Fall 2009 - Problem 5

construction, measure theory

Construct a Borel subset EE of the real line R\R such that for all intervals [a,b]\br{a, b} we have

0<m(E[a,b])<ba0 < m\p{E \cap \br{a, b}} < b - a

where mm denotes Lebesgue measure.

Solution.

Let {rn}n\set{r_n}_n be an enumeration of Q\Q. Define In=(rn2n,rn+2n)I_n = \p{r_n - 2^{-n}, r_n + 2^{-n}},

Gn=knIk.G_n = \bigcup_{k \geq n} I_k.

GnG_n is open since it is a union of open sets. Also GnG_n contains all but possibly finitely many rational numbers, so it is dense. By construction, Gn+1GnG_{n+1} \subseteq G_n and

m(Gn)k=n22k=42nn0.m\p{G_n} \leq \sum_{k=n}^\infty 2 \cdot 2^{-k} = 4 \cdot 2^{-n} \xrightarrow{n\to\infty} 0.

Fix k1k \geq 1. We will construct disjoint sets which remove a small, positive measure portion of IkI_k.

By density of Q\Q again, for any n1n \geq 1, there exists jnj \geq n such that qjIk    IjIkq_j \in I_k \implies I_j \cap I_k \neq \emptyset. Since IjGnI_j \subseteq G_n,

m(IkGn)m(IkIj)>0.m\p{I_k \cap G_n} \geq m\p{I_k \cap I_j} > 0.

Thus, since m(G)0m\p{G_\ell} \to 0 as \ell \to \infty, there exists >n\ell > n such that

0<m(G)<m(IkGn)    m(IkGnG)>0,0 < m\p{G_\ell} < m\p{I_k \cap G_n} \implies m\p{I_k \cap G_n \setminus G_\ell} > 0,

since GGnG_\ell \subseteq G_n. Intuitively, any interval IkI_k will intersect with GnG_n with positive measure. To get infinitely many of these, we can remove the tail GG_\ell of GnG_n and continue inductively.

Let n1=1n_1 = 1, and inductively pick n1<n2<n_1 < n_2 < \cdots as follows: Given n2k1n_{2k-1},

n2k>n2k1 such that m(IkGn2k1Gn2k)>0, andn2k+1>n2k such that m(IkGn2kGn2k+1)>0.\begin{aligned} \exists &&n_{2k} &> n_{2k-1} &&\text{ such that } m\p{I_k \cap G_{n_{2k-1}} \setminus G_{n_{2k}}} &&> 0, \text{ and} \\ \exists &&n_{2k+1} &> n_{2k} &&\text{ such that } m\p{I_k \cap G_{n_{2k}} \setminus G_{n_{2k+1}}} &&> 0. \end{aligned}

Moreover, because nm    GmGnn \subseteq m \implies G_m \subseteq G_n, it follows that the Gn2k1Gn2kG_{n_{2k-1}} \setminus G_{n_{2k}} are pairwise disjoint. Set

E=k=1Gn2k1Gn2k.E = \bigcup_{k=1}^\infty G_{n_{2k-1}} \setminus G_{n_{2k}}.

Now let I=[a,b]RI = \br{a, b} \subseteq \R be such that a<ba < b. If nn is large enough, then [a+2n,b2n]\br{a + 2^{-n}, b - 2^{-n}} \neq \emptyset. Thus, by density, there exists knk \geq n such that qk[a+2n,b2n]q_k \in \br{a + 2^{-n}, b - 2^{-n}}, and so

a+2nqkb2n    aqk2nqk+2nb,a + 2^{-n} \leq q_k \leq b - 2^{-n} \implies a \leq q_k - 2^{-n} \leq q_k + 2^{-n} \leq b,

which means IkII_k \subseteq I.

Notice that E(Gn2kGn2k+1)=E \cap \p{G_{n_{2k}} \setminus G_{n_{2k+1}}} = \emptyset by construction. Indeed, Gn21G_{n_{2\ell-1}} is always odd, whereas Gn2kG_{n_{2k}} is even, and the GnmGnm+1G_{n_{m}} \setminus G_{n_{m+1}} are pairwise disjoint. Consequently, Gn2kGn2k+1EcG_{n_{2k}} \setminus G_{n_{2k+1}} \subseteq E^\comp, which means

m(IE)m(IkE)m(IkGn2kGn2k+1)>0.\begin{aligned} m\p{I \setminus E} \geq m\p{I_k \setminus E} \geq m\p{I_k \cap G_{n_{2k}} \setminus G_{n_{2k+1}}} > 0. \end{aligned}

On the other hand,

m(EI)m(EIk)m(IkGn2k1Gn2k)>0.m\p{E \cap I} \geq m\p{E \cap I_k} \geq m\p{I_k \cap G_{n_{2k-1}} \setminus G_{n_{2k}}} > 0.

Hence, 0<m(EI)<m(I)0 < m\p{E \cap I} < m\p{I}, which was what we wanted.