6
$\begingroup$

Suppose for every $y$, $f(x,y)$ is strictly convex in $x$. $f(x,y)$ is continuous in $y$ and let $\mathcal X$ be compact (in my problem, $\mathcal X$ is an interval). Can anyone suggest any theorems or references for the problem of whether $x^*(y) = \text{argmin}_{x \in \mathcal X} f(x,y)$ is continuous in $y$?

$\endgroup$
4
$\begingroup$

As no one has answered, I will give this a try.

To begin with, together with compactness, and of course nonemptiness, of $X$, which you have assumed, you need $f(x,y)$ to be continuous in $x$ in order to ensure, via Weierstrass extrem value theorem, that $$\theta(y)=\text{argmin}_{x\in X}f(x,y)\tag{1}$$ exists for any $y$. It is good, as you do, to assume strict convexity in $x$, for this gives us a unique solutions to the minimization problems, and thus $(1)$ is a singleton.

Under these assumptions, I will try to show that $(1)$ is continous in the following sense, where I am essentially using the definition of continuity in terms of limits of sequences. The fact that the argmin is a set makes it complicated to directly apply the notion of continuity of function which seldom, in elementary calculus, maps elements to some set of sets.

Theorem. Consider a nonempty compact set $X\subseteq \mathbb{R}$ and a continuous function $f:\mathbb{R}^2\to\mathbb{R}$. Furthermore, suppose that $$\theta(\bar{y})=\{z|z\text{ minimizes }f(x,\bar{y})\text{ over }x\in X\}$$ is the singleton $\{\bar{z}\}$ (to ensure this, assume strict convexity in the first argument of $f$). Then, for sequences $z_n\in \theta(y_n)$, where $\{y_n\}_{\mathbb{N}}$ is a sequence converging to $\bar{y}$, we have $z_n\to \bar{z}$.

Proof. I will give a short proof. It is by contradiction, and is a variant of the proof of Lemma 6.3.2. in Nonlinear Programming: Theory and Algorithms, 3rd Edition, 2006, by Bazaara et al.

So suppose $z_n\to\bar{z}$, $z_n\in \theta(y_n)$, and suppose $|z_n-\bar{z}|>\epsilon>0$ for all $n\in I$ where $I$ is some index set. Since $X$ is compact and $z_n\in X$, the sequence $\{z_n\}_I$ has a convergent subsequence $\{z_n\}_{I'}$ with limit $z_0$ in $X$. By assumption, $|z_0-\bar{z}|\geq\epsilon>0$ and thus $z_0\neq \bar{z}$. In addition, for $\{y_n\}_{I'}$ it holds that $$f(z_n,y_n)\leq f(\bar{z},y_n)\tag{2}$$ since $z_n$ minimizes for $y_n$, but $\bar{z}$ does not minimize for $y_n$ necessarily.

Now, use continuity of $f$ in both of its arguments, and take the limit over $I'$ in $(2)$, to get that $$f(z_0,\bar{y})\leq f(\bar{z},\bar{y})\tag{3}.$$ But as $z_0\neq\bar{z}$ and $\bar{z}$ minimized $f(x,\bar{y})$, it follows that $z_0$ is another, distinct minimizer of $f(x,\bar{y})$ over $X$. Hence, $\theta(\bar{y})$ is not a singleton. But this contradicts the fact that $\theta(\bar{y})$ was a singleton. QED.

Reference suggestion. The Theorem of the Maximum. This theorem is more general and "provides conditions for the continuity of an optimized function and the set of its maximizers as a parameter changes". It involves the concept of a multivalued function (note that the assumption of the argmin being a singleton simplified the proof of the theorem above).

| improve this answer | |
$\endgroup$

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.