An impulse response is the output produced by a system when its input is a unit impulse, with the system initially at rest. For a linear time-invariant system, this single response completely determines the output for any admissible input through a mathematical operation called convolution. It provides a common description of phenomena as different as echoes in a room, transients in an electrical circuit, vibration in a structure, and the spreading of a pulse through a communications channel.
The concept is fundamental to signal processing, control theory, and electrical engineering. Its usefulness comes from connecting a physically interpretable time-domain measurement with frequency-domain descriptions, differential equations, and digital implementations. An impulse response can be calculated from a model or estimated experimentally, although no physical experiment can generate the ideal continuous-time impulse exactly. Its ability to characterize an entire system depends on linearity, time invariance, and the treatment of initial conditions; outside those assumptions, a more general description is necessary.
Unit impulses and the definition of the response
In continuous time, the ideal unit impulse is the Dirac delta, denoted . It is not an ordinary function with a finite height. Rather, it is a distribution defined by its action under an integral:
for suitable test functions . Informally, it represents an excitation concentrated at a single instant with unit area. A rectangular pulse of width and height approaches this ideal in the distributional sense as tends to zero. Its height grows without bound while its area remains one.
If a system is represented by an operator , its continuous-time impulse response is
with all independent initial conditions set to zero. This qualification separates the response caused by the input from motion or stored energy already present in the system. For example, a charged capacitor can produce a circuit output even when no external signal is applied. That output belongs to the initial-condition response, not to the zero-state impulse response.
In discrete time, the unit impulse is the ordinary sequence
The impulse response is correspondingly . Unlike its continuous-time counterpart, this impulse has a finite amplitude and can be represented exactly in digital arithmetic, subject to numerical precision. A single nonzero sample is therefore a direct way to test a digital linear system.
The units of an impulse response require care. Since the continuous-time delta has units of inverse time, has units of output divided by input and by time when ordinary physical input and output quantities are used in the convolution integral. For a force-driven mechanical system, a physical impulse has units of force multiplied by time. The response to an impulse of magnitude is when is normalized to a unit force impulse. In discrete-time convolution, the summation index is dimensionless, so no explicit time factor appears unless the sequence is being used to approximate a continuous-time integral.
The modern delta notation is associated with Paul Dirac’s 1930 treatment of quantum mechanics, although impulsive forcing and operational methods had been used earlier in mechanics and electrical theory. Distribution theory, developed systematically by Laurent Schwartz in the 1940s, supplied a rigorous framework for manipulating impulses, their derivatives, and responses containing instantaneous components.
Convolution and complete system characterization
The reason an impulse response is so powerful is that an arbitrary input can be decomposed into shifted, scaled impulses. In discrete time, this decomposition is exact:
Linearity means that the response to a weighted sum of inputs is the corresponding weighted sum of responses. Time invariance means that delaying an input merely delays its response. Consequently, the response to is , and the complete output is
The analogous continuous-time result follows from the impulse representation of an input:
which gives
These expressions require the signals and system to satisfy the relevant convergence or distributional conditions. They are not unrestricted formulas for every conceivable input. Within their domain, however, they turn the task of finding a response to each new waveform into a single convolution calculation.
For a causal system initially at rest, with an input beginning at time zero, the continuous-time expression reduces to
This equation also gives a physical interpretation: each earlier input contribution creates a shifted copy of the impulse response, and the present output is the sum of all surviving contributions. A reverberant room retains contributions from earlier sounds; a circuit with energy storage retains contributions from earlier electrical excitation.
Several interconnection rules follow directly. Two linear time-invariant systems in cascade have impulse response . Systems connected in parallel, with their outputs added, have response . The identity system has impulse response , and a pure continuous-time delay has response . Although scalar convolution is commutative, this does not imply that arbitrary physical components can always be interchanged: loading, operating-point changes, and interactions may invalidate the independent-system models.
An impulse response describes the zero-state response, not necessarily the entire observed output. In a linear system with nonzero initial conditions, the output is the sum of the convolution response and a zero-input response determined by those conditions. A complete model must retain both when previously stored energy or internal state matters.
Frequency response, transfer functions, and transform methods
The impulse response and the frequency response are two representations of the same linear time-invariant input–output relation. When the Fourier transform exists in the appropriate sense,
The convolution theorem then gives . A computation that requires integration over all earlier input values in the time domain becomes multiplication in the frequency domain. This relationship underlies spectral analysis, filter design, and many efficient numerical algorithms.
For a system whose impulse response is absolutely integrable, a complex sinusoid is an eigenfunction: its output is . Thus specifies the amplitude scaling and the phase shift. A real sinusoid is handled by combining positive- and negative-frequency components. For an unstable system, the familiar interpretation as a settled sinusoidal response can fail even when a formal algebraic expression can be written.
The transfer function extends this description through the Laplace transform:
For a causal system, a unilateral transform is often used to solve initial-value problems, with initial-condition terms handled explicitly. Under zero initial conditions, a constant-coefficient differential equation becomes an algebraic equation, and the transfer function is the output transform divided by the input transform. For rational transfer functions, poles describe natural exponential modes, while zeros describe frequencies or complex exponential inputs suppressed by the input–output relation.
The transform expression must be interpreted together with its region of convergence. For example,
corresponds to the causal response when , where is the unit step. The same algebraic fraction with a different region of convergence describes a different, left-sided response. A rational expression alone therefore does not always uniquely specify a bilateral transform.
In discrete time, the corresponding transform is the z-transform,
The discrete-time frequency response is obtained on the unit circle, , when that circle lies in the region of convergence. Sampling the continuous-time impulse response does not automatically produce a digital system with an identical frequency response. Sampling, reconstruction, scaling, and aliasing must be considered separately, especially when continuous-time frequency content extends beyond half the sampling rate.
Causality, stability, memory, and invertibility
Several important system properties can be read directly from an impulse response. A linear time-invariant system is causal if and only if
in continuous time, or for in discrete time. An output preceding the applied impulse would imply dependence on future input. Causality is essential for real-time physical operation, but not for every computational task: offline processing can use samples on both sides of the time being analyzed.
For an ordinary continuous-time impulse-response function, bounded-input bounded-output stability is equivalent to absolute integrability:
The discrete-time criterion is
These conditions ensure that every bounded input produces a bounded output. A finite direct-feedthrough term is also compatible with stability; in the more general distributional formulation, the impulse response must define a finite measure rather than necessarily an ordinary integrable function. By contrast, an ideal differentiator, with response , is not bounded-input bounded-output stable.
A memoryless scalar linear time-invariant system has , or . Any response extending beyond the instant of excitation indicates memory. The duration and decay of that response describe how long past inputs continue to influence the output. An exponentially decaying response has theoretically infinite memory even if its later values become negligible in practice.
For a causal, proper rational continuous-time transfer function, bounded-input bounded-output stability requires its uncancelled poles to lie strictly in the left half-plane. The discrete-time counterpart requires them to lie strictly inside the unit circle. These are input–output statements. An internal unstable mode that is unobservable or uncontrollable can be absent from a reduced transfer function, so external stability does not by itself establish internal stability of every realization.
An inverse system, when it exists, has impulse response satisfying . Algebraically this suggests , but existence of a formal reciprocal does not guarantee a causal, stable, or noise-tolerant implementation. A delay requires an advance for exact inversion, which is not causal. Zeros on or near the frequency range of interest can make inversion impossible or highly sensitive to measurement noise. A minimum-phase system has the zero locations needed for a stable causal inverse, subject also to properness, delay, and the precise continuous- or discrete-time definition being used.
Differential equations, state space, and physical examples
Many physical impulse responses follow from ordinary differential equations. A simple first-order low-pass model is
where is a time constant and is the static gain. Its causal impulse response is
For a series resistor–capacitor circuit with output measured across the capacitor, and . If and , the time constant is . After five time constants, the exponential has fallen to approximately of its initial value.
A mass–spring–damper system provides a second-order example:
For an underdamped system, define , , and . The displacement response per unit force impulse is
An impulse changes momentum instantaneously: a force impulse of magnitude produces a velocity jump , while displacement remains continuous. The subsequent oscillation reveals natural frequency and damping. Such responses are central to structural vibration testing, machinery diagnostics, and the design of mechanical isolation.
A general finite-dimensional state-space representation is
Its causal matrix-valued impulse response is
Here the step function in the first term enforces causality, while represents instantaneous feedthrough. The matrix exponential describes the evolution of internal modes. With multiple inputs and outputs, each entry gives the response at output to an impulse at input , with the other inputs set to zero.
The corresponding discrete-time model,
has and for . These coefficients, often called Markov parameters, connect experimentally measured responses to internal models. The impulse response is therefore not limited to a graphical trace: it can encode the same dynamics as a set of differential equations or a state-space model, while emphasizing only externally visible behavior.
Digital filters and computational implementation
A digital filter is frequently specified directly by its impulse-response samples. A finite impulse response filter, or FIR filter, has only finitely many nonzero coefficients. For a causal filter of length ,
Its implementation stores recent input samples, multiplies them by the coefficients, and adds the products. Any finite set of finite-valued coefficients gives an absolutely summable response and hence bounded-input bounded-output stability.
A moving average illustrates the connection between coefficient values and behavior. For over , each output is the mean of the latest inputs. The response smooths rapid fluctuations but also modifies desired signal components and introduces delay. More selective FIR filters use carefully designed coefficients rather than equal weights. Real symmetric or antisymmetric coefficients produce a linear phase response, apart from phase discontinuities at zeros, with a nominal group delay of samples.
An infinite impulse response filter, or IIR filter, generally uses a recursive difference equation:
With the denominator normalized to have constant coefficient one,
Feedback allows a small number of coefficients to produce a long decaying response. This can achieve sharp frequency selectivity with less arithmetic than a comparable FIR filter, but stability and numerical sensitivity require attention. A recursive implementation does not necessarily imply an infinite response in every special case, since cancellations can reduce the effective transfer function.
Direct convolution with an -sample response requires roughly multiplications per output sample. For long responses, block methods based on the fast Fourier transform can be more efficient. Overlap-add and overlap-save divide the input into blocks, multiply spectra, and reconstruct the linear convolution. Sufficient padding or overlap is essential because an unmodified discrete Fourier transform computes circular rather than linear convolution.
Coefficient quantization and finite arithmetic can alter the implemented response. In IIR filters, small coefficient errors can move poles, and rounding can generate persistent low-level oscillations called limit cycles. In FIR filters, coefficient errors alter gain and frequency selectivity but cannot create an infinite tail in a nonrecursive realization. Practical implementations must also manage overflow, scaling, latency, and block boundaries; these are part of realizing a desired impulse response rather than properties of the ideal mathematical sequence alone.
Experimental measurement and system identification
An impulse response can be measured by applying a short excitation and recording the output, but a real pulse has finite duration and amplitude. The measured output is the true impulse response convolved with that pulse. Treating it directly as is accurate only over a frequency range in which the excitation spectrum is approximately constant after normalization. A very narrow pulse covers a wide bandwidth but can require an impractically high peak amplitude, exceed an actuator limit, or drive the system outside its linear regime.
Mechanical testing often uses an instrumented impact hammer. A force sensor measures the actual excitation, while accelerometers or other transducers record the response. Because the input is measured rather than assumed ideal, a frequency-response estimate can account for the hammer pulse shape. Hammer-tip stiffness affects excitation bandwidth: a harder tip generally produces a shorter impact and more high-frequency content. The sensors, fixtures, and data-acquisition system also contribute dynamics that must be distinguished from the test object.
An alternative is system identification using broadband excitation. White noise, pseudorandom binary sequences, and swept sinusoids distribute energy over a frequency band without concentrating all of it into a single instant. Under suitable assumptions, input–output cross-correlation identifies the impulse response. If , then
For a zero-mean white input with variance , , giving , provided additive output noise is uncorrelated with the input.
Swept-sine techniques are widely used in acoustics. The response to a known sweep is processed by deconvolution to estimate the impulse response. Exponential sweeps can also separate contributions from different orders of harmonic distortion in the recovered time record under suitable conditions. This is useful because a loudspeaker driven too strongly may generate nonlinear products that should not be mistaken for a linear room response.
Every estimate has limits set by bandwidth, record duration, signal-to-noise ratio, and system variability. A short record truncates slow decay; a long record may include changes in the environment or operating conditions. Spectral division becomes unreliable where the excitation is weak. Averaging reduces uncorrelated noise but does not remove systematic errors, and windowing trades reduced leakage or unwanted reflections against altered spectral resolution. Reliable interpretation therefore depends on calibration, timing alignment, repeatability, and a clear definition of the input and output measurement locations.
Acoustics, communications, and control applications
In room acoustics, an impulse response describes sound propagation from a particular source position to a particular receiver position. It typically contains a direct arrival, discrete early reflections, and a denser reverberant tail. Its initial delay reflects propagation time, while later features depend on room geometry, absorption, scattering, and source and microphone directivity. Moving either endpoint can change the response, especially at high frequencies, where wavelengths are short relative to ordinary position changes.
The squared room impulse response can be integrated backward in time to form an energy-decay curve, a method introduced by Manfred Schroeder in 1965. Decay slopes are used to estimate reverberation time, conventionally associated with a 60-decibel reduction in level. Because actual measurement noise may prevent observation of the full decay, shorter ranges such as 20 or 30 decibels are often used for extrapolation. Impulse responses also support convolution reverberation in sound recording: dry audio is convolved with a measured or synthesized response to reproduce selected acoustic characteristics.
In telecommunications, a multipath channel is often modeled as
where each path has a delay and a gain , possibly complex in a baseband representation. Delayed copies of transmitted symbols can overlap later symbols, causing intersymbol interference. Channel estimation, equalization, and waveform design use the response to compensate for or accommodate this spreading. In multicarrier systems, a sufficiently long cyclic prefix makes the useful part of a block behave as a circular convolution, enabling frequency-domain equalization under the assumed channel model.
In control engineering, impulse-response shape reveals delay, damping, oscillation, and dominant time scales. A slowly decaying tail indicates persistent dynamics, while an oscillatory tail suggests complex poles. Model-based controllers can use convolution or an equivalent state-space model to predict future outputs. In process control, finite impulse or step-response models are particularly useful for representing systems with transport delays and multiple interacting inputs and outputs.
Impulse responses also occur in radar and imaging, where finite bandwidth and aperture spread an idealized point target into a finite response. A matched filter uses a conjugated, time-reversed waveform to maximize the output signal-to-noise ratio for a known signal in additive white noise. This does not automatically recover an ideal delta response: the resulting peak has a width and sidelobe structure determined by the waveform. Across these applications, the response is meaningful only for the specified channel, operating conditions, bandwidth, and input–output variables.
Time-varying, nonlinear, and spatial generalizations
A single one-dimensional impulse response is insufficient for a general linear time-varying system. Instead, an impulse applied at time produces a response depending separately on the observation time and excitation time:
Time invariance is the special case . For a causal time-varying system, the kernel vanishes when . Moving radio channels, periodically switched circuits, and systems whose material properties change during observation may require this two-time description. A slowly varying system is sometimes approximated by a sequence of locally time-invariant models, but that approximation depends on the relative time scales of excitation, response decay, and change.
Nonlinear systems violate the superposition principle that makes convolution sufficient. Doubling an input need not double its output, and combining two inputs can generate interactions absent from the separate responses. A nonlinear system can still have an observable response to a pulse, but that response generally depends on pulse amplitude and shape and cannot predict arbitrary inputs through ordinary convolution.
One extension is the Volterra series, which represents a nonlinear input–output relation using a hierarchy of kernels. In schematic continuous-time form,
The first-order kernel is the linear impulse response; higher-order kernels describe interactions between input values at different times. Such expansions have convergence and modeling limitations, and their dimensionality grows rapidly with order. Another approach is local linearization around an operating point, producing an impulse response valid only for sufficiently small perturbations about that point.
The concept also extends from time to space. In a linear shift-invariant imaging system, a point object produces a point spread function. An image is then represented as a spatial convolution of the object with that function. Real optical systems can be shift-variant, in which case the response depends on both source and observation positions rather than only their separation.
For differential equations, the related Green’s function describes the response to a delta source while incorporating specified boundary or initial conditions. It can depend on several spatial coordinates and time. The causal impulse response of a linear dynamical system is a particular instance of this broader construction. The shared principle is the reconstruction of a distributed response by superposing responses to localized sources; the precise kernel, admissible inputs, and boundary conditions determine whether that reconstruction is valid.