Precoding

Zero-Forcing (ZF) precoding eliminates multi-user interference by multiplying the data vector with the pseudo-inverse of the channel matrix. This forces the received signal at each intended user to be a scaled version of the transmitted symbol, with zero contribution from other users' data streams.

• Mechanism: W = H^H (H H^H)^-1, where H is the combined channel matrix
• Trade-off: Achieves perfect interference cancellation but can amplify noise at low SNR
• Application: High-SNR scenarios in massive MIMO where noise amplification is negligible
• Computational Cost: Requires matrix inversion, scaling as O(K^3) with number of users K

SLNR precoding maximizes the ratio of desired signal power at the intended user to the interference power leaked to all other users plus noise. Unlike pure ZF, it balances interference suppression against noise enhancement.

• Objective: Maximize SLNR_k = ||H_k w_k||^2 / (∑_{i≠k} ||H_i w_k||^2 + σ^2)
• Advantage: Decouples into per-user generalized eigenvalue problems, enabling parallel computation
• Robustness: Performs well even with imperfect CSI, gracefully degrading rather than catastrophically failing
• Closed Form: Each user's precoder is the dominant generalized eigenvector of a matrix pair

Regularized Zero-Forcing (RZF) adds a regularization parameter proportional to the inverse SNR to the ZF matrix inversion, mitigating noise amplification at low signal-to-noise ratios. It converges to matched filtering at low SNR and ZF at high SNR.

• Formulation: W = H^H (H H^H + α I)^-1, where α scales with noise variance
• Optimality: Approaches the MMSE solution when α = Kσ²/P, where P is transmit power
• Massive MIMO: As antenna count grows, RZF approaches optimal dirty paper coding performance
• Implementation: Dominant computational cost is the matrix inversion of size K×K

Dirty Paper Coding (DPC) is the information-theoretic optimal precoding strategy that achieves the MIMO broadcast channel capacity. It pre-subtracts known interference at the transmitter using successive encoding, as if the interference were not present.

• Principle: If interference is known non-causally at the transmitter, it can be canceled without power penalty
• Implementation: Tomlinson-Harashima Precoding (THP) approximates DPC using modulo arithmetic and feedback filters
• Performance: Sets the theoretical upper bound against which all linear precoders are benchmarked
• Complexity Barrier: Practical DPC requires complex sphere encoding, limiting real-world deployment

In FDD systems without channel reciprocity, the UE selects the optimal precoding matrix from a standardized codebook and reports only the index back to the base station, dramatically reducing feedback overhead.

• 3GPP Type-I: DFT-based codebook for single-user MIMO with wideband and subband granularity
• 3GPP Type-II: High-resolution codebook using linear combination of multiple DFT beams with amplitude and phase quantization
• Selection Criterion: Maximizes mutual information or signal-to-interference-plus-noise ratio
• Trade-off: Larger codebooks improve precision but increase search complexity and feedback bits

Deep learning models learn to generate precoding matrices directly from imperfect or compressed CSI, outperforming model-based approaches when channel models are mismatched or hardware impairments are present.

• Architectures: CNNs process spatial channel structure; transformers capture long-range antenna correlations
• Training: Supervised learning on channel realizations with the true precoder as label, or unsupervised via maximizing sum-rate
• Advantage: Robust to non-linear hardware effects like power amplifier distortion and I/Q imbalance
• Deployment: Inference is feed-forward and parallelizable, enabling real-time precoder computation per slot