You are on page 1of 1

IV.22.

Set Theory 631

guess that every game is determined, but actually it is 9 Projective Sets and
quite easy, using AC, to prove the existence of a game Descriptive Set Theory
that is not determined.
It turns out that the determinacy of the games asso- As we have seen, very basic questions about sets of
ciated with certain classes of sets of reals implies that real numbers can be extremely hard to answer. How-
all sets in the class have properties similar to those of ever, it often turns out to be possible to answer them
the Borel sets. For example, the axiom of determinacy for sets that occur “in nature,” or that can be explicitly
(AD), which asserts that all sets of reals are determined, described. This raises the hope that one might be able
implies that every set of reals is Lebesgue measurable, to prove facts about definable sets of reals that cannot
has the property of Baire (i.e., differs from an open set be proved for arbitrary sets.
by a set of first category), and has the perfect set prop- The study of the structure of definable sets of reals is
erty (i.e., contains a perfect set if it is uncountable). To the subject of descriptive set theory. Examples of such
give the flavor of a typical argument, let us indicate why sets are the Borel sets, and also the projective sets,
every set A of reals is Lebesgue measurable. which are sets that can be obtained from Borel sets by
First, one observes that it is enough to show that if taking continuous images and complements. An equiv-
all measurable subsets of A are null, then A itself must alent definition of the projective sets is that they are
be null. And for this one plays, for every ε > 0, the subsets of R that can be obtained from closed subsets
covering game for A and ε. In this game, player I plays of Rn by a mixture of projecting to a lower dimension
so that the sequence a = n0 , n2 , n4 , . . .  represents and taking complements. To see how this relates to
an element of A, and player II plays (binary encodings definability, consider projecting a subset A ⊂ R2 down
of) finite unions of rational intervals, with measures to the x-axis. The result will be the set of all x such
adding up to at most ε, while attempting to cover a. that there exists y with (x, y) ∈ A. Thus, projection
It can be shown that if every measurable subset of A is corresponds to existential quantification. Taking com-
null, then player I cannot have a winning strategy. So by plements corresponds to negation, so one can combine
AD there must be a winning strategy for II. Using this the two and obtain universal quantification as well. One
strategy one can show that the outer measure of A is can therefore think of a projective set as a set that is
at most ε. And since this works for all ε > 0, A must definable from a closed set.
be null. Since analytic sets are continuous images of Borel
While AD rules out the existence of badly behaved sets, they are projective. And so are the complements
sets of reals, it implies the negation of AC, so AD is of the analytic sets, the coanalytic sets, and the con-
inconsistent with ZFC. However, weaker versions of AD tinuous images of coanalytic sets, the Σ21 sets. More
are compatible with, and even follow from, ZFC. Indeed, complex projective sets are obtained by taking com-
Donald Martin proved in 1975 that ZFC implies that plements of Σ21 sets, the so-called Π21 sets, their contin-
every Borel set is determined. Moreover, if there exists a uous images, called Σ31 , etc. The projective sets form a
measurable cardinal, then every analytic set, and there- hierarchy of increasing complexity, in accordance with
fore also every coanalytic set, is determined. A natural the number of steps (always finite) that are necessary
question, therefore, is whether the existence of larger to obtain them from the Borel sets. Many sets of reals
cardinals implies the determinacy of more complex that appear naturally in usual mathematical practice
sets such as the Σ21 sets. are projective. Moreover, the results and techniques of
The intimate connection between large cardinals and descriptive set theory, although originally developed
the determinacy of simple sets of reals was first made for the study of sets of reals, also apply to definable
explicit by Leo Harrington, who showed that the deter- sets in any Polish space (a separable and complete-
minacy of all analytic sets is in fact equivalent to a metrizable space). These include basic examples such
large-cardinal principle slightly weaker than the exis- as Rn , C, separable banach spaces [III.62], etc., where
tence of a measurable cardinal. As we shall shortly see, projective sets arise in a very natural way. For example,
large cardinals imply the determinacy of certain sim- in the space C[0, 1] of continuous real-valued functions
ply definable sets of reals, the so-called projective sets, on [0, 1] with the sup norm, the set of everywhere dif-
while the determinacy of those sets implies in turn the ferentiable functions is coanalytic, and the set of func-
existence of the same kind of large cardinals in some tions that satisfy the mean value theorem is Π21 . Thus,
inner models. since descriptive set theory deals with rather natural

You might also like