ProbXiv
Explorersign in
Problem archiveProblem record

Statement

The fact that sometimes linear operators A are available only via black-box implementations, i.e. one can give some input x to a program and gets back a vector Ax, and that the respective adjoint are available via other black-box implementations which give V ∗ y for vectors y raises some other questions: How can one compute kAk (which is needed in the bounds for the step-sizes) if only black boxes are available and there is no access to an adjoint? An even more difficult question is, how to compute kA − V k if only a black-box for A and another one for V ∗ is available. For the first question we propose to use projected stochastic gradient ascent for the problem (1) max kAvk2 . kvk=1 The gradient of the objective is AT Av, but under our assumption, we can not compute it. However, we can get an unbiased estimate by using a trick: Let x be a random vector of appropriate size such that E(xxT ) = I. Then it holds that E (Av)T (Ax)x = E xxT AT Av = AT Av.

Record

Comments

Nothing has been published against this problem yet, and nothing has been said about it.

Sign in with an institutional address to take part in the discussion. Reading every thread stays open to everyone.

Sign in

Solve with an agent

Open the statement in a chat, with the problem and the ground rules already written into the prompt.

This opens a third-party site. Nothing is posted back to ProbXiv and nothing you write there is recorded here — what a model gives you is an attempt, which a person still has to check.