Summary

Rick Lyons disputes the common claim that using the discrete Fourier transform (DFT) as a filter inherently frequency-translates an input spectrum to DC. The paper explains how the DFT actually represents spectral samples, shows how DFT-based filtering is correctly implemented (zero-padding, overlap methods, frequency-domain multiplication), and highlights pitfalls such as circular convolution and spectral leakage.

Key Takeaways

• Differentiate between spectral analysis and filtering when using the DFT: the DFT samples spectrum bins, it does not automatically shift the entire spectrum to DC.
• Use zero-padding and proper block-processing (overlap-add/overlap-save) to obtain linear convolution with FFT-based filters and avoid circular convolution artifacts.
• Apply frequency-domain multiplication followed by an inverse DFT to implement filters efficiently, while managing windowing and bin-resolution trade-offs.
• Recognize and mitigate spectral leakage and resolution limits through window choice and segment length selection.
• Assess aliasing, bin-centered frequency errors, and how interpolation or longer DFTs improve filter selectivity and accuracy.

Who Should Read This

DSP engineers, graduate students, and applied researchers working on FFT-based filtering, spectral analysis, or implementation of digital filters in audio, radar, or communications who want to clarify DFT-based filtering principles and avoid common mistakes.

TimelessIntermediate

Topics