Math 145a Lecture Notes

This is Harvard College’s Math 145a, instructed by Peter Koellner. The

formal name for this class is “Set Theory I”.

Contents

1 September 2, 2014 5
1.1 Broad Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.2 Introduction to Independence . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.3 Example of Independence . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.4 More Surprising Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.5 Escaping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.6 Models and Compactness . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.7 Looking Ahead . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.8 So what happens in this class? . . . . . . . . . . . . . . . . . . . . . . . . 8

2 September 4, 2014 9
2.1 Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.2 Subsets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.3 Cantor’s Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.4 Informal comprehension principle . . . . . . . . . . . . . . . . . . . . . . . 10
2.5 Empty set . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.6 Well-Orderings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.7 Ordinals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.8 The Hierarchy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

3 September 9, 2014 13
3.1 Housekeeping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.2 Last Time . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.3 Quasi-philosophical interlude . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.4 Categoricity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.5 Language of Set Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.6 The First Five Axioms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.7 The Sixth Axiom, Infinity . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.8 The Seventh Axiom, Foundation . . . . . . . . . . . . . . . . . . . . . . . 16
3.9 The Last Infinitely Many Axioms, Replacement . . . . . . . . . . . . . . . 16

1
Evan Chen
[email protected] (Harvard College Contents
Fall 2014)

4 September 11, 2014 17

4.1 Defining the nonnegative integers . . . . . . . . . . . . . . . . . . . . . . . 17
4.2 Comprehension and Collection . . . . . . . . . . . . . . . . . . . . . . . . 17
4.3 Set abstraction and the “set of all sets” . . . . . . . . . . . . . . . . . . . 17
4.4 Some abuse of notation with sets and classes . . . . . . . . . . . . . . . . 18
4.5 Some more basic notions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
4.5.1 Intersection and Difference . . . . . . . . . . . . . . . . . . . . . . . 19
4.5.2 Ordered Pairs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
4.5.3 Products . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
4.5.4 Relations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
4.5.5 Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
4.6 Philosophy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
4.7 Axiom of Choice . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

5 September 16, 2014 21

5.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
5.2 Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
5.3 Transfinite Induction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
5.4 Transfinite Recursion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
5.5 Transfinite things can be BIG . . . . . . . . . . . . . . . . . . . . . . . . . 23

6 September 18, 2014 25

6.1 What does a well-ordering look like? . . . . . . . . . . . . . . . . . . . . . 25
6.2 Well-orderings look roughly the same . . . . . . . . . . . . . . . . . . . . 25
6.3 Stuff with classes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
6.4 How to preserve transitive sets . . . . . . . . . . . . . . . . . . . . . . . . 27
6.5 Ordinals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

7 September 23, 2014 29

7.1 Recap . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
7.2 Order Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
7.3 Arithmetic on Ordinals . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
7.3.1 Addition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
7.3.2 Multiplication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
7.3.3 Exponentiation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
7.3.4 A nontrivial theorem . . . . . . . . . . . . . . . . . . . . . . . . . . 31
7.4 The hierarchy of sets, done with ordinals . . . . . . . . . . . . . . . . . . 31
7.5 The ordinals grab all the sets . . . . . . . . . . . . . . . . . . . . . . . . . 32

8 September 25 33
8.1 Axiom of Choice is equivalent to WO . . . . . . . . . . . . . . . . . . . . 33
8.2 Zorn’s Lemma . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
8.3 Some weakenings of AC . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34

9 September 30, 2014 35

9.1 Equinumerous sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
9.2 Creating cardinals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
9.3 Properties of cardinals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
9.4 Recurse . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

2
Evan Chen
[email protected] (Harvard College Contents
Fall 2014)

10 October 2, 2014 38
10.1 Cardinal arithmetic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
10.2 Finite sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
10.3 Infinite cardinal arithmetic is boring . . . . . . . . . . . . . . . . . . . . . 39
10.4 Cardinal exponentiation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
10.5 Fundamental theorems of cardinal arithmetic . . . . . . . . . . . . . . . . 40
10.6 Some exponentiation is not interesting . . . . . . . . . . . . . . . . . . . . 40

11 October 7, 2014 41
11.1 Cofinality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
11.2 Lemmas on cofinality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
11.3 Given AC, cofinalities and successor cardinals are regular . . . . . . . . . 42
11.4 Inaccessible cardinals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44

12 October 9, 2014 45
12.1 Linear Orders . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
12.2 Intermission . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
12.3 Separable complete linear orders . . . . . . . . . . . . . . . . . . . . . . . 46
12.4 Suslin property . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47

13 October 14, 2014 49

13.1 Sparse Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49

14 October 16, 2014 50

14.1 Partitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
14.2 Infinite Ramsey Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
14.3 Some negative results on partitions . . . . . . . . . . . . . . . . . . . . . . 51
14.4 Large Cardinals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
14.5 Trees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53

15 October 21, 2014 54

15.1 Definability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
15.2 Fragments of the Universe . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
15.3 Definable sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55

16 October 23, 2014 57

16.1 Model theory notions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
16.2 Reflection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
16.3 Variations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59

17 October 28, 2014 60

18 October 30, 2014 61

18.1 Completing the proof of a theorem . . . . . . . . . . . . . . . . . . . . . . 61
18.2 The statement Lκ  V = L . . . . . . . . . . . . . . . . . . . . . . . . . . 61
18.3 A Nice application of Reflection . . . . . . . . . . . . . . . . . . . . . . . 63
18.4 A sketch of the proof that Lκ  AC . . . . . . . . . . . . . . . . . . . . . 63

19 November 4, 2014 64
19.1 Axiom of Choice . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
19.2 GCH . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
19.3 Hulls . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65

3
Evan Chen
[email protected] (Harvard College Contents
Fall 2014)

19.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66

20 November 6, 2014 67
20.1 β-Gaps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
20.2 A Taste of Fine Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
20.3 The Guessing Sequence . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68

21 November 11, 2014 70

21.1 Problems with trying to set up Forcing . . . . . . . . . . . . . . . . . . . 70
21.2 Coding; good and bad reals . . . . . . . . . . . . . . . . . . . . . . . . . . 71
21.3 Ground Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71

22 November 13, 2014 72

22.1 The generic extension M [G] . . . . . . . . . . . . . . . . . . . . . . . . . . 72
22.2 Existence of generic extensions . . . . . . . . . . . . . . . . . . . . . . . . 72
22.3 Splitting posets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
22.4 Defining the generic extension . . . . . . . . . . . . . . . . . . . . . . . . 73
22.5 Defining M [G] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74

23 November 18, 2014 75

23.1 Review of Names . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
23.2 Filters of Transitive Models . . . . . . . . . . . . . . . . . . . . . . . . . . 75
23.3 Using the Generic Condition . . . . . . . . . . . . . . . . . . . . . . . . . 77

24 November 20, 2014 78

24.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
24.2 Semantics and Desired Conditions . . . . . . . . . . . . . . . . . . . . . . 78
24.3 Defining the Relation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
24.4 Consequences of the Definition . . . . . . . . . . . . . . . . . . . . . . . . 80

25 December 2, 2014 81
25.1 Forcing tells us that M [G] satisfies ZF . . . . . . . . . . . . . . . . . . . . 81
25.2 Forcing V 6= L is really easy . . . . . . . . . . . . . . . . . . . . . . . . . . 82
25.3 Forcing ZFC + ¬CH . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83

26 December 4, 2014 84
26.1 The Countable Chain Condition . . . . . . . . . . . . . . . . . . . . . . . 84
26.2 Possible Values Argument . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
26.3 Preserving Cardinals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
26.4 Infinite Combinatorics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
26.5 Finishing Off . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
26.6 Concluding Remark . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87

4
Evan Chen
[email protected] (Harvard College 1 September 2, 2014
Fall 2014)

§1 September 2, 2014
§1.1 Broad Overview
The text for this course is Set Theory, by Koellner, which I just downloaded. The
grading for this course is 100% problem sets.
The course has three parts,

• an introduction to set theory,

• the constructible universe,

• the Solovay model and forcing.

The goal of the course is the second and third part, in which we prove some things are
independent. For example, CH is independent of ZFC.1
Today we’ll be talking about the notion of independence.

§1.2 Introduction to Independence

This is “chapter 0”, not yet in the text.
There are, very broadly, two kinds of mathematics.

1. Algebraic, which is not in general concerned with a specific structure. For example,
group theory or theory of rings, topology, and so on. These topics are not about
specific structures like N.

2. In contrast, non-algebraic stuff is concerned with a specific structure. For example,

number theory cares a lot about N. (In a bit, we will show that (N, 0, S, +×),
where S is a successor function, is unique up to isomorphism.) As another example,
analysis, which cares about (PN, ∈). Note that 2N ∼ R. As a third example,
functional analysis cares about (PPN, ∈).

Set theory is the second type, and we care about iterating P. Iterating n times gives Vn .

• Vn is finite combinatorics.

• Vω ∼ N

• Vω+1 ∼ PN.

§1.3 Example of Independence

Take the axioms of group theory: given (G, ·) we require

(1) ∀x, y, xy ∈ G

(2) ∀x, y, z we have x(yz) = (xy)z

(3) ∃1 ∈ G: ∀x, x · 1 = 1 = 1 · x

(4) ∀x ∈ G, ∃y ∈ G with xy = 1 = yx.

Consider (5), G is abelian (meaning xy = yx)

1
Axioms of infinity makes certain things more answerable. ZFC “can’t articulate large enough objects”.
We can introduce larger and larger infinities, much like a “name the larger number” game that small
children play.

5
Evan Chen
[email protected] (Harvard College 1 September 2, 2014
Fall 2014)

Fact. The “abelian” statement is independent of the group axioms (1) through (4).

Proof. There exist both abelian (Z) and non-abelian (D6 ) groups. Duh.

So this is not a big deal in algebra, because there are tons of groups, and no one was
trying to determine whether “THE group” was abelian.
On the other hand, independence in Euclidean geometry might be a bit more surprising
– we now know the parallel postulate is independent of the other four of Euclid’s axioms.2
That’s because we’re talking about ONE THING. Of course now we have like non-
Euclidean geometry, and geometry as a formal discipline became just like group theory.
So once we discovered the parallel postulate was independent of the other four, geometry
fractured into “formal” geometry and “applied” geometry.
Philosophical question now: do we have independence with “fixed” structures?

§1.4 More Surprising Examples

Let’s take number theory! This has a fixed structure, (N, 0, S).
The axioms of this structure are denoted PA, the Peano axioms. This again has five
axioms.

(1) 0 is a number.

(2) If x is a number, so is S(x).

(3) S(x) 6= 0 for any x.

(4) S(x) = S(y) =⇒ x = y.

(5) (Least Criminal) If 0 has property ϕ (denoted ϕ(0)), and ∀x : ϕ(x) =⇒ ϕ(Sx),
then ϕ(x) holds for all x.

The last axiom is the one that’s doing all the work – ∞ is packaged in the first four
axioms.
Again, the principle is that the resulting structure is FIXED. In the “second order”
form of the axioms, PA determines N up to isomorphism. So there “should” not be
undecidable things!

Question 1.1. Is PA complete?3 In other words, is it the case that for any sentence ϕ
(in the language of PA) that either PA ` ϕ or PA ` ¬ϕ?

Here PA ` ϕ means PA proves ϕ.

So the answer is Gödel’s incompleteness theorems – no.

Theorem 1.2 (Gödel’s First Incompleteness Theorem)

Assume that PA is consistent. Then there exists a ϕ such that PA can neither prove
ϕ nor ¬ϕ.

Remark. Gödel had to assume a tiny bit more. The above statement is Rösser’s version.

2
Note that Euclid’s axioms suck. There are twenty-one Tarski axioms.
3
Tarski’s geometry is complete – there is only “one” Euclidean geometry.

6
Evan Chen
[email protected] (Harvard College 1 September 2, 2014
Fall 2014)

Theorem 1.3 (Gödel’s Second Incompletness Theorem)

Assume PA is consistent. Then PA cannot prove its own consistency.

One can express “PA is consistent” in the language of Peano arithmetic. This is a lot like
the “liar” paradox: namely “this sentence is false”. Similarly, we construct a statement
“I’m not provable”.
This is like bad. Outside of PA we can see the truth of PA, but not from within itself.

§1.5 Escaping
It’s not enough to add Con(PA) to the system. Gödel’s Theorem generalizes as fol-
lows.

Theorem 1.4
For any extension T of PA, T cannot prove Con(T ).

§1.6 Models and Compactness

Theorem 1.5 (Gödel, Compactness)

Suppose T is a formal system such that every finite fragment of T has a model.
Then T has a model (i.e. it is consistent).

Consider the following: let us add a constant symbol ċ to the language of PA. Let T be
the following list of axioms:

• PA

• ċ 6= 0

• ċ 6= 1

• ċ 6= 2

• ...

Obviously every finite fragment of T has a model. So T has a model! Let’s call this
model
(M, 0, S, ċ).
The axioms of Peano arithmetic implies we have S(ċ), S(S(ċ)). That means there are
“strange” numbers in addition to this chain of fake numbers hanging out on the side.

§1.7 Looking Ahead

For set theory,we have a system, ZFC. By Gödel, we know if ZFC is consistent, then
ZFC cannot prove its own consistency.
However, we can find worse examples.

Question 1.6 (Continuum Hypothesis). Is there an infinite subset A ⊂ R such that

A 6∼ N and A 6∼ R?

7
Evan Chen
[email protected] (Harvard College 1 September 2, 2014
Fall 2014)

Theorem 1.7 (Gödel, 1938)

If ZFC is consistent, then ZFC cannot refute CH.

Theorem 1.8 (Cohen, 1965)

If ZFC is consistent, then ZFC cannot prove CH.

This is really terrible because CH does not refer to the system of axioms at all (unlike
Con(PA) or Con(ZFC)).
We also have the following fact.

Fact 1.9. Up to isomorphism, (R, <) is the unique dense linear ordering without endpoints
that is separable (i.e. has a countable dense subset).

Hint: it’s Q.

Remark. Separable implies every disjoint sequence of intervals is countable.

Question 1.10 (Suslin’s Hypothesis). Is the above theorem still true if we replace “sepa-
rable” with the weaker condition that “every disjoint sequence of intervals is countable”?

This is again independent of ZFC.

These two questions are pretty unfortunate because this time, we have picked the
question “before” the choice of axioms (namely ZFC).

§1.8 So what happens in this class?

1. Part 1 of the course will talk about ZFC.

2. Part 2 of the course (“the constructible universe”) will deal with Gödel’s proof that
ZFC cannot refute CH.

3. Part 3 of the course (“forcing”) will deal with Cohen’s proof that ZFC cannot prove
CH.

8
Evan Chen
[email protected] (Harvard College 2 September 4, 2014
Fall 2014)

§2 September 4, 2014
This class is pretty flexible. Anyways, here is an informal introduction to the notions of
set theory.

§2.1 Sets
First we contrast sets and concepts. Examples of concepts are:

• “is red” – applies to red things

• “is a human being”

• “is a featherless biped”

These concepts apply to various things. By the way, the two latter things apply to the
same thing, Homo Sapiens. However, they are different concepts.
Sets do NOT have this feature.

{Adam, Eve, . . . } = {featherless bipeds} .

Hence given the set, we cannot tell what concept created it. In other words:

Sets are determined by their elements, while concepts are not determined by
the things falling under them.

Remark 2.1. Sets are extensional, which means

A = B iff ∀x (x ∈ A ⇐⇒ x ∈ B) .

This is the formalization of the above notion.

§2.2 Subsets
Consider N = {0, 1, 2, 3, . . . , }. We can consider subsets – write

A ⊂ B iff ∀x (x ∈ A =⇒ x ∈ B) .

Let E = {0, 2, 4, . . . } ⊂ N and O = {1, 3, 5, . . . } ⊂ N. Note that E ∩ O = ∅ = {}.

The key operation is considering all subsets. Consider all subsets N, including E and
O. Notice that E and O are definable. Unfortunately there are only countably many
definable sets (while 2N ∼ R is uncountable). However, we are still interested in the
undefinable sets.
Put
P(N) = {A : A ⊆ N}
as the power set of N. (Here P(X) is defined analogously.)

§2.3 Cantor’s Theorem

Some useful notions (er, more like review):

• Given f : A → B, we say f is onto or surjective if f (A) = B.

• Moreover, f is one-to-one injective if it never collapses elements; ∀x, y ∈ A (x 6= y =⇒ f x 6= f y).

• Finally, f is bijective if f is both one-to-one and onto.

9
Evan Chen
[email protected] (Harvard College 2 September 4, 2014
Fall 2014)

Definition 2.2. Two sets A and B have the same size if and only if ∃f : A → B which
is one-to-one and onto.

This lifts this criterion up to infinite sets. If you leave a knife and fork for each person
at a table (possibly infinitely large) then there are the same number of knives and
forks.

Example 2.3
As above, E ∼ N ∼ O. Also, Q ∼ N.

Theorem 2.4 (Cantor)

N 6∼ R.

Proof. Hmm this is a new one. Assume for contradiction we can enumerate the reals as
x1 , . . . .
Consider the following sequence of nested closed intervals. Start with I0 = [0, 1]. Let
I1 = [a1 , b1 ] be such that b1 − a1 ≤ 11 and x1 ∈
/ [a1 , b1 ]. Then set I2 = [a2 , b2 ] ⊆ I1 with
b2 − a2 ≤ 21 and x2 ∈ / I2 . Continue.
This gives rise to a nested sequence of intervals

I0 ⊃ I1 ⊃ . . . .
T
By design, xn ∈
/ In for any n. However these intervals are all closed, so n≥0 In = {x}
for some x. Hence this x is not of the form xk , contradiction.

This generalizes as follows.

Theorem 2.5
For any set X, P(X) 6∼ X.

Proof. Barber paradox. Also on our Math 55 pset.

Now that the review is done, let’s state some philosophy.

The idea of P(X) is one of the two key ideas of set theory.

Now what’s the second idea?

§2.4 Informal comprehension principle

The informal comprehension principle states the following: if X is a set and P as a
definite4 property of elements in X, then there exists a set Y which contains all and only
the elements P holds for. In other words,

Y = {x ∈ X : P (x)}

is a set. Another name is the “informal separation principle”.

4
An example of non-definite property is “is bald”. The hallmark of a vague property is that it doesn’t
obey induction. If a person with n hairs is bald, is someone with n + 1 hairs bold?

10
Evan Chen
[email protected] (Harvard College 2 September 4, 2014
Fall 2014)

Example 2.6
Take X = N. Then D is even.

Remark 2.7. Note that this does NOT imply that all sets are generated by compre-
hension. Indeed, if there are only countably many properties, so some subsets of N are
necessarily not generated by properties.

§2.5 Empty set

Definition 2.8. Define ∅ = {x : x 6= x}.
The empty set has the very important property that it is a subset of everything.
We could have started with N. Specifically,
• N is where all number theory happens.

• P(N) ∼ R is where all algebra happens.

• P(P(N)) ∼ {f : R → R} is where all algebra happens.

But we start lower. We have

V1 = P(∅) = {∅} = {{}}

V2 = P(P(∅)) = {∅, {∅}}
V3 = P(P(P(∅))) = {∅, {∅} , {{∅}} , {∅, {∅}}} .

Notice that in general, the nth iteration is a power tower of 2’s:

 5  2
P (∅) = 222 = 216 = 65536.

These numbers get quite big quite quickly.

Let’s give these things a name.
Definition 2.9. Let V0 = ∅ and Vn+1 = P(Vn ).
Despite how large these numbers are, the collection of {Vn }n is pretty countable. Define
[
Vω = Vn .
n<ω

Then N ∼ Vω , R ∼ Vω+1 , and we recover our ladder.

Note that Cantor’s Continuum Hypothesis is that there are no cardinalities between
|Vω | and |Vω+1 |.

§2.6 Well-Orderings
Definition 2.10. Suppose ≺ is a binary relation (whatever that means) on a set X.
Then ≺ is well-founded if and only if every non-empty subset Y of X has an ≺-minimal
element. Moreover, ≺ is well-ordered if it is well-founded, strict, linear.

Example 2.11
The relation “is an ancestor of” on humans is well-founded. Similarly, < is well-
founded on N. However, > is not well-founded over N.

11
Evan Chen
[email protected] (Harvard College 2 September 4, 2014
Fall 2014)

§2.7 Ordinals
Are there larger examples?
Let ω be the next ordinal after all the nonnegative integers. If we want to be formal,
ω = {0, 1, 2, 3, . . . }. We can then posit ω + 1, and keep going. Thus we have
0 < 1 < 2 < ··· < ω < ω + 1 < ...
Another similar vein is re-ordering the integers as
0 < 2 < 4 < 6 < ··· < 1 < 3 < 5 < ...
which is certainly a valid ordering. These have the same order type – there is an obvious
order-preserving bijection between them. Thus we focus on the former, which are called
the ordinals.
Exercise. Check that these are well ordered! It’s an finite way down, but possibly an
infinite way up.
Let’s start talking about ordinals (keeping in mind everything up to now is still
informal).
0,1, 2, 3, . . . , ω
ω + 1, ω + 2, . . . , ω + ω
2ω + 1, 2ω + 2, . . . , 3ω
..
.
ω 2 + 1, ω 2 + 2, . . .
..
.
ω3, . . . , ω4, . . . , ωω
..
.,

Here I’m abusing notation. The notion 2ω should be ω + ω; the notion ω 2 should be
ω · ω, but you get the idea.
..
.
We finish with ω ω here ω many times, which is called ε0 . In general, εk+1 is εk stacked
k times. Then we get to εε0 , and so on. . .
Remark 2.12. Compute ω ε0 = ε0 . Actually, ε0 is the least α such that ω α = α.
Transfinite induction with ε0

§2.8 The Hierarchy

Let V be the universe of sets. Iterate P along all the well-orderings.
Later we’ll see the following:
• V0 = ∅
• Vα+1 = P(Vα ) (successor)
S
• Vλ = α<λ Vα if λ is a limit
S
• V = α∈On Vα .
Here On is the order types of everything. We’ll show soon On 6⊂ On and V ∈
/ V.

12
Evan Chen
[email protected] (Harvard College 3 September 9, 2014
Fall 2014)

§3 September 9, 2014
§3.1 Housekeeping
• Office hours: Tuesday 2PM - 3PM, 414 2 Arrow Street

• Book: A Course in Set Theory, by Ernest S

• Email professor: background (logic, ST, math) and what you would like to see

Exercises coming next week.

§3.2 Last Time

Subject matter: set theory hierarchy.
Three fundamental ideas:

• Well-ordering (ordinals)

• Power set

• Absolute infinity, which we’ll discuss more below.

The hierarchy V is built up in stages from ∅, by taking the power set of the previous
stage. This picture has shown up so many times I will actually draw it.

S
Vω = Vn

Vn+1 = P(Vn )
Vn

V2 = {∅, {∅}}
V1 = {∅}
V0 = ∅

ε0 : this is what forever really is. :D Except ε0 is supposedly “piddling”. darn.

§3.3 Quasi-philosophical interlude

Some people (potentialists) think that

• The power set of an infinite set (like N) is “open-ended”. We call these width
potentialists.

13
Evan Chen
[email protected] (Harvard College 3 September 9, 2014
Fall 2014)

• The height of the universe (ordinals) is “open-ended”. We call these height poten-
tialists.
Why would someone think 2N is open-ended? We showed that |P(N)| > |N| already.
Some people think the idea P(N) does not even make sense. The potentialists think
that subsets of N which are definable make sense but are skeptical of the notion of “all”
subsets of N.
Consider a binary branching tree which represents decisions on whether to include an
element of N. (Going right at the nth step means including n in.)
Anyways the opposite of a potentialist is an actualist.

§3.4 Categoricity
We now show the philosophy is a moot point. Suppose we have two hierarchies V0 , V1 , . . .
and V00 , V10 , . . . . Since N is isomorphic in these things, there as an isomorphism of Vω to
Vω0 . Now in my sense
Vω+1 = P V (Vω ) .
In your sense

0 0
Vω+1 = PV Vω0 .
But any A specified in Vω leads to a set of Vω0 .
OK I have no idea what this is. But it was informal. The point is going to be
categoricity. Regardless of what makes sense and what does not, apparently everything
is isomorphic?

§3.5 Language of Set Theory

OK whatever, let’s consider the universe V .
In number theory, we defined the integers through a successor function. The analog
here is the power set operator P.
We now enter the domain of Chapter 1, the language of set theory. Firstly it ni the
language of logic. This includes ∀, ∃, an infinite supply of variables, as well as negation
¬, ∧, ∨, and finally → and ↔. Moreover, we have a symbol =.
You can’t say too much with this; in Number Theory we solve this problem by defining
a few extra things, 0 and the successor function S, among other things. In set theory we
add only one symbol: the membership ∈. For example, consider

∀x∃y(x ∈ y).

This has important implications: we can already generate N by

x ∈ y0 ∈ y1 ∈ y2 ∈ . . .

Now we can prepare the axioms.

§3.6 The First Five Axioms

First, we distinguish a set from a concept.
Axiom I (Extensionality). A set is determined by its elements, meaning

∀x∀y∀z ((z ∈ x ↔ z ∈ y) → x = y) .

Now we need to begin somewhere.

14
Evan Chen
[email protected] (Harvard College 3 September 9, 2014
Fall 2014)

Axiom II (Empty Set). The empty set exists: ∃x∀y(y ∈

/ x).

By Extensionality, there is only one such empty set, and we denote it by ∅.

Axiom III (Pairing). Informally, for any x, y we can create {x, y}. Formally,

∀x∀y∃a∀z (z ∈ a ↔ (z = x ∨ z = y)) .

Again by Extensionality, this set is unique, and we denote it by {x, y}.

S
Axiom IV (Union). Informally, we can take a0 ∈a . Formally,

∀x∃a∀z (z ∈ a ↔ ∃y (z ∈ y ∈ x)) .
S
Again by Extensionality (see a pattern here?), this is unique, so we call it x.

Example 3.1
S
If x = {N}, x = N. So that means using union, we can generate large sets from
small ones.

Now for the power set.

Axiom V (Power Set). We can construct P(x). Formally,

∀x∃a∀y(y ∈ a ↔ y ⊆ x)

where y ⊆ x is short for ∀z(z ∈ y → z ∈ x).

Again, we do not have a lot of symbols, so we define lots of shorthands like ⊆. The
important thing is that this in principle can be expanded. Anyways, we denote this by
P(x) since it is unique by Extensionality.

§3.7 The Sixth Axiom, Infinity

Note that the first five axioms are already sufficient to construct

P(P(. . . (∅))).

This lets us get up to the natural

S numbers.
Moreover, remark that Vω = n∈N Vn is a model for these axioms. Just check all the
axioms hold.
But Vω is countable, and all the sets are freaking finite. Boring! Let’s force infinitely
large sets.

Axiom VI (Infinity).

∃x (∅ ∈ x ∧ ∀y (y ∈ x → y ∪ {y} ∈ x)) .

This time we can’t conclude x is unique. But we can deduce some things:

∅ ∈ x =⇒ {∅} ∈ x
=⇒ {∅, {∅}} ∈ x
=⇒ {∅, {∅} , {∅, {∅}}} ∈ x
...

15
Evan Chen
[email protected] (Harvard College 3 September 9, 2014
Fall 2014)

Actually, each line is just the union of the objects on all the previous lines. This is a lot
like ω, where we define ω = {0, 1, 2, . . . , } and ω + 1 = {0, 1, 2, . . . , ω}. Basically, the way
we generate a new object is to take the set of all previous objects.
Anyways, x is not unique. We can throw in {{∅}} in or something, and suddenly we
get more junk in x. However, there is nonetheless a minimal x. Here’s how we get our
hands on it: Call a set x inductive if

∅ ∈ x ∧ ∀y (y ∈ x → y ∪ {y} ∈ x) .
T T
Then define ω = {A | A inductive}. (We’ll define later).

§3.8 The Seventh Axiom, Foundation

Now we will make induction work by ruling out infinite chains.

Axiom VII (Foundation).

∀x (x 6= ∅ → ∃y ∈ x∀z ∈ x(z ∈
/ y)) .

Equivalently, “there are no infinite descending ∈ chains”.

This is not too hard to see. Suppose

x1 3 x2 3 x3 3 . . .

Then take x = {xk }k≥1 , which fails Foundation. Conversely, if Foundation fails then it is
not hard to generate an infinite 3 chain.
Anyways, what is a model for this? If we take the universe Vω+1 = PVω , it turns out
that this satisfies Infinity, because Vω ∈ Vω+1 . However, it doesn’t satisfy Power Set
anymore, because P(Vω+1 ) is not in there. Nonetheless, we can extend this up to Vω+ω ,
and now we have a model!
You might ask why we don’t just stop at Vω+ω . Mainly, because we can. Like, FLT
was not proved in ZFC. If you give yourself more machinery, you can prove more stuff.

§3.9 The Last Infinitely Many Axioms, Replacement

ZFC does not have eight axioms, it has infinitely many! We just like to condense this as
replacement.

Axiom VIII (Replacement). For each formula ϕ in the Language of Set Theory, and
for every positive integer n,

∀p1 ∀p2 . . . ∀pn ∀a [∀y ∈ a ∃!z ϕ(y, z, p1 , p2 , . . . , pn )]

→ [∃b ∀z (z ∈ b ↔ ∃y ∈ a ϕ (y, z, p1 , . . . , pn ))] .

What is this saying? Really ϕ defines a function F (relative to the parameters pi ) on

domain a, then we can collect them together in a set b = F (a).
It turns out the Replacement does not hold in Vω+ω because there is no set

{Vω , Vω+1 , Vω+2 . . . } .

Actually, Replacement blows the top off this, making ε0 seem small. Oops.

16
Evan Chen
[email protected] (Harvard College 4 September 11, 2014
Fall 2014)

§4 September 11, 2014

Happy 18th birthday, Jessica!
In what follows, we will just abbreviate ϕ(x, p1 , . . . , pn ) as ϕ(x).

§4.1 Defining the nonnegative integers

Last time we introduced ZF and the language of set theory.
First, let us actually define

0=∅
1 = {∅}
2 = {∅, {∅}}
3 = {∅, {∅} , {∅, {∅}}}
S
and in general, n + 1 = {0, 1, . . . , n}. Then ω = n∈N n.

§4.2 Comprehension and Collection

Here are two more schemes, not part of ZFC.
Axiom (Comprehension). For any ϕ in LST,

∀a∃b∀x (x ∈ b ↔ x ∈ a ∧ ϕ(x)) .

Very loosely, this states that you can do something like

{x ∈ A | ϕ(x)}

Axiom (Collection). For any ϕ in LST,

∀a [∀y ∈ a∃!zϕ(y, z)] → ∃b∀y ∈ ∃z ∈ bϕ(y, z)

In other words, if ϕ defines a “function”.

Collection is “looser” than replacement. Given a set A, replacement states that it’s
possible to get f (A) exactly, “replacing” every element of a ∈ A with f (a). Collection
only lets you collect them in a big blob.
Exercise 4.1. In ZF without replacement, Comprehension and Collection together are
equivalent to replacement.
Sketch of Proof. It’s easy to see Comprehension and Collection together give Replacement.
Also, Replacement implies Collection. To see Replacement gives Comprehension, assume
B 6= ∅ and contrive a function.

§4.3 Set abstraction and the “set of all sets”

Because Comprehension holds in ZF, we can write the shorthand

{x ∈ a | ϕ(x)} .

This is called a set abstract The condition x ∈ a is crucial; it must be a restricted abstract.
Otherwise everything burns.
What happens when this restriction is dropped? The resulting expression may then
fail to denote a set.

17
Evan Chen
[email protected] (Harvard College 4 September 11, 2014
Fall 2014)

Proposition 4.2
The expression {x | x = x} is not a set.

Proof. Assume not, and let v denote it. Clearly v = v, so v ∈ v. Then

v ∈ v ∈ v ∈ ...

which breaks Foundation.

In other words, the universe V = {x | x = x} is not itself a set.

The way to think of this is that “I am in this room” is not the same as “this room is
contained in the physical universe”. The universe is not really a container; it is a “stage”
on which the actual sets lie.
Nonetheless, we will use this notation for non-sets as well. Just be careful. . . In the
above, V is not a set, it is a class. But for example, consider the expression

{z | ∃y (z ∈ {y} ∧ y ∈ a)} .

This is a set by replacement; roughly, consider a function x 7→ {x}. To write it as a

restricted comprehension, we must use

{z ∈ P(a) | ∃y (z ∈ {y} ∧ y ∈ a)} .

§4.4 Some abuse of notation with sets and classes

Despite the warning above, we will write

b ∈ {x | ϕ(x)}

which doesn’t necessarily make sense, since the RHS is not really a set, but this is
shorthand for ϕ(b). (Note we can’t replace ∈ with 3 here!)
Similarly, we may write
{x | ϕ(x)} = {x | ψ(x)}
as shorthand for
∀x (ϕ(x) ↔ ψ(x)) .
Similarly,
{x | ϕ(x)} ⊆ {x | ψ(x)}
is shorthand for
∀x (ϕ(x) → ψ(x)) .
Just distinguish with sets and classes. . .
That means we can write the following shorthands.

• Comprehension states that for all sets x and classes C, C ∩ x is a set.

• Replacement states that For all sets a and all class functions F , the image F (a)
is a set.

• Collection is similar to replacement but only gives you that F (a) is a subset of
some set.

18
Evan Chen
[email protected] (Harvard College 4 September 11, 2014
Fall 2014)

§4.5 Some more basic notions

4.5.1 Intersection and Difference
Definition 4.3 (Intersection). Given x, we define the intersection by
[
x = {z | ∀y ∈ x(z ∈ y)} .
S
This is a set because we can put z ∈ y.

Definition 4.4 (Difference). Given x and y, define the difference by

x \ y = {z ∈ x | z ∈
/ y} .

4.5.2 Ordered Pairs

We would like to define ordered pairs as well; {x1 , x2 } = {x2 , x1 } so we need a new
notion.

Definition 4.5. Let (x1 , x2 ) denote {x1 , {x1 , x2 }}.

Notice that (x1 , x2 ) 6= (x2 , x1 ) if x1 =

6 x2 , and (x, x) = {{x} , {x, x}} = {{x} , {x}} =
{{x}}.

Exercise 4.6. Show that (x1 , x2 ) = (y1 , y2 ) if and only if x1 = y1 and x2 = y2 .

4.5.3 Products
We can define the direct product.

Definition 4.7. Define

X × Y = {(x, y) | x ∈ X ∧ y ∈ Y } .
S
We claim this is a set. Just notice we can carve it out from P(P(X ∪ Y )).
Naturally, we can also define

X1 × X2 × · · · × Xn+1 = (X1 × · · · × Xn ) × Xn+1 .

We also write X n = X × X × · · · × X.

4.5.4 Relations
An n-ary relation R on a class X is a subclass of X n .
Suppose n = 2. The domain of R is

dom(R) = {x ∈ X | ∃y(x R y)} .

Similarly,the range is
ran(R) = {y ∈ Y | ∃x(x R y)} .
Then we let
field(R) = dom(R) ∪ ran(R).
We can define
R−1 = {(y, x) | (x, y) ∈ R} .
Finally, we say that R is

19
Evan Chen
[email protected] (Harvard College 4 September 11, 2014
Fall 2014)

• reflexive if ∀x ∈ X, x R x..

• symmetric if ∀x, y ∈ X, x R y → y R x.

• transitive if ∀x, y, z ∈ X we have x R y ∧ y R z → x R z.

It is an equivalence relation if all three are true. This gives us a partition into equivalence
classes [a]R .

Exercise 4.8. Show that

S
(i) X = {[a]R | x ∈ X}.

(ii) ∀a, b ∈ X (¬ (a R b) → [a]R ∩ [b]R = ∅)

4.5.5 Functions
A function is a binary relation R such that

∀x ∈ dom(R) ∃!y ∈ ran(R) : x R y.

We’ll use f and g, thanks. Also, for sets X and Y write

X
Y = {f | f : X → Y } .

Thus ω ω is the functions from N to N. We sometimes abuse notation if there is no

possibility of confusion and write this as Y X .
Restriction, images, inverses.

§4.6 Philosophy
Everything is a set. N is a set. R = P(N) is a set. And so on.
OK let’s move on.

§4.7 Axiom of Choice

Axiom IX (Choice). We have

∀x (∀y ∈ x (y 6= ∅)) → ∃f (f is a function, dom(f ) = x, ∀y ∈ x (f y ∈ y)) .

In other words, given a bunch of nonempty sets x = {y1 , y2 , . . . }, we can construct a

choice function f with f (yi ) ∈ yi for all i.
Note that this can be proven in the case x is finite. Let x = {y1 }, where y1 6= ∅. Look
at all functions f with domain x and f (y1 ) ∈ y1 .
In the case x = {y1 , y2 } with y1 , y2 6= ∅.
Then y1 × y2 6= ∅.
In other words, Choice states that y1 × y2 × . . . is nonempty, even in the infinite case.

Exercise 4.9. Show that Choice is equivalent to

∀x∀y∀f : x → y ∃(g : ran(f ) → x) such that g ⊆ f −1 .

20
Evan Chen
[email protected] (Harvard College 5 September 16, 2014
Fall 2014)

§5 September 16, 2014

Summary so far:
(1) informal introduction to the subject matter of set theory
• V
• two fundamental notions: well-ordering (stages) and power sets (what we do
to stages)

(2) axioms of set theory

Now, we redo our (1) formally using (2).
Recall (informally) our ordinals, 0, . . . , ε0 , . . . . We’re going to make this all rigorous
now.

§5.1 Motivation
Our motivation is that we are going to try and generalize the natural numbers ω. Recall
we have induction: if ϕ(0) and ∀n (ϕ(n) → ϕ(n + 1)) then ∀nϕ(n). This is not enough
to get up to ω. Now we’re going to generalize this massively a la “strong induction”.
We are also going to generalize recursion. Given a point p, if we can get f (p) from all
predecessors of p, then we have a recursion.
We’re going to do this vastly more generally, on well-founded rather than just well-
ordered relations.

§5.2 Setup
Let R be a binary relation on X.
Definition 5.1. We say R is strict or irreflexive if ∀x ∈ X, x R x does NOT hold. We
say R is linear if ∀x, y ∈ X either x R y or y R x.
Then for each x ∈ X, let

extR (x) = {y ∈ X | y R x} .

Example 5.2
Let R denote “is a parent of” and X be the set of humans. Then R is strict but not
linear. Moreover, extR (x) are the parents of x.

Definition 5.3. The transitive closure TC(R) is the intersection of all transitive
relations containing R.
By taking the intersection of them all, we get rid of the “junk” and get the smallest
possible.

Example 5.4
TC(“is a parent of ”) = “is an ancestor of”.

Definition 5.5. The predecessors are predR (x) = extTC(R) (x). A subset Y of X is
R-transitive if for all y ∈ Y , predR (y) ⊆ Y .

21
Evan Chen
[email protected] (Harvard College 5 September 16, 2014
Fall 2014)

Example 5.6
The set predR (Fred Koellner) ∪ predR (Obama) is R-transitive.

Definition 5.7. An element y ∈ Y is R-minimal if there is no z ∈ Y such that z 6= y

and z R y.
Then a relation R (on X, even if it is a class rather than a set) is well-founded if
(i) Every nonempty subset Y has an R-minimal element
(ii) ∀x ∈ X, extR (x) is a set.
Note that this last relation is obvious for X a set. However, we can consider X a class as
well, so long as locally everything looks like a set (for example, “set of all sets”).
Definition 5.8. A well-ordering is a well-founded, strict, linear relation.

§5.3 Transfinite Induction

Now we can generalize induction and recursion to well-founded relations. (If you look at
the proof of induction, neither “linear” nor “countable” is needed.)

Theorem 5.9 (Transfinite Induction)

Suppose R is a well-founded relation on X. If Y ⊆ X is such that

∀x ∈ X (predR (x) ⊆ Y → x ∈ Y )

then Y = X.

In other words, let Y be some subset of X. Suppose that any time the predecessors of x
are in Y , then x is in Y too. It follows that, actually, all of X is in Y .
The “base cases” are handled by predR (x) = ∅ for any R-minimal element x.
Proof. Suppose that Y 6= X and consider X − Y . Then let a be an R-minimal element
in X − Y . Now all the guys in predR (a) cannot be in X − Y , so they are contained in Y .
Contradiction.

§5.4 Transfinite Recursion

Likewise, we can also build objects on well-founded structures.

Theorem 5.10 (Transfinite Recursion)

Suppose R is a well-founded relation on X. Suppose

G:X ×V →V

is a class function. Then there’s a unique function

F :X→V

so that for all x ∈ X,

F (x) = G(x, F |predR (x) ).
Here F |predR (x) is a restriction.

22
Evan Chen
[email protected] (Harvard College 5 September 16, 2014
Fall 2014)

Here G is a “generating function” in the sense of the recursion. Our G takes in an x and
the predecessors, and spits out F (x). Intuitively, G has way, way more information than
we need.

Proof. First, we show uniqueness. Suppose F1 and F2 are two distinct such functions.
Use transfinite induction to show they are equal: consider the R-minimal point a with
F1 (a) 6= F2 (a). Then F1 |predR (a) = F2 |predR (a) , contradicting that G is a function.
For existence, take an R-minimal element x. Our G gives G(x, ∅), so now we know
F (x). Intuitively, it’s clear how to proceed.
Let us say a function f : D → V is good if

(i) D ⊆ X and ∀x ∈ D(predR (x) ⊆ D)

(ii) ∀x ∈ D we have f (x) = G(x, f |pred(x) ).

In other words, f is doing the right thing on a predecessor-closed subset of X.

First, suppose f1 : D1 → V and f2 : D2 → V . We claim f1 |D1 ∩D2 = f2 |D1 ∩d2 . The
proof is exactly the same as that for uniqueness.
The second claim is that [
{dom(P ) | f is good}

is X. Now suppose f : predR (x) → V is good for some x. Define f 0 : predR (x) ∪ {x} → V
by (
f (y) if y ∈ predR (x)
y 7→
G(x, f ) otherwise.
This is clearly good. So we have shown that
[ [
predR (x) ⊆ {dom(f ) | f is good} → x ∈ {dom(f ) | f is good} .

The second claim then

S follows by transfinite induction.
Finally, let F = {f | f good}. By Claim 1 this is a function. By Claim 2, dom(F ) =
X. Hence F is good.

§5.5 Transfinite things can be BIG

Let G : V × V → V , and let R be the membership relation ∈. For any set x, we have

ext∈ (x) = x.

Then pred∈ (x) consists of a very nested sequence of things. In particular, ∅ ∈ pred∈ (x).
It really goes ALL THE WAY DOWN.
A second example is
( S
P ( ran(f )) if α 6= ∅
G : On × V → V by (α, f ) 7→
∅ otherwise.

with R as <. The transitive recursion gives us an F such that for any α ∈ On we have

F (α) = G(α, Fα )

and you can check that in fact,

F (0) = F (∅) = G(∅, ∅) = ∅ = V0 .

23
Evan Chen
[email protected] (Harvard College 5 September 16, 2014
Fall 2014)

Oops okay this doesn’t actually work. OK, the correct version is

∅ S
 if α = ∅
G : On × V → V by (α, f ) 7→ P( ran(x)) if α is a successor

S
ran(x) if α is a limit.

24
Evan Chen
[email protected] (Harvard College 6 September 18, 2014
Fall 2014)

§6 September 18, 2014

Last time we looked at transfinite induction as well-founded relations.
Basically, a well-founded relation (X, R) is just one step shy of being a sub-chunk of
the ordinals. A well-founded relation looks a lot like a poset with a base. You can have
infinite ascending sequences, just not descending ones (limit points).
A special case is a well-orderings. The main topic for today.

§6.1 What does a well-ordering look like?

Imagine a poset with a base in which every two things are comparable. Oh wait, we have
a total order.
We can also add limit points.
Anyways, the basic claim is that any well-ordering is just isomorphic to sub-orderings
of the standard ordering of the ordinals.

§6.2 Well-orderings look roughly the same

Definition 6.1. Suppose (X, R) is a well-ordering. For x ∈ X, we define another
well-ordering

IxR = predR (x), R ∩ predR (x)2 .
Here IxR and R itself are called the initial segments of (R, X).

Intuitively, all we’re doing is cutting off our relation past x.

Remark. A different notation of a well-ordering is (A, <A ). We sometimes also abbreviate

(X, R) as just R.

The following proof is just a generalization of the proof that the natural numbers
are unique up to isomorphism. First, let us make sure we agree on the definition of
isomorphism:

Definition 6.2. Suppose (X, R) and (X 0 , R0 ) are well-orderings. We say that (X, R) ∼
=
0 0
(X , R ) are isomorphic if f is bijective.

Remark. We can weaken “bijective” to “surjective” because R0 is strict and R is linear.

6 x2 , then they are R-related, so f x1 and f x2 are R-related and hence
Explicitly, if x1 =
not equal.

Theorem 6.3
Suppose (X, R) and (Y, S) are well-orderings. Then either

(i) R ∼
= IyS for some y ∈ Y ,

(ii) IxR ∼
= S for some x ∈ X,

(iii) R ∼
= S.

Again, the proof is just to walk upwards.

25
Evan Chen
[email protected] (Harvard College 6 September 18, 2014
Fall 2014)

Proof. For A ⊆ X, f : A → Y and x ∈ X. We use transfinite recursion. Let

(
s-least element in Y − f (predR (x)) if possible
G(x, f ) = .
undefined otherwise.

OK this is pretty sloppy. But lol. We’re allowed to be sloppy because in set theory,
functions are just pairs, and so “undefined” means we just don’t add a pair. But
consequently, G might not have the full domain X.
But the point is that if we exhaust S, then we’re done, and if we haven’t run out of
things in S, we just keep going.
Let F be defined by transfinite recursion from G. By transfinite induction we have

∀x ∈ dom(F )F (predR (x)) = predS (F (x)) .

Also it’s easy to see F is order-preserving (both ways). Blah.

OK we finish by considering three cases.
(1) If dom(F ) = X and ran(F ) = Y , then R ∼
= S.
(2) If dom(F ) = X and ran(F ) 6= Y , then just take the minimal element y of Y −ran(F ),
then this must be equal to predS (y), thus R ∼ = IyS .

(3) The case dom(F ) 6= X is handle similarly and we obtain R ∼

= IxR .

§6.3 Stuff with classes

For a given well-ordering (X, R) consider the class of all well-orderings (X1, R0 ) such
that
(X 0 , R0 ) ∼
= (X, R).
This gives a massive equivalence class.

Example 6.4
The equivalence class of (ω, ∈) has a lot of things. For example, consider (Even, <),
(Odd, <). This is not even a set. Given S, we can construct a copy of ω by putting
∈ on
S, {S}, {{S}}, . . .
and things blow up!

Also the equivalence classes are stacked too. For C1 and C2 are equivalence classes of
a well-ordering. Pick (x1 , R1 ) ∈ C1 or (X2 , R2 ) ∈ C2 . WLOG, assume R1 is an initial
segment of R2 . Then every guy in C1 is an initial segment of C2 . That means the
equivalence classes are well-ordered too.
Now we want to pick a specific canonical representative from each equivalence class.
Here’s how you do it. Given (X, R) a representative of a class, we want to add a single
element. There are tons of elements not in X, but the canonical one is X. Hence we do

X 0 = X ∪ {X}

and
R0 = R ∪ {(y, X) | y ∈ X} .
So starting with the empty set, we have a trivial relation (∅, ∅), and we recover the
ordinals.

26
Evan Chen
[email protected] (Harvard College 6 September 18, 2014
Fall 2014)

§6.4 How to preserve transitive sets

Recall that a set A is transitive if a ∈ A =⇒ a ⊆ A.

Lemma 6.5
If X is transitive then

(a) X ∪ {X} is transitive

(b) ∪X is transitive

(c) P(X) is transitive.

Proof. (a) Consider x ∈ X and do cases on x ∈ X versus x = X.

(b) Suppose X is transitive. Then ∪X is transitive. Suppose a ∈ ∪X. Then a ∈ b ∈ X

for some b. By transitivity, a ∈ X, and hence by union a ⊆ ∪X.

(c) Check it.

Of course, ∅ is transitive. So Vα are all transitive by invoking the lemma. The stem
On consists of transitive sets too.

§6.5 Ordinals
Definition 6.6. An ordinal is a transitive set that is well-ordered by ∈.

Let On denote the class of ordinals.

Every transitive set is closed under membership ∈. Ordinals are additionally well-
ordered under ∈.
You can show that any transitive well-ordering is isomorphic to some ordinal.

Exercise 6.7. If α ∈ On and β ∈ α then β ∈ On.

Proof. To show β is transitive, use both the fact that α is transitive and well-ordered.
(Neither alone is sufficient.)
To show that β is well-ordered by ∈, then. . .

Exercise 6.8. If α ∈ On then

α = {β | β < α} .
Here < means ∈.

Lemma 6.9
Suppose α ∈ β ∈ On. The following are equivalent. Then α ∈ β if and only if α ( β.

Proof. Clearly α ∈ β =⇒ α ⊆ β, and α 6= β is not possible (since otherwise β ∈ β).

The reverse is more tricky.
Now suppose α ( β. Since β is well-ordered under membership, we can consider
an ∈-least element b in β − α. By the first exercise, b ∈ On. By the second exercise,
b = {γ | γ < b} = α. So α ∈ β.

27
Evan Chen
[email protected] (Harvard College 6 September 18, 2014
Fall 2014)

Theorem 6.10 (Linearity)

If α and β are ordinals then either α ≤ β or β ≤ α.

Proof. Check α ∩ β ∈ On. We claim that α ∩ β is either α or β. If not, α ∩ β ( α, β. By

the lemma, α ∩ β ∈ α and α ∩ β ∈ β. Hence α ∩ β ∈ α ∩ β and that is a contradiction.

Theorem 6.11
On is not a set.

Proof. Assume Ω is the set of all ordinals. Then Ω is well-ordered by ∈. Moreover, if

a ∈ b ∈ Ω, then b is an ordinal, so a is on ordinal, hence Ω is transitive.

28
Evan Chen
[email protected] (Harvard College 7 September 23, 2014
Fall 2014)

§7 September 23, 2014

§7.1 Recap
• Well-orderings and well-founded relations

• Transfinite induction

• Transfinite reduction

• Well-ordering theorems – isomorphism.

• On is not an ordinal (not a set)

Recall that an ordinal is a transitive set which is well-ordered by the relation ∈.

§7.2 Order Types

We will now show that ordinals are indeed a canonical representative.

Theorem 7.1 (Comprehensiveness)

Fro every well-ordering (X, R) there is an ordinal α such that

(X, R) ∼
= (α, ∈).

This proof is a little subtle. Actually, Schimmerling does not give a full proof.

Proof. Fix a well-ordering (X, R).

Claim. For any a ∈ X there is a unique ordinal α with IaR ∼
= (α, ∈).
Proof of lemma. First, we prove existence. Take the R-minimal one a ∈ X. Then look
at the things below it:


∀ā ∈ X āRa → ∃!x ∈ On I R a ∼
= (α, ∈) .

Now pick using Replacement


α = α | ∃aRa IaR ∼
= (α, ∈) .

To prove α is an ordinal, it suffices to prove that it is is transitive (since it is clearly

well-ordered by ∈). You can check this. Fix α ∈ α ∈ α. Then α has an a ∈ IaR ∼ = (α, ∈).
R
Now α ∈ (α, ∈), so a ∈ Ia . That gives α ∈ α.
Hence there exists such an α with

IaR ∼
= (α, ∈) .

Of course, this α is unique, since

(α, ∈) ∼
= IaR ∼
= α0 , ∈

implies α ∼
= α0 . 

Hence, for any initial segment in (X, R) we can get a corresponding ordinal. Now to
hit all of (X, R) we can set

β = β | ∃a ∈ XIaR ∼= (β, ∈) .

29
Evan Chen
[email protected] (Harvard College 7 September 23, 2014
Fall 2014)

Definition 7.2. Suppose (X, R) is a well-ordering. Then the order type of (X, R) is the
unique ordinal α such that (X, R) ∼
= (α, ∈).
Exercise 7.3. Show that there are arbitrarily large limit ordinals.
This is not true without Replacement, since Vω+ω is a model of ZFC minus Replacement.
You need Replacement to show that λ + ω is a set for any limit ordinal λ.

§7.3 Arithmetic on Ordinals

There are two approaches,
• transfinite induction,
• explicit construction.
We want to get α + β, α · β, αβ .

7.3.1 Addition
The transfinite recursion method is as follows.
α+0=α
α + (β + 1) = (α + β) + 1
[
α+λ= (α + β).
β<λ

Here λ 6= 0.
For explicit construction, we work by concatenation. Suppose we have (X, R), (Y, S)
(so we do this with any well-orderings rather than just with ordinals, even though it’s
the same concept.). Basically we want to stick Y on top of X, so we do a quick hack by
using ordered pairs. Define
(X, R) + (Y, S) = (({0} × X) ∪ ({1} × Y ) , <lex )
Yay. Since we can add well-orderings, we can add ordinals by just looking at order types.
It is not too hard to check that these definitions coincide (just use transfinite induction).

Example 7.4
Note that 2014 + ω = ω 6= ω + 2014.

Remark 7.5. Ordinal addition is not commutative. However, it is associative.

7.3.2 Multiplication
Again, it suffices to define things using transfinite recursion.
α·0=α
α · (β + 1) = (α · β) + α
[
α·λ= α · β.
β<λ

We can also do explicit construction.

α · β = (β × α, <lex ) .
Again transfinite induction lets you show these are the same.

30
Evan Chen
[email protected] (Harvard College 7 September 23, 2014
Fall 2014)

Example 7.6
Check that 2014 · ω = ω 6= ω · 2014 for similar reasons as addition.

Remark 7.7. Like addition, multiplication is associative (since we just get ordered
triples in the second definition), but not commutative.

7.3.3 Exponentiation
The recursion is clear.

α0 = 1
αβ+1 = αβ · α
[
αλ = αβ .
β<λ

The explicit construction is more complicated here, so tl;dr, have fun with the exercises.

7.3.4 A nontrivial theorem

Exercise 7.8 (Division Algorithm). Suppose we have a “base” ordinal β > 0. Let α be
an ordinal. Then there exists a unique choice of γ1 , γ2 such that

α = β · γ1 + γ2

and γ2 < β.
Proof. Mimic the proof of the division algorithm.

Theorem 7.9 (Cantor Normal Form)

Suppose α > 0 is an ordinal. Then α can be written uniquely as

α = ω β1 · κ1 + ω β2 · κ2 + · · · + ω βn · κn

where α > β1 > · · · > βn ≥ 0 are ordinals, and κ1 , . . . , κn , n are positive integers.

Proof. Transfinite induction. Just greedily do it. LOL.

§7.4 The hierarchy of sets, done with ordinals

Definition 7.10. By transfinite induction, we set

V0 = ∅
Vα+1 = P(Vα )
[
Vλ = Vα
α<λ

Finally, let V be the following class.

[
V = Vα .
α∈On

This is the “cumulative hierarchy of sets”.

31
Evan Chen
[email protected] (Harvard College 7 September 23, 2014
Fall 2014)

Lemma 7.11
For all α, β ∈ On such that α < β,

(1) Vα is transitive,

(2) Vα ⊆ Vβ .

Proof. The proof of (1) is just transfinite induction on α. For (ii), Vα ∈ Vβ =⇒ Vα ⊆ Vβ .

(The first part is just mass power set.)

§7.5 The ordinals grab all the sets

Exercise 7.12. Show that
def
[
{x | x = x} = V = Vα
α∈On
S
In other words, every set is picked up in α∈On (Vα ).

Proof. If some set doesn’t appear, take an ∈-minimal set not in any ordinal (possible
since Foundation tells us ∈ is well-founded). Thus the key is Foundation.

The exercise implies the following notion is well-defined.

Definition 7.13. The rank of a set x, written rank(x), is the least α such that x ∈ Vα+1 .

Remark 7.14. Notice that rank(α) = α and rank(Vα ) = α for every α ∈ On.

Here is an alternative definition.

Definition 7.15. Show that an equivalent definition is as follows. By transfinite recursion

on (V, ∈), we define for each x ∈ V with x 6= ∅ by

ρ(∅) = 0
[
ρ(x) = (ρ(y) + 1)
y∈x

Remark 7.16. This is why we want to have transfinite recursion on well-founded

structures, not just well-orderings.

Exercise 7.17. Show that rank and ρ coincide.

32
Evan Chen
[email protected] (Harvard College 8 September 25
Fall 2014)

§8 September 25
Today we discuss cardinals. In ordinals, order matters; it doesn’t for cardinals. Note that
ordinals and cardinals coincide in the finite case.
Ordinals are a canonical representative for well-orderings. The idea is to do the same
thing with equivalence classes of bijections: we want to pick a canonical representative
for each of these guys.
At the end of this lecture, we state CH and then shoot for it.

§8.1 Axiom of Choice is equivalent to WO

Recall that our axiom states that there exists a choice function for any collection of sets
X.
Cantor implicitly assumes that every set can be well-ordered. (Can R be well-ordered?)
This isn’t quite true.

Theorem 8.1 (Zermelo, 1904)

The following are equivalent.

(1) AC, the Axiom of Choice.

(2) WO, which states that every set X can be well-ordered.

Proof. First, let us show that AC implies WO. Let X be an arbitrary set. Pick a choice
function f on P(X) − {∅} (this follows from AC). Hence every subset Y ⊆ X has a
choice f (y) ∈ X.
First, pick x1 = f (X) ∈ X. Now we have a first element. Select x2 = f (X − x1 ) ∈
X − x1 , and so on. In general, we wish to define f ∗ : β → X such that for each α < β,
[
f ∗ (α) = f (X) − f ∗ (x).
x∈α

This process cannot continue forever, since otherwise X is bijected to the ordinals, which
is impossible (recall that X is a set). Hence we can find β as above, so we have a
well-ordering of X, namely β.
Now assume WO; we wish to show AC.SLet X be an arbitrary set whose elements are
not ∅. Now take a well-ordering R on X. Then given any set Y ∈ X, we just the
R-minimal element.

Yeah OK so AC is equivalent to the assertion that everything is well-ordered. GG.

§8.2 Zorn’s Lemma

Definition 8.2. Let (X, EX ) be a poset or partial ordering; i.e. a reflexive transitive
relation on the set X An element of x ∈ X is maximal if


@x0 ∈ X x / x0 .

Definition 8.3. A subset C ⊆ X is a chain if C is linearly ordered by EX . An element

is x ∈ X is an upper bound for a chain C if ∀c ∈ C, c EX x.

Remark. Note that this upper bound does not nead to lie in C.

33
Evan Chen
[email protected] (Harvard College 8 September 25
Fall 2014)

Definition 8.4 (Zorn’s Principle). If (X, ≤X ) is a partial ordering such that every chain
has an upper bound then (X, EX ) has a maximal element.

Theorem 8.5 (Zorn’s Lemma)

AC is equivalent to Zorn’s Principle.

Proof. Assume AC and WO. We wish to show ZP. Let (X, EX ) be a partial order such
that all chains have an upper bound.
By WO we can put a well-order ≺ with order type β on X. Pick first the ≺-minimal
point x0 = f (0). Then let x1 be the ≺-least point above x0 . Repeat for the successors
(we’re using our contradiction assumption). Then we’ll set xω to be the ≺-least z above
the chain x0 / x1 / x2 / . . . . In general set xλ to be the ≺-least number above xα for
every α < λ. This process must terminate eventually, since the resulting chain C can be
embedded in β. It can only terminate when we have reached a maximal element.
Now to show that ZP implies AC. Look at the set of partial choice functions P, ordered
by EP be the poset of partial choice functions on X “ordered by extension”; i.e.
n [ o
P = f :X→ X is a choice function | X 6= ∅ and X ⊆ X .

g if and only if f ⊆ g (not dom f ⊆ dom g!). Observe that for any
The order is just f EP S
chain C, the function C is an upper bound for the chain C. By Zorn, we now have
a maximal element in P , denoted F . It follows that F is the required choice function,
because the only possible maximal elements of F are complete choice functions.

Moral: partial orders should approximate the order that you want
Reverse mathematics: our lemmas prove the axioms.

§8.3 Some weakenings of AC

Definition 8.6 (Dependent Choice, DC). Given a relation ∼ such that for every x there
is something related to it, we can construct a chain {xi : i < ω} such that for all i < w,
we have
xi ∼ xi+1 .

This is called “dependent” since previous steps matter for future steps.

Exercise 8.7. Show that AC implies DC.

Fact 8.8. DC does not imply AC.

Definition 8.9. A set X is countable if and only if X can be bijected to ω.

Definition 8.10 (Countable Choice, ACω ). Every countable set of nonempty sets has a
choice function.

Exercise 8.11. Show that DC implies ACω .

Fact 8.12. ACω does not imply DC.

34
Evan Chen
[email protected] (Harvard College 9 September 30, 2014
Fall 2014)

§9 September 30, 2014

Last time we showed AC, WO, and Zorn are equivalent.
In what follows, AC is not assumed until further notice.

§9.1 Equinumerous sets

The following definition is actually really important; historically, it was not recognized as
significant.

Definition 9.1. Two sets A and B are equinumerous, written A ≈ B, if there is a

bijection between them.

Definition 9.2. We say A  B if we can injection A into B.

Clearly, N  Q  R, and N ≈ Q 6≈ R.

Example 9.3
Set E = {0, 2, 4, . . . } and O = {1, 3, 5, . . . }. Then E ≈ N ≈ O.

Example 9.4
N × N ≈ N.

The following fundamental theorem shows that this notation makes sense.

Theorem 9.5 (Cantor-Bernstein)

Suppose A  B and B  A. Then A ≈ B.

Proof. We are given injective functions f : A → B and g : B → A and desire a bijection

h : A → B.
Define X0 = A, Y0 = B, and recursively set

Xn+1 = g(Yn ) and Yn+1 = f (Xn ).

This gives us a set of chains

X0 ⊇ X1 ⊇ X2 ⊇ . . .
T
and similarly for Y . Let X ∞ = n∈N Xn . Define now h : A → B by


f (x) if x ∈ Xn \ Xn+1 for some even n
x 7→ g −1 (x) if x ∈ Xn \ Xn+1 for some odd n


f (x) if x ∈ X ∞ .

The point is that f (X ∞ ) = g −1 (X ∞ ) = Y ∞ , so in the last sentence we could replace f

with g −1 (this would lead to a different h, but both are still bijections).
Check that this works.

35
Evan Chen
[email protected] (Harvard College 9 September 30, 2014
Fall 2014)

§9.2 Creating cardinals

Definition 9.6. A cardinal is an ordinal κ such that for no α < κ do we have α ≈ κ.
Remark 9.7. These are also called initial ordinals.
You can check that all the ordinals ω + ω, ω 2 , ω ω , and even 0 are all countable sets,
so none of them are cardinals.
Remark 9.8. The first few cardinals are 0, 1, . . . . Then, ω is a cardinal as well, and we
set ℵ0 = ω.
Our goal is to prove the following proposition.

Proposition 9.9
There are arbitrarily large cardinals.

If we had WO (equivalently AC) right now, for X we could define

|X| = least ordinal κ with X ≈ κ.

This must be a cardinal, by definition. (The existence of such a κ is equivalent to WO.

In fact, AC is equivalent to every set having a cardinality.) Now we could construct a
cardinal greater than ω by taking |P(ω)|, which is a cardinal.
However, one can show this without WO.
Exercise 9.10. Given an ordinal κ, let κ+ be the least ordinal cardinal greater than κ.
Show that κ+ exists.
Sketch. Take the supremum of all the guys that have a 1 − 1 correspondence with κ.
Show that the guys form a set. Use Replacement?

This shows that the class of cardinals C is not a set. For otherwise, ∪C has a cardinality
κ, but there’s a cardinal greater than κ in C.

§9.3 Properties of cardinals

Lemma 9.11
S
Given any set A of cardinals, A is a cardinal.

This can let us construct ℵω .

S
Proof. If A has a maximal element κ, then A = κ and we’re done. S
Now suppose there is no largest cardinal. Suppose forScontradiction that ∃α < A
which can be bijected to it by some function f . So let κ ∈ A be such that α < κ. Then
κ can be injected to α by restricting f , contradiction.

Corollary 9.12
ℵω exists.

Proof. Set A = {ℵ0 , ℵ1 , . . . }, and note that A is a set by Replacement. Now apply the
above lemma.

36
Evan Chen
[email protected] (Harvard College 9 September 30, 2014
Fall 2014)

§9.4 Recurse
Summary: given a cardinalSκ we can construct κ+ a “successor” cardinal, and given a
set of cardinals A we have A is a cardinal.
Hence by transfinite recursion we can define the following.

Definition 9.13. For any α ∈ On, we may define

ℵ0 = ω
ℵα+1 = (ℵα )+
[
ℵλ = ℵα .
α<λ

Example 9.14
Here are some big ordinals.
ℵω , ℵ0
and
ℵℵω , ℵℵℵω , . . .
If we take this to the limit, we get a κ such that ℵκ = κ. This is called an “ℵ-fixed
point”.

It’s interesting to note that looking at the first few cardinals ℵ0 , ℵ1 , . . . , it seems that
the cardinal is much bigger than its index. Evidently, the κ mentioned above is the union
of ℵℵℵ... , where we have n ℵ’s.
Then we can take κ+ , . . .
So now we have even showed there are arbitrarily large ℵ-fixed points, say κ0 , κ1 , . . . .

Lemma 9.15
If κ is a cardinal then either κ is finite (i.e. κ ∈ ω) or κ = ℵα for some α ∈ On.

Proof. Assume κ is infinite. Take α minimal with ℵα ≥ κ. For contradiction suppose

ℵα > κ. S
Now, consider two cases. If α is a limit ordinal λ, then ℵλ is the supremum γ<λ ℵγ .
So some γ < λ has ℵγ > κ.
On the other hand, suppose α = α + 1. Then

ℵᾱ < κ < ℵα

are three cardinals. This is a contradiction.

Remark 9.16. Still no AC!

The Continuum Hypothesis (CH) says that if X ⊂ P(ω) is infinite, then either X ≈ ω
or X ≈ P(ω). Given AC, we can talk about |R|, in which case CH states |R| = ℵ1 .

37
Evan Chen
[email protected] (Harvard College 10 October 2, 2014
Fall 2014)

§10 October 2, 2014

More on cardinals.

§10.1 Cardinal arithmetic

Recall the way we set up ordinal arithmetic. Note that in particular, ω + ω > ω and
ω 2 > ω.
But cardinals count cardinality. We want to have

ℵ0 + ℵ0 = ℵ0
ℵ0 × ℵ0 = ℵ0

because ω + ω and ω · ω are countable. In the case of cardinals, we simply “ignore order”.
The following definition uses the Axiom of Choice.

Definition 10.1. The cardinality |X| of a set X is the least ordinal α such that |X| ≈ α.

Recall that ≈ means “equinumerous”, and this is well-defined. (AC implies that X
has a well-ordering, so there is at least one such ordinal).

Definition 10.2 (Cardinal Arithmetic). Given cardinals κ and λ, define

κ + λ = |({0} × κ) ∪ ({1} × λ)|

and
k × λ = |λ × κ| .

§10.2 Finite sets

Definition 10.3. A set X is

• finite if |X| < ω.

• countable if |X| ≤ ω.

• uncountable if |X| > ω.

Here are two nice characterizations of finite-ness.

Definition 10.4. A set is Dedekind-finite (abbreviated D-finite) if every injective

function from X to itself is also surjective.

Exercise 10.5. Show that a set is finite if and only if it is Dedekind finite.

Definition 10.6. A set X is D∗ -finite if there exists f : X → X such that for no subset
∅ ( Y ( X is f closed under Y (meaning f (Y ) ⊆ Y ).

Exercise 10.7. Show that this is also a characterization of finite sets.

Partial Proof. Look at 

a, f (a), f 2 (a), . . .
and let this be Y . If Y is finite done. Otherwise take Y − a.

Without AC you can find models of set theory for which the above equivalences break.

38
Evan Chen
[email protected] (Harvard College 10 October 2, 2014
Fall 2014)

§10.3 Infinite cardinal arithmetic is boring

For finite cardinals, we can show that
n × m > n + m > n, m.
But now we have the following property of cardinal arithmetic.

Theorem 10.8
Let κ be an infinite cardinal. Then κ × κ = κ.

Proof. By transfinite induction (which is fine because we are working over ordinals), we
may assume that the theorem holds for all infinite cardinals κ < κ. Put an ordering
on κ × κ as follows: for (α, β) and (γ, δ) in κ × κ set (α, β) / (γ, δ) if and only if
max(α, β) < max(γ, δ) or the maximums are equal and (α, β) <lex (γ, δ). (This gives a
“diagonal” argument.)
We claim that the order type of (κ × κ, /) equals κ.
Count the /-predecessors of some (α, β) ∈ κ × κ. We claim there are at most
|(max(α, β) + 1) × (max(α, β) + 1)|
such points. Set κ = |max(α, β) + 1| < κ. By induction, |κ × κ| = κ.
Hence we’ve shown that there are at most κ points /-below any pair, which implies
that ot(κ × κ, /) = κ.

Corollary 10.9
Given cardinals κ and λ, one of which is infinite, we have

κ × λ = κ + λ = max (κ, λ) .

Proof. Compute
max (κ, λ) ≤ κ + λ
≤κ×λ
≤ max (κ, λ) × max (κ, λ)
= max (κ, λ) .

§10.4 Cardinal exponentiation

Unlike in addition and multiplication, we do not ignore order for exponentiation.
Definition 10.10. Suppose κ and λ are cardinals. Then
κλ = |F (λ, κ)| .
Here F (A, B) is the set of functions from A to B.

Example 10.11
2ℵ0 gives us binary numbers, R.

Continuum Hypothesis states that 2ℵ0 = ℵ1 .

39
Evan Chen
[email protected] (Harvard College 10 October 2, 2014
Fall 2014)

§10.5 Fundamental theorems of cardinal arithmetic

Theorem 10.12 (Cantor)

For all κ, κκ > κ.

Proof. We already have seen (from Cantor) that

2ℵα > ℵα

for every ordinals α. For all κ, since κκ = 2κ (why?), we have κκ > κ.

Theorem 10.13 (Konig)

For all finite κ, we have κcof(κ) > κ.

Here cof(κ) is cofinality, which we haven’t defined yet.

We are going to show that these two theorems are “all we can prove” in ZFC.
There are some trivial things we can prove, like

Lemma 10.14
We have
(κλ )µ = κλ·µ .

Proof. This was like a Math 55 problem. Biject

F (µ, F (λ, κ)) → F (λ × µ, κ)

in the obvious way.

§10.6 Some exponentiation is not interesting

Lemma 10.15
Suppose 2 ≤ κ ≤ λ, and λ is infinite. Then

κλ = 2λ .

Proof. We have
2λ ≤ κλ ≤ (2κ )λ = 2κ·λ = 2λ .

So we only care about κ > λ and 2κ .

40
Evan Chen
[email protected] (Harvard College 11 October 7, 2014
Fall 2014)

§11 October 7, 2014

Last time, we saw that cardinal addition and multiplication were trivial, and that
exponentiation is hard to prove stuff about. We only know that

κλ = 2λ ∀2 ≤ κ ≤ λ, λ > ω.

It appears that (λ+ )λ can be much bigger than λ. For example, if CH is true then
2ℵ0 = ℵ1  ℵℵ1 0 .

§11.1 Cofinality
Put in a magazine with ω many bullets. . .

Definition 11.1. Let α be a limit ordinal. A function f : β → α is called cofinal if for

every α < α, there is a β < β such that f (β) > α. In other words, im f gets arbitrarily
high.
The cofinality of α, denoted cof(α), is the least ordinal β such that ∃f : β → α which
is cofinal in α.
Loosely, the cofinality is the smallest ordinal β which can be “stretched out” to get α.
Observe that in particular, cof(α) ≤ α.

Example 11.2
cof(ℵω ) = ω, because we can stretch ω = {0, 1, . . . } as ℵ0 , ℵ1 , . . . which grows
arbitrarily large up to ℵω .

Example 11.3
We have cof(ℵ1 ) = ℵ1 = ω1 . In general, any successor cardinal has cofinality equal
to itself.

Definition 11.4. We say that a limit ordinal α is regular if cof(α) = α and singular
if cof(α) < α.
The case of interest is when α is a cardinal.

Example 11.5
ℵ0 = ω, ℵn and ℵα+1 are all regular. However, ℵω is singular, as is ℵω+ω and so on.

Remark 11.6. There may exist large cardinals ℵλ , with λ a limit ordinal, such that ℵλ
is regular. The question of whether they exist is independent of ZFC. In fact, it turns
out their existence establishes the consistency of ZFC; very roughly, they are so big that
cutting the universe off at them gives a model of ZFC.

§11.2 Lemmas on cofinality

Lemma 11.7
Suppose α is a limit ordinal, and let β = cof(α). Then there exists a map from β to
α which is cofinal and moreover is strictly increasing.

41
Evan Chen
[email protected] (Harvard College 11 October 7, 2014
Fall 2014)

Clearly we have a function β → α cofinal, so the lemma just lets us assume WLOG it is
increasing.

Proof. Take the cofinal function g : β → α; we wish to modify it so that it’s increasing.
Define by transifnite recursion the map f : β → α with
 
γ 7→ max g(γ), 1 + sup f (γ) .
γ<γ

This forces f to be strictly increasing and obviously it remains cofinal. Note that we are
OK at the top since α is a limit ordinal.

Lemma 11.8
Let α, β be limit ordinals. Suppose f : α → β is cofinal and non-decreasing. Then
cof(α) = cof(β).

This is pretty clear if you look at a picture.

Proof. First, we prove cof(β) ≤ cof(α). Let g : cof(α) → α be cofinal.

g f
cof(α) →
− α−
→β

You can check that f ◦ g is cofinal, so cof(β) ≤ cof(α) by definition (since cof(β) is the
minimal one).
Now let’s prove cof(α) ≤ cof(β). This time let g : cof(β) → β be cofinal, and assume
WLOG that it’s strictly increasing. We want to construct a cofinal map h : cof(β) → α.
Let’s have h map β to the least α such that f (α) > g(β); this is OK since f is cofinal.
one can check that this works.

§11.3 Given AC, cofinalities and successor cardinals are regular

This gives the following corollary, which we’ll call a lemma for kicks.

Lemma 11.9
For any limit ordinal α we have cof(α) is a regular cardinal.

Proof. First, we prove that cof(α) is actually a limit ordinal. If not, suppose cof(α) = γ+1.
Then we can map γ cofinally into α by shifting everything up by one, contradiction.
Now, we have a cofinal increasing map

cof(α) → α.

By applying the lemma, we have cof(cof(α)) = cof(α). This implies (by definition) that
cof(α) is regular.
Finally, we want to show cof(α) is a cardinal. If not, there’s a bijection between
cof(α) and a smaller ordinal γ, say g. Then f ◦ g is a cofinal map into α which is a
contradiction.

42
Evan Chen
[email protected] (Harvard College 11 October 7, 2014
Fall 2014)

Proposition 11.10
Given the Axiom of Choice (or just Countable Choice), cof(ℵ1 ) is regular.

Proof. If not, then cof(ℵ1 ) is a cardinal less than ℵ1 ; this can only mean cof(ℵ1 ) = ℵ0 .
Let f : ℵ0 → ℵ1 be cofinal. We observe that f (n) < ℵ1 , which means that f (n) is
countable. For each n, we can pick a gn from ℵ0 onto f (n). By “union”-ing all the gn ,
we get a surjection from ℵ0 onto ℵ1 . Contradiction.

More generally, we have the following.

Lemma 11.11
Assume Choice. Given that |A| = κ+ , there is no collection of sets

{Aα | α < κ}

such that

(1) ∀α < κ we have |Aα | ≤ κ.

S
(2) A = α<κ Aα .

Proof. Basically the same. I’ll copy it down, but it’s not especially worth reading.
Assume not. Each Aα (for α < κ) has size ≤ κ. So there exists surjections for each
α < κ with codomain
∃fα : κ → Aα .
By AC we can pick such an fα for each α. S S 
We will now produce a surjection from κ onto α<κ Aα . This will imply  α<κ Aα  ≤ κ.
Let π : κ → κ × κ be a bijection, and set
[
g :κ×κ→ Aα
α<κ

by
(α, β) 7→ fα (β).
Clearly g is surjective, since the fα are all surjective. Thus
[
g◦π :κ→ Aα
α<κ

is surjective, completing the proof.

This gives us the following theorem.

Theorem 11.12
Assume Choice. Then κ+ is a regular cardinal for every κ.

Proof. It suffices to show that cof(κ+ ) = κ+ , so assume for contradiction that cof(κ+ ) <
κ+ . Consider the corresponding map f . Set Aα = f (α); now {Aα }α≤κ is a partition
described in the above lemma, contradiction.

43
Evan Chen
[email protected] (Harvard College 11 October 7, 2014
Fall 2014)

Theorem 11.13 (Konig)

For every infinite cardinal κcof(κ) > κ.

Proof. This is just a diagonal construction.

Suppose f : cof(k) → κ is cofinal. Suppose also for contradiction that

κcof(κ) = κ.

That means we can enumerate all functions cof(κ) → κ as (gα )α≤κ .

But let
g : cof(α) → κ
by sending β to be the least element in

κ − {gα (β) | α < f (β)} .

What we’ve done is that for every β ∈ cof(κ), our selection of g(β) doesn’t match gα for
any α < f (β). Since f is cofinal, f (β) gets arbitrarily large, and so g does not agree for
any gα . Then g ∈
/ {gα }α , contradiction.

§11.4 Inaccessible cardinals

Consider the monstrous cardinal
κ = ℵℵℵ .
..
.
This might look frighteningly huge, as κ = ℵκ , but its cofinality is ω as it is the limit of
the sequence
ℵ0 , ℵℵ0 , ℵℵℵ0 , . . .

Definition 11.14. We say κ is weakly inaccessible if

(i) κ > ω, and

(ii) κ is a regular limit cardinal.

You can show that ℵκ = κ in this case.

Definition. We say κ is a strong limit if for all λ < κ, we have 2λ < κ. If κ is strong
limit and weakly inaccessible, we say it is strongly inaccessible.

Aside: You can check that if κ is strongly inaccessible, then Vκ is a model for ZFC.
(The limit implies its closed under Power Set, and inaccessibility gives Replacement).
That implies that proving the existence of strongly inaccessible cardinals.

44
Evan Chen
[email protected] (Harvard College 12 October 9, 2014
Fall 2014)

§12 October 9, 2014

We discuss trees and Suslin’s Hypothesis (SH).

§12.1 Linear Orders

Definition 12.1. A linear order is a pair (L, ≤L ) which is reflexive, symmetric, tran-
sitive, and linear (meaning any two x and y are comparable). It is dense if ∀x <L y ∈ L
we can find z such that
x <L < z <L y
(here <L means ≤L but not equal). Finally, we say it is without endpoints if ∀x ∈ L,
there is a y1 <L x, and y2 >L x.

We had the following theorem from Cantor, which shows that in the countable case,
any guy which has all the above properties is actually isomorphic to Q.

Theorem 12.2 (Cantor)

Suppose (A, ≤A ) and (B, ≤B ) are countable dense linear orders without endpoints.
Then
(A, ≤A ) ∼
= (B, ≤B ) .

Proof. We need a bijection f from A to B which is order-preserving. We’ll construct f

using what is called a “back and forth argument”.
Let A = {an : n ∈ N} and B = {bn : n ∈ N}. First set f (a0 ) arbitrarily. Now we define
f −1 (b0 ) as follows:

• If b = f (a0 ) then we’re done, f −1 (b0 ) = a0 .

• Otherwise, since ≤A is without endpoints we can let f −1 (b0 ) be such that

a0 ≤A f −1 (b0 )

iff f (a0 ) ≤B b0 .

And just rinse and repeat with a1 . If it’s already selected, ignore it. Otherwise, since
both sides are without endpoints and dense, regardless of which “interval” we can map it
preserving order.
a0 f −1 (b0 )

f (a0 ) b0

We keep going. This exhausts both A and B, because on the even steps we use up
the elements in B, and on the odd steps we use up all the elements in A. This gives a
bijection f .

Remark 12.3. The theory of dense linear orders is actually complete.

45
Evan Chen
[email protected] (Harvard College 12 October 9, 2014
Fall 2014)

§12.2 Intermission
Question 12.4. Can you generalize the proof to arbitrary cardinalities?
First, let us construct a dense linear order without endpoints of cardinality ℵ1 . Consider
the set
A = Q × ω1
sorted lexicographically. This clearly works. This has a structure of ω1 embedded in it.
So we have like a chain
A = Q0 , Q1 , . . . , Qα , . . .
On the other hand we can just run this the other way.
B = . . . , Qα , . . . , Q1 , Q0 .
1
Consider ( 2014 , 0). In A, it has countably many things to its left. But every guy in B
has uncountably many things to its right. Hence no bijection exists.
We can also do Q × (ω1 + 1), which is not isomorphic to A. You have a “maximal”
Q-blob Qω1 , and there’s intuitively no way to map this into A. (Check this!)

§12.3 Separable complete linear orders

Definition 12.5. A linear order (L, ≤L ) is complete if whenever A ⊆ L has an upper
bound5 , then there is a least upper bound, and similarly for sets bounded below.
Definition 12.6. We say (L, ≤L ) is separable if there exists a countable D ⊆ L for
which (D, ≤L ) is dense.
Now we can prove a generalization of something Cantor did.

Theorem 12.7 (Cantor)

Suppose (A, ≤A ) and (B, ≤B ) are dense linear orders without endpoints, which are
both complete and separable. Then (A, ≤A ) ∼ = (B, ≤B ).

Here we’ve replaced “countable” with “complete and separable”.

Proof. Take A ⊂ A which is countable and dense,6 and similarly define B. Evidently our
first theorem implies
(A, ≤A ) ∼
= (B, ≤B )
(just check all the conditions). Let f be the bijection from A to B. We want to lift this
to f : A → B. This is easy enough: for any a ∈ /A
f (a) = sup {f (a) | a ≤A a} .
≤B

We can take sup’s by completeness.

To prove this is order-preserving, just notice that if a1 <A a2 then we get some a ∈ A
between them, so f (a) separates them. (In particular f is injective.)
To prove f is surjective, look at any b ∈ B, take the downwards interval, and then use
the fact that A is complete.

In particular, (R, ≤R ) is the unique dense linear order without endpoints which is
complete and separable (note that (Q, ≤Q ) witness separability).
5
Meaning there exists a z ∈ L such that a ≤L z for all a ∈ A.
6
Think of these as “glowing special points” that are scattered all throughout A.

46
Evan Chen
[email protected] (Harvard College 12 October 9, 2014
Fall 2014)

§12.4 Suslin property

Definition 12.8. A linear order (L, ≤L ) has the Suslin property if any collection of
disjoint open intervals is countable. Here an open interval is one of the form
def
(a, b) = {z ∈ L | a <L z <L b} .

It might seem that you should be able to pick ω1 into R, but in fact R does have the
Suslin property – any open interval contains a rational number, so you can pick at most
countably many. (But you can embed 0 into R!)

Lemma 12.9
If a linear order (L, ≤) is separable then it has the Suslin property.

Proof. Let D be a collection of disjoint open intervals, and let L be the dense countable
subset. By Separability, each open interval contains a distinct element of L.

So we can’t embed (ω1 , ∈) into there, but we can get order-preserving maps from
(ω, ∈), (ω ω , ∈), (0 , ∈), . . . . Now is it true that you can embed any α < ω1 in there?
Not sure. We can try transfinite induction. Embedding in (R, ≤R ) is the same as
((−∞, 0), ≤R ), giving us space on the right to work with. For the transfinite induction
the successor case is clear, but for limit ordinals λ we must somehow use the fact that λ
is countable (since the induction fails at λ = ω1 ).
Suppose λ is a limit less than ω1 .7 So for each α we have an fα : α → R and a bijection
π : λ → N.

Question 12.10 (Suslin’s Hypothesis). Is (R, ≤R ) the unique dense linear order without
endpoints which is both complete and has the Suslin property?

Notice that SP is weaker than separable, and if we replace SP with separable then the
answer is “yes”.

Definition 12.11. We call a Suslin line any other ordering with this property; that is,
a dense linear order which is complete, has the Suslin property, but is not separable.

SH is the conjecture that no Suslin lines exist. This would be weird. We can only lie
down countably many disjoint union intervals, but the reason is not because we can pick
a “rational” in every set (by “rational” we mean an element of a separable set).

Theorem 12.12
If ZFC is consistent, then

(a) ZFC + SH is consistent

(b) ZFC + ¬ SH is consistent.

This is proved using something called a Suslin tree. We have the following theorems,
which we won’t prove.

7
This can terribly be written as λ ∈ Lim ∩ ω1 .

47
Evan Chen
[email protected] (Harvard College 12 October 9, 2014
Fall 2014)

Theorem 12.13
There is a Suslin line if and only if there is a Suslin tree.

Theorem 12.14
Let L be the constructible universe (“up next”). Then if V = L, we have a Susline
tree ¬ SH.

But we can use a trick called “forcing” as follows: we can put a branch through it to kill
it, but then more trees come up, and then it’s just massive deforestation. That’s how we
get another direction.

48
Evan Chen
[email protected] (Harvard College 13 October 14, 2014
Fall 2014)

§13 October 14, 2014

Sick day. What happened?

§13.1 Sparse Review

Given a model (M, E), an assignment to

VAR = {(7, i) | i ∈ ω}

is a function s : VAR → M . We think of s(“x000 ) = s((7, 0)) = a ∈ M . Basically, we

assign every variable to something in M .
Given a formula ϕ, ϕ(x0 , . . . , xn ) where the free variables are x0 , . . . , xn and an
assignment s : VAR → M , the only part that matters for is s(“x000 ), . . .
We’re going to be lazy and write things like x ∈ y, even though x and y are not actually
official variables.
Often we write
(M, E)  φ[~a]
where ~a is really just the part of s : VAR → M that is relevant, namely

~a = (s(x0 ), . . . , s(xn )) .

(Remember  means “satisfies”.)

Definitions of ∆0 , Σn , Πn .
∆0 corresponds to having no unbounded ∀ and ∃. It means you don’t have to look
around things.

49
Evan Chen
[email protected] (Harvard College 14 October 16, 2014
Fall 2014)

§14 October 16, 2014

Guest lecture! Today we discuss infinite combinatorics. Unfortunately in my illness I
arrived late and failed to get the guest’s name.
In finite combinatorics we studied partitions of numbers and rooted trees. Today we
will be discussion the generalizations of these to the transfinite.

§14.1 Partitions
We can think of a partition of S as writing

S = S0 t S1 t · · · t Sm−1

as just a function f : S → {0, . . . , m − 1} (that is, f : S → m) specifying here to send

each element of S.

Definition 14.1. For a set A, let [A]n be the subsets of size n; i.e.

[A]n = {x ⊆ A | |A| = n} .

Definition 14.2. Let δ be an infinite cardinal. We define a function f : S → δ to be a

δ-partition of S. Given a δ-partition of the [A]n , a set H ⊆ A is called homogeneous
for f if f is constant on [H]n .

Remark 14.3. It’s helpful to think of f as a coloring function on the n-subsets. Homo-
geneous corresponds to monochromatic.

§14.2 Infinite Ramsey Theorem

Example 14.4
It’s well-known from Ramsey theory that any six people, either three are mutual
friends or three are mutual stranger. We can rephrase this as saying: any 2-partition
of [6]2 has a homogeneous subset of size 3. (Here [6]2 are the edges, and we put them
into two colors.)

This is a consequence of the finite Ramsey theorem, but we’ll prove an infinite version.

Theorem 14.5 (Infinite Ramsey Theorem)

If A is an infinite set, and f : [A]n → m is an m-partition of the n-subsets of A, then
there is some countably infinite H ≤ A homogeneous.

Here (and for the rest of the lecture), m and n will denote natural numbers.

Remark 14.6. When n = 2, this states that any m-coloring of an infinite graph has a
monochromatic coloring.

Proof. We are going to repeatedly use Infinite Pigeonhole. If n = 1, then we are trying
to partition [A]1 = A into a finite pieces, so some part must be infinite, QED.
Now suppose it holds for n. Consider f : [A]n+1 → m. For a ∈ A, we define fa from
[A − {a}]n → m by
fa (x) = f (x ∪ {a}).

50
Evan Chen
[email protected] (Harvard College 14 October 16, 2014
Fall 2014)

By inductive hypothesis, each fa has an infinite homogeneous subset. Let H0 = A and

a0 be arbitrary. For i < ω we define

Hi+1 ⊆ Hi infinite and homogeneous for fai

and then pick ai+1 ∈ Hi+1 − {aj | j ≤ i}.

Now consider {ai : i < ω}. The point is that {ai : i > k} is homogeneous for fak . Let
the constant value of fak on {ai : i ≥ k} be g(i). By infinite pigeonhole there is an infinite
subset I on which g is constant, as necessary. Then {ai : i ∈ I} is the desired set.

The special case of the proof with m = 3 colors (RGB, anyone?) and n = 2 is more
suitable for an olympiad, but makes the above proof much more illuminating. Consider
an infinite graph. Take a vertex v0 . Some color emanates infinitely often from v0 . Throw
away everything else. Take a new v1 . Ignoring v0 , some color emanates infinitely from v1 .
Throw away the other guys (but keep v1 )! In this way we get a sequence v0 , v1 , . . . of
vertices, each with a color it really likes. By Pigeonhole, we can take an infinite subset
that all like the same color. Done.

§14.3 Some negative results on partitions

Definition 14.7. If κ and δ are cardinals, and n ∈ N, and α is an ordinal we write
κ → (α)nδ if for every δ-partition of [k]n , there is a homogeneous set H ⊆ κ for the
partition, and H has size α.

Example 14.8
6 → (3)22 and ω → (ω)nm .

Note that you can increase κ and decrease α, so the → is “like” an inequality ≥.
Why do we assume n ∈ N? It turns out if AC is given, then we have the follow-
ing.

Lemma 14.9
Assume AC. Then κ 6→ (ω)ω2 for any infinite cardinal κ.

Remark 14.10. People often drop the subscript 2.

Proof. Take a well-ordering ≺ of [κ]ω . We’ll split κ into two parts using f : [κ]ω → 2 by
(
1 A is ≺-minimal among the infinite subsets of A
f (A) =
0 otherwise.

Assume for contradiction there’s some homogeneous H ⊂ κ of size ω; WLOG it’s

countable.
If H isn’t ≺-minimal, then consider an infinite A ∈ H which is. Then f colors A by 1.
But H is also an ω-subset of H, so f colors H by 0, contradiction.
So H must be ≺-minimal, and since H is homegeneous, so must every infinite subset
of H. Select an infinite subset A0 ∈ H such that H − A0 is also infinite. By looking at
the elements missed by A0 , we can create a chain

A0 ⊆ A1 ⊆ A2 ⊆ . . .

51
Evan Chen
[email protected] (Harvard College 14 October 16, 2014
Fall 2014)

But now we can obtain a decreasing ≺ chain

H ⊇ A0  A1  A2  . . . .

Indeed, each Ai is infinite but is ≺-minimal with respect to all its infinite subsets,
including Ai−1 . Boom!

Lemma 14.11
Assume AC. For any κ, 2κ 6→ (κ+ )22 .


Proof. Let xα | α < 2k enumerate the binary sequences of κ; i.e. the functions from κ
to 2. For each “edge” α, β consider the coloring where

α < β → (f ({α, β}) ⇐⇒ xα <lex xβ .

Assume for contradiction there’s a homogeneous (think monochromatic) subsequence

hxα : α ∈ Hi, and WLOG it is increasing.
For each i we define

βi = min β | β ≥ 1 + sup βj and
j<i

∃α > 1 + sup αj , α ∈ H
j<i

and β has 1 in the α digit.

αi = min α ∈ H | α > 1 + sup αj
j<i

and βi has 1 in the α digit .

Set the sup’s as 0 when i = 0.

Basically what this does is that as i grows, more digits get fixed. Supremum exists by
regularity of κ+ . So you get screwed by the time you get to xκ .

Remark 14.12. This is highly similar to the proof that you can’t have an ℵ1 increasing
sequence of real numbers.

Corollary 14.13
2κ 6→ (2κ )22 .

Proof. 2κ ≥ κ+ .

§14.4 Large Cardinals

Definition 14.14. A cardinal κ is called weakly compact if κ → (κ)22 .

Corollary 14.15
If κ is weakly compact, κ is inaccessible.

52
Evan Chen
[email protected] (Harvard College 14 October 16, 2014
Fall 2014)

Proof. By above stuff, we clearly have κ > 2λ for all λ < κ. S

Assume for contradiction that κ is singular. Then κ = i<λ Ai where λ < κ and
|Ai | < κ. Define f : [k]2 → 2 by

f ({α, β}) = 1 ⇐⇒ ∃i < λ : α, β ∈ Ai .

§14.5 Trees
Definition 14.16. A tree is a poset (T, <) such that for every x ∈ T , the set

predT (x) = {y ∈ T | y < x}

is well-ordered by <. The height of any x is ht(x) = ot(predT (x)) and Tα is the set of
x’s with height α, meaning

Tα = {x ∈ T | ht(x) = α} .

The height of the whole tree is the smallest α with Tα 6= 0, i.e.

ht(T ) = min {α : Tα = ∅} .

A branch through T is a maximal linear subset. A κ-tree is a tree of height κ whose

levels each have less than κ.

Example 14.17
The tree 2<ω is an ω-tree. The tree ω <ω has size ω, but it is very fat and not an
ω-tree.

Definition 14.18. An Aronszajn tree (pronounced EH-rin-shine) is a κ-tree without

a branch of size κ.

Lemma 14.19 (Konig’s Lemma)

There is no ℵ0 -Aronszajn tree.

Proof. At each level there’s finitely many guys. So some branch is infinite. Now go up
the tree greedily.

Lemma 14.20 (Aronszajn)

There is an ℵ1 -Aronszajn tree.

53
Evan Chen
[email protected] (Harvard College 15 October 21, 2014
Fall 2014)

§15 October 21, 2014

Constructability, Model Theory, . . .

§15.1 Definability
Definition 15.1. Given (M, E) a model in LST, we say that a relation X ⊆ M n is
(M, E)-definable by ϕ(x1 , . . . , xn ) if

X = {~a ∈ M n | (M, E)  φ[~a]} .

We say X is (M, E)-definable if such a ϕ exists.

Definition 15.2. We say X is definable in (M, E) with parameters if there is a
formula φ(x1 , . . . , xn , y1 , . . . , ym ) and a parameter sequence ~b ∈ M m such that
n o
X = ~a ∈ M n | (M, E)  ϕ[~a~b] .

Remark 15.3. Everything in M is definable in (M, E) with parameters. Take b ∈ M

and consider the formula “x = b”. This has one free variable and one parameter.
Remark 15.4. We’ll say a ∈ M is definable if {a} is definable.

Example 15.5
We can write down the formula defining “is an ordinal”. We can expand the machine
code for transitive: ∀y ∈ x(y ⊆ x), or ∀y ∈ x(∀z ∈ y(z ∈ x)). This is ∆0 because it
has no unbounded parameters. Similarly expand “well-ordered by ∈”.

Example 15.6
∅ is ∆0 definable. Write the formula ∀y ∈ x(y 6= y).

We are mainly interested in ∈-models, where M is transitive and ∈ |M ×M = E. We

will often write this as (M, ∈).

§15.2 Fragments of the Universe

Let’s see what is definable over the structure (Vω , ∈).

Example 15.7
ω = {0, 1, . . . } is ∆0 definable without parameters in (Vω , ∈). Get

{a ∈ Vω | (Vω , ∈)  ϕ[a]} = {0, 1, . . . } .

where ϕ means “is an ordinal”.

Example 15.8
{a ∈ Vω1 | (Vω1 , ∈)  ϕ[a]} = ω1 and {a ∈ V | (V, ∈)  ϕ[a]} = On.

54
Evan Chen
[email protected] (Harvard College 15 October 21, 2014
Fall 2014)

Lemma 15.9
Suppose (M, ∈) and (N, ∈) are two transitive ∈-models. Moreover, suppose M ⊆ N .
Let ϕ(x1 , . . . , xn ) be a formula in LST. Let ~a ∈ M n be a parameter sequence. Then

(a) If ϕ ∈ ∆0 , then (M, ∈)  φ[~a] ⇐⇒ (N, ∈)  φ[~a].

(b) If ϕ ∈ Σ1 , then (M, ∈)  φ[~a] implies (N, ∈)  φ[~a], but the converse fails.

(c) If ϕ ∈ Π1 , then (N, ∈)  φ[~a] implies (M, ∈)  φ[~a], but the converse fails.

These are called ∆0 -absoluteness, upward Σ1 -absoluteness, and downward Π1 -absoluteness.

This is actually intuitively clear. In the first case, since ϕ ∈ ∆0 , then truth depends only
on things internal to ~a, and M and N have the same information in this case. In the
second case, we have some existence qualifiers, but setting these to things in M is fine in
N as well. Similarly for the third case.

Proof. We prove (a) by induction on formula complexity. In the base case, we have
several cases. The first is φ is “x1 ∈ x2 ”. Given (a1 , a2 ) = ~a ∈ M 2 , we have

(M, ∈)  a1 ∈ a2 ⇐⇒ (N, ∈)  a1 ∈ a2 .

Well, duh. We also have “x1 = x2 ”, which is clear.

For the inductive step, we assume the statement holds for φ(x1 , . . . , xn ) and ψ(y1 , . . . , yn ).
We just check ¬φ(x1 , . . . , xn ) and φ(x1 , . . . , xn ) ∧ ψ(y1 , . . . , ym ).
Do stuff with the other quantifiers. We need the transitivity condition at the bounded
quantifier step. If M thinks there exists x in the parameter a, we need N to also think
that. Writing it out
(M, ∈)  ∃y ∈ x1 φ(y, x1 , , , .xn )[~a]
means by definition
∃b ∈ a1 : b ∈ M and (M, ∈)  φ[b, a1 ]. (1)
On the other end we have

(N, ∈)  ∃y ∈ x1 φ(y, x1 , , , .xn )[~a]

means by definition
∃b ∈ a1 : b ∈ N and (N, ∈)  φ[b, a1 ]. (2)
So we need to show the witness statements (1) and (2) are equivalent. To show (1) implies
(2), we only need M ⊆ N . For the other direction, though, we do need transitivity; we
know b ∈ a1 , b ∈ N , a1 ∈ M , but we want b ∈ M and that doesn’t follow unless M is
transitive!
For (b), (c) do similar things. Invoke (a) for convenience.

§15.3 Definable sets

Definition 15.10. For (M, ∈), with M transitive, and define

DefM = {X ⊆ M | X definable on (M, ∈) with parameters} .

Remark 15.11. We clearly have M ⊆ DefM ⊆ P(M ).

 
Fact 15.12. If |M | = n ∈ N, then DefM  = 2n .

55
Evan Chen
[email protected] (Harvard College 15 October 21, 2014
Fall 2014)

Fact 15.13. If |M | = α ≥ ω, then

 
DefM  = α × ℵ0 = α.

Exercise 15.14. Suppose (M, ∈) is an ∈-model with M transitive.

(1) (M, ∈)  Extensionality, Foundation.

(2) (M, ∈)  ∅Emptyset if and only if ∅ ∈ M , i.e. M 6= ∅. (Just note that “x = ∅00 is
∆0 .)

(3) M  Pairing if and only if ∀x, y ∈ M we have {x, y} ∈ M .

(4) . . .

(5) . . .

(6) (M, ∈)  Infinity when ω ∈ M .

(7) (M, ∈)  Comprehension if and only if for every C ∈ DefM and for all a ∈ M , we
have C ∩ a ∈ M .

The key is ∆0 – this lets us move back and forth between (M, ∈) and V .

56
Evan Chen
[email protected] (Harvard College 16 October 23, 2014
Fall 2014)

§16 October 23, 2014

Today we cover reflection. In what follows, the LST models are not necessarily transitive.

§16.1 Model theory notions

Definition 16.1. Suppose M = (M, E) and M 0 = (M 0 , E 0 ) are LST-models. We say
M is a substructure of M 0 , written M ⊆ M 0 , if M ⊆ M 0 and E = E 0 ∩ (M × M ).

Definition 16.2. We say M is an elementary substructure of M 0 , written M ≺ M 0 ,

if for all φ in LST and ~a ∈ M n , M  φ[~a] ⇐⇒ M 0  φ[~a].

Theorem 16.3 (Tarski-Vaught Test)

Suppose M , M 0 are LST-models such that M ⊆ M 0 . Then M ≺ M 0 if and only if
the following condition holds: Let φ be a sentence in LST and ~b ∈ M n . If ∃a ∈ M 0
such that M 0  φ[a, ~b] then there exists a ∈ M such that M 0  φ[a, ~b].

Proof. First, we show (1) implies (2). Assume M ≺ M 0 . Fix φ in LST and ~b ∈ M n .
Suppose ∃a ∈ M 0 such that M 0  φ[a, ~b]. Then M 0  ∃aφ[a, ~b]. Since M ≺ M 0 ,

M  ∃aφ[a, ~b].

Fix such an a ∈ M . So M  φ[~a, ~b]. But M ≺ M 0 , so

M 0  φ[a, ~b]

as needed.
The converse is the interesting direction. We’ll prove (1) by induction on formula
complexity. The base case is Σ0 formulas. a = b means

M  a = b ⇐⇒ a = b ⇐⇒ M 0  a = b.

Now a ∈ b is

M  a ∈ b ⇐⇒ (a, b) ∈ E ⇐⇒ (a, b) ∈ E 0 ⇐⇒ M 0  a ∈ b.

Then for some last things. Negation ¬ and or ∨ is immediate. We also need bounded
quantifiers. Assume it’s true for φ, so

M  φ[~b] ⇐⇒ M 0  φ[~b]

Show that for a ∈ M ,

M  ∃x ∈ aφ(x, b) ⇐⇒ ∃a ∈ M, a ∈ a, M  φ[a, ~b].

Hence ∃a ∈ M , a ∈ a, such that M 0 φ[a, ~b].

Then for the inductive step. Suppose that M and M 0 satisfy the same formulas for
Πn formulas (the Σn case follows by negation). Want to show Σn+1 Fix φ ∈ Σn+1 , say

φ = ∃xψ(x, y1 , . . . , yn )

where ψ ∈ Πn . Suppose
M  ∃xψ[x, ~b].

57
Evan Chen
[email protected] (Harvard College 16 October 23, 2014
Fall 2014)

Pick a ∈ M such that

M  ψ[a, ~b]
so since M ≺Πn M 0 we have
M 0  ψ[a, ~b].
Hence
M 0  ∃xψ[x, ~b].
Hence we have prove than

M  ∃xψ[x, ~b] =⇒ M 0  ∃xψ[x, ~b].

We need the converse now. So far we haven’t used the Tarski-Vaught condition; we invoke
it now and then repeat the proof in the opposite direction. Indeed, just take a ∈ M 0 such
that M 0  [a, ~b]. Then ∃a ∈ M such that M 0  ψ[a, ~b]. The inductive hypothesis then
finishes the proof.

Intuitively, what this is saying that M ≺ M 0 if for some formula φ, if we have a

witness a in M 0 then we can also get a witness in M .

§16.2 Reflection
Now for the cool stuff.
Definition 16.4. Suppose κ is a regular uncountable cardinal. Then C ⊆ κ is a club
(short for “closed unbounded”) if the following holds for all α < κ
• There exists β ∈ C with β > α.

• If α = sup(C ∩ α), then α ∈ C.

Theorem 16.5 (Reflection)

Let κ be a regular uncountable cardinal. Let hMα : α < κi be such that

(1) Each Mα = (Mα , Eα ) is an LST-model.

(2) Mα ⊆ Mβ for all α < β < κ.

(3) |Mα | < κ for any α.

S
(4) Mλ = α<λ Mα for any limit λ.
S S
We define M = (M, E), where M = α<κ Mα and E = α<κ Eα . Then the set

C = {α < κ | Mα ≺ M }

is a club in κ.

Remark 16.6. It follows from this that if α < β and α, β ∈ C, then Mα ≺ Mβ .

Proof. Let λ be a limit cardinal less than κ. We wish to show that C ∩ λ is unbounded
in κ. We wish to show λ ∈ C, i.e. that Mλ ≺ M . In other words, for any φ in LST and
~b ∈ M n we want the equivalence
λ

Mλ  φ[~b] ⇐⇒ M  φ[~b].

58
Evan Chen
[email protected] (Harvard College 16 October 23, 2014
Fall 2014)

Well, ~b gets covered by by some Mα since λ is a limit dude. Now we want to check the
Tarski-Vaught test. Suppose a ∈ M , M  φ[a, ~b]. We want to show there is a ∈ Mλ ,
M  φ[a, ~b]. So we go down to Mα :

M  ∃xϕ[x, ~b] =⇒ Mα  ∃xϕ[x, ~b]

and blah blah blah.

So this shows that C is closed. That’s not too interesting, because for all we know C
might just be empty. But now we’re going to show C is unbounded, which is far more
interesting. For φ in LST, let
fφ : κ → κ
by sending α to the least β < κ such that for all ~b ∈ Mα , if there exists a ∈ M such that
M  φ[a, ~b] then ∃a ∈ Mβ such that M  φ[a, ~b].
We claim this is well-defined. There are only |Mα |n many possible choices of ~b, and in
particular there are fewer than κ of these (this is why we assume |Mα | < κ). Otherwise,
we can construct a cofinal map from |Mαn | into κ by mapping each vector ~b into a β for
which the proposition fails. And that’s impossible since κ is regular!
In other words, what we’ve done is fix φ and then use Tarski-Vaught on all the ~b ∈ Mαn .
Now let g : κ → κ be defined by

α 7→ sup fφ (α).
φ∈LST

Since κ is regular and there are only countably many formulas, g(α) is well-defined.
Check that if α has the property that g maps α into itself (in other words, α is closed
under g), then by the Tarski-Vaught test, we have Mα ≺ M .
So it suffices to show there are arbitrarily large α < κ which are closed under g. Fix
α0 . Let α1 = g(α0 ), et cetera and define

α = sup αn .
n<ω

This α is closed under g, and by making α0 arbitrarily large we can make α as large as
we like.

§16.3 Variations

Theorem 16.7
Suppose φ is in LST and ~b ∈ V n . If V  φ[~b], then there exists α such that Vα ` φ[~b].

Why is this called reflection? Basically it means any statement φ one tries to make about
V is destined to fail in the sense that some level Vα will also have the statement.

Theorem 16.8
C = {α ∈ On | Vα ≺Σn V } is a club in On.

59
Evan Chen
[email protected] (Harvard College 17 October 28, 2014
Fall 2014)

§17 October 28, 2014

Laptop sick day. TODO: fill in.

60
Evan Chen
[email protected] (Harvard College 18 October 30, 2014
Fall 2014)

§18 October 30, 2014

Continuation of last time. (Guess whose laptop was dead?)

§18.1 Completing the proof of a theorem

We were proving Lκ  ZF

Proof.SSee notes. TODO: fill in later. Reflection Theorem is useful because the setup is
Lκ = α<κ Lα .
...
Let F ∈ DefLκ and F : Lκ → Lκ . We want to show for any set a, F applied pointwise
to a lives inside κ.
Assume not. Let α < κ be such that a ∈ Lα . We can shoot each element of Lα onto
|α| < κ, and then we get a cofinal map from |α| into κ, contradiction.

§18.2 The statement Lκ  V = L

We will show Lκ gives V = L, AC, and CH.
We need to prove that
∀x∃α (x ∈ Lα ) .
S
This might seem since by definition Lκ = α<κ Lα . But this particular statement is in
V . We need to prove it in Lκ ; i.e. while it’s obvious that V  V = L, we don’t have
Lκ  V = L.
Let x = Lα . We want to decide whether Lκ  (x = Lα ). In other words, we defined
Lα in V , and we want to check we get the same thing as if we run the definition in L (or
Lκ ).
Like assuming ZFC is consistent, we can construct a model M ∗ of ZFC which is not
well-founded by adding on the condition α1 > α2 , α2 > α3 , . . . . This is consistent by
compactness (look it up): for the model to be inconsistent, it must have a finite set of
inconsistent statements. And if we run the definition of L in here, we get something else
In still other words, what we really want to do is show that if we run the definition of
Lκ inside itself, we get all of Lκ . So our goal it look at the definition of hLα : α < κi and
check that it is simple enough that we have absoluteness. Since Lκ is transitive, we’d
be done if this was ∆0 – but it’s not. In fact, we will show it is Σ1 . This is sufficient
because we then have
Lκ  (x ∈ Lα ) =⇒ V  (x ∈ Lα ) .
For this we use something like Godël encoding. First we need to get the language, which
we’ve already done. Then we have to define what a term is. Then we need to define a
formula sequence; it’s just going to be a finite sequence where you build up the formula.
Next you define the length of a formula. It turns out that all these things so far are ∆0 .
Finally you define satisfaction sequence and definability in Σ1 .
In the end by encoding syntax (with some fairly scary computations) that
• “x is the αth level of the L-hierarchy”, that is, Lα , is Σ1 -definable.
• “x is constructible” is Σ1 since it is ∃y, α (Lev(y, α) ∧ x ∈ y).
• V = L is the statement ∀x (y is constructible), so it is Π2 .
Exercise 18.1. Assume λ > ω is a limit ordinal. Show that for all α such that ω < α < λ,
we have
Lλ  DEF(Lα+1 , Lα ).

61
Evan Chen
[email protected] (Harvard College 18 October 30, 2014
Fall 2014)

Here the Lα+1 and Lα are the real thing.

Proof. Check that DEF(x, y) is Σ1 . Check that Lα+1 is a Σ1 formula.

The converse of what we want is also a thing.

Theorem 18.2
If M is transitive and M  V = L then M = Lλ for some limit ordinal λ.

Proof. Since M  V = L, and α first appears in Lα+1 , the sup of the ordinals in M must
be a limit ordinal, say λ.
Consider x ∈ M . Since M  V = L, there exists b, α such that M  LEV(b, α) ∧ x ∈ b.
But (and this is the point), the statement LEV(b, α) is Σ1 . So by Σ1 -upward absoluteness,
we get V  LEV(b, α). Hence b = Lα .
Thus x ∈ b = Lα . Hence we showed that for all x ∈ M , x ∈ Lα for some α < λ.
Consequently M ⊆ Lλ . Thus M = Lλ .

Now here’s what we want.

Theorem 18.3
If λ is a limit ordinal, then Lλ  V = L.

Proof. Fix x ∈ Lλ . Let α < λ be such that x ∈ Lα . It suffices to show that Lλ 

LEV(Lα , α).
We expand the definition of LEV. Use the fact that Lλ has the function sequence
witnessing this Σ1 fact.
Like basically, it turns out there’s a ϕ, f such that

V  ∃f (ϕ[f, Lα , α)

and you can show f ∈ Lλ .

The key point is the following lemma.

Lemma 18.4
If λ is a limit ordinal then for all α < λ,

hLβ | β < αi ∈ Lλ .

In fact,
hLβ | β < λi = DefLλ .

Corollary 18.5
In ZFC+ , Lκ  V = L. So we have Lκ  ZF + (V = L).

Remark 18.6. We only need the κ condition to guarantee that V = L holds, but we
need regularity to cause it satisfy ZF. As an example, both models Lω , Lω+ω implies
V = L but not ZF.

62
Evan Chen
[email protected] (Harvard College 18 October 30, 2014
Fall 2014)

§18.3 A Nice application of Reflection

Proposition 18.7
Assuming ZFC+ , there exists a countable α for which Lα satisfies ZF.

Proof. By Reflection, the set D = {α < κ | Lα ≺ Lκ } is a club in κ. Hence for each

α ∈ D, we have Lα  ZF.
Using Choice (Lκ  AC), we can get a model M ≺ Lκ such that M ∈ Lκ and |M | = ℵ0 .
(This is the same argument as last time right before we did the Mostowski collapse.)
Now exactly like last time we can let M be the transitive collapse of M ; hence M is
transitive and M satisfies ZF and V = L. The only transitive sets with V = L are Lα by
the earlier theorem, and since M is countable, we have α < ω1 .

§18.4 A sketch of the proof that Lκ  AC

Induction. Assume we have a well-ordering of Lα . We wish to get an ordering of Lα+1 .
But n o
Lα+1 = DefLα = x ∈ Lα | Lα  ϕ[x, ~b] .

Here φ is in LST and ~b is a vector of Lα We can well order the countably many formulas
of ϕi , we can order ~b lexicographically, and then lexicographically order the entire set.
The limit ordinals are just unions.
In fact we get more that Choice in Lκ – we can order the entire universe. More on
that next time.

63
Evan Chen
[email protected] (Harvard College 19 November 4, 2014
Fall 2014)

§19 November 4, 2014

Today we will finish constructibility. The goal is to show that Lκ implies AC, GCH, and
a principle  to be defined later.

§19.1 Axiom of Choice

Theorem 19.1
Hence we have shown that ZF + (V = L) =⇒ AC. Hence Lκ  AC.

You need κ to be inaccessible to get κ  ZF. But you only need κ a limit to get V = L.

Proof. We want a function F from the ω, ω + 1, . . . to V sending α to a <α such that

for α ≥ ω, the following conditions hold:

(1) <α is a well-ordering of Lα .

(2) α < β =⇒ <α ⊆<β . Moreover, for any x ∈ Lα and y ∈ Lβ − Lα , we have x <β y.

(3) There exists a Σ1 formula φ(x0 , x1 , x2 ) such that for every limit ordinal λ ≥ ω, we
have
(Lλ  φ[a, x, Vω ]) ⇐⇒ (α ∈ On ∨ x =<α ).

Here’s a sketch of how to do this. We can define <ω because Vω consists of a bunch of
finite layers. Let hϕi : i < ωi be a Σ1 -definable well-ordering of all the formulas in LST
(possible since they live in Vω ). So you can just lex order everything!
To be explicit – suppose we have <α defined and want to define <α+1 . We have by
definition 
Lα+1 = x ⊆ Lα | x ∈ DefLα
So every element of Lα+1 is of the form

x = {z ∈ Lα | ϕi [z, a1 , . . . , an ]}

so we can associate with each x a tuple (i, a1 , . . . , an ) and then just lex-order it. (Since
you can possibly get multiple tuples, you
S should take the smallest one.)
For λ a limit ordinal, we write <λ = α<λ <α .

§19.2 GCH

Lemma 19.2 (Condensation)

Assume ZF + (V = L). Suppose X ≺ Lλ is an elementary substructure which is
transitive. Then
(X, ∈) ∼
= Lλ
for some limit ordinal λ.

Remark 19.3. It follows that if X is an elementary substructure of Lω (not necessarily

transitive) then X ∼
= Lω .

64
Evan Chen
[email protected] (Harvard College 19 November 4, 2014
Fall 2014)

Proof. Since ∈ is well-founded and extensional, we can take the Mostowski collapse of
(X, ∈) to get π : (X, ∈) → (X, ∈), where X is transitive. We know Lλ  (V = L), so
X  (V = L) (by the assumption X ≺ Lλ ), thus X  (V = L). But we already proved
that the transitive sets which satisfy V = L are precisely Lλ .
It remains to show that λ is a limit. But this follows from the fact that Lλ thinks “∀α,
α + 1 exists”, and this propagates down to Lλ , QED.

Theorem 19.4
ZF + (V = L) =⇒ GCH, and hence Lκ  GCH. Here GCH is the statement
κ+ = 2κ for any infinite cardinal κ.

Proof. Note that ZF + (V = L) implies AC, so we use Choice freely here.

By Cantor, 2κ ≥ κ+ , so it suffices to prove that 2κ ≤ κ+ . But we also showed in
homework that
|Lκ+ | = κ+ .
(Actually, we proved |Lα | = |α| for any α.) Hence, it suffices to show that any subset
A ⊆ κ in L is actually in Lκ+ , because then we’ve squeezed 2κ into a space of cardinality
κ+ . In other words, we will show P(A) ⊆ Lκ+ where the power set is taken over L.
Consider any A ⊆ κ. Let λ > κ be a limit. As in the proof of Reflection, we can
take X ≺ Lλ such that |X| = κ, A ∈ X, and κ ⊆ X. Let π : (X, ∈) → (Lλ , ∈) be the
transitive collapse in Condensation. We claim that A ∈ Lλ . Indeed,

π(A) = {π(γ) | γ ∈ A} .

But π acts as the identity on κ (check this by transfinite induction), and it follows that
π(A) = A.
However,  
L  = |X| = κ < κ+ .
λ
Thus A = π(A) ∈ Lλ .

§19.3 Hulls
Let me write down the details for “as in the proof of reflection”, since I’ve messed this
up twice. Here’s what we do. By Tarski-Vaight it suffices to find an X such that for all
ϕ in LST and for all ~b ∈ X, if we have witnesses a ∈ Lλ establishing Lλ  ϕ[a, ~b], then
we can find a witness a ∈ X establishing

Lλ  ϕ[a, ~b].

Now we define X by recursion as follows. Set

X0 = κ ∪ {A}.

This is our “initial set” of things we want X to have at the end of the day. Then we let
[ n o
X1 = a ∈ Lλ | a is <λ -minimal andLλ  [a, ~b] .
φ∈LST,~b∈X0ω
Lλ ∃zφ[z,~b]

OK, that’s retarded because the union is over a bunch of singleton sets. But I think
that makes the point. Basically, we’ve thrown in parameters in Lλ such that X1 satisfies

65
Evan Chen
[email protected] (Harvard College 19 November 4, 2014
Fall 2014)

Tarski-Vaught for all φ when the parameters are in X0 . The problem is now we have
some elements in X1 \ X0 which can also be parameters. So we construct Xn+1 based on
Xn in the same way.
S Xn+1 works for parameters in Xn , and moreover Xn ⊆ Xn+1 . By taking
Hence
X = n<ω Xn , now X satisfies the Tarski-Vaught test.
Now we claim |X| = |X0 | = κ ≥ ω. This just follows that any step, we only added in
ω · κω = κ many things.
Apparently these things are called hulls.

Theorem 19.5
Assume ZFC+ , so κ is a strongly inaccesible cardinal. There is a countable transitive
model of ZFC.

Proof. As in the proof of reflection, one can exhibit X ≺ Vκ countable. Then take the
Mostowski collapse of X.
Details: let X0 = {∅}. We construct a sequence of countable sets Xn inductively as
follows. Consider a sentence φ ∈ LST and for a parameter sequence b ∈ Xn such that
Vκ  ∃zφ[z, b]. Then we take the any such z, and add it into Xn . We let Xn+1 be the
result.
Since Xn was countable, the set φ × Xn<ω of possible pairs of sentences and parameters
is countable, and hence we add in at most countably many new elements; hence Xn+1 is
countable. S
Then X = Xn is countable and is an elementary substructure of Vκ . Now take the
Mostowski collapse of X.

§19.4 Summary
We have shown the following.

• ZF + (V = L) implies AC and GCH.

• Lκ  ZF + (V = L).

• Lκ  ZF + (CH).

So assuming ZFC+ , we have demonstrated that ZFC cannot establish ¬CH. Later we
will use forcing to show that ZFC cannot prove CH either.
But actually we showed more than L  CH. We showed that every real number (subsets
of N) appears in Lω1 . So we ask, where do the real numbers appear? For which α do we
have get a new real at some level? That’s searching for solutions to

P(ω) ∩ Lα+1 ) P (ω) ∩ Lα .

Fact 19.6. Suppose P(ω) ∩ Lλ+1 6= P (ω) ∩ Lλ . Then there exists a function f ∈ Lλ+1
such that f : ω → Lλ which is onto.

In other words, the witness that Lλ is countable appears immediately after a new real
number appears.
Next time: β-gap theorem.

66
Evan Chen
[email protected] (Harvard College 20 November 6, 2014
Fall 2014)

§20 November 6, 2014

We’ll finish with some more consequences of condensation.

§20.1 β-Gaps
Definition 20.1. For ordinals α, β a β-gap is a segment

[α, α + β] = {γ ordinal | α ≤ γ ≤ α + β}

such that P (ω) ∩ Lγ is empty for each γ in the interval. In other words, P (ω) ∩
(Lα+β − Lα ).

Theorem 20.2
Assume ZF and V = L. Then for all β < ω1 there is a β-gap [α, α + β] such that
α + β < ω1 .

Proof. We’re interested in gaps less than ω1 since all the reals appear by Lω1 . Like clearly
we can select a gap
[α0 , α0 + β] ∈ Lω2 .
Actually, we are going to use the obvious statement to prove the non-obvious statement.
Using hulls, we can let X ≺ Lω2 be countable such that [α0 , α0 + β] ⊆ X. Hence,


Lω2  P (ω) ∩ (Lα0 +β − Lα0 ) = ∅ .

Hence X satisfies this.

Denoting by π : X → Lλ be the Mostowski collapse; we have λ < ω1 . We set
α = π(α0 ) an ordinal (since “is an ordinal” is ∆0 ). By the definition of π, we see that
α + 1 = π(α0 + 1), and so on. Hence the segment is preserved. Now since Lλ is an
elementary substructure of Lω1 it follows that [α, α + β] is still a gap. To be pedantic,



Lλ  P(π(ω)) ∩ Lπ(α0 )+π(β) \ Lπ(α0 ) = ∅ .

§20.2 A Taste of Fine Structure

Work in ZF + (V = L) and let λ be a limit ordinal. In our proof of GCH, we showed
that for any cardinal κ, if A ⊆ κ appears in Lλ+1 \ Lλ , then in L there is a function
f : κ → Lλ which is onto. This implies that |Lλ | = κ; from this CH follows.
Now we actually claim that the collapsing function exists in Lλ+1 . In other words,
the moment a new subset appears above Lλ , we immediately also see a witness f that
collapses all of Lλ onto κ.

Theorem 20.3
Assume ZF and V = L. Let κ be a cardinal and A any subset of κ such that
A = Lλ+1 \ Lλ for some limit ordinal λ. Then there exists a function f definable
over Lλ which collapses κ onto Lλ .

67
Evan Chen
[email protected] (Harvard College 20 November 6, 2014
Fall 2014)

Proof. We have an order <L of Lλ which is Σ1 -definable. Let ψ be the formula giving
Lλ . By assumption,
A = {x ∈ Lλ | Lλ  ϕA [x, pA ]}
for some formula φA and parameters pA in Lλ . It is important right now that we select
pA to be the <L -least parameter which works for the choice of ϕA .
Using hulls, we let X be an elementary substructure of Lλ such that κ ∪ {pA } ⊆ X.
(Note that X may be a class here.) Via condensation again we take the Mostowski
collapse
π : X → Lλ .
Claim. λ = λ! That means the collapse did not actually collapse anything.

Proof. We wish to show A is in Lλ+1 − Lλ ; hence it suffices to show A is definable over

Lλ . Since κ ⊆ X, π acts as the identity on κ, thus π(A) = A.
Recalling that A consists of ordinals α < κ, we have
α ∈ A ⇐⇒ Lλ  ϕA [α, pA ]
⇐⇒ X  ϕA [α, pA ]
⇐⇒ Lλ  ϕA [π(α), π(pA )]
⇐⇒ Lλ  ϕA [α, π(pA )].
Thus A is definable over Lλ+1 . 

Claim. The map π is actually the identity, implying X = Lλ .

Proof. We use the fact that pA is <L -minimal. By looking at π, we say that π(pA ) is the
<L -least set of parameters for which

π(A) = π(α) ∈ Lλ | Lλ  ϕπ(A) [ϕ(α), ϕ(pA ) .
We showed π(A) = A and π(α) = α. Thus pA = π(pA ). The point is that because
we select pA to be <L -minimal, then when we look at the projected version and notice
π(A) = A and π(α) = α, we find that the resulting π(pA ) gets exactly the same thing.
But in any case we defined Lλ based solely on the images of π(α) for α < κ and the
image π(pA ). Since π is the identity here. Hence X = Lλ . 

This gives, in Lλ+1 , a function f which maps κ onto Lλ .

CORRECTION: Let n ∈ ω such that the formula φA and ψ (which is Σ1 ) are Σn .
Take a X ≺Σn Lλ substructure, not an arbitrary one.

§20.3 The Guessing Sequence

This is .
As a reminder, here is the definition of a club.
Definition. Suppose κ is a cardinal which is regular and uncountable. Then C ⊆ κ is a
club if
• C is closed, meaning it contains all its limit points. Precisely, if λ is a limit ordinal
less than κ, if C ∩ λ is unbounded then λ ∈ C.
• C itself is unbounded.
Exercise 20.4. The intersection of (Cα )γα=0 is a sequence of clubs in κ (where γ < κ) is
also a club in κ.

68
Evan Chen
[email protected] (Harvard College 20 November 6, 2014
Fall 2014)

Example 20.5
Let κ = ω1 . Examples of clubs in κ are ω1 , (α, ω1 ), Lim ∩ ω1 .

Think of clubs in κ as “measure 1 sets”. They are pretty fat.

Definition 20.6. A set S is stationary if S ∩ C 6= ∅ for a clubs C.
Remark 20.7. Since the intersection of two clubs is a club (which is nonempty), every
club is stationary.

Example 20.8
Set κ = ω2 . Then
Sωω2 = {α < ω2 | cof(α) = ω}

Definition 20.9. We define  as the following claim: there exists a sequence (Aα )α<ω1
such that for any A ⊆ ω1 , the set
{α < ω1 | A ∩ α = Aα }
is stationary, and moreover we have the identity Aα ⊆ α. Such a sequence is called a
-sequence or guessing sequence.

Proposition 20.10
 =⇒ CH.

Proof. Let x ∈ ω correspond to a real number. The set of correct guesses Aα is a

stationary set; i.e.
{α < ω1 | x ∩ α = Aα } .
Take any α in here such that α ≥ ω. Thus
Aα = x ∩ α = x.
Thus every x is an element of the universal -sequence
hAα : α < ω1 i .
This is insanely powerful. Basically, every subset A of ω1 is approximated by a
stationary set SA as described. In other words,
[
A= {Aα | α ∈ SA }
and the Aα are guess the initial segments.
We won’t have time to prove this, but this in fact implies the following.

Theorem 20.11
ZF and V = L implies .

Theorem 20.12
 implies the existence of a Suslin tree; hence Suslin’s hypothesis fails.

Hence, assuming ZFC is consistent, ZFC cannot prove SH.

69
Evan Chen
[email protected] (Harvard College 21 November 11, 2014
Fall 2014)

§21 November 11, 2014

Today we start forcing!
We have already shown that

• (Gödel) ZFC + (V = L) =⇒ CH.

• (Jensen) ZFC + (V = L) =⇒ ¬SH (through ).

and moreover, if ZFC is consistent so is ZFC + (V = L).

We now move on the following theorems, assuming ZFC is consistent.

• (Cohen) ZFC does not imply CH.

• (Martin and Solowry) ZFC does not imply ¬SH.

§21.1 Problems with trying to set up Forcing

Showing Cohen’s theorem is very hard; he apparently nearly lost his mind.
To show that ZFC cannot prove ¬CH, we used L. These are the obstructions this
time.

• We can’t use L again for this direction, so we need something new.

• Moreover, L is already the minimal inner model which (1) satisfies ZFC, (2) is
transitive, and (3) contains the ordinals. So we can’t construct some “inner model”
of ZFC.

• It is not possible to prove (in ZFC) that one can extend L to L0 . If we somehow
knew V 6= L we could do this, but the problem is that V = L is in fact consistent
with ZFC.

In light of problem 3, suppose we are in a worst-case scenario in which we are in a

model of ZFC + (V = L). To make things simpler let’s assume there is an inaccessible κ,
so we have models (such as Lκ ) of ZFC.
Let X ≺ Lκ be such that |X| = ω and let π : X → Lλ be the Mostowski collapse. So,
we have Lλ  ZFC + (V = L). But Lλ is countable, so there exists a real number x which
is not in Lλ .
Let’s define Lλ [x] by setting by transfinite recursion

• L0 [x] = x
L [x]
• Lα+1 [x] = Defx α
S
• Lλ [x] = α<λ Lα [x].
S
Finally, we set L[x] = α∈On Lα [x]. Clearly if x ∈ L then L[x] = L. But for our choice
of λ we have Lλ [x] 6= Lλ .
Now suppose we select x ∈ Lα+1 \ Lα (here α > λ), meaning Lα+1 [x] = Lα+1 and
Lγ [x] 6= Lγ for every γ ≤ α.
We cross our fingers that Lλ [x] satisfies ZFC + (V 6= L) Only then do we even have a
hope of refuting CH in this model. Let’s prove the easy part now.

70
Evan Chen
[email protected] (Harvard College 21 November 11, 2014
Fall 2014)

Proposition 21.1
As defined above, Lλ [x]  (V 6= L).

Proof. The constructible part of L in Lλ [x] is the true Lλ , so Lλ [x] can see that x is not
in Lλ .

But the real issue is showing that Lλ [x] actually satisfies ZFC. In fact, this is in general
not even true. To see why, note that |Lλ | is countable, so there is some x ⊆ ω such that
there exists a bijection from x to Lλ . For such an x, we claim that

Lλ [x] 6 ZFC.

Assume not. Then we could decode x to get a surjection f : ω → λ which violates

Replacement, as λ ∈
/ Lλ [x].

§21.2 Coding; good and bad reals

What does coding mean? First, let f ; ω → ω × ω × ω be a canonical bijection. For n ∈ ω,
we write f (n) = (n1 , n2 , n3 ). Now we can use the first coordinate to code the domain
and the second coordinates to code the ∈-relation. Consider real x ⊆ ω such that
def
(D, E) = ({n1 | n ∈ ω} , {(n2 , n3 ) | n ∈ ω})

so that E is a well-founded extensional on D. Since (Lλ , ∈) is countable there exists an

x ⊆ ω such that (D, E) ∼ = (Lλ , ∈) via the Mostowski collapse. Then when Lλ [x] takes
decrypts x and takes the transitive collapse, it gets the entire universe Lλ , which is a
terrible violation of Replacement.
Thus, there exist “bad” reals x over Lλ . But good reals can do anything. You might
find reals x such that Lλ [x]  ZFC, Lλ [x]  ZFC + ¬CH, Lλ [x]  ZFC + (2ℵ0 = ℵ17 ).
And in fact, there are!
Our goal is to show that Lλ knows everything about Lλ [x] other than x. In other
words, we want Lλ can approximate truth in Lλ [x].

§21.3 Ground Models

Definition 21.2. More generally, we work with arbitrary transitive models (M, ∈) of
ZFC, called the ground model, with M being countable. We will let M contain a
partial order P = (P, ≤P ) which will exhibit a generic extension M [G].

We will show that

(A) (Existence) M [G] exists, in the same way we showed Lλ [x] exists. This will be the
easy part.

(B) (Truth) Truth in M [G], meaning truth in M [G] can be suitably approximated in
M.

(C) (Preservation) M [G]  ZFC, which follows from above.

(D) (Control) By picking P ∈ M carefully, we can arrange M [G]  ZFC + whatever.

71
Evan Chen
[email protected] (Harvard College 22 November 13, 2014
Fall 2014)

§22 November 13, 2014

§22.1 The generic extension M [G]
Let M be a transitive model of ZFC (this is stronger than assuming ZFC is consistent).
Let P = (P, ≤) ∈ M be a partial order with a maximal element 1P which lives inside a
model M . The elements of P are called conditions; because they will force things to be
true in M [G].

Definition 22.1. A subset D ⊆ P is dense if for all p ∈ P, there exists a q ∈ D such

that q ≤ p.

Examples of dense subsets include the entire P as well as any downwards slice.

Definition 22.2. For p, q ∈ P we write p k q, saying “p is compatible with q”, if there

is exists r ∈ P with r ≤ p and r ≤ q. Otherwise, we say p and q are incompatible and
write p ⊥ q.

Definition 22.3. We say that A ∈ P is an antichain if for any distinct p and q in A,

we have p ⊥ q. An antichain is called a maximal antichain if it cannot be extended;
i.e. for every r ∈ P, r k p for some p ∈ A.

Definition 22.4. A set G ⊆ P is M -generic if

(a) If p ∈ G, then for all q ≥ p, q ∈ G. In other words, we require that G is

upwards-closed.

(b) For any p, q ∈ G, there exists a point r ∈ G such that r ≤ p and r ≤ q holds
simultaneously. In other words, any pair of elements in G is compatible.

(c) For all D in the model M , if D is dense then G ∩ D 6= ∅.

A set which only satisfies the first two conditions is called a filter.

From the last condition we deduce that G 6= ∅ (since D = P is dense), and hence it
follows 1P ∈ G.
Notice that if p ≥ q, then the sentence q ∈ G tells us more information than the
sentence p ∈ G. In that sense q is a stronger condition. In another sense 1P is the weakest
possible condition, because it tells us nothing about G; we always have 1P ∈ G since G is
upwards closed.

Exercise 22.5. Show that in the definition of “M -generic” we can replace “for all dense
D” with “for all maximal antichains A”.

Sketch of Solution. Show that for every dense set, you can construct an antichain through
it. Need AC.

§22.2 Existence of generic extensions

Theorem 22.6 (Rasiowa-Sikorski Lemma)

Suppose M is a countable transitive model of ZFC and P is a partial order. Then
there exists an M -generic filter G.

72
Evan Chen
[email protected] (Harvard College 22 November 13, 2014
Fall 2014)

Proof. Since M is countable, there are only countably many dense sets (they live in M !),
say
D1 , D2 , . . . , Dn , · · · ∈ M.
Using AC, let p1 ∈ D1 , and then let p2 ≤ p1 such that p2 ∈ D2 (this is possible since D2
is dense). In this way we can inductively exhibit a chain

p1 ≥ p2 ≥ p3 ≥ . . .

with pi ∈ Di for every i.

Hence, we want to generate a filter from the (pi ). Just take the upwards closure – let
G be the set of q ∈ P such that q ≥ pn for some n. By construction, G is a filter (this is
actually trivial). Moreover, G intersects all the dense sets by construction.

This is really just the proof of the Baire category theorem.

§22.3 Splitting posets

This is very important: we are interested in P ∈ M such that if G ⊆ P is an M -generic
then G ∈/ M.

Definition 22.7. A partial order P is splitting if for all p ∈ P, there exists q, r ≤ p

such that q ⊥ r.

Exercise 22.8. Suppose P is splitting. Then if F is a filter such that F ⊆ P and F ∈ M ,

then P \ F is dense in P.

Proof. Consider p ∈ / P \ F ⇐⇒ p ∈ F . Then there exists q, r ≤ p which are not

compatible. Since F is a filter it cannot contain both; we must have one of them outside
F , say q. Hence every element of p ∈ P \ (P \ F ) has an element q ≤ p in P \ F . That’s
enough to prove P \ F is dense.

Of note is the case that G ⊆ P is an M -generic. We must have G ∈

/ M in that case.
This is very important – do not forget this fact!

§22.4 Defining the generic extension

Definition 22.9. Suppose M is a transitive model of ZFC, P = (P, ≤) ∈ M is a partial
order, and G ⊆ P is M -generic. We define the P-names recursively by

Name0 = ∅
Nameα+1 = P(Nameα × P)
[
Nameλ = Nameα .
α<λ
S
Finally, Name = α Nameα .

Definition 22.10. For τ ∈ Name, let rank(τ ) be the least α such that τ ∈ Nameα . We
define the interpretation of τ , denote τ G , using the transfinite recursion

τ G = σ G | hσ, pi ∈ τ and p ∈ G .

73
Evan Chen
[email protected] (Harvard College 22 November 13, 2014
Fall 2014)

Example 22.11
Let us compute

Name0 = ∅
Name1 = P(∅ × P)
= {∅}
Name2 = P({∅} × P)
= P ({h∅, pi | p ∈ P}) .

Compare the corresponding von Neuman universe.

V0 = ∅, V1 = {∅}, V2 = {∅, {∅}} .

Example 22.12
For τ = {∅}, τ G = ∅.
Now suppose
τ = {(∅, p1 ), (∅, p2 ), . . . (∅, pn ), } .
Then (
{∅} if some pi ∈ G
τG = {∅ | h∅, pi ∈ τ and p ∈ G} =
∅ otherwise.
So under interpretation, we get V2 back. In fact, we have in general that
 G
τ | τ ∈ Namen = Vn

for any n ∈ N.

Things change at Nameω .

Example 22.13 (Rasiowa-Sikorski Lemma)

Let M be a countable transitive model. Let P be the partial order on finite binary
sequences ordered by ⊇; this is possible since M satisfies ZFC. This gives a binary
tree starting with 1P as the empty sequence. This is as splitting of a poset as we can
possibly get. A filter then corresponds to some “branch”.
Then a generic G consists of an infinite branch down, which corresponds to an
infinite binary sequence. This G is not in M since P is splitting.
Now this generic G has the property that every binary sequence occurs in G. This
follows since the set of points “ends in 101010” is a dense set.

§22.5 Defining M [G]


Definition 22.14. Set M [G] = τ G | τ ∈ NameM .
Definition 22.15. For x ∈ M , define by transfinite recursion a name
x̌ = {hy̌, 1P i | y ∈ x} .
Moreover, we define another name
Ġ = {hp̌, pi | p ∈ P} .

74
Evan Chen
[email protected] (Harvard College 23 November 18, 2014
Fall 2014)

§23 November 18, 2014

§23.1 Review of Names
Recall that we defined
x̌ = {hy̌, 1P i | y ∈ x} .

Example 23.1
Compute

0̌ = 0


1̌ = 0̌, 1P .

So

(0̌)G = 0
(1̌)G = 1.

In general, (x̌)G = x.

Basically, x̌ is just a copy of x where we add check marks and tag every element with 1P .
Recall also that we define
Ġ = {hp̌, pi | p ∈ P} .
Thus
n o
(Ġ)G = σ G | hσ, pi ∈ Ġ ∧ p ∈ G

= (p̌)G | hp̌, pi , p ∈ P
= {p | p ∈ G}
= G.
Note that x̌, Ġ ∈ M .
Exercise 23.2. Show that rank σ G ≤ n-rank(σ) for all names σ in M .
All the names live inside M . Some names, when interpreted, live inside M . But you
can find names which when interpreted which fall outside M ; an example is Ġ.
The whole point of this construction is that G lives outside M .

§23.2 Filters of Transitive Models

Lemma 23.3
Suppose M is a transitive model of ZFC and P ∈ M is a partial order. Let G ⊆ P is
a filter. Then

(1) M ⊆ M [G]

(2) G ∈ M [G]

(3) M [G] is transitive.

(4) The ordinals of M are the same as the ordinals of M [G].

75
Evan Chen
[email protected] (Harvard College 23 November 18, 2014
Fall 2014)

x)G ∈ M [G]. For (2), note that G = (Ġ)G ∈ M [G].

Proof. For (1), note that x = (b
For (3), consider x ∈ M [G] and y ∈ x. Let σ ∈ NameM such that x = σ G . Then

y ∈ σ G = τ G | hτ, pi ∈ σ ∧ p ∈ G .

Hence y = τ G for some τ . We have τ ∈ σ ∈ M , so transitivity of M gives τ ∈ M . Thus

y = τ G ∈ M [G].
For (4), by absoluteness the ordinals of M are precisely the ordinals in the universe V
which live inside M . In other words,

(On)M = OnV ∩ M.

Analogously
(On)M [G] = OnV ∩ M [G].
Since M ⊆ M [G] we have On ∩ M ≤ On ∩ M [G]. But we also know that for any
σ ∈ NameM we have
rank(σ G ) ≤ n-rank(σ)
and so the inequality is sharp.

Lemma 23.4
Suppose M is a transitive model of ZFC and P ∈ M is a partial order. Let G ⊆ P is
a filter. Then

M [G]  ZF − (Comprehension + Replacement + Powerset).

The missing axioms (including AC) will be recovered when we assume G is in fact a
generic, not just a filter.

Proof. Extensionality and Foundation follow from the fact that M [G] is transitive.
Emptyset and Infinity get inherited from M , since M ⊆ M [G].
M
For Pairing, suppose x, y ∈ M [G]. Let σ1 , σ2 ∈ Name
be such that σ1G = x and
M
σ2G = y. We seek a σ ∈ Name such that σ G = σ1G , σ2G ; the choice

σ = {hσ1 , 1P i , hσ2 , 1P i}

works. And this is why we’re happy; 1P is always in the generic. (Note that we used
Pairing in M .)
Finally, we show Union. Let σ G = x ∈ M [G]; we seek a σ 0 such that σ 0G = ∪σ G . Now
 
∪σ G = ∪ τ G | hτ, pi ∈ σ ∧ p ∈ G . = x | x ∈ τ G for some hτ, pi ∈ σ, p ∈ G .

Now let

σ0 = hτ , ri |∃τ ∃p, q ∈ P such that
hτ , pi ∈ τ ∧
hτ, qi ∈ σ∧

r ≤ p, q .

You can check that this works.

76
Evan Chen
[email protected] (Harvard College 23 November 18, 2014
Fall 2014)

§23.3 Using the Generic Condition

So far we have assumed that

(a) M is a transitive model of ZFC.

(b) G ⊆ P ∈ M is a filter. Assuming G exists, we have the bunch of axioms proven

above.

(c) By further assuming M is countable, we showed that an M -generic filter G exists

for any partial order P. By taking P to be splitting, we deduced that G ∈
/ M.

We want to now show that M [G] implies PowerSet, Comprehension, Replacement,

AC; this will give M [G]  ZFC. Then by carefully choosing G we will try to force
M [G]  ¬CH.
Suppose we want to show Comprehension holds in M [G]. Let σ G ∈ M [G] and
φ(x0 , x1 , . . . , xn ) is a formula in LST. Let τ1G , τ2G , . . . , τnG be parameters. We wish to
show 

M [G]  ∃y y = x | x ∈ σ G and ϕ[x, τ1G , . . . , τnG ] .
To build the appropriate name y, we need to be able to determine

M [G]  ϕ[x, τ1G , . . . , τnG ]

in order to do anything. So we have to use the truths in M (which can’t see G at all) in
order to get (or at least approximate) truths in M [G]. As G varies, the truths in M [G]
vary, and somehow M can’t see these.
We’re running a bit out of time because I asked some intuition questions (some of
the answers which I filled in retroactively to the last lecture), so I’ll just state the big
theorem.

Theorem 23.5 (Fundamental Theorem of Forcing)

Suppose M is a transitive model of ZF. Let P ∈ M be a poset, and G ⊆ P is an
M -generic filter. Then,

(1) Consider σ1 , . . . , σn ∈ NameM , Then

M [G]  ϕ[σ1G , . . . , σnG ]

if and only if there exists a condition p ∈ G such that p forces the sentence
ϕ(σ1 , . . . , σn ). We denote this by p ϕ(σ1 , . . . , σn ).

(2) This forcing relation is (uniformly) definable in M .

77
Evan Chen
[email protected] (Harvard College 24 November 20, 2014
Fall 2014)

§24 November 20, 2014

Note: no class next Tuesday.
Today we will define the forcing relation

p M
P φ(σ1 , . . . , σn )

where p ∈ P and φ ∈ LST in M .

§24.1 Motivation
Recall Cohen’s Fundamental Theorem of Forcing from the previous lecture.
P such that for p ∈ P and σ1 , . . . , σn ∈
Hope: in M we define a semantic relation #M
M
Name the following are equivalent:

(1) p#M
P ϕ(σ1 , . . . , σn )

(2) for every M -generic G ⊆ P if p ∈ G then M [G]  ϕ[σ1G , . . . , σnG ].

Our goal is to show # is .

Compare the completeness theorem: Recall the semantic notation of logical consequence
: we write Γ  ϕ (where Γ ⊆ LST) if any model M which satisfies Γ also satisfies ϕ.
The sentences of Γ themselves are encoded as natural numbers. We’re looking through
all the models M ; this is hence a Π1 -definable notion. Now the completeness theorem
says that Γ  φ if and only if Γ ` φ (i.e. there exists a proof, which can also be encoded
as a natural number N). The former is a Π1 notion over V ; the latter is a Σ1 notion of N.
This conversion is the power here. When we prove Fermat’s Last Theorem in ZFC, we
didn’t go through every model of V ; rather we simply wrote down the proof, which can
be encoded in a natural number. Our hope is that we can do the same thing with the
clause “for all generics G ⊆ P”.

§24.2 Semantics and Desired Conditions

We desire the following conditions on our .

(1) Persistence: If q ≤P p and p φ then q φ.

(2) Consistency: If p φ then p 6 ¬φ.

(3) Completeness (the interesting one): If p 6 φ then ∃q ≤P p such that q ¬φ. In

other words, the set of conditions which force φ are dense.

Obviously # satisfies all these properties. In fact, we claim that along with the first part
of the Fundamental Theorem of Forcing, this gives ∼ = #.

Proposition 24.1
Suppose we’ve defined a that satisfies the three relations above. Then part (1) of
the Fundamental Theorem of Forcing implies

(p ϕ ⇐⇒ p#ϕ).

78
Evan Chen
[email protected] (Harvard College 24 November 20, 2014
Fall 2014)

Proof. First, suppose p ϕ; we want to prove p#ϕ. Assume for contradiction that
p 6 #ϕ. Then there exists a M -generic G ⊆ P such that p ∈ G and

M [G]  ¬φ[σ1G , . . . , σnG ].

Then by the Fundamental Theorem of Forcing there exists q ∈ G so that

q ¬ϕ(σ1 , . . . , σn ).

But now since G is a filter we have some r ≤ p, q. Since p ϕ and q ¬ϕ, Persistence
gives r ϕ, ¬ϕ which is impossible.
Conversely, suppose p#ϕ. We wish to show p ϕ. Suppose not, so p 6 ϕ. By
Completeness, there exists q ≤ p such that q ¬φ. Let G be an M -generic such that
q ∈ G. Thus
M [G]  ¬ϕ[σ1G , . . . , σnG ].
But p ≥ q =⇒ p ∈ G (by upwards closure), and then M [G]  ϕ[σ1G , . . . , σnG ].

§24.3 Defining the Relation

Now that we’re done with that tautology, we need to show satisfies the Fundamental
Theorem of Forcing as well as the three properties Persistence, Consistency, Completeness.
We’ll go by induction on the formula complexity. It turns out in this case that the
atomic formula (base cases) are hardest and themselves require induction on ranks.
For some motivation, let’s consider how we should define p τ1 ∈ τ2 given that we’ve
already defined p  τ1 = τ2 . We need to ensure this holds iff

∀M -generic G with p ∈ G : M [G]  τ1G ∈ τ2G .

So it suffices to ensure that any generic G 3 p hits a condition q which forces τ1G to equal
a member τ G of τ2G . In other words, we want to choose the definition of p τ1 ∈ τ2 to
hold if and only if

{q ∈ P | ∃ hτ, ri ∈ τ2 (q ≤ r ∧ q (τ = τ1 ))}

is dense below in p. In other words, if the set is dense, then the generic must hit q, so it
must hit r, meaning that hτr i ∈ τ2 will get interpreted such that τ G ∈ τ2G , and moreover
the q ∈ G will force τ1 = τ .
Now let’s write down the definition. . . In what follows, the omits the M and P.

Definition 24.2. Let M be a countable transitive model of ZFC. Let P ∈ M be a partial

order. For p ∈ P and ϕ(σ1 , . . . , σn ) a formula in LST, we write τ ϕ(σ1 , . . . , σn ) to
mean the following, defined by induction on formula complexity plus rank.

(1) p τ1 = τ2 means
(i) For all hσ1 , q1 i ∈ τ1 the set
def
Dσ1 ,q1 = {r | r ≤ q1 → ∃ hσ2 , q2 i ∈ τ2 (r ≤ q2 ∧ r (σ1 = σ2 ))} .

is dense in p. (This encodes “τ1 ⊆ τ2 ”.)

(ii) For all hσ2 , q2 i ∈ τ2 , the set Dσ2 ,q2 defined similarly is dense below p.

79
Evan Chen
[email protected] (Harvard College 24 November 20, 2014
Fall 2014)

(2) p τ1 ∈ τ2 means

{q ∈ P | ∃ hτ, ri ∈ τ2 (q ≤ r ∧ q (τ = τ1 ))}

is dense below p.

(3) p ϕ ∧ ψ means p ϕ and p ψ.

(4) p ¬ϕ means ∀q ≤ p, q 6 ϕ.

(5) p ∃xϕ(x, σ1 , . . . , σn ) means that the set

{q | ∃τ (q ϕ(τ, σ1 , . . . , σn )}

is dense below p.

This is definable in M ! All we’ve referred to is P and names, which are in M . (Note
that being dense is definable.) Actually, in parts (3) through (5) of the definition above,
we use induction on formula complexity. But in the atomic cases (1) and (2) we are doing
induction on the ranks of the names.

§24.4 Consequences of the Definition

Exercise 24.3. Show that

(1) If p ϕ and q ≤ p then q ϕ.

(2) If {q | q ϕ} is dense below p then p ϕ.

These are just inductions on the five parts of the definition. The first guy is Persistence;
the second is a sort of converse. Consistency is immediate from the definition (take q = p
in the definition). We can use this to establish Completeness as a corollary.

Corollary 24.4
If p 6 ϕ then for some p ≤ p we have p ¬ϕ.

Proof. By the exercise {q | q ϕ} is not dense below p, meaning for some p ≤ p, every
q ≤ p gives q 6 ϕ. By the definition of p  ¬ϕ, we have p  ¬ϕ.

Now we are in the position to prove the Fundamental Theorem of Forcing. We have
already shown (2) of the Fundamental Theorem. Next lecture we will prove (1).

80
Evan Chen
[email protected] (Harvard College 25 December 2, 2014
Fall 2014)

§25 December 2, 2014

Today we’ll be doing the independence of CH.
Reproduced here again is the fundamental theorem of forcing.
Suppose M is a transitive model of ZF. Let P ∈ M be a poset, and G ⊆ P is
an M -generic filter. Then,
(1) Consider σ1 , . . . , σn ∈ NameM , Then
M [G]  ϕ[σ1G , . . . , σnG ]
if and only if there exists a condition p ∈ G such that p forces the
sentence ϕ(σ1 , . . . , σn ). We denote this by p ϕ(σ1 , . . . , σn ).
(2) This forcing relation is (uniformly) definable in M .
Here “uniformly” means that the definition can be written down without relying on
properties of M ; in other words, there is some fixed definition that we can run in more
models.

§25.1 Forcing tells us that M [G] satisfies ZF

Theorem 25.1
Suppose M is a transitive model of ZF. Let P ∈ M be a poset, and G ⊆ P is an
M -generic filter. Then
M [G]  ZF.
Moreover if M satisfies Choice, so does M [G].

Proof. (Partial) The easy stuff is a lemma from the previous lecture. We always have
M [G] satisfies Extensionality and Foundation. Also check it satisfies EmptySet, Infinity,
Pairing, Union. (We did this in an earlier lecture). Up to here we did not need the
Fundamental Theorem of Forcing.
The issue is different with Comprehension because we need to actually understand
when a formula is true in M [G]. Let’s do Comprehension.
Proof of Comprehension. Suppose σ G , σ1G , . . . , σnG ∈ M [G] are a set and parameters, and
ϕ(x, x1 , . . . , xn ) is an LST formula. We want to show that the set

A = x ∈ σ G | M [G]  ϕ[x, σ1G , . . . , σnG ]
is in M [G].
Note that every element of σ G is of the form ρG for some ρ ∈ dom(σ) (since σ is a
bunch of pairs of names and p’s). So by the Fundamental Theorem of Forcing, we may
write 
A = ρG | ρ ∈ dom(σ) and ∃p ∈ G (p ρ ∈ σ ∧ ϕ(ρ, σ1 , . . . , σn )) .
To show A ∈ M [G] we have to write down a τ such that the name τ G coincides with A.
We claim that
τ = {hρ, pi ∈ dom(σ) × P | p ∧ϕ(ρ, σ1 , . . . , σn )}
is the correct choice. It’s actually clear that τ G = A by construction; the “content” is
showing that τ is in actually a name of M , which follows from Comprehension in M .
So really, the point of the Fundamental Theorem of Forcing is just to let us write down
this τ ; it lets us show that τ is in NameM without actually referencing G. 
Everything else is similar.

81
Evan Chen
[email protected] (Harvard College 25 December 2, 2014
Fall 2014)

§25.2 Forcing V 6= L is really easy

Given M , we are now going to pick a partial order such that M [G] has V 6= L. Note the
contrast here: given a ZFC model M ,

• We could construct LM by running the definition of L in M , obtaining an “internal

model” LM  ZFC + V = L. (By “internal” I mean LM lives inside M )

• We are now going to construct an external model M [G] for which V 6= L.

Note that this is one step on the way to CH; since V = L along with ZFC implies CH,
this is the first necessary step.
Let P ∈ M be a splitting poset like the infinite complete binary tree P = (<ω 2, ⊇)
(look this up).

Theorem 25.2
Let M be a countable transitive model of ZFC. Let P ∈ M be any splitting poset,
and let G ⊆ P be M -generic. Then M [G]  (V 6= L).

Proof. Since P is splitting, G ∈

/ M , so there are elements of M [G] not in M .
Since L has a Σ1 definition, we have

LM [G] = LM

the latter which is contained in M . So LM [G] is contained in M and hence cannot

encompass all of M [G]. Hence M [G] satisfies the sentence ∃x(x ∈
/ L), QED.

You can do this even more dramatically. Given any Lλ  ZFC, take P be a splitting
poset; then Lλ [G] does what we want. Hence we’ve shown that if M is a countable
transitive model of ZFC then we get an M [G] which is a countable transitive model of
ZFC and V 6= L. We’ll strengthen this to the following result.

Theorem 25.3
Suppose ZFC is consistent. Then so is ZFC + (V 6= L).

The proof will use the Compactness Theorem.

Proof. Let N be a model of ZFC (not necessarily transitive or even well-founded). We

want to build an N 0 obeying ZFC + (V 6= L). By the Compactness Theorem it suffices
to show that for any finite fragment F of ZFC, F + (V 6= L) has a model MF .
Issue: N could be ill-founded but the Fundamental Theorem of Forcing is phrased in
terms of transitive models, so we cannot just take the Mostowski collapse. However, we
can just drop into N . Let F be a fragment; let F 0 be a sufficiently bigger fragment of
ZFC which ensures that MF [G]  +(V 6= L). By reflection there exists a λ with VλN  F 0 .
Now N thinks VλN is transitive, so we can take an elementary substructure MF . Even
though MF might be ill-founded, it’s still a model. Use forcing:

MF  F 0 =⇒ MF [G]  F + (V 6= L).

Using compactness, from MF for every F then we get N 0 .

82
Evan Chen
[email protected] (Harvard College 25 December 2, 2014
Fall 2014)

§25.3 Forcing ZFC + ¬CH

Starting with a countable transitive model M , pick P in M such that if G ⊆ P is
M -generic, then M [G]  ¬CH.
We want to choose P ∈ M such that (ℵ2 )M many real numbers appear. However, we
need to make sure that (ℵ2 )M 7→ (ℵ2 )M [G] doesn’t get collapsed. In other words, we want
to make sure that
(ℵ2 )M [G] = (ℵ2 )M .
Then, M [G] will satisfy the sentence “there are at least (ℵ2 )M many real numbers”.
Recall the earlier situation where we set P to be the infinite complete binary tree; its
nodes can be thought of as partial functions n → 2 where n < ω. Then G itself is a path
down this tree; i.e. it can be encoded as a total function G : ω → 2. Thus it corresponds
to a real number G ∈ / M . Now what we’d like to do is add in tons of reals by considering
reals as subsets of N.
Consider in M the following poset:
def
Add (ω2 , ω) = ({p : ω2 × ω → 2, dom(p) < ω} , ⊇)

these elements (conditions) are “partial functions”: we take some finite subset of ω × ω2
and map it into 2 = {0, 1}. Moreover, p ≤P q if dom(p) ⊇ dom(q) and the two functions
agree over dom(q). We claim that this adds ω2 many reals.
Let G ⊆ Add(ω2 , ω) be an M -generic. We claim that, like in the binary case, G can
be encoded as a function ω2 × ω → 2. To see this, consider α ∈ ω2 and n ∈ ω; we have
the dense set
Dα,n = {p ∈ Add(ω2 , ω) | (α, n) ∈ dom(p)}
(this is obviously dense, given any p add in (α, n) if it’s not in there already). So G hits
this dense set, meaning that for every (α, n) there’s a function in G which defines it.
Using the fact that G is upwards closed and a filter, we may as before we may interpret
G as a function ω2 × ω → 2.
For a fixed α define
Gα = {n | G (hα, ni) = 0} ∈ P(N)
which is a real number for each α < ω2 .

Claim 25.4. Let α, β ∈ ω2 . Then Gα 6= Gβ for any α 6= β.

Proof. We claim the following set is dense:

D = {q | ∃n ∈ ω : q ((α, n)) 6= q ((β, n))} .

This is pretty easy to see. Consider p ∈ Add(ω2 , ω), Then you can find an n such that
neither (α, n) nor (β, n) is defined, just because dom(p) is finite. Then you make p0 as p
plus p0 ((α, n)) = 1 and p0 ((β, n)) = 0. Hence the set is dense.
Since G is an M -generic it hits this dense set D. Hence Gα 6= Gβ .

So, start with a countable transitive model M of ZFC. Let

P = (Add(ω2 , ω))M

and let G ⊆ P be an M -generic. Then G adds (ℵ2 )M many new real numbers. Why?
Note G itself is an element of M [G] (interpret Ġ), and hence each Gα is an element of
M [G], but from every Gα we can construct a new real.

83
Evan Chen
[email protected] (Harvard College 26 December 4, 2014
Fall 2014)

§26 December 4, 2014

We’re trying to force not CH now!
Let M be a countable transitive model of ZFC, and we let P be the partial order
P = (Add(ω, ω2 ))M . We showed that M [G] contains (ω2 )M many new subsets of ω. So
it just remains to show that
(ω2 )M [G] = (ω2 )M .
In fact, we will show that M satisfies “κ is a cardinal” if and only if M [G] satisfies the
same sentence. This will entirely resolve the problem of CH, since we can just climb the
chain of ordinals; this will cause all the cardinals to match up.

§26.1 The Countable Chain Condition

Definition 26.1. A poset P has the κ-chain condition (where κ is a cardinal) if all
antichains in P have size less than κ. The special case κ = ℵ1 is called the countable
chain condition.
Here are antichain means that no two pairs of elements are compatible (not comparable,
like the usual meaning).
We are going to show that if the poset has the κ-chain condition then it preserves
all cardinals greater than κ. In particular, the countable chain condition will show that
P preserves all the cardinals. Then, we’ll show that Add(ω, ω2 ) does indeed have this
property. This will complete the proof.

§26.2 Possible Values Argument

The following argument is important enough that we isolate it separately.

Lemma 26.2 (Possible Values Argument)

Suppose M is a transitive model of ZFC and P is a partial order such that P has
the κ-chain condition in M . Let X, Y ∈ M and let

f :X→Y

be some function in M [G], but not necessarily in M . Then there exists a function
F ∈ M such that for any x ∈ X, we have

f (x) ∈ F (x)

and
|F (x)|M < κ.

Intuitively, F tells us that we can narrow down the possible values of f to a set of size κ .

Proof. Let f˙ be a name for f , hence f˙G = f . Similarly, let X̌ and Y̌ be the canonical
names for X and Y . By the Fundamental Theorem of Forcing, we have that for some p,
p ∃ function fˆ : X̌ → Y̌
We work in M now. For each x ∈ X, we let A(x) be a maximal set of pairwise incompatible
conditions conditions q ≤P p such that
 
∀q ∈ A(x)∃y ∈ Y q f˙(x̌) = y̌ .

84
Evan Chen
[email protected] (Harvard College 26 December 4, 2014
Fall 2014)

Note this uses the Axiom of Choice. What this is doing is as follows. We let q0 be a guy
which forces f (x) to be some y0 , q1 to be a guy which forces f (x) to be some other y1 ,
and so on. Then let

F (x) = {y ∈ Y | ∃q ∈ A(x) : q f (x̌) = y̌} .

The chain condition means that |F (x)|M < κ. So now we just have to show we got all
possible values.
Suppose f (x) = y, meaning (f˙)G (x) = y. Let q 0 ∈ G be such that

q 0 (f˙)(x̌) = y̌.

WLOG q 0 ≤ p. Since A(x) is a maximal, there must exist q ∈ A(x) such that q and q 0
are compatible. Let r ≤ q, q 0 accordingly. Hence

r f˙(x̌) = y̌.

But q f˙(x̌) = ž for some z ∈ Y . Thus f (x) = y ∈ F (x).

§26.3 Preserving Cardinals

Definition 26.3. For M a transitive model of ZFC and P ∈ M a poset, we say P
preserves cardinals if ∀G ⊆ P an M -generic, the model M and M [G] agree on the
sentence “κ is a cardinal” for every κ.
In the same way we will talk about P preserving cofinalities, et cetera.
Exercise 26.4. Let M be a transitive model of ZFC. Let P ∈ M be a poset. Show that
the following are equivalent for each λ:
(1) P preserves cofinalities less than or equal to λ.

(2) P preserves regular cardinals less than or equal to λ.

Moreover the same holds if we replace “less than or equal to” by “greater than or equal
to”.
Thus, to show that P preserves cardinality and cofinalities it suffices to show that P
preserves regularity.

Theorem 26.5
Suppose M is a transitive model of ZFC, and P ∈ M is a poset. Suppose M satisfies
the sentence “P has the κ chain condition and κ is regular”. Then P preserves
cardinals and cofinalities greater than or equal to κ.

Proof. It suffices to show that P preserves regularity greater than or equal to κ. Consider
λ > κ which is regular in M , and suppose for contradiction that λ is not regular in M [G].
That’s the same as saying that there is a function f ∈ M [G], f : λ → λ cofinal, with
λ < λ. Then by the Possible Values Argument, there exists a function F ∈ M from
λ → P(λ) such that f (α) ∈ F (α) and |F (α)|M < κ for every α.
Now we work in M again. Note for each α ∈ λ, F (α) is bounded in λ since λ is regular
in M and greater than |F (α)|. Now look at the function λ → λ in M by just

α 7→ ∪F (α) < λ.

This is cofinal in M , contradiction.

85
Evan Chen
[email protected] (Harvard College 26 December 4, 2014
Fall 2014)

§26.4 Infinite Combinatorics

In particular, if P has the countable chain condition then P preserves all the cardinals
(and cofinalities). Therefore, it remains to show that Add(ω, ω2 ) satisfies the countable
chain condition. And this is going to be infinite combinatorics.
Definition 26.6. Suppose C is an uncountable collection of finite sets. C is a ∆-system
if there exists a root R with the condition that for any distinct X and Y in C, we have
X ∩ Y = R.

Lemma 26.7 (∆-System Lemma)

Suppose C is an uncountable collection of finite sets. Then ∃C ⊆ C such that

(1) C is uncountable.

(2) C is a ∆-system.

Proof. There exists an integer n such that C has uncountably many guys of length n.
So we can throw away all the other sets, and just assume that all sets in C have size n.
We now proceed by induction on n. The base case n = 1 is trivial, since we can just
take R = ∅. For the inductive step we consider two cases.
First, assume there exists an a ∈ C contained in uncountably many F ∈ C. Throw
away all the other guys. Then we can just delete a, and apply the inductive hypothesis.
Now assume that for every a, only countably many members of C have a in them.
We claim we can even get a C with R = ∅. First, pick F0 ∈ C. It’s straightforward to
construct an F1 such that F1 ∩ F0 = ∅. And we can just construct F2 , F3 , . . .

Lemma 26.8
For all κ, Add(ω, κ) satisfies the countable chain condition.

Proof. Assume not. Let

{pα : α < ω1 }
be an antichain. Let
C = {dom(pα ) : α < ω1 } .
Let C ⊆ C be such that C is uncountable, and C is a ∆-system which root R. Then let

B = {pα : dom(pα ) ∈ R} .

Each pα ∈ B is a function pα : R → {0, 1}, so there are two that are the same.

§26.5 Finishing Off

Let’s wrap this all up.

Theorem 26.9
Let M be a countable transitive model of ZFC. Then there exists a generic extension

M [G]  ZFC + ¬CH.

86
Evan Chen
[email protected] (Harvard College 26 December 4, 2014
Fall 2014)

Proof. Let P = (Add(ω, ω2 ))M . Let G ⊆ P. Then M [G]  2ω > (ω2 )M . But we also
know that
(ω2 )M [G] = (ω2 )M .
Thus
M [G]  2ω ≥ ℵ2
hence
M [G]  ZFC + ¬CH.

With our work before we showed that we don’t need the countable transitive model
because if we just have any old model N of ZFC, we can construct a countable transitive
model internally.
Hence V = L and CH are both totally independent of ZFC.

§26.6 Concluding Remark

You can use forcing to prove even more things. One can construct models for Suslin’s
Hypothesis, its negation, 2ω = ℵ17 , and so on.
Next semester, we’ll study large candinals as a system of new axioms, and these settle
questions like V = L, and the HOD dichotomy.
OK, fine, a teaser.

Definition 26.10. A cardinal κ is measurable if there exists an elementary embedding

(i.e. preserving truth or satisfaction)

j:V →M

where M is a transitive class, j is not the identity, and κ is the first ordinal moved.

This κ is very powerful and very big. It is regular, inaccessible, . . . Moreover, the
definition can be made first order.

Theorem 26.11 (Scott)

Assume ZFC plus the sentence “there is a measurable cardinal κ”. Then V 6= L.

Proof. Assume for contradiction that V = L. Let κ be the least measurable cardinal in
V = L. Take j : L = V → M witness that κ is measurable.
But V  (V = L), M  (V = L), yet M is transitive. That means M has to be L.
Since L satisfies “κ is the least measurable cardinal” it follows that L also satisfies “j(κ)
is the least measuarble cardinal”. But κ < j(κ). Everything breaks.

87