Physically Feasible Sampling Trajectories for Signal Acquisition

50 concepts34 papers3 proposals

In the theoretical world of compressed sensing, the most efficient way to reconstruct a signal is often to sample it at random. By capturing a small, incoherent subset of data, algorithms can perfectly recover high-resolution images or complex waveforms. However, as Donoho et al. point out, there is a significant "sparsity penalty" when moving from these idealized random matrices to the structured, anisotropic sampling required by real-world devices.

The fundamental challenge is that physical hardware—whether an MRI scanner, a CT gantry, or a robotic arm—cannot teleport between arbitrary data points. It is bound by the laws of physics: inertia, maximum acceleration, and energy dissipation. Research in hardware-constrained sampling seeks to bridge this gap, ensuring that the "optimal" sampling patterns suggested by mathematics are actually executable by machines.

The Gap Between Points and Paths

Most signal acquisition hardware is optimized for continuity. In Magnetic Resonance Imaging (MRI), for instance, data is collected by varying magnetic gradients. As Jeromin et al. demonstrate, purely stochastic sampling patterns can exceed the physical capabilities of these gradients, inducing mechanical vibrations and image artifacts. Instead, the authors show that deterministic spiral geometries often provide a better balance between mathematical efficiency and hardware stability.

This constraint is not limited to medical imaging. In Electron Microscopy, Pjescic-Emedji et al. note that hardware is typically designed to scan contiguous blocks rather than isolated pixels. Similarly, in X-ray Fluorescence (XRF) imaging, Dai explains that while adaptive sampling can identify "informative" pixels to skip, the physical scanning probes are often restricted to rigid raster paths. When the algorithm asks for a jump, the hardware must either slow down or risk inaccuracy.

Integrating Physics into the Optimization Loop

To solve this, researchers have moved toward "hardware-aware" optimization, where the physical limits of the machine are treated as differentiable constraints within the learning process.

One prominent approach is to parameterize the sampling trajectory itself. Wang et al. introduced BJORK, a framework that uses B-splines to define MRI sampling paths. By optimizing these splines, the system ensures that the resulting trajectories are inherently smooth. To handle the strict limits of the hardware, Shor et al. utilize projection algorithms within their Multi-PILOT framework to keep learned paths within the scanner's maximum gradient strength and slew-rate (the speed at which gradients can change).

A more sophisticated method involves treating the sampling process as a dynamical system. Peng et al. employ Neural Ordinary Differential Equations (Neural ODEs) to learn k-space trajectories. By baking the physics of the scanner directly into the ODE, the optimization process acts as a regularizer, preventing the model from ever proposing a trajectory that would be physically impossible or damaging to the hardware.

Beyond Imaging: Robotics and Quantum Control

The tension between optimal information gain and physical feasibility extends to any system controlled by actuators. In robotics, Batke et al. found that teaching a bipedal robot to run requires incorporating "transferability constraints"—such as ground friction and force change rates—into the trajectory optimization. Without these, the reinforcement learning agent might discover a mathematically "fast" gait that relies on unrealistic physical interactions, causing the robot to fail in the real world.

Similarly, in quantum state preparation, Goldschmidt et al. use Model Predictive Control to navigate hardware limits like pulse amplitudes. By integrating these constraints natively, the optimization avoids generating control sequences that the quantum hardware cannot physically produce.

The Path Forward: Joint Optimization

The emerging consensus across these fields is that acquisition and reconstruction cannot be treated as separate steps. Frameworks like LOUPE, proposed by Dalca et al., and LARO, developed by Zhang et al., jointly optimize the sampling mask and the reconstruction network.

By forcing the neural network to "see" the hardware constraints during training—such as restricting sampling to full lines of data (Weiss et al.) or using fan-beam ordering to minimize gradient jumps—researchers are creating systems that are not just theoretically superior, but clinically and mechanically ready for deployment. The goal is no longer just to sample less, but to sample smarter within the boundaries of the physical world.

Key Foundational Work

HYR2PICS: Hybrid regularized reconstruction for combined parallel imaging and compressive sensing in MRI (2012) — contributed 2 concept elements to this direction.
Optimal compressed sensing reconstructions of fMRI using 2D deterministic and stochastic sampling geometries (2012) — contributed 2 concept elements to this direction.
Variable Density Sampling with Continuous Trajectories (2014) — contributed 1 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_hardware_constrained_sampling_f81e307c,
  title = {hardware-constrained sampling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/hardware-constrained-sampling/rdr_f81e307ca1e140ee8e477ac78d88bed3},
  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_hardware_constrained_sampling_f81e307c_v1,
  title = {hardware-constrained sampling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/hardware-constrained-sampling/rdr_f81e307ca1e140ee8e477ac78d88bed3/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