Multiscale Mapper: Robust Topological Analysis

• Multiscale Mapper is a topological data analysis technique that constructs towers of nerve complexes from pullback covers, providing a robust multi-resolution summary of data topology.
• The method overcomes the instability of classical Mapper by tracking the evolution of topological features across continuous cover scales through persistent homology.
• Extensions such as bi-filtration and 2-Mapper enhance feature detection, enabling capture of multidimensional structures and guiding precise parameter selection.

Multiscale Mapper is a topological data analysis (TDA) technique generalizing the Mapper framework by encoding the evolution of topological summaries over a continuous range of cover scales. Developed to overcome the instability and ad hoc parameter dependence of classical Mapper, Multiscale Mapper constructs a tower of covers on the codomain of a data-associated map, pulls these covers back through the map, and analyzes the resulting tower of nerve complexes via persistent homology. This approach creates a robust, multi-resolution summary reflecting the topological structures present in the data across scale parameters, with provable stability guarantees under cover and data perturbation (Dey et al., 2015, Bungula et al., 2024). Recent work has extended these methods to bi-filtration and 2-Mapper constructions, enabling the capture of multidimensional features and providing practical guidance for parameter selection (Fritze, 26 Sep 2025, Bungula et al., 2024).

1. Mathematical Foundations

Let $Z$ be a topological space and $\mathcal{U} = \{U_\alpha\}_{\alpha \in A}$ an open cover of $Z$. The nerve of the cover,

$$\mathrm{Nerve}(\mathcal{U}) = \left\{\sigma \subset A \mid \bigcap_{\alpha \in \sigma} U_\alpha \neq \emptyset\right\},$$

is an abstract simplicial complex encoding how the sets in $\mathcal{U}$ intersect. Given a continuous map $f : X \to Z$, the pullback cover $f^*(\mathcal{U})$ is the collection of path-connected components of $f^{-1}(U_\alpha)$ for all $\alpha \in A$. The classical Mapper is defined as the nerve of $f^*(\mathcal{U})$.

A tower of covers is a sequence $\mathcal{U} = \{U_\alpha\}_{\alpha \in A}$0 indexed by resolution parameter $\mathcal{U} = \{U_\alpha\}_{\alpha \in A}$1, equipped with structure maps $\mathcal{U} = \{U_\alpha\}_{\alpha \in A}$2 for $\mathcal{U} = \{U_\alpha\}_{\alpha \in A}$3 satisfying functoriality. The Multiscale Mapper is the associated tower of nerves:

$\mathcal{U} = \{U_\alpha\}_{\alpha \in A}$4

together with the induced simplicial maps.

Applying homology $\mathcal{U} = \{U_\alpha\}_{\alpha \in A}$5 to this tower yields a persistence module, and the collection of barcodes or persistence diagrams summarizes the emergence and disappearance of topological features across scales (Dey et al., 2015, Bungula et al., 2024).

2. Algorithmic Construction and Combinatorial Approximations

For a finite simplicial complex $\mathcal{U} = \{U_\alpha\}_{\alpha \in A}$6 and a piecewise-linear map $\mathcal{U} = \{U_\alpha\}_{\alpha \in A}$7, Multiscale Mapper can be computed exactly from the 1-skeleton $\mathcal{U} = \{U_\alpha\}_{\alpha \in A}$8 under the minimum-diameter condition: for every simplex $\mathcal{U} = \{U_\alpha\}_{\alpha \in A}$9, $Z$0, the minimal diameter of a cover element (Dey et al., 2015). The tower of nerves can thus be efficiently constructed by:

1. For each scale $Z$1, construct $Z$2.
2. Pull back to $Z$3: mark vertices/edges covered by each interval, extract connected components.
3. Build the nerve complex of these components.
4. Feed the resulting directed system of simplicial complexes into a persistence computation.

For more general settings or massive data, the combinatorial multiscale mapper constructs the pullback cover at the graph (1-skeleton) level, using only vertex connectivity, and approximates the full multiscale mapper with a bottleneck bound controlled in $Z$4-scale by the cover parameters and the minimum-diameter constant.

3. Extensions: Bi-Filtration and Multiscale 2-Mapper

When applying Mapper to point-cloud data clustered by DBSCAN, the parameterization must include both the cover diameter ($Z$5) and the DBSCAN radius ($Z$6), which each induce monotone refinements of the nerve complexes. The full construction is a bi-filtration: the collection of nerve complexes $Z$7 indexed by $Z$8, ordered so that $Z$9 iff $$\mathrm{Nerve}(\mathcal{U}) = \left\{\sigma \subset A \mid \bigcap_{\alpha \in \sigma} U_\alpha \neq \emptyset\right\},$$0 and $$\mathrm{Nerve}(\mathcal{U}) = \left\{\sigma \subset A \mid \bigcap_{\alpha \in \sigma} U_\alpha \neq \emptyset\right\},$$1 (Bungula et al., 2024). Along monotone paths in the parameter plane, the corresponding filtration yields 1-parameter persistent homology—generalizing to 2D persistent homology for the full bi-filtration.

The Multiscale 2-Mapper extracts higher-dimensional features by encoding not just the 1-skeleton (graph) of the pullback nerve but its 2-skeleton, recording 2-simplices corresponding to triple overlaps. The 2-Mapper complex $$\mathrm{Nerve}(\mathcal{U}) = \left\{\sigma \subset A \mid \bigcap_{\alpha \in \sigma} U_\alpha \neq \emptyset\right\},$$2 is the subcomplex of the nerve of $$\mathrm{Nerve}(\mathcal{U}) = \left\{\sigma \subset A \mid \bigcap_{\alpha \in \sigma} U_\alpha \neq \emptyset\right\},$$3 comprised of all cells of dimension at most 2. This is crucial for accurately recovering $$\mathrm{Nerve}(\mathcal{U}) = \left\{\sigma \subset A \mid \bigcap_{\alpha \in \sigma} U_\alpha \neq \emptyset\right\},$$4 (the first Betti number), which counts loops (Fritze, 26 Sep 2025). Parameter selection (notably the overlap fraction $$\mathrm{Nerve}(\mathcal{U}) = \left\{\sigma \subset A \mid \bigcap_{\alpha \in \sigma} U_\alpha \neq \emptyset\right\},$$5 of a cubical cover) is guided by maximizing the persistence intervals of $$\mathrm{Nerve}(\mathcal{U}) = \left\{\sigma \subset A \mid \bigcap_{\alpha \in \sigma} U_\alpha \neq \emptyset\right\},$$6 along the cover tower, distinguishing genuine cycles from artifacts.

4. Stability Theory

Multiscale Mapper is designed for stability under both data and cover perturbations, formalized via interleavings. Two towers of covers are $$\mathrm{Nerve}(\mathcal{U}) = \left\{\sigma \subset A \mid \bigcap_{\alpha \in \sigma} U_\alpha \neq \emptyset\right\},$$7-interleaved if there exist cover maps between scales $$\mathrm{Nerve}(\mathcal{U}) = \left\{\sigma \subset A \mid \bigcap_{\alpha \in \sigma} U_\alpha \neq \emptyset\right\},$$8 and $$\mathrm{Nerve}(\mathcal{U}) = \left\{\sigma \subset A \mid \bigcap_{\alpha \in \sigma} U_\alpha \neq \emptyset\right\},$$9 respecting the tower structure. This induces interleavings on the associated towers of nerves and, in turn, on their homology modules. The main theorem states:

If two towers of covers are $\mathcal{U}$0-interleaved and $\mathcal{U}$1 satisfy $\mathcal{U}$2, then the collections $\mathcal{U}$3 and $\mathcal{U}$4 are $\mathcal{U}$5-interleaved whenever the Hausdorff distance between $\mathcal{U}$6 and $\mathcal{U}$7 is at most $\mathcal{U}$8 and either $\mathcal{U}$9 in DBSCAN or no free-border points exist (Bungula et al., 2024).

For each $f : X \to Z$0, the bottleneck distance of the persistence diagrams is then controlled by the interleaving parameter, ensuring that the persistent topological summaries computed by Multiscale Mapper are robust to small perturbations:

$f : X \to Z$1

for $f : X \to Z$2-good towers (Dey et al., 2015).

5. Practical Issues: Cover Construction, Clustering, and Parameter Selection

The choice of cover family affects both theoretical guarantees and computational behavior. The standard approach uses cubical covers in $f : X \to Z$3 with adjustable resolution and overlap, with overlap fraction $f : X \to Z$4 and covering number $f : X \to Z$5 per coordinate controlling both granularity and intersections (Fritze, 26 Sep 2025). For point-cloud data, the clustering of pullback sets is typically done by DBSCAN, with parameters ($f : X \to Z$6) affecting the monotonicity required for consistent filtrations. When $f : X \to Z$7, DBSCAN guarantees monotone refinement for both cover scale and cluster scale and the absence of free-border points avoids artifacts.

Parameter selection is guided by tracking the persistence diagram of $f : X \to Z$8 as a function of overlap fraction $f : X \to Z$9: small $f^*(\mathcal{U})$0 fails to recover genuine loops (few 2-simplices), while large $f^*(\mathcal{U})$1 fills in cycles prematurely. Stability theorems guarantee that long intervals in the barcode reflect true topological features (Fritze, 26 Sep 2025).

6. Comparison with Related Constructions and Applications

Multiscale Mapper unifies and extends classical Mapper, Reeb graphs, and merge trees under the persistent homology perspective. With appropriately constructed towers and reindexing (typically in $f^*(\mathcal{U})$2 scale), it yields single persistence diagrams capturing the evolution of connectivity, cycles, and higher-dimensional holes, as opposed to single-scale Mapper which is highly parameter-dependent and lacks stability.

Persistent summaries from Multiscale Mapper have been applied to multivariate data analysis, shape summarization (e.g., recovering the true Betti numbers for torus and Klein-bottle point clouds), and scientific data summarization across biology, neuroscience, and materials science domains (Dey et al., 2015, Fritze, 26 Sep 2025). The bi-filtration and 2-Mapper extensions provide the machinery to handle higher-order features and inform scale choices in exploratory analysis.

7. Open Questions and Limitations

Ongoing directions include statistical convergence guarantees for approximating multiscale mapper from noisy samples, the study of nonmonotonic “zigzag” towers to better encode dynamics, understanding the exact relationship to pullback Čech complexes via pseudometrics, and the study of limit behavior as the cover scale tends to zero. Feature attribution within the data domain and connections with other topological summarization tools remain active research areas (Dey et al., 2015). The non-existence of filtrations for certain clustering parameters (e.g., large $f^*(\mathcal{U})$3 with free-border points), and the non-monotonic dependence of Betti numbers on cover overlap, present important caveats for practical deployments (Bungula et al., 2024, Fritze, 26 Sep 2025).