Written and maintained by CASRAI Editorial Board
Last updated
Signal processing is the branch of engineering and applied mathematics concerned with representing, analyzing, transforming, and extracting information from signals — quantities that vary over time, space, or another variable and carry information. A signal can be a sound wave, a radio transmission, a heart’s electrical activity, a photograph, or a sensor reading logged once a second. Signal processing supplies the mathematical and computational tools to clean such signals up, compress them, detect what is in them, and turn them into something a person or another system can use. It sits inside electrical engineering and applied mathematics, and today it overlaps heavily with computer science, statistics, and machine learning.
What Signal Processing Studies
Almost every signal-processing problem falls into one of a few recurring tasks. Filtering removes or emphasizes parts of a signal, such as suppressing noise or isolating one radio channel. Transformation re-expresses a signal in a different domain, most often frequency, where its structure is easier to see. Estimation and detection infer hidden quantities or decide whether something is present, such as a heartbeat in a noisy recording. Compression and coding represent a signal with fewer bits while keeping what matters. Reconstruction rebuilds a signal or image from indirect, incomplete, or distorted measurements.
What unites these tasks is a shared toolkit: linear systems theory, Fourier analysis, probability and statistics, linear algebra, and optimization. A researcher who learns the toolkit once can apply it to audio, seismic recordings, medical images, or wireless links, which is why the field is usually described by its methods rather than by one subject matter.
Analog and Digital Signals
An analog signal is defined for every instant of time and can take any value in a continuous range — the voltage from a microphone, for example. A digital signal is a sequence of numbers: the signal has been sampled at discrete times and each sample quantized to one of a finite set of values. Analog processing uses physical circuits such as resistors, capacitors, and amplifiers. Digital signal processing (DSP) performs the same kinds of operations as arithmetic on the numbers, using general-purpose processors, specialized DSP chips, FPGAs, or graphics processors.
Digital processing largely displaced analog processing for most tasks because it is repeatable, programmable, and not subject to component drift, and because the same hardware can run different algorithms. Analog electronics have not disappeared: sensors, amplifiers, and the converters at the boundary between the physical world and the computer are all analog design problems, and some high-frequency or ultra-low-power tasks are still better done in analog.
Sampling Theory and the Nyquist-Shannon Theorem
Converting an analog signal to digital form raises an immediate question: how often must it be sampled so that nothing is lost? The answer is the Nyquist-Shannon sampling theorem. In its standard form, a signal containing no frequency components above B hertz is completely determined by samples taken at a rate greater than 2B samples per second, and it can in principle be reconstructed exactly from those samples. The threshold of 2B is called the Nyquist rate.
The theorem’s practical consequence is aliasing. If a signal is sampled too slowly, frequency content above half the sampling rate does not disappear; it folds back and masquerades as lower-frequency content that was never in the original. This is why analog-to-digital converters are almost always preceded by an anti-aliasing filter that removes content above the range that the sampling rate can represent. The same effect appears in everyday life as the apparent backward spinning of wheels in film.
The theorem’s history is usually told in two steps. Harry Nyquist’s 1928 paper on telegraph transmission showed that a system of bandwidth B can carry up to 2B independent pulse samples, though he did not explicitly treat sampling and reconstruction of continuous signals. Claude Shannon proved the sampling theorem in the form used today in “Communication in the Presence of Noise,” published in the Proceedings of the IRE in January 1949. Idealized sampling assumes perfectly band-limited signals, and real signals never are exactly, so practical design always involves margins, filters, and an accounting of quantization error.
Fourier Methods
The Fourier transform expresses a signal as a sum of sinusoids of different frequencies. It moves the analysis from the time domain, where one asks what the signal does at each instant, to the frequency domain, where one asks which frequencies are present and how strongly. Many operations that are laborious in time, such as filtering, become simple multiplication in frequency. For sampled signals the working tool is the discrete Fourier transform (DFT), and its efficient implementation is the fast Fourier transform (FFT).
The FFT is a landmark of the field. J. W. Cooley and J. W. Tukey described their algorithm in “An algorithm for the machine calculation of complex Fourier series,” Mathematics of Computation 19 (1965), pp. 297–301. It reduces the cost of computing a transform of N points from roughly proportional to N squared to roughly proportional to N log N, which is what made spectral analysis practical on real data sizes. FFT-based methods now sit under spectrum analyzers, audio codecs, radar processing, and the multicarrier modulation used in many wireless systems.
The classical Fourier transform has a limitation: it describes which frequencies occur but not when, so it suits stationary signals whose character does not change. For signals whose frequency content evolves, such as speech, practitioners use the short-time Fourier transform, which analyzes the signal through a sliding window and produces a spectrogram.
Wavelet Methods
Wavelets address the same time-and-frequency problem differently. Instead of infinitely long sinusoids, a wavelet transform decomposes a signal using short, localized oscillating functions that are scaled and shifted. Because the analysis window narrows for high-frequency detail and widens for slow trends, wavelets are well suited to signals with sharp transients, edges, or features at several scales at once. Ingrid Daubechies’s paper “Orthonormal bases of compactly supported wavelets,” Communications on Pure and Applied Mathematics 41 (1988), pp. 909–996, presented a class of wavelets that has been applied in many settings. Typical uses include denoising, image compression, and detecting abrupt changes in a recording. A choice between Fourier and wavelet analysis is a modeling decision: Fourier methods are the natural fit for stationary, periodic structure; wavelets for localized and multi-scale structure.
Filtering
A filter is a system that modifies a signal by passing some components and attenuating others. Low-pass, high-pass, band-pass, and band-stop filters are defined by which frequency ranges they keep. In digital form there are two main families. Finite impulse response (FIR) filters compute each output as a weighted sum of a finite number of recent input samples; they are always stable and can be designed to have linear phase, which preserves waveform shape. Infinite impulse response (IIR) filters also feed back previous outputs, which achieves sharp frequency responses with far fewer coefficients but requires attention to stability and phase distortion.
Beyond fixed filters, adaptive filters adjust their coefficients as data arrives, which is how echo cancellation and noise-cancelling headphones track changing conditions. Statistical estimation filters are a further family: the Wiener filter is the classical optimal linear filter for separating a signal from noise given their statistics, and the Kalman filter, introduced by Rudolf Kalman around 1960, recursively estimates the state of a dynamic system from noisy measurements. It underlies navigation, tracking, and sensor-fusion systems.
Major Subfields
- Audio, speech, and acoustic processing — coding, enhancement, recognition, and synthesis of sound.
- Image and video processing — enhancement, restoration, compression, segmentation, and computational imaging.
- Communications signal processing — modulation, equalization, channel estimation, and synchronization.
- Radar, sonar, and array processing — extracting direction, range, and velocity from sensor arrays.
- Biomedical signal processing — analysis of electrical, acoustic, optical, and imaging signals from the body.
- Statistical signal processing — detection and estimation theory, spectral estimation.
- Sparse and compressive methods, and learning-based processing — recovering signals from fewer measurements and learning representations from data.
Applications
Biomedical signals
Clinical and research recordings are noisy by nature. An electrocardiogram (ECG) is contaminated by baseline drift from breathing, muscle artifacts, and interference from mains power at 50 or 60 Hz; an electroencephalogram (EEG) records microvolt-scale activity that must be separated from eye-blink and muscle artifacts. Filtering, artifact removal, time-frequency analysis, and detection algorithms are standard steps in turning such recordings into measurements. For the engineering field that builds the devices around these recordings, see CASRAI’s guide to what biomedical engineering is.
Imaging
Signal processing is built into how many images are formed, not just how they are cleaned up afterward. Magnetic resonance imaging acquires data in the spatial-frequency domain and forms an image with a Fourier transform; computed tomography reconstructs cross-sections from many projections using filtered back-projection. In microscopy, restoration algorithms undo blur: Richardson-Lucy deconvolution is one worked example. For the clinical side, see what medical imaging is.
Spectroscopy and instrumentation
Instruments that measure spectra rely on the same mathematics. Fourier-transform infrared (FTIR) instruments measure an interferogram in one domain and recover the spectrum with a Fourier transform; the practical differences between techniques are laid out in CASRAI’s comparison of Raman and FTIR spectroscopy. Interpreting spectra also involves baseline correction, smoothing, and peak detection — see the guides on reading an IR spectrum, fluorescence spectroscopy, and mass spectrometers.
Communications
Every wireless and wired link depends on signal processing: modulating information onto a carrier, filtering out-of-band energy, estimating and equalizing the distortion introduced by the channel, synchronizing receiver and transmitter, and applying error-correcting codes. Multicarrier techniques widely used in modern wireless standards rely on the FFT to modulate and demodulate many subcarriers efficiently.
Data science and machine learning
Feature extraction from audio, vibration, or sensor streams usually starts with signal-processing steps, and many neural-network layers are themselves filters. The boundary with data science and machine learning is therefore porous, and increasingly the two are combined: classical models supply structure and interpretability; learned models supply flexibility.
A Short History of the Discipline
Several threads converge. Fourier’s analysis of heat flow in the early nineteenth century supplied the mathematics of frequency decomposition. Twentieth-century telephony and radio supplied the engineering problems; Nyquist’s 1928 work and Shannon’s 1949 paper gave the theory of sampling and, more broadly, information. Wartime and postwar work on radar, filtering, and prediction fed the statistical methods. Then digital computers made it practical to process signals as numbers, and the 1965 FFT paper made frequency analysis cheap enough to use routinely. The 1980s brought wavelet theory.
The profession’s institutional history shows how late the label arrived. The IEEE Signal Processing Society began on June 2, 1948 as the Professional Group on Audio of the Institute of Radio Engineers, the first professional group of the IRE. It became the IEEE Group on Acoustics, Speech, and Signal Processing in 1974, the Acoustics, Speech, and Signal Processing Society in 1976, and took its present name, the IEEE Signal Processing Society, on October 6, 1989 — at which point signal processing was recognized as a discipline in its own right.
Societies, Journals, and Conferences
The principal professional home is the IEEE Signal Processing Society, which publishes journals such as IEEE Transactions on Signal Processing and sponsors the annual International Conference on Acoustics, Speech and Signal Processing (ICASSP). Signal-processing papers also appear in the IEEE transactions on image processing, audio and speech, and biomedical engineering, and in applied-mathematics venues. As with any field, authorship and data-sharing practices are set by the journal and publisher, and funders increasingly expect shared code and datasets so that results can be reproduced.
Training Paths
Most signal-processing researchers train in electrical engineering, computer engineering, applied mathematics, physics, or computer science; see what electrical engineering is and what engineering is for the parent discipline, and the branches of science for the wider map. A typical undergraduate path covers calculus, linear algebra, probability, circuits, and a core course on signals and systems, followed by electives in DSP, communications, image processing, and estimation. Graduate study adds statistical signal processing, optimization, and machine learning. Because the methods are portable, graduates work in communications, medical devices, audio, defense and sensing, imaging, and data-driven research roles. Practical fluency means being able to implement algorithms in a numerical language, reason about sampling and noise, and validate results against known signals.
How Signal-Processing Research Is Funded
In the United States, core signal-processing research is supported by the National Science Foundation’s Computing and Communication Foundations division. Its Communications and Information Foundations (CIF) program supports work on the theoretical underpinnings of information acquisition, transmission, and processing, and describes its scope as including communications, information theory, coding theory, and signal and image processing. Stated areas of interest include signal processing on graphs and networks, geometric methods in signal processing and machine learning, and computational imaging. Applied work is also funded through domain agencies, for example through health agencies for biomedical signals and through defense and energy agencies for sensing, and elsewhere through national research councils. Program names, deadlines, and award sizes change, so applicants should confirm current solicitations with the funder. For the broader grant lifecycle, see CASRAI’s material on grants management.
Signal Processing and Research Administration
For research administrators, signal-processing projects have recognizable features. They are often computation-heavy, so budgets lean on personnel, computing, and sometimes specialized instrumentation rather than consumables. Projects that record human signals such as EEG or ECG involve human-subjects review and data-protection obligations even when the analysis is purely computational. Software and datasets are first-class outputs, which raises questions of licensing, data management plans, and long-term access. Collaborations with clinical, industrial, or defense partners add subaward, intellectual-property, and, for some sensing and communications technologies, export-control questions that should be raised at proposal stage rather than discovered at award.
Frequently Asked Questions
What is signal processing in simple terms?
It is the set of mathematical and computational techniques for cleaning up, analyzing, compressing, and extracting information from signals such as sound, images, radio waves, and sensor readings.
What is the difference between analog and digital signal processing?
Analog processing operates on continuously varying physical quantities with electronic circuits. Digital processing operates on sampled, quantized numbers using algorithms, which makes it programmable and repeatable.
What does the Nyquist-Shannon theorem say?
A signal with no frequency content above B hertz can be reconstructed exactly from samples taken at more than 2B samples per second. Sampling slower causes aliasing.
Fourier transform or wavelet transform: which should I use?
Use Fourier methods for stationary signals where frequency content is the question. Use wavelets, or a short-time Fourier transform, when frequency content changes over time or when features are localized.
Is signal processing the same as electrical engineering?
No. It began within electrical engineering and is still taught there, but it is a distinct field with its own society and journals, and its methods are used in mathematics, physics, biology, and computer science.
How does signal processing relate to machine learning?
They overlap. Machine learning often relies on signal-processing features and ideas such as convolution, and signal-processing systems increasingly use learned components.
Who funds signal-processing research?
In the US, the NSF’s Communications and Information Foundations program is one core source; domain agencies fund application-driven work. Check current solicitations for specifics.
