Tame Topology and O-Minimal Structures

Charles Steinhorn

Vassar College, Poughkeepsie, NY 12604 [email protected]

I first want to thank Professor Brown for his kind introduction. My goal is to
give everyone a sense of the subject of o-minimality. Ten hours is a lot of time
to lecture, but the literature on this topic has grown vast, so I will ignore a
lot of things in order to get at issues relevant for this audience. I confess that
this is the first time I have spoken to economists.
I want to start with an outline of the talks. The topic is tame topology
and o-minimal structures, but today I’m not going to tell you what either one
really is. Today I’m going to give you a sense of what’s important so that you
can understand things better in my subsequent lectures.

Outline of Lectures

1. An introduction to definability and quantifier elimination

2. The semialgebraic case
3. O-minimality and some basic properties
4. Examples and some further properties
5. VC dimension and applications

1 An Introduction to Definability
For which x do we have

ax + by > e1
∃y (∗)
cx + dy ≤ e2
where a, b, c, d, e1 , e2 ∈ R? We want to know

{x ∈ R | statement (1) holds for x} .

We also can replace the constants a, b, c, d, e1 , e2 by variables u, v, w, z, s1 ,

s2 and ask for which u, v, w, z, s1 , s2 ∈ R do we have
166 Charles Steinhorn


ux + vy > s1
∃x∃y . (∗∗)
wx + zy ≤ s2

That is, we want to know

{(u, v, w, z, s1 , s2 ) ∈ R6 | (u, v, w, z, s1 , s2 ) satisfies (1).

These are simple examples of definable sets. I want to give a couple more
examples before I start to get precise.

Examples

1. S = {x ∈ R | sin πx = 0}. Observe S = Z.

2. Let L = {(m, n) ∈ R2 | m, n ∈ Z} be the integer lattice in R2 . View L as
a two-place relation on pairs of real numbers:

L(m, n) ⇐⇒ m, n ∈ Z.

Define
) *
C = (x, y) ∈ R2 | ∃u∃v[L(u, v) ∧ (x − u)2 + (y − v)2 ≤ 1/16] .

The set C is the set of all closed discs of radius 1/4 whose centers are
points in L.
You can see that as you introduce quantifiers, you have to think a bit
about what you are defining.

1.1 Structures on the Real Numbers R

Throughout these lectures, I am going to limit my discussion to structures on

the real line, for concreteness. Fix a family of
1. F of basic functions f : Rk → R, for k = 1, 2, 3, ....
2. R of basic relations (subsets) R ⊆ Rk , for k = 1, 2, 3, ..., such that the
“less than” relation <⊂ R2 (and equality) always is included in R.
Refer to F and R together as a language L. The real numbers R together
with the functions and relations included in L is called an L-structure that in
general we denote by RL . Special structures will have special names.

Examples

1. F = {+, ·, −} where − : R → R is given by x −→ −x and R = {<}. This

is the language Lalg of the ordered field of real numbers, the structure
denoted by Ralg .
 Tame Topology and O-Minimal Structures 167

2. F contains + and, for each r ∈ R, the scalar multiplication function

μr : R → R given by x −→ rx, and R = {<}. This is the ordered
real vector space language Llin for the real numbers, whose corresponding
structure is denoted by Rlin .
3. F = {+, ·, −, exp} where exp : R → R is the exponential function x −→ ex
and R = {<}. This is the language Lexp of the ordered exponential field
of real numbers; the structure is denoted by Rexp .

1.2 Terms

Construct an expanded class of functions by repeated composition starting

from the functions in F . These are called terms.

Examples

1. The Lalg -terms are all integer coefficient polynomial functions p(x1 , ..., xk )
for k = 1, 2, 3, ...,
2. The Llin -terms are all R-linear functions L(x1 , ..., xk ) = μr1 x1 + · · · +
μrk xk .
The obvious kinds of rules apply in these examples, for instance μr1 (x) +
μr2 (x) = μr1 +r2 (x).

1.3 Formulas

If you think of the definable sets I wrote down at the beginning of the lecture,
they were defined in terms of equalities and inequalities. The basic formulas
are
1. t1 = t2 for terms t1 , t2 .
2. R(t1 , ..., tk ) where R is a k-place relation and t1 , ..., tk are terms. Note
that t1 < t2 is a special case.
So if you think of Ralg , we can write down polynomial equalities and in-
equalities, and for Rlin we can write down linear equalities and inequalities.
Next, from the basic formulas recursively construct
1. by Boolean operations from formulas ϕ and ψ :

ϕ ∧ ψ read as “ϕ and ψ,”

ϕ ∨ ψ read as “ϕ or ψ,”
¬ϕ read as “not ϕ.”

2. by existential quantification over R from a formula ϕ:

∃υϕ read as “there exists υ such that ϕ.”

168 Charles Steinhorn

Caution: to get to the point more quickly I have mixed syntax and semantics.
When you look at a formal text on model theory you will see a distinction
between symbols in a formal language and their interpretations.
Now what I’m going to do is go back to what I was vague about earlier:
L-definable sets.

L-definable sets

For our purposes these are subsets of Rk , for k = 1, 2, 3, ..., specified as follows.
For each L-formula ϕ, certain variables are bound to quantifiers and others
are not. Call the latter free variables, and it is for these that we can substitute
real numbers.
For an L-formula ϕ list its free variables as x1 , ..., xk , z1 , ..., zm . Choose
c1 , ..., cm ∈ R and substitute them for z1 , ..., zm , respectively. Write c for
(c1 , ..., cm ). The set

Dϕ,c = {(x1 , ..., xk ) ∈ Rk | ϕ(x1 , ..., xk , c̄) is true} ⊆ Rk

is an L-definable subset of Rk . If L is clear from context, we usually shall drop

the L.

Examples

Let’s go back to our earlier examples of definable sets. The set

{x ∈ R | statement (1) holds for x}

is Llin -definable. The set

{(u, v, w, z, s1 , s2 ) ∈ R6 | (u, v, w, z, s1 , s2 ) satisfies (1)}

is Lalg -definable. The set

{x ∈ R | sin πx = 0}

is (F , R)-definable, where F = {·, sin}, R = {<}. The set

{(x, y) ∈ R2 | ∃u∃v, [L(u, v) ∧ (x − u)2 + (y − v)2 ≤ 1/16]}

is (F , R)-definable, where F = {+, ·, −}, R = {L, <}.

Comments

• We call the c̄ parameters; the division of the free variables in a formula

into those which are parameter variables and those that are not can be
made arbitrarily.
 Tame Topology and O-Minimal Structures 169

• A function F : A ⊆ Rm → Rn is said to be L-definable if its graph, as a

subset of Rm × Rn , is L-definable.
• If a function F (x1 , ..., xm ) is definable, we can treat it as if it is a term
for the purposes of constructing more complicated definable functions and
formulas.
• Similarly, we can treat a definable subset of Rk as a basic relation in the
language for such purposes.
What I want to do now is consider some further examples of definability,
to explore the capabilities of the language we are dealing with.

Some Further Examples

1. Let f : R → R be L-definable, where L contains Lalg . Then

{x ∈ R | f is convex in an interval aroundx}

is L-definable, since “f is convex in an interval around x” can be written

as

∃x1 ∃x2 (x1 < x ∧ x < x2 ∧ “f is convex in the interval (x1 , x2 )”),

and “f is convex in the interval (x1 , x2 )” can be written as

 
x2 − t t − x1
∀t (x1 < t ∧ t < x2 ) → f (t) < f (x1 ) + f (x2 ) .
x2 − x1 x2 − x1

Note that ∀t := ¬∃t¬ and P → Q := ¬P ∨ Q.

2. Let f : A ⊂ Rk → R be L-definable, where L contains Lalg . Then {x̄ ∈
A | f is continuous at x̄} is L-definable, since “f is continuous at x” can
be written as

∀ε(ε > 0 → ∃δ(δ > 0 ∧ ∀y(|y − x| < δ) → |f (y) − f (x)| < ε)).

Same for differentiability. Note that |y − x| < δ := −δ < y − x ∧ y − x < δ.

3. Let f : R → R be L-definable, where L contains Lalg . If f is differentiable,
then f  is L-definable. Same for functions of several variables.
4. Let f : R → R be L-definable, where L contains Lalg . Then

{x ∈ R | f is Lipschitz in an interval around x}

is L-definable.
5. Let A ⊂ Rk be Lalg -definable. Then there is a formula that expresses that
“A is convex.” Same for any L containing Lalg .
6. Let A ⊂ Rk be L-definable, where L is arbitrary. Then the topological clo-
sure of A in Rk is L-definable. Many other notions from point-set topology
also are definable.
170 Charles Steinhorn

7. Let A ⊂ Rk be L-definable. Then all level sets of A are L-definable.

8. Closed polyhedra in Rk are Llin -definable.
Some guidelines for which sets are definable:
• Defining formulas cannot be infinitely long. The formula “X ⊆ R is finite,”
which we can write as “X contains 1 element ∨X contains 2 elements ∨...,”
is not definable.
• Quantification over real numbers only is allowed. Quantifying over the class
of all polynomials is not permitted, but quantifying over all polynomials
of degree ≤ n is allowed, since such polynomials are defined by n + 1
coefficients.

1.4 Set-theoretic Definability

The L-definable subsets of Rn for n = 1, 2, 3, ... is the smallest collection

D = {Dn | n ≥ 1} such that
1. Each D ∈ Dn is a subset of Rn ;
2. Rn ∈ Dn ;
3. The graph of each f : Rn → R in F is in Dn+1 ;
4. Each R ⊆ Rn in R is in Dn ;
5. For all 1 ≤ i, j ≤ n, {(x1 , ..., xn ) ∈ Rn | xi = xj } ∈ Dn ;
6. Each Dn is closed under intersection, union, and complement;
7. If π : Rn → Rm is a projection map (x1 , ..., xn ) −→ (xi1 , ..., xim ) and
X ∈ Dn then π(X) ∈ Dm ;
8. If π is as above and Y ∈ Dm then π −1 (Y ) ∈ Dn ;
9. If X ∈ Dn+m and b̄ ∈ Rm , then {ā ∈ Rn | (ā, b̄) ∈ X} ∈ Dn .

1.5 Finer structure of definable sets

A definable set typically has several different definitions. We have the Lalg
definitions of the unit interval [−1, 1]:

{x ∈ R | (−1 < x ∧ x < 1) ∨ x = −1 ∨ x = 1}

{x ∈ R | ∃y∃z x2 + y 2 + z 2 = 1}
{x ∈ R | ∀y∃z x2 + (y − z)2 = 1}.

These are three different definitions of the same set. We could construct in-
finitely many. The maxim here is that quantification generates conceptual
complexity. One goal of model theory is to attempt to analyze definable sets
in a specific context by showing that definable sets can be defined by simple
formulas.
 Tame Topology and O-Minimal Structures 171

1.6 Quantifier elimination for Llin -definable sets

Theorem. For every n = 1, 2, 3, ..., every Llin -definable subset of Rn can be

defined by a quantifier-free Llin -formula.

For convenience, write rx instead of μr (x), where r ∈ R. The theorem tells

us that every Llin -definable subset of Rn is a finite Boolean combination (i.e.,
finitely many intersections, unions, and complements) of sets of the form

{(x1 , ..., xn ) ∈ Rn | a1 x1 + · · · + an xn > b}

where a1 , ..., an , b are fixed but arbitrary real numbers.

The Llin -definable sets are called the semilinear sets. By routine set the-
oretic manipulation, semilinear sets can be written as a finite union of the
intersection of finitely many sets defined by conditions of the form

a1 x1 + · · · + an xn + b = 0,
c1 x1 + · · · + cn xn + d > 0.

Thus Llin -definability reduces (basically) to linear algebra, which we under-

stand well.

Llin -definable subsets of R

All Llin -definable subsets of R are finite Boolean combinations of sets of the
form {x ∈ R | ax > b}. Geometrically these are the union of finitely many
(possibly unbounded) open intervals and points. Consequently, neither Z nor
Q is Llin -definable.

Idea of the Proof

Eliminate quantifiers one at a time (proceed inductively).

Example

Eliminate the quantifier for the Llin -definable set

{x ∈ R | ∃y, [2x − 3y > 2 ∧ 4x − 2y ≤ 0]}.

High school algebraic elimination leads to {x ∈ R | x < 1/2}.

Question: What about Lalg -definability? I will discuss this in my next lecture.
172 Charles Steinhorn

2 The Semialgebraic Case

Today I’m going to be discussing the semialgebraic case, that is, the case of
Lalg . To construct this language, you start with polynomials with real coef-
ficients, and build up through the operations of +, ·, −. The focus of today’s
lecture is the Tarski–Seidenberg theorem.
Theorem (Tarski–Seidenberg Theorem [Tar51]). For every n = 1, 2,
3, ..., every Lalg -definable subset of Rn can be defined by a quantifier-free Lalg -
formula.
Thus every Lalg -definable subset of Rn is a finite Boolean combination
(i.e., finitely many intersections, unions, and complements) of sets of the form
{(x1 , ..., xn ) ∈ Rn | p(x1 , ..., xn ) > 0}
where p(x1 , ..., xn ) is a polynomial with coefficients in R. These are called the
semialgebraic sets. A function f : A ⊂ Rn → Rm is semialgebraic if its graph
is a semialgebraic subset of Rn × Rm .
As I did yesterday, I want to draw special attention to the definable sets
in one variable: the Lalg -definable subsets of R. You start off with sets of
the form p(x) > 0, where p(x) is a polynomial. This is nothing other than
finitely many open intervals. Consider the negation p(x) ≤ 0; this defines
finitely many closed intervals. So, the Lalg -definable subsets of R are simply
the union of finitely many intervals and points. In particular, Z and Q are
not Lalg -definable. This is the same as in the case of Llin , which is surprising,
because the structure of Lalg seems more complex. This complexity manifests
itself in higher dimensions.
As for the semilinear sets, every semialgebraic set can be written as a
finite union of the intersection of finitely many sets defined by conditions of
the form
p(x1 , ..., xn ) = 0
q(x1 , ..., xn ) > 0
where p(x1 , ..., xn ) and q(x1 , ..., xn ) are polynomials with coefficients in R.
This is just standard set theoretic manipulation of the kind I talked about
yesterday.
Let me tell you now about the proof, which is the substance of what
today’s lecture is about. There are two steps. First we prove a geometric
structure theorem that shows that any semialgebraic set can be decomposed
into finitely many semialgebraic generalized cylinders and graphs. Then we
deduce quantifier elimination from this. I will define what I mean by cylinders
and graphs a little later on. We start with a nice result called Thom’s lemma.
Lemma (Thom’s Lemma [Tho65]). Let p1 (X), ..., pk (X) be polynomials
in the variable X with coefficients in R such that if pj (X)
= 0 then pj (X) is
included among p1 , ..., pk . Let S ⊂ R have the form
 Tame Topology and O-Minimal Structures 173
+
S= pj (X) ∗j 0
j

where ∗j is one of <, >, or =, then S is either empty, a point, or an open

interval. Moreover, the (topological) closure of S is obtained by changing the
sign conditions (changing < to ≤ and > to ≥).

Note: There are 3k such possible sets, and these form a partition of R.

Proof. The proof is by induction on k. When k = 1, p1 (x) is a constant

polynomial, so ∗j gives either R or ∅. Assume the theorem is true for k − 1;
we will show that it must be true for k. Without loss of generality, suppose
pk (x) has the largest degree of the polynomials. Then {p1 , ..., pk−1 } must also
satisfy the ,k−1conditions of Thom’s lemma. Let ∗1 , ..., ∗k be given, and form the
set S  = j=1 pj (X)∗j 0. If S  = ∅ or {r}, then it clear that S = S  ∩pk (X)∗k 0
has the right form. Now suppose S  is an interval I. Note that pk is among
p1 , ..., pk−1 . On I, pk (X) > 0 or pk (X) < 0 or pk (X) = 0. So pk is either
monotone or constant on I, and so S has the right form.

We now need some “tricks” to continue with our proof. First, we identify
the complex numbers C with R2 via

a + bi ←→ (a, b)
√
where a, b ∈ R and i = −1. With this identification, multiplication of com-
plex numbers is a semialgebraic function from R2 × R2 to R2 . More generally,
Cn is identified with R2n . Second, the collection of polynomials in the variable
X with coefficients in R of degree not greater than n can be identified with
Rn+1 via
a0 + a1 X + · · · + an X n ←→ (a0 , a1 , ..., an ).
Similarly for polynomials with coefficients in C. Addition, multiplication, dif-
ferentiation of polynomials are semialgebraic functions.
Let Bkn (R) denote (as a subset of Rn+1 ) the collection of polynomials in
the variable X with real coefficients of degree not greater than n that have
exactly k distinct complex roots. Let Mkn (R) ⊂ Bkn (R) be those polynomials
of degree n with this property.
I’m now going to state a lemma that is vital for our proof.

Lemma (“Continuity of Roots”). Suppose that A ⊂ Mkn (R) is connected.

For each ā ∈ A let rā be the number of distinct real roots of the polynomial
pā (X) associated with ā. Then
1. rā = r is constant on A;
2. There are continuous functions f1 , ..., fr : A → R such that for all ā ∈ A
we have fi (ā) < fi+1 (ā) for i = 1, ..., r−1 and pā (fi (ā)) = 0 for i = 1, ..., r.
174 Charles Steinhorn

I’m not going to prove this result, but I’ll try to give you some intuition by
going back to the quadratic case. Consider the polynomial p(x) = ax2 +bx+c.
The determining factor here is the discriminant b2 − 4ac. Consider the sets
obtained by setting the discriminant equal to zero and fixing a. This is a
parabola in the plane x = a in R3 (with b and c serving as the y and z
coordinates). We have zero distinct real roots below the parabola, two distinct
real roots above the parabola, and one distinct real root on the parabola. So
these three cases constitute connected sets.
Continuity of roots takes a little while to prove. You first prove local con-
tinuity, then use connectedness to show that the number of roots is constant
across a set. This proof is involved, and I will not show it here.
Lemma. The subsets Bkn (R) and Mkn (R) of Rn+1 are semialgebraic.
The idea here is that the polynomial p(X) has a repeated root if and only
if p(X) and its derivative p (X) have a common factor. This can be expressed
by the condition that the determinant of a matrix constructed from the coef-
ficients of the so-called resultant of p and p (also called the discriminant of
p), has value 0. This is a semialgebraic condition on the coefficients. We can
extend this idea to capture Bkn (R) and Mkn (R) semialgebraically.
I’m going to go back to the case of the quadratic equation again. Let
p(x) = ax2 + bx + c, so that p (x) = 2ax + b. The discriminant can be written
as
 
c b a 
 
D(p, p ) =  b 2a 0 
 0 b 2a 
= a(4ac − b2 ).

So we ask that a(4ac − b2 ) = 0. This condition ensures that p has a repeated

root. So we have a semialgebraic condition for p to have a repeated root.
What we’ve seen is that the sets Bkn (R) and Mkn (R) are semialgebraic, and
as long as we stay on a connected subset of Mkn (R), we get the same number
of real roots. And these root functions are continuous.

2.1 Graphs and Cylinders

The structure theorem shows that a semialgebraic set S ⊆ Rn can be parti-

tioned into finitely many sets of two kinds, all of which are semialgebraic.

Graphs

Let A ⊂ Rk and f : A → R be continuous. The graph of f is the subset of

Rk+1 given by

Graph(f ) = {(x̄, y) ∈ Rk+1 | x̄ ∈ A and y = f (x̄)}.

 Tame Topology and O-Minimal Structures 175

Generalized Cylinders

Let A ⊂ Rk , and let f, g : A → R be continuous and satisfy f (x̄) < g(x̄) for
all x̄ ∈ A. The cylinder determined by f, g, and A is the subset of Rk+1 given
by
(f, g)A = {(x̄, y) ∈ Rk+1 | x̄ ∈ A and f (x̄) < y < g(x̄)}.
If A is connected, then graphs and cylinders based on A are connected.
Theorem (Structure Theorem). Let S be semialgebraic. Then:
In S has finitely many connected components and each one is semialgebraic
IIn There is a finite partition P of Rn−1 into connected semialgebraic sets
such that for each A ∈ P there is kA ∈ N and fiA : A → R ∪ {±∞} for
i = 0, 1, ..., kA + 1 satisfying
(a) f0A = −∞, fkAA +1 = ∞, fiA is continuous for 1 ≤ i ≤ kA , and
A
fi−1 (x̄) < fiA (x̄) for all 1 ≤ i ≤ kA + 1 and x̄ ∈ A;
(b) all graph sets Graph(fiA ) for 1 ≤ i ≤ kA and generalized cylinders
A
(fi−1 , fiA )A are semialgebraic.
The graphs and cylinders in (b) for all A ∈ P partitions Rn and S.
The essence of the theorem is that S can be constructed as a finite union
of semialgebraic cylinders and graphs.
Proof. The proof is by induction on n, and I shall outline the induction step.
Most broadly, the argument is as follows: show In−1 ⇒ IIn and IIn ⇒ In .
IIn ⇒ In is evident, because the graphs are connected since the functions
are semialgebraic, and the cylinders are connected since the base sets are
connected. The crux is In−1 ⇒ IIn .
Split the coordinates of Rn as (x1 , ..., xn−1 , t). Using standard set theory,
write S as the union of finitely many finite intersections of polynomial equal-
ities and inequalities. Extend the finite collection of polynomials in the given
definition of S by including all iterated partial derivatives with respect to t.
Let this expanded list of polynomials be q1 , ..., qr . The nice thing about poly-
nomials is that we only have to do this finitely many times to obtain closure
under differentiation.
For each subset S ⊂ {1, ..., r}, form the polynomial

QS (x̄, t) = qj (x̄, t).
j∈S

View x̄ as parameter variables and consider the polynomial as QS,x̄ (t), a

polynomial in the variable t whose coefficients are polynomials in x̄. For each
 ≤ degreeQS,x̄ (t) and k ≤ , let

MS,k ={x̄ ∈ Rn−1 | degreeQS,x̄ (t) =  and it has exactly
k distinct real roots}.
176 Charles Steinhorn

It should come as no surprise that you can in fact show that this subspace of
Rn−1 is semialgebraic.
Next we partition Rn−1 by taking all intersections of all MS,k l
. This still
is a finite semialgebraic partition of R n−1
. Refine this partition further to
obtain a partition P0 by taking the connected components of the sets in the
partition above. By In−1 this again is a finite semialgebraic partition of Rn−1 .
For A ∈ P0 , let QA,x̄ (t) be the product of those qj (x̄, t) which are nonzero for
(all) x̄ ∈ A. It can be shown that the number of roots of QA,x̄ (t) is uniform
as x̄ ranges over A and that the 1st , 2nd ,... root functions are continuous on
A, as A is connected.
Form the corresponding graph and generalized cylinder sets above each
set A ∈ P0 . It can be shown that each such set has the form
+
r
{(x̄, t) | x̄ ∈ A and qj (x̄, t) ∗j 0}
j=1

where ∗j is one of <, >, or =. This step uses Thom’s lemma. 

That’s the rough idea of the proof of the structure theorem. Now we want
to use it to deduce the Tarski-Seidenberg theorem.
Theorem (Tarski–Seidenberg Theorem Redux). Let f : X ⊂ Rn → Rm
be semialgebraic. Then the image of f ,

f (X) = {ȳ ∈ Rm | ȳ = f (x̄) for some x̄ ∈ X},

is semialgebraic.
Proof. Let π : Rn × Rm → Rm be the projection map onto the last m co-
ordinates. Then f (x) = π(graph(f )). It thus suffices to show that the image
under π of a semialgebraic set S ⊆ Rm+n is semialgebraic. We will show this
by induction on n. When n = 1, our desired result follows directly from the
structure theorem. Suppose the result is true for n − 1; we will show that it
holds for n. Let π1 : Rn−1 × R1+m → R1+m be the projection onto the last
m + 1 coordinates, and let π2 : R × Rm → Rm be the projection onto the last
m coordinates. Then π(S) = π2 (π1 (S)). But π1 (S) is semialgebraic by the
induction hypothesis, so π(S) is semialgebraic by the structure theorem. 
The Tarski–Seidenberg theorem gives quantifier elimination because if we
take the formula ∃y, ϕ(x1 , ..., xn , y), where ϕ is a semialgebraic quantifier-free
formula, then this is just the projection to Rn of a semialgebraic subset of
Rn+1 , and is thus semialgebraic.

2.2 Algorithmic Cruelty

It turns out that everything I’ve explained can be done algorithmically. That
is, we can take any definable set and algorithmically give it a semialgebraic
 Tame Topology and O-Minimal Structures 177

description. The proof I’ve given you today relates to work done by George
Collins on cylindrical algebraic decomposition. Tarski’s original proof gives an
algorithm for quantifier-elimination: given an Lalg -formula as input, it out-
puts a quantifier-free formula that defines the same set as the input formula.
Computational efficiency of a quantifier elimination algorithm thus becomes
important for applications (e.g., robot motion planning). Cylindrical algebraic
decomposition-based quantifier elimination, such as described above and de-
veloped in 1975 by Collins [Col75], has played an important role.
Quantifier elimination for Ralg is, unfortunately, an inherently computa-
tionally intensive problem. It is known that there is a doubly exponential
lower bound in the number of quantifiers for worst-case time complexity. So,
quantifier elimination is something that is do-able in principle, but not by
any computer that you and I are ever likely to see. Well, I’ll retract that last
statement because it’s probably false.

3 O-minimal Structures
An L-structure R is o-minimal if every definable subset of R is the union
of finitely many points and open intervals (a, b), where a < b and a, b ∈
R ∪ {±∞}. Thus far we have seen two examples of o-minimal structures: Rlin ,
the semilinear context, and Ralg , the semialgebraic context. We showed by
quantifier elimination that these structures are o-minimal. O-minimal is short
for ordered minimal. We use this name because, for an o-minimal structure,
the definable subsets of R are exactly those that must be there because of the
presence of <. The hypothesis of o-minimality combined with the power of
definability have remarkable consequences.

3.1 Minimal Structures

Consider the field of complex numbers (C, ·, +). There is a theorem due to
Chevalley which says that (C, ·, +) has the quantifier elimination property. All
definable subsets of C can be defined by finite Boolean combinations of poly-
nomials equalities p(x) = 0 or inequalities p(x)
= 0, where p has coefficients
in C. As long as p is not the zero polynomial, p(x) will consist of finitely many
points, and p(x)
= 0 will consist of cofinitely many points. So all definable
subsets of C are either finite or cofinite. Notice that these are the sets you
have to be able to define using equality. So (C, ·, +) is minimal.
Minimal structures were first know by model theorists. One of my contri-
butions was to look at the properties of minimal structures when we have an
ordering. The theme of what I’m going to be talking about today and in the
remaining lectures is o-minimal structures.
178 Charles Steinhorn

3.2 O-minimal Structures

The first result, which I’m going to spend quite a bit of time on, is one of the
most important results on o-minimality.

Theorem (Monotonicity Theorem). Let R be an L-structure that is o-

minimal. Suppose that f : R → R is R-definable. Then there are −∞ =
a0 < a1 < · · · < ak−1 < ak = ∞ in R ∪ {±∞} such that for each j < k
either f  (aj , aj+1 ) is constant or is a strictly monotone bijection of (possibly
unbounded) open intervals in R.

In particular, all definable f : R → R are piecewise continuous. Now we can

see the power of o-minimality. We are not confined to looking at semialgebraic
sets.
I’m going to give you some idea of what’s in the proof of the monotonicity
theorem. We need to show that for every open interval I ⊆ R, there is an open
interval I ∗ ⊆ I on which f is constant or strictly montone. This will give us
a local version of the theorem, which we need to globalize. For now, assume
the local version and consider the task of globalization.
Let θ(x) say “x is the left endpoint of an interval on which f is constant
or strictly monotone and this interval cannot be extended properly on the left
while preserving this property.” You should check that this is a meaningful
expression in the context of our structure. Clearly θ(x) defines a finite subset of
R, say b1 , ..., bp . We know that f  (bi , bi+1 ) is strictly monotone or constant,
because otherwise the local version of the theorem implies that (bi , bi+1 ) ∩
θ(x)
= ∅. To finish working from local to global, we note that if f  (bi , bi+1 ) is
strictly monotone or constant then f ((bi , bi+1 )) consists of finitely many points
and intervals, and so we can partition each (bi , bi+1 ) into a finite collection of
points and intervals in such a way that the image under f of each interval in
the partition is itself an open interval.
Now I’ll try to give you some idea of what’s in the proof of the local version
of the theorem. Define the following formulas:

ϕ0 (x) says “∃ open interval J with x as an endpoint on which f

is constantly equal to f (x).”
ϕ1 (x) says “∃ open interval J containing x such that f (y) < f (x)
for y ∈ J, y < x and f (y) > f (x) for y ∈ J, y > x.”
ϕ2 (x) says “∃ open interval J containing x such that f (y) < f (x)
for y ∈ J, y > x and f (y) > f (x) for y ∈ J, y < x.”
ϕ3 (x) says “∃ open interval J containing x such that f (y) < f (x)
for y ∈ J, y
= x.”
ϕ4 (x) says “∃ open interval J containing x such that f (y) > f (x)
for y ∈ J, y
= x.”

These five formulas exhaust all possibilities, by o-minimality. Also by o-

minimality, possibilities 3 and 4 cannot hold, so ∃I1 ⊆ I contained in the
 Tame Topology and O-Minimal Structures 179

set defined by one of ϕ0 , ..., ϕ2 . If I1 is contained in the set defined by ϕ0 ,

then f is constant on I1 . If I1 is contained in the set defined by ϕ1 , then f is
strictly monotone increasing on I1 . And if I1 is contained in the set defined
by ϕ2 , then f is strictly monotone decreasing on I1 .

3.3 Cells

What I want to do next is talk about a particular kind of definable set: cells.
Cells in R are either points (0-cells) or intervals (1-cells). Cells in R2 can
be constructed as follows: begin with cells in R, and take the graph of a
continuous function defined on those cells, as well as the cylinders defined by
pairs of continuous functions.
More formally,-let R be an L-structure. The collection of R-cells is a
subcollection C = ∞ n=1 Cn of the R-definable subsets of R for n = 1, 2, 3, ...
n

defined recursively as follows.

Cells in R

The collection of cells C1 in R consists of all single point sets {r} ⊂ R and all
open intervals (a, b) ⊆ R, where a < b and a, b ∈ R ∪ {±∞}.

Cells in Rn+1

Assume the collection of cells Cn in Rn have been defined. The collection Cn+1
of cells in Rn+1 consist of two different kinds: graphs and cylinders.

Graphs

Let C ∈ Cn and let f : C ⊆ Rn → R be R-definable and continuous. Then

Graph(f ) ⊆ Rn+1 is a cell.

Generalized Cylinders

Let C ∈ Cn . Let f, g : C ⊆ Rn → R be R-definable and continuous such that

f (x̄) < g(x̄) for all x̄ ∈ C. Then the cyclinder set (f, g)C ⊂ Rn+1 is a cell.
Cells are R-definable and connected. There is a concept of dimension for
cells: for each cell C ⊆ Rn there is a largest k ≤ n and i1 , ..., ik ∈ {1, 2, ..., n}
such that if π : Rn → Rk is the projection mapping given by π(x1 , ..., xn ) =
(xi1 , ..., xik ), then π(C) ⊆ Rk is an open cell in Rk . This value of k we call the
dimension of C. Basically speaking, the dimension of a cell is the number of
times we construct cylinders in the “bottom-up” construction of a cell.
180 Charles Steinhorn

3.4 Cell Decomposition

An R-decomposition D of Rn is a partition of Rn into finitely many R-cells

satisfying:
1. If n = 1, then D consists of finitely many open intervals and points.
2. If n > 1 and πn : Rn → Rn−1 denotes projection onto the first n − 1
coordinates, then {πn (C) : C ∈ D} is a decomposition of Rn−1 .
This is just a generalized version of the cylindrical algebraic decomposition.

Theorem (Cell Decomposition Theorem). Let R be o-minimal and let

S ⊂ Rn be definable. Then there is a decomposition D of Rn that partitions
S into finitely many cells. In particular, if f : A ⊂ Rn → R is definable, then
there is a partition of A into cells such that the restriction of f to each cell is
continuous.

Some Obvious Consequences

• Using the definition of a cell as defined above, we obtain a good geometric

definition of the dimension of a definable set. Namely, the dimension of
a set is the maximium dimension of the cells in a decomposition that
partitions it.
• Since cells are connected, it follows that every definable set has finitely
many connected components. This is one aspect of the “tameness” alluded
to in the name “tame topology.”
• The topological closure of a definable set consists of finitely many con-
nected components; same for the the interior and the frontier (or bound-
ary). Even if you start with a connected set, this won’t necessarily be true
unless the set is definable. Consider a comb with infinitely many teeth,
where the points of those teeth are not in the set; the frontier of this
connected set consists of infinitely many connected components.

3.5 Definable Families

Let S ⊂ Rn+p be a definable set in the o-minimal structure R. For each b̄ ∈ Rn

define Sb̄ := {ȳ ∈ Rp | (b̄, ȳ) ∈ S}. (Note that some Sb̄ may be empty.) The
family {Sb̄ | b̄ ∈ Rn } of subsets of Rp is called a definable family.
The next result is quite surprising; we did not expect to find it.

Theorem (Uniform Bounds Theorem). Let R be o-minimal and let S ⊂

Rn+p be a definable set such that Sb̄ is finite for all b̄ ∈ Rn . Then there is a
fixed K ∈ N satisfying |Sb̄ | ≤ K for all b̄ ∈ M n .
 Tame Topology and O-Minimal Structures 181

Note that a definable subset of R in an o-minimal structure R is infinite

if and only if it contains an interval. So the theorem actually is stronger. It
can be thought of as a generalization of our earlier result on the roots of
polynomial equations.
Recall our discussion of minimality and the field of complex numbers. We
can also obtain a version of the uniform bounds theorem for subsets of Cn+1 .

3.6 Sketch Proof of the Cell Decomposition Theorem

Next I want to give you a quick sketch of the proof of the cell decomposi-
tion theorem. The proof is by induction on n. Define the following induction
hypotheses:
In There exists a cell decomposition that respects finitely many definable
subsets of Rn .
IIn Let f : A ⊆ Rn → R be definable. Then there is a decomposition D of
Rn that partitions A such that f  C is continuous for all C ∈ D, C ⊆ A.
IIIn The uniform finiteness property holds for all definable families {Sb̄ | b̄ ∈
Rn }, where S ⊆ Rn+p is definable.
I1 is true by the definition of o-minimality, II1 is just the monotonicity
theorem, and III1 requires an intricate direct argument that we will not go
into. By assuming Im , IIm , IIIm for all m < n, we can use the monotonicity
theorem to show that In holds, and then that IIn holds, and finally that IIIn
holds.

3.7 Quantifier Elimination

Theorem (van den Dries [Van98]). Let I be an index set and for each
i ∈ I let fi : Rni → R be (total) analytic functions. Then the structure
(Ralg , {fi : i ∈ I}) admits quantifier elimination if and only if each fi is
semialgebraic.
So, e.g., Rexp , the real exponential field does not have quantifier elimina-
tion. Quantifier elimination is not easy to come by.

3.8 Partial Elimination

Suppose that a structure R has the property that every definable set is defin-
able by an existential formula, that is, a formula having the form
∃x1 ∃x2 · · · ∃xk ϕ
where ϕ is a quantifier-free L-formula. How can this help? Suppose that the
R-definable sets that are definable using quantifier-free formulas can be ana-
lyzed, and that all such have finitely many connected components. The contin-
uous image of a connected set is connected (elementary topology). Existential
182 Charles Steinhorn

quantification corresponds to projection, and projection is a continuous map.

Thus all R-definable subsets of R have finitely many connected components,
that is, all such are the union of finitely many points and open intervals. We
conclude that R is o-minimal, and so all the geometric and topological proper-
ties available as consequences of o-minimality apply. We will talk more about
this tomorrow. Tomorrow I want to give you a deeper understanding of what
tame topology means, and take you on a tour of structures that we now know
to be o-minimal.

4 Examples and Some Further Properties

I’m going to start off by giving you a tour of some o-minimal structures, and
I’ll finish today by giving you some of the finer tame topological results.

4.1 Examples of O-minimal Structures

So far we have two examples of o-minimal structures: Rlin , the semilinear con-
text, and Ralg , the semialgebraic context. As I shall discuss later, o-minimality
implies a wealth of good analytic and topological properties. This provided
ample motivation to seek out o-minimal structures that expand Ralg to include
transcendental data.
We now survey some of the remarkable results that have been obtained
beginning in the mid-1980s.

4.2 Expansions of Ralg

Consider the class of restricted analytic functions, an, where g : Rn → R ∈ an 

open
if there is some analytic f : U ⊂ R → R such that [0, 1]n ⊂ U ,
n

g  [0, 1]n = f  [0, 1]n , and g(x̄) = 0 otherwise. (The point of doing something
like this is that you can avoid considering behaviour at ±∞ by appropriate re-
striction of the functions’ domain.) Let Ran be the expansion of Ralg by adding
as basic functions all g ∈ an. Then the structure Ran admits elimination down
to existential formulas and is o-minimal (van den Dries [Van86]).
Let me give you a sense of what kind of functions you can actually get
here. You can not only get bounded functions, but also functions that live on
all of R. For instance, we can obtain arctan x from the restriction of sin x and
cos x to (−π, π). Everything I said about definable functions—monotonicity
and the cell decomposition—works for these functions. This result depends
on the work of Lojasiewicz and Gabrielov (in the 1960s). Another useful fact
about Ran is that it is polynomially bounded.
Polynomial growth: Let f : (a, ∞) → R be definable in Ran . Then there is
some N ∈ N such that |f (x)| < xN for sufficiently large x.
Adjoin to Ran the function −1 given by x −→ 1/x for x
= 0 and 0−1 = 0.
 Tame Topology and O-Minimal Structures 183

Theorem (Denef–van den Dries [DV88]). (Ran ,−1 ) admits elimination

of quantifiers.

Whereas in the semialgebraic context we know all the basic functions (they
are just polynomials), in this case we have a much larger collection of basic
functions, so that our descriptive language is much richer. The languages in
(1) and (2) are large, but nonetheless natural. Quantifier elimination always
can be achieved by enlarging the language, but no advantage is gained: in
general, the quantifier-free sets thus obtained can be horribly badly behaved.
The next theorem is really quite spectacular, and was a breakthrough for
the subject.

Theorem (Wilkie [Wil96]). Rexp admits elimination down to existential

formulas.

O-minimality then follows by a result of Khovanskii [Kho80] (which Wilkie

also uses in his proof). Recall that yesterday I showed that quantifier elimi-
nation is not possible in Rexp . This result tells us “the best that we can do.”
This was the first o-minimal structure with functions growing faster than
polynomials.
This theorem addresses a question posed originally by Tarski. He asked
if his results on Ralg could be extended to Rexp . Wilkie’s result from the
syntactic and topological points of view is the best possible.
Macintyre and Wilkie [MW96] link decidability of the theory of the real
exponential field to the following conjecture.

Conjecture (Schanuel’s Conjecture [Sch91]). Let r1 , ..., rn ∈ R be lin-

early independent over Q. Then the transcendence degree over Q of Q(r1 , ..., rn ,
er1 , ..., ern ) is at least n.

This has implications about the transcendence of various things, which I

might talk more about tomorrow. Schanuel’s conjecture is now regarded by
mathematicians as being intractable, so I don’t know if we will ever see it
verified. But most mathematicians seem to believe that it is true.

Theorem (Macintyre–Wilkie [MW96]). Schanuel’s conjecture implies

that the theory of the real exponential field is decidable.

A natural question is to ask what happens if we combine the restricted

analytic and the exponential functions in our basic functions. Van den Dries
and Miller adapt Wilkie’s techniques to prove the following.

Theorem (van den Dries–Miller [VM94]). Ran,exp admits elimination

down to existential formulas and (by Khovanskii) is o-minimal.

Inspired by the work of Ressayre [Res93], these authors analyze Ran,exp

further.
184 Charles Steinhorn

Theorem (van den Dries–Macintyre–Marker [VMM94]). Ran,exp,log

admits elimination of quantifiers.

Their analysis further shows that every definable function in one variable
is bounded by an iterated exponential. Macintyre–Marker [MM97] show that
the logarithm is necessary for the quantifier elimination. In a second paper
van den Dries–Macintyre–Marker [VMM97] develop tools that enable them to
obtain several further results.
What I want to do now is talk about results of definability and undefin-
ability. Let f (x) = (log x)(log log x) and let g(x) be a compositional inverse
to f defined on some interval (a, ∞). Hardy [Har12] conjectured in 1912 that
g is not asymptotic to a composition of exp, log, and semialgebraic functions.

Theorem (van den Dries–Macintyre–Marker [VMM97]). Hardy’s con-

jecture is true.

Let me talk about some other undefinability results which people will
perhaps find more down to earth. Building on some remarkable ideas and
results of Mourges–Ressayre [MR93], van den Dries–Macintyre–Marker derive
some “undefinability” results also.

Theorem (van den Dries–Macintyre–Marker [VMM97]). None of the

following functions is definable in Ran,exp.
∞
i. the restriction of the gamma function Γ (x) = 0 e−t tx−1 dt to (0, ∞),
x 2
ii. the error function 0 e−t dt,
∞ −1 −t
iii. the logarithmic integral x t e dt, ∞
iv. the restriction of the Riemann zeta function ζ(s) = n=1 n−s to (1, ∞).

These are functions that we deal with all the time, but are not definable
in the context of Ran,exp. So our work is not yet complete.
For r ∈ R let xr denote the real power function
 r
x if x > 0
xr =
0 if x ≤ 0.

Let RR
an denote the expansion of Ran by all power functions x for r ∈ R.
r

Theorem (Miller [Mil94]). RR

an has elimination of quantifiers.

Again, this is a structure with functions whose growth at infinity is

bounded by polynomials. An expansion R of Ralg is polynomially bounded
if for every definable f : R → R there is some N ∈ N so that |f (x)| < xN for
sufficiently large x. I’ve given you examples of functions that are polynomially
bounded (an) and those that are not (exp).
 Tame Topology and O-Minimal Structures 185

Theorem (Growth Dichotomy Theorem). Let R be an o-minimal ex-

pansion of Ralg . Then either the exponential function ex is definable in R or
R is polynomially bounded. In the second case, for every definable f : R → R
in R not ultimately identically zero, there are c ∈ R\{0} and r ∈ R such that
f (x) = cxr + o(xr ) as x → ∞.
This amazing theorem shows that there is no “middle ground”: if functions
are not bounded by polynomial growth, then exponential growth must be
possible.
Let ∂Φ be a collection of restricted analytic functions that is closed under
differentiation. Since derivatives are definable in Ran (definability of deriva-
tives was mentioned in Section 1), all of the functions in ∂Φ are definable in
this structure.
Theorem. The structure (Ralg , f )f ∈∂Φ has elimination down to existential
formulas.
This result does not give us new functions, but tells us how we can think
of what we have in a nicer way. The next result gives us new functions.
Using (delicate) generalized power series methods new expansions of Ralg
are constructed in van den Dries–Speissegger [VS98]. There are two polyno-
mially bounded versions that have elimination down to existential formulas:
generalized convergent power series (using real, rather than integer, powers)
and multisummable series. Moreover, the exponential function can be added
while preserving o-minimality. If, in addition the logarithmic function is ad-
joined as a basic function, these expansions admit quantifier elimination. In
one of these expansions, the gamma function on (0, ∞) is definable, and in
the second, the Riemann zeta function on (1, ∞) is definable.
Now we come to some really beautiful results of Wilkie [Wil96]. I have to
introduce another class of functions, which again is quite natural. A function
f : Rn → R is said to be Pfaffian if there are functions f1 , ..., fk : Rn → R
and polynomials pij : Rn+i → R such that
∂fi
(x̄) = pij (x̄, f1 (x̄), ..., fi (x̄))
∂xj
for all i = 1, ..., k, j = 1, ..., n, and x̄ ∈ Rn . So this is a chaining procedure,
going one step at a time to generate more complicated functions. Wilkie proved
(by quite different methods than used previously) that the expansion of Ralg
by all Pfaffian functions is o-minimal. The introduction of Pfaffian functions
allows us to integrate and retain o-minimality. Our class of functions is now
even richer.
Speissegger [Sep99] extends Wilkie’s methods to obtain the “Pfaffian clo-
sure” of an o-minimal expansion of Ralg . In particular, such a structure is
closed under integration (antidifferentiation) of functions in one variable.
I have one more comment before I move on to the second part of this
talk. What about quantifier elimination in these expansions? Unfortunately,
186 Charles Steinhorn

this is completely unknown. We have seen that many geometric results are
obtainable for these o-minimal strcutures, but results on quantifier elimination
are not available.

4.3 Finer Analytic and Topological Consequences of O-minimality

For this section, assume throughout that we work in some o-minimal expansion
R of Ralg . We showed earlier in our discussion of the cell decomposition that
our functions are continuous on each cell. A natural question to ask is: can
we do better than this? The next theorem shows that the answer is yes.
Theorem (C k Cell Decomposition Theorem). For each definable set
X ⊂ Rm and k = 1, 2, ..., there is a decomposition of Rm that respects X
and for which the data in the decomposition are C k .
In the next theorem, “definably homeomorphic” means that the homeo-
morphism between structures is itself a definable function.
Theorem (Triangulation Theorem). Every definable set X ⊂ Rm is de-
finably homeomorphic to a semilinear set. More precisely, X is definably home-
omorphic to a union of simplices of a finite simplicial complex in Rm .
Note that these results are “nice” topological results, in the spirit of the
term “tame topology” coined by Grothendieck. The next theorem is another
nice result. It says that if we look at the fibers of some set, then the fibers
corresponding to a given connected component of that set are homeomorphic.
Theorem (Number of Homeomorphism Types). Let S ⊂ Rm+n be de-
finable, so that {Sā | ā ∈ Rm } is a definable family of subsets of Rn . Then
there is a definable partition {B1 , ..., Bp } of Rm such that for all ā1 , ā2 ∈ Rm ,
the sets Sā1 and Sā2 are homeomorphic if and only if there is some j = 1, ..., p
such that ā1 , ā2 ∈ Bj .
Uniform finiteness combined with Wilkie’s theorem yields Khovanskii’s
theorem. You all know the theorem that says that a polynomial of degree k
has no more than k distinct real roots. One of the implications of Khovanskii’s
theorem is that there is a uniform bound on the number of distinct real roots
of pl,k (x) = axl + bxk as l, k vary.
Theorem (Khovanskii [Kho91]). There exists a bound in terms of m and
n for the number of connected components of a system of n polynomial in-
equalities with no more than m monomials.
There is a trick to proving this theorem. Replace xm by em log x , and let
m vary over R. The set of (a, b, m, n, x) such that aem log x + ben log x = 0 is in
Rexp . Now fix x and let the other parameters vary. Uniform finiteness gives
us a bound on the size of the fibers.
The next theorem gives an o-minimal improvement of the previous result.
 Tame Topology and O-Minimal Structures 187

Theorem. There is a bound in terms of m and n for the number of homeo-

morphism types of the zero sets in Rn of polynomials p(x1 , ..., xn ) over R with
no more than m monomials.

Theorem (Marker–Steinhorn [MS94]). Let R be an o-minimal expansion

of Ralg , and let gā : B ⊆ Rm → R for ā ∈ A ⊂ Rm be an R-definable family
G of functions. Then every f : B ⊆ Rm → R which is in the closure of G is
definable in R.

Here, closure refers to closure in the product topology RB . This result is

very surprising. Consider it in the case of semialgebraic functions: the point-
wise limit of semialgebraic functions is semialgebraic.

4.4 The Euler Characteristic

Now what I want to talk about is how we can begin to do algebraic topol-
ogy from the viewpoint of o-minimality. We would like to consider the Euler
characteristic in an o-minimal context.
Let S ⊂ Rn be definable and P be a partition of S into cells.
Let n(P, k) be
the number of cells of dimension k in P, and define EP (S) = (−1)k n(P, k).

Proposition. If P and P  are partitions of S into cells, then EP (S) =

EP  (S).

So we define E(S) = EP (S) for any partition P. The Euler characteristic

has played a very interesting role in o-minimal theory generally. Its properties
include the following:
1. Let A and B be disjoint definable subsets of Rn . Then E(A∪B) = E(A)+
E(B).
2. Let A ⊂ Rm and B ⊂ Rn be definable. Then E(A × B) = E(A)E(B).
3. Let f : A ⊂ Rm → Rn be definable and injective. Then E(A) = E(f (A)).

5 VC Dimension and Applications

Before I begin, I want to thank the organizer, Don. I know it was talked about
on Wednesday night—the wonderful attention you pay to graduate students—
and I can see that this week. And for me too, not a graduate student, it has
been a wonderful week. So I just want to thank Don for that.

5.1 Vapnik-Čhervonenkis Dimension

A collection C of subsets of a set X shatters a finite subset F if {F ∩ C | C ∈

C} = P(F ), where P(F ) is the set of all subsets of F . The collection C is a
188 Charles Steinhorn

VC-class if there is some n ∈ N such that no set F containing n elements is

shattered by C, and the least such n is the VC-dimension, V(C), of C.
Let C ∩ F := {C ∩ F | C ∈ C} and for n = 1, 2, ..., let

fC (n) := max{|C ∩ F | | F ⊂ X and |F | = n}.

 
Also, let pd (n) = i<d ni . The next theorem gives us a polynomial bound
on the growth of fC for VC-classes.

Theorem (Sauer [Sau72]). Suppose that fC (d) < 2d for some d. Then
fC (n) ≤ pd (n) for all n.

To prove Sauer’s theorem we need the following proposition.

Proposition. Let |F | = n, and let D be a collection of subsets of F such that

|D| > pd (n), where d ≤ n. Then there is E ⊆ F , |E| = d, so that D shatters
E.
Proof. The proof is by induction on n. The result is clear when d = 0 or d = n,
so assume 0 < d < n. Fix x ∈ F and let F  = F \{x}, D = {D\{x} | D ∈ D}.
Consider the map π(D) = D\{x}. Note that π −1 (D ) has either one or two
elements (i.e., D , D ∪ {x}), depending on whether or not x ∈ D. Write
D = D1 ∪ D2 , where D1 is the class of all sets with one preimage under π,
and D2 is the class of all sets with two preimages under π. If |D | > pd (n − 1),
then by the induction hypothesis we have E  ⊆ F  , |E  | = d, so that D
shatters E  . It follows that D shatters E  . If |D | ≤ pd (n − 1), then we have
|D| = |D1 | + 2|D2 | = |D | + |D2 |. But |D| > pd (n) = pd (n − 1) + pd−1 (n − 1),
and so |D2 | > pd−1 (n−1). Thus by the induction hypothesis we have E  ⊆ F  ,
|E  | = d − 1 so that D2 shatters E  . It follows that D shatters E  ∪ {x}. 
Now we can use our proposition to prove Sauer’s theorem.
Proof. If d > n then pd (n) = 2n , and the inequality holds trivially. So let
d ≤ n. Consider an arbitrary set F ⊂ X with |F | = n. If |C ∩ F | > pd (n),
then by our proposition there exists E ⊆ F , |E| = d such that C shatters E.
But this contradicts fC (d) < 2d . Thus for all F we must have |C ∩ F | ≤ pd (n),
implying that fC (d) ≤ pd (n). 
Now I’m going to try to connect back to the logical issues we have been
discussing through the week. An L-formula ϕ(x1 , ..., xk ; y1 , ..., ym ) has the
independence property with respect to the L-structure R if for every n =
1, 2, ..., there are b̄1 , ..., b̄n ∈ Rm such that for every X ⊆ {1, ..., n}, there is
some āX ∈ Rk satisfying

ϕ(āX ; b̄i ) is true in R ⇔ i ∈ X.

If ϕ does not have the independence property with respect to R, we let I(ϕ)
be the least n for which the property above fails.
 Tame Topology and O-Minimal Structures 189

For an L-formula ϕ(x̄; ȳ) and a structure R, let S ⊆ Rk+m be the set
defined by ϕ. We let Cϕ := {Sb̄ | b̄ ∈ Rm } denote the family of subsets of Rk
determined by S.
Now what I want to do is come to the connection between the concept I
have just defined, and VC dimension.
Theorem (Laskowski [Las92a]). The definable family Cϕ is a VC-class if
and only if ϕ does not have the independence property. Moreover, if V(Cϕ ) = d
and I(ϕ) = n, then n ≤ 2d and d ≤ 2n (and these bounds are sharp).
Let ψ(ȳ; x̄) := ϕ(x̄; ȳ) be the dual formula of ϕ. That is, ψ and ϕ are the
same formula (and so define the same set) with the roles of x̄ and ȳ reversed.
The theorem follows from the next two lemmas.
Lemma 1. With the notation as above, V(Cϕ ) = d if and only if I(ψ) ≥ d.
Proof. By definition, V(Cϕ ) ≥ d if and only if there exist ā1 , ā2 , ..., ād ∈ Rk
such that for every X ⊆ {1, ..., d} there is bX ∈ Rm for which ϕ(āj , b̄X ) true
⇔ j ∈ X. This exactly says: I(ψ) ≥ d. 
Understanding Lemma 1 is just a matter of understanding the definitions
of VC dimension and the independence number.
Lemma 2. Let the notation be as above. Then I(ϕ) = n implies I(ψ) ≤ 2n .
Proof. Suppose I(ψ) > 2n . By definition, there are b̄s ∈ Rk for each s ⊆
{1, ..., n} so that for every X ⊆ {s | s ⊆ {1, ..., n}} there is āX ∈ Rm such that
ϕ(āX , b̄s ) true ⇔ s ∈ X. For i = 1, 2, ..., n, let Xi = {s ⊆ {1, ..., n} | i ∈ s}. We
have āX1 , āX2 , ..., āXn ∈ Rm . Now for each s ⊆ {1, ..., n}, we have ϕ(b̄s ; āXi )
true ⇔ s ∈ Xi ⇔ i ∈ s. 
Now we can use these two lemmas to do a quick proof of Laskowski’s
theorem.
Proof.
Cϕ is a VC-class ⇔ V(Cϕ ) = d for some d ∈ N
⇔ I(ψ) = d (by Lemma 1)
⇒ I(ϕ) ≤ 2d (by Lemma 2)
⇒ ϕ does not have the independence property.
Conversely,
ϕ does not have the independence property ⇔ I(ϕ) = n for some n ∈ N
⇒ I(ϕ) ≤ 2n (by Lemma 2)
⇔ V(Cϕ ) ≤ 2n (by Lemma 1)
⇒ Cϕ is a VC-class.

190 Charles Steinhorn

We say that the L-structure R has the independence property if there is

a formula ϕ(x; ȳ) with just the single variable x that has the independence
property with respect to R.
Applying model theoretic methods, Laskowski gives a clear combinatorial
proof of the following theorem due to Shelah.

Theorem (Shelah [She71]). An L-structure R has the independence prop-

erty if and only if there is a formula ϕ(x̄; ȳ) (in any number of x variables)
that has the independence property with respect to R.

This is a very hard theorem, and attests to Shelah’s incredible combinato-

rial genius. Interestingly, if you look at the paper with Sauer’s theorem, there
is a footnote that quotes a referee as saying that these results had been proved
earlier by Shellah.
Laskowski’s theorem combined with the following result provides the link
between o-minimality and VC-classes.

Proposition (Pillay–Steinhorn [PS86]). O-minimal structures do not

have the independence property.

Theorem (Laskowski [Las92a]). Let R = (R, <, ...) be o-minimal and let
S ⊂ Rk+m be definable. Then the collection C = {Sx̄ | x̄ ∈ Rm } is a VC-class.

Thus any definable family of subsets in an o-minimal structure constitutes

a VC-class. This theorem gives us a “black box” to generate an enormous
variety of VC-classes.
Note that many structures are known not to have the independence prop-
erty (by work of Shelah), and thus Laskowski’s theorem provides significantly
more examples of VC-classes. To illustrate, the field of complex numbers,
(C, +, ·) does not have the independence property, and thus any definable
family of sets in this structure is a VC-class.

5.2 Probably Approximately Correct (PAC) Learning

Begin with an instance space X that is supposed to represent all instances

(or objects) in a learner’s world. A concept c is a subset of X, which we can
identify with a function c : X → {0, 1}. A concept class C is a collection of
concepts.
A learning algorithm for the concept class C is a function L which takes
as input m-tuples ((x1 , c(x1 )), ..., (xm , c(xm ))) for m = 1, 2, ... and outputs
hypothesis concepts h ∈ C that are consistent with the input. If X comes
equipped with a probability distribution, then we can define the error of h to
be err(h) = P (hΔc).
The learning algorithm L is said to be PAC if for every ε, δ ∈ (0, 1) there
is mL (ε, δ) so that for any probability distribution P on X and any concept
c ∈ C, we have for all m ≥ mL(ε,δ) that
 Tame Topology and O-Minimal Structures 191

P ({x̄ ∈ X m | err(L((xi , c(xi ))i≤m )) ≤ ε}) ≥ 1 − δ.

It can be shown that an algorithm that outputs a hypothesis concept h con-

sistent with the sample data is PAC provided that C is a VC-class. Moreover,
for given # and δ, the number of sample points needed is, roughly speaking,
proportional to the VC-dimension V(C).

5.3 Neural Networks

Macintyre–Sontag [MS93] and Karpinski–Macintyre [KM95] apply Laskowski’s

result and the uniform bounds available in o-minimal structures to answer
questions about neural networks. The output in a sigmoidal neural network is
the result of computing a quantifier-free formula whose atomic formulas have
the form τ (x̄, w̄) > 0 or τ (x̄, w̄) = 0, where τ is built from polynomials and
exp, x̄ are input values, and w̄ represent a tuple of programmable parameters.
Varying the parameters gives rise to a definable family in an o-minimal struc-
ture and hence Laskowski’s theorem applies, which tells us that it is possible
to PAC learn the architecture of such a network.
The first results of Macintyre and Sontag applied Laskowski’s theorem to
prove finite VC-dimension. Using quantitative results of Khovanskii, Karpinski
and Macintyre give an upper bound for the VC-dimension that is O(m4 ),
where m is the number of weights. Koiran and Sontag [KS97] have established
a quadratic lower bound (in the number of weights) for the VC-dimension.

Acknowledgments

We would like to thank Brendan Beare for his accurate reporting of Professor
Steinhorn’s five lectures on “Tame Topology and O-minimal Structures.”