Abstract
In this paper, we show that the adaptive projected subgradient method (APSM) is bounded perturbation resilient. To illustrate a potential application of this result, we propose a set-theoretic framework for MIMO detection, and we devise algorithms based on a superiorized APSM. Various low-complexity MIMO detection algorithms achieve excellent performance on i.i.d. Gaussian channels, but they typically incur high performance loss if realistic channel models (e.g., correlated channels) are considered. Compared to existing low-complexity iterative detectors such as individually optimal large-MIMO approximate message passing (IO-LAMA), the proposed algorithms can achieve considerably lower symbol error ratios over correlated channels. At the same time, the proposed methods do not require matrix inverses, and their complexity is similar to IO-LAMA.
Introduction
In modern Multiple-Input Multiple-Output (MIMO) wireless communications, the receiver's primary task is to estimate transmitted signals from noisy, scrambled observations across multiple antennas. While Maximum Likelihood (ML) detection provides the theoretical "gold standard" for accuracy, its computational complexity is NP-hard, making it impractical for large-scale systems.
Engineers often rely on iterative methods like Approximate Message Passing (AMP). While AMP-based methods like IO-LAMA are highly effective for idealized i.i.d. Gaussian channels, they frequently fail in realistic environments where channels are spatially correlated—such as dense urban areas where signals reflect off the same structures. More robust alternatives like OAMP exist but require computationally expensive matrix inversions. This research introduces a middle ground: a detector that maintains low computational complexity (no matrix inverses) while providing robustness against realistic channel correlations.
The Superiorized APSM Framework
The core of the proposed solution lies in the Adaptive Projected Subgradient Method (APSM) and the concept of Superiorization. APSM is a flexible iterative framework designed to find a point within the intersection of multiple constraint sets.
The authors prove a critical theoretical property: APSM is bounded perturbation resilient. This means the algorithm can be "nudged" during each iteration without compromising its convergence guarantees. By applying these nudges strategically, one can steer the algorithm toward a "superior" solution—one that not only satisfies the basic data-matching constraints but also aligns with the discrete nature of digital symbol alphabets.
Perturbation Strategies
The paper explores two specific superiorization strategies for MIMO detection:
- APSM-L2: Employs hard projections onto the constellation set.
- APSM-L1: Utilizes a "soft-thresholding" approach inspired by compressed sensing. This method penalizes estimates that fall outside the valid symbol set, allowing the algorithm to explore the solution space more effectively before converging.
Performance on Correlated Channels
The proposed algorithms were evaluated against standard benchmarks using realistic 3GPP spatial channel models, which incorporate the non-Gaussian effects and correlations that typically degrade iterative detectors.
As shown in the figure above, the proposed APSM-L1 and APSM-L2 variants demonstrate significant performance gains over traditional methods in correlated environments. While IO-LAMA (an AMP-based method) fails to reduce the error rate effectively, the superiorized APSM methods eventually converge to a Symbol Error Ratio (SER) nearly an order of magnitude lower than basic APSM and significantly lower than the LMMSE baseline. Notably, the L1-norm based "soft" perturbations prove more effective than the "hard" L2 projections.
Complexity and Practicality
A key advantage of the APSM-based approach is its computational efficiency. Unlike OAMP or LMMSE, which require matrix inversions with $O(K^3)$ complexity, the proposed method relies only on matrix-vector multiplications. This results in a per-iteration complexity of $O(NK)$, making it as efficient as IO-LAMA and highly suitable for high-speed hardware implementations like FPGAs and ASICs.
However, there is a trade-off in convergence speed. While AMP-based methods typically converge in 10-20 iterations on simple channels, the APSM-based methods require approximately 150-175 iterations to reach peak performance on correlated channels.
Conclusion and Future Work
By proving the bounded perturbation resilience of APSM, this work provides a robust framework for designing matrix-agnostic MIMO detectors. The resulting algorithms offer a compelling balance between the low complexity of AMP and the robustness of matrix-inversion-based methods. Future research may focus on "Deep Unfolding," where the iterations are mapped to neural network layers to learn optimal step sizes and perturbation weights automatically, potentially reducing the required iteration count for real-time applications.
Summary updated