Constellation Precoding for MIMO Systems

• Constellation precoding is a transformative technique that applies algebraic or numerical rotations to modulation vectors before transmission in MIMO systems.
• It maximizes spatial diversity and coding gain by optimizing the distribution of symbol differences across subchannels, crucial for full diversity gains.
• The design leverages criteria such as minimum distance and product distance, effectively balancing complexity, throughput, and error performance in both fully and partially precoded scenarios.

Constellation precoding is a multi-antenna transmission framework in which algebraic or numerical transformations are applied to vectors of modulation symbols prior to transmission, with the aim of controlling diversity, coding gain, and the distribution of symbol differences after spatial channel transformation. Constellation precoding is particularly significant in the context of multiuser MIMO and SVD-based beamforming, where it enables the achievement of full spatial diversity in scenarios where straightforward beamforming or spatial multiplexing alone would suffer performance loss. The design of constellation precoders is deeply connected to criteria based on minimum distance, product distance, and algebraic rotations, and forms a cornerstone of modern uncoded and coded SVD-MIMO systems pursuing high reliability at high data rates (0903.4738).

1. Mathematical Model: SVD MIMO and Precoder Insertion

Consider a flat-fading MIMO channel $H \in \mathbb{C}^{M \times N}$ with i.i.d. Gaussian entries and perfect CSI at both ends. The channel can be factorized by SVD: $H = U \Sigma V^H$ where $U$ and $V$ are unitary, and $$\Sigma = \text{diag}(\lambda_1, \dotsc, \lambda_S, 0, ..., 0)$$, $S = \min\{M,N\}$.

When transmitting $S$ symbols in parallel, the transmitter uses the first $S$ columns of $V$ for precoding, and the receiver uses the first $S$ columns of $H = U \Sigma V^H$0 for combining. The equivalent channel post beamforming is $H = U \Sigma V^H$1, leading to decoupled spatial subchannels. Sending $H = U \Sigma V^H$2 yields the received vector: $H = U \Sigma V^H$3 where $H = U \Sigma V^H$4.

Constellation precoding augments this framework by applying an $H = U \Sigma V^H$5 unitary (or semi-unitary) matrix $H = U \Sigma V^H$6 to $H = U \Sigma V^H$7, so the actual transmitted vector is $H = U \Sigma V^H$8. More generally, $H = U \Sigma V^H$9 can be written as

$U$0

where $U$1 is an $U$2 rotation, $U$3 is a permutation, and $U$4 allows the design of fully or partially precoded systems (0903.4738).

2. Diversity Analysis and Design Criteria

Diversity Order and Pairwise Error Probability

When SVD is used for beamforming without precoding, full diversity is achieved for single-stream systems, but the diversity order reduces for uncoded spatial multiplexing ($U$5), since dominant error events are tied to lower singular values. The pairwise error probability (PEP) between codewords $U$6 and $U$7 is: $U$8 Averaging over the channel, the diversity order is governed by the first index $U$9 such that the $V$0-th coordinate of $V$1 is nonzero. The asymptotic diversity order is then $V$2 (0903.4738).

Full diversity (order $V$3) is achieved if for every $V$4, the first coordinate of $V$5 is nonzero, i.e., $V$6. This leads to key design criteria for $V$7:

• Φ₁: Maximize the minimum over all codeword pairs and coordinates: $V$8
• Φ₂: Maximize the minimum of the dominant coordinate: $V$9
• Φ₃: Maximize the minimum geometric mean: $$\Sigma = \text{diag}(\lambda_1, \dotsc, \lambda_S, 0, ..., 0)$$0

Designs based on algebraic rotations, IFFT matrices, or full parameterizations of unitary matrices (via Givens rotations, etc) can be deployed to meet these criteria (0903.4738).

Partial Precoding

If only $$\Sigma = \text{diag}(\lambda_1, \dotsc, \lambda_S, 0, ..., 0)$$1 symbols are precoded, those symbols are mapped onto selective subchannels with indices $$\Sigma = \text{diag}(\lambda_1, \dotsc, \lambda_S, 0, ..., 0)$$2 (precoded) and $$\Sigma = \text{diag}(\lambda_1, \dotsc, \lambda_S, 0, ..., 0)$$3 (non-precoded). The diversity order is

$$\Sigma = \text{diag}(\lambda_1, \dotsc, \lambda_S, 0, ..., 0)$$4

with $$\Sigma = \text{diag}(\lambda_1, \dotsc, \lambda_S, 0, ..., 0)$$5. Placing precoded symbols on the least favorable (largest) subchannel indices minimizes $$\Sigma = \text{diag}(\lambda_1, \dotsc, \lambda_S, 0, ..., 0)$$6 and maximizes achievable diversity (0903.4738).

3. Performance and Trade-Offs

Extensive simulation demonstrates that fully precoded multiple beamforming (FPMB) with well-designed $$\Sigma = \text{diag}(\lambda_1, \dotsc, \lambda_S, 0, ..., 0)$$7 provides:

• Full $$\Sigma = \text{diag}(\lambda_1, \dotsc, \lambda_S, 0, ..., 0)$$8 diversity order, as indicated by parallel BER-vs-SNR slopes to single-beamforming.
• Coding gain improvements up to 6 dB at high SNR for the geometric mean design criterion Φ₃.
• For partially precoded multiple beamforming (PPMB), the diversity can exceed that of plain multiple beamforming, approaching or reaching full diversity when subchannel assignment is optimized.
• Uncoded multiple-beamforming without precoding exhibits a strictly lower diversity order and thus inferior BER slope (0903.4738).

The complexity-diversity-throughput space is stratified as:

• PSB (S=1): minimal complexity, full diversity, low throughput.
• MB (S>1), no precoding: high throughput, low diversity, low complexity.
• FPMB: full diversity and high throughput, at the cost of an S-dimensional lattice decoding problem (sphere decoding or ML search).
• PPMB: intermediate diversity, reduced ML dimension (and complexity), suitable for complexity-constrained systems (0903.4738).

4. Implementation and Detection Algorithms

ML detection for constellation precoding entails minimizing $$\Sigma = \text{diag}(\lambda_1, \dotsc, \lambda_S, 0, ..., 0)$$9 over all $S = \min\{M,N\}$0.

• For the fully precoded subspace, this is a lattice search of dimension S over the rotated constellation.
• Sphere decoding is used to make ML practical, with additional complexity reductions possible by
  • Precomputing certain products,
  • Exploiting zero patterns (block-orthogonality) in the QR decomposition of the real-valued equivalent,
  • Efficient initial radius selection based on ZF-DFE estimates (0911.0709, Li et al., 2010).

Simulation shows orders-of-magnitude reduction in computational complexity for both bit and symbol-metric evaluations, making high-dimensional (e.g., S=4, 64-QAM) ML detection feasible (0911.0709).

5. Channel Coding Interaction: Coded Systems

In bit-interleaved coded multiple beamforming (BICMB), the diversity order with no constellation precoding is subject to the code rate $S = \min\{M,N\}$1 and the number of subchannels: $S = \min\{M,N\}$2 Full diversity ($S = \min\{M,N\}$3) is guaranteed only when $S = \min\{M,N\}$4 (0908.3702).

If constellation precoding is included (BICMB-CP), full diversity is restored for any $S = \min\{M,N\}$5 and $S = \min\{M,N\}$6, under the same first-row nonvanishing criterion for $S = \min\{M,N\}$7. The key implementation insight is that constellation precoding spreads coded bits across all subchannels, ensuring that every bit sees every singular value and thus reaps the full spatial diversity, independent of code rate. This effect is provably maintained, both theoretically and in simulation, for a range of MIMO sizes and QAM orders (0908.3702, 0911.0709, Li et al., 2010).

6. Broader Impact and Practical Implications

Constellation precoding enables the combination of spatial multiplexing (high-rate) with full spatial diversity (robustness) in uncoded and coded SVD-MIMO transmission, using only a linear transformation at the transmitter and no FEC. Practical systems can choose fully precoded, partially precoded, or unprecoded options to trade off diversity against complexity and latency.

The underlying theory of constellation difference transformation to ensure error events are always linked to the largest singular value (and thus maximal diversity) has ramifications for robust multiantenna design in wireless standards, especially when channel coding is impractical or code rate/diversity constraints would otherwise limit system throughput (0903.4738). The design framework (Φ₁–Φ₃) also forms a foundation for further studies in symbol-level and nonlinear precoding.

• Park & Ayanoglu, “Constellation Precoded Beamforming” (0903.4738)
• Park & Ayanoglu, “Constellation Precoded Multiple Beamforming” (0911.0709)
• Park & Ayanoglu, “Bit-Interleaved Coded Multiple Beamforming with Constellation Precoding” (0908.3702)
• Iyengar et al., “Reduced Complexity Decoding for Bit-Interleaved Coded Multiple Beamforming with Constellation Precoding” (Li et al., 2010)