Random-initialization success conjecture for alternating projections in phase retrieval

Let x0Cnx_0\in\mathbb{C}^n be the target vector, let ACm×nA\in\mathbb{C}^{m\times n} be the sensing matrix, and let alternating projections reconstruct x0x_0 from the phaseless measurements Ax0|Ax_0|. An initialization is isotropic when its law is invariant under linear unitary transformations. Random-initialization success conjecture. For every fixed ϵ>0\epsilon>0, there is a sufficiently large constant C>0C>0 such that, whenever mCnm\geq Cn, alternating projections started from a random isotropic initialization converge to the true solution with probability at least 1ϵ1-\epsilon. The conjecture predicts that, although stagnation points remain in the linear-measurement regime, their attraction basins are small enough that random initialization succeeds with probability arbitrarily close to one when the measurement-to-dimension ratio is sufficiently large.

Sources & referencesView supporting material

Primary source

Irène Waldspurger, “Phase retrieval with random Gaussian sensing vectors by alternating projections”, arXiv:1609.03088 (2016).

Progress summary

Never refreshed

Nothing recorded yet. Refresh searches the literature and the public web for attempts on this problem, and writes the first summary here.

Solutions 0

No solutions have been posted yet.