As wireless systems evolve toward 6G, the number of antennas at base stations is scaling from dozens to hundreds or even thousands—a trend known as Massive MIMO and Extremely Large-Scale MIMO (XL-MIMO). While more antennas provide higher data rates and better beamforming, they also create a "curse of dimensionality." Processing signals across thousands of antenna elements in real-time is computationally prohibitive and requires an enormous amount of pilot overhead.
The core insight across recent research is that while the antenna-domain signal is high-dimensional, the underlying physical environment is sparse. By transforming signals from the spatial antenna domain into a "transform domain"—such as the angular, beamspace, or wavenumber domain—researchers can concentrate signal energy into a few coefficients, enabling dramatic reductions in complexity without sacrificing performance.
From Antennas to Angles
In a traditional antenna array, a signal is represented as a vector of complex numbers, one for each antenna element. In this representation, the signal is "spread out." However, Ko-Feng Chen and Yu T. Su observe that because wireless signals typically arrive from a limited number of directions, the channel exhibits high spatial correlation.
By applying a fixed unitary basis, such as the Discrete Cosine Transform (DCT) or Discrete Fourier Transform (DFT), the signal is mapped into the angular domain. In this domain, the signal is no longer a dense vector; it becomes a sparse set of peaks corresponding to the Angles of Arrival (AoA). Ko-Feng Chen and Yu T. Su demonstrate that this "energy compaction" allows for reduced-rank channel estimation, where the system only needs to estimate a small number of significant coefficients rather than the entire high-dimensional matrix.
Denoising in the Beamspace
The sparsity of the angular domain does more than just reduce dimensionality; it provides a natural mechanism for noise suppression. In the antenna domain, noise is present on every element. When transformed into the "beamspace" (the angular domain for millimeter-wave signals), the signal power is concentrated into a few "bins," while the noise remains distributed across all bins.
This is the intuition behind the BEACHES (BEAmspace CHannel EStimation) algorithm developed by Alexandra Gallyas-Sanhueza, Seyed Hadi Mirfarshbafan, Christoph Studer, and Ramina Ghods. By using a statistical tool called Stein’s Unbiased Risk Estimator (SURE), the algorithm adaptively identifies which beamspace components contain signal and which contain only noise. This allows for near-optimal denoising at a fraction of the cost of traditional methods—running up to 58,000x faster than state-of-the-art iterative solvers. Seyed Hadi Mirfarshbafan further showed that this sparsity can be exploited for linear equalization, slashing base station power consumption by up to 66%.
Expanding to the Near Field
As arrays grow larger (XL-MIMO) and frequencies get higher (mmWave/THz), the "far-field" assumption—where signals arrive as flat plane waves—begins to fail. In the "near-field," signals arrive as spherical wavefronts, meaning the distance to the user matters as much as the angle.
To handle this, researchers have moved beyond simple angular transforms:
- The Wavenumber Domain: Hao Jiang, Yuanwei Liu, and Zhaolin Wang propose transforming near-field signals into the wavenumber domain using Fourier plane-wave expansions. They found that near-field channels, which look complex in the spatial domain, exhibit a sparse and nearly diagonal structure in the wavenumber domain. Their "Sense-Then-Train" (STT) framework exploits this to perform beam alignment with significantly lower overhead.
- The Polar Domain: Jiayi Zhang and Huahua Xiao introduce polar-domain transforms that sample both angle and distance. This representation allows deep learning models, like their P-MSRDN network, to efficiently estimate channels by exploiting the sparsity of spherical wavefronts.
- Wavenumber-Division Multiplexing (WDM): Luca Sanguinetti, Antonio Alberto D'Amico, and Mérouane Debbah treat space similarly to how OFDM treats time. By using Fourier basis functions to represent spatial signals, they enable low-complexity communication over continuous electromagnetic surfaces (holographic MIMO).
Enhancing Generative Models
Transform-domain sparsity is also proving to be a powerful "pre-conditioner" for machine learning. In channel estimation tasks, Zehua Jiang and Fenghao Zhu show that transforming data into the angular domain via FFT simplifies the learning task for generative models. By focusing the neural network on a compressed set of features rather than raw antenna data, their Average Velocity Field (AVF) model achieves higher accuracy with 90% lower latency. Similarly, Amar Kasibovic and Benedikt Fesl use angular-domain representations within Variational Autoencoders (VAEs) to distinguish between desired signals and interference, effectively mitigating pilot contamination.
Summary of Intuition
The shift from antenna-domain processing to transform-domain processing is analogous to the shift from raw pixel editing to JPEG compression. By moving to a domain where the signal is sparse (Angular, Beamspace, or Wavenumber):
- Complexity drops because we only process the "peaks" (high-energy components).
- Noise is filtered because it doesn't share the signal's structured sparsity.
- Learning is faster because the model sees a simplified, low-dimensional version of the world.