Block Diagonalization (BD)

The core mechanism of BD is the null space projection. For each user k, the precoder Wk must satisfy HjWk = 0 for all j ≠ k, where Hj is the channel matrix of user j.

• This constraint forces the signal intended for user k to be invisible to all other users.
• The precoder is constructed from the right singular vectors corresponding to zero singular values of the aggregate interference channel matrix.
• The existence of a non-trivial null space requires the transmitter to have more antennas than the total number of receive antennas of all other users combined.

BD transforms the multiuser MIMO broadcast channel into a set of parallel, non-interfering single-user MIMO channels. Once inter-user interference is eliminated via null space projection, the remaining per-user channel can be further diagonalized using Singular Value Decomposition (SVD).

• This two-stage process—nulling followed by eigen-beamforming—achieves the full degrees of freedom of the channel.
• The resulting parallel subchannels allow standard water-filling power allocation to maximize sum capacity.
• BD is a suboptimal but computationally tractable alternative to the capacity-achieving Dirty Paper Coding (DPC).

BD critically depends on accurate Channel State Information at the Transmitter (CSIT). The transmitter must know the channel matrices of all users to compute the correct null spaces.

• In Frequency Division Duplex (FDD) systems, CSI is obtained via limited feedback using codebooks and Precoding Matrix Indicators (PMI).
• In Time Division Duplex (TDD) systems, channel reciprocity is exploited to estimate the downlink channel from uplink pilots.
• Imperfect or outdated CSI leads to residual inter-user interference, which is the primary performance bottleneck for BD in mobile environments.

After null space projection, the effective channel for each user is a single-user MIMO channel. BD enables independent optimization of each user's transmission.

• Water-filling power allocation is applied per-user across the eigenmodes of the effective channel to maximize throughput.
• The number of spatial streams per user is bounded by the dimension of the null space, which is determined by the excess of transmit antennas over interfering receive antennas.
• This decoupling simplifies scheduling and link adaptation, as each user's Modulation and Coding Scheme (MCS) can be selected independently.

BD requires computing the SVD of the aggregate interference channel matrix for each user, which has a computational complexity of O(Nt³) where Nt is the number of transmit antennas.

• For large antenna arrays, this cubic scaling becomes a practical limitation.
• Block Diagonalization with Zero-Forcing (BD-ZF) variants reduce complexity by combining null space projection with simpler linear precoding.
• In Massive MIMO systems, simpler linear precoders like Maximum Ratio Transmission (MRT) or Zero-Forcing (ZF) are often preferred due to channel hardening effects.

BD occupies a middle ground between optimal non-linear and simpler linear precoding strategies.

• vs. Dirty Paper Coding (DPC): BD is suboptimal but computationally feasible; DPC achieves the capacity region but is non-causal and practically unrealizable.
• vs. Zero-Forcing (ZF): BD generalizes ZF to users with multiple antennas; ZF treats each receive antenna as an independent user.
• vs. MMSE Precoding: BD eliminates interference completely, while Minimum Mean Square Error (MMSE) precoding allows controlled interference to improve signal-to-noise ratio at low transmit power.