2
5


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







m
o
d
i
f
i
e
d
:
2
0
0
3
-
1
0
-
1
4


4 Varia
By appropriate modications of the constructions, the following can be achieved:
1. For any given innite cardinal , a -saturated elementary extension of R, de-
nable with as the only parameter of denition.
2. A special elementary extension of R, of as large cardinality as desired. For in-
stance, take, in stage of the construction considered in Section 3, ultralters
on

. Then the result will be a denable special structure of cardinality

1
.
Recall that special models of equal cardinality are isomorphic [1, Theorem 5.1.17].
Therefore, such a modication admits an explicit model-theoretical characteriza-
tion up to isomorphism.
3. A class-size denable elementary extension of R, -saturated for any cardinal .
4. A class-size denable elementary extension of the whole set universe, -saturated
for any cardinal . (Note that this cannot be strengthened to Ord-saturation, i. e.,
saturation with respect to all class-size families. For instance, Ord
M
-saturated
elementary extensions of a minimal transitive model M [= ZFC, denable in M,
do not exist see [2, Theorem 2.8].)
The authors thank the anonimous referee for valuable comments and corrections.
References
[1] C. C. Chang and H. J. Keisler, Model Theory, 3rd ed., North Holland, Amsterdam, 1992,
xiv + 650 pp. (Studies in logic and foundations of mathematics, 73).
[2] V. Kanovei and M. Reeken, Internal approach to external sets and universes, part 1,
Studia Logica, 1995, 55, no. 2, pp. 229257.
[3] H. J. Keisler, The hyperreal line, in P. Erlich (ed.) Real numbers, generalizations of
reals, and theories of continua, Kluwer Academic Publishers, 1994, pp. 207237.
[4] W. A. J. Luxemburg, What is nonstandard analysis ? Amer. Math. Monthly 1973, 80
(Supplement), pp. 3867.
[5] R. M. Solovay, A model of set theory in which every set of reals is Lebesgue measurable,
Ann. of Math. 1970, 92, pp. 156.
Vladimir Kanovei
Institute for information transmission problems (IPPI), Russian academy of sciences,
Bol. Karetnyj Per. 19, Moscow 127994, Russia
E-mail address : kanovei@mccme.ru
Saharon Shelah
Institute of Mathematics, The Hebrew University of Jerusalem, 91904 Jerusalem, Israel,
and Department of Mathematics, Rutgers University, New Brunswick, NJ 08854, USA
E-mail address : shelah@math.huji.ac.il
URL: http://www.math.rutgers.edu/~shelah
6