Abstract
The integration of sensing and communication (ISAC) within massive multiple-input multiple-output (MIMO) systems is a cornerstone for 6G wireless networks. While promising significant gains in spectral and energy efficiency, the deployment of large antenna arrays introduces substantial challenges in acquiring accurate channel state information (CSI) and target parameter information (TPI). This paper presents a unified tensor-based framework that addresses both channel estimation and target sensing. By parameterizing the communication channel through physical attributes—angular, delay, and Doppler dimensions—we formulate the estimation problem as a canonical polyadic decomposition (CPD). We further leverage the inherent Vandermonde structure of antenna arrays to establish uniqueness conditions and resolve a higher number of targets. Our approach also introduces a segment-based training pattern to mitigate "beam squint" in wideband scenarios. Numerical results demonstrate that the proposed method approaches the Cramer-Rao Bound (CRB) and outperforms traditional matrix-based estimation techniques.
1. Introduction
As we transition toward 6G, the convergence of communication and radar sensing into a single Integrated Sensing and Communication (ISAC) system has emerged as a key paradigm. Massive MIMO is essential to this vision, providing the spatial degrees of freedom necessary for high-resolution sensing and high-capacity communication. However, the "dimensionality curse" associated with massive antenna arrays often leads to prohibitive training overhead. Furthermore, in wideband systems, the frequency-dependent nature of spatial steering vectors—known as "beam squint"—complicates the estimation of physical parameters.
This research proposes a unified mathematical framework that treats communication channel estimation and radar target sensing as a single problem. By modeling received signals as multi-dimensional tensors, we can exploit the inherent structure of the data to achieve high-precision parameter estimation with reduced overhead.
2. A Unified Tensor Framework
The core innovation of this work lies in the physical parameterization of the communication channel. Rather than treating the channel as an unstructured matrix, we model it using the same parameters used in radar sensing: Angle of Arrival (AoA), Angle of Departure (AoD), time delay, and Doppler shift.
2.1 Tensor Modeling and CPD
We formulate the received signals as a third-order tensor across the dimensions of Space, Time, and Frequency. This multi-dimensional representation preserves the correlations between different signal components that are often lost in traditional matrix "flattening." We apply Canonical Polyadic Decomposition (CPD) to this tensor, which allows us to factorize the signal into its constituent physical parameters.
2.2 Exploiting Vandermonde Structure
By utilizing Uniform Linear Arrays (ULAs), the steering vectors in our model exhibit a Vandermonde structure. This mathematical property is crucial; it allows us to prove a relaxed uniqueness condition for the tensor decomposition. Consequently, the system can resolve more targets or multipath components than the number of physical antennas would typically allow in a standard matrix-based framework.
3. Addressing Wideband Challenges: Beam Squint
In wideband massive MIMO, the phase shift across the array varies with frequency, causing the beam direction to "squint." This phenomenon degrades the performance of traditional estimation algorithms. To counter this, we propose a segment-based training pattern. By repeating precoders over specific segments, we "freeze" the frequency-dependent steering vectors within the tensor framework. This design enables the decoupling of wideband effects and ensures robust estimation even at high bandwidths (e.g., 1 GHz).
4. Algorithmic Refinement
We developed a unified algorithm that iteratively refines the estimates of AoD and Doppler shifts. These parameters are often coupled in the received signal, making them difficult to distinguish. Our iterative approach decouples these dimensions, significantly improving the resolution and accuracy of the final parameter extraction.
5. Performance Evaluation
The proposed framework was evaluated through extensive numerical simulations at 28 GHz (mmWave). The setup included a Base Station with 64 transmit antennas and a User Equipment (UE) with 8 antennas.
5.1 Accuracy and Resolution
The results indicate that the Root Mean Square Error (RMSE) for angle, range, and velocity estimation closely approaches the Cramer-Rao Bound (CRB), the theoretical limit for estimation accuracy. Notably, the tensor-based approach maintains a near 100% success rate in target resolution even as the number of targets increases. In contrast, traditional methods like MUSIC (Multiple Signal Classification) fail once the number of targets exceeds the number of receive antennas.
5.2 Efficiency and Robustness
The unified approach achieves superior performance with significantly fewer training subcarriers ($K=16$) compared to benchmark algorithms requiring $K=32$. Furthermore, in 1 GHz wideband scenarios, our segment-based algorithm successfully prevented the "performance floor" that limited the accuracy of traditional methods due to beam squint.
6. Conclusion and Future Work
This paper demonstrates that a unified tensor-based approach can effectively bridge the gap between communication and sensing in massive MIMO systems. By leveraging the physical geometry of the signal environment, we achieve high-resolution parameter estimation with reduced overhead and robustness to wideband impairments.
Future research will focus on target classification—distinguishing between targets of interest and environmental clutter—and investigating the impact of hardware non-idealities, such as phase noise and power amplifier non-linearities, on the stability of the tensor decomposition.
Summary updated