Is the product of positive real numbers convex?

620 Views Asked by At

This maybe a trivial question, but I don't seem to find the answer. I want to show that for two vectors $x'=(x_1', ...)$ and $x''=(x_1'', ...)$, such that all elements are in $[0,1]$, the following statement is true:

$\Pi_i(\lambda x' + (1-\lambda) x'')\leq \lambda \Pi_i( x') + (1-\lambda)\Pi_i( x'')$,

where $\Pi$ is the product operatorand $\lambda \in [0,1]$.

Inspired by this post, I tried for $x'=(x_1', x_2')$ (product of only two elements) and found, by computing the Hessian matrix that $x'x''$ is indeed convex on $[0,1]$.

My intuition is that it is still true for more general products but I am not sure how to show it.

Also, if I am wrong I would actually be curious to know if the product is actually concave.

2

There are 2 best solutions below

2
On

No in general it's not true unfortunately (in $\mathbb{R}^{n}$) but it is indeed convex on $\mathbb{R}_{++}^{n}$. Consider $f(x_1,\cdots,x_n)=x_1x_2\cdots x_n>0$. Then $\frac{\partial^2f}{\partial x_i \partial x_j} = x_1x_2\cdots x_{i-1}x_{i+1}\cdots x_{j-1}x_{j+1}\cdots x_n=\frac{f}{x_ix_j}$. Therefore the elements of Hessian are $[H]_{ij}=\frac{f}{x_ix_j}$. For an arbitrary vector $v\in \mathbb{R}_{++}^{n}$ we have

$$v^THv=\sum_{i=1}^{n}\sum_{j=1}^{n} f\frac{v_i}{x_i}\frac{v_j}{x_j}(1-\delta(i-j))>0$$

Where $\delta(i-j)=\begin{cases}1 &i=j\\ 0 &O.W.\end{cases}$ is the Kronecker delta function. Since $v^THv$ is a positive form on $\mathbb{R}_{++}^{n}$, thus the function is convex in $\mathbb{R}_{++}^{n}$.

To generalize a little further, it is still convex in $\mathbb{R}_{+}^{n}$ but not strictly convex, because there is the possibility that one or more of $x_i$'s be zero, therefore the Hessian is P.S.D. .

0
On

Note that $$ \mathbf{H} = \frac{1}{\mathbf{x}}\left[\frac{1}{\mathbf{x}}\right] ^ \text{T} -\text{diag}\{\frac{1}{\mathbf{x} ^ 2}\}. $$ In the equation above, the scaling factor equaling the product of $\mathbf{x}$ has been omitted.
Then we have $$\begin{eqnarray} s \triangleq \mathbf{y} ^ \text{T} \mathbf{H} \mathbf{y} & = & \mathbf{y} \frac{1}{\mathbf{x}} \left[ \frac{1}{\mathbf{x}}\right] ^ \text{T} \mathbf{y} - \mathbf{y} ^ \text{T} \text{diag}\{\frac{1}{\mathbf{x} ^ 2}\} \mathbf{y} \\ & = & \left(\sum_{i=1}^{n}{\frac{y_i}{x_i}} \right) ^ 2 - \sum_{i=1}^{n}{\left(\frac{y_i}{x_i}\right)^2}\text{.} \end{eqnarray}$$ Consider two cases:
Case 1: $\mathbf{x} = [1, 1] ^ \text{T}$, $\mathbf{y} = [1, -1] ^ \text{T}$.
Case 1: $\mathbf{x} = [1, 1] ^ \text{T}$, $\mathbf{y} = [2, 3] ^ \text{T}$.
In case 1, $s < 0$, while in case 2, $s > 0$.
In summary, $\mathbf{H}$ is indefinite.