Ferkans — Interactive Telecom Tutor

When Matrix Inversion Is Too Expensive

The MMSE receiver requires inverting the $K \times K$ matrix $\mathbf{H}^{H} \mathbf{H} + \frac{\sigma^2}{P}\mathbf{I}$ at a cost of $\mathcal{O}(K^{3})$ . For massive MIMO with $K = 16$ , this is fast. But in emerging scenarios — cell-free massive MIMO with hundreds of users, or millimeter-wave systems with hybrid beamforming — the matrix dimension can grow to the point where exact inversion becomes a bottleneck.

This section presents three strategies to approximate the MMSE filter without explicit matrix inversion: the Neumann series, local MMSE (for distributed systems), and iterative (conjugate gradient) methods.

Definition:
Neumann Series Approximation

For a matrix $\mathbf{A}$ with spectral radius less than 1, the Neumann series gives

$\mathbf{A}^{-1} = \sum_{n=0}^{\infty} (\mathbf{I} - \mathbf{A})^n.$

Applied to the regularized Gram matrix $\mathbf{A} = \mathbf{H}^{H} \mathbf{H} + \frac{\sigma^2}{P}\mathbf{I}$ , we decompose $\mathbf{A} = \mathbf{D} + \mathbf{E}$ where $\mathbf{D}$ is the diagonal part and $\mathbf{E}$ is the off-diagonal. Then

$\mathbf{A}^{-1} = \mathbf{D}^{-1} \sum_{n=0}^{\infty} (-\mathbf{D}^{-1}\mathbf{E})^n \approx \mathbf{D}^{-1} \sum_{n=0}^{L-1} (-\mathbf{D}^{-1}\mathbf{E})^n,$

where $L$ is the truncation order. The $L$ -th order approximation has complexity $\mathcal{O}(L K^{2})$ — avoiding the cubic cost of exact inversion.

The convergence rate depends on the spectral radius of $\mathbf{D}^{-1}\mathbf{E}$ , which is small when the Gram matrix is diagonally dominant. In the massive MIMO regime with favorable propagation, diagonal dominance holds and even $L = 2$ or $L = 3$ gives near-optimal performance.

Theorem: Neumann Series Convergence for Massive MIMO

Let $\mathbf{A} = \mathbf{H}^{H} \mathbf{H} + \alpha \mathbf{I}$ with $\alpha = \sigma^2/P > 0$ and i.i.d. Rayleigh fading. As $N_t/K \to \infty$ , the spectral radius of $\mathbf{D}^{-1}\mathbf{E}$ satisfies

$\rho(\mathbf{D}^{-1}\mathbf{E}) \to 0,$

and the $L$ -th order Neumann approximation satisfies

$\|\mathbf{A}^{-1} - \hat{\mathbf{A}}^{-1}_{L}\| = \mathcal{O}\left(\left(\frac{K}{N_t}\right)^L\right).$

Favorable propagation makes the Gram matrix approximately diagonal, so the off-diagonal perturbation $\mathbf{E}$ is small relative to $\mathbf{D}$ . More antennas relative to users means faster convergence.

Proof

Bound the off-diagonal elements

By favorable propagation, $\frac{1}{N_t}|\mathbf{h}_k^H \mathbf{h}_j| = \mathcal{O}(1/\sqrt{N_t})$ for $k \neq j$ , while the diagonal elements $\frac{1}{N_t}\|\mathbf{h}_k\|^2 \to \beta_k$ .

Bound the spectral radius

The Gershgorin circle theorem gives $\rho(\mathbf{D}^{-1}\mathbf{E}) \leq \max_k \sum_{j \neq k} |[\mathbf{D}^{-1}\mathbf{E}]_{kj}| = \mathcal{O}(K/\sqrt{N_t})$ , which vanishes as $N_t \to \infty$ with $K$ fixed.

Error bound

The truncation error is $\|\sum_{n=L}^{\infty}(-\mathbf{D}^{-1}\mathbf{E})^n\| \leq \rho^L / (1 - \rho)$ , giving the stated convergence rate. $\blacksquare$

Example: Second-Order Neumann Approximation

For $N_t = 128$ , $K = 8$ , $\text{SNR} = 10$ dB with i.i.d. Rayleigh fading, compute the average SINR loss of the 2nd-order Neumann approximation relative to exact MMSE.

Solution

Compute the diagonal and off-diagonal parts

The diagonal entries are $[\mathbf{D}]_{kk} = \|\mathbf{h}_k\|^2 + \sigma^2/P \approx N_t \beta_k + 1/\text{SNR}$ . For $\beta_k = 1$ : $[\mathbf{D}]_{kk} \approx 128 + 0.1 = 128.1$ .

Estimate the spectral radius

The off-diagonal entries have magnitude $\mathcal{O}(\sqrt{N_t}) = \mathcal{O}(11.3)$ , so $|[\mathbf{D}^{-1}\mathbf{E}]_{kj}| \approx 11.3/128.1 \approx 0.088$ . With $K - 1 = 7$ off-diagonal terms per row, $\rho \lesssim 7 \times 0.088 = 0.62$ .

Assess SINR loss

The 2nd-order error is $\mathcal{O}(\rho^2) \approx 0.38$ , suggesting the Neumann approximation introduces roughly $10\log_{10}(1/(1-0.38^2)) \approx 0.7$ dB SINR loss. Monte Carlo simulations confirm a loss of approximately 0.3–0.5 dB for this configuration, making $L = 2$ sufficient.

Neumann Series Convergence vs. Truncation Order

Explore how the Neumann series approximation converges to the exact MMSE SINR as the truncation order $L$ increases. Observe that convergence is faster for larger $N_t/K$ ratios.

Parameters

N_t

(BS antennas)64

K

(users)8

\text{SNR}

(dB)10

Max Neumann order

L

5

Local MMSE for Distributed Antenna Systems

In cell-free massive MIMO (MIMO Ch. 11–16), the BS antennas are distributed across many access points (APs), each with a local processor and limited fronthaul capacity. The local MMSE approach partitions the antenna array into groups and applies MMSE within each group independently:

$\hat{x}_k^{(g)} = (\mathbf{g}_k^{(g)})^H \mathbf{y}^{(g)},$

where $\mathbf{y}^{(g)}$ and $\mathbf{g}_k^{(g)}$ are the received signal and combining vector at AP group $g$ . The final estimate is a weighted combination:

$\hat{x}_k = \sum_{g=1}^{G} w_{k,g} \hat{x}_k^{(g)},$

where $w_{k,g}$ are large-scale fading decoding (LSFD) weights.

Local MMSE avoids centralizing all received signals and requires only local CSI at each AP. The performance gap to centralized MMSE depends on the spatial diversity of the distributed array.

Conjugate Gradient Method

The MMSE detection problem $(\mathbf{H}^{H} \mathbf{H} + \frac{\sigma^2}{P}\mathbf{I})\hat{\mathbf{x}} = \mathbf{H}^{H} \mathbf{y}$ is a positive definite linear system that can be solved iteratively by the conjugate gradient (CG) method. CG converges in at most $K$ iterations (exact solution), but in practice 5–10 iterations suffice because the condition number of the regularized Gram matrix is small in massive MIMO.

Each CG iteration costs $\mathcal{O}(K^{2})$ (one matrix-vector multiply with the Gram matrix), giving a total cost of $\mathcal{O}(L K^{2})$ for $L$ iterations — the same scaling as the Neumann series.

Common Mistake: Neumann Series Diverges When the Antenna-to-User Ratio Is Small

Mistake:

Applying the Neumann series with $L = 2$ or $L = 3$ when $N_t/K$ is close to 1 (e.g., $N_t = 16$ , $K = 12$ ).

Correction:

The spectral radius $\rho(\mathbf{D}^{-1}\mathbf{E})$ can exceed 1 when $N_t/K$ is small, causing the Neumann series to diverge. In this regime, use exact matrix inversion or the conjugate gradient method instead. As a rule of thumb, the Neumann series is reliable for $N_t/K \geq 4$ .

Neumann Series

An iterative expansion $\mathbf{A}^{-1} = \sum_{n=0}^{\infty}(\mathbf{I} - \mathbf{A})^n$ that avoids explicit matrix inversion. Converges when the spectral radius of $(\mathbf{I} - \mathbf{A})$ is less than 1.

Local MMSE

An MMSE detection strategy for distributed antenna systems where each access point applies MMSE locally using only its own received signal and CSI. Results are combined centrally using LSFD weights.

Low-Complexity Approximations