A Synthesizer Based on Frequency-Phase Analysis and Square Waves

This article introduces an effective generalization of the polar flavor of the Fourier Theorem based on a new method of analysis. Under the premises of the new theory an ample class of functions become viable as bases, with the further advantage of using the same basis for analysis and reconstruction. In fact other tools, like the wavelets, admit specially built nonorthogonal bases but require different bases for analysis and reconstruction (biorthogonal and dual bases) and vectorial coordinates; this renders those systems unintuitive and computing intensive. As an example of the advantages of the new generalization of the Fourier Theorem, this paper introduces a novel method for the synthesis that is based on frequency-phase series of square waves (the equivalent of the polar Fourier Theorem but for nonorthogonal bases). The resulting synthesizer is very efficient needing only few components, frugal in terms of computing needs, and viable for many applications.

💡 Research Summary

The paper proposes a novel synthesis technique that replaces the traditional sinusoid‑based additive synthesis with a square‑wave‑based frequency‑phase decomposition. The authors argue that while the Fourier theorem provides a mathematically elegant framework, its reliance on sine and cosine functions makes real‑time digital implementation costly: each sinusoid requires many samples, lookup tables, or dedicated DSP hardware. To overcome this, they introduce a generalization of the polar (phase) form of the Fourier theorem that allows any sufficiently regular periodic function to serve as a basis, provided the analysis algorithm is tailored to that basis.

The core contribution is a recursive analysis algorithm that expresses an arbitrary L² periodic signal as a sum of square waves of different frequencies, each described by an amplitude and a phase. Starting with a fundamental‑frequency square wave, the algorithm subtracts its contribution, then iteratively removes the residual error using higher‑frequency square waves whose frequencies are odd multiples (or, more generally, prime multiples) of the fundamental. Each iteration pushes the remaining error to higher frequencies, creating a high‑frequency “spike” spectrum that can be eliminated with simple oversampling or low‑pass filtering. Because the basis is non‑orthogonal, the error does not cancel as in the orthogonal Fourier case, but the authors demonstrate that the error energy concentrates at frequencies far above the audible range, making it practically negligible after filtering.

Mathematically, the authors invert the classic Fourier series of a square wave to show that a single sinusoid can be reconstructed from an infinite series of square waves. By linearity, any L² function can therefore be represented as a square‑wave series, establishing the square wave as a complete basis for the space. The paper includes an appendix (not reproduced here) that formalizes this claim and extends it to a broader class of non‑orthogonal bases.

Experimental results illustrate the method on several signals: a pure sine wave, a composite signal built from multiple square waves, and a more complex waveform approximated with 21 and 36 square‑wave components. With as few as 21 components (up to 55× the fundamental frequency) the reconstructed sine closely matches the original, with only high‑frequency noise. Increasing to 36 components (up to 50× the fundamental) further reduces the noise. The authors compare the square‑wave reconstruction to a Haar wavelet reconstruction of the same sine; despite using fewer components, the square‑wave method yields lower RMS error, though it requires two parameters (amplitude and phase) per component versus one for the wavelet.

The paper highlights several practical advantages:

1. Computational Efficiency – Only two scalar operations per component (multiply by amplitude and add phase) are needed, halving the arithmetic load compared with complex‑exponential Fourier synthesis.
2. Hardware Simplicity – Square waves are trivial to generate digitally (just toggle between +1 and –1), eliminating the need for trigonometric function generators or large lookup tables.
3. Unified Analysis/Synthesis Basis – The same set of square waves is used for both forward analysis and inverse synthesis, avoiding the bi‑orthogonal or dual bases required by many wavelet and frame methods.

The authors acknowledge limitations: the non‑orthogonal nature introduces high‑frequency spurious components that must be filtered, and precise phase estimation can be sensitive to quantization error. Nevertheless, they argue that in low‑power or low‑cost environments (e.g., embedded audio chips, RF signal generators, or battery‑operated devices) the trade‑off is favorable.

Finally, the paper suggests broader implications: the same framework could be applied to nonlinear signal processing, generic filtering by manipulating the “square‑wave spectrum,” and even cross‑instrument timbre morphing by representing one instrument’s waveform as a square‑wave basis and re‑synthesizing another. Future work will expand the theory to other non‑orthogonal bases and explore applications beyond audio synthesis.