Deterministic Sort-Free Candidate Pruning for Scalable MIMO Box Decoding

Box Decoding is a sort-free tree-search MIMO detector whose complexity is independent of the QAM order, achieved by searching a fixed candidate box around a zero-forcing (ZF) estimate. However, without pruning, the number of visited nodes grows exponentially with the MIMO dimension, limiting scalability. This work proposes two deterministic, low-complexity, sort-free pruning strategies to control node growth. By exploiting the geometric symmetry of the QAM grid and the relative displacement between the ZF estimate and nearby constellation points, the proposed methods eliminate unnecessary metric evaluations while preserving QAM-order independence. The resulting detector achieves substantial complexity reduction with negligible error-rate degradation and enables fully parallel, hardware-efficient implementations for large-scale MIMO and higher-order QAM systems.

💡 Research Summary

The paper addresses a fundamental scalability issue in Box Decoding, a sort‑free tree‑search MIMO detector whose complexity is independent of the QAM order because it searches a fixed‑size candidate “box” around a zero‑forcing (ZF) estimate. While this property eliminates the need for constellation‑wise enumeration, the number of visited nodes still grows exponentially with the number of transmit/receive antennas (the MIMO dimension), making the original algorithm impractical for large‑scale systems.
To overcome this limitation, the authors propose two deterministic, low‑complexity, sort‑free pruning strategies that exploit the geometric symmetry of square QAM constellations and the relative displacement between the ZF estimate and nearby constellation points. The first strategy, Single‑Step Candidate Pruning (SCP), applies a simple “local PED minimum” rule (Rule 1) within each box cluster. By comparing only the real and imaginary offsets (Δq and δ) of candidates, SCP can select the single child node with the smallest partial Euclidean distance increment without performing any full Euclidean distance calculations or sorting. This reduces the per‑layer candidate count from B to 1, dramatically cutting the node expansion workload. However, because SCP makes decisions independently for each cluster, it can skip globally optimal child nodes—a problem the authors call the “optimal child‑node skip.”
The second strategy, Iterative Candidate Pruning (ICP), resolves this issue by treating each box cluster as a pre‑sorted source list and applying a multi‑way merge (m‑Merge) algorithm. Before merging, candidates inside each cluster are ordered using a second set of comparison rules (Rule 2) that also rely only on simple offset arithmetic. The m‑Merge repeatedly extracts the global minimum among the local minima of all clusters and replaces it with the next best element from the same cluster, continuing until K child nodes are selected. This approach guarantees that the K most promising nodes across all clusters are retained, while still avoiding exhaustive distance calculations and sorting.
Recognizing that applying ICP at every layer may introduce unnecessary iterations, the authors introduce a hybrid scheme called SCP‑ICP (SICP). In SICP, ICP is applied only in the first t layers (typically t = 1), where pruning has the greatest impact on overall complexity, while SCP is used in the remaining layers. Simulation results show that even with t = 1, the performance gap to the optimal K‑Best detector is negligible, and increasing t beyond 1 yields diminishing returns.
The paper provides a detailed complexity analysis in terms of real multipliers (RMUL) and real adders (RADD). Key observations include:

• Candidate list expansion (CLE) for box‑based detectors requires only 4 RMUL and √B RADD per box, independent of the QAM constellation size A.
• SCP adds no extra pruning overhead; ICP adds only (L − 1) RADD per merge, where L is the number of clusters.
• Compared with K‑Box, DKB, and conventional K‑Best, the proposed methods achieve orders‑of‑magnitude reductions in both RMUL and RADD counts, especially as N and A increase.
  Simulation results for 4 × 4 and 8 × 8 MIMO systems with 64‑QAM confirm the analytical findings. All box‑based variants provide roughly a 5 dB SNR gain over linear ZF/LMMSE detectors and closely track the BER of K‑Best (K = B = 4). Box‑SCP suffers a modest ~1.3 dB loss, while Box‑ICP shows virtually no degradation. The hybrid Box‑SICP 1 improves performance by about 0.7 dB relative to Box‑SCP and is essentially indistinguishable from K‑Best; further increasing t yields negligible additional benefit.
  In summary, the authors deliver deterministic, sort‑free pruning mechanisms that retain the QAM‑order independence of Box Decoding while dramatically curbing exponential node growth. The resulting detector is highly amenable to fully parallel, hardware‑efficient implementations, making it a strong candidate for large‑scale massive‑MIMO and high‑order QAM deployments in next‑generation wireless systems.