2
$\begingroup$

I am having trouble proving the following result:

Let $X$ be a nonnegative random variable and $g:\mathbb{R}\rightarrow\mathbb{R}$ a nonnegative strictly increasing differentiable function. Then

$$\mathbb{E}g(X)=g(0)+\int_{0}^{\infty}g^{\prime}(x)\mathbb{P}(X>x)dx$$

I know that it should follow using integration by parts, but using integration by parts in the more abstract setting of probability is a bit confusing to me. Details would be appreciated.

$\endgroup$
2
$\begingroup$

$\mathbb E[g(X)] = \mathbb E[\int_0^{g(X)}dt] = \mathbb E[\int_{g(0)}^{g(X)}dt + \int_0^{g(0)} dt ] = \mathbb E[\int_{g(0)}^{g(X)}dt ] + \mathbb E[g(0)] = \mathbb E[\int_{g(0)}^{g(X)}dt ] + g(0) $

Now using the definition of expectation, we get: \begin{align*} \mathbb E\left[\int_{g(0)}^{g(X)}dt \right] &= \int_\Omega \int_{g(0)}^{g(X)}dt d\mathbb P(\omega) = \int_\Omega \int_0^\infty \chi_{(g(0),g(X(\omega))}(t)dtd\mathbb P(\omega)\\& = \int_0^\infty \int_\Omega \chi_{(g(0),g(X(\omega))}(t)d\mathbb P(\omega) dt =\int_0^\infty \mathbb P( g(0) < t <g(X)) dt \\&= \int_{g(0)}^\infty \mathbb P( g(0) < t <g(X)) dt \end{align*}

Use of Fubini due to all things being nonnegative (so we can swap order of integration).

Now, last thing $\mathbb P( t \in (g(0),g(X)) = \mathbb P( 0 <g^{-1}(t) < X) $

So, we get $\int_{g(0)}^\infty \mathbb P( g(0) < t <g(X)) dt = \int_{g(0)}^\infty \mathbb P(0 < g^{-1}(t) < X)dt = \int_0^\infty g'(s)\mathbb P( s < X) ds$

And we get $\mathbb E[g(X)] = g(0) + \int_0^\infty g'(s)\mathbb P(X>s)ds$

$\endgroup$
7
  • 1
    $\begingroup$ (+1)Nice answer using the change of variable in the last step to produce the derivative of $g$. $\endgroup$
    – Feng
    Aug 6 '19 at 2:25
  • $\begingroup$ I am with until this line: $\int_{g(0)}^{\infty}\mathbb{P}(0<g^{-1}(t)<X)dt=\int_{0}^{\infty}g^{\prime}(s)\mathbb{P}(s<X)ds$ There is a substitution happening here, but I don't see how you are getting from a preimage to a derivative. $\endgroup$ Aug 6 '19 at 2:27
  • $\begingroup$ Is the substitution $s=g^{-1}(t)$? $\endgroup$ Aug 6 '19 at 2:29
  • 1
    $\begingroup$ @RobertThingum Yes. It is the subsitution $s=g^{-1}(t)$. Then $t=g(s)$. $\endgroup$
    – Feng
    Aug 6 '19 at 2:52
  • 1
    $\begingroup$ @RobertThingum Yes, sorry. Maybe I should have explained more. $g^{-1}$ exists, cause $g$ is stricly monotone. If $c = \sup\{X(\omega) : \omega \in \Omega\}$, then $\int_{g(0)}^{\infty} \mathbb P( g^{-1}(t) \in (0,X) )dt = \int_{g(0)}^{g(c)} \mathbb P(g^{-1}(t) \in (0,X)) dt$. Now after substitution $t = g(s)$, we get $dt = g'(s)ds$ and limits of integral from $0$ to $c$, so $\int_0^c g'(s)\mathbb P( s \in (0,X))ds$. But $s>0$, so it's the same as $ \int_0^c g'(s) \mathbb P(s < X) ds$. As $c = \sup\{X(\omega) : \omega \in \Omega\}$, it holds that for $x>c: \mathbb P(x<X)=0$.Hencelimit $\infty$ $\endgroup$ Aug 6 '19 at 14:15
1
$\begingroup$

The general result is:

Claim: Let $g$ be differentiable. If $g$ and $g'$ are bounded, then $$ \bbox[5px,border:2px solid red] {E[g(X)]=g(0) +\int_0^\infty g'(x)P(X>x)\,dx-\int_{-\infty}^0 g'(x)P(X\le x)\, dx.}$$ The result also holds if $g$ is monotonic, provided the RHS is not $\infty-\infty$.

Proof: First suppose that $X$ is nonnegative. Write $$ g(X)-g(0)\stackrel{(1)}=\int_0^X g'(t)\,dt\stackrel{(2)}=\int_0^\infty g'(t) I_{X>t}\,dt.$$ Equality (1) is the fundamental theorem of calculus (remember $g$ is differentiable), while (2) is valid because the indicator random variable $I_{X>t}$ has value $1$ when $t<X$, and equals zero otherwise. Take expectation: $$ E[g(X)-g(0)]=E\left[\int_0^\infty g'(t) I_{X>t}\,dt\right]\stackrel{(3)}=\int_0^\infty g'(t)E[I_{X>t}]\,dt\stackrel{(4)}=\int_0^\infty g'(t)P(X>t)\,dt.$$ Identity (3) is the result of Fubini's theorem. In (4) we recognize that the expectation of the indicator of an event is the probability of the event.

Next, suppose $X$ is nonpositive. A similar argument shows $$E[g(X)-g(0)] = -\int_{-\infty}^0 g'(t)P(X\le t)\,dt.$$

For general $X$, write $$g(X)-g(0) = g(X^+)-g(0)+g(-X^-)-g(0)$$ where $X^+ := XI(X>0)$ is the positive part of $X$ and $X^-:=-XI(X<0)$ is the negative part. Apply the previous special cases to obtain $$ E[g(X^+) -g(0)]= \int_0^\infty g'(t)P(X^+>t)\,dt$$ and $$ E[g(-X^-)-g(0)]=-\int_{-\infty}^0 g'(t)P(-X^- \le t)\,dt.$$ To conclude, note that $\{X^+>t\}=\{X>t\}$ when $t>0$, and $\{-X^-\le t\} = \{X\le t\}$ when $t<0$.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.