In modern wireless communications and remote sensing—ranging from Massive MIMO and millimeter-wave (mmWave) imaging to Time-of-Flight (ToF) depth sensing—systems must process massive amounts of environmental data. To prevent pilot signals from clogging network bandwidth or to avoid the slow speeds of traditional raster scanning, engineers rely heavily on Compressed Sensing (CS).
The core intuition behind compressed sensing is that the physical world is often sparse. In high-frequency wireless bands, for instance, signals are modeled as bouncing off only a few discrete objects, meaning the underlying channel matrix can be compressed into just a handful of dominant propagation paths. By leveraging mathematical frameworks like $L_1$ optimization or low-rank matrix completion, systems can theoretically reconstruct high-dimensional signals from a fraction of the measurements.
As demonstrated by Gan et al. and Shen et al., this sparsity assumption allows for drastically shorter pilot signals and reduced signaling overhead in Frequency Division Duplex (FDD) Massive MIMO systems. Similarly, Mrozack et al. utilize scene sparsity to accelerate millimeter-wave imaging through adaptive sensing frameworks that pinpoint targets on the fly.
However, a growing body of research reveals that this elegant mathematical assumption frequently collides with physical reality, leading to severe model mismatch when deployed in real-world environments.
The Physical Reality of Rich Scattering
The primary vulnerability of compressed sensing frameworks is that real-world environments are rarely perfectly sparse. Instead, they are cluttered, non-idealized, and characterized by rich scattering.
When a wireless signal or light pulse encounters a dense urban environment or a highly cluttered room, it does not travel along a few clean, specular paths. Instead, it undergoes diffuse scattering and multi-bounce reflections.
Idealized Sparse Channel (Few Discrete Paths)
Transmitter ----> Path 1 ----> Receiver
----> Path 2 ---->
Real-World Rich-Scattering Channel (Continuous/Diffuse)
Transmitter --/-> Diffuse Clutter --\-> Receiver
--/-> Multi-bounce --\->
--/-> Dense Reflections -\->
This physical reality violates the core mathematical requirements of sparse recovery algorithms across multiple domains:
- ToF Depth Sensing: In optical depth sensing, continuous and diffuse scattering means light returns via a dense spectrum of paths rather than a few discrete reflections, causing standard $L_1$ optimization models to fail (Freedman et al.).
- Angle-Domain Estimation: Simplified directional estimation models assume a few dominant angles characterize the channel. In dense multipath environments with significant interference and no clear line-of-sight, these angle-based shortcuts fail to accurately represent the complex channel state (Zhong et al.).
- Spectrum Sensing: In shared radio environments, the presence of numerous secondary users increases signal density, violating the sparsity assumptions required for effective sub-Nyquist wideband spectrum sensing (Gao et al.).
- Movable Antennas: The efficiency of directional sparsity models in 6D movable antenna systems drops significantly when encountering extremely rich scattering environments (Shao et al.).
Why Rigid Mathematical Priors Fail
To understand why traditional recovery algorithms degrade in these environments, one must look at how they enforce sparsity. Classical compressed sensing relies on fixed mathematical transforms and rigid priors (such as the $L_1$-norm or fixed maximum path constraints) to force a signal into a sparse representation.
When the physical channel deviates from this mold, the algorithms suffer from severe model mismatch. For example, low-rank tensor decomposition methods and parametric feedback frameworks assume a strict upper bound on the number of signal paths. As shown by Qian et al. and Ju et al., when the environment introduces more scattering or clutter than the model assumes, these algebraic shortcuts suffer from massive reconstruction errors.
Furthermore, techniques that attempt to actively manipulate or align individual paths run into fundamental physical limits. Zeng argues that path-alignment modulation techniques are inherently restricted to sparse bands like mmWave or Terahertz. In rich-scattering environments, the sheer number of multipath components quickly exceeds the available spatial degrees of freedom, making individual path management mathematically and physically impossible.
Ultimately, when real-world wireless channels do not exhibit perfect sparsity in standard domains, conventional optimization methods break down because they cannot adapt to unstructured noise or unmodeled path continuums (Gao et al.; Hu et al.; Ma et al.).
Beyond Rigid Priors: Moving to Data-Driven Adaptivity
Recognizing the brittleness of the channel sparsity assumption, recent literature focuses on moving away from fixed mathematical dictionaries toward data-driven architectures. Instead of forcing a channel matrix into an artificial sparse constraint, deep learning models can learn implicit, latent statistical structures directly from data.
Gunduz et al. highlight how deep learning architectures can uncover latent statistical patterns that traditional compression methods overlook, allowing for accurate data reconstruction even in non-idealized environments. When channels exhibit complex, non-linear characteristics—such as the near-field regions of Extremely Large-Scale MIMO (XL-MIMO) systems—conventional compressed sensing is highly brittle, whereas polar-domain multi-scale residual networks can robustly map the channel (Zhang et al.).
Similarly, in dense urban environments where multiple reflections create non-sparse conditions, frameworks like Graph Transformers (Singh et al.) and Hierarchical Fourier Neural Operators (Xiao et al.) outperform traditional model-based recovery. By replacing rigid sparsity assumptions with learned neural operators, these frameworks maintain high estimation accuracy regardless of whether the physical environment is sparse or densely scattered.
Key Foundational Work
Deterministic Sequences for Compressive MIMO Channel Estimation (2013) — contributed 2 concept elements to this direction.
Compressive and Adaptive Millimeter-wave SAR (2014) — contributed 1 concept elements to this direction.
SRA: Fast Removal of General Multipath for ToF Sensors (2014) — contributed 1 concept elements to this direction.