Signal Processing

Elements and Operations

Signal processing studies how information carried by physical or digital waveforms can be measured, transformed, filtered, and reconstructed. The irreducible elements are the signal itself (a function of time or space) and its spectrum (the distribution of energy across frequencies).

Waveform primitives (amplitude, frequency, phase) compose the time-domain signal. The Fourier transform (and its fast implementation, the FFT) converts between domains. Convolution with a kernel implements filtering — the central operation that selectively attenuates or passes frequency components. Sampling discretizes the continuous signal; when the Nyquist frequency is violated, aliasing irreversibly corrupts the spectrum.

Higher-order structures such as the Wiener filter (optimal statistical estimator) and image pyramids (multi-scale representations built by iterated filtering + decimation) are composed from these atoms. Cross-links to wave mechanics and optics appear naturally: electromagnetic and acoustic waves are signals; diffraction and interference are frequency-domain phenomena; graphics rendering pipelines are heavy users of exactly these operations.

Axiomatic Foundations and Theorems

The field rests on a small set of powerful axioms and their consequences:

• Linearity and superposition allow complex signals to be decomposed into simpler ones whose transforms are easy to combine.
• The convolution theorem turns expensive time-domain convolution into cheap multiplication in the frequency domain — the practical basis for fast filtering.
• The Nyquist-Shannon sampling theorem gives the exact rate at which a band-limited signal must be sampled for perfect reconstruction; violation produces aliasing (a hard information-loss boundary).
• Parseval’s relation guarantees energy preservation across domains, enabling power-spectrum calculations and Wiener optimality proofs.
• The Wiener filter formula is the unique linear estimator that minimizes mean-squared error given known signal and noise spectra.

These are not engineering heuristics; they are mathematical certainties that delimit what any subsequent algorithm or hardware can achieve.

Measurement, Spectra, and Causal Structure

Everything in signal processing is measurable and causal:

• Frequency (Hz), power spectral density, SNR, and PSNR are the primary observables. They are estimated from data (periodograms, Welch averages) or computed exactly for synthetic test signals.
• Filter frequency response (magnitude and phase) is measured by driving the filter with sinusoids or impulses and observing the output.
• Causal links are direct: increasing noise power strictly decreases SNR; widening a low-pass kernel strictly lowers the cutoff frequency; omitting the anti-alias pre-filter before downsampling strictly increases aliasing energy; applying the Wiener filter strictly increases PSNR relative to blind Gaussian smoothing (for the assumed spectra).

These quantities and links drive both the design of algorithms and the evaluation of real systems (audio codecs, medical imaging, wireless receivers, computational photography).

Core Procedures

Three representative, fully-specified procedures illustrate the field:

1. Separable Convolution — the workhorse of image and volume filtering (directly from the raw exercises).
2. FFT-based Wiener Filtering — optimal statistical restoration with explicit frequency-domain design and IFFT reconstruction.
3. Gaussian / Laplacian Pyramid Construction — the multi-scale representation that powers blending, compression, and scale-space analysis.

Each procedure lists concrete inputs, numbered steps with cost and correctness arguments, and outputs. All are implementable and appear in production DSP, graphics, and scientific computing stacks.

(See the algorithmic YAML for the detailed step lists derived from textbook exercises and standard DSP practice.)

Stocks, Flows, and Multi-Scale Feedback

A signal or image is a dynamical system whose state is the distribution of energy across frequency bands. Filtering and sampling are flows that move energy between bands or to coarser scales.

• Balancing loop (noise-detail): noise increases residual error → Wiener correction applies stronger attenuation in noisy bands → residual drops, but at the cost of lost high-frequency detail. The system equilibrates at the point where marginal detail loss equals marginal noise reduction (exactly the Wiener optimum).
• Balancing loop (pyramid consistency): each pyramid level is a low-pass + decimate of the level above; reconstruction (upsample + add Laplacian) must recover the original within the Nyquist limits of each band. Any mismatch produces visible artifacts or drift across scales.

These stock/flow structures explain why pyramid and wavelet methods are so effective: they explicitly represent the equilibria that linear filtering alone leaves implicit.

Objectives, Trade-offs, and Hard Constraints

The engineering problem is: given a signal corrupted by noise or blur, and given limits on compute, latency, power, and numeric precision, design a filter or sampler that extracts the maximum usable information.

Objectives are information-theoretic or perceptual (max SNR/PSNR, max structural similarity, minimal detectable artifact).

Constraints are unforgiving:

• Computational: FFT is O(N log N); direct convolution for large kernels or 3D data is prohibitive on embedded or real-time hardware.
• Temporal: group delay, pipeline depth, and buffering must fit the application (audio < 10 ms, video 16-33 ms).
• Stability & numeric: IIR filters can oscillate or amplify quantization noise; FFT bin leakage and windowing must be controlled.
• Sampling physics: once aliasing has folded high frequencies into baseband, no post-processing can recover them.

Practical systems (JPEG, MP3, medical ultrasound beamformers, radar, graphics texture pipelines) are all explicit negotiations of these objectives against these constraints, using exactly the atoms and procedures declared above.