Anchor-Based Distance Encodings

• Anchor-based distance encodings are defined by representing objects with fixed-length vectors computed from their distances to a finite set of reference anchors, ensuring invariance and interpretability.
• They employ various domain-specific anchor selection and placement strategies to capture key geometric and topological features in areas such as geospatial analysis, 3D vision, and graph learning.
• Empirical evaluations demonstrate that these encodings boost performance metrics and computational efficiency compared to traditional benchmarks across multiple application domains.

Anchor-based distance encodings are a broad class of geometric and graph-based embedding techniques in which the representation of an object is defined by its metric-based relationships (typically distances or distance transforms) to a finite, fixed set of reference points called anchors. These encodings transform structural, spatial, or abstract entities—such as geospatial geometries, point clouds, graph nodes, or even class logits—into fixed-length vectors, enabling them to interface directly with neural and statistical models. The mathematical formulation, interpretability, and downstream utility of anchor-based distance encodings have made them foundational in various domains, including geospatial machine learning, 3D computer vision, graph representation learning, and open set recognition. Theoretical and empirical results demonstrate that anchor-based distance encodings provide invariances, preserve salient topological information, and can approximate a range of classical geometric and spectral features.

1. Mathematical Formulation and Variants

The general form of an anchor-based distance encoding for a domain $\mathcal{X}$ (e.g., $\mathbb{R}^2$, $\mathbb{R}^3$, node set of a graph, or logit space) is as follows:

• Select a set of $k$ anchor points $A = \{a_1, \dots, a_k\}$ in the space of interest.
• For each object $x \in \mathcal{X}$, compute the feature vector $f(x) = [d(x,a_1), \dots, d(x,a_k)]$, where $d(\cdot, \cdot)$ is a domain-appropriate metric or generalized distance.

Domain-specific instantiations include:

• Multi-Point Proximity (MPP) Encoding for Vector Geometries: For vector-mode geospatial features (points, lines, polygons), MPP encoding assigns each geometry $g$ in a planar region of interest (ROI) the vector $[\exp(-\operatorname{dist}(g, r_i)/s)]_{i=1}^N$, where $r_i$ are grid- or tessellation-based anchors and $s$ is a scaling factor (Collins, 5 Jun 2025).
• ESCAPE Anchor Point Encoding for 3D Shapes: For point clouds, ESCAPE selects $k$ shape-adaptive anchors via farthest-point sampling refined by local curvature, and represents each point $x$ in the cloud as $[\|x - a_j\|_2]_{j=1}^k$ (Bekci et al., 2024).
• Anchor Distance for 3D Object Detection: Instead of 2D bounding-box anchors, distance-based anchors cluster the training data by object depth and associate each predictor with a reference anchor distance, predicting a multiplicative residual from this anchor (Yu et al., 2021).
• Anchor-based Distance Encoding (DE) in Graphs: Each node $v$ is mapped to $[d_\mathrm{sp}(v,a_1), ..., d_\mathrm{sp}(v,a_k)]$, with $d_\mathrm{sp}$ the shortest-path distance; other variants use powers of random-walk matrices or PageRank scores (Li et al., 2020).

2. Anchor Selection and Placement Strategies

Anchor selection determines the representational power, invariance properties, and resolution of the encoding.

• Uniform or Structured Layouts: For geometries, anchors may be distributed on a grid, hexagonal tessellation, or via k-means clustering in the ROI (Collins, 5 Jun 2025).
• Greedy K-center/Farthest-Point Sampling in Graphs and Point Clouds: Greedy k-center in graphs ensures spread over the topology; in 3D shapes, farthest-point sampling (FPS) combined with curvature maximization yields anchors that both cover and capture salient geometry (Bekci et al., 2024, Li et al., 2020).
• Clustering in Label or Metric Space: For anchor distances in object detection, k-means is performed in target distance space to minimize intra-cluster variance (Yu et al., 2021).
• Fixed Coordinate Anchors in Representation Space: In open set recognition, class anchors are fixed to scaled one-hot vectors in logit space, imposing isotropic geometry (Miller et al., 2020).

A table summarizing common anchor placement strategies:

Domain	Anchor Placement	Rationale
Geospatial	Grid/hex/k-means	Homogeneous spatial coverage
Point Clouds	FPS + curvature	Geometric salience, coverage
Graphs	Random, high-degree, greedy k-center	Topological coverage, diversity
Logit/Class Space	Orthonormal basis vectors	Tight per-class clustering

3. Mathematical and Representational Properties

Anchor-based distance encodings generally inherit key properties from their construction:

• Invariance: Depending on the metric and domain, encodings can be invariant to transformations such as rotation, translation, permutation, or rigid motion. For example, the ESCAPE distance encoding is strictly rotation-invariant and injective up to rigid motion for $k\geq 4$ anchors in general position (Bekci et al., 2024).
• Continuity and Stability: Encodings using continuous kernels (e.g., $\exp(-d/s)$ in MPP) yield representations that vary continuously as the geometry changes, unlike indicator- or raster-based encodings (Collins, 5 Jun 2025).
• Shape-Centricity and Discriminativity: By design, anchor distances depend only on the structure and not on the ordering or sampling density of the representation (e.g., invariance to vertex ordering in MPP (Collins, 5 Jun 2025)).
• Expressive Power: In the context of graphs, DE-GNNs using anchor-based SPD features can distinguish substructures that are indistinguishable by standard 1-WL GNNs, and this discrimination is provably better than higher-order WL in certain regimes (Li et al., 2020).

4. Integration with Machine Learning Architectures

Anchor-based distance encodings interface with models via concatenation, feature embedding, or as core components of the prediction pipeline:

• Geospatial ML Pipelines: MPP vectors are concatenated with tabular attributes and fed to models such as random forests, MLPs, GNNs, or spatio-temporal transformers (Collins, 5 Jun 2025).
• Point Cloud Transformers: Distance-encoded points are linearly embedded and processed by self-attention networks for shape completion, with decoders predicting distance matrices for reconstructing 3D structures (Bekci et al., 2024).
• Single-Shot Object Detection: Predictors specialized to anchor distances output regression residuals, enabling real-time 3D distance estimation (Yu et al., 2021).
• Graph GNNs: Anchor distance vectors are appended as node features (DE-GNN), or intervene in message aggregation functions (DEA-GNN), leading to more expressive set and subgraph representations (Li et al., 2020).
• Open Set Recognition: Neural nets trained with Class Anchor Clustering (CAC) minimize tuple and absolute anchor distance losses, and replace softmax with distance-based rejection at inference (Miller et al., 2020).

5. Theoretical Connections to Spectral and Diffusion Geometry

Recent work formally interprets anchor-based distance encodings as low-rank surrogates for spectral or diffusion-based positional encodings in graphs (Yan et al., 8 Jan 2026).

• Trilateration Map: Given $m+1$ anchors and their truncated Laplacian eigen-coordinates $P_i := \Phi^{(m)}(a_i)$, the truncated spectral embedding $\Phi^{(m)}(v)$ of any node $v$ can be reconstructed (up to error) from the set of transformed anchor distances via solving a system of quadratic (then linearized) equations: $A \cdot z = b^*$, with $A$ the anchor-difference matrix (Yan et al., 8 Jan 2026).
• Nyström Approximation: Anchor distance encodings approximate the full spectral or heat-kernel geometry via a three-step Nyström procedure, greatly reducing computational burden while preserving accuracy.
• Error Bounds: With appropriate choices of the radial transform $\psi$, separation of anchor positions, and match to the diffusion distance, O$(\varepsilon)$-close recovery of spectral coordinates is guaranteed for all nodes within a neighborhood radius on random regular graphs (Yan et al., 8 Jan 2026).

6. Empirical Evaluation and Comparative Performance

Across application domains, anchor-based distance encodings robustly outperform traditional or naive benchmarks:

• Geospatial Learning: MPP encodings enable MLPs to retrieve length/area/orientation/complexity of random vector shapes with $R^2 \approx 0.98$, substantially exceeding raster tile-indicator (DIV) encodings ($R^2 \lesssim 0.85$), and yield uniformly higher ROC-AUC for spatial predicates even at coarse resolutions (Collins, 5 Jun 2025).
• 3D Shape Completion: ESCAPE achieves rotation-robust completion with Chamfer-$L_1$ distances of 10.58 (vs 26.65–92.15 for non-equivariant baselines), and exhibits stability to noise and missing data (Bekci et al., 2024).
• Graph ML: DE-GNN-anchored node embeddings drive gains of up to 15% accuracy over GIN on structural tasks, meaningfully outperforming specialized baselines on link and triangle prediction (Li et al., 2020). In molecular graphs, anchor-based Nyström approximations reach AUROC/F1 $\geq$ 0.976/0.927, close to Laplacian PE (Yan et al., 8 Jan 2026).
• Single-Shot 3D Distance Estimation: In KITTI object detection, anchor-distance regression achieves the lowest reported RMSE (2.08 m), with error curves invariant to object distance, and executes at $\sim 30$ FPS (Yu et al., 2021).
• Open Set Recognition: CAC boosts AUROC by up to 17% over conventional distance-based classifiers on challenging datasets such as TinyImageNet, without compromising closed-set accuracy (Miller et al., 2020).

7. Implementation Considerations and Future Directions

Practical guidelines for anchor-based distance encodings emphasize:

• Choosing Anchor Number and Layout: Tradeoff between encoding dimension and geometric detail; cross-validation or domain heuristics (e.g., anchor spacing $\Delta$ for MPP, $k$ clusters for object detection, coverage for GNNs) (Collins, 5 Jun 2025, Yu et al., 2021, Li et al., 2020).
• Handling Large Domain Size: Sparse or localized computation, thresholding of kernel responses, hierarchical tiling, or subgraph-limited aggregation for scalability (Collins, 5 Jun 2025, Li et al., 2020).
• Extension to New Modalities: Variations under active research include anchor-based encodings for full-3D bounding boxes, orientational or scale anchors, and dynamic or learned anchor adaptation during training (Yu et al., 2021).
• Limiting Factors: Isotropic/axis-aligned anchor arrangements may not capture semantic inter-class relations; regular graphs or symmetric domains may reduce discriminative power unless combined with node/edge features (Miller et al., 2020, Li et al., 2020).
• Algorithmic Optimizations: Radial transforms, rank reduction, and explicit trilateration enable approximation of diffusion geometry at near-linear complexity, with empirical accuracy close to full spectral embeddings (Yan et al., 8 Jan 2026).

Anchor-based distance encodings provide a unifying framework for geometric and relational representation in machine learning, synthesizing invariance, interpretability, and computational tractability across diverse tasks and data manifolds.