As wireless networks transition toward 6G, antenna architectures are expanding dramatically. We are moving from standard multi-antenna systems to Massive MIMO, extremely large-scale antenna arrays (ELAAs), and even Continuous-Aperture Arrays (CAPAs) that turn entire environmental surfaces into continuous electromagnetic sensors.
However, this spatial explosion introduces a severe computational bottleneck. In an array with $N$ elements, the search space for optimizing beamforming vectors or matrices typically scales quadratically ($O(N^2)$) or cubically ($O(N^3)$) with the number of antennas. For continuous surfaces, the problem theoretically scales to infinite dimensions.
Subspace-constrained beamforming resolves this paradox. The core intuition is simple: while the physical antenna array may have hundreds of elements or infinite degrees of freedom, the underlying wireless environment is mathematically sparse. Signals propagate along a limited number of spatial paths to reach a small number of users or targets. Consequently, the optimal beamforming solution does not require a high-dimensional search; it is inherently trapped within a low-dimensional subspace spanned by the physical channel responses. By projecting the optimization problem into this lower-dimensional subspace, we can decouple computational complexity from the physical antenna count entirely.
Shifting Complexity from Antennas to Users
To understand why this works, consider a multi-group multicast scenario where a base station with a massive antenna array sends data to multiple groups of users. Traditionally, finding the optimal beamforming vector for each group requires searching across the entire high-dimensional space of the antenna array.
In Multi-group Multicast Beamforming: Optimal Structure and Efficient Algorithms, Min Dong et al. proved that the optimal beamformer possesses an inherent low-dimensional structure: it is always a weighted minimum mean square error (MMSE) filter that resides strictly within the subspace spanned by the channel responses of the users within that specific group.
$$\mathbf{w}^* \in \text{span}\left(\{\mathbf{h}_i\}_{i \in \mathcal{G}}\right)$$
Because any component of the beamforming vector orthogonal to this user-channel subspace contributes only to wasted power or unintended interference, the optimal solution can be expressed purely as a linear combination of the user channels. This mathematical reality transforms the optimization landscape:
- Before: The search space dimension equals the number of physical antennas ($N$).
- After: The search space dimension collapses to the number of users per group ($K$), which is typically orders of magnitude smaller.
Leveraging this structural insight, Chong Zhang et al. developed fast first-order algorithms for max-min fair multicasting. By constraining the search space to this weighted MMSE filter structure, the computational burden shifts from scaling with the physical array size to scaling with the user count, enabling near-optimal beamforming to be calculated up to 30 times faster. A similar principle applies to multi-user interference channels; Juho Park et al. demonstrated that the search space for Pareto-optimal beams can be restricted to the union of the eigen-spaces of the channels to all receivers, rendering the optimization complexity independent of the transmit antenna count once the array becomes sufficiently large.
Collapsing the Search Space to Geometric Subspaces
When dealing with interference constraints—such as in cognitive radio networks where a secondary user must avoid disrupting a primary user—the dimensionality can be reduced even further into tight, low-dimensional geometric constructs.
In Robust Cognitive Beamforming With Partial Channel State Information, Lan Zhang et al. analyzed a system where the transmitter only possesses partial channel knowledge (the statistical mean and covariance) of the interference link. The authors proved that the optimal robust beamforming vector is strictly constrained to a two-dimensional subspace defined entirely by two vectors: the transmitter's own direct channel and the mean channel of the primary user.
By projecting the high-dimensional antenna optimization problem onto this 2D plane, a complex, constrained optimization problem collapses into a simple geometric task. This allows for closed-form analytical solutions that achieve identical performance to standard optimization toolboxes at a fraction of the computational cost.
Similarly, in the context of cooperative relaying, Rui Zhang et al. showed that by applying singular value decomposition (SVD) to the channel matrices of a two-way multi-antenna relay, the optimization of a large beamforming matrix can be restricted to a small, fixed-size subspace. This reduction collapses the optimization complexity from the square of the antenna count down to a constant number of variables.
Taming Infinite Dimensions in Continuous Apertures
The power of subspace constraints becomes most apparent when moving from discrete antenna elements to Continuous-Aperture Arrays (CAPAs). In CAPA systems, the traditional beamforming vector is replaced by a continuous source current distribution function across an entire surface, introducing theoretically infinite degrees of freedom.
Remarkably, the same subspace principles hold true. In Optimal Beamforming for Multi-User Continuous Aperture Array (CAPA) Systems, Zhaolin Wang et al. proved that the optimal continuous beamforming function is always a linear combination of the active users' continuous spatial channel responses. This allows researchers to map an infinite-dimensional functional optimization problem directly onto a finite-dimensional matrix calculation.
Infinite-Dimensional Continuous Surface
│
▼ [Project onto User/Target Spatial Responses]
Low-Dimensional Signal Subspace (Size ∝ Number of Users/Targets)
│
▼ [Solve Manageable Matrix Operations]
Optimal Continuous Beamforming Function
This reduction makes real-time processing viable for both communication and sensing applications:
- Linear Receive Beamforming: Chongjun Ouyang et al. demonstrated that continuous receive beamforming techniques (like MRC, ZF, and MMSE) can be simplified into finite-dimensional problems by focusing exclusively on the signal subspace spanned by the users' spatial responses.
- Near-Field Sensing: For holographic near-field sensing, Hao Jiang et al. showed that the optimal source current distribution is a linear combination of the array responses associated with the targets. Consequently, the optimization complexity scales with the number of detected objects rather than the physical resolution of the continuous surface.
- Integrated Sensing and Communication (ISAC): In joint communication and radar systems, Boqun Zhao et al. established that projecting continuous-aperture signals into a low-dimensional subspace spanned by the communication and sensing channels reduces computational complexity by several orders of magnitude, keeping processing times independent of the physical aperture size or the carrier frequency.
Subspace Tracking and Codebook Design
Beyond calculating instantaneous beamformers, constraining operations to a dominant subspace streamlines hardware tracking and feedback loops.
For instance, in Frequency Division Duplexing (FDD) Massive MIMO systems, base stations require channel feedback from mobile users to construct accurate beams. In Channel Feedback Based on AoD-Adaptive Subspace Codebook in FDD Massive MIMO Systems, Wenqian Shen et al. exploited the fact that physical angles of departure (AoDs) change much more slowly than instantaneous channel gains. By restricting quantization codebooks to the low-dimensional subspace spanned by these dominant propagation paths, the authors ensured that feedback bits are only spent describing relevant signal dimensions, drastically reducing feedback overhead.
On the hardware side, tracking these subspaces does not require expensive, full-scale matrix decompositions. L. Ehrenberg et al. demonstrated that projection approximation subspace tracking (PASTd) algorithms can approximate principal eigenvectors directly from data statistics. This shifts the computational complexity of channel tracking from a cubic function of the antenna count ($O(N^3)$) to a quadratic function ($O(N^2)$), allowing mobile devices with strict battery and processing constraints to perform real-time spatial filtering.
Ultimately, subspace-constrained beamforming works because it respects the underlying physics of wave propagation. By discarding the vast "orthogonal complement" of empty space where no signals travel, it strips away redundant degrees of freedom—proving that the complexity of controlling next-generation wireless interfaces depends not on how many antennas we build, but on how many targets we serve.