4
$\begingroup$

Suppose we have a symmetric PD or PSD matrix M which induces an inner product $\langle \cdot, \cdot \rangle_M$. If we have that $\langle x, y \rangle > 0$ for two unit vectors $x$, $y$, are there any sufficient conditions on $x, y$, and/or $M$ that ensure that $\langle{x}, {y}\rangle_M > 0$ (other than the obvious $x=y$ or $M = I$)?.

$\endgroup$
9
$\begingroup$

Since $M$ is PSD, it can be decomposed as $M=A^T A$, where $A$ is another matrix; see https://math.stackexchange.com/questions/1801403/decomposition-of-a-positive-semidefinite-matrix

Now $\langle x,y\rangle=x^T M y=x^T A^T Ay=\langle Ax,Ay\rangle$ and so your question is equivalent to: Which matrices $A$ preserve positive correlations between vectors? It is known that the matrices that preserve orthogonality are precisely those that are scalar multiples of orthogonal transformations: https://math.stackexchange.com/questions/2355551/linear-transformations-that-preserve-orthogonality

Certainly if a matrix does not preserve orthogonality, it does not preserve positive correlations (it can fix $x$ and map $y$ to a vector orthogonal to $x$). Conversely, scalar multiples of orthogonal matrices do preserve positive correlations -- more generally, they preserve dot products: https://math.stackexchange.com/questions/2161729/how-orthogonal-matrices-preserve-dot-product-and-volume-proof

We conclude that $A$ must be a scalar multiple of an orthogonal matrix, and hence $M$ must be a positive multiple of the identity $I$.

$\endgroup$
  • $\begingroup$ Minor edit: $M$ must be a positive multiple of $I$ (previously I had "scalar"). $\endgroup$ – Aryeh Kontorovich Sep 17 '18 at 20:35
  • $\begingroup$ Thank you! Are there any assumptions on $x$ and $y$ we could make instead? $\endgroup$ – B Merlot Sep 17 '18 at 20:46
  • $\begingroup$ For all positively correlated $x,y$, you can find a "bad" $M$ that will cause them to be orthogonal under the induced inner product. $\endgroup$ – Aryeh Kontorovich Sep 17 '18 at 20:53

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.