8

How one can find all functions $f$ that there exists $f'$ satisfying $(f(x))^2+(f'(x))^2=1$? I just found solutions $f(x)=1,-1,\sin x$.

curious
  • 91
  • 3
    $\cos(x)$ is also a solution – Sanderr Aug 10 '16 at 11:48
  • 3
    And $\sin(x+\alpha)$ for all $\alpha\in\Bbb R$. I think it's a bit sketchy, because you may as well pick $$f(x)=\begin{cases}\sin x&\text{if }x<\frac\pi2\ 1&\text{if }x\ge \frac\pi2\end{cases}$$ –  Aug 10 '16 at 11:49
  • $sin(x+c)$ is another solution – Zau Aug 10 '16 at 11:51

5 Answers5

16

enter image description here

We are going to obtain in two steps all $C^1$ solutions of

$$\tag{0}(f(x))^2+(f'(x))^2=1.$$

Step 1: Let us follow a method similar to that given either by @David Quinn for example or @Ian Eerland or @Battani, with some supplementary precision on the intervals of validity.

Let $f$ be a solution to $(0)$.

Let us consider a point $x_0$. Either

  • case 1: $|f(x_0)|=1$ or

  • case 2: $|f(x_0)|<1$, which implies $|f'(x_0)|>0$.

Let us detail case 2.

By continuity of $f$, there exist $\epsilon_1>0$ such that on $I_1:=(x_0-\epsilon_1,x_0+\epsilon_1)$ we have $|f(x)|<1$.

By continuity of $f'$, there exist $\epsilon_2>0$ such that on $I_2:=(x_0-\epsilon_2,x_0+\epsilon_2)$ we have $|f'(x)|>0$.

Let us transform $(0)$ into

$$\tag{1}\dfrac{y'}{\sqrt{1-y^2}}=s$$

Differential equation (1) needs to be considered for values of variable $x$ in $I:=I_1 \cap I_2$ in order that the ratio is well defined and that the sign $s=\pm1$ doesn't change throughout this interval $I$.

Taking a primitive function on $I$ of both sides of (1):

$$\arcsin(y)=s x+\varphi \ \ \ \text{for some constant} \ \varphi.$$

Thus $y=f(x)=\sin(s x+ \varphi)$ on interval $I$ with either $s=1$ or $s=-1$.

This solution can be presented under the simpler form: $f(x)=\sin(x+ \varphi)$ (one finds back case $s=-1$ through replacement of $\varphi$ by $\varphi+\pi$).

Conclusion of step 1: a solution $f$ and a particular point $x_0$ being given, there are two cases:

\begin{cases} (a) \ \text{Either} \ f(x_0) = \pm1 \ \text{or}\\ (b) \ \text{There exist an interval} \ J \ \text{containing} \ x_0 \ \text{ such that,} \forall x \in J, \ f(x)=\sin(x+\varphi) \ \end{cases}

for a certain $\varphi$ ($J$ being any interval of validity of solution $f$, including interval $I$ as defined before).

Step 2: We are going to obtain these intervals $J$ by using a topological argument (A recall about topology can be found there). From (0), we know that the values taken by $f(x)$ are in the closed interval $[-1,1]$. Now consider the set of values of $x$ such that their image is in the open interval $(-1,1)$:

$$S:=\{x \ | -1<f(x)<1 \}=f^{-1}((0,1)). $$

$S$ is an open set, being the preimage of an open set by a continuous function; as such, it can be written as a countable union of open intervals $S=\bigcup_k (a_k,b_{k})$ (see this). In this way ;

(b) will be valid on open intervals $(a_k,b_k)$

(a) will be valid on closed intervals $[b_k,a_{k+1}]$.

At junction point $b_k$, the only possible way to preserve $C^1$ continuity is by taking, according to the value of $\lim_{x\rightarrow b_k} f(x)$ value $1$ or $-1$ for the whole closed interval $[b_k,a_{k+1}]$. Same procedure at the initial junction point $a_k$.

Let the curve of $y=sin(x)$ restricted

  • to $[0,\pi/2]$ (increasing) be called an $s_+$ arc.

  • to $[\pi/2,\pi]$ (decreasing) be called an $s_-$ arc.

Let a line segment of any positive length with equation $y=1$ (resp. $y=-1$) be called a $1_+$ (resp a $1_-$).

Thus one can describe the possible $C^1$ integral curves of (0) by a $C^1$ "glueing" of curves as a (finite or infinite) repetition of the following "motives":

$$\cases{1_+^*[s_-1_-^*s_+1_+^*]^*\\(1_-)^*[s_+1_+^*s_-1_+^*]^*}$$

(mathematical "grammars" notations, where exponant * means 0 or more times). See figure.

  • Intervals where the solution is constant may have any length.

  • "Phase shifts" $\varphi_k$ of the sine curves are specific to each interval.

  • It is to be noted that closed intervals $[b_k,a_{k+1}]$ can be reduced to a point.

Appendix : I had, at first, obtained a certain family of solutions by differentiating both sides of $(0)$:

$$2f'(x)(f(x)+f''(x))=0.$$

After a certain number of hours spent on this method, which assumes $f \in C^2$, I realized that it is impossible to go back to the initial equation dealing with $C^1$ solutions.

Added on August 30th: this differential geometry question gives an application of (0).

Jean Marie
  • 81,803
  • 1
    (I think) This also hints why piecewise-defined counterexamples can be found only for the $C^1$ case. (+1) –  Aug 10 '16 at 12:30
  • @G. Sassatelli. Yes. I just provided through a graphic example an idea of the huge variety of $C^1$ solutions. – Jean Marie Aug 10 '16 at 12:54
  • You seem to be saying that $\forall x, 2f'(x)(f(x)+f''(x))=0$ implies either $\forall x,f'(x)=0$ or $\forall x, f(x)+f''(x)=0$. How is that true? – xavierm02 Aug 10 '16 at 15:16
  • You are right. I must correct this flaw. As I proceed by necessary condition, it may open the door to (many) other solutions. – Jean Marie Aug 10 '16 at 15:58
  • @xavierm02 Correction done (see the end). Finally, there are "new" solutions but are not that different. Thank you again. – Jean Marie Aug 10 '16 at 18:29
  • @Mehrdad Yes "deriving" is surely preferable. – Jean Marie Aug 10 '16 at 18:32
  • 1
    @Mehrdad Thanks ! I am a difficult student ! I have indeed problems with this word (in French : "dériver" is one thing, "différentier" another) – Jean Marie Aug 10 '16 at 19:03
  • 1
    @xavierm02 I think that after many successive versions I arrived at a final stage that looks rigorous. – Jean Marie Aug 10 '16 at 19:33
  • @JeanMarie: It's okay, even native speakers mess it up! – user541686 Aug 10 '16 at 21:11
  • @Martin Argerami Thanks for editing – Jean Marie Aug 10 '16 at 21:13
  • It does look rigorous. I'd just rename your "Global solutions" part to "Simple solutions", "First solutions" or something similar, to avoid making the reader think that the other solutions will be local. – xavierm02 Aug 10 '16 at 22:40
  • I don't understand the "but with the difference that now the intervals where the solution is constant may have any length (i.e. are not restricted to have a length multiple of $\pi$)" part. Why are the piecewise-defined functions of the previous part restricted to lengths that are a multiple of $\pi$ for the constant parts? – xavierm02 Aug 10 '16 at 22:41
  • @xavierm02 Answers to your 2 questions 1) I agree with a transformation of the title 2) It is true that the previous part could be cancelled because, as you say it, one could have guessed, already at that stage, that there exist a more general solution (I didn't...). I kept it because it was a milestone in my understanding of the problem, but now it is more an obstacle than an aid... Thanks again for your very constructive interaction. I will be absent now for some hours; I have to sleep a while and it's late (1:00 CET) – Jean Marie Aug 10 '16 at 22:59
  • @xavierm02 I have "revamped" my presentation 1) by placing its original contribution (reasoning on $S$ and $\bar{S}$) in the forefront, and 2) by reducing to a small heuristic addendum the form $2f(x)(f(x)+f''(x))=0$, which has a majoi drawback : being obliged to assume $f \in C^2$. – Jean Marie Aug 11 '16 at 21:40
  • @xavierm02 still a big change (this time in the first part): the "step one" and "step 2" appear now in a natural setting. May I ask you your opinion ? – Jean Marie Aug 12 '16 at 07:30
  • @JeanMarie I think it's converging towards perfection :-) | You have a few $phi$ instead of $\phi$. | I'd be tempted to use $S^c$ instead of $\overline{S}$ because the latter is (in a topological context) sometimes used to denote the closure. | You say that you are going to obtain the maximal intervals $J$ but the $J_k$ are not maximal (because, as you've noted, the interval where $(a)$ is valid can be reduced to a point, and in this case, $(b)$ is valid on the two intervals and the point between them). – xavierm02 Aug 13 '16 at 12:27
  • @xavierm02 Thank you very much for your nice appreciation and remarks (I will take into account). I thought also on my side yesterday that "maximal" is not at all adapted. Thank you again for your fundamental interaction that has allowed to transform to such an extend an akward (and inexact) argumentation. – Jean Marie Aug 13 '16 at 12:39
  • @xavierm02 If you are still interested in this issue, I have added in my text a connection of the differential equation $f^2+f' ^2=1$ with a problem of differential geometry. – Jean Marie Aug 30 '16 at 05:05
  • May ask you, "curious", why were you interested by this question ? Because (see the last sentence I just added to my answer) it seems there are connections of this issue with differential geometry. – Jean Marie Aug 30 '16 at 06:01
3

If $y=f(x)$, you have $$\frac{dy}{dx}=\pm\sqrt{1-y^2}$$ Either $y=\pm 1$, or if $y\neq \pm 1$, $$\Rightarrow\int\frac{1}{\sqrt{1-y^2}}dy=\pm x+c$$ $$\Rightarrow y=\sin(\pm x+c)$$

David Quinn
  • 34,121
2

$$f(x)^2+f'(x)^2=1\Longleftrightarrow$$ $$f'(x)^2=1-f(x)^2\Longleftrightarrow$$ $$f'(x)=\pm\sqrt{1-f(x)^2}\Longleftrightarrow$$ $$\frac{f'(x)}{\sqrt{1-f(x)^2}}=\pm1\Longleftrightarrow$$ $$\int\frac{f'(x)}{\sqrt{1-f(x)^2}}\space\text{d}x=\int\pm1\space\text{d}x\Longleftrightarrow$$


For, the integral on the LHS substitute $u=f(x)$ and $\text{d}u=f'(x)\space\text{d}x$:

$$\int\frac{f'(x)}{\sqrt{1-f(x)^2}}\space\text{d}x=\int\frac{1}{\sqrt{1-u^2}}\space\text{d}u=\arcsin(u)+\text{C}=\arcsin(f(x))+\text{C}$$

For, the integral on the RHS:

$$\int\pm1\space\text{d}x=\pm\int1\space\text{d}x=\pm x+\text{C}=\text{C}\pm x$$


$$\arcsin(f(x))=\text{C}\pm x$$

Your solutions, can be found by setting $\text{C}$, when we have $f(a)=b$:

$$\arcsin(b)=\text{C}\pm a\Longleftrightarrow\text{C}=\arcsin(b)\mp a$$

So:

$$\arcsin(f(x))=\left(\arcsin(b)\mp a\right)\pm x$$

Jan Eerland
  • 28,671
1

$$ y^2+y^{\prime^2}=1 $$

Differentiate to get

$$ 2 y y^{\prime} + 2 y^{\prime} y^{\prime \prime } =0$$

Factorize, two different/separate solutions

$$ y^{\prime} = 0 , \quad y = C $$

$$ y^{\prime \prime } + y = 0, $$

with particular solution non-homogeneous DE

$$ y = A \sin (x + B) +1 $$

First solution

$$ C= 1, \quad \rightarrow y=1 ; $$

For second solution when plugged into given equation $ B=0, A=1$,

so that only $ y= 1+ \sin x $ is admissible.

EDIT1

Squaring introduces extraneous solutions. But how an elementary direct solution got missed!

$$ y^2+y^{\prime^2}=1 ,\quad y^{\prime }= \pm \sqrt{1-y^2} $$

$$\pm \dfrac{dy}{\sqrt{1-y^2}}= dx $$

$$\pm \int\dfrac{dy}{\sqrt{1-y^2}}= x +C = \pm \sin^{-1}y $$

$$ y = \pm \sin(x+C) $$

which includes all solutions.

CHECK: $ \, y^{\prime} = \pm \cos(x+C),\quad y^2+y^{\prime^2}=1. $

and it includes both the envelopes mentioned by OP:

$$ y= \pm 1 $$

And I believe this envelope solution aspect of differential equations theory is often discarded as being trivial, which definitely it is not.

Narasimham
  • 40,495
0

$$(f(x))^{ 2 }+(f'(x))^{ 2 }=1\\ { y }^{ 2 }+{ \left( y\prime \right) }^{ 2 }=1\\ y\prime =\pm \sqrt { 1-{ y }^{ 2 } } \\ \int { \frac { dy }{ \sqrt { 1-{ y }^{ 2 } } } =\pm \int { dx } } \\ \arcsin { y } =\pm x+C\\ y=\sin { \left( \pm x+C \right) } $$

haqnatural
  • 21,578