Minimizing the function.

Minimizing the following function

$f(x_1,x_2,\cdots,x_n)=\prod\limits_i^n x_i^{x_i}$

such that

$x_1+x_2+\cdots+x_n=P, 2\le x_i$ and $x_i$ are integers.

My attempt: In my opinion we obtain the result when all $x_i's$ are almost equal i.e. $|x_i-x_j|\le 0$ for all $i$ and $j$. I am trying to solve by Lagrange's multiplier and I obtained a system of equation which looks messy.

try as a logarithm: $$f(x_1,x_2,\cdots,x_n)=\prod_i^nx_i^{x_i}$$ log: $$g(\vec x)=\log(f(\vec x))=\sum_i^n x_i \log(x_i)$$ Then lagrange method: $$\mathcal L(\vec x)=g(\vec x)+\lambda(P-\sum_i^n x_i)=\sum_i^n x_i \log(x_i)+\lambda(P-\sum_i^n x_i)=$$ $$=\sum_i^n x_i( \log(x_i) -\lambda)+\lambda P$$ now impose $\delta \mathcal L=0$.

The logarithm is strickly monotone so extremas are preserved.

EDIT: Full calculation: $$0=\delta\mathcal L=\delta\left\{\sum_i^n x_i( \log(x_i) -\lambda)+\lambda P\right\}=$$ $$=\sum_i^n \delta \left\{x_i( \log(x_i) -\lambda)\right\}+\delta\lambda P=$$ $$=\sum_i^n \left\{\delta x_i( \log(x_i) -\lambda) +x_i( \frac{\delta x_i}{x_i} -\delta \lambda) \right\}+\delta\lambda P=$$ $$=\sum_i^n \delta x_i\left( \log(x_i) -\lambda +1\right)+\delta\lambda(P-\sum_i^n x_i)$$ This implies both: $$\begin{cases} \log(x_i) -\lambda +1=0\\ P-\sum_i^n x_i=0 \end{cases}\implies\begin{cases} x_i =e^{\lambda-1}\equiv\hat x\\ P-\sum_i^n x_i=0 \end{cases}\implies x_i=\frac{P}{n}$$

• Does this show that $x_i=P/n$ are integers? @Francesco – kamran jamil Aug 31 '17 at 9:32
• obviously not, in general the result is rational. If the $x_i$ must be integers then this is an integer programming problem and the solution is obtained case by case by very specific algorithms. Unless you choose $P$ and $n$ such that $P/n$ is integer. – Francesco Alem. Aug 31 '17 at 14:35
• may you help in that case? – kamran jamil Sep 1 '17 at 6:09
• the optimal solution in this case is to consider the integer part of $\frac{P}{n}$ and set all $x_i$ to that value, then add $+1$ in a uniform way until the constraint is satisfied. Example $P=4$ and $n=3$, the solutions are all permutations of $1,1,2$. However i cannot demonstrate this procedure. – Francesco Alem. Sep 1 '17 at 15:25

Using weighted AM-GM inequality, $$\frac{P}{\dfrac{x_1}{x_1}+\dfrac{x_2}{x_2}+\dots+\dfrac{x_n}{x_n}}\leq\Big(\prod_{i=1}^{n}x^{x_i}_i\Big)^\dfrac{1}{P}\leq \frac{x_1\cdot x_1+x_2\cdot x_2+\dots+x_n\cdot x_n}{P}\\\Rightarrow\frac{P}{n}\leq \Big(\prod_{i=1}^{n}x^{x_i}_i\Big)^\dfrac{1}{P}\leq\frac{x_1^2+x^2_2+\dots+x_n^2}{P}\leq \frac{(x_1+x_2+\dots+x_n)^2}{P}=P$$

Hence $$\Big(\dfrac{P}{n}\Big)^P\leq\Big(\prod_{i=1}^{n}x^{x_i}_i\Big)\leq P^P$$