LECTURE NOTES ON NONSTANDARD ANALYSIS

UCLA SUMMER SCHOOL IN LOGIC

ISAAC GOLDBRING

Contents
1. The hyperreals 3
1.1. Basic facts about the ordered real field 3
1.2. The nonstandard extension 4
1.3. Arithmetic in the hyperreals 5
1.4. The structure of N∗ 7
1.5. More practice with transfer 8
1.6. Problems 9
2. Logical formalisms for nonstandard extensions 10
2.1. Approach 1: The compactness theorem 11
2.2. Approach 2: The ultrapower construction 12
2.3. Problems 16
3. Sequences and series 17
3.1. First results about sequences 17
3.2. Cluster points 19
3.3. Series 21
3.4. Problems 22
4. Continuity 23
4.1. First results about continuity 23
4.2. Uniform continuity 25
4.3. Sequences of functions 27
4.4. Problems 30
5. Differentiation 33
5.1. The derivative 33
5.2. Continuous differentiability 35
5.3. Problems 36
6. Riemann Integration 38
6.1. Hyperfinite Riemann sums and integrability 38
6.2. The Peano Existence Theorem 41
6.3. Problems 43
7. Weekend Problem Set #1 44
8. Many-sorted and Higher-Type Structures 47
8.1. Many-sorted structures 47

Date: November 10, 2014.

1
2 ISAAC GOLDBRING

8.2. Higher-type sorts 48

8.3. Saturation 51
8.4. Useful nonstandard principles 53
8.5. Recap: the nonstandard setting 54
8.6. Problems 54
9. Metric Space Topology 55
9.1. Open and closed sets, compactness, completeness 55
9.2. More about continuity 63
9.3. Compact maps 64
9.4. Problems 65
10. Banach Spaces 67
10.1. Normed spaces 67
10.2. Bounded linear maps 68
10.3. Finite-dimensional spaces and compact linear maps 69
10.4. Problems 71
11. Hilbert Spaces 73
11.1. Inner product spaces 73
11.2. Orthonormal bases and `2 75
11.3. Orthogonal projections 79
11.4. Hyperfinite-dimensional subspaces 82
11.5. Problems 83
12. Weekend Problem Set #2 85
13. The Spectral Theorem for compact hermitian operators 88
13.1. Problems 93
14. The Bernstein-Robinson Theorem 94
15. Measure Theory 101
15.1. General measure theory 101
15.2. Loeb measure 102
15.3. Product measure 103
15.4. Integration 104
15.5. Conditional expectation 104
15.6. Problems 105
16. Szemerédi Regularity Lemma 106
16.1. Problems 108
References 110

Nonstandard analysis was invented by Abraham Robinson in the 1960s as

a way to rescue the naı̈ve use of infinitesimal and infinite elements favored by
mathematicians such as Leibniz and Euler before the advent of the rigorous
methods introduced by Cauchy and Weierstrauss. Indeed, Robinson realized
that the compactness theorem of first-order logic could be used to provide
fields that “logically behaved” like the ordered real field while containing
“ideal” elements such as infinitesimal and infinite elements.
 LECTURE NOTES ON NONSTANDARD ANALYSIS 3

Since its origins, nonstandard analysis has become a powerful mathemat-

ical tool, not only for yielding easier definitions for standard concepts and
providing slick proofs of well-known mathematical theorems, but for also
providing mathematicians with amazing new tools to prove theorems, e.g.
hyperfinite approximation. In addition, by providing useful mathematical
heuristics a precise language to be discussed, many mathematical ideas have
been elucidated greatly.
In these notes, we try and cover a wide spectrum of applications of non-
standard methods. In the first part of these notes, we explain what a non-
standard extension is and we use it to reprove some basic facts from calculus.
We then broaden our nonstandard framework to handle more sophisticated
mathematical situations and begin studying metric space topology. We then
enter functional analysis by discussing Banach and Hilbert spaces. Here we
prove our first serious theorems: the Spectral Theorem for compact hermit-
ian operators and the Bernstein-Robinson Theorem on invariant subspaces;
this latter theorem was the first major theorem whose first proof was non-
standard. We then end by briefly discussing Loeb measure and using it
to give a slick proof of an important combinatorial result, the Szemerédi
Regularity Lemma.
Due to time limitations, there are many beautiful subjects I had to skip.
In particular, I had to omit the nonstandard hull construction (although this
is briefly introduced in the second weekend problem set) as well as applica-
tions of nonstandard analysis to Lie theory (e.g. Hilbert’s fifth problem),
geometric group theory (e.g. asymptotic cones), and commutative algebra
(e.g. bounds in the theory of polynomial rings).
We have borrowed much of our presentation from two main sources: Gold-
blatt’s fantastic book [2] and Davis’ concise [1]. Occasionally, I have bor-
rowed some ideas from Henson’s [3]. The material on Szemerédi’s Regularity
Lemma and the Furstenberg Correspondence come from Terence Tao’s blog.
I would like to thank Bruno De Mendonca Braga and Jonathan Wolf for
pointing out errors in an earlier version of these notes.

1. The hyperreals
1.1. Basic facts about the ordered real field. The ordered field of real
numbers is the structure (R; +, ·, 0, 1, <). We recall some basic properties:
• (Q is dense in R) for every r ∈ R and every  ∈ R>0 , there is q ∈ Q
such that |r − q| < ;
• (Triangle Inequality) for every x, y ∈ R, we have |x + y| ≤ |x| + |y|;
• (Archimedean Property) for every x, y ∈ R>0 , there is n ∈ N such
that nx > y.
Perhaps the most important property of the ordered real field is
Definition 1.1 (Completeness Property). If A ⊆ R is nonempty and bounded
above, then there is a b ∈ R such that:
• for all a ∈ A, we have a ≤ b (b is an upper bound for A);
4 ISAAC GOLDBRING

• if a ≤ c for all a ∈ A, then b ≤ c (b is the least upper bound for A).

Such b is easily seen to be unique and is called the least upper bound of A,
or the supremum of A, and is denote sup(A).
Exercise 1.2. Show that if A is nonempty and bounded below, then A has
a greatest lower bound. The greatest lower bound is also called the infimum
of A and is denoted inf(A).
1.2. The nonstandard extension. In order to start “doing” nonstandard
analysis as quickly as possible, we will postpone a formal construction of the
nonstandard universe. Instead, we will pose some postulates that a nonstan-
dard universe should possess, assume the existence of such a nonstandard
universe, and then begin reasoning in this nonstandard universe. Of course,
after we have seen the merits of some nonstandard reasoning, we will return
and give a couple of rigorous constructions of nonstandard universes.
We will work in a nonstandard universe R∗ that has the following prop-
erties:
(NS1) (R; +, ·, 0, 1, <) is an ordered subfield of (R∗ ; +, ·, 0, 1, <).
(NS2) R∗ has a positive infinitesimal element, that is, there is  ∈ R∗ such
that  > 0 but  < r for every r ∈ R>0 .
(NS3) For every n ∈ N and every function f : Rn → R, there is a “natural
extension” f : (R∗ )n → R∗ . The natural extensions of the field
operations +, · : R2 → R coincide with the field operations in R∗ .
Similarly, for every A ⊆ Rn , there is a subset A∗ ⊆ (R∗ )n such that
A∗ ∩ Rn = A.
(NS4) R∗ , equipped with the above assignment of extensions of functions
and subsets, “behaves logically” like R.
The last item in the above list is, of course, extremely vague and imprecise.
We will need to discuss some logic in order to carefully explain what we mean
by this. Roughly speaking, any statement that is expressible in first-order
logic and mentioning only standard numbers is true in R if and only if it
is true in R∗ . This is often referred to as the Transfer Principle, although,
logically speaking, we are just requiring that R∗ , in a suitable first-order
language, be an elementary extension of R. We will explain this in more
detail later in these notes.
That being said, until we rigorously explain the logical formalism of non-
standard analysis, we should caution the reader that typical transferrable
statements involve quantifiers over numbers and not sets of numbers. For
example, the completeness property for R says that “for all sets of num-
bers A that are nonempty and bounded above, sup(A) exists.” This is an
example of a statement that is not transferrable; see Exercise 1.8 below.
Definition 1.3. R∗ is called the ordered field of hyperreals.
Remark. If f : A → R is a function, where A ⊆ Rn , we would like to also
consider its nonstandard extension f : A∗ → R∗ . We will take care of this
matter shortly.
 LECTURE NOTES ON NONSTANDARD ANALYSIS 5

1.3. Arithmetic in the hyperreals. First, let’s discuss some immediate

consequences of the above postulates. Since R∗ is an ordered field, we can
start performing the field operations to our positive infinitesimal . For
example,  has an additive inverse −, which is then a negative infinitesimal.
Also, we can consider π · ; it is reasonably easy to see that π ·  is also a
positive infinitesimal. (This will also follow from a more general principle
that we will shortly see.)
Since  6= 0, it has a multiplicative inverse −1 . For a given r ∈ R>0 ,
since  < 1r , we see that −1 > r. Since r was an arbitrary positive real
number, we see that −1 is a positive infinite element. And of course, −−1
is a negative infinite
√ element. But now we can continue playing, considering
numbers like 2 · −3 and so on...
And besides algebraic manipulations, we also have transcendental matters
to consider. Indeed, we have the nonstandard extension of the function
sin : R∗ → R∗ ; what is sin()? All in due time... First, let’s make precise
some of the words we have been thus far freely tossing around.
Definition 1.4.
(1) The set of finite hyperreals is
Rfin := {x ∈ R∗ | |x| ≤ n for some n ∈ N}.
(2) The set of infinite hyperreals is Rinf := R∗ \ Rfin .
(3) The set of infinitesimal hyperreals is
1
µ := {x ∈ R∗ | |x| ≤ for all n ∈ N>0 }.
n
The notation µ comes from the more general notion of monad, which we
will encounter later in the notes.
Observe that µ ⊆ Rfin , R ⊆ Rfin , and µ ∩ R = {0}. Also note that if
δ ∈ µ \ {0}, then δ −1 ∈
/ Rfin .
Lemma 1.5.
(1) Rfin is a subring of R∗ : for all x, y ∈ Rfin , x ± y, x · y ∈ Rfin .
(2) µ is an ideal of Rfin : µ is a subring of R∗ and for all x ∈ Rfin and
y ∈ µ, we have xy ∈ µ.
Proof. (1) Fix x, y ∈ Rfin . Choose r, s ∈ R>0 such that |x| ≤ r and |y| ≤ s.
Then |x ± y| ≤ r + s and |xy| ≤ rs, whence x ± y, xy ∈ Rfin .
(2) Suppose x, y ∈ µ. We need to show that x ± y ∈ µ. Fix r ∈ R>0 ;
we need |x ± y| ≤ r. Well, since x, y ∈ µ, we have that |x|, |y| ≤ 2r . Then,
|x ± y| ≤ |x| + |y| ≤ 2r + 2r = r.
Now suppose x ∈ Rfin and y ∈ µ. We need xy ∈ µ. Fix r ∈ R>0 ; we need
|xy| ≤ r. Choose s ∈ R>0 such that |x| ≤ s. Since y ∈ µ, we have |y| ≤ rs .
Thus, |xy| = |x||y| ≤ s · rs = r. 
A natural question now arises: What is the quotient ring Rfin /µ? The
answer will arrive shortly.
6 ISAAC GOLDBRING

Definition 1.6. For x, y ∈ R∗ , we say x and y are infinitely close, written,

x ≈ y, if x − y ∈ µ.
Exercise 1.7.
(1) ≈ is an equivalence relation on R∗ , namely, for all x, y, z ∈ R∗ :
• x ≈ x;
• x ≈ y implies y ≈ x;
• x ≈ y and y ≈ z implies x ≈ z.
(2) ≈ is a congruence relation on Rfin , namely, ≈ is an equivalence re-
lation on Rfin and, for all x, y, u, v ∈ Rfin , if x ≈ u and y ≈ v, then
x ± y ≈ u ± v and xy ≈ uv.
Exercise 1.8. Show that R is a nonempty set that is bounded above in R∗
but that sup(R) does not exist in R∗ . Thus, the completeness property is
not true for R∗ .
The next theorem is of extreme importance.
Theorem 1.9 (Existence of Standard Parts). If r ∈ Rfin , then there is a
unique s ∈ R such that r ≈ s. We call s the standard part of r and write
st(r) = s.
Proof. Uniqueness is immediate: if r ≈ s1 and r ≈ s2 , with s1 , s2 ∈ R, then
s1 ≈ s2 , so s1 − s2 ∈ µ ∩ R = {0}, so s1 = s2 . We now show existence.
Without loss of generality, we can assume r > 0. (Why?) We then set
A = {x ∈ R | x < r}. Since r ∈ Rfin , A is bounded above. Also, 0 ∈ A,
so A is nonempty. Thus, by the Completeness Property, sup(A) exists. Set
s := sup(A). We claim that this is the desired s. Fix δ ∈ R>0 . Since s is an
upper bound for A, s + δ ∈ / A, so r ≤ s + δ. If r ≤ s − δ, then s − δ would
be an upper bound for A, contradicting the fact that s was the least upper
bound for A. Consequently, r ≥ s − δ. It follows that |r − s| ≤ δ. Since
δ ∈ R>0 is arbitrary, it follows that r − s ∈ µ. 
Remark. In proving that standard parts exist, we used the Completeness
Property for R. Later, we will show that the Completeness Property is a
consequence of the existence of standard parts. Thus, an equivalent way to
state the completeness property for R is that standard parts exist.
Remark. Later, when we start studying metric space topology from the
nonstandard perspective, we will call an element nearstandard if it is within
an infinitesimal distance from a standard element. Some metric spaces also
have a natural notion of finite, e.g. R, and, more generally, normed vector
spaces. The equivalence between finite and nearstandard asserted in the
previous theorem will also hold for finite-dimensional normed spaces. In
general, nearstandard points are always finite, but in infinite-dimensional
settings, this inclusion is often strict. This topic will be discussed later in
these notes.
Exercise 1.10. Let x, y ∈ Rfin .
 LECTURE NOTES ON NONSTANDARD ANALYSIS 7

(1) x ≈ y if and only if st(x) = st(y).

(2) If x ≤ y, then st(x) ≤ st(y). The converse of this statement is false;
give an example.
(3) If x ∈ R, then st(x) = x.
Theorem 1.11. st : Rfin → R is a surjective ring homomorphism: for all
x, y ∈ Rfin , st(x + y) = st(x) + st(y) and st(xy) = st(x) st(y).
Proof. This follows immediately from Exercises 1.7(2) and 1.10(3). 
Corollary 1.12. Rfin /µ ∼
= R.
Proof. The kernel of st is precisely µ; now use the First Isomorphism The-
orem for rings. 
Corollary 1.13. µ is a maximal ideal of Rfin .
Proof. This follows from the fact that Rfin /µ is a field. 
Exercise 1.14. Give a direct proof of the last corollary, that is, show di-
rectly that µ is a maximal ideal of Rfin .
1.4. The structure of N∗ . In this subsection, let’s take a brief look at the
picture of N∗ . First, let’s establish that N∗ \ N 6= ∅. To see this, let y ∈ R∗
be positive infinite. Since the statement “for all x ∈ R, if x > 0, then there
is n ∈ N such that x ≤ n” is true in R, the statement “for all x ∈ R∗ , if
x > 0, then there is n ∈ N∗ such that x ≤ n” is true in R∗ by the transfer
principle. Thus, there is N ∈ N∗ such that y ≤ N . However, if N ∈ N,
then y is finite, a contradiction. Thus, N ∈ N∗ \ N. Also note that the same
argument implies that N is positive infinite.
The last sentence of the previous paragraph holds for all N ∈ N∗ \ N:
if N ∈ N∗ \ N, then N is positive infinite. Indeed, the statement “for all
n ∈ N, n ≥ 0” is true in R, so the statement “for all n ∈ N∗ , n ≥ 0” is
true in R∗ , whence N ≥ 0. Also, if N ∈ Rfin , then there is n ∈ N such that
n ≤ N ≤ n + 1. However, the statement “for all m ∈ N, if n ≤ m ≤ n + 1,
then m = n or m = n+1” is true in R; applying the transfer principle to this
statement, we have N = n or N = n + 1, whence N ∈ N, a contradiction.
Thus, N ∈ Rinf .
Now that we know that all nonstandard natural numbers are positive
infinite, let’s ask the question: “How many nonstandard natural numbers
are there?” To examine this question, let’s first establish some notation and
terminology. For N ∈ N∗ , set γ(N ) := {N ± m | m ∈ N}, the galaxy or
archimedean class of N . Clearly, N ∈ N if and only if γ(N ) = Z; we call
this the finite galaxy, while all other galaxies will be referred to as infinite
galaxies.
Lemma 1.15. If N ∈ N∗ \ N, then γ(N ) ⊆ N∗ .
Proof. By transfer, N +1 ∈ N∗ ; by induction, this shows that N +m ∈ N∗ for
all m ∈ N. We now show, also inductively, that N − m ∈ N∗ for all m ∈ N.
8 ISAAC GOLDBRING

Suppose that the result is true for a given m. Notice that N − m 6= 0, else
N = m ∈ N. Applying transfer to the statement “for all n ∈ N, if n 6= 0,
then n − 1 ∈ N,” we see that (N − m) − 1 = N − (m + 1) ∈ N∗ . 
Since we know we have at least one nonstandard natural number, we now
know that we have an entire galaxy of them. Notice that a galaxy looks just
like a copy of Z and that γ(M ) = γ(N ) if and only if |M − N | ∈ N.
Observe that if γ(M ) = γ(M 0 ) and γ(N ) = γ(N 0 ) and γ(M ) 6= γ(N ),
then M < N if and only if M 0 < N 0 . Consequently, we can define an
ordering on galaxies: if γ(M ) 6= γ(N ), then we say γ(M ) < γ(N ) if and
only if M < N . When γ(M ) < γ(N ), we think of M as being infinitely less
than N .
What can be said about the ordering of the set of galaxies? In particular,
are there more than just two galaxies?
Lemma 1.16. The set of infinite galaxies is densely ordered without end-
points, meaning:
(1) there is no largest infinite galaxy, that is, for every M ∈ N∗ \ N,
there is N ∈ N∗ \ N such that γ(M ) < γ(N );
(2) there is no smallest infinite galaxy, that is, for every M ∈ N∗ \ N,
there is N ∈ N∗ \ N such that γ(N ) < γ(M );
(3) between any two infinite galaxies, there is a third (infinite) galaxy,
that is, for every M1 , M2 ∈ N∗ \ N such that γ(M1 ) < γ(M2 ), there
is N ∈ N∗ \ N such that γ(M1 ) < γ(N ) < γ(M2 ).
Proof. (1) Given M ∈ N∗ \ N, we claim that γ(M ) < γ(2M ). Otherwise,
2M = M + m for some m ∈ N, whence M = m, a contradiction.
(2) Since γ(M ) = γ(M −1), we may as well suppose that M is even. Then
γ( M M ∗
2 ) < γ(M ) from the proof of (1); it remains to note that 2 ∈ N \ N.
(3) Again, we may as well assume that M1 and M2 are both even. In this
case, arguing as before, one can see that γ(M1 ) < γ( M1 +M
2
2
) < γ(M2 ). 
Under suitable richness assumptions on the nonstandard extension (to
be discussed later), one can go even further: if (Nα )α<κ is a descending
sequence of nonstandard natural numbers, then there is N ∈ N∗ \ N such
that N < Nα for all α < κ.
1.5. More practice with transfer. In order to get some practice with the
Transfer Principle, we will prove the assertion made in Remark 1.3. More
precisely:
Theorem 1.17. The statement “every finite element of R∗ has a standard
part” implies the Completeness Property of the ordered real field.
Proof. Suppose that A ⊆ R is nonempty and bounded above. We must
show that sup(A) exists. Let b ∈ R be an upper bound for A. Let’s define
a function f : R → R as follows: If r ∈ R \ N, set f (r) = 0. Otherwise, set
f (n) = the least k ∈ Z such that nk is an upper bound for A; such a k exists
 LECTURE NOTES ON NONSTANDARD ANALYSIS 9

by the Archimedean property. Observe that, if m, n ∈ N and n ≤ m, then

f (n) ≤ f (m).
We now consider the nonstandard extension of the function f : R∗ → R∗ .
Fix N ∈ N∗ \ N. The key idea is to understand f (N ). By the transfer
principle, we know that f (N ) is the least element of Z∗ such that f (N
N
)
is
∗
an upper bound for A . Also, by the transfer principle applied to the last
observation of the previous paragraph, if m ∈ N, then f (m) ≤ f (N ).

Claim 1: f (N ) f (N )
N ∈ Rfin : Suppose this is not the case, so, N ∈ Rinf . Since
f (1) ≤ f (N ), we must have that f (N ) is a positive infinite element. Since
the statement “for all n ∈ N, there is a ∈ A such that f (n)−1
n < a” is true in
R, by the Transfer Principle, the statement for all n ∈ N , there is a ∈ A∗
∗

such that f (n)−1

n < a” is true in R∗ . Thus, we may fix a0 ∈ A∗ such that
f (N )−1
N < a0 . Since f (NN)−1 is a positive infinite element (as it differs from
the positive infinite element f (N ) 1
N by the infinitesimal amount N ), we have
that a0 is also a positive infinite element of R∗ . However, the statement “for
all a ∈ A, we have a ≤ b” is true in R, whence the statement “for all a ∈ A∗ ,
we have a ≤ b” is true in R∗ . Consequently, a0 ≤ b, contradicting the fact
that a is a positive infinite element of R∗ .

By Claim 1 and the assumption of the theorem, st( f (N )

N ) exists. Set r :=
st( f (N )
N ).

Claim 2: r is an upper bound for A: To see this, fix a ∈ A. Then, by the

transfer principle, r ≤ f (N ) f (N )
N , whence a = st(a) ≤ st( N ) = r.

Claim 3: r = sup(A): To see this, fix δ ∈ R>0 . We must find a ∈ A such

that r − δ < a. As we showed in the proof of Claim 1, there is a ∈ A∗ such
that f (NN)−1 < a. Since N1 < δ, we have r − δ < f (NN)−1 , whence r − δ < a.
In other words, the statement “there is a ∈ A∗ such that r − δ < a” is true
in R∗ . By the transfer principle, the statement “there is a ∈ A such that
r − δ < a” is true in R, which is precisely what we needed. 

1.6. Problems.
Problem 1.1. Let x, x0 , y, y 0 ∈ Rfin .
(1) Show that x ≈ y if and only if st(x) = st(y).
(2) Show that if x ∈ R, then x = st(x).
(3) Show that x ≤ y implies st(x) ≤ st(y). Show that the converse is
false.
(4) Show that if st(x) < st(y), then x < y. In fact, show that if st(x) <
st(y), then there is r ∈ R such that x < r < y.
(5) Suppose that x ≈ x0 and y ≈ y 0 . Show that:
(a) x ± y ≈ x0 ± y 0 ;
10 ISAAC GOLDBRING

(b) x · y ≈ x0 · y 0 ;
0
(c) xy ≈ xy0 if y 6≈ 0.
Show that (c) can fail if y, y 0 ∈ µ \ {0}.
Problem 1.2. Suppose x, y ∈ R∗ and x ≈ y. Show that if b ∈ Rfin , then
bx ≈ by. Show that this can fail if b ∈
/ Rfin .
Problem 1.3.
(1) Show that R∗ is not complete by finding a nonempty subset of R∗
which is bounded above that does not have a supremum.
(2) Show that if A ⊆ R is unbounded, then A has no least upper bound
when considered as a subset of R∗ . (This may even be how you
solved part (a).)
Problem 1.4. Let F be an ordered field. F is said to be archimedean if for
any x, y ∈ F with x, y > 0, there is n ∈ N such that y < nx. Show that R∗ is
not archimedean. (It is a fact that archimedean ordered fields are complete,
so this problem strengthens the result of the previous problem.)
Problem 1.5. Construct a sequence of subsets (An ) of R such that
∞
[ [∞
∗
( An ) 6= A∗n .
n=1 n=1

Problem 1.6. If F is a field and V is an F -vector space, let dimF (V ) denote

the dimension of V as an F -vector space.
(1) Observe that R∗ is an R∗ -vector space. (More generally, any field F
is naturally an F -vector space.) What is dimR∗ (R∗ )?
(2) Observe that R∗ is also a vector space over R. Show that dimR (R∗ ) =
∞. (Hint: Let x ∈ R∗ \ R. Show that {1, x, x2 , . . . , xn , . . .} is an R-
linearly independent set.)
(3) Show that Rfin is an R-subspace of R∗ . What is dimR (Rfin )?
(4) Show that µ is an R-subspace of R∗ . What is dimR (µ)?
(5) What is dimR (R∗ /Rfin )?
(6) What is dimR (Rfin /µ)?
Problem 1.7. Show that card(N∗ ) ≥ 2ℵ0 ; here card(A) denotes the cardi-
nality of the set A. (Hint: First show that card(Q∗ ) ≥ 2ℵ0 ).
Problem 1.8. Give a direct proof that µ is a maximal ideal of Rfin , that
is, show that if I is an ideal of Rfin such that µ ⊆ I, then I = µ or I = Rfin .

2. Logical formalisms for nonstandard extensions

At this point, you might be wondering one of two things:
(1) What else can I do with these wonderful postulates for nonstandard
extensions? or
(2) Does such a nonstandard extension exist or was everything done in
Section 1 all magical nonsense?
 LECTURE NOTES ON NONSTANDARD ANALYSIS 11

If you asked the former question, you can safely skip this section and
discover the wonders of the nonstandard calculus to come in the follow-
ing sections. (But please, at some point, return and read this section!) If
you asked the latter question, we will ease your trepidations by offering
not one, but two, different logical formalisms for nonstandard extensions.
The first formalism will rely heavily on the Compactness Theorem from
first-order logic, but, modulo that prerequisite, this route is the quickest
way to obtain nonstandard extensions. The second formalism is the Ultra-
product Approach, which is the most algebraic and “mainstream” way to
explain nonstandard methods to “ordinary” mathematicians. Of course, at
some point, some logic must be introduced in the form of Los’ (pronounced
“Wash’s”) theorem, which will be discussed as well.

2.1. Approach 1: The compactness theorem. In this section, some

familiarity with first-order logic is assumed. We let L denote the first-order
language consisting of the following symbols:
• for every r ∈ R, we have a constant symbol cr ;
• for every n > 0 and every f : Rn → R, we have a n-ary function
symbol Ff ;
• for every n > 0 and every A ⊆ Rn , we have a n-ary relation symbol
PA .
We let R be the L-structure whose universe is R and whose symbols are
interpreted in the natural way, namely (cr )R = r, (Ff )R = f , and (PA )R =
A. Let Γ = Th(R) ∪ {c0 < v < cr | r ∈ R>0 }, a set of L-formulae. Then
Γ is finitely satisfiable: for any r1 , . . . , rn ∈ R>0 , choose r ∈ R>0 such that
r < ri for i = 1, . . . , n. Then R |= Th(R) ∪ {c0 < v < cri | i = 1, . . . , n}JrK.
By the Compactness Theorem, Γ is satisfiable, say by some L-structure A
and a ∈ A. Intuitively, a will be our desired positive infinitesimal. But first,
we need R to be a substructure of A; some “abstract nonsense” can help
take care of this.

Exercise 2.1.
(1) The function h : R → A given by h(r) = (cr )A is an injective
homomorphism of L-structures.
(2) Use (1) to find some L-structure A0 isomorphic to A such that R is
a substructure of A0 .

By the result of the previous exercise, we may suppose that R is a sub-

structure of A. In this case, we denote A by R∗ and denote the universe
∗
of R∗ by R∗ . For f : Rn → R, we let f : (R∗ )n → R∗ denote (Ff )R . For
∗
A ⊆ Rn , we set A∗ := (PA )R . These are the extensions that axiom (NS3)
postulates. Since being an ordered field is part of Th(R), we have that R∗
is an ordered field and, since R is a substructure of R∗ , we have that R is
an ordered subfield of R∗ , verifying postulate (NS1). Let  ∈ R∗ be such
12 ISAAC GOLDBRING

that, for every r ∈ R>0 , R∗ |= c0 < v < cr JK. Then  is a positive infinites-
imal, verifying postulate (NS2). Finally, the fact that R∗ |= Th(R) is the
rigorous, precise meaning of postulate (NS4). We have thus proven:
Theorem 2.2. There is a nonstandard universe, namely R∗ .
Exercise 2.3. Suppose that A ⊆ Rn and f : A → R is a function.
(1) Let f1 : Rn → R be an arbitrary extension of f to all of Rn . Define
f ∗ := f1∗ |A∗ . Show that f ∗ is independent of the choice of extension
f1 . (This provides a way to define the nonstandard extensions of
partial functions.)
(2) Set Γf := {(x, y) ∈ Rn+1 | f (x) = y}. Show that Γ∗f is the graph of
a function g : A∗ → R∗ . Show also that g = f ∗ .
The following observation is very useful.
Proposition 2.4. Suppose that ϕ(x1 , . . . , xm ) is an L-formula. Set S :=
{~r ∈ Rm | R |= ϕJ~rK}. Then S ∗ = {~r ∈ (R∗ )m | R∗ |= ϕJ~rK}.
Proof. Just observe that ∀~v (PS ~v ↔ ϕ(~v )) belongs to Th(R). 
Corollary 2.5. N is not a definable set (even with parameters) in R∗ .
Proof. Suppose, towards a contradiction, that there is an L-formula ϕ(x, ~y )
and ~r ∈ R∗ such that N = {a ∈ R∗ | R∗ |= ϕJa, ~rK}. Write down an
L-sentence σ which says that for all ~y , if ϕ(x, ~y ) defines a nonempty set
of natural numbers that is bounded above, then ϕ(x, ~y ) has a maximum.
(Remember you have a symbol PN for the set of natural numbers.) Since
R |= σ, we have R∗ |= σ. Now N is bounded above in R∗ (by an infinite
element). Thus, N should have a maximum in R∗ , which is clearly ridiculous.

Since N∗ is a definable set in R∗ (defined by PN ), we obtain the following
Corollary 2.6. N∗ \ N is not definable in R∗ .
In modern nonstandard analysis parlance, the previous two results would
be phrased as “N and N∗ \ N” are not internal sets. We will discuss internal
sets later in these notes.
2.2. Approach 2: The ultrapower construction. In this approach to
nonstandard analysis, one gives an “explicit” construction of the nonstan-
dard universe in a manner very similar to the explicit construction of the
real numbers from the rational numbers. Recall that a real number can be
viewed as a sequence of rational numbers which we view as better and better
approximations to the real number. Similarly, an element of R∗ should be
viewed as a sequence of real numbers. For example, the sequence (1, 2, 3, . . .)
should represent some infinite element of R∗ .
However, many different sequences of rational numbers represent the same
real number. Thus, a real number is an equivalence class of sequences of
 LECTURE NOTES ON NONSTANDARD ANALYSIS 13

rational numbers (qn ), where (qn ) and (qn0 ) are equivalent if they “represent
the same real number,” or, more formally, if limn→∞ (qn − qn0 ) = 0. We run
into the same issue here: many sequences of real numbers should represent
the same hyperreal number. For instance, it should hopefully be clear that
the sequence (1, 2, 3, . . . , n, n + 1, . . .) and (π, e, −72, 4, 5, 6, . . . , n, n + 1, . . .)
should represent the same (infinite) hyperreal number as they only differ in
a finite number of coordinates.
More generally, we would like to say that two sequences of real numbers
represent the same hyperreal number if they agree on “most” coordinates.
But what is a good notion of “most” coordinates? A first guess might be
that “most” means all but finitely many; it turns out that this guess is
insufficient for our purposes. Instead, we will need a slightly more general
notion of when two sequences agree on a large number of coordinates; this
brings in the notion of a filter.
Definition 2.7. A (proper) filter on N is a set F of subsets of N (that is,
F ⊆ P(N)) such that:
• ∅∈ / F, N ∈ F;
• if A, B ∈ F, then A ∩ B ∈ F ;
• if A ∈ F and A ⊆ B, then B ∈ F.
We think of elements of F as “big” sets (because that’s what filters do,
they catch the big objects). The first and third axioms are (hopefully)
intuitive properties of big sets. Perhaps the second axiom is not as intuitive,
but if one thinks of the complement of a big set as a “small” set, then the
second axiom asserts that the union of two small sets is small (which is
hopefully more intuitive).
Exercise 2.8. Set F := {A ⊆ N | N \ A is finite}. Prove that F is a filter
on N, called the Frechet or cofinite filter on N.
Exercise 2.9. Suppose that D is a set of subsets of N with the finite intersec-
tion property, namely, whenever D1 , . . . , Dn ∈ D, we have D1 ∩ · · · ∩ Dn 6= ∅.
Set
hDi := {E ⊆ N | D1 ∩ · · · Dn ⊆ E for some D1 , . . . , Dn ∈ D}.
Show that hDi is the smallest filter on N containing D, called the filter
generated by D.
If F is a filter on N, then a subset of N cannot be simultaneously big and
small (that is, both it and it’s complement belong to F), but there is no
requirement that it be one of the two. It will be desirable (for reasons that
will become clear in a second) to add this as an additional property:
Definition 2.10. If F is a filter on N, then F is an ultrafilter if, for any
A ⊆ N, either A ∈ F or N \ A ∈ F (but not both!).
Ultrafilters are usually denoted by U. Observe that the Frechet filter on
N is not an ultrafilter since there are sets A ⊆ N such that A and N \ A
14 ISAAC GOLDBRING

are both infinite (e.g. the even numbers). So what is an example of an

ultrafilter on N?
Definition 2.11. Given m ∈ N, set Fm := {A ⊆ N | m ∈ A}.
Exercise 2.12. For m ∈ N, prove that Fm is an ultrafilter on N, called the
principal ultrafilter generated by m.
We say that an ultrafilter U on N is principal if U = Fm for some m ∈
N. Although principal ultrafilters settle the question of the existence of
ultrafilters, they will turn out to be useless for our purposes, as we will see
in a few moments.
Exercise 2.13. Prove that an ultrafilter U on N is principal if and only if
there is a finite set A ⊆ N such that A ∈ U.
Exercise 2.14. Suppose that U is an ultrafilter on N and A1 , . . . , An are
pairwise disjoint subsets of N such that A1 ∪ · · · ∪ An ∈ U. Prove that there
is a unique i ∈ {1, . . . , n} such that Ai ∈ U.
We are now ready to explain the ultrapower construction. Fix an ul-
trafilter U on N. If (an ) and (bn ) are infinite sequences of real numbers,
we say that (an ) and (bn ) are equal modulo U, written (an ) ∼U (bn ), if
{n ∈ N| | an = bn } ∈ U. (This is the precise meaning of when two sequences
agree on “most” coordinates.)
Exercise 2.15. Show that ∼U is an equivalence relation on the set of infinite
sequences of real numbers. (The “ultra” assumption is not used in this
exercsie.)
For an infinite sequence (an ), we write [(an )]U , or simply [(an )], for the
equivalence class of (an ) with respect to ∼U . We let RU denote the set of
∼U -equivalence classes.
We want to turn RU into an ordered field. The natural guess for the field
operations are:
• [(an )] + [(bn )] := [(an + bn )]
• [(an )] · [(bn )] := [(an bn )].
Of course, since we are working with equivalence classes, we need to make
sure that the above operations are well-defined. For example, if (an ) ∼U (a0n )
and (bn ) ∼U (b0n ), we need to know that (an + bn ) ∼U (a0n + b0n ) (and
similarly for multiplication). However, if we set A := {n ∈ N | an = a0n }
and B := {n ∈ N | bn = b0n }, then for n ∈ A ∩ B, an + bn = a0n + b0n ; since
A ∩ B ∈ U and A ∩ B ⊆ {n ∈ N | an + bn = a0n + b0n }, this shows that
(an + bn ) ∼U (a0n + b0n ).
Next we define an order on RU by declaring [(an )] < [(bn )] if and only if
{n ∈ N | an < bn } ∈ U. One needs to verify that this is in fact a linear
order. For example, to check the “linearity” axioms, fix [(an )], [(bn )] ∈ RU .
Set A := {n ∈ N | an < bn }, B := {n ∈ N | an = bn }, and C = {n ∈
N| an > bn }. Then N = A ∪ B ∪ C and A, B, C are pairwise disjoint.
 LECTURE NOTES ON NONSTANDARD ANALYSIS 15

Thus, by Exercise 2.14, exactly one of A, B, C is in U, corresponding to the

situations [(an )] < [(bn )], [(an )] = [(bn )], and [(an )] > [(bn )]. (This is the
first time that we have used the ultra- assumption.)
Set 0 := [(0, 0, 0, . . .)] and 1 := [(1, 1, 1, . . .)].
Theorem 2.16. (RU , +, ·, 0, 1, <) is an ordered field.
Proof. Checking the ordered field axioms is mainly routine. Let us only
verify the existence of multiplicative inverses. Suppose that [(an )] 6= 0, that
is, that {n ∈ N | an 6= 0} ∈ U. Define a new sequence (bn ) by bn = an if
an 6= 0 and bn = 1 if an = 0. (The choice of 1 is irrelevant here; all that
matters is we defined bn 6= 0.) Observe that [(an )] = [(bn )] and bn 6= 0 for all
n. Now [(an )]·[(b−1 −1
n )] = [(bn )]·[(bn )] = 1, whence [(an )] has a multiplicative
inverse. 
We claim that RU (for suitable U) will serve as a nonstandard extension.
First, in order to verify postulate (NS1), we need R to be an ordered subfield
of RU . This can’t literally be true as R is not a subset of RU . Instead, we are
going to define an ordered field embedding d : R → RU (and then, following
ordinary mathematical practice, we pretend that R “is” d(R) ⊆ RU , that is,
we identify R with its image d(R)). The embedding d is given by d(r) =
[(r, r, r, . . .)]. For example, d(0) = 0 and d(1) = 1. It is straightforward to
verify that d is an ordered field embedding, called the diagonal embedding.
In this sense, (NS1) is satisfied.
What about (NS2)? Here, we need to specify an extra condition on our
ultrafilter U. To see this, we first explain why principal ultrafilters are
boring:
Exercise 2.17. Prove that the diagonal embedding d : R → RU is surjective
(that is, an isomorphism) if and only if U is principal.
The whole point of nonstandard extensions is to get new elements, so
principal ultrapowers (that is RU for U principal) do not help us achieve the
goals of nonstandard analysis. So, we should use nonprincipal ultrafilters.
But do they exist?
Theorem 2.18. There exists a nonprincipal ultrafilter on N.
Proof. Let F be the Frechet filter on N. By Zorn’s Lemma, there is a
maximal filter U on N containing F. We leave it as an exercise to the reader
to show that U is a nonprincipal ultrafilter. (Hint: Use Exercise 2.9 to show
that U is an ultrafilter.) 
For the rest of this subsection, fix a nonprincipal ultrafilter U on N. We
can now verify (NS2). For notational reasons, it will be simpler to construct
a positive infinite element α ∈ RU ; then  := α1 will be the desired positive
infinitesimal. We claim that α := [(1, 2, 3, . . .)] is a positive infinite element.
To see this, fix r ∈ R>0 ; we need to verify that d(r) < α, that is, we need
to verify X := {n ∈ N | r < n} ∈ U. Well, if X ∈ / U, then N \ X ∈ U; but
16 ISAAC GOLDBRING

N \ X is finite, whence U is principal by Exercise 2.13. Consequently X ∈ U

and α is infinite.
For axiom (NS3), we define nonstandard extensions of sets and functions
as follows:
• For f : Rn → R, we define f : (RU )n → RU by
f ([(a1m )], . . . , [(anm )]) := [(f (a1m , . . . , anm ))].
• For A ⊆ Rn , we define A∗ ⊆ (RU )n by
([(a1m )], . . . , [(anm )]) ∈ A∗ if and only if {m ∈ N | (a1m , . . . , anm ) ∈ A} ∈ U.
As before, one must check that these operations are well-defined (we leave
this to the reader) and, after identifying R with d(R), these functions and
relations really do “extend” the original functions and relations (again rele-
gated to the lucky reader to verify).
Finally, what about (NS4)? It is here that logic must reenter the picture
in some shape or form. (Up until this point, the nonlogician aiming to use
nonstandard methods via the ultrapower approach has been content.) Let L
be the first-order language described in the previous subsection. We make
RU into an L-structure RU by interpreting Ff and PA as the extensions
defined above. Then the precise formulation of (NS4) is the following:
Theorem 2.19 (Los’). Suppose that ϕ(v1 , . . . , vn ) is an L-formula and
[(a1m )], . . . , [(anm )] ∈ RU . Then

RU |= ϕJ[(a1m )], . . . , [(anm )]K if and only if {m ∈ N | R |= ϕJa1m , . . . , anm K} ∈ U.

Proof. A useful exercise in logic; proceed by induction on the complexity of
ϕ. 

Observe that, as a corollary of Los’ theorem, that RU is an ordered field

(as these axioms are first-order). The analyst trying to refrain from logic
surely avoids Los’ theorem, but in practice, ends up repeatedly verifying its
conclusion on a case-by-case basis.
In summary, we obtain:
Theorem 2.20. RU is a nonstandard universe.
2.3. Problems.
Problem 2.1. Discuss how to make define the nonstandard extension of
functions f : A → B with A ⊆ Rm and B ⊆ Rn .
Problem 2.2. Let A ⊆ Rm and B ⊆ Rn .
(1) Suppose f : A → B is 1-1. Show that if a ∈ A∗ \ A, then f (a) ∈
B ∗ \ B.
(2) Show that A is finite if and only if A∗ = A.
 LECTURE NOTES ON NONSTANDARD ANALYSIS 17

3. Sequences and series

3.1. First results about sequences. OK, so let’s start doing some calcu-
lus nonstandardly. We start by studying sequences and series. A sequence
is a function s : N → R. We often write (sn | n ∈ N) or just (sn ) to de-
note a sequence, where sn := s(n). By Exercise 2.3, we have a nonstandard
extension s : N∗ → R∗ , which we also often denote by (sn | n ∈ N∗ ).

Notation: We write N > N to indicate N ∈ N∗ \ N.

Definition 3.1. (sn ) converges to L, written (sn ) → L or limn→∞ sn = L,
if: for all  ∈ R>0 , there is m ∈ N such that, for all n ∈ N, if n ≥ m, then
|sn − L| < .
We now give our first nonstandard characterization of a standard concept.
Theorem 3.2. sn → L if and only if sN ≈ L for all N > N.
This theorem lends solid ground to the heuristic expression: sn → L if
and only if, for really large N , sN is really close to L.
Proof. (⇒) Suppose sn → L. Fix N > N. We want sN ≈ L. Fix  ∈ R>0 .
We want |sN − L| < . By assumption, there is m ∈ N such that
R |= ∀n ∈ N(n ≥ m → |sn − L| < ).
(Here, we are mixing formal logic with informal notation. If we were being
polite, we would write
R |= ∀n((PN n ∧ P≥ nm) → P< F|∗−∗| Fs ncL c ).
After seeing the formal version, hopefully you will forgive our rudeness and
allow us to write in hybrid statements as above!) Thus, by the Transfer
Principle, R∗ |= ∀n ∈ N∗ (n ≥ m → |sn − L| < ). Since m ∈ N and N > N,
we have N ≥ m. Thus, |sN − L| < , as desired.
(⇐) We now suppose sN ≈ L for N > N. Fix  ∈ R>0 . We need
m ∈ N such that, n ∈ N and n ≥ m implies |sn − L| < . But how
are we to find such m? Well, R∗ knows of such an m (satisfying the ∗-
version of the desired condition). Indeed, if m > N, then n ∈ N∗ and
n ≥ m implies n > N, whence sn ≈ L and, in particular, |sn − L| < . So,
R∗ |= ∃m ∈ N∗ ∀n ∈ N∗ (n ≥ m → |sn − L| < ). Thus, by the Transfer
Principle, R |= ∃m ∈ N∀n ∈ N(n ≥ m → |sn − L| < ), as desired. 
We used transfer in each direction of the previous proof. The first appli-
cation is often called “Upward Transfer” as a fact from below (in the “real
world”) was transferred up to the nonstandard world. Similarly, the second
application is often called “Downward Transfer” for a similar reason.
Theorem 3.3 (Monotone Convergence). Let (sn ) be a sequence.
(1) Suppose (sn ) is bounded above and nondecreasing. Then (sn ) con-
verges to sup{sn | n ∈ N}.
18 ISAAC GOLDBRING

(2) Suppose (sn ) is bounded below and nonincreasing. Then (sn ) con-
verges to inf{sn | n ∈ N}..
Proof. We only prove (1); the proof of (2) is similar. By the previous theo-
rem, it suffices to prove the following

Claim: For any N > N, sN ∈ Rfin and st(sN ) = sup{sn | n ∈ N}.

Proof of Claim: Let b be an upper bound for (sn ), so
R |= ∀n ∈ N(s1 ≤ sn ≤ b).
Thus, by transfer, R∗ |= ∀n ∈ N∗ (s1 ≤ sn ≤ b). Thus, sN ∈ Rfin for
all N ∈ N∗ . Fix N > N and let L := st(sN ). We need to show L =
sup{sn | n ∈ N}. By the transfer principle applied to the nondecreasing
assumption, sn ≤ sN for all n ∈ N, whence sn ≤ st(sN ) = L for all n ∈ N.
Why is L the least upper bound of {sn | n ∈ N}? Let r be an upper bound
for {sn | n ∈ N}. Then R |= ∀n ∈ N(sn ≤ r). By the transfer principle,
R∗ |= ∀n ∈ N∗ (sn ≤ r); in particular, sN ≤ r, whence L = st(sN ) ≤ r. 
Corollary 3.4. Suppose c ∈ (0, 1). Then cn → 0.
Proof. (cn | n ∈ N) is bounded below (by 0) and nonincreasing since c ∈
(0, 1). Thus, cn → L for some L. Fix N > N; by Theorem 3.2, cN ≈ L and
cN +1 ≈ L. So
L ≈ cN +1 = c · cn ≈ c · L.
Since L, cL are both standard numbers, we have cL = L. Since c 6= 1, we
have L = 0. 
Definition 3.5. (sn ) is bounded if there is b ∈ R such that |sn | < b for all
n ∈ N.
Proposition 3.6. (sn ) is bounded if and only if sN ∈ Rfin for all N ∈ N∗ .
Proof. The (⇒) direction follows immediately from transfer. For the con-
verse, suppose that sN ∈ Rfin for all N ∈ N∗ . Fix M > N. Then |sN | ≤ M
for all N ∈ N∗ , so R∗ |= (∃M ∈ N∗ )(∀n ∈ N∗ )(|sN | ≤ M ). Applying the
transfer principle to this statement yields a standard bound for (sn ). 
Definition 3.7. (sn ) is a Cauchy sequence if for all  ∈ R>0 , there is k ∈ N
such that, for all m, n ∈ N, if m, n ≥ k, then |sm − sn | < .
Lemma 3.8. If (sn ) is Cauchy, then (sn ) is bounded.
Proof. We use the nonstandard criteria for boundedness. Fix N > N; we
must show sN ∈ Rfin . Set A := {n ∈ N∗ | |sn − sN | < 1}, so A is a definable
set in R∗ . Since (sn ) is Cauchy, there is k ∈ N such that, for m, p ∈ N, if
m, p ≥ k, then |sm − sp | < 1. Applying the transfer principle to this last
statement, if n > N, then |sn − sN | < 1. Thus, N∗ \ N ⊆ A. Since N∗ \ N
is not definable, we have N∗ \ N ( A. In other words, there is n ∈ A ∩ N.
Since sn ∈ R for n ∈ N, we see that sN ∈ Rfin . 
 LECTURE NOTES ON NONSTANDARD ANALYSIS 19

Exercise 3.9. (sn ) is Cauchy if and only if, for all M, N > N, sM ≈ sN .
Proposition 3.10. (sn ) converges if and only if (sn ) is Cauchy.
Proof. The (⇒) direction is an easy exercise, so we only prove the (⇐)
direction. Suppose (sn ) is Cauchy; then (sn ) is bounded. Fix M > N; then
sM ∈ Rfin . Set L := st(sM ). If N > N is another infinite natural number,
then by the previous exercise, sN ≈ sM , so sN ≈ L. Thus, (sn ) converges
to L. 
3.2. Cluster points. If (sn ) is a sequence and L ∈ R, then we say that
L is a cluster point of (sn ) if, for each  ∈ R>0 , the interval (L − , L + )
contains infinitely many sn ’s. It will be useful for us to write this in another
way: L is a cluster point of (sn ) if and only if, for every  ∈ R>0 , for every
m ∈ N, there is n ∈ N such that n ≥ m and |sn − L| < .
We can also recast this notion in terms of subsequences. A subsequence
of (sn ) is a sequence (tk ) such that there is an increasing function α : N → N
satisfying tk = sα(k) . We often write (snk ) for a subsequence of (sn ), where
nk := α(k).
Exercise 3.11. L is a cluster point of (sn ) if and only if there is a subse-
quence (snk ) of (sn ) that converges to L.
Recall that (sn ) converges to L if sN ≈ L for all N > N. Changing the
quantifier “for all” to “there exists” gives us the notion of cluster point:
Proposition 3.12. L is a cluster point of (sn ) if and only if there is N > N
such that sN ≈ L.
Proof. (⇒): Apply the transfer principle to the definition of cluster point.
Fix  ∈ µ>0 and M > N. Then there is N ∈ N∗ , N ≥ M , such that
|sN − L| < . This is the desired N since N > N and  is infinitesimal.
(⇐): Fix N > N such that sN ≈ L. Fix  ∈ R>0 , m ∈ N. Then
R∗ |= (∃n ∈ N∗ )(n ≥ m ∧ |sn − L| < );
indeed, N witnesses the truth of this quantifier. It remains to apply transfer
to this statement. 
We immediately get the famous:
Corollary 3.13 (Bolzano-Weierstraß). Suppose that (sn ) is bounded. Then
(sn ) has a cluster point.
Proof. Fix N > N. Then since (sn ) is bounded, sN ∈ Rfin . Let L = st(sN );
then L is a cluster point of (sn ) by the last proposition. 
Suppose s = (sn ) is a bounded sequence. Let C(s) denote the set of
cluster points of s. Then, by the previous proposition, we have
C(s) = {L ∈ R | sN ≈ L for some N > N}.
Exercise 3.14. C(s) is a bounded set.
20 ISAAC GOLDBRING

We may thus make the following definitions:

Definition 3.15.
(1) lim sup sn := sup C(s);
(2) lim inf sn := inf C(s).
It turns out that lim sup sn and lim inf sn are themselves cluster points of
s, that is, these suprema and infima are actually max and min:
Proposition 3.16. lim sup sn , lim inf sn ∈ C(s).
Proof. We only prove this for lim sup sn ; the proof for lim inf sn is similar.
For simplicity, set r := lim sup sn . Fix  > 0 and m ∈ N. We need to find
n ∈ N such that |r−sn | < . Since r := sup C(s), there is L ∈ C(s) such that
r −  < L ≤ r. Take N > N such that L = st(sN ). Then r −  < sN < r + .
Now apply transfer. 
Theorem 3.17. (sn ) → L if and only if lim sup sn = lim inf sn = L.
Proof. (sn ) → L if and only if st(sN ) = L for all N > N if and only if
C(s) = {L} if and only if lim sup sn = lim inf sn = L. 
There is an alternate description of lim sup sn and lim inf sn in terms of
tail sets which we now explain. For n ∈ N, set Tn := {sn , sn+1 , sn+2 , . . .},
the nth tailset of s. Set Sn := sup Tn . Since Tn+1 ⊆ Tn , we have Sn+1 ≤ Sn ,
whence (Sn ) is a nonincreasing sequence. Since (sn ) is bounded, so is (Sn ).
Thus, by the Monotone Convergence Theorem, (Sn ) converges to inf Sn .
Theorem 3.18. lim sup sn = lim Sn = inf Sn .
Before we can prove this result, we need a preliminary result:
Lemma 3.19. Let  ∈ R>0 . Then there is m ∈ N such that, for all n ∈ N,
if n ≥ m, then sn < lim sup sn + .
Proof. Fix N > N. Then st(sN ) ≤ lim sup sn , so sN < lim sup sn + . Thus,
any M > N witness that
R∗ |= (∃M ∈ N∗ )(∀N ∈ N∗ )(N ≥ M ⇒ sN < lim sup sn + ).
Now apply the transfer principle. 
Proof. (of Theorem 3.18) Fix m ∈ N. Then by definition, if n ∈ N and
n ≥ m, then sn ≤ Sm . By transfer, this holds for all n ∈ N∗ with n ≥ m.
Take N > N such that lim sup sn = st(sN ). Then sN ≤ Sm , so lim sup sn =
st(sN ) ≤ Sm . Since m ∈ N is arbitrary, we see that lim sup sn ≤ inf Sn . Now
suppose, towards a contradiction, that lim sup sn < inf Sn . Choose  ∈ R>0
such that lim sup sn + < inf Sn . By the previous lemma, there is m ∈ N such
that, for all n ≥ m, sn < lim sup sn + . Thus, Sm ≤ lim sup sn +  < inf Sn ,
a contradiction. 
Exercise 3.20. State and prove an analog of Theorem 3.18 for lim inf sn .
 LECTURE NOTES ON NONSTANDARD ANALYSIS 21

3.3. Series. Recall that an infinite series ∞

P
k=0 ak is
Pconvergent if the se-
quence ofP partial sums (sn ) converges, where sn := nk=0 ak . In this case,
we write ∞ k=0 ak = L if (sn ) converges to L. It is useful to observe that
P n
k=m ak = sn − sm−1 .
Some notation: If N > N, thenPwe have an element sN P of the extended
sequence. We use this to define N a
k=0 k , that is, we set N
a := sN .
PN k=0 k
Similarly, if M > N is also infinite and M ≤ N , we set k=M ak := sN −
sM −1 .
Proposition 3.21.
P∞
only if N
P
(1) k=0 ak = L if andP k=0 ak ≈ L for all N > P
N.
∞
(2) (Cauchy Criteria) k=0 ak converges if and only if N k=M ak ≈ 0
for all M, N > N with M ≤ N .
Proof. (1) follows immediately from the nonstandard characterization of
convergent sequences, while (2) follows immediately from the fact that a
sequence converges if and only if it is Cauchy together with the nonstan-
dard characterization of Cauchy sequence given in Exercise 3.9. 
P
Corollary 3.22. If ak converges, then ak → 0 as k → ∞.
Proof. Let N > N. Then aN = sN − sN −1 ≈ 0 by part (2) of the previous
proposition. 
The usual warning
P is relevant here: the converse of the above corollary is
not true. Indeed, 1/k diverges even though 1/k → 0 as k → ∞.
P P
Definition 3.23. ak converges absolutely if |ak | converges.
P
Exercise 3.24. If ak converges absolutely, then it converges.
Theorem 3.25 (Ratio Test). Suppose ak 6= 0 for all k.
|a | P
(1) If lim sup |ak+1
k|
< 1, then ak converges absolutely.
|ak+1 | P
(2) If lim inf |ak | > 1, then ak diverges.
|ak+1 | |aN +1 |
Proof. (1) Choose N > N such that lim sup |ak | = st( |aN | ). Fix c ∈ R
|ak+1 | |a | |a |
satisfying lim sup < c < 1. If M > N, then
|ak | st( |aMM+1| )
≤ lim sup |ak+1
k|
<
aM +1
c, whence | aM | < c and hence |aM +1 | < c|aM |. By transfer, there is m ∈ N
such that, for all k ≥ m, |ak+1 | < c|ak |. By induction, |ak+m | < ck |am | for
all k ∈ N, whence k+n
P Pn i
k+m |aj | ≤ ( m c )|am | for all m, n ∈ N with m ≤ n.
Fix M, N > N with M ≤ N . We will show that N
P
PN PN −m j M |aj | ≈ 0. Indeed,
P i
M |aj | ≤ ( M −m c )|am |. Use the fact that the geometric series c
P −m j
converges and M − m, N − m > N to conclude that N M −m c ≈ 0.
|ak+1 | |aN +1 |
(2) Choose N > N such that lim inf |ak | = st( |aN | ). For any M > N,
|a | |a |
we have st( |aMM+1| ) ≥ st( |aNN+1| ) > 1, whence |aM +1 | > |aM |. Thus, R∗ |=
22 ISAAC GOLDBRING

(∃N ∈ N∗ )(∀M ∈ N∗ )(M ≥ N → |aM +1 | ≥ |aM |). Apply transfer to this

statement, yielding n ∈ N such that |am+1 | ≥ |am | for all m ≥ n. By
induction, this yields |ak | > |an | for all k > n. Thus, by transfer again,
P if
K > N, then |aK | > |an | > 0, so aK 6≈ 0, whence ak 6→ 0 and thus ak
diverges. 
3.4. Problems.
Problem 3.1. Suppose that sn → L and tn → M . Show that:
(1) (sn ± tn ) → L ± M ;
(2) (csn ) → cL for any c ∈ R;
(3) (sn tn ) → LM ;
(4) ( stnn ) → M
L
if M 6= 0;
(5) If sn ≤ tn for all n ∈ N, then L ≤ M .
Problem 3.2.
(1) Let (sn ) be a sequence. Show that (sn ) is bounded above if and only
if sN is not positive infinite for all N > N. Likewise, show that (sn )
is bounded below if and only if sN is not negative infinite for all
N > N.
(2) Say that (sn ) converges to positive infinity if for every r ∈ R, there
is n ∈ N such that, for all m ∈ N, if m ≥ n, then sm > r. Show that
(sn ) converges to positive infinity if and only if for every N > N,
sN is positive infinite. Define what it means for (sn ) to converge to
negative infinity and give an analogous nonstandard description.
Problem 3.3. Suppose that ∞
P P∞
0 ai = L and 1 bi = M . Show that:
P∞
(1) P0 (ai ± bi ) = L ± M ;
∞
(2) 0 (cai ) = cL for any c ∈ R.

Problem 3.4. Let c ∈ R. Recall the identity

1 − cn+1
1 + c + c2 + · · · + cn = .
1−c
Use this identity to show that ∞
P i
0 c converges if |c| < 1. Conclude that
P n i
m c is infinitesimal if |c| < 1 and m, n > N are such that m ≤ n.
Problem 3.5. Suppose that (rn ), (sn ), (tn ) are three sequences such that
rn ≤ sn ≤ tn for all n ∈ N. Further suppose that L ∈ R is such that rn → L
and tn → L. Show that sn → L. (This is often referred to as the Squeeze
Theorem.)
Problem 3.6. Suppose that s, t are bounded sequences. Show that
lim sup(s + t) ≤ lim sup s + lim sup t.
Problem 3.7. Suppose that ai ≥ 0 for all i ∈ N. Prove that the following
are equivalent:
P∞
(1) 0 ai converges;
 LECTURE NOTES ON NONSTANDARD ANALYSIS 23

PN
(2) a ∈ Rfin for all N > N;
P0N i
(3) 0 ai ∈ Rfin for some N > N.
(Hint: Use the Monotone Convergence Theorem.)
Problem 3.8. Let (sn ) be a sequence. Show that (sn ) is Cauchy if and
only if for every M, N > N, we have sM ≈ sN .
Problem 3.9 (Advanced).
(1) Let ∞
P P∞
0 ai and 0 bi be two series, where an , bn ≥ 0 for all P
i ∈ N.
Suppose that an ≤ bnP for all n > N. Further suppose that ∞ 0 bi
converges. Show that ∞ 0 a i converges.
(2) Show that the condition “an ≤ bn for all n > N” is equivalent to the
condition “there exists k ∈ N such that an ≤ bn for all n ∈ N with
n ≥ k,” i.e. (bn ) eventually dominates (an ).
The result established in this problem is usually called the Comparison
Test.

4. Continuity
4.1. First results about continuity. Let A ⊆ R, f : A → R a function,
and c ∈ A. We say that f is continuous at c if: for all  ∈ R>0 , there is δ ∈
R>0 such that, for all x ∈ R, if x ∈ A and |x − c| < δ, then |f (x) − f (x)| < .
We say that f is continuous if f is continuous at c for all c ∈ A. Here is the
nonstandard characterization of continuity:
Theorem 4.1. Suppose f : A → R and c ∈ A. The following are equivalent:
(1) f is continuous at c;
(2) if x ∈ A∗ and x ≈ c, then f (x) ≈ f (c);
(3) there is δ ∈ µ>0 such that, for all x ∈ A∗ , if |x − c| < δ, then
f (x) ≈ f (c).
The equivalence between (1) and (2) backs up our usual heuristic that f
is continuous at c if, for all x ∈ A really close to c, we have f (x) is really
close to f (c).
Proof. (1) ⇒ (2): Suppose that x ≈ c; we want f (x) ≈ f (c). Fix  ∈ R>0 ;
we want |f (x) − f (c)| < . By (1), there is δ ∈ R>0 such that
R |= ∀x((x ∈ A ∧ |x − c| < δ) → |f (x) − f (c)| < ).
Applying transfer to this statement and realizing that x ≈ c implies |x−c| <
δ yields that |f (x) − f (c)| < .
(2) ⇒ (3) follows by taking δ ∈ µ>0 arbitrary.
(3) ⇒ (1): Fix δ as in (3). Fix  ∈ R>0 . Since x ∈ A∗ and |x − c| < δ
implies f (x) ≈ f (c), and, in particular, |f (x) − f (c)| < , we have
R∗ |= (∃δ ∈ R>0 )(∀x ∈ A)(|x − c| < δ → |f (x) − f (c)| < ).
Apply transfer. 
24 ISAAC GOLDBRING

Example 4.2. Since | sin x| ≤ |x| for x small, we see that, by transfer,
sin  ∈ µ for  ∈ µ. (In other words, sin is continuous at 0.) A similar
argument, shows that cos is continuous at 0, that is, cos  ≈ 1 for  ∈ µ.
Using this and the transfer of the usual trigonometric identities, we can
prove that sin is continuous: if c ∈ R and x ≈ c, write x = c +  for  ∈ µ.
Then
sin x = sin(c + ) = sin c cos  + cos c sin  ≈ sin c · 1 + cos c · 0 = sin c.
Example 4.3. Let (
sin( x1 ) if x 6= 0
f (x) =
0 if x = 0.
Then f is not continuous at 0. Indeeed, let N > N be odd and set x :=
2 Nπ
N π ≈ 0. Then f (x) = sin 2 = 1 6≈ f (0).
However, the function g defined by
(
x sin(1/x) if x 6= 0
g(x) =
0 if x = 0
is continuous at 0. Indeed, suppose x ≈ 0, x 6= 0. Then since | sin(1/x)| ≤ 1,
|g(x)| = |x|| sin(1/x)| ≈ 0.
How about the usual connection between continuity and limits? We say
that limx→c f (x) = L if: for all  ∈ R>0 , there is δ ∈ R>0 such that, for all
x ∈ A, if 0 < |x − c| < δ, then |f (x) − L| < .
Exercise 4.4. Prove that limx→c f (x) = L if and only if, for all x ∈ A∗ , if
x ≈ c but x 6= c, then f (x) ≈ L.
Corollary 4.5. f is continuous at c if and only if limx→c f (x) = f (c).
Proposition 4.6. Suppose that f is continuous at c and g is continuous at
f (c). Then g ◦ f is continuous at c.
Proof. If x ≈ c, then f (x) ≈ f (c), whence g(f (x)) ≈ g(f (c)). 
The following result is fundamental:
Theorem 4.7 (Intermediate Value Theorem). Suppose that f : [a, b] → R
is continuous. Let d be a value strictly in between f (a) and f (b). Then there
is c ∈ (a, b) such that f (c) = d.
The nonstandard proof of the Intermediate Value Theorem will be our
first example of using so-called “hyperfinite methods;” in this case, we will
be using hyperfinite partitions. The idea is to partition the interval [a, b]
into subintervals of width N1 for N > N. Then, logically, this partition will
behave as if it were finite; in particular, we will be able to detect “the last
time” some particular phenomenon happens. However, since we are taking
infinitesimal steps, the change in function value at this turning point will be
infinitesimal. Here are the precise details:
 LECTURE NOTES ON NONSTANDARD ANALYSIS 25

Proof. Without loss of generality, f (a) < f (b), so f (a) < d < f (b). Define a
sequence (sn ) as follows: for n > 0, let {p0 , . . . , pn } denote the partition of
[a, b] into n equal pieces of width b−a
n , so p0 = a and pn = b. Since f (p0 ) < d,
we are entitled to define the number sn := max{pk | f (pk ) < d}, so pk is the
“last time” that f (pk ) < d. Observe that sn < b.
We now fix N > N and claim that c := st(sN ) ∈ [a, b] is as desired,
namely, that f (c) = d. (Note that sN ∈ [a, b], whence st(sN ) is defined.)
Indeed, by transfer, sN < b, whence sN + b−a N ≤ b. Again, by transfer,
f (sN ) < d ≤ f (sN + N ). However, sN + b−a
b−a
N ≈ sN ≈ c. Since f is
continuous at c, we have
b−a
f (c) ≈ f (sN ) < d ≤ f (sN + ) ≈ f (c),
N
whence f (c) ≈ d. Since f (c), d ∈ R, we get that f (c) = d. 
The next fundamental result is also proven using hyperfinite partitions.
Theorem 4.8 (Extreme Value Theorem). Suppose that f : [a, b] → R is
continuous. Then there are c, d ∈ [a, b] such that f (c) ≤ f (x) ≤ f (d) for all
x ∈ [a, b]. (In other words, f achieves it maximum and minimum.)
Proof. We will only prove the existence of the maximum of f . Once again, let
{p0 , p1 , . . . , pn } denote the partition of [a, b] into n equal pieces. This time,
we define sn to be some pk such that f (pj ) ≤ f (pk ) for all j = 0, . . . , n. Fix
N > N and set d := st(sN ) ∈ [a, b]. We claim that f achieves its maximum
at d. (Please appreciate the beauty of this claim: We are partitioning [a, b]
into hyperfinitely many pieces, looking for where the function achieves its
maximum on this hyperfinite set, and claiming that this element is infinitely
close to where the function achieves its maximum on [a, b]. Magic!)
Fix x ∈ [a, b]. We need f (x) ≤ f (d). First, we need to “locate” x in our
hyperfinite partition. Since
k(b − a) (k + 1)(b − a)
R |= (∀n ∈ N)(∃k ∈ N)(0 ≤ k < n ∧ a + ≤ x ≤ a+ ),
n n
by transfer, we can find k ∈ N∗ , 0 ≤ k < N , such that a + k(b−a) N ≤x≤
(k+1)(b−a) k(b−a) (k+1)(b−a)
a+ N . Since the interval [a + N , a + N ] has infinitesimal
width N and f is continuous at x, we see that f (a + k(b−a)
b−a
N ) ≈ f (x) ≈
(k+1)(b−a) k(b−a)
f (a + N ). However, by transfer, f (a + N ) ≤ f (sN ). Since
f (sN ) ≈ f (d) (since f is continuous at d), it follows that f (x) ≤ f (d). 

4.2. Uniform continuity. Recall that f : A → R is continuous if:

(∀x ∈ A)(∀ ∈ R>0 )(∃δ ∈ R>0 )(∀y ∈ A)(|x − y| < δ → |f (x) − f (y)| < ).
The crucial point here is that a particular δ (a priori) depends both on
the given  and the location x. Uniform continuity is what one gets when
one demands that the δ depend only on the given  and not on the location
26 ISAAC GOLDBRING

x. In other words, it is what one gets when one slides the first universal
quantifier in the above display over to the spot after the ∃δ. More formally:
Definition 4.9. f : A → R is uniformly continuous if, for all  ∈ R>0 , there
exists δ ∈ R>0 such that, for all x, y ∈ A, if |x−y| < δ, then |f (x)−f (y)| < .
Using purely symbolic language as above, we see that f : A → R is
uniformly continuous if and only if:
(∀ ∈ R>0 )(∃δ ∈ R>0 )(∀x, y ∈ A)(|x − y| < δ → |f (x) − f (y)| < ).
While uniform continuity is no more difficult to state than ordinary con-
tinuity (as it results from a simple permutation of quantifiers), it is often
difficult for students to first digest. For this reason, uniform continuity
is perhaps one of the best examples of elucidating a standard concept by
nonstandard means. Here is the nonstandard equivalence:
Proposition 4.10. f : A → R is uniformly continuous if and only if,
whenever x, y ∈ A∗ are such that x ≈ y, then f (x) ≈ f (y).
Please make sure that you see how this is different from ordinary continu-
ity. Indeed, f : A → R is continuous if and only if, whenever x, y ∈ A∗ are
such that x ≈ y, and at least one of x and y are standard, then f (x) ≈ f (y).
Thus, uniform continuity demands continuity for the extended part of A as
well.
Exercise 4.11. Prove Proposition 4.10.
Example 4.12. It is now easy to see why f : (0, 1] → R given by f (x) = x1
is not uniformly continuous. Indeed, fix N > N. Then N1 , N 1+1 ∈ (0, 1]∗ ,
1 1 1 1
N ≈ N +1 , but N = f ( N ) 6≈ f ( N +1 ) = N + 1. Note that this calculation
does not contradict the fact that f is continuous. Indeed, since 0 is not in
the domain of f , we would never be calculating f (x) for infinitesimal x when
verifying continuity.
Recall our above heuristic, namely that f : A → R is uniformly continuous
when f is continuous on the “extended part” of A. But sometimes A doesn’t
have an extended part. For example, if A = [a, b], then A∗ is such that
every element is infinitely close to an element of A. It is for this reason
that continuous functions whose domains are closed, bounded intervals are
automatically uniformly continuous:
Theorem 4.13. Suppose f : [a, b] → R is continuous. Then f is uniformly
continuous.
Proof. Suppose x, y ∈ [a, b]∗ , x ≈ y. Then c := st(x) = st(y) ∈ [a, b]. Since
f is continuous at c, we have that f (x) ≈ f (c) ≈ f (y). 
Please compare the proof of the previous theorem with the usual standard
proof. In particular, compare the lengths of these proofs!
 LECTURE NOTES ON NONSTANDARD ANALYSIS 27

4.3. Sequences of functions. We now would like to consider sequences

of functions and different types of convergence of such sequences. For each
n ∈ N, let fn : A → R be a function. Also, let f : A → R be another
function.
Definition 4.14. We say that fn converges to f pointwise if: for each x ∈ A,
the sequence of numbers (fn (x)) converges to the number f (x).
In other words: fn converges to f pointwise if and only if:
(∀x ∈ A)(∀ ∈ R>0 )(∃m ∈ N)(∀n ∈ N)(n ≥ m → |fn (x) − f (x)| < ).
We would like to get a nonstandard characterization of this concept. First,
how should we define the extended sequence (fn )? Clearly, for each n ∈ N,
we get an extended function fn : A∗ → R∗ . But what about infinite N ?
Here’s, the trick: define a function F : N × R → R by F (n, x) = fn (x). We
thus have a nonstandard extension F : N∗ × R∗ → R∗ . We now define, for
n ∈ N∗ , fn : A∗ → R∗ by fn (x) = F (n, x). We should be a bit worried at
this point as, for standard n ∈ N, we have used fn : A∗ → R∗ to denote
two, possibly different functions. Thankfully, by transfer, there is no issue
here: we leave it to the reader to check that, for standard n ∈ N, both uses
of fn : A∗ → R∗ agree. And now we have extended the sequence (fn ) by
defining fn : A∗ → R∗ for n ∈ N∗ \ N.
Exercise 4.15. Prove that fn converges to f pointwise on A if and only if,
for all x ∈ A, for all N > N, fN (x) ≈ f (x).
Once again, we will define the uniform concept by moving the quantifiers
and asking the m in the above definition of pointwise convergence to depend
only on the  and not on the location x.
Definition 4.16. We say that fn converges uniformly to f if, for all  ∈ R>0 ,
there exists m ∈ N, such that, for all n ∈ N, if n ≥ m, then, for all x ∈ A,
|fn (x) − f (x)| < .
In symbols: fn converges uniformly to f if and only if:
(∀ ∈ R>0 )(∃m ∈ N)(∀n ∈ N)(n ≥ m → (∀x ∈ A)(|fn (x) − f (x)| < )).
The way to visualize uniform convergence is the following: for a given ,
draw the strip around f of radius ; then fn converges uniformly to f if and
only if, eventually, all of the functions fn live in the strip of radius  around
f.
Here’s the nonstandard equivalence; notice the similarity in the nonstan-
dard comparison between continuity and uniform continuity.
Proposition 4.17. fn converges uniformly to f if and only if, for all x ∈
A∗ , for all N > N, fN (x) ≈ f (x).
Exercise 4.18. Prove Proposition 4.17.
28 ISAAC GOLDBRING

Example 4.19. Consider fn : [0, 1] → R given by fn (x) = xn . Let f :

[0, 1] → R be defined by
(
0 if x 6= 1
f (x) =
1 if x = 1.

Then fn converges pointwise to f . However, fn does not converge uniformly

to f . Indeed, fix x ∈ [0, 1]∗ such that x < 1 but x ≈ 1. We will find N > N
such that fN (x) = xN 6≈ 0 = f (x). To do this, set
1
X = {n ∈ N∗ | |xn − 1| < }.
2
Note that X is definable and, by continuity of fn for n ∈ N, we have N ⊆ X.
Since N is not definable, there is N ∈ X \N. For this N , we have |xN −1| < 12 ,
whence xN 6≈ 0.
Notice that, in the above example, each fn was continuous but f was not
continuous. This cannot happen if the convergence is uniform:
Theorem 4.20. Suppose that fn : A → R is continuous for each n and that
fn converges to f uniformly. Then f is continuous.
Proof. Fix c ∈ A; we want to show that f is continuous at c. Fix x ∈ A∗ .
By the triangle inequality, we have, for any n ∈ N∗ , that
|f (x) − f (c)| ≤ |f (x) − fn (x)| + |fn (x) − fn (c)| + |fn (c) − f (c)|.
We would like to make all three of the quantities in the right hand side of the
above inequality infinitesimal. Well, if n > N, then by uniform continuity, we
have f (x) ≈ fn (x) and f (c) ≈ fn (c). But we only know that fn (x) ≈ fn (c)
for n ∈ N. Can we arrange for the middle term to be infinitesimal for infinite
n as well? Yes, but we need to be a bit more careful.
Since each fn is continuous at c, we have that the following is true in R:
(∀n ∈ N)(∀ ∈ R>0 )(∃δ ∈ R>0 )(∀y ∈ A)(|y − c| < δ → |fn (y) − fn (c)| < ).
Apply the transfer principle to this result with N > N and  ∈ µ>0 . We
then get a δ ∈ (R∗ )>0 such that, whenever x ∈ A∗ satisfies |x − c| < δ, then
fN (x) ≈ fN (c). Without loss of generality, we may assume that δ ∈ µ>0 .
Thus, for all x ∈ A∗ , whenever |x − c| < δ, we have f (x) ≈ f (c). By the
direction (3) ⇒ (1) of Theorem 4.1, we see that f is continuous at c. 

In the above proof, we actually stumbled upon some important concepts.

Definition 4.21. Suppose that f : A∗ → R∗ is a function (not necessarily
the nonstandard extension of a standard function). Then f is internally
continuous or ∗ continuous if: for all x ∈ A∗ , for all  ∈ (R∗ )>0 , there is
δ ∈ (R∗ )>0 such that, for all y ∈ A∗ , if |x − y| < δ, then |f (x) − f (y)| < .
 LECTURE NOTES ON NONSTANDARD ANALYSIS 29

The terminology internally continuous will make more sense later in these
notes when we define internal sets and functions. Notice that being inter-
nally continuous is just like ordinary continuity, but with everything deco-
rated by stars. By the transfer principle, if f : A → R is continuous, then
its nonstandard extension f : A∗ → R∗ is internally continuous. Another
example is provided by the above proof: if each element of the sequence fn
is continuous, then fn : A∗ → R∗ is internally continuous for all n ∈ N∗ .
While we are on the topic, let’s define another notion of continuity for
nonstandard functions.
Definition 4.22. f : A∗ → R∗ is S-continuous if, for all x, y ∈ A∗ , if x ≈ y,
then f (x) ≈ f (y).
So, for example, if f : A → R is a standard function, then f is uniformly
continuous if and only if f : A∗ → R∗ is S-continuous.
Here is another nice convergence theorem:
Theorem 4.23 (Dini). Suppose that fn : [a, b] → R is continuous, f :
[a, b] → R is continuous and fn converges pointwise to f . Further suppose
that fn (x) is nonincreasing for each x ∈ [a, b]. Then fn converges uniformly
to f .
Proof. By subtracting f from all of the functions involved, we may as well
assume that f is the zero function. Fix N > N and c ∈ [a, b]∗ . We need
to show that fN (c) ≈ 0. Fix n ∈ N. By transfer, 0 ≤ fN (c) ≤ fn (c).
Set d := st(c) ∈ [a, b]. Since fn is continuous, we see that fn (c) ≈ fn (d).
Consequently, fN (c) ∈ Rfin . Taking standard parts, we see that
0 ≤ st(fN (c)) ≤ st(fn (c)) = fn (d).
Since fn (d) → 0 as n → ∞, we see that st(fN (c)) = 0. 
In Theorem 4.20, if we assumed that each fn was uniformly continuous,
could we conclude that f was uniformly continuous? Unfortunately, this is
not true in general. To explain when this is true, we need to introduce a
further notion. First, suppose that each fn : A → R is uniformly continuous.
Then, in symbols, this means:
(∀n ∈ N)(∀ ∈ R>0 )(∃δ ∈ R>0 )(∀x, y ∈ A)(|x−y| < δ → |fn (x)−fn (y)| < ).
Thus, δ can depend on both the  and the particular function fn . Here’s
what happens when we only require the δ to depend on the :
Definition 4.24. The sequence (fn ) is equicontinuous if: for all  ∈ R>0 ,
there is δ ∈ R>0 such that, for all x, y ∈ A, for all n ∈ N, if |x − y| < δ, then
|fn (x) − fn (y)| < .
In other words, (fn ) is equicontinuous if it is uniformly uniformly contin-
uous in the sense that each fn is uniformly continuous and, for a given ,
there is a single δ that witnesses the uniform continuity for each fn .
30 ISAAC GOLDBRING

Exercise 4.25. Prove that (fn ) is equicontinuous if and only if fn : A∗ → R∗

is S-continuous for each n ∈ N∗ .
Equicontinuity guarantees that the limit function is uniformly continuous:
Theorem 4.26. Suppose fn : A → R and f : A → R are functions such that
fn converges pointwise to f . If (fn ) is equicontinuous, then f is uniformly
continuous.
Proof. Fix x, y ∈ A∗ with x ≈ y. We need to prove that f (x) ≈ f (y). Fix
 ∈ µ>0 . Choose n1 , n2 ∈ N∗ such that, for all n ∈ N∗ :
• if n ≥ n1 , then |fn (x) − f (x)| < ;
• if n ≥ n2 , then |fn (y) − f (y)| < .
Such n1 , n2 exist by the transfer of the pointwise converge assumption.
Without loss of generality, we may assume that n1 , n2 > N. Now fix N ∈ N∗
with N > max(n1 , n2 ). Then fN (x) ≈ f (x) and fN (y) ≈ f (y). But, by
Exercise 4.25, fN (x) ≈ fN (y). Thus, f (x) ≈ f (y), as desired. 
Here’s one last equivalence to round out the discussion:
Theorem 4.27. Suppose each fn : [a, b] → R is continuous and fn converges
pointwise to f . Then fn converges uniformly to f if and only if (fn ) is
equicontinuous.
Proof. First suppose that fn converges uniformly to f . Then f is continuous.
Since the domain of each fn and f is a closed, bounded interval, we know
that, by Theorem 4.13, that each fn and f are uniformly continuous. Thus,
for n ∈ N, we have that fn : [a, b]∗ → R∗ is S-continuous. Now suppose that
N > N and x, y ∈ [a, b]∗ are such that x ≈ y. By Exercise 4.25, it suffices to
prove that fN (x) ≈ fN (y). However,
fN (x) ≈ f (x) ≈ f (y) ≈ fN (y),
where the first and last ≈ are true by uniform convergence, while the middle
≈ holds by uniform continuity of f .
Conversely, suppose that (fn ) is equicontinuous; we want to show that fn
converges to f uniformly. To do this, we need to show that, if N > N and
x ∈ [a, b]∗ , that fN (x) ≈ f (x). Let y := st(x). Since fN is S-continuous (by
Exercise 4.25), we know that fN (x) ≈ fN (y). By Theorem 4.26, we know
that f is (uniformly) continuous, whence f (x) ≈ f (y). Since y is standard
and fn converges pointwise to f , we know that fN (y) ≈ f (y). Thus,
fN (x) ≈ fN (y) ≈ f (y) ≈ f (x),
as desired. 
4.4. Problems.
Problem 4.1. Suppose f : A → R and c, L ∈ R.
(1) Show that limx→c+ f (x) = L iff f (x) ≈ L for all x ∈ A∗ with x ≈ c
and x > c.
 LECTURE NOTES ON NONSTANDARD ANALYSIS 31

(2) Show that limx→c− f (x) = L iff f (x) ≈ L for all x ∈ A∗ with x ≈ c
and x < c.
(3) Show that limx→c f (x) = L iff f (x) ≈ L for all x ∈ A∗ with x ≈ c
and x 6= c.
(4) Show that limx→c+ f (x) = L iff f (x) ≈ L for all x ∈ A∗ with x ≈ c
and x > c.
(5) Show that limx→c f (x) = +∞ iff f (x) > N for all x ∈ A∗ with x ≈ c
and x 6= c.
(6) Show that limx→c f (x) = −∞ iff −f (x) > N for all x ∈ A∗ with
x ≈ c and x 6= c.
(7) Show that limx→+∞ f (x) = L iff there is x ∈ A∗ such that x > N
and f (x) ≈ L for all x ∈ A∗ with x > N.
(8) Show that limx→−∞ f (x) = L iff there is x ∈ A∗ such that −x > N
and f (x) ≈ L for all x ∈ A∗ with −x > N.
Problem 4.2. Suppose f, g : A → R and c, L, M ∈ R. Show that:
(1) limx→c f (x) = L iff limx→c+ f (x) = L and limx→c− f (x) = L.
(2) If limx→c f (x) = L and limx→c g(x) = M , then:
(a) limx→c [(f + g)(x)] = L + M ;
(b) limx→c [(f g)(x)] = LM ;
(c) limx→c [ fg (x)] = M
L
if M 6= 0.
Problem 4.3. Suppose that  ∈ µ. Show that:
(1) sin() ≈ 0;
(2) cos() ≈ 1;
(3) tan() ≈ 0.
Problem 4.4. Determine the points of continuity for each of the functions
below.
(1) (
1 if x ∈ Q
g(x) =
0 if x ∈/ Q.
(2) (
x if x ∈ Q
h(x) =
−x if x ∈
/ Q.
(3)
(
0 if x ∈
/Q
j(x) = 1
n if x = m
n ∈ Q in simplest form with n ≥ 1.

Problem 4.5. Suppose f : R → R is uniformly continuous and (sn ) is a

Cauchy sequence. Show that (f (sn )) is also a Cauchy sequence.
Problem 4.6. Suppose f : [a, b] → R is monotonic on [a, b]. Further sup-
pose that f satisfies the Intermediate Value Property, i.e. for any d in
32 ISAAC GOLDBRING

between f (a) and f (b), there is c ∈ [a, b] such that f (c) = d. Show that f is
continuous.

For c ∈ R, define the monad of c to be µ(c) := {x ∈ R∗ | x ≈ c}.

Problem 4.7. There is a meta-principle, called Cauchy’s principle, which

states that if c ∈ R and a property holds for all x within some infinitesimal
distance of c, then the property holds for all x within some real distance of
c. Once we have the logical framework described precisely, we can give this
a more definite meaning. This problem investigates some concrete instances
of this principle.
(1) Suppose c ∈ R and A ⊆ R. Show that µ(c) ⊆ A∗ if and only if there
is  ∈ R>0 such that (c − , c + ) ⊆ A.
(2) Suppose f : A → R and µ(c) ⊆ A∗ . Show that if f is constant on
µ(c), then there is  ∈ R>0 such that (c − , c + ) ⊆ A and f is
constant on (c − , c + ).
(3) Suppose f : A → R is continuous and f (x) ∈ R for all x ∈ A∗ (i.e.
the extension of f to A∗ only takes standard values). Show that f
is locally constant. (Hint: Use part (2).)

Problem 4.8. A function f : R → R is said to be Lipshitz with constant

c if c ∈ R>0 and for any x, y ∈ R, |f (x) − f (y)| ≤ c|x − y|. f is said to
be a contraction mapping if f is Lipshitz with constant c for some c < 1.
The Contraction Mapping Theorem says that a contraction mapping has a
unique fixed point, that is, there is a unique y ∈ R such that f (y) = y. This
theorem has many applications, for example in the proof of the existence
and uniqueness of solutions of differential equations. In this exercise, we will
prove the Contraction Mapping Theorem.
(1) Use nonstandard reasoning to show that a Lipshitz function is uni-
formly continuous.
(2) Show that a contraction mapping cannot have two distinct fixed
points (this proves the uniqueness part of the Contraction Mapping
Theorem).
Let f be a contraction mapping with Lipshitz constant c ∈ (0, 1). Fix x ∈ R.
Recursively define a sequence (sn ) by s0 = x and sn+1 = f (sn ). The goal is
to show that (sn ) converges to a fixed point of f .
(3) Use induction to show that |sn − sn+1 | ≤ cn |s0 − s1 | for all n ∈ N.
(4) Use the formula for the partial sums of a geometric series to show
n
that |s0 − sn | ≤ 1−c 1
1−c |s0 − s1 | ≤ 1−c |s0 − s1 | for all n ∈ N.
(5) Use transfer of the result in (4) to show that sn ∈ Rfin for n > N.
Fix n > N and set y := st(sn ). Show that sn ≈ sn+1 . Conclude that
f (y) = y.
It is interesting to remark that one can start with any x ∈ R and iterate f
to approach the fixed point. The standard proof of the contraction mapping
 LECTURE NOTES ON NONSTANDARD ANALYSIS 33

theorem involves proving the more careful estimate that

cm
|sm − sn | ≤ |s0 − s1 |
1−c
to show that (sn ) is Cauchy and hence converges.

5. Differentiation
5.1. The derivative. We suppose that f : A → R is a function and c ∈ A
is an interior point, that is, there is an interval around c contained in A.
Definition 5.1. f is differentiable at c if limh→0 f (c+h)−f
h
(c)
exists; in this
0
case, the limit is denoted by f (c) and is called the derivative of f at c.
The nonstandard characterization of limits (Exercise 4.4) immediately
gives the following nonstandard characterization of differentiability:
Proposition 5.2. f is differentiable at c with derivative D if and only if
for every  ∈ µ \ {0}, we have f (c+)−f

(c)
≈ D.
Suppose f is differentiable at c. Fix dx ∈ µ \ {0}. (Here we are calling
our infinitesimal dx to match with the usual verbiage of calculus.) Set
df
df := df (c, dx) = f (c + dx) − f (c). Then f 0 (x) = st( dx ). In this way, the
derivative is, in some sense, an actual fraction. (Recall in calculus we are
df
warned not to take the notation dx to seriously and not to treat this as an
actual fraction.)
If the domain of f is an open (possibly infinite) interval, we say that f is
differentiable if it is differentiable at all points of its domain.
Example 5.3. If f (x) = x2 , then
df = f (x + dx) − f (x) = (x + dx)2 − x2 = x2 + 2xdx + (dx)2 − x2 .
df df
Thus, dx = 2x + dx ≈ 2x, whence f 0 (x) exists and f 0 (x) = st( dx ) = 2x.
Proposition 5.4. If f is differentiable at x, then f is continuous at x.
Proof. Suppose y ≈ x; we need f (y) ≈ f (x). Write y = x + dx with dx ∈ µ.
Without loss of generality, dx 6= 0. Then
f (y) − f (x) = df ≈ f 0 (x)dx ≈ 0.

Here are some fundamental properties of derivatives:
Theorem 5.5. Suppose that f, g are differentiable at x and c ∈ R. Then
f + g, cf , and f g are also differentiable at x. If g(x) 6= 0, then fg is also
differentiable at x. Moreover, the derivatives are given by:
(1) (f + g)0 (x) = f 0 (x) + g 0 (x);
(2) (cf )0 (x) = cf 0 (x);
(3) (Product Rule) (f g)0 (x) = f 0 (x)g(x) + f (x)g 0 (x);
34 ISAAC GOLDBRING

f 0 (x)g(x)−f (x)g 0 (x)

(4) (Quotient Rule) ( fg )0 (x) = [g(x)]2
.

Proof. We’ll only prove the Product Rule, leaving the others as exercises.
Fix dx ∈ µ \ {0}. Then

d(f g) = f (x + dx)g(x + dx) − f (x)g(x) = (f (x) + df )(g(x) + dg) − f (x)g(x),

so
d(f g) dg df dg
= f (x) + g(x) + (df ) .
dx dx dx dx
Since f is continuous at x, df ≈ 0. Thus, taking standard parts of the above
display yields (f g)0 (x) = f (x)g 0 (x) + g(x)f 0 (x). 

Now, the Chain Rule is notorious for having many incorrect proofs in
textbooks. Hopefully our nonstandard proof is correct:

Theorem 5.6 (Chain Rule). Suppose that f is differentiable at x and g is

differentiable at f (x). Then g ◦ f is differentiable at x and (g ◦ f )0 (x) =
g 0 (f (x)) · f 0 (x).

Proof. Fix dx ∈ µ \ {0}. Then df = f (x + dx) − f (x). Set h := g ◦ f , so

dh = h(x + dx) − h(x). Then

dh = g(f (x + dx)) − g(f (x)) = g(f (x) + df ) − g(f (x)).

Since f is continuous at x (by Proposition 5.4), df ≈ 0. First suppose that

df 6= 0. Then
dh dh df g(f (x) + df ) − g(f (x)) df
= · = · ≈ g 0 (f (x)) · f 0 (x).
dx df dx df dx

Now suppose that df = 0. Then f 0 (x) = 0 and dh = 0, so dh 0

dx = g (f (x)) ·
0 dh 0 0
f (x). Either way, dx ≈ g (f (x)) · f (x). Since dx was arbitrary, this shows
that h is differentiable at x and h0 (x) = g 0 (f (x)) · f 0 (x). 

The following theorem is the reason we use derivatives in connection with

extrema of functions:

Theorem 5.7. Suppose that f : (a, b) → R is differentiable and f achieves

a local max at c ∈ (a, b), that is, there is  ∈ R>0 such that f (x) ≤ f (c) for
all x ∈ (c − , c + ). Then f 0 (c) = 0.

Proof. Suppose, towards a contradiction, that f 0 (c) > 0. (The case that
f 0 (c) < 0 is similar.) Fix  as in the statement of the theorem and fix
dx ∈ µ>0 . Then c + dx ∈ (c − , c + )∗ , so f (c + dx) ≤ f (c). However,
f 0 (c) ≈ f (c+dx)−f
dx
(c)
, whence f (c+dx)−f
dx
(c)
> 0. Since dx > 0, this forces
f (c + dx) − f (c) > 0, a contradiction. 
 LECTURE NOTES ON NONSTANDARD ANALYSIS 35

5.2. Continuous differentiability. In this subsection, f : (a, b) → R is a

function and c ∈ (a, b). We aim to give a nonstandard characterization of f
being continuously differentiable, that is f 0 exists and is continuous. First,
let’s give an alternate nonstandard characterization of differentiability.
Proposition 5.8. f is differentiable at c if and only if, for all u, u0 ∈ (a, b)∗
with u ≈ u0 ≈ c and u 6= c and u0 6= c, we have
f (u) − f (c) f (u0 ) − f (c)
≈ .
u−c u0 − c
Observe that this proposition has the desirable feature that it does not
need to know the actual value of f 0 (c) in order to verify differentiability;
compare this with Proposition 5.2.
Proof. The (⇒) direction is immediate as both fractions are ≈ f 0 (c). We
now prove the (⇐) direction. First, we show that f is continuous at c. Fix
0
u0 ≈ c, u0 6= c. Set L := | f (uu)−f
0 −c
(c)
| ∈ R∗ . Suppose L ∈ Rfin . Then whenever
u ≈ c, u 6= c, since we have,
f (u) − f (c) f (u0 ) − f (c)
− ≈ 0,
u−c u0 − c
0
it follows that f (u) − f (c) − ( f (uu)−f
0 −c
(c)
)(u − c) ≈ 0, and hence f (u) ≈ f (c).
Now suppose that L ∈ / Rfin . Fix  ∈ µ, so L ∈ µ. Suppose |u − c| < L . Then
f (u0 ) − f (c) f (u0 ) − f (c)
|f (u) − f (c)| ≤ |f (u) − f (c) − ( 0
)(u − c)| + |( )(u − c)|
u −c u0 − c
The first term on the right hand side is infintiesimal and the second term is
≤ . Thus, if |u − c| < L , it follows that f (u) ≈ f (c). By the equivalence of
(1) an (3) in Theorem 4.1, we see that f is continuous at c in this case too.
Now, fix δ ∈ R>0 such that, whenever u ∈ (a, b) and |u − c| < δ, then
|f (u) − f (c)| < 1. Suppose u ∈ (a, b)∗ and |u − c| < δ. If u 6≈ c, then
f (u)−f (c)
u−c ∈ Rfin . If u ≈ c, then f (u)−fu−c
(c)
might be infinite, but all such
quotients (for different u’s infinite close to c) are all infinitely close to one
another. Thus:
f (u) − f (c)
R∗ |= (∃M ∈ R∗ )(∀u ∈ (a, b)∗ )(|u − c| < δ → | | ≤ M ).
u−c
Applying transfer to this statement, we get a real bound M ∈ R>0 such that
f (u) − f (c)
R |= (∀u ∈ (a, b))(|u − c| < δ → | | ≤ M ).
u−c
But now, transfer this statement back to R∗ : for all u ∈ (a, b)∗ , if |u−c| < δ,
then | f (u)−f
u−c
(c)
| ≤ M . In particular, for all u ∈ (a, b)∗ , if u ≈ c, then
f (u)−f (c)
u−c ∈ Rfin . The common standard part of all these fractions is then
f 0 (c). 
We can now give a similar nonstandard criteria for continuous differen-
tiability:
36 ISAAC GOLDBRING

Theorem 5.9. f : (a, b) → R is continuous differentiable if and only if, for

all u, u0 , v, v 0 ∈ (a, b)∗ with u ≈ u0 ≈ v ≈ v 0 and u 6= v and u0 6= v 0 and
st(u) ∈ (a, b), we have
f (u) − f (v) f (u0 ) − f (v 0 )
≈ .
u−v u0 − v 0
Proof. (⇒) Take c ∈ (a, b) and fix u, u0 , v, v 0 ≈ c with u 6= v and u0 6= v 0 .
Then by the transfer of the Mean Value Theorem, there are w (resp. w0 )
between u and v (resp. between u0 and v 0 ) such that
f (u) − f (v) f (u0 ) − f (v 0 )
f 0 (w) = and f 0 (w0 ) = .
u−v u0 − v 0
Since w, w0 ≈ c and f 0 is continuous at c, we have that f 0 (w) ≈ f 0 (w0 ),
yielding the desired result.
(⇐) Fix c ∈ (a, b). By considering u, u0 ≈ c, u 6= c, u0 6= c and v = v 0 = c,
the assumption and the previous proposition imply that f 0 (c) exists. We
now must show that f 0 is continuous at c. Fix u ∈ (a, b)∗ , u ≈ c; we
need f 0 (u) ≈ f 0 (c). Fix  ∈ µ>0 . By the transfer of the definition of
f 0 , there is δ ∈ (R∗ )>0 such that, for all v ∈ (a, b)∗ , if 0 < |u − v| < δ,
then | f (u)−f
u−v
(v)
− f 0 (u)| < . Without loss of generality, we may assume
that δ ∈ µ. Likewise, by the transfer of the fact that f 0 (c) exists, we may
further assume that δ is chosen small enough so that, for all u0 ∈ (a, b)∗ , if
0
0 < |u0 − c| < δ, then | f (uu)−f0 −c
(c)
− f 0 (c)| < . Choose v, u0 ∈ (a, b)∗ such that
0 < |u − v|, |u0 − c| < δ. Then:
f (u) − f (v) f (u0 ) − f (c)
|f 0 (u) − f 0 (c)| ≤  + | − | + .
u−v u0 − c
By assumption, the middle term is infinitesimal, whence f 0 (u) ≈ f 0 (c). 
5.3. Problems.
Problem 5.1. Suppose f : (a, b) → R and c ∈ (a, b). Show that f is
differentiable at c with derivative D if and only if for every  ∈ µ \ {0}, we
have
f (c + ) − f (c)
≈ D.

Problem 5.2. Suppose that  ∈ µ \ {0}. Show that:
sin()
(1)  ≈ 1.
cos()−1
(2)  ≈ 0.
d
(3) dx (cos x) = − sin x.

Problem 5.3. Suppose that f and g are differentiable at x and c ∈ R.

Show that:
(1) cf is differentiable at x and (cf )0 (x) = cf 0 (x).
0 (x)g 0 (x)
(2) If g(x) 6= 0, then fg is differentiable at x and ( fg )0 (x) = g(x)f (x)−f
[g(x)]2
.
 LECTURE NOTES ON NONSTANDARD ANALYSIS 37

Problem 5.4. Suppose f : R → R. We say that f has a local maximum at

c if there is  ∈ R>0 such that f (c) ≥ f (x) for all x ∈ (c − , c + ). Show
that f has a local maximum at c if and only if f (c) ≥ f (x) for all x ≈ c.
Define what it means for f to have a local minimum at c and give a similar
nonstandard description.
Problem 5.5. Suppose f : R → R and c ∈ R. Suppose that f 0 (c) = 0 and
f 00 (c) < 0. Prove that f has a local maximum at c. Hint: Use transfer and
the Mean Value Theorem to conclude that, for any x ≈ c, x 6= c, there is t
in between x and c such that
f (c) − f (x)
f 0 (t) = .
c−x
(This result is often called the Second Derivative Test.)
The next few problems deal with the Taylor Series of a function. Suppose
f : R → R and a ∈ R. Further suppose that the nth derivative of f at a
exists for all n. Then the Taylor series for f centered at a is the power series
∞ 00
X f (i) (x) i 0 f (a)
(x − a) = f (a) + f (a)(x − a) + (x − a)2 + · · · .
i! 2
i=0

The Taylor series for f need not converge at some (or even any) x. Even
if the Taylor series for f does converge at some x, it need not converge to
f (x). For example, suppose
( −1
e x2 if x 6= 0
f (x) =
0 if x = 0.

Then f (n) (x) = 0 for all n ∈ N and x ∈ R, whence the Taylor series for f
is identically 0.
Fix n ≥ 0. The nth degree Taylor polynomial for f centered at a is the
polynomial
n
X f (i) (x) f (n) (x)
pn (x) = (x − a)i = f (a) + f 0 (a)(x − a) + · · · + (x − a)n .
i! n!
i=0

Problem 5.6. For a given x, show that the Taylor series for f at x converges
to f if and only if pn (x) ≈ f (x) for all n > N.
Set Rn (x) := f (x) − pn (x). It follows from the previous problem that
the Taylor series for f at x converges to f (x) if and only if Rn (x) ≈ 0
for all n > N. There is a theorem due to Lagrange that says that if f is
(n + 1)-times differentiable on some open interval I containing a, then for
each x ∈ I, there is a c between a and x such that
f (n+1) (c)
Rn (x) = (x − a)n+1 .
(n + 1)!
38 ISAAC GOLDBRING

Problem 5.7. Suppose that f (n) (x) exists for all n ∈ N and x ∈ I (we say
f is infinitely differentiable on I in this situation). Discuss how to make
sense of f (n) (x) for n ∈ N∗ and x ∈ I ∗ . Show that for all x ∈ I ∗ and n ∈ N∗
with n ≥ 1, we have
f (n) (a)
pn (x) − pn−1 (x) = (x − a)n .
n!
Problem 5.8. Suppose that f is infinitely differentiable on I and x ∈ I.
(1) Suppose that
f (n+1) (c)
(x − a)n+1
(n + 1)!
is infinitesimal for every n > N and every c ∈ R∗ such that c is
between a and x. Show that the Taylor series for f at x converges
to f (x).
(2) Use part (1) to show that the Taylor series for cos x centered at
a = 0 (otherwise known as the Maclaurin series for cos x) converges
to cos x for all x ∈ R.
(3) Use part (1) to show that the Maclaurin series for ex converges to
ex for all x ∈ R.
Problem 5.9. Suppose that f (n) exists for all real numbers in some open
interval I. Further suppose that f (n) is continuous at x ∈ I. Show that for
any infinitesimal ∆x, there is an infinitesimal  such that
f 00 (x) f (n) (x)
f (x + ∆x) = f (x) + f 0 (x)∆x + + ··· + (∆x)n + (∆x)n .
2 n!
(Hint: Consider the Lagrange form of the remainder Rn .)
Problem 5.10. There is another form of the remainder Rn , which states
that
f (n) (c) − f (n) (a)
Rn (x) = (x − a)n+1
(c − a)(n + 1)!
for some c in between a and x. Apply this form of the remainder to Rn−1 to
prove the result from Problem 5.9 without using the assumption that f (n)
is continuous.

6. Riemann Integration
6.1. Hyperfinite Riemann sums and integrability. The Riemann in-
tegral has a particularly slick description in terms of hyperfinite partitions.
Indeed, one takes a Riemann sum with respect to rectangles of infinitesi-
mal width; this Riemann sum will then be infinitely close to the Riemann
integral.
First, let’s recall some standard notions. Throughout, we assume that
f : [a, b] → R is a bounded function. A partition of [a, b] is a finite ordered
set P = {x0 , . . . , xn } such that a = x0 < x1 < . . . < xn = b. A partition P2
is a refinement of a partition P1 if P1 ⊆ P2 . (So P2 is obtained from P1 by
 LECTURE NOTES ON NONSTANDARD ANALYSIS 39

further subdivision.) For a partition P = {x0 , . . . , xn } of [a, b], we define,

for i = 1, . . . , n:
• Mi := Mi (P ) := sup{f (x) | x ∈ [xi−1 , xi ]}
• mi := mi (P ) := inf{f (x) | x ∈ [xi−1 , xi ]}
• ∆xi := xi − xi−1 .
Definition 6.1.
(1) The upper Riemann sum of f with respect to P is
n
X
U (f, P ) := Mi ∆xi .
i=1

(2) The lower Riemann sum of f with respect to P is

n
X
L(f, P ) := mi ∆xi .
i=1

(3) The left-hand Riemann sum of f with respect to P is

n
X
S(f, P ) := f (xi−1 )∆xi .
i=1

Set M := sup{f (x) | x ∈ [a, b]} and m := inf{f (x) | x ∈ [a, b]}.
Exercise 6.2.
(1) For any partition P , we have
m(b − a) ≤ L(f, P ) ≤ S(f, P ) ≤ U (f, P ) ≤ M (b − a).
(2) If P2 is a refinement of P1 , then
L(f, P1 ) ≤ L(f, P2 ) ≤ U (f, P2 ) ≤ U (f, P1 ).
(3) For any two partitions P1 and P2 , L(f, P1 ) ≤ U (f, P2 ).
Part (2) of the previous exercise motivates the following
Definition 6.3.
(1) The lower Riemann integral of f is
L(f ) := sup{L(f, P ) | P a partition of [a, b]}.
(2) The upper Riemann integral of f is
U (f ) := inf{L(f, P ) | P a partition of [a, b]}.
By Exercise 6.2(1) and (3), we see that
m(b − a) ≤ L(f ) ≤ U (f ) ≤ M (b − a).
We say that f is Riemann integrable if U (f ) = L(f ). In this case, we set
Rb
a f dx := U (f ) = L(f ).
The following Cauchy-type criterion for integrability is quite useful:
40 ISAAC GOLDBRING

Exercise 6.4 (Riemann Lemma). f is Riemann integrable if and only if, for
every  ∈ R>0 , there is a partition P of [a, b] such that U (f, P )−L(f, P ) < .
Fix ∆x ∈ R>0 . Set P∆x = {x0 , . . . , xn }, where [a, x1 ], [x1 , x2 ], . . . , [xn−2 , xn−1 ]
all have equal length ∆x and [xn−1 , xn ] is the “leftover” piece. (If ∆x ≥ b−a,
then P∆x = {a, b}.) We thus get a function U (f, ·) : R>0 → R by setting
U (f, ∆x) := U (f, P∆x ). In a similar manner, we get functions L(f, ·) and
S(f, ·).
Here is another Cauchy-type criterion for integrability:
Exercise 6.5. f is Riemann integrable if and only if, for every  ∈ R>0 ,
there is δ ∈ R>0 such that, for all ∆x ∈ R>0 , if ∆x < δ, then U (f, ∆x) −
L(f, ∆x)) < .
The functions U (f, ·), L(f, ·), S(f, ·) have nonstandard extensions U (f, ·) :
(R∗ )>0 → R∗ , etc... So, for example, if ∆x ∈ µ>0 , then U (f, ∆x) equals the
upper Riemann sum of f with respect to a hyperfinite partition, where each
interval in the partition has infinitesimal length.
Theorem 6.6. f is Riemann integrable if and only if U (f, ∆x) ≈ L(f, ∆x)
for any ∆x ∈ µ>0 . In this case, for any ∆x ∈ µ>0 , we have
Z b
f dx = st(U (f, ∆x)) = st(L(f, ∆x)) = st(S(f, ∆x)).
a

Proof. First suppose that f is Riemann integrable. Fix  ∈ R>0 . Take

δ ∈ R>0 satisfying the conclusion of Exercise 6.5. By transfer, if ∆x ∈
µ>0 , then U (f, ∆x) − L(f, ∆x) < . Since  was arbitrary, this shows that
L(f, ∆x) ≈ U (f, ∆x).
For the converse, we verify the criterion given by the Riemann Lemma.
Let  > 0. By the assumption of the theorem,
R∗ |= (∃∆x ∈ (R∗ )>0 )(U (f, ∆x) − L(f, ∆x) < ).
Now apply transfer.
It remains to verify the statement about the value of the integral. Fix
∆x ∈ µ>0 . Then L(f, ∆x) ≤ S(f, ∆x) ≤ U (f, ∆x) and L(f, ∆x) ≈ U (f, ∆x).
Thus, st(U (f, ∆x)) = st(L(f, ∆x)) = st(S(f, ∆x)). Fix an ordinary parti-
tion P of [a, b]. Then, by the transfer of Exercise 6.2(3), we have
L(f, P ) ≤ U (f, ∆x) ≈ L(f, ∆x) ≤ U (f, P ).
Thus, L(f ) ≤ st(U (f, ∆x)) ≤ U (f ). Since f is Riemann integrable, we have
Z b
f dx = U (f ) = L(f ) = st(U (f, ∆x)).
a


We now verify that certain classes of functions are integrable.

 LECTURE NOTES ON NONSTANDARD ANALYSIS 41

Theorem 6.7. If f is continuous on [a, b], then for any ∆x ∈ µ>0 , L(f, ∆x) ≈
U (f, ∆x). Consequently, by the previous theorem, continuous functions are
Riemann integrable.
Proof.
Pn Let’s give the idea of the proof first. Note that U (f, ∆x)−L(f, ∆x) =
i=1 (M i − mi )∆xi . We will find an upper bound for this sum of the form
[f (c) − f (d)](b − a), where |c − d| < ∆x. If ∆x ∈ µ>0 , then c ≈ d, so
f (c) ≈ f (d) by the uniform continuity of f . Since b − a ∈ R, this will show
that U (f, ∆x) − L(f, ∆x) ≈ 0.
Now for the details: for ∆x ∈ R>0 , define Mi (∆x) := Mi (P∆x ) and
mi (∆x) := mi (P∆x ). We then define the oscillation of f with respect to ∆x
to be the quantity
ω(∆x) := max{Mi (∆x) − mi (∆x) | i = 1, . . . , n}.
Suppose that j ∈ {1, . . . , n} is such that ω(∆x) = Mj (∆x) − mj (∆x). Fix
c∆x , d∆x ∈ [xj−1 , xj ] such that f (c∆x ) = Mj (∆x) and f (d∆x ) = mj (∆x);
this is possible since continuous functions achieve their max and min. Clearly,
|c∆x − d∆x | ≤ ∆x. Also,
Xn
U (f, ∆x) − L(f, ∆x) = (Mi (∆x) − mi (∆x))(∆xi ) ≤ ω(∆x)(b − a).
i=1
Fix ∆x ∈ By transfer, there are c, d ∈ [a, b]∗ such that |c−d| ≤ ∆x and
µ>0 .
U (f, ∆x)−L(f, ∆x) ≤ (f (c)−f (d))(b−a), whence U (f, ∆x) ≈ L(f, ∆x). 
Exercise 6.8. Prove Theorem 6.7 with the assumption of “continuity” re-
placed by “monotonicity.”
Now that integrals are infinitely close to hyperfinite sums, properties of
integrals follow almost immediately from properties of sums. For example:
Proposition 6.9. Suppose that f is integrable and c ∈ R. Then cf is
Rb Rb
integrable and a (cf )dx = c a f dx.
Proof. Fix ∆x ∈ µ>0 . First suppose that c ≥ 0. Then U (cf, ∆x) =
cU (f, ∆x) and L(cf, ∆x) = cL(f, ∆x). Since U (f, ∆x) ≈ L(f, ∆x), we
have that U (cf, ∆x) ≈ L(cf, ∆x), whence cf is integrable. Moreover,
Z b Z b
(cf )dx = st(U (cf, ∆x)) = c st(U (f, ∆x) = c f dx.
a a
If c < 0, then U (cf, ∆x) = cL(f, ∆x) and L(cf, ∆x) = cU (f, ∆x). The
proof then proceeds as in the previous paragraph. 
6.2. The Peano Existence Theorem. Here is an application of the non-
standard approach to integration to differential equations:
Theorem 6.10 (Peano Existence Theorem). Suppose that g : [0, 1]×R → R
is a bounded, continuous function. Then for any a ∈ R, there is a differen-
tiable function f : [0, 1] → R satisfying f (0) = a and f 0 (t) = g(t, f (t)) for
all t ∈ [0, 1].
42 ISAAC GOLDBRING

Proof. The idea is quite simple: we find a “polygonal” solution by starting

at (0, a) and taking infinitesimal steps with slope according to g; we then
take the standard part of the polygonal solution to obtain a curve solving
the differential equation. Now for the gory details:
Let Y : R × N → R be such that, for all 0 ≤ k < n, we have Y (0, n) = a
and Y ( k+1 k k k 1
n , n) = Y ( n , n) + g( n , Y ( n , n)) · n . Let Z : R → R be such that
Z( nk ) = Y ( nk , n). Fix N > N. We will show that Z( Nl ) ∈ Rfin for all
0 ≤ l ≤ N and that f : [0, 1] → R defined by f (t) = st(Z( Nl )), where t ≈ Nl ,
is the desired solution to the differential equation.
Fix a bound M ∈ R>0 for g. Fix 0 ≤ k ≤ l ≤ N . Then:
l−1
l k X n n 1
|Z( ) − Z( )| = | (g( , Z( )) · )|
N N N N N
n=k
l−1
X n n 1
≤ (|g( , Z( ))| · )
N N N
n=k
l−1
X 1
≤M·
N
n=k
l−k
=M· .
N
This shows two things: first, by setting k = 0, we see that |Z( Nl )| ≤ |a| +
M · Nl for all 0 ≤ l ≤ N , whence Z( Nl ) ∈ Rfin . Secondly, if l−k N ≈ 0, then
Z( Nl ) ≈ Z( Nk ). These observations allow us to define f : [0, 1] → R by
setting f (t) = st(Z( Nl )) for any l ∈ {0, . . . , N } with t ≈ Nl .
Exercise 6.11. Use the above calculation to verify that f is continuous.
It remains to verify that f is a solution of the differential equation f 0 (t) =
g(t, f (t)). By
R xthe fundamental theorem of calculus, it is enough to show that
f (x) = a + 0 g(t, f (t))dt for all x ∈ [0, 1].
Let W : R → R be such that W ( nk ) = g( nk , Z( nk )). Define h : [0, 1] → R
by h(r) = g(r, f (r)), a continuous function. Then, for 0 ≤ k ≤ N , we have
k k k k k
W( ) ≈ g(st( ), f (st( )) = h(st( )) ≈ h( ).
N N N N N
For n ∈ N, define sn := max{|W ( nk ) − h( nk )| | 0 ≤ k ≤ n}. Thus, by the
above observation, sN ∈ µ. Consequently, for 0 ≤ k ≤ N , we have
k k k
X n 1 X n 1 X n n 1
| (W ( ) · ) − (h( ) )| ≤ |W ( ) − h( )| ·
N N N N N N N
n=0 n=0 n=0
k
X 1
≤ sN ·
N
n=0
 LECTURE NOTES ON NONSTANDARD ANALYSIS 43

k+1
= sN · ∈ µ.
N
Therefore, for any x ∈ [0, 1], writing x = st( Nk ) with Nk < x, we have
k
f (x) ≈ Z()
N
k−1
X n n 1
=a+ (g( ), Z( )) · )
N N N
n=0
k−1
X n 1
=a+ (W ( )· )
N N
n=0
k−1
X n 1
≈a+ (h( )· )
N N
n=0
Z x
≈a+ h(t)dt.
0
The last step follows from Theorem 6.6.
R x Since the beginning and end are
standard numbers, we have f (x) = a + 0 h(t)dt. 
6.3. Problems.
Problem 6.1. Let f : [a, b] → R be continuous. For n ∈ N>0 and i =
0, . . . , n − 1, define xi := a + i(b−a)
n . Define
f (x0 ) + f (x1 ) + · · · + f (xn−1 )
Av(n) := ,
n
which is often referred to as a sample average for f . Prove that if n > N,
1
Rb
we have Av(n) ≈ b−a a f (x)dx.
1
Rb
Remark. b−a a f (x)dx is often referred to as the average of f on [a, b]. This
exercises illustrates a common phenomenon in nonstandard analysis, namely
approximating continuous things by hyperfinite discrete things.
Problem 6.2 (Both). Suppose that f, g : [a, b] → R are Riemann integrable.
Show that:
Rb Rb Rb
(1) f + g is integrable and a (f (x) + g(x))dx = a f (x)d(x) + a g(x)dx.
(2) f is Riemann integrable on both [a, c] and [c, b] for any c ∈ [a, b] and
Rb Rc Rc
f (x)dx = f (x)dx + b f (x)dx.
Rab Rab
(3) a f (x)dx ≤ a g(x)dx if f (x) ≤ g(x) for all x ∈ [a, b].
Rb
(4) m(b − a) ≤ a f (x)dx ≤ M (b − a) if m ≤ f (x) ≤ M for all x ∈ [a, b].
Problem 6.3. Suppose thatRf : [a, b] → R is Riemann integrable. Define
x
F : [a, b] → R by F (x) := a f (t)dt. Prove that F is continuous (even
though f may not be).
Problem 6.4.
44 ISAAC GOLDBRING

(1) Suppose f : [a, ∞) → R is continuous. Define

Z ∞ Z b
f (x)dx := lim f (x)dx
a b→∞ a
R∞
if this limit exists; in this case the improper integral a f (x)dx is
said to
R ∞converge. Otherwise, the improper integral diverges. Show
that a f (x)dx = L if and only if for every b ∈ R∗ such that b > N,
Rb Rb
we have a f (x)dx ≈ L. (But first: what does a f (x)dx for b > N
even mean?)
(2) There are three other kinds of improper integrals, namely when f
is continuous on intervals (−∞, a] as well as [a, b) and (a, b]. Dis-
cuss how to define these improper integrals and give nonstandard
characterizations of them as in (1).
Problem 6.5. Suppose f : [a, ∞) R ∞→ R is continuous and f (x) ≥ 0 for
all x ∈ [a, ∞). Show that either a f (x)dx Rconverges or diverges to +∞.
x
(Hint: Define F : [a, ∞) → R by F (x) := a f (t)dt. Consider the cases
when F is bounded and unbounded respectively.)
Problem 6.6. Suppose that (an : n ≥ 1) is a sequence from R. Suppose
that f : [1, ∞) → R is a continuous, nonnegative,
P∞ nonincreasing function
such
R∞ that f (n) = an for all n ≥ 1. Show that 1 an converges if and only
if 1 f (x)dx converges. (This result is often called the Integral Test.)

7. Weekend Problem Set #1

Problem 7.1. (Sierpinski) Suppose r, a1 , a2 , . . . , an ∈ R>0 . Show that the
equation
a1 a2 an
+ + ··· + =r
x1 x2 xn
has only finitely many solutions (x1 , . . . , xn ) in N>0 .
Problem 7.2.
(1) Show there is N ∈ N∗ such that N is divisible by k for every k ∈ N.
(2) Let P ⊆ N denote the set of primes. Show that for every n ∈ N∗ ,
there is p ∈ P ∗ such that p divides n.
(3) Use parts (1) and (2) to show that P is infinite (Euclid).
(4) Show that Z∗ is a subring of R∗ .
(5) Let p ∈ P ∗ \ P . Let (p) be the ideal of Z∗ generated by p. Show
that the quotient ring Z∗ /(p) is a field. What is the characteristic
of Z∗ /(p)?
Problem 7.3. (Limit Comparison Test) Let ∞
P P∞
0 ai and 0 bi be two se-
ries, where ai , bi > 0 for all i ∈ N. Suppose that ( abii ) converges. Show that
Pn
µ if and only if nm bi ∈ µ.
P
for m, n > N with P∞ m ≤ n, we have m ai ∈ P
Conclude that 0 ai converges if and only if ∞ 0 bi converges.
 LECTURE NOTES ON NONSTANDARD ANALYSIS 45

Problem 7.4. Let (sn ) denote a sequence in R. For each n ∈ N, define

s0 + · · · sn
σn := .
n+1
Show that if sn → L, then σn → L. (Hint: If N > N, there exists M > N
such that M
N is infinitesimal.)

We should say the sequence (σn ) is called the sequence of Cesáro means
of the sequence (sn ). It is possible that (σn ) converges when (sn ) diverges.
When (sn ) is the sequence of partial sums of an infinite series, this leads to
the notion of Cesáro summability, which is useful in the theory of Fourier
series.
Problem 7.5. Suppose f : R → R is a function and f (x + y) = f (x) + f (y)
for all x, y ∈ R (i.e. f is an additive group homomorphism).
(1) Show that f (0) = 0.
(2) Show that f (nx) = nf (x) for all x ∈ R and n ∈ N.
(3) Show that f (−x) = −f (x) for all x ∈ R.
(4) Show that f (kx) = kf (x) for all k ∈ Z.
(5) Show that f (qx) = qf (x) for all q ∈ Q.
(6) (Cauchy) Suppose f is continuous. Show that f (x) = f (1) · x for all
x ∈ R.
Problem 7.6. In this problem, we prove a strengthening of Cauchy’s result
from Problem 6 by showing that if f : R → R is an additive group homo-
morphism and there is an inteveral I ⊆ R such that f is bounded on I, then
f (x) = f (1) · x for all x ∈ R. (This result is due to Darboux.) Fix x0 ∈ I
and M ∈ R>0 such that |f (x)| ≤ M for all x ∈ I.
(1) Show that if x ≈ 0, then |f (x)| ≤ M + |f (x0 )|.
(2) Show that if x ≈ 0, then f (x) ≈ 0. (Hint: If x ≈ 0, then nx ≈ 0 for
all n ∈ N.
(3) Show that f (x) = f (1) · x for all x ∈ R. (Hint: Use the fact that
any x ∈ R is infinitely close to an element of Q∗ .)
Problem 7.7.
(1) Suppose f : (a, b) → R is C 1 . Suppose x ∈ (a, b)∗ is such that
st(x) ∈ (a, b). Suppose ∆x ≈ 0. Prove that there exists  ∈ µ such
that
f (x + ∆x) = f (x) + f 0 (x) + ∆x.
(2) Define f : R → R by
(
x2 sin( x1 ) if x 6= 0
f (x) =
0 if x = 0.
(a) Show that f 0 (x) exists for all x but that f 0 is not continuous at
0.
46 ISAAC GOLDBRING

1
(b) Let N > N and x = 2πN . Show that there is an infinitesimal
∆x such that there is no  ∈ µ making the conclusion of (a)
true.
(c) Discuss why parts (1) and (2)(b) don’t contradict the fact that
f 0 is continuous on (0, 1).
Problem 7.8. A Dirac delta function is a definable function D : R∗ → R∗
such that:
• D(x) ∗
R +∞ ≥ 0 for all x ∈ R ;
• −∞ D(x)dx = 1;
Rδ
• there is a positive infinitesimal δ > 0 such that −δ D(x)dx ≈ 1.
(1) Make sense of the above properties, i.e. explain how to precisely
state the above properties of a Dirac delta function.
R +∞delta function and f : R → R a standard function.
(2) Let D be a Dirac
Show that st( −∞ f (x)D(x)dx) = f (0).
R∞
(3) Suppose f : R → R is standard and −∞ f (x)dx = 1. Fix n ∈ N∗ \ N
and define D : R∗ → R∗ by D(x) := nf (nx). Show that D is a Dirac
delta function. In particular, the following functions are Dirac delta
functions:
• (
1
n if |x| ≤ 2n
D(x) =
0 otherwise.
n
• D(x) = π(1+n2 x2 )
.
2 2
• D(x) = ne−πn x .
(4) Suppose that f : R → R is a standard function such that f (x) ≥ 0
R +∞ R1
for all x ∈ R and −∞ f (x) = 1. For n ∈ N>0 , define an := −n 1 f (x).
n
Show that an ≈ 0 for all n > N. Conclude that a Dirac delta function
can never be the nonstandard extension of a standard nonnegative,
integrable function.
Problem 7.9. For a, b ∈ N with a ≤ b, we set
[a, b] := {a, a + 1, . . . , b − 1, b} ⊆ N.
Suppose that A ⊆ N. We say that:
• A is thick if for all k ∈ N>0 , there is x ∈ N such that [x+1, x+k] ⊆ A.
• A is syndetic if N \ A is not thick, that is, there is k ∈ N>0 such that,
for all x ∈ N, [x + 1, x + k] ∩ A 6= ∅.
• A is piecewise syndetic if A = B ∩ C, where B is thick and C is
syndetic.
(1) Prove that A is thick if and only if A∗ contains an infinite interval,
that is, there are M, N ∈ N∗ with N − M ∈ N∗ \ N such that
[M, N ] ⊆ A∗ .
 LECTURE NOTES ON NONSTANDARD ANALYSIS 47

(2) Prove that A is syndetic if and only if A∗ has finite gaps, that is, for
all intervals [M, N ] ⊆ N∗ , if [M, N ] ∩ A∗ = ∅, then N − M ∈ N.
(3) (Standard) Prove that A is piecewise syndetic if and only if there is
a finite set F ⊆ N such that A + F is thick, where
A + F := {a + f : a ∈ A, f ∈ F }.
(4) Prove that A is piecewise syndetic if and only if there is an infinite
interval on which A∗ has only finite gaps.
(5) Use the nonstandard characterization of piecewise syndeticity to
prove that piecewise syndeticity is a partition regular notion, that
is, if A = A1 ∪ · · · ∪ An is piecewise syndetic, then Ai is piecewise
syndetic for some i ∈ {1, . . . , n}.
The notions appearing in the previous problem are present in additive
combinatorics and combinatorial number theory.

8. Many-sorted and Higher-Type Structures

We would now like to start applying nonstandard methods to areas of
mathematics more complex than calculus. To do this, we will need a slightly
more elaborate nonstandard framework.
8.1. Many-sorted structures. In many areas of mathematics, we study
many different sets at a time as well as functions between these various sets.
Example 8.1 (Linear Algebra). A vector space is a set V together with
two functions: vector addition, which is a function + : V × V → V , and
scalar multiplication, which is a function · : F × V → V , where F is some
field.
Example 8.2 (Topology). A metric space is a set X together with a metric,
which is a function d : X × X → R satisfying certain axioms (to be defined
in the next section).
Example 8.3 (Measure Theory). A measure space is a triple (X, B, µ),
where X is a set, B is a σ-algebra of subsets of X (so B ⊆ P(X)), and
µ : B → R is a measure.
We now develop a nonstandard framework suitable for studying such sit-
uations. Before, we were working with a structure consisting of just a single
“sort,” namely a sort for R. Now, we will work in a structure M with a
(nonempty) collection of sorts S. For each s ∈ S, we have a set Ms , the
universe of the sort s in M . So, for example, in the linear algebra situation,
we might have S = {s, t}, with Ms = V and Mt = F. Often we will write
a many-sorted structure as M = (Ms | s ∈ S). Thus, we might write the
linear algebra example as (V, F), suppressing mention of the names of the
sorts.
For any finite sequence ~s = (s1 , . . . , sn ) of sorts, we have the product set
M~s := Ms1 × · · · × Msn .
48 ISAAC GOLDBRING

We now consider a language which was just as expressive as before.

Namely, we have:
• For every finite sequence ~s of sorts and every A ⊆ M~s , we have a
predicate symbol PA .
• For every sort s and every a ∈ Ms , we have a constant symbol ca .
• For every finite sequence ~s of sorts, every sort t, and every function
f : M~s → Mt , we have a function symbol Ff .
One now builds terms and formulae just as in ordinary logic, with the
understanding that each sort comes equipped with its own collection of
variables. If we need to be clear, we might decorate a variable with the
name of the sort it is intended to range over, e.g. xs .
Example 8.4. Returning to the vector space example, let’s see how we
might write the distributive law c · (x + y) = c · x + c · y. Recall that
S = {s, t}, with Ms = V and Mt = F. Let f : Ms × Ms → Ms denote
vector addition and g : Mt × Ms → Ms denote scalar multiplication. Then
the axiom for the distributive law would be written as:
∀xs ∀y s ∀z t (g(z, f (x, y)) = f (g(z, x), g(z, y))).
Of course, for the purpose of sanity, in practice we will continue to write
things as they might naturally be written in ordinary mathematics; however,
one must be aware of the formal way that such sentences would be written.
As in the earlier part of these notes, we obtain a “nonstandard exten-
sion” by considering an embedding M → M ∗ , where the universe of M ∗
corresponding to the sort s is Ms∗ and the map is given by a 7→ a∗ . We
will demand more from this embedding later, but for now we do not even
assume that the embedding is an inclusion, that is, we do not assume that
Ms ⊆ Ms∗ . As before, we write A∗ for the interpretation of the predicate
symbol PA in M ∗ .
To be of any use, we assume that the nonstandard extension is proper,
meaning that, for every s ∈ S and every infinite A ⊆ Ms , we assume that
there is b ∈ A∗ such that b 6= a∗ for any a ∈ A. (In other words, we are
postulating the existence of many nonstandard elements.) Of course, such
nonstandard extensions exist by the Compactness Theorem (Exercise!).
Exercise 8.5. Suppose the nonstandard extension is proper. Then for any
finite sequence ~s of sorts and any infinite A ⊆ M~s , A∗ contains a nonstandard
element.
8.2. Higher-type sorts. As we saw with the measure theory example in
the previous subsection, it will often be convenient to have a sort for P(X)
whenever X is itself a sort. For simplicity of discussion, let us consider
the many-sorted structure (X, P(X)); what we say now is easily adapted
to the more general situation that X and P(X) are sorts in a many-sorted
structure containing other sorts.
 LECTURE NOTES ON NONSTANDARD ANALYSIS 49

We have the nonstandard extension (X, P(X)) → (X ∗ , P(X)∗ ). We must

be careful not to confuse P(X)∗ with P(X ∗ ), the latter retaining its usual
meaning as the set of subsets of X ∗ . At the moment, P(X)∗ is some abstract
set, perhaps having no affiliation with an actual powerset. We now discuss
how to relate P(X)∗ and P(X ∗ ).
Set E = {(x, A) ∈ X × P(X) | x ∈ A}, the symbol for the membership
relation.
Lemma 8.6 (Normalization). We may assume that our nonstandard exten-
sion satisfies the additional two conditions:
(N1) X ⊆ X ∗ and x = x∗ for all x ∈ X;
(N2) P(X)∗ ⊆ P(X ∗ ) and E ∗ is the membership relation restricted to
X ∗ × P(X)∗ .
Proof. We begin with some abstract nonsense: let Y be a set and h : X ∗ → Y
a bijection such that X ⊆ Y and h(x∗ ) = x for all x ∈ X.
For A ∈ P(X)∗ , set Φ(A) = {h(x) | x ∈ X ∗ and (x, A) ∈ E ∗ } ⊆ Y .
We claim that Φ is injective. Indeed, suppose that A1 , A2 ∈ P(X)∗ and
A1 6= A2 . By the transfer principle, there is x ∈ X ∗ such that either
(x, A1 ) ∈ E ∗ and (x, A2 ) ∈
/ E ∗ ; or (x, A2 ) ∈ E ∗ and (x, A1 ) ∈
/ E ∗ . Then
either h(x) ∈ Φ(A1 ) \ Φ(A2 ) or h(x) ∈ Φ(A2 ) \ Φ(A1 ); either way, Φ(A1 ) 6=
Φ(A2 ).
Now make (Y, Φ(P(X)∗ )) into a structure in the unique way so that the
map (h, Φ) : (X ∗ , P(X)∗ ) → (Y, Φ(P(X)∗ )) is an isomorphism. (Exercise!)
Note that (Y, Φ(P(X)∗ )) has the desired properties (N1) and (N2). 
Recap: We have the many-sorted structure (X, P(X)) and its nonstandard
extension (X ∗ , P(X)∗ ). Furthermore, we assume that X ⊆ X ∗ and for all
Y ∈ P(X)∗ , we view Y ⊆ X ∗ by declaring, for x ∈ X ∗ :
x ∈ Y ↔ (x, Y ) ∈ E ∗ .
There is some potential confusion that we should clear up now. Suppose
that A ⊆ X. Then we have A∗ ⊆ X ∗ from the interpretation of the symbol
PA . However, A ∈ P(X), so it is mapped by the embedding to an element
of P(X)∗ , which we temporarily denote by (A)∗ . Fortunately, all is well:
Lemma 8.7. A∗ = (A)∗ .
Proof. By the transfer principle, we have that, for x ∈ X ∗ , x ∈ A∗ if and
only if (X ∗ , P(X)∗ ) |= PA (x). By the normalization assumption, we have
that, for x ∈ X ∗ , x ∈ (A)∗ if and only if (x, (A)∗ ) ∈ E ∗ . Fortunately,
(X, P(X)) |= ∀x ∈ X(PA (x) ↔ PE (x, A)), so the desired result follows from
transfer. 
Definition 8.8. A subset A of X ∗ is called internal if A ∈ P(X)∗ ; otherwise,
A is called external.
Thus, the transfer principle applies to the internal subsets of X ∗ .
50 ISAAC GOLDBRING

Example 8.9. Let us consider (N, P(N)) and its nonstandard extension
(N∗ , P(N)∗ ). We claim that N is an external subset of N∗ . To see this, note
that the following sentence is true in (N, P(N)):
∀A ∈ P(N)((∃x ∈ N(PE (x, A)) ∧ ∃y ∈ N∀z ∈ N(PE (z, A) → z ≤ y))
→ ∃y ∈ N(PE (y, A) ∧ ∀z ∈ N(PE (z, A) → z ≤ y).
This sentence says that if A ⊆ N is bounded above, then A has a maximum
element. By transfer, the same holds true for any A ∈ P(N)∗ , that is, for
any internal subset of N∗ . If N were internal, then since it is bounded above
(by an infinite element), it would have a maximum, which is clearly not true.
Example 8.10. We continue to work with the set-up of the previous ex-
ample. Since
(N, P(N)) |= ∀n ∈ N∃A ∈ P(N)∀m ∈ N(PE (m, A) ↔ m ≤ n),
by transfer we have
(N∗ , P(N)∗ ) |= ∀n ∈ N∗ ∃A ∈ P(N)∗ ∀m ∈ N∗ (PE (m, A) ↔ m ≤ n).
Fixing N ∈ N, we suggestively let {0, 1, . . . , N } denote the internal subset
of N∗ consisting of all the elements of N∗ that are no greater than N . This
is a prototypical example of a hyperfinite set.
The following principle is useful in practice; it says that sets defined (in
the first-order logic sense) from internal parameters are internal.
Theorem 8.11 (Internal Definition Principle). Let ϕ(x, x1 , . . . , xm , y1 , . . . , yn )
be a formula, where x, x1 , . . . , xm range over the sort for X and y1 , . . . , ym
range over the sort for P(X). Suppose that a1 , . . . , am ∈ X ∗ and A1 , . . . , An ∈
P(X)∗ . Set
B := {b ∈ X ∗ | (X ∗ , P(X)∗ ) |= ϕ(b, a1 , . . . , am , A1 , . . . , An )}.
Then B is internal.
Proof. The following sentence is true in (X, P(X)):
∀x1 , . . . , xm ∀y1 , . . . , yn ∃z∀x(PE (x, z) ↔ ϕ(x, x1 , . . . , xm , y1 , . . . , yn )).
By transfer, this remains true in (X ∗ , P(X)∗ ). Plugging in ai for xi and Aj
for yj , we see that
(X ∗ , P(X)∗ ) |= ∃z∀x(PE (x, z) ↔ ϕ(x, a1 , . . . , am , A1 , . . . , An )).
The set asserted to exist is B, which then belongs to P(X)∗ , that is, B is
internal. 
Example 8.12. For any finite collection a1 , . . . , am ∈ X ∗ , the set {a1 , . . . , am }
is internal. Indeed, let ϕ(x, x1 , . . . , xm ) be the formula x = x1 ∨· · ·∨x = xm .
Then
{a1 , . . . , am } = {b ∈ X ∗ | (X ∗ , P(X)∗ ) |= ϕ(b, a1 , . . . , am )}.
 LECTURE NOTES ON NONSTANDARD ANALYSIS 51

It will also prove useful to have a notion of internal function. To do this,

we need to expand our set-up a bit. We now consider the many-sorted struc-
ture (X, P(X), P(X × X)) with an embedding into a nonstandard extension
(X ∗ , P(X)∗ , P(X × X)∗ ). We set:
• E1 := {(x, A) ∈ X × P(X) | x ∈ A}, and
• E2 := {(x, y, A) ∈ X × X × P(X × X) | (x, y) ∈ A}.
The proof of the following lemma is exactly like the proof of Lemma 8.6.
Lemma 8.13. We may assume that our nonstandard extension satisfies the
additional three conditions:
(N1) X ⊆ X ∗ and x = x∗ for all x ∈ X;
(N2) P(X)∗ ⊆ P(X ∗ ) and E1∗ is the membership relation restricted to
X ∗ × P(X)∗ ;
(N3) P(X × X)∗ ⊆ P(X ∗ × X ∗ ) and E2∗ is the membership relation re-
stricted to X ∗ × X ∗ × P(X × X)∗ .
Definition 8.14. B ⊆ X ∗ × X ∗ is internal if B ∈ P(X × X)∗ . If A, B ⊆ X ∗
and f : A → B is a function, then we say that f is internal if the graph of
f , Γ(f ) := {(x, y) ∈ X ∗ × X ∗ | x ∈ A and f (x) = y} ⊆ X ∗ × X ∗ , is internal.
At this point, the reader should verify that they would know how to es-
cape the friendly confines of considering many-sorted structures of the form
(X, P(X)) or (X, P(X), P(X × X)) and instead be able to consider much
wilder many-sorted structures that might contain many sets and their power-
sets. For example, in studying vector spaces, it will be convenient to consider
a many-sorted structure of the form (V, P(V ), P(V × V ), F, P(F), P(F × V )).
As an exercise, make sure you understand how to speak of an internal linear
transformation T : V ∗ → V ∗ or an internal norm k · k : V ∗ → R∗ .
In a similar vein, if s : N∗ → X ∗ is an internal function, then we refer to
the “sequence” (sn | n ∈ N∗ ) as an internal sequence.
Definition 8.15. Set Pfin (X) := {A ∈ P(X) | A is finite } ⊆ P(X). We
then say that B ⊆ X ∗ is hyperfinite if B ∈ Pfin (X)∗ .
Exercise 8.16. Assume that (N, P(N), P(N × N)) is part of the structure.
(1) Prove that hyperfinite sets are internal.
(2) Prove that an internal subset of a hyperfinite set is hyperfinite.
(3) Prove that B ⊆ X ∗ is hyperfinite if and only if there is an internal
function f : B → N∗ such that f is a bijection between B and
{0, 1, . . . , N } for some N ∈ N∗ ; we then refer to N +1 as the internal
cardinality of B.
(4) Prove that finite subsets of X ∗ are hyperfinite.
8.3. Saturation. For various nonstandard arguments, it will not suffice to
merely assume that the extension is proper; we will need to assume a further
richness condition on our nonstandard extensions.
52 ISAAC GOLDBRING

Definition 8.17. Suppose that κ is an uncountable cardinal. We say that

the nonstandard extension is κ-saturated if whenever (Ai | i ∈ I) is a family
of internal sets with |I| < κ satisfying the finite intersection property,
T that is,
the intersection of any finite number of Ai ’s is nonempty, then i∈I Ai 6= ∅.
Example 8.18. Suppose that the nonstandard extension is ℵ1 -saturated
and that R is a basic sort. For each q ∈ Q>0 , set Aq := {r ∈ R∗ | 0 < r < q}.
By the internal definition principle, Aq is an internal set. Moreover, it is
easy to verify that the family (Aq | q T ∈ Q>0 ) has the finite intersection
property. By ℵ1 -saturation, there is r ∈ q∈Q>0 Aq ; this r is then a positive
infinitesimal.
Although this example is modest (as we already have ways of constructing
infinitesimals), saturation will prove crucial in the analysis to come; see also
the problems below.
Exercise 8.19. Assume that the nonstandard extension is ℵ1 -saturated and
that (Nm | m ∈ N) is a sequence of elements of N∗ \ N. Prove that there is
N > N such that N < Nm for each m ∈ N.
In any research article, one always assumes at least ℵ1 -saturation of the
nonstandard extension. For other applications, κ-saturation for larger κ is
often needed. However, how can we be assured that κ-saturated nonstandard
extensions exist?
Theorem 8.20. For any uncountable cardinal κ, there is a κ-saturated non-
standard extension.
Proof. See any textbook on model theory. 
Saturation is closely related to another richness concept, namely compre-
hension.
Theorem 8.21 (Saturated extensions are comprehensive). Suppose that the
nonstandard extension is κ-saturated. Suppose A and B are internal sets
and suppose that A0 ⊆ A is a (possibly external) set with |A0 | < κ. Suppose
f0 : A0 → B is a function. Then there is an internal function f : A → B
extending f0 .
Proof. For x ∈ A0 , define
Dx := {f | f : A → B is internal and f (x) = f0 (x)}.
By the internal definition principle, each Dx is internal. By κ-saturation,
it remains to verify that (Dx | x ∈ A0 ) has the finite intersection property.
Fix x1 , . . . , xn ∈ A0 . Define f : A → B by
(
f0 (xi ) if y = xi for some i ∈ {1, . . . , n};
f (y) =
f0 (x1 ) otherwise.
Then f is internal by the internal definition principle. 
 LECTURE NOTES ON NONSTANDARD ANALYSIS 53

Definition 8.22. The nonstandard extension is said to be countably com-

prehensive if whenever B is internal and (bn | n ∈ N) is a countable sequence
from B, then there is an internal f : N∗ → B such that f (n) = bn for all
n ∈ N.
In other words, countable comprehension says that sequences indexed by
N can be internally extended to sequences indexed by N∗ .
Corollary 8.23. If the nonstandard extension is ℵ1 -saturated, then it is
also countably comprehensive.
The converse to the previous corollary is also true; see Problem 8.8 below.
8.4. Useful nonstandard principles. In this subsection, we collect some
useful principles that are often used in nonstandard proofs.
Theorem 8.24 (Overflow). Suppose that A ⊆ R∗ is internal and suppose
that, for every n ∈ N, there is a ∈ A such that a > n. Then there is a ∈ A
such that a > N.
Proof. Set B := {n ∈ N∗ | there is a ∈ A such that a > n}. B is internal
by the internal definition principle. By assumption, N ⊆ B. Since N is
external, we have N ( B, whence there is n ∈ B with n > N. Take a ∈ A
such that a > n; it follows that a > N. 
Theorem 8.25 (Underflow). Let A ⊆ (R∗ )>0 be internal and suppose that,
for every n ∈ N∗ \ N, there is a ∈ A such that a < n. Then A ∩ Rfin 6= ∅.
Proof. Set B := {n ∈ N∗ | there is a ∈ A such that a < n}. Then B is
internal and contains N∗ \ N. Note that N∗ \ N is external, else, since N∗ is
internal, we would see that N is internal. Thus, we have B ∩ N 6= ∅, which
yields the desired result. 
Theorem 8.26 (Internal Induction). Let A ⊆ N∗ be internal and suppose
that:
• 0 ∈ A;
• for all n ∈ N∗ , if n ∈ A, then n + 1 ∈ A.
Then A = N∗ .
Proof. Suppose A 6= N∗ . Since N∗ \ A is internal, it has a minimum element
(by transfer). Let n = min(N∗ \ A). By assumption, n > 0. Then n − 1 ∈ A,
whence, by assumption, n ∈ A, a contradiction. 
As a warning, we really must assume that A is internal in the previous
theorem. Indeed, N satisfies the two assumptions of the previous theorem,
but N 6= N∗ .
Theorem 8.27 (Infinitesimal Prolongation). Suppose that (sn | n ∈ N∗ ) is
an internal sequence from R∗ . Suppose that sn ≈ 0 for each n ∈ N. Then
there is N > N such that sn ≈ 0 for each n ≤ N .
54 ISAAC GOLDBRING

Incorrect First Attempt. Let A = {n ∈ N∗ | sn 6≈ 0}. If A = ∅, we are done.

Otherwise, let N = min A. By assumption, N > N, whence we are done.
This proof is incorrect since A is external, whence we cannot conclude that
it has a minimum! 
Proof. For n ∈ N∗ , define tn := n · sn ; then the sequence (tn | n ∈ N∗ ) is
internal by assumption. By assumption, tn ≈ 0 for n ∈ N; again, this is
an external statement. However, it is enough to replace that statement by
the weaker, internal statement |tn | < 1 for all n ∈ N. More precisely, let
A = {n ∈ N∗ | |tm | < 1 for all m ≤ n}. By overflow, there is N ∈ N∗ \ N
such that N ∈ A. This N is as desired. 
The Infinitesimal Prolongation Theorem is often used in conjunction with
countable comprehension as follows. Suppose that (xn | n ∈ N∗ ) is an
internal sequence from R∗ such that xn ∈ Rfin for n ∈ N. For n ∈ N, define
yn := st(xn ). By countable comprehension, we can internally extend the
sequence (yn | n ∈ N) to an internal sequence (yn | n ∈ N∗ ) from R∗ . For
n ∈ N, we know that xn ≈ yn . By infinitesimal prolongation, we can find
N > N such that xn ≈ yn for all n ≤ N .
The last principle is perhaps the most important of all in applications, as
we will see later in these notes.
Theorem 8.28 (Hyperfinite Approximation). Suppose that the nonstandard
extension is κ-saturated. Let A ⊆ X ∗ be internal and let B ⊆ A be a (possibly
external) set with |B| < κ. Then there is a hyperfinite set C ⊆ A such that
B ⊆ C.
Proof. For x ∈ B, set
Dx := {C ∈ Pfin (X)∗ | C ⊆ A and x ∈ C}.
Note that each Dx is internal by the internal definition principle. Moreover,
(Dx | x ∈ B) has the finite intersection property. Indeed, given x1 , . . . , xn ∈
T {x1 , . . . , xn } ∈ Dx1 ∩ · · · ∩ Dxn . Thus, by κ-saturation,
B, we have that
there is C ∈ x∈B Dx ; such C is as desired. 
8.5. Recap: the nonstandard setting. From now on, we will proceed
under the following assumptions: we assume that our nonstandard extension
contains as sorts all sets relevant to the mathematics we are about to study.
Moreover, we assume that enough cartesian products and powersets are also
sorts. We assume that the nonstandard extension satisfies the normalization
assumptions from Lemma 8.6. Finally, we assume that the nonstandard
extension is κ-saturated for κ large enough for our purposes.
8.6. Problems.
Problem 8.1. Suppose that A and B are internal subsets of X ∗ .
(1) Show that X ∗ \ A, A ∪ B and A ∩ B are internal.
(2) Suppose that f : X ∗ → X ∗ is an internal function. Show that f (A)
and f −1 (A) are internal sets and f  A is an internal function.
 LECTURE NOTES ON NONSTANDARD ANALYSIS 55

Problem 8.2.
(1) Suppose r, s ∈ R∗ and r < s. Set [r, s] := {t ∈ R∗ | r ≤ t ≤ s}. Show
that [r, s] is internal.
(2) Show that µ is external.
(3) Show that Rfin is external.
Problem 8.3. Discuss what it should mean for a function f : A → B to be
internal, where A ⊆ M~s and B ⊆ M~t.
Problem 8.4. Suppose that f : N∗ × X ∗ → X ∗ is an internal function. Fix
x ∈ X ∗ . Show that there exists a unique internal function F : N∗ → X ∗
such that F (0) = x and F (n + 1) = f (n + 1, F (n)). (This is the principle of
Internal Recursion.)
Problem 8.5. Suppose that the nonstandard extension is κ-saturated. Show
that every infinite internal set has cardinality at least κ.
Problem 8.6.
(1) Suppose that N ∈ N∗ \ N. Fix r ∈ (0, 1) (so r is standard). Show
there is a smallest k ∈ N∗ such that N r ≤ k.
(2) Show that any infinite hyperfinite set has cardinality at least 2ℵ0 .
(3) Show that any infinite internal set has cardinality at least 2ℵ0 . (This
improves the result from the previous exercise.)
Problem 8.7. Suppose the nonstandard extension satisfies the Countable
Comprehension Principle. Further suppose that (Kn | n ∈ N) is a sequence
of elements of N∗ such that Kn > N for all n ∈ N. Show that there is
K ∈ N∗ \ N such that K < Kn for all n ∈ N.
Problem 8.8. Show that a nonstandard extension satisfying the Countable
Comprehension Principle must be ℵ1 -saturated. (Hint: you might find the
previous problem useful.)
Problem 8.9. Fix k ∈ N. Suppose that G = (V, E) is a (combinatorial)
graph such that every finite subgraph of G is k-colorable. Prove that G is
k-colorable. (Hint: hyperfinite approximation!)

9. Metric Space Topology

In this section, we will start using the new nonstandard framework de-
veloped in the previous section to develop metric space topology from a
nonstandard point of view. Although general topological spaces can be
treated in the nonstandard framework, we have made a conscience decision
to discuss the (important) special case of metric spaces.
9.1. Open and closed sets, compactness, completeness.
Definition 9.1. A metric space is a pair (X, d) such that X is a nonempty
set and d : X × X → R is a function (the metric or distance function)
satisfying, for all x, y, z ∈ X:
56 ISAAC GOLDBRING

(1) d(x, y) ≥ 0;
(2) d(x, y) = 0 if and only if x = y;
(3) d(x, y) = d(y, x);
(4) (Triangle Inequality) d(x, z) ≤ d(x, y) + d(y, z).
If (2) in the above list is replaced by the weaker condition
(2’) x = y implies d(x, y) = 0,
then d is called a pseudometric and (X, d) is called a pseudometric space.
As usual, we often speak of “the metric space X,” suppressing mention
of the metric. Keep in mind: the same set X can be equipped with many
different metrics, yielding many different metric spaces. (See the examples
below.)
Example 9.2.
pPn
(1) For ~x, ~y ∈ Rn , define d(~x, ~y ) := 2 n
i=1 (xi − yi ) . Then (R , d) is
a metric space. This metric is usually referred to as the euclidean
metric on Rn .
(2) For ~x, ~y ∈ Rn , define d∞ (~x, ~y ) := maxi=1,...,n |xi −yi |. Then (Rn , d∞ )
is a metric space.
(3) Set C([0, 1]), R) to be the set of continuous functions f : [0, 1] → R.
Define a metric d on C([0, 1], R) by setting
d(f, g) := max |f (x) − g(x)|.
x∈[0,1]

Note that this maximum exists by continuity.

Until otherwise specified, we fix a metric space X. For a ∈ X and r ∈ R>0 ,
we set B(a; r) := {x ∈ X | d(a, x) < r}, the open ball in X centered at a
with radius r. We can also consider the closed ball in X centered at a with
radius r: B̄(a; r) := {x ∈ X | d(a, x) ≤ r}.
We will work in a nonstandard extension containing N, R, X, and what-
ever else we might need to refer to. We also assume that the extension is
κ-saturated for κ > max(2ℵ0 , |X|) (although we can get away with an of-
ten smaller level of saturation; we will discuss this later). Observe that the
metric extends to a function d : X ∗ × X ∗ → R∗ satisfying the axioms in
Definition 9.1 above. Thus, (X ∗ , d) is almost a metric space; it’s only defect
is that the metric takes values in R∗ rather than R.
We begin by giving a nonstandard characterization of the topological
notion of a set being open. First, the standard definition:
Definition 9.3. O ⊆ X is open if for any a ∈ O, there is r ∈ R>0 such that
B(a; r) ⊆ O.
To give the nonstandard equivalent of open, we first make a very impor-
tant nonstandard definition:
Definition 9.4. For a ∈ X, the monad of a is the set
µ(a) := {x ∈ X ∗ | d(a, x) ≈ 0}.
 LECTURE NOTES ON NONSTANDARD ANALYSIS 57

Notice that if X = R is given the euclidean metric, then µ(0) = µ is the

set of infinitesimals.
Proposition 9.5. O ⊆ X is open if and only if, for all a ∈ O, µ(a) ⊆ O∗ .
Proof. First assume that O is open. Fix a ∈ O. By assumption, there is
r ∈ R>0 such that, for all x ∈ X, if d(a, x) < r, then x ∈ O. Thus, by
transfer, if x ∈ X ∗ and d(a, x) < r, then x ∈ O∗ . In particular, if x ∈ µ(a),
then d(a, x) ≈ 0, so d(a, x) < r and hence x ∈ O∗ . Thus, µ(a) ⊆ O∗ .
For the converse, fix a ∈ X. By assumption, µ(a) ⊆ O∗ . Fix δ ∈ µ>0 .
Then if x ∈ X ∗ is such that d(a, x) < δ, then x ∈ µ(a), whence x ∈ O∗ .
Thus, (∃δ ∈ (R>0 )∗ )(∀x ∈ X ∗ )(d(a, x) < δ → x ∈ O∗ ). Now apply the
transfer principle to obtain δ ∈ R>0 such that, for all x ∈ X, if d(a, x) < δ,
then x ∈ O, whence B(a; δ) ⊆ O. Since a ∈ O was arbitrary, this shows
that O is open. 
Before we state the next corollary, we first note that, by the transfer
principle, B(a; r)∗ := {x ∈ X ∗ | d(a, x) < r}.
Corollary 9.6. For any a ∈ X and r ∈ R>0 , B(a; r) is open. (So the
terminology open ball is appropriate.)
Proof. Fix b ∈ B(a; r); we need µ(b) ⊆ B(a; r)∗ . Suppose c ∈ µ(b), so
d(b, c) ≈ 0. Then, by the (transfer of the) triangle inequality, d(a, c) ≤
d(a, b) + d(b, c) < r, whence c ∈ B(a; r)∗ . 
Exercise 9.7. Prove that µ(a) is internal if and only if a is an isolated point
of X, that is, {a} is an open set.
Some notation: for x, y ∈ X ∗ , we write x ≈ y to indicate d(x, y) ≈ 0. So,
for a ∈ X, µ(a) = {b ∈ X ∗ | a ≈ b}. Also, if x, y ∈ X, then x ≈ y if and
only if x = y.
Exercise 9.8. The relation ≈ is an equivalence relation on X ∗ .
Definition 9.9. C ⊆ X is closed if X \ C is open.
Corollary 9.10. C is closed if and only if, whenever p ∈ X and q ∈ C ∗ are
such that p ≈ q, then p ∈ C.
Proof. X \ C is open if and only if for all p ∈ X \ C, for all q ∈ X ∗ with
p ≈ q, we have q ∈ (X \ C)∗ = X ∗ \ C ∗ . 
Digression: Before we continue studying metric space topology, let us just
mention briefly how the above set-up generalizes to an arbitrary topological
space. First, what is a topological space? A topological space is a nonempty
set X, together with a collection τ of subsets of X, called the open subsets
of X, satisfying the following criteria:
• ∅, X ∈ τ ;
• If
S (Ui | i ∈ I) is a family of subsets of X, each of which is in τ , then
i∈I Ui ∈ τ ;
58 ISAAC GOLDBRING

• If U, V ∈ τ , then U ∩ V ∈ τ .
So, for example, a metric space, equipped with its collection of open sets
(as defined above), is a topological space. There are a plethora of topological
spaces not arising from metric spaces and the notion of topological space
is central to most areas of mathematics (even logic!). Given aTtopological
space X and a ∈ X, we define the monad of a to be µ(a) = {U ∗ | U ∈
τ, a ∈ U }. (Double-check that this agrees with the definition in the metric
space context.) With a little more effort, the results we have established in
this section (that do not refer to metric notions) hold in the more general
context of topological spaces. In fact, this is true of the majority of the
results to come (at least the ones that do not mention metric notions).

We return to metric space topology. The following definition, while awk-

ward at first site, is fundamental in topology:
Definition 9.11. K ⊆ X is compact if and only if:S whenever (Oi | i ∈ I)
is a family of open subsets of X such that K ⊆ i∈I Oi , then there are
i1 , . . . , in ∈ I such that K ⊆ Oi1 ∪ · · · ∪ Oin .
In English: every open covering of K has a finite subcover. Observe
that, in the above definition of compactness, we can restrict attention to
open coverings of K whose index set I has cardinality at most |K|. (Why?)
Consider the following elegant nonstandard characterization of compactness:
Proposition 9.12 (Robinson’s characterization of compactness). K is com-
pact if and only if, for every p ∈ K ∗ , there is q ∈ K with p ≈ q.
Proof. First assume that K is compact and yet, towards a contradiction,
that there is p ∈ K ∗ such that p 6≈ q for every q ∈ K. For each q ∈ K, there
is then rq ∈ R>0 such that d(p, q) ≥ rq . Notice that {B(q; rq ) | q ∈ K} is an
open coverSof K. Thus, since K is compact, there are q1 , . . . , qn ∈ K such
that K ⊆ ni=1 B(qi ; rqi ). In logical terms, the following is true:
(∀x ∈ K)(d(x, q1 ) < rq1 ∨ d(x, q2 ) < rq2 ∨ · · · ∨ d(x, qn ) < rqn ).
Applying the transfer principle to the displayed statement, we see that
d(p, qi ) < rqi for some i ∈ {1, . . . , n}, which is a contradiction.
For the converse, assume that K is not compact. Thus, there is an open
cover (Oi | i ∈ I) of K with no finite subcover. For i ∈ I, consider the
internal set Ai := K ∗ \ Oi∗ . By assumption, each Ai is nonempty and the
family (Ai | i ∈ I) has the finite intersection property: indeed, if Ai1 ∩ · · · ∩
Ain = ∅, then for all x ∈ K ∗ , there is S i ∈ {1, . . . , n} such that x ∈ Oi∗ . By
transfer, this would imply that K ⊆ nj=1 Oij , contradicting the choice of
T
the open cover (Oi ). Thus, by saturation, there is p ∈ i∈I Ai . (Our first
serious use of saturation!) Then p ∈ K ∗ and p 6≈ q for any q ∈ K: indeed,
if p ≈ q with q ∈ K, then taking i ∈ I such that q ∈ Oi , we would have
p ∈ µ(q) ⊆ Oi∗ , contradicting p ∈ Ai . 
 LECTURE NOTES ON NONSTANDARD ANALYSIS 59

Example 9.13. It is easy to see, using the nonstandard characterization of

compactness, that [a, b] is compact. However, if  ∈ µ>0 , then a +  ∈ (a, b]∗ ,
but not in the monad of any element of (a, b], whence (a, b] is not compact.
Similarly, if N > N, then N ∈ [0, ∞)∗ but not in the monad of any element of
[0, ∞), whence [0, ∞) is not compact. Compactness is meant to generalize
the notion of closed, bounded intervals [a, b], but we’ll soon see that this
analogy breaks down in “infinite-dimensional” examples.
Corollary 9.14.
(1) If C ⊆ K, where C is closed and K is compact, then C is compact.
(2) If K is compact, then K is closed.
Proof. For (1), suppose p ∈ C ∗ ; we need to find q ∈ C such that p ≈ q.
Since p ∈ K ∗ as well, by compactness of K, p ≈ q for some q ∈ K. By
the nonstandard characterization of closed, we have q ∈ C. For (2), suppose
that p ∈ K ∗ , q ∈ X are such that p ≈ q; we need q ∈ K. Since K is compact,
there is q 0 ∈ K such that p ≈ q 0 . By Exercise 9.8, q ≈ q 0 ; since q, q 0 ∈ X, we
have that q = q 0 , whence q ∈ K. 

Here’s a question: What is the analog of Rfin for our metric space X? If we
use just the definition of Rfin , then we should make the following definition:
Definition 9.15. The set of finite points of X ∗ is
Xfin = {a ∈ X ∗ | d(a, b) ∈ Rfin for some b ∈ X}.
However, by Theorem 1.9, every element of Rfin is infinitely close to a
(standard) real number. This motivates:
Definition 9.16. The set of nearstandard elements of X ∗ is
Xns := {a ∈ X ∗ | a ≈ b for some b ∈ X}.
S
In other words, Xns = b∈X µ(b). Some remarks are in order:
Remarks 9.17.
(1) The “for some” in Definition 9.15 can be replaced with “for all,”
that is,
Xfin := {a ∈ X ∗ | d(a, b) ∈ Rfin for all b ∈ X}.
Indeed, suppose that a ∈ X ∗ and b ∈ X are such that d(a, b) ∈ Rfin .
For any other c ∈ X, we have
d(a, c) ≤ d(a, b) + d(b, c) ∈ Rfin + R ⊆ Rfin .
(2) It is immediate to see that Xns ⊆ Xfin . Sometimes we have equality;
for example, Theorem 1.9 says that Rns = Rfin . However, we often
have a strict inclusion Xns ( Xfin . For example, let X = C([0, 1], R)
from Example 9.2. By transfer, an element of X ∗ is a function
60 ISAAC GOLDBRING

f : [0, 1]∗ → R∗ that satisfies the -δ definition of continuity for

, δ ∈ R∗ . So consider f : [0, 1]∗ → R∗ given by

0
 if x ≤ 12
f (x) =  (x − 2 ) if 21 ≤ x ≤ 12 + 
1 1

1 otherwise,


where  ∈ µ>0 . Then f ∈ X ∗ , but since f makes an “appreciable”

jump in an infinitesimal time period, f cannot be infinitely close to
a standard, continuous function g : [0, 1] → R. We will later clarify
exactly when Xns = Xfin .
(3) Suppose p ∈ Xns . Then there is a unique q ∈ X such that p ≈ q:
indeed, if p ≈ q and p ≈ q 0 , where q, q 0 ∈ X, then q ≈ q 0 , so q = q 0 . In
analogy with earlier in these notes, we call this unique q the standard
part of p, denoted st(p).
Robinson’s characterization of compactness can now be phrased as:
Corollary 9.18. X is compact if and only if X ∗ = Xns .
Definition 9.19. B ⊆ X is bounded if there is a ∈ X and r ∈ R>0 such
that B ⊆ B(a; r).
Proposition 9.20. B ⊆ X is bounded if and only if B ∗ ⊆ Xfin .
Proof. Suppose that B is bounded, say B ⊆ B(a; r). Then B ∗ ⊆ B(a; r)∗ ;
clearly B(a; r)∗ is contained in Xfin . Conversely, suppose that B ∗ ⊆ Xfin .
Fix a ∈ X and N > N. Then the following is true in the nonstandard
universe:
(∃N ∈ N∗ )(∀x ∈ B ∗ )(d(a, x) < N ).
Now apply transfer. 

In particular, X is bounded if and only if X ∗ = Xfin .

Corollary 9.21. If K ⊆ X is compact, then K is bounded.
Proof. This follows from the inclusions K ∗ ⊆ Xns ⊆ Xfin . 
Definition 9.22. X is a Heine-Borel (or proper ) metric space if, for all
K ⊆ X, we have K is compact if and only if K is closed and bounded.
The name Heine-Borel metric space comes from the Heine-Borel Theorem
(to be proven below), which states that a subset of Rn is compact if and
only if it is closed and bounded. We have already seen that compact sets
are always closed and bounded, so the meat of the definition is the other
implication.
Proposition 9.23. X is a Heine-Borel metric space if and only if Xns =
Xfin .
 LECTURE NOTES ON NONSTANDARD ANALYSIS 61

Proof. First suppose that X is a Heine-Borel metric space and suppose that
a ∈ Xfin ; we need a ∈ Xns . Fix b ∈ X; then d(a, b) ∈ Rfin , say d(a, b) < r
with r ∈ R>0 . Then a ∈ B̄(b; r)∗ . Since X is Heine-Borel, we have B̄(b; r)
is compact, whence a ≈ c for some c ∈ B̄(b; r). In particular, a ∈ Xns .
Conversely, suppose that Xns = Xfin . Suppose that K ⊆ X is closed and
bounded; we need K to be compact. Fix a ∈ K ∗ ; we need a ≈ b for some
b ∈ K. Since K is bounded, we have K ∗ ⊆ Xfin , whence a ∈ Xfin = Xns .
Thus, there is (unique) b ∈ X such that a ≈ b. It remains to verify that
b ∈ K; but this follows immediately from the fact that K is closed (and the
nonstandard characterization of closed). 
Corollary 9.24. Rn is a Heine-Borel metric space. C([0, 1], R) is not a
Heine-Borel metric space.
We can define continuity between metric spaces. Suppose that Y is also
a metric space and f : X → Y is a function. For p ∈ X, we say that f is
continuous at p if whenever O ⊆ Y is open and f (p) ∈ O, then there is an
open O0 ⊆ X such that f (O0 ) ⊆ f (O). We say that f is continuous if f is
continuous at p for all p ∈ X.
The astute observer will notice that this is not the direct translation of
continuity for functions on R. However, the following exercise will make
them feel better:
Exercise 9.25. The following are equivalent:
(1) f is continuous at p;
(2) For all  ∈ R>0 , there is δ ∈ R>0 such that, for all q ∈ X, if
d(p, q) < δ, then d(f (p), f (q)) < ;
(3) f (µ(p)) ⊆ µ(f (p)), that is, if q ≈ p, then f (q) ≈ f (p).
We use the above definition for continuity as it makes sense in an arbitrary
topological space and not just for metric spaces. The equivalence of (1) and
(3) in the previous exercise will still hold in this more general context.
Proposition 9.26. Suppose that f : X → Y is continuous and K ⊆ X is
compact. Then f (K) ⊆ Y is compact.
Proof. Suppose y ∈ f (K)∗ ; we need y ∈ f (K)ns . By transfer, we have
y = f (x) for some x ∈ K ∗ . Since K is compact, st(x) exists and belongs to
K. Since f is continuous at st(x), y = f (x) ∈ µ(f (st(x)), so y ∈ f (K)ns . 
Exercise 9.27. Suppose that f : X → Y is a function.
(1) Define what it means for f to be uniformly continuous. Then state
and prove a nonstandard characterization of uniform continuity.
(2) Suppose that f is continuous and X is compact. Prove that f is
uniformly continuous.
For the purpose of the next exercise, define Xinf := X ∗ \ Xfin . A (not
necessarily continuous) function f : X → Y is said to be proper if f −1 (K) ⊆
X is compact for every compact K ⊆ Y .
62 ISAAC GOLDBRING

Exercise 9.28. Suppose that X and Y are Heine Borel metric spaces and
f : X → Y is continuous. Prove that f is proper if and only if f (Xinf ) ⊆ Yinf .
We can also bring the notions of sequences and convergence of sequences
into the metric space setting. For example, a sequence (an ) from X converges
to a ∈ X if and only if, for every  ∈ R>0 , there is m ∈ N such that, for all
n ∈ N, if n ≥ m, then d(an , a) < .
Here is the metric space version of Bolzano-Weierstrauss:
Theorem 9.29. If X is a compact metric space and (an ) is a sequence in
X, then an has a convergent subsequence.
Proof. Fix N > N. Then aN ∈ X ∗ = Xns . Then st(aN ) is a limit point of
(an ). 
Definition 9.30. X is a complete metric space if every Cauchy sequence in
X converges.
Corollary 9.31. Compact metric spaces are complete.
Proof. Suppose that (an ) is a Cauchy sequence in X, so aM ≈ aN for all
M, N > N. Since X is compact, aN ∈ Xns for all N > N. Thus, if L =
st(aN ) for N > N, then aM ≈ L for all M > N, whence (an ) converges to
L. 
Exercise 9.32. Suppose that X is complete and C ⊆ X is closed. Prove
that C is also complete.
In order to explain the nonstandard characterization of completeness, it
is convenient at this point to introduce another important set of points in
X ∗:
Definition 9.33. The set of pre-nearstandard points of X ∗ is
Xpns := {a ∈ X ∗ | for each  ∈ R>0 , there is b ∈ X such that d(a, b) < }.
Immediately, we see that Xns ⊆ Xpns ⊆ Xfin .
Theorem 9.34. X is complete if and only if Xns = Xpns .
Proof. First suppose that X is complete and p ∈ Xpns . Then, for every
n ≥ 1, there is qn ∈ X such that d(p, qn ) < n1 . It follows that (qn ) is Cauchy,
whence converges to q ∈ X. It follows that p ≈ q, whence p ∈ Xns .
Towards the converse, suppose that Xns = Xpns and suppose that (xn )
is Cauchy. Fix N > N; it suffices to show that xN ∈ Xns . If not, then
xN ∈/ Xpns , whence there is  ∈ R>0 such that d(xN , q) ≥  for all q ∈ X.
In particular, d(xN , xn ) ≥  for all n ∈ N. But (xn ) is Cauchy, so for some
n ∈ N big enough, d(xN , xn ) < , a contradiction. 
Corollary 9.35. If X is Heine-Borel, then X is complete.
The following theorem on “remoteness” will prove useful later in these
notes:
 LECTURE NOTES ON NONSTANDARD ANALYSIS 63

Theorem 9.36. Suppose that (pn | n ∈ N∗ ) is an internal sequence from X ∗ .

Suppose that r ∈ R>0 is such that d(pm , pn ) ≥ r for all distinct m, n ∈ N.
Then pn ∈/ Xns for some n ∈ N∗ .
Proof. Suppose, towards a contradiction, that pn ∈ Xns for all n ∈ N∗ . For
n ∈ N set qn := st(pn ). We then get the nonstandard extension of (qn ),
namely (qn | n ∈ N∗ ). We must be careful here: just because pn ≈ qn for
all n ∈ N does not imply that pn ≈ qn for all n ∈ N∗ (as the relation ≈
is external). Nevertheless, the sequence (d(pn , qn ) | n ∈ N∗ ) is internal and
infinitesimal for n ∈ N. Thus, by the Infinitesimal Prolongation Theorem,
there is N > N such that d(pn , qn ) ≈ 0 for all n ≤ N . Fix M > N with
M < N and set q := st(pM ) = st(qM ) (which is possible by our standing
assumption). Thus, there is a subsequence (qnk ) converging to q. Choose
n0 ∈ N such that, for j, k ≥ n0 : d(qnj , qnk ) < 2r . It follows that d(pnj , pnk ) <
r, a contradiction. 
9.2. More about continuity.
Definition 9.37. Suppose that f : X ∗ → Y ∗ is a function.
(1) f is S-continuous if, for all x, x0 ∈ X ∗ , if x ≈ x0 , then f (x) ≈ f (x0 ).
(2) f is δ-continuous if, for all p ∈ X ∗ and  ∈ R>0 , there is δ ∈ R>0
such that, for all q ∈ X ∗ , if d(p, q) < δ, then d(f (p), f (q)) < .
The important point in (2) is that both  and δ are standard. It is easy
to see that δ-continuity implies S-continuity. The converse need not hold:
Exercise 9.38. Consider f : R∗ → R∗ defined by f (x) = 0 if x ≈ 0 while
f (x) = 1 otherwise. Show that f is S-continuous but not δ-continuous.
Observe that the function f in the previous exercise is external. For
internal functions, the above notions coincide:
Proposition 9.39. Suppose that f : X ∗ → Y ∗ is internal and S-continuous.
Then f is δ-continuous.
Proof. Suppose, towards a contradiction, that there is  ∈ R>0 and p ∈ X ∗
such that, for every δ ∈ R>0 , there is q ∈ X ∗ such that d(p, q) < δ while
d(f (p), f (q)) ≥ . By saturation (how?), there is q ∈ X ∗ such that d(p, q) ≈
0 while d(f (p), f (q)) ≥ , contradicting S-continuity. 
The following construction is crucial in defining standard continuous func-
tions from internal functions.
Theorem 9.40. Suppose X and Y are metric spaces with X compact. Sup-
pose that f : X ∗ → Y ∗ is an internal, S-continuous function. Suppose fur-
ther that f (p) ∈ Yns for each p ∈ X. Define F : X → Y by F (p) = st(f (p)).
Then F is continuous and F (p) ≈ f (p) for all p ∈ X ∗ .
Proof. Fix p ∈ X; we show that F is continuous at p. Fix  ∈ R>0 . By
Proposition 9.39, there is δ ∈ R>0 witnessing that f is δ-continuous for
64 ISAAC GOLDBRING


2. Suppose q ∈ X is such that d(p, q) < δ. Then d(f (p), f (q)) < 2 . Since
f (p) ≈ F (p) and f (q) ≈ F (q), this shows that d(F (p), F (q)) < .
Now suppose that p ∈ X ∗ ; we need F (p) ≈ f (p). Let p0 := st(p); this
is possible since X is compact. Then f (p) ≈ f (p0 ) by S-continuity of f .
Meanwhile, F (p0 ) ≈ f (p0 ) by definition of F and F (p0 ) ≈ F (p) by continuity
of F . 

We recall the definition of equicontuity, this time in the metric space

setting: a sequence of functions (fn ) from X to Y is equicontinuous if, for
all  ∈ R>0 , there is δ ∈ R>0 such that, for all p, q ∈ X and n ∈ N: if
d(p, q) < δ, then d(fn (p), fn (q)) < .
Exercise 9.41. Suppose that X and Y are metric spaces and X is compact.
Set Z := C(X, Y ) to be the space of continuous functions from X to Y .
Define a function d on Z by setting d(f, g) := supx∈X d(f (x), g(x)).
(1) Show that d is a metric on Z. (In particular, this means showing
that the supremum is never infinite.)
(2) Show that (fn ) converges to f (in the sense of the metric space Z)
if and only if (fn ) converges to f uniformly.
If X is a compact metric space and fn : X → R is a continuous function
for each n ∈ N, we say that the sequence (fn ) is uniformly bounded if there
is M ∈ R such that |fn (x)| ≤ M for all x ∈ X and all n ∈ N.
Corollary 9.42 (Arzela-Ascoli). If X is a compact metric space and (fn )
is a uniformly bounded, equicontinuous sequence of functions from X to R,
then (fn ) has a uniformly convergent subsequence.
Proof. Fix N > N. As in an earlier part of the notes, fN : X ∗ → R∗ is
S-continuous (and internal). Since (fn ) is uniformly bounded, it follows
that fN (X) ⊆ Rns .Define F : X → R (as in the last theorem) by setting
F (p) := st(fN (p)). By the last exercise, it suffices to show that F is a
limit point of (fn ); to do this, we will show that F ≈ fN as elements of
C(X, R)∗ . Well, by transfer, d(F, fN ) <  if and only if d(F (p), fN (p)) < 
for all p ∈ X ∗ ; this follows immediately from the conclusion of the previous
theorem. 

9.3. Compact maps. We now discuss an important class of functions that

will appear later in the functional analysis section. Once again, X and Y
are metric spaces.
Definition 9.43. f : X → Y is compact if, for every bounded B ⊆ X, we
have a compact K ⊆ Y such that f (B) ⊆ K.

In other words, f is compact if and only if f (B) is compact.

Theorem 9.44. f : X → Y is compact if and only if f (Xfin ) ⊆ Yns .
 LECTURE NOTES ON NONSTANDARD ANALYSIS 65

Proof. First suppose that f is compact. Fix p ∈ Xfin ; we need f (p) ∈ Yns .
∗
Well, p ∈ B := B(a; r) for some a ∈ X and r ∈ R>0 , whence f (p) ∈ f (B) ⊆
Yns since f (B) is compact.
Conversely, suppose that f (Xfin ) ⊆ Yns . Fix B ⊆ X bounded; we must
∗
show that f (B) is compact. Take q ∈ f (B) ; we must find q 0 ∈ f (B)
such that q ≈ q 0 . Fix  ∈ µ>0 ; by transfer, there is y ∈ f (B)∗ such that
d(q, y) < . Write y = f (x) for x ∈ B ∗ . By assumption, f (x) ∈ Yns , so
f (x) ≈ q 0 for some q 0 ∈ Y . It remains to show that q 0 ∈ f (B). Fix δ ∈ R>0 .
By assumption, there is z ∈ f (B)∗ such that d(q, z) < δ, whence it follows
that d(q 0 , z) < δ. Applying transfer to this last fact, we see that there is
z ∈ f (B) such that d(q 0 , z) < δ. 
Corollary 9.45. Suppose that f : X → Y is a function. If Y is compact,
then f is compact.
Corollary 9.46. Suppose that (fn ) is a sequence of compact functions from
X to Y . Further assume that Y is complete and that (fn ) converges uni-
formly to f . Then f is compact.
Proof. Suppose x ∈ Xfin ; we need f (x) ∈ Yns . Since Y is complete, it
suffices to prove that f (x) ∈ Ypns . Fix  ∈ R>0 . Fix m ∈ N such that
d(fm (p), f (p)) < 2 for all p ∈ X. By transfer, d(fm (x), f (x)) < 2 . Since fm
is compact, we have fm (x) ∈ Yns , say fm (x) ≈ y with y ∈ Y . It follows that
d(f (x), y) < . Since  was arbitrary, this shows that f (x) ∈ Ypns . 
9.4. Problems. You may assume any level of saturation that you need in
any given problem.
Problem 9.1. Suppose that X is a metric space and A is a subset of X.
The interior of A, denoted A◦ , is defined by
A◦ := {x ∈ A | there exists r ∈ R>0 such that B(x, r) ⊆ A}.
(1) Show that A is open ◦
◦
S iff A = A . (Standard reasoning)
(2) Show that A = {O | O is open and O ⊆ A}. (Standard reason-
ing)
(3) Show that, for any x ∈ X, we have x ∈ A◦ iff y ∈ A∗ for any y ∈ X ∗
with y ≈ x.
Problem 9.2. Suppose that X is a metric space and A is a subset of X.
The closure of A, denoted A. is defined by
A := {x ∈ X | for any r ∈ R>0 , there is a ∈ A such that d(x, a) < r}.
(1) Show that A = {x ∈ X | there is (an ) from A such that an → x}.
(Standard reasoning)
T
(2) Show that A = {F | F is closed and A ⊆ F }. (Standard reasoning)
(3) Show that A is closed iff A = A. (Standard reasoning)
(4) Show that, for any x ∈ X, we have x ∈ A iff there is y ∈ A∗ such
that x ≈ y.
66 ISAAC GOLDBRING

Problem 9.3. Let C ⊆ Rfin be internal. Define st(C) := {st(c) | c ∈ C} ⊆

R. Prove that st(C) is closed.
Problem 9.4.
(1) Suppose f : X → Y is continuous. Show that f : X ∗ → Y ∗ is
∗-continuous.
(2) Suppose that f : X ∗ → Y ∗ is δ-continuous. Show that f is S-
continuous.
(3) Consider the function f : R∗ → R∗ defined by
(
0 if x ≈ 0
f (x) =
1 otherwise.
Show that f is S-continuous, but not δ-continuous.
(4) Let f : R → R be defined by f (x) = x2 . Show that f : R∗ → R∗ is
∗-continuous, but not S-continuous.
(5) Fix α ∈ µ>0 . Consider the function f : R∗ → R∗ defined by
(
α sin( x1 ) if x 6= 0
f (x) =
0 if x = 0.
Show that f is S-continuous at 0 but not ∗-continuous at 0.
Problem 9.5. Suppose that A ⊆ X ∗ is internal. Let
st(A) = {x ∈ X | x ≈ y for some y ∈ A}.
(1) Show that st(A) is closed.
(2) Suppose that A ⊆ Xns is internal. Show that st(A) is compact.
(3) Suppose that A ⊆ X. Show that A = st(A∗ ).
(4) We say that A ⊆ X is relatively compact if A is compact. Show that
A ⊆ X is relatively compact iff A∗ ⊆ Xns .
Problem 9.6. . Let C([0, 1], R) denote the set of all continuous functions
from [0, 1] to R. For f, g ∈ C([0, 1], R), set
d(f, g) := sup{|f (x) − g(x)| | x ∈ [0, 1]}.
(1) Show that C([0, 1], R)∗ is the set of ∗-continuous functions from
[0, 1]∗ to R∗ .
(2) Suppose f ∈ C([0, 1], R)∗ . Show that f ∈ C([0, 1], R)ns iff f is S-
continuous and f (x) ∈ Rfin for all x ∈ [0, 1].
(3) Show that C([0, 1], R)ns ( C([0, 1], R)fin .
(4) Convince yourself that the results of this problem remains true when
[0, 1] is replaced by any compact metric space X.
Problem 9.7. (Arzela-Ascoli Theorem-reformulated) Suppose that F ⊆
C([0, 1], R). Show that the following are equivalent:
(1) F is relatively compact;
(2) F ∗ ⊆ C([0, 1], R)ns ;
 LECTURE NOTES ON NONSTANDARD ANALYSIS 67

(3) F is equicontinuous and, for all x ∈ [0, 1], the set

Fx := {f (x) | f ∈ F}
is relatively compact.
(The notion of equicontinuity is exactly as in Section 3, namely, F is equicon-
tinuous iff for every  ∈ R>0 , there is δ ∈ R>0 such that for all x, y ∈ [0, 1]
and all f ∈ F, if |x − y| < δ, then |f (x) − f (y)| < . You will need to show
again that F is equicontinuous iff each f ∈ F ∗ is S-continuous.)
By (4) of Problem 9.6, the equivalence in the previous problem remains
true when [0, 1] is replaced by any compact metric space.

10. Banach Spaces

In this section, F denotes one of the two fields R or C. Let’s say a word
about C∗ . By transfer, the ∗
p elements of C are of the form z = x + iy for
∗ 2 2
x, y ∈ R and then |z| = x + y . It is then straightforward to verify that
Cfin = {z ∈ C∗ | |z| ∈ Rfin } = {x + iy | x, y ∈ Rfin } = Cns since Rfin = Rns .
If z = x + iy ∈ Cns , then st(z) = st(x) + i st(y).
10.1. Normed spaces.
Definition 10.1. If V is a vector space over F, then a norm on V is a
function k · k : V → R satisfying, for all x, y ∈ V and α ∈ F:
(1) kxk ≥ 0;
(2) kxk = 0 if and only if x = 0;
(3) kαxk = |α| · kxk;
(4) (Triangle Inequality) kx + yk ≤ kxk + kyk.
A normed space is a vector space equipped with a norm.
Example 10.2.
(1) For any p ≥ 1, the p-norm on Fn is given by kxkp := p ni=1 |xi |p .
pP

(2) There is an infinitary analog of the previous example. Fix p ≥ 1 and

set
P∞`p to be the set of all infinite sequences (αn ) from F such that
|α |p < ∞. Then ` is a vector space over F and we define the
n=1 n p
p-norm on `p in the exact same way: k(αn )kp := p ∞
pP
p
n=1 |αn | .
(3) Fix a compact metric space X and set C(X, F) := {f : X →
F | f is continuous}. Then C(X, F) is a (usually infinite-dimensional)
vector space over F and kf k := supx∈X |f (x)| defines a norm on
C(X, F).
Exercise 10.3. Suppose that (V, k·k) is a normed space. Define d : V ×V →
R by d(x, y) := kx − yk. Then d is a metric on V and d(x, y) = d(x − y, 0)
for any x, y ∈ V .
We will always treat a normed space as a metric space as in the previous
exercise. A normed space is called a Banach space if the associated metric
is complete.
68 ISAAC GOLDBRING

Exercise 10.4. Verify that all of the normed spaces from Example 10.2
are Banach spaces. (Hints: Don’t forget our nonstandard characterization
of completeness. Also, for showing that C(X, F) is a Banach space, don’t
forget about our nonstandard characterization of C(X, F)ns from Problem
9.6.)
Until otherwise stated, we fix normed spaces V and W ; we write d for
both of the associated metrics on V and W . For x ∈ V ∗ , we say that x is
infinitesimal if x ≈ 0, that is, d(x, 0) ∈ µ (equivalently, kxk ∈ µ). It follows
immediately that x ≈ y if and only if x − y is infinitesimal.
Lemma 10.5. If x, y ∈ V ∗ and x ≈ y, then kxk ≈ kyk. (The converse fails
miserably!)
Proof. We may suppose that kxk ≤ kyk Write y = x + (y − x). Then
kyk ≤ kxk + ky − xk ≈ kxk since y − x is infinitesimal. Thus, kxk ≈ kyk. 
Exercise 10.6.
(1) If α ∈ Ffin and x, y ∈ V ∗ are such that x ≈ y, show that αx ≈ αy.
(2) Prove that the addition and scalar multiplication maps + : V × V →
V and · : F × V → V are continuous (with respect to the metric d).
Please use the nonstandard characterization of continuity.
10.2. Bounded linear maps.
Proposition 10.7. Suppose that T : V → W is a linear transformation and
T is continuous at some x0 ∈ V . Then T is uniformly continuous.
Proof. We use the nonstandard characterization of uniform continuity: sup-
pose x, y ∈ V ∗ and x ≈ y. We show that T x ≈ T y. Well, x0 + (x − y) ≈ x0 ,
so by the continuity of T at x0 , we have T (x0 + x − y) ≈ T (x0 ). Thus,
T (x0 ) + T (x) − T (y) ≈ T (x0 ), whence T (x) ≈ T (y). (We have used the
transfer principle to infer that the nonstandard extension of T is also lin-
ear.) 
Exercise 10.8. Suppose that T : V → W is a linear transformation that is
continuous. Prove that ker(T ) := {x ∈ V | T (x) = 0} is a closed subspace
of V .
Definition 10.9. We say that a linear transformation T : V → W is
bounded if there is M ∈ R>0 such that kT xk ≤ M kxk for all x ∈ V .
The terminology in the above definition corresponds to the next fact:
Proposition 10.10. T : V → W is bounded if and only if {T (x) | kxk = 1}
is a bounded subset of W .
Proof. Let A := {T (x) | kxk = 1}. For the (⇒) direction, if kT xk ≤ M kxk
for all x ∈ V , then A is contained in the closed ball around 0 (in W ) of radius
M . Conversely, suppose A is contained in the closed ball around 0 of radius
M . We claim that kT xk ≤ M kxk for all x ∈ V . Indeed, for x ∈ V \ {0},
1 1
k kxk xk = 1, so kT ( kxk x)k ≤ M , whence kT xk ≤ M kxk. 
 LECTURE NOTES ON NONSTANDARD ANALYSIS 69

If T : V → W is bounded, set kT k := sup{kT (x)k | kxk = 1}, which is a

real number by the previous proposition. Taking M = kT k in the proof of
the (⇐) direction of the previous proposition, we obtain the following:
Corollary 10.11. kT xk ≤ kT kkxk for all x ∈ V .
Definition 10.12. We set B(V, W ) to be the set of all bounded linear
transformations from V to W .
Exercise 10.13. Show that B(V, W ) is a vector subspace of the set of all
linear transformations from V to W . Further show that B(V, W ) is a normed
space (with the above definition of kT k). If V = W , prove that whenever
T, U ∈ B(V, V ), then T ◦ U ∈ B(V, V ) and kT ◦ U k ≤ kT k · kU k. (This shows
that B(V, V ) is a normed algebra.)
Theorem 10.14. Suppose that T : V → W is linear. The following are
equivalent:
(1) T is continuous at x0 for some x0 ∈ V ;
(2) T is uniformly continuous;
(3) T is bounded;
(4) T (Vfin ) ⊆ Wfin ;
(5) T (Vns ) ⊆ Wns .
Proof. The equivalence of (1) and (2) follows from Proposition 10.7 and
(2) ⇒ (5) is clear. Suppose that T is not bounded. Let
X = {n ∈ N∗ | kT (x)k > n for some x ∈ V ∗ with kxk = 1},
an internal subset of N∗ that contains N by assumption. Thus, by overflow,
there is N ∈ X \ N. Choose x ∈ V ∗ with kT xk > N and kxk = 1. At this
1
point we have established the implication (4) ⇒ (3). Set y := kT (x)k x, so
1
kyk < N , whence y ≈ 0. But kT (y)k = 1, so T (y) 6≈ y, that is, T is not
continuous. This proves the direction (2) ⇒ (3). Set z := √ 1 x; since
kT xk
kzk ∈ µ, we have z ∈ Vns . But kT zk = √ 1 kT xk ∈ Rinf , so T z ∈
/ Wns ;
kT xk
this shows (5) ⇒ (3). For (3) ⇒ (1), if x ≈ 0, then kT xk ≤ kT k · kxk ≈ 0, so
T x ≈ 0 and T is continuous at 0. For (3) ⇒ (4), assume that T is bounded
and fix x ∈ Vfin . Then kT (x)k ≤ kT k · kxk ∈ Rfin , whence T (x) ∈ Wfin . 
10.3. Finite-dimensional spaces and compact linear maps. We now
aim to understand what happens for finite-dimensional normed spaces. First,
some lemmas.
Lemma 10.15. Suppose that x1 , . . . , xn ∈ Vfin , α1 , . . . , αn ∈ F∗ are such
that α1 , . . . , αn ≈ 0. Then α1 x1 + · · · + αn xn ≈ 0.
Proof. Immediate from the triangle inequality. 
Lemma 10.16. Suppose that x1 , . . . , xn ∈ V are linearly independent and
α1 , . . . , αn ∈ F∗ are such that α1 x1 + · · · + αn xn ∈ Vfin . Then αi ∈ Ffin for
all i = 1, . . . , n.
70 ISAAC GOLDBRING

Proof. Without loss of generality (by renumbering if necessary), we may

assume that max{|α1 |, . . . , |αn |} = |α1 |. Suppose, towards a contradiction,
that α1 ∈ / Ffin . Then kx1 + αα21 x2 + · · · + ααn1 xn k = |α11 | kα1 x1 + · · · + αn xn k ≈ 0
since | α11 | ≈ 0 and kα1 x1 + · · · + αn xn k ∈ Rfin . Since |αi | ≤ |α1 | for each
i = 1, . . . , n, st(| αα1i |) exists. By the previous lemma, we have
α2 αn α2 αn
0 ≈ x1 + x2 + · · · + xn ≈ x1 + st( )x2 + · · · + st( )xn .
α1 α1 α1 α1
Since the term on the right of the above display is standard, it equals 0.
This contradicts the fact that x1 , . . . , xn are linearly independent. 
Corollary 10.17. If V is finite-dimensional, then Vfin = Vns
Proof. Let {x1 , . . . , xn } be a basis for V . By transfer, every element of V ∗ is
a F∗ -linear combination of x1 , . . . , xn . If x = α1 x1 + . . . + αn xn ∈ Vfin , then
each αi ∈ Ffin = Fns by the previous lemma, whence x is infinitely close to
st(α1 )x1 + · · · + st(αn )xn ∈ V . 
Corollary 10.18. If V is finite-dimensional, then V is a Banach space.
Corollary 10.19. If T : V → W is a linear transformation and V is finite-
dimensional, then T is bounded.
Proof. By Theorem 10.14, it suffices to prove that T (Vfin ) ⊆ Wfin . Fix
x = α1 x1 + · · · + αn xn ∈ Vfin , so each αi ∈ Ffin . Thus,
T (x) = α1 T (x1 ) + · · · + αn T (xn ) ≈ st(α1 )T (x1 ) + · · · + st(αn )T (xn ) ∈ W,
so T (x) ∈ Wns ⊆ Wfin . 
We now introduce a very important class of linear transformations that
will play a prominent role in the rest of these notes. In some sense, they
are the transformations on infinite-dimensional spaces that behave most like
transformations between finite-dimensional spaces.
Definition 10.20. If T : V → W is a linear transformation, we say that T
is a compact transformation if T is a compact map of the associated metric
spaces, that is, for every bounded B ⊆ V , there is a compact K ⊆ W such
that T (B) ⊆ K. (Or in nonstandard terms, T (Vfin ) ⊆ Wns .)
By Theorem 10.14, a compact linear transformation is automatically
bounded.
Exercise 10.21. Let B0 (V, W ) denote the set of compact linear transfor-
mations from V to W . Show that B0 (V, W ) is a subspace of B(V, W ).
Observe that the proof of Corollary 10.19 actually shows
Corollary 10.22. If T : V → W is a linear transformation and V is finite-
dimensional, then T is compact.
In fact, we can generalize the previous corollary, but first we need a defi-
nition.
 LECTURE NOTES ON NONSTANDARD ANALYSIS 71

Definition 10.23. If T : V → W is a linear transformation, we say that T

is finite-rank if T (V ) is a finite-dimensional subspace of W .
Certainly, if V is finite-dimensional, then T : V → W is of finite-rank.
Lemma 10.24. If T : V → W is of finite rank and bounded, then T is
compact.
Proof. We know that T (Vfin ) ⊆ Wfin . Set W 0 := T (V ). Then Wfin
0 =
0 0 0
Wns by Corollary 10.17. Thus T (Vfin ) ⊆ Wfin = Wns ⊆ Wns , whence T is
compact. 
Exercise 10.25. Let B00 (V, W ) denote the set of finite-rank bounded trans-
formations from V to W . Show that B00 (V, W ) is a subspace of B0 (V, W ).
We now consider convergence of transformation.
Lemma 10.26. Suppose that B ⊆ V is a bounded set and Tn , T : V → W
are all bounded linear transformations. Further assume that kTn − T k → 0
as n → ∞. Then Tn → T uniformly on B.
Proof. Choose M ∈ R>0 such that kxk ≤ M for all x ∈ B. We need
to show that TN (x) ≈ T (x) for all x ∈ B ∗ and all N ∈ N∗ \ N. Well,
kTN (x)−T (x)k ≤ kTN −T kkxk ≤ M ·kTN −T k ≈ 0 since kTN −T k ≈ 0. 
Exercise 10.27. Suppose that W is a Banach space. Prove that the normed
space B(V, W ) is also a Banach space.
Proposition 10.28. Suppose that W is a Banach space, Tn , T : V → W
are bounded linear operators and each Tn is compact. Further suppose that
kTn − T k → 0 as n → ∞. Then T is compact.
Proof. Let x ∈ Vfin ; we need T (x) ∈ Wns . Let B ⊆ V be a ball around 0
such that x ∈ B ∗ . Since Tn converges uniformly to T on B, we know that
T |B : B → W is compact. Thus, T (B) is contained in a compact subset of
W , whence T (B)∗ ⊆ Wns . In particular, T (x) ∈ Wns . 
The previous proposition can be rephased as saying that, when W is a Ba-
nach space, that B0 (V, W ) is a closed subspace of B(V, W ), whence also a Ba-
nach space. By Exercise 10.25, the closure of B00 (V, W ) is a closed subspace
of B0 (V, W ), in symbols: B00 (V, W ) ⊆ B0 (V, W ). In the next section, we will
encounter a certain class of Banach spaces where B00 (V, W ) = B0 (V, W ).
10.4. Problems. Throughout, V denotes a normed space. F denotes either
R or C.
Problem 10.1. Suppose that x ∈ V ∗ is such that x ≈ 0. Show that there
is N ∈ N∗ \ N such that N x ≈ 0.
Problem 10.2. Suppose that
P∞V is a Banach space andP(x n | n ∈ N) is a
∞
sequence from V such that n=1 kxn k < ∞. Show that n=1 xn converges
in V .
72 ISAAC GOLDBRING

For the next problem, you will need to use the following:
Fact 10.29. If V is a normed space, then {x ∈ V | kxk ≤ 1} is compact if
and only if V is finite-dimensional.
Problem 10.3. Suppose that V is a Banach space and T : V → V is a
compact linear operator.
(1) Show that the identity operator I : V → V is compact if and only if
V is finite-dimensional.
(2) Suppose that U : V → V is any bounded linear operator. Show that
T ◦ U and U ◦ T are also compact.
(3) Suppose that T is invertible. Show that T −1 is compact if and only
if V is finite-dimensional.
Problem 10.4. Suppose K : [0, 1] × [0, 1] → R is a continuous function.
Suppose that T : C([0, 1], R) → C([0, 1], R) is defined by
Z 1
T (f )(s) := f (t)K(s, t)dt.
0
Show that T is a compact linear operator. (Hint: Use our earlier charac-
terization of C([0, 1], R)ns .) Such an operator is called a Fredholm Integral
Operator.
Problem 10.5. Suppose that k · k1 and k · k2 are both norms on a vector
space W . We say that k · k1 and k · k2 are equivalent if there exist constants
c, d ∈ R>0 such that, for all x ∈ W , we have
ckxk1 ≤ kxk2 ≤ dkxk1 .
For x, y ∈ W∗and i = 1, 2, let us write x ≈i y to mean kx − yki ≈ 0.
(1) Suppose that, for all x ∈ W ∗ , if x ≈1 0, then x ≈2 0. Show that
{kxk2 | x ∈ W, kxk1 ≤ 1} is bounded.
(2) Suppose that {kxk2 | x ∈ W, kxk1 ≤ 1} is bounded. Let
d := sup{kxk2 | x ∈ W, kxk1 ≤ 1}.
Show that kxk2 ≤ dkxk1 for all x ∈ W .
(3) Show that k · k1 and k · k2 are equivalent iff for all x ∈ W ∗ , we have
x ≈1 0 iff x ≈2 0.
(4) Show that (W, k · k1 ) is a Banach space iff (W, k · k2 ) is a Banach
space.
(5) Suppose A ⊆ W . For i = 1, 2, say that A is openi if A is open with
respect to the metric associated to k · ki . Show that k · k1 and k · k2
are equivalent iff for all A ⊆ W , we have A is open1 iff A is open2 .
(In fancy language, this exercise says that two norms are equivalent
if and only if they induce the same topology on W .)
Problem 10.6.
(1) Let k · k be any norm on Rn (not necessarily the usual norm on Rn ).
Suppose x ∈ (R∗ )n . Show that kxk ≈ 0 iff |xi | ≈ 0 for i = 1, . . . , n.
 LECTURE NOTES ON NONSTANDARD ANALYSIS 73

(2) Conclude that any two norms on a finite-dimensional vector space

are equivalent. R1
(3) For f ∈ C([0, 1], R), define kf k1 := 0 |f (x)|dx. Show that k · k1 is
a norm on C([0, 1], R).
(4) Show that (C([0, 1], R), k·k1 ) is not a Banach space. Thus k·k1 is not
equivalent to the norm k·k∞ on C([0, 1], R) considered earlier. Hence,
for infinite-dimensional vector spaces, there can exist inequivalent
norms.
(5) Give a direct proof that the identity function
I : (C([0, 1], R), k · k1 ) → (C([0, 1], R), k · k∞ )
is an unbounded linear operator. (This follows from earlier problems,
but it would be nice to find f ∈ C([0, 1], R)∗ such that kf k1 ∈ Rfin
but kf k∞ ∈/ Rfin .)

11. Hilbert Spaces

Once again, F denotes either R or C.

11.1. Inner product spaces.

Definition 11.1. An inner product space (i.p.s.) is a vector space V over
F equipped with an inner product, which is a function h·, ·i : V × V → F
satisfying, for all x, y, z ∈ V and α ∈ F:
• hx, xi ∈ R and hx, xi ≥ 0;
• hx, xi = 0 ⇔ x = 0;
• hx, yi = hy, xi (complex conjugate);
• hαx + y, zi = αhx, zi + hy, zi.
Throughout, (V, h·, ·i) denotes an inner product space over F.
Exercise 11.2. Show that, for all x, y, z ∈ V and α ∈ F:
(1) hx, αy + zi = αhx, yi + hx, zi.
(2) hx, 0i = 0.
p
Exercise 11.3. Define k · k : V → R by kxk := hx, xi. Show that k · k is
a norm on V .
By the previous exercise, we can consider an i.p.s. over F as a normed
space over F, and hence as a metric space as well. We will often consider V
as an i.p.s., normed space, and metric space all at the same time. (In fact,
it is the interplay between these three structures on V that is what is most
interesting.)
The following is encountered in a first course in linear algebra:
Theorem 11.4 (Cauchy-Schwarz Inequality). For all x, y ∈ V , we have
|hx, yi| ≤ kxk · kyk.
74 ISAAC GOLDBRING

Corollary 11.5. Equip V × V with the metric d given by

d((x1 , x2 ), (y1 , y2 )) := max{kx1 − y1 k, kx2 − y2 k}.
Then the inner product h·, ·i : V × V → F is continuous.
Proof. Fix x, y, u, v ∈ V . Then
khx, yi − hu, vik = khx, yi − hx, vi + hx, vi − hu, vik
≤ khx, y − vik + khx − u, vik
≤ kxk · ky − vk + kx − uk · kvk.
Now suppose that (x, y) ∈ V × V and (u, v) ∈ µ((x, y)). Then x ≈ u and
y ≈ v. Since kxk, kvk ∈ Rfin , it follows from the transfer of the above
inequality that hx, yi ≈ hu, vi. 
Example 11.6. The main example of an i.p.s. encountered in a first course
on linear algebra is Fn , equipped with the standard inner product h~x, ~y i :=
P n n
i=1 xi yi . Observe that this induces the norm on F introduced in the
previous section.
Example 11.7. The infinite-dimensional analogue of the previous example
is `2 as defined inP the previous section. The inner product on `2 is given
by h(xn ), (yn )i := ∞ n=1 xn yn . To see that this sum converges, use Cauchy-
m
Schwarz for F (applied to (|x1 |, . . . , |xm |) and (|y1 |, . . . , |ym |)) to see that
v v
m um m u∞ ∞
X uX X uX X
|xn yn | ≤ t 2
|xn | · 2
|yn | ≤ t |xn | ·2 |yn |2 .
n=1 n=1 n=1 n=1 n=1

Now let m → ∞. It is now easy to verify that the axioms for an i.p.s. hold.
We should remark that of all the `p spaces, `2 is the only one that carries the
structure of an i.p.s. and the above inner product on `2 induces the norm
on `2 introduced in the previous section.
Example 11.8. Let V = C([0, 1], F). Then V becomes an i.p.s. when
R1
equipped with the inner product given by hf, gi := 0 f (x)g(x)dx. How
does the norm on V induced by the inner product compare with the norm
placed on V in the previous section?
Definition 11.9. V is called a Hilbert space if the metric associated to V
is complete.
In other words, an i.p.s. is a Hilbert space if the associated normed space
is a Banach space.
Exercise 11.10.
(1) Prove that the inner product spaces in Exercises 11.6 and 11.7 are
Hilbert spaces. (Your proof for `2 should probably be standard as
we have yet to characterize `2ns .)
(2) Prove that the inner product space in Exercise 11.8 is not a Hilbert
space.
 LECTURE NOTES ON NONSTANDARD ANALYSIS 75

11.2. Orthonormal bases and `2 .

Definition 11.11.
(1) If x, y ∈ V , we say that x and y are perpendicular or orthogonal if
hx, yi = 0. We sometimes write x ⊥ y to indicate that x and y are
orthogonal.
(2) A set {ei | i ∈ I} from V is called orthonormal if
• kei k = 1 for each i ∈ I, and
• ei , ej are orthogonal for all i 6= j.
Lemma 11.12. If {ei | i ∈ I} is orthonormal, then it is also linearly inde-
pendent.
Proof. Suppose m
P
n=1 cn ein = 0. Then
m
X m
X
ck = cn hein , eik i = h cn ein , eik i = h0, eik i = 0.
n=1 n=1


Definition 11.13. An orthonormal basis for V is a maximal orthonormal
sequence of vectors for V .
By Zorn’s lemma, every inner product space has an orthonormal basis.
One must be careful with the word basis here: while in finite-dimensional
inner product spaces, an orthonormal basis is a basis (in the usual linear
algebra sense), for infinite-dimensional inner product spaces, an orthonormal
basis is never a basis. (In this setting, the usual notion of “basis” is called
“Hamel basis” to help make the distinction.)
Fact 11.14. Let (en | n ∈ N) be an orthonormal set of vectors for the Hilbert
space H. Then the following are equivalent:
(1) (en ) is an orthonormal basis for V ;
(2) If v ∈ V is such that v ⊥ en for each n, then v = 0;
(3) For all v ∈ V , there is a sequence (αn ) from F such that m
P
n=0 αn en
converges to v as m → ∞.
You will prove the
P previous fact in the exercises. In (3) of the previous
fact, we write v = ∞ n=0 αn en .

Example 11.15. For n ∈ N, let bn ∈ `2 be defined by

bn := (0, 0, . . . , 0, 1, 0, . . . , 0),
where the 1 is in the nth spot. Then certainly (bn ) is an orthonormal se-
quence. Fix a = (a1 , a2 , . . .) ∈ `2 . Then
m
X ∞
X
ka − an bn k2 = kan k2 → 0
n=1 n=m+1
76 ISAAC GOLDBRING

as m → ∞ since a ∈ `2 . It follows that (bn ) is an orthonormal basis for

`2 , often referred to as the standard orthonormal basis for `2 . Observe that
(bn ) is not a Hamel basis as the vector ( n12 ) is not in the span of (bn ).
Fact 11.16. Every orthonormal basis for `2 is countable.
More generally, for any inner product space V , every two orthonormal
bases for V have the same cardinality. We will often refer to an inner
product space as being separable when its orthonormal bases are countable.
(It is a fact that all separable Hilbert spaces are isomorphic to `2 , but we
will not need this fact.)
Until further notice, let us fix a separable Hilbert space H (which you
may think of as `2 ). We also fix an orthonormal basis (en | n ∈ N) for H.
Lemma 11.17. For a = ∞ αn en and b = ∞
P P
P∞ n=0 n=0 βn en , we have ha, bi =
n=0 α n βn .
Proof. Fix m ∈ N. Then it is straightforward to verify that
m
X m
X m
X
h αn en , βn en i = αn βn .
n=0 n=0 n=0
It remains to let m → ∞ and use the fact that the inner product is contin-
uous. 
Corollary 11.18. For a = n=0 αn en , we have kak2 = ∞
P∞ P 2
n=0 |αn | .
Corollary P 11.19. For a ∈ H, there is a unique sequence (αn ) from F such
that a = ∞ n=0 αn en .
P∞ P∞ 0
P∞
Proof. Suppose that n=0 αn en = a = n=0 αn en . Then n=0 (αn −
αn0 )en = 0, whence ∞ 0 |2 = k0k2 = 0.
P
n=0 |αn − α n 
Let’s bring the nonstandard description into the picture: We get a non-
standard extension of the orthonormal basis (en | n ∈ N∗ ); this set is still
orthonormal by transfer. Also, for every P a ∈ H ∗ , there is a unique sequence
(αn | n ∈ N∗ ) from F∗ such that a = n∈N∗ αn en in the sense that, for
every  ∈ (R>0P )∗ , there is m0 ∈ N∗ such that, for all m ∈ N∗ with
Pm m ≥ m0 ,
we have ka − m α
n=0 n ne k < . But what do we mean by n=1 αn en if
m > N? Well, we have the set FinSeq(V ) of finite sequences from V , which
is a subset of P(N × V ), and the corresponding function Σ : FinSeq(V ) → V
given by adding the elements of the finite sequence. Consequently, we get
a function Σ : FinSeq(V )∗ → V ∗ , given by “adding” the elements of the
hyperfinite sequence.
In a similar vein, we have kak2 = n∈N∗ |αn |2 , with the interpretation of
P
the sum as in the previous paragraph. We have thus proven:
Proposition 11.20. For a = n∈N∗ αn en ∈ H ∗ , we have a ∈ Hfin if and
P

only if n∈N∗ |αn |2 ∈ Rfin .

P

Characterizing the nearstandard points is a bit more subtle:

 LECTURE NOTES ON NONSTANDARD ANALYSIS 77
P
Theorem 11.21. Suppose that a = n∈N∗ αn en is in Hfin . Then a ∈ Hns
2
P
if and only
P if, for every k > N, we have n>k |αn | ≈ 0. In this case,
st(a) = n∈N st(αn )en .
P
Proof. First suppose that a ∈ H, so a = n∈N αn en . For k ∈ N, set
P∞ 2
rk = n=k+1 |αn | . Thus, rk → 0 as k → ∞ and hence rk ≈ 0 if k > N;
the result for a ∈ H. Now suppose that a ∈ Hns , say a ≈ b ∈ H.
this proves P
Write b = ∞ n=1 βn en . Fix k > N. Then, by the transfer of the triangle
inequality for `2 , we have
sX sX sX
2
|αn | ≤ 2
|αn − βn | + |βn |2 = (†) + (††).
n>k n>k n>k

ka − bk2
= n∈N∗ |αn − βn |2 ≈ 0, whence (†) ≈ 0.
P
Since a ≈ b, we have that
It remains to prove that (††) ≈ 0. However, since b ∈ H, by the first part of
P (††) ≈ 0.
the proof, we know that
Now suppose that n>k |αn |2 ≈ 0 for every k > N; we must show that a ∈
Hns . Since a ∈ Hfin , we know that n∈N∗ |αn |2 ∈ Rfin ; say n∈N∗ |αn |2 ≤
P P

M ∈ R>0 . In particular, αn ∈ P Ffin for each n ∈ N∗ . For n ∈ N, set

βn := st(αn ). We claim that b := n∈N βn en defines an element of H. For
m ∈ N, define sm := m
P 2
Pm 2
n=0 |βn | . Then sm ≈ n=0 |αn | ≤ M . Hence,
(sm ) is a bounded, nondecreasing sequence in R, whence convergent. For
m, k ∈ N with m ≤ k, we have
m
X k
X k
X
k βn en − β n e n k2 ≤ |βn |2 .
n=0 n=0 n=m+1
P∞
Since (sm ) is Cauchy, it follows that
P∞ the partial sums of n=0 βn en are
Cauchy, whence, by completeness, n=0 βn en defines an element of H.
It remains to verify that a ≈ b. To see this, observe that, for any k ∈ N∗ ,
we have
X X X
ka − bk2 = |αn − βn |2 = |αn − βn |2 + |αn − βn |2 .
n∈N∗ n≤k n>k

|2
P
If k ∈ N, then n≤k |αn − βn ≈ 0; thus, by the Infinitesimal Prolongation
Theorem, there is k > N such that n≤k |αn − βn |2 ≈ 0. On the other hand,
P
X X X
|αn − βn |2 ≤ |αn |2 + |βn |2 .
n>k n>k n>k

|2 2
P P
By assumption, n>k |αn ≈ 0, whilst n>k |βn | ≈ 0 by the forward
direction of the theorem and the fact that b is standard. Consequently,
ka − bk2 ≈ 0. 
P∞ Pm
For a = n=0 αn en ∈ H and m ∈ N, set P (m, a) := Pm (a) = n=0 αn en ∈
H. We thus have maps P : N × H → H and, for n ∈ N, Pn : H → H.
Exercise 11.22. For n ∈ N, Pn is a bounded linear transformation with
kPn k = 1.
78 ISAAC GOLDBRING

Extending to the nonstandard universe, we get a map P : N∗ × H ∗ → H ∗ .

For N ∈ N∗ , we define PN : H ∗ → H ∗ by PN (a) = P (N, a). For a =
P ∗
PN
n∈N∗ αn en ∈ H , we may view PN (a) as the hyperfinite sum n=0 αn en ∈
H ∗.

Theorem 11.23. Suppose N > N and a ∈ Hns . Then PN (a) ≈ a.

P
Proof. Write a = n∈N∗ αn en . Then, by Theorem 11.21, we have
N
X X X
ka − αn en k2 = k αn en k2 = |αn |2 ≈ 0.
n=0 n>N n>N

Recall that we have a norm k · k on B(H). Going to the nonstandard

universe, we get an internal norm k · k : B(H)∗ → R∗ . By transfer, for
T ∈ B(H)∗ , we have kT k is the internal suprememum of the set {kT (x)k | x ∈
H ∗ , kxk = 1}. By the transfer of Exercise 11.22, for N > N, we have
PN ∈ B(H)∗ and kPN k = 1.

Corollary 11.24. Suppose that T : H → H is a compact operator and

N > N. Then kT − PN T k ≈ 0.

Proof. Suppose a ∈ H ∗ and kak = 1. Then a ∈ Hfin , whence T (a) ∈

Hns by the compactness of T . Thus, by Theorem 11.23, T (a) ≈ PN T (a).
Since the internal supremum of an internal set of infinitesimals is once again
infinitesimal, this shows that kT − PN T k ≈ 0. 

The following lemma will be crucial in the proof of the Bernstein-Robinson

Theorem in Section 14.

Lemma 11.25. Suppose that T ∈ B(H) is compact, {e1 , e2 , . . .} is an or-

thonormal basis for H, and [ajk ] is the “matrix”
P for T with respect to this
basis, that is, for all k ≥ 1, we have T (ek ) = j∈N ajk ej . Then ajk ≈ 0 for
all j ∈ N∗ \ N and k ∈ (N∗ )>0 . (So the “infinite” rows of the matrix for T
consist entirely of infinitesimals.)

Proof. Fix k ∈ (N∗ )>0 . Since

P T is compact and ek is finite, T (ek ) ∈ Hns ,
say T (ek ) ≈ y. Write y = i∈N∗ yi ei . Set
X X
η := kT (ek ) − yk2 = k (aik − yi )ei k2 = |aik − yi |2 .
i∈N∗ i∈N∗

Since T (ek ) ≈ y, we have η ≈ 0. Now suppose j ∈ N∗ \ N. Then |ajk − yj |2 ≤

η, whence |ajk − yj | ≈ 0. Since y ∈ H, yj ≈ 0. Since |ajk | ≤ |ajk − yj | + |yj |,
we get that ajk ≈ 0. 
 LECTURE NOTES ON NONSTANDARD ANALYSIS 79

11.3. Orthogonal projections. In this subsection, we assume that H is an

arbitrary Hilbert space. For u, v ∈ H, we have the Parallelogram Identity:
ku + vk2 + ku − vk2 = 2(kuk2 + kvk2 ).
If, in addition, u and v are orthogonal, we have the Pythagorean Theorem:
ku + vk2 = kuk2 + kvk2 . By induction, we see that whenever u1 , . . . , un are
mutually orthogonal, then ku1 + · · · + un k2 = ku1 k2 + · · · + kun k2 .

Lemma 11.26. If {t1 , . . . , tk } is an orthonormal set in H, then sp(t1 , . . . , tk )

is a closed subspace of H.

Later we will be able to remove the “orthonormal” assumption.

Proof. Set E := sp(t1 , . . . , tk ). Suppose x ∈ E ∗ and y ∈ H are such that

x ≈ y; we need y ∈ E. By transfer, there are α1 , . . . , αk ∈ F∗ such that
x = α1 t1 + · · · + αk tk . By the (transfer of the) Pythagorean theorem, we see
that kxk2 = kα1 t1 k2 +· · ·+kαk tk k2 = |α1 |2 +· · ·+|αk |2 . Since x ∈ Hns , we see
that kxk ∈ Rfin , whence each αi ∈ Ffin . Set z := st(α1 )t1 +· · ·+st(αk )tk ∈ E.
By the Pythagorean theorem again, we see that
kx − zk2 = |α1 − st(α1 )|2 + · · · + |αk − st(αk )|2 ≈ 0.
By uniqueness of standard part, we see that y = z ∈ E. 

The next result is crucial in the study of Hilbert spaces.

Theorem 11.27 (Existence of Orthogonal Projection). Suppose that E is

a closed subspace of H. Then for each x ∈ H, there is a unique y ∈ E such
that kx − yk ≤ kx − zk for all z ∈ E.

Proof. The result is obvious if x ∈ E (take y = x). Thus, we may assume

that x ∈/ E. Set α := inf{kx − zk | z ∈ E}; note that α > 0 since E is
/ E. Fix  ∈ µ>0 . By transfer, there is z ∈ E ∗ such that
closed and x ∈
α ≤ kx − zk < α + ; in particular, kx − zk ≈ α.

Claim: z ∈ Ens .
Proof of Claim: Suppose, towards a contradiction, z ∈ / Ens . Since H is
complete and E is closed in H, we have that E is complete. Consequently,
since z ∈
/ Ens , we have that z ∈ whence there is r ∈ R>0 such that
/ Epns ,q
2
kz − wk ≥ r for all w ∈ E. Since α < α2 + r4 , we have w ∈ E such that
q
2
kx − wk < α2 + r4 . By the Parallelogram Identity, we have:

k(x − w) + (x − z)k2 + kw − zk2 = 2(kx − wk2 + kx − zk2 ).

Now k(x − w) + (x − z)k = 2kx − 12 (w + z)k ≥ 2α by transfer and the fact

2
that 12 (w + z) ∈ E ∗ . Since kx − zk ≈ α, we have that 2kx − zk2 < 2α2 + r2 .
80 ISAAC GOLDBRING

Thus:
kw − zk2 = 2(kx − wk2 + kx − zk2 ) − k(x − w) + (x − z)k2
r2
< 2(α2 + ) + 2kx − zk2 − 4α2
4
r2 r2
< 2(α2 + ) + 2α2 + − 4α2
4 2
< r2 .
This contradicts the fact that kz − wk ≥ r.

Set y := st(z) ∈ E. Since x − z ≈ x − y, we see that kx − yk = α, whence

kx − yk ≤ kx − zk for all z ∈ E.
To prove uniqueness: suppose that y 0 ∈ E also satisfies the conclusion of
the theorem. In particular, kx − yk = kx − y 0 k = α. Then
1
ky − y 0 k2 = 2kx − yk2 + 2kx − y 0 k2 − 4kx − (y + y 0 )k2 ≤ 4α2 − 4α2 = 0,
2
whence y = y . 0 
Definition 11.28. If E is a closed subspace of H and x ∈ E, we let PE (x)
denote the unique element of y fulfilling the conclusion of the previous the-
orem; we refer to PE (x) as the orthogonal projection of x onto E.
Definition 11.29. For E a closed subspace of H, we set
E ⊥ := {x ∈ H | x ⊥ z for all z ∈ E}.
Until further notice, E denotes a closed subspace of H.
Lemma 11.30. E ⊥ is a closed subspace of H.
Proof. We leave it to the reader to check that E ⊥ is a subspace of H. To
check that E ⊥ is closed, suppose that x ∈ (E ⊥ )∗ and y ∈ H are such that
x ≈ y; we need to show that y ∈ E ⊥ . Fix z ∈ E; we need hy, zi = 0. Well,
by continuity of the inner product, hy, zi ≈ hx, zi = 0 by transfer. Since
hy, zi ∈ R, it follows that hy, zi = 0. 
Much of what follows is easy and standard but will be needed in the next
sections. We will thus omit many proofs.
Lemma 11.31. For any x ∈ H, x − PE (x) ∈ E ⊥ .
Lemma 11.32. For x, y ∈ H, we have y = PE (x) if and only if y ∈ E and
x − y ∈ E ⊥ . Thus, if x ∈ E ⊥ , then PE (x) = 0.
Theorem 11.33 (Gram-Schmidt Process). Let (un | n ∈ I) be a sequence
of linearly independent vectors in H, where I = {1, . . . , m} for some m ∈ N
or I = N. Then there is an orthonormal sequence (wn | n ∈ I) such that,
for each k ∈ I, sp(u1 , . . . , uk ) = sp(t1 , . . . , tk ).
 LECTURE NOTES ON NONSTANDARD ANALYSIS 81

Proof. Set w1 := kuu11 k . Inductively, suppose that w1 , . . . , wk have been con-

structed. Set vk+1 := uk+1 − PE (uk+1 ), where E := sp(w1 , . . . , wk ), a closed
vk+1 ⊥
subspace of H by Lemma 11.26. Set wk+1 := kvk+1 k . Then wk+1 ∈ E ,
whence {w1 , . . . , wk+1 } is orthonormal. 
Corollary 11.34. Any finite-dimensional subspace of H is closed.
Proof. By the Gram-Schmidt process, any finite-dimensional subspace of H
has an orthonormal basis that is also a Hamel basis; thus, the subspace is
closed by Lemma 11.26. 
Exercise 11.35. Suppose that u1 , . . . , uk are orthonormal and E = sp(u1 , . . . , uk ).
Then, for all x ∈ H, we have PE (x) = ki=1 hx, ui iui .
P

Let us momentarily return to the discussion of the previous subsection.

Suppose that H is separable and that (en | n ≥ 1) is an P orthonormal basis
m
P∞Set Pm : H → H as before, namely Pm (a) = n=1 αn en , where
for H.
a = n=1 P αn en . Set Em := sp(e1 , . . . , em ), a closed subspace of H. Then,
for a = ∞ n=1 αn en , we have:
m
X
PEm (a) = ha, en ien
n=1
Xm Xk
= h lim ( αj ej ), en ien
k→∞
n=1 j=1
m
X k
X
= lim h αj ej , en ien
k→∞
n=1 j=1
m X
X k
= lim αj hej , en ien
k→∞
n=1 j=1
m
X
= αj ej = Pm (a).
j=1

Lemma 11.36. PE is a bounded linear operator. If E 6= {0}, then kPE k =

1.
Lemma 11.37. Suppose that F is also a closed subspace of H and E ⊥ F ,
that is, x ⊥ y for all x ∈ E and y ∈ F . Set
G := E + F := {x + y | x ∈ E, y ∈ F }.
Then G is a closed subspace of H and for all z ∈ G, there are unique x ∈ E
and y ∈ F such that z = x + y.
In the situation of the previous lemma, we write G = E ⊕ F , the direct
sum of E and F .
Lemma 11.38. H = E ⊕ E ⊥ .
82 ISAAC GOLDBRING

Lemma 11.39. (E ⊥ )⊥ = E.
Lemma 11.40. Suppose that G = E ⊕ F . Then PG = PE + PF . Conse-
quently, I = PE + PE ⊥ .
Now suppose that E1 , E2 , E3 are closed subspaces such that E1 ⊥ E2 ,
E2 ⊥ E3 and E1 ⊥ E3 . Then E1 ⊥ (E2 ⊕ E3 ), (E1 ⊕ E2 ) ⊥ E3 , and
E1 ⊕ (E2 ⊕ E3 ) = (E1 ⊕ E2 ) ⊕ E3 . We may thus unambiguously write
E1 ⊕ E2 ⊕ E3 . Ditto for any finite number of mutually perpendicular closed
subspacesLE1 , . . . , En of H; we often write the direct sum in the compact
notation ni=1 Ei .
Corollary 11.41. If G = ni=1 Ei , then PG = PE1 + · · · + PEn .
L

Now suppose that we have countably many mutually perpendicular sub-

spaces (En | n ≥ 1). How should we define their direct sum? First, two
lemmas, the second of which we give a short nonstandard proof.
Lemma 11.42. Suppose that (Gn | n ≥ 1)Sis a sequence of subspaces of H
with Gn ⊆ Gn+1 for all n ≥ 1. Then G = n≥1 Gn is a subspace of H.
Lemma 11.43. Suppose that G is a subspace of H. Then G is also a
subspace of H.
Proof. Suppose that x, y ∈ G; we show that x + y ∈ G. (The proof for
closure under scalar multiplication is similar.) Since x ∈ G, there is x0 ∈ G∗
such that x ≈ x0 . Similarly, there is y 0 ∈ G∗ such that y ≈ y 0 . Thus,
x + y ≈ x0 + y 0 ∈ G∗ (by transfer), whence x + y ∈ G. 

Ln | n ≥
Returning to the situation preceding the lemmas, suppose that (E
1) are mutually perpendicular closed subspaces of H. Set Gn := ni=1 Ei ,
L∞ S
a closed subspace of H. We then define n=1 Ei := n≥1 Gn , a closed
subspace of H by the previous two lemmas.
Lemma 11.44. Suppose that (Gn | n ≥ 1) is a sequence of subspaces of H
S
with Gn ⊆ Gn+1 for all n ≥ 1 and G = n≥1 Gn . Then, for all x ∈ H, we
have PGn (x) → PG (x) as n → ∞.
L∞
Corollary
Pn 11.45. Suppose that E = i=1 Ei . Then, for all x ∈ H, we
have i=1 PEi (x) → PE (x) as n → ∞.
11.4. Hyperfinite-dimensional subspaces. Once again, H denotes an
arbitrary Hilbert space. Let E denote the set of finite-dimensional subspaces
of H, whence E ⊆ P(H). We will refer to elements of E ∗ as hyperfinite-
dimensional subspaces of H ∗ . We have a map P : E × H → H given by
P (E, x) := PE (x). We thus get a nonstandard extension P : E ∗ × H ∗ → H ∗ ,
whence it makes sense to speak of the orthogonal projection map PE : H ∗ →
H ∗ for E ∈ E ∗ . Similary, we have a map dim : E → N, whence we get a map
dim : E ∗ → N∗ .
If H is separable and (en | n ≥ 1) is an orthonormal basis for H, then
by transfer, for N ∈ N∗ , we have HN := sp(e1 , . . . , eN ) ∈ E ∗ , the internal
 LECTURE NOTES ON NONSTANDARD ANALYSIS 83

subspace of H ∗ spanned by e1 , . . . , eN . The elements of HN are internal linear

combinations (with coefficients in F∗ ) of e1 , . . . , eN . By transfer, dim(HN ) =
N.
The following theorem is crucial in applications of nonstandard methods
to the study of infinite-dimensional linear algebra:
Theorem 11.46. If H is an infinite-dimensional Hilbert space, then there
is E ∈ E ∗ such that H ⊆ E.
Proof. Exercise. 
Recall from Problem 9.5 the notion of the standard part of an internal
set.
Exercise 11.47. Suppose that E ∈ E ∗ and x ∈ H. Then x ∈ st(E) if and
only if PE x ≈ x.
Lemma 11.48. If E ∈ E ∗ , then st(E) is a closed linear subspace of H.
Proof. st(E) is closed by Problem 9.5. It is straightforward to verify that
st(E) is a subspace of H. 
11.5. Problems.
Problem 11.1. Suppose (λn | n ≥ 1) is a sequence from C. Consider the
function D : CN → CN given by D((xn )) := (λn xn ).
(1) Show that D ∈ B(`2 ) if and only if the sequence (λn ) is bounded.
(We are sort of abusing notation and letting D denote its restriction
to `2 .)
(2) Show that D ∈ B◦ (`2 ) if and only if limn→∞ λn = 0.
Problem 11.2. Consider the function T : `2 → `2 given by
T (x1 , x2 , x3 , . . .) = (x2 , x3 , . . .).
Show that T ∈ B(`2 ) / B◦ (`2 ). What is kT ||? (T is called the left
and T ∈
shift operator.)
Problem 11.3. Suppose that T is a compact operator on the separable
Hilbert space H. Let  ∈ R>0 . Show that there is a bounded finite-rank
linear operator T 0 on `2 such that kT − T 0 k < . In fancier language, we are
showing that the set of finite-rank linear operators on `2 , B◦◦ (`2 ), is dense
in the set of compact linear operators on `2 , B◦ (`2 ). (Hint: The desired T 0
will be of the form Pk T for some k ∈ N. Consider k > N and use underflow.)

P x ∈ Hns .
Problem 11.4. Suppose that H is a separable Hilbert space and
Suppose that (en ) is an orthonormal basis for `2 and x = n∈N∗ xn en .
Further suppose that xn ≈ 0 for all n ∈ N. Show that x ≈ 0.
Problem 11.5. Suppose P : H → H is an idempotent bounded operator on
a Hilbert space H. Show that P is compact if and only if P is a finite-rank
operator.
84 ISAAC GOLDBRING

Problem 11.6. Suppose that E is any linear subspace of any unitary space
U . Show that E ⊥ is a linear subspace of U .
We will need the following notation for the next problem. If w = a + bi ∈
C, then Re(w) := a and Im(w) := b.
Problem 11.7. Suppose that E is a closed linear subspace of the Hilbert
space H. Suppose x ∈ / E. We aim to show that x − PE (x) ∈ E ⊥ . Set
α := kx − PE (x)k > 0. Fix z ∈ E \ {0}.
(1) Fix λ ∈ R \ {0}. Show that λ2 kzk2 − 2λ Re(hx − PE (x), zi) > 0.
(Hint: Start with α2 < kx − (PE (x) + λz)k2 .)
(2) Considering
Re(hx − PE (x), zi)
λ := ,
kzk2
conclude that Re(hx − PE (x), zi) = 0.
(3) Show that hx − PE (x), zi = 0, and thus x − PE (x) ∈ E ⊥ .
Problem 11.8. Suppose (e1 , . . . , ek ) is an orthonormal sequence of vectors
in the Hilbert space H. Suppose E = span(e1 , . . . , ek ). Show that, for all
v ∈ H, we have
X k
PE (x) = hv, ei iei .
i=1

Problem 11.9. Suppose that (en ) is an orthonormal sequence from the

Hilbert space H. Fix v ∈ H. Prove Bessel’s inequality:
∞
X
|hv, en i|2 ≤ kvk2 .
n=1
(Hint: Use the last problem and the fact that a projection Pn has norm 1.)
Problem 11.10. Suppose that (en ) is an orthonormal sequence from the
Hilbert space H. Show that the following are equivalent:
(1) (en ) is an orthonormal basis for `2 ;
(2) For every v ∈ `2 , if hv, en i = 0 for all n ≥ 1, then v = 0.
(3) For every v ∈ `2 , there is a sequence (αn ) such that
∞
X
v= αn e n .
n=1

Problem 11.11. Suppose that H is a finite-dimensional Hilbert space.

Show that an orthonormal basis for H is also a Hamel basis.
Problem 11.12. Show that the orthonormal basis (bn ) for `2 defined in
lecture is not a Hamel basis.
It can be shown that any orthonormal basis for an infinite dimensional
Hilbert space is not a Hamel basis. We outline the proof of this here for `2 .
We first need a definition. If X is a metric space and A ⊆ X, we say that
 LECTURE NOTES ON NONSTANDARD ANALYSIS 85

A is nowhere dense if (A)◦ = ∅; in English, if the interior of the closure of

A is empty.
Problem 11.13. Suppose that H is a Hilbert space and (en | n ≥ 1) is a
linearly independent set of vectors. For n ≥ 1, set
En := span(e1 , . . . , en ).
Show that each En is a nowhere dense subset of V . (Hint: We already know
that each En is closed. (Why?) It remains to show that each En has empty
interior. Use the nonstandard characterization of the interior to do this.)
The Baire Category Theorem states that in a complete metric space, the
union of a countable number of nowhere dense sets is nowhere dense.
Problem 11.14. Use the Baire category theorem and the previous problem
to show that a Hamel basis of an infinite dimensional Hilbert space must
be uncountable. Conclude that no orthonormal basis for `2 can be a Hamel
basis.
In working with Hilbert spaces, orthonormal bases are more useful than
Hamel bases. One can show that any two orthonormal bases for a Hilbert
space have the same cardinality, called the dimension of the Hilbert space.
One can then show that two Hilbert spaces are isomorphic (that is, there is a
bijective linear transformation between them preserving the inner product)
if and only if they have the same dimension. This is not true if we use
dimension to mean the cardinality of a Hamel basis (which is well-defined).
Indeed, for s ∈ R, let us : R → C be defined by us (t) := eist . Let V be
the vector space consisting of all finite linear combinations of the us . For
f, g ∈ H, one can show that
Z N
1
hf, gi := lim f (t)g(t)dt
N →∞ 2N −N

exists and defines an inner product on V . Let H denote the completion of

V . Then H is naturally a Hilbert space. One can show that `2 and H both
have Hamel dimension 2ℵ0 . However, (us | s ∈ R) is an orthornormal basis
for H, so H has dimension 2ℵ0 and cannot be isomorphic to `2 as a Hilbert
space.

12. Weekend Problem Set #2

Problem 12.1. Assume the nonstandard extension is ℵ1 -saturated. Let A
be an internal set and B a (not necessarily internal) subset of A. We say
0
S1 if there is a sequence (C0k | k ∈ N) of internal subsets of A such
that B is Σ
that B = k Ck . We say that B is Π1 ifTthere exists a sequence (Ck | k ∈ N)
of internal subsets of A such that B = k Ck .
(1) Show that Rfin is a Σ01 subset of R∗ and µ is a Π01 subset of R∗ .
(This is why Σ01 sets are sometimes called galaxies and Π01 sets are
sometimes called monads.)
86 ISAAC GOLDBRING

(2) Suppose (Ck | k S

∈ N) is a sequence of internal subsets of an internal
set A. Let B = k Ck . Suppose that B is internal. Show that there
is n ∈ N such that B = C1 ∪ · · · ∪ Cn . (Hint: Reduce to the case
that Cn ⊆ Cn+1 for all n ∈ N. Use the Countable Comprehension
Principle and Underflow.)
(3) Suppose (Ck | k T
∈ N) is a sequence of internal subsets of an internal
set A. Let B = k Ck . Suppose that B is internal. Show that there
is n ∈ N such that B = C1 ∩ · · · ∩ Cn .
(4) Suppose that B1 is a Σ01 subset of the internal set A and B2 is a Π01
subset of A. Suppose that B1 ⊆ B2 . Show that there is an internal
subset E of A such that B1 ⊆ E ⊆ B2 . Conclude that any subset of
A which is both Σ01 and Π01 must be internal. (This is a handy way
of showing that certain sets are internal.)
We should remark that (2) from the previous exercise is the key observa-
tion behind the Loeb measure concept, which will be discussed later in these
notes. This concept is one of the central ideas in applications of nonstan-
dard analysis to probability theory, stochastic analysis, and mathematical
finance (to name a few).

There is an ultrapower construction for our more general nonstandard

framework. If U is a nonprincipal ultrafilter on N, then one can set X ∗ :=
X U , where X U is defined just as RU . This construction has the usual ad-
vantage of being “concrete.” In particular, we can identify the internal sets
in a concrete fashion as follows.
Suppose that weQhave a familyQ(Xn | n ∈ N) of sets. Then we can define
their ultraproduct U Xn to be n Xn / ∼U . (Thus, the ultrapower is the
special case of the ultraproduct where all Xn ’s are equal.)

Problem 12.2. Prove that A ⊆ X U isQinternal if and only if there are

subsets An ⊆ X for n ∈ N such that A = U An . (So the internal subsets of
X U are the ones that are almost everywhere given coordinatewise by subsets
of X.)

Problem 12.3. Show that A ⊆ X U is Qhyperfinite if and only if there are

finite subsets An ⊆ X such that A = U An . Use this fact to explain how
to define the internal cardinality |A| of a hyperfinite subset A of X U .
Problem 12.4. Prove that a nonstandard extension constructed using a
nonprincipal ultrafilter U on N is automatically ℵ1 -saturated. (Hint: this is
essentially a diagonal argument.)
Problem 12.5. (Assume as much saturation as you like) Suppose that G
is a group. Then G is said to be nilpotent if there exists a finite sequence of
subgroups G1 , . . . , Gn of G such that
{1} ≤ G1 ≤ G2 ≤ . . . ≤ Gn = G
 LECTURE NOTES ON NONSTANDARD ANALYSIS 87

and such that for all i ∈ {1, . . . , n}, all g ∈ G, and all h ∈ Gi , we have
ghg −1 h−1 ∈ Gi−1 . If n is the smallest length of such a sequence of subgroups
for G, we call n the nilpotency class of G.
(1) G is nilpotent of nilpotency class 1 if and only if G is nontrivial (i.e.
G 6= {1}) and abelian (i.e. xy = yx for all x, y ∈ G.)
(2) Show that G∗ is a group, as is any subset of G∗ closed under the
extension of the group operations.
(3) Suppose A ⊆ G. The subgroup of G generated by A, denoted hAi, is
\
hAi := {H | H ≤ G and A ⊆ H}.
A subgroup H of G is called finitely generated if H = hAi for some
finite A ⊆ G. Discuss how the nonstandard extension of the set of
finitely generated subgroups of G is the set of hyperfinitely generated
subgroups of G∗ .
(4) Show that there is a hyperfinitely generated subgroup H of G∗ such
that G ≤ H.
(5) Suppose that G is locally nilpotent, that is, suppose that every finitely
generated subgroup of G is nilpotent. Further suppose that the
nilpotency class of any finitely generated subgroup of G is less than
or equal to n. Show that G is nilpotent of nilpotency class less than
or equal to n. (You will have to use the fact that a subgroup of a
nilpotent group of nilpotency class ≤ n is itself a nilpotent group of
nilpotency class ≤ n. Try proving this fact!)
(6) Why doesn’t your proof in (5) work if G is locally nilpotent with un-
bounded nilpotency class? (There are some difficult open problems
about locally nilpotent groups of unbounded nilpotency class that
some group theorists hope might be solved by nonstandard meth-
ods.)
Problem 12.6. (Assume ℵ1 -saturation) Suppose that X is a normed space
over F and E ⊆ X ∗ is an internal subspace, that is, E is internal, 0 ∈ E, E
is closed under addition, and E is closed under multiplication by elements
of F∗ . For example, E := X ∗ is an internal subspace of X ∗ . Define
Efin := {x ∈ E | kxk ∈ Ffin }
and
µE := {x ∈ E | kxk ∈ µ(0)}.
(1) Show that Efin is a vector space over F and that µE is a subspace
of Efin . We set Ê := Efin /µE and call it the nonstandard hull of E.
For x ∈ Efin , we often write x̂ instead of x + µE .
(2) For x̂ ∈ Ê, define kx̂k := st(kxk). Show that this definition is
independent of the coset representative and that k · k : Ê → R is a
norm on Ê.
(3) Show that Ê is a Banach space. (Notice as a consequence that even
if X was incomplete, X c∗ is automatically complete.)
88 ISAAC GOLDBRING

(4) Show that there is a map ι : Ê → X c∗ given by ι(x + µE ) = x + µX ∗ .

Show that ι is a norm-preserving linear map whose image is closed
in Xˆ∗ . (This exercise allows us to treat Ê as a Banach subspace of
Xc∗ .)

The nonstandard hull construction is extremely useful in applications of

nonstandard methods to functional analysis. Of particular interest is when
E is a hyperfinite-dimensional subspace of X ∗ containing X, for then one
can often apply facts about finite-dimensional normed spaces to Ê and then
have these facts “trickle down” to X.

13. The Spectral Theorem for compact hermitian operators

In this section, H denotes an arbitrary Hilbert space over C.
Definition 13.1. If T : H → H is linear, then λ ∈ C is called an eigenvalue
for T if there is a nonzero x ∈ H such that T x = λx; such an x will be
called an eigenvector for T corresponding to λ.
If λ is an eigenvalue for T , then {x ∈ H | T x = λx} is called the eigenspace
of T corresponding to λ; as the name indicates, it is a subspace of H. It is
also easy to verify that eigenspaces are closed in H.
Definition 13.2. T : H → H is Hermitian if hT x, yi = hx, T yi for all
x, y ∈ H.
Exercise 13.3. Suppose that E is a closed subspace of H. Prove that PE
is a Hermitian operator.
The following theorem is standard fare for a first undergraduate course
in linear algebra:
Theorem 13.4 (Spectral Theorem for Hermitian Operators on Finite-Di-
mensional Hilbert Spaces). Suppose T : H → H is a Hermitian operator
and H is finite-dimensional. Then there is an orthonormal basis for H con-
sisting of eigenvectors of T . Moreover, if λ1 , . . . , λk ∈ C are the distinct
eigenvalues of T with corresponding eigenspaces W1 , . . . , Wk , then Wi ⊥ Wj
for all i 6= j, H = W1 ⊕· · ·⊕Wk , and T = λ1 PW1 +· · ·+λk PWk , the so-called
spectral resolution of T .
Proof. (Sketch) Set n := dim(H).
Step 1: Show that T has an eigenvalue λ1 with corresponding eigenvector
x1 ; we may further assume kx1 k = 1. (This part does not use the
fact that T is Hermitian, but only that we are working over C so the
characteristic polynomial of T splits.)
Step 2: Let W := sp(x1 ). If H = W , we are done. Otherwise, W ⊥ is a
Hilbert space of dimension n−1. Furthermore, since T is Hermitian,
one can easily verify that T (W ⊥ ) ⊆ W ⊥ , whence we can consider
the Hermitian operator T |W ⊥ : W ⊥ → W ⊥ . By induction, W ⊥ has
 LECTURE NOTES ON NONSTANDARD ANALYSIS 89

an orthonormal basis x2 , . . . , xn consiting of eigenvectors of T . Since

W ⊥ W ⊥ , {x1 , . . . , xn } is an orthonormal basis for H consisting of
eigenvectors of T .
Step 3: Show that each λi is a real number. We leave this as an exercise for
the reader.
Step 4: Show that Wi ⊥ Wj for i 6= j.
Step 5: Since we have a basis of eigenvectors, H = W1 ⊕ · · · ⊕ Wk .
Step 6: Just compute
T = T I = T (PW1 +· · · +PWk ) = T PW1 +· · · +T PWk = λ1 PW1 +· · · +λk PWk .

What about Hermitian operators on infinite-dimensional Hilbert spaces?
We immediately run into problems:
Exercise 13.5. Let H = C([0, 1], C) with the inner product
Z 1
hf, gi := f (t)g(t)dt.
0
Define T : H → H by T (f ) = t · f . Show that T is a bounded operator on
H that is not compact and does not have any eigenvalues.
In order to obtain eigenvalues, we will need to study compact Hermitian
operators. For such operators, there is a suitable version of the Spectral The-
orem that is a rather straightforward generalization of the finite-dimensional
case. In fact, we will use the method of hyperfinite approximation and the
transfer of the finite-dimensional spectral theorem to obtain the spectral the-
orem for compact Hermitian operators. We should mention that there are a
plethora of spectral theorems for other sorts of operators on Hilbert spaces,
but the functional analysis needed to study them (e.g. spectral measures)
is beyond the scope of these notes.
For the rest of this section, we assume that H is an infinite-dimensional
Hilbert space. Moreover, we fix a hyperfinite-dimensional subspace E of H ∗
(that is, E ∈ E ∗ ) such that H ⊆ E; see Theorem 11.46. We let P denote
PE .
Lemma 13.6. If x ∈ Hns , then P x ≈ x.
Proof. Write x = y + , where y ∈ H and  ≈ 0. Then P x = P y + P ; since
kP k = 1, P  ≈ 0. Since y ∈ H ⊆ E, P y = y. Thus, P x ≈ P y = y ≈ x. 
If T ∈ B(H), we define T 0 : E → E by T 0 (x) = P T (x). It is straightfor-
ward to check that kT 0 k ≤ kT k and T 0 x = T x for x ∈ H. A slightly less
trivial observation is:
Lemma 13.7. If T is Hermitian, so is T 0 .
Proof. For x, y ∈ E, we calculate:
hT 0 x, yi = hT x, P yi = hT x, yi = hx, T yi = hP x, T yi = hx, P T yi = hx, T 0 yi.

90 ISAAC GOLDBRING

Lemma 13.8. If x ∈ Hns , then T 0 P x ≈ T x.

Proof. By Lemma 13.6, P x ≈ x. Since T is continuous, we have T P x ≈ T x.
Since T x ∈ Hns (again, by continuity of T ), we have P T x ≈ T x. Thus,
T 0 P x = P T P x ≈ P T x ≈ T x. 
For the rest of this section, we assume that T ∈ B(H) is a nonzero,
compact Hermitian operator and T 0 : E → E is defined as above. In this
case, T 0 is internally compact in the following precise sense:
Lemma 13.9. If x ∈ E ∩ Hfin , then T 0 x ∈ Hns .
Proof. Since T is compact, T x ∈ Hns . By Lemma 13.6, T 0 x = P T x ≈ T x,
whence T 0 x ∈ Hns as well. 
Exercise 13.10. Suppose that λ ∈ C∗ is an eigenvalue of T 0 . Show that
|λ| ≤ kT 0 k, whence λ ∈ Cfin .
Lemma 13.11. Suppose that λ ∈ C∗ is an eigenvalue of T 0 with correspond-
ing eigenvector x ∈ E of norm 1. Further suppose that λ 6≈ 0. Then x ∈ Hns
and st(x) is an eigenvector of T of norm 1 corresponding to the eigenvalue
st(λ).
Proof. By the internal compactness of T 0 , we know that T 0 x ∈ Hns . But
T 0 x = λx, so since λ1 ∈ Cfin , we have x = λ1 T 0 x ∈ Hns . Also, by Lemma
13.8, we have
T (st(x)) ≈ T x ≈ T 0 P x = T 0 x = λx ≈ st(λ) st(x).
Consequently, T (st(x)) = st(λ) st(x). It remains to observe that st(x) has
norm 1; however, this follows from the fact that x has norm 1 and k st(x)k
is a real number. 
At this point, we apply the (transfer of the) finite-dimensional Spectral
Theorem to T 0 ; in particular, we get a hyperfinite sequence of eigenvalues
λ1 , . . . , λν (perhaps with repetitions), where ν := dim(E). Without loss of
generality (by reordering if necessary), we may assume that
|λ1 | ≥ |λ2 | ≥ · · · ≥ |λν |.
We let {x1 , . . . , xν } be an orthonormal basis for E, where xi is an eigen-
value for T 0 corresponding to λi .
We now worry about how many repetitions we might have amongst the
eigenvalues. We show that a noninfinitesimal eigenvalues can only be re-
peated a finite number of times. First, we need:
Lemma 13.12. Suppose  ∈ R>0 and |λk | ≥ . Then k ∈ N.
√
Proof. By the Pythagorean Theorem, kxi − xj k = 2 for i, j = 1, . . . , k.
We define the internal sequence (si : i ∈ N∗ ) by si = xi for i = 1, . . . , k,
while si = 0√for i > k. Suppose, towards a contradiction, that k > N. Then
ksi − sj k = 2 for distinct i, j ∈ N, whence, by Theorem 9.36, si = xi ∈
/ Hns
for some i ∈ N∗ , contradicting Lemma 13.11. 
 LECTURE NOTES ON NONSTANDARD ANALYSIS 91

Corollary 13.13. Suppose that λl+1 = λl+2 = · · · = λl+k 6≈ 0. Then k ∈ N.

Proof. By the previous lemma, l + k ∈ N, whence k ∈ N. 
We now remove repetitions: let κ1 , . . . , κη enumerate all the distinct eigen-
values of T 0 , again ordered so that |κ1 | ≥ |κ2 | ≥ · · · ≥ |κη |. For i = 1, . . . , η,
we let {xi1 , . . . , ximi } be an orthonormal basis for the eigenspace Wi cor-
responding to κi . By the previous corollary, we know that mi ∈ N for
each i ∈ {1, . . . , η} with κi 6≈ 0. Moreover, we have E = W1 ⊕ · · · ⊕ Wη ,
P = PW1 + · · · + PWη , and T 0 = κ1 PW1 + · · · + κη PWη .
Which of the κi ’s are noninfinitesimal? To answer this question, we set
X := {i ∈ N∗ | 1 ≤ i ≤ η and κi 6≈ 0}.
Lemma 13.14. X 6= ∅.
Proof. Suppose, toward a contradiction, that X = ∅. Fix b ∈ H such that
T b 6= 0, this is possible since T 6= 0. Write b = b1 + · · · + bη , with bi ∈ Wi
for each i. Then T b = T 0 b = T 0 b1 + · · · + T 0 bη = κ1 b1 + · · · + κη bη . However,
by the Pythagorean Theorem,
η
X η
X
2 2 2 2
kκ1 b1 + · · · + κη bη k = |κi | kbi k ≤ |κ1 | kbi k2 = |κ1 |2 kbk2 ≈ 0,
i=1 i=1
a contradiction. 
Lemma 13.15. X ⊆ N.
Proof. Fix n ∈ X and choose  > 0 such that |κn | ≥ . Recall that κn = λm
for some m ∈ {1, . . . , ν}. By Lemma 13.12, m ∈ N. Since n ≤ m, we have
n ∈ N. 
We now know that one of the two situations occurs: either X = N>0 or
X = {1, . . . , k} for some k ∈ N. Both situations are in fact possible and
must be treated slightly differently.
We set yij := st(xji ) for j ∈ X and i ∈ {1, . . . , mj }. If i 6= i0 , then
by continuity of the inner product, we know that hyij , yij0 i ≈ hxji , xji0 i = 0,
whence the yij ’s are pairwise perpendicular. We set W̃j := sp(y1j , . . . , ymj
j ),

so dim(W̃j ) = mj . In an analogous fashion, one shows that W̃j ⊥ W˜j 0 for

distinct j, j 0 ∈ X.
Lemma 13.16. For x ∈ H and j ∈ X, we have PW̃j (x) ≈ PWj (x).
Proof.
mj mj
hx, yij iyij hx, xji ixji = PWj (x).
X X
PW̃j (x) = ≈
i=1 i=1

L
We now set K := j∈X W̃j ; this is either a finite or infinite direct sum,
depending on whether X is finite or infinite. We set W̃0 := K ⊥ .
92 ISAAC GOLDBRING

For simplicity, set αj := st(κj ). We can now state the Spectral Theorem
in case W is finite.
Theorem 13.17 (Spectral Theorem- Case 1). If X = {1, . . . , k} for some
k ∈ N, then the eigenvalues of T are 0, α1 , . . . , αk and
T = α1 PW̃1 + · · · + αk TW̃k
is the spectral resolution of T .
Proof. Fix x ∈ H; then T (x) = T 0 (x). Consequently,
η
X k
X η
X
0
T (x) = T (x) κj PWj = κj PWj (x) + κj PWj (x) = (†) + (††).
i=1 j=1 j=k+1

By the previous lemma, (†) ≈ kj=1 αj PW̃j (x) ∈ H. We need to show that
P

(††) ≈ 0. This again follows from a Pythagorean Theorem calculation, as

k(††)k2 ≤ |κk+1 |2 kxk2 ≈ 0. This proves the Spectral Resolution part of the
Theorem. We next show that 0 is an eigenvalue. First, since K is finite-
dimensional while H is infinite-dimensional, we have W̃0 6= {0}. If x ∈ W̃0 \
{0}, then by the Spectral Resolution, we have T (x) = kj=1 αj PWj (x) = 0,
P
whence x is an eigenvector of T corresponding to 0. It remains to show that
{0, α1 , . . . , αk } are all of the eigenvalues of T ; this will again follow from the
Spectral Resolution of T . To see this, suppose that α 6= 0 is an eigenvalue
of T with corresponding eigenvector x. Then
k
X k
X
αPW̃j (x) = αx = T x = αi PW̃i (x).
j=0 i=1

Thus, αPW̃0 (x) = 0 and αPW̃j (x) = αj PW̃j (x) for j = 1, . . . , k (since ele-
ments of W̃0 , W̃1 , . . . , W̃k are pairwise perpendicular and thus linearly inde-
pendent). Since x 6= 0, there is j ∈ {1, . . . , k} such that PW̃j (x) 6= 0; for this
j, we have α = αj . 
In order to deal with the case that X = N>0 , we need one final lemma.
Lemma 13.18. If X = N>0 , then limj→∞ αj = 0.
Proof. Since |αn | is nonincreasing and bounded below, it suffices to prove
that 0 is a limit point of |αn |. We consider the extension of the sequence
(αn | n ∈ N>0 ) to a sequence (αn | n ∈ (N>0 )∗ ). By the Infinitesimal
Prolongation Theorem, there is N ∈ N∗ \ N, N ≤ η such that αN ≈ κN .
But κN ≈ 0 since N ∈ / X. Thus, αN ≈ 0 and hence 0 is a limit point of
(αn ). 
We are now ready to state:
Theorem 13.19 (Spectral Theorem-Case >0
P∞ 2). If X = N , then the nonzero
eigenvalues are α1 , α2 , . . . and T = j=1 αj PW̃j is the spectral resolution of
T , that is, T (x) = ∞
P
i=1 αj PW̃j (x) for all x ∈ H.
 LECTURE NOTES ON NONSTANDARD ANALYSIS 93

Proof. Fix x ∈ H. Set rn := kT (x)− nj=1 αj PW̃j (x)k; we need to show that
P

rn → 0 as n → ∞. Again, we know that nj=1 αj PW̃j (x) ≈ nj=1 κj PWj (x)

P P

and T (x) = T 0 (x) = ηj=1 κj PWj (x). Thus, rn ≈ k ηj=n+1 κj PWj (x)k for
P P
all n ∈ N. We play a Pythagorean game again:
η
X η
X η
X
k κj PWj (x)k2 = |κj |2 kPWj (x)k2 ≤ |κn+1 |2 kPWj (x)k2 = |κn+1 |2 kxk2 .
j=n+1 j=n+1 j=1

Thus rn ≤ |αn+1 |2 kxk2 for all n ∈ N. Since αn → 0 as n → ∞, we see that

rn → 0 as n → ∞. The claim about the nonzero eigenvalues of T being
{α1 , α2 , . . . , } follows from the Spectral Resolution of T in a manner similar
to Case 1 and is left as an exercise. 

13.1. Problems.
Problem 13.1. Suppose that T is a Hermitian operator on a unitary space
V . Show that every eigenvalue of T is a real number.
Problem 13.2. Let V be the unitary space C([0, 1], C) endowed with the
inner product
Z 1
hf, gi := f (t)g(t)dt.
0
Define T : V → V by T (f ) := tf , i.e. T (f )(t) := tf (t).
(1) Show that T is a bounded Hermitian operator on V .
(2) Show that T has no eigenvalues.
(3) Show that T is not compact. (Don’t just say if it were compact,
it would have eigenvalues by the Spectral Theorem. Show directly
that T is not compact.)
Problem 13.3. For each of the following bounded linear operators T given
below, explain why they are compact and Hermitian. Then find the eigen-
values, eigenspaces, and projections, yielding the spectral decomposition for
T.
(1) Let A be the matrix
 
3 −4
.
−4 3
Define T : C2 → C2 by T (x) = Ax.
(2) Define T : `2 → `2 by T ((xk | k ≥ 1)) = ( xkk | k ≥ 1).
There is a Spectral Theorem for Hermitian operators which need not be
compact. The statement of the spectral theorem is more involved (and a
little less satisfying), so we will not give it here. However, the results in
the remaining problems are ingredients towards proving this more general
Spectral Theorem. The interested reader can find a complete discussion of
this in Chapter 5, Section 5 of [1].
94 ISAAC GOLDBRING

We let H be a Hilbert space and T a (not necessarily compact) Hermitian

operator on H. E, P , T 0 and the sequence (λi | i ≤ ν) are defined exactly
as they are in the notes. We also set J := [−kT k, kT k].
Pν
Problem 13.4. For n ≥ 0, show that (T 0 )n = P n
i=1 λi Pi . Conclude that
for any polynomial p(z) ∈ R∗ [z], we have p(T 0 ) = i=1ν p(λi )Pi .
For any internal function f : J ∗ → R∗ , set f (T 0 ) = νi=1 f (λi )Pi .
P

Problem 13.5. For any internal f : J ∗ → R∗ , show that f (T 0 ) is Hermitian.

Problem 13.6. Suppose f, g : J ∗ → R∗ are internal. Show that:
(1) (f + g)(T 0 ) = f (T 0 ) + g(T 0 );
(2) (f · g)(T 0 ) = f (T 0 ) · g(T 0 );
(3) (cf )(T ) = cf (T ) for any c ∈ R∗ .
These properties are often referred to as the Operational Calculus.
Problem 13.7. Suppose that f, g : J ∗ → R∗ are internal functions and
f (c) ≤ g(c) for all c ∈ J ∗ . Show that f (T 0 ) ≤ g(T 0 ).
We call an internally bounded linear operator (what does this mean?) U
on E nearstandard if for every x ∈ E ∩ Hns , we have U (x) ∈ Hns .
Problem 13.8. Suppose that U is nearstandard. Show that kU k ∈ Rfin .
(Careful: This is not immediate from the definition)
If U is nearstandard, define st(U ) : H → H by st(U )(x) = st(U (x)) for
all x ∈ H.
Problem 13.9. Suppose that U is nearstandard. Show that st(U ) is a
bounded linear operator on H and k st(U )k ≤ st(kU k).
Problem 13.10. Show that, for any n ∈ N, (T 0 )n is nearstandard. Conclude
that for any p(z) ∈ R[z], we have p(T 0 ) is nearstandard.
Problem 13.11. Suppose that f : J → R is continuous.
(1) Show that f (T 0 ) is nearstandard. (Hint: Suppose, towards a contra-
diction, that x ∈ E ∩ Hns is such that f (T 0 )(x) ∈
/ Hns . Thus, there
>0 0
is r ∈ R such that kf (T )(x) − bk > r for all b ∈ H. By the Weier-
strauss Approximation Theorem, there is a polynomial p(z) ∈ R[z]
such that |f (c) − p(c)| < r for all c ∈ J.)
(2) Show that st((f (T 0 ))) is Hermitian.
(3) Suppose that f (c) ≥ 0 for all c ∈ J. Show that st(f (T 0 )) ≥ 0, that
is, hst(f (T 0 ))x, xi ≥ 0 for all x ∈ H.

14. The Bernstein-Robinson Theorem

In this section, we will prove a theorem which was the first serious theorem
proven using nonstandard methods. Before we can state this theorem, we
need some background. Throughout this section, all normed spaces are over
C.
 LECTURE NOTES ON NONSTANDARD ANALYSIS 95

Definition 14.1. Suppose that X is a Banach space, T ∈ B(X), and Y ⊆ X

is a subspace. We say that Y is T -invariant if T (Y ) ⊆ Y .
If X is a Banach space and T ∈ B(X), then clearly {0} and X are T -
invariant; these are said to be the trivial T -invariant subspaces. Clearly, if
dim(X) = 1, then X has no nontrivial T -invariant subspaces.
Theorem 14.2. If X is a finite-dimensional Banach space with dim(X) ≥ 2
and T ∈ B(X), then X has a nontrivial T -invariant subspace.
Proof. Set n := dim(X) ≥ 2. Fix x ∈ X \ {0}. Since {x, T x, . . . , T n x} has
n + 1 vectors, it must be linearly dependent, whence we can find scalars
α0 , . . . , αn ∈ C, not all 0, such that α0 x + α1 T x + . . . + αn T n x = 0. Factor
the polynomial α0 +α1 z+. . .+αn z n = α(z−λ1 ) · · · (z−λn ), where α, λi ∈ C.
Then α(T − λ1 I) · · · (T − λn I)x = 0. It follows that there is i ∈ {1, . . . , n}
such that ker(T − λi I) 6= {0}. Observe that ker(T − λi I) = {x ∈ X | T (x) =
λi x} is T -invariant. If ker(T − λi I) 6= X, then it is a nontrivial T -invariant
subspace. Otherwise, T x = λi x for all x ∈ X, whence it follows that
T = λi I. But then, every subspace is T -invariant, so any 1-dimensional
subspace will be a nontrivial T -invariant subspace. 
Thus, for “small” Banach spaces, non-trivial T -invariant subspaces always
exist. At the opposite extreme, if X is a non-separable Banach space, then
for any x ∈ X \ {0}, the closed linear span of {x, T x, T 2 x, . . .} is a nontrival
T -invariant subspace. What about separable, infinite-dimensional Banach
spaces? Well, it was first shown by Enflo that there are quite exotic separable
Banach spaces that possess bounded linear operators without any nontrivial,
closed invariant subspaces. But what about the least exotic Banach spaces:
Hilbert spaces? Well, the following is still open:

The Invariant Subspace Problem: Suppose that H is a separable,

infinite-dimensional Hilbert space and T ∈ B(H). Does H have a nontrivial,
closed T -invariant subspace?

In the rest of this section, H will denote a separable, infinite-dimensional

Hilbert space.
Theorem 14.3. If T ∈ B(H) is a compact, Hermitian operator, then H
has a nontrivial, closed T -invariant subspace.
Proof. By the Spectral Theorem, there is an eigenvalue λ for T . The as-
sociated eigenspace is a closed, T -invariant subspace that is not {0}. If
the associated eigenspace were all of H, then T = λI and once again any
1-dimensional subspace is a nontrivial, closed T -invariant subspace. 
Notice that the proofs of the preceding theorems utilized eigenvalues in an
essential way. Even so, the next example and exercises show that eigenvalues
are not necessary for invariant subspaces.
96 ISAAC GOLDBRING

Example 14.4. Let S : `2 → `2 be the “right-shift” operator, that is,

S(x1 , x2 , . . .) = (0, x1 , x2 , . . .). Notice kSk = 1, so S ∈ B(`2 ). For n > 0,
set Mn := {(x1 , x2 , . . .) ∈ `2 | xk = 0 for k = 1, 2, . . . , n}. Then Mn is a
nontrivial, closed S-invariant subspace.
Exercise 14.5. For S as in the previous example, show that S is not com-
pact and has no eigenvalues.
Exercise 14.6. Let W : `2 → `2 be given by
1 1
W (x1 , x2 , . . .) = (0, x1 , x2 , x3 , . . .).
2 3
Show that W is compact but does not have any eigenvalues. Once again,
the various Mn are nontrivial W -invariant subspaces.
Before nonstandard methods, the state-of-the-art result was the following
strengthening of Theorem 14.3:
Theorem 14.7 (von Neumann, 1930s). If T ∈ B(H) is compact, then H
has a nontrivial, closed T -invariant subspace.
Definition 14.8. T ∈ B(H) is said to be polynomially compact if there is
a nonzero polynomial p(z) ∈ C[z] such that p(T ) is compact.
Our goal is to prove the following:
Theorem 14.9 (Bernstein-Robinson 1966). If T ∈ B(H) is polynomially
compact, then H has a nontrivial, closed T -invariant subspace.
To be fair, we should say that after the Bernstein-Robinson theorem,
a more general theorem was proven whose proof is much more elementary
(modulo some standard facts from a first graduate course in functional anal-
ysis):
Theorem 14.10 (Lomonosov, 1973). If T ∈ B(H) is such that there is a
nonzero K ∈ B(H) such that K is compact and T K = KT , then H has a
nontrivial, closed T -invariant subspace.
(To see how Lomonosov implies Bernstein-Robinson, observe that T com-
mutes with p(T ).)
The main idea behind the proof of the Bernstein-Robinson Theorem is
fairly easy to explain. Once again, it is an instance of the method of Hy-
perfinite Approximation. Fix an orthonormal basis {e1 , e2 , . . .} for H and
fix ν ∈ N∗ \ N. Set Hν := sp(e1 , . . . , eν ). Let Pν : H ∗ → Hν be the orthog-
onal projection onto Hν . Set Tν : H ∗ → H ∗ by Tν = Pν T Pν ; observe that
Tν (H ∗ ) ⊆ Hν . We also set T̃ := Tν |Hν = (Pν T )|Hν , so T̃ is an internally
linear operator on Hν . Observe also that
kT̃ k ≤ kTν k = kPν T Pν k ≤ kPν kkT kkPν k = kT k.
We recall the following standard fact from linear algebra (whose proof we
include for the sake of completeness).
 LECTURE NOTES ON NONSTANDARD ANALYSIS 97

Theorem 14.11 (Existence of Upper Triangular Representations). Suppose

that V is a finite-dimensional vector space of dimension m and T : V → V
is linear. Then there is a chain {0} = V0 ⊆ V1 ⊂ · · · ⊂ Vm = V such that
dim(Vi ) = i and such that each Vi is T -invariant.
Proof. We prove the theorem by induction on m, the case m = 0 being
trivial. Assume now that m ≥ 1 and the theorem is true for vector spaces
of dimension m − 1. Suppose dim(V ) = m and T : V → V is linear. Since
we are working over C, we can find an eigenvector u of T . Consider a
basis {u, w1 , . . . , wm−1 } of V . Set U = sp(u), W = sp(w1 , . . . , wm−1 ) and
P : V → W the projection onto W . Define S : W → W by S(w) = P (T (w)).
By the induction hypothesis, we have a basis {z1 , . . . , zm−1 } for W such
that, for each i ∈ {1, . . . , m − 1}, S(sp(z1 , . . . , zi )) ⊆ sp(z1 , . . . , zi ). We
leave it to the reader to check that T (sp(u, z1 , . . . , zi )) ⊆ sp(u, z1 , . . . , zi ) for
each i ∈ {1, . . . , m − 1}. Set V1 := U and, for i = 1, . . . , m − 1, Vi+1 :=
sp(u, z1 , . . . , zi ). 

We now apply the (transfer of the) previous theorem to Hν and T̃ , yielding

a chain
{0} = E0 ⊂ E1 ⊂ · · · ⊂ Eν = Hν ,
where each Ei ∈ E ∗ , dim(Ei ) = i, and each Ei is T̃ -invariant. For ap-
propriately chosen {e1 , e2 , . . .} and ν ∈ N∗ \ N, we will have that each
st(Ei ) is T -invariant. Since each st(Ei ) is closed, it will remain to find
some i ∈ {1, . . . , ν} such that st(Ei ) is nontrivial.
Let’s get started: fix v ∈ H, kvk = 1. Let A = {v, T v, T 2 v, . . .}.
Let’s eliminate some trivial cases early on. First, suppose that A is lin-
early dependent. Then there are α0 , α1 , . . . , αk−1 ∈ C, not all 0, such that
T k v = α0 v + α1 T v + · · · + αk−1 T k−1 v. Set Y := sp(v, T v, . . . , T k−1 v).
Then Y is finite-dimensional (hence closed), nontrivial since v ∈ Y and H
is infinite-dimensional, and T -invariant.
We may thus assume that A is linearly independent. Let F be the closed
linear span of A. Suppose that F 6= H. We claim that F is T -invariant,
whence F is a nontrivial T -invariant subspace of H. First, suppose that
w ∈ sp(A), so w = α0 v + α1 T v + · · · + αk T k v for some α0 , . . . , αk ∈ C. But
then T w = α0 T v + α1 T 2 v + · · · + αk T k+1 v ∈ sp(A). For the more general
situation, w = limn→∞ wn with each wn ∈ sp(A). But then, since T is
continuous, T w = limn→∞ T wn ∈ F since T wn ∈ sp(A) for each n.
We may thus suppose F = H. In this case, we apply the Gram-Schmidt
process to A to get an orthonormal basis {e1 , e2 , . . .} of H with the property
that sp(x, T x, . . . , T n−1 x) = sp(e1 , e2 , . . . , en ) for each n ≥ 1.
Let’s observe something interesting about the “matrix representation”
of T with respect to the basis {e1 , e2 , . . .}. For k ∈ N>0 , we have ek ∈
sp(x, T x, . . . , T k−1 x), so T (e
Pk∞) ∈ sp(T x, T 2 x, . . . , T k x) ⊆ sp(e1 , . . . , ek+1 ).
Thus, if we write T (ek ) = j=1 ajk ek , we see that ajk = 0 for j > k + 1.
We then say that the matrix [ajk ] for T is almost superdiagonal.
98 ISAAC GOLDBRING

It will prove desirable later on to know that T and T̃ are “close.” More
precisely, we will want to know that, for y ∈ Hν ∩ Hfin , T (y) ≈ T̃ (y). We
can achieve this goal by choosing ν appropriately, as we proceed to show
now.
First, we would like to know a predictable form for the matrix repre-
sentation of T n with respect to our basis {e1 , e2 , . . .}. Usually, this is not
possible, but thankfully, [ajk ] is almost superdiagonal. Recursively define
(n) (1) (n+1) (n) Pk+1 (n)
:= ∞
P
ajk as follows: Set ajk := ajk and ajk i=1 aji aik = i=1 aji aik .

Exercise 14.12. For n, j, k > 0, we have:

(n)
(1) aj+k,j = 0 if n < k, and
(k)
(2) aj+k,j = k−1
Q
i=0 aj+i+1,j+i .

For the rest of this section, fix a polynomial p(z) = α0 + α1 z + · · · + αk z k

with coefficients from C such that αk 6= 0 and p(T ) is compact. We can now
choose the “appropriate” ν ∈ N∗ \ N.
Lemma 14.13. There is ν ∈ N∗ \ N such that aν+1,ν ≈ 0.
Proof. Let [bjk ] be the matrix for p(T ). Then by the previous two exercises,
(1) (2) (k)
bj+k,j = 0 + α1 aj+k,j + α2 aj+k,j + · · · + αk aj+k,j
(k)
= αk aj+k,j
k−1
Y
= αk aj+i+1,j+i .
i=0
By Lemma 11.25, bN +k,N ≈ 0 for all N > N. Thus, there is i ∈ {0, . . . , k −1}
such that aN +i+1,N +i ≈ 0. Set ν = N + i for this i. 
For the remainder of this section, we fix ν as in the previous lemma. Let
us show how this ν has the promised property.
Lemma 14.14. If y ∈ Hν ∩ Hfin , then T (y) ≈ T̃ (y).
Proof. Write y = P νk=1 αk ek with each αk ∈ C∗ . By the Pythagorean
P
ν
Theorem, kyk2 = 2 2
k=1 |αk | ≥ |αν | , whence αν ∈ Cfin since y ∈ Hfin .
Using that [ajk ] is almost superdiagonal, we have
ν
X ν
X k+1
X
T (y) = αk T (ek ) = αk ajk ej
k=1 k=1 j=1
ν
X ν+1
X
= αk ajk ej
k=1 j=1
ν+1 X
X ν
= ( αk ajk )ej .
j=1 k=1
 LECTURE NOTES ON NONSTANDARD ANALYSIS 99
Pν Pν
Thus, T̃ (y) = Pν T (y) = j=1 ( k=1 αk ajk )ej , whence
ν
X
T (y) = Pν T (y) + αk aν+1,k eν+1
k=1
= Pν T (y) + αν aν+1,ν eν+1
≈ Pν T (y),
since aν+1,ν ≈ 0 and αν ∈ Cfin . 
Corollary 14.15. For any n ∈ N>0 and y ∈ Hν ∩ Hfin , we have T n (y) ≈
T̃ n (y). Consequently, for any polynomial q(z) and any y ∈ Hν ∩ Hfin ,
q(T )y ≈ q(T̃ )(y).
Proof. The proof is by induction on n, the case n = 1 being the content of
the previous lemma. For the induction step, assume that n > 1 and that
the corollary holds for n − 1. Since T̃ is internally bounded and y ∈ Hfin ,
we have that T̃ n−1 (y) ∈ Hfin . Consequently,
T n (y) = T (T n−1 (y)) ≈ T (T̃ n−1 (y)) ≈ T̃ (T̃ n−1 (y)) = T̃ n (y),
where the first ≈ holds by the continuity of T and the inductive assumption
that T n−1 (y) ≈ T̃ n−1 (y), while the second ≈ holds by the base case of the
induction applied to the vector T̃ n−1 (y). 
Lemma 14.16. For i = 0, 1, . . . , ν, we have T (st(Ei )) ⊆ st(Ei ).
Proof. Fix i ∈ {0, 1, . . . , ν} and set E := Ei . Fix x ∈ st(E); we need
T (x) ∈ st(E). By Exercise 11.47, we have PE (x) ≈ x. Since T is continu-
ous, T PE (x) ≈ T x. Consequently, T PE (x) ∈ Hns . Thus, by Lemma 13.6,
Pν (T PE (x)) ≈ T PE (x). Consequently, we have Tν PE (x) = Pν T Pν PE (x) =
Pν T PE (x) ≈ T PE (x) ≈ T x. Since Tν (E) = T̃ (E) ⊆ E, it follows that
T (x) ∈ st(E). 
It thus remains to find some i such that st(Ei ) is nontrivial. To achieve this
goal, we define the internal sequence r0 , r1 , . . . , rν of numbers from R∗ by
rj := kp(Tν )(v) − p(Tν )PEj (v)k. (Recall that we have fixed a vector v ∈ H
of norm 1.) We first show that r0 6≈ 0 while rν ≈ 0. To do this, we first
prove:
Lemma 14.17. p(Tν ) is S-continuous.
Proof. Observe that
k
X k
X
kp(Tν )k ≤ |αi |kTν ki ≤ |αi |kT ki ,
i=0 i=0
whence kp(Tν )k ∈ Rfin . Now suppose that x, y ∈ H ∗ and x ≈ y. Then
kp(Tν )x − p(Tν )yk ≤ kp(Tν )kkx − yk ≈ 0. 
For simplicity, we write Pj instead of PEj .
100 ISAAC GOLDBRING

Corollary 14.18. r0 6≈ 0 and rν ≈ 0.

Proof. By Lemma 13.6, Pν (v) ≈ v. Since p(T ) is continuous, p(T )(Pν (v)) ≈
p(T )(v). However, since Pν (v) ∈ Hν , we have
p(T )(Pν (v)) ≈ p(T̃ )(Pν (v)) = p(Tν )(Pν (v)) ≈ p(Tν )(v).
Thus, p(Tν )(v) ≈ p(T )(v) 6= 0 since {v, T v, T 2 v, . . .} is linearly independent.
Since E0 = {0}, r0 = kp(Tν (v))k 6≈ 0. On the other hand,
rν = kp(Tν )(v) − p(Tν )Pν (v)k ≈ 0
since Pν (v) ≈ v and p(Tν ) is S-continuous.

Observe that in the previous two results, we did not use the fact that p(T )
is compact, that is, the previous two results are valid for any polynomial
q(z).
Choose r ∈ R>0 such that r < r0 . Set η to be the least k ∈ {0, 1, . . . , ν}
such that rk < 21 r; since rν ≈ 0, rν < 12 r, so 0 < η ≤ ν. We will now show
that either st(Eη−1 ) is nontrivial or st(Eη ) is nontrivial.
Lemma 14.19. v ∈
/ st(Eη−1 ) (whence st(Eη−1 ) 6= H).
Proof. Suppose that v ∈ st(Eη−1 ). Then by Exercise 11.47, v ≈ Pη−1 v.
But then, by Lemma 14.17, p(Tν )v ≈ p(Tν )Pη−1 (v), whence rη−1 ≈ 0. But
rη−1 ≥ 21 r, a contradiction. 
Lemma 14.20. st(Eη ) 6= {0}.
Proof. Set y := p(Tν )Pη (v). Since Eη is T̃ -invariant, it is also p(T̃ )-invariant,
whence y ∈ Eη . By Lemma 14.15, y = p(Tν )Pη (v) ≈ p(T )Pη (v). Since p(T )
is compact and Pη (v) ∈ Hfin , we have that p(T )Pη (v) ∈ Hns , whence y ∈
Hns . Let z := st(y) ∈ H, whence z ∈ st(Eη ). If z = 0, then y ≈ 0, whence
rη ≈ kp(Tν )(v)k = r0 > r. This contradicts the fact that rη < 12 r. 
Thus, the only way for both st(Eη−1 ) and st(Eη ) to both be trivial is for
st(Eη−1 ) = {0} and st(Eη ) = H. We now show that this cannot happen for
“dimension reasons.”
Lemma 14.21. Suppose that E, F ∈ E ∗ , E ⊆ F , and dim(F ) = dim(E)+1.
Then for any x, y ∈ st(F ), there is λ ∈ C and z ∈ st(E) such that x = λy +z
or y = λx + z.
Proof. Fix x0 , y 0 ∈ F such that x ≈ x0 and y ≈ y 0 . By the transfer principle,
there is λ ∈ C∗ and z ∈ E such that either y 0 = λx0 + z or x0 = λy 0 + z.
Without loss of generality, we assume y 0 = λx0 + z. If λ ∈ Cfin , then since
z = y 0 − λx0 ∈ E, we have z ∈ Hns and st(z) ∈ st(E). If λ ∈ / Cfin , then
1 0 0 1
λ y = x + λ z, reverting us back to the previous case. 
Suppose, towards a contradiction, that st(Eη−1 ) = {0} and st(Eη ) = H.
By the previous lemma, for any x, y ∈ H, there is λ ∈ C such that x = λy
or y = λx, that is, dim(H) = 1, a serious contradiction. 
 LECTURE NOTES ON NONSTANDARD ANALYSIS 101

15. Measure Theory

In this section, we will describe the Loeb measure construction. There is
a lot to say here, but we will cover only enough to discuss the application
to combinatorics in the final section.
15.1. General measure theory. Fix a set X. A nonempty set Ω ⊆ P(X)
is an algebra if it is closed under unions, intersections, and complements,
that is, if A, B ∈ Ω, then A ∪ B, A ∩ B, and X \ A all belong to Ω. If Ω is
an algebra of subsets of X, then ∅, X ∈ Ω. An algebra Ω on X is said to be
a σ-algebra if it is also closed under countable unions, that is, if A1 , A2 , . . .
all belong to Ω, then so does ∞
S
n=1 n . A σ-algebra is then automatically
A
closed under countable intersections.
Exercise 15.1. Suppose that X is a set and O ⊆ P(X) is an arbitrary
collection of subsets of X. Prove that there is a smallest σ-algebra Ω con-
taining O. We call this σ-algebra the σ-algebra generated by O and denote
it by σ(O).
Remark. When trying to prove that every element of σ(O) has a certain
property, one just needs to show that the set of elements having that prop-
erty contains O and is a σ-algebra.
Suppose that Ω is an algebra on X. A pre-measure on Ω is a function
µ : Ω → [0, +∞] satisfying the following two axioms:
• µ(∅) = 0;
• (Countable Additivity) If A1 , A2 , . . . , all belong
S∞ to Ω, P are pairwise
disjoint, and ∞ ∞
S
A
n=1 n belongs to Ω, then µ( A
n=1 n ) = n=1 µ(An ).
If Ω is a σ-algebra, then a pre-measure is called a measure. If µ is a
measure on X and µ(X) = 1, then we call µ a probability measure on X.
Example 15.2. Fix n ∈ N and suppose that X = {1, 2, . . . , n}. Let Ω :=
P(X). Then Ω is an algebra of subsets of X that is actually a σ-algebra for
trivial reasons. Define the function µ : Ω → [0, 1] by µ(A) = |A|
n . Then µ is
a probability measure on Ω, called the normalized counting measure.
Exercise 15.3. Suppose that µ : Ω → [0, +∞] is a pre-measure. Prove that
µ(A) ≤ µ(B) for all A, B ∈ Ω with A ⊆ B.
For subsets A, B of X, we define the symmetric difference of A and B to
be A4B := (A \ B) ∪ (B \ A).
Exercise 15.4. Suppose that Ω is an algebra and µ : σ(Ω) → [0, ∞] is a
measure. Prove that, for any A ∈ σ(Ω) with µ(A) < ∞ and any  ∈ R>0 ,
there is B ∈ Ω such that µ(A4B) < .
For our purposes, it will be of vital importance to know that a pre-measure
µ on an algebra Ω can be extended to a measure on a σ-algebra Ω0 extending
Ω. We briefly outline how this is done; the interested reader can consult any
good book on measure theory for all the glorious details.
102 ISAAC GOLDBRING

Fix an algebra Ω of subsets of X and a pre-measure µ on Ω. For arbitrary

A ⊆ X, we define the outer measure of A to be
X [
µ+ (A) := inf{ µ(Bn ) | A ⊆ Bn , each Bn ∈ Ω}.
n∈N n∈N

Note that µ+ (A)= µ(A) for all A ∈ Ω.

Now although µ+ is defined on all of P(X) (which is certainly a σ-algebra),
it need not be a measure. However, there is a canonical σ-sub-algebra Ωm of
P(X), the so-called µ+ -measurable subsets of X, on which µ+ is a measure.
(The precise definition of µ+ -measurable is not important here; again, any
standard measure theory text will have the details.) Let us collect the
relevant facts here:
Fact 15.5. Let X be a set, Ω an algebra of subsets of X, and µ : Ω → [0, ∞]
a pre-measure on Ω with associated outer measure µ+ and σ-algebra of µ+ -
measurable setsSΩm . Further suppose that µ is σ-finite, meaning that we
can write X = n∈N Xn with each Xn ∈ Ω and µ(Xn ) < ∞.
(1) σ(Ω) ⊆ Ωm and µ+ |Ω = µ.
(2) (Uniqueness) If Ω0 is another σ-algebra on X extending Ω and µ0 :
Ω0 → [0, ∞] is a measure on Ω0 extending µ, then µ+ and µ0 agree
on Ωm ∩ Ω0 (and, in particular, on σ(Ω)).
(3) (Completeness) If A ⊆ B ⊆ X are such that B ∈ Ωm and µ+ (B) =
0, then A ∈ Ωm and µ+ (A) = 0.
(4) (Approximation Results)
(a) If A ∈ Ωm , then there is B ∈ σ(Ω) containing A such that
µ+ (B \ A) = 0. (So Ωm is the completion of σ(Ω).)
(b) If A ∈ Ωm is such that µ+ (A) < ∞, then for every  ∈ R>0 ,
there is B ∈ Ω such that µ(A4B) < .
(c) Suppose that A ⊆ X is such that, for every  ∈ R>0 , there is
B ∈ Ω such that µ(A4B) < . Then A ∈ Ωm .
Example 15.6 (Lebesgue measure). Suppose that X = R and Ω is the col-
lection of elementary sets, namely the finite unions of intervals. PDefine µ :
Ω → [0, ∞] by declaring µ(I) = length(I) and µ(I1 ∪ · · · ∪ In ) = ni=1 µ(Ij )
whenever I1 , . . . , In are pairwise disjoint. The above outer-measure proce-
dure yields the σ-algebra Ωm , which is known as the σ-algebra of Lebesgue
measurable subsets of R and usually denoted by M. The measure µ+ is often
denoted by λ and is referred to as Lebesgue measure. The σ-algebra σ(Ω)
in this case is known as the σ-algebra of Borel subsets of R. It can also be
seen to be the σ-algebra generated by the open intervals.
15.2. Loeb measure. How do we obtain pre-measures in the nonstandard
context? Well, we obtain them by looking at normalized counting measures
on hyperfinite sets. Suppose that X is a hyperfinite set. We set Ω to be
the set of internal subsets of X. Then Ω is an algebra of subsets of X that
is not (in general) a σ-algebra. For example, if X = {0, 1, . . . , N } for some
 LECTURE NOTES ON NONSTANDARD ANALYSIS 103
S
N > N, then for each n ∈ N, An := {n} belongs to Ω, but n An = N does
not belong to Ω as N is not internal.
If A ∈ Ω, then A is also hyperfinite. We thus define a function µ : Ω →
|A|
[0, 1] by µ(A) := st( |X| ). We claim that µ is a pre-measure. It is easily
seen to be finitely additive, that is, µ(A1 ∪ · · · ∪ An ) = ni=1 µ(Ai ) whenever
P
A1 , . . . , An ∈ Ω are disjoint. But how do we verify countable additivity?

Exercise 15.7. If A1 , A2 , . . . all belong to Ω and ∞

S
S∞ Sk n=1 An also belongs to
Ω, then there is k ∈ N such that n=1 An = n=1 An .

Thus, by the exercise, countable additivity is a trivial consequence of finite

additivity in this context. We may thus apply the outer measure procedure
in the previous section to obtain a probability measure µ+ : Ωm → [0, 1]
extending µ. µ+ is called the Loeb measure on X and is often denoted µL .
The elements of Ωm are referred to as the Loeb measurable subsets of X and
this set is usually denoted by ΩL .
There are many interesting things to say about Loeb measure. It is crucial
for applications of nonstandard analysis to many different areas of mathe-
matics. Let us just mention its connection to Lebesgue measure. We will
prove the following theorem in the exercises.

Theorem 15.8. Suppose that N > N and X = {0, N1 , N2 , . . . , N N

= 1}, a
hyperfinite set. Consider st : X → [0, 1]. Define a σ-algebra Ω on [0, 1] by
A ∈ Ω if and only if st−1 (A) ∈ ΩL . For A ∈ Ω, define ν(A) := µL (st−1 (A)).
Then Ω = M and ν = λ. (In other words, Lebesgue measure can be con-
structed nonstandardly from a hyperfinite normalized counting measure.)

15.3. Product measure. Suppose that (X, ΩX , µX ) and (Y, ΩY , µY ) are

two probability measure spaces. We can then form their product as follows:
first, set Ω to be the set of finite unions of rectangles of the form A × B,
where A ∈ ΩX and B ∈ ΩY . It is an exercise to show that Ω is an algebra
of subsets of X × Y and that every element of Ω can be written as a finite
unionSof disjoint such rectangles. We can then define a pre-measure µ on Ω
by µ( ni=1 (Ai ×Bi )) := ni=1 (µX (Ai )·µY (Bi )). Applying the outer measure
P
procedure, we get a measure µX ⊗ µY : Ωm → [0, 1] extending µ. We denote
Ωm by ΩX ⊗ ΩY .
The following situation will come up in the final section: suppose that
X and Y are hyperfinite sets and we construct the Loeb measure spaces
(X, ΩX,L , µX,L ) and (Y, ΩY,L , µY,L ). We are thus entitled to consider the
product measure space (X × Y, ΩX,L ⊗ ΩY,L , µX,L ⊗ µY,L ). However, X × Y
is itself a hyperfinite set, whence we can consider its Loeb measure space
(X × Y, ΩX×Y,L , µX×Y,L ). There is a connection:

Exercise 15.9. Show that ΩX,L ⊗ ΩY,L is a sub-σ-algebra of ΩX×Y,L and

that µX×Y,L |(ΩX,L ⊗ ΩY,L ) = µX,L ⊗ µY,L .
104 ISAAC GOLDBRING

15.4. Integration. Once one has a measure space (X, Ω, µ), one can inte-
grate as follows. First, for a function f : X → R, we say that f is measurable
if f −1 (U ) ∈ Ω for any open U ⊆ R.
Next, define a simple function g : X → R to be a measurable function
with finite range. Given a set A ⊆ X, we define 1A : X → R by 1A (x) = 1 if
x ∈ A and 0 otherwise. If A P∈ Ω, then 1A is measurable. Any simple function
g can be written as g = ni=1 ri 1Ai with ri ∈ R and R Ai ∈ Ω.PnFor such a
simple function g, we define the integral of g to be gdµ := i=1 ri µ(Ai ).
For an arbitrary positive
R R function f : X → R, we define the
measurable
integral of f to be f dµ R:= sup{ gdµ | g ≤ f, g a simple function}. f
is said to be integrable if f dµ < ∞. For an arbitrary function f , we set
f + := max(f, 0) and f − := max(−f, 0). We say that R f is integrableR if both
f + and f − are integrable, in Rwhich case we defineR f dµ = f + dµ− f − dµ.
R

If A ∈ Ω, then we write A f dµ to indicate f · 1A dµ and call this the

integral of f on A.

15.5. Conditional expectation. Let (X, Ω, µ) be a measure space. We

define an inner-product space L2 (X, Ω, µ) as follows:R the set of vectors is
the set of measurable functions f : X → R such that |f |2 dµ < ∞. We add
vectors and multiply
R by scalars in the usual way. We define the inner product
to be hf, gi := f gdµ. Of course, one has to check that this integral is finite
and this satisfies the requirements of an inner product, but this follows from
the appropriate properties of the integral. (O.K. We are telling a bit of
a lie. Technically speaking, we should be considering equivalence classes
of measurable functions with respect to the equivalence relation f ≡ g if
{x ∈ X | f (x) 6= g(x)} is contained in a set in Ω of measure 0. We will
ignore this important subtlety.)
Theorem 15.10 (Riesz-Fisher Theorem). L2 (X, Ω, µ) is a Hilbert space.
The proof of the Riesz-Fisher Theorem requires some basic convergence
theorems for integrals, but we will not get into this now.
Now suppose that Ω0 is a sub-σ-algebra of Ω. We then get a subspace
L (X, Ω0 , µ) of L2 (X, Ω, µ) consisting of the Ω0 -measurable functions.
2

Exercise 15.11. L2 (X, Ω0 , µ) is a closed subspace of L2 (X, Ω, µ).

We thus have the orthogonal projection P : L2 (X, Ω, µ) → L2 (X, Ω0 , µ).
Consider A ∈ Ω0 with µ(A) < ∞, so 1A ∈ L2 (X, Ω0 , µ). Then for any
g ∈ L2 (X, Ω, µ), we have hg, 1A i = hP (g), 1A i, or, in integral notation:
Z Z
gdµ = P (g)dµ.
A A

In other words, P (g) is an Ω0 -measurable function which has the same in-
tegral as g over subsets of Ω0 . For probability reasons, P (g) is called the
conditional expectation of g with respect to Ω0 and is often denoted E[g|Ω0 ].
 LECTURE NOTES ON NONSTANDARD ANALYSIS 105

15.6. Problems. We are aiming to prove the following:

Theorem: Suppose that N > N and X = {0, N1 , N2 , . . . , N

N = 1}, a hyperfi-
nite set. Consider st : X → [0, 1]. Define a σ-algebra Ω on [0, 1] by A ∈ Ω
if and only if st−1 (A) ∈ ΩL . For A ∈ Ω, define ν(A) := µL (st−1 (A)). Then
Ω = M and ν = λ.
We will do this in steps.
Problem 15.1. Prove that Ω is a σ-algebra and ν is a measure on Ω.
Problem 15.2. Fix a, b ∈ [0, 1] with a < b.
(1) Prove that X ∩ (a, b)∗ ∈ ΩSL and µL (X ∩ (a, b)∗ ) = b − a.
(2) Prove that st−1 ((a, b)) = n∈N (X ∩ (a + n1 , b − n1 )∗ ).
(3) Prove that (a, b) ∈ Ω and ν((a, b)) = b − a.
Let B denote the σ-algebra of Borel subsets of [0, 1]. We now use the fact
that λ is the only measure on B satisfying λ(a, b) = b − a to conclude that
B ⊆ Ω and ν|B = λ|B.
Problem 15.3. Conclude that M ⊆ Ω and ν|M = λ|M. (Hint: Use Fact
15.5.)
It remains to show that Ω ⊆ M. We will need the following fact, whose
proof we will return to after we finish the Lebesgue measure business:
Fact 15.12. If B ∈ ΩX , then, for every  ∈ R>0 , there are internal subsets
C, D of X such that C ⊆ B ⊆ D and µ(D \ C) < .
Now suppose that B ∈ Ω. Fix  ∈ R>0 . By the previous fact, there are
internal C, D ⊆ X such that C ⊆ st−1 (B) ⊆ D and µ(D \ C) < . Set
C 0 := st(C) and D0 := [0, 1] \ st(X \ D). Notice that C 0 is closed and D0 is
open (why?), whence C 0 , D0 ∈ B ⊆ Ω.
Problem 15.4. Prove that C ⊆ st−1 (C 0 ) and st−1 (D0 ) ⊆ D. Conclude that
B ∈ M.
We now explain how to prove Fact 15.12. Recall that the outer measure
procedure yields the following formula for µL :
X [
µL (B) = inf{ µ(An ) | each An is internal and B ⊆ An }.
n∈N n∈N

Problem 15.5. In this problem, we aim to prove, if B ∈ ΩL , then

µL (B) = inf{µL (A) | A is internal and B ⊆ A}.
The inequality ≤ is clear. Fix  ∈ R>0 ; we need to find internal A such that
B ⊆A S and µL (A) ≤ µS L (B) + . Fix internal sets (An | n ∈ N) such that
B ⊆ n∈N An and µL ( n∈N An ) < µL (B) + . Without loss of generality,
we may assume that An ⊆ An+1 for all n. By countable comprehension, we
extend this sequence to an internal sequence (An | n ∈ N∗ ).
106 ISAAC GOLDBRING

(1) Prove that, for each k ∈ N, we have

(∀n ∈ N∗ )(n ≤ k → (An ⊆ Ak and µ(An ) < µL (B) + )).
(2) Prove that there is K > N such that µL (AK ) ≤ µL (B) + , finishing
the proof.
Problem 15.6. Prove a dual version of the previous problem, namely, if
B ∈ ΩL , then
µL (B) = sup{µL (A) | A is internal and A ⊆ B}.
Problem 15.7. Use the previous two problems to prove Fact 15.12.

16. Szemerédi Regularity Lemma

A nice recent application of nonstandard methods is a slick proof of the
celebrated Szemerédi Regularity Lemma for graphs. To state this result, we
need to introduce some terminology. Suppose that (V, E) is a finite graph.
For two disjoint, nonempty subsets X, Y of V , we define the density of
arrows between X and Y to be the quantity
|E ∩ (X × Y )|
δ(X, Y ) := .
|X||Y |
For example, if every element of X is connected to every element of Y by
an edge, then δ(X, Y ) = 1. Fix  ∈ R>0 . We say that X and Y as above are
-pseudorandom if whenever A ⊆ X and B ⊆ Y are such that |A| ≥ |X|
and |B| ≥ |Y |, then |δ(A, B) − δ(X, Y )| < . In other words, as long as A
and B contain at least an  proportion of the elements of X and Y , then
δ(A, B) is essentially the same as δ(X, Y ), so the edges between X and Y
are distributed in a sort of random fashion.
If X = {x} and Y = {y} are singletons, then clearly X and Y are -
pseudorandom for any . Thus, any finite graph can trivially be partitioned
into a finite number of -pseudorandom pairs by partitioning the graph into
singletons. Szemerédi’s Regularity Lemma essentially says that one can
do much better in the sense that there is a constant C() such that any
finite graph has an “-pseudorandom partition” into at most C() pieces.
Unfortunately, the previous sentence isn’t 100% accurate; there’s a bit of
error that we need to account for.
Suppose that V1 , . . . , Vm is a partition of V into m pieces. Set
R := {(i, j) | 1 ≤ i, j ≤ m, Vi and Vj are -pseudorandom}.
|V ||V |
We say that the partition is -regular if (i,j)∈R |Vi |2j > (1 − ). This says
P

that, in some sense, almost all of the pairs of points are in -pseudorandom
pairs. We can now state
Theorem 16.1 (Szemerédi’s Regularity Lemma). For any  ∈ R>0 , there
is a constant C() such that any graph (V, E) admits an -regular partition
into m ≤ C() pieces.
 LECTURE NOTES ON NONSTANDARD ANALYSIS 107

Exercise 16.2. Prove that Szemerédi’s Regularity Lemma is equivalent to

the following statement: for any  ∈ R>0 and any hyperfinite graph (V, E),
there is a finite partition V1 , . . . , Vm of V into internal sets and a subset
R ⊆ {1, . . . , m}2 such that:
• for (i, j) ∈ R, Vi and Vj are internally -pseudorandom: for all
internal A ⊆ Vi and B ⊆ Vj with |A| ≥ |Vi | and |B| ≥ |Vj |, we
have |δ(A, B) − δ(Vi , Vj )| < ; and
P |Vi ||Vj |
• (i,j)∈R |V |2 > (1 − ).

We will thus prove the above nonstandard equivalent of the Szemerédi

Regularity Lemma. Fix  ∈ R>0 and hyperfinite graph (V, E). We consider
three different probability spaces:
• The Loeb measure space (V, ΩV,L , µV,L );
• The product measure space (V × V, ΩV,L ⊗ ΩV,L , µV,L ⊗ µV,L );
• The Loeb measure space (V × V, ΩV ×V,L , µV ×V,L ).
For ease of notation, we will write ΩV for ΩV,L , ΩV ×V for ΩV ×V,L , µV for
µV,L , and µV ×V for µV ×V,L .
As mentioned in the previous section, ΩV ⊗ΩV is a σ-subalgebra of ΩV ×V
and µV ⊗µV is the restriction of µV ×V to ΩV ⊗ΩV . Since E ⊆ V ×V , we can
consider f := E[1E |ΩV ⊗ ΩV ]. The following calculation will prove useful:
|A| |B|
Suppose that A, B ⊆ V are internal and |V | and |V | are noninfinitesimal.
Then (†):
Z Z
f d(µV ⊗ µV ) = 1E dµV ×V by the definition of f
A×B A×B
|E ∩ (A × B)|
= st( )
|V |2
|E ∩ (A × B)| |A||B|
= st( ) st( )
|A||B| |V |2
|A||B|
= st(δ(A, B)) st( ).
|V |2
Fix η ∈ R>0 , to be determined later. Now, since Rf is µV ⊗ µV -integrable,
there is a µV ⊗ µV -simple function g ≤ f such that (f − g)d(µV ⊗ µV ) < η.
√
Set C := {ω ∈ V ×V | f (ω)−g(ω) ≥ η} ∈ ΩV ⊗ΩV . Then (µV ⊗µV )(C) <
√
η, for otherwise
√ √ √
Z Z Z
(f −g)d(µV ⊗µV ) ≥ (f −g)d(µV ⊗µV ) ≥ ηd(µV ⊗µV ) ≥ η η = η.
C C
By Fact 15.12, there is a set D ∈ ΩV ⊗ ΩV that is a finite, disjoint union
of rectangles of the form V 0 × V 00 , with V 0 , V 00 ⊆ V internal sets, such that
√
C ⊆ D and (µV ⊗ µV )(D) < η. In a similar way, we may assume that
the level sets of g (that is, the sets on which g takes constant values) are
finite disjoint unions of rectangles (Exercise). We now take a finite partition
V1 , . . . , Vm of V into internal sets such that g and 1D are constant on each
108 ISAAC GOLDBRING

rectangle Vi × Vj . For ease of notation, set dij to be the constant value of g

on Vi × Vj .

Claim: If µV (Vi ), µV (Vj ) 6= 0 and (Vi × Vj ) ∩ D = ∅, then Vi and Vj are

√
internally 2 η-pseudorandom.

Proof of Claim: Since C ⊆ D, we have that (Vi × Vj ) ∩ C = ∅, whence

√
dij ≤ f (ω) < dij + η for ω ∈ Vi × Vj . (††).
√
Now suppose that A ⊆ Vi and B ⊆ Vj are such that |A| ≥ 2 η|Vi | and
√ |A| |B|
|B| ≥ 2 η|Vj |. In particular, |V i|
and |V j|
are noninfinitesimal. Since
|A| |B|
µV (Vi ), µV (Vj ) > 0, it follows that |V | and |V | are noninfinitesimal and
the calculation (†) applies. Integrating the inequalities (††) on A × B yields:
|A||B| |A||B| √ |A||B|
dij st( ) ≤ st(δ(A, B)) st( ) < (dij + η) st( ).
|V |2 |V |2 |V |2
We thus get:
√
|δ(A, B) − δ(Vi , Vj )| ≤ |δ(A, B) − dij | + |δ(Vi , Vj ) − dij | < 2 η.

By the Claim, we see that we should choose η < ( 2 )2 , so Vi and Vj are

internally -pseudorandom when Vi and Vj are non-null and satisfy (Vi ×
Vj ) ∩ D = ∅. It remains to observe that the -pseudorandom pairs almost
cover all pairs of vertices. Let R := {(i, j) | Vi and Vj are -pseudorandom}.
Then
X |Vi ||Vj | [
st( ) = µ V ×V ( (Vi × Vj ))
|V |2
(i,j)∈R (i,j)∈R
≥ µV ×V ((V × V ) \ D)
√
>1− η
> 1 − .

16.1. Problems. The Szemerédi Regularity Lemma proven above was inte-
gral in Szemerédi’s 1975 proof of the following theorem (henceforth referred
to as Szemerédi’s Theorem), which was originally conjectured by Erdős and
Turán in 1936:
Theorem 16.3. If A ⊆ Z has positive upper Banach density (to be defined
below), then A contains arbitrary long nontrivial arithmetric progressions,
that is, for every k ∈ N>0 , there is n, m ∈ Z with m > 0 such that n, n +
m, . . . , n + (k − 1)m ∈ A.
In 1977, Furstenberg gave a different proof of the Szemerédi’s Theorem.
Rather than using difficult combinatorics a la Szemerédi, Furstenberg trans-
lated the question into a question about ergodic theory (roughly defined as
 LECTURE NOTES ON NONSTANDARD ANALYSIS 109

the study of measure preserving transformations of a probabilty space) and

then proved the ergodic version of Szemerdi’s Theorem. This translation is
now known as the Furstenberg Correspondence Principle. In these problems,
we will outline a nonstandard proof of the Furtstenburg Correspondence
Principle. First, some terminology:
Definition 16.4. If A ⊆ Z, the upper Banach density of A is
|[n, m] ∩ A|
d(A) = lim sup .
m−n→∞ m−n
Definition 16.5. If (X, B, µ) is a probability space, we say that T : X → X
is a measure-preserving transformation if, for all A ∈ B, both T (A) and
T −1 (A) belong to B and µ(T (A)) = µ(T −1 (A)) = µ(A).
Here is the theorem:
Theorem 16.6 (Furstenburg Correspondence Principle). Suppose that A ⊆
Z has positive upper Banach density. Then there is a probability space
(X, B, µ), a measure-preserving transformation T : X → X, and a mea-
surable set A0 ∈ B such that µ(A0 ) = d(A) and such that, for any finite set
U ⊆ Z, we have:
\ \
d( (A − i)) ≥ µ( T −i (A0 )).
i∈U i∈U

Before proving the Furstenberg Correspondence Princple, let us mention

the ergodic-theoretic fact that Furstenburg proved:
Theorem 16.7. Suppose that T : X → X is a measure preserving transfor-
mation on the probability space X, µ(A) > 0, and k ∈ N. Then there exists
n ∈ N such that µ(A ∩ T −n (A) ∩ T −2n (A) ∩ · · · ∩ T −(k−1)n (A)) > 0.
Notice that the above theorem, coupled with the Furstenberg Correspon-
dence Principle, yields a proof of the Szemerédi Theorem.
Problem 16.1. If A ⊆ Z, prove that
|A∗ ∩ I|
d(A) = max{st( ) | I ⊆ Z∗ is an interval of hyperfinite length}.
|I|
(Here, when we say “hyperfinite,” we mean “hyperfinite, but not finite”.)
Problem 16.2. Fix A ⊆ Z with positive upper Banach density. Fix I ⊆ Z∗ ,
∗
an interval of hyperfinite length such that |A|I|∩I| ≈ d(A). Show that:
• the Loeb measure space (I, ΩI , µI ),
• the map T : E → E given by T (x) = x + 1(mod I), and
• the set A0 := A∗ ∩ I
satisfy the conclusion of the Furstenberg Correspondence Princple.
110 ISAAC GOLDBRING

References
[1] M. Davis, Applied nonstandard analysis, John Wiley and Sons Inc., 1977.
[2] R. Goldblatt, Lectures on the hyperreals: an introduction to nonstandard analysis,
volume 188 of Graduate Texts in Mathematics. Springer-Verglag, 1998
[3] C. W. Henson, Foundations of Nonstandard Analysis: A Gentle Introduction to Non-
standard Extensions; Nonstandard Analysis: Theory and Applications, L.O. Arkeryd,
N.J. Cutland, and C.W. Henson, eds., NATO Science Series C; Springer 2001.

Department of Mathematics, Statistics, and Computer Science, University

of Illinois at Chicago, Science and Engineering Offices M/C 249, 851 S. Mor-
gan St., Chicago, IL, 60607-7045
E-mail address: [email protected]
URL: http://www.math.uic.edu/~isaac