Take the 2-minute tour ×
Mathematics Stack Exchange is a question and answer site for people studying math at any level and professionals in related fields. It's 100% free, no registration required.

In the course of my research I came across the following integral:


where $$\operatorname{erf}(x)=\frac{2}{\sqrt{\pi}}\int_{0}^{x}e^{-t^2}dt$$ is the familiar error function.

Does anyone know if this is solvable? If so, a suggestion on how to do this would be very much appreciated. I can tolerate a solution in terms of $\operatorname{erf}$ functions of $a$. I have tried my usual sources in these matters (G&R, the tables on the Wolfram website, Abramowitz's Handbook of Mathematical Formulas, the web) but couldn't find anything of use.

Thank you!

share|improve this question
Perhaps you are interested in $$\int_{-\infty}^\infty \left( \text{erf}(a+x) \text{erf}(a-x)+1 \right)\ dx = \frac{2 \sqrt{2}}{\sqrt{\pi}} e^{-2 a^2} + 4 a\ \text{erf}(\sqrt{2} a)$$ –  Robert Israel Sep 9 '11 at 5:45
This came from trying to find an expression the differential entropy of a random variable that is distributed as the sum of a uniform random variable $\mathcal{U}[-\delta,\delta]$ and a zero-mean Gaussian. I would be satisfied to find a meaningful lower bound on differential entropy. Perhaps I should ask about that in a separate question? –  M.B.M. Sep 9 '11 at 20:12
Let $X∼\mathcal{U}[-\delta,\delta]$ and $Y \sim \mathcal{N}(0,\sigma^2)$ be independent. By the entropy-power inequality (lower-bound) and the maximum-entropy property of the normal distribution (upper bound), we get that $$\frac{1}{2}\log_2(4 \delta^2 + 2\pi e\sigma^2) \leq h(X+Y) \leq \frac{1}{2} \log_2( 2\pi e \delta^2 / 3 + 2\pi e \sigma^2) \> .$$ How much tighter of a bound do you need? (NB: $2 \pi e /3 \approx 5.693$.) –  cardinal Sep 10 '11 at 5:54
The integral you request does not converge. $\lim\limits_{x\to-\infty}\operatorname{erf}(x)=-1$ and $\lim\limits_{x\to\infty}\operatorname{erf}(x)=1$, thus when $|x|$ is big, the product $\operatorname{erf}(a-x)\operatorname{erf}(a+x)\cong-1$. What you must want is the integral suggested by Robert Israel. –  robjohn Sep 10 '11 at 9:06
@cardinal: Special thank you with help on the lower bound for differential entropy of the sum of a uniform and a Gaussian! I did not know about the entropy-power inequality, so, thank you for adding that tool to my arsenal! –  M.B.M. Sep 12 '11 at 0:13
show 1 more comment

2 Answers

up vote 8 down vote accepted

$\operatorname{erf}(x)$ is an odd function, therefore, $$ \begin{align} \int_{-\infty}^\infty(\operatorname{erf}(a+x)+\operatorname{erf}(a-x))\;\mathrm{d}x &=\lim_{L\to\infty}\;\int_{-L}^L(\operatorname{erf}(x+a)-\operatorname{erf}(x-a))\;\mathrm{d}x\\ &=\lim_{L\to\infty}\;\int_{-L+a}^{L+a}\operatorname{erf}(x)\;\mathrm{d}x-\lim_{L\to\infty}\;\int_{-L-a}^{L-a}\operatorname{erf}(x)\;\mathrm{d}x\\ &=\lim_{L\to\infty}\;\int_{L-a}^{L+a}\operatorname{erf}(x)\;\mathrm{d}x-\lim_{L\to\infty}\;\int_{-L-a}^{-L+a}\operatorname{erf}(x)\;\mathrm{d}x\\ &=4a\tag{1} \end{align} $$ since $\lim\limits_{x\to\infty}\operatorname{erf}(x)=1$ and $\lim\limits_{x\to-\infty}\operatorname{erf}(x)=-1$.

Furthermore, $$ \begin{align} \int_{-\infty}^\infty\left(\operatorname{erf}(a+x)\operatorname{erf}(a-x)+1 \right)\;\mathrm{d}x &=\int_{-\infty}^\infty(\operatorname{erf}(a+x)+1)(\operatorname{erf}(a-x)+1)\;\mathrm{d}x\\ &-\int_{-\infty}^\infty(\operatorname{erf}(a+x)+\operatorname{erf}(a-x))\;\mathrm{d}x\tag{2} \end{align} $$ To evaluate $$ \begin{align} \int_{-\infty}^\infty(\operatorname{erf}(a+x)+1)(\operatorname{erf}(a-x)+1)\;\mathrm{d}x &=\frac{4}{\pi}\int_{-\infty}^\infty\int_{-\infty}^{a+x}\int_{-\infty}^{a-x}e^{-s^2-t^2}\;\mathrm{d}s\;\mathrm{d}t\;\mathrm{d}x \end{align} $$ note that $s\le a+x$ and $t\le a-x$; i.e. $s-a\le x\le a-t$ and $s+t\le2a$. Thus, $$ \begin{align} \frac{4}{\pi}\int_{-\infty}^\infty\int_{-\infty}^{a+x}\int_{-\infty}^{a-x}e^{-s^2-t^2}\;\mathrm{d}s\;\mathrm{d}t\;\mathrm{d}x &=\frac{4}{\pi}\int\int_{s+t\le2a}\int_{s-a}^{a-t}e^{-s^2-t^2}\;\mathrm{d}x\;\mathrm{d}s\;\mathrm{d}t\\ &=\frac{4}{\pi}\int_{-\infty}^\infty\int_{-\infty}^\infty(2a-s-t)_+\;e^{-s^2-t^2}\;\mathrm{d}s\;\mathrm{d}t \end{align} $$ Change variables: $u=(s+t)/\sqrt{2}$ and $v=(s-t)/\sqrt{2}$ so that $s=(u+v)/\sqrt{2}$ and $t=(u-v)/\sqrt{2}$: $$ \begin{align} \frac{4}{\pi}\int_{-\infty}^\infty\int_{-\infty}^\infty(2a-s-t)_+\;e^{-s^2-t^2}\;\mathrm{d}s\;\mathrm{d}t &=\frac{4}{\pi}\int_{-\infty}^\infty\int_{-\infty}^\infty(2a-\sqrt{2}u)_+\;e^{-u^2-v^2}\;\mathrm{d}u\;\mathrm{d}v\\ &=\frac{4}{\sqrt{\pi}}\int_{-\infty}^{\sqrt{2}a}(2a-\sqrt{2}u)\;e^{-u^2}\;\mathrm{d}u\\ &=4a(\operatorname{erf}(\sqrt{2}a)+1)-\frac{4}{\sqrt{\pi}}\int_{-\infty}^{\sqrt{2}a}\sqrt{2}u\;e^{-u^2}\;\mathrm{d}u\\ &=4a(\operatorname{erf}(\sqrt{2}a)+1)-\frac{2\sqrt{2}}{\sqrt{\pi}}\int_{-\infty}^{\sqrt{2}a}\;e^{-u^2}\;\mathrm{d}u^2\\ &=4a(\operatorname{erf}(\sqrt{2}a)+1)+\frac{2\sqrt{2}}{\sqrt{\pi}}\;e^{-2a^2} \end{align} $$ Therefore, $$ \int_{-\infty}^\infty\left(\operatorname{erf}(a+x)+1\right)\left(\operatorname{erf}(a-x)+1\right)\;\mathrm{d}x =4a\left(\operatorname{erf}(\sqrt{2}a)+1\right)+\frac{2\sqrt{2}}{\sqrt{\pi}}\;e^{-2a^2}\tag{3} $$ Thus, the convolution of $\operatorname{erf}(x)+1$ with itself is $2x(\operatorname{erf}(x/\sqrt{2})+1)+\frac{2\sqrt{2}}{\sqrt{\pi}}e^{-x^2/2}$.

Subtract $4a$ from $(3)$ using $(1)$ and $(2)$ to get $$ \int_{-\infty}^\infty\left(\operatorname{erf}(a+x)\operatorname{erf}(a-x)+1 \right)\;\mathrm{d}x =4a\operatorname{erf}(\sqrt{2}a)+\frac{2\sqrt{2}}{\sqrt{\pi}}\;e^{-2a^2}\tag{4} $$ My guess is you want either $(3)$ or $(4)$.

share|improve this answer
Thank you for the neat derivation of an interesting definite integral -- and, yes, as you pointed out, the original integral does diverge. However, cardinal's comment above helped me solve the underlying problem. Thank you all for your help with this! –  M.B.M. Sep 12 '11 at 0:17
add comment

$\int_{-\infty}^\infty 4/\pi \int_0^{a+x} e^{-t^{2}} dx $ . $ \int_0^{a-x} e^{-t^{2}} $ dx $= \int_{-\infty}^\infty 4/\pi [-(1/2t) e^{-t^{2}}]_0^{a+x} [-(1/2t) e^{-t^{2}}]_0^{a-x}$ = $ \int_{-\infty}^\infty (4/\pi) ((1/2a) e^{-a^{2}})^{2} - ((1/2x) e^{-x^{2}})^{2} $ = $[-((1/2x) e^{-x^{2}})^{2}]_{-\infty}^\infty$ = $\int_{-\infty}^\infty -((1/2x) e^{-x^{2}})^{2}$= $\int_{-\infty}^\infty -((1/4x^{2}) e^{-2x^{2}})] $ = $ [-(1/4x) e^{-2x^{2}}]_{-\infty}^\infty - \int_{-\infty}^\infty e^{-2x^{2}}$ = $[-(1/4x) e^{-2x^{2}}$+$(1/4x) e^{-2x^{2}}]_{-\infty}^\infty$ =0

share|improve this answer
Um, what? The antiderivative of $e^{-t^2}$ is not $\frac{-1}{2t} e^{-t^2}$. –  Rahul Sep 10 '11 at 5:41
@Angela, at the very least, it would be good of you to rewrite what you've posted in a (more) readable format. At present, it reminds me of this. –  cardinal Sep 10 '11 at 5:51
add comment

Your Answer


By posting your answer, you agree to the privacy policy and terms of service.

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