When trying to solve $\int\frac{dx}{\sqrt{x^2-a^2}}$ we can substitue $x$ in two ways:

METHOD 1: $x=a\sec(\theta)$ and $dx=a\sec(\theta)\tan(\theta) d\theta$ and $\theta=\sec^{-1}(\frac{x}{a})$.

METHOD 2: $x=a\cosh(u)$ and $dx=a\sinh(u)du$ and $u=\cosh^{-1}(\frac{x}{a})$

My calculus-book states that we can use a similar method when we assume $x\le -a$.

I can see how that would work with method 1 because $\sec^{-1}(\frac{x}{a})$ is indeed defined for $x\le -a$.

QUESTION: But how would that work in method 2? $\cosh^{-1}(\frac{x}{a})$ is only defined for $\frac{x}{a}\ge1$ and if $x\le-a$ then $\frac{x}{a}$ is negative. So in this case $\cosh^{-1}(\frac{x}{a})$ is not defined!!!

What is happening here? Thanks!

  • $\begingroup$ Maybe it's just a typo and it should be $x=-a\cosh u$. $\endgroup$ – Saad Apr 20 '18 at 0:32
  • $\begingroup$ @AlexFrancisco I think I paraphrased my calculus-book wrongly. I edited the question. Your comment seems to imply we have different substitutions for different domains of $x$. Is this true? Why? How? $\endgroup$ – GambitSquared Apr 20 '18 at 7:10
  • $\begingroup$ Does your textbook ever invoke complex numbers? Also note it says "similar", meaning you might have to play around with signs or constants or whatever. $\endgroup$ – GFauxPas Apr 22 '18 at 19:09
  • $\begingroup$ @GFauxPas No complex numbers. Can we use different substitutions for different parts of the domain? $\endgroup$ – GambitSquared Apr 22 '18 at 20:11

Given that $a>0$ you cannot substitute $x=a \cosh u$ when you are interested in $x$-values $<-a$. In order to use the basic idea of method 2 also when $x<-a$ we just set up the substitution $$x:=-a\cosh u\quad (u>0),\qquad dx=-a\sinh u\>du\ ,$$ and obtain $$\eqalign{\int{dx\over\sqrt{x^2-a^2}}&=\int{-a\sinh u\>du\over a\sinh u}\Biggr|_{u:={\rm arcosh}(-x/a)}\cr &=(-u+C)\biggr|_{u:={\rm arcosh}(-x/a)}\cr &=-{\rm arcosh}{-x\over a}\qquad(x<-a)\ .\cr}$$ Note that this is not the same as the answer for $x>a$, since ${\rm arcosh}$ is not an odd function.

| cite | improve this answer | |
  • $\begingroup$ Which is the same as the answer for x>a isn't it? $\endgroup$ – GambitSquared Apr 23 '18 at 10:35

The method $2$ only works for $x > a$. That's why your book suggests there is a similar method for $x < -a$. As we shall see below, since the integrand's maximal real domain consists of the union of two disjoint intervals, it is necessary to separately consider each interval together with two different(but indeed similar) variable substitutions, one for each.

What does actually mean to "solve" $\int f(x) \,dx$? It means finding an "integral-free expression" for the function $$ F(y) = \int_r^y f(x) \,dx $$ where $r$ and $y$ are elements of $f$'s domain and $r$ is arbitrarily fixed.

The change of variables theorem says that if $\varphi : [p,q] \to [r,y]$ is a continuously differentiable bijective function and $f : [r,y] \to \mathbb{R}$ is a continuous function then $$ \int_{r}^{y} f(x) \,dx = \int_{\varphi{-1}(r)}^{\varphi^{-1}(y)} f(\varphi(u)) \varphi'(u) \,du \tag{1} $$

In this case we have $$ f : \left]-\infty, -a \right[ \cup \left]a, \infty \right[ \to \mathbb{R} : x \mapsto \frac{1}{\sqrt{x^2-a^2}} $$ where $a > 0$ and $\sqrt{\cdot}$ is the nonnegative real square root.

Since $f$'s domain is the union of two open intervals, at some moment we must choose whether $r, y \in \left]-\infty, -a \right[$ or $r, y \in \left]a, \infty \right[$ for we can't integrate over the gap between these two intervals because $f$ can't be defined or assumes complex values there.

By wanting to subtitute $$ \varphi : \left]-\infty, \infty \right[ \to \left[a, \infty \right[ : u \mapsto a \cosh(u) $$ we are forced to assume $r, y \in \left]a, \infty \right[$. Therefore consider $f$'s domain to be only the right interval: $\left]a, \infty \right[$. Furthermore this implies that at a first moment $F$ can only be defined on $\left]a, \infty \right[$.

To make $\varphi$ invertible, it is necessary to restrict its domain to either $\left]-\infty, 0 \right]$ or $\left[0, \infty \right[$. Without loss of generality choose $\left[0, \infty \right[$. Then its inverse is $$ \varphi^{-1} : \left[a, \infty \right[ \to \left[0, \infty \right[ : x \mapsto \cosh^{-1}\left(\frac{x}{a}\right) $$

Now that all the hypothesis of the theorem are satisfied, we can use $(1)$ to compute the integral on the right interval, $\left]a, \infty \right[$.

The similar method your book states to compute $F$ when $x < -a$ is to integrate over the left interval, namely $\left] -\infty, -a \right]$, and subtitute $-\varphi$ instead of $\varphi$ using the same reasoning as above, with the suitable modifications.

Let $\psi = -\varphi$. That is $$ \psi : \left]-\infty, \infty \right[ \to \left] -\infty, -a \right] : u \mapsto - a \cosh(u) $$ If we want to substitute $\psi$ we must assume that $s, y \in \left]-\infty, -a \right[$ instead of $\left]a, \infty \right[$, where $s$ is playing the role of $r$. Note that $s \neq r$ since they belong to diferent intervals. So in this case we restrict $f$'s domain to $\left]-\infty, -a \right[$ instead. To make $\psi$ invertible we also restrict its domain to $\left[0, \infty \right[$. So $$ \psi^{-1} : \left] -\infty, -a \right] \to \left[0, \infty \right[ : x \mapsto \cosh^{-1}\left(-\frac{x}{a}\right) $$

And thus by $(1)$ we are able to compute the integral $$ \int_s^y f(x) \,dx = \int_{\psi{-1}(p)}^{\psi^{-1}(q)} f(\psi(u)) \psi'(u) \,du $$ this time on the left interval $\left]-\infty, -a \right[$.

Hence we can extend $F$ to $f$'s entire domain by "gluing" the solutions as $$ F : \left]-\infty, -a \right[ \cup \left]a, \infty \right[ \to \mathbb{R} : y \mapsto \begin{cases} \int_r^y f(x) \,dx &, y \in \left]a, \infty \right[\\ \int_s^y f(x) \,dx &, y \in \left]-\infty, -a \right[ \end{cases}$$

Computing both integrals we obtain $$ F(y) = \begin{cases} \log \left(y + \sqrt{y^2-a^2}\right) + C_1(r) &, y \in \left]a, \infty \right[\\ -\log \left(-y + \sqrt{y^2-a^2}\right) + C_2(s) &, y \in \left]-\infty, -a \right[ \end{cases} $$

where $C_1(r)$ and $C_2(s)$ are constants that respectively depends on $r$ and $s$.

Which can also be written as $$ F(y) = \text{sgn}(y) \log \left( |y| + \sqrt{y^2-a^2} \right) + \frac{\text{sgn}(y) + 1}{2}C_1(r) + \frac{1 - \text{sgn}(y)}{2}C_2(s) $$ where $\text{sgn}$ is the sign function.

| cite | improve this answer | |
  • $\begingroup$ Isn't the last step too complicated? Can't we rewrite the solution for the case $x<-a$ to match the solution for the case $x>a$ (though with different C? $\endgroup$ – GambitSquared Apr 23 '18 at 10:44
  • $\begingroup$ @GambitSquared Yes, we can say that if $F$ is a solution for the case when $x>a$ and $G$ for the case when $x<-a$, then $G(x) = -F(-x) + C$ for some constant $C$. However this is the most general solution and the last expression is there just to show that the most general $F$ can be written as an one-liner expression. It is just a matter of a two case checking. If $x>a$, then $|x|=x$ and $\text{sgn}(x)=1$, otherwise, $x<−a$, then $|x|=−x$ and $\text{sgn}(x) = −1$. $\endgroup$ – mucciolo Apr 23 '18 at 17:30
  • $\begingroup$ Injectivity of $\phi$, interestingly, can be relaxed in 1-dimensional case. If you look at the classic proofs of integration by substitution, they don't invoke injectivity in the 1D case. $\endgroup$ – GFauxPas Apr 23 '18 at 17:54

You have to make the substitution $t=-x$. Then $\;\mathrm dx=-\mathrm dt$, and $\DeclareMathOperator{\acosh}{argcosh}$ \begin{alignat}{2} \int\frac{\mathrm dx}{\sqrt{x^2-a^2}}= -\int\frac{\mathrm dt}{\sqrt{t^2-a^2}}&=-\acosh\frac ta=-\acosh\frac{|x|}a \\ &=-\ln\Bigl(|x|+\sqrt{x^2-a^2}\Bigr)&\text{(modulo a constant term).} \end{alignat}

For method $1$, I don't see where it can lead, since the derivative of $\operatorname{arcsec} x$ is $$(\operatorname{arcsec})'(x)=\frac1{|x|\sqrt{x^2-1}}\ne\frac1{\sqrt{x^2-1}}.$$

| cite | improve this answer | |
  • $\begingroup$ This doesn't answer the question what is wrong with my way nor why it isn't a problem that arccosh is only defined for x>1 $\endgroup$ – GambitSquared Apr 17 '18 at 12:21
  • $\begingroup$ It is isn't a problem because the antiderivative exists, anyway, since the integrand is continuous. The only problem was to express it terms of $\;\operatorname{argcosh}$, not as. I'll add a word on your method. $\endgroup$ – Bernard Apr 17 '18 at 12:36
  • $\begingroup$ What do we need the derivative of arcsec for? We can just say: $dx=a\tan\theta\sec\theta d\theta$ $\endgroup$ – GambitSquared Apr 17 '18 at 12:46

For method 1 :

Let $x=asec(u)$, $dx=atan(u)sec(u)du$. Then $\sqrt{x^2-a^2}=\sqrt{a^2sec^2(u)-a^2}=atan(u)$ and $u=sec^{-1}(x/a)$. The integral then becomes $\int{sec(u)du}$. The trick is then to multiply numerator and denominator by $tan(u)+sec(u)$ and then do the change of variable (#another one) $s=tan(u)+sec(u)$ and $ds=(sec^2(u)+tan(u)sec(u))du$. Then you are left with $\int{\frac{1}{s}}ds$. You just need to subtitute back in all the change of variables we did. Note that $tan(arcsec(z))=\sqrt{1-\frac{1}{z^2}}z$.

| cite | improve this answer | |

Antiderivative exists also in the case $x<a.$

To show this, let us consider the definite integral $$J=\int\limits_{-3}^{-2}\dfrac{dx}{\sqrt{x^2-1}}.$$ The substitution $$x= \cosh u, \quad u = \cosh^{-1}x$$ allows to obtain the value $$J=\int\limits_{\cosh^{-1}(-3)}^{\cosh^{-1}(-2)}\ du = \cosh^{-1}(-2) - \cosh^{-1}(-3),$$ wherein $$\cosh^{-1}x = \ln(x+\sqrt{x^2-1}).$$ Easy to see that the imaginary phase additive can not change the real difference.

So $$\cosh^{-1}(-2) - \cosh^{-1}(-3) = \ln\dfrac{-2+\sqrt{2^2-1}}{-3+\sqrt{3^2-1}} = \ln\dfrac{2-\sqrt3}{3-\sqrt8} = \ln\dfrac{3+\sqrt8}{2+\sqrt3}\approx 0.445789$$ (see also Wolfram Alpha).

| cite | improve this answer | |
  • $\begingroup$ arccosh(-2) does not have a real solution, does it? wolframalpha.com/input/?i=arccosh(-2) $\endgroup$ – GambitSquared Apr 20 '18 at 21:30
  • $\begingroup$ @GambitSquared It has the imaginary phase additive $i\cdot\pi,$ such as $\cosh^{-1}(-3).$ Really, $\ln\ re^{i\phi} = ln|r| + i\phi. $ $\endgroup$ – Yuri Negometyanov Apr 20 '18 at 22:52
  • $\begingroup$ @GambitSquared That's why we can consider this phase additive as the imaginary part of integration constant. $\endgroup$ – Yuri Negometyanov Apr 24 '18 at 0:39

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.