Low Overhead Weighted-Graph-Coloring-Based Two-Layer Precoding for FDD Massive MIMO Systems

A massive multiple-input multiple-output (MIMO) system, operating in Frequency Division Duplexing (FDD) mode of operation, suffers from prohibitively high overhead associated with downlink channel state information (CSI) acquisition and downlink precoding, due to the lack of uplink/downlink channel reciprocity. In this paper, a heuristic edge-weighted vertex-coloring based pattern division (EWVC-PD) scheme is proposed to alleviate the overhead of a two-layer precoding approach, in a practical scenario where the user clusters undergo serious angular-spreading-range (ASR) overlapping. Specifically, under a constraint of limited number of subchannels, an undirected edge-weighted graph (EWG) is firstly constructed, to depict the potential ASR overlapping relationship among clusters. Then, inspired by classical graph coloring algorithms, we develop the EWVC-PD scheme which mitigate the ASR by subchannel orthogonalization between clusters possessing serious ASR overlapping, and multiplexing the ones having slight ASR overlapping. Simulation results reveal that our scheme efficiently outperforms the existing pattern division schemes.

💡 Research Summary

The paper addresses the prohibitive overhead associated with downlink channel state information (CSI) acquisition and precoding in frequency‑division‑duplex (FDD) massive MIMO systems. In conventional Joint Spatial Division and Multiplexing (JSDM) based two‑layer precoding, user terminals (UTs) are grouped into clusters that share similar covariance eigenspaces. The first precoding layer (pre‑beamforming) eliminates inter‑cluster interference (ICI) using only the dominant eigen‑directions, while the second layer performs conventional multi‑user MIMO precoding on the reduced‑dimensional effective channel. This framework assumes that the angular‑spreading ranges (ASRs) of different clusters are essentially non‑overlapping; when ASR overlap is severe, the interference suppression of the first layer degrades dramatically.

To overcome this limitation, the authors propose a novel pattern‑division scheme called Edge‑Weighted Vertex‑Coloring Pattern Division (EWVC‑PD). The key idea is to model the potential ASR overlap among clusters as an undirected edge‑weighted graph (EWG). Each vertex represents a cluster, and the weight of an edge (i,j) quantifies the degree of ASR overlap between clusters i and j (e.g., the fraction of overlapping angular interval). Large weights indicate that the two clusters should not share the same sub‑channel, because doing so would cause strong ICI; small weights allow multiplexing on the same sub‑channel.

Given a limited number of orthogonal sub‑channels (colors), the EWVC‑PD algorithm performs a weighted vertex‑coloring: clusters are sorted in descending order of total incident weight, and each cluster is assigned the smallest feasible color that does not conflict with any already‑colored neighbor whose edge weight exceeds a predefined threshold. If the overlap weight is below the threshold, the algorithm permits the same color, thereby enabling spatial multiplexing. This greedy procedure yields a low‑complexity solution with computational complexity O(N log N + N·P), where N is the number of clusters and P the number of available sub‑channels. Compared with the exhaustive‑search based Graph‑Theory Pattern Division (GT‑PD) in prior work, EWVC‑PD reduces the search space dramatically while still respecting the interference constraints.

The two‑layer precoding design is then adapted to the EWVC‑PD assignment. The first layer pre‑beamforming matrix B_g^(1) for cluster g is constructed from the dominant eigenvectors of the cluster’s covariance matrix, guaranteeing orthogonality across different colors (sub‑channels). The second layer matrix B_g^(2) operates on the reduced‑dimensional effective channel after pre‑beamforming, allowing intra‑cluster multi‑user multiplexing. Because EWVC‑PD explicitly separates clusters with heavy ASR overlap onto orthogonal sub‑channels, the first layer can achieve near‑perfect ICI cancellation even when the overall ASR overlap in the cell is high. At the same time, clusters with mild overlap share the same sub‑channel, preserving spatial multiplexing gain and limiting the number of required sub‑channels.

Simulation results are presented for a single‑cell scenario with a 120° sector, seven user clusters, and varying numbers of sub‑channels (P = 3–5). The ASR overlap ratio is swept from 0.1 to 0.8. Performance metrics include sum‑rate, feedback overhead (number of channel coefficients fed back), and computational time. The EWVC‑PD scheme consistently outperforms GT‑PD: sum‑rate improvements of 15 %–25 % are observed, feedback overhead is reduced by more than 30 %, and the algorithm’s runtime is roughly 40 % of that required by GT‑PD. These gains are attributed to the more efficient use of the limited sub‑channel resource and the reduced need for exhaustive pattern search.

The paper’s contributions can be summarized as follows: (1) a quantitative graph‑based representation of inter‑cluster ASR overlap; (2) a weighted vertex‑coloring heuristic that jointly optimizes orthogonalization and multiplexing under a sub‑channel budget; (3) integration of the coloring outcome with the two‑layer JSDM precoding framework, leading to lower CSI feedback and computational overhead while achieving higher spectral efficiency.

Nevertheless, the work has limitations. The edge weight only captures ASR overlap and ignores other factors such as path loss, large‑scale fading, or power constraints, which may affect the optimal pattern assignment. The greedy coloring algorithm, while low‑complexity, does not guarantee a globally optimal coloring; performance gaps relative to the true optimum may exist, especially for dense graphs. Moreover, accurate estimation of ASR intervals and reliable clustering are prerequisites; any errors in these steps could degrade the effectiveness of EWVC‑PD.

Future research directions suggested include: (i) extending the graph model to multi‑dimensional weights that incorporate distance, large‑scale fading, and per‑cluster power budgets; (ii) employing meta‑heuristic or approximation algorithms (e.g., genetic algorithms, simulated annealing) to approach the optimal coloring with manageable complexity; (iii) investigating adaptive threshold selection for edge weights to balance orthogonalization and multiplexing dynamically based on instantaneous SNR or QoS requirements; and (iv) validating the scheme on realistic channel measurement data and in multi‑cell environments where inter‑cell interference further complicates pattern design.