Let $\rho : \mathbb{R}^n\to \mathfrak{so}(2m)$ be a faithful representation of the commutative Lie algebra $\mathbb{R}^n$ into the Lie algebra of skew-symmetric matrices. There is an orthonormal basis $d_1,\ldots,d_{2m}$ of $\mathbb{R}^{2m}$ such that for all $i=1,\ldots,n$ and $j=1,\ldots,m$ $$\rho(e_i)(d_{2j-1})=\lambda_{ij}d_{2j}\ \text{and}\ \rho(e_i)(d_{2j})=-\lambda_{ij}d_{2j-1}$$ for some real numbers $\lambda_{ij}$. The degeneracy property of this representation is as follows \begin{equation}\label{deg}\tag{1} \lambda_{ki}^2=\lambda_{kj}^2\ \text{for all}\ k=1,\ldots,n\ \text{and for all basis of}\ \mathbb{R}^n. \end{equation} If $v_k=\sum_{p=1}^na_{pk}e_p$ $k=1,\ldots,n$ is another basis of $\mathbb{R}^n$, then \begin{align*} \text{$\eqref{deg}$}& \Longleftrightarrow \left(\sum_{p=1}^na_{pk}\lambda_{pi}\right)^2-\left(\sum_{p=1}^na_{pk}\lambda_{pj}\right)^2=0\ \text{for all}\ k=1,\ldots,n\\ & \Longleftrightarrow \sum_{p=1}^na_{pk}^2(\lambda_{ki}^2-\lambda_{kj}^2)+2\sum_{1\leq p<q\leq n}a_{pk}a_{qk}\left(\lambda_{pi}\lambda_{qi}-\lambda_{pj}\lambda_{qj}\right)=0\ \text{for all}\ k=1,\ldots,n\\ & \Longleftrightarrow \sum_{1\leq p<q\leq n}a_{pk}a_{qk}\left(\lambda_{pi}\lambda_{qi}-\lambda_{pj}\lambda_{qj}\right)=0\ \text{for all}\ k=1,\ldots,n \end{align*}

So I am studying the following question: $$\sum_{1\leq i<j\leq n}a_{ik}a_{jk}x_{ij}=0\ \text{for all invertible matrix}\ A=(a_{ij})$$

  1. Does this imply that $x_{ij}=0$ for all $1\leq i<j\leq n$?
  2. Is it possible to write the equality above using some quadratic form?

1 Answer 1


Suppose $V = \mathbb{R}^n$ has a basis $(e_1,\dots,e_n)$. Your assumption is that you have a family of linear maps $\lambda_1,\dots,\lambda_m \in V^*$ which are defined such that $\lambda_r(e_i) = \lambda_{ir}$ for any $1 \leqslant r \leqslant m$. Working inside the algebra of all functions $\mathrm{Fun}(V,\mathbb{R})$, with multiplication given by pointwise evaluation, we have a function

$$q_{rs} = \lambda_r\lambda_r - \lambda_s\lambda_s$$

for any $1 \leqslant r,s \leqslant m$. This is, indeed, a quadratic form on $V$.

If $(e_1^*,\dots,e_n^*)$ is the dual basis of $V^*$ then $\lambda_r = \sum_{i=1}^n \lambda_{ir}e_i^*$ and

$$\lambda_r\lambda_r = \sum_{i=1}^n \lambda_{ir}^2e_i^* + 2\sum_{1 \leqslant i < j \leqslant n} \lambda_{ir}\lambda_{jr}e_i^*e_j^*$$

Your assumption that $\lambda_{ir}^2 = \lambda_{is}^2$ for any $1 \leqslant r,s \leqslant m$ means that

$$q_{rs} = 2\sum_{1 \leqslant i < j \leqslant n} (\lambda_{ir}\lambda_{jr}-\lambda_{is}\lambda_{js})e_i^*e_j^*$$

As you do, we could divide by $2$ to get a quadratic form $q_{rs}'$ so that $q_{rs} = 2q_{rs}'$.

If $g \in \mathrm{End}(V)$ then $ge_k = \sum_{i=1} a_{ki}e_i$ for some $a_{ki} \in \mathbb{R}$ and we have $e_i^*(ge_k)e_j^*(ge_k) = a_{ki}a_{kj}$ so you have

$$q_{rs}'(ge_k) = \sum_{1 \leqslant i < j \leqslant n} a_{ki}a_{kj}(\lambda_{ir}\lambda_{jr}-\lambda_{is}\lambda_{js})$$

So your condition reads that $q_{rs}'(ge_k) = 0$ for all $g \in \mathrm{GL}(V)$.

Now let $q : V \to \mathbb{R}$ be a quadratic form. Your question amounts to asking whether $q = 0$ given that $q(ge_i) = 0$ for all $g \in \mathrm{GL}(V)$ and $1 \leqslant i \leqslant n$. If we were working over $\mathbb{C}$ then we could show this by viewing $V$ as an affine space and using that $q$ is a polynomial function. So as $\mathrm{GL}(V)$ is dense in $\mathrm{End}(V)$ we must have $q = 0$. Probably there is an equivalent argument over $\mathbb{R}$ but I'll leave this to someone else.

One can instead argue as follows. For any $i \neq j$ we see that the map $\tau_{ij} \in \mathrm{End}(V)$ given by $\tau_{ij}e_k = e_k + \delta_{ki}e_j$ is invertible and $\tau_{ij}e_i = e_i+e_j$. Hence $q(e_i+e_j) = 0$ for any $1 \leqslant i,j \leqslant n$ (if $i=j$ then this is $q(2e_i) = 4q(e_i) = 0$).

Now let $\beta(u,v) = q(u+v) - q(u) - q(v)$ be the symmetric bilinear form determined by $q$. Then for any $1 \leqslant i,j \leqslant n$ we have

$$\beta(e_i,e_j) = q(e_i+e_j)-q(e_i) - q(e_j) = 0$$

Using the bilinearity we see that $\beta(u,v) = 0$ for any $u,v \in V$ which shows that $q(v) = \frac{1}{2}\beta(v,v) = 0$ for any $v \in V$. Thus $q = 0$ is identically zero.

  • $\begingroup$ Many thanks Jay, but $q_{rs}$ is already identically zero: $q_{rs}(e_i)=0$ for all $i=1,\ldots,n$ also $e_i^*e_j^*=0$ $\endgroup$
    – user56980
    Jul 18 at 16:38
  • 1
    $\begingroup$ That doesn't imply $q_{rs} = 0$ because $q_{rs}$ is not linear. Take $n=2$ and let $\beta$ be the symmetric bilinear form with $\beta(e_1,e_2) = \beta(e_2,e_1) = 1$ and $\beta(e_1,e_1) = \beta(e_2,e_2) = 0$. We have a quadratic form given by $q(v) = \frac{1}{2}\beta(v,v)$ for all $v \in V$. Certainly $q(e_1) = q(e_2) = 0$ but $q(e_1+e_2) = \frac{1}{2}\beta(e_1+e_2,e_1+e_2) = 1 \neq 0$. So it is not enough to check this on a basis. $\endgroup$
    – Jay Taylor
    Jul 18 at 16:51
  • 1
    $\begingroup$ Relatedly $e_i^*e_j^*$ is also not zero. Let $v = e_i + e_j$ then $e_i^*(v) = e_j^*(v) = 1$ and so $(e_i^*e_j^*)(v) = e_i^*(v)e_j^*(v) = 1\cdot 1 = 1$. Recall we evaluate the product pointwise. In fact, in the example in the previous comment the quadratic form is $q = e_1^*e_2^*$. $\endgroup$
    – Jay Taylor
    Jul 18 at 16:55
  • $\begingroup$ Thank you. I reacted quickly, of course $q_{rs}$ is not linear! $\endgroup$
    – user56980
    Jul 18 at 17:32
  • $\begingroup$ I can't edit the nice answer $\lambda_r^2=...e_i^{*2}$ instead of $e_i^*$ $\endgroup$
    – user56980
    Jul 18 at 18:04

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.