Sparse Phase Retrieval from Magnitude Measurements

20 concepts11 papers2 proposals

In many physical imaging fields—such as X-ray crystallography, astronomy, and optical coherence tomography—detectors can only record the intensity or magnitude of a arriving wave, completely losing its phase information. Reconstructing the original signal from these intensity-only measurements is the classic phase retrieval problem.

Mathematically, this is fundamentally ill-posed. In one dimension, losing the phase means we only observe the signal's autocorrelation. Multiple distinct signals can produce the exact same Fourier transform magnitude, a ambiguity often conceptualized as "zero-flipping" in the z-transform domain. To recover the true signal, traditional methods rely on heavy oversampling, requiring several times more measurements than the signal's ambient dimension.

However, a powerful paradigm shift occurs when we know a priori that the underlying signal is sparse. By combining phase retrieval with compressive sensing, we can exploit this low-dimensional structure to eliminate phase ambiguities entirely and reconstruct signals from a fraction of the measurements traditionally required.


How Sparsity Resolves Phase Ambiguity

To understand why sparsity makes phase retrieval solvable, it helps to look at how phase ambiguities arise. When you take the Fourier magnitude of a 1-D signal, the underlying mathematical representation allows you to flip the roots (zeros) of the signal's z-transform across the unit circle without changing its Fourier magnitude. Each flip creates a completely different valid signal, leading to an exponential number of possible solutions.

Sparsity fundamentally breaks this ambiguity. If a signal is sparse and its non-zero elements are distributed irregularly, it possesses what is known as non-uniform support. As demonstrated by Jaganathan et al. in Recovery of Sparse 1-D Signals from the Magnitudes of their Fourier Transform, this non-uniformity severely constrains the allowable locations of the zeros in the z-transform. If you try to flip a zero to create an alternate signal, the resulting signal will almost always be dense or violate the strict support constraints of the original signal.

Consequently, prior knowledge of sparsity transforms an ill-posed problem into a well-posed one. Jaganathan et al. established that almost all sparse signals with aperiodic support can be uniquely recovered from their Fourier magnitudes alone, up to trivial ambiguities like global phase rotation or time shifts. Similarly, Robert Beinert showed in Sparse Phase Retrieval of One-Dimensional Signals by Prony's Method that leveraging this compact representation guarantees almost sure unique recovery for continuous-domain sparse signals like spikes or splines.


Algorithmic Frameworks: Lifting vs. Direct Recovery

Translating this theoretical uniqueness into practical algorithms requires overcoming the non-convexity of magnitude measurements ($|Ax|^2 = b$). Researchers have converged on two primary algorithmic strategies: convex relaxation via "lifting" and direct greedy support recovery.

1. Convex Relaxation and Lifting

The core idea behind lifting is to transform a non-linear, non-convex problem in a lower-dimensional space into a linear, convex problem in a higher-dimensional space. If we define a matrix $X = xx^*$, the quadratic measurement equation $b_i = |a_i^* x|^2$ can be rewritten as a linear measurement of the matrix: $b_i = \text{Tr}(a_i a_i^* X)$.

The authors of CPRL -- An Extension of Compressive Sensing to the Phase Retrieval Problem introduced the CPRL (Compressive Phase Retrieval via Lifting) framework, which combines this lifting technique with semidefinite programming (SDP). By adding an $\ell_1$-regularization penalty to the objective function, the framework minimizes both the rank of $X$ and the sparsity of the underlying signal. As explored by Ohlsson et al. in Compressive Phase Retrieval From Squared Output Measurements Via Semidefinite Programming, this regularization allows for accurate signal reconstruction in sub-Nyquist sampling regimes where traditional phase-less recovery methods completely fail.

Non-Convex Vector Space             Higher-Dimensional Convex Matrix Space
   (Find sparse vector x)                 (Find low-rank, sparse matrix X = xx*)

    |Ax|^2 = b  ───[ Lifting ]───>  Tr(A_i A_i* X) = b_i  +  Min(l1-norm)

While lifting provides strong theoretical guarantees, optimizing over an $N \times N$ matrix is computationally expensive. To mitigate this, Waldspurger et al. proposed PhaseCut, a convex relaxation that reformulates phase retrieval as a complex variant of the classic MaxCut problem. Instead of optimizing the entire matrix, PhaseCut explicitly separates the known measurement amplitudes from the unknown phases. By ignoring the phases associated with zero-amplitude measurements, the algorithm drastically reduces the variable count and stabilizes the optimization process for sparse measurement vectors.

2. Greedy and Multi-Stage Recovery

Rather than lifting the problem to a higher dimension, direct recovery methods operate in the original signal space by iteratively identifying the signal's sparse support. For instance, Baechler et al. developed a continuous-domain framework in Super Resolution Phase Retrieval for Sparse Signals. By combining Finite Rate of Innovation (FRI) theory with a greedy support recovery algorithm, they achieved "super-resolution" phase retrieval that bypasses the discretization artifacts common to grid-based methods.


Enhancing Recovery via Redundancy and Dictionary Learning

While sparsity alone restricts the solution space, combining it with structural measurement redundancy or learned representations makes recovery far more robust to noise and data constraints.

Short-Time Fourier Transform (STFT) Redundancy

Instead of taking a single global Fourier transform, Eldar et al. proposed utilizing the Short-Time Fourier Transform (STFT) in Sparse Phase Retrieval from Short-Time Fourier Measurements. The STFT applies a sliding window across the signal, creating highly overlapping, redundant local measurements. When this structural redundancy is paired with a greedy sparse recovery algorithm, it significantly eases the uniqueness conditions and dramatically boosts the probability of perfect signal reconstruction compared to standard Fourier magnitude retrieval.

Dictionary Learning for Hidden Sparsity

Signals are not always sparse in the standard time or frequency domains. To address this, Tillmann et al. introduced DOLPHIn (DictiOnary Learning for PHase retrIeval).

Instead of assuming a fixed mathematical basis, the algorithm simultaneously reconstructs the signal and learns an adaptive, sparsifying dictionary from image patches. This uncovers "hidden" sparsity within the underlying data. By enforcing this learned prior, the authors demonstrated that the oversampling ratio required for successful recovery can be drastically reduced, enabling stable convergence even in highly noisy, data-constrained environments where traditional gradient-based methods fail.


Broad Applications of the Sparsity Prior

The mathematical machinery of sparse phase retrieval extends naturally to any domain where time, depth, or spectral information must be recovered from phase-less or incomplete data:

Key Foundational Work

Sparse Phase Retrieval from Short-Time Fourier Measurements (2015) — contributed 6 concept elements to this direction.
CPRL -- An Extension of Compressive Sensing to the Phase Retrieval Problem (2012) — contributed 2 concept elements to this direction.
Compressive Phase Retrieval From Squared Output Measurements Via Semidefinite Programming (2012) — contributed 2 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_sparsity_driven_phase_retrieval_2e01d9ad,
  title = {sparsity-driven phase retrieval},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/sparsity-driven-phase-retrieval/rdr_2e01d9ad9c6b4d499f14ab18e7ba829f},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_sparsity_driven_phase_retrieval_2e01d9ad_v1,
  title = {sparsity-driven phase retrieval},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/sparsity-driven-phase-retrieval/rdr_2e01d9ad9c6b4d499f14ab18e7ba829f/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake