Abstract
Accurate estimation of sparse signals in noisy, unpredictable environments is a fundamental challenge in wireless communications. Traditional adaptive filtering methods based on the Minimum Mean Square Error (MMSE) criterion, such as the Least Mean Squares (LMS) algorithm, perform poorly in the presence of non-Gaussian impulsive noise. This paper introduces the Correntropy Induced Metric penalized Maximum Correntropy Criterion (CIMMCC) algorithm. By combining the Maximum Correntropy Criterion (MCC) for outlier robustness and the Correntropy Induced Metric (CIM) for effective sparsity exploitation, CIMMCC offers a robust, low-complexity solution. We provide theoretical convergence analysis and demonstrate through simulations that CIMMCC outperforms standard sparse adaptive filters in various impulsive noise scenarios.
Introduction
In broadband wireless communications, channel impulse responses are often sparse, characterized by a few significant coefficients among many near-zero values. While sparse adaptive filters are designed to exploit this structure, their performance is heavily dependent on the noise model. The industry-standard MMSE criterion assumes Gaussian noise; however, real-world environments frequently encounter impulsive noise—short-duration, high-power bursts. Because the MSE criterion squares the error, these outliers significantly distort the estimation process. There is a critical need for algorithms that are both sparsity-aware and robust to non-Gaussian interference.
Methodology
The proposed Correntropy Induced Metric penalized Maximum Correntropy Criterion (CIMMCC) algorithm addresses these challenges through two primary mechanisms:
1. Maximum Correntropy Criterion (MCC)
Instead of minimizing the squared error, CIMMCC maximizes correntropy, a local similarity measure between two random variables. Correntropy employs a Gaussian kernel; as the difference between the desired signal and the filter output increases (as in the case of an impulsive noise spike), the kernel value approaches zero. This effectively "ignores" outliers, providing inherent robustness that MSE lacks.
2. Correntropy Induced Metric (CIM) Penalty
To exploit signal sparsity, a penalty term is added to the cost function. While the $\ell_1$-norm is a common choice, it can lead to biased estimates by shrinking all coefficients. CIMMCC utilizes the CIM, which serves as a smooth, continuous approximation of the $\ell_0$-norm. This provides a stronger zero-attraction for insignificant coefficients without heavily biasing the larger, significant ones.
The iterative update rule for the filter weights is summarized in the following pseudocode:
Theoretical analysis establishes a conservative upper bound for the step size ($\mu$), ensuring the algorithm's stability despite its non-linear nature.
Performance Evaluation
The CIMMCC algorithm was evaluated against several benchmarks, including standard MCC and sparse variants like ZAMCC and RZAMCC, across multiple scenarios:
- Synthetic Noise Models: Testing included Mixed Gaussian noise and Alpha-stable noise to simulate various impulsive environments.
- Sparse Echo Cancellation: A high-dimensional task ($M=1024$) with only 52 non-zero coefficients was used to test sparsity exploitation.
- Real-World Data: A 2-second fragment of real human speech served as the input signal to verify performance in non-stationary conditions.
Key Results
- Robustness to Impulsive Noise: In Alpha-stable noise environments where traditional algorithms like ZALMS failed, CIMMCC maintained stability and achieved superior Mean Square Deviation (MSD).
- Convergence Speed: CIMMCC demonstrated faster convergence to steady-state accuracy compared to standard MCC in mixed Gaussian noise.
- High-Dimensional Performance: In the 1024-tap echo cancellation test, CIMMCC achieved an MSD of approximately -32 dB, outperforming competitors by several decibels.
- Computational Efficiency: The algorithm is highly efficient, requiring only $3M$ additions and $2M$ multiplications per iteration, making it suitable for real-time implementation.
Conclusion
The CIMMCC algorithm provides a powerful framework for sparse adaptive filtering in non-Gaussian environments. By integrating MCC and CIM, it achieves a balance between outlier robustness and sparsity exploitation. Future work will focus on automated tuning of the kernel width ($\sigma$) and extending the framework to non-linear system identification.
Summary updated