A novel and efficient parameter estimation of the Lognormal-Rician turbulence model based on k-Nearest Neighbor and data generation method

Maoke Miao, Xinyu Zhang, Bo Liu, Rui Yin, Jiantao Yuan, Feng Gao, Xiao-Yu ChenarXiv 2024 · DOI · arXiv

Abstract

Atmospheric turbulence significantly impacts the performance of free-space optical (FSO) communications. While the Lognormal-Rician (LR) model accurately describes these effects, its complexity makes parameter estimation computationally intensive. This paper proposes a novel estimator combining $k$-Nearest Neighbor (kNN) density estimation with synthetic data generation and Genetic Algorithms (GA). This approach bypasses the need for complex integrals and Bessel functions, achieving high accuracy with significantly fewer data samples than traditional methods.

Introduction

Free-space optical and quantum communications offer high-speed data transmission but are susceptible to atmospheric turbulence, which causes rapid light intensity fluctuations known as scintillation. The Lognormal-Rician (LR) model is widely used to characterize these fluctuations. However, the LR probability density function (PDF) involves intricate integrals and Bessel functions, making standard Maximum Likelihood Estimation (MLE) difficult. Existing alternatives, such as the Generalized Method of Moments (GMM), often require up to $10^6$ samples to achieve stability, leading to high latency. This research introduces a "middle ground" that balances computational speed and data efficiency.

Proposed Methodology

The proposed method avoids the mathematical complexity of the LR model by using a non-parametric approach:

  1. kNN Density Estimation: The $k$-Nearest Neighbor algorithm approximates the density of received signal samples, replacing the theoretical PDF.
  2. Data Generation: Synthetic data is generated based on candidate parameters ($r, \sigma_z^2$) to compare against the observed signal.
  3. Log-Likelihood Function (LLF): An objective function is constructed to measure the similarity between synthetic and received data.
  4. Genetic Algorithm (GA) Optimization: Since the LLF surface can be non-convex and noisy with limited samples, a GA is employed to find the global maximum, ensuring robust parameter estimation.

Experimental Setup

The algorithm's performance was validated using simulated LR samples generated in MATLAB. The study compared the proposed kNN+GA method against benchmarks like the Hansen GMM, Saddlepoint Approximation (SAP), and Expectation-Maximization (EM) methods across sample sizes ranging from $10^3$ to $10^6$.

Key Insights: The Optimal $k$ and Optimization Landscape

A critical factor in kNN estimation is the choice of $k$, which governs the bias-variance tradeoff. Through Kolmogorov-Smirnov (KS) testing, the researchers identified that for sample sizes between $10^3$ and $10^4$, the optimal $k$ typically falls between 8 and 15.

Optimal k selection and KS test results for different sample sizes.

Furthermore, while increasing the number of synthetic samples ($L$) can smooth the optimization landscape, the use of a Genetic Algorithm allows for accurate estimation even on "rough" surfaces. This capability enables the method to operate effectively with as few as $10^3$ samples, drastically reducing the data collection time compared to the $10^6$ samples required by GMM.

Results and Discussion

The kNN+GA estimator demonstrates Mean Square Error (MSE) performance comparable to the computationally intensive EM and SAP estimators. Notably, for the variance parameter $\sigma_z^2 \ge 0.6$, the proposed method outperforms the SAP estimator.

MSE performance comparison for the variance parameter $\sigma_z^2$ and coherence parameter $r$ across different estimation methods.

By eliminating the need for Bessel functions and complex integrals, this method is highly suitable for implementation on digital signal processing (DSP) hardware, offering a practical solution for real-time FSO link adaptation. However, the estimation for the coherence parameter $r$ is still approximately twice as high in MSE compared to the EM estimator, representing a trade-off for the significantly reduced computational demand.

Conclusion

While the proposed method offers significant gains in efficiency and variance estimation, the coherence parameter $r$ estimation remains a point for further refinement. Future research will focus on improving $r$ accuracy and validating the algorithm with real-world atmospheric data. The framework's flexibility suggests potential applications in other complex fading models within wireless and optical communications.

Summary updated

BibTeX
@misc{miao2024novel,
  author = {Maoke Miao and Xinyu Zhang and Bo Liu and Rui Yin and Jiantao Yuan and Feng Gao and Xiao-Yu Chen},
  title = {{A novel and efficient parameter estimation of the Lognormal-Rician turbulence model based on k-Nearest Neighbor and data generation method}},
  year = {2024},
  eprint = {2409.01694},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2409.01694},
}
Figure