Applied Nonstandard Analysis

ANALYSIS

INTRODUCTION

1.WHY NONSTANDARD ANALYSIS?

Nonstandard analysis is a technique rather than a subject. Aside from theorems that tell us that nonstandard notions are equivalent to corresponding standard notions, all the results we obtain can be proved by standard methods. Therefore, the subject can only be claimed to be of importance insofar as it leads to simpler, more accessible expositions, or (more important) to mathematical discoveries.

As to the first, the reader must be the judge. The best evidence for the second is the Bernstein–Robinson theory of invariant subspaces of infinite dimensional linear spaces, which settled a question that had remained open for many years. Quite simple standard proofs of their results now exist. Nevertheless, we develop part of their theory not only because this was the path of discovery, but also because it gives us an opportunity to exhibit the truly beautiful idea of approximating an infinite dimensional space from above by a space to which the results of finite dimensional linear algebra are applicable.

2.INFINITESIMALS AS IDEAL ELEMENTS

The study of nonstandard analysis locates itself, mathematically, in the context of the method of ideal elements. This is a time-honored and significant mathematical idea. One simplifies the theory of certain mathematical objects by assuming the existence of additional ideal objects as well. Examples are the embedding of algebraic integers in ideals, the construction of the complex number system, and the introduction of points at infinity in projective geometry.

Nonstandard analysis involves introducing ideal elements that are infinitely close to the objects we are interested in and also ideal elements that are infinitely far away. Leibniz first made use of these ideas in developing differential and integral calculus. In spite of the fact that this was intuitively a very appealing way to think, it seemed to be impossible to provide a sound mathematical basis for the theory of infinitesimals.

There is no difficulty with the purely algebraic problem of embedding the real numbers in a field containing infinitesimals (i.e., what is called in algebraic language a non-Archimedean ordered field). But difficulties arise as soon as one deals with transcendental functions: Thus to differentiate sin x à la Leibniz, one might wish to write:

where dx is infinitesimal. But to write this assumes not only that sine has been defined for numbers of the form real + infinitesimal, but also that this has been done in such a way that the addition law for the sine continues to hold. (In complex function theory analogous problems lead to analytic continuation and the principle of permanence of form.) It is this problem to which modern logic (more precisely, model theory) turns out to hold the key.

3.THE ROLE OF LOGIC

Leibniz postulated a system of numbers having the same properties as ordinary numbers but which included infinitesimals. So Leibniz had no problem with the differentiation of sin x as discussed above. Yet Leibniz' position seems absurd on its face. The ordinary real numbers obviously have at least one property not shared by Leibniz' desired extension. Namely, in the real numbers, there are no infinitesimals.

This paradox is avoided by specifying a formal language in the sense of modern logic (mercilessly precise in the same way that programming languages for computers are). Leibniz' principle is then reinterpreted: there is an extension of the reals that includes infinitesimal elements and has the same properties as the real numbers insofar as those properties can be expressed in the specified formal language. One concludes that the property of being infinitesimal cannot be so expressed, or, as we shall learn to say: the set of infinitesimals is an external set.

4.THREE TECHNIQUES

In nonstandard analysis one is working with two structures, the standard universe and the nonstandard universe. In addition there is a formal language that can be used to make assertions that are ambiguous in that they can refer to either of the two structures.

MATHEMATICAL LOGIC AND RIGOR

There are three main tools in nonstandard analysis. One is the transfer principle, which roughly states that the same assertions of the formal language are true in the standard universe as in the nonstandard universe. It is typically used by proving a desired result in the nonstandard universe, and then, noting that the result is expressible in the language, concluding that it holds in the standard universe as well.

Another technique is concurrence. This is a logical technique that guarantees that the extended structure contains all possible completions, compactifications, and so forth.

The third technique is internality. A set s of elements of the nonstandard universe is internal if s itself is an element of the nonstandard universe ; otherwise, s is external. A surprisingly useful method of proof is one by reductio ad absurdum in which the contradiction is that some set one knows to be external would in fact be internal under the assumption being refuted.

Of course, the above discussion is only approximate. The reader should not expect these matters to be really clear until the detailed exposition has been presented.

5.MATHEMATICAL LOGIC AND RIGOR

Mathematicians who are perfectly content with the usual level of rigor in mathematical exposition in their own speciality and who are rarely troubled by philosophical qualms, sometimes exhibit symptoms of acute anxiety when the same standards are followed in applying mathematical logic. Hence this warning: Although logic is important in discussions of fundamental issues in the foundations of mathematics, these issues need not be injected when mathematical logic is being used just as another mathematical technique.

The use of logic in nonstandard analysis is in some ways analogous to the use of geometric language in mathematics. The intuition enabling us to read a meaning into a sentence of a formal language (technically speaking, a sentence is simply a finite sequence of objects called symbols) can be usefully compared to the intuition that enables us to perceive geometric truths by checking a diagram. Similar expository problems arise in the two cases. Just as writing detailed analytic proofs of things that are obvious from a diagram quickly degenerates into a tedious unrewarding task once one has learned how, so giving detailed proofs that specific sentences in a formal language say what they intuitively seem to say, quickly becomes superfluous. It is recommended that the reader ignore such proofs when they seem unnecessary and supply his own when they are omitted, if there is the least doubt about the matter.

6.NUMBERING OF THEOREMS

A reference to Theorem 2–8.1 is to Theorem 8.1 of Chapter 2, that is, to the first theorem of Section 8 of Chapter 2. When a reference is to a theorem in the same chapter, the chapter reference is omitted. Thus a reference to Theorem 7.2 within Chapter 2 is to Theorem 2–7.2.

1

UNIVERSES

AND LANGUAGES

1.SETS AND RELATIONS

Our development of nonstandard analysis makes use of the convenient fact that the various objects and relations with which mathematics deals can all be construed as sets. Therefore, we begin with a brief survey of the notions of elementary set theory we will be using.

We write x A to indicate that x is an element of the set A, x ∉ A to indicate the contrary. For sets A, B we write A ⊆ B or B ⊇ A to mean that x A implies x B, and A ⊂ B or B ⊃ A to mean that A ⊆ B but A ≠ B.

We use the usual notations

for defining particular sets. The finite set whose members are precisely a1, …, an is written {a1, …, an}. The empty set, written Ø is the only set that has no members. We also consider other objects, called individuals, which have no members, but then these cannot be sets. For any set A, we write

(A) is called the power set of A.

For any objects a,b we write:

〈a,b〉 is called the ordered pair of a and b. Although this definition seems quite arbitrary, the name ordered pair is justified by the

LEMMA. 〈a,b〉 = 〈a′,b′〉 if and only if a = a′ and b = b′.

PROOF. We are given

Equality of these sets means that they have the same members. So we are led to two cases:

CASE 1

{a} = {a′}, {a,b) = {a′,b′}. In this case a = a′, so {a,b) = {a,b′}. If a ≠ b, {a,b} has two members. Hence so does {a,b′} and b = b′. If a = b then a = b′ so that again b = b′.

CASE 2

{a} = {a′, b′}, {a,b} = {a′}. Then, a = a′ = b′ and a = b = a′. So a = a′ and b = a′ = b

We define the (ordered) n-tuple 〈x1, x2, …, xn〉, n > 2 by the recursion

It then follows by a trivial induction that 〈x1, …, xn〉 = 〈y1, …, yn〉 if and only if x1 = x2 = y2, …, xn = yn. In order to include the case n = 1, we write 〈x〉 = x. For any set X, we set

If A and B are sets, we write

If R ⊆ A × B, then R is called a relation. When B = A, we sometimes speak of a relation on A. For R a relation, we sometimes write xRy for 〈x,yR. By the domain of the relation R (written dom(R)), we mean the set of all x such that xRy for some y.

g is called a function (or a mapping) from A into B if g ⊆ A × B (g is a relation) and if for each x A there is exactly one y B such that 〈x,yg. For x A we write g(x) for the unique element of B with 〈x,g(xg. For C ⊆ A, we set

g[C] is called the image of C under g. Here A = dom(g), the domain of g, and the set g[A] is called the range of g. If g[A] = B, g is said to map A onto B. If g(x) = g(y) implies x = y, then g is called one-one.

If f maps Xn into Y we call f an n-ary function (or a function of n arguments) from X into Y, and we write f(x1, …, xn) for f(〈x1, …, xn〉).

Throughout, we write N for the set {0,1,2,3,…} of natural numbers or nonnegative integers. We write N+ for the set {1,2,3,…} of positive integers.

Sometimes, when we are less interested in the domain of a function than its range, we shift our notation and language: we call the domain of the function an index set, and instead of a function we speak of an indexed family. Thus an indexed family with index set I is just a mapping X with domain I. In such a case, for each i I we write Xi instead of X(i) and we write {Xi|i I} to indicate the family itself (i.e., the mapping). When the index set is clear from the context, we sometimes write simply {Xi}. If the index set is of the form I = {1,2,…, n} or {0,1, …, n), the family is called a finite sequence, and if the index set is N or N+, the family is called an infinite sequence.

Let {Xi|i I} be an indexed family of sets, that is, each Xi is a set. Then we write

and speak of the union, the intersection, and the Cartesian product, respectively, of the family. In the special case, I = {1,2, …, n}, we write

If A and B are sets, we write:

Let Ai ⊆ U for i I. Then one readily verifies the so-called de Morgan identities:

2.FILTERS

We use some very simple properties of filters in the construction of our nonstandard universe, and we now proceed to develop the necessary material.

DEFINITION. If I is any nonempty set, then F (I) is called a filter on I if

(1)A F and A ⊆ B implies B F,

(2)A,B F implies A ∩ B F, and

(3)Ø ∉ F, I F.

Of course, (2) implies that the intersection of any finite number of elements of a filter is also in the filter.

EXAMPLE. Let A0 ⊆ I, A0 ≠ Ø. Then as it is very easy to see,

is a filter on I.

DEFINITION. F is an ultrafilter on I if

(1)F is a filter on I.

(2)If F ⊆ F1 and F1 is a filter on I, then F = F1.

EXAMPLE. For any x I, let A0 = {x}. Then the family

is an ultrafilter. For, if F ⊆ F1, and A F1 – F, then x ∉ A (since otherwise A ⊇ {x} and hence A F); thus A ∩ {xF1, so that F1 cannot be a filter.

The basic fact about the existence of ultrafilters is :

THEOREM 2.1. If F0 is a filter on I, then there is an ultrafilter F on I such that F ⊇ F0.

This theorem is, technically speaking, a weak form of the so-called axiom of choice. More precisely, it is known that :

(1)Theorem 2.1 cannot be proved from the axioms of set theory without the axiom of choice;

(2)it can be proved using the axiom of choice;

(3)it is weaker than the axiom of choice, that is, Theorem 2.1 does not imply the axiom of choice.

The proof given below uses Zorn’s lemma, a convenient equivalent of the axiom of choice. The reader who chooses to ignore the proof and to simply use Theorem 2.1 itself as an axiom of set theory is of course welcome to do so. In fact this has some technical advantages: namely, a result whose classical proof uses the full axiom of choice (e.g., the existence of a non-Lebesgue measurable set of real numbers) can sometimes be seen to depend only on the weaker Theorem 2.1 when a nonstandard proof of it is given. However, we are not concerned with such matters here.

= {G ⊇ F0|G is a filter on Iis linearly ordered under ⊆ (i.e., G1,Gimplies G1 ⊆ G2 or G2 ⊆ G⊇ F0, since for each G , G ⊇ Fis a filter on I as we proceed to show:

(1)If A and, B ⊇ A, then A G for some G , where (since G is a filter) B G. Thus B .

(2)If A,B , then A G1, B G2, G1, Gis linearly ordered, either A,B G1 or A,B G2; hence A ∩ B G1 or A ∩ B G2, so that A ∩ B .

G for some G , which is impossible; since I G for all G , we have I .

has a maximal element F. That is, F and F ⊆ Fimplies F = F1. Since F , we have F ⊇ F0 and F is a filter. If F ⊆ F1 where F1 is a filter on I, then clearly F, so that F = F

DEFINITION. G (I) is called a filter basis on I if

(1)Ø ∉ G,

(2)A,B G implies A ∩ B G,

(3)G ≠ Ø.

THEOREM 2.2. If G is a filter basis on I, then there is a filter F on I such that F ⊇ G.

PROOF. Let F = {A ⊆ I|A ⊇ C for some C G}. Clearly G ⊆ F. We now see that F is a filter.

(1)Let A F and B ⊇ A. Then for some C G, A ⊇ C. Since B ⊇ A, we have B ⊇ C, and thus B F.

(2)Let A,B F. Then for C1,CG, A ⊇ C1 and B ⊇ C2, and, therefore, A ∩