Fall 2019 - Problem 1

Lebesgue-Radon-Nikodym derivative, measure theory

Given σ\sigma-finite measures μ1,μ2,ν1,ν2\mu_1, \mu_2, \nu_1, \nu_2 on a measurable space (X,X)\p{X, \mathcal{X}}, suppose that μ1ν1\mu_1 \ll \nu_1 and μ2ν2\mu_2 \ll \nu_2. Prove that the product measures μ1μ2\mu_1 \otimes \mu_2 and ν1ν2\nu_1 \otimes \nu_2 on (X×X,XX)\p{X \times X, \mathcal{X} \otimes \mathcal{X}} satisfy μ1μ2ν1ν2\mu_1 \otimes \mu_2 \ll \nu_1 \otimes \nu_2 and the Radon-Nikodym derivatives obey

d(μ1μ2)d(ν1ν2)(x,y)=dμ1dν1(x)dμ2dν2(y)\deriv{\p{\mu_1 \otimes \mu_2}}{\p{\nu_1 \otimes \nu_2}}\p{x, y} = \deriv{\mu_1}{\nu_1}\p{x} \deriv{\mu_2}{\nu_2}\p{y}

for ν1ν2\nu_1 \otimes \nu_2-almost every (x,y)X×X\p{x, y} \in X \times X.

Solution.

Let EXXE \in \mathcal{X} \otimes \mathcal{X} and suppose (ν1ν2)(E)\p{\nu_1 \otimes \nu_2}\p{E}. Then by Fubini's theorem (all measures are σ\sigma-finite)

(ν1ν2)(E)=χE(x,y)dν1(x)dν2(y)=ν1(Ey)dν2(y)=0.\p{\nu_1 \otimes \nu_2}\p{E} = \int \chi_E\p{x, y} \,\diff\nu_1\p{x} \,\diff\nu_2\p{y} = \int \nu_1\p{E_y} \,\diff\nu_2\p{y} = 0.

where Ey={xX(x,y)E}E_y = \set{x \in X \mid \p{x, y} \in E}. Since ν10\nu_1 \geq 0, it follows that ν1(Ey)=0\nu_1\p{E_y} = 0 for ν2\nu_2-almost every yXy \in X, so by absolute continuity, it follows that μ1(Ey)=0\mu_1\p{E_y} = 0 for ν2\nu_2-almost every yXy \in X, and by absolute continuity again,

ν2({yXμ1(Ey)>0})=0    μ2({yXμ1(Ey)>0})=0,\nu_2\p{\set{y \in X \mid \mu_1\p{E_y} > 0}} = 0 \implies \mu_2\p{\set{y \in X \mid \mu_1\p{E_y} > 0}} = 0,

i.e., μ1(Ey)=0\mu_1\p{E_y} = 0 for μ2\mu_2-almost every yy. Hence,

(μ1μ2)(E)=χE(x,y)dμ1(x)dμ2(y)=μ1(Ey)dμ2(y)=0,\p{\mu_1 \otimes \mu_2}\p{E} = \int \chi_E\p{x, y} \,\diff\mu_1\p{x} \,\diff\mu_2\p{y} = \int \mu_1\p{E_y} \,\diff\mu_2\p{y} = 0,

so μ1μ2ν1ν2\mu_1 \otimes \mu_2 \ll \nu_1 \otimes \nu_2. For the second claim, by absolute continuity, we have

d(μ1μ2)=d(μ1μ2)d(ν1ν2)d(ν1ν2),dμ1=dμ1dν1dν1anddμ2=dμ2dν2dν2.\begin{gathered} \diff\p{\mu_1 \otimes \mu_2} = \deriv{\p{\mu_1 \otimes \mu_2}}{\p{\nu_1 \otimes \nu_2}} \,\diff\p{\nu_1 \otimes \nu_2}, \\ \diff\mu_1 = \deriv{\mu_1}{\nu_1} \,\diff\nu_1 \quad\text{and}\quad \diff\mu_2 = \deriv{\mu_2}{\nu_2} \,\diff\nu_2. \end{gathered}

Observe that if A×BXXA \times B \in \mathcal{X} \otimes \mathcal{X},

A×Bd(μ1μ2)d(ν1ν2)d(ν1ν2)=(μ1μ2)(A×B)=χA×Bdμ1dμ2=(Adμ1)(Bdμ2)=(Adμ1dν1dν1)(Bdμ2dν2dν2)=A×Bdμ1dν1(x)dμ2dν2(y)dν1(x)dν2(y).\begin{aligned} \int_{A \times B} \deriv{\p{\mu_1 \otimes \mu_2}}{\p{\nu_1 \otimes \nu_2}} \,\diff\p{\nu_1 \otimes \nu_2} &= \p{\mu_1 \otimes \mu_2}\p{A \times B} \\ &= \int \chi_{A \times B} \,\diff\mu_1 \,\diff\mu_2 \\ &= \p{\int_A \diff\mu_1}\p{\int_B \diff\mu_2} \\ &= \p{\int_A \deriv{\mu_1}{\nu_1} \,\diff\nu_1} \p{\int_B \deriv{\mu_2}{\nu_2} \,\diff\nu_2} \\ &= \int_{A \times B} \deriv{\mu_1}{\nu_1}\p{x} \deriv{\mu_2}{\nu_2}\p{y} \,\diff\nu_1\p{x} \,\diff\nu_2\p{y}. \end{aligned}

It follows that equality holds when we replace A×BA \times B with countable unions of disjoint measurable rectangles. Hence, by Carathéodory's theorem, we see that d(μ1μ2)d(ν1ν2)d(ν1ν2)\deriv{\p{\mu_1 \otimes \mu_2}}{\p{\nu_1 \otimes \nu_2}} \,\diff\p{\nu_1 \otimes \nu_2} and dμ1dν1(x)dμ2dν2(y)dν1(x)dν2(y)\deriv{\mu_1}{\nu_1}\p{x} \deriv{\mu_2}{\nu_2}\p{y} \,\diff\nu_1\p{x} \,\diff\nu_2\p{y} agree on all measurable sets. Thus,

d(μ1μ2)d(ν1ν2)(x,y)=dμ1dν1(x)dμ2dν2(y)\deriv{\p{\mu_1 \otimes \mu_2}}{\p{\nu_1 \otimes \nu_2}}\p{x, y} = \deriv{\mu_1}{\nu_1}\p{x} \deriv{\mu_2}{\nu_2}\p{y}

ν1ν2\nu_1 \otimes \nu_2-almost everywhere.