Finite Horizon Optimization for Large-Scale MIMO

Yi Feng, Kaiming ShenarXiv 2025 · DOI · arXiv

In the evolving landscape of 5G and 6G communications, Large-Scale MIMO (Multiple-Input Multiple-Output) stands as a pivotal technology. By deploying thousands of antennas at a base station, network operators can achieve unprecedented spectral efficiency. However, this hardware advancement introduces a significant computational challenge: the standard beamforming algorithms used to coordinate these antennas become prohibitively slow as the system scales.

A recent study by Yi Feng and Kaiming Shen from the Chinese University of Hong Kong, Shenzhen, introduces a novel Finite Horizon Optimization approach. By pivoting from traditional "infinite convergence" goals to optimizing performance within a strictly limited number of steps, they have developed a method that significantly outperforms industry-standard algorithms in real-time scenarios.

The Computational Bottleneck in Large-Scale MIMO

The current gold standard for beamforming is the Weighted Minimum Mean-Square-Error (WMMSE) algorithm. While theoretically robust, WMMSE faces a major scalability hurdle. Each iteration requires the inversion of an $M \times M$ matrix, where $M$ is the number of transmit antennas. In large-scale systems where $M$ reaches into the thousands, this $O(M^3)$ operation becomes a massive bottleneck, preventing real-time implementation.

The authors propose a shift in perspective: instead of seeking perfect convergence through costly matrix inversions, can we achieve near-optimal results using only a few computationally cheap iterations?

Methodology: Finite Horizon and Chebyshev Acceleration

The proposed framework rests on three key pillars:

  1. Quadratic Reformulation via Fractional Programming (FP): The complex, non-convex "weighted sum-rate" maximization problem is first recast into a series of unconstrained quadratic sub-problems.
  2. Gradient Descent (GD) Integration: To avoid matrix inversions, the authors replace exact sub-problem solutions with Gradient Descent. This reduces the per-iteration complexity to $O(M^2)$ by relying on matrix-vector multiplications.
  3. Finite Horizon Step Sizes: This is the core innovation. Traditional optimization focuses on asymptotic convergence—how an algorithm behaves as iterations approach infinity. However, in practical wireless networks, time constraints often limit optimization to 5 or 10 iterations. Using Chebyshev polynomial theory, the authors calculate optimal, varying step sizes ($\eta_t$) designed to minimize error specifically at a target $T$-th iteration.

Key Insights: Asymptotic vs. Finite Horizon Performance

The paper highlights a critical trade-off in algorithm design. An algorithm that is mathematically "superior" due to faster asymptotic convergence might be outperformed by a "simpler" algorithm if the process is halted early.

Conceptual trade-off between asymptotic convergence (Algorithm A) and finite-horizon performance (Algorithm B).

By utilizing Chebyshev acceleration, the algorithm effectively "spreads" its optimization effort across the spectrum of the matrix's eigenvalues. This prevents the gradient descent from stalling and ensures the best possible performance within the allocated "finite horizon" of iterations.

Experimental Results and Scalability

The researchers evaluated their approach using Monte Carlo simulations based on 3GPP channel models, including a single-cell setup with 2048 transmit antennas serving multiple users.

CPU Time Efficiency: The proposed method demonstrates a clear advantage in clock-time performance. In a 2048-antenna scenario, the proposed algorithm reached a sum-rate of 100 bps/Hz in just 1.6 seconds, compared to 2.6 seconds for the standard WMMSE.

Sum of data rates versus elapsed time, illustrating that the proposed algorithm achieves high performance significantly faster than conventional methods.

Scalability: The performance gap widens as the number of antennas increases. While WMMSE is competitive at $M=512$, the $O(M^3)$ cost becomes untenable at $M=2048$. The Finite Horizon method, with its $O(M^2)$ complexity, scales much more gracefully, providing a higher performance-per-second of computation.

Limitations and Future Directions

While promising, the approach has areas for further refinement:

  • Eigenvalue Estimation: Optimal step sizes depend on knowing the system matrix's eigenvalue range. Efficiently estimating these values without significant overhead remains an open challenge.
  • Dynamic Power Control: Current multi-cell evaluations assume fixed power. Integrating dynamic power allocation into the finite horizon framework is a logical next step.
  • Hardware Acceleration: Future research could explore mapping these matrix-vector operations to specialized hardware like GPUs or FPGAs, where the avoidance of matrix inversions would yield even greater speedups.

By prioritizing immediate performance over theoretical limits, Finite Horizon Optimization offers a practical and powerful path forward for the next generation of large-scale wireless systems.

Summary updated

BibTeX
@misc{feng2025finite,
  author = {Yi Feng and Kaiming Shen},
  title = {{Finite Horizon Optimization for Large-Scale MIMO}},
  year = {2025},
  eprint = {2503.11356},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2503.11356},
}
Figure