Optimization of MIMO Device-to-Device Networks via Matrix Fractional Programming: A Minorization-Maximization Approach

Kaiming Shen, Wei Yu, Licheng Zhao, Daniel P. Palomar

TL;DR

This paper introduces FPLinQ, a high-performance optimization framework for managing interference in MIMO Device-to-Device (D2D) networks. By developing a novel Matrix Fractional Programming (FP) technique and linking it to the Minorization-Maximization (MM) framework, the authors provide a mathematically robust way to jointly optimize link scheduling, power control, and beamforming without the need for manual parameter tuning.

What Problem Does This Paper Address?

In Device-to-Device (D2D) networks, devices communicate directly rather than through a base station. While this increases spectral efficiency, it creates a massive interference headache because many nearby pairs reuse the same frequencies.

Managing this requires solving three interconnected problems:

  1. Scheduling: Deciding which links should be active (ON/OFF).
  2. Power Control: Deciding how much power each active transmitter should use.
  3. Beamforming: In MIMO systems (multiple antennas), deciding the spatial "shape" of the signal to maximize the desired signal and minimize interference.

This joint problem is NP-hard and non-convex. Existing heuristics like FlashLinQ or ITLinQ often rely on "magic numbers" (thresholds) that must be manually tuned for different environments and are typically limited to single-antenna scenarios.

Methods

The authors propose a strategy called FPLinQ, built on two major pillars:

  • Matrix Fractional Programming (FP): Traditional FP handles ratios of scalars or vectors. This paper generalizes the theory to Matrix FP, allowing it to handle multiple data streams per link—a necessity for modern MIMO systems. They introduce a "Matrix Quadratic Transform" and a "Matrix Lagrangian Dual Transform" to decouple the signal and interference terms in the Shannon capacity formula.
  • Minorization-Maximization (MM) Interpretation: A key theoretical contribution is proving that these FP transforms are actually a way of constructing surrogate functions. In the MM framework, you optimize a simpler "surrogate" that sits below the complex objective function. This connection provides a formal proof of convergence for the algorithm.

The FPLinQ Algorithm: The algorithm iterates through three main steps:

  1. Update Auxiliary Variables: Closed-form updates for variables representing the "state" of the interference.
  2. Beamforming Update: Optimizing the transmit vectors (resembling an MMSE beamformer structure).
  3. Scheduling via Bipartite Matching: The discrete ON/OFF decision is transformed into a weighted bipartite matching problem, which can be solved optimally in polynomial time using the Hungarian or Auction algorithms.

Datasets

The paper evaluates the algorithm using numerical simulations rather than a static dataset:

  • Environment: A 1km × 1km square area with randomly dropped D2D links.
  • Channel Model: ITU-1411 short-range outdoor model at 2.4 GHz.
  • Benchmarks: Compared against FlashLinQ, ITLinQ, ITLinQ+, and Block Coordinate Descent (BCD).
  • Configurations: Tested various antenna setups, from $1 \times 1$ (SISO) up to $8 \times 8$ MIMO.

Key Insights and Ideas

  • Alleviating "Premature Turn-off": Standard optimization methods (like BCD) often turn a link OFF early in the process and can never turn it back ON because the "gradient" disappears. Because FPLinQ uses a surrogate function that mimics the original objective more gracefully, it allows links to "wake up" later in the optimization if the interference environment improves.
  • No Tuning Required: Unlike previous state-of-the-art methods that require a threshold ($\theta$) to decide if a link is "too noisy," FPLinQ is entirely optimization-driven. It adapts to the network geometry automatically.
  • Flexible Association: Most D2D models assume fixed pairs. FPLinQ allows for "Flexible Association," where a transmitter can choose the best receiver from a set of candidates, further boosting network utility.

Key Results

  • Throughput Gains: In high-density scenarios (900 links), FPLinQ achieved nearly double the sum-rate of FlashLinQ and significantly outperformed ITLinQ+.
  • Fairness: When optimizing for log-utility (proportional fairness), FPLinQ provided a much better balance, significantly improving the data rates for the "bottom 50%" of users compared to BCD.
  • MIMO Scaling: The Matrix FP approach showed massive gains as the number of antennas increased. For an $8 \times 8$ system, FPLinQ outperformed simpler "Vector FP" approaches by a wide margin because it could effectively manage multiple data streams per link.
  • Convergence: The algorithm is efficient, with most of the sum-rate gains occurring within the first 10 iterations.

Limitations and Future Work

  • Computational Complexity: The complexity is $O(N^4 L^2)$, where $N$ is the number of antennas and $L$ is the number of links. While manageable for metropolitan-scale networks, the $N^4$ term makes it computationally heavy for "Massive MIMO" arrays with hundreds of antennas.
  • Centralization: The current algorithm assumes a centralized controller with full channel state information (CSI). While the authors mention that FP can be decentralized, a fully distributed, low-overhead version of FPLinQ remains a challenge for future implementation in ad-hoc networks.

Summary updated

Abstract

Interference management is a fundamental issue in device-to-device (D2D) communications whenever the transmitter-and-receiver pairs are located in close proximity and frequencies are fully reused, so active links may severely interfere with each other. This paper devises an optimization strategy named FPLinQ to coordinate the link scheduling decisions among the interfering links, along with power control and beamforming. The key enabler is a novel optimization method called matrix fractional programming (FP) that generalizes previous scalar and vector forms of FP in allowing multiple data streams per link. From a theoretical perspective, this paper provides a deeper understanding of FP by showing a connection to the minorization-maximization (MM) algorithm. From an application perspective, this paper shows that as compared to the existing methods for coordinating scheduling in the D2D network, such as FlashLinQ, ITLinQ, and ITLinQ+, the proposed FPLinQ approach is more general in allowing multiple antennas at both the transmitters and the receivers, and further in allowing arbitrary and multiple possible associations between the devices via matching. Numerical results show that FPLinQ significantly outperforms the previous state-of-the-art in a typical D2D communication environment.
BibTeX
@article{shen2019optimization,
  author = {Kaiming Shen and Wei Yu and Licheng Zhao and Daniel P. Palomar},
  title = {{Optimization of MIMO Device-to-Device Networks via Matrix Fractional Programming: A Minorization-Maximization Approach}},
  journal = {IEEE/ACM Trans. Networking},
  year = {2019},
  url = {https://openreview.net/forum?id=WpaTVrciFLn},
}

Loading related content...

Figure