Notes on Ultrafilters

Alex Kruckman

Toolbox Seminar 11/7/12

1 Basic theory
Let X be a set. An (ultra)filter on X is a consistent choice of which subsets of X are “large”.
Definition. A filter on X is F ⊆ P(X) such that
1. X ∈ F (the whole set is large).

2. ∅ 6∈ F (the empty set is not large).

3. If A ∈ F and A ⊆ B, then B ∈ F (any set containing a large set is large).

4. If A, B ∈ F, then A ∩ B ∈ F (large sets have large intersection).

Examples. • The trivial filter: F = {X}

• The principal filter generated by x ∈ X: F = {A ⊆ X | x ∈ A}

• The cofinite filter (X infinite): F = {A ⊆ X | X \ A is finite}

Note that filters have the finite intersection property (FIP):
\
If A1 , . . . , An ∈ F, then Ai 6= ∅.
i

Lemma (FIP). Any subset S ⊆ P(X) with the FIP has a minimal filter containing it, the
filter generated by S.
Proof. Close S downward under finite intersections, then upward under supersets.
Definition. A filter F on X is an ultrafilter if
5. For any A ⊆ X, A ∈ F or X \ A ∈ F (every set is either large or co-large).
Exercise 1. Given a filter F, show that F is an ultrafilter iff ni=0 Ai ∈ F implies that
S
Ai ∈ F for some i (a large set cannot be a finite union of small sets).
Note that principal filters are ultrafilters, but the trivial filter and the cofinite filter are
not. Are there any nonprincipal ultrafilters? The answer is yes, but not constructively! The
following lemma is a weak form of the Axiom of Choice, i.e. it is not provable in ZF.

1
Lemma (Ultrafilter lemma). Every filter is contained in an ultrafilter.

Proof. Zorn’s Lemma on the poset of filters on X containing F.

We need to check the following:

• The union of a chain of filters is a filter. (Easy.)

• A maximal filter is an ultrafilter. (Let F be maximal. If F is not an ultrafilter, take A

with A ∈ / F and X \ A ∈ / F. Then F ∪ {A} has the FIP, and F is properly contained
in the filter generated by F ∪ {A}, contradicting maximality.)

Corollary. Any subset S ⊆ P(X) with the FIP is contained in an ultrafilter.

Exercise 2. Every nonprincipal ultrafilter contains the cofinite filter.

2 Generalized limits
Consider a function f : N → [0, 1], determining a sequence a0 , a1 , a2 , . . . We say that
lim an = L if for every open set U containing L, all but finitely many natural numbers
n→∞
are mapped into U by f , i.e. f −1 (U ) is in the cofinite filter on N.
Now given distinct points a and b in [0, 1], the sequence a, b, a, b, . . . does not converge. If
we take disjoint neighborhoods a ∈ U , b ∈ V , f −1 (U ) and f −1 (V ) are both infinite, coinfinite
subsets of N. An ultrafilter on N would give preference to one of these two sets and decide
whether a or b should be the limit of the sequence.

Definitions
We’ll take a more general perspective by transporting (ultra)filters from the index set to the
target space (from N to [0, 1] in the example above).

Exercise 3. Let F be an (ultra)filter on the set X, and let f : X → Y be a map of sets.

Then f∗ F = {A ⊆ Y | f −1 (A) ∈ F} is an (ultra)filter on Y .

Definition. A filter F on a topological space Y converges to a point y ∈ Y if for all open

sets U containing y, U ∈ F.

Now we can define generalized limit points of maps. With this definition, the usual notion
of convergence of a sequence is F-convergence, where F is the cofinite filter on N.

Definition. Let X be a set, F an filter on X, Y a topological space, and f : X → Y a map

of sets. Then y ∈ Y is an F-limit point of f if f∗ F converges to Y .

Filter convergence works best with ultrafilters on compact Hausdorff spaces, as illustrated
by the following theorem.

2
Theorem (Ultrafilter convergence theorem). Let Y be a topological space.
1. Y is compact if and only if every ultrafilter F on Y converges to at least one point.

2. Y is Hausdorff if and only if every ultrafilter F on Y converges to at most one point.

Proof. 1. Suppose for contradiction that Y is compact, but F has no limit points.SThen
for all y ∈ Y , there is an open
S set Uy containing y such that Uy ∈/ F. So Y = y Uy ,
and by compactness, Y = ni=1 Uyi . But Y ∈ F, so some Uyi ∈ F, contradiction.
S
Conversely, suppose that Y isTnot compact. Then there is an open cover Y = i Ui
with no finite subcover. So i (Y \ Ui ) = ∅, but no finite subintersection is empty.
Then {(Y \ Ui )}i has the FIP, so we can take an ultrafilter F containing it. Now for
any point y ∈ Y , y is contained in some Ui , and Ui ∈
/ F, since (Y \ Ui ) ∈ F. So y is
not a limit point of F.

2. Suppose for contradiction that Y is Hausdorff, but y 6= y 0 are both limit points of F.
Take disjoint open sets y ∈ U , y 0 ∈ U 0 . Now U, U 0 ∈ F, but U ∩ U 0 = ∅, contradition.
Conversely, suppose that Y is not Hausdorff. Then there are points y 6= y 0 such that
every open neighborhood of y intersects every open neighborhood of y 0 . This means
that {U | y ∈ U open} ∪ {U 0 | y 0 ∈ U 0 open} has the FIP. Let F be an ultrafilter
containing it. Then y and y 0 are both limit points of F.

So if f : X → Y is a function from a set to a compact Hausdorff space, f has a unique

F-limit for every ultrafilter F on X. If F is the principal filter generated by x ∈ X, then
the F-limit of f is f (x).

Application: Stone-Cech compactification

The Stone-Cech compactification of a set X is a compact Hausdorff space βX along with
a map of sets η : X → βX satisfying the following universal property: Given a compact
Hausdorff space Y and a map of sets f : X → Y , there is a unique continuous map φ :
βX → Y such that φ ◦ η = f .
η
XC / βX
CC
CC
C ∃! φ
f CC
! 
Y
For those initiated into category theory, the Stone-Cech compactification is a functor
β : Set → CHaus which is left-adjoint to the forgetful functor CHaus → Set, and βX can
be seen as the free compact Hausdorff space on X. As usual, βX is unique up to unique
isomorphism.
The construction is as follows.
Underlying set: βX = {F | F is an ultrafilter on X}.
Topology: Given A ⊆ X, let UA = {F | A ∈ F}. The sets UA are a basis (of clopen sets).
The universal map η: x 7→ the principal ultrafilter generated by {x}.
Given f : X → Y , define φ : βX → Y by F 7→ the F-limit of f .

3
Exercise 4. Check the details.
• βX is a compact Hausdorff space for all X. In fact, it is a Stone space: compact,
Hausdorff, and totally disconnected.

• β is a functor (given a map g : X → X 0 , there’s only one reasonable choice for

βg : βX → βX 0 taking ultrafilters on X to ultrafilters on X 0 )

• The diagram commutes.

• φ is continuous.

• φ is unique.
Note that since every ultrafilter on a finite set is principal, if X is finite, βX = X with
the discrete topology.

Application: Tychenoff ’s theorem

The convergence theorem allows us to give a snappy proof of Tychenoff’s theorem.
Theorem (Tychenoff’s theorem). A product of compact spaces is compact.
Q
Proof. Let hXi ii be a collection of compact spaces. To show that X = i Xi is compact, it
suffices to show that every ultrafilter F on X has a limit point.
Let πi : X → Xi be the ith projection map. Then (πi )∗ F is an ultrafilter on Xi , and since
Xi is compact, it has a limit point xi ∈ Xi . I claim that x = (xi ) ∈ X is a limit point of F.
The topology on X is generated by sets of the form V = πi−1 (U ), where U is an open set
in Xi . So any open set containing x contains a finite intersection of sets of this form which
contain x. Since F is closed under supersets and finite intersections, it suffices to show that
if x ∈ V = πi−1 (U ), then V ∈ F.
Suppose x ∈ V . Then xi ∈ U , so U ∈ (πi )∗ F, since xi is a limit point of (πi )∗ F. But
then by definition of (πi )∗ , V ∈ F. This completes the proof.

3 Ultraproducts
Application: Prime ideals in products of rings
Exercise 5. In a finite product of rings ni=0 Ai , all prime ideals have the following form:
Q
{(a0 , . . . , ai , . . . , an ) | ai ∈ p}, where p is a prime ideal in Ai for some i. Hence a prime
ideal Q in the product `n is determined by a choice of a prime idea in one of the factors, and
Spec( ni=0 Ai ) ∼ = i=0 Spec(Ai ) (you can think of this as an isomorphism of sets, topological
spaces, or schemes).
The situation is not so simple for an infinite product of rings. For example, if {Ai | i ∈ I}
Q infinite collection of rings, the set {(ai ) | ai = 0 for all but finitely many i} is an ideal
is an
in i∈I Ai . It can be extended to a maximal ideal, which cannot be contained in any prime
ideal of the form described in the exercise.

4
 This situation is analagous to one we’ve seen before: on a finite set, all ultrafilters are
principal. But on an infinite set, there are other ultrafilters, obtained by extending the
cofinite filter. One can give a complete description of the primes in an infinite product of
rings using ultrafilters, but to keep things simple, we’ll focus on products of fields here.
Q
Theorem. Let {Fi | i ∈ I} be a collection of fields. The prime ideals in the ring i∈I Fi
are in bijection with the ultrafilters on I. The ultrafilter F corresponds to the prime ideal
{(ai ) | the set of indices i such that ai = 0 is in F}. In the same way, the proper ideals in
the ring are in bijection with the filters on I.
Q
Proof. Given an element a = (ai ) ∈ i Fi , let Za = {i ∈ I | ai = 0}. Notice that Zab =
Za ∪ Zb , and Za+b ⊇ Za ∩ Zb .
Let F be a filter on I. We claim that p = {a | Za ∈ F} is a proper ideal. Indeed, 0 ∈ p
since I ∈ F, and 1 6∈ p since ∅ 6∈ F.
Suppose a, b ∈ p. Then Za , Zb ∈ F. Now Za+b ⊇ Za ∩ Zb , so Za+b ∈ F, and a + b ∈ p.
Suppose a ∈ p and c is another ring element. Then Zca = Zc ∪ Za ⊇ Za , so Zca ∈ F, and
ca ∈ p.
We have established that p is an ideal. Suppose now that F is an ultrafilter, and let a
and b be ring elements such that ab ∈ p. Now Zab = Za ∪ Zb is in F. Then Za ∈ F or
Zb ∈ F, so a ∈ p or b ∈ p, and hence p is prime.

Conversely, let p be a proper ideal. If Zb = Za , then a and b differ by a unit: define

c = (ci ) by ci = 1 if i ∈ Za and ci = bi /ai otherwise. Then ac = b. So a ∈ p if and only if
b ∈ p, and p is completely determined by the set Fp = {Za ⊆ I | a ∈ p}, as p = {a | Za ∈ Fp }.
We’ll check that Fp is a filter on I.

1. I ∈ Fp , since 0 ∈ p.

2. ∅ ∈ Fp , since 1 6∈ p.

3. Suppose A ∈ Fp and A ⊆ B. Then there is an element a ∈ p such that Za = A. Define

the element c by ci = 0 if i ∈ B and ci = 1 otherwise. Then Zca = A ∪ B = B, and
ca ∈ p, so B ∈ Fp .

4. Suppose A, B ∈ Fp . Take a, b ∈ p with Za = A and Zb = B. Then Za+b ⊇ A ∩ B.

This may be a proper containment because of unexpected cancellation in some indices
i such that bi = −ai 6= 0. To remedy this problem, multiply b by c with ci = 0 for
the problem indices i and cj = 1 for the other indices j. Then Za+cb = A ∩ B, and
a + cb ∈ p, so A ∩ B ∈ Fp .

5. We have established that Fp is a filter. Suppose now that p is a prime ideal, and let
A be any subset of I. Let a and b be ring elements such that Za = A and Zb = I \ A.
Then ab = 0 ∈ p, so one of a or b is in p, and hence one of A or I \ A is in Fp , so Fp is
an ultrafilter.

5
Syntax and semantics
The ultraproduct construction is a very general method for putting together algebraic struc-
tures to obtain structures with specified properties. To speak in the correct level of generality,
we need a quick introduction to the language of universal algebra and model theory.
Definition. A language L is a set of symbols. Each symbol is specified to be a constant
symbol, a function symbol (of a specified finite arity), or a relation symbol (of a specified
finite arity). An L-structure is a set M , together with interpretations of the symbols in L:
• For each constant symbol c, an element cM ∈ M .

• For each function symbol f of arity n, a function f M : M n → M .

• For each relation symbol R of arity m, a subset RM ⊆ M m .

Examples. • Posets and linear orders are L≤ -structures, where L≤ = {≤}, and ≤ is a
binary relation.

• Rings and fields are Lr -structures, where Lr = {+, −, ·, 0, 1}, + and · are binary
functions, − is a unary function, and 0 and 1 are constants.
There is an obvious notion of homomorphism of L-structures (a map of underlying sets
which preserves the interpretations of the symbols). There is also a natural way to define
a product of L-structures, which, as one might hope, gives the categorical product in the
category of L-structures.
But this situation is unsatisfying: the class of L-structures is too broad. At this point,
we have no way of specifying that an Lr -structure is a ring, or that ≤ is interpreted as an
order in an L≤ -structure.
In order to impose more interesting conditions on our structures, we need more expressive
syntax. We’ll use the syntax of first-order logic.
Definition. Fix a language L. A term is built up from the constant symbols of L and a
supply of variables x0 , x1 , x2 , . . . by applications of the function symbols of L.
An atomic formula is of the form t1 = t2 or R(t1 , . . . , tn ), where the ti are terms and R
is a relation symbol of arity n.
A formula is built up from atomic formulas by Boolean connectives ∧, ∨, and ¬, and
quantifiers ∀xi and ∃xi .
A sentence is a formula with no free variables. That is, every variable appearing in the
formula is quantified.
Sentences in a language L have a familiar semantics in L-structures, in which they express
properties of these structures. In particular, given a sentence φ and an L-structure M , φ is
either true or false in M . In the same way, formulas with free variables express properties
of elements. Given a formula φ(x) with free variable x and an element a ∈ M , φ(a) is either
true or false in M .
For example, the following sentence asserts that the interpretation of ≤ is a transitive
relation:
∀x1 ∀x2 ∀x3 (x1 ≤ x2 ∧ x2 ≤ x3 ) → x1 ≤ x3 .

6
 Here φ → ψ is being used as a convenient abbreviation for ¬φ ∨ ψ.
The following formula in the free variable x1 expresses that x1 has a multiplicative inverse:

∃x2 x1 · x2 = 1.

Definition. If the sentence φ is true in the L-structure M , we say that M satisfies φ and
write M |= φ.
A theory is a set of sentences in a language L.
If T is a theory and the L-structure M |= φ for all φ ∈ T , then M is a model for T ,
written M |= T .

Exercise 6. Write down the field axioms as a first-order theory. Now find a theory whose
models are exactly the algebraically closed fields.

Note that the definition of formulas has an inductive structure: each formula is built
from strictly simpler formulas. This property allows us to conduct arguments by induction
on the complexity of formulas.
Since φ ∨ ψ is equivalent to ¬(¬φ ∧ ¬ψ) and ∀x φ(x) is equivalent to ¬∃x ¬φ(x), it suffices
to consider formulas built up by ¬, ∧, and ∃x.

Ultraproducts and Loś’s theorem

Now that we have the syntax of first-order logic at our disposal, we can talk about the class
of models for any theory. Unfortunately, given an arbitrary theory, its class of models may
lack nice categorical properties. For example, the class of fields does not have products,
coproducts, initial or final objects, etc.
As a compromise, there is a nice class of theories, the algebraic theories, which give rise to
well-behaved categories. The study of algebraic theories and their models is called universal
algebra, and it’s really a delight!
But what we’re interested in here is the ultraproduct construction, which will allow us
to build models for arbitrary (consistent) first-order theories.

Definition. Let {Mi | i ∈ Q I} be a collection of L-structures, and let F be an ultrafilter on

I. The ultraproduct M = I Mi /F is an L-structure defined as follows:
Q
• The underlying set is I Mi / ∼, where (ai ) ∼ (bi ) if they agree on a large set, i.e.
{i ∈ I | ai = bi } ∈ F. We will denote the equivalence class of an element (ai ) by [(ai )].

• If c is a constant symbol, cM = [(cMi )].

• If f is a function symbol of arity n, f M ([(ai1 )], . . . , [(ain )]) = [(f Mi (ai1 , . . . , ain ))].

• If R is a relation symbol of arity n, ([(ai1 )], . . . , [(ain )]) ∈ RM if and only if {i ∈

I | (ai1 , . . . , ain ) ∈ RMi } ∈ F.

Exercise 7. Check that this is well-defined: that ∼ is an equivalence relation and that the
interpretations of f and R are independent of the choice of representative for each equivalence
class.

7
Theorem (Loś’s theorem). Let {Mi | i ∈ I} be a collection of L-structures, and let F be an
ultrafilter on I. Let φ(x) be a first-order Q
formula in the free variables
Q x, and let [(ai )] be a
tuple of elements from the ultraproduct I Mi /F |= φ. Then I Mi /F |= φ([(ai )]) if and
only if {i ∈ I | Mi |= φ(ai )} ∈ I.

Exercise 8. Prove Loś’s theorem. The proof is by induction on the complexity of φ.

As a consequence of Loś’s theorem, if {Mi | i ∈ I} is a collection of models of some first-

order theory T , then their ultraproduct will also be a model of T . So an ultraproduct of
fields is a field, and ultraproduct of linear orders is a linear order, etc.

Exercise 9. Let F be the principal ultrafilter on I generated by j ∈ I. Show that

Q
Mi /F ∼
= Mj .
I

Exercise 10. Let {Fi | i ∈ I} be a collection of fields. Explain the connection between
the ultraproduct of these fields with respect to the ultrafilter F and the prime ideal in the
product of these fields corresponding to F identified in the previous section.

Application: Building funny structures

Given an L-structure M , the theory of M , written Th(M ), is the set of all sentences true in
M . One of the most famous applications of ultraproducts is the construction of nonstandard
models for Th(R).
Consider R as a structure in the language Lr . Let {R | i ∈ N} be a countable collection Q
of copies of R, and let F be a nonprincipal ultrafilter on N. The ultraproduct R = i R/F
is called an ultrapower of R, since all the factors are the same. Note that since each factor
satisfies Th(R), R |= Th(R).
But R has “infinitesimal” elements. Consider the element  = [(1, 21 , 13 , 14 , . . .)] ∈ R. For
any integer n, R |=  < n1 , since the number of factors in which this is true is cofinite, and
hence in F.
Abraham Robinson was able develop the theory of calculus in fields like R by dispensing
with the limit definition of derivative and interpreting the concept of infinitesimal change
quite literally. His theory is called “Nonstandard Analysis”. It is important to note that while
an ultrapower like R satisfies the same first-order sentences as R, many of the important
properties of R are second-order in nature, that is, they quantify over subsets of the domain,
not just elements. In order to get his hands on enough of the second order properties
to develop analysis, Robinson actually worked in an expansion of R with extra elements
representing subsets of the line, sets of subsets of the line, etc.

Exercise 11. Use the ultrapower construction to exhibit a nonstandard model for Th(N),
i.e. a semiring which satisfies all the same first-order statements as N in the language of
rings, but which has infinite elements.

Exercise 12. Use the ultraproduct construction to construct a field of characteristic 0 which
has exactly one algebraic extension of each degree.

8
Application: The compactness theorem of first-order logic
The compactness theorem is a powerful tool for showing that first-order theories are sat-
isfiable, i.e. that structures with certain properties exist. There are several proofs of this
theorem, of which the ultraproduct proof given here is the slickest.
Theorem (Compactness theorem). A theory T is satisfiable if and only if every finite subset
of T is satisfiable.
Proof. If T is satisfiable, then there is a model M |= T . Now M |= ∆ for any finite ∆ ⊆ T ,
so every finite subset of T is satisfiable.
Conversely, suppose every finite subset of T is satisfiable. Then we have a collection of
structures {M∆ | ∆ ∈ I} indexed by the collection of all finite subtheories of T , with M∆ |= ∆
for all ∆ ∈ I.
The idea is to take the ultraproduct of Q the M∆ with respect to some ultrafilter on I, in
such a way that the ultraproduct M = I M∆ /F is a model for T . By Loś’s theorem, it
suffices to find an ultrafilter F such that for all φ ∈ T , {∆ | M∆ |= φ} ∈ F.
Now certainly {∆ | M∆ |= φ} ⊇ {∆ | φ ∈ ∆}, since M∆ |= ∆. So we just need to pick F
so that Aφ =T {∆ | φ ∈ ∆} ∈ F for all φ ∈ T . We can do this if {Aφ | φ ∈ T } has the FIP.
And it does: ni=0 Aφi 6= ∅, since it contains the finite theory {φ0 , . . . , φn }.
To apply the compactness theorem, just write down any theory you like. If you can show
that there are no inconsistencies arising from finite pieces of the theory, the entire theory
has a model. Write down what you want, and then get what you want!
Exercise 13. For each of the examples in the “Building funny structures” section, show
how to use the compactness theorem to prove that a structure with the desired properties
exists.

4 Two more perspectives on ultrafilters

Ultrafilters on Boolean algebras
A Boolean algebra is an algebraic structure with two binary operations (∧, ∨), one unary
operation (¬), and two distinguished elements (1, 0), satisfying the rules of classical propo-
sitional logic.
Given any set X, the power set algebra (P(X), ∩, ∪,c , X, ∅) is a Boolean algebra. This
fact leads to an observation and a question.

Observation: The definition of an ultrafilter on X does not refer to the elements of

X, only to the Boolean algebra operations on the powerset algebra (A ⊆ B is equivalent to
A ∪ B = B). So we can extend the definition in a natural way to consider ultrafilters on
arbitrary Boolean algebras, not just on powerset algebras.
In fact, if we view the elements of a Boolean algebra B as logical propositions, an ultra-
filter on B is just a consistent assignment of truth values to the propositions (the elements
in the ultrafilter are true, the others are false). Equivalently, it is the preimage of 1 (true)
under a Boolean homomorphism to the 2 element Boolean algebra, B → {0, 1}.

9
 Question: Is every Boolean algebra isomorphic to the powerset algebra of a set X?

The answer to the question is no, but the same idea that led to the construction of the
Stone-Cech compactification can lead us to a positive result: every Boolean algebra can be
embedded as a subalgebra of a powerset algebra.

Theorem (Stone representation theorem). There is an equivalence of categories β : Bool →

Stone between the category of Boolean algebras and the category of Stone spaces (compact,
Hausdorff, totally disconnected topological spaces with continuous maps).
The functor β takes a boolean algebra B to the space of ultrafilters on B, with a basis for
the topology given by (the clopen sets) Ux = {F | x ∈ F} for all x ∈ B. B can be recovered
from βB as the Boolean algebra of clopen sets in βB, so B is a subalgebra of P(βB).

Note that if B is the powerset algebra of the set X, then βB is the Stone-Cech compact-
ification of X.

Measurable cardinals
Another way to view ultrafilters is as finitely additive 0-1-valued measures on the σ-algebra
P(X) (assigning measure 1 to “large” subsets and measure 0 to “small” subsets).
Since the usual convention in measure theory is to consider countably additive measures,
it is natural to ask whether we can find countably additive 0-1-valued measures. These would
be ultrafilters satisfying the countable intersection property (CIP), instead of the FIP.
It turns out that any ultrafilter with the CIP in fact must be an ultrafilter on a set X
of cardinality κ and satisfy the (<κ)IP (any intersection of fewer than κ many subsets is
nonempty), where κ is a measurable cardinal.
Measurable cardinals are one of the mysterious large cardinals studied in set theory. They
must necessarily be much much larger than any of the cardinalities encountered in everyday
mathematics, so large that their existence cannot be proven in ZFC. Think of the comparison
of cardinals like ℵ0 and 2ℵ0 with κ as being analagous to the comparison of cardinals like 2
and 47 with ℵ0 .
In fact, the situation is not just that we cannot prove measurable cardinals exist: we
cannot prove that the consistency of ZFC implies the consistency of ZFC + “there exists
a measurable cardinal” (so adding this axiom could make the system inconsistent), and
moreover, we can prove that such a thing cannot be proven. It’s entirely possible that ZFC
proves that there are no measurable cardinals.
But we have no such proof, and most set theorists believe that this and other “large
cardinal axioms” are not only consistent, but also philosophically motivated, and provide an
extension of the ZFC set theory which is too elegant and well-structured to be ignored. But
that’s another story...

10