Standard Topological Data Analysis (TDA) typically relies on a single filtration parameter—usually a spatial scale or distance threshold—to track how the "shape" of data evolves. While powerful, this one-dimensional view is notoriously sensitive to noise. A single outlier in a dataset can act as a "bridge," prematurely connecting components or closing loops, which effectively destroys the topological signal.
Multiparameter Persistent Homology (MPH) addresses this by filtering data across multiple dimensions simultaneously. By considering parameters like scale and local density together, researchers can distinguish between significant structural features and transient noise, creating a far more robust "topological fingerprint" of the data.
From a Line to a Grid: The Bifiltration
In standard persistence, we grow balls around data points and watch as they merge to form clusters or holes. This is a 1D filtration. As Magnus Bakke Botnan and Michael Lesnick explain in An Introduction to Multiparameter Persistence, MPH generalizes this to a multi-dimensional grid.
The most common form is the bifiltration, which often uses distance ($r$) and density ($k$). Instead of just asking "what is the shape at scale $r$?", we ask "what is the shape formed by the densest points at scale $r$?". This allows us to ignore low-density outliers that would otherwise distort the topology. Michael Kerber and colleagues demonstrate in Computing the Multicover Bifiltration that this approach can recover the true underlying topology of a noisy point cloud where 1D methods fail.
Capturing Interactions Between Features
The power of MPH lies in its ability to capture the interaction between different data properties. For example, in medical imaging, a single threshold of pixel intensity might not be enough to identify a tumor. Fulya Tastan and Baris Coskunuzer show in Topological Feature Fusion for Dermoscopic Skin Cancer Detection that by filtering across multiple color channels simultaneously, they can capture connectivity patterns—like how specific pigments cluster—that are invisible to single-parameter analysis.
Similarly, in network science, Dejan Govc and Ran Levi utilize bifiltrations to analyze directed graphs. By combining different measures of "directionality" (like vertex flow and cycle frequency), their tournaplex framework can distinguish complex brain network dynamics that appear identical under simpler topological lenses.
Integrating Topology into Machine Learning
A major hurdle for MPH has been its complexity; unlike 1D persistence, it doesn't have a simple "barcode" representation. Recent work has focused on "vectorizing" these complex 2D structures so they can be fed into neural networks:
- Rank Functions and Landscapes: Cheng Xin and Soham Mukherjee introduced GRIL, a method that converts 2-parameter data into stable, differentiable landscapes. This allows topological features to be optimized directly within a machine learning pipeline.
- Graph-based Encodings: In Graphcode, Florian Russold and Michael Kerber represent the evolution of topological features as an embedded graph, allowing Graph Neural Networks (GNNs) to learn directly from the multi-parameter relationships.
- Differentiable Layers: Luis Scoccola and Steve Oudot proved in Differentiability and Optimization of Multiparameter Persistent Homology that many MPH descriptors are semilinearly determined on grids, enabling their use as differentiable layers that outperform 1D methods in noise-robust point cloud optimization.
Beyond Static Shapes: Multiscale Dynamics
MPH is also being used to track how structures change across non-hierarchical scales. Juni Schindler and Mauricio Barahona developed MCbiF, which uses 2-parameter persistence to measure "topological autocorrelation." This framework captures how clusters merge, split, and overlap across different resolutions, providing a more nuanced view of data partitions than traditional information-theoretic measures.
By moving from a single filtration to a multi-dimensional grid, this research direction provides the mathematical tools to see through noise and capture the intricate, multi-faceted geometry of complex datasets.