Title:Complex phase retrieval from subgaussian measurements

Abstract:Phase retrieval refers to the problem of reconstructing an unknown vector $x_0 \in \mathbb{C}^n$ or $x_0 \in \mathbb{R}^n $ from $m$ measurements of the form $y_i = \big\vert \langle \xi^{\left(i\right)}, x_0 \rangle \big\vert^2 $, where $ \left\{ \xi^{\left(i\right)} \right\}^m_{i=1} \subset \mathbb{C}^m $ are known measurement vectors. While Gaussian measurements allow for recovery of arbitrary signals provided the number of measurements scales at least linearly in the number of dimensions, it has been shown that ambiguities may arise for certain other classes of measurements $ \left\{ \xi^{\left(i\right)} \right\}^{m}_{i=1}$ such as Bernoulli measurements or Fourier measurements. In this paper, we will prove that even when a subgaussian vector $ \xi^{\left(i\right)} \in \mathbb{C}^m $ does not fulfill a small-ball probability assumption, the PhaseLift method is still able to reconstruct a large class of signals $x_0 \in \mathbb{R}^n$ from the measurements. This extends recent work by Krahmer and Liu from the real-valued to the complex-valued case. However, our proof strategy is quite different and we expect some of the new proof ideas to be useful in several other measurement scenarios as well. We then extend our results $x_0 \in \mathbb{C}^n $ up to an additional assumption which, as we show, is necessary.