Let $n\geq 2$ and $\displaystyle f:(x_1,\ldots,x_n) \mapsto \prod_{i=1}^n x_i^{\alpha_i}\quad$ be defined for positive $x_i$

Suppose $\sum_{i=1}^n \alpha_i > 1$. Prove that $f$ is neither concave nor convex.

The Hessian matrix of $f$ is easy to compute. It suffices to prove that it is neither positive semi-definite, neither negative semi-definite at some point.

Let $H(f)(x)$ denote the Hessian matrix of $f$ at $x=(x_1,\ldots,x_n)$. It suffices to find some $x$ such that $\det H(f)(x)<0$ or one may look for $u$ and $v$ such that $u^T H(f)(x)u> 0$ and $v^T H(f)(x)v< 0$.

I've noted that $$\frac{\partial^2 f}{\partial^2 x_i}(x) = \alpha_i(\alpha_i-1)\frac{f(x)}{x_i^2}$$ and $$\frac{\partial^2 f}{\partial x_j \partial x_i}(x) = \alpha_i \alpha_j\frac{f(x)}{x_i x_j}$$

I think the approach with the determinant is the most tractable. $x=(\alpha_1,\ldots, \alpha_n)$ seems like a promising point, although I'm not through with the computations.

EDIT: it seems that the approach with the determinant would require that $n$ is even...

EDIT 2: Provided my computation is correct, at $x=(\alpha_1,\ldots, \alpha_n)$, $$\det H(f)(x)=(f(x))^n\frac{(-1)^{n-1}}{\prod_{i=1}^{n}\alpha_i} \left[\left(\sum_{i=1}^n \alpha_i \right) -1 \right]$$

This solves the problem, with $n$ even. It remains to take care of odd $n$.

  • $\begingroup$ I imagine that $n>1$, right? If not, the result is wrong. $\endgroup$ – mathcounterexamples.net Mar 9 '17 at 18:45
  • $\begingroup$ @mathcounterexamples.net yes, it's the Cobb-Douglas function in economics. $\endgroup$ – Gabriel Romon Mar 9 '17 at 18:47
  • $\begingroup$ It seems that the claim is not true for $n$ odd. Please see below. $\endgroup$ – mathcounterexamples.net Mar 10 '17 at 10:45
  • $\begingroup$ Below a general proof regarding convexity. $\endgroup$ – mathcounterexamples.net Mar 10 '17 at 14:02

If $f$ was convex or concave, its graph would sit on a single side of its tangent plane $$\varphi(x_1, \dots, x_n) = 0$$ with $$\varphi(x_1, \dots, x_n)=\alpha_1(x_1-1) + \dots + \alpha_n (x_n-1)$$ at point $U=(1, \dots, 1)$

For the origin $O$ we have $$\varphi(O)-f(O)=- (\alpha_1 + \dots + \alpha_n) < -1 \le 0$$

Now consider $i_0$ for which $\alpha_{i_0} > 0$ and a point $P_t$ whose coordinates are all vanishing except the $i_0$-one which is equal to $t>0$. You have $\varphi(P_t)-f(P_t)= \alpha_{i_0}t - \sum_i \alpha_i > 0$ for $t>\frac{\sum_i \alpha_i}{\alpha_{i_0}}$.

Hence the graph of $f$ is not on a single side of its tangent plane at $U$ proving that $f$ is not convex nor concave.

  • $\begingroup$ A slight variation of this reasoning may be written as follows: $$\forall x\in (\mathbb R_+^*)^n, \;f(x) \;?\; f(1)+\nabla f(1)\cdot (x-1)$$ Evaluating at $x=0$ (although $f$ may be not be differentiable at that point, you can evaluate it for some vector with small positive coordinates and then use the continuity of $f$), we can replace the $?$ sign with $\geq $ and write $$\forall x, \; f(x) \;\geq\; 1+\alpha_1(x_1-1) +\ldots +\alpha_n(x_n-1)$$ The end of the proof is similar to what you have written. $\endgroup$ – Gabriel Romon Mar 10 '17 at 15:28

This is a partial answer, proving that the determinant of the Hessian is of constant sign.

Looking at specific cases

Case $n = 2$

At least a partial solution for $n = 2$. I haven't try to extend it for $n>2$.

For $x=(1,1)$ you have $\det H(f)(1,1)=\alpha_1 \alpha_2 (1-\sum \alpha_i)<0$.

And it could be interesting to still look at the value of $\det H(f)(x)$ for $x = (1, \dots,1)$ and $n>2$.

Case $n = 3$

Take $f(x_1,x_2,x_3)=x_1 x_2 x_3$. You have in that case $\det H(f)(x)=2x_1 x_2 x_3 >0$, proving that determinant of the Hessian is always positive in that case.

The general case

Denoting $\alpha^\prime_i = \frac{\alpha_i}{x_i}$ and $\beta_i = \frac{1}{x_i}$, we have: $$ \begin{aligned}\det H(f)(x)&= (f(x))^n \begin{vmatrix} (\alpha^\prime_1 - \beta_1) \alpha^\prime_1 & \alpha^\prime_1 \alpha^\prime_2 & \dots & \alpha^\prime_1 \alpha^\prime_n\\ \alpha^\prime_1 \alpha^\prime_2 & (\alpha^\prime_2 - \beta_2) \alpha^\prime_2 & \dots & \alpha^\prime_2 \alpha^\prime_n\\ \vdots & \ddots & & \vdots\\ \alpha^\prime_1 \alpha^\prime_n & \dots & \alpha^\prime_{n-1} \alpha^\prime_n & (\alpha^\prime_n - \beta_n) \alpha^\prime_n\\ \end{vmatrix}\\ &=(f(x))^n (\alpha^\prime_1 \dots \alpha^\prime_n)^2 \begin{vmatrix} 1-\frac{1}{\alpha_1} & 1 & \dots & 1\\ 1 & 1-\frac{1}{\alpha_2} & \dots & 1\\ \vdots & \ddots & & \vdots\\ 1 & \dots & 1 & 1-\frac{1}{\alpha_n}\\ \end{vmatrix} \end{aligned}$$ proving that the sign of the determinant of the Hessian is independent from the point $x$. And according to your computation for $(x_1, \dots,x_n)=(\alpha_1, \dots,\alpha_n)$ this sign is positive for $n$ odd and negative for $n$ even.

  • $\begingroup$ Interesting! Thank you for computing the determinant in the general case. However, I'm quite dubious of what you derived. I agree the determinant is $>0$ for odd $n$ regardless of $x$. But this doesn't imply that, for a given $x$ all the eigenvalues share the same sign, let alone that for all $x$, the eigenvalues have the same sign. Unless I've missed something, this does not prove that $f$ is concave or $f$ is convex. $\endgroup$ – Gabriel Romon Mar 10 '17 at 11:14
  • $\begingroup$ You're right. My answer is focused on your statement regarding the sign of the determinant of the Hessian which is constant. This is not sufficient to make a conclusion on the convexity of the function. I'll update my answer accordingly. $\endgroup$ – mathcounterexamples.net Mar 10 '17 at 11:22

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.