Abstract
This paper considers parameter estimation of superimposed exponential signals in multiplicative and additive noise which are all independent and identically distributed. A modified Newton-Raphson algorithm is used to estimate the frequencies of the considered model, which is further used to estimate other linear parameters. It is proved that the modified Newton-Raphson algorithm is robust and the corresponding estimators of frequencies attain the same convergence rate with Least Squares Estimators (LSEs) under the same noise conditions, but it outperforms LSEs in terms of the mean squared errors. Finally, the effectiveness of the algorithm is verified by some numerical experiments.
Introduction
In fields such as radar, sonar, and wireless communications, the fundamental task is to extract the parameters—frequencies, amplitudes, and phases—of a signal. While traditional models often assume corruption only by additive noise, real-world environments frequently introduce multiplicative noise. This type of noise acts as a random, rapid fluctuation in signal amplitude, characteristic of fading channels or Doppler spreading.
The Least Squares Estimator (LSE) is the standard for statistical efficiency in these scenarios. However, the LSE is computationally intensive, requiring the solution of complex nonlinear optimization problems and frequent matrix inversions. This paper introduces a modified Newton-Raphson algorithm that achieves LSE-level accuracy with significantly lower computational overhead.
Methodology
The proposed approach utilizes a Modified Newton-Raphson (NR) algorithm based on an Approximate Least Squares Estimator (ALSE) criterion. Key innovations include:
- Modified Step Factor: The algorithm employs a fixed step factor of 1/4. This specific value is mathematically proven to ensure convergence toward the true frequency value, preventing the algorithm from stalling in local minima.
- Varying Sample Size Technique: To overcome the limitations of initial estimates (e.g., from a periodogram), the authors use a "bootstrapping" method. The algorithm initially processes a subset of the data ($N_1 = N^{6/7}$) before moving to the full dataset. This allows the estimator to reach the optimal $O(N^{-3/2})$ convergence rate efficiently.
- Simultaneous Estimation: The algorithm estimates all frequencies in a superimposed signal simultaneously, avoiding the error propagation common in sequential estimation methods.
Experimental Evaluation
The algorithm was tested using synthetic numerical experiments to compare performance against known ground truths. Two primary models were used:
- Model 1: A signal with two well-separated frequencies.
- Model 2: A "dense" signal with closely spaced frequencies to test resolution limits.
Noise conditions were varied, testing different levels of additive noise ($\sigma_0 = 0.5, 1.0, 1.5$) and multiplicative noise variances to verify robustness.
Results and Discussion
The numerical simulations confirmed several key advantages of the modified NR algorithm:
- Theoretical Efficiency: Frequency estimators achieved the theoretical "speed limit" convergence rate of $O_p(N^{-3/2})$.
- Superior MSE: In many cases, the algorithm produced Mean Squared Errors (MSE) lower than the asymptotic variances of traditional LSEs.
- Computational Speed: Convergence typically occurred within 30 iterations. By avoiding $O(p^3)$ matrix inversions, the method is well-suited for real-time applications.
Conclusion
The modified Newton-Raphson algorithm provides a robust and efficient alternative to traditional LSE for signals corrupted by both multiplicative and additive noise. While it relies on a reasonably accurate initial guess from a periodogram, its performance in terms of convergence rate and MSE is exceptional. Future work may involve integrating higher-resolution initial estimators like MUSIC or ESPRIT to handle even denser frequency clusters and extending the model to non-identically distributed noise.
Summary updated