Singular Value Decomposition (SVD)

SVD factorizes the channel matrix H into H = U Σ V^H, where Σ is a diagonal matrix of singular values. This transformation creates rank(H) independent spatial subchannels, each with a distinct gain equal to its singular value. By precoding with V and shaping at the receiver with U^H, the cross-stream interference is completely eliminated, converting a complex MIMO channel into a bank of non-interfering scalar channels.

Once the channel is diagonalized, optimal capacity is achieved through water-filling. More power is allocated to eigenmodes with higher singular values (stronger subchannels) and less—or none—to weaker ones.

• The water level is determined by the total power constraint
• Modes with singular values below a noise threshold receive zero power
• This maximizes the sum of log(1 + SNR) across all active eigenmodes

The condition number of the channel matrix—the ratio of the maximum to minimum singular value—directly quantifies how well-conditioned a MIMO channel is for spatial multiplexing.

• A condition number near 1 (all singular values equal) indicates a well-conditioned channel where all spatial streams perform similarly
• A high condition number signals that some eigenmodes are severely attenuated, limiting multiplexing gain
• This metric guides adaptive modulation and coding scheme selection

The number of non-zero singular values defines the rank of the channel matrix, which equals the number of usable spatial streams. The rank is fundamentally limited by:

• min(N_t, N_r) — the minimum of transmit and receive antennas
• Scattering richness — a sparse environment reduces rank
• Antenna correlation — closely spaced elements reduce independence

A rank-deficient channel cannot support full spatial multiplexing regardless of SNR.

When channel state information is available at the transmitter, the right singular vectors in V define the optimal beamforming directions. Transmitting along the dominant eigenvector concentrates energy into the strongest spatial path, maximizing received SNR.

This is particularly effective in low-rank channels or when prioritizing link reliability over multiplexing gain, as it achieves full diversity order without sacrificing array gain.

SVD-based precoding provides the theoretical upper bound for MIMO performance, but practical systems often use suboptimal methods:

• Zero-Forcing inverts the channel but amplifies noise on weak eigenmodes
• MMSE balances interference suppression and noise enhancement
• Codebook-based precoding (LTE/5G) approximates SVD with quantized feedback

SVD requires full CSI at the transmitter, which is costly in FDD systems but feasible in TDD via channel reciprocity.

The complete characterization of a wireless link's propagation properties, including scattering, fading, and power decay. SVD-based precoding requires accurate CSI at the transmitter to decompose the channel matrix into independent eigenmodes. Without precise CSI, the parallel subchannels created by SVD suffer from cross-talk, destroying orthogonality. CSI is typically obtained through pilot-based channel estimation or explicit feedback in FDD systems.

A metric defined as the ratio of the largest to smallest singular value of the channel matrix. A high condition number indicates an ill-conditioned channel where spatial multiplexing gain collapses because one eigenmode dominates. SVD reveals this directly: if the smallest singular value approaches zero, the corresponding subchannel becomes unusable, reducing the effective rank and capacity of the MIMO link.

The linear increase in data rate achieved by transmitting independent streams over parallel spatial paths. SVD mathematically creates these non-interfering eigenmodes by diagonalizing the channel matrix. The number of usable streams equals the rank of the channel matrix—the count of non-negligible singular values. In a rich scattering environment with uncorrelated antennas, this scales with min(N_t, N_r).

A scalable architecture where base stations employ hundreds of antenna elements. SVD-based analysis becomes computationally expensive at this scale, motivating hybrid beamforming that splits precoding between analog phase shifters and low-dimensional digital baseband processing. However, the theoretical foundation remains: as the number of antennas grows, the singular values of the channel matrix harden toward deterministic values due to channel hardening.

The optimal power distribution strategy applied after SVD decomposes the channel. Once the channel is diagonalized into parallel subchannels, water-filling allocates more power to eigenmodes with higher singular values (stronger subchannels) and less to weaker ones, subject to a total power budget. This maximizes the Shannon capacity of the MIMO channel and is the second step after SVD-based eigen-beamforming.