Skip to article frontmatterSkip to article content

The proofs in the introduction section are quite general, that is, they make little use of specific properties of R\Real. If we scan the arguments carefully, we see that we can replace R\Real by any metric space that is complete and contains a countable basis of the topology.

Review of some concepts from topology

Basis

Let (X,O)(X, \mathcal{O}) be a topological space. A family BO\mathcal{B} \subseteq \mathcal{O} of subsets if XX is a basis for the topology if every open set from O\mathcal{O} is the union of elements of B\mathcal{B}. For example, the open intervals with rational endpoints form a basis of the standard topology of R\Real. A family SO\mathcal{S} \subseteq \mathcal{O} is a subbasis if the set of finite intersections of sets in S\mathcal{S} is a basis for the topology.

Finally, if S\mathcal{S} is any family of subsets of XX, the topology generated by S\mathcal{S} is the smallest topology on XX containing S\mathcal{S}. It consists of all unions of finite intersections of sets in S{X,}\mathcal{S} \cup \{X,\emptyset\}.

Density

A set DXD \subset X is dense if for any open UU \neq \emptyset there exists zDUz \in D \cap U. If a topological space (X,O)(X, \mathcal{O}) has a countable dense subset, the space is called separable.

Products

If (Xi)iI(X_i)_{i \in I} is a family of topological spaces, one defines the product topology on ΠiIXi\Pi_{i \in I} X_i to be the topology generated by the sets πj1(U)\pi_j^{-1}(U), where jIj \in I, UXjU \subseteq X_j is open, and πj:ΠiIXiXj\pi_j: \Pi_{i \in I} X_i \to X_j is the jjth projection.

Subspaces

If (X,O)(X,\mathcal{O}) is a topological space and YXY \subseteq X, we can put a topology on YY by intersecting it with all open sets in O\mathcal{O}, that is, the open sets are

OY={YU ⁣:UO}.\mathcal{O}_Y = \{Y \cap U \colon U \in \mathcal{O} \}.

This is called the subspace topology on YY.


Now suppose (X,d)(X,d) is a metric space. With each point xXx \in X and every ε>0\eps > 0 we associate an ε\eps-neighborhood or ε\eps-ball

Uε(x)={yX ⁣:d(x,y)<ε}.U_\eps(x) = \{y \in X \colon d(x,y)<\eps\}.

The topology generated by the ε\eps-neighborhoods is called the topology of the metric space (X,d)(X,d). If this topology agrees with a given topology O\mathcal{O} on XX, we say the metric dd is compatible with the topology O\mathcal{O}. If for a topological space (X,O)(X, \mathcal{O}) there exists a compatible metric, (X,O)(X, \mathcal{O}) is called metrizable.

If a topological space (X,O)(X,\mathcal{O}) is separable and metrizable, then the balls with center in a countable dense subset DD and rational radius form a countable basis of the topology.

Polish spaces – the basics

There may be many different compatible metrics that make XX complete. If XX is already given as a complete metric space with countable dense subset, then we call XX a Polish metric space.

The standard example is, of course, R\Real, the set of real numbers. One can obtain other Polish spaces using the following basic observations. (We leave the proof as an exercise.)

We conclude that Rn\Real^n, C\C, Cn\C^n, the unit interval [0,1][0,1], the unit circle T={zC ⁣:z=1}\Ci = \{z \in \C \colon |z| = 1\}, and the infinite dimensional spaces RN\Real^\Nat and [0,1]N[0,1]^\Nat (the Hilbert cube) are Polish spaces.

Any countable set with the discrete topology is Polish, by means of the discrete metric d(x,y)=1xyd(x,y) = 1 \: \Leftrightarrow \: x \neq y.

Some subsets of Polish spaces are Polish but not closed.

We will later characterize all subsets of Polish spaces that are Polish themselves.

Product spaces

In a certain sense, the most important Polish spaces are of the form ANA^\Nat, where AA is a countable set carrying the discrete topology. The standard cases are

2N\Cant, the Cantor space, \qquad and \qquad NN\Baire, the Baire space.

We will, for now, denote elements from ANA^\Nat by lowercase greek letters from the beginning of the alphabet. The nn-th term of α\alpha we denote by either α(n)\alpha(n) or αn\alpha_n, whichever is more convenient.

We endow AA with the discrete topology. The product topology on these spaces has a convenient characterization. Given a set AA, let A<N\Astr{A} be the sets of all finite sequences over AA. Given σ,τA<N\sigma, \tau \in A^{<\Nat}, we write στ\sigma \Sleq \tau to indicate that σ\sigma is an initial segment of τ\tau. \Sle means the initial segment is proper. This notation extends naturally to hold between elements of A<N\Astr{A} and ANA^\Nat, σα\sigma \Sle \alpha meaning that σ\sigma is a finite initial segment of α\alpha.

A basis for the product topology on ANA^\Nat is given by the cylinder sets

Nσ={αAN ⁣:σα},\Cyl{\sigma} = \{\alpha \in A^\Nat \colon \sigma \Sle \alpha \},

that is, the set of all infinite sequences extending σ\sigma. The complement of a cylinder is a union of cylinders and hence open. Therefore, each set Nσ\Cyl{\sigma} is clopen.

A compatible metric is given by

d(α,β)={2N where N is least such that αNβN 0 if α=β.d(\alpha,\beta) = \begin{cases} 2^{-N} & \text{ where } N \text{ is least such that $\alpha_N \neq \beta_N$ }\\ 0 & \text{ if $\alpha = \beta$}. \end{cases}

The representation of the topology via cylinders (which are characterized by finite objects) allows for a combinatorial treatment of many questions and will be essential later on.

Via the mapping

αi=02αi3i+1,\alpha \mapsto \sum_{i = 0}^\infty \frac{2\alpha_i}{3^{i+1}},

2N\Cant is homeomorphic to the middle-third Cantor set in R\Real, whereas the continued fraction mapping

ββ0+1β1+1β2+1β3+\beta \mapsto \beta_0 + \cfrac{1} {\beta_1 + \cfrac{1}{ \beta_2 + \cfrac{1}{ \beta_3 + \ldots}}}

provides a homeomorphism between Z×(N{0})N\Z \times (\N\setminus\{0\})^\N and the irrational real numbers.

The universal role played by the discrete product spaces is manifested in the following results. A Polish (or, more generally, topological) space is perfect if it does not have any isolated points (no singleton set {x}\{x\} is open).

The proof is very similar to the proof of Theorem (Cantor, 1884). Note that this proof actually constructs an embedding of 2N\Cant. The continuity of the mapping is straightforward.

In a similar way we can adapt the proof of Cantor-Bendixson Theorem to show that the perfect subset property holds for closed subsets of Polish spaces.

The special role of Baire space NN\Baire is underlined by the following fact.

Proof

Let dd be a compatible metric on XX, and let D={xi ⁣:iN}D = \{x_i \colon i \in \Nat\} be a countable dense subset of XX. Every point in XX is the limit of a sequence in DD. We could try to define a mapping g:NNXg:\Baire \to X by putting

α=α(0)α(1)α(2)limnxα(n).\alpha = \alpha(0)\, \alpha(1)\, \alpha(2)\dots \mapsto \lim_n x_{\alpha(n)}.

The problem is, of course, that the limit on the right-hand side does not necessarily exist. We have to proceed more carefully. Given αN\alpha \in \Nat, we put y0α=xα(0)y^\alpha_0 = x_{\alpha(0)} and define iteratively

yn+1α={xα(n+1) if d(ynα,xα(n+1))<2n,ynα otherwise .y^\alpha_{n+1} = \begin{cases} x_{\alpha(n+1)} & \text{ if $d(y^\alpha_n,x_{\alpha(n+1)}) < 2^{-n}$}, \\ y^\alpha_n & \text{ otherwise }. \end{cases}

The resulting sequence (ynα)(y^\alpha_n) is clearly Cauchy in XX and hence converges to some point yαXy^\alpha \in X, by completeness. We define

f(α)=yα.f(\alpha) = y^\alpha.

ff is continuous, since if α{\alpha} and β{\beta} agree up to length NN (that is, their distance is at most 2N2^{-N} with respect to the above metric), then the sequences (ynα)(y^\alpha_n) and (ynβ)(y^\beta_n) will agree up to index NN, and all further terms are within 2N2^{-N} of yNαy^\alpha_N and yNβy^\beta_N, respectively.

Finally, since DD is dense in XX, ff is a surjection.

Continuous mappings on product spaces

Let f:ANANf: A^\Nat \to A^\Nat be continuous. We define a mapping ϕ:A<NA<N\phi: \Astr{A} \to \Astr{A} by setting

ϕ(σ)= the longest τ with τσ such that Nσf1(Nτ).\phi(\sigma) = \text{ the longest $\tau$ with $|\tau| \leq |\sigma|$ such that } \Cyl{\sigma} \subseteq f^{-1}(\Cyl{\tau}).

This mapping has the following properties:

  1. It is monotone, i.e. στ\sigma \Sleq \tau implies ϕ(σ)ϕ(τ)\phi(\sigma) \Sleq \phi(\tau).
  2. For any αAN\alpha \in A^\Nat we have limnϕ(αn)=\lim_n |\phi(\alpha\Rest{n})| = \infty. This follows directly from the continuity of ff: For any neighborhood Nτ\Cyl{\tau} of f(α)f(\alpha) there exists a neighborhood Nσ\Cyl{\sigma} of α{}\alpha such that f(Nσ)Nτf(\Cyl{\sigma}) \subseteq \Cyl{\tau}. But τ{}\tau has to be of the form τ=f(α)m\tau = f(\alpha)\Rest{m}, and σ{}\sigma of the form αn\alpha\Rest{n}. Hence for any mm there must exist an nn such that ϕ(αn)f(α)m\phi(\alpha\Rest{n}) \Sgeq f(\alpha)\Rest{m}.

On the other hand, if a function ϕ:A<NA<N\phi: \Astr{A} \to \Astr{A} satisfies (1.) and (2.), it induces a function ϕ:ANAN\phi^*: A^\Nat \to A^\Nat by letting

ϕ(α)=limnϕ(αn)= the unique sequence extending all ϕ(αn).\phi^*(\alpha) = \lim_n \phi(\alpha\Rest{n}) = \text{ the unique sequence extending all $\phi(\alpha\Rest{n})$}.

This ϕ\phi^* is indeed continuous: The preimage of Nτ\Cyl{\tau} under ϕ\phi^* is given by

(ϕ)1(Nτ)={Nσ ⁣:ϕ(σ)τ},(\phi^*)^{-1}(\Cyl{\tau}) = \bigcup \{\Cyl{\sigma} \colon \phi(\sigma) \Sgeq \tau \},

which is an open set.

We have shown the following.

Note that we can completely describe a topological concept, continuity, through a relation between finite strings.