LSRC: Near-Field Localization via Ray Clustering

• The paper introduces LSRC, which leverages grid-free pursuit and Newton refinement to cluster subregion rays for super-resolution target localization.
• It employs FFT-based atomic selection and energy maximization to efficiently extract target parameters from sparse, resource-constrained measurements.
• Performance evaluations indicate near-perfect detection probability and resolution of off-grid targets, with RMSE approaching the Cramér–Rao bound using iterative refinement.

Near-field localization via Subregion Ray Clustering (LSRC) refers to the process of estimating target parameters—such as range and velocity—with super-resolution in scenarios where measured electromagnetic or acoustic rays interact within subregions of the sensing domain, typically under non-ideal, resource-sparse conditions like those encountered in modern radars, integrated sensing-and-communications (ISAC), and MIMO channel estimation. The algorithmic approach in recent high-resolution methods is dominated by atom clustering and Newtonized grid-free pursuit, encapsulated by the generalized Newtonized Orthogonal Matching Pursuit (NOMP) family. While the literature does not mention "LSRC" by name, all elements of near-field, subregional clustering of measurement rays as foundational to super-resolution detection and localization are fully realized in the high-resolution, grid-free, multistage frameworks surveyed below (Shah et al., 2024, Zhu et al., 2018, Mamandipoor et al., 2015, Zeng et al., 2024).

1. Measurement Model and Subregion Ray Abstraction

Localization in the near-field is fundamentally governed by the system’s expansion of the measurement domain into discrete but often sparse subregions (e.g., nonzero subsets of time-frequency, space-angle, or sensor snapshots) upon which the data acquisition occurs. In sparse OFDM ISAC systems, received data after cyclic-prefix removal and DFT yields

$$Y_{n,m} = \sum_{k=1}^K \beta_k X_{n,m} e^{-j2\pi n \Delta f \tau_k} e^{j2\pi m T_o \alpha_k} + Z_{n,m}$$

where each $(n,m)\in\Omega_s$ pertains to a subregion defined by resource occupancy constraints (e.g., sparse time-frequency slots in V2X sidelink radar) (Shah et al., 2024). Stacking nonzero entries forms the measurement vector $\mathbf h_s$, while the corresponding continuous atomic responses

$$\mathbf a(\tau, \alpha) = [e^{-j2\pi n \Delta f \tau} e^{j2\pi m T_o \alpha}]_{(n,m)\in\Omega_s}$$

provide a parametric mapping of the subregion rays in the phase domain. In MMV regimes and generalized channel models, equivalent stacking and response formation applies, supporting direct mapping of rays to subregions in array, time, or frequency domains (Zhu et al., 2018, Zeng et al., 2024).

2. Atom Selection, Clustering Strategy, and Energy Maximization

Super-resolving targets masked by mutual interference or resource sparseness requires greedy selection of atomic responses corresponding to true subregion rays. NOMP provides this via maximization of energy capture: $$(\hat\tau_\ell, \hat\alpha_\ell) = \arg\max_{(\tau, \alpha)} |\langle \mathbf r_{\ell-1}, \mathbf a(\tau, \alpha) \rangle|^2$$ where $\mathbf r_{\ell-1}$ denotes the current residual and the atom $\mathbf a(\tau, \alpha)$ encapsulates subregion ray response. Clustering occurs implicitly as each iteration adds a dominant atomic response—which, in the geometric sense, clusters candidate rays into subregions of high target energy. The selection step is computationally optimized by using 2D-FFT convolutions over the subregion keys, reducing complexity from $O(|\Omega_s|NM)$ to $O(NM \log N + NM \log M)$ (Shah et al., 2024, Mamandipoor et al., 2015).

3. Newton Refinement: Off-grid Ray Localization

Crucial to subregion ray clustering is Newton refinement, which enables continuous localization beyond a coarse grid. For each detected atom, the local Newton update solves

$$S(\tau, \alpha, \beta) = \|\mathbf r - \beta \mathbf a(\tau, \alpha)\|_2^2$$

with respect to $(\tau, \alpha)$, utilizing the gradient and Hessian of the energy function with respect to atomic parameters. This step moves estimates off-grid to their true location within the subregion, circumventing basis mismatch and clustering rays precisely to target loci. In recent extensions, global joint Newton steps over all detected targets and parameters further refine ray clusters, often leveraging block-diagonal Hessian approximation to maintain scalability (Shah et al., 2024, Zeng et al., 2024).

4. Workflow, Algorithmic Complexity, and Subregion Resolution

The general NOMP (and QNOMP) workflow for near-field localization by subregion ray clustering follows:

1. Initialization: Stack measurements over sparse subregions, precompute dictionary atoms.
2. Iterative Atom Selection: Cluster rays by selecting dominant atoms via FFT-correlation.
3. Local Newton Refinement: Update parameters within the current subregion using Newton’s method.
4. Add to Estimated Set: Aggregate refined parameters for further clustering.
5. Global Joint Refinement: Apply blockwise quasi-Newton or Newton steps.
6. Least-squares Amplitude Update: Re-estimate reflection coefficients or amplitudes.
7. CFAR Stopping: Terminate when maximum correlation across subregions falls below statistical threshold.

Complexity per iteration is dominated by FFT-based correlation ($O(NM\log NM)$), with refinement and amplitude update at $O(R_s |\Omega_s|)$ and $O(|\Omega_s|\ell^2)$ respectively (Shah et al., 2024, Zeng et al., 2024). For MMV and MIMO scenarios, computational cost is $O(K N T \log N + R_s K N T + R_c K^2 N T)$ (Zhu et al., 2018).

5. Performance, Resolution, and Experimental Outcomes

Simulation and measurement results demonstrate that subregion ray clustering by NOMP achieves:

• Near-perfect detection probability for weak targets, robust up to strong-to-weak-peak ratios of $\approx23$ dB, outperforming OMP/SALSA/FFT-based methods.
• Resolution of two closely spaced off-grid targets (range separation as fine as 0.5 m, velocity separation $1$ m/s) identifiable only by NOMP.
• RMSE for range/velocity estimation approaches the Cramér–Rao bound when using sufficient local Newton steps ($R_s=5$–$10$).
• Convergence is rapid (1–4 s for seven iterations with typical grid resolution) and per-iteration cost is minimized by FFT acceleration (Shah et al., 2024, Mamandipoor et al., 2015, Zeng et al., 2024).
• Real-world bi-static radar trials verified LSRC principles by tracking rotating metallic spheres using 1% resource occupancy, with error minimized for higher subregion energy (larger reflector cross-section, higher resource occupancy) (Shah et al., 2024).

6. Limitations, Practicalities, and Extensions

Subregion ray clustering, as realized via high-resolution NOMP, faces key limitations:

• Computational cost at scale demands block-diagonal Hessian approximations.
• Reliable coarse initialization requires moderate grid oversampling, as Newton steps are local.
• Severe interference and extremely tight spacing degrade clustering and localization; multiple joint Newton passes can incrementally aid convergence.
• Accurate CFAR thresholding mandates a priori noise variance knowledge, though recent CFAR-NOMP variants mitigate this via reference-based estimation (Xu et al., 2022).

Extensions plausible for the LSRC paradigm include:

• Joint angle-of-arrival estimation in MIMO ISAC, leading to full 3D subregion ray clustering (Zeng et al., 2024).
• Adaptive resource selection in V2X informed by sensitivity maps derived from localization results.
• GPU/parallel acceleration of subregion-based correlation and refinement for real-time operation.
• Integration with track-before-detect for low-SNR, moving target environments.

7. Relation to Generalized Grid-Free Pursuit and Subregion Clustering

The essential linkage between LSRC and modern grid-free pursuit algorithms is that each detected and refined atom corresponds to a cluster of rays—within a measurement subregion—whose energy or likelihood maximization encapsulates the presence and location of a true target. Newton refinement and global joint updates constitute the clustering engine, driving off-grid localization and supporting super-resolution against resource and measurement sparsity. LSRC is thus not a standalone algorithm but a descriptive abstraction of the subregion and clustering logic formalized by NOMP, QNOMP, and their variants (Shah et al., 2024, Mamandipoor et al., 2015, Zeng et al., 2024).