Geometric Deep Learning

The foundational design philosophy stating that a neural network's architecture should be equivariant to the symmetry group of its input domain. For cellular topology graphs, this means the model's output must transform predictably under permutations of node ordering. A GNN processing a base station deployment graph will produce identical resource allocation decisions regardless of how the network management system indexes the cell sites. This principle unifies CNNs (translation symmetry) and GNNs (permutation symmetry) under a single mathematical framework.

Explicit constraints baked into a model's architecture that encode the known structure of the data, dramatically reducing the search space of learnable functions. Key priors for cellular networks include:

• Permutation invariance: The ordering of nodes in the input matrix is arbitrary
• Locality: A base station's interference footprint is primarily local, not global
• Scale separation: Network-wide optimization can be decomposed hierarchically These priors prevent the model from wasting capacity learning symmetries already known to exist in the physical world.

A specialized form of equivariance ensuring predictions remain consistent under arbitrary local reference frame transformations. In MIMO beamforming, the optimal precoding vector should rotate covariantly with the choice of coordinate system at each base station. A gauge equivariant GNN guarantees that if all antenna orientations are rotated by a local phase angle, the predicted beamforming weights rotate identically, preserving the physical meaning of the solution across the entire cellular topology.

The assumption that high-dimensional wireless channel data lies on a low-dimensional manifold embedded in the ambient measurement space. Channel state information vectors, despite having hundreds of dimensions, are constrained by the physics of electromagnetic propagation and scatterer geometry. Geometric deep learning exploits this by learning the intrinsic coordinates of the manifold, enabling robust channel prediction and compression with far fewer parameters than naive fully-connected architectures.

A critical distinction in geometric model design:

• Invariance: The output is completely unchanged by a symmetry transformation. Example: Predicting total network throughput from a graph—reordering nodes must yield the identical scalar value.
• Equivariance: The output transforms in a predictable way with the input. Example: Predicting per-node power allocations—if nodes are permuted, the allocation vector must permute identically. Choosing the correct property for each task prevents architectural misspecification and ensures physically consistent predictions.

The generalization of the standard convolution operation to arbitrary symmetry groups. While a CNN convolves over the translation group, a GNN convolves over the permutation group of the graph's nodes. The graph Laplacian's eigenvectors serve as the Fourier basis, enabling spectral filtering operations that are provably equivariant to node permutations. This mathematical formalism guarantees that message-passing operations are the only linear equivariant layers possible on graph-structured data.

A fundamental property ensuring the output for a graph or node is unchanged regardless of the arbitrary ordering of input nodes. This symmetry is the defining constraint for all graph neural network layers.

• Mechanism: Achieved through aggregation functions like sum, mean, or max that are commutative.
• Significance: Guarantees a consistent representation of network topology regardless of how base stations are indexed in a database.
• Contrast: A standard multi-layer perceptron (MLP) is not permutation invariant; feeding it the same graph with shuffled node order produces a different output.

A matrix derived from a graph's structure (degree matrix minus adjacency matrix) that encodes fundamental topological properties. It serves as the basis for spectral graph convolutions and Fourier transforms on graphs.

• Definition: L = D - A, where D is the degree matrix and A is the adjacency matrix.
• Eigendecomposition: The eigenvectors of the Laplacian define the graph's Fourier basis, enabling frequency-domain filtering.
• Application: Used to design spectral filters that smooth signals across the graph, analogous to low-pass filters in image processing.

A failure mode in deep GNNs where node representations become indistinguishable after too many layers of aggregation. Each message-passing step acts as a low-pass filter, and repeated application collapses all node features toward a common value.

• Cause: The receptive field of a node grows exponentially with depth, eventually covering the entire graph.
• Impact: Destroys local information, making tasks like node classification in large cellular topologies impossible.
• Mitigation: Residual connections, skip connections, or limiting the number of convolutional layers.

A phenomenon where information from an exponentially-growing receptive field is compressed into a fixed-size vector. This bottleneck prevents the GNN from learning long-range dependencies between distant nodes in a network graph.

• Mechanism: The number of nodes in a k-hop neighborhood grows exponentially with k, but the message vector size remains constant.
• Consequence: Critical interactions between distant base stations in a large cellular topology may be lost.
• Solutions: Graph Transformers with global attention, or adding virtual nodes that connect to all nodes in the graph.

A GNN architecture that applies the global self-attention mechanism of a Transformer to a graph's nodes. Unlike message-passing networks limited to local neighborhoods, each node can attend to all other nodes directly.

• Advantage: Mitigates the over-squashing problem by providing direct, long-range communication paths.
• Requirement: Requires positional encoding to inject topological information, since pure attention is permutation invariant and structure-agnostic.
• Use Case: Modeling interference relationships between distant cells that are not directly adjacent in the physical topology.

A technique for injecting information about a node's absolute or relative position within a graph structure into its initial features. This allows structure-agnostic models like Transformers to capture topological context.

• Laplacian PE: Uses the eigenvectors of the graph Laplacian to encode a node's spectral position.
• Random Walk PE: Encodes the probability of a random walk starting at a given node landing on other nodes.
• Necessity: Without positional encoding, a Graph Transformer cannot distinguish between isomorphic nodes and loses all structural awareness.