Introduction
In high-speed signal processing applications such as radar, the ability to estimate the frequency of a complex sinusoid quickly and accurately is paramount. While the Discrete Fourier Transform (DFT) provides a foundational frequency estimate, its resolution is inherently limited by the number of samples ($N$). To achieve higher precision, engineers often employ interpolation techniques between DFT bins.
This paper presents a novel frequency estimator designed for use with the Sliding DFT (SDFT), a computationally efficient algorithm that updates the spectrum on a sample-by-sample basis. By addressing the inherent stability issues of the SDFT and deriving a specific bias-correction formula, the authors demonstrate superior performance compared to industry-standard methods.
The Stability Challenge in Sliding DFT
The standard SDFT is "marginally stable," meaning its poles reside exactly on the unit circle in the z-domain. In practical implementations, numerical rounding errors can inadvertently push these poles outside the circle, leading to system instability. To mitigate this, a damping factor ($r < 1$) is typically introduced to move the poles slightly inside the unit circle, creating a "guaranteed stable" SDFT.
However, this damping factor introduces a new problem: it distorts the DFT bin values. Standard fine-frequency estimation formulas, such as those proposed by Jacobsen or Candan, do not account for this damping, resulting in significant estimation bias.
Proposed Estimator and Bias Correction
The authors propose a mathematical estimator for the frequency offset ($\delta$) that explicitly incorporates the damping factor $r$. The methodology involves:
- Damped SDFT Implementation: Treating the SDFT as an Infinite Impulse Response (IIR) filter with a damping factor $r$.
- Geometric Series Analysis: Deriving the frequency response of the damped system by analyzing the DFT bin calculation as a geometric progression.
- Closed-Form Expression: By evaluating the differences in frequency response around the peak bin and applying approximations for large $N$, the authors arrived at a new estimator $\hat{\delta}$. This formula includes a correction term $\frac{(1+r)^2}{4r}$ and a scaling factor $\frac{\tan(\pi/N)}{\pi/N}$, ensuring accuracy even in the presence of damping.
Performance Evaluation
The proposed method was evaluated using synthetic simulations of a single complex sinusoid corrupted by additive white Gaussian noise (AWGN). The simulation parameters included $N=128$ samples and a sampling frequency of 128 MHz, with signal frequencies varied between 30.1 MHz and 30.9 MHz.
Bias Reduction
In noise-free environments, the proposed estimator demonstrated a significant reduction in bias. While the Jacobsen and Candan (labeled as "Candace" in figures) methods showed substantial errors due to the damping factor, the proposed method maintained a bias of less than 1000 Hz.
Robustness to Noise
Under low Signal-to-Noise Ratio (SNR) conditions (2 dB to 3 dB), the proposed estimator continued to outperform existing techniques. It exhibited both lower bias and lower variance, proving its robustness in practical, noisy environments.
Conclusion
The research demonstrates that filter stability in SDFT implementations does not have to come at the cost of frequency accuracy. By mathematically predicting and canceling the bias introduced by the damping factor, the proposed closed-form estimator provides a highly efficient and accurate solution for real-time hardware like FPGAs and DSPs.
Future work aims to extend this derivation to multi-tone signals with closely spaced frequencies, further enhancing the utility of the stable SDFT in complex signal environments.
Summary updated