0
$\begingroup$

Consider a twice differentiable strongly convex function $f:\mathbb{R}^n \rightarrow \mathbb{R^+}$ that attains its minimum value at the point $x^*$. I am wondering if one can compute a direction of slowest ascent $u$ from the point $x^*$.

Intuitively, the direction of slowest ascent would be a unit vector $u$ such that $f(x^*+u)$ is as close as possible to $f(x^*)$ (or at least, we would expect $f(x^*+u)$ and $f(x^*)$ to be close based on the information provided by the Hessian at $x^*$).

I am having some difficulty formally defining $u$ since it depends on the size of the step that we take away from $x^*$. That said, a poor man's definition would be:

$$ u = \text{argmin}_{\|w\| \neq 0} ~~f(x^* + w) - f(x^*)$$

$\endgroup$
1
  • 4
    $\begingroup$ Near the minimum $f-f(x^*)$ is equivalent to a positive quadratic form, given by the Hessian. This defines a family of hyper-ellipsoids, with coinciding center and axis. The corresponding directions form a diagonalizing orthonormal basis. One candidate vector is probably a unit eigenvector corresponding to smallest eigenvalues. $\endgroup$ Sep 15, 2015 at 19:51

1 Answer 1

1
$\begingroup$

There may not be a unique 'direction of slowest ascent', take for example the simple case of $f(\vec{x}) = x_0^2 + x_1^2$, in which case the minimum is clearly (0,0), and any direction is equally fast. That said, near enough to the minimum, the you can define a 'subspace of slowest ascent' which is basically just the eigenspace associated with the lowest eigenvalue of the Hessian. If the lowest eigenvalue is non-degenerate, then the direction you seek is well defined.

$\endgroup$

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.