$$R[\alpha(x,y)] = \left[ (y-\alpha(x,y)x)^2 + c\left(\frac{\partial \alpha(x,y)x}{\partial y}\right)_+\right]$$

$$c > 0$$

I would like to find the function $\alpha(x,y)$ which minimizes $R$. As additional information I know that $R$ is always greater than or equal to 0. I think this is a subject of calculus of variations, which I have no clue about.

I found the Functional Derivative article on Wikipedia.


Am I on the right track? Which rule should be applied here?

For example $\alpha(x,y) = \frac{y}{x}$ is a possible candidate function, but it only minimizes the first term (the square) but not both terms (square and the partial derivative) simultaneously. In this scenario $R = 0 + c = c$ assuming $x \ne 0$. If I try functions with predefined structure like $\alpha(x,y) = m\frac{y}{x}$ and solve for $m$ then it is easy, but I would like to find the best minimizing function not one with a particular structure (analytically, if not numerically for a particular $x$ and $y$).

Above is a simplified version of the following, which I plan to solve once I understand the simple one. I am also OK to start without the positive part requirement for the derivative term, in order not to introduce discontinuity. I just would like to have some kickstart for the derivations.

$$R[\alpha(x,y)] = \left[ (y-f(\alpha(x,y),x))^2 + c\left(\large\frac{\partial f(\alpha(x,y),x)}{\partial y}\right)_+\right]$$

  • $\begingroup$ Is $R$ a function of not only $\alpha$ but also of $x$ and $y$? Do you want to find $\alpha$ that minimizes $R$ for given $x$ and $y$? $\endgroup$ – md2perpe Oct 24 '18 at 20:36
  • $\begingroup$ Normally in calc of variation you would be looking to find $\alpha$ which minimises R integrated over x and y, subject to appropriate boundary conditions. In what you have you could set R=some large negative constant and treat it as a differential equation in $\alpha$ $\endgroup$ – user121049 Oct 25 '18 at 8:45
  • 2
    $\begingroup$ I am not sure what you mean by a minimum of R. It could take on different values at every x & y. You could possibly arrange that R is very small at some point but at the cost of being large somewhere else. Do you mean average of R over the domain of x & y? Also what do x & y range over and is R fixed on the boundary? $\endgroup$ – user121049 Oct 25 '18 at 10:29
  • $\begingroup$ Let me point it out. Say e.g. $x_0=y_0=1$. So you want to minimize $R$ at $(1,1)$ and do not care for values of $R$ at other $(x,y)$, right? This is what you say above. This is a simple interpolation where $R$ at that point can be made $0$. For example, pick $$\alpha(x,y)=\frac{y_0}{x_0}+(y-y_0)^2.$$ $\endgroup$ – A.Γ. Oct 26 '18 at 6:40
  • $\begingroup$ Why not? $y_0$ is simply a constant, e.g. if $x_0=y_0=1$ then $\alpha(x,y)=1+(y-1)^2$. Could you provide more information on the problem? $\endgroup$ – A.Γ. Oct 26 '18 at 7:16

Too long for a comment so I write as an answer. I am just taking another guess at what you want. Let's guess that for the right choice of $\alpha$, $R(x,y)=r$ i.e. a constant everywhere. Also note in your simple example you can scale $x$ away so write $z=x \alpha$.

The equation becomes $c \frac{dz}{dy} + (y-z)^2=r$. Being lazy I've asked Wolfram Alpha to solve this for me. https://www.wolframalpha.com/input/?i=Solve+(y-+a(y))%5E2+%2B+c++da(y)%2Fdy+%3D+r

Let me call $b=\sqrt{\frac{r-c}{c^2}}$

Then $\alpha(x,y)=\frac{z}{x}=\frac{c b +y}{x} -\frac{b c}{x(b k \exp{(2 b y)} - 1)}$

Now the integration constant $k$ can depend on $x$, i.e. $k(x)$ and should be fixed by any boundary conditions you may have.

For $b$ to be real then $r\geq c$. When $r=c$ then $b=0$ and hence $\alpha=\frac{y}{x}$ as you already have. So the minimum value of $r$ and hence $R$ is given by the solution you already have.

| cite | improve this answer | |

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.