Abstract
This paper introduces Multiple Illumination Phaseless Super-Resolution (MIPS), a mathematical framework designed to reconstruct high-resolution signals from low-frequency intensity-only measurements. By addressing the simultaneous challenges of the diffraction limit and the phase retrieval problem, MIPS enables stable signal recovery using a set of five physically realizable masks. We demonstrate the effectiveness of this approach through semidefinite relaxation and convex optimization, providing theoretical guarantees for unique recovery and validating the method via numerical simulations in applications such as Direction-of-Arrival (DoA) estimation and Coherent Diffraction Imaging (CDI).
Introduction
In advanced sensing applications ranging from medical imaging to astronomical observation, researchers frequently encounter two fundamental physical constraints. The first is the diffraction limit, which prevents the capture of high-frequency details, resulting in blurred, low-pass filtered signals. The second is the phase problem, where sensors like optical detectors or standard radio receivers can only record wave intensity (magnitude) while losing critical phase information.
Recovering high-resolution signals from low-resolution, phaseless data—a combination of super-resolution and phase retrieval—is a significant challenge. While theoretical solutions involving signal modulation via "masks" have been proposed, they often relied on physically impractical configurations. MIPS bridges this gap by introducing a flexible masking framework that is both mathematically sound and physically implementable.
Methodology
The MIPS framework utilizes a specific set of five masks constructed from simple diagonal matrices. These masks act as "building blocks" corresponding to phase shifts, ensuring that the signal's phase information can be uniquely recovered from intensity measurements.
To solve the resulting non-linear, quadratic recovery problem, the authors employ a technique known as semidefinite relaxation or "lifting." This process transforms the problem of finding a signal vector $x$ into a linear problem of finding a rank-one matrix $X = xx^*$. The recovery is then performed using convex optimization, specifically $l_1$-norm minimization to promote signal sparsity and semidefinite programming (SDP) to satisfy measurement constraints. A key theoretical contribution is Theorem 2.1, which proves that if a signal is sufficiently sparse and the masks satisfy specific "co-prime" conditions, the recovery is unique and stable.
Experimental Validation
The MIPS framework was validated through extensive numerical simulations. Synthetic sparse signals of lengths $N=20$ and $N=40$ were generated with non-zero entries at random positions, adhering to a minimum separation condition. To test robustness, i.i.d. standard normal noise was added to the measurements across various signal-to-noise ratios (SNR).
Results and Discussion
The simulations revealed two critical performance characteristics:
- Phase Transition: In noiseless environments, the algorithm exhibits a distinct "phase transition." Successful reconstruction occurs with high probability when the number of low-frequency measurements ($K$) is sufficiently large relative to the signal's sparsity and the minimum separation ($\Delta$) between components.
- Noise Robustness: Under noisy conditions, the Mean-Squared Error (MSE) decreases linearly as the SNR increases on a log-log scale. This linear relationship confirms that MIPS is a stable reconstruction method suitable for real-world applications where measurement errors are inevitable.
Physical Realizability
A major insight of this work is the translation of mathematical masking into physical setups. In Direction-of-Arrival (DoA) estimation, the masks correspond to placing five transmitters at specific locations with defined phase offsets. In Coherent Diffraction Imaging (CDI), the same effect is achieved by illuminating a sample from five strategic angles. The use of "co-prime" mask parameters ensures that the underlying measurement graph is connected, allowing the algorithm to "lock in" the correct signal phase.
Conclusion and Future Work
MIPS provides a robust framework for overcoming the "double whammy" of diffraction limits and missing phase data. While the current implementation relies on signal sparsity and a minimum separation between components, it offers a clear path toward practical high-resolution imaging. Future research will focus on scaling the method for dense or complex 2D/3D signals by developing more efficient first-order optimization algorithms to replace the computationally intensive semidefinite programming currently employed.
Summary updated