Manifold Learning

Unlike linear methods such as Principal Component Analysis (PCA), manifold learning algorithms do not assume data lies on a flat Euclidean subspace. They model the data as a curved, lower-dimensional manifold embedded in the high-dimensional space. This is critical for RF signals where the relationship between features like cyclostationary signatures and higher-order statistics is inherently non-linear.

• Isomap preserves geodesic distances along the manifold.
• Locally Linear Embedding (LLE) reconstructs each point from its neighbors.
• Laplacian Eigenmaps focus on preserving local proximity.

t-distributed Stochastic Neighbor Embedding (t-SNE) is a probabilistic manifold learning technique that excels at visualizing high-dimensional RF feature vectors in 2D or 3D scatter plots. It converts high-dimensional Euclidean distances into conditional probabilities representing similarities, then minimizes the Kullback-Leibler divergence between the high- and low-dimensional distributions.

• Heavily used for exploratory analysis of spectrogram embeddings.
• Reveals natural clusters of modulation types or emitter hardware fingerprints.
• The perplexity hyperparameter controls the balance between local and global structure.

Uniform Manifold Approximation and Projection (UMAP) is a modern manifold learning algorithm that competes with t-SNE for visualization quality but scales significantly better to large datasets and preserves more global structure. It is built on rigorous Riemannian geometry and algebraic topology.

• Assumes data is uniformly distributed on a locally connected Riemannian manifold.
• Constructs a fuzzy simplicial set representation of the high-dimensional data.
• Optimizes a low-dimensional embedding by minimizing the cross-entropy between fuzzy sets.
• In RF applications, UMAP is used to cluster IQ sample features and identify distinct emitter populations.

A critical preprocessing step for manifold learning is estimating the intrinsic dimensionality of the RF feature space—the minimum number of parameters needed to capture the signal's degrees of freedom. Overestimating leads to noise retention; underestimating collapses distinct clusters.

• Maximum Likelihood Estimation (MLE) approaches analyze neighbor distances.
• Correlation Dimension methods exploit the scaling of the correlation integral.
• In automatic modulation classification, the intrinsic dimension often correlates with the modulation order.

A foundational concept in manifold learning is the distinction between straight-line Euclidean distance in the ambient high-dimensional space and geodesic distance along the curved manifold surface. Two signals that appear far apart in raw IQ space may be close neighbors on the signal manifold.

• Isomap explicitly approximates geodesic distances by constructing a neighborhood graph and computing shortest paths (e.g., using Dijkstra's algorithm).
• This is essential for disentangling the non-linear effects of channel impairments like multipath fading and Doppler shift from the intrinsic signal structure.

Manifold learning techniques have known failure modes that must be managed in spectrum sensing contexts. They are generally non-parametric and do not provide an explicit mapping function for out-of-sample extension, requiring retraining or approximation methods like Nyström extension for new signals.

• t-SNE is sensitive to hyperparameter choice (perplexity) and can produce misleading clusters if misconfigured.
• The crowding problem in t-SNE can artificially separate continuous manifolds.
• UMAP is preferred for production RF pipelines due to its speed, deterministic output with a fixed random seed, and support for inverse transforms.

A non-linear dimensionality reduction technique that converts pairwise high-dimensional Euclidean distances into conditional probabilities representing similarities. It then minimizes the Kullback-Leibler divergence between these probabilities in the high-dimensional and low-dimensional spaces. t-SNE excels at preserving local structure, making it ideal for visualizing clusters of RF signal features, but its stochastic gradient descent optimization is non-convex and computationally intensive. The perplexity hyperparameter critically controls the effective number of local neighbors.

A manifold learning technique grounded in Riemannian geometry and algebraic topology. UMAP constructs a fuzzy topological representation of the high-dimensional data using k-nearest neighbor graphs, then optimizes a low-dimensional embedding by minimizing the cross-entropy between the two topological representations. It preserves more global structure than t-SNE, scales significantly better to large datasets, and has a robust mathematical foundation in category theory, making it the preferred choice for exploratory analysis of large RF signal databases.

A global non-linear dimensionality reduction method that extends classical Multidimensional Scaling (MDS). Isomap estimates the geodesic distance between all pairs of points by constructing a neighborhood graph and computing the shortest path distances through it using Dijkstra's algorithm. It then applies MDS to find a low-dimensional embedding that preserves these intrinsic manifold distances. Isomap is effective for datasets with a single, well-sampled, and developable manifold but is sensitive to short-circuit errors caused by noise or outliers.

An unsupervised manifold learning algorithm that assumes each data point and its neighbors lie on or near a locally linear patch of the manifold. LLE computes reconstruction weights that best reconstruct each point from its k-nearest neighbors, then finds a low-dimensional embedding that preserves these weights. It is an eigenvector-based method that solves a sparse eigenvalue problem, making it computationally efficient. LLE is particularly useful for unwrapping folded manifolds, such as the Swiss roll dataset, but can struggle with non-uniform sample densities.

A neural network approach to manifold learning where an encoder compresses high-dimensional input into a low-dimensional bottleneck representation, and a decoder reconstructs the original input. The latent space learned by the autoencoder implicitly captures the data manifold. Variational Autoencoders (VAEs) enforce a continuous, probabilistic latent space, while denoising autoencoders learn robust representations. In RF applications, autoencoders can learn compact representations of IQ samples for downstream tasks like anomaly detection and signal compression.

A manifold learning technique rooted in spectral graph theory. It constructs a weighted adjacency graph from data points and computes the graph Laplacian. The low-dimensional embedding is given by the eigenvectors corresponding to the smallest non-zero eigenvalues of this Laplacian. This approach preserves local neighborhood relationships and has a closed-form solution, avoiding iterative optimization. It is closely related to spectral clustering and provides a theoretical bridge between manifold learning and the analysis of heat diffusion on graphs.