Papers
Topics
Authors
Recent
Search
2000 character limit reached

IIR Systems: Definition, Stability, & Applications

Updated 13 November 2026
  • Infinite-Impulse-Response (IIR) systems are discrete-time systems whose outputs depend recursively on previous outputs and current inputs, characterized by rational transfer functions and potentially infinite duration impulse responses.
  • IIR systems are widely used in adaptive signal processing for applications such as spectral analysis, control, neural feature extraction, and scientific instrumentation, offering advantages in low-latency operation, reduced storage, and efficient implementation.
  • Common design methods for IIR filters include continuous-to-discrete mappings (e.g., bilinear transformation, matched $z$-transformation), optimization criteria, and data-driven objectives, with stability-improvement techniques such as cascaded low-order sections and pole placement near the unit circle to enhance performance.

Infinite-Impulse-Response (IIR) systems are discrete-time systems whose outputs depend recursively on previous outputs as well as current and delayed inputs. Their transfer functions are rational functions of the delay operator z−1z^{-1}, and their impulse responses generally remain nonzero for infinitely many samples. A generic causal IIR filter is described by

∑k=0Naaky[n−k]=∑k=0Nbbkx[n−k],\sum_{k=0}^{N_a} a_k y[n-k] = \sum_{k=0}^{N_b} b_k x[n-k],

or, with a0=1a_0=1,

y[n]=−∑k=1Naaky[n−k]+∑k=0Nbbkx[n−k].y[n] = -\sum_{k=1}^{N_a}a_k y[n-k] + \sum_{k=0}^{N_b}b_k x[n-k].

The feedback terms generate finite-dimensional memory with potentially infinite-duration response. For a causal rational system, BIBO stability requires every pole to lie strictly inside the unit circle, ∣pi∣<1|p_i|<1. IIR systems are used in recursive filtering, spectral analysis, control, adaptive signal processing, graph learning, neural feature extraction, acoustic rendering, and scientific instrumentation.

1. Mathematical foundations and system structure

An IIR system can be represented by a transfer function

H(z)=∑k=0Mbkz−k1+∑k=1Nakz−k,H(z)= \frac{\displaystyle\sum_{k=0}^{M}b_kz^{-k}} {\displaystyle 1+\sum_{k=1}^{N}a_kz^{-k}},

where the numerator determines feedforward behavior and the denominator determines recursive feedback. Its impulse response is the inverse zz-transform,

h[n]=Z−1{H(z)}.h[n]=\mathcal{Z}^{-1}\{H(z)\}.

For a causal system,

y[n]=∑k=0∞h[k]x[n−k].y[n]=\sum_{k=0}^{\infty}h[k]x[n-k].

The defining distinction from an FIR system is the nonzero denominator. An FIR filter contains only delayed input samples and has finite impulse-response support. An IIR filter feeds delayed outputs back into the recursion, allowing a finite-dimensional state to continue generating output after the input impulse has vanished. Consequently, infinite impulse response does not imply an infinite-dimensional state: classical IIR systems are finitely generated linear recurrence systems (Essl, 2020).

A first-order low-pass IIR filter illustrates the relationship between recursion and memory:

H(z)=11−cz−1,H(z)=\frac{1}{1-cz^{-1}},

with causal impulse response

h[n]=cnu[n].h[n]=c^n u[n].

For $0<|c|<1$, the response decays geometrically but never becomes exactly zero. As cc approaches one, the memory becomes longer and the system approaches an accumulator. A first-order recursive high-pass filter can be written as

H(z)=1−z−11−az−1,H(z)=\frac{1-z^{-1}}{1-az^{-1}},

which combines a zero at z=1z=1 with a pole at z=az=a. The zero suppresses the DC component, while the pole controls the decay and memory.

The state-space form makes the finite-dimensional nature of IIR systems explicit:

x[n+1]=Ax[n]+Bu[n],\mathbf{x}[n+1]=\mathbf{A}\mathbf{x}[n]+\mathbf{B}u[n],

y[n]=Cx[n]+Du[n].y[n]=\mathbf{C}\mathbf{x}[n]+Du[n].

The transfer function is

H(z)=D+C(zI−A)−1B,H(z)=D+\mathbf{C}(z\mathbf{I}-\mathbf{A})^{-1}\mathbf{B},

and the poles are the eigenvalues of A\mathbf{A}. Feedback is represented by the state-transition matrix, while feedforward coefficients appear in the input and output maps.

IIR behavior can also be expressed using sheaves over simplicial topologies. In that formulation, neighboring simplices carry finite-dimensional state spaces connected by linear consistency maps. The feedback recursion is encoded in the state-transition map, and compatible global sections correspond to complete filter trajectories. The construction translates conventional pole-zero LTI filters into topological filters without requiring infinite-dimensional local state spaces (Essl, 2020).

2. Poles, zeros, stability, and numerical behavior

For a causal rational IIR system, BIBO stability is equivalent to absolute summability of the impulse response,

∑n=−∞∞∣h[n]∣<∞,\sum_{n=-\infty}^{\infty}|h[n]|<\infty,

which requires the region of convergence to include the unit circle. In the usual causal realization this means

∣pi∣<1|p_i|<1

for every pole. A pole inside the unit circle produces a decaying mode. A pole on the unit circle produces a nondecaying mode and loss of BIBO stability. A pole outside the unit circle produces a growing response.

The pole location determines both temporal and spectral behavior. A real pole produces monotonic or alternating decay, while a complex pole produces a damped oscillation. A pole close to the unit circle gives long memory and narrow frequency selectivity, but also increases sensitivity to coefficient quantization, roundoff, initialization, and transient buildup. These issues are especially important in recursive frequency analyzers, where nominal pole-zero cancellation can conceal marginally stable implementation states. The conventional sliding DFT, for example, can have an overall FIR response in exact arithmetic while relying internally on unit-circle resonators whose imperfect cancellation causes long-term drift (Kennedy, 2014).

A stable realization must therefore be evaluated at the implementation level rather than solely from an ideal transfer function. In the stabilized IIR sliding DFT, unit-circle cancellation is avoided by using leaky resonators with poles

zp,k=rejωk,0<r<1,z_{p,k}=re^{j\omega_k}, \qquad 0<r<1,

and a nonrecursive mixing matrix. Because the mixing stage introduces no poles, all recursive poles remain strictly inside the unit circle (Kennedy, 2014).

Pole placement also determines the trade-off between memory and adaptation. In low-order regression-based smoothers and differentiators, the pole is parameterized by

p=eσ,0<p<1.p=e^\sigma, \qquad 0<p<1.

Small pp produces short memory and rapid adaptation; pp near one produces longer memory, stronger noise attenuation, and narrower low-pass behavior. A shape parameter KK can modify the weighting from a conventional exponentially decaying memory to a delayed, more nearly symmetric weighting. The resulting causal filters have repeated poles at pp and impulse responses of the form

h[n]=PB,K(n,q)pnu[n],h[n]=P_{B,K}(n,q)p^n u[n],

where BB is the polynomial degree and qq is the synthesis delay (Kennedy, 2015).

Finite precision can change pole locations and therefore stability. Direct-form implementations are particularly sensitive because coefficient quantization changes the roots of the numerator and denominator polynomials simultaneously. Factored cascades or first- and second-order sections reduce this sensitivity. In a detector deconvolution system, a third-order direct-form implementation required more than 34-bit coefficient precision for satisfactory behavior, whereas a factored cascade reduced the required DSP resources and improved quantization robustness (Födisch et al., 2016).

3. Design methods and approximation trade-offs

IIR filters may be designed from continuous-time prototypes, rational approximations, optimization criteria, or data-driven objectives. Common continuous-to-discrete mappings include forward difference, backward difference, bilinear transformation, and matched-zz transformation. The matched-zz method maps a continuous pole or zero according to

zk=eskT,z_k=e^{s_kT},

and directly preserves sampled exponential decay. The bilinear transformation maps the stable left half of the ss-plane inside the unit circle and generally provides favorable frequency-response behavior, although it introduces frequency warping (Födisch et al., 2016).

Classical analog prototypes include Butterworth, Chebyshev type I, Chebyshev type II, and elliptic filters. Butterworth filters are monotonic and maximally flat in the passband. Chebyshev type I filters have passband ripple, whereas Chebyshev type II filters have a monotonic passband and equiripple stopband. Elliptic filters have ripple in both passband and stopband. These prototypes provide different compromises among transition sharpness, ripple, phase, order, and coefficient sensitivity.

Continued Fraction Expansion (CFE) provides another route to low-order rational approximation. For fractional differ-integrators, a continuous-time operator such as

G(s)=sαG(s)=s^\alpha

or

G(s)=s−γG(s)=s^{-\gamma}

is mapped to a rational discrete-time operator and truncated to a finite CFE convergent,

GN(z−1)=BN(z−1)AN(z−1).G_N(z^{-1})=\frac{B_N(z^{-1})}{A_N(z^{-1})}.

Genetic-algorithm optimization can tune an interpolation parameter in the generating function to minimize a weighted magnitude-phase discrepancy. For the Al-Alaoui family, the parameter controls the balance between Euler and Tustin behavior; for the Chen–Vinagre family, it controls the balance between Simpson and Tustin behavior. The resulting filters remain ordinary low-order IIR systems (Das et al., 2012).

Orthogonal-polynomial constructions provide a direct alternative to iterative optimization. Desired low-pass and high-pass characteristics are transformed into object functions and expanded using Legendre polynomials,

fN(x)=∑n=0Na2nP2n(x),fD(x)=∑n=0Nb2nP2n(x).f_N(x)=\sum_{n=0}^{N}a_{2n}P_{2n}(x), \qquad f_D(x)=\sum_{n=0}^{N}b_{2n}P_{2n}(x).

The IIR response is then formed as a ratio,

HIIR(ω)=HN,a(ω)HD,a(ω).H_{\mathrm{IIR}}(\omega)=\frac{H_{N,a}(\omega)}{H_{D,a}(\omega)}.

The expansion coefficients are orthogonal-projection coefficients, so the approximation is least-mean-square with respect to the selected polynomial basis. However, denominator roots can lie outside the unit circle and require explicit stabilization. Moving unstable poles to the origin makes the realization stable but changes magnitude and phase slightly; therefore, the method produces approximately rather than exactly linear phase (Kumar et al., 2013).

Min-max semidefinite programming offers a different design framework when an IIR filter is to be approximated by a stable FIR filter. Given a stable IIR filter P(z)P(z), an FIR filter Q(z)Q(z) can be selected by solving

min⁡Q∈FN∥(P−Q)W∥∞,\min_{Q\in\mathcal{F}_N}\|(P-Q)W\|_\infty,

where W(z)W(z) is a weighting filter. Inverse FIR filtering similarly solves

min⁡Q∈FN∥(QP−1)W∥∞.\min_{Q\in\mathcal{F}_N}\|(QP-1)W\|_\infty.

Using the KYP lemma or generalized KYP lemma, the frequency-domain bound is converted into a finite-dimensional linear matrix inequality and solved as a semidefinite program. The full-band KYP formulation controls the worst-case error over [0,π][0,\pi], while the generalized KYP formulation restricts the bound to a selected frequency interval (Nagahara, 2013).

4. IIR architectures and implementation strategies

IIR systems can be implemented as direct-form recursions, cascades of low-order sections, parallel filter banks, state-space systems, or feedback delay networks. Direct forms are compact but can be numerically sensitive at high order. Cascaded first- or second-order sections generally provide better coefficient conditioning and localized control over pole-zero behavior.

Parallel structures use multiple low-order IIR sections whose outputs are combined. In summed parallel IIR gravitational-wave filtering, a chirp is decomposed into exponentially windowed local sinusoids,

U(t)≈∑lul(t),U(t)\approx \sum_l u_l(t),

and each component is assigned a single-pole filter. The filter-bank recursion is

yk,l=a1,lyk−1,l+b0,lxk−dl,y_{k,l}=a_{1,l}y_{k-1,l}+b_{0,l}x_{k-d_l},

with coherently summed output

zk≃2∑lyk,l.z_k\simeq 2\sum_l y_{k,l}.

This converts long matched-filter correlations into recursive state updates. The method achieved approximately 99% waveform overlap for a reported 1.4+1.4 solar-mass template using 687 IIR filters and recovered more than 99% of the injections recovered by the traditional matched filter at the stated false-alarm conditions (Hooper et al., 2011).

Feedback delay networks (FDNs) extend the IIR concept to reverberation. The early reflection component is a finite delayed sum,

yearly[n]=∑ibix[n−κi],y_{\mathrm{early}}[n]=\sum_i b_i x[n-\kappa_i],

while the reverberant tail is generated by recursive delay lines,

s[n+1]=AFDNs[n]+BFDNx[n],\mathbf{s}[n+1]=\mathbf{A}_{\mathrm{FDN}}\mathbf{s}[n]+\mathbf{B}_{\mathrm{FDN}}x[n],

y[n]=CFDNs[n]+DFDNx[n].y[n]=\mathbf{C}_{\mathrm{FDN}}\mathbf{s}[n]+D_{\mathrm{FDN}}x[n].

The poles are the eigenvalues of AFDN\mathbf{A}_{\mathrm{FDN}}, and stability requires

ρ(AFDN)<1.\rho(\mathbf{A}_{\mathrm{FDN}})<1.

A 16-loop FDN has been proposed for real-time room-impulse-response rendering, with differentiable optimization of decay and scale parameters to match clarity, definition, center time, and T30T_{30} while avoiding the latency and computational cost of long convolution (Gerami et al., 30 Sep 2025).

IIR systems can also be embedded in adaptive and distributed algorithms. An augmented complex adaptive IIR filter models widely linear ARMA processes using ordinary and conjugate-valued autoregressive and moving-average terms. The adaptive output includes both y[n−m]y[n-m] and y∗[n−m]y^*[n-m], enabling representation of proper and improper complex signals. In an incremental network, each node updates the coefficient estimate and passes it to the next node. The resulting incremental augmented complex IIR algorithm uses recursive sensitivity updates and improves prediction gain relative to non-cooperative adaptation in the reported experiments (Khalili et al., 2016).

5. Applications and modern extensions

IIR filters are used when long memory, sharp spectral shaping, low latency, or compact recursive representation is important.

In adaptive optics, independently optimized modal IIR controllers are designed from measured disturbance spectra. For each mode,

Wi(z)=11+Ci(z)P(z)W_i(z)=\frac{1}{1+C_i(z)P(z)}

is the rejection transfer function, and the controller coefficients minimize integrated residual modal variance subject to a pole-location constraint. Third-order filters are used for turbulence-dominated modes, while seventh-order filters are used for Tip and Tilt in the presence of vibration lines near 13 and 22 Hz. The IIR controller improves reported Strehl ratios relative to common-gain and optimized modal-gain integrators, particularly for faint visible guide stars and vibration-contaminated conditions (Agapito et al., 2012).

In detector electronics, IIR deconvolution inverts the exponential decay of a charge-sensitive amplifier. The continuous-time amplifier response is

HQ(s)=sτ1+sτ,H_Q(s)=\frac{s\tau}{1+s\tau},

so the inverse is

HQ−1(s)=1+sτsτ.H_Q^{-1}(s)=\frac{1+s\tau}{s\tau}.

A matched-zz realization has the form

H(z)=G1−e−T/τz−11−z−1,H(z)=G\frac{1-e^{-T/\tau}z^{-1}}{1-z^{-1}},

converting the exponential tail into a step-like signal that can subsequently be transformed into a rectangular or trapezoidal pulse. Higher-order system identification is required when the measured amplifier deviates from the ideal first-order model (Födisch et al., 2016).

In plant electrophysiology, high-pass IIR filters are selected by optimizing wavelet-packet energy distributions. The objective minimizes the dispersion of pre-stimulus energy centroids across experiments while preserving stimulus-dependent post-stimulus differences. Among Butterworth, Chebyshev type I, Chebyshev type II, and elliptic filters, the Chebyshev type-II high-pass filter produced the lowest reported objective values for the tested segmentation lengths (Das et al., 2016).

In embedded machine learning, first-order learnable IIR filters act as recurrent feature-extraction layers. For each accelerometer axis,

a~d[n]=γda~d[n−1]+aˉd[n]−aˉd[n−1],\tilde a_d[n] = \gamma_d\tilde a_d[n-1] +\bar a_d[n]-\bar a_d[n-1],

with transfer function

Hd(z)=1−z−11−γdz−1.H_d(z)=\frac{1-z^{-1}}{1-\gamma_dz^{-1}}.

The learned pole controls temporal memory while the zero at z=1z=1 suppresses quasi-static gravity and orientation components. Combined with short FIR filters and an MLP, the architecture performs animal-behavior classification on an embedded collar device with low parameter count and real-time inference (Arablouei et al., 2021).

In graph neural networks, IIR graph filters replace polynomial FIR graph filters. A first-order recursion is

Yt+1=aL~Yt+bX,\mathbf{Y}^{t+1} = a\widetilde{\mathbf{L}}\mathbf{Y}^{t} +b\mathbf{X},

with steady-state response

hARMA1(λ~)=b1−aλ~.h_{\mathrm{ARMA}_1}(\widetilde{\lambda}) = \frac{b}{1-a\widetilde{\lambda}}.

Parallel first-order branches form higher-order rational graph responses. In cyberattack localization on a 300-bus smart grid, IIR graph neural networks provide sharper graph-frequency approximation than same-order FIR graph filters and improve the reported sample-wise and bus-wise localization results relative to an FIR-GNN (Boyaci et al., 2022).

Neural sequence models have also used adaptive second-order IIR filters before local attention. The Focus layer generates causal, input-dependent filter coefficients from preceding temporal bins and applies order-two denominators before chunked local attention. Its recurrence is

y[t]=b0x[t]+b1x[t−1]+b2x[t−2]−a1y[t−1]−a2y[t−2].y[t] = b_0x[t]+b_1x[t-1]+b_2x[t-2] -a_1y[t-1]-a_2y[t-2].

The adaptive filter provides long-range propagation, while local attention models short-range interactions. The reported architecture has sub-quadratic sequence complexity and performs competitively on associative recall, language modeling, and sequential MNIST (Lutati et al., 2023).

In spatial audio, neural IIR filter fields represent HRTFs using cascades of low-shelf, peaking, and high-shelf sections. The neural field maps source direction to interpretable filter parameters such as center frequency, bandwidth, and gain. Structured parameterization is used to maintain stable sections while producing compact resonant representations. The approach is competitive with magnitude-based neural fields and is particularly effective when HRTF measurements are sparse (Masuyama et al., 2024).

6. Advantages, limitations, and design considerations

The principal advantage of an IIR system is that a compact recursive state can represent long temporal memory or sharp spectral structure. A low-order rational response can approximate behavior that would require a higher-order FIR polynomial, while recursive updates avoid revisiting the full history. This supports low-latency operation, reduced storage, and efficient implementation on processors, GPUs, FPGAs, and embedded systems.

The same feedback that provides efficiency introduces risks. Stability depends on pole locations, and poles close to the unit circle create long transients, sensitivity to finite precision, and slow adaptation. Nominal pole-zero cancellation can conceal marginally stable states, as in sliding DFT implementations. Adaptive IIR systems additionally require stability-aware training or coefficient constraints; the distributed augmented complex IIR algorithm, for example, uses sensitivity recursions but does not provide a formal stability theorem (Khalili et al., 2016).

IIR systems also impose structural assumptions. A parametric cascade may be highly efficient for resonant HRTFs but inefficient for responses that lack modal structure. A rational graph filter may approximate sharp graph-frequency transitions efficiently, but its recursive convergence depends on the graph operator and coefficients. An FDN can reproduce perceptual reverberation metrics without reproducing every measured RIR sample, but its acoustic fidelity depends on the chosen delay structure and feedback matrix.

Zero-phase processing requires separate qualification. Forward-backward filtering cancels phase distortion and produces an approximately magnitude-squared response, but it is noncausal as an offline operation. This is useful in eye-tracking and plant-signal preprocessing, but it cannot be used unchanged when strict real-time causality is required. In EyeLink 1000 recordings, a seventh-order zero-phase Butterworth IIR produced a substantially sharper roll-off than heuristic and Savitzky–Golay filters, although the FIR alternative had the sharpest roll-off and the IIR increased temporal autocorrelation (Raju et al., 2023).

Practical IIR design therefore requires joint consideration of:

  • Pole stability: all causal poles should remain strictly inside the unit circle.
  • Realization structure: cascaded low-order sections are generally less sensitive than high-order direct forms.
  • Initialization: recursive states can produce startup transients, especially for narrowband filters.
  • Finite precision: coefficient and state quantization can move poles and alter decay.
  • Bandwidth and memory: poles near the unit circle improve selectivity but slow adaptation.
  • Causality: zero-phase forward-backward filtering requires future samples.
  • Model mismatch: inverse filtering and data-driven controllers depend on the fidelity of the modeled plant or disturbance spectrum.
  • Approximation criterion: waveform overlap, H∞H^\infty error, residual variance, wavelet-energy dispersion, perceptual metrics, and classification performance optimize different notions of quality.

The continuing development of IIR systems is centered on making recursion both expressive and reliable. Relevant directions include stability-aware optimization, differentiable pole parameterizations, robust finite-precision realizations, adaptive and graph-structured recursions, hybrid IIR/FIR architectures, and neural systems that combine recursive long-range memory with local attention or other nonlinear processing. Recent work on mixed-signal neural interfaces further illustrates the importance of fixed-point IIR design in constrained hardware, although detailed filter equations and implementation results must be established from the complete device architecture rather than inferred from the filter class alone (Antoniadis et al., 16 Nov 2025).

Definition Search Book Streamline Icon: https://streamlinehq.com
References (18)

Topic to Video (Beta)

No one has generated a video about this topic yet.

Whiteboard

No one has generated a whiteboard explanation for this topic yet.

Follow Topic

Get notified by email when new papers are published related to Infinite-Impulse-Response (IIR) Systems.