Matched-Filter Residual Parameterization

• Matched-filter residual parameterization is a method that extends classical filtering by jointly estimating amplitude and waveform deformations, enabling compensation for model discrepancies.
• It employs analytical two-parameter models, Student-t distributions, and neural network adaptations to robustly manage non-Gaussian noise and structured waveform deviations.
• Key results demonstrate up to 2–10× improvement in shape discrimination at low SNR and enhanced detection rates with efficient real-time computational strategies.

Matched-filter residual parameterization is a framework in which the standard matched filter, originally designed for optimal amplitude estimation in additive Gaussian noise, is analytically extended or adapted to enable the estimation or compensation of residual waveform deformations, non-Gaussian noise structure, or physical model discrepancies. This method provides explicit parameterization and inference for waveform perturbations or model inadequacies by modifying the filter architecture, likelihood function, or adaptation mechanism. Recent developments span analytical extensions for pulse shape discrimination (Cappelli et al., 2024), robustifying the filter to non-Gaussian noise via a Student-t residual model (Röver, 2011), memory-optimized ratio filtering for structured waveform deviations (Nitz et al., 25 Jan 2026), and neural-network-assisted residual learning for dynamic channel effects (Haigh, 6 Feb 2026).

1. Signal and Residual Parameterization Models

A canonical matched-filter scenario assumes detector data $w(t)$ contains a reference template pulse $T(t)$ with unknown amplitude $A$ in additive noise:

$w(t) = A\,T(t) + n(t).$

Matched-filter residual parameterization generalizes $T(t)$ to a family that covers small shape or temporal deformations, e.g.,

$$s(t;A,\varepsilon) = A\,T(t) + \varepsilon\,\Delta(t),$$

where $\Delta(t)$ is a known residual mode (such as a time-dilation deformation $\Delta(t) \approx \partial_g T(g\,t)\big|_{g=1}$) and $\varepsilon$ is a dimensionless parameter quantifying departure from the nominal template. In this two-parameter model, both amplitude $A$ and shape residual $T(t)$0 are estimated jointly from $T(t)$1 (Cappelli et al., 2024).

Alternate parameterizations include modeling residuals as arising from non-Gaussian, heavy-tailed noise (impulse- or glitch-prone environments) using a Student-t distribution for the spectral-domain residuals $T(t)$2, introducing a degree-of-freedom parameter $T(t)$3 to control the tail thickness and adaptation (Röver, 2011).

In bandlimited channels, residual parameterization is carried out in the filter domain by expressing the effective matched filter $T(t)$4 as

$T(t)$5

with $T(t)$6 learned end-to-end for each channel condition, using physically motivated waveform features as conditioning variables (Haigh, 6 Feb 2026).

2. Analytical Frameworks and Estimation Procedures

Two-Parameter Matched Filter

The extended matched filter for simultaneous amplitude and residual estimation involves the joint minimization of the frequency-domain chi-squared,

$T(t)$7

with inner product

$T(t)$8

where $T(t)$9 is the noise power spectral density. The resulting estimators $A$0 are found by solving the $A$1 normal equations:

$A$2

(Cappelli et al., 2024).

Student-t Matched Filter

For robust detection under non-Gaussian noise, the likelihood is specified as

$A$3

leading to iteratively re-weighted least-squares updates where the weights $A$4 depend on the current residuals and $A$5 (Röver, 2011). This method infers the optimal waveform parameters while adapting the influence of high-leverage spectral outliers.

Filter-domain Neural Residual Learning

For dynamic waveform compensation, the complex-valued residual filter coefficients $A$6 are parameterized by a neural network $A$7 mapping a set of $A$8 physically interpretable waveform features $A$9 to the correction vector, producing

$w(t) = A\,T(t) + n(t).$0

with training performed end-to-end over error vector magnitude (EVM) loss, and smoothness regularization to ensure physical realizability of $w(t) = A\,T(t) + n(t).$1 (Haigh, 6 Feb 2026).

3. Statistical Properties and Discrimination Capability

In the analytic two-parameter framework, $w(t) = A\,T(t) + n(t).$2 are jointly unbiased with covariance given by the inverse Fisher information matrix:

$w(t) = A\,T(t) + n(t).$3

yielding Cramér–Rao lower bounds

$w(t) = A\,T(t) + n(t).$4

with $w(t) = A\,T(t) + n(t).$5.

Compared to time-domain rise time, decay time, or reduced chi-square computed post-matched-filter, the residual parameter $w(t) = A\,T(t) + n(t).$6 captures shape deformations in a statistically efficient manner. Discrimination potential $w(t) = A\,T(t) + n(t).$7, defined as

$w(t) = A\,T(t) + n(t).$8

was shown to be at least a factor 2 greater for $w(t) = A\,T(t) + n(t).$9 than for classical metrics across low and high SNR regimes (Cappelli et al., 2024). In the Student-t case, robustness to outliers achieves higher detection probabilities in real noise at a fixed false-alarm rate compared to the standard matched filter (Röver, 2011).

4. Computational and Implementation Aspects

The cost per event for the two-parameter analytic extension is dominated by a single data FFT plus three inner products ($T(t)$0, $T(t)$1, $T(t)$2) just as in the scalar matched filter, allowing real-time implementation. For time-shift recovery, FFT-based convolution and parabolic interpolation on the $T(t)$3 surface yield the optimal delays (Cappelli et al., 2024).

In Student-t robust matched filtering, the computational cost is a small multiple (typically $T(t)$4–$T(t)$5) of the classical filter due to the need for iteratively re-weighted evaluation, converging rapidly for practical $T(t)$6 (Röver, 2011).

A distinct computational strategy is adopted in "ratio-filter dechirping," where the matched filtering operation is split into a reference template and a finite impulse response (FIR) convolution implementing the residual ratio $T(t)$7 between target and reference waveforms. This yields a dramatic reduction in memory bandwidth, shifting the computational bottleneck from DRAM/FFT to cache/FIR convolution, and achieving measured $T(t)$8 throughput improvement with straightforward GPU acceleration (Nitz et al., 25 Jan 2026).

Neural residual filter-learners in digital communications are implemented with online feature extraction, forward neural inference, and backpropagation-compatible end-to-end loss (EVM and smoothness). Hardware acceleration via FPGA (e.g., Xilinx ZCU102 + AD-FMC-DAQ3) enables real-time adaptation (Haigh, 6 Feb 2026).

5. Comparative Performance and Application Domains

Matched-filter residual parameterization has been validated in several contexts:

• Pulse shape discrimination (detector physics): Two-parameter matched filter improves shape discrimination at low $T(t)$9 ($$s(t;A,\varepsilon) = A\,T(t) + \varepsilon\,\Delta(t),$$0) by a factor of $$s(t;A,\varepsilon) = A\,T(t) + \varepsilon\,\Delta(t),$$1–$$s(t;A,\varepsilon) = A\,T(t) + \varepsilon\,\Delta(t),$$2 relative to traditional metrics and retains a factor $$s(t;A,\varepsilon) = A\,T(t) + \varepsilon\,\Delta(t),$$3 improvement at $$s(t;A,\varepsilon) = A\,T(t) + \varepsilon\,\Delta(t),$$4 (Cappelli et al., 2024).
• Gravitational-wave search (robust detection): Student-t parameterized matched filter achieves $$s(t;A,\varepsilon) = A\,T(t) + \varepsilon\,\Delta(t),$$5 higher detection rates in realistic interferometer noise at fixed low false-alarm rates, compared to the Gaussian model, through effective suppression of glitch-induced outliers (Röver, 2011).
• Waveform banks with structured physical residuals: The ratio-filter framework efficiently handles high-dimensional template banks with eccentricity, spin, or mode harmonics, transforming the main filtering operation into compact FIR kernels that enable cache-resident high-throughput filtering (Nitz et al., 25 Jan 2026).
• Bandlimited optical communications: ML-enabled residual parameterization enables adaptive compensation for bandwidth-induced pulse distortion, learning a compact, channel-specific correction to the analytical matched filter using low-dimensional waveform statistics, achieving superior EVM under severe constraints without additional latency (Haigh, 6 Feb 2026).

6. Limitations and Considerations

The analytic two-parameter expansion assumes the deformation amplitude $$s(t;A,\varepsilon) = A\,T(t) + \varepsilon\,\Delta(t),$$6 for validity; larger deformations necessitate higher-order Taylor expansion or iterative updates of base templates (Cappelli et al., 2024). In robust Student-t residual filtering, selection of $$s(t;A,\varepsilon) = A\,T(t) + \varepsilon\,\Delta(t),$$7 is data-dependent and may require Q–Q residual fit or hierarchical modeling. Bandlimit compensation by neural parameterization is bounded by feature expressivity and filter smoothness constraints; performance gracefully recovers to the analytical filter in near-ideal channels (Haigh, 6 Feb 2026).

Student-t residual models and neural filter learners introduce several hyperparameters (degrees-of-freedom $$s(t;A,\varepsilon) = A\,T(t) + \varepsilon\,\Delta(t),$$8, regularizers, network architectures). Optimal settings are application- and data-dependent, impacting the computational burden and convergence rate.

7. Connections to Broader Signal Processing Frameworks

Matched-filter residual parameterization formalizes and extends the role of residual analysis in detection, estimation, and discrimination problems. By embedding physically or statistically motivated residuals into the estimation procedure, these frameworks bridge the gap between template-based optimality and flexibility for real-world deviations—whether as analytic deformations, adaptive weighting for non-Gaussianity, FIR representations of waveform structure, or learned corrections via conditioning on efficient waveform measures. The approach generalizes classical matched filtering as a special case of $$s(t;A,\varepsilon) = A\,T(t) + \varepsilon\,\Delta(t),$$9-only parameterization and provides a mathematically grounded route for adaptive or model-augmented waveform filtering in a wide range of domains (Cappelli et al., 2024, Röver, 2011, Nitz et al., 25 Jan 2026, Haigh, 6 Feb 2026).