Signal processing

Signal processing is the discipline of extracting information from signals — time-varying, spatial, or spectral representations of physical or abstract phenomena. The field is built on the mathematical tools of Fourier analysis, filter theory, and sampling theory, and it underlies virtually every modern technology that converts between the analog and digital worlds: audio, video, telecommunications, radar, medical imaging, and seismic exploration.

The foundational theorem of signal processing is the Nyquist-Shannon sampling theorem, which states that a bandlimited signal can be perfectly reconstructed from its samples if the sampling rate exceeds twice the maximum frequency. This theorem is not merely a technical result; it is a structural claim about the relationship between continuous and discrete representations. It says that information is not lost in the transition from analog to digital, provided the sampling rate is high enough. The theorem is the basis of the digital revolution.

Signal processing also encompasses the theory of filters — systems that modify signals by attenuating or amplifying particular frequency components. Filter design is an optimization problem: given a desired frequency response, find a system (typically a linear time-invariant system) that approximates it within specified tolerances. The Wiener filter and the Kalman filter are optimal estimators that minimize mean squared error under different assumptions about noise and signal statistics.

Modern signal processing extends beyond linear methods to machine learning and deep learning, where neural networks are trained to perform denoising, source separation, and feature extraction. The Tucker decomposition and other tensor methods have become essential tools for analyzing multi-dimensional signals. The field has also expanded into the analysis of network signals — time series on graphs — where the structure of the network influences the properties of the signal.

The fundamental tension in signal processing is between representation and interpretation. A signal is a representation; processing is interpretation. The Nyquist theorem guarantees that representation is lossless; it says nothing about whether the interpretation is correct. The field's confidence in its mathematical foundations has sometimes obscured the fact that the choice of what to measure, how to sample, and what to filter is always a choice about what matters.

Signal processing is not the science of recovering information from noise. It is the science of deciding what counts as information and what counts as noise — and that decision is not technical but epistemological. The Fourier transform does not reveal the frequencies that are "really there"; it reveals the frequencies that are there given the assumption that the signal is periodic and infinite. The sampling theorem does not guarantee perfect reconstruction; it guarantees reconstruction under the assumption that the signal is bandlimited. Every theorem in signal processing carries a hidden ontology, and the field's greatest weakness is its tendency to treat that ontology as natural rather than chosen.