Understanding the FFT Algorithm (2013)

The spark

Jake VanderPlas dug into something many of us take for granted: the Fast Fourier Transform. He admits — honest and human — that despite years of using FFTs he never stopped to ask how the algorithm actually speeds up the Discrete Fourier Transform. That moment of curiosity is the emotional core of the piece. It’s a reminder: even seasoned practitioners can rediscover delight in fundamentals. Who hasn’t had that “wait, how does that work?” jolt?

What the post explains

VanderPlas walks through the Cooley–Tukey trick from 1965 and lays out the math behind why the FFT drops computation from O(N^2) to O(N log N). He doesn’t stop at theory: the post includes straightforward Python implementations and timing comparisons. The simple matrix-based DFT he builds is more than 1,000× slower than NumPy’s fft on a 1024-point example — a crisp, practical demonstration of why the algorithm matters. He also notes common libraries (numpy.fft, scipy.fftpack) and points readers toward FFTW; it has been reported that FFTW is widely regarded as a top performer.

Why it matters now

This isn’t just a history lesson. FFTs underpin audio, imaging, astronomy, and many machine‑learning pipelines. VanderPlas ties the discussion to real use cases — from power spectra to solving differential equations — and offers runnable code so data scientists can peek under the hood. The piece reads like a short course: hands-on, curious, and unpretentious. Want to stop treating black‑box tools like magic? Start here.

Sources: jakevdp.github.io, Hacker News