Sequential Matched Filter (SMF) Overview

• Sequential Matched Filter (SMF) is a technique that sequentially applies sliding correlation with a known template to maximize SNR, enabling flexible implementations across time and transform domains.
• Various SMF designs, such as streaming FIR filters for chaotic signal detection and multi-stage transform approaches, illustrate practical trade-offs in quantization, thresholding, and real-time performance.
• Adaptive and learned SMF variants use reinforcement learning and least-squares optimization to adjust filter parameters dynamically, reducing bit error rates despite challenges in noise modeling and terminological ambiguity.

Sequential Matched Filter (SMF) denotes matched-filter architectures in which correlation with a known template is carried out sequentially rather than as a single batch operation. In the cited literature, this sequentiality appears in several forms: sample-by-sample sliding convolution in a streaming FIR, stage-by-stage realization of a matched filter after an intermediate transform, structured line-by-line search over a time–frequency grid, path-by-path matched filtering with residual cancellation, and a learned sequence of filters generated by a reinforcement-learning policy (Bailey et al., 2015, Movafagh et al., 2021, Fish et al., 2011, Li et al., 12 Jul 2025, Tian et al., 29 Aug 2025). In the classical communication-theory sense, the matched filter is the linear time-invariant filter $h(t)=k\,x(t_x-t)$ that maximizes the output SNR for a known signal in additive white Gaussian noise (AWGN), and its discrete-time FIR form,

$Y[n]=\sum_{k=0}^{N-1}H[k]\,X[n-k],$

is already an explicitly sequential sliding convolution (Bailey et al., 2015). A distinct and unrelated literature uses the same acronym for Set-Membership Filter in bounded-noise state estimation (Cong et al., 2020).

1. Matched-filter foundations

The theoretical core of SMF remains the classical matched filter. For a received signal

$v(t)=A\,x(t-t_p)+n(t),$

with known waveform $x(t)$, amplitude scaling $A$, propagation delay $t_p$, and AWGN $n(t)$, the matched filter is the linear time-invariant filter whose impulse response is the time-reversed known signal,

$h(t)=k\,x(t_x-t),$

chosen to maximize output SNR at a decision time. In discrete Gaussian detection, the same object appears as the Neyman–Pearson test statistic

$T(x)=x^T C^{-1}s=x^T f_s,\qquad f_s=C^{-1}s,$

with $C$ the noise covariance matrix. Under stationary Gaussian noise, Neyman–Pearson optimality and SNR optimality coincide, the output statistic is Gaussian under both hypotheses, and the effective output SNR is $Y[n]=\sum_{k=0}^{N-1}H[k]\,X[n-k],$0 (Bailey et al., 2015, Vio et al., 2021).

Sequential matched filtering does not alter this optimum kernel; it changes how the kernel is evaluated. When the signal position is unknown, the matched filter is realized as a sliding correlation over candidate positions,

$Y[n]=\sum_{k=0}^{N-1}H[k]\,X[n-k],$1

In this sense, an SMF is the operational, streaming realization of the same matched-filter principle: the statistic is updated over successive windows or hypotheses, and detection is based on thresholding the evolving correlation output (Vio et al., 2021).

2. Streaming FIR SMFs and reverse-time chaos

A direct hardware-level formulation of SMF appears in digital matched filtering for reverse-time chaos. There, the reverse-time chaotic oscillator

$Y[n]=\sum_{k=0}^{N-1}H[k]\,X[n-k],$2

admits a closed-form solution through a basis pulse $Y[n]=\sum_{k=0}^{N-1}H[k]\,X[n-k],$3, and the chaotic waveform is written as

$Y[n]=\sum_{k=0}^{N-1}H[k]\,X[n-k],$4

Because the waveform is a superposition of shifted basis pulses, the matched filter is chosen as the time-reversed basis pulse $Y[n]=\sum_{k=0}^{N-1}H[k]\,X[n-k],$5. Sampling this response yields FIR coefficients $Y[n]=\sum_{k=0}^{N-1}H[k]\,X[n-k],$6, and the digital realization

$Y[n]=\sum_{k=0}^{N-1}H[k]\,X[n-k],$7

is described as exactly a sequential matched filter: each new sample $Y[n]=\sum_{k=0}^{N-1}H[k]\,X[n-k],$8 is processed with a sliding convolution over the coefficient sequence $Y[n]=\sum_{k=0}^{N-1}H[k]\,X[n-k],$9 (Bailey et al., 2015).

The implementation details are equally explicit. The chaotic oscillator operates at $v(t)=A\,x(t-t_p)+n(t),$0; the FIR has 100 coefficients; and the HDL implementation runs at $v(t)=A\,x(t-t_p)+n(t),$1. Input samples correspond to 10-bit ADC resolution, while internal computations use 32-bit fixed-point arithmetic. Coefficients are represented by SOPOT decomposition,

$v(t)=A\,x(t-t_p)+n(t),$2

with $v(t)=A\,x(t-t_p)+n(t),$3, so coefficient multiplications become shift-and-add networks rather than general multipliers. The direct-form FIR is implemented as a tapped delay line of unit delays $v(t)=A\,x(t-t_p)+n(t),$4, weighted taps, and summed outputs, supporting real-time sequential decoding with one new output sample per clock cycle (Bailey et al., 2015).

Detection in this system proceeds by matched filtering followed by threshold-based symbol reconstruction. For isolated basis pulses, the filter produces a strong correlation peak at the pulse location and low output elsewhere. For full chaotic waveforms, the output is post-processed with three thresholds $v(t)=A\,x(t-t_p)+n(t),$5, $v(t)=A\,x(t-t_p)+n(t),$6, and $v(t)=A\,x(t-t_p)+n(t),$7: crossings of the midpoint $v(t)=A\,x(t-t_p)+n(t),$8 followed by hitting $v(t)=A\,x(t-t_p)+n(t),$9 or $x(t)$0 indicate a change in the symbol sequence $x(t)$1, while values between such events are held constant. BER is the principal metric,

$x(t)$2

and simulation showed that truncating the basis pulse at $x(t)$3 cycles produced the lowest BER among the tested lengths $x(t)$4. At $x(t)$5, the matched filter still clearly detects both single basis pulses and chaotic waveforms; the simulated BER is slightly worse than the theoretical continuous-time matched filter because of FIR truncation, quantization, and thresholds chosen by visual inspection (Bailey et al., 2015).

3. Sequentiality in transform, search, and path domains

Sequential matched filtering also arises when a full matched filter is decomposed into processing stages. In FMCW radar and synthetic aperture radar, a matched-filter-based method was proposed in which stretch processing is used first to keep the ADC rate low, after which the sampled IF data are up-sampled, multiplied by $x(t)$6 to reconstruct the original chirp-domain signal, and then passed through a conventional matched filter. The compression gain is given by

$x(t)$7

so for $x(t)$8 and $x(t)$9, the analytical gain is about $A$0. In the reported simulations, the matched-filter-based method achieved about $A$1–$A$2, whereas stretch processing alone produced about $A$3–$A$4. The additional $A$5 is attributed to SAR coherent integration through Circular Global Back Projection (GBP). The paper does not use the term “Sequential Matched Filter,” but it explicitly interprets the architecture as a sequential, multi-stage realization of the matched filter effect (Movafagh et al., 2021).

A different form of sequentiality appears in discrete time–frequency shift estimation. For signals on $A$6, the matched filter matrix is

$A$7

and classical computation over the full $A$8 grid costs $A$9. By designing waveforms through Heisenberg and Weil representations, the flag and cross algorithms concentrate the matched-filter output on one or two shifted lines in the time–frequency plane. The search then becomes sequential and structured: first probe a carefully chosen line by FFT, then refine along the candidate shifted line, yielding complexity $t_p$0 under the regime $t_p$1. The paper does not use SMF as a formal label, but it describes the method as “very much in the spirit of sequential and structured matched filtering” (Fish et al., 2011).

In AFDM over doubly selective channels, sequentiality is tied to sparse path estimation in the DAFT domain. Each path induces a kernel centered at the equivalent shift

$t_p$2

and the pilot observation is written as

$t_p$3

Because the columns of $t_p$4 are approximately orthogonal, the joint ML problem simplifies to sums of per-path matched-filter energies. The estimator proceeds sequentially: find the strongest path by MF, estimate the integer and fractional parts of its equivalent shift, reconstruct that path, subtract it from the residual, and repeat. With two consecutive AFDM transmissions using different chirp rates $t_p$5 and $t_p$6, the delay and Doppler are recovered from

$t_p$7

The MF-GFS variant replaces uniform grid search over the fractional parameter by Generalized Fibonacci Search, reducing complexity from $t_p$8 to $t_p$9. The paper also reports that the maximum SINR loss due to fractional effects can be about $n(t)$0 and cannot be significantly reduced by increasing $n(t)$1 (Li et al., 12 Jul 2025).

4. Adaptive and learned sequential matched filters

A related development appears in full-duplex communication, where a conventional symbol-domain Hammerstein canceller becomes structurally mismatched once pulse shaping and the receiver matched filter are included. The proposed remedy is a new matched filter $n(t)$2 trained directly on distorted self-interference pilots. With pilot matrix $n(t)$3 built from the MF input samples and desired symbol vector $n(t)$4, the coefficients are chosen by least squares,

$n(t)$5

This filter is “adjusted according to the interfering signal” rather than matched to the ideal transmit pulse, and the paper presents it as directly relevant to sequential or multi-stage matched-filter architectures. Reported results show lower residual SI power than Hammerstein for all tested SNR values, a cancellation gain on the order of $n(t)$6 around $n(t)$7, and execution complexity

$n(t)$8

versus

$n(t)$9

for Hammerstein plus conventional MF (Lari, 2023).

An explicit modern use of the name appears in low-SNR cardiac pattern localisation. There, the Sequential Matched Filter is defined as a paradigm that “replaces the conventional single matched filter with a sequence of filters designed by a Reinforcement Learning agent.” For a signal $h(t)=k\,x(t_x-t),$0 and template $h(t)=k\,x(t_x-t),$1, each stage computes

$h(t)=k\,x(t_x-t),$2

and the process is cast as an MDP $h(t)=k\,x(t_x-t),$3 with final-step reward

$h(t)=k\,x(t_x-t),$4

The policy outputs a Gaussian-distributed filter $h(t)=k\,x(t_x-t),$5 using a lightweight network of roughly 150k parameters, approximately $h(t)=k\,x(t_x-t),$6, and training is performed with PPO or SAC. In evaluation, the final output is passed to find_peaks with height $h(t)=k\,x(t_x-t),$7 and distance $h(t)=k\,x(t_x-t),$8 samples, and detections are matched to ground truth with a tolerance of $h(t)=k\,x(t_x-t),$9 samples. On ear-ECG, SMF-SAC achieved precision $T(x)=x^T C^{-1}s=x^T f_s,\qquad f_s=C^{-1}s,$0, recall $T(x)=x^T C^{-1}s=x^T f_s,\qquad f_s=C^{-1}s,$1, and $T(x)=x^T C^{-1}s=x^T f_s,\qquad f_s=C^{-1}s,$2-1 $T(x)=x^T C^{-1}s=x^T f_s,\qquad f_s=C^{-1}s,$3; on arrhythmia ECG, it achieved precision $T(x)=x^T C^{-1}s=x^T f_s,\qquad f_s=C^{-1}s,$4, recall $T(x)=x^T C^{-1}s=x^T f_s,\qquad f_s=C^{-1}s,$5, and $T(x)=x^T C^{-1}s=x^T f_s,\qquad f_s=C^{-1}s,$6-1 $T(x)=x^T C^{-1}s=x^T f_s,\qquad f_s=C^{-1}s,$7. Average processing times per $T(x)=x^T C^{-1}s=x^T f_s,\qquad f_s=C^{-1}s,$8 segment were $T(x)=x^T C^{-1}s=x^T f_s,\qquad f_s=C^{-1}s,$9 for SMF-PPO and $C$0 for SMF-SAC (Tian et al., 29 Aug 2025).

5. Detection logic, noise models, and implementation trade-offs

Across SMF variants, detection theory and threshold setting remain central. In the Gaussian matched-filter model, the decision statistic has known distributions under $C$1 and $C$2, allowing analytic false-alarm and detection probabilities. When the signal position is unknown, however, detection is based on the maximum of the sliding matched-filter output, and the distribution of local maxima is not Gaussian. For long scans, the relevant probability is the probability that the maximum exceeds threshold,

$C$3

not the pointwise Gaussian tail alone. The same review also emphasizes that the matched filter maximizes SNR independently of the noise distribution, whereas Neyman–Pearson optimality holds only under Gaussian noise; for non-Gaussian noise, a local optimal detector uses

$C$4

This establishes an important boundary condition for SMF practice: sequential evaluation does not remove the need for correct noise modeling (Vio et al., 2021).

Implementation trade-offs are equally visible in hardware SMFs. In the reverse-time-chaos FIR, BER degradation relative to the theoretical continuous-time matched filter is explicitly attributed to finite basis-pulse truncation, quantization of data and coefficients, and thresholds chosen by visual inspection. Yet the same implementation demonstrates that a 100-tap, 32-bit fixed-point, SOPOT-based direct-form FIR at $C$5 can detect chaotic signals in AWGN at $C$6 and reconstruct the underlying symbol sequence in agreement with numerical simulation. This suggests that practical SMF design is often dominated by coefficient quantization, truncation length, post-processing thresholds, and the balance between hardware simplicity and correlation fidelity (Bailey et al., 2015).

6. Terminological ambiguity and conceptual boundaries

SMF is not a stable acronym across fields. In the bounded-noise state-estimation literature, SMF means Set-Membership Filter rather than Sequential Matched Filter. That framework is built on uncertain variables, the law of total range, and an equivalent Bayes’ rule; under a non-stochastic Markov condition it is fundamentally equivalent to the Bayes filter, while outside that condition the classical SMF is not optimal and only gives an outer bound on the optimal estimation (Cong et al., 2020). This usage concerns recursive feasible-state-set estimation, not matched filtering for detection.

Within matched-filter research itself, the acronym is also unevenly used. The FMCW-SAR work explicitly states that it does not use the term “Sequential Matched Filter,” even though it interprets its stretch-process, reconstruction, and matched-filter chain as a kind of SMF; the fast time–frequency shift work similarly frames its contribution as a structured sequential search rather than a narrow SMF formalism (Movafagh et al., 2021, Fish et al., 2011). The term is therefore best understood as a conceptual umbrella for matched-filter realizations that are sequential in time, stage, hypothesis space, or path decomposition, rather than as a single universally standardized architecture.