Abstract
We consider the joint constellation design problem for the two-user non-coherent multiple-access channel (MAC). Based on an analysis on the non-coherent maximum-likelihood (ML) detection error, we propose novel design criteria so as to minimize the error probability. Based on these criteria, we propose a simple and efficient construction consisting in partitioning a single-user constellation. Numerical results show that our proposed metrics are meaningful, and can be used as objectives to generate constellations through numerical optimization that perform better than other schemes for the same transmission rate and power.
Introduction
In modern MIMO wireless communications, receivers typically rely on coherent detection, using pilot signals to estimate Channel State Information (CSI). However, in high-mobility scenarios or massive MIMO arrays, the pilot overhead can become prohibitive. This motivates non-coherent communication, where data is transmitted without explicit channel estimation. While single-user non-coherent design is well-established, the Multiple-Access Channel (MAC)—where multiple users transmit to a single receiver simultaneously—presents significant challenges. The receiver must distinguish between symbols from different users while handling inter-user interference without knowing channel coefficients.
Methodology and Design Metrics
The authors focus on the Maximum Likelihood (ML) detector and analyze the Pairwise Error Probability (PEP). They derive a design metric based on the expected Pairwise Log-Likelihood Ratio (PLLR). At high Signal-to-Noise Ratios (SNR), this simplifies to a tractable metric, $d_{min}(\mathcal{X})$, measuring the "distance" between joint symbol matrices.
Two primary solutions are proposed for the resulting non-convex optimization:
- Numerical Optimization: Gradient descent on the "oblique manifold" using the
manopttoolbox. - Constellation Partitioning: A heuristic where a high-quality single-user constellation is partitioned into disjoint sets for each user.
The designs were evaluated via Monte Carlo simulations of a MIMO MAC with isotropic Rayleigh block fading, comparing against standard pilot-based schemes and existing non-coherent benchmarks.
Key Insights
- Subspace Identifiability: Information in non-coherent MIMO is carried in the subspace of the transmitted matrix. In the multi-user case, the joint signal subspace must remain identifiable despite random channel gains.
- Error Types: Detection failures occur as "simultaneous" errors (both users misidentified) or "one-sided" errors (one user misidentified). The $d_{min}$ metric balances protection against both.
- Partitioning Efficiency: Partitioning a single-user Grassmannian constellation is highly effective, transforming a complex joint optimization into a simpler graph-partitioning problem.
Results
Numerical results confirm that the derived $d_{min}$ metric is a robust proxy for actual ML error probability at high SNR. The optimized constellations consistently achieve lower Symbol Error Rates (SER) than pilot-based schemes, particularly in low-to-mid SNR ranges.
The partitioning construction performed remarkably well and demonstrated better scalability to larger constellation sizes compared to full numerical optimization.
Conclusion and Future Work
While effective for two users, extending this framework to $K$ users increases complexity exponentially. Future research will address asymmetrical rates, power optimization, and the computational cost of finding global optima in non-convex Grassmannian spaces.
Summary updated