Skip to article frontmatterSkip to article content

Subsets with respect to the subspace topology.

Closed subsets of Polish spaces (with the subspace topology) are Polish (Proposition 1 in the Section Polish Spaces).

What about other subsets of a Polish space? In a previous exercise, we saw (0,1)R(0,1) \subset \Real is Polish, but we had to pick a different compatible metric. It turns out we can do this for arbitrary open subsets of a Polish space.

Proof

Use

d(x,y)=d(x,y)1+d(x,y).d'(x,y) = \frac{d(x,y)}{1+d(x,y)}.
Proof

For a set AXA \subseteq X and a metric dd on XX, we define the distance of xx from AA as

d(x,A)=inf{d(x,a) ⁣:aA}d(x,A) = \inf \{ d(x,a) \colon a \in A\}

Let UXU \subseteq X be open. Then d(x,XU)d(x , X \setminus U) is strictly positive.

By Lemma 1, we can choose a compatible metric dd on XX with d<1d < 1. Let

dU(x,y)=d(x,y)+1d(x,XU)1d(y,XU).d_U(x,y) = d(x,y) + \left | \frac{1}{d(x,X\setminus U)} - \frac{1}{d(y,X\setminus U)} \right|.

This is a metric on UU compatible with the subspace topology and with respect to which UU is complete (exercise!).

Proof

Consider the mapping f:XXNf: X \to X^\Nat given by x(x,x,x,)x \mapsto (x, x, x, \dots). The restriction of ff to nYn\bigcap_n Y_n is a homeomorphism between nYn\bigcap_n Y_n and the diagonal ΔnYn\Delta \subseteq \prod_n Y_n,

Δ={(x,x,x,) ⁣:xYn for all nN}.\Delta = \{(x,x,x, \dots) \colon \: x \in Y_n \text{ for all } n \in \Nat \}.

Δ{}\Delta is closed in the product space nYn\prod_n Y_n and hence Polish, and this property extends to nYn\bigcap_n Y_n (see Proposition 1).

Hence every GδG_\delta subset of a Polish space is Polish. This is as far as we can get.

We have already established the “if” direction of this result. For the other direction, we need a lemma that is interesting in its own right.

Compare this with the last lecture, where we showed that the points of continuity of a function is always a GδG_\delta set (Theorem 2).

Proof

We can adapt the ε\eps-oscillation set CεC_\eps used in the proof of Theorem 2 to the domain AA:

CεA={xX ⁣:δ>0a,bA  [a,bUδ(x)d(f(a),f(b))<ε]}.()\tag{$*$} C^A_\eps = \left \{ x \in X \colon \exists \delta > 0 \: \forall a,b \in A \; [ a,b \in U_\delta(x) \: \Rightarrow \: d(f(a),f(b)) < \eps ] \right \}.

As before, CεAC^A_\eps is open and hence

G=AnC1/nAG = \Cl{A} \cap \bigcap_n C^A_{1/n}

is GδG_\delta and since ff is continuous, AGAA \subseteq G \subseteq \Cl{A}.

To extend ff to GG, let xGx \in G. Since xAx \in \Cl{A}, there exists a sequence (an)(a_n) in AA with x=limnanx = \lim_n a_n. As xnC1/nAx \in \bigcap_n C^A_{1/n}, the sequence (f(an))(f(a_n)) is Cauchy. YY is complete, so there exists y=limnf(an)Yy = \lim_n f(a_n) \in Y. It is straightforward to verify that yy is independent of the choice of (an)(a_n) and agrees with f(x)f(x) for xAx \in A. Hence we can put

g(x)=y,g(x) = y,

which yields the desired continuous extension.

Now assume YXY \subset X is Polish. Then, by the previous lemma, the identity mapping id:YY\operatorname{id}: Y \to Y has a continuous extension g:GYg: G \to Y to a GδG_\delta set GG with YGYY \subseteq G \subseteq \Cl{Y}. Let xGx \in G. YY is dense in GG, so there exists (yn)(y_n) in YY with x=limnynx = \lim_n y_n. By continuity of gg,

x=limnyn=limng(yn)=g(x)Y.x = \lim_n y_n = \lim_n g(y_n) = g(x) \in Y.

It follows that GYG \subseteq Y and therefore Y=GY = G. This completes the proof of Theorem 1.

Borel set as clopen sets

More complicated Borel sets in Polish spaces are not Polish anymore in the subspace topology, as we just saw. But what if we are allowed to change the topology? In the process, we would like to “preserve” as much as possible of the original space. It turns out we can change the topology so that a given Borel set becomes clopen while inducing the same family of Borel sets overall.

We start with closed sets.

Proof

By Proposition 1 of the section on Polish spaces and Proposition 1 above, respectively, FF and XFX \setminus F are Polish spaces with compatible metrics dFd_F and dXFd_{X\setminus F}, respectively. By Lemma 1, we may assume dF,dXF<1d_F, d_{X\setminus F} < 1. We form the disjoint union of the spaces FF and XFX \setminus F: This is the set X=F(XF)X = F \,\sqcup\, (X \setminus F) with the following topology O\mathcal{O}': UF(XF)U \subseteq F \,\sqcup\, (X \setminus F) is in O\mathcal{O}' if and only if UFU \cap F is open (in FF) and U(XF)U \cap (X\setminus F) is open (in XFX\setminus F).

The disjoint union is Polish, as witnessed by the following metric.

d(x,y)={dF(x,y)if x,yF,dXF(x,y)if x,yXF,  2otherwise.d_\sqcup(x,y) = \begin{cases} d_F(x,y) &\text{if } x,y \in F, \\ d_{X\setminus F}(x,y) &\text{if } x,y \in X\setminus F, \\ \; 2 &\text{otherwise}. \end{cases}

It is straightforward to check that dd_\sqcup is compatible with O\mathcal{O}'. Furthermore, let (xn)(x_n) be Cauchy in (X,d)(X,d_\sqcup). Then the xnx_n are completely in FF or in XFX\setminus F from some point on, and hence (xn)(x_n) converges.

Under the disjoint union topology, FF is is clopen. Moreover, an open set in this topology is a disjoint union of an open set in XFX\setminus F, which is also open in the original topology O\mathcal{O}, and an intersection of an open set from O\mathcal{O} with FF. Such sets are Borel in (X,O)(X,\mathcal{O}) and hence (X,O)(X,\mathcal{O}) and (X,O)(X,\mathcal{O}') have the same Borel sets.

Proof

Let S\mathcal{S} be the family of all subsets AA of XX for which a finer Polish topology exists that has the same Borel sets as O\mathcal{O} and in which AA is clopen.

We will show that S\mathcal{S} is a σ\sigma-algebra, which by the previous lemma contains the closed sets. Hence S\mathcal{S} must contain all Borel sets, and we are done.

S\mathcal{S} is clearly closed under complements, since the complement of a clopen set is clopen in any topology.

So assume now that {An}\{A_n\} is a countable family of sets in S\mathcal{S}. Let On\mathcal{O}_n be a Polish topology on XX that makes AnA_n clopen and does not introduce new Borel sets.

Let O\mathcal{O}_\infty be the topology generated by nOn\bigcup_n \mathcal{O}_n. Then nAn\bigcup_n A_n is open in (X,O)(X, \mathcal{O}_\infty), and we can apply Lemma 3. For this to work, however, we have to show that (X,O)(X, \mathcal{O}_\infty) is Polish and does not introduce any new Borel sets.

We know that the product space n(X,On)\prod_n (X,\mathcal{O}_n) is Polish. Once again we consider the mapping ϕ:XnX\phi: X \to \prod_n X

x(x,x,x,).x \mapsto (x,x,x, \dots).

Observe that ϕ{}\phi is a continuous mapping between (X,O)(X,\mathcal{O}_{\infty}) and nX\prod_n X. The preimage of a basic open set U1×U2××Un×X×X×U_1 \times U_2 \times \cdots \times U_n \times X \times X \times \cdots under ϕ{}\phi is just the intersection of the UiU_i. Furthermore, ϕ{}\phi is clearly one-to-one, and the inverse mapping between ϕ(X)\phi(X) and XX is continuous, too.

If we can show that ϕ(X)\phi(X) is closed in nX\prod_n X, we know it is Polish as a closed subset of a Polish space, and since (X,O)(X,\mathcal{O}_\infty) is homeomorphic to ϕ(X)\phi(X) (in the subspace topology for n(X,On)\prod_n (X,\mathcal{O}_n)), we can conclude it is Polish.

To see that ϕ(X)\phi(X) is closed in nX\prod_n X, let (y1,y2,y3,)¬ϕ(X)(y_1,y_2,y_3, \dots) \in \Co{\phi(X)}. Then there exist i<ji < j such that yiyjy_i \neq y_j. Since (X,O)(X, \mathcal{O}) is Polish, we can pick U,VU,V open, disjoint such that yiUy_i \in U, yjVy_j \in V. Since each On\mathcal{O}_n refines O\mathcal{O}, UU is open in Oi\mathcal{O}_i, and VV is open in Oj\mathcal{O}_j. Therefore,

X1×X2××Xi1×U×Xi+1××Xj1×V×Xj+1×Xj+2×X_1 \times X_2 \times \cdots \times X_{i-1} \times U \times X_{i+1} \times \cdots \times X_{j_1} \times V \times X_{j+1} \times X_{j+2} \times \cdots

where Xk=XX_k = X for ki,jk \neq i,j, is an open neighborhood of (y1,y2,y3,)(y_1,y_2,y_3, \dots) completely contained in ¬ϕ(X)\Co{\phi(X)}.

Finally, to see that the Borel sets of (X,O)(X, \mathcal{O}_\infty) are the same as the ones of (X,O)(X,\mathcal{O}), for each nn, let {Ui(n)}iN\{U^{(n)}_i\}_{i \in \Nat} be a basis for On\mathcal{O}_n. By assumption, all sets in On\mathcal{O}_n are Borel sets of (X,O)(X, \mathcal{O}). The collection {Ui(n)}i,nN\{U^{(n)}_i\}_{i,n \in \Nat} is a subbasis for O\mathcal{O}_\infty. This means that any open set in (X,O)(X, \mathcal{O}_\infty) is a countable union of finite intersections of the Ui(n)U^{(n)}_i. Since every Ui(n)U^{(n)}_i is Borel in (X,O)(X, \mathcal{O}), this means that any open set in O\mathcal{O}_\infty is Borel in (X,O)(X, \mathcal{O}). Since the Borel sets are closed under complementation and countable unions, this in turn implies that every Borel set of (X,O)(X, \mathcal{O}_\infty) is already Borel in (X,O)(X, \mathcal{O}).

Proof

Let (X,O)(X,\mathcal{O}) be Polish, and assume BXB \subseteq X is Borel. We can choose a finer topology OO\mathcal{O}' \supseteq \mathcal{O} so that BB becomes clopen, but the Borel sets stay the same. By Theorem 1, BB is Polish with respect to the subspace topology OB\mathcal{O}'|_B

Suppose BB is uncountable. By Theorem 1 there exists a continuous injection ff from 2N\Cant (with respect to the standard topology) into (B,OB)(B,\mathcal{O}'|_B).

Since O\mathcal{O}' is finer than O\mathcal{O}, ff is continuous with respect to O\mathcal{O}, too. Since 2N\Cant is compact, f(2N)f(\Cant) is closed with respect to O\mathcal{O}. Finally, f(2N)f(\Cant) has no isolated points with respect to O\mathcal{O}', which then also holds for the coarser topology O\mathcal{O}.

Therefore, BB has a perfect subset.