3
$\begingroup$

This emerged while I was investigating this question, i.e. the solution to the definite integral $$I_x = \int_0^\infty\left(5x^5+x\right)\operatorname{erfc}\left(x^5+x\right)\,dx$$

In a comment, its value is given as 0.2119539... using the Maple solver, and I have verified this value by using another on-line definite integral calculator.

But manipulating the integral leads me to a different solution. Specifically,

define $y=h(x) = x^5 +x$, and we have $dy = h'dx$, with $h' = 5x^4+1$. Also, $h(x)$ admits an inverse, so $x = h^{-1}(y)$. Then

$$\int_0^\infty\left(5x^5+x\right)\operatorname{erfc}\left(x^5+x\right)\,dx = \int_0^\infty x\left(5x^4+1\right)\operatorname{erfc}\left(x^5+x\right)\,dx$$ $$= \int_0^\infty xh'\operatorname{erfc}\left(y\right)\,dx = \int_0^\infty h^{-1}(y)\operatorname{erfc}\left(y\right)\,dy$$

We also have $\operatorname{erfc}(y) = 1-\operatorname{erf}(y) = 1- \left(2\Phi(\sqrt 2 y)-1\right) \Rightarrow \operatorname{erfc}(y) = 2\Phi(-\sqrt 2 y)$ where $\Phi()$ is the standard normal cdf.

Then $$\int_0^\infty h^{-1}(y)\operatorname{erfc}\left(y\right)\,dy = \int_0^\infty h^{-1}(y)2\Phi(-\sqrt 2 y)\,dy = 2\int_0^\infty \int_{-\infty}^{-\sqrt 2 y}h^{-1}(y)\phi(t)\,dtdy$$ where $\phi()$ is the standard normal pdf. Interchange the limits of integration from

\begin{matrix} 0 \le y \le \infty\\ -\infty \le t \le - \sqrt 2 y \end{matrix}

to \begin{matrix} -\infty \le t \le 0\\ 0 \le y \le - t/\sqrt 2 \end{matrix} to obtain

$$I_x = 2\int_{-\infty}^0\phi(t) \int_{0}^{- t/\sqrt 2}h^{-1}(y)\,dydt$$

using inverse function integration we have (HERE COMES THE MISTAKE)

$$\int_{0}^{- t/\sqrt 2}h^{-1}(y)dy = xh(x)\Big |_0^{- t/\sqrt 2} - \int_{0}^{- t/\sqrt 2}h(x)dx $$ $$= (x^6 + x^2)\Big|_0^{- t/\sqrt 2} - \left((\frac 16 x^6 + \frac 12 x^2)\Big|_0^{- t/\sqrt 2}\right) = \left(\frac 56 x^6 + \frac 12 x^2\right)\Big|_0^{- t/\sqrt 2} = \frac {5}{48} t^6 + \frac 14 t^2$$

The correct expression for this case is $$\int_{0}^{- t/\sqrt 2}h^{-1}(y)dy = xh(x)\Big |_0^{h^{-1}(- t/\sqrt 2)} - \int_{0}^{h^{-1}(- t/\sqrt 2)}h(x)dx $$

Inserting into the integral we have

$$I_x = \frac {5}{48}\cdot 2\int_{-\infty}^0t^6\phi(t)dt + \frac {1}{4}\cdot 2\int_{-\infty}^0t^2\phi(t)dt$$

The integrals now represent even half-moments of the standard normal distribution, which are symmetric around zero and since they are multiplied by $2$ they equal whole moments. So $$I_x = \frac {5}{48}E_{\Phi}(t^6) + \frac {1}{4}E_{\Phi}(t^2) = \frac {5}{48}(3\cdot 5) + \frac {1}{4} \cdot 1 = \frac{25}{16} + \frac{4}{16} = \frac {29}{16} = 1.8125$$

Either there is a silly calculating mistake somewhere, or at some point I must be doing something illegal (or both).

I would greatly appreciate if somebody would point these to me, or just give me a hint.

$\endgroup$
2
$\begingroup$

You misapplied the inverse function integration. The identity is

$$\int f(x)\,dx = x\cdot f(x) - G(f(x)),$$

where $G$ is the indefinite integral of $f^{-1}$.

With $f = h^{-1}$, and $H(x) = \int h(t)\,dt = \frac16 x^6 + \frac12 x^2$, you get

$$\begin{align} \int_0^{-t/\sqrt{2}} h^{-1}(y)\,dy &= yh^{-1}(y)\Big\lvert_0^{-t/\sqrt{2}} - H(h^{-1}(y))\Big\lvert_0^{-t/\sqrt{2}}\\ &= -\frac{t}{\sqrt{2}}h^{-1}\left(-\frac{t}{\sqrt{2}}\right) - H\left(h^{-1}\left(-\frac{t}{\sqrt{2}}\right)\right), \end{align}$$

which I don't see yielding a nice expression.


In response to your comment, when you apply the inverse function integration with $f = h$, you still have - with $H' = h$ and $K' = h^{-1}$

$$H(x) = x\cdot h(x) - K(h(x)),$$

so in order to evaluate

$$\int_u^v h^{-1}(y)\,dy = K(v) - K(u)$$

you still must apply $h^{-1}$ to the bounds to obtain the values to plug into $x\cdot h(x)$ and $H$. That leads (unsurprisingly) to the same expression as above.

$\endgroup$
3
  • $\begingroup$ Thank you for your answer, I am still trying to figure this one out. However, I am applying the inverse function integration using $ f = h$. Then what I have is $G(f(x)) = G(y)$ and so I look at $G(f(x)) = G(y) = \int f^{-1}(y)\,dy = x\cdot f(x) - \int f(x)\,dx$. Is this not correct? (I have also corrected notation in the question to make this clear). One more try please! $\endgroup$ – Alecos Papadopoulos Jan 3 '14 at 15:39
  • $\begingroup$ But then your $x\cdot f(x)$ becomes $h^{-1}(y)\cdot y$, and you also must apply $h^{-1}$ to the limits of $\int h$. It leads to the same expression. $\endgroup$ – Daniel Fischer Jan 3 '14 at 16:04
  • $\begingroup$ Thank you. After my comment I have started to suspect that it has to be a matter of "internal consistency" and integration limits, in the application of inverse function integration -and now I see it. $\endgroup$ – Alecos Papadopoulos Jan 3 '14 at 17:12

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.