27

Let $f\in C^{3}[0,1]$such that $f(0)=f'(0)=f(1)=0$ and $\big|f''' (x)\big|\le 1$.Prove that $$\left|30240\int_{0}^{1}x(1-x)f(x)f'(x)dx\right|\le1 .$$

I couldn't make much progress on this problem. I thought that maybe I should try using polynomial interpolation since I have a bound for $|f'''|$, but I can't determine the interpolation polynomial and I am quickly stuck (there is also the problem that I am dealing with both $f$ and $f'$ under the integral). Apart from this, I don't think that there is much we can do, the solution probably relies on this technique, but I can't make further progress.

JustAnAmateur
  • 801
  • 5
  • 10
  • Try integration by parts – Integrand Aug 07 '20 at 13:13
  • 2
    @Integrand I most certainly would, but the result looks horrendous to me. Do you have any specific idea? – JustAnAmateur Aug 07 '20 at 13:20
  • 2
    Integrating by parts once (to get $f^2$ in the integrand instead of having $ff'$) and then use polynomial interpolation on $f$ you can easily get bounds like 5/110592/5 or 101/2580480 instead of 1/30240, but it is also clear those are unattainable. I can't think of a way to get 30240 at this point. – user10354138 Aug 07 '20 at 15:58
  • 5
    at least the equality is attained when $f(x) = \pm\dfrac{x^3-x^2}{6}.$ I will try to work out a solution in a couple days. – dezdichado Aug 09 '20 at 20:56
  • 1
    Another question: https://artofproblemsolving.com/community/c7h1951001p13467104 – River Li Aug 10 '20 at 03:13
  • if we add the condition $f'(1)=0$ the problem does not change under the substitution $x=1-y$ – Erik Satie Aug 10 '20 at 15:50
  • Following dezdichado I think that $\frac{x^n-x^{n-1}}{n!}$ is valid .Maybe it's the only one . – Erik Satie Aug 10 '20 at 16:21
  • Well there is another one $\frac{x\sin(\pi x)}{32}=f(x)$ – Erik Satie Aug 10 '20 at 16:41
  • I can only claim that $x (1 - x) f(x) f'(x) = o(x^2)$ as $x \to 0$ by Taylor's theorem with the Peano form of the remainder. The given conditions $f(1) = 0$ and $|f'''(x)| \le 1$ are not used. Can this direction be helpful ? – taiwanjizhan Aug 10 '20 at 19:43
  • Prove this for polynomials, then use Stone Weierstrass to approximate and obtain your result, perhaps (in which case the upper bound will be given exactly by @dezdichado's suggestion) – Romain S Aug 10 '20 at 22:00
  • Also, @user698573, I don't think that either of your suggestions actually reach equality (via Mathematica) – Romain S Aug 10 '20 at 22:01
  • Just out of interest where did you find this problem? –  Aug 16 '20 at 10:49
  • @user I found it here https://artofproblemsolving.com/community/u150054h2090303p15104143, but since there was no solution I decided to post it here too. I also tried contacting the OP from AoPS, but he hasn't responded yet. – JustAnAmateur Aug 16 '20 at 12:00
  • 1
    @JustAnAmateur Many thanks! –  Aug 16 '20 at 12:05

4 Answers4

10

At first,

$$I=\int_0^1x(1-x)f(x)f'(x)\text{d}x \,\overset{IBP}{=\!=\!=}\,\frac12x(1-x)f^2(x)\bigg|_0^1+\frac12\int_0^1 (2x-1)f^2(x)\text{ d}x,$$ $$I=\frac12\int_0^1 (2x-1)f^2(x)\text{ d}x.\tag1$$

The given conditions mean that $f(x)$ has a root of multiplicity $2$ at $x=0$ and a simple root at $x=1.$

Taking in account the Besou theorem and assuming the condition $|f'''(x)| = 1,$ one can get $$|f_{opt}(x)| = a(x-0)^2(1-x) = a(x^2-x^3),\quad a = \frac16,\tag2$$

$$|I|\le\dfrac1{72}\int_0^1(2x-1)(x^2-x^3)^2\text{ d}x = \dfrac1{30240}.$$

$\color{green}{\textbf{EDIT of 12.08.20.}}$

Let us integrate the given inequality $$-1\le f'''(x) \le 1\tag{1n}$$ under the given conditions on the interval $(0,x),$ then $$ \begin{cases} -x\le f''(x) - f''(0)\le x\\[4pt] -\frac12x^2 \le f'(x) - xf''\left(0\right) \le \frac12x^2\\[4pt] -\frac16x^3 \le f(x) - \frac12x^2f''\left(0\right) \le \frac16x^3._{\Large\mathstrut} \end{cases}\tag{2n}$$

From $(2n.3)$ should $$|6f(x)-3f''(0)x^2| \le x^3,\quad -1\le -3f''(0)\le1,$$ $$6|f(x)|\le |x^3+ax^2|,\quad |a|\le 1,\quad f(1)=0,$$ $$6|f(x)|\le x^2-x^3 = h(x).$$

This confirms formulas $(2).$

Similarly, from $(2n.2)$ $$|f'(x)| \le \frac12 x^2 -\frac13x = h'(x),$$ wherein $h'(x)$ is synchronized with $h(x).$

Therefore, $$|I| \le \left|\int_0^1 x(1-x) h(x) h'(x) \text{ d}x\right| = \frac1{30240}.$$

Yuri Negometyanov
  • 27,332
  • 1
  • 29
  • 59
  • Thank you! Could you please elaborate on how you came up with (2)? I initially thought that you used the estimate for the interpolation error from here https://en.wikipedia.org/wiki/Polynomial_interpolation#Interpolation_error, but it doesn't look like it – JustAnAmateur Aug 10 '20 at 23:11
  • @JustAnAmateur You are welcome! $x^2-x^3 = x^2(1-x).$ Quadratic factor provides the conditions near $0,$ and the second factor near $1.$ – Yuri Negometyanov Aug 10 '20 at 23:13
  • Oh, I see. And this basically made you think about considering the auxiliary function $g(x)=f(x)-x^2+x^3$ and that's how you got the inequality. – JustAnAmateur Aug 10 '20 at 23:16
  • It seems that there is some more work to do before finally getting (2), I will look into it tomorrow since now it is late here. – JustAnAmateur Aug 10 '20 at 23:22
  • 3
    I am curious how you derived $(2).$ Taylor's remainder gives $6f(x) = ax^2+b_xx^3$ with $|a|, |b_x|\leq 1$ and the triangle inequality gives $|f(x)|\leq\dfrac{1}{6}(x^2+x^3).$ – dezdichado Aug 11 '20 at 00:59
  • @JustAnAmateur Thank you for the comments! Just edited. – Yuri Negometyanov Aug 11 '20 at 06:44
  • @dezdichado Thank you for the comment. Clarified. – Yuri Negometyanov Aug 11 '20 at 06:46
  • 1
    I still don't see how things would be this immediate. The Bezout I know only concerns polynomials and it says nothing about the derivatives. – dezdichado Aug 11 '20 at 06:51
  • @dezdichado I agree with you. I tried to develop on my comment from last night and it isn't possible to prove that inequality in the usual manner with the derivative(s) of that auxiliary function – JustAnAmateur Aug 11 '20 at 07:07
  • @dezdichado 1) Besou's theorem had applied 3 times; 2) The polynomial bounds the function. – Yuri Negometyanov Aug 11 '20 at 07:07
  • 1
    But you need your function to be polynomial in the first place in order to apply Bezout. Besides, that is not even the interesting part - the interesting part is how in the world the third derivative come out so cleanly whilst there is no second derivative in your estimation. – dezdichado Aug 11 '20 at 07:13
  • @YuriNegometyanov I think that where both me and dezdichado have our doubts is why the polynomial bounds the function. Could you please develop on this? I tried to prove it in the straightforward way, but it didn't work out, I would need some bounds on $f''$, so I am curious to see how you got the inequality, it doesn't seem to be so obvious. – JustAnAmateur Aug 11 '20 at 07:14
  • @JustAnAmateur Detalized. – Yuri Negometyanov Aug 11 '20 at 21:30
  • @dezdichado Detalized. – Yuri Negometyanov Aug 11 '20 at 21:31
  • 2
    The part until $|a|\le 1$ looks OK, but I don't see how you go from $|a|\le 1$ to your final bound – MoonKnight Aug 11 '20 at 21:49
  • 4
    Another point. Even if $|f(x)| \le x^2-x^3$, why the integral takes the maximum when $|f(x)|$ takes the maximum? especially noticing that $(2x-1)$ is negative in the first half the integral and positive in the second half of the integral – MoonKnight Aug 11 '20 at 22:45
  • @MoonKnight Thank you for the comments. New edit is ready. – Yuri Negometyanov Aug 12 '20 at 12:28
  • 1
    This solution should be right, but the use of notation $b'(x)$ is confusing. Just call it $c(x)$ or something and this will be easier to read at the end. Nice work. – Jake Mirra Aug 12 '20 at 13:04
  • @JakeMirra Thank you for advice! $h(x)$ looks the best choice. – Yuri Negometyanov Aug 12 '20 at 13:22
  • @dezdichado New edit is ready. – Yuri Negometyanov Aug 12 '20 at 13:24
  • @JustAnAmateur New edit is ready. – Yuri Negometyanov Aug 12 '20 at 13:29
  • 2
    @YuriNegometyanov. Thanks, it makes more sense now. But I am still confused about one step. Starting from $|6f(x) + ax^2|\le x^3$ and $|a|\le 1$. But how did you get $6|f(x)| \le x^2 - x^3$ from here? I can only get $6|f(x)| \le x^2 + x^3$. You may argue that $f(x)$ should have the same sign of $f''(0)$ (which is $a/3$) when $x$ is small. But why $f(x)$ cannot change sign? – MoonKnight Aug 12 '20 at 14:33
  • 2
    Yeah i had the same concern since the beginning. – dezdichado Aug 12 '20 at 15:30
  • 3
    Another tricky point here, $f'(x)$ must change sign somewhere as $f(0)= f(1)=1$. So changing the integral from IBP (which involves $2x-1$ part that changes the sign) back to the original form (which has a $f'(x)$ part that changes the sign) does not resolve the issue. The upper bound you set for $|f'(x)|$ is not correct, since $x/3-x^2/2$ is actually negative around $x=1$. – MoonKnight Aug 12 '20 at 16:17
  • 4
    now I am surprised by the fact that this answer keeps getting upvoted. The flaw here seems not rectifiable and similar questions [here](https://math.stackexchange.com/questions/3654560/given-int-frac13-frac23fxdx-0-how-to-prove-4860-int-01fxdx2-l?noredirect=1&lq=1) and [here](https://math.stackexchange.com/questions/1179154/prove-the-following-integral-inequality-int-01fx2dx-ge-1920-left) have very complicated solutions whilst only having much smaller coefficients, $4860$ and $1920$, and only having to deal with second order derivatives. – dezdichado Aug 14 '20 at 23:29
  • @dezdichado The optimal function is correct, the assuming can be refined (see the new answer). I think, this answer is a step in the right direction. – Yuri Negometyanov Aug 15 '20 at 08:04
9

First, using integration by parts, and noting that $f(0)=f'(0)=0$ we get $$ f(x)=\frac{x^2}{2}f''(0)+\frac{1}{2}\int_0^x(x-t)^2f^{(3)}(t)dt\tag{1} $$ And from $f(1)=0$ we get also $$ 0=f(1)=\frac{1}{2}f''(0)+\frac{1}{2}\int_0^1(1-t)^2f^{(3)}(t)dt $$ Hence $$\eqalign{ 0&=\frac{x^2}{2}f''(0)+ \frac{1}{2}\int_0^xx^2(1-t)^2f^{(3)}(t)dt\cr &+ \frac{1}{2}\int_x^1x^2(1-t)^2f^{(3)}(t)dt \tag{2}} $$ Subtracting from $(1)$ from $(2)$ we get $$f(x)=-\int_0^1k(x,t)f^{(3)}(t)dt\tag{3}$$ with $$k(x,t)=\cases{\dfrac{x^2(1-t)^2-(x-t)^2}{2}& if $t\leq x$\cr \dfrac{x^2(1-t)^2}{2}& if $ x\leq t$ }$$ (Note that $k(x,t)\ge 0$ for $0\leq x,t\leq 1$).

Now, we note that $$\eqalign{I&=\int_0^1x(1-x)f(x)f'(x)dx=\int_0^1\left(x-\frac{1}{2}\right)(f(x))^2dx\\ &=\int_0^1\int_0^1\int_0^1\left(x-\frac{1}{2}\right)k(x,t)k(x,s)f^{(3)}(t)f^{(3)}(s)dtdsdx\\ &=\int_0^1\int_0^1H(t,s)f^{(3)}(t)f^{(3)}(s)dtds\tag{4} } $$ with $$H(t,s)=\int_0^1\left(x-\frac{1}{2}\right)k(x,t)k(x,s)dx\tag{5}$$ Now, $H(s,t)=H(t,s)$ and for $0\leq t\leq s\leq 1$ we find $$H(t,s)=\frac{1}{240} (s-1)^2 t \left(s^4 (t-2)+s^3 (t+2)+s^2 (1-4 t)+s t+(t-1) t^4\right)$$ It is not very easy but we can show (see the remark below,) that $H$ is nonnegative, and this is the crucial point. As it is depicted in the next figure

1

It follows from the condition $|f^{(3)}|\leq 1$ that $$|{I}|\leq \int_0^1\int_0^1H(t,s)dtds=\frac{1}{30240}.$$

Remark. Let me add a proof of the positivity of $H(t,s)$. Indeed, It is a matter of verification that for $0<t\leq s\leq 1$ we have $$\eqalign{\frac{240 }{(1-s)^2 t}H(t,s)&= (1-t) \left(s^2-t^2\right)+2 s^2 (1-s) (s-t)\\& +s (s+1) (1-s)^2 t+(1-t)^2 t^2 (t+1)}$$

which is a sum of positive terms.

Omran Kouba
  • 28,320
  • 1
  • 51
  • 96
  • 1
    +1. now this is an answer. – dezdichado Aug 16 '20 at 19:32
  • 1
    @Omran Kouba congratulations, this looks correct to me, you will most likely receive the bounty :) – JustAnAmateur Aug 16 '20 at 19:51
  • 1
    Your solution is impressive. I used Maple to check it. (+1 already) – River Li Aug 17 '20 at 05:38
  • 2
    @Omran Kouba I am curious, what was the intuition behind your solution? – JustAnAmateur Aug 17 '20 at 07:23
  • 3
    @JustAnAmateur, (1) The basic difficulty in this problem is the $2x-1$ changes sign on $[0,1]$ and direct majorization obliges us to replace it with $|2x-1|$ and this is not recommended here. (2) We need to express $f$ in terms of $f^{(3)}$ and here comes Taylor's formula with integral rest. This leads to the kernel $k(x,t)$, (3) Interchanging the order of integration will hopefully help us get rid of (2x-1). (4)So, we obtain $I$ with a kernel $H(t,s)$, and win if this kernel is positive. First, I made a plot, to be sure, This I proved that. Glad that you found it interesting. – Omran Kouba Aug 17 '20 at 12:51
  • 1
    @OmranKouba, this kernel trick is eye-opening for me. Thanks! – MoonKnight Aug 17 '20 at 16:42
  • 1
    @OmranKouba I know it's been a while, but I was rereading this solution and I have a little bit of a hard time understanding how you found that kenel. Could you please develop more on this? – JustAnAmateur Aug 25 '20 at 22:29
  • 2
    @JustAnAmateur, I made an edit, maybe it is clearer now. – Omran Kouba Aug 26 '20 at 16:59
  • 2
    @JustAnAmateur Can we write it as follows: ... we get $f(x) = -\int_0^1 \frac{x^2(1-t)^2}{2} f'''(t) \mathrm{d} t + \int_0^x \frac{(x-t)^2}{2} f'''(t) \mathrm{d} t.$ (see next comment). – River Li Aug 27 '20 at 00:57
  • 2
    @JustAnAmateur Let $u(x, t) = \left\{ \begin{array}{cc} \frac{(x-t)^2}{2} & t \le x \\ 0 & t > x \\ \end{array} \right.$ Then $\int_0^x \frac{(x-t)^2}{2} f'''(t) \mathrm{d} t = \int_0^1 u(x, t) f'''(t) \mathrm{d} t$. – River Li Aug 27 '20 at 00:57
  • 2
    @JustAnAmateur Then $f(x) = -\int_0^1 (\frac{x^2(1-t)^2}{2} - u(x, t)) f'''(t) \mathrm{d} t$. Let $k(x, t) = \frac{x^2(1-t)^2}{2} - u(x, t)$. Then $f(x) = -\int_0^1 k(x,t) f'''(t) \mathrm{d} t$. – River Li Aug 27 '20 at 00:59
  • 1
    Thank you, now it is clearer! – JustAnAmateur Aug 27 '20 at 09:56
  • 2
    @JustAnAmateur You are welcome. – River Li Aug 27 '20 at 09:57
  • Wonderful answer +1 – Felix Klein Aug 02 '21 at 21:21
4

Just an idea put $f(x)=\frac{g(x^2(1-x)^2)}{x(1-x)}$ we have :

$$f'(x)=-\frac{(2x-1)(2(x-1)^2x^2g'((x-1)^2x^2)-g((x-1)^2x^2))}{x^2(1-x)^2}$$

So we have :

$$f'(x)f(x)x(1-x)=-\frac{(2x-1)(2(x-1)^2x^2g'((x-1)^2x^2)-g((x-1)^2x^2))}{x^2(1-x)^2}x(1-x)\frac{g(x^2(1-x)^2)}{x(1-x)}$$

Or :

$$f'(x)f(x)x(1-x)=-(2x-1)\Big(2g'((1-x)^2x^2)g((1-x)^2x^2)-\frac{g^2(x^2(1-x)^2)}{x^2(1-x)^2}\Big)$$

Now we can integrate by parts but I can't go further .

Update :

As I'm stuck with the substitution above I propose another one :

Put :

$$f(x)=g(\ln(x)-\ln(1-x))$$

Then :

$$f'(x)f(x)x(1-x)=g(\ln(x)-\ln(1-x))g'(\ln(x)-\ln(1-x))$$

Now we can make the substitution $y=\ln(x)-\ln(1-x)$ or $x=\frac{e^y}{e^y+1}$ in the integral . Another substitution and we see that there is a link with the Laplace transform .

Update 2:

It's really my last idea on this problem .

We have the condition :

$$|f'''(x)|\leq 1\quad (1)$$

but :

$$1=(x+(1-x))^n$$

So we can use binomials theorem to expand the terms .

Now we put :

$$f(x)=\operatorname{B}_{i,n}(x)$$

Where we speak about the Bernstein polynomials .

See here (3.1) for the derivative of a Bernstein polynomial

Remains to compare the two quantities .

If it's really useless I shall delete .

Hope it inspire someone .

Erik Satie
  • 3,376
  • 3
  • 9
  • 36
1

$\color{green}{\textbf{Version of 15.08.20.}}$

At first,

$$I=\int_0^1x(1-x)f(x)f'(x)\text{d}x \,\overset{IBP}{=\!=\!=}\,\frac12x(1-x)f^2(x)\bigg|_0^1+\frac12\int_0^1 (2x-1)f^2(x)\text{ d}x,$$ $$I=\frac12\int_0^1 (2x-1)f^2(x)\text{ d}x.\tag1$$

Let $$x = \frac{y+1}2,\quad \text{ d}x = \frac12\text{ d}y,\quad y = 2x-1,\quad g(y) = f\left(\frac{y+1}2\right), \tag2$$

then $$f(x) = g(2x-1) = g(y),\quad f'''(x) = 8g'''(2x-1) = g'''(y),\tag3$$ $$I = \frac14\int\limits_{-1}^{1} y g^2(y)\text{ d}y,\tag4$$ under the conditions $$g(-1) = g'(-1) = g(1) = 0,\quad |g'''(y)| \le \frac18.\tag5$$

Decomposition to the even and the odd parts $$g(y)=g^\,_+(y)+g^\,_-(y),\quad g^\,_\pm(y) = \frac12(g(y)\pm g(-y)),\quad g^\,_\pm(-y) = \pm g^\,_\pm(y),\tag6$$ gives $$I = \int\limits_{0}^{1} y g^\,_+(y)\,g^\,_-(y)\text{ d}y.\tag7$$

In accordance with the Shwartz inequality, $$I^2 \le \int\limits_{0}^{1} \big(y g^\,_+(y)\big)^2\text{ d}y\cdot \int\limits_{0}^{1} g^2_-(y)\text{ d}y,\tag8$$

wherein $(8)$ becames the equality if $$|g^\,_-(y)| = y\,g^\,_+(y).$$ Then from $(5)$ should $$g^\,_+(y) = (1-y^2)h(y),\quad \big|g^\,_-(y) \big| =(y-y^3)h(y)\tag{9}$$ Therefore, the function $$g(y)=(1+y)(1-y^2) h(y)$$ maximizes $|I|$ under the conditions $(5)$ near $y=\pm1.$

Taking in account the rest of the conditions, one can get $h(y) = \text{constant} =\frac1{48},$ $$g(y) = \frac1{48}(1+y)(1-y^2),\tag{10}$$

$$48^2I_{opt} = \int\limits_0^1 (y^3-y)^2\text{ d}y = \int\limits_0^1 (y^6-2y^4+y^2)\text{ d}y = \frac17-\frac25+\frac13 = \frac8{105},\tag{11}$$ $$\color{brown}{\mathbf{|I| \le \frac1{30240}}},$$ $$f_{opt}(x) = \pm g_{opt}(2x-1) = \pm \frac16 (x^2-x^3).$$

Yuri Negometyanov
  • 27,332
  • 1
  • 29
  • 59
  • 1
    @JustAnAmateur, refined version is there – Yuri Negometyanov Aug 15 '20 at 08:14
  • 1
    @MoonKnight refined version is there. – Yuri Negometyanov Aug 15 '20 at 08:17
  • 6
    sorry but again that sentence between $(7)$ and $(8)$ is the meat of the proof and it has no justification. Look, I do want this problem to have simple solution. However, we are talking about optimizing a functional $I(f) = \int_0^1x(1-x)f(x)f'(x)dx$ over $C^3([0,1])$ and but you just kind of hand-wave your way through proving that the suspected value is indeed the maximum. A person of your reputation on this site surely must know what a complete proof would be. – dezdichado Aug 15 '20 at 16:33
  • 1
    @dezdichado You are right. Appendium added. – Yuri Negometyanov Aug 16 '20 at 10:25
  • 2
    There isn't a reason to assume why $\theta$ would be a constant and moreover $|g'''|$ wouldn't be differentiable, unless $g$ does not change sign, in which case the problem is almost trivial. – dezdichado Aug 16 '20 at 17:49