Centralized MMSE Beamforming

• Centralized MMSE beamforming is a signal processing technique that uses global CSI to minimize mean-squared error and optimize multiuser communication performance.
• It formulates and solves convex or structured nonconvex quadratic optimization problems to jointly mitigate interference and noise, thereby elevating sum-rate and QoS metrics.
• Widely applied in multiuser MIMO, massive MIMO, satellite, and CAPA systems, it balances optimal performance with considerations of computational complexity and hardware constraints.

Centralized minimum mean-squared error (MMSE) beamforming is a core methodology in multiuser MIMO, massive MIMO, satellite, and continuous-aperture array (CAPA) systems for jointly suppressing interference and noise while optimizing quality-of-service (QoS) metrics such as sum-rate and (sum) MSE. Centralized MMSE designs are characterized by the use of all available global channel state information (CSI) to solve a convex or structured nonconvex quadratic optimization, yielding globally optimal or stationary-point beamformers for a broad class of architectures including spatially-discrete arrays, continuous-aperture systems, multicast and unicast networks, and hybrid analog-digital transceivers.

1. System Models and Problem Formalization

Centralized MMSE beamforming can be formulated in discrete, continuous, or hybrid array architectures. In canonical MU-MIMO uplink and downlink settings, users transmit $K$ independent data streams to a centralized base station (BS) equipped with $N$ antennas or a continuous-aperture region. The vector channel model:

$y = Hx + n$

where $H\in\mathbb{C}^{N\times K}$ (or continuous kernel $h_k(r)$ for CAPA), $x\in\mathbb{C}^{K}$ is the transmit symbol vector, and $n\sim \mathcal{CN}(0,\sigma^2I_N)$ is spatial white Gaussian noise, underlies the MMSE optimization (Ouyang et al., 2024, Xing et al., 2012).

In continuous-aperture uplink (CAPA), the received field $y(r)$ at $r\in A$ is

$y(r)=\sum_{k=1}^K \sqrt{P_k} h_k(r)s_k + n(r)$

with MMSE receive filters $N$0 acting on $N$1 to recover $N$2 (Ouyang et al., 2024).

For massive MIMO, the system may involve broadcast/multicast (e.g., one BS, many users per group), joint transmission, or hybrid analog/digital front-ends (Yin et al., 2024, Lin et al., 2019). Multibeam satellite and holographic arrays follow similar principles, with adaptations for array geometry and coupling (Li et al., 7 Jan 2025).

The MMSE problem can be succinctly written as: for a given system model, design transmit/receive beamformers $N$3 (and/or power allocations) to minimize

$N$4

possibly under individual or sum power constraints, with the MSE operator depending on the transmit and receive strategies (Xing et al., 2012).

2. MMSE Beamforming: Optimization and Closed-Form Structure

The centralized MMSE beamformer is obtained by minimizing the mean-squared error between estimated and true data symbols, typically under transmit power constraints. In standard discrete MIMO, for fixed channel $N$5 and transmit precoder $N$6:

$N$7

$N$8

where $N$9 is a dual variable chosen to enforce a power constraint (Xing et al., 2012, Lin et al., 2019).

In the Uplink MMSE receive context:

$y = Hx + n$0

For continuous-aperture (CAPA) beamforming, the optimal MMSE receive filter is a function in the $y = Hx + n$1-dimensional subspace spanned by all user channel responses:

$y = Hx + n$2

where $y = Hx + n$3 is the channel correlation matrix $y = Hx + n$4 (Ouyang et al., 2024). In downlink CAPA, the continuous MMSE beamformer for user $y = Hx + n$5 is constructed as a linear combination of all channel responses, solving for the optimal weights using the continuous covariance matrix (Wang et al., 2024).

Closed forms also arise in multicast/coordination contexts: the MMSE update for each BS exploits the global channel Gram and weighted interference-plus-noise structure (Yin et al., 2024). Hybrid analog/digital systems alternate MMSE digital designs with manifold or beamspace projection constraints on the analog front-end (Lin et al., 2019).

3. Performance Analysis and Optimality Properties

Centralized MMSE beamforming provably yields optimal tradeoffs between signal, interference, and noise. Key analytic results:

• MMSE is rate- and MSE-optimal: In CAPA and discrete MIMO, the MMSE beamformer minimizes per-user MSE and maximizes sum-rate, exceeding or matching the performance of maximum ratio combining (MRC) and zero-forcing (ZF) (Ouyang et al., 2024, Wang et al., 2024).
• SINR and MSE characterizations: The closed-form achieved SINR for user $y = Hx + n$6 takes the structure:

$y = Hx + n$7

with MSE $y = Hx + n$8 (Ouyang et al., 2024).

• Signal subspace property: All proposed beamformers (MMSE, MRC, ZF) operate within the subspace spanned by the user spatial channel responses, limiting the effective DOF to $y = Hx + n$9 regardless of aperture granularity (Ouyang et al., 2024).
• Asymptotic regimes: At low SNR, MMSE converges to MRC; at high SNR, MMSE approaches ZF, reflecting the interpolation between noise-dominated and interference-dominated regimes (Ouyang et al., 2024, Wang et al., 2024).

4. Algorithmic Realizations and Computational Aspects

Centralized MMSE realization depends on array structure, system size, and hardware constraints:

• Classical matrix inversion: For $H\in\mathbb{C}^{N\times K}$0 antennas and $H\in\mathbb{C}^{N\times K}$1 users, the core operations involve $H\in\mathbb{C}^{N\times K}$2 (forming $H\in\mathbb{C}^{N\times K}$3) and $H\in\mathbb{C}^{N\times K}$4 (inversion) complexity (Feng et al., 2024, Lin et al., 2019).
• Dimension reduction: Beamspace methods such as convolutional beamspace (CBS) pre-filtering project received signals into a lower-dimensional subspace, reducing MMSE computation from $H\in\mathbb{C}^{N\times K}$5 scaling to linear or quadratic in $H\in\mathbb{C}^{N\times K}$6, especially effective when $H\in\mathbb{C}^{N\times K}$7 or in bandwidth-rich scenarios (Feng et al., 2024, Guvensen et al., 2016).
• Iterative algorithms: Alternating minimization between transmit and receive MMSE designs, with KKT updates and power constraint enforcement, yields fast convergence for large-scale systems (Xing et al., 2012, Yin et al., 2024).
• Hybrid architectures: In mmWave HBF, MMSE digital designs are alternated with analog beamformer updates via manifold optimization, generalized eigenvector methods, or OMP-based sparse support (Lin et al., 2019).
• Continuous/discretized CAPA: CAPA MMSE implementations employ numerical quadrature to approximate the effect of continuum summations, with complexity $H\in\mathbb{C}^{N\times K}$8 for $H\in\mathbb{C}^{N\times K}$9 quadrature points (Wang et al., 2024).

5. Extensions: Channel Estimation, Coordination, and Robustness

MMSE beamforming is central not only for data transmission/combining, but also for channel estimation and multi-cell coordination:

• Separation principle: For Rayleigh fading massive MIMO, MMSE channel estimation followed by MMSE beamforming achieves the full information-theoretic lower-bounds—joint non-linear mapping offers no gain over separation (Miretti et al., 11 Jul 2025).
• Coordinated multicell/multicast: Centralized weighted MMSE (WMMSE) algorithms solve power-constrained SINR-constrained optimization for joint multicell transmission. Each transmitter solves for its beamformer by inverting a regularized interference-plus-noise covariance reflecting all users’ global channels and MSE weights (Yin et al., 2024).
• CSI reduction, robustness: Reduced-dimension MMSE channel estimation and statistical pre-beamforming harness long-term spatial covariance to suppress pilot contamination and minimize required pilot overhead without large-dimensional matrix inversions (Guvensen et al., 2016). In satellite/LEO/HMA systems, low-complexity MMSE can be achieved by replacing sample covariances with their statistical means based on stochastic geometry, circumventing full CSI acquisition for all interferers (Li et al., 7 Jan 2025).

6. Comparative Analysis: MMSE versus MRC and ZF

In all analyzed settings—discrete, continuous, satellite, cellular—centralized MMSE consistently outperforms MRC and ZF:

Scheme	Interference Suppression	Noise Sensitivity	Asymptotics
MRC	None	Minimum	Optimal at low SNR
ZF	Full	Strongest	Optimal at high SNR
MMSE	Optimal tradeoff	Regularizes ZF/MRC	Matches best in both regimes

MMSE always achieves $h_k(r)$0 (Ouyang et al., 2024, Wang et al., 2024). In practical/empirical studies, CAPA systems using MMSE/regularized ZF beamforming outperform their spatially-discrete counterparts in both sum-rate and sum-MSE (Ouyang et al., 2024, Wang et al., 2024). Hybrid MMSE typically incurs only minor degradation relative to full digital MMSE, e.g., $h_k(r)$1–$h_k(r)$2 dB loss in MSE or $h_k(r)$3–$h_k(r)$4 in rate for moderate $h_k(r)$5 (Lin et al., 2019).

7. Practical Impact and Implementation Considerations

Centralized MMSE beamforming, utilizing global CSI and closed-form or efficiently iterative algorithms, constitutes the practical foundation for state-of-the-art uplink/downlink and joint transmission in massive MIMO, CAPA, LEO satellite, and coordinated cellular systems.

• Complexity: MMSE dimension reduction (CBS, RR-MMSE, GEB) decreases the computational burden from cubic in user number to quadratic or linear, enabling scalability as system size grows (Feng et al., 2024, Guvensen et al., 2016).
• Robustness: Stochastic-statistical MMSE handles partial CSI, non-ideal conditioners, and pilot contamination gracefully (Li et al., 7 Jan 2025, Guvensen et al., 2016).
• Flexibility: MMSE/WMMSE unifies unicast, multicast, hybrid, and coordinated multi-cell frameworks under regularized quadratic programming, with consistent signal subspace structure and optimality properties (Yin et al., 2024).
• CAPA advancements: Continuous beamforming architectures generalize and outperform traditional SPDAs given the same aperture, with MMSE design ensuring maximal sum-rate/sum-MSE efficiency (Ouyang et al., 2024, Wang et al., 2024).

These features anchor centralized MMSE beamforming as a canonical solution across an array of advanced wireless, satellite, and array processing domains.