Is there a variational problem that can provide the following class of variational derivative?

63 Views Asked by At

Suppose I have the variational problem

$$ E(y) = \frac{1}{2}\int_{a}^{b} y^2 + \alpha y'^2dx $$

Variational derivative will provide

$$ \frac{\delta E}{ \delta y} = y -\alpha y'', $$

Is there a variational problem known in literature that can give me something like as a variational derivative the following

$$ \frac{\delta E}{\delta y} = \left(1-\alpha( y\right)) y - \alpha(y)y'' $$

?

The reason is the following, if I write down a gradient descent I can get something like

$$ \partial_t y = - (1 - \alpha(y))y + \alpha(y)y'' $$

And what I want to describe with $\alpha = \alpha(y) \in [0,1]$ is some measure of some feature of $y$, and I want the update to weight more one term of the gradient rather than the other, based on such function.

Is there anything like this known in literature? The closest think I could come up with was the EM algorithm, but it doesn't seem to me there's a variational framework for this.

More specifically $\alpha(y)$ could be a sigmoid or some gaussian depending on $y$.