Robust Sparse Recovery under Sensing Matrix Uncertainty

This direction is split into other directions: sensing matrix perturbation. Its last version stays readable and citable here.

Version 1 of 171 papers80 conceptsfrozen 2026-10-04

In the idealized world of signal processing, we assume that the mathematical model of a sensor perfectly matches its physical reality. We write down a simple linear equation, $y = Ax$, where $x$ is the signal we want, $y$ is the data we measured, and $A$ is the "forward operator"—a fixed, perfectly known matrix representing our hardware.

In practice, $A$ is often a lie.

As Henry Arguello and colleagues note in their overview of Computational Spectral Imaging, the mathematical models used to train reconstruction algorithms frequently fail to account for fabrication tolerances, lens aberrations, and alignment errors. Whether it is clock jitter in a Random Demodulator, light leakage in a lensless camera, or misaligned masks in hyperspectral imaging, the discrepancy between the theoretical model and the physical hardware—often called matrix uncertainty or model mismatch—is the primary bottleneck preventing lab-grade algorithms from working in the real world.

The Fragility of the Perfect Model

Traditional compressive sensing relies on the "sparsity" of a signal to reconstruct it from very few measurements. However, these algorithms are notoriously brittle. If the sensing matrix $A$ used during reconstruction is even slightly different from the $A_{phys}$ that actually captured the data, the optimization process tries to "force" the signal to fit the wrong grid.

Maarten De Vos and Yipeng Liu explain that this matrix noise causes standard recovery algorithms to fail by overfitting to an inaccurate model. Instead of a clean reconstruction, you get artifacts, blur, or total signal collapse. The research direction emerging across these 66 papers shifts the focus from "better signal priors" to "better hardware awareness."

Blind Calibration: Learning the Sensor on the Fly

One powerful intuition for solving this is Blind Calibration. If we don't know the exact state of our hardware, why not treat the hardware parameters as unknowns to be solved alongside the signal?

Christophe Schülke, Florent Krzakala, and Lenka Zdeborová demonstrated this with Cal-AMP, an algorithm that simultaneously infers the sparse signal and unknown sensor gains. The intuition here is that while we have two unknowns (the signal and the matrix), the signal's sparsity provides enough constraint to "anchor" the reconstruction, allowing the algorithm to wiggle the matrix parameters until the measurements make sense.

This concept has been extended to more complex systems. For instance, Mingyang Xie and Ulugbek S. Kamilov proposed a framework for joint reconstruction and calibration that uses deep-learning-based denoisers to "fix" a blurry image and a misaligned sensor at the same time.

Residual Learning and Physics-Informed Networks

A more recent trend involves using neural networks to bridge the gap between theory and reality. Rather than throwing away the physical model, researchers are using it as a "base" and training networks to learn the residual degradation.

In the RDLUF-MixS² framework, Dahua Gao and Yubo Dong use a deep unfolding approach. Instead of assuming the sensing matrix is perfect, they include a module that explicitly learns the "residual" errors—the lens aberrations and alignment shifts that the mathematical model missed.

Similarly, Jiamian Wang and Xin Yuan tackle mask uncertainty by treating the physical hardware mask as a variable within a Graph-based Self-Tuning network. This allows a single model to adapt to different physical hardware systems, even if they were manufactured with slightly different tolerances.

Why Robustness is Possible

The theoretical foundation for why these methods work lies in the stability of certain optimization problems. Alex Gutiérrez and Sam Stewart proved that for many common sparse recovery objectives, the solution set is "calm"—meaning small errors in the measurement matrix don't cause the solution to jump uncontrollably.

Furthermore, researchers like Laming Chen and Yuantao Gu have shown that greedy algorithms like Orthogonal Matching Pursuit are remarkably robust to "general perturbations." As long as the hardware error is bounded, the algorithm can still identify the "support" (the most important features) of the signal.

Summary of the Shift

The research in this area represents a transition from idealized signal processing to hardware-in-the-loop reconstruction. By acknowledging that the sensing matrix is an estimate rather than a ground truth, researchers are creating imaging systems that are:

  1. Self-Calibrating: Able to fix their own alignment and gain errors during the reconstruction process.
  2. Model-Agnostic: Capable of generalizing across different physical units of the same camera model.
  3. Physically Grounded: Using deep learning not as a black box, but as a tool to correct the specific "residuals" of physical laws.

Papers in this version

Showing 60 of 71 papers.

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_robust_sparse_recovery_under_sensing_mat_e4b95d9d,
  title = {Robust Sparse Recovery under Sensing Matrix Uncertainty},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/robust-sparse-recovery-under-sensing-matrix-uncertainty/rdr_e4b95d9df39248b8a74c5bea1bf32e28},
  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_robust_sparse_recovery_under_sensing_mat_e4b95d9d_v1,
  title = {Robust Sparse Recovery under Sensing Matrix Uncertainty},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/robust-sparse-recovery-under-sensing-matrix-uncertainty/rdr_e4b95d9df39248b8a74c5bea1bf32e28/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}