Probability Theory 1에서 가장 중요한 Part이다. 이 내용들을 이해하기 위해 Product measure 파트가 필요하다.



Theorem (Fubini’s Theorem)

If \((\Omega_1,\mathcal{F}_1, \mu_1)\) and \((\Omega_2\mathcal{F}_2, \mu_2)\) are \(\sigma\)-finite measure spaces and if \(f:\Omega_1\times\Omega_2\rightarrow \bar{\mathbb{R}}\) is and \(\mathcal{F}_1\times\mathcal{F}_2\)-measurable function for which \(\int_{\Omega_1\times\Omega_2}f\mbox{ }d(\mu_1\times\mu_2)\) exists, then \[\begin{eqnarray*} \int_{\Omega_1\times\Omega_2}f(\omega_1, \omega_2)\mbox{ }d(\mu_1\times\mu_2)(\omega_1, \omega_2)&=&\int_{\Omega_1}\left[\int_{\Omega_2}f(\omega_1,\omega_2)\mbox{ } d\mu_2(\omega_2)\right]d\mu_1(\omega_1) \\ &=&\int_{\Omega_2}\left[\int_{\Omega_1}f(\omega_1,\omega_2)\mbox{ } d\mu_1(\omega_1)\right]d\mu_2(\omega_2). \end{eqnarray*}\] Here, \(\int_{\Omega_2}f(\omega_1,\omega_2)\mbox{ } d\mu_2(\omega_2)\) is defined for \(\mu_1\)-almost all \(\omega_1\) and we take its value to be zero(or any other arbitrary constant) for those \(\omega_1\) for which it is not defined.

Similarly, \(\int_{\Omega_1}f(\omega_1,\omega_2)\mbox{ } d\mu_1(\omega_1)\) is defined for \(\mu_2\)-almost all \(\omega_2\) and we take its value to be zero for those \(\omega_2\) for which it is not defined.

In particular, the result holds if \(f\) is either nonnegative or integrable w.r.t \(\mu_1\times\mu_2\), and in the latter case, \(\int_{\Omega_2}f(\omega_1,\omega_2)\mbox{ } d\mu_2(\omega_2)\) is finite for \(\mu_1\)-almost all \(\omega_1\) and \(\int_{\Omega_1}f(\omega_1,\omega_2)\mbox{ } d\mu_1(\omega_1)\) is finite for \(\mu_2\)-almost all \(\omega_2\) and



Definition(Sections)

To set notation, for \(A\subset\Omega_1\times\Omega_2\), let \[ A_{\omega_1}^{(1)}=\{\omega_2:(\omega_1,\omega_2)\in A \}\subset \Omega_2, \] and \[ A_{\omega_2}^{(2)}=\{\omega_1:(\omega_1,\omega_2)\in A \}\subset \Omega_1 \] be the \(\omega_1\) and \(\omega_2\) sections of \(A\), respectively.



Lemma

If \(A\in \mathcal{F}_1\times \mathcal{F}_2\), then \(A_{\omega_1}^{(1)}\in \mathcal{F}_2\) for all \(\omega_1\in \Omega_1\).




증명을 위해 필요

Lemma(Fubini’s theorem for indicators)

If \((\Omega_1,\mathcal{F}_1, \mu_1)\) and \((\Omega_2,\mathcal{F}_2, \mu_2)\) are \(\sigma\)-finite measure spaces, and if \(A\in \mathcal{F}_1\times\mathcal{F}_2\), then

  1. for fixed \(\omega_1\), the function \(\omega_2 \mapsto I_A(\omega_1,\omega_2)\) is \(\mathcal{F}_2\)-measurable;

  2. the function \(\omega_1\mapsto \int_{\Omega_2}I_A(\omega_1,\omega_2)d\mu_2(\omega_2)\) is \(\mathcal{F}_1\)-measurable;

    • Note \(\int_{\Omega_2}I_A(\omega_1,\omega_2)d\mu_2(\omega_2)\)\(\omega_2\)에 대한 적분이기 때문에 \(\omega_1\)은 적분과 관계가 없다. \(\omega_1\)을 fix되었다고 생각한다면 \[ \int_{\Omega_2}I_A(\omega_1,\omega_2)d\mu_2(\omega_2)=\int_{\Omega_2}I_{A_{\omega_1}^{(1)}}(\omega_2)d\mu_2(\omega_2)=\mu_2(A_{\omega_1}^{(1)}) \] 이기 때문에 2번을 \(\omega_1\mapsto \mu_2(A_{\omega_1}^{(1)})\) 라고 쓸 수 있다.
  3. \((\mu_1\times\mu_2)(A)=\int_{\Omega_1}\left[\int_{\Omega_2}I_A(\omega_1, \omega_2)d\mu_2(\omega_2)\right]d\mu_1(\omega_1)\).







Simple function \(f\) \(\rightarrow\) nonnegative function \(f\) \(\rightarrow\) general function \(f\) 순서로 증명할 것이다.

Proof of Fubini’s Theorem(simple function \(f\))

Suppose that \(f=\sum_{i=1}^m a_iI_{A_i}\) is a nonnegative simple function, with \(A_1,\ldots, A_m \in \mathcal{F}_1\times\mathcal{F}_2\). Then, for each fixed \(\omega_1\), the function \[ \omega_2\mapsto f(\omega_1,\omega_2)=\sum_{i=1}^m a_i I_{A_i}(\omega_1,\omega_2) \] is \(\mathcal{F}_2\)-measurable, being a nonnegative linear combination of \(\mathcal{F}_2\)-measurable functions. Also, by linearity of the integral, \[ \int_{\Omega_2} f(\omega_1, \omega_2)d\mu_2(\omega_2)= \sum_{i=1}^m a_i\int_{\Omega_2}I_{A_i}(\omega_1,\omega_2)I_{A_i}(\omega_1,\omega_2)d\mu_2(\omega_2) \] so that the function \[ \omega_1\mapsto \int_{\Omega_2} f(\omega_1,\omega_2)d\mu_2(\omega_2) \] is \(\mathcal{F}_1\)-measurable, being a nonnegative linear combination of \(\mathcal{F}_1\)-measurable functions.

Finally by linearity of the integral and Fubini’s theorem for indicator functions, \[\begin{eqnarray*} \int_{\Omega_1\times\Omega_2} f(\omega_1,\omega_2)d(\mu_1\times\mu_2)(\omega_1, \omega_2)&=&\int_{\Omega_1\times\Omega_2} \sum_{i=1}^m a_iI_{A_i}(\omega_1,\omega_2) d(\mu_1\times\mu_2)(\omega_1, \omega_2)\mbox{ }\mbox{ }\mbox{ by the definition of }f\\ &=&\sum_{i=1}^m a_i\int_{\Omega_1\times\Omega_2} I_{A_i}(\omega_1,\omega_2) d(\mu_1\times\mu_2)(\omega_1, \omega_2) \\ &=&\sum_{i=1}^m a_i\int_{\Omega_1}\left[\int_{\Omega_2} I_{A_i}(\omega_1,\omega_2) d\mu_2(\omega_2)\right]d\mu_1(\omega_1) \mbox{ }\mbox{ }\mbox{ by }3 \mbox{ of the above lemma}\\ &=&\int_{\Omega_1}\left[\sum_{i=1}^m a_i\int_{\Omega_2} I_{A_i}(\omega_1,\omega_2) d\mu_2(\omega_2)\right]d\mu_1(\omega_1)\\ &=&\int_{\Omega_1}\left[\int_{\Omega_2}\sum_{i=1}^m a_i I_{A_i}(\omega_1,\omega_2) d\mu_2(\omega_2)\right]d\mu_1(\omega_1)\\ &=&\int_{\Omega_1}\left[\int_{\Omega_2}f(\omega_1,\omega_2) d\mu_2(\omega_2)\right]d\mu_1(\omega_1)\mbox{ }\mbox{ }\mbox{ by the definition of }f. \end{eqnarray*}\]



Proof of Fubini’s Theorem(Nonnegative \(f\))

For \(\mathcal{F}_1\times\mathcal{F}_2\)-measurable \(f\ge 0\), let \(f_n, n\ge 1\) be simple function satisfying \(0\le f_n \uparrow f\). Then, for each fixed \(\omega_1\), the function \[ \omega_2\mapsto f(\omega_1,\omega_2)= \lim_n f_n(\omega_1, \omega_2) \] is \(\mathcal{F}_2\)-measurable, being a limit of \(\mathcal{F}_2\)-measurable functions.

Also, by the monotone convergence theorem, \[ 0\le \int_{\Omega_2}f_n(\omega_1,\omega_2)d\mu_2(\omega_2)\uparrow \int_{\Omega_2}f(\omega_1,\omega_2)d\mu_2(\omega_2) \mbox{}\mbox{ }\mbox{ for all }\omega_1, \] so that the function \[ \omega_1\mapsto \int_{\Omega_2}f(\omega_1,\omega_2)d\mu_2(\omega_2) \] is \(\mathcal{F}_1\)-measurable, being a limit of \(\mathcal{F}_1\)-measurable functions ($ _{_2}f_n(_1,_2)d_2(_2)$이 \(\mathcal{F}_1\)-measurable이기 때문).

Thus by monotone convergence and the result for the simple function again, \[\begin{eqnarray*} \int_{\Omega_1\times\Omega_2} f(\omega_1,\omega_2)d(\mu_1\times\mu_2)(\omega_1, \omega_2)&=&\lim_n \int_{\Omega_1\times\Omega_2} f_n(\omega_1,\omega_2) d(\mu_1\times\mu_2)(\omega_1, \omega_2)\mbox{ }\mbox{ }\mbox{ by M.C.T}(0\le f_n\uparrow f, \mbox{ take integral w.r.t } \mu_1\times\mu_2) \\ &=& \lim_n \int_{\Omega_1}\left[\int_{\Omega_2} f_n(\omega_1,\omega_2) d\mu_2(\omega_2)\right]d\mu_1(\omega_1)\mbox{ }\mbox{ }\mbox{ }\mbox{ simple function }f_n\\ &=& \int_{\Omega_1}\left[\lim_n\int_{\Omega_2} f_n(\omega_1,\omega_2) d\mu_2(\omega_2)\right]d\mu_1(\omega_1)\\ &=&\int_{\Omega_1}\left[\int_{\Omega_2} f(\omega_1,\omega_2) d\mu_2(\omega_2)\right]d\mu_1(\omega_1)\mbox{ }\mbox{ }\mbox{ by M.C.T}(0\le f_n\uparrow f, \mbox{ take integral w.r.t } \mu_2). \end{eqnarray*}\]

Proof of Fubini’s Theorem(general \(f\))

Finally, suppose that \(f\) is a general \(\mathcal{F}_1\times\mathcal{F}_2\)-measurable function.

Then, by the result for nonnegative functions, the function \[ \omega_2\mapsto f^+(\omega_1,\omega_2) \] is \(\mathcal{F}_2\)-measurable for each fixed \(\omega_1\), and the function \[ \omega_1\mapsto \int_{\Omega_2}f^+(\omega_1,\omega_2)d\mu_2(\omega_2) \] is \(\mathcal{F}_1\)-measurable, and \[ \int_{\Omega_1\times\Omega_2} f^+(\omega_1,\omega_2)d(\mu_1\times\mu_2)(\omega_1, \omega_2)=\int_{\Omega_1}\left[\int_{\Omega_2} f^+(\omega_1,\omega_2) d\mu_2(\omega_2)\right]d\mu_1(\omega_1) \] Of course the analogous results hold for \(f^-\).

Now if \[ \int_{\Omega_1\times\Omega_2} f(\omega_1,\omega_2)d(\mu_1\times\mu_2)(\omega_1, \omega_2)=\int_{\Omega_1\times\Omega_2} f^+(\omega_1,\omega_2)d(\mu_1\times\mu_2)(\omega_1, \omega_2)-\int_{\Omega_1\times\Omega_2} f^-(\omega_1,\omega_2)d(\mu_1\times\mu_2)(\omega_1, \omega_2) \] exists(well-defined), then at least one of the integrals on the right-hand side must be finite. WLOG, suppose that the negative part is finite. Then, \[ \int_{\Omega_1\times\Omega_2} f^-(\omega_1,\omega_2)d(\mu_1\times\mu_2)(\omega_1, \omega_2)=\int_{\Omega_1}\left[\int_{\Omega_2} f^-(\omega_1,\omega_2) d\mu_2(\omega_2)\right]d\mu_1(\omega_1) \] which implies that \[ 0\le \int_{\Omega_2} f^-(\omega_1,\omega_2) d\mu_2(\omega_2) < \infty \mbox{ }\mbox{ }\mbox{ }\mu_1\mbox{-a.e. } \] by the \(5^{th}\) part of the Basic Properties of the Integral.

Thus, \[ \int_{\Omega_2} f(\omega_1,\omega_2)d\mu_2(\omega_2)=\int_{\Omega_2} f^+(\omega_1,\omega_2)d\mu_2(\omega_2)-\int_{\Omega_2} f^-(\omega_1,\omega_2)d\mu_2(\omega_2) \] is defined \(\mu_1\)-a.e., as claimed.

Thus, finally we have \[\begin{eqnarray*} \int_{\Omega_1}\left[\int_{\Omega_2} f(\omega_1,\omega_2) d\mu_2(\omega_2)\right]d\mu_1(\omega_1)&=&\int_{\Omega_1}\left[\int_{\Omega_2} f^+(\omega_1,\omega_2) d\mu_2(\omega_2)\right]d\mu_1(\omega_1) -\int_{\Omega_1}\left[\int_{\Omega_2} f^-(\omega_1,\omega_2) d\mu_2(\omega_2)\right]d\mu_1(\omega_1)\\ &=&\int_{\Omega_1\times\Omega_2} f^+(\omega_1,\omega_2)d(\mu_1\times\mu_2)(\omega_1, \omega_2)-\int_{\Omega_1\times\Omega_2} f^-(\omega_1,\omega_2)d(\mu_1\times\mu_2)(\omega_1, \omega_2)\\ &=&\int_{\Omega_1\times\Omega_2} f(\omega_1,\omega_2)d(\mu_1\times\mu_2)(\omega_1, \omega_2). \end{eqnarray*}\]



정리

Fubini’s theorem은 두 주어진 measurable spaces에서 \(f\)가 measurable-\(\mathcal{F}_1/\mathcal{F}_2\)일 때(가장 중요) product measure에 대한 indefinite integral이 각 measure에 대한 indefinite integral로 쪼개질 수 있고, 그 순서도 바뀔 수 있다는 것을 보여준다. 즉 \[\begin{eqnarray*} \int_{\Omega_1\times\Omega_2}f(\omega_1, \omega_2)\mbox{ }d(\mu_1\times\mu_2)(\omega_1, \omega_2)&=&\int_{\Omega_1}\left[\int_{\Omega_2}f(\omega_1,\omega_2)\mbox{ } d\mu_2(\omega_2)\right]d\mu_1(\omega_1) \\ &=&\int_{\Omega_2}\left[\int_{\Omega_1}f(\omega_1,\omega_2)\mbox{ } d\mu_1(\omega_1)\right]d\mu_2(\omega_2). \end{eqnarray*}\] 이고, 매우 많이 쓰이기 때문에 반드시 암기하자.



back