GeoPE: Geometric Embeddings Across Domains

• GeoPE is a comprehensive framework uniting geometric positional encoding, spatial contextual embedding, and quantum control methodologies to advance various scientific applications.
• It employs quaternion-based 3D rotations and context-aware embeddings to capture true Euclidean distances, significantly enhancing performance in vision and graph neural networks.
• GeoPE supports precomputed geospatial embeddings and quantum gate synthesis, achieving state-of-the-art accuracy and accelerated convergence across multiple domains.

GeoPE is an acronym found in multiple advanced research contexts, including geometric positional encodings for neural architectures, context-aware geographic embeddings for spatial graph learning, precomputed geospatial embeddings for satellite data products, and geodesic pulse engineering for quantum optimal control. The term GeoPE therefore encompasses fundamental innovations in structured representations of spatial, geometric, and manifold-valued data across computer vision, geospatial AI, graph learning, and quantum information. This article delineates the principal methodologies and technical foundations for each major usage, referencing the key published works.

1. Unified Geometric Positional Embedding for Structured Tensors

GeoPE, as formulated in "GeoPE: A Unified Geometric Positional Embedding for Structured Tensors" (Yao et al., 4 Dec 2025), addresses the fundamental shortcoming of standard positional embeddings in vision transformers (ViTs) operating on structured grid data. Traditional methods such as Rotary Positional Embedding (RoPE) and axis-wise 2D RoPE variants fail to encode true Euclidean distances and non-axis-aligned relationships due to their sequence-centric, axis-factorized construction.

Quaternion-Based 3D Rotations

GeoPE introduces a geometric framework in which each spatial location is encoded as a 3D rotation parameterized by quaternions. For each image patch at position $(h, w)$, angular parameters $\theta_h$, $\theta_w$ are assigned per-frequency, and axis-specific unit quaternions ($\mathbf{j}$ for height, $\mathbf{k}$ for width) are defined: $$r_h(\theta_h) = \cos\left(\frac{\theta_h}{2}\right) + \sin\left(\frac{\theta_h}{2}\right) \mathbf{j}, \quad r_w(\theta_w) = \cos\left(\frac{\theta_w}{2}\right) + \sin\left(\frac{\theta_w}{2}\right) \mathbf{k}$$

Symmetric Rotational Combination

To ensure geometric coupling and commutativity, GeoPE computes the geometric mean in the Lie algebra $\mathfrak{so}(3)$:

• Take $\log$ maps to obtain axis vectors
• Compute their arithmetic mean $$u=\tfrac{1}{4} (\theta_h \mathbf{j} + \theta_w \mathbf{k})$$
• The composite rotation's angle is $\Theta = \frac{1}{2}\sqrt{\theta_h^2 + \theta_w^2}$ with axis $\theta_h$0
• Exponential map yields the unified rotation $\theta_h$1

This construction yields relative rotation operators whose angle increases proportionally to the Euclidean distance $\theta_h$2, correctly penalizing attention between distant patches and enabling true 2D manifold encoding.

Empirical Performance

GeoPE achieves state-of-the-art accuracy on ImageNet (e.g., ViT-Small: 81.2%, surpassing APE and axis RoPE), object detection (COCO), and 3D semantic segmentation (S3DIS). It also increases network shape bias and decorrelates long-range sequential artifacts. Ablations confirm the necessity of the log-exp mean for isotropy; axis-multiplied variants degrade geometric fidelity (Yao et al., 4 Dec 2025).

2. Positional Encoder for Graph Neural Networks on Geographic Data

In geographic graph learning, GeoPE refers to the context-aware vector embedding that augments node features in a GNN with learnable mappings from geodetic coordinates (Klemmer et al., 2021).

Methodology

• Each point's latitude-longitude pair $\theta_h$3 is passed through a multi-scale sinusoidal encoder at $\theta_h$4 log-spaced scales, generating a tensor of $\theta_h$5 frequencies.
• The stacked sinusoids are processed through a shallow MLP to produce a $\theta_h$6-dimensional embedding $\theta_h$7.
• This is concatenated to other node features and input into a graph backbone (GCN, GraphSAGE, GAT).
• Training incorporates an auxiliary spatial autocorrelation task: predict local Moran's $\theta_h$8 for the target variable, regularizing the backbone to respect local spatial smoothness.

Results

On regression benchmarks (California Housing, global air temperature), PE-GNN architectures match or exceed Kriging and Gaussian Processes, with up to 80–84% MSE reduction over vanilla GNNs. The auxiliary autocorrelation objective provides an additional 1–3% improvement in select regimes (Klemmer et al., 2021).

3. Precomputed Geospatial Embeddings (GeoPEs) as Earth Data Products

GeoPE is used in the context of large-scale geospatial precomputed embeddings—the static representations output by frozen geospatial foundation models (GFMs) such as SatCLIP or Major TOM—for use in downstream Earth observation workflows (Fang et al., 19 Jan 2026). These embeddings are rapidly emerging as essential data assets, decoupling the inference-intensive computation from subsequent applications.

Three-Layer Taxonomy

• Data: Precomputed vectors at location, patch, or pixel granularity. Examples: location $\theta_h$9 (SatCLIP), patch-level (Earth Index), and pixel-level (Presto, Tessera).
• Tools: Benchmarks and open challenges (NeuCo-Bench, GeoINRID) analyze clustering, intrinsic dimension, and retrieval fidelity of GeoPEs.
• Value: Enables dense mapping tasks (cropland, land cover) and zero-shot semantic retrieval.

Interoperability challenges

GeoPE products are hindered by:

• Heterogeneous storage formats (GeoParquet, GeoTIFF, .pt, .npy), metadata inconsistencies, and CRS misalignments.
• Gaps in reproducibility—lack of code, pipeline, and data versioning.
• Cumbersome embedding generation—users must manually gather models, weights, and preprocessing scripts.

TorchGeo API Standardization

Extension of TorchGeo provides drop-in loading, stacking, and querying of major GeoPE products as first-class raster datasets. This enables unified downstream analysis and benchmarking without bespoke I/O or pipeline logic, supporting workflows such as semantic patch retrieval and land-cover classification.

Design Principles

Calls for the field include expanding input data diversity, integrating uncertainty and provenance metadata, standardizing cloud-native formats (COG, GeoZarr), and open publication of benchmarks and model weights to support transparent evaluation (Fang et al., 19 Jan 2026).

4. Geodesic Pulse Engineering (GEOPE) for Quantum Optimal Control

GeoPE also designates Geodesic Pulse Engineering, an algorithm for quantum gate synthesis on $\theta_w$0 with hardware constraints (Lewis et al., 22 Aug 2025).

Formalism and Algorithm

• Control goal: Synthesize a target unitary $\theta_w$1 by tuning time-dependent controls $\theta_w$2 over $\theta_w$3 intervals.
• The system state propagates via a sequence of exponentials $\theta_w$4.
• At each optimization step:
  • Compute the exact geodesic direction in $\theta_w$5 from current $\theta_w$6 to $\theta_w$7 by taking $\theta_w$8.
  • Project this direction onto the span of accessible control Jacobians $\theta_w$9 via least-squares
  • Normalize and take an optimal step size by line search in fidelity $\mathbf{j}$0.
  • If the step stalls, execute a Gram–Schmidt perturbation orthogonal to $\mathbf{j}$1.
• Principal advantage: Each update maximally decreases the Riemannian distance to the target unitary along available directions.

Performance

GEOPE converges to high-fidelity gate solutions up to an order of magnitude more rapidly than GRAPE, particularly for large multi-qubit gates (e.g., Toffoli, QFT). In numerical experiments, GEOPE achieved $\mathbf{j}$2 success within 13–50 iterations for 3–5 qubit gates, compared to 200+ for GRAPE; GRAPE frequently failed outright on high-dimensional targets. The method is readily implemented in autodiff frameworks with per-iteration complexity $\mathbf{j}$3 (Lewis et al., 22 Aug 2025).

5. Distinctions, Limitations, and Use Contexts

Table: Core GeoPE Variants

Context	Key Principle	Core Reference
Geometric Positional Embedding (ViT)	Symmetric 3D quaternion rotation	(Yao et al., 4 Dec 2025)
Graph Neural Network GeoCoding	Contextual geo-embedding + Moran's	(Klemmer et al., 2021)
Precomputed Geospatial Embeddings	GFM-produced, dataset/product APIs	(Fang et al., 19 Jan 2026)
Quantum Geodesic Pulse Engineering	Riemannian geodesic control update	(Lewis et al., 22 Aug 2025)

Each variant is tailored to substantially distinct data modalities and architecture classes: grid tensors, non-Euclidean graphs, raster geoscience, or quantum unitary groups. A commonality is the explicit modeling of geometric or manifold structure, but the mathematical framework—quaternion algebra, spatial context learning, data product pipelines, or differential geometry—differs.

Limitations and Outlook

• Quaternion-based positional encodings increase per-query compute and memory but provide symmetry and manifold awareness not accessible to axis-factorized alternatives (Yao et al., 4 Dec 2025).
• Graph-based GeoPEs are limited by auxiliary task hyperparameter selection and the smoothness of spatial autocorrelation in real data (Klemmer et al., 2021).
• Precomputed geospatial embeddings require global efforts for true format and benchmark standardization; many sources remain proprietary or unversioned (Fang et al., 19 Jan 2026).
• Geodesic quantum control (GEOPE) scaling is restricted by $\mathbf{j}$4 least-squares projection and the absence of global optimality guarantees (Lewis et al., 22 Aug 2025).

A plausible implication is that future GeoPE research may further unify geometric encoding and manifold-aware optimization principles across deep learning, geospatial science, and quantum information, subject to the constraints inherent to each application domain.