Allen Downey (@AllenDowney)

I am a professor at Olin College in Needham, MA, author of several books, including Think Bayes and Think DSP, and I write a blog called "Probably Overthinking It".

In Search of The Fourth Wave

Allen Downey ● September 25, 2021●4 comments

While working on Think DSP the presenter ran into a curious spectral pattern: sawtooth waves have all harmonics with amplitudes that scale like 1/f, square waves keep only odd harmonics with 1/f, and triangle waves keep odd harmonics with 1/f^2. That observation motivates a simple question: is there a basic waveform that has all integer harmonics but a 1/f^2 rolloff? The talk walks through four solution approaches, a fifth idea from the audience, and links to a runnable Colab notebook.

Autocorrelation and the case of the missing fundamental

Allen Downey ● January 21, 2016●10 comments

A short hands-on exploration shows why we perceive the fundamental pitch even when it's absent from the spectrum. Using saxophone recordings, high-pass filtering, and autocorrelation plots, the post demonstrates that the highest ACF peak often predicts perceived pitch rather than the strongest spectral line. The experiments also show that removing high harmonics eliminates the effect, and that autocorrelation is a useful but incomplete model of pitch perception.

Generating pink noise

Allen Downey ● January 20, 2016●1 comment

This post implements a stochastic Voss-McCartney pink-noise generator in Python, tackling why incremental per-sample algorithms do not map well to NumPy batch operations. It presents a practical NumPy/Pandas approach that uses geometric-distributed update events and pandas' fillna for column-wise zero-order hold to make batch generation efficient. The generated noise shows a power-spectrum slope near -1, matching expected 1/f behavior.

Amplitude modulation and the sampling theorem

Allen Downey ● December 18, 2015●6 comments

Amplitude modulation turns out to be a neat way to build intuition for the Nyquist-Shannon sampling theorem. In this draft chapter from Think DSP, the author shows how multiplying by a carrier shifts spectra, why sampling creates repeated copies in frequency, and how low-pass filtering can recover the original signal when those copies do not overlap.

Differentiating and integrating discrete signals

Allen Downey ● December 14, 2015●2 comments

Think DSP's new chapter digs into discrete differentiation and integration, using first differences, convolution, and FFTs to compare time and frequency domain views. The author reproduces diff via convolution then explores cumsum as its inverse and runs into two puzzling mismatches: noisy FFT amplitude ratios for nonperiodic data, and a time-domain convolution that does not reproduce cumsum for a sawtooth despite matching frequency responses. The post includes IPython notebooks and invites troubleshooting.

Bayes meets Fourier

Allen Downey ● October 26, 2015

Bayes filters and Fourier transforms turn out to have a neat symmetry: prediction uses convolution, while measurement update uses multiplication. In this post, Allen Downey shows how the characteristic function ties Bayes filtering to the Fourier domain, then uses that connection to sketch an FFT-based implementation that can speed up the predict-update cycle. If you like Bayesian estimation and signal processing, this is a satisfying crossover.