High-resolution frequency estimation using generalized coprime sampling

Si Qin, Yimin D. Zhang, Moeness G. AminProceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE 2015 · DOI

TL;DR

This paper introduces a generalized coprime sampling technique that significantly enhances the ability to estimate signal frequencies from undersampled data. By extending the original coprime sampling method, it achieves finer frequency resolution and identifies more frequencies without substantially increasing the time it takes to collect data.

What Problem Does This Paper Address?

Modern applications, from radar to communications, often deal with signals that have very high bandwidths. Sampling these signals at the Nyquist rate (twice the highest frequency) requires extremely fast and expensive analog-to-digital converters (ADCs). This paper tackles the challenge of accurately estimating signal frequencies when using sub-Nyquist sampling, meaning the sampling rate is much lower than required by the Nyquist theorem. Specifically, it addresses the limitations of existing "coprime sampling" techniques, which, while effective, can struggle to resolve closely spaced frequencies or identify a large number of them without introducing significant delays in data processing.

Methods

The core innovation is a "generalized coprime sampling" technique. The original coprime sampling method uses two samplers with coprime down-sampling rates, say $M$ and $N$, to generate $MN$ virtual samples from $M+N$ physical samples. This paper introduces an integer factor $p$ to generalize this. Instead of using $2N$ samples from one sampler, the generalized method uses $pN$ samples. This allows for the generation of $pMN$ virtual samples, effectively creating a much larger set of correlation lags. This expanded set of lags is then used to construct a larger covariance matrix, which is the basis for frequency estimation using standard techniques like MUSIC. The paper provides analytical derivations for the number of virtual samples, frequency resolution, and latency time as a function of $p$.

Datasets

The paper uses simulated sinusoidal signals with varying frequencies and signal-to-noise ratios (SNRs) to evaluate the proposed technique. No real-world datasets are used.

Key Insights and Ideas

  • Generalization is Key: The original coprime sampling is shown to be a special case ($p=2$) of a broader framework, allowing for systematic improvement.
  • More Lags, Better Resolution: Increasing the parameter $p$ directly leads to a larger number of available correlation lags. This, in turn, significantly improves the frequency resolution, allowing for the separation of much closer frequency components.
  • More Lags, More Frequencies: A larger number of lags also enables the identification of a greater number of distinct frequencies within the signal.
  • Latency Trade-off: While increasing $p$ improves performance, the latency time (time to collect data) is shown to increase only slightly, especially when a large number of time blocks ($L$) are used for correlation estimation. This makes the generalized approach attractive for real-time applications.

Key Results

  • Improved Frequency Identification: For a scenario with 61 uniformly distributed frequencies, the generalized method with $p=30$ successfully identified all frequencies, whereas the conventional method ($p=2$) and $p=5$ failed to resolve many.
  • Finer Frequency Resolution: When tested with closely spaced frequencies (e.g., 45 MHz and 45.75 MHz), the generalized method ($p=30$) could resolve all pairs, while $p=5$ could only resolve wider pairs, and $p=2$ could resolve none.
  • Enhanced Identification Performance: Monte Carlo simulations showed that the Root Mean Square Error (RMSE) of frequency estimates improved with increasing $p$, indicating more accurate frequency localization.
  • Manageable Latency: The latency time increased only marginally with $p$, particularly as the number of data blocks ($L$) for correlation estimation grew.

Limitations and Future Work

  • Computational Complexity: While not explicitly detailed, increasing $p$ will inherently increase the size of the covariance matrix, leading to higher computational complexity for estimation.
  • Bounded $p$: The maximum value of $p$ is limited by the system's minimum data updating rate, which depends on $M$, $N$, $L$, and $T_s$.
  • Future Work: The paper implicitly suggests exploring optimal choices of $p$, $M$, and $N$ for specific applications and further investigating the impact of noise on the generalized scheme. The authors also mention extending this concept to array configurations for direction-of-arrival estimation, which has been explored in their prior work.

Summary updated

Abstract

The coprime sampling scheme allows signal frequency estimation through two sub-Nyquist samplers where the down-sampling rates M and N are coprime integers. By considering the difference set of this pair of O(M + N ) physical samples, O(MN ) consecutive virtual samples can be generated. In this paper, a generalized coprime sampling technique is proposed by using O(M + pN ) samples to generate O(pMN ) virtual samples, where p is an integer argument. As such, the existing coprime sampling techniques are represented as a special case of a much broader and generalized scheme. The analytical expressions of the number of virtual samples, frequency resolution and the corresponding latency time are derived. The effectiveness of the proposed technique is verified using simulation results.
BibTeX
@inproceedings{qin2015high,
  author = {Si Qin and Yimin D. Zhang and Moeness G. Amin},
  title = {{High-resolution frequency estimation using generalized coprime sampling}},
  booktitle = {Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE},
  year = {2015},
  doi = {10.1117/12.2178341},
}
Figure