Let $V$ be an $n$-dimensional real inner product space. Consider the space $V \otimes V$, endowed with the tensor product metric, i.e.

$$ \langle v_1 \otimes v_2 , w_1 \otimes w_2\rangle := \langle v_1, w_1 \rangle_V \cdot \langle v_2 ,w_2 \rangle_V.$$

Define $$S=\{ \sigma \in V \otimes V \, | \, \langle \sigma ,v \otimes v \rangle \ge 0 \, \, \text{ for every $v \in V$} \}$$

Question: Does every $\sigma \in S$ admit a representation in the form of $\sigma=\sum_i a_i v_i \otimes v_i$, for some $a_i \ge 0$ and $v_i \in V$?

Clearly every such element belong to $S$. Note that $S$ is a convex cone inside $V \otimes V$.

I tried to start with some orthonormal basis $e_i$ for $V$, and write $\sigma=a_{ij}e^i \otimes e^j$ (with the summation convention); testing against $w \otimes w$ for $w=e^i \pm e^j$ I deduced that $a_{ii} \ge 0 $ and $|a_{ij}+a_{ji}| \le a_{ii}+a_{jj} $. I am not sure however whether $\sigma \in S$ implies that $a_{ij}=0$ for $i \neq j$ in general.


No, because $\sigma$ "might be alternating". However each $\sigma$ has such representation up to an arbitrary alternating tensor. Consider the maps $\operatorname{Alt}, \operatorname{Sym} \colon V \otimes V \rightarrow V \otimes V$ given (on elementary tensors and extended linearly) by

$$ \operatorname{Alt}(v \otimes w) = \frac{v \otimes w - w \otimes v}{2}, \,\,\, \operatorname{Sym}(v \otimes w) = \frac{v \otimes w + w \otimes v}{2}. $$

One can verify that $\operatorname{Alt},\operatorname{Sym}$ are orthogonal complementary projections onto the space of alternating and symmetric tensors respectively.

Now, given $\sigma \in S$, consider the symmetric bilinear form $g_{\sigma}(v,w) := \left< \sigma, \operatorname{Sym}(v\otimes w) \right>$. Note that $$ g_{\sigma}(v,w) = \left< \operatorname{Alt}(\sigma) + \operatorname{Sym}(\sigma), \operatorname{Sym}(v \otimes w) \right> = \left< \operatorname{Sym}(\sigma), \operatorname{Sym}(v \otimes w) \right> = g_{\operatorname{Sym}(\sigma)}(v,w) $$ so that $g_\sigma$ depends only on the symmetric part of $\sigma$. By assumption, we have $$g_{\sigma}(v,v) = \left< \sigma, \operatorname{Sym}(v \otimes v) \right> = \left< \sigma, v \otimes v \right> \geq 0 $$ so $g_{\sigma}$ is positive semidefinite. Hence, we can find an orthonormal basis $(e_i)_{i=1}^n$ for $V$ such that $g_{\sigma}(e_i,e_j) = 0$ if $i \neq j$ and $g_{\sigma}(e_i,e_i) = a_i \geq 0$. If we write $\sigma$ with respect to this basis as $\sigma = \sigma^{kl} e_k \otimes e_l$ we get

$$ g_{\sigma}(e_i, e_j) = \left< \sigma^{kl} e_k \otimes e_l, \frac{e_i \otimes e_j + e_j \otimes e_i}{2} \right> = \frac{\sigma^{ij} + \sigma^{ji}}{2} = 0, \,\, i \neq j, \\ g_{\sigma}(e_i,e_i) = \left< \sigma^{kl} e_k \otimes e_l, e_i \otimes e_i \right> = \sigma^{ii} = a_i \geq 0.$$

Hence, we can write $\sigma$ as

$$ \sigma = \operatorname{Alt}(\sigma) + \sum_{i=1}^n a_i e_i \otimes e_i $$

with $a_i \geq 0$.

  • $\begingroup$ In other words, your condition puts a restriction only on the symmetric part of $\sigma$ leaving the alternating part completely arbitrary. $\endgroup$ – levap Aug 18 at 14:17
  • $\begingroup$ This is similar to how the condition $\left< Av, v \right> \geq 0$ doesn't imply that $A$ is symmetric as this condition doesn't depend on the antisymmetric part of $A$ which can be arbitrary while the symmetric part can be written as $\sum_{i=1}^n a_i e^i \otimes e_i$ with $a_i \geq 0$. $\endgroup$ – levap Aug 18 at 14:35
  • $\begingroup$ I want to add a specific counterexample: take $V = \mathbb{R}^2$ and $\sigma = e_1 \otimes e_1+e_1 \otimes e_2 + e_2 \otimes e_2$. Take any $v \in V$, we have $ \langle \sigma, v \otimes v \rangle = (v_1)^2 + v_1v_2 +(v_2)^2$ which you can write as either $(v_1-v_2)^2 + 3v_1v_2$ or $(v_1+v_2)^2 -v_1v_2$ to see that the inner product is positive for all $v_1,v_2 \in \mathbb{R}$. Secondly you cannot represent $\sigma$ in the way you want since every linear combination of tensors of the form $v \otimes v$ is symmetric and $\sigma$ is not. $\endgroup$ – abcdef Aug 18 at 14:58
  • 1
    $\begingroup$ I would just go with the counterexample $\sigma = e_1 \otimes e_2 - e_2 \otimes e_1$ as $\left< \sigma, (v_1 e_1 + v_2 e_2) \otimes (v_1 e_1 + v_2 e_2) \right> = v_1 v_2 - v_2 v_1 \equiv 0.$ and, as you write, it is not a linear combination of tensors of the form $v \otimes v$ because it is alternating and not symmetric. $\endgroup$ – levap Aug 18 at 15:10
  • $\begingroup$ @levap Thanks! Just to be sure: The fact that "we can find an orthonormal basis such that..." is essentially the theorem that every real symmetric matrix is orthogonally diagonalizable, right? $\endgroup$ – Asaf Shachar Aug 18 at 15:15

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.