Quantifier Elimination and Applications

David Marker
Mathematical Logic
Fall 2015

1 Quantifier Elimination
In model theory we try to understand structures by studying their definable
sets. Recall that if M is an L-structure, then X ⊆ M n is definable if there is
an L-formula φ(v1 , . . . , vn , w1 , . . . , wm ) and b1 , . . . , bm ∈ M such that

X = {a ∈ M n : M |= φ(a, b)}.

The study of definable sets is often complicated by quantifiers. For example,

in the structure (N, +, ·, <, 0, 1) the quantifier-free definable sets are defined by
polynomial equations and inequalities. Even if we use only existential quantifiers
the definable sets become complicated. By the Matijasevič–Robinson–Davis–
Putnam solution to Hilbert’s 10th problem every recursively enumerable subset
of N is defined by a formula

∃v1 . . . ∃vn p(x, v1 , . . . , vn ) = 0

for some polynomial p ∈ N[X, Y1 , . . . , Yn ]. As we allow more alternations of

quantifiers, we get even more complicated definable sets.
Not surprisingly, it will be easiest to study definable sets that are defined by
quantifier-free formulas. Sometimes formulas with quantifiers can be shown to
be equivalent to formulas without quantifiers. Here are two well-known exam-
ples. Let φ(a, b, c) be the formula

∃x ax2 + bx + c = 0.

By the quadratic formula,

R |= φ(a, b, c) ↔ [(a 6= 0 ∧ b2 − 4ac ≥ 0) ∨ (a = 0 ∧ (b 6= 0 ∨ c = 0))],

whereas in the complex numbers

C |= φ(a, b, c) ↔ (a 6= 0 ∨ b 6= 0 ∨ c = 0).

In either case, φ is equivalent to a quantifier-free formula. However, φ is not

equivalent to a quantifier-free formula over the rational numbers Q.

1
 For a second example, let φ(a, b, c, d) be the formula

∃x∃y∃u∃v (xa + yc = 1 ∧ xb + yd = 0 ∧ ua + vc = 0 ∧ ub + vd = 1).

The formula φ(a, b, c, d) asserts that the matrix

 
a b
c d

is invertible. By the determinant test,

F |= φ(a, b, c, d) ↔ ad − bc 6= 0

for any field F .

Definition 1.1 We say that a theory T has quantifier elimination if for every
formula φ there is a quantifier-free formula ψ such that

T |= φ ↔ ψ.

Our goal in this section is to give a very useful model theoretic test for
elimination of quantifiers. In the next section we will show that this method
can be applied to the theory of algebraically closed fields and develop some rich
consequences. We begin by introducing some preliminary tools.

Diagrams
We begin by giving a way to construct L-embeddings.
Definition 1.2 Suppose that M is an L-structure. Let LM be the language
where we add to L constant symbols m for each element of M . The atomic dia-
gram of M is {φ(m1 , . . . , mn ) : φ is either an atomic L-formula or the negation
of an atomic L-formula and M |= φ(m1 , . . . , mn )}. We let Diag(M) denote the
atomic diagram of M

Lemma 1.3 Suppose that N is an LM -structure and N |= Diag(M); then,

viewing N as an L-structure, there is an L-embedding of M into N .

Proof Let j : M → N be defined by j(m) = mN ; that is, j(m) is the interpre-

tation of this constant symbol m in N . If m1 , m2 are distinct elements of M ,
then m1 6= m2 ∈ Diag(M); thus, j(m1 ) 6= j(m2 ) so j is an embedding. If f is a
function symbol of L and f M (m1 , . . . , mn ) = mn+1 , then f (m1 , . . . , mn ) =
mn+1 is a formula in Diag(M) and f N (j(m1 ), . . . , j(mn )) = j(mn+1 ). If
R is a relation symbol and m ∈ RM , then R(m1 , . . . , mn ) ∈ Diag(M) and
(j(m1 ), . . . , j(mn )) ∈ RN . Hence, j is an L-embedding.

2
Quantifier Elimination Tests
Theorem 1.4 Suppose that L contains a constant symbol c, T is an L-theory,
and φ(v) is an L-formula. The following are equivalent:
i) There is a quantifier-free L-formula ψ(v) such that T |= ∀v (φ(v) ↔ ψ(v)).
ii) If M and N are models of T , A is an L-structure, A ⊆ M, and A ⊆ N ,
then M |= φ(a) if and only if N |= φ(a) for all a ∈ A.

Proof i)⇒ ii) Suppose that T |= ∀v (φ(v) ↔ ψ(v)), where ψ is quantifier-free.

Let a ∈ A, where A is a common substructure of M and N and the latter
two structures are models of T . In Proposition ??, we saw that quantifier-free
formulas are preserved under substructure and extension. Thus

M |= φ(a) ⇔ M |= ψ(a)
⇔ A |= ψ(a) (because A ⊆ M)
⇔ N |= ψ(a) (because A ⊆ N )
⇔ N |= φ(a).

ii) ⇒ i) First, if T |= ∀v φ(v), then T |= ∀v (φ(v) ↔ c = c). Second, if

T |= ∀v ¬φ(v), then T |= ∀v (φ(v) ↔ c 6= c).
Thus, we may assume that both T ∪ {φ(v)} and T ∪ {¬φ(v)} are satisfiable.
Let Γ(v) = {ψ(v) : ψ is quantifier-free and T |= ∀v (φ(v) → ψ(v))}. Let
d1 , . . . , dm be new constant symbols. We will show that T ∪ Γ(d) |= φ(d). Then,
by compactness, there are ψ1 , . . . , ψn ∈ Γ such that
n
!
^
T |= ∀v ψi (v) → φ(v) .
i=1

Thus !
n
^
T |= ∀v ψi (v) ↔ φ(v)
i=1
n
^
and ψi (v) is quantifier-free. We need only prove the following claim.
i=1
Claim T ∪ Γ(d) |= φ(d).
Suppose not. Let M |= T ∪ Γ(d) ∪ {¬φ(d)}. Let A be the substructure of
M generated by d.
Let Σ = T ∪ Diag(A) ∪ φ(d). If Σ is unsatisfiable, then there are quantifier-
free formulas ψ1 (d), . . . , ψn (d) ∈ Diag(A) such that
n
!
^
T |= ∀v ψi (v) → ¬φ(v) .
i=1

But then !
n
_
T |= ∀v φ(v) → ¬ψi (v) ,
i=1

3
 n
_ n
_
so ¬ψi (v) ∈ Γ and A |= ¬ψi (d), a contradiction. Thus, Σ is satisfiable.
i=1 i=1
Let N |= Σ. Then N |= φ(d). Because Σ ⊇ Diag(A), A ⊆ N , by Lemma
1.3 i). But M |= ¬φ(d); thus, by ii), N |= ¬φ(d), a contradiction.
The proof above can easily be adapted to the case where L contains no
constant symbols. In this case, there are no quantifier-free sentences, but for
each sentence we can find a quantifier-free formula ψ(v1 ) such that T |= φ ↔
ψ(v1 ).
The next lemma shows that we can prove quantifier elimination by getting
rid of one existential quantifier at a time.

Lemma 1.5 Let T be an L-theory. Suppose that for every quantifier-free L-

formula θ(v, w) there is a quantifier-free formula ψ(v) such that T |= ∃w θ(v, w) ↔
ψ(v). Then, T has quantifier elimination.

Proof Let φ(v) be an L-formula. We wish to show that T |= ∀v (φ(v) ↔

ψ(v)) for some quantifier-free formula φ(v). We prove this by induction on the
complexity of φ(v).
If φ is quantifier-free, there is nothing to prove. Suppose that for i = 0, 1,
T |= ∀v (θi (v) ↔ ψi (v)), where ψi is quantifier free.
If φ(v) = ¬θ0 (v), then T |= ∀v (φ(v) ↔ ¬ψ0 (v)).
If φ(v) = θ0 (v) ∧ θ1 (v), then T |= ∀v (φ(v) ↔ (ψ0 (v) ∧ ψ1 (v))).
In either case, φ is equivalent to a quantifier-free formula.
Suppose that T |= ∀v(θ(v, w) ↔ ψ0 (v, w)), where ψ0 is quantifier-free and
φ(v) = ∃wθ(v, w). Then T |= ∀v (φ(v) ↔ ∃w ψ0 (v, w)). By our assumptions,
there is a quantifier-free ψ(v) such that T |= ∀v (∃w ψ0 (v, w) ↔ ψ(v)). But
then T |= ∀v (φ(v) ↔ ψ(v)).
Combining Theorem 1.4 and Lemma 1.5 gives us the following simple, yet
useful, test for quantifier elimination.

Corollary 1.6 Let T be an L-theory. Suppose that for all quantifier-free for-
mulas φ(v, w), if M, N |= T , A is a common substructure of M and N , a ∈ A,
and there is b ∈ M such that M |= φ(a, b), then there is c ∈ N such that
N |= φ(a, c). Then, T has quantifier elimination.

Theories with Quantifier Elimination

We conclude with several observations about theories with quantifier elimina-
tion.
Definition 1.7 An L-theory T is model-complete M ≺ N whenever M ⊆ N
and M, N |= T .
Stated in terms of embeddings: T is model-complete if and only if all em-
beddings are elementary.

Proposition 1.8 If T has quantifier elimination, then T is model-complete.

4
Proof Suppose that M ⊆ N are models of T . We must show that M is
an elementary submodel. Let φ(v) be an L-formula, and let a ∈ M . There
is a quantifier-free formula ψ(v) such that M |= ∀v (φ(v) ↔ ψ(v)). Because
quantifier-free formulas are preserved under substructures and extensions, M |=
ψ(a) if and only if N |= ψ(a). Thus

M |= φ(a) ⇔ M |= ψ(a) ⇔ N |= ψ(a) ⇔ N |= φ(a).

There are model-complete theories that do not have quantifier elimination,

but model completeness implies that we can eliminate all but the last existential
quantifiers.

Proposition 1.9 If T is model complete, then for any formula φ(v), there is a
quantifier free formula ψ(v, w) such that

T |= ∀v [φ(v) ↔ ∃w ψ(v, w)] .

Let us just point out the following test for completeness of model-complete
theories.

Proposition 1.10 Let T be a model-complete theory. Suppose that there is

M0 |= T such that M0 embeds into every model of T . Then, T is complete.

Proof If M |= T , the embedding of M0 into M is elementary. In particular

M0 ≡ M. Thus, any two models of T are elementarily equivalent.
We will use Proposition 1.10 below in cases where Vaught’s test does not
apply.
We have provided a number of proofs of quantifier elimination without ex-
plicitly explaining how to take an arbitrary formula and produce a quantifier
free one. In all of these cases, one can give explicit effective procedures. After
the fact, the following lemma tells us that there is an algorithm to eliminate
quantifiers.

Proposition 1.11 Suppose that T is a decidable theory with quantifier elimi-

nation. Then, there is an algorithm which when given a formula φ as input will
output a quantifier-free formula ψ such that T |= φ ↔ ψ.

Proof Given input φ(v) we search for a quantifier-free formula ψ(v) such that
T |= ∀v (φ(v) ↔ ψ(v)). Because T is decidable this is an effective search.
Because T has quantifier elimination, we will eventually find ψ.

5
2 Algebraically Closed Fields
We now return to the theory of algebraically closed fields. In Proposition ??,
we proved that the theory of algebraically closed fields of a fixed characteristic
is complete. We begin this section by showing that algebraically closed fields
have quantifier elimination. For convenience we will formulate ACF in the
language L = {+, −, ·, 0, 1}. We add − to the language, so that substructures
are integral domains. Without − we would have weaker structures that are a
bit more cumbersome to deal with.

Theorem 2.1 ACF has quantifier elimination.

Proof
Suppose K and L are algebraically closed fields and A is an integral domain
with A ⊆ K ∩L. By Corollary 5.6, we need to show that if φ(v, w) is a quantifier
free formula, a ∈ A, b ∈ K and K |= φ(b, a), then there is c ∈ L such that
L |= φ(c, a).
Let F be the algebraic closure of the fraction field of A . We, may without
loss of generality, assume that F ⊆ K ∩L. It will be enough to show that , a ∈ F ,
and K |= φ(b, a) for some b ∈ K, then there is c ∈ F such that F |= φ(c, a), for
then, by Proposition 1.8, L |= φ(c, a).
We first note that φ can be put in disjunctive normal form, namely there
are atomic or negated atomic formulas θi,j (v, w) such that:
n ^
_ m
φ(v, w) ↔ θi,j (v, w).
i=1 j=1
Vm
Because K |= φ(a, b), K |= j=1 θi,j (a, b) for some i. Thus, without loss
of generality, we may assume that φ is a conjunction of atomic and negated
atomic formulas. In our language atomic formulas θ(v1 , . . . , vn ) are of the form
p(v) = 0, where p ∈ Z[X1 , . . . , Xn ]. If p(X, Y ) ∈ Z[X, Y ], we can view p(X, a) as
a polynomial in F [X]. Thus, there are polynomials p1 , . . . , pn , q1 , . . . , qm ∈ F [X]
such that φ(v, a) is equivalent to
n
^ m
^
pi (v) = 0 ∧ qi (v) 6= 0.
i=1 i=1

If any of the polynomials pi are nonzero, then b is algebraic over F . In this case,
because F is algebraically closed, b ∈ F . Thus, we may assume that φ(v, a) is
equivalent to
^m
qi (v) 6= 0.
i=1

But qi (X) = 0 has only finitely many solutions for each i ≤ m. Thus, there are
only finitely many elements of F that do not satisfy F . Because algebraically
closed fields are infinite, there is a c ∈ F such that F |= φ(c, a).

6
Corollary 2.2 ACF is model-complete and ACFp is complete where p = 0 or
p is prime.
Proof Model-completeness is an immediate consequence of quantifier elimina-
tion.
The completeness of ACFp was proved in Proposition ??, but it also follows
from quantifier elimination. Suppose that K, L |= ACFp . Let φ be any sentence
in the language of rings. By quantifier elimination, there is a quantifier-free
sentence ψ such that
ACF |= φ ↔ ψ.
Because quantifier-free sentences are preserved under extension and substruc-
ture,
K |= ψ ⇔ Fp |= ψ ⇔ L |= ψ,
where Fp is the p-element field if p > 0 and the rationals if p = 0. Thus,

K |= φ ⇔ K |= ψ ⇔ L |= ψ ⇔ L |= φ.

Thus K ≡ L and ACFp is complete.

Definable Sets and Constructible Sets

Quantifier elimination has a geometric interpretation. We begin by looking at
the sets defined by quantifier free formulas.

Lemma 2.3 Let K be a field. The subsets of K n defined by atomic formulas

are exactly those of the form V (p) = {x for some p ∈ K[X]. A subset of K n
is quantifier-free definable if and only if it is a Boolean combination of Zariski
closed subsets.

Proof If φ(x, y) is an atomic Lr -formula, then there is q(X, Y ) ∈ Z[X, Y ]

such that φ(x, y) is equivalent to q(x, y) = 0. If X = {x : φ(x, a)}, then
X = V (q(X, a)) and q(X, a) ∈ K[X]. On the other hand, if p ∈ K[X], there is
q ∈ Z[X, Y ] and a ∈ K m such that p(X) = q(X, a). Then, V (p) is defined by
the quantifier-free formula q(X, a) = 0.
If X ⊆ K n is a finite Boolean combination of Zariski closed sets we call
X constructible. If K is algebraically closed, the constructible sets have much
stronger closure properties.

Corollary 2.4 Let K be an algebraically closed field.

i) X ⊆ K n is constructible if and only if it is definable.
ii) (Chevalley’s Theorem) The image of a constructible set under a poly-
nomial map is constructible.

Proof i) By Lemma 2.3, the constructible sets are exactly the quantifier-free
definable sets, but by quantifier elimination every definable set is quantifier-free
definable.

7
 ii) Let X ⊆ K n be constructible and p : K n → K m be a polynomial map.
Then, the image of X = {y ∈ K m : ∃x ∈ K n p(x) = y}. This set is definable
and hence constructible.
Quantifier elimination has very strong consequences for definable subsets of
K.

Corollary 2.5 If K is an algebraically closed field and X ⊆ K is definable,

then either X or K \ X is finite.

Proof By quantifier elimination X is a finite Boolean combination of sets of

the form V (p), where p ∈ K[X]. But V (p) is either finite or (if p = 0) all of K.
We say that a theory T is strongly minimal if for any M |= T and any
definable X ⊆ M either X or M \X is finite. This is a very powerful assumption.
For example, it can be shown that any strongly minimal theory in a countable
language is κ-categorical for every uncountable κ.
The model-completeness of algebraically closed fields can be used to give a
proof of the Nullstellensatz.

Theorem 2.6 (Hilbert’s Nullstellensatz) Let K be an algebraically closed

field. Suppose that I and J are radical ideals in K[X1 , . . . , Xn ] and I ⊂ J. Then
V (J) ⊂ V (I). Thus X 7→ I(X) is a bijective correspondence between Zariski
closed sets and radical ideals.

Proof Let p ∈ J \ I. By Primary Decomposition, there is a prime ideal P ⊇ I

such that p 6∈ P . We will show that there is x ∈ V (P ) ⊆ V (I) such that
p(x) 6= 0. Thus V (I) 6= V (J). Because P is prime, K[X]/P is a domain and we
can take F , the algebraic closure of its fraction field.
Let q1 , . . . , qm ∈ K[X1 , . . . , Xn ] generate I. Let ai be the element Xi /P in
F . Because each qi ∈ P and p 6∈ P ,
m
^
F |= qi (a) = 0 ∧ p(a) 6= 0.
i=1

Thus
m
^
F |= ∃w qi (w) = 0 ∧ p(w) 6= 0
i=1

and by model-completeness
m
^
K |= ∃w qi (w) = 0 ∧ p(w) 6= 0.
i=1

Thus there is b ∈ K n such that q1 (b) = . . . = qm (b) = 0 and p(b) 6= 0. But then
b ∈ V (P ) \ V (J).
The next corollary is a simple consequence of model completeness.

8
Corollary 2.7 Suppose K ⊆ L are algebraically closed fields, V and W are
varieties defined over K and f : V → W is a polynomial isomorphism defined
over L. Then there is an isomorphism defined over K.
Proof Suppose f : V → W is a polynomial isomorphism defined over L and f
and f −1 both have degree at most d. As in the proof of Ax’s Theorem we can
write down an L-formula Ψ with parameters from K saying that for some choice
of coefficients there is a polynomial bijection from V between V and W where
the polynomials have degree at most d. Since L |= Ψ, by model completeness,
K |= Ψ. Thus we can choose an isomorphism defined over K.
Quantifier elimination gives us a powerful tool for analyzing definability in
algebraically closed fields. For example, we will give the following characteriza-
tion of definable functions.
Definition 2.8 Let X ⊆ K n . We say that f : X → K is quasirational if either
i) K has characteristic zero and for some rational function q(X) ∈ K(X1 , . . . , Xn ),
f (x) = q(x) on X, or
ii) K has characteristic p > 0 and for some rational function q(X) ∈ K(X),
1
f (x) = q(x) pn .
Rational functions are easily seen to be definable. In algebraically closed
1
fields of characteristic p, the formula x = y p defines the function x 7→ x p ,
because every element has a unique pth -root. Thus, every quasirational function
is definable.
Proposition 2.9 If X ⊆ K n is constructible and f : X → K is definable, then
there are constructible
S sets X1 , . . . , Xm and quasirational functions ρ1 , . . . , ρm
such that Xi = X and f |Xi = ρi |Xi .
Proof Let Γ(v1 , . . . , vn ) = {f (v) 6= ρ(v) : ρ a quasirational function} ∪{v ∈
X}∪ ACF ∪ Diag(K).
Claim Γ is not satisfiable.
Suppose that Γ is consistent. Let L |= ACF +Diag(K) with b1 , . . . , bn ∈ L
such that for all γ(v) ∈ Γ, L |= γ(b).
Let K0 be the subfield of L generated by K and b. Then, K0 is the closure
of B = {b1 , . . . , bn } under the rational functions of K. Let K1 be the closure of
B under all quasirational functions. If K has characteristic 0, then K0 = K1 .
S 1n
If K has characteristic p > 0, K1 = K0p , the perfect closure of K0 .
By model-completeness, K ≺ L, thus f L , the interpretation of f in L, is
a function from X L to L, extending f . Because L |= Γ(b), f (b) is not in K1 .
Because K1 is perfect there is an automorphism α of L fixing K1 pointwise such
that α(f L (b)) 6= f L (b). But f L is definable with parameters from K; thus, any
automorphism of L which fixes K and fixes a must fix f (a), a contradiction.
Thus Γ is unsatisfiable.
Thus, by compactness, there are quasirational functions ρ1 , . . . , ρm such that
^
K |= ∀x ∈ X f (x) = ρi (x).

9
Let Xi = {x ∈ X : f (x) = ρi (x)}. Each Xi is definable.
We end by stating two more far reaching definability results for algebraically
closed. They are a bit more involved–and ideally best understood using the
model theoretic tool of ω-stability that we will not discuss in these lectures.
Let K be algebraically closed.

Theorem 2.10 (Elimination of Imaginaries) Suppose X ⊆ K n is definable

and E is a deifnable equivalence relation on X. There is a definable f : X → K m
for some m such that xEy if and only if f (x) = f (y).

This is related to the existence of fields of definitions. It is a useful tool for

viewing projective, quasiprojective or abstract varieties (at least in the style of
Weil) as constructible objects.

Theorem 2.11 Let G ⊆ K n be a definable group. Then G is definably isomor-

phic to an algebraic group.

Combining these we could conclude that if G is an algebraic group and H is

a normal algebraic subgroup, then G/H is an algebraic group.

10
3 Real Closed Fields and o-minimality
In this section, we will concentrate on the field of real numbers. Unlike alge-
braically closed fields, the theory of the real numbers does not have quantifier
elimination in Lr = {+, 1, ·, 0, 1}, the language of rings. The proof of Corollary
2.5 shows that any field with quantifier elimination is strongly minimal, whereas
in R, if φ(x) is the formula ∃z z 2 = x, then φ defines an infinite coinfinite de-
finable set. In fact, algebraically closed fields are the only infinite fields with
quantifier elimination.
In fact, the ordering is the only obstruction to quantifier elimination. We will
eventually analyze the real numbers in the language Lor = {+, −, · · · , <, 0, 1}
and show that we have quantifier elimination in this language. Because the
ordering x < y is definable in the real field by the formula

∃ z (z 6= 0 ∧ x + z 2 = y),

any subset of Rn definable using an Lor -formula is already definable using an

Lr -formula). We will see that quantifier elimination in Lor leads us to a good
geometric understanding of the definable sets.
We begin by reviewing some of the necessary algebraic background on or-
dered fields. All of the algebraic results stated in this chapter are due to Artin
and Schreier. These results are all proved in the appendix
Definition 3.1 We say that a field F is orderable if there is a linear order <
of F making (F, <) an ordered field.
Although there are unique orderings of the fields R and Q, orderable fields
may have many possible orderings. The field of rational functions Q(X) has 2ℵ0
distinct orderings. To see this, let x be any real number transcendental over
Q. The evaluation map f (X) 7→ f (x) is a field isomorphism between Q(X) and
Q(x), the subfield of R generated by x. We can lift the ordering of the reals to
an ordering Q(X) by f (X) < g(X) if and only if f (x) < g(x). Because X < q if
and only if x < q, choosing a different transcendental real would yield a different
ordering. These are not the only orderings. We can also order Q(X) by making
X infinite or infinitesimally close to a rational.
There is a purely algebraic characterization of the orderable fields.
Definition 3.2 We say that F is formally real if −1 is not a sum of squares.
In any ordered field all squares are nonnegative. Thus, every orderable field
is formally real. The following result shows that the converse is also true.

Theorem 3.3 If F is a formally real field, then F is orderable. Indeed, if a ∈ F

and −a is not a sum of squares of elements of F , then there is an ordering of
F where a is positive.

Because the field of complex numbers is the only proper algebraic exten-
sion of the real field, the real numbers have no proper formally real algebraic
extensions. Fields with this property will play a key role.

11
Definition 3.4 A field F is real closed if it is formally real with no proper
formally real algebraic extensions.
Although it is not obvious at first that real closed fields form an elementary
class, the next theorem allows us to axiomatize the real closed fields.

Theorem 3.5 Let F be a formally real field. The following are equivalent.
i) F is real closed.
ii) F (i) is algebraically closed (where i2 = −1).
iii) For any a ∈ F , either a or −a is a square and every polynomial of odd
degree has a root.

Corollary 3.6 The class of real closed fields is an elementary class of Lr -

structures.

Proof We can axiomatize real closed fields by:

i) axioms for fields
ii) for each n ≥ 1, the axiom

∀x1 . . . ∀xn x21 + . . . + x2n + 1 6= 0

iii) ∀x∃y (y 2 = x ∨ y 2 + x = 0)
iv) for each n ≥ 0, the axiom
2n
X
∀x0 . . . ∀x2n ∃y y 2n+1 + xi y i = 0.
i=0

Although we can axiomatize real closed fields in the language of rings, we

already noticed that we do not have quantifier elimination in this language.
Instead, we will study real closed fields in Lor , the language of ordered rings. If
F is a real closed field and 0 6= a ∈ F , then exactly one of a and −a is a square.
This allows us to order F by

x < y if and only if y − x is a nonzero square.

It is easy to check that this is an ordering and it is the only possible ordering
of F .
Definition 3.7 We let RCF be the Lor -theory axiomatized by the axioms above
for real closed fields and the axioms for ordered fields.
The models of RCF are exactly real closed fields with their canonical order-
ing. Because the ordering is defined by the Lr -formula

∃z (z 6= 0 ∧ x + z 2 = y),

the next result tells us that using the ordering does not change the definable
sets.

12
Proposition 3.8 If F is a real closed field and X ⊆ F n is definable by an
Lor -formula, then X is definable by an Lr -formula.

Proof Replace all instances of ti < tj by ∃v (v 6= 0 ∧ v 2 + ti = tj ), where ti

and tj are terms occurring in the definition of X.
The next result suggests another possible axiomatization of RCF.

Theorem 3.9 An ordered field F is real closed if and only if whenever p(X) ∈
F [X], a, b ∈ X, a < b, and p(a)p(b) < 0, there is c ∈ F such that a < c < b and
p(c) = 0.

Definition 3.10 If F is a formally real field, a real closure of F is a real closed

algebraic extension of F .
By Zorn’s Lemma, every formally real field F has a maximal formally real
algebraic extension. This maximal extension is a real closure of F .
The real
√ closure of a formally
√ real field may not be unique. Let F = Q(X),
F0 = F ( X), and F1 = F ( −X). By Theorem 3.3, F0 and F1 are formally
real. Let Ri be a real closure of Fi . There is no isomorphism between R0 and
R1 fixing F because X is a square in R0 but not in R1 . Thus, some work needs
to be done to show that any ordered field (F, <) has a real closure where the
canonical order extends the ordering of F .

Lemma 3.11 If (F, <) is an ordered field, 0 <√x ∈ F , and x is not a square in
F , then we can extend the ordering of F to F ( x).
√ √
Proof We can extend the ordering to F ( x) by 0 < a + b x if and only if
i) b = 0 and a > 0, or
2
ii) b > 0 and (a > 0 or x > ab2 ), or
2
iii) b < 0 and (a < 0 and x < ab2 ).

Corollary 3.12 If (F, <) is an ordered field, there is a real closure R of F such
that the canonical ordering of R extends the ordering on F .

Proof
By successive applications of Lemma 3.11, we can find an ordered field (L, <)
extending (F, <) such that every positive element of F has a square root in L.
We now apply Zorn’s Lemma to find a maximal formally real algebraic extension
R of L. Because every positive element of F is a square in R, the canonical
ordering of R extends the ordering of F .

Although a formally real field may have nonisomorphic real closures, if (F, <
) is an ordered field there will be a unique real closure compatible with the
ordering of F .

13
Theorem 3.13 If (F, <) is an ordered field, and R1 and R2 are real closures
of F where the canonical ordering extends the ordering of F , then there is a
unique field isomorphism φ : R1 → R2 that is the identity on F .

Note that because the ordering of a real closed field is definable in Lr , φ also
preserves the ordering. We often say that any ordered field (F, <) has a unique
real closure. By this we mean that there is a unique real closure that extends
the given ordering.

Quantifier Elimination for Real Closed Fields

We are now ready to prove quantifier elimination.

Theorem 3.14 The theory RCF admits elimination of quantifiers in Lor .

Proof We use the quantifier elimination tests of §5. Suppose K and L are real
closed ordered fields and A is a common substructure. Then A is an ordered
integral domain. We extend the ordering on A to its fraction field to obtain an
ordered subfield F0 ⊆ K ∩ L. Let F be the real closure of F0 . By uniqueness of
real closures, F is isomorphic, as an ordered field, to the algebraic closure of F0
inside K and L. Without loss of generality we may assume F ⊆ K ∩ L.
It suffices then to show that if φ(v, w) is a quantifier-free formula, a ∈ F ,
b ∈ K and K |= φ(b, a), then there is b0 ∈ F such that F |= φ(b0 , a).
Note that
p(X) 6= 0 ↔ (p(X) > 0 ∨ −p(X) > 0)
and
p(X) 6> 0 ↔ (p(X) = 0 ∨ −p(X) > 0).
With this in mind, we may assume that φ is a disjunction of conjunctions of
formulas of the form p(v, w) = 0 or p(v, w) > 0. As in Theorem 2.1, we may
assume that there are polynomials p1 , . . . , pn and q1 , . . . , qm ∈ F [X] such that
n
^ m
^
φ(v, a) ↔ pi (v) = 0 ∧ qi (v) > 0.
i=1 i=1

If any of the polynomials pi (X) is nonzero, then b is algebraic over F . Be-

cause F has no proper formally real algebraic extensions, in this case b ∈ F .
Thus, we may assume that
m
^
φ(v, a) ↔ qi (v) > 0.
i=1

The polynomial qi (X) can only change signs at zeros of qi and all zeros of qi
are in F . Thus, we can find ci , di ∈ F such that ci < b < di and qi (x) > 0 for
all xV∈ (ci , di ). Let c = max(c1 , . . . , cm ) and d = min(d1 , . . . , dm ). Then, c < d
m
and i=1 qi (x) > 0 whenever c < x < d. Thus, we can find b0 ∈ F such that
F |= φ(b0 , a).

14
Corollary 3.15 RCF is complete, model complete and decidable. Thus RCF is
the theory of (R, +, ·, <) and RCF is decidable.

Proof By quantifier elimination, RCF is model complete.

Every real closed field has characteristic zero; thus, the rational numbers are
embedded in every real closed field. Therefore, Ralg , the field of real algebraic
numbers (i.e., the real closure of the rational numbers) is a subfield of any real
closed field. Thus, for any real closed field R, Ralg ≺ R, so R ≡ Ralg .
In particular, R ≡ Ralg ≡ R.
Because RCF is complete and recursively axiomatized, it is decidable.

Semialgebraic Sets
Quantifier elimination for real closed fields has a geometric interpretation.
Definition 3.16 Let F be an ordered field. We say that X ⊆ F n is semial-
gebraic if it is a Boolean combination of sets of the form {x : p(x) > 0}, where
p(X) ∈ F [X1 , . . . , Xn ].
By quantifier elimination, the semialgebraic sets are exactly the definable
sets. The next corollary is a geometric restatement of quantifier elimination. It
is analogous to Chevalley’s Theorem (2.4) for algebraically closed fields.

Corollary 3.17 (Tarski–Seidenberg Theorem) The semialgebraic sets are

closed under projection.

The next corollary is a typical application of quantifier elimination.

Corollary 3.18 If F |= RCF and A ⊆ F n is semialgebraic, then the closure

(in the Euclidean topology) of A is semialgebraic.

Proof We repeat the main idea of Lemma ??. Let d be the definable function
n
X
d(x1 , . . . , xn , y1 , . . . , yn ) = z if and only if z ≥ 0 ∧ z 2 = (xi − yi )2 .
i=1

The closure of A is
{x : ∀ > 0 ∃y ∈ A d(x, y) < }.
Because this set is definable, it is semialgebraic.
We say that a function is semialgebraic if its graph is semialgebraic. The
next result shows how we can use the completeness of RCF to transfer results
from R to other real closed fields.
Corollary 3.19 Let F be a real closed field. If X ⊆ F n is semialgebraic, closed
and bounded, and f is a continuous semialgebraic function, then f (X) is closed
and bounded.

15
Proof If F = R, then X is closed and bounded if and only if X is compact.
Because the continuous image of a compact set is compact, the continuous image
of a closed and bounded set is closed and bounded.
In general, there are a, b ∈ F and formulas φ and ψ such that φ(x, a) defines
X and ψ(x, y, b) defines f (x) = y. There is a sentence Φ asserting:
∀u, w [if ψ(x, y, w) defines a continuous function with domain φ(x, u)
and φ(x, u) is a closed and bounded set, then the range of the func-
tion is closed and bounded].
By the remarks above, R |= Φ. Therefore, by the completeness of RCF,
F |= Φ and the range of f is closed and bounded.
Model-completeness has several important applications. A typical applica-
tion is Abraham Robinson’s simple proof of Artin’s positive solution to Hilbert’s
17th problem.
Definition 3.20 Let F be a real closed field and f (X) ∈ F (X1 , . . . , Xn ) be
a rational function. We say that f is positive semidefinite if f (a) ≥ 0 for all
a ∈ F n.

Theorem 3.21 (Hilbert’s 17th Problem) If f is a positive semidefinite ra-

tional function over a real closed field F , then f is a sum of squares of rational
functions.

Proof Suppose that f (X1 , . . . , Xn ) is a positive semidefinite rational function

over F that is not a sum of squares. By Theorem 3.3, there is an ordering of
F (X) so that f is negative. Let R be the real closure of F (X) extending this
order. Then
R |= ∃v f (v) < 0
because f (X) < 0 in R. By model-completeness

F |= ∃v f (v) < 0,

contradicting the fact that f is positive semidefinite.

We will show that quantifier elimination gives us a powerful tool for under-
standing the definable subsets of a real closed field.
Definition 3.22 Let L ⊇ {<}. Let T be an L-theory extending the theory
of linear orders. We say that T is o-minimal if for all M |= T if X ⊆ M is
definable, then X is a finite union of points and intervals with endpoints in
M ∪ {±∞}.
We can think of o-minimality as an analog of strong minimality for ordered
structures. Strong minimality says that the only definable subsets in dimension
one can be defined using only equality–i.e., the ones that can be defined in any
structure. O-minimality says the only sets that can be defined in one dimension
are the ones definable in any ordered structure.

Corollary 3.23 RCFis an o-minimal theory.

16
Proof Let R |= RCF. We need to show that every definable subset of R is a
finite union of points and intervals with endpoints in R ∪ {±∞}. By quantifier
elimination, very definable subset of R is a finite Boolean combination of sets
of the form
{x ∈ R : p(x) = 0}
and
{x ∈ R : q(x) > 0}.
Solution sets to nontrivial equations are finite, whereas sets of the second form
are finite unions of intervals. Thus, any definable set is a finite union of points
and intervals.
Next we will show that definable functions in one variable are piecewise
continuous. The first step is to prove a lemma about R that we will transfer to
all real closed fields.

Lemma 3.24 If f : R → R is semialgebraic, then for any open interval U ⊆ R

there is a point x ∈ U such that f is continuous at x.

Proof
case 1: There is an open set V ⊆ U such that f has finite range on V .
Pick an element b in the range of f such that {x ∈ V : f (x) = b} is infinite.
By o-minimality, there is an open set V0 ⊆ V such that f is constantly b on V .
case 2: Otherwise.
We build a chain U = V0 ⊃ V1 ⊃ V2 . . . of open subsets of U such that the
closure V n+1 of Vn+1 is contained in Vn . Given Vn , let X be the range of f
on Vn . Because X is infinite, by o-minimality, X contains an interval (a, b) of
length at most n1 . The set Y = {x ∈ Vn : f (x) ∈ (a, b)} contains a suitable open
interval Vn+1 . Because R is locally compact,
∞
\ ∞
\
Vi = Vi 6= ∅.
i=1 i=1
T∞
If x ∈ i=1 Vi , then f is continuous at x.

The proof above makes essential use of the completeness of the ordering of
the reals. However, because the statement is first order, it is true for all real
closed fields, by the completeness of RCF.

Corollary 3.25 Let F be a real closed field and f : F → F is a semialgebraic

function. Then, we can partition F into I1 ∪ . . . ∪ Im ∪ X, where X is finite
and the Ij are pairwise disjoint open intervals with endpoints in F ∪ {±∞} such
that f is continuous on each Ij .

Proof Let

D = {x : F |= ∃ > 0 ∀δ > 0 ∃y |x − y| < δ ∧ |f (x) − f (y)| > }

17
be the set of points where f is discontinuous. Because D is definable, by o-
minimality D is either finite or has a nonempty interior. By Corollary 3.23,
D must be finite. Thus, F \ D is a finite union of intervals on which F is
continuous.
If F is real closed, then o-minimality tells us what the definable subsets of
F look like. Definable subsets of F n are also relatively simple.
Definition 3.26 We inductively define the collection of cells as follows.
• X ⊆ F n is a 0-cell if it is a single point.
• X ⊆ F is a 1-cell if it is an interval (a, b), where a ∈ F ∪ {−∞}, b ∈
F ∪ {+∞}, and a < b.
• If X ⊆ F n is an n-cell and f : X → F is a continuous definable function,
then Y = {(x, f (x)) : x ∈ X} is an n-cell.
• Let X ⊆ F n be an n-cell. Suppose that f is either a continuous definable
function from X to F or identically −∞ and g is either a continuous definable
function from X to F such that f (x) < g(x) for all x ∈ X or g is identically
+∞; then
Y = {(x, y) : x ∈ X ∧ f (x) < y < g(x)}
is an n + 1-cell.
In a real closed field, every nonempty definable set is a finite disjoint union
of cells. The proof relies on the following lemma.
Lemma 3.27 (Uniform Bounding) Let X ⊆ F n+1 be semialgebraic. There
is a natural number N such that if a ∈ F n and Xa = {y : (a, y) ∈ X} is finite,
then |Xa | < N .
Proof First, note that Xa is infinite if and only if there is an interval (c, d)
such that (c, d) ⊆ Xa . Thus {(a, b) ∈ X : Xa is finite} is definable. Without
loss of generality, we may assume that for all a ∈ F n , Xa is finite. In particular,
we may assume that
F |= ∀x∀c∀d¬[c < d ∧ ∀y(c < y < d → y ∈ Xa )].
Consider the following set of sentences in the language of fields with constants
added for each element of F and new constants c1 , . . . , cn . Let Γ be
   
 ^ ^m 
RCF + Diag(F ) + ∃y1 , . . . , ym  yi 6= yj ∧ yi ∈ Xc  : m ∈ ω
 
i<j i=1

Suppose that Γ is satisfiable. Then, there is a real closed field K ⊇ F and

elements c ∈ K n such that Xc is infinite. By model-completeness, F ≺ K.
Therefore
K |= ∀x∀c, d ¬[c < d ∧ ∀y (c < y < d → y ∈ Xa )].
This contradicts the o-minimality of K. Thus, Γ is unsatisfiable and there is an
N such that
  
^ ^N
RCF + Diag(F ) |= ∀x ¬ ∃y1 , . . . , yN  yi 6= yj ∧ yi ∈ Xx  .
i<j i=1

18
 In particular, for all a ∈ F n , |Xa | < N .
We now state the Cell Decomposition Theorem and give the proof for subsets
of F 2 . In the exercises, we will outline the results needed for the general case.

Theorem 3.28 (Cell Decomposition) Let X ⊆ F m be semialgebraic. There

are finitely many pairwise disjoint cells C1 , . . . , Cn such that X = C1 ∪ . . . ∪ Cn .

Proof (for m = 2) For each a ∈ F , let

Ca = {x : ∀ > 0∃y, z ∈ (x − , x + ) [(a, y) ∈ X ∧ (a, z) 6∈ X]}.

We call Ca the critical values above a. By o-minimality, there are only finitely
many critical values above a. By uniform bounding, there is a natural number
N such that for all a ∈ F , |Ca | ≤ N . We partition F into A0 , A1 , . . . , AN , where
An = {a : |Ca | = n}.
For each n ≤ N , we have a definable function fn : A1 ∪ . . . ∪ An → F by
fn (a) = nth element of Ca . As above, Xa = {y : (a, y) ∈ X}.
For n ≤ N and a ∈ An , we define Pa ∈ 22n+1 , the pattern of X above a, as
follows.
If n = 0, then Pa (0) = 1 if and only if Xa = F . Suppose that n > 0.
Pa (0) = 1 if and only if x ∈ Xa for all x < f1 (a).
Pa (2i − 1) = 1 if and only if fi (a) ∈ X.
For i < n, Pa (2i) = 1 if and only if x ∈ Xa for all x ∈ (fi (a), fi+1 (a)).
P (2n) = 1 if and only if x ∈ Xa for all x > fn (a).
For each possible pattern σ ∈ 22n+1 , let An,σ = {a ∈ An : Pa = σ}.
Each An,σ is semialgebraic. For each An,σ , we will give a decomposition of
{(x, y) ∈ X : x ∈ An,σ } into disjoint cells. Because the An,σ partition F , this
will suffice.
Fix one An,σ . By Corollary 3.25, we can partition An,σ = C1 ∪ . . . ∪ Cl ,
where each Cj is either an interval or a singleton and fi is continuous on Cj for
i ≤ n, j ≤ l. We can now give a decomposition of {(x, y) : x ∈ An,σ } into cells
such that each cell is either contained in X or disjoint from X.
For j≤ l, let Dj,0 = {(x, y) : x ∈ Cj and y < f (1)}.
For j≤ l and 1 ≤ i ≤ n, let Dj,2i−1 = {(x, fi (x)) : x ∈ Cj }.
For j≤ l and 1 ≤ i < n, let Dj,2i = {(x, y) : x ∈ Cj , fi (x) < y < fi+1 (x)}.
For j≤ l, let Dj,2n = {(x, y) : x ∈ Cj , y > fn (x)}.
S
Clearly, each Dj,i is a cell, Dj,i = {(x, y) : x ∈ An,σ }, and each Dj,i
is either contained in X or disjoint from X. Thus, taking the Dj,i that are
contained in X, we get a partition of {(x, y) ∈ X : x ∈ An,σ } into disjoint cells.

o-minimal Expansions of R
The proofs above used very little about semialgebraic sets beyond o-minimality.
Indeed, they would work in any o-minimal expansion of the real field. Indeed,
there is a rich theory of definable sets in o-minimal expansions of the reals. We

19
will survey some of the results in this section. For full details, see van den Dries
book Tame topology and o-minimal structures.
Let R = (R, +, ·, <, . . .) be an o-minimal expansion of the reals, i.e., a struc-
ture obtained by adding extra structure to the reals such that Th(R) is o-
minimal. Below by “definable” we will mean definable in R.
Theorem 3.29 Assume R is an o-minima expansion of R.
i) Every definable subset of Rn is a finite union of cells.
ii) If f : X → Rn is definable, there is a finite partition of X into cells
X1 , ∪, Xn such that f |Xi is continuous for each i. Indeed, for any r ≥ 0, we
can choose the partition such that f |Xi is C r for each i.
An easy consequence of ii) is that definable sets have only finitely many
connected components. Much more is true, for example:
• Definable bounded sets can be definably triangulated.
• Suppose X ⊆ Rn+m is definable. For a ∈ Rm let
Xa = {x = (x1 , . . . , xn ) ∈ Rn : (x, a) ∈ X}.
There are only finitely many definable homeomorphism types for the sets Xa .
• (Curve selection) If X ⊆ Rn is definable and a is in the closure of X, then
there is a continuous definable f : (0, 1) → X such that
lim f (x) = a.
x→1

• If G is a definable group, then G is definably isomorphic to a Lie group.

• If we assume in addition that all definable functions are majorized by poly-
nomials, then many of the metric properties of semialgebraic sets and asymptotic
properties of semialgebraic functions also generalize.
Of course, this leads to the question: are there interesting o-minimal expan-
sions of R?

Ran and subanalytic sets

Most of the results on o-minimal structures mentioned above were proved before
we knew of any interesting o-minimal structures other than the real field. The
first new example of an o-minimal theory was given by van den Dries.
Let Lan = L ∪ {fb : for some open U ⊃ [0, 1]n , f : U → R is analytic}.
We define fb : Rn → R by
n
fb(x) = f (x) x ∈ [0, 1]
n
0 otherwise.
We let Ran be the resulting Lan -structure. Denef and van den Dries proved that
Ran is o-minimal and that Ran has quantifier elimination if we add a function

D(x, y) = x/y if 0 ≤ |x| ≤ |y|

n
0 otherwise

20
to the language. Quantifier elimination is proven by using the Weierstrass prepa-
ration theorem to replace arbitrary analytic functions of several variables by an-
alytic functions that are polynomial in one of the variables. Tarski’s elimination
procedure is then used to eliminate this variable.
Denef and van den Dries also showed that if f : R → R is definable in
Ran , then f is asymptotic to a rational function. In particular, although we
can define the restriction of the exponential function to bounded intervals, we
cannot define the exponential function globally. It is also impossible to define
the sine function globally; for its zero set would violate o-minimality.
Although Ran may seem unnatural, the definable sets form an interesting
class.
We say that X ⊆ Rn is semi-analytic if for all x in Rn there is an open
neighborhood U of x such that X ∩ U is a finite Boolean combination of sets
{x ∈ U : f (x) = 0} and {x ∈ U : g(x) > 0} where f, g : U → R are analytic.
We say that X ⊆ Rn is subanalytic if for all x in Rn there is an open U and
Y ⊂ Rn+m a bounded semianalytic set such that X ∩ U is the projection of Y
into U . It is well known that subanalytic sets share many of the nice properties
of semialgebraic sets.
If X ⊂ Rn is bounded, then X is definable in Ran if and only if X is sub-
analytic. Indeed Y ⊆ Rn is definable in Ran if and only if it is the image of a
bounded subanalytic set under a semialgebraic map. Most of the known prop-
erties of subanalytic sets generalize to sets defined in any polynomial bounded
o-minimal theory.

Exponentiation
The big breakthrough in the subject came in 1991. While quantifier elimination
for Rexp is impossible, Wilkie proved the next best thing.

Theorem 3.30 (Wilkie) Let φ(x1 , . . . , xm ) be an Lexp formula. Then there

is n ≥ m and f1 , . . . , fs ∈ Z[x1 , . . . , xn , ex1 , . . . , exn ] such that φ(x1 , . . . , xn ) is
equivalent to

∃xm+1 . . . ∃xn f1 (x1 , . . . , xn , ex1 , . . . , exn ) = . . . = fs (x1 , . . . , xn , ex1 , . . . , exn ) = 0.

Thus every formula is equivalent to an existential formula (this property is

equivalent to model completeness) and every definable set is the projection of
an exponential variety.
Wilkie’s proof depends heavily on the following special case of a theorem of
Khovanski. Before Wilkie’s theorem, Khovanski’s result was the best evidence
that Rexp is o-minimal; indeed Khovanski’s theorem is also the crucial tool
needed to deduce o-minimality from model completeness.

Theorem 3.31 (Khovanski) If f1 , . . . , fm : Rn → R are exponential poly-

nomials, then {x ∈ Rn : f1 (x) = . . . fn (x) = 0} has finitely many connected
components.

21
 If X ⊆ R is definable in Rexp then by Wilkie’s Theorem there is an expo-
nential variety V ⊆ Rn such that X is the projection of V . By Khovanski’s
Theorem V has finitely many connected components and X is a finite union of
points and intervals. Thus Rexp is o-minimal.
Using the o-minimality of Rexp one can improve some of Khovanski’s results
on “fewnomials”. From algebraic geometry we know that we can bound the
number of connected components of a hypersurface in Rn uniformly in the degree
of the defining polynomial. Khovanski showed that it is also possible to bound
the number of connected component uniformly in the number of monomials in
the defining polynomial. We will sketch the simplest case of this. Let Fn,m be
the collection of polynomials in R[X1 , . . . , Xn ] with at most m monomials. For
p ∈ Fn,m let
n
^
+ n
V (p) = {x = (x1 , . . . , xn ) ∈ R : xi ≥ 0 ∧ p(x) = 0}.
i=1

We claim that there are only finitely many homeomorphism types of V + (p) for
p ∈ Fn,m . Let Φm,n (x1 , . . . , xn , r1,1 . . . , r1,n , . . . , rm,1 , . . . , rm,n , a1 , . . . , am ) be
the formula
^ m
X n
Y
∃w1 . . . , wm (( ew i = xi ) ∧ ai ewi ri,j = 0).
i=1m i=1 j=1

We see that Φ expresses

m n
r
X Y
ai xj i,j = 0.
i=1 j=1
n
Let Xr,a denote the set of x ∈ R such that Φ(x, r, a) holds. By o-minimality,
{Xr,a : r ∈ Rmn , a ∈ Rm } represents only finitely many homeomorphism types.
In addition to answering the question of o-minimality, some headway has
been made on the problem of decidability. Making heavy use of Wilkie’s methods
and Khovanski’s theorem, Macintyre and Wilkie have shown that if Schanuel’s
Conjecture in is true then the first order theory of Rexp is decidable. Where
Schanuel’s Conjecture is the assertion that if λ1 , . . . , λn are complex numbers
linearly independent over Q, then the transcendence degree of the field

Q(λ1 , . . . , λn , eλ1 , . . . , eλn )

is at least n.
Miller provided an interesting counterpoint to Wilkie’s theorem. Using ideas
of Rosenlicht he showed that if R is any o-minimal expansion of the real field that
contains a function that is not majorized by a polynomial, then exponentiation
is definable in R.
Let Lan,exp be Lan ∪ {ex } and let Ran,exp be the real numbers with both ex-
ponentiation and restricted analytic functions. Using the Denef-van den Dries

22
quantifier elimination for Ran and a mixture of model-theoretic and valuation
theoretic ideas, van den Dries, Macintyre, and I were able to show that Ran,exp
has quantifier elimination if we add log to the language. Using quantifier elim-
ination and Hardy field style arguments (but avoiding the geometric type of
arguments used by Khovanski) we were able to show that Ran,exp is o-minimal.
Since the language Lan,exp has size 2ℵ0 , one would not expect to give a sim-
ple axiomatization of the first order theory of Ran,exp . Ressayre noticed that
the model-theoretic analysis of Ran,exp uses very little global information about
exponentiation. This observation leads to a “relative” axiomatization. The the-
ory T h(Ran,exp ) is axiomatized by the theory of Ran and axioms asserting that
exponentiation is an increasing homomorphism from the additive group onto
the multiplicative group of positive elements that majorizes every polynomial.
Using this axiomatization and quantifier elimination one can show that any
definable function is piecewise given by a composition of polynomials, exp, log,
and restricted analytic functions on [0, 1]n . For example, the definable function
x 2
f (x) = ee − ex − 3x is eventually increasing and unbounded. Thus for some
large enough r ∈ R there is a function g : (r, +∞) → R such that f (g(x)) = x for
y 2
x > r. The graph of g is the definable set {(x, y) : x > r and ee −ey −3y = x}.
Thus g is a definable function and there is some way to express g explicitly as
a composition of rational functions, exp, log, and restricted analytic functions.
In most cases it is in no way clear how to get these explicit representations of
an implicitly defined function. One important corollary is that every definable
function is majorized by an iterated exponential.

23
A Real Algebra
We prove some of the algebraic facts needed in Section 3. All of these results
are due to Artin and Schreier. See Lang’s Algebra §XI for more details.
All fields are assumed to be of characteristic 0.
Definition A.1 A field K is real if −1
P can not be expressed as a sum of squares
of elements of K. In general, we let K 2 be the sums of squares from K.
If F is orderable, then F is real because squares are nonnegative with respect
to any ordering.

Lemma A.2 Suppose that F is real and a ∈ F \ {0}. Then, at most one of a
and −a is a sum of squares.

Proof If a and b are both sums of squares, then ab P = ba2 b is a sum of squares.
Thus, if F is real, at least one of a and −a is not in F 2.

P 2 √
Lemma A.3 If F is real √ and −a ∈ F \ F √ , then F ( a) is real. Thus, if F
is real and a ∈ F , then F ( a) is real or F ( −a) is real.
√ √
Proof We may assume that a 6∈ F . If F ( a) is not real, then there are
bi , ci ∈ F such that
X √ X √
−1 = (bi + ci a)2 = (b2i + 2ci bi a + c2i a).
√ √
Because a and 1 are a vector space basis for F ( a) over F ,
X X
−1 = b2i + a c2i .

Thus P 2
P 2
P 2

1 + b2i
P
bi ci + ci
−a = P 2 = P 2 2
ci ( ci )
F 2 , a contradiction.
P
and −a ∈

Lemma A.4 If F is real, f (X) ∈ F [X] is irreducible of odd degree n, and

f (α) = 0, then F (α) is real.

Proof We proceed by induction on n. If n = 1, this is clear. Suppose, for

purposes of contradiction, that n > 1 is odd, f (X) ∈ F [X] is irreducible of
degree n, f (α) = 0, and F (α)
P is not real. There are polynomials gi of degree at
most n − 1 such that −1 = gi (α)2 . Because F is real, some gi is nonconstant.
Because F (α) ∼= F [X]/(f ), there is a polynomial q(X) ∈ F [X] such that
X
1= gi2 (X) + q(X)f (X).

24
 P 2
The polynomial gi (X) has a positive even degree at most 2n − 2. Thus, q
has odd degree at most n − 2. Let β P be the root of an irreducible factor of q.
By induction, F (β) is real, but −1 = gi2 (β), a contradiction.
Definition A.5 We say that a field R is real closed if and only if R is real and
has no proper real algebraic extensions.

If R is real closed and a ∈ R, then, by Lemmas A.2 and A.3, either a ∈ R2

or −a ∈ R2 . Thus, we can define an order on R by

a ≥ 0 ⇔ a ∈ R2 .

Moreover, this is the only way to define an order on R because the squares must
be nonnegative. Also, if R is real closed, every polynomial of odd degree has a
root in R.

Lemma A.6 Let F be a real field. There is R ⊇ F a real closed algebraic

extension. We call R a real closure of F .

Proof Let I = {K ⊇ F : K real, K/F algebraic}. The union of any chain of

real fields is real; thus, by Zorn’s Lemma, there is a maximal R ∈ I. Clearly, R
has no proper real algebraic extensions; thus, R is real closed.

Corollary A.7 If F is any real field, then F is orderable. Indeed, if a ∈ F and

−a 6∈ F 2 , then there is an ordering of F , where a > 0.
P

√
Proof By Lemma A.3, F ( a) is real. Let R be a real closure of F . We order
F by restricting the ordering of R because a is a square in R, a > 0.
The following theorem is a version of the Fundamental Theorem of Algebra.

Theorem A.8 Let R be a√real field √ such that

i) for all a ∈ R, either a or −a ∈ R and
if f (X) ∈ R[X] has odd degree, then f has a root in R.
ii) √
If i = −1, then K = R(i) is algebraically closed.

Proof
Claim 1 Every element of K has √a square root in K.
2 2
Let a + bi ∈ K. Note that a+ a2 +b is nonnegative for any ordering of R.
Thus, by i), there is c ∈ R with
√
2 a + a2 + b2
c = .
2
b
If d = 2c , then (c + di)2 = a + bi.
Let L ⊇ K be a finite Galois extension of R. We must show that L = K.
Let G = Gal(L/R) be the Galois group of L/R. Let H be the 2-Sylow subgroup
of G.

25
Claim 2 G = H.
Let F be the fixed field of H. Then F/R must have odd degree. If F = R(x),
then the minimal polynomial of x over R has odd degree, but the only irreducible
polynomials of odd degree are linear. Thus, F = R and G = H.
Let G1 = Gal(L/K). If G1 is nontrivial, then there is G2 a subgroup of G1
of index 2. Let F be the fixed field of G2 . Then, F/K has degree 2. But by
Claim 1, K has no extensions of degree 2. Thus, G1 is trivial and L = K.
Corollary A.9 Suppose that R is real. Then R is real closed if and only if R(i)
is algebraically closed.
Proof
(⇒) By Theorem A.8.
(⇐) R(i) is the only algebraic extension of R, and it is not real.
Let (R, <) be an ordered field. We say that R has the intermediate value
property if for any polynomial p(X) ∈ R[X] if a < b and p(a) < 0 < p(b), then
there is c ∈ (a, b) with p(c) = 0.

Lemma A.10 If (R, <) is an ordered field with the intermediate value property,
then R is real closed.

Proof Let a > 0 and let p(X) = X 2 − a. Then p(0) < 0, and p(1 + a) > 0;
thus, there is c ∈ R with c2 = a.
Let
n−1
X
f (X) = X n + ai X i
i=0

where n is odd. For M large enough, f (M ) > 0 and f (−M ) < 0; thus, there is
a c such that f (c) = 0.
By Theorem A.8, R(i) is algebraically closed. Because R is real, it must be
real closed.

Lemma A.11 Suppose that R is real closed and < is the unique ordering, then
(R, <) has the intermediate value property.

Proof Suppose f (X) ∈ R[X], a < b, and f (a) < 0 < f (b). We may assume
that f (X) is irreducible (for some factor of f must change signs). Because R(i)
is algebraically closed, either f (X) is linear, and hence has a root in (a, b), or

f (X) = X 2 + cX + d,

where c2 − 4d < 0. But then

c2
 
 c 2
f (X) = X + + d−
2 4

and f (x) > 0 for all x.

We summarize as follows.

26
Theorem A.12 The following are equivalent.
i) R is real closed.
ii) For all a ∈ R, either a or −a has a square root in R and every polynomial
of odd degree has a root in R.
iii) We can order R by a ≥ 0 if and only if a is a square and, with respect
to this ordering, R has the intermediate value property.

Finally, we consider the question of uniqueness of real closures. We first

note that there are √
some subtleties. For example, there are nonisomorphic real
closures of F = Q(√2). The field√ of real algebraic numbers √ is one real closure
P 2
of F . Because a + b 2 7→ a√ − b 2 is an automorphism of F , 2 is not in F .
Thus,
√ by Corollary B.5, F ( −2) is real. Let R be a real closure of F containing
F ( −2). Then, R is not isomorphic to the real algebraic numbers over F .
This is an example of a more general phenomenon. It is proved by successive
applications of Lemmas A.2 and A.3.

Lemma A.13 If (F, <) is an ordered field, then there is a real closure of F in
which every positive element of F is a square.
√
Because Q( 2) has two distinct orderings, it has two nonisomorphic real
closures. The field Q(t) of rational functions over Q has 2ℵ0 orderings and
hence 2ℵ0 nonisomorphic real closures.
The next theorem shows that once we fix an ordering of F , there is a unique
real closure that induces the ordering.

Theorem A.14 Let (F, <) be an ordered field. Let R0 and R1 be real closures
of F such that (Ri , <) is an ordered field extension of (F, <). Then, R0 is
isomorphic to R1 over F and the isomorphism is unique.

The proof of Theorem A.14 uses Sturm’s algorithm.

Definition A.15 Let R be a real closed field. A Sturm sequence is a finite
sequence of polynomials f0 , . . . , fn such that:
i) f1 = f00 ;
ii) for all x and 0 ≤ i ≤ n − 1, it is not the case that fi (x) = fi+1 (x) = 0;
iii) for all x and 1 ≤ i ≤ n − 1, if fi (x) = 0, then fi−1 (x) and fi+1 (x) have
opposite signs;
iv) fn is a nonzero constant.
If f0 , . . . , fn is a Sturm sequence and x ∈ R, define v(x) to be the number
of sign changes in the sequence f0 (x), . . . , fn (x).
Suppose that f ∈ R[X] is nonconstant and does not have multiple roots.
We define a Sturm sequence as follows:
f0 = f ;
f1 = f 0 .
Given fi nonconstant, use the Euclidean algorithm to write

fi = gi fi−1 − fi+1

27
where the degree of fi+1 is less than the degree of fi−1 . We eventually reach a
constant function fn .

Lemma A.16 If f has no multiple roots, then f0 , . . . , fn is a Sturm sequence.

Proof
iv) If fn = 0, then fn−1 |fi for all i. But f has no multiple roots; thus f and
f 0 have no common factors, a contradiction.
ii) If fi (x) = fi+1 (x) = 0, then by induction fn (x) = 0, contradicting iv).
iii) If 1 ≤ i ≤ n − 1 and fi (x) = 0, then fi−1 (x) = −fi+1 (x). Thus, fi−1 (x)
and fi+1 (x) have opposite signs.

Theorem A.17 (Sturm’s Algorithm) Suppose that R is a real closed field,

a, b ∈ R, and a < b. Let f be a polynomial without multiple roots. Let f =
f0 , . . . , fn be a Sturm sequence such that fi (a) 6= 0 and fi (b) 6= 0 for all i.
Then, the number of roots of f in (a, b) is equal to v(a) − v(b).

Proof Let z1 < . . . < zm be all the roots of the polynomials f0 , . . . , fn that are
in the interval (a, b). Choose c1 , . . . , cm−1 with zi < ci < zi+1 . Let a = c0 and
b = cm . For 0 ≤ i ≤ P m − 1, let ri be the number of roots of f in the interval
(ci , ci+1 ). Clearly, ri is the number of roots of f in the interval (a, b). On
the other hand,
m−1
X
v(a) − v(b) = (v(ci ) − v(ci+1 )).
i=0

Thus, it suffices to show that if c < z < d and z is the only root of any fi in
(c, d), then 
v(c) − 1 z is a root of f
v(d) = .
v(c) otherwise
If fi (b) and fi (c) have different signs, then fi (z) = 0. We need only see what
happens at those places.
If z is a root of fi , i > 0, then fi+1 (z) and fi−1 (z) have opposite signs and
fi+1 and fi−1 do not change signs on [c, d]. Thus, the sequences fi−1 (c), fi (c), fi+1 (c)
and fi−1 (d), fi (d), fi+1 (d) each have one sign change. For example, if fi−1 (z) >
0 and fi−1 (z) < 0, then these sequences are either +, +, − or +, −, +, and in
either case both sequences have one sign change.
If z is a root of f0 , then, because f 0 (z) 6= 0, f is monotonic on (c, d). If f
is increasing on (c, d), the sequence at c starts −, +, . . . and the sequence at d
starts +, +, . . .. Similarly, if f is decreasing, the sequence at c starts +, −, . . .,
and the sequence at b starts −, −, . . .. In either case, the sequence at c has one
more sign change than the sequence at d. Thus, v(c) − v(d) = 1, as desired.

Corollary A.18 Suppose that (F, <) is an ordered field. Let f be a nonconstant
irreducible polynomial over F . If R0 and R1 are real closures of F compatible
with the ordering, then f has the same number of roots in both R0 and R1 .

28
Proof Let f0 , . . . , fn be the Sturm sequence from Lemma A.16. Note that
each fi ∈ F [X].
P We can find M ∈ F such that any root of fi is in (−M,P M ) (if
g(X) = X n + ai X i , then any root of g has absolute value at most 1 + |ai |,
for example). Then, the number of roots of f in Ri is equal to v(−M ) − v(M ),
but v(M ) depends only on F .

Lemma A.19 Suppose (F, <) is an ordered field and R0 and R1 are real clo-
sures of F such that (Ri , <) is an ordered field extension of (F, <). If α ∈ R0 \F ,
there is an ordered field embedding of F (α) into R1 fixing F .

Proof Let f ∈ F [X] be the minimal polynomial of α over F . Let α1 < . . . < αn
be all zeros of f in R0 . By Corollary B.18, f has exactly n zeros β1 < . . . <
βn ∈ R1 . Let
σ : F (α1 , . . . , αn ) → F (β1 , . . . , βn )
be the map obtained by sending αi to βi . We claim that σ is an ordered field
isomorphism.
√
For i = 1, . . . , n − 1, let γi = αi+1 − αi ∈ R0 . By the Primitive Element
Theorem, there is a ∈ F such that

F (a) = F (α1 , . . . , αn , γ1 , . . . , γn−1 ).

Let g ∈ F [X] be the minimal polynomial of a over F . By Corollary B.18, g has

a zero b ∈ R1 and there is a field isomorphism φ : F (a) → F (b). Because F (a)
contains n zeros of F , so does F (b). Thus β1 , . . . , βn ∈ F (b) and for each i there
is a j such that φ(αi ) = βj . But

φ(γi )2 = φ(αi+1 ) − φ(αi ).

Thus φ(αi ) = βi for i = 1, . . . , n. We still must show that σ is order preserving.

Suppose c ∈ F (α1 , . . . , αn ) and c > 0. There is d ∈ R0 such that d2 = c.
Arguing as above, we can find a field embedding

ψ : F (α1 , . . . , αn , d) ⊆ R1

fixing F . As above, ψ(αi ) = βi and ψ ⊇ σ. Because

ψ(d)2 = ψ(c) = σ(c),

we have σ(c) > 0. Thus σ is order preserving.

Proof of Theorem A.14 Let P be the set of all order preserving σ : K → R1
where F ⊆ K → R0 and σ|F is the identity. By Zorn’s Lemma, there is a
maximal σ : K → R1 in P. By identifying K and σ(K) and applying the
previous lemma, we see that K = R0 . A similar argument shows that σ(K) =
R1 .
Uniqueness follows because the ith root of f (X) in R0 must be sent to the
ith root of f (X) in R1 .

29