Low-Complexity Near-ML Decoding of Large Non-Orthogonal STBCs Using PDA

Non-orthogonal space-time block codes (STBC) from cyclic division algebras (CDA) having large dimensions are attractive because they can simultaneously achieve both high spectral efficiencies (same spectral efficiency as in V-BLAST for a given number of transmit antennas) {\em as well as} full transmit diversity. Decoding of non-orthogonal STBCs with hundreds of dimensions has been a challenge. In this paper, we present a probabilistic data association (PDA) based algorithm for decoding non-orthogonal STBCs with large dimensions. Our simulation results show that the proposed PDA-based algorithm achieves near SISO AWGN uncoded BER as well as near-capacity coded BER (within about 5 dB of the theoretical capacity) for large non-orthogonal STBCs from CDA. We study the effect of spatial correlation on the BER, and show that the performance loss due to spatial correlation can be alleviated by providing more receive spatial dimensions. We report good BER performance when a training-based iterative decoding/channel estimation is used (instead of assuming perfect channel knowledge) in channels with large coherence times. A comparison of the performances of the PDA algorithm and the likelihood ascent search (LAS) algorithm (reported in our recent work) is also presented.

💡 Research Summary

The paper tackles the long‑standing problem of decoding large‑dimensional, non‑orthogonal space‑time block codes (STBCs) constructed from cyclic division algebras (CDA). Such codes are attractive because they achieve full‑rate transmission (the same number of complex symbols per channel use as V‑BLAST) while providing full transmit diversity, and they can reach very high spectral efficiencies (tens of bits per second per Hertz) when the code dimension grows to hundreds of real symbols. However, maximum‑likelihood (ML) detection for these codes becomes infeasible as the dimensionality increases, since the search space grows exponentially.

The authors propose a probabilistic data association (PDA) algorithm adapted to the MIMO detection problem. The complex‑valued system model is first converted to an equivalent real‑valued linear model y = H′x + n. Each transmitted complex symbol is expressed as a linear combination of its constituent bits, which are mapped to a binary vector b ∈ {+1, –1}. The detection problem is then to estimate b. In each iteration, the algorithm updates the a‑posteriori probability of every bit sequentially. For a given bit, the interference from all other bits is approximated as Gaussian; consequently, the conditional distribution of the received vector y given the bit value is Gaussian with a mean µ and covariance C that can be written in closed form. The likelihood ratio Λ is computed from these Gaussian densities, and the bit probabilities are updated accordingly. This process is repeated for a predetermined number of iterations, after which hard decisions (or soft values for a channel decoder) are produced.

A major contribution is the reduction of computational complexity. Direct inversion of the covariance matrix C for each bit would require O(N³) operations (N = 2N_r p, the dimension of the real‑valued system). By exploiting the matrix inversion lemma, the authors show that updating the inverse after a single bit change amounts to a rank‑one update, which can be performed in O(N²) time. The same technique is applied to compute the inverse of C from a pre‑computed matrix D, further lowering the per‑iteration cost. Overall, the algorithm’s complexity scales as O(2qk·I·N²), where q = log₂√M (bits per real symbol), k is the number of complex symbols, and I is the number of iterations—making it practical for codes with several hundred dimensions.

Performance is evaluated under several realistic conditions. With i.i.d. Rayleigh fading, perfect channel state information at the receiver (CSIR), a 12×12 CDA code (256 complex symbols), 4‑QAM, and a rate‑3/4 turbo code (≈18 bps/Hz), the PDA decoder achieves uncoded bit‑error rates (BER) that are essentially indistinguishable from the SISO AWGN benchmark. Coded BER lies within about 5 dB of the Shannon capacity, confirming near‑ML behavior. When spatial correlation is introduced (using the model of Gesbert et al.), the loss can be mitigated by increasing the number of receive antennas; the simulations show that adding receive dimensions restores performance close to the i.i.d. case. For channels with large coherence time, a training‑based iterative scheme that jointly refines channel estimates and PDA detection is presented; after a few iterations the BER matches that obtained with perfect CSIR, demonstrating robustness to imperfect channel knowledge.

The paper also compares PDA with the previously proposed likelihood ascent search (LAS) algorithm. PDA outperforms LAS at low SNRs, especially for higher‑order constellations such as 16‑QAM, and it is more resilient to spatial correlation. Moreover, PDA’s sequential bit‑wise updates lead to a simpler memory access pattern and lower overall computational load, making it more attractive for hardware implementation.

In summary, the authors deliver a comprehensive solution for the practical decoding of large non‑orthogonal CDA‑based STBCs: (1) an adapted PDA algorithm that yields near‑ML performance, (2) a systematic method to reduce matrix‑inverse complexity via rank‑one updates, (3) extensive analysis of robustness to spatial correlation and channel‑estimation errors, and (4) a clear performance advantage over existing low‑complexity alternatives such as LAS. This work paves the way for deploying high‑spectral‑efficiency, full‑diversity MIMO systems with realistic receiver complexity.