Nonrecursive (FIR) Filter

A causal discrete-time linear filter realized as a finite weighted sum of current and delayed input samples, with no previous-output feedback, so its finite tap vector completely determines both impulse and frequency response.

Version

v1 · 2026-08-30 · History

Domain-specific #

2390

Origin domain

digital signal processing

Subdomain

finite impulse response filtering

Aliases

Non-recursive digital filter

Core Idea

A nonrecursive filter in the retained digital-signal-processing sense computes each output sample as a finite weighted sum of current and delayed input samples, without using any previous output as an input to the next calculation:

\[ y[n]=\sum_{k=0}^{M} b_k x[n-k]. \]

The coefficient vector \((b_0,\ldots,b_M)\), often called the taps, is both the filter's impulse response and the fixed convolution kernel. An impulse at the input produces exactly those \(M+1\) values and then zero, hence finite impulse response (FIR). ARM's production CMSIS-DSP implementation exposes the structure directly as a multiply–accumulate sequence over a coefficient array and delayed input-state buffer.^[1]

“Nonrecursive” names a direct feedforward realization; “FIR” names an input–output property. They coincide for the canonical finite-tap direct form but should not be treated as perfect logical synonyms in every implementation. An FIR response can sometimes be computed with a recursion plus exact pole–zero cancellation, such as a running-sum moving average; finite-precision arithmetic can then disturb the cancellation. The node therefore retains the robust professional identity finite-tap, input-only linear convolution, while keeping realization and response concepts explicit.

The abstraction matters because the small tap vector licenses strong deductions before processing any signal: finite memory, bounded-input bounded-output stability for finite coefficients, a polynomial transfer function with no finite poles, a finite startup/flush transient, and—when taps have the required symmetry—exact linear phase.

Structural Signature

The recurring structure is:

sampled input stream -> finite delay line -> fixed finite tap vector -> parallel multiply–accumulate -> output sample, with no prior-output return path.

Eight roles are mandatory:

1. A discrete index. Input and output are sequences sampled over time, space, or another regular index.
2. A finite input history. At output index \(n\), only \(x[n],\ldots,x[n-M]\) are retained.
3. A fixed tap vector. The coefficients define how each lag contributes. Time-varying taps require a separately qualified adaptive/time-varying extension.
4. Delay alignment. Each tap multiplies the input sample at its designated lag.
5. Linear accumulation. Products are summed to form \(y[n]\).
6. No previous-output term. The difference equation contains no \(a_k y[n-k]\) feedback terms.
7. Finite impulse support. The response to \(\delta[n]\) is \(b_n\) for the finite tap interval and zero outside it.
8. Frequency-response dual. The transfer function is \(H(z)=\sum_{k=0}^{M}b_kz^{-k}\), and on the unit circle its Fourier response specifies gain and phase by frequency.

The invariant is: the next output is determined entirely by a bounded window of input samples and one finite coefficient vector; no computed output re-enters the signal path.

What It Is Not

It is not an IIR or recursive digital filter. A general rational difference equation includes previous outputs; its denominator creates poles and an impulse response that ordinarily persists indefinitely. A stable IIR may be safe and efficient, but it fails the input-only invariant.

It is not every moving window operation. A median filter uses a finite window and no output feedback but is nonlinear; morphological filters and rank filters likewise lack the linear weighted-sum/convolution identity. They are nonrecursive in a broad computational sense, not instances of this retained LTI FIR abstraction.

It is not linear phase by definition. FIR filters can achieve exact linear phase when their taps are symmetric or antisymmetric; generic taps can have nonlinear phase. MathWorks distinguishes four linear-phase FIR types according to tap symmetry and order and notes the corresponding endpoint restrictions.^[2]

It is not stateless. The algorithm stores delayed inputs, so it has finite state even though it does not feed back outputs. Nor is it automatically safe in every implementation: finite mathematical coefficients guarantee BIBO stability, but fixed-point accumulators can overflow or saturate, a boundary explicitly noted by CMSIS-DSP.^[1]

Scope of Application

Nonrecursive FIR filters are used for low-pass, high-pass, band-pass, and band-stop filtering; smoothing and differentiation; Hilbert transformation; equalization; matched filtering; pulse shaping; decimation and interpolation; and spatial image kernels. SciPy and MATLAB expose window, least-squares, and equiripple design tools, while embedded libraries implement the same tap-and-delay structure across floating-point and fixed-point types.^[3][2][1]

The abstraction spans offline and streaming operation. A short audio equalizer may run a direct multiply–accumulate loop. A very long impulse response can be evaluated by FFT overlap-add/overlap-save while retaining the same FIR input–output system. A polyphase decomposition rearranges the taps to avoid computing samples discarded during sample-rate conversion. Sparse FIR forms store only nonzero taps and their delays. These are computational realizations of the same finite convolution if their outputs agree.

The scope assumes linear time-invariant coefficients unless explicitly extended. Adaptive FIR filters such as LMS preserve the finite input-only architecture at each step but update coefficients from an error signal; their learning rule and stability analysis are additional domain machinery.

Clarity

Three quick tests identify the abstraction.

Difference-equation test: after expansion, does \(y[n]\) depend only on finitely many \(x[n-k]\) terms? Impulse test: does a unit impulse yield a response that becomes exactly zero after finitely many samples? realization test: does the implemented signal path avoid using previous computed outputs? A direct-form candidate passes all three.

The tests also surface the response/realization distinction. The running average

\[ y[n]=\frac1N\sum_{k=0}^{N-1}x[n-k] \]

is an FIR system. It can be implemented directly and nonrecursively, or updated as \(y[n]=y[n-1]+(x[n]-x[n-N])/N\). The latter uses recursion but algebraically cancels it in exact arithmetic. If the question is filter class, answer FIR; if it is implementation topology, answer recursive realization. The Encyclopedia node emphasizes their canonical intersection and records this exception rather than hiding it.

Manages Complexity

The filter reduces an entire stream transformation to one inspectable vector. The tap support fixes memory and arithmetic cost; coefficient sum gives DC gain; alternating sum gives Nyquist gain; symmetry reveals phase behavior; the Fourier transform of the taps reveals magnitude response; and the \(\ell_1\) norm supplies a simple output bound.

It also separates design from execution. Design methods choose coefficients approximating a desired frequency response under constraints. Execution is the same delay–multiply–accumulate mechanism no matter whether taps came from a window, least-squares criterion, frequency sampling, or Parks–McClellan minimax optimization. MathWorks documentation explicitly separates these design families and their error criteria.^[2]

Finally, absence of output feedback localizes numerical error. Rounding at one output does not become a state that circulates through later outputs. This eliminates feedback limit cycles and coefficient-sensitive pole movement, although input quantization and accumulator errors remain.

Abstract Reasoning

Several deductions follow directly.

BIBO stability. If \(|x[n]|\le B\), then

\[ |y[n]|\le B\sum_{k=0}^{M}|b_k|, \]

which is finite for a finite tap vector. Stability requires no pole-location calculation.

Finite forgetting. An input perturbation at index \(n_0\) can affect only outputs \(n_0\) through \(n_0+M\). The system forgets exactly after its support length, making warm-up and flush intervals calculable.

Linear-phase condition. Symmetric taps \(b_k=b_{M-k}\) or antisymmetric taps \(b_k=-b_{M-k}\) factor the frequency response into a real amplitude times a pure delay. In the passband, symmetric linear-phase filters preserve waveform shape while delaying by \(M/2\) samples; the parity/symmetry combination constrains DC and Nyquist response.^[2]

Cascade inference. Cascading two FIR filters convolves their finite tap vectors, producing another FIR filter. Orders add and supports remain finite. Conversely, factorization can split one long response into stages.

Cost inference. Direct execution uses roughly \(M+1\) multiplications and additions per output. Stronger attenuation or a narrower transition band often requires greater order and delay than a comparable IIR design, the central resource tradeoff.^[2]

Knowledge Transfer

The identity transfers literally wherever samples and fixed finite convolutions occur. In audio, taps implement equalizers and crossover filters. In communications, they shape pulses, match known waveforms, and compensate channel distortion. In images, a finite 2-D kernel blurs, sharpens, or detects edges; MathWorks identifies finite spatial support as 2-D FIR and notes the same stability and linear-phase benefits.^[4]

In multirate processing, polyphase FIR banks reorganize one tap vector around decimation or interpolation phases without changing the underlying response. In embedded computing, ARM CMSIS uses the same equation across Q7, Q15, Q31, and floating-point data, with implementation-specific overflow rules.^[1] In scientific computing, SciPy's firwin returns a finite, linear-phase coefficient vector and applies ordinary discrete filtering.^[3]

Outside sampled linear systems, the portable residue is Convolution and Feedforward. A finite weighted local mixture in an image or neural layer can instantiate Convolution, but it is not a digital FIR filter unless the sampling, fixed kernel, LTI input–output interpretation, and impulse response are meaningful.

Examples

Three-tap smoothing filter

Let

\[ y[n]=0.25x[n]+0.5x[n-1]+0.25x[n-2]. \]

The roles are explicit: a three-sample input delay line, taps \((0.25,0.5,0.25)\), and one weighted sum. The impulse response is exactly those three values. The tap sum is one, so DC is preserved. The symmetric taps give a one-sample constant group delay; the response attenuates high-frequency alternation because \(0.25-0.5+0.25=0\) at Nyquist. A bounded input stays bounded by its maximum magnitude because the absolute tap sum is one.

Embedded block implementation

An embedded microphone pipeline applies a 63-tap low-pass FIR to blocks of samples before decimation. CMSIS-DSP stores 63 coefficients and a state buffer containing the delayed input history, then performs the documented multiply–accumulate equation for every output.^[1] Coefficients may be shared across channels, but each channel requires separate input state. The implementation designer must budget 63 tap operations per direct output, group delay of 31 samples if taps are symmetric, and sufficient accumulator width for fixed-point data. A polyphase realization can avoid unnecessary computations after decimation while producing the same mathematical output.

Recursive-realization boundary

For a 1000-sample moving average, direct convolution is FIR and nonrecursive but expensive. A running-sum update adds the newest sample and subtracts the oldest using a previous running result. The response still has 1000 taps in exact arithmetic, but the implementation is recursive. This boundary case demonstrates why “FIR” and “nonrecursive” are close professional labels without being perfectly interchangeable predicates.

Structural Tensions

Guaranteed mathematical stability versus implementation overflow. Finite taps give BIBO stability, but a fixed-width accumulator can saturate or wrap. Diagnostic: does the stated stability concern the ideal input–output system or the numeric realization?

Linear phase versus delay/order. Symmetric taps can preserve waveform shape, but a length-\(M+1\) linear-phase filter carries delay \(M/2\), and demanding sharp transitions raises \(M\). Diagnostic: is phase fidelity worth latency and computation in this application?

Direct transparency versus efficient restructuring. Direct form makes the finite convolution obvious; FFT, polyphase, sparse, or recursive-cancellation implementations can be cheaper but obscure the nonrecursive topology. Diagnostic: is the classification about mathematical response or about the actual dependency graph?

No feedback-error circulation versus coefficient count. FIR avoids pole sensitivity and feedback limit cycles, but often requires more coefficients than IIR for similar magnitude specifications. Diagnostic: which resource is scarcer—numerical robustness/phase control or memory, multiply rate, and latency?

Finite history versus sharp frequency selectivity. A short kernel localizes response in time but broadens transitions and ripple constraints in frequency. Diagnostic: what transition width can the available support length actually realize?

Structural–Framed Character

The nonrecursive FIR filter is mixed-structural. Its mathematical core—finite discrete convolution—is formal and evaluatively neutral, and physical implementations can be recognized from their dependency graph. Yet its distinctive vocabulary and operational commitments are those of sampled-data signal processing: taps, impulse response, passband, stopband, group delay, fixed-point accumulator, and frequency response.

The construct travels literally across audio, communications, image, measurement, and embedded systems because those domains share the same sampled linear substrate. Beyond that substrate, the reusable skeleton is the prime Convolution rather than “FIR filter.” It is therefore an unusually structural domain-specific abstraction, not a new substrate-independent prime.

Structural Core vs. Domain Accent

The structural core is one fixed finite kernel sliding over a position-indexed input and producing weighted local sums. That is exactly prime:convolution, with an added feedforward dependency graph and bounded support.

The domain accent is what makes the object a digital filter: causal sample indexing, delay-line implementation, impulse and frequency response, filter-design specifications, phase, latency, coefficient quantization, and multirate realization. Remove those commitments and any small convolution kernel would count; retain them and the node predicts stability, finite forgetting, group delay, operation count, and FIR/IIR tradeoffs. The residual is coherent and recurrent within signal processing, so the node clears the domain-specific bar while its portable content is already covered.

Instantiates / Related Primes

The minimal prospective DAG parent is prime:convolution through strict subsumption. A retained nonrecursive FIR filter is a particular finite, causal, discrete convolution whose kernel is the tap vector. Convolution can be infinite, continuous, noncausal, spatial, or probabilistic, so the parent occurs without the child; the child cannot occur without the parent's fixed-kernel weighted sum.

Feedforward describes the no-output-return topology, Fourier Transform diagonalizes the convolution into frequency multiplication, and Signal Extraction explains many applications. Recursive Attenuating Amplification is a contrast rather than coverage. Particle Filter and Quotient Filter are lexical collisions: one is sequential Bayesian estimation and the other an approximate-membership data structure.

Relationships to Other Abstractions

Current abstraction Nonrecursive (FIR) Filter Domain-specific

Parents (1) — more general patterns this builds on

• Nonrecursive (FIR) Filter is a kind of Convolution Prime

  The minimal prospective DAG parent is prime:convolution through strict subsumption.

Hierarchy path (1) — routes to 1 parentless root

• Nonrecursive (FIR) Filter → Convolution → Function (Mapping)

Neighborhood in Abstraction Space

Nonrecursive (FIR) Filter 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 — Unclustered & Miscellaneous (1565 abstractions)

Nearest neighbors

• Modified Discrete Cosine Transform — 0.83
• Simulation Decomposition — 0.81
• Bilinear Transform — 0.80
• Lyapunov Exponent — 0.78
• Variogram — 0.78

Computed from structural-signature embeddings · 2026-09-08

Not to Be Confused With

• IIR/recursive filter: uses previous outputs and ordinarily has an infinite impulse response.
• Any FIR system: FIR names response support; some FIR responses have recursive computational realizations.
• Median or rank filter: finite-window and feedforward but nonlinear, so not convolution with taps.
• Moving average statistic: is an FIR filter when used as a fixed sampled convolution, but broader statistical usage may not invoke filter design.
• Adaptive FIR filter: retains finite convolution while adding a coefficient-learning loop and its own convergence conditions.
• Particle filter: a sequential Monte Carlo estimator, not a digital frequency-selective filter.
• Quotient/Bloom filter: data structures for approximate membership, unrelated to signal convolution.
• Linear-phase filter: a property requiring tap symmetry, not every FIR filter.

References

[1] Arm, CMSIS-DSP: “Finite Impulse Response (FIR) Filters”. Authoritative implementation documentation supporting the tap equation, delay state, block processing, coefficient storage, and fixed-point cautions. registry ↩a ↩b ↩c ↩d ↩e

[2] MathWorks, “FIR Filter Design”. Authoritative technical documentation supporting stability, exact linear-phase possibility, finite transients, design methods, symmetry types, order/delay, and FIR–IIR tradeoffs. registry ↩a ↩b ↩c ↩d ↩e

[3] SciPy, Signal Processing Tutorial: “FIR Filter”. Maintained scientific-computing documentation supporting FIR/IIR classification and window-designed linear-phase FIR examples. registry ↩a ↩b

[4] MathWorks, “Design Linear Image Filters in the Frequency Domain”. Supports the 2-D finite-support FIR extension and its stability and phase-preservation properties. registry ↩

[5] H. D. Helms, “Nonrecursive Digital Filters: Design Methods for Achieving Specifications on Frequency Response”, IEEE Transactions on Audio and Electroacoustics 16, no. 3 (1968): 336–342. Historical primary source for nonrecursive-filter design. registry

[6] Alan V. Oppenheim and Ronald W. Schafer, Discrete-Time Signal Processing, 3^rd ed. (Pearson, 2010). Standard authority for LTI convolution, FIR/IIR systems, z transforms, structures, and linear phase. registry