CDWMMSE: Covariance Decomposition for Satellite MIMO

• CDWMMSE is a precoding strategy that exploits covariance decomposition to convert sum-rate maximization into a generalized weighted MMSE problem.
• It leverages statistical CSI in multi-satellite beamspace MIMO systems to jointly optimize satellite clustering, beam selection, and synchronization under practical constraints.
• The framework achieves near-optimal performance with reduced computational overhead, demonstrating robustness against phase errors and limited instantaneous CSI.

Covariance decomposition-based weighted minimum mean square error (CDWMMSE) is a precoding strategy for multi-satellite multi-stream (MSMS) beamspace massive MIMO systems, wherein statistical channel state information (sCSI) is leveraged to jointly optimize satellite clustering, beam selection, and transmit precoding under practical constraints of satellite cooperation and synchronization. CDWMMSE precoding casts the sum-rate maximization under a tractable upper-bound approximation as a generalized weighted MMSE problem, solvable via covariance decomposition and iterative minimization. This framework is especially suited to scenarios with limited instantaneous CSI—such as LEO and GEO satellite constellations—wherein statistical large-scale geometrical and propagation information are available, but instantaneous multi-node synchronization is not feasible (Wang et al., 26 Dec 2025).

1. Theoretical Basis and Signal Model

In MSMS beamspace transmission, multiple satellites—each with a phased array—jointly serve multi-antenna user terminals (UTs) by exploiting the sparsity of the line-of-sight (LoS) dominant MIMO channel in the angular (beam) domain. The frequency-domain channel including synchronization errors is modeled as:

$$\mathbf{H}_{s,k} = \sqrt{\frac{\kappa_{s,k}\gamma_{s,k}}{\kappa_{s,k}+1}}\,\mathbf{u}_{s,k}\mathbf{v}^T_{s,k} + \sqrt{\frac{\gamma_{s,k}}{\kappa_{s,k}+1}}\,\tilde{\mathbf{u}}_{s,k}\mathbf{v}^T_{s,k}$$

where $\gamma_{s,k}$ is the large-scale path gain, $\kappa_{s,k}$ the Rician factor, $\mathbf{u}_{s,k}$ and $\mathbf{v}_{s,k}$ the receive/transmit steering vectors, and $\tilde{\mathbf{u}}_{s,k}$ a NLoS component with covariance $\boldsymbol\Sigma_{s,k}$. Phase drifts and synchronization errors are incorporated as random phase factors in the received signal model.

Upon mapping to the beamspace using satellite codebook $\mathbf{F}_s$ and beam selection $\mathbf{A}_s$, the aggregate MSMS channel is constructed. This model naturally supports distributed cooperation, asynchronous transmission, and the exploitation of LoS sparsity.

2. sCSI-Based Sum-Rate Upper Bound

Given the absence of instantaneous CSI, the achievable ergodic sum-rate is approximated by a statistical upper-bound:

$$\bar R_k = \log_2\det\left(\mathbf{I} + \bar{\mathbf{R}}^{-1}_{\rm other,k}\,\bar{\mathbf{R}}_{\rm sig,k}\right)$$

with $\bar{\mathbf{R}}_{\rm sig,k}$ and $\bar{\mathbf{R}}_{\rm other,k}$ denoting, respectively, the mean signal and interference covariance matrices, computable in closed-form under the LoS-dominated and phase-randomized assumptions:

$$\bar{\mathbf{R}}_{\rm sig,k} = \sum_{s_1,s_2\in\mathcal S_k} (\mathbf{q}_{s_1,k,k}^H\mathbf{q}_{s_2,k,k})\, \boldsymbol\Delta_{s_1,s_2,k}$$

$$\bar{\mathbf{R}}_{\rm other,k} = \sum_{j\neq k}\sum_{s\in\mathcal S_j} (\mathbf{q}_{s,j,k}^H\mathbf{q}_{s,j,k})\,\boldsymbol\Delta_{s,s,k} + \sigma_k^2\,\mathbf{I}$$

where $\boldsymbol\Delta_{s_1,s_2,k}$ is parameterized by statistical LoS/NLoS and phase error statistics (Wang et al., 26 Dec 2025).

3. CDWMMSE Problem Formulation

The core of the CDWMMSE approach is to equivalently recast the sum-rate maximization given fixed satellite clustering and beam selection as a weighted MMSE problem, whose variables are the per-user precoders $\{W_{s,k}\}$, equalizers $\{D_k\}$, and MSE weight matrices $\{C_k\}$. The optimization reads:

$$\min_{\{W_{s,k}\},\{C_k,D_k\}} \sum_k \beta_k \left[ \mathrm{Tr}(C_k E_k) - \log\det(C_k) \right]$$

$$\text{s.t.} \quad \sum_k \mathrm{Tr}(W_{s,k}W_{s,k}^H) \le P_s$$

with user weight $\beta_k$ and mean-squared error matrix (for user $k$):

$$E_k = D_k^H(\bar R_{\rm sig,k}+\bar R_{\rm other,k}) D_k - D_k^H \bar R_{\rm sig,k}^{1/2} - \bar R_{\rm sig,k}^{1/2,H} D_k + I$$

where $\bar R_{\rm sig,k}^{1/2}$ is any matrix square root of $\bar R_{\rm sig,k}$. The covariance decomposition of $\bar R_{\rm sig,k}$ into a square-root enables tractable optimization (Wang et al., 26 Dec 2025).

4. Covariance Decomposition and Iterative Algorithm

Closed-form decomposition is achieved using Lemma 1 (in (Wang et al., 26 Dec 2025)), allowing the computation of a square-root:

$$R_{\rm sig,k}^{1/2} = \left[ \sum_{s=1}^S \bar\varphi_{s,k} \sqrt{\rho_{s,k}}\,\mathbf{u}_{s,k} \mathbf{q}_{s,k,k}^H,\quad \tilde{\boldsymbol\Sigma}_k \mathbf{Q}_k^H \right]$$

The alternating minimization algorithm (MS$^2$CDWM) proceeds iteratively:

• Update the MMSE receiver $$D_k = (\bar R_{\rm sig,k} + \bar R_{\rm other,k})^{-1}\bar R_{\rm sig,k}^{1/2}$$
• Update the MSE weight $C_k = E_k^{-1}$
• Update the precoder $W_k$ in closed form by solving the Lagrangian, using auxiliary variables gathered from $C_k$, $D_k$, and the decomposed covariances.

Convergence is typically rapid; per-iteration complexity is $O(KS^3N_R^3M^3 + K^2 \tilde B^2 S + K \tilde B^3)$ and does not depend on the full array dimension $N_T$ (Wang et al., 26 Dec 2025).

5. Heuristic Closed-Form and Low-Overhead Approximations

To further reduce complexity, the MS$^2$CDM heuristic fixes $D_k$ and $C_k$ as identity matrices, allowing a non-iterative, fully closed-form update for $W_k$:

$$\tilde W_k = (\tilde \Xi_k + \breve \beta_k I)^{-1} \breve V_k^H \breve T_k\,, \qquad W_k = \tilde \eta_k \tilde W_k$$

This approximation achieves $>$85% of the full CDWMMSE performance at one tenth the computational cost. For single-stream per user, a satellite-side “location-information-based” (LIB) precoder uses only GNSS-provided UT angles with complexity $O(K \sum_s Q_s N_T)$ (Wang et al., 26 Dec 2025).

6. Applications, Performance, and Implications

CDWMMSE enables tractable optimization in large-scale distributed MIMO systems where centralized instantaneous CSI is infeasible. Simulations under 3GPP NTN channel models with practical system parameters (satellites, $16 \times 16$ UPA, $1$--$4$ UT receive antennas, codebook size $Q_s = 256$) demonstrate that:

• MS$^2$CDWM precoding closes over 90% of the gap to full-dimensional MIMO at 2–3 dB lower power, with strictly reduced dimension and complexity.
• The sum rate increases linearly with the number of served satellites $S_k$ until saturated by $N_R$.
• Moderate beam selection ($B_s=48$) achieves $>$95% of the performance of exhaustive selection.
• Pragmatic LIB schemes and heuristics attain near-optimal multiplexing for $M_k=1$ at negligible signaling overhead.
• The framework is robust to phase/synchronization errors and leverages only large-scale geometry, supporting scalability and deployment in LEO swarms and GEO constellations (Wang et al., 26 Dec 2025).

A plausible implication is that covariance decomposition–based MMSE precoding generalizes conventional codebook-based approaches, serving as a basis for scalable, high-spectral-efficiency satellite MIMO with realistically available sCSI.

7. Integration in MSMS Beamspace Frameworks and Broader Impact

CDWMMSE serves as the cornerstone for practical beam-domain linear precoding in MSMS systems, directly transitioning concepts from terrestrial massive MIMO to the satellite domain, where distributed transmitters and channel synchronization are fundamentally more challenging. It complements satellite clustering algorithms (user-centric competition) and two-stage beam selection mechanisms optimized for LoS power and multi-user orthogonality.

By unifying covariance-based statistical optimization with the beamspace approach, CDWMMSE supports the rapid, robust, and low-overhead realization of distributed MIMO for next-generation non-terrestrial networks (NTN), approaching the performance of fully coordinated MIMO—yet requiring only knowledge of position, large-scale path parameters, and statistical channel models (Wang et al., 26 Dec 2025).

For further details, see "Multi-Satellite Multi-Stream Beamspace Massive MIMO Transmission" (Wang et al., 26 Dec 2025).