Loading field map
Signals and systems is the branch of electrical engineering that studies how information-bearing quantities—voltages, currents, light intensities, sound pressures, and countless other physical variables—change over time or space, and how devices or algorithms transform those quantities. Its central concern is not any particular physical medium but the abstract relationship between an input and an output: given a signal entering a system, what signal emerges, and what general laws govern that transformation? The field provides the mathematical language and analytical tools used across communications, control, audio processing, image analysis, and virtually every other area where engineers manipulate information carried by physical quantities.
A signal is a function that conveys information. In the most common formulation, a signal is a function of one independent variable—usually time—mapping each instant to a real or complex number. A continuous-time signal is defined for every instant of time; a discrete-time signal is defined only at equally spaced instants, typically produced by sampling a continuous signal at a fixed rate. Signals can also depend on more than one variable: an image is a signal of two spatial variables, and a video is a signal of two spatial variables plus time. The field treats all these cases with the same underlying mathematics, though the one-dimensional time case dominates introductory treatments.
A system is any process that maps an input signal to an output signal. A resistor, a digital filter running on a microprocessor, a room that adds echoes to a sound, and an algorithm that sharpens a photograph are all systems in this sense. The field's power comes from classifying systems by their mathematical properties and then deriving general consequences from those properties. The most important classification is linearity and time-invariance. A system is linear if scaling the input scales the output by the same factor and if the response to a sum of inputs is the sum of the responses to each input individually. A system is time-invariant if shifting the input in time shifts the output by the same amount without otherwise changing it. Systems that are both linear and time-invariant—called LTI systems—form the backbone of the field because they admit a complete and elegant theory.
The central question of signals and systems is: for a given class of signals and a given class of systems, can we predict the output efficiently and understand what the system does to the information in the signal? The answer for LTI systems is the theory of convolution and the Fourier transform, which together constitute the field's core achievement.
For an LTI system, the entire behavior is captured by its impulse response: the output produced when the input is a single, infinitely brief pulse (the Dirac delta function in continuous time, the unit impulse in discrete time). Because of linearity, any input signal can be viewed as a sum of many such pulses, each scaled by the signal's value at its own instant. Because of time-invariance, the response to each pulse is just the impulse response shifted to that instant. Adding all these responses together gives the output as a sum of shifted, scaled copies of the impulse response—an operation called convolution. In continuous time this is an integral; in discrete time it is a sum. Convolution completely solves the input-output problem for LTI systems: given the impulse response, the output for any input is determined.
The second pillar is the Fourier transform, which represents a signal not as a function of time but as a function of frequency. The Fourier transform decomposes a signal into a continuum of pure sinusoidal components, each with its own amplitude and phase. For an LTI system, a sinusoidal input always produces a sinusoidal output at the same frequency, with the amplitude scaled and the phase shifted by amounts that depend only on the frequency. The function describing this scaling and shifting is the frequency response of the system. The Fourier transform of the impulse response equals the frequency response, so the two representations—time-domain and frequency-domain—are two views of the same object. The practical consequence is profound: instead of computing a convolution integral for every input, one can transform the input, multiply by the frequency response, and transform back. More importantly, the frequency view reveals what the system does to the information in the signal: which frequency components it passes, which it attenuates, and which it shifts in time.
The Laplace transform (for continuous time) and the z-transform (for discrete time) generalize the Fourier transform by introducing a complex variable. These transforms turn differential and difference equations into algebraic equations, making system analysis far more tractable. They also expose the system's poles and zeros—the complex frequencies where the system's response becomes infinite or zero—which determine stability and shape the frequency response. A system is stable if all its poles lie in the left half of the complex plane (continuous time) or inside the unit circle (discrete time). The pole-zero view is the working language of filter design and control system analysis.
The field's modern form was shaped by the rise of digital computation. Discrete-time signals and systems were studied mathematically long before digital computers were common, but the practical importance of the discrete case exploded when microprocessors became cheap enough to process signals in real time. The key enabling result was the fast Fourier transform (FFT), an algorithm that computes the discrete Fourier transform in far fewer operations than a direct calculation would require. The FFT made frequency-domain processing of long signals practical and turned digital filtering from a mathematical curiosity into a ubiquitous engineering tool.
Discrete-time theory is not merely a sampled version of continuous-time theory; it has its own phenomena. The most important is aliasing: when a continuous signal is sampled, frequencies that differ by multiples of the sampling rate become indistinguishable. If the signal contains energy at frequencies above half the sampling rate—the Nyquist frequency—that energy is folded down into lower frequencies, corrupting the sampled signal irreversibly. The sampling theorem states that a signal whose frequencies are all below the Nyquist frequency can be reconstructed exactly from its samples. This theorem, associated with Harry Nyquist and Claude Shannon, is the theoretical guarantee that digital representation of analog signals is possible without loss of information, provided the signal is band-limited and the sampling rate is high enough.
The discrete-time framework also introduced the digital filter as a design object in its own right. Finite impulse response (FIR) filters have an impulse response that ends after a finite number of samples; they are always stable and can have exactly linear phase, meaning they delay all frequency components by the same amount. Infinite impulse response (IIR) filters have feedback and can achieve sharper frequency selectivity with fewer computations, but they require careful stability analysis. The design of these filters—choosing coefficients to meet specifications on passband ripple, stopband attenuation, and transition width—is a mature engineering discipline with well-established methods.
A second major tradition within signals and systems focuses not on the input-output relationship but on the internal condition of the system. The state-space representation describes a system by a set of first-order differential or difference equations governing a collection of state variables, together with an output equation that maps the state and input to the output. This representation, developed in the mid-twentieth century, is more general than the impulse-response description: it handles systems with multiple inputs and multiple outputs, time-varying systems, and nonlinear systems. It also connects naturally to computation, since the state equations can be simulated step by step.
The state-space view is the natural setting for control theory, which asks not just what a system does but how to make it do something desired. A controller measures the output, compares it to a reference, and adjusts the input to drive the error toward zero. The analysis of such closed-loop systems—stability, responsiveness, robustness to modeling error—is a central engineering activity. While control theory is often treated as a separate field, its mathematical foundations are the same as those of signals and systems, and the two share the core concepts of transfer functions, poles, stability, and frequency response.
The state-space approach also led to deeper theoretical results. Observability asks whether the internal state can be deduced from measurements of the output; controllability asks whether the state can be driven to any desired value by some choice of input. These concepts, formalized in the 1960s, revealed that the internal structure of a system matters in ways that the input-output view cannot capture. Two systems with identical input-output behavior can differ in controllability or observability, with practical consequences for control design.
Many signals are not deterministic but random. Noise in electronic circuits, interference in communication channels, and speech or music all have unpredictable components. The field therefore includes a substantial probabilistic branch. A random signal is a function whose values are described by probability distributions rather than by definite numbers. The relevant tools are correlation functions, which measure how much a signal at one time resembles itself at another time, and power spectral density, which describes how the signal's power is distributed across frequency. For LTI systems, the relationship between input and output statistics is clean: the output power spectral density equals the input power spectral density multiplied by the squared magnitude of the frequency response. This result is the foundation of noise analysis in every electronic system.
When the statistics of the signal or the system are unknown, engineers use adaptive filters, which adjust their own coefficients based on the incoming data. The most widely used algorithm, least mean squares (LMS), updates the filter coefficients in the direction that reduces the mean squared error between the filter output and a desired response. Adaptive filters are used for echo cancellation in telephones, noise reduction in headphones, channel equalization in modems, and many other applications where the environment changes or is unknown in advance. This adaptive tradition extends the field from analyzing fixed systems to designing systems that learn from their inputs.
Contemporary signals and systems is a mature field whose core theory is settled and taught uniformly across the world. The LTI framework, the Fourier and Laplace/z transforms, sampling theory, and state-space methods constitute a fixed body of knowledge that every electrical engineer learns. The field's continuing evolution comes from three directions.
First, the scale and speed of digital processing have made possible systems that were once only theoretical. Real-time audio and video processing, software-defined radio, and sensor arrays with hundreds of elements all rely on the classical theory but push its computational limits. The theory of multirate signal processing—changing the sampling rate of a signal by rational factors—has become practically important in audio conversion and data compression. Wavelet transforms, which decompose signals into components localized in both time and frequency, address a limitation of the Fourier transform: it reveals which frequencies are present but not when they occur. Wavelets and related time-frequency methods are now standard tools for analyzing non-stationary signals such as speech, seismic data, and biomedical recordings.
Second, the field has absorbed ideas from statistics and information theory. Compressed sensing showed that a signal with few nonzero coefficients in some basis can be recovered from far fewer measurements than the Nyquist rate would require, provided the measurements are chosen appropriately. This result, developed in the 2000s, has influenced imaging, radar, and data acquisition. Sparse signal processing more broadly seeks representations of signals using few basis functions, with applications in compression and denoising.
Third, the boundary between signals and systems and machine learning has become porous. Neural networks can be viewed as cascades of simple nonlinear systems, and the training of a network is an optimization problem over the parameters of a very large system. Conversely, classical signal processing concepts—filtering, spectral analysis, and system identification—are used to design and interpret machine learning models. The relationship is not a replacement of one tradition by another but a convergence: the mathematical habits of signals and systems—thinking in terms of frequency, linearity, stability, and uncertainty—remain valuable even when the systems in question are learned from data rather than designed from first principles.
The field's durability lies in its abstraction. By stripping away the physical details of specific devices and focusing on the mathematical relationship between input and output, signals and systems provides tools that apply equally to a mechanical oscillator, an electronic filter, a radio channel, and a neural network. Its central ideas—linearity, superposition, frequency, sampling, stability—are not tied to any particular technology and have survived every change in the underlying hardware. An engineer who understands convolution and the Fourier transform can approach almost any problem involving information carried by a physical quantity, whether the implementation is analog, digital, optical, or biological.