Chapter Summary

Chapter Summary

Key Points

• 1.

  XL-MIMO arrays break the spatial stationarity assumption that every earlier chapter relied on. Each user illuminates only a subset of the array — its visibility region $\mathcal{V}_k$ — and the per-user spatial covariance varies across antennas because of aperture-range geometry, blockage, and multipath clustering. The binary mask $\mathbf{m}_k \in \{0,1\}^{N_t}$ is the central new random object introduced in this chapter.

• 2.

  Failing to detect the VR is expensive on both sides: an oversized $\hat{\mathcal{V}}_k$ pumps noise on dead antennas ($|\hat{\mathcal{V}}_k|/\text{SNR}$), while an undersized one throws away signal energy ($|\mathcal{V}_k \setminus \hat{\mathcal{V}}_k|/|\mathcal{V}_k|$). The optimal detector is the true VR, and every VR-aware estimator is tuned to navigate this two-sided trade-off.

• 3.

  The CommIT contribution of Xu and Caire models $\mathbf{m}_k$ as a 2D Ising / Markov random field with coupling $J$ and external field $h$. The prior is Markov on the antenna lattice: each spin's conditional is a sigmoid of its four neighbours. Calibration with $J \in [0.7, 1.0]$ (just below the critical coupling) produces typical samples that match empirical VR statistics — single large contiguous clusters with smooth boundaries.

• 4.

  Given a pilot observation, the posterior over $\mathbf{m}_k$ is again a 2D Markov field with antenna-dependent external fields driven by the matched- filter LLRs. Loopy belief propagation on this posterior converges within 10–20 sweeps even on $64 \times 64$ arrays and is the workhorse inference routine of the whole chapter.

• 5.

  Full-aperture MMSE costs $\mathcal{O}(N_t^{3})$ flops per user, which is prohibitive beyond $N_t \sim 256$. Partitioning the array into $S$ subarrays of size $M = N_t/S$ drops the cost to $\mathcal{O}(N_t^{3}/S^2)$ — a factor $S^2$ speedup. Adding VR-aware pruning ($|\mathcal{A}_k| \ll S$ active subarrays per user) provides another order of magnitude. Subarray partitioning is the computational backbone of XL-MIMO channel estimation.

• 6.

  In the near field ($r < d_F = 2D^2/\lambda$), the steering vector picks up a quadratic phase term from spherical wavefront curvature. A polar dictionary $\mathbf{A}_{\text{polar}}$ indexed by (azimuth, range) pairs makes a near-field channel with $L$ scatterers $L$-sparse; polar-OMP recovers it from $\mathcal{O}(L \log G)$ pilot measurements instead of the $N_t$ required by far-field DFT dictionaries.

• 7.

  Wrapping the VR mask and the polar channel in a joint EM loop exploits feedback between them: the E-step runs loopy BP on the MRF with LLRs driven by the current channel estimate; the M-step solves a weighted sparse LS on the polar dictionary. EM is guaranteed to monotonically increase the ELBO, and in practice 3–5 iterations suffice to reach within 1 dB of the genie estimator — 4–7 dB better than sequential VR-detect-then-estimate schemes.

• 8.

  The 2D Markov prior partially substitutes for pilot orthogonality: users with disjoint VRs can share pilots because their spatial evidence disambiguates them, yielding a form of spatial pilot decontamination specific to XL-MIMO that complements the covariance-based decontamination of Chapter 3.