Why $\widetilde m = argmin_{m \in \mathcal P(X)} E[m]$ implies $\widetilde m (\arg \min \frac{\delta E}{\delta m}[\widetilde m]) =1$?

38 Views Asked by At

Consider $E: \mathcal P (X) \rightarrow \mathbb R \cup \{ \infty \}$ a functional (with a convex and dense domain, $E< +\infty$) over $\mathcal P(X)$ the set of probability measures of a metric compact space $X$.

Then consider the minimisation problem $\underset{m \in \mathcal P(X) }{\inf} E[m] $

Suppose that there is a $\widetilde m \in \mathcal P(X)$ solves the minimisation problem and that $E$ is differentiable in $\widetilde m$ is the sense that there is a function $\frac{\delta E}{\delta m}[\widetilde m]$ such that for all $m$ in the domain of $E$

$$ 0 \leq \lim_{\epsilon \rightarrow 0^+} \frac{E[\widetilde m + \epsilon (m-\widetilde m)]-E[\widetilde m]}{\epsilon}\leq \int_X \frac{\delta E}{\delta m}[\widetilde m] ~d(m -\widetilde m)$$

Then we have for all $m$ in the domain of $E$ $$ \int_X \frac{\delta E}{\delta m}[\widetilde m] ~d \widetilde m \leq \int_X \frac{\delta E}{\delta m}[\widetilde m] ~dm .$$

How to conclude that $\widetilde m$ is concentrated on $(\arg \min \frac{\delta E}{\delta m}[\widetilde m]) $?

Here is my idea. Since $$ \int_X \frac{\delta E}{\delta m}[\widetilde m] ~d \widetilde m \leq \int_X \frac{\delta E}{\delta m}[\widetilde m] ~dm $$

folows for any $m$ so by taking $m(x) = \begin{cases} 1, ~\text{if } x = \arg \min \frac{\delta E}{\delta m} [\widetilde m]\\ 0, \text{otherwise}\end{cases}$

we have

$$ \int_X \frac{\delta E}{\delta m}[\widetilde m] ~d \widetilde m \leq \arg \min \frac{\delta E}{\delta m} [\widetilde m]$$

but $$ \arg \min \frac{\delta E}{\delta m} [\widetilde m]\leq \int_X \frac{\delta E}{\delta m}[\widetilde m] ~d \widetilde m \leq \arg \max \frac{\delta E}{\delta m} [\widetilde m]$$

therefore

$$ \int_X \frac{\delta E}{\delta m}[\widetilde m] ~d \widetilde m = \arg \min \frac{\delta E}{\delta m} [\widetilde m]$$

and so $\widetilde{m}(x) = \begin{cases} 1, ~\text{if } x = \arg \min > \frac{\delta E}{\delta m} [\widetilde m]\\ 0, > \text{otherwise}\end{cases}$

Many thanks