Integral Transforms from Finite Data: An Application of Gaussian Process Regression to Fourier Analysis

Luca Ambrogioni, Eric Maris

Abstract

Traditional spectral estimation via the Discrete Fourier Transform (DFT) relies on rigid assumptions of periodicity and band-limiting, which often lead to spectral leakage and artifacts in real-world, finite data. This paper introduces the Bayes-Gauss-Fourier (BGF) transform, a novel framework that treats Fourier analysis as a Bayesian inference problem using Gaussian Process (GP) regression. By learning a data-driven covariance structure, the BGF transform provides a continuous, analytic spectral estimate that significantly outperforms the DFT in sidelobe suppression and precision across synthetic and real-world datasets, including atmospheric $CO_2$ and brain MEG signals.

Introduction: The Limitations of the DFT

The Fourier transform is a cornerstone of signal processing, but its digital implementation—the DFT—comes with significant theoretical baggage. According to the Nyquist–Shannon sampling theorem, the DFT provides an unbiased estimate only if the signal is perfectly periodic and band-limited.

In practice, signals are rarely periodic or strictly band-limited. When non-periodic data is forced into a DFT, "spectral leakage" occurs: the implicit assumption that the signal repeats infinitely creates artificial sidelobes that obscure true spectral features. Estimating the continuous analog spectrum from finite samples is an ill-posed problem. The authors argue that instead of making rigid periodic assumptions, we should employ a Bayesian framework to assign a prior probability over the space of possible functions.

Methodology: The Bayes-Gauss-Fourier (BGF) Transform

The authors leverage Gaussian Process (GP) regression to bridge the gap between discrete samples and continuous integral transforms.

GP Regression as a Foundation

In GP regression, the posterior expectation of a signal is expressed as a linear combination of covariance kernels. Because the Fourier transform is a linear operator, the transform of the signal is simply the sum of the transforms of these kernels. This allows for a closed-form, analytic expression of the spectrum.

Hierarchical Covariance Learning

The "magic" of the BGF transform lies in how the kernel is selected. Rather than using a fixed kernel, the authors learn the spectral density directly from the data using a constrained gradient ascent method to find a Maximum-a-Posteriori (MAP) estimate. By learning the underlying covariance, the model effectively "imagines" how the signal continues beyond the observed window, naturally suppressing the sidelobes that plague the DFT.

Experimental Results and Key Insights

The BGF transform was validated using synthetic signals, Mauna Loa $CO_2$ data, and human magnetoencephalography (MEG) signals.

Precision and Sidelobe Suppression

In noise-free simulations, the BGF transform achieved sidelobe suppression up to $10^6$ times higher than a Hann-tapered DFT. As illustrated in Figure 1, the BGF method (labeled BFG in the figure) demonstrates superior performance in both the "passband" (the main signal) and the "stopband" (the noise floor) compared to standard and tapered DFT approaches.

Figure 1: Performance comparison in passband and stopband for BGF (BFG), DFT, and tapered DFT.

Visualizing Spectral Clarity

The visual difference in spectral estimation is stark. While the DFT produces broad peaks and significant leakage artifacts, the BGF transform provides a sharp, accurate representation of the true power spectrum.

Figure 2: Comparison of (A) signal sampling and (B) power spectrum estimation between BGF and various DFT windowing methods.

Real-World Applications

  • Atmospheric Data: On $CO_2$ records, the BGF identified subtle 3/year and 4/year frequency peaks that were nearly invisible in the DFT.
  • Neuroscience: On MEG data, the BGF produced significantly sharper peaks for alpha oscillations (approx. 10 Hz) against $1/f$ background noise compared to the popular multitaper (DPSS) method.

Discussion and Future Work

The BGF transform offers a powerful alternative to the DFT, particularly for short or noisy datasets where spectral resolution is critical. One of the paper's most elegant findings is the theoretical proof that the classical DFT is actually a special case of GP Fourier analysis with a specific, degenerate periodic prior.

However, challenges remain. The $O(N^3)$ computational complexity of GP regression makes it slower than the Fast Fourier Transform (FFT) for massive datasets. Additionally, the method currently assumes stationarity. Future research into non-stationary kernels could extend this Bayesian approach to signals with time-varying frequency content, further expanding the reach of Gaussian Processes in signal processing.

Summary updated

Abstract

Computing accurate estimates of the Fourier transform of analog signals from\ndiscrete data points is important in many fields of science and engineering.\nThe conventional approach of performing the discrete Fourier transform of the\ndata implicitly assumes periodicity and bandlimitedness of the signal. In this\npaper, we use Gaussian process regression to estimate the Fourier transform (or\nany other integral transform) without making these assumptions. This is\npossible because the posterior expectation of Gaussian process regression maps\na finite set of samples to a function defined on the whole real line, expressed\nas a linear combination of covariance functions. We estimate the covariance\nfunction from the data using an appropriately designed gradient ascent method\nthat constrains the solution to a linear combination of tractable kernel\nfunctions. This procedure results in a posterior expectation of the analog\nsignal whose Fourier transform can be obtained analytically by exploiting\nlinearity. Our simulations show that the new method leads to sharper and more\nprecise estimation of the spectral density both in noise-free and\nnoise-corrupted signals. We further validate the method in two real-world\napplications: the analysis of the yearly fluctuation in atmospheric CO2 level\nand the analysis of the spectral content of brain signals.\n
BibTeX
@inproceedings{ambrogioni2018integral,
  author = {Luca Ambrogioni and Eric Maris},
  title = {{Integral Transforms from Finite Data: An Application of Gaussian Process Regression to Fourier Analysis}},
  booktitle = {International Conference on Artificial Intelligence and Statistics},
  year = {2018},
}
Figure