In probability, I have seen some examples for which both Fubini's theorem and integration by parts (for Riemann-Stieltjes integrals with cdf as integrator) provide different but correct solutions. For example
- In proving $E(|X|)=\int_0^\infty P(|X| > t)dt$, Edvin and Did used Fubini's theorem, while
Ben used integration by parts; - In proving $\operatorname{median}(X)$ solves $\min_{c \in \mathbb{R}} E |X-c|$, Did used Fubini's theorem, while Sivaram
used integration by parts in Edit.
So I wonder if the two are related somehow?
For example, in some cases (especially the two examples above), can one lead to the other?
-
A wide guess for going from Fubini's theorem to integration by parts is:
Integration by parts says
$$
\begin{align} f(b)g(b) – f(a)g(a) & = \int_a^b g(x) \, df(x) + \int_a^b f(x) \, dg(x). \end{align} $$If there is some $c \in \mathbb{R}$ such that $g(c)=0$, then $$
\int_a^b g(x) \, df(x) = \int_a^b \int_c^x dg(t) \, df(x ) $$ If
Fubini's theorem or some of its variants can apply, then for some $d
\in \mathbb{R}$, $$ \int_a^b \int_c^x dg(t) \, df(x ) = \int_c^b
\int_d^t df(x) \, dg(t ) = \int_c^b \int_d^x df(t) \, dg(x ) $$
one step closer to $\int_a^b f(x) \, dg(x)$, but still far away from
integration by parts. - No idea yet about going from integration by parts to Fubini's
theorem.
Best Answer
Fubini implies integration by parts. To see this in dimension $1$, consider some $C^1$ functions $f$ and $g$ and the integral $$ I=\int\limits_a^bf'(x)g(x)\mathrm dx. $$ Then, the fundamental theorem of calculus shows that, for every $x$, $$ g(x)=g(a)+\int\limits_a^xg'(t)\mathrm dt, $$ hence $I=J+K$ with $$ J=g(a)\int\limits_a^bf'(x)\mathrm dx=g(a)(f(b)-f(a)), $$ and, using Fubini theorem to justify the second equality sign below, $$ K=\int\limits_a^bf'(x)\int\limits_a^xg'(t)\mathrm dt\mathrm dx=\int\limits_a^bg'(t)\int\limits_t^bf'(x)\mathrm dx\mathrm dt=\int\limits_a^bg'(t)(f(b)-f(t))\mathrm dt. $$ One sees that $K=L-M$ with $$ L=f(b)\int\limits_a^bg'(t)\mathrm dt=f(b)(g(b)-g(a)),\qquad M=\int\limits_a^bg'(t)f(t)\mathrm dt. $$ Finally, $I=(J+L)-M$ with $$ J+L=g(a)(f(b)-f(a))+f(b)(g(b)-g(a))=f(b)g(b)-f(a)g(a). $$