Problem

The landscape of wireless traffic is rapidly transforming due to the proliferation of AI-driven services, machine-type communications, and the massive growth of IoT devices. This shift results in dense, bursty, highly uplink-oriented, and latency-critical communication patterns, demanding frequent and rapid scheduling decisions. Multi-User Multiple-Input Multiple-Output (MU-MIMO) is crucial for supporting numerous concurrent connections through spatial multiplexing in these environments. However, the inherent combinatorial complexity of MU-MIMO user selection presents significant scalability barriers. Existing user selection approaches, which involve selecting a subset of users from a large pool, face computational complexity that grows prohibitively with the number of candidate users and spatial layers. This makes near-optimal heuristic methods impractical for the dense and dynamic scenarios characteristic of emerging AI-native and machine-driven traffic, where parallel execution and low latency are paramount. Traditional methods like Semi-orthogonal User Selection (SUS), greedy approaches, and subspace compatibility methods often rely on sequential operations, heavy matrix computations, or exhaustive searches, hindering their scalability and real-time applicability.

Method

This work introduces the Space Splitting-based User Selection (SS-US) algorithm, a novel low-complexity method designed for scalable multi-user scheduling in delay-critical, massive-connectivity scenarios. SS-US departs from traditional subset-based selection by operating on three core principles: partitioning spatial degrees of freedom into orthonormal subspaces, evaluating user compatibility with a one-shot correlation per spatial direction, and enabling intrinsic parallelism by constructing multiple independent spatial hypotheses. Unlike existing techniques that involve iterative or combinatorial exploration, SS-US reformulates user selection as a directional matching problem in the spatial domain. The algorithm begins by identifying the user with the strongest single-stream rate, whose normalized channel vector becomes the initial basis vector. From this, SS-US constructs L independent orthonormal bases. For each basis vector, it evaluates user compatibility using a correlation metric and selects users that maximize the product of this correlation and their single-stream rate, provided the correlation exceeds a defined threshold (α). This process avoids repeated subset evaluations, matrix inversions, and iterative orthogonality refinements, instead enabling concurrent processing. The final selected user set is the one among the L bases that exhibits the highest average channel norm, balancing spatial separability with signal strength.

Results

Simulations conducted using a MATLAB link-level simulator with the 3GPP CDL-B channel model demonstrated the effectiveness of SS-US. The algorithm achieved a significant reduction in computational complexity, by over three orders of magnitude (more than 2500 times lower than mCore+ in larger configurations, half of SUS, and over 10 times lower than GZF for M=8). This massive complexity reduction is a direct result of its parallelizable design and the replacement of costly iterative operations with lightweight correlation analyses. Regarding spectral efficiency, SS-US showed comparable performance to state-of-the-art baselines, with observed differences typically within a ±3-8% range and often statistically insignificant. The study also explored the impact of two key tunable parameters: L (number of parallel orthonormal bases) and α (correlation threshold). Increasing L generally improved spectral efficiency but linearly increased complexity, offering a controllable performance-complexity trade-off. A higher α enforced stricter orthogonality, which could yield performance gains in scenarios with multiple parallel bases and large user pools by maintaining both high orthogonality and high user norms. While some performance gaps were noted in high-SNR regimes for M=8 compared to GZF and mCore+, SS-US consistently remained close to the top-performing methods across diverse MIMO configurations, user densities, and channel conditions, despite its drastically lower complexity.

Implications

The Space Splitting-based User Selection (SS-US) algorithm presents a compelling solution for the evolving challenges in wireless communication. By offering over three orders of magnitude lower computational complexity and inherent massive parallelization capabilities, SS-US addresses the critical scalability issues faced by traditional MU-MIMO user selection schemes in the context of dense, uplink-oriented, and latency-sensitive AI-native traffic. Its ability to maintain spectral efficiency comparable to high-performing baselines, while drastically reducing computational burden, positions SS-US as a strong candidate for next-generation MU-MIMO deployments. This method enables more frequent and responsive scheduling decisions, crucial for supporting the massive connectivity and stringent latency requirements of future AI-driven services and IoT ecosystems. Future research will explore advanced orthonormal matrix generation strategies, extend support to single-user MIMO, and integrate SS-US into comprehensive system schedulers for joint optimization of various communication resources, considering hardware-software flexibility.