Given some definite integral over $[a,b]$ make the substitution $ u = x(x-a-b)$. The integral is then transformed to an integral over $[-ab,-ab]$ which is zero. What is wrong with this reasoning? When can and when can we not make a given u-substitution? Thanks.

  • 3
    $\begingroup$ If $f(x)$ can be written as $g(x(x-a-b))(2x-a-b)$ for some $g$, then the $\int_a^b f(x)\,dx$ is, in fact, zero. $\endgroup$ Aug 3 '17 at 0:34
  • 5
    $\begingroup$ You can only make the substitution $u=f(x)$ if $f(x)$ is one to one on the given domain of $u,x$. One usually fixes this by splitting the integral: $[a,b]=[a,(a+b)/2]\cup((a+b)/2,b]$, or some variation of this. $\endgroup$ Aug 3 '17 at 0:36
  • $\begingroup$ I'm talking about in general u-subsitutions? There are many more examples where the integral is "zero" when making a certain u-substituion. $\endgroup$ Aug 3 '17 at 0:36
  • 3
    $\begingroup$ @SimplyBeautifulArt Not exactly. We don't need 1-1, we just need the end points to match up. $\endgroup$
    – zhw.
    Aug 3 '17 at 0:38
  • 4
    $\begingroup$ @SimplyBeautifulArt If $f$ is continuous on $[a,b]$ and $g:[c,d]\to [a,b]$ is $C^1$ with $g(c) = a, g(d)=b,$ then $\int_a^b f(x)\,dx = \int_a^bf(g(t))\,g'(t)\,dt.$ $\endgroup$
    – zhw.
    Aug 3 '17 at 0:54

In integration theory when you want to use a u-subsitution you have to use a function which is a $C^1$-diffeomorphism. Otherwise you could end up with funny things.

Edit Actually a diffeomorphism is not necessary as pointed out by @zhw. To apply the substitution $x=\psi(u)$ from $u \in [c,d]$ to $x \in [a,b]$ you "only" need to have $\psi$ to be $C^1$ and $\psi([c,d])$ included in the domain of $f$, with $\psi(c)=a$ and $\psi(d)=b$.

For example if you want to compute $\int_{-1}^1 x^2\mathrm{d}x$, using a u-substitution $u=x^2$ will change the bounds to $1$ and $1$ and give 0, which is not true. This is because the function $x \in [-1,1] \mapsto x^2$ is not a bijection.

In your case your function $x \in [a, b] \mapsto x(x-a-b)$ has the following graph for $a=-1$ and $b=1$. You clearly see that this function is not 1-1. enter image description here

The fact that your substitution is not 1-1 is clearly due to the square. Let $\phi: x \in [a,b] \mapsto x(x-a-b)$. The function $\phi$ reaches its minimum in $c=(a+b)/2$ and is bijective on $[a, c]$ and on $[c, b]$. If you split your integral into 2 integrals on these intervals you can use your substitution.

You would get $\int_a^b f(x) \mathrm{d}x=\int_a^c f(x)\mathrm{d}x + \int_b^c f(x)\mathrm{d}x$

Now use $u=\phi(x)=x(x-a-b)$. You have therefore $x=\dfrac{a+b-\sqrt{(a+b)^2+4u}}{2}=\psi_1(u)$ on $[a,c]$ and $x=\dfrac{a+b+\sqrt{(a+b)^2+4u}}{2}=\psi_2(u)$ on $[c,b]$

This would give $$\int_a^b f(x) \mathrm{d}x=\int_{-ab}^{-(a+b)^2/4}f(\psi_1(u))\psi_1'(u)\mathrm{d}u + \int_{(a+b)^2/4}^{-ab}f(\psi_2(u))\psi_2'(u)\mathrm{d}u$$

And you see that you cannot "merge" these two integrals because their content is now different.

Of course this u-substitution does not seem to help a lot to compute the integral, particularly because $f$ is general. But if $f$ has a particular form, maybe this could help.

  • $\begingroup$ If you try $u=x^2$, the real problem is that $x^2\,dx = \pm 2\sqrt{u}\,du$. That $\pm$ is the key to the problem. $\endgroup$ Aug 3 '17 at 0:41
  • $\begingroup$ Or this is where you should break it up into $[-1,0]$ and $[0,1]$, intervals that are bijective. $\endgroup$ Aug 3 '17 at 0:43
  • $\begingroup$ Aherm, why not use $a=-1$ and $b=1$ to match up with your example? $\endgroup$ Aug 3 '17 at 0:46
  • 2
    $\begingroup$ You don't need a diffeomorphism, you just need the end points to match up. See my comment above. $\endgroup$
    – zhw.
    Aug 3 '17 at 0:57
  • 1
    $\begingroup$ You don' need $C^{1}$ either. $\psi'$ just needs to be Riemann integrable. $\endgroup$
    – AlohaSine
    Aug 3 '17 at 1:45

The only way you could do such a substitution to compute $\int_a^b f(x)\,dx$ is if there is some function $g$ such that: $$f(x)\,dx=g(u)\,du=g(x(x-a-b))(2x-a-b)\,dx$$

So if $f(x)=g(x(x-a-b))(2x-a-b)$, then let $h(x)=f\left(\frac{a+b}{2}+x\right)$, then $h$ is defined on $\left[-\frac{b-a}{2},\frac{b-a}{2}\right]$, and $h(-x)=-h(x)$. So you get that if $f(x)$ is of this form, then:

$$\int_{a}^{b} f(x)\,dx \int_{-\frac{b-a}{2}}^{\frac{b-a}{2}}h(t)\,dt= 0$$ since $h$ is an odd function.

So when the substitution can be done, you do get $0$. The problem is that, for the general function $f$, you can't find a function $g$ - you end up, instead, with a multi-valued function $g$.

Similarly, if you use $u=\sin x$, to calculate $\int_{0}^{\pi} f(x)\,dx$ then there should be a $g$ such that $f(x)=g(\sin x)\cos x$. And you can show that $$\int_{0}^{\pi}g(\sin x)\cos x\,dx = 0.$$ This is because $f(\pi-x)=-f(x)$.

More generally, if $g(x)$ is a continuous function and $u(x)$ a differentiable function and $u(a)=u(b)$ then any: $$\int_{a}^{b} g(u(x))u'(x)\,dx = 0$$

Proof: Let $G(x)=\int_{a}^{x} g(t)\,dt$. Then $g(u(x))u'(x)=\frac{d}{dx}(G(u(x))$ by the chain rule and the property that $G'=g$. So, by the Fundamental Theorem of Calculus, $$\int_{a}^{b} g(u(x))u'(x)\,dx = G(u(x))\big\vert_{a}^{b}=0$$

This is essentially the long form of substitution - substitution is doing the reverse of the chain rule.


If I use this example : $\displaystyle \int_{-1}^1x^2\cdot dx$

And let

$$\boxed{\begin{array}{ll}f:&\mathbb{R}_+&\to \mathbb{R}\\&X&\to\sqrt{\dfrac{X}{4}}\end{array}\qquad \text{AND }\qquad \begin{array}{ll}\varphi:&\mathbb{R}&\to \mathbb{R}_+\\&x&\to x^2\end{array}}$$

$f$ is continuous and $\varphi([-1,1])\subset \mathcal{D}_f=\mathbb{R}_+$ which are sufficient

$\begin{array}{ll}\displaystyle \int_{-1}^1x^2\cdot dx &=\displaystyle\int_{-1}^1 \sqrt{\dfrac{x^2}{4}}\cdot2|x|\cdot dx=\displaystyle\int_{-1}^1 f(\varphi(x))\;|\varphi'(x)|\cdot dx\\\\ &=\displaystyle\int_{-1}^0 \sqrt{\dfrac{x^2}{4}}\cdot2(-x)\cdot dx+\displaystyle\int_{0}^1 \sqrt{\dfrac{x^2}{4}}\cdot2x\cdot dx\\\\ &=\displaystyle\int_{0}^{-1} \sqrt{\dfrac{x^2}{4}}\cdot2x\cdot dx+\displaystyle\int_{0}^1 \sqrt{\dfrac{x^2}{4}}\cdot2x\cdot dx\qquad \textbf{We apply the substitution: }s=\varphi(x)\\\\ &=\displaystyle \int_{\varphi(0)}^{\varphi(-1)}f(s)\cdot ds \quad +\int_{\varphi(0)}^{\varphi(1)}f(s)\cdot ds =2\int_0^1\dfrac{\sqrt{s}}{2}\cdot ds =\dfrac{2}{3}\end{array}$

So $\varphi$ doesn't need to be a bijection


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.