1

Does there exist a function $f\colon \mathbb{R}^2 \rightarrow \mathbb{R}$ such that

(1) for all $(x,y) \in \mathbb{R}^2$, f is convex in the x-direction and y-direction

(2) $f$ has multiple non-degenerate minima? Non-degenerate: for any path P between minima $m_1$ and $m_2$, there exists $p \in P$ s.t. $f(p) > f(m_1)$ and $f(p) > f(m_2)$.

A function satisfying (1) is for example $f(x,y) = (xy)^2$ and a function satisfying (2) is for example $$f(x,y) = -\left(e^{-\left((x-1)^2 + (y-1)^2\right)} + e^{-\left((x+1)^2 + (y+1)^2\right)}\right).$$

If not (as I suspect) prove so, and how does this generalise for $f\colon \mathbb{R}^n \rightarrow \mathbb{R}$?

With $f(x,y) = (x-y)^2$ we come close, but minima are degenerate in the sense described in (2).

newling
  • 183
  • 1
    A rigourous statement would require you to specifically take care of the degenerate case $f = cst.$, which in a certain sens satisfy your two constraints. – jibe Jun 27 '14 at 15:14

2 Answers2

2

This function should satisfy your requirements $$ f(x,y) =((x-y)^2-1)^2 +2(x+y)^2. $$ It has global minima at $(-\frac12,\frac12)$, $(\frac12,-\frac12)$ - according to wolframalpha.

The second derivatives are non-negative: $$ \frac{\partial^2}{\partial x^2} f(x,y) = \frac{\partial^2}{\partial y^2} f(x,y)= 12(x-y)^2. $$

daw
  • 49,113
  • 2
  • 38
  • 76
  • Out of curiosity, could you elaborate on how you managed to make and tune this example ? Could such a function be generalized to the $\mathbb{R}^{n}$ case ? – jibe Jun 27 '14 at 22:14
  • I tried to lift the function $f(x) = (x^2-1)^2$ to the 2d-case. Then played with wolframalpha ;) – daw Jun 28 '14 at 08:28
  • 1
    @newling, please check your calculations. The second derivative is indeed $\partial^2 f/\partial^2 x = 12(x-y)^2$. – daw Jun 30 '14 at 05:34
  • I had calculated $\partial^2f / {\partial x \partial y}$, apologies – newling Jul 01 '14 at 08:36
1

I guess that with the current writing of your constraints, (1) and (2) are compatible. Indeed, we consider the function

$$ f(x,y) \mapsto (x - y)^{2} $$

One immediately has $\frac{\partial^{2} f}{\partial x^{2}} = \frac{\partial^{2} f}{\partial y^{2}} = 1 > 0$, so that $f$ is always strictly convex in the $x$ and $y$ direction. However, the minimum value of $f$ is $0$ reached for all the points $f (t , - t) = 0$, so that $f$ reaches its minimal value along the entire axis $y = -x$, having an infinite number of minima.

jibe
  • 892