1
$\begingroup$

I am following this tutorial and in it they make a statement that the jacobian is symmetric. So, I have $\partial f(\mathbf{x}) / \partial {x}$, and I want to mathematically prove that it is symmetric. Here, $f(\mathbf{x})$ is softmax function.

I know that $f(\mathbf{x}) = f(\mathbf{x})^T$ must be satisfied to prove that two matrices are symmetric. However, I'm not sure how to mathematically begin doing this. Should I be re-deriving softmax in a different way to incorporate the "transposed" portion of it?

$\endgroup$

1 Answer 1

1
$\begingroup$

The off-diagonal $i,j$ entries of the Jacobian in this specific case are $$ -s_i(\mathbf x)s_j(\mathbf x) $$ as mentioned in the paragraph just before where they state that the Jacobian is symmetric. This is clearly equal to $$ -s_j(\mathbf x)s_i(\mathbf x) $$ showing that the Jacobian in this specific case is symmetric.

In general, Jacobian matrices (even square ones) are not symmetric.

$\endgroup$
1
  • $\begingroup$ Ahh, I see, that was actually a simple and straightforward way, thank you $\endgroup$
    – Jonathan
    Commented Sep 28, 2019 at 21:32

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .