Covariance-Guided Beam Selection

• Covariance-guided beam selection is a method that uses statistical covariance matrices to construct transmit or receive beams in multi-antenna systems.
• It employs techniques like convex programming, manifold optimization, and greedy search to optimize performance metrics such as sum-rate and DoA estimation precision.
• The approach has been validated in various applications including FDD massive MIMO, hybrid mmWave systems, and MIMO radar, offering scalable and robust solutions.

A covariance-guided beam selection strategy is a class of methods in array signal processing and multi-antenna communication or radar systems in which statistical information—specifically signal or channel covariance matrices—is directly exploited to select or design transmit or receive beams. These methodologies optimize key system objectives (such as ergodic sum-rate, beampattern fidelity, inter-beam orthogonality, or direction-of-arrival (DoA) estimation MSE) based on long-term channel or signal subspace statistics, rather than instantaneous observations. The resulting strategies combine convex programming, manifold optimization, and greedy or structured search algorithms, enabling robust and low-complexity solutions that are highly effective in scenarios with partial channel state information, finite RF resources, or demanding real-time constraints.

1. Fundamental Principles and System Models

Covariance-guided beam selection refers to beamforming strategies that utilize the signal, noise, or channel covariance matrices as their principal source of information for beam construction or selection. Distinct from instantaneous channel state-based beamforming, these approaches are particularly effective in settings where feedback or estimation overhead severely limits the availability of instantaneous channel state information (e.g., FDD massive MIMO, hybrid analog-digital mmWave systems, or dynamic MIMO radar platforms).

Architectures and problem contexts include:

• Multi-user broadcast channels, where only the second-order channel statistics (covariances) are known at the transmitter—each user has perfect local CSI, but transmitter optimization must rely on $$\Sigma_i = \mathbb{E}[ \mathbf{h}_i \mathbf{h}_i^H ]$$ (Raghavan et al., 2011).
• MIMO radar transmit beampattern design, in which a transmit covariance $R$ and possibly an antenna selection vector $p$ must be found to approximate a desired beam-pattern and minimize inter-target correlation under hardware constraints (Bose et al., 2020, Bose et al., 2019).
• Large-scale mmWave beamspace processing, where receiver-side hybrid architectures use DFT-based analog combiners and select a sparse subset of beams based on a fitted signal covariance to maximize DoA estimation performance (Şenyuva, 30 Nov 2025).
• MIMO interference channels, where transmit covariance matrices must reside in subspaces determined by aggregate channel statistics to satisfy Pareto-optimality of rate objectives (Park et al., 2012).

The overarching principle is that, for a wide class of quadratic objectives (e.g., mean signal/noise ratio, rate, beampattern error), a properly chosen or optimized covariance structure or its dominant eigenmodes encode nearly all the actionable information for robust beam selection.

2. Convex and Nonconvex Covariance-Based Optimization Formulations

Most covariance-guided strategies formulate beam selection as either a convex or a nonconvex optimization over the feasible set of covariance matrices and, in some cases, binary selection vectors.

For example, in MIMO radar beamforming, the joint transmit covariance and antenna selection problem is cast as: $$\min_{p,R,\alpha} J(p, R, \alpha) \quad \text{s.t. } R \succeq 0,\, \operatorname{diag}(R)=\tfrac{c}{M} \mathbf{1}_M,\, p \in \{0,1\}^M,\, \|p\|_1=N,\, \alpha>0$$ where $J$ penalizes both beampattern mismatch and target cross-correlation (Bose et al., 2020, Bose et al., 2019): $$J(p, R, \alpha) = \frac{1}{K} \sum_{k=1}^K w_k |p^T [R \odot (a(\theta_k)a(\theta_k)^H)^*] p - \alpha \phi(\theta_k)|^2 + \frac{2\omega_c}{\tilde K(\tilde K - 1)} \sum_{p<q}|p^T \Re\{R \odot (a(\tilde{\theta}_p)a(\tilde{\theta}_q)^H)^*\} p|^2$$ This mixed Boolean and semidefinite program is solved by splitting into convex subproblems (SDP for $R,\alpha$ with fixed $p$; combinatorial optimization or local search for $p$ with fixed $R$).

In hybrid mmWave receivers, Toeplitz-constrained denoised signal covariance matrices $R$0 are computed by solving a quadratic program under positive semidefinite and Toeplitz constraints, and contiguous DFT beam subsets are scored via a covariance-capture functional (Şenyuva, 30 Nov 2025).

Covariance-guided multiuser MIMO beamforming problems are formulated as fractional programs or manifold-constrained optimizations over covariance matrix factorizations, leveraging generalized eigenvalue problems or product Stiefel manifolds (Raghavan et al., 2011, Park et al., 2012).

3. Algorithmic Realizations: Eigenstructure, Manifold Optimization, and Greedy Search

Algorithmic solutions for covariance-guided beam selection exploit the intrinsic structure of the optimization landscape:

• Generalized Eigenvector Beamformers: For the two-user broadcast channel, the optimal beamforming vectors $R$1 maximize a statistically averaged SINR metric, solvable as a dominant generalized eigenvector (GEV) problem:

$R$2

The principal GEV forms the optimal beam (Raghavan et al., 2011).

• Cyclic Alternating Minimization: In radar settings, joint transmit covariance matrix and sparsity pattern optimization are performed iteratively—first solving a convex SDP or QP for $R$3 given $R$4, then updating $R$5 via greedy (Hamming-distance-1) local search, potentially with random perturbations to escape shallow minima. Each iteration guarantees non-increasing objective value and practical convergence in few cycles (Bose et al., 2020, Bose et al., 2019).
• Product Manifold Parameterizations: For multiuser MIMO interference channels, Pareto-optimal transmit covariances are shown to reside in the union of channel eigenspaces. The parameterization

$R$6

allows optimization via Riemannian or projected first/second-order optimization on a low-dimensional product manifold (Park et al., 2012).

• Covariance-Fitting and Beam Scoring: In hybrid DFT-beamspace architectures, the signal-plus-noise covariance fitted on a virtual subarray is projected to produce a full-aperture, Toeplitz-PSD covariance. Sector-based beam selection is achieved by maximizing a covariance-capture score over contiguous DFT beam blocks; selection strategies favor blocks that maximize signal covariance projection while penalizing ill-conditioning (Şenyuva, 30 Nov 2025).

4. Analytical Performance and Empirical Results

Covariance-guided beam selection strategies demonstrate performance close to systems with full instantaneous CSI, especially at moderate SNR or when the statistical envelope dominates the channel structure. For two-user broadcast channels, covariance-only schemes nearly match the sum-rate of optimal instantaneous CSI-based beamforming and substantially outperform fixed codebooks (Raghavan et al., 2011). In MIMO radar beampattern synthesis, such strategies yield mainlobe fidelity, sidelobe suppression, and inter-target decorrelation at substantially reduced computational cost compared to prior approaches such as ADMM-based methods (e.g., 4500 s vs. 17 s in a specified scenario) (Bose et al., 2020).

In hybrid mmWave MIMO, the covariance-guided DFT beam selection method closes the gap to the Cramér–Rao bound for DoA estimation—achieving root-MSE within 1–2 dB of the CRB at moderate SNR, with strictly lower failure rates and better accuracy-runtime trade-offs than sectorization (Şenyuva, 30 Nov 2025).

Multi-user MIMO manifold-based covariance parameterizations strictly enlarge the efficient rate region versus classical eigen-beamforming and attain robust performance gains (10–25% in intermediate SNR, up to 3 dB in sum-rate), adapting degrees of freedom allocation to instantaneous SNR conditions (Park et al., 2012).

5. Complexity, Scalability, and Applicability

Closed-form or low-complexity structure underpins the appeal of covariance-guided methods. Dominant eigenvector and GEV problems can be solved in $R$7 per user (with $R$8 antennas); cyclic or greedy search over selection vectors converges in $R$9 practical time for $p$0; and manifold optimization leverages both dimensionality reduction (from $p$1 to $p$2) and first/second order geometry-aware updates (Raghavan et al., 2011, Park et al., 2012, Bose et al., 2020).

Covariance-based approaches are particularly suitable for:

• FDD or hybrid systems with limited feedback
• Scenarios where channel statistics change more slowly than instantaneous values
• Sensor and antenna selection in array processing and compressive sensing.

Extensions cover detection and localization in mmWave systems, multi-sector beamforming, robust distributed or online implementations (where low-rank projections or stochastic gradients enable adaptation during system motion or environmental change) (Bose et al., 2020, Şenyuva, 30 Nov 2025, Park et al., 2012).

6. Connections to Broader Research Themes

Covariance-guided beam selection unifies several core themes across array signal processing and wireless communications:

• Statistical signal processing and robust optimization via covariance fitting and subspace methods
• Exploitation of low-rank structures and manifold geometry to reduce complexity and ensure efficient implementation in high-dimensional settings
• Practical algorithm design integrating convex relaxations, greedy/discrete optimization, and separable updates suitable for parallelization and real-time constraints.

These methodologies overlap with compressed sensing via structured sparsity, MIMO codebook design under statistical reciprocity, and advanced multi-user interference management. Additionally, covariance-guided approaches can be adapted to minimize cross-correlation in sensing (radar), maximize system throughput (wireless communication), and enable high-fidelity parametric estimation (DoA, channel state) in compact, resource-limited hardware.

7. Impact and Future Directions

Covariance-guided beam selection strategies provide a rigorous, tractable, and broadly applicable framework essential in modern large-scale array and hybrid system design. By shifting emphasis from instantaneous to statistical CSIT/CSIR, these methods support scalable robust beamforming, near-optimal sensing, and efficient resource allocation in environments where acquiring full high-dimensional state information is impractical.

Anticipated research directions include:

• Online and distributed covariance adaptation under non-stationary or adversarial channel dynamics
• Integrated hardware-in-the-loop optimization for sparse, non-uniform, or reconfigurable array geometries
• Joint design across multiple array layers (e.g., subarray structures, mutual coupling)
• Cross-layer applications spanning communications, sensing, and networked control via covariance-aware joint transceiver architectures.

Covariance-guided frameworks thus stand as a foundational component of statistically robust and algorithmically scalable beamforming in current and emerging multi-antenna wireless and sensing systems (Raghavan et al., 2011, Bose et al., 2020, Bose et al., 2019, Park et al., 2012, Şenyuva, 30 Nov 2025).