Lecture
Echoes of Fourier
- The presentation intends to bridge the relationship between echoes and the Fourier transform by first reviewing complex numbers in polar form, where radiuses multiply and angles add, leading to Euler's formula ($e^{i\theta} = \cos\theta + i\sin\theta$) which describes circular motion as the natural form of periodic motion.
- Time is defined as discrete for computational purposes with 44,100 moments per second, and signals will be decomposed into a sum of complex sinusoids that are linearly independent and form a basis for $C^n$, where coefficient magnitudes represent frequency loudness and angles represent temporal phase.
- The mapping from the time domain to the frequency domain is identified as the Discrete Fourier Transform (DFT) and its inverse, which can be computed efficiently in $n \log n$ time using the Fast Fourier Transform (FFT) algorithm, a divide-and-conquer method utilizing even and odd indices.
- For real-valued signals, high frequencies appear as negative low frequencies due to mirror image symmetry, while complex sound phases are expected to be difficult to interpret, and filtering by retaining frequencies between 300 and 400 cycles per second will yield a real signal upon inversion with minor numerical error requiring playback of the real part.
- Echoes are modeled as the convolution of an original sound with an impulse response, a process computationally accelerated from $n^2$ to $n \log n$ multiplications via the convolution theorem; synthetic impulse responses for simulated environments, such as a 30-meter hallway or a 3x3x3 meter room, will demonstrate resonant frequencies corresponding to specific wavelengths.
- Practical challenges include the need to rescale convolution values to avoid clipping when they exceed the -1 to 1 range, and the inherent difficulty of deconvolution, where small Fourier coefficients can cause instability and large inverse coefficients, making tasks like digital room correction or image sharpening approximate rather than exact.
- Simulations involving a three-dimensional lattice of points require $O(n^3)$ calculations, and while deconvolving by the initial part of an impulse response has been attempted, it is predicted to fail significantly, necessitating techniques to smear frequency components in time to avoid driving microphones into nonlinearity.