2
$\begingroup$

Let $n\in\mathbb{N}$. Let $\sigma_k$ ($0\leq k\leq n$) be the elementary symmetric polynomials: \begin{align} \sigma_0(x_1,\ldots,x_n):&=1 \\ \sigma_k(x_1,\ldots,x_n):&=\sum_{1\leq i_1<\cdots<i_k\leq n}x_{i_1}\cdots x_{i_k},\qquad\forall 1\leq k\leq n \end{align} I am interested in showing the quotients \begin{align} f=\frac{\sigma_k}{\sigma_{k-1}},\qquad 1\leq k\leq n \end{align} is monotone, in the sense that \begin{align} \frac{\partial f}{\partial x_i}>0,\qquad\forall 1\leq i\leq n \end{align} (Surely, I do not intend for this to hold everywhere in $\mathbb{R}^n$. Let us thus assume that the domain of $f$ is the open set $\mathcal{C}:=\left\{x\in\mathbb{R}^n:\sigma_k(x)>0\right\}$. By the Maclaurin's inequality this will imply that $\sigma_{k-1}>0$, so $f$ is well-defined on $\mathcal{C}$.)

By quotient rule, we have \begin{align} \frac{\partial f}{\partial x_i}&=\frac{1}{\sigma_{k-1}^2}\left(\sigma_{k-1}\frac{\partial\sigma_k}{\partial x_i}-\sigma_k\frac{\partial\sigma_{k-1}}{\partial x_i}\right) \\ %%% &=\frac{1}{\sigma_{k-1}^2}\left[\sigma_{k-1}\sigma_{k-1}(x_1,\ldots,\hat{x_i},\ldots,x_n)-\sigma_k\sigma_{k-2}(x_1,\ldots,\hat{x_i},\ldots,x_n)\right] \end{align} where the hat symbol denotes the absence of that component. To show the monotonicity, it remains to show that \begin{align} \sigma_{k-1}\sigma_{k-1}(x_1,\ldots,\hat{x_i},\ldots,x_n)>\sigma_k\sigma_{k-2}(x_1,\ldots,\hat{x_i},\ldots,x_n) & & (*) \end{align} If all $x_j$'s are nonnegative, then (*) can be shown easily by a simple counting argument. However, our domain is the larger set $\mathcal{C}$, which may contain some element $(x_1,\ldots,x_n)$ in which some $x_j$'s are negative. Thus I would like to seek for some approach to tackle this problem.

Any comment, hint, suggestion and answer is greatly appreciated.

Edit: Since Maclaurin's inequality can only be used if we assume that all $x_j$'s are already positive, let us modify the domain of $f$ to be $\{x\in\mathbb{R}^n:\sigma_k(x)>0\text{ and }\sigma_{k-1}(x)>0\}$. This is an intersection of two open sets, hence it is still open.

$\endgroup$
  • $\begingroup$ Wait, don't the Maclaurin's inequalities themselves hold for positive $x_i$ only? $\endgroup$ – Ivan Neretin Aug 21 '19 at 7:40
  • $\begingroup$ @IvanNeretin Ah... You are absolutely right. Thanks a lot. I have edited the question. $\endgroup$ – Hopf eccentric Aug 21 '19 at 9:44
0
$\begingroup$

For simplicity let us denote for each $1\leq i\leq n$, that $\sigma_k(\hat{i}):=\sigma_k(x_1,\ldots,\hat{x_i},\ldots,x_n)$. Then for each $1\leq k\leq n-1$, one has the identity \begin{align} \sigma_k=\sigma_{k-1}(\hat{i})x_i+\sigma_k(\hat{i}) \end{align} (Just split the sum in the definition of $\sigma_k$ into two parts, depending on whether it has $x_i$ as a factor.) For $k=n$ one has even simpler identity: \begin{align} \sigma_n=\sigma_{n-1}(\hat{i})x_i \end{align}


For $2\leq k\leq n-1$, the numerator on the R.H.S. is then \begin{align} \sigma_{k-1}\sigma_{k-1}(\hat{i})-\sigma_k\sigma_{k-2}(\hat{i}) &=\big[\sigma_{k-2}(\hat{i})x_i+\sigma_{k-1}(\hat{i})\big]\sigma_{k-1}(\hat{i}) -\big[\sigma_{k-1}(\hat{i})x_i+\sigma_k(\hat{i})\big]\sigma_{k-2}(\hat{i}) \\ &=\sigma_{k-1}(\hat{i})^2-\sigma_k(\hat{i})\sigma_{k-2}(\hat{i}) +x_i\big[\sigma_{k-2}(\hat{i})\sigma_{k-1}(\hat{i})-\sigma_{k-1}(\hat{i})\sigma_{k-2}(\hat{i})\big] \\ &=\sigma_{k-1}(\hat{i})^2-\sigma_k(\hat{i})\sigma_{k-2}(\hat{i}) \end{align} Note that the last expression is now a polynomial which does not involve $x_i$ at all. Thus let us rename $x_1,\ldots,\hat{x_i},\ldots,x_n$ as $y_1,\ldots, y_{n-1}$ so that \begin{align} \sigma_k(\hat{i})=\sigma_k(y_1,\ldots,y_{n-1}) \end{align} hence in the remaining, I will not write '$(\hat{i})$' anymore. Then, let us denote the normalization of $\sigma_k$ by \begin{align} S_k:=\frac{\sigma_k}{C_{n-1,k}}\qquad\text{where }C_{n-1,k}:=\frac{(n-1)!}{k!(n-1-k)!} \end{align} Being understood as a polynomial of $y_1,\ldots,y_{n-1}$, the R.H.S. is now \begin{align} \sigma_{k-1}^2-\sigma_k\sigma_{k-2} &=\big(C_{n-1,k}S_{k-1}\big)^2-\big(C_{n-1,k}S_k\big)\big(C_{n-1,k-2}S_{k-2}\big) \\ &=\underbrace{\frac{[(n-1)!]^2}{k!(n-1-k)!(k-2)!(n-k+1)!}}_{:=M}\left[\frac{k}{k-1}\frac{n-k+1}{n-k}S_{k-1}^2-S_kS_{k-2}\right] \end{align} In our domain of $f$, we have $\sigma_{k-1}>0$ (note: to be more explicit, we have \begin{align} \sigma_{k-1}(y_1,\ldots,y_{n-1})=\sigma_{k-1}(x_1,\ldots,x_n)\bigg|_{x_i=0}>0 \end{align} as long as $(x_1,\ldots,0,\ldots,x_n)\in \mathcal{C}$), hence $S_{k-1}>0$ as well. Thus \begin{align} R.H.S.>M(S_{k-1}^2-S_kS_{k-2}) \end{align} Now this last expression is nonnegative by the Newton's inequalities. This proves that $\displaystyle\frac{\partial f}{\partial x_i}>0$.


For $k=n$, things are even simpler. The numerator is \begin{align} \sigma_{n-1}\sigma_{n-1}(\hat{i})-\sigma_n\sigma_{n-2}(\hat{i}) &=\big[\sigma_{n-2}(\hat{i})x_i+\sigma_{n-1}(\hat{i})\big]\sigma_{n-1}(\hat{i}) -\big[\sigma_{n-1}(\hat{i})x_i\big]\sigma_{n-2}(\hat{i}) \\ &=\sigma_{n-1}(\hat{i})^2 +x_i\big[\sigma_{n-2}(\hat{i})\sigma_{n-1}(\hat{i})-\sigma_{n-1}(\hat{i})\sigma_{n-2}(\hat{i})\big] \\ &=\sigma_{n-1}(\hat{i})^2 \\ &>0 \end{align} so again, we get $\displaystyle\frac{\partial f}{\partial x_i}>0$.


Finally, the case $k=1$ is easy, since now $f=x_1+\cdots+x_n$ which clearly implies $\displaystyle\frac{\partial f}{\partial x_i}=1>0$.

$\endgroup$

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.