Loading field map
How should an engineer represent a signal—a voltage waveform, a sound, a radio transmission—so that its information can be extracted, modified, or transmitted? The answer has changed repeatedly over the past 130 years, and each change has brought a different set of mathematical commitments, computational tools, and practical limitations. The history of signals and systems is not a smooth accumulation of techniques but a series of competing frameworks, each offering a distinct answer to the same question: what matters most about a signal, and how should we process it?
The first systematic framework for signal analysis emerged from the practical pressure of alternating-current power systems. In the 1890s, Charles Proteus Steinmetz showed that steady-state AC circuits could be analyzed by representing sinusoidal voltages and currents as complex numbers—phasors—and treating capacitors and inductors as frequency-dependent impedances. This Frequency-Domain Analysis (1893) transformed a set of differential equations into algebraic manipulations, but it came with a strict assumption: the system must be linear and time-invariant (LTI), and the signals must be periodic or at least stationary in their frequency content. Engineers could now design filters and transmission lines using transfer functions, but they could not easily handle transient behavior or time-varying signals.
Almost simultaneously, engineers began building dedicated analog circuits—filters, amplifiers, modulators—that performed signal-processing operations in continuous time. Analog Signal Processing (1900) became the first practical infrastructure for shaping signals. Its methods were entirely hardware-based: resistors, capacitors, inductors, and later operational amplifiers realized mathematical operations like integration, differentiation, and frequency selection. Analog processing was fast and continuous, but it suffered from component drift, temperature sensitivity, and the difficulty of reconfiguring a circuit once built. For most of the first half of the twentieth century, frequency-domain analysis and analog hardware formed a tightly coupled pair: the theory assumed LTI behavior, and the hardware delivered it.
By the 1930s, control engineers needed to analyze systems that were not easily captured by transfer functions—systems with multiple inputs and outputs, or with internal dynamics that could not be reduced to a single input-output relationship. Linear Systems Theory (1930) provided a time-domain alternative. Instead of representing a system by its frequency response, engineers described it by a set of first-order differential equations—the state-space representation. This framework could handle initial conditions, transient responses, and multivariable systems in a way that frequency-domain methods could not. Yet it did not replace frequency-domain analysis; the two coexisted, with state-space methods dominating control theory and transfer functions remaining central to filter design and communications. The key difference was methodological: frequency-domain analysis treated signals as sums of sinusoids, while state-space analysis treated them as trajectories in a vector space.
By the mid-1940s, the limitations of pure frequency-domain thinking had become clear. Many real-world signals—speech, radar returns, seismic vibrations—change their frequency content over time. A single Fourier transform could not capture when a particular frequency occurred. Time-Frequency Analysis (1946), pioneered by Dennis Gabor, introduced the spectrogram and the short-time Fourier transform, which sliced a signal into short windows and computed the frequency content of each window. This framework preserved both time and frequency information, but at the cost of a fundamental trade-off: narrow windows gave good time resolution but poor frequency resolution, and vice versa. Time-frequency analysis did not replace frequency-domain methods; it extended them to non-stationary signals, and it later provided the conceptual foundation for wavelet theory and multirate filter banks.
At nearly the same moment, Norbert Wiener and others introduced a radically different approach. Statistical Signal Processing (1949) treated signals not as deterministic functions but as random processes. Wiener's theory of optimal filtering (the Wiener filter) assumed that signals and noise were stationary random processes with known autocorrelation functions. This framework brought probability and statistics into signal processing, enabling engineers to design filters that minimized mean-square error even when the signal was buried in noise. Statistical signal processing did not reject frequency-domain analysis; it absorbed it, using power spectral densities to characterize random processes. But its probabilistic assumptions were a sharp departure from the deterministic worldview of earlier frameworks.
The 1960s saw three frameworks emerge in quick succession, each responding to a different limitation of the existing toolkit.
Adaptive Signal Processing (1960), introduced by Bernard Widrow and others, addressed the problem of environments that change over time. Instead of designing a fixed filter based on assumed statistics, adaptive algorithms—such as the least-mean-squares (LMS) algorithm—adjusted filter coefficients in real time using the incoming data. This framework preserved the statistical perspective of Wiener filtering but dropped the assumption of stationarity: the filter could track a changing signal or interference. Adaptive methods became essential for echo cancellation, beamforming, and equalization.
Model-Based Signal Processing (1960) took a different path. Rudolf Kalman's state-space formulation of linear filtering—the Kalman filter—combined a dynamic model of the system (how the state evolves over time) with a measurement model (how observations relate to the state). Unlike Wiener filtering, which required the entire past signal, the Kalman filter operated recursively, processing each new measurement as it arrived. Model-based signal processing assumed that the engineer could write down a reasonably accurate mathematical model of the signal generation process. When the model was good, the results were excellent; when the model was poor, the filter could diverge.
Digital Signal Processing (1965) transformed the entire field. The discovery of the fast Fourier transform (FFT) by Cooley and Tukey made frequency-domain analysis computationally feasible on digital computers. Digital signal processing (DSP) replaced analog circuits with programmable algorithms: a filter could be implemented as a difference equation, a Fourier transform as a sequence of arithmetic operations. The advantages were enormous. Digital systems were immune to component drift, could be reprogrammed without changing hardware, and could achieve arbitrary precision by increasing word length. DSP rapidly displaced analog signal processing in most applications—telecommunications, audio, image processing, radar. But analog processing did not disappear entirely. It retained advantages in high-speed RF front-ends, sensor interfaces, and applications where power consumption or latency was critical. The competition between digital and analog was not a clean victory; it was a division of labor, with digital dominating baseband and analog persisting at the physical interface.
Once digital processing became the dominant infrastructure, engineers began to push beyond its original assumptions. Nonlinear Signal Processing (1980) challenged the linearity assumption that underlay both frequency-domain and state-space methods. Nonlinear filters—median filters, morphological operators, Volterra series—could suppress impulsive noise, preserve edges in images, and model systems that linear methods could not. Nonlinear signal processing did not replace linear frameworks; it coexisted with them, applied where linear assumptions broke down.
Multirate Signal Processing (1983) addressed the assumption that signals are sampled at a single uniform rate. By using decimation (downsampling) and interpolation (upsampling), engineers could process different parts of a signal at different rates, reducing computational load and enabling efficient filter banks. Multirate methods became essential for audio compression (MP3), subband coding, and wavelet transforms. They also deepened the connection between time-frequency analysis and digital processing: the discrete wavelet transform could be implemented as a multirate filter bank.
In 2006, Compressed Sensing (2006) overturned a fundamental assumption that had guided signal acquisition since Nyquist: that a signal must be sampled at twice its highest frequency to be perfectly reconstructed. Candès, Romberg, Tao, and Donoho showed that if a signal is sparse in some basis (i.e., most of its coefficients are zero), it can be recovered from far fewer measurements than the Nyquist rate requires. Compressed sensing did not reject frequency-domain analysis; it used it as a tool (the Fourier basis is one possible sparsity basis). But it shifted the focus from uniform sampling to the structure of the signal itself. The framework assumed sparsity and incoherence between the measurement and sparsity bases—assumptions that hold for many natural signals but not all.
Most recently, Data-Driven Signal Processing (2010) has begun to challenge the model-based paradigm that dominated from the 1960s onward. Instead of writing an explicit mathematical model of the signal (as in Kalman filtering) or assuming a known statistical distribution (as in Wiener filtering), data-driven methods learn representations directly from large datasets using deep neural networks. This framework assumes that enough training data is available to capture the signal's structure implicitly. It has achieved remarkable results in speech recognition, image restoration, and biomedical signal analysis. But it also raises new questions: data-driven models can be opaque, require enormous computational resources, and may fail when the test data differs from the training distribution. The relationship between data-driven and model-based frameworks is not yet settled; they coexist, with data-driven methods excelling where models are hard to specify and model-based methods remaining preferable where interpretability and guarantees are required.
No single framework has unified signals and systems. Instead, the subfield is a pluralistic toolkit. Frequency-domain analysis and linear systems theory remain the foundational language for LTI systems, taught to every undergraduate. Statistical signal processing provides the probabilistic backbone for detection and estimation. Adaptive and model-based methods are standard in control and navigation. Digital signal processing is the universal implementation platform, with analog processing confined to specialized niches. Nonlinear and multirate methods extend the digital toolkit. Compressed sensing has changed how we think about sampling, and data-driven methods are rapidly expanding what is possible.
Today's leading frameworks agree on one thing: the choice of representation matters enormously. They disagree on whether that representation should be hand-designed (frequency, time-frequency, state-space) or learned from data, whether it should assume linearity or embrace nonlinearity, and whether it should rely on explicit models or statistical regularities. The central tension that opened this history—how to represent a signal—remains unresolved, and that is precisely what keeps the field alive.