Abstract:This work takes the first steps towards solving the "phaseless subspace tracking" (PST) problem. PST involves recovering a time sequence of signals (or images) from phaseless linear projections of each signal under the following structural assumption: the signal sequence is generated from a much lower dimensional subspace (than the signal dimension) and this subspace can change over time, albeit gradually. It can be simply understood as a dynamic (time-varying subspace) extension of the low-rank phase retrieval problem studied in recent work.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is the "Phaseless Subspace Tracking (PST)" problem. Specifically, PST involves recovering a series of time - varying signals (or images) from the phaseless linear projections of each signal, and it is assumed that these signal sequences are generated from a low - dimensional subspace, which can change slowly over time.
### Problem Background
The traditional Phase Retrieval (PR) problem occurs in many applications, such as ptychography, crystallography, and astronomy. The goal of PR is to recover the original signal from the magnitudes of the discrete Fourier transform (DFT) coefficients of the signal. Generalized PR uses the inner product of an arbitrary measurement vector $\mathbf{a}_i$ instead of DFT. Therefore, the goal is to recover $\mathbf{x}$ from $|\langle \mathbf{a}_i, \mathbf{x}\rangle|$, where $i = 1, 2,\cdots, m$.
To reduce the required number of measurements $m$, recent studies have introduced structural assumptions, such as sparsity and low - rankness. Low - rankness is used in two ways:
1. Assume that a single signal (or image) is low - rank after being rearranged into a matrix.
2. Assume that the matrix composed of a series of signals (or images) is low - rank, and each signal / image is a column of the matrix. The measurement is the phaseless linear projection of each signal or image.
### Dynamic Low - Rank Phase Retrieval (Dynamic LRPR or PST)
PST is a dynamic extension of Low - Rank Phase Retrieval (LRPR). Unlike LRPR, PST assumes that the subspace generating the signal can change slowly over time. Specifically, assume that the signal $\mathbf{x}_t$ can be represented as $\mathbf{x}_t=\mathbf{U}_t\mathbf{b}_t$, where $\mathbf{U}_t$ is a basis matrix and $\mathbf{b}_t$ is a coefficient vector. For each $t$, we have $m$ phaseless measurement values $y_{i,t}=|\langle \mathbf{a}_{i,t},\mathbf{x}_t\rangle|$.
### Research Contributions
The main contributions of this paper are:
1. **Phaseless Subspace Change Detection**: Given an accurate estimate of the previous subspace, detect whether the subspace has changed.
2. **Phaseless PCA with Partial Subspace Knowledge**: Given an accurate estimate of the previous subspace and knowing that the subspace has changed, estimate the new subspace with as few measurements as possible (with as short a tracking delay as possible).
### Solutions
The paper proposes two methods to solve the above two sub - problems:
- **Automatic Phaseless Subspace Change Detection**: Detect whether the subspace has changed by calculating the largest eigenvalue $\lambda_1(\tilde{\mathbf{Y}}_U)$ of the matrix $\tilde{\mathbf{Y}}_U$ and comparing it with the threshold $C\cdot\text{tr}(\bar{\Lambda})$.
- **Phaseless PCA with Partial Subspace Knowledge**: After detecting the change, estimate the current subspace by an alternating minimization algorithm, use the previously estimated subspace $\hat{\mathbf{U}}_{j - 1}$ to construct a larger subspace matrix $\tilde{\mathbf{U}}_j$, and gradually optimize the subspace and coefficients.
### Experimental Results
The experimental results show that the PST algorithm performs well in phaseless subspace change detection and phaseless PCA with partial subspace knowledge, especially when the initial subspace error is small. Compared with the existing LRPR - AltMin and TWF algorithms, PST - PCA in the reduction