Information Theory
MIMO Channel Capacity
Цели урока
- Derive MIMO capacity via SVD and water-filling over singular values
- Distinguish ergodic and outage capacity for fading channels
- Explain the diversity-multiplexing tradeoff and its practical implications
- Understand channel hardening and Massive MIMO capacity
Предварительные знания
- Shannon capacity for AWGN single-antenna channels
- SVD and eigenvalue decomposition of complex matrices
- Water-filling power allocation as a Lagrangian KKT solution
- Random matrix theory: Marchenko-Pastur and free probability basics
Why do n antennas give n-fold capacity growth rather than just n-fold SNR gain? How does SVD turn a complex MIMO channel into r independent AWGN channels?
- Wi-Fi 6 (8x8 MIMO): linear capacity growth to 9.6 Gbps vs 1.2 Gbps for SISO
- 5G NR Ericsson AIR 6468: 64 antennas, channel hardening, 100 users in one band
- RIS (Reconfigurable Intelligent Surfaces): +10 dB SNR via programmable reflections
- DetNet (Samsung Research): neural MIMO detector replacing exponential ML search
From One Antenna to 64: 25 Years of MIMO
Foschini and Gans (1996) and Telatar (1995/1999) independently discovered linear capacity growth for MIMO. BLAST (Bell Laboratories Layered Space-Time): first practical MIMO algorithm (1996). Alamouti (1998): space-time coding for 2x1 MIMO. IEEE 802.11n (2009): first Wi-Fi standard with MIMO (4x4). LTE (2011): 4x4 MIMO mandatory. Massive MIMO (Marzetta, 2010): channel hardening theory at n_T >> 1. 5G NR (2019): 64 TRX Massive MIMO. Telatar received the IEEE IT Society Paper Award in 2000.
Telatar's Formula and SVD Decomposition
Emre Telatar computed MIMO channel capacity in 1995 (Bell Labs technical memo, published 1999). The result shocked engineers: n antennas give linear capacity growth - C ~ n * log(1 + SNR). Compare to SISO: C ~ log(1 + n*SNR) - only logarithmic growth. A difference in kind. Wi-Fi 6 implements 8x8 MIMO - 8x linear growth.
Wi-Fi 6 (802.11ax): 8x8 MIMO at 80 MHz, 256-QAM. Number of sub-channels r = min(8,8) = 8. Each sub-channel at SNR = 30 dB: C_i ~ 10 bits/use. Total: 8 * 10 = 80 bits/use * 80 MHz = 6.4 Gbps. The standard specifies 9.6 Gbps - accounting for 4 spatial streams MU-MIMO and multiuser beamforming.
Why does SVD decompose MIMO into independent channels?
H = U*Sigma*V^H. After transforms x_tilde = V^H*x and y_tilde = U^H*y: y_tilde = Sigma*x_tilde + z_tilde. Sigma is diagonal, so y_tilde_i = sigma_i*x_tilde_i + z_tilde_i - independent SISO channels. U^H and V are unitary: they do not change signal or noise power.
Ergodic Capacity, Outage, and DMT
Telatar's formula is for deterministic H. A real wireless channel is random: Rayleigh fading, Doppler shift, inter-cell interference. For random H, two capacity concepts arise depending on how fast the channel changes relative to the codeword length.
LTE MIMO: Theory Meets Production
LTE Cat.15 (4x4 MIMO, 4 CA, 256-QAM): theoretical peak 800 Mbps. Actual capacity: 150-300 Mbps in urban networks. Gap: 3 dB additional losses from imperfect channel estimation (CSI feedback overhead), antenna correlation (d < lambda/2 = 15 cm at 1 GHz), and inter-cell interference (SINR instead of SNR). DMT: LTE selects a point closer to d=4 (diversity) rather than r=4 (multiplexing) for reliability in bad channel conditions.
Channel hardening in Massive MIMO: when n_T >> n_R, the product H*H^H/n_T approaches I_{n_R} by the law of large numbers. The channel becomes deterministic. Ergodic capacity converges to n_R * log2(1 + P*n_T/sigma^2) - linear growth in both n_T and n_R.
What is the fundamental difference between ergodic capacity and outage capacity for MIMO?
Fast fading: block covers many H realizations -> capacity = average E[log det]. Slow fading: H fixed over block -> cannot average. Need C_out: rate guaranteed in (1-epsilon) fraction of cases. For mobile communications (100 km/h, 2 GHz): coherence time ~0.1 ms = fast fading. For fixed access (1 ms block): slow fading.
Massive MIMO and Neural Detectors
Massive MIMO: n_T or n_R >> 1. Nokia and Ericsson deploy 64-antenna 5G NR base stations. As n_T grows large, HH^H/n_T converges to I_{n_R} - channel hardening. The channel becomes nearly deterministic, enabling K simultaneous users without inter-user interference.
Neural MIMO Detectors: DetNet
DetNet (Samuel et al., 2019): MIMO ML detection approximated by a T-layer neural network. Each layer: W_e * y + W_s * s + projected gradient step. Training: minimize reconstruction error on (H, s, y) pairs. At n_T=64 antennas with QPSK: ML detector complexity O(4^64) - computationally infeasible. DetNet with T=30: complexity O(T*n_T^2), within 1 dB of ML. Samsung Research uses this in 5G base-station silicon.
Reconfigurable Intelligent Surfaces (RIS): programmable reflecting surfaces with N phase-shift elements. Channel with RIS: H_eff = H2 * Theta * H1, where Theta = diag(e^{j*phi_1}, ..., e^{j*phi_N}). Capacity: C(Theta) = log2 det(I + rho * H_eff * H_eff^H). Optimization of Theta is a non-convex problem. Theoretical gain at N=100: up to 10 dB in SNR-equivalent under fading. China Mobile (2023) deployments: +3 dB in dead zones.
What is channel hardening in Massive MIMO and why does it matter?
Channel hardening: randomness of H averages out at large n_T. Rows of H (channels of different users) become nearly orthogonal - favorable propagation. Consequences: (1) simple beamforming = zero-forcing; (2) capacity close to AWGN without fading; (3) K users in the same band without interference.
Connection to Other Topics
MIMO capacity combines: Shannon's theorem (AWGN channel as building block), SVD (linear algebra), water-filling (from R(D) theory), and random matrix theory (Marchenko-Pastur for Massive MIMO). In engineering: OFDM + MIMO = OFDMA, the foundation of 4G/5G. Beamforming = transmission in the V-subspace of the channel matrix. In ML: neural detectors (DetNet, OAMP-Net) replace computationally heavy ML decoding. RIS: programming matrix H to maximize C.
- Channel Capacity — MIMO capacity is the multi-antenna extension of the AWGN Shannon formula
- Rate-Distortion Theory — Water-filling power allocation reuses the dual variable structure from R(D)
- Network Information Theory — Multi-user MIMO is the antenna-domain analogue of the MAC/BC problems
Итоги
- MIMO via SVD: r = min(n_T,n_R) independent AWGN channels; water-filling over sigma_i^2
- Linear growth: C ~ r*log(1+SNR) - multiplexing gain r; vs log(1+r*SNR) for SISO
- DMT (Zheng-Tse): diversity d + multiplexing r <= n_T*n_R, a fundamental tradeoff
- Massive MIMO: channel hardening as n_T grows; C ~ n_R*log(1 + P*n_T/sigma^2)
Вопросы для размышления
- Why is linear capacity growth from MIMO fundamentally better than logarithmic growth from increasing power in SISO?
- How does the diversity-multiplexing tradeoff manifest in real LTE/5G standard parameter choices?
- What limits real Massive MIMO systems from achieving the theoretical capacity bound?