Localization-Based Sparse Merging

• Localization-Based Sparse Merging is a technique that extracts spatially or functionally localized subregions to reconstruct global models with enhanced accuracy and interpretability.
• The approach employs sparse selection and merging strategies—such as L&S model merging, diffusion-based radio map reconstruction, and covariance-domain recovery—to optimize performance while reducing redundancy.
• Applications span deep neural networks, wireless localization, super-resolution microscopy, and MIMO processing, offering significant improvements in efficiency, parameter reduction, and signal recovery.

Localization-Based Sparse Merging refers to a class of techniques in which sparse, spatially or functionally localized subregions—determined via a localization criterion—are extracted from multiple sources and merged to reconstruct or synthesize useful global structures or models with enhanced efficiency, interpretability, or accuracy. These methods exploit the fact that informative content or discriminative signal often resides in compact regions (either in parameter, measurement, or spatial domains), permitting synergy through judicious merging of only essential, non-redundant components. Prominent instantiations appear in domains ranging from neural network model merging, physical radio map reconstruction, multi-emitter wireless localization, super-resolution microscopy, to sparse MIMO array processing.

1. Core Principles and Rationale

Localization-based sparse merging is grounded in the observation that for a range of scientific and engineering problems, either: (i) only a small fraction of the parameters or measurements define the key features of interest, or (ii) spatial or parameter support is sparse but information-rich. Central objectives are to:

• Localize support: Identify regions (parameter subsets, physical sensor locations, or high-contrast domains) where essential information for task performance or signal reconstruction concentrates.
• Sparse selection: Impose constraints or regularization favoring minimal, often binary or highly-sparse masks, patches, or measurement sets.
• Merging/synthesis: Combine support—often from multiple localized regions, models, or measurements—into a unified output while managing information overlap, normalization, or ambiguity.
• Maintain fidelity: Ensure that the merged construction recovers the core content (e.g., task performance, localization accuracy, signal sense) of each constituent, despite aggressive sparsification.

These principles drive algorithmic choices such as greedy, mutual- or Fisher-information–maximizing sampling; $\ell_0$ or $\ell_1$-regularized sparse selection; and compositional fusion rules (e.g., averaging, clustering, patching).

2. Algorithms and Mathematical Formulations

2.1 Localized Parameter Merging in Deep Models

The "Localize-and-Stitch" (L&S) approach exemplifies parameter-level localization and merging for multi-task model composition (He et al., 2024). Given pretrained weights $W_0 \in \mathbb{R}^d$ and finetuned models $W_i$, L&S computes task vectors $\tau_i = W_i - W_0$ and solves, for each task $i$, a sparse mask $\gamma_i \in \{0,1\}^d$:

$$\min_{S \in \mathbb{R}^d} \ell_i\left(W_0 + \sigma(S)\odot\tau_i\right) + \lambda\|\sigma(S)\|_1$$

where $\sigma(S)$ is the elementwise sigmoid, and $\odot$ is the Hadamard product. If validation data is absent, a dataless variant retains top-$k$ entries by $|\tau_i|$. Masks are then stitched using overlap-aware normalization:

$$\gamma_i'[j] = \gamma_i[j] / \left(\sum_{\ell=1}^n \gamma_\ell[j]\right)$$

yielding the merged model:

$$W_{\text{merged}} = W_0 + \sum_{i=1}^n \gamma_i' \odot \tau_i$$

2.2 Sparse Radio Map Merging via Geometric Sampling and Generative Models

In "RadioDiff-Loc" (Wang et al., 2 Sep 2025), sparse physical RSS samples are localized to obstacle vertices maximizing Fisher and mutual information. With unknown transmit power, measurements $r_i$ are normalized to $r_i^N = r_i / \max_j r_j$. A conditional diffusion model reconstructs a dense radio map from the sparse, normalized vector:

• The U-Net denoiser receives $[H, \text{mask} \cdot r^N]$ (environment and sparse data masks).
• Iterative denoising merges sparse evidence into a full-resolution map.
• Localization is performed by $\hat\theta = \arg\max_{x \in S_r} \hat M(x)$.

Sampling strategy guarantees (by submodularity) near-optimal information gain with $O(\log(1/\delta))$ vertices.

2.3 Sparse Recovery for Multi-Source Localization

In multi-source wireless localization (Chu et al., 2021), sparse merging proceeds via:

• Discretization over a grid $G$; sparse vector $s \in \mathbb{R}^N$ models $K$ sources.
• Relaxed sparse recovery:

$$\min_{0 \le s \le 1} \frac{1}{2}\|\hat r - \Phi s\|_2^2 + \lambda\|s\|_1$$

followed by adaptive thresholding (ADT) and $k$-means clustering over support. Cluster centroids and weights are merged to initialize parameter estimates for MLE refinement.

2.4 Covariance-Based Sparse Merging in Microscopy

The COL0RME approach (Stergiopoulou et al., 2021) leverages emitter stochasticity for $\ell_0$-domain support localization, then merges these supports for intensity estimation:

• Sample covariance $R_y = \Psi R_x \Psi^\top + s I$; objective:

$$\min_{r_x \ge 0,\, s \ge 0} \frac{1}{2} \|r_y - A r_x - s v_I\|_2^2 + \lambda \|r_x\|_0$$

• Alternating-minimization solves for $r_x$ (sparse support), then intensity and background are estimated on $\Omega = \{i: r_{x,i}>0\}$.

2.5 Sparse MIMO Virtual Array Construction

In sparse MIMO (Min et al., 25 Feb 2025), physical antenna elements are merged into virtual arrays via the co-array principle:

$$z = \text{vec}(R) = (A^* \odot A) r + \sigma^2 \mathbf{1}$$

Rows of $A^*\odot A$ enumerate the difference set $\mathcal{V}$. Virtual array processing reconstructs high-DoF spatial manifolds for angle estimation (e.g., via MUSIC), with fusion limited to the localization pipeline due to SNR/ambiguity tradeoffs.

3. Applications Across Domains

3.1 Multi-Task Deep Model Merging

Localization-based sparse merging allows construction of multi-task neural models from a base plus minimal, interpretable "task patches," supporting continual learning and drastic reduction in storage (e.g., 99% parameter reduction for additional tasks) (He et al., 2024). The approach outperforms global merging, e.g., RoBERTa-base 12-task merging achieves up to 75.9% average accuracy with only 1% parameter patches versus 67.5% for global arithmetic.

3.2 NLoS Radio Localization

Sparse merging at obstacle-vertex locations enables robust emitter localization in NLoS scenarios. RadioDiff-Loc achieves $<3$ m localization error with $<1\%$ sparse sampling, outperforming classical random or uniform RSS sampling by $>30\times$ in measurement cost (Wang et al., 2 Sep 2025). Generative diffusion models enhance map fidelity from highly incomplete data.

3.3 Multi-Source Wireless Sensing

SR-WAC in wireless sensor networks combines grid-based sparse recovery and weighted centroid merging to robustly estimate multiple transmitters under heavy shadow fading, improving localization RMSE by factors of 2–4 compared to MMSE baselines even with low sensor densities (Chu et al., 2021).

3.4 Super-Resolution Microscopy

Sparse merging via covariance-domain $\ell_0$ recovery in COL0RME combines localized emitter identification and merging for sub-diffraction imaging at increased SNR and reduced artifacts, with robust parameter selection and efficient implementation (Stergiopoulou et al., 2021).

3.5 High-Resolution Angle-of-Arrival Estimation

Sparse MIMO virtual arrays (nested, co-prime) merge physical antennas into high-DoF virtual manifolds, enabling $\mathcal{O}(1/M^2)$ angle resolution. The merged support expands beyond compact ULAs while virtual–physical hybrid processing optimally allocates merging for localization and avoids SNR losses for communication (Min et al., 25 Feb 2025).

4. Theoretical Guarantees and Trade-offs

Key theoretical properties established for localization-based sparse merging include:

• Submodular mutual information yields $(1-1/e)$-approximate sensor configurations—error decays logarithmically with sample count (Wang et al., 2 Sep 2025).
• In deep model merging, localized patches overlap minimally (Jaccard index $<0.05$ for most task pairs), indicating task disentanglement and reduced interference (He et al., 2024).
• In MIMO virtual arrays, main-lobe beam width scales as $\mathcal{O}(1/M^2)$, far surpassing traditional arrays. Trade-offs include SNR loss and grating lobes, remediable in the virtual localization domain (Min et al., 25 Feb 2025).
• Covariance-based merging achieves high recall and intensity fidelity in microscopy, with convergence properties traceable to non-convex regularization theory (Stergiopoulou et al., 2021).
• Storage and computation scale linearly or sub-linearly with number of merged tasks or sources owing to efficient sparse support (He et al., 2024, Chu et al., 2021).

5. Limitations, Extensions, and Open Challenges

Despite broad applicability, limitations persist:

• Model merging approaches (e.g., L&S) may require per-task finetuned weights and, for best accuracy, modest amounts of validation data; purely dataless variants need higher sparsity (e.g., $s=5-10\%$) for equivalent performance (He et al., 2024).
• Physical sparsity assumptions (vertex/edge dominance in NLoS, independence in emitter blinking in microscopy) may not hold in all settings; performance may degrade outside these regimes (Wang et al., 2 Sep 2025, Stergiopoulou et al., 2021).
• In wireless sensor networks, complexity scales cubically with grid resolution, though this is less severe than previous MMSE methods (Chu et al., 2021).
• Virtual array merging must be restricted to localization domains—applying similar merging schemes to communication signals destroys critical phase information and impairs SNR (Min et al., 25 Feb 2025).

Extensions include adaptation to Poissonian noise models in microscopy, gridless sparse recovery strategies, and improved overlap management in continual task merging.

6. Comparative Summary Table

Domain/Application	Localization Mechanism	Sparsity/Merging Rule
Model merging (He et al., 2024)	Sigmoid+$\ell_1$ mask opt./threshold	Overlap-averaged patch addition
NLoS localization (Wang et al., 2 Sep 2025)	Vertex Fisher info. maximization	Diffusion model for dense map
WSN RSS multi-source (Chu et al., 2021)	$\ell_1$-BPDN sparse recovery	k-means weighted centroid fusion
Microscopy (Stergiopoulou et al., 2021)	Covariance-domain $\ell_0$ prior	Support-based least-squares merge
Sparse MIMO (Min et al., 25 Feb 2025)	Difference co-array support	Virtual array manifold fusion

7. Implications and Future Directions

Localization-based sparse merging demonstrates that global problems—whether in learning (parameter merging), sensing (radio/optical maps), or array processing (direction-of-arrival)—can be efficiently reconstructed from compact, high-yield supports. This reduces storage, computation, and sampling requirements, while retaining or even enhancing core task performance. Anticipated future advances include dynamic, data-driven support localization, adaptive sparsity allocation, and domain-specific merging rules that further minimize overlap and redundancy while expanding applicability to new modalities and non-ideal conditions.