I feel like this should be straightforward, but does anyone have a proof of the following?

Let $f: \mathbb{R}^n \to \mathbb{R}$ satisfy the following. For each coordinate $i$, for an arbitrary vector $x_{i}$, define $ f_i(y)$ to be $f$ restricted to the $i^{th}$ parameter, fixing the others $ x_{i}$; then $ f_i(y)$ is convex. Then $f$ is a convex function.

Also, if there is a more standard math notation I should use to describe this problem, what would that be?

(Edit: or a counterexample...)


I should add what I've tried.

$g(\lambda \vec{x} + (1-\lambda)\vec{y}) = g(\lambda x_1 + (1-\lambda)y_1,\dots,\lambda x_n + (1-\lambda)y_n)$

$\leq \lambda g(x_1,\lambda x_2 + (1-\lambda)y_2,\dots) + (1-\lambda) g(y_1,\lambda x_2 + (1-\lambda)y_2,\dots)$

I don't see this going anywhere good, we end up with something nasty like

$\lambda^n g(\vec{x}) + \lambda^{n-1}(1-\lambda) \sum_i g(y_i,\vec{x}_{i}) + \dots + (1-\lambda)^n g(\vec{y})$

I think. Maybe we can get it from there?


Simple counterexample (assuming I understood your setup): $xy$ is nonconvex while trivially convex (linear) when fixing one of the variables.

  • $\begingroup$ Ha, yep. Thanks! $\endgroup$
    – mathison
    Oct 12 '12 at 8:49

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.