Infinite impulse response¶

An IIR filter has an ideal impulse response with nonzero support unbounded in time, commonly realized through persistent physical state or digital recursion.

Version: v1 · 2026-10-07 · History
Domain-specific #: 13915
Domain group: Applied Sciences & Engineering
Origin domain: Engineering & Design (beyond software)
Subdomains: Signal Processing, Linear Systems → Engineering & Design (beyond software)
Aliases: IIR, Infinite impulse response filter

Core Idea¶

An infinite impulse response (IIR) filter is a linear time-invariant signal filter whose ideal response to a unit impulse has nonzero support unbounded in time. In a causal filter, the response remains nonzero at arbitrarily late future times; a noncausal realization can instead have unbounded support toward negative time. This is an input–output property. It does not say that every response decays, that every IIR is stable, or that every realization contains a visible feedback wire. For LTI systems, the impulse response also determines the response to other inputs by convolution.^[1]

Persistent physical state or a digital recurrence often realizes IIR behavior. A digital first-order low-pass filter with \(y[n]=a\,y[n-1]+x[n]\), zero initial state and \(0<a<1\) has \(h[n]=a^n u[n]\): every future sample is nonzero in the ideal model even though its magnitude shrinks. An analog RC high-pass filter has a different physical realization and a continuous-time exponential tail. Both satisfy the same response-duration test.^[1]^[2]

Structural Signature¶

LTI signal operator — admissible carrier. Under stated initial-state and time-index conventions, the filter maps input signals to outputs with linearity and shift invariance. Its impulse response can represent other zero-state responses through convolution.^[1]
Unit impulse and its response — recognition probe. Apply a discrete unit sample or continuous Dirac impulse and inspect \(h[n]\) or \(h(t)\). An ordinary finite record of samples cannot by itself prove an ideal mathematical tail; an equation or model supplies that conclusion.^[1]^[2]
Unbounded nonzero support — defining differentia. No finite time interval contains every nonzero part of the ideal impulse response. For a causal case, nonzero response persists arbitrarily far forward; it may decay, remain constant or grow. If the response becomes exactly zero outside a finite interval, the filter is FIR.^[1]
Persistent mode or state — common realization. A digital past-output term or analog capacitor state can produce the tail. In a rational transfer model, a mode that survives pole–zero cancellation matters for the overall response. Internal recursion that cancels completely can still yield an FIR filter, so visible implementation alone is not the definition.^[1]^[2]^[3]
Separate stability and design assessment — derived question. After classifying response duration, check bounded-input bounded-output stability and any phase, coefficient or implementation constraints. These are important engineering tests but not admission conditions for the IIR class.^[1]^[4]

The first three roles define the named response class. State, recurrence and design checks explain how instances are built and evaluated without replacing that identity.

What It Is Not¶

IIR is not a synonym for feedback. Digital IIR filters usually use previous outputs, but an analog RC high-pass circuit has an exponential impulse tail through capacitor state without needing an explicit returned output wire. Conversely, MIT's recursive frequency-sampling FIR structure has internal poles that cancel exactly with zeros, leaving a finite overall impulse response. The classification follows the total input–output response, not the appearance of a circuit or signal-flow diagram.^[2]^[3]

It is not “a decaying filter” or “a stable filter.” For \(h[n]=a^n u[n]\), the MIT notes show that \(|a|\geq 1\) makes the first-order example unstable; the response still has infinite support when \(a\neq0\). The same notes show that finite-duration FIR responses are absolutely summable and BIBO stable under finite coefficients. Neither stability nor decay is the test that separates FIR from IIR.^[1]

Nor is every nonlinear, adaptive or time-varying signal operation classified by a single LTI impulse response. Without the LTI assumptions, one observed impulse trace may fail to characterize other inputs. The present entry concerns that response class, not all systems informally said to have “memory.”^[1]

Scope of Application¶

The identity is used in analog and digital signal processing: frequency-selective circuits, sampled-data filters and other LTI response-bearing systems. For causal discrete-time rational filters, recurrences and transfer-function poles are common tools for deriving the tail. For continuous-time LTI circuits, storage elements and transfer functions play a similar analytical role. A noncausal LTI filter can also have unbounded response support in the negative-time direction; causal examples should not be mistaken for the whole definition.^[1]^[2]

The filter task is separate from its response-duration class. A low-pass or high-pass goal can be met by different realizations; it is IIR only if the resulting ideal impulse response has unbounded nonzero support. Analog Devices discusses common IIR advantages such as less memory and fewer multiply–accumulates in compared designs and an analog-like response, alongside possible instability and phase limitations. Those are design-dependent comparisons, not properties every IIR instance must optimize or suffer.^[4]

Clarity¶

Write down the time domain, initial-state convention, and impulse response before saying “IIR.” In discrete time, FIR means there is some finite interval outside which \(h[n]=0\); IIR means no such finite interval contains its nonzero samples. For a causal digital filter, this reduces to asking whether nonzero values occur at arbitrarily large \(n\). An anticausal or two-sided filter requires checking the whole time axis. MIT distinguishes causal from anticausal solutions of a difference equation, so a recurrence alone does not fix the direction of support.^[1]

The digital recurrence \(y[n]=a\,y[n-1]+x[n]\) gives \(h[n]=a^n u[n]\) for a causal zero-state realization. With \(a=0.9\), every \(n\geq0\) has a nonzero ideal value; with \(a=0\), the tail vanishes and the same written family reduces to \(h[n]=\delta[n]\), an FIR case. The boundary is exact mathematical zero, not whether a late output falls below a measurement threshold or is rounded to zero by a finite-precision machine.^[1]

Manages Complexity¶

A compact recurrence can represent a response with indefinitely many nonzero samples. In the first-order example, the one stored previous output plus coefficient \(a\) determines the whole geometric tail; listing every response sample is unnecessary. In an analog RC circuit, one time constant similarly determines the exponential part. This compression is why response, state and transfer descriptions can be compared across unlike implementations.^[1]^[2]

The compression has limits. An internal recurrence can cancel out at the whole-system level, and a pole or feedback term in one block does not prove the complete filter is IIR. After computing the overall response, stability must be checked separately. Design guides often compare IIR and FIR implementation cost and phase, but the useful comparison holds only for specified response requirements and realizations.^[3]^[1]^[4]

Abstract Reasoning¶

Start from the LTI operator and impulse convention. Determine \(h\) from its difference equation, differential equation or transfer function; reduce any exact pole–zero cancellations in the overall transfer; then test whether nonzero support fits inside a finite interval. If it does, the response is FIR even if a recursive structure was used. If it does not, it is IIR, regardless of whether the tail decreases. Only then ask whether the system is causal and BIBO stable under the relevant time direction and pole conditions.^[1]^[3]

This order prevents three tempting shortcuts: “there is feedback, so IIR,” “the response decays, so stable and IIR,” and “a trace looks zero after a while, so FIR.” The first can fail through exact cancellation; the second confuses duration with stability; the third mistakes numerical or measurement resolution for ideal response support.^[3]^[1]

Knowledge Transfer¶

The causal digital low-pass and analog RC high-pass examples have different carriers, implementation technologies and frequency objectives. Yet each has an LTI input–output relation, an impulse probe and a nonzero response tail extending arbitrarily far forward in its own time domain. The precise same response-duration test transfers within signal processing from discrete samples to continuous-time circuits. The state realization differs: a stored digital prior output versus capacitor relaxation.^[1]^[2]

That literal transfer does not make “a lasting effect” in psychology or organizations an IIR filter. Such analogies omit the signal operator, unit impulse, LTI assumptions and exact support test. The named filter remains domain-specific even though its parent notions of a signal filter and an LTI system have broader descendants within technical domains.

Examples¶

Causal digital first-order low-pass¶

Greenberg and Delgutte give \(y[n]=a\,y[n-1]+x[n]\) with \(0<a<1\) and show \(h[n]=a^n u[n]\); their illustrated case takes \(a=0.9\). Mapped back: the fixed-coefficient equation supplies the LTI signal operator; \(x[n]=\delta[n]\) is the impulse probe; \(0.9^n\) at every \(n\geq0\) gives unbounded nonzero support; the delayed prior output is the state realization; and \(|a|<1\) establishes stability for this case, not for IIR generally. At \(a=0\) the system would have a finite one-sample response instead.^[1]

Causal analog RC high-pass¶

Rowell's MIT lecture presents an RC high-pass with \(H(s)=\tau s/(1+\tau s)\), where \(\tau=RC>0\). Rewriting it as \(1-1/(1+\tau s)\) and taking the causal inverse Laplace transform yields \(h(t)=\delta(t)-e^{-t/\tau}u(t)/\tau\). That impulse formula is an explicit derivation from the lecture's transfer function, not a quoted equation. Mapped back: the circuit's transfer relation supplies the LTI signal operator; an ideal Dirac impulse is the probe; the exponential term is nonzero at every finite \(t>0\), establishing unbounded support; capacitor state is the physical realization; and positive \(R,C\) give a decaying tail for this circuit. Its high-pass objective and continuous time differ from the digital low-pass example, while the response-duration identity is the same.^[2]

Structural Tensions¶

Some specified design tasks trade a shorter implementation against phase behavior or stability margin, and the Analog Devices guide compares typical IIR and FIR options on those axes. Those are contingent engineering choices: they depend on the desired response, coefficients and hardware. They are not an intrinsic two-sided tension defining every IIR filter. Likewise, “infinite duration” and “stable” can coexist, as the \(a=0.9\) example shows; treating them as opposed would invent a conflict. The diagnostic is to state the response-support classification first and then evaluate the actual design constraints.^[4]^[1]

Structural–Framed Character¶

Evaluative weight: IIR is a mathematical response classification, while whether a particular implementation is preferable depends on a filter specification. Human-practice dependence: an engineer chooses signal units, initial-state assumptions and measurement horizon, but ideal support and stability follow from the declared model. Institutional origin: textbooks and design guides use a shared technical term, yet a vendor's circuit is only one instance. Vocabulary travel: “impulse,” “filter” and “infinite” here refer to signal-system operations, not every delayed consequence. Import versus recognition: recognize IIR by an LTI impulse response with unbounded nonzero support, not by calling an unrelated persistent process a filter.^[1]^[4]

The live-parent skeleton is signal conditioning plus LTI impulse-response representation. The infinite-support differentia survives changes from analog circuitry to digital equations, while the signal-system assumptions remain essential. Its character: structurally precise within signals and systems, but framed by the domain's impulse, response and filter semantics rather than a substrate-independent Prime.

Structural Core vs. Domain Accent¶

Live Filter (Signal Processing) supplies the signal-to-signal conditioning genus; live Linear Time-Invariant System supplies the independent input–output structure in which one impulse response represents arbitrary zero-state inputs. IIR adds the exact support condition. Those broader roles do not collapse the child because an LTI filter can have a finite response, and a filter need not be LTI at all. Through the live LTI parent, Prime Linearity supplies the portable superposition role; through the live Filter parent, Prime Transformation supplies the input-to-output mapping. These are inherited live graph paths, not extra direct IIR edges. Prime Convolution is related to LTI response calculation, but is not a claimed inherited edge in that path.^[1]

The domain accent is load-bearing. A generic system that retains information after a disturbance is not IIR unless it meets the signal, LTI, impulse and ideal-support tests. Digital output feedback is common, but the analog RC case shows why it cannot replace the response criterion; MIT's recursive FIR example shows why even a digital loop is insufficient on its own.^[2]^[3]

This entry is a kind of Filter (Signal Processing) and is a kind of Linear time-invariant system.

Filter (Signal Processing) is an approved strict subsumption parent: IIR is a signal filter, while the parent also admits finite-response, nonlinear and time-varying filters. Linear time-invariant system is an independent strict subsumption parent: IIR depends on LTI impulse-response representation, while many LTI systems have finite response. The two edges encode signal-conditioning purpose and mathematical response form, respectively.^[1]

Live Prime Feedback describes a common digital return path but is not an all-instance IIR parent. Live Nonrecursive Filter is the finite-tap contrast, not a parent. Live Prime Convolution is relevant to all LTI response calculations through the accepted LTI parent; a redundant direct edge would add no necessary independent role. Internal poles that cancel, phase tendencies and source-domain resemblance do not justify extra direct parents.^[1]^[3]^[2]

Relationships to Other Abstractions¶

Current abstraction Infinite impulse response Domain-specific

Parents (2) — more general patterns this builds on

Infinite impulse response is a kind of Filter (Signal Processing) Domain-specific

An IIR filter is a signal-conditioning filter distinguished by unbounded impulse-response support.
Infinite impulse response is a kind of Linear time-invariant system Domain-specific

The IIR classification depends on an LTI operator characterized by its impulse response.

Hierarchy paths (2) — routes to 2 parentless roots

Infinite impulse response → Filter (Signal Processing) → Transformation → Function (Mapping)

Show alternative path (1)

Neighborhood in Abstraction Space¶

Infinite impulse response sits in a sparse region of the domain-specific corpus (87^th percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.

Family — Wavelets & Function-Space Approximation (19 abstractions)

Nearest neighbors

Computed from structural-signature embeddings · 2026-10-08

Not to Be Confused With¶

An FIR filter has a response supported in a finite interval; it may even use an exactly canceling recursive implementation. A stable filter maps bounded inputs to bounded outputs under the chosen LTI conditions, but an IIR can be unstable. An analog circuit with state may be IIR without a named digital feedback algorithm. A finite measured tail does not settle whether an ideal model has exact zero beyond the observation horizon. A noncausal IIR can have unbounded support toward past time, so causal future-tail examples must not be promoted to the universal definition.^[1]^[3]^[2]

References¶

[1] Julie Greenberg and Bertrand Delgutte, “Biomedical Signal and Image Processing,” Harvard–MIT Division of Health Sciences and Technology course materials, MIT OpenCourseWare, Spring 2007, Chapter 2 “Digital Filters”. §§2.1.1–2.1.3 and 2.2.1–2.2.2, printed pp.2–5, 8; Fig.2.5 printed p.13; §2.3.2 printed p.14. Consulted for the LTI impulse/convolution identity, finite versus infinite response, anticausal boundary, geometric low-pass example and stability limits. https://ocw.mit.edu/courses/hst-582j-biomedical-signal-and-image-processing-spring-2007/b60ddec52ee7fc26607f8afc8a77daf5_ch2_dfilt.pdf registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g ↩h ↩i ↩j ↩k ↩l ↩m ↩n ↩o ↩p ↩q ↩r ↩s ↩t ↩u ↩v ↩w ↩x

[2] Derek Rowell, “Introduction to Time-Domain Digital Signal Processing,” Lecture 13 of 2.161 Signal Processing: Continuous and Discrete, Massachusetts Institute of Technology, Fall 2008, printed pp.13-1–13-2/PDF pp.1–2. Printed p.13-1 gives the continuous RC high-pass \(H(s)=RCs/(RCs+1)\) and \(\tau=RC\); the displayed inverse-Laplace impulse formula in this entry is a curator derivation from that source, not quoted source text. https://ocw.mit.edu/courses/2-161-signal-processing-continuous-and-discrete-fall-2008/13532904b5dcaa23f42cb00bb1b79ad4_lecture_13.pdf registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g ↩h ↩i ↩j ↩k

[3] MIT OpenCourseWare, “Network Structures for Finite Impulse Response (FIR) Systems and Parameter Quantization Effects in Digital Filter Structures,” Lecture 13 transcript in Digital Signal Processing RES.6-008, instructor Alan V. Oppenheim, Spring 2011, PDF pp.5–6. The instructor explains a recursive frequency-sampling FIR realization whose internal poles cancel zeros so the overall response remains finite; the transcript, rather than a formal journal article, is the consulted source. https://ocw.mit.edu/courses/res-6-008-digital-signal-processing-spring-2011/4260578ccde71f509549a31a9d2e3e97_AsSsGjaBbas.pdf registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g ↩h

[4] Analog Devices, “Mixed-Signal and DSP Design Techniques,” Analog Devices 2000 edition, Section 6 “Digital Filters” by Walt Kester, printed pp.6.5–6.6 and 6.30/PDF zero-index pp.6–7 and 31, especially Fig.6.38. Manufacturer design guide consulted for bounded FIR/IIR implementation, phase and stability comparisons; no such tendency is treated as an IIR identity condition. https://www.analog.com/media/en/training-seminars/design-handbooks/MixedSignal_Sect6.pdf registry ↩a ↩b ↩c ↩d ↩e