Spring 2021 - Problem 1

measure theory, Radon-Nikodym derivative

Let μ\mu be a positive Borel probability measure on [0,1]\br{0, 1}, and let

C=sup{μ(E)|E[0,1] with m(E)=12}.C = \sup\,\set{\mu\p{E} \st E \subseteq \br{0, 1} \text{ with } m\p{E} = \frac{1}{2}}.

Show that there exists a Borel set F[0,1]F \subseteq \br{0, 1} such that

m(F)=12andμ(F)=C.m\p{F} = \frac{1}{2} \quad\text{and}\quad \mu\p{F} = C.

Hint: When dμ=fdx\diff\mu = f \,\diff{x}, one can sometimes take F={x[0,1]|f(x)>λ}F = \set{x \in \br{0, 1} \st f\p{x} > \lambda}, for a suitable λ0\lambda \geq 0.

Solution.

First suppose μm\mu \ll m so that dμ=fdx\diff\mu = f \,\diff{x} with fL1([0,1])f \in L^1\p{\br{0,1}}. For t0t \geq 0, let Ft={f(x)>t}F_t = \set{f\p{x} > t} and set

λ=inf{t|m(Ft)12}\lambda = \inf\, \set{t \st m\p{F_t} \leq \frac{1}{2}}

so that m(Fλ)12m\p{F_\lambda} \leq \frac{1}{2}. Observe that

m({f(x)λ})=m(n=1{f(x)>λ1n})=m(n=1Fλ1n)12.m\p{\set{f\p{x} \geq \lambda}} = m\p{\bigcap_{n=1}^\infty \set{f\p{x} > \lambda - \frac{1}{n}}} = m\p{\bigcap_{n=1}^\infty F_{\lambda - \frac{1}{n}}} \geq \frac{1}{2}.

Set Eλ={f(x)=λ}E_\lambda = \set{f\p{x} = \lambda}, and notice that xm(Eλ[0,x])x \mapsto m\p{E_\lambda \cap \br{0, x}} is a continuous function by continuity of measures. In particular, by the intermediate value theorem and our estimate above, we see that there exists AEλA \subseteq E_\lambda such that m(A)=12m(Fλ)m\p{A} = \frac{1}{2} - m\p{F_\lambda}. Thus, if we let F=FλAF = F_\lambda \cup A, then m(F)=12m\p{F} = \frac{1}{2}. We claim that μ(F)=C\mu\p{F} = C as well. Indeed, suppose E[0,1]E \subseteq \br{0, 1} satisfies m(E)=12m\p{E} = \frac{1}{2}. Observe that

m(E)=m(EF)+m(EF)m(F)=m(EF)+m(FE),\begin{aligned} m\p{E} &= m\p{E \cap F} + m\p{E \setminus F} \\ m\p{F} &= m\p{E \cap F} + m\p{F \setminus E}, \end{aligned}

so because m(E)=m(F)m\p{E} = m\p{F}, we have m(EF)=m(FE)m\p{E \setminus F} = m\p{F \setminus E}. But on EFE \setminus F, we have fλf \leq \lambda, and on FEF \setminus E, we have fλf \geq \lambda. Thus,

μ(E)=Efdx=EFfdx+EFfdxEFfdx+FEfdx=Ffdx=μ(F),\begin{aligned} \mu\p{E} &= \int_E f \,\diff{x} \\ &= \int_{E \cap F} f \,\diff{x} + \int_{E \setminus F} f \,\diff{x} \\ &\leq \int_{E \cap F} f \,\diff{x} + \int_{F \setminus E} f \,\diff{x} \\ &= \int_{F} f \,\diff{x} \\ &= \mu\p{F}, \end{aligned}

so because EE was arbitrary, it follows that μ(F)=C\mu\p{F} = C.

Finally, in the most general case, we may apply the Lebesgue-Radon-Nikodym theorem to write dμ=fdx+dλ\diff\mu = f \,\diff{x} + \diff\lambda, where λ\lambda is singular with respect to Lebesgue measure. Then by the special case, there exists FF such that m(F)=12m\p{F} = \frac{1}{2} and Ffdx=C\int_F f \,\diff{x} = C. Let SS be such that m(S)=λ(Sc)=0m\p{S} = \lambda\p{S^\comp} = 0, and let F=FScF' = F \cap S^\comp. Then λ(F)=0\lambda\p{F'} = 0 and so

μ(F)=FScfdx+Sfdx=Fdx=C,\mu\p{F'} = \int_{F \cap S^\comp} f \,\diff{x} + \int_S f \,\diff{x} = \int_{F'} \,\diff{x} = C,

which was what we wanted to show.