A function $$f:\mathbb{R}\rightarrow [-1,1]$$, has its integral vanish over the interval $x\in[1,3]$.

Find the maximum value of the integral $$\displaystyle\int_{1}^{3}\dfrac{f(x)}{x}\mathrm{d}x$$

I tried to throw the Euler-Lagrange equation at this functional but got some non-sensical condition that said $x=\mathrm{constant}$.

My teacher gave a non constructive proof that f is actually a non-continuous function, and proved the upper bound.

My question is:

  1. What does this condition mean? What should you do in the case the EL equation yields no function?
  2. Is there any motivation for considering a discontinuous function?
  • $\begingroup$ Using Euler Lagrange imposes some differentiability conditions on $f$. Without having checked I would speculate that the solution is a 'bang-bang' solution, so I would guess $+1$ on $[1,2]$ and $-1$ on $[2,3]$. $\endgroup$ – copper.hat Feb 10 at 5:43
  • $\begingroup$ You're right. Can you explain why? And what you mean by 'bang-bang'? $\endgroup$ – Archimedesprinciple Feb 10 at 5:46
  • $\begingroup$ It means that the control ($f$) is at a limit most of the time. The constraints here are $\int f = 0$, $-1 \le f(x) \le 1$. $\endgroup$ – copper.hat Feb 10 at 5:53

There are three constraints on $f$, $-1 \le f(x) \le 1$ and $\int f = 0$. This combination of constraints makes this slightly different than a classical optimal control setting.

Intuitively, since $x \mapsto {1 \over x}$ is decreasing, we expect $f$ to have value $+1$ and then transition to $-1$. Since we have the zero average constraint the transition would occur at $t=2$.

Let $f^*$ be $1$ on $[1,2]$ and $-1$ on $(2,3]$ and let $f$ satisfy the constraints. We have $\int_1^2 f+ \int_2^3 f = 0$ and $\int_1^2 f^*+ \int_2^3 f^* = 0$, so $\int_1^2 (f^*-f) = -\int_2^3 (f^*-f) $, or equivalently $\int_1^2 (1-f) = \int_2^3 (1+f)$.

Since ${1 \over x} \ge {1 \over 2} \ge {1 \over y}$ for $x \in [1,2], y \in [2,3]$, we have $\int_1^2 {(1-f(x)) \over x}dx \ge \int_1^2 {(1-f(x)) \over 2}dx = \int_2^3 {(1+f(x)) \over 2}dx \ge \int_2^3 {(1+f(x)) \over x}dx$ from which we get $\int_1^3 {f^*(x) \over x} dx \ge \int_1^3 {f(x) \over x} dx $.


Here is a slightly different approach that might offer some more intuition (or just moves the complication somewhere else).

Let $C= \{f:[1,3]\to [-1,1] | f \text{ is measurable}, \int f = 0 \}$, and consider $C \subset L^2[1,3]$. It is straightforward to show that $C$ is convex & compact (in $L^2$) and the function $c(f)= \int_1^3 {f(x) \over x}dx$ is continuous and linear. Hence there is a minimiser $f^*$ because of compactness.

Finally, we can show that if $f \in C$, then if the measure of $\{ x \in [1,2] |f(x) < 1 \}$ is non zero, there is some $f' \in C$ such that $c(f') > c(f)$. Hence the function $1_{[1,2]}-1_{(2,3]}$ (or any ae. equivalent function) is a minimiser.

Another way of doing this would be to note that since $c$ is linear, we can take a minimiser $f^*$ to be an extreme point of $C$ and we can show that the extreme points of $C$ are the measurable zero average functions that satisfy $f(x) \in \{-1,+1\}$ for ae. $x$. In particular, you can see a simple form of a 'bang bang' control, and the problem reduces to determining the 'on set' (where $f(x) = 1$).


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.