A linear time-invariant system, commonly abbreviated LTI system, is a mathematical model of a process whose input–output relationship satisfies two conditions: responses obey superposition, and shifting an input in time shifts its output by the same amount without otherwise changing it. These properties make it possible to predict the response to a broad class of inputs from a single characteristic function, the system’s impulse response. LTI models are fundamental to signal processing, control theory, electrical engineering, acoustics, and mechanical dynamics.
The importance of LTI theory lies in the connection it establishes between several descriptions of the same system. Time-domain convolution, differential or difference equations, frequency response, transfer functions, and state-space equations reveal complementary aspects of system behavior. Although physical systems generally depart from exact linearity or time invariance, LTI models often describe them accurately within specified operating ranges. Their mathematical structure supports systematic analysis of filtering, resonance, transient response, feedback, and stability.
Linearity and time invariance
A system may be represented by an operator that maps an input signal to an output signal . It is linear if, for admissible inputs and scalar coefficients ,
This combines additivity and homogeneity, together called the superposition principle. Doubling an input doubles its response, and the response to a sum of inputs is the sum of their individual responses. In particular, a linear input–output operator maps the zero input to the zero output. A device described by is therefore affine, not linear, because of its constant offset.
Time invariance means that the system’s behavior does not depend on the absolute time at which an experiment occurs. For continuous-time signals, if , then
for every admissible shift . For discrete-time signals, the corresponding shifts are integer multiples of the sampling interval. Time invariance does not mean that signals remain constant: an LTI system can produce oscillations, delays, and decaying transients. It means that its governing input–output relationship does not change with the clock.
The two properties are independent. The system is linear but time varying, since its gain depends explicitly on time. The system is time invariant but nonlinear. A fixed delay, , satisfies both properties. A modulator multiplying an input by a prescribed sinusoid is linear in that input, but generally time varying.
Initial conditions require particular care. A linear differential equation with a fixed nonzero initial state generally produces an input–output map that is affine rather than linear: its output includes a zero-input response that does not scale with the applied input. Consequently, statements about an LTI input–output system normally refer to its zero-state response, with stored energy initially absent, or explicitly treat the initial state as an additional independent argument.
Linearity and time invariance are always relative to a chosen model and operating range. An electronic amplifier may be approximately LTI for small signals but become nonlinear when its output clips. A mechanical structure may have effectively constant parameters during a short measurement but change behavior as its temperature varies. Such qualifications determine where an LTI approximation is meaningful.
Impulse response and convolution
The central time-domain description of an LTI system follows from decomposing an input into shifted elementary impulses. In continuous time, the ideal impulse is the Dirac delta distribution, written . Formally,
If , time invariance identifies the response to as . Linearity then permits the weighted responses to be combined. Under appropriate continuity and convergence assumptions, the result is the convolution integral:
The discrete-time derivation uses the unit sample , which equals one at and zero elsewhere. Since
the output is
These representations show why one impulse response can characterize an entire LTI input–output system. Each input value excites a scaled, shifted copy of the same response; the output combines all those contributions. For a causal system, only present and past inputs contribute.
Consider the causal first-order system
Its impulse response is
where is the unit step. Convolving this response with a unit step gives for . The parameter is the time constant: after one time constant the response has reached approximately of its final value. This behavior appears in resistor–capacitor circuits, thermal equilibration models, and viscous mechanical systems.
Convolution also describes system interconnection. Two scalar LTI systems in cascade have an overall impulse response ; two systems in parallel have response . Because scalar convolution is commutative, reversing the order of ideal scalar LTI blocks does not change their zero-state input–output relationship. Physical loading between actual devices can invalidate the assumption that the blocks retain their separate characteristics.
For multiple-input, multiple-output systems, the impulse response becomes a matrix of functions, with one response for each input–output pair. Convolution then includes matrix multiplication, which is generally not commutative. The scalar cascade property therefore does not extend unchanged to multivariable systems.
Differential equations, difference equations, and physical models
Many finite-dimensional continuous-time LTI systems are governed by constant-coefficient linear differential equations:
The coefficients determine the dynamics and remain independent of time. Linearity follows because neither the dependent variables nor their derivatives are multiplied together or passed through nonlinear functions. Constant coefficients distinguish this class from linear time-varying equations, although boundary conditions and the definition of the input–output map must also be considered.
A series circuit consisting of a resistor and a capacitor, with output measured across the capacitor, satisfies
Its time constant is , measured in seconds. If and , then . This circuit suppresses rapid input variations because changing capacitor voltage requires charge transfer through the resistor. Ideal component models make the circuit exactly LTI; temperature dependence, dielectric effects, and voltage-dependent capacitance introduce departures from that model.
A mass–spring–damper system provides a second standard example:
Here is displacement, is applied force, is mass, is viscous damping, and is spring stiffness. Its undamped natural angular frequency is , and its damping ratio is . For , free motion oscillates with a decaying envelope; critical damping occurs at . The model assumes a linear spring and damping force proportional to velocity, rather than dry friction or displacement-dependent stiffness.
Discrete-time finite-dimensional systems are commonly described by constant-coefficient difference equations:
Previous outputs and present or previous inputs determine each new output. Such equations are directly implementable in digital processors, although finite numerical precision adds effects absent from the ideal equations.
Not every LTI system has a finite-order differential or difference equation. An ideal continuous-time delay has impulse response and cannot be represented exactly by a finite-dimensional ordinary differential equation. Distributed systems, including wave propagation and heat diffusion, can also be LTI while requiring infinitely many modes. Finite-dimensional models approximate their behavior over selected time and frequency ranges.
Transform methods and transfer functions
Transforms convert convolution into multiplication, often making system equations easier to analyze. For continuous-time systems, the bilateral Laplace transform of an impulse response is
where . Wherever the relevant transforms and convolution theorem apply,
The function is the system’s transfer function. For a constant-coefficient differential equation under zero initial conditions, it is typically a rational function,
Roots of the numerator are transmission zeros in the scalar rational case, while roots of the denominator that remain after cancellation are poles. Poles describe natural modes visible in the input–output relationship; zeros describe frequencies or complex exponential modes at which transmission vanishes. Repeated poles can produce terms such as , rather than simple exponentials alone.
For a discrete-time system, the corresponding z-transform is
Applying it to a difference equation gives
The variable captures discrete-time exponential behavior just as captures continuous-time exponential behavior.
A transform expression must be interpreted together with its region of convergence. The algebraic function , for , can describe either the causal response , with , or the anticausal response , with . These systems have different causality and stability properties despite sharing the same algebraic formula.
Transfer functions describe zero-state input–output behavior, not every aspect of a physical realization. Nonzero initial conditions contribute additional terms when differential equations are transformed. Furthermore, internal modes can disappear from a reduced transfer function through pole–zero cancellation, even though those modes remain in the underlying state equations.
For finite-dimensional causal systems without input differentiation, rational transfer functions are proper: the numerator degree does not exceed the denominator degree. A strictly proper transfer function has lower numerator degree and no instantaneous direct-feedthrough term. Ideal differentiators have improper transfer functions and are useful mathematical models, but their unbounded high-frequency gain prevents exact realization as ordinary finite-bandwidth physical devices.
Frequency response and sinusoidal behavior
Complex exponentials are eigenfunctions of LTI convolution operators. If an input is , and the relevant integral converges, then
The system changes only the exponential’s complex amplitude. Setting gives the frequency response . For a real system driven by a sinusoid, the steady-state output has the same frequency:
produces
This statement concerns sinusoidal steady state. If excitation begins at a finite time, a transient response is generally present as well. For an asymptotically stable system, the transient decays.
The Fourier transform expresses a broad class of signals as combinations of complex exponentials. An LTI system acts on each frequency component independently, multiplying its spectrum by . In exact steady-state operation, it does not generate harmonics or shift one input frequency to another. New frequencies produced by clipping, switching, or frequency conversion therefore indicate behavior outside an ordinary LTI model, apart from effects associated with transients or finite observation windows.
Magnitude and phase are often displayed in a Bode plot, with logarithmic frequency and magnitude expressed as decibels. For the first-order resistor–capacitor low-pass circuit,
At , the magnitude is , approximately decibels below its low-frequency value, and the phase is . At frequencies much higher than , the magnitude falls approximately decibels per decade.
Frequency-dependent phase can alter waveform shape even when the magnitude response is nearly flat. Group delay is defined, where the phase is differentiable, by
A pure delay has constant group delay and unit magnitude. Other systems delay different frequency components by different amounts, potentially spreading pulses or distorting modulated signals.
In discrete time, the frequency response is evaluated on the unit circle, , and is periodic in normalized angular frequency with period . This periodicity reflects the fact that discrete-time complex exponentials whose frequencies differ by are identical. Continuous-time and discrete-time frequency descriptions must therefore be related through the sampling interval rather than treated as interchangeable.
Causality, stability, and inversion
An input–output system is causal if its output at a given time depends only on present and past inputs. For an LTI convolution system, causality is equivalent to
in continuous time, or for in discrete time. A pure positive delay is causal; a time advance is not. Noncausal processing remains useful for stored data, where future samples are already available relative to the point being processed. For example, symmetric smoothing filters can use samples on both sides of a time index.
Bounded-input bounded-output stability, usually called BIBO stability, means that every bounded input produces a bounded output. For an ordinary continuous-time convolution kernel, the standard condition is
For discrete time, it is
These conditions prevent the system from accumulating bounded excitation into an unbounded output. Generalized impulse responses containing delta functions require a broader formulation: a finite direct-feedthrough term is compatible with BIBO stability even though it is not an ordinary integrable function.
For a causal rational transfer function after exact cancellations, continuous-time BIBO stability requires all poles to lie strictly in the left half of the complex -plane. In discrete time, all poles must lie strictly inside the unit circle. A continuous-time integrator, , is not BIBO stable because a constant bounded input produces a ramp. An undamped oscillator can likewise produce an unbounded response to a bounded resonant input.
Internal stability is distinct from input–output stability. A system can contain an unstable internal mode absent from its reduced transfer function because that mode is unobservable or unreachable from the chosen input. Asymptotic internal stability of a finite-dimensional continuous-time state model requires every eigenvalue of its state matrix to have negative real part. The discrete-time requirement is that every eigenvalue have magnitude less than one.
Inversion introduces additional restrictions. Formally, a scalar inverse has transfer function , but it may be noncausal, unstable, or excessively sensitive to noise. A delay requires an advance to undo it exactly. A zero in the right half-plane becomes an unstable pole of a continuous-time causal inverse; a discrete-time zero outside the unit circle creates the analogous difficulty. Such zeros are called nonminimum-phase zeros. Even when a stable causal inverse exists, strong attenuation in the original system can require large inverse gain, amplifying measurement noise and numerical errors.
State-space models and feedback interconnections
A finite-dimensional continuous-time LTI system can be represented by state-space equations:
Here is the internal state, the input vector, and the output vector. The constant matrices specify state evolution, input coupling, output measurement, and direct feedthrough. The matrix formulation handles multiple interacting inputs and outputs without requiring a separate high-order scalar equation for every signal.
Given an initial state , the state evolves as
The first term is the zero-input response; the integral is the zero-state response. For zero initial state, the transfer matrix is
For positive time, its ordinary impulse-response component is , while direct feedthrough contributes .
A realization is controllable if its state can be driven between arbitrary states by a suitable input over a finite interval, and observable if the initial state can be inferred from the output over a finite interval when the input is known. A realization that is both controllable and observable is minimal. Nonminimal realizations can contain internal modes absent from the transfer function, explaining why a reduced input–output model may conceal internal instability.
LTI models also make feedback analysis algebraically tractable. For scalar plant and controller in a standard negative-feedback arrangement, the closed-loop transfer function from reference to output is
The denominator determines candidate closed-loop poles, but internal stability requires examining the complete interconnection rather than only one reduced transfer function. Algebraic loops involving direct-feedthrough terms must also be well posed.
A PID controller combines proportional, integral, and derivative actions. With constant gains and an appropriate linear derivative filter, it can be incorporated into an LTI model. Saturation, rate limits, dead zones, and mode switching make the implemented loop nonlinear or time varying. Integral windup, for example, arises through the interaction of an integrating controller with actuator saturation.
Frequency-domain stability tests and state-space methods offer different views of the same closed-loop dynamics. Robust control extends this framework by examining how uncertainty in model parameters or unmodeled dynamics affects stability and performance. An LTI nominal model is often the starting point, not a claim that the physical plant is known exactly.
Sampling, digital filters, and numerical realization
Digital LTI systems operate on sequences, but many are embedded in continuous physical processes. Sampling theory establishes how a band-limited continuous-time signal can be represented by uniformly spaced samples. If its highest frequency is strictly below , where is the sampling frequency, ideal reconstruction is possible under the theorem’s assumptions. Sampling does not itself make an arbitrary analog signal band limited; practical systems use an anti-aliasing filter to suppress frequencies that would otherwise fold into the sampled band.
A continuous-time LTI plant sampled uniformly with a zero-order hold at its input has the discrete state model
where, for sampling interval ,
These equations exactly describe the state at sampling instants when each input value is held constant during its interval. The continuous evolution between samples remains important. The complete hybrid arrangement is not generally time invariant under arbitrary continuous-time shifts, even though its sample-to-sample model is discrete-time LTI.
Finite impulse response filters have only finitely many nonzero impulse-response coefficients:
Their BIBO stability follows directly from the finite coefficient sum. Symmetric or antisymmetric real coefficients can provide linear phase, subject to the filter type and frequencies where phase is defined. Recursive filters usually have infinite impulse responses and can achieve sharp frequency selectivity with fewer coefficients, but their poles must satisfy stability constraints.
Numerical implementation can depart from the mathematical model. Rounding and quantization are nonlinear operations. In recursive fixed-point filters, they can produce limit cycles even with zero input, while overflow can cause much larger deviations. Coefficient quantization shifts poles and zeros, potentially changing the behavior of a filter designed to be close to instability. Cascades of second-order sections are therefore commonly used instead of a single high-order polynomial realization.
Long convolutions can be computed efficiently using a fast Fourier transform. Because multiplication of discrete Fourier transforms corresponds to circular convolution, linear convolution requires appropriate zero padding or block methods such as overlap-add and overlap-save. For finite sequences of lengths and , a transform length of at least avoids circular wraparound in a direct zero-padded calculation. This computational technique changes how the output is calculated, not the underlying LTI input–output relationship.
Modeling, identification, and applications
An LTI model can be derived from physical laws or estimated from observations. Physical derivation uses conservation relations and constitutive equations, such as force balance, electrical circuit laws, or linear heat transfer. Experimental system identification instead estimates dynamics from measured input and output. Impulses, steps, swept sinusoids, broadband random inputs, and pseudorandom sequences excite different aspects of the response. Their usefulness depends on bandwidth, noise, physical constraints, and how fully the input excites the relevant modes.
A physical impulse is always finite in duration and amplitude. It approximates an ideal impulse only over a frequency range in which its spectrum is sufficiently broad. Step tests are easier to produce in many processes, but differentiating a measured step response to estimate an impulse response amplifies noise. Frequency-response measurements can estimate gain and phase directly, while parametric identification fits differential equations, difference equations, or state-space models.
In telecommunications, an approximately LTI channel models attenuation, delay, and multipath propagation over a suitable observation interval. An equalizer attempts to compensate for channel distortion, subject to noise amplification and nonminimum-phase limitations. A mobile radio channel may vary too rapidly for one fixed model; it is then represented as time varying or as a succession of approximately stationary models.
In acoustics and sound recording, a room impulse response describes reflections and reverberation between a specified source position and receiver position. Convolution can reproduce that response for another recording. The approximation assumes fixed geometry, approximately linear sound propagation, and sufficiently constant environmental conditions. Moving the source or receiver changes the relevant system, while loudspeaker distortion can introduce nonlinear behavior.
Mechanical vibration testing uses transfer functions between applied force and measured displacement, velocity, or acceleration. Resonant peaks reveal structural modes, and their widths provide information about damping. In process control, heating, mixing, flow, and pressure dynamics are often approximated by low-order models with delays. These models support controller analysis while omitting spatial detail or fast dynamics that have little influence in the intended operating range.
Linearization explains why LTI models apply to many nonlinear plants. For smooth dynamics , expansion around a fixed equilibrium gives
The derivatives are constant at that equilibrium. Linearization around a changing trajectory instead generally produces a linear time-varying model. Large excursions, hysteresis, friction, saturation, aging, and changing operating conditions restrict the accuracy of a fixed LTI approximation; model validation therefore concerns the intended amplitudes, frequencies, and duration of use.
Historical development and theoretical boundaries
LTI theory emerged from several mathematical and engineering traditions rather than a single invention. Eighteenth-century work on linear differential equations established methods for superposition and natural modes. Joseph Fourier’s 1822 treatise on heat developed trigonometric representations that became central to frequency-domain analysis. The connection between linear dynamics and decomposition into elementary functions gradually supplied a common language for otherwise different physical systems.
During the late nineteenth century, Oliver Heaviside developed operational techniques for electrical circuits and transmission lines. His manipulation of differentiation operators anticipated much of the algebra later expressed systematically through Laplace transforms. Work on long-distance telegraphy and telephone transmission made attenuation, dispersion, and transient behavior practical engineering concerns. Distributed transmission lines also demonstrated that useful linear systems need not be finite dimensional.
Feedback theory became increasingly important with electronic amplification. Harold Stephen Black conceived negative-feedback amplification in 1927. Harry Nyquist’s 1932 paper on regeneration established a frequency-domain stability criterion that related closed-loop behavior to the open-loop frequency response. Hendrik Bode’s work developed logarithmic gain-and-phase techniques, consolidated in his 1945 book on network analysis and feedback amplifier design. These methods made transfer functions a practical basis for engineering design.
Digital computation and sampled-data control expanded the discrete-time side of the theory in the mid-twentieth century. Rudolf Kalman’s 1960 work on filtering and prediction helped establish state-space methods as central tools for systems with multiple variables and stochastic disturbances. James Cooley and John Tukey’s 1965 fast Fourier transform paper accelerated practical spectral analysis and convolution. These developments connected continuous-time circuit and control methods with programmable digital processing.
Several distinctions continue to define the theory’s scope. A deterministic LTI operator can act on random signals without becoming time varying, and stationarity is a statistical property of a signal rather than a synonym for system time invariance. A spatial imaging system may be called linear shift invariant when translations in position, rather than time, commute with its operation. Adaptive filters with changing coefficients and gain-scheduled controllers generally fall outside a single fixed LTI description.
Infinite-dimensional systems require additional mathematical care. Delays, distributed media, ideal differentiators, and generalized impulse responses cannot all be treated as ordinary finite-order rational transfer functions. Function spaces, convergence regions, and operator domains determine when convolution and transform arguments are valid. At this boundary, LTI theory connects to functional analysis and partial differential equations, while retaining the defining requirements of superposition and invariance under shifts.