Stated simply, the question is:

Consider two elementary symmetric polynomials $\sigma_{k}$ and $\sigma_{k+1}$ on $\mathbb{R}^{n}$ with zero sets $U_{k}$ and $U_{k+1}$. Let $V_{i_{1}i_{2}\dotsb i_{j}}$ be the coordinate linear space $\{x\in\mathbb{R}^{n}: x_{i_{1}} = \dotsb = x_{i_{j}} = 0\}$ and $W_{n-k+1} = \cup_{i_{1} < i_{2} < \dotsb < i_{n-k+1}}V_{i_{1}\dotsb i_{n-k+1}}$, the set of all points in $\mathbb{R}^{n}$ with at least $n-k+1$ coordinates equal to zero. Does \begin{equation} (\*) \qquad U_{k} \cap U_{k+1} = W_{n-k+1} \end{equation} ? Why?

Note: I am interested in a slightly more specific question. Namely, if $\Gamma_{k}^{+}$ is the component of $\sigma_{k} > 0$ containing the point $(1,1,\dotsb,1)$, then is

\begin{equation} (\*\*) \qquad \overline{\Gamma_{k}^{+}} \cap \overline{\Gamma_{k+1}^{+}} \subset W_{n-k+1} \end{equation}

? It is well known (see "An Inequality for Hyperbolic Polynomials", Lars Garding) that \begin{equation} (\*\*\*)\qquad \Gamma_{k}^{+} \supset \Gamma_{k+1}^{+} \end{equation} and that \begin{equation} (****)\qquad \{\sigma_{k+1} > 0\} \cap \Gamma_{k}^{+} = \Gamma_{k+1}^{+} \end{equation} If $(\*)$ is true, then $(\*\*)$ follows. The equation $(\*)$ is simpler to state and appears to be true. Of course $(\*\*)$ may hold with $(\*)$ failing, but I think this is unlikely.

Background and motivation:

Numerical simulations in mathematica suggest that $(\*)$ holds. Certainly $W_{n-k+1} \subset U_{k} \cap U_{k+1}$. I can prove the reverse inclusion in certain cases. For example, denoting $x^{j} = (x_{1}^{j},\dotsb,x_{n}^{j})$ for $x\in\mathbb{R}^{n}$, notice that $W_{n-k+1}$ is exactly the zero set of $\sigma_{k}(x^{2})$. So we have

n arbitrary, k=1: From Newton's identities (see wikipedia), $\sigma_{1}^{2}(x) = \sigma_{1}(x^{2})+2\sigma_{2}(x)$. So if $\sigma_{1}(x) = \sigma_{2}(x) = 0$, then $\sigma_{1}(x^{2}) = 0$ which means $x = (0,\dotsb,0)$ so that $x \in W_{n}$.

n arbitrary, k = 2, $(\*\*)$ only: can be proved similarly to $k=1$ by writing $0 \le \sigma_{2}(x^{2}) = \sigma_{2}^{2}(x) -2\sigma_{1}(x)\sigma_{3}(x)+2\sigma_{4}(x)$. So if $\sigma_{2}(x) = \sigma_{3}(x) = 0$, then $\sigma_{2}(x^{2}) = 2\sigma_{4}(x)$. Now, $\sigma_{4}(x) \le 0$ because of $(\*\*\*)$ and $(\*\*\*\*)$. Hence $\sigma_{2}(x^{2}) = 0$ so that $x \in W_{n-1}$.

n arbitrary, $k=n-2$ and $k =n-1$ can be proved in a similar fashion. However, because of the additional terms in the equation

\begin{equation} \sigma_{k}(x^{2}) = \sigma_{k}^{2}(x)-2\sigma_{k-1}(x)\sigma_{k+1}(x)+2\sigma_{k-2}(x)\sigma_{k+2}(x)+\dotsb+(-1)^{k}2\sigma_{0}(x)\sigma_{2k}(x) \end{equation}

for other values of $k$, the above approach fails in general.

I'm actually trained as a differential geometer, so I may be approaching this problem in the wrong way (perhaps there's a technique in real algebraic geometry?). If you can suggest an alternative approach I would be grateful to hear it. This result seems rather elementary to state so I would be surprised if the result is not known. Thank you.


To establish that $U_k \cap U_{k+1} \subset W_{n-k+1}$:

Consider the polynomial $P(t)=\prod (t-x_i)$. The elementary symmetric polynomials $\sigma_i$ are its coefficients, up to the sign.

Suppose that $\sigma_k=\sigma_{k+1}=0$. This means that $0$ is a multiple root of the derivative of order $(n-k-1)$ of $P$. Now you can conclude that $0$ is a root of $P$ of multiplicity at least $n-k+1$ (that is: at least $n-k+1$ of the numbers $x_i$ are zero) by means of the following facts:

If a non-constant polynomial $Q$ has all its roots real, then so does its derivative $Q'$. Then the roots of $Q'$ are the multiple roots of $Q$, plus one root between each pair of consecutive distinct roots of $Q$, necessarily simple (otherwise the sum of the multiplicities of the roots of $Q'$ would exceed its degree). In particular, if $x$ is a multiple root of $Q'$ then it is also necessarily a root of $Q$.


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.