Persistence Barcodes vs Diagrams Explained | TDA Tutorial

Persistence images provide a stable vectorization method for converting persistence diagrams into numerical feature vectors, enabling their use with standard machine learning techniques like support vector machines and logistic regression; this is achieved by centering differentiable probability distributions (such as normalized symmetric Gaussians) at each point in the diagram, applying weighting functions (linear or bump functions), and integrating over a grid to produce a stable representation that is provably stable with respect to perturbations in the input data.

This section covers advanced topological vectorization techniques including persistence landscapes, persistence images, and path signatures. Persistence landscapes provide hierarchical vector space representations through iterative layer peeling, enabling neural network integration and maintaining bijection with original diagrams. Persistence images transform diagrams into 2D grids via Gaussian convolution after rotating the diagram, creating natural representations suitable for CNNs. For graph classification, one can learn weight functions on persistence images to emphasize relevant topological features. Summary statistics (infinity norm, p-norms) provide efficient scalar measures for time-varying topological data analysis. Path signatures offer universal nonlinear representations for topological paths, connecting various vectorization approaches as special cases. Each method balances computational efficiency against representational completeness, with persistence images excelling for neural network applications while landscapes provide theoretical guarantees.

To apply persistence diagrams in machine learning, they must be converted to vector representations. Two major approaches are: (1) Persistence Image: Applies linear transformation to shift points away from the diagonal, then computes weighted sums of Gaussian functions centered at each point, integrating over regions of interest. (2) Persistence Landscape: Constructs piecewise linear functions by drawing triangles to the diagonal and taking lower envelopes. These vectorization methods enable persistence diagrams as inputs to machine learning models. Applications include material science (distinguishing liquid vs glass states of SiO₂) and protein analysis, demonstrating practical utility across scientific domains.

Converting persistence diagrams to vectors enables machine learning applications. Persistence images rasterize diagrams onto grids, apply Gaussian blurring, and weight pixels by persistence (death-birth difference). Parameters include birth range, persistence range, pixel size, and weight factors. Infinite-living components are discarded as irrelevant. Alternative methods include persistence landscapes (functional summaries), silhouette, simple statistics (counts, averages, minima/maxima), neural network layers, and kernel methods (enabling averaging). Each approach balances interpretability, computational efficiency, and suitability for different machine learning architectures.

This section addresses the theoretical foundations and practical applications of persistent homology. The bottleneck distance measures similarity between persistence diagrams by finding optimal matchings between points (plus diagonal points for unmatched features). A fundamental stability theorem guarantees that if two point clouds are ε-close, their persistence diagrams are bottleneck distance at most 2ε apart, validating persistent homology for data science applications. However, this guarantee fails when adding/removing points, motivating research into multi-parameter persistence approaches. The section then presents two primary vectorization methods for integrating persistent homology into machine learning pipelines: persistence images (applying Gaussian weightings to transformed diagram points, creating smooth surfaces discretized onto grids) and persistence landscapes (constructing sequences of triangular functions where λ_k(t) takes the k-th largest value across all functions at each time t). Both methods enable averaging across samples, statistical hypothesis testing via central limit theorems, and integration into machine learning pipelines for topological data analysis.

The stability of persistence diagrams is captured by the bottleneck distance, which measures how much barcodes can differ under perturbations. A delta-matching requires matched intervals differ by at most delta in endpoints, with unmatched intervals shorter than 2delta. The interleaving distance provides an equivalent formulation: two diagrams are delta-interleaved if there are natural transformations between them and their shifted versions. The equivalence between barcodes and matching diagrams shows these distances coincide, establishing that small perturbations in input data yield small changes in persistence diagrams. This isometry theorem forms the mathematical foundation guaranteeing the robustness of topological data analysis pipelines.

Each feature has a birth time (when it first appears) and a death time (when it gets filled in or merges away). You record all this as a persistence diagram or persistence barcode. Features that persist for a long range of scales (born early and die late) are likely to represent genuine structure in the data. Features that flicker in and out quickly are likely to be noise. By looking at the persistent features, you're extracting the topological shape of the data in a way that's robust to small perturbations. This is powerful because topology is naturally scale-free and coordinate-free—you don't have to choose a particular scale to look at your data, and you don't have to define distances in a particular way. The topology will be stable under small changes to the data. The stability theorem (proved by Cohen-Steiner, Edelsbrunner, Harer, and Carlsson) says that small perturbations to the data produce small changes in the persistence diagram as measured by a specific metric called the bottleneck distance. This gives the whole framework a rigorous statistical foundation.

The stability theorem states that the bottleneck distance between persistence diagrams is Lipschitz continuous with respect to the Gromov-Hausdorff distance between metric spaces. This means small changes in the data produce small changes in the persistence diagram. This stability enables efficient sampling strategies and robustness to noise in the data.

The algebraic stability theorem states that for pointwise finite-dimensional persistence modules, the bottleneck distance between their persistence diagrams is equal to the interleaving distance between the modules, establishing that these two distances are isometric; this theorem was proven by Bauer and Lesnick in 2015 and provides a fundamental guarantee that small perturbations in the input data or functions result in correspondingly small changes in the topological features captured by persistence diagrams.

A fundamental result establishes that small function perturbations yield small changes in persistence diagrams: if ||f-g||∞ ≤ ε, then bottleneck distance between diagrams ≤ ε. This stability ensures robustness. The bottleneck distance finds optimal point matchings minimizing maximum displacement, while Wasserstein distance sums p-th powers of displacements. These metrics enable comparing persistence diagrams and form theoretical foundations for TDA applications.

Topological Data Analysis (TDA) is a powerful data analysis approach that focuses on the shape and structure of data. The core philosophy is that 'data has shape, shape has meaning, and meaning has value.' TDA has been applied across diverse fields including materials science, genomics, life sciences, information technology, and business strategy. The two main methods introduced are Persistent Homology (for detailed topological analysis) and Mapper (for high-dimensional data visualization). TDA enables researchers to extract meaningful patterns from complex, high-dimensional data by focusing on its underlying topological structure rather than just numerical values.

Topological Data Analysis (TDA) applies algebraic topology concepts to data analysis, revealing hidden structures in complex datasets. Three key applications demonstrate its power: (1) Chemistry - analyzing molecular conformation spaces by mapping molecules to high-dimensional points and discovering geometric structures like spheres and cylinders; (2) Biology - studying grid cell activity patterns in brains to reveal toroidal structures underlying spatial perception; (3) Medicine - classifying breast cancer subtypes by transforming patient genomic data into high-dimensional point clouds and identifying branching tree structures that revealed previously unknown survival-positive patient groups. TDA bridges pure mathematics with practical data science by focusing on the underlying topology of data rather than raw numerical values.

This section demonstrates TDA's practical power and integration with machine learning. Mapper extends TDA through three steps: define filter functions, construct overlapping intervals, and connect clusters across partitions. This reveals patterns like flares (diverging paths), loops (periodicity), and bifurcations (branching). In practice, Mapper detected a July 4th state change in 27 million data points invisible through traditional analysis. Bottleneck distance measures similarity between persistence diagrams. For time series, Takens' theorem converts data to point clouds. Applications span bioinformatics, finance, economics, anomaly detection, and explainable AI. TDA provides robust feature summarization by transforming from local to global perspectives. Outputs serve as inputs for supervised learning classifiers, improving feature selection and model interpretability. The fundamental insight is that reasoning becomes possible because data has shape and shape has meaning.

TDA has been successfully applied across diverse fields: (1) Image analysis—revealing that certain image collections have underlying structures like Klein bottles (non-orientable surfaces); (2) Cancer research—identifying cycles and structural patterns in cancer cell data; (3) Political science—analyzing US House voting patterns over 10 years, revealing complex subgroup structures beyond simple party divisions; (4) Sports analytics—NBA player data analysis revealed nine distinct player types instead of the traditional five positions, with practical implications for team building. These applications demonstrate how TDA can uncover hidden patterns that traditional methods miss, though current computational limitations prevent it from being a true Big Data solution.

TDA has applications in classification, defect detection, brain analysis, and shape analysis. It provides a new way to analyze data by focusing on shape properties rather than traditional statistical features. This makes it particularly useful for complex, high-dimensional datasets where traditional methods may fail. TDA is a rapidly developing field with applications in machine learning, neuroscience, and data science. It enables analysis of the essential shape properties of data, providing insights that complement traditional statistical approaches.