Abstract: Harmonic retrieval (HR) has a wide range of applications in the scenes where signals are modelled as a summation of sinusoids. Past works have developed a number of approaches to recover the original signals. Most of them rely on classical singular value decomposition, which are vulnerable to unexpected outliers. In this paper, we present new decomposition algorithms of third-order complex-valued tensors with $L_1$-principle component analysis ($L_1$-PCA) of complex data and apply them to a novel random access HR model in presence of outliers. We also develop a novel subcarrier recovery method for the proposed model. Simulations are designed to compare our proposed method with some existing tensor-based algorithms for HR. The results demonstrate the outlier-insensitivity of the proposed method.
Introduction
Harmonic Retrieval (HR) is a fundamental task in signal processing, focused on identifying the frequencies and amplitudes of sinusoids within a noisy signal. Traditionally, this is achieved by organizing signals into a multi-dimensional array, or tensor, and applying Tucker decomposition (a higher-order version of PCA). However, standard Tucker decomposition relies on the Frobenius ($L_2$) norm, which minimizes squared errors. This makes the model extremely sensitive to "bursty" outliers—sudden, high-intensity interference that can pull the entire decomposition away from the true signal.
Proposed Methodology
To address this vulnerability, the authors propose a two-pronged solution to make HR robust by leveraging the $L_1$-norm, which is mathematically less sensitive to extreme values than the $L_2$-norm.
$L_1$-Tucker Decomposition
The paper introduces two new algorithms: $L_1$-TOTD (Third-Order Tucker Decomposition) and $L_1$-TOOI (Third-Order Orthogonal Iteration). The core of these methods is a robust $L_1$-PCA routine designed for complex-valued data.
As shown in Algorithm 1, the $L_1$-PCA routine iteratively optimizes the solution to minimize the impact of outliers. This routine is then utilized within the $L_1$-TOTD framework (Algorithm 2) to decompose third-order tensors by processing their mode-$i$ unfoldings.
Subcarrier Sorting Method (SCSM)
In addition to the decomposition, the authors propose the Subcarrier Sorting Method (SCSM) for frequency recovery. While traditional methods like ESPRIT rely on shift-invariance—a property easily disrupted by outliers—SCSM identifies active subcarriers by projecting the signal onto a dictionary of possible frequencies. This geometric approach proves significantly more stable than solving the matrix equations required by ESPRIT.
Experimental Evaluation
The proposed methods were evaluated using synthetic simulations modeled after real-world random access communication scenarios. The signal model included $K_a=6$ active transmitters among $K-1=49$ available subcarriers using QPSK modulation. To simulate bursty interference, the researchers added sparse, high-variance Gaussian outliers ($\sigma_o^2 = 20$) affecting 5% to 10% of the signal samples.
Key Results and Insights
- Outlier Robustness: In tests with 10% outlier density, the $L_1$-based methods outperformed standard HOSVD and HOOI by several decibels (dB) in terms of Root Mean Square Error (RMSE).
- Subspace Stability: The $L_1$ approach effectively "ignores" outliers that do not fit the low-rank structure of the main signal, maintaining a stable signal subspace even when the tensor decomposition is not perfect.
- SCSM vs. ESPRIT: SCSM consistently yielded more accurate frequency estimations than ESPRIT, particularly in low Signal-to-Noise Ratio (SNR) environments where outliers are most disruptive.
- Superiority over IR-HOSVD: The proposed method also outperformed other existing robust methods like iteratively reweighted HOSVD (IR-HOSVD), especially as the SNR increased.
Conclusion and Future Work
By replacing standard $L_2$-norm tensor decompositions with $L_1$-norm versions and introducing the SCSM, the researchers achieved signal recovery that is significantly more resilient to outliers. While $L_1$-PCA is more computationally intensive than standard SVD, the gains in robustness are substantial. Future work will focus on optimizing the computational cost for very large tensors and validating the algorithm on real-world radio frequency (RF) or biomedical data.
Summary updated