A Block Term Decomposition Model Based Algorithm for Tensor Completion of Multidimensional Harmonic Signals

Lei Wang, Xiao‐Feng Gong, X. Liu, Wei Feng, Qiu‐Hua LinarXiv 2025 · DOI · arXiv

Abstract

We consider tensor data completion of an incomplete observation of multidimensional harmonic (MH) signals. Unlike existing tensor-based techniques for MH retrieval (MHR), which mostly adopt the canonical polyadic decomposition (CPD) to model the simple "one-to-one" correspondence among harmonics across difference modes, we herein use the more flexible block term decomposition (BTD) model that can be used to describe the complex mutual correspondences among several groups of harmonics across different modes. An optimization principle that aims to fit the BTD model in the least squares sense, subject to rank minimization of hankelized MH components, is set up for the tensor completion task, and an algorithm based on alternating direction method of multipliers is proposed, of which the effectiveness and applicability are validated through both numerical simulations and an application in Sub-6GHz channel state information (CSI) completion.

Introduction

Multidimensional Harmonic Retrieval (MHR) is a cornerstone of signal processing, used in everything from radar and spectroscopy to wireless channel estimation. Often, researchers must deal with incomplete data due to hardware limitations or sparse sampling. Existing tensor-based completion methods typically rely on Canonical Polyadic Decomposition (CPD), which assumes that each harmonic in one dimension corresponds to exactly one harmonic in another. However, real-world signals—especially wireless Channel State Information (CSI)—are more complex. In a wireless environment, a single "cluster" in the frequency domain might be associated with multiple "paths" in the space-time domain, creating a "one-to-many" correspondence that CPD cannot model effectively.

Methods

The authors propose a framework based on the MLR-($L_r, L_r, 1$) BTD model. This approach shifts from the rigid 1:1 structure of CPD to a flexible BTD structure that captures the "cluster-path" dynamics of physical signals. To ensure the recovered signal remains harmonic, the authors utilize Hankelization, exploiting the property that a hankelized version of a harmonic vector is rank-1.

The resulting optimization problem minimizes the rank of these hankelized components (relaxed to the nuclear norm for convexity) and is solved using the Alternating Direction Method of Multipliers (ADMM). Two specific updating methods were evaluated within this framework: Alternating Least Squares (ALS) and Gauss-Newton Nonlinear Least Squares (GN-NLS).

Key Results

The researchers evaluated their algorithm using numerical simulations and realistic CSI data generated via the QuaDRiGa v2.0.0 simulator. The proposed BTD-based algorithms (MHR-TC-BTD-ALS/NLS) consistently achieved lower Relative Least Normalized Error (RLNE) compared to CPD-based methods and general tensor completion techniques.

RLNE versus SNR (dB)

As shown in Figure 1, the proposed methods maintained high performance even at low SNR levels (e.g., -5dB to 5dB), where other methods failed to recover the signal structure.

RLNE versus proportion of sampled entries (%)

Furthermore, Figure 2 demonstrates that the BTD model can recover the full tensor with significantly fewer observations than its competitors. Interestingly, the simpler ALS-based update actually outperformed the more complex GN-NLS approach when data was extremely sparse (sampling rates below 8%), suggesting that ALS is more robust in "low-information" regimes.

Conclusion

The shift from CPD to BTD aligns the mathematical model more closely with the underlying physics of wireless propagation. By enforcing harmonic structure through the rank of Hankel matrices, the proposed method provides a powerful way to regularize the completion process. Future work will address automated strategies for selecting weighting parameters and reducing the computational complexity of handling large-scale Hankel matrices for real-time applications.

Summary updated

Abstract

We consider tensor data completion of an incomplete observation of multidimensional harmonic (MH) signals. Unlike existing tensor-based techniques for MH retrieval (MHR), which mostly adopt the canonical polyadic decomposition (CPD) to model the simple "one-to-one" correspondence among harmonics across difference modes, we herein use the more flexible block term decomposition (BTD) model that can be used to describe the complex mutual correspondences among several groups of harmonics across different modes. An optimization principle that aims to fit the BTD model in the least squares sense, subject to rank minimization of hankelized MH components, is set up for the tensor completion task, and an algorithm based on alternating direction method of multipliers is proposed, of which the effectiveness and applicability are validated through both numerical simulations and an application in Sub-6GHz channel state information (CSI) completion.
BibTeX
@misc{wang2025block,
  author = {Lei Wang and Xiao‐Feng Gong and X. Liu and Wei Feng and Qiu‐Hua Lin},
  title = {{A Block Term Decomposition Model Based Algorithm for Tensor Completion of Multidimensional Harmonic Signals}},
  year = {2025},
  eprint = {2501.15166},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2501.15166},
  howpublished = {arXiv},
}
Figure