Delay–Doppler Matched Filtering

Updated 7 March 2026

Delay–Doppler matched filtering is a technique that uses the 2D ambiguity function to pinpoint target delay and Doppler shifts from known or pseudo-random waveforms.
Algorithmic innovations, including sparsity exploitation and separable estimation, significantly reduce computational complexity while preserving high estimation accuracy.
Optimized waveform and filter designs, such as 2D-sinc interpolation for fractional delay–Doppler estimation, enhance performance in radar, wireless, and OTFS/ISAC transceivers.

Delay–Doppler matched filtering is the core operation for extracting target delay and Doppler parameters in radar, wireless channel estimation, and integrated sensing and communications (ISAC) systems. The method exploits the parametric structure of time-delayed and frequency-shifted echoes of known or pseudo-random waveforms, producing a multidimensional correlation surface—known as the ambiguity function—whose peaks correspond to physical channel parameters. Algorithmic and architectural refinements have enabled dramatic reductions in computational complexity and enhanced estimation performance, particularly for high-dimensional grids, highly sparse channels, and distributed or passive systems.

1. Delay–Doppler Matched Filter Fundamentals

The generic signal model considers the received signal as a superposition of scaled, delayed, and Doppler-shifted versions of a known transmit sequence. In continuous time, with baseband transmit waveform $s(t)$ , a target or multipath component with complex amplitude $\alpha$ , delay $\tau$ , and Doppler shift $\nu$ produces a return:

$r(t;\tau,\nu) = \alpha\, s(t - \tau)\, e^{j 2\pi \nu t}$

The canonical matched filter forms the two-dimensional (2D) cross-ambiguity function,

$A_{r,s}(\tau, \nu) = \int r(t)\, s^*(t-\tau)\, e^{-j2\pi\nu t}\,dt$

The ambiguity function quantifies the response of the matched filter at each delay-Doppler test point, sharply peaking at the actual target parameters for well-designed ( $\approx$ impulse-like ambiguity) waveforms (Jitsumatsu, 2023).

In discrete domains, let $x[n]$ denote the length- $N$ transmit sequence, and $R[n]$ the observed signal. The matched filter matrix is:

$\mathrm{MF}(R, x)[\tau, \nu] = \sum_{n=0}^{N-1} R[n]\, x^*[n+\tau]\, e^{-2\pi i\, \nu n / N}$

Under typical random or pseudo-random waveform designs, $\mathrm{MF}(x, x)[\tau, \nu]$ exhibits a sharp peak at $(0,0)$ and low side-lobes elsewhere (Fish et al., 2012).

2. Classical and Modern Complexity Reduction

The direct computation of the 2D matched filter surface across an $N \times N$ delay-Doppler grid is $\mathcal{O}(N^2 \log N)$ using FFT-based “line-by-line” evaluation. For sparse multipath or radar channels, group-theoretic sequence design and the “flag method” permit further acceleration to $\mathcal{O}(m N \log N)$ for $m$ -path scenarios, by restricting the search to lines (Heisenberg–Weyl group structure) and exploiting orthogonality along group-invariant subspaces (Fish et al., 2012).

For active radar and ISAC scenarios, similar acceleration is achieved by batchwise processing, sifting out dominant interference components, and restoring parameter separation via sequential estimation (cf. passive radar separable estimation below).

3. Matched Filtering in Distributed and Passive Radar: Separable Estimation

In distributed passive radar, the reference (direct-path) and surveillance (echo) signals are available at each spatial node. Following suppression of direct-path and clutter via least squares projection, the post-cancellation ambiguity function for target delay $\tau$ and Doppler $\omega$ is formulated as

$P_m(\tau, \omega) = \frac{|\hat{a}_m(\tau, \omega)^H \Pi_m^\perp y_m|^2}{\hat{a}_m(\tau, \omega)^H \Pi_m^\perp \hat{a}_m(\tau, \omega)}$

where $\Pi_m^\perp$ projects onto the subspace orthogonal to interference, and $\hat{a}_m(\tau,\omega)$ is the steering vector for batch $m$ .

Conventionally, a 2D search over $(\tau, \omega)$ is required—computationally prohibitive for large grid sizes. Under a slow-target (small Doppler) approximation, the delay and Doppler can be efficiently estimated separably:

Perform a 1D maximization over $\tau$ using a Doppler-collapsed cost function
Estimate Doppler by linear regression of complex coefficients across batches at fixed $\hat{\tau}$

This yields computational and data-fusion savings of factor $N_\omega$ per node (reduction from $\mathcal{O}(M Q N_\tau N_\omega)$ to $\mathcal{O}(M Q N_\tau)$ for $M$ batches of length $Q$ ) and enables local Doppler estimation with reduced backhaul (Viberg et al., 22 Jan 2026).

Processing Approach	Complexity	Accuracy (Delay/Doppler)	Communication Overhead
2D Matched Filter (incoherent batch)	$\mathcal{O}(M Q N_\tau N_\omega)$	CRLB (delay and Doppler within each batch)	2D grid per node
Separable 1D+1D Approach	$\mathcal{O}(M Q N_\tau) + \mathcal{O}(M)$ (regression)	CRLB (delay $\approx$ 2D), Doppler improved over 2D	1D map per node; local regression

For sufficiently large $Q$ (batch size), the CRLB for delay is attained, and Doppler estimation is improved by coherent accumulation across all batches (effective aperture increase).

4. Waveform and Filter Optimization for Delay–Doppler Estimation

Optimal delay–Doppler matched filtering requires waveform design that ensures sharp ambiguity function concentration and low sidelobes. Joint transmit/receive filter optimization under a weighted mean squared error (WMSE) or Bayesian CRLB criterion yields nearly rectangular Pareto fronts for the achievable $(\text{CRLB}_\tau, \text{CRLB}_\nu)$ . Increasing the transmit bandwidth $B$ predominantly benefits delay estimation, while increasing observation duration $T_0$ quadratically improves Doppler estimation accuracy.

Transmit and receive filters can be designed via an alternating eigendecomposition approach. For sub-Nyquist scenarios ( $B > f_s$ ), it is possible to concentrate transmit energy at high frequencies (for minimum delay CRLB) and at time edges (for minimum Doppler CRLB), or to hybridize for joint optimality (Lenz et al., 2017).

Practical rules:

Delay estimation: maximize waveform bandwidth
Doppler estimation: maximize observation duration, energy at temporal endpoints
Near-optimal joint designs result from alternating optimization in the frequency domain, verified with Monte Carlo simulations (delay NMSE gain $> 20$ dB, Doppler NMSE gain $>4$ dB over rectangular reference pulses).

5. Fractional Delay–Doppler Estimation

Finite sampling yields main lobe widening in the 2D ambiguity function, limiting grid-based estimation precision. Local 2D-sinc (or Dirichlet) interpolation exploits the analytical form of the ambiguity near its maximum to estimate fractional offsets $\epsilon_t, \epsilon_f$ :

$\taû= (\hat{\ell}_d + \hat{\epsilon}_t) T_s, \qquad \nû = (\hat{k}_D + \hat{\epsilon}_f)\, \Delta f$

where $(\hat{\ell}_d, \hat{k}_D)$ are the peak grid indices, and fractional offsets are estimated by paraboloid fitting or nonlinear least squares on the local ambiguity patch (Jitsumatsu, 2023).

This approach approaches the actual main-lobe–determined resolution limit, with RMSE reaching $0.0061T_s$ in delay and $0.0676 \Delta f$ in Doppler at high SNR, and offers computational advantages—closed-form quadratic interpolation is especially efficient ( $<$ 10 µs per estimate).

6. Delay–Doppler Matched Filtering in OTFS and ISAC Transceivers

Orthogonal Time Frequency Space (OTFS) and Zak-OTFS transceivers natively operate in the discrete Delay–Doppler (DD) domain. The system is modeled as a twisted convolutional operation, where transceiver filtering and the channel are all described as (possibly quasi-periodic) DD-domain convolutions:

$y[k, l] = \sum_{k', l'} h_\text{eff}[k-k', l-l']\, e^{j2\pi k' (l - l') / (MN)}\, x[k', l'] + n[k, l]$

The effective channel $h_\text{eff}$ results from the convolution of the transmit/receive pulse shapes and the physical DD-spread channel.

Matched filtering in the DD domain can be implemented via correlation or twisted convolution; exact closed-form I/O relations are available for relevant pulse shapes (sinc, Gaussian). Channel-matched receive filtering yields the best SNR and BER, with 1–2 dB gains over matched or identical filters for highly time-varying Vehicular-A channels (Das et al., 26 Apr 2025).

7. Applications and Outlook

Delay–Doppler matched filtering is foundational for:

Passive or distributed radar networks (with batch-based separable estimation for scalability)
Next-generation communication systems leveraging high-resolution parametric channel models (AFDM, OTFS, Zak-OTFS)
ISAC applications requiring simultaneous data transmission and accurate environment sensing

Key advances include:

Separable estimation in passive radar yielding order-of-magnitude complexity and data reduction (Viberg et al., 22 Jan 2026)
Filter/waveform co-design for near-fundamental accuracy, even under sub-Nyquist constraints (Lenz et al., 2017)
Fractional parameter estimation via 2D-sinc fitting for sub-bin accuracy (Jitsumatsu, 2023)
Twisted convolution formalism for unified comms-radar transceiver architectures (Das et al., 26 Apr 2025, Mohammed et al., 2023)
Group-theoretic methods for sparse channel estimation at almost linear complexity (Fish et al., 2012)

Ongoing research includes low-SNR robust estimation (e.g., deep-learning guides for path detection), further acceleration via search algorithms (generalized Fibonacci, hierarchical refinement), and tight integration of delay–Doppler matched filtering with modern MIMO and massive sensing architectures.

Primary References:

"Separable Delay And Doppler Estimation In Passive Radar" (Viberg et al., 22 Jan 2026)
"Joint Transmit and Receive Filter Optimization for Sub-Nyquist Delay–Doppler Estimation" (Lenz et al., 2017)
"2D Sinc Interpolation-Based Fractional Delay and Doppler Estimation Using Time and Frequency Shifted Gaussian Pulses" (Jitsumatsu, 2023)
"Delay-Doppler Channel Estimation with Almost Linear Complexity" (Fish et al., 2012)
"Closed-Form Expressions for I/O Relation in Zak-OTFS with Different Delay-Doppler Filters" (Das et al., 26 Apr 2025)