Abstract
Frequency estimation is a fundamental problem in many areas. The well-known A&M and its variant estimators have established an estimation framework by iteratively interpolating the discrete Fourier transform (DFT) coefficients. In general, those estimators require two DFT interpolations per iteration, have uneven initial estimation performance against frequencies, and are incompetent for small sample numbers due to low-order approximations involved. Exploiting the iterative estimation framework of A&M, we unprecedentedly introduce the Padé approximation to frequency estimation, unveil some features about the updating function used for refining the estimation in each iteration, and develop a simple closed-form solution to solving the residual estimation error. Extensive simulation results are provided, validating the superiority of the new estimator over the state-the-art estimators in wide ranges of key parameters.
Introduction
Frequency estimation is a cornerstone of modern signal processing, essential for applications ranging from vehicular radar to 5G communications. A popular family of algorithms, derived from the Aboutanios and Mulgrew (A&M) framework, works by iteratively refining a frequency estimate using Discrete Fourier Transform (DFT) interpolations.
However, traditional A&M variants suffer from two primary limitations. First, they rely on first-order Taylor Series (linear) approximations, which assume a large number of signal samples ($N$). When $N$ is small (e.g., $N=8$ in small antenna arrays), these linear models break down. Second, the computational cost is high; achieving high accuracy typically requires numerous iterations, each involving multiple complex multiplications.
Methodology
This research introduces a novel iterative framework that replaces standard linear approximations with the Padé Approximation (PA). Unlike the Taylor Series, which uses a single polynomial, a Padé Approximation uses a ratio of two polynomials. This rational function provides a significantly tighter fit to the true signal behavior over a wider range of frequencies.
Key innovations in the proposed method include:
- Symmetry Exploitation: The authors proved the "updating function" for frequency refinement is an odd function. This allowed the elimination of even-powered terms, simplifying a complex 6th-order approximation into a manageable cubic equation.
- Smart Initialization: A "sign test" determines the initial frequency region, allowing the algorithm to start closer to the true value. This reduces the required DFT interpolations in the first step from two to one.
- Closed-Form Solution: The method utilizes a direct, closed-form solution to solve the cubic equation for residual error, avoiding computationally expensive search-based fits.
Experimental Results
The method was evaluated using synthetic simulations of complex single-tone exponential signals in Additive White Gaussian Noise (AWGN). Performance was benchmarked against A&M, Generalized A&M (GAM), and the Hybrid A&M and q-Shift Estimator (HAQSE) across $5 \times 10^4$ Monte-Carlo trials.
Performance in Small Sample Scenarios
As shown in the simulation results, the proposed Padé-based estimator demonstrates remarkable stability in "small $N$" scenarios. While benchmark methods like HAQSE show performance degradation as parameters shift, the proposed method remains robust.
Convergence and Efficiency
A critical finding is the speed of convergence. The proposed estimator reaches the Cramér-Rao Lower Bound (CRLB)—the theoretical limit of estimation accuracy—in just two iterations.
As illustrated in the normalized MSE plots, after only one iteration, the proposed estimator achieves an MSE just 1.079 times the CRLB for $N=8$, significantly outperforming benchmarks which typically hover around 1.14 times the CRLB. Because it requires fewer iterations to reach peak accuracy, the total computational complexity—dominated by $(2I - 1)N$ complex multiplications—is lower than that of simpler algorithms requiring more iterations.
Conclusion
By moving beyond linear approximations to the more robust Padé Approximation, this research provides a frequency estimator that is both more accurate and more efficient, particularly for small datasets. Future work will focus on extending this framework to multi-tone scenarios using Estimate-and-Subtract Strategies (ESS) and validating the cubic solver's latency on real-time hardware like FPGAs.
Summary updated