Computational Multiscale Mapper

• Computational Multiscale Mapper is a framework that extends the traditional Mapper algorithm by using a tower of covers to capture the evolution of topological features across scales.
• It employs exact and combinatorial computation methods on pullback covers and clustering techniques to generate persistence diagrams approximating topological summaries.
• Advanced extensions such as density-sensitive refinement, higher-skeleton constructions, and parallel execution enhance both local adaptability and computational scalability.

Computational Multiscale Mapper denotes the family of algorithmic frameworks that extend Mapper from a single fixed cover to a scale-indexed, locally adaptive, or multi-parameter collection of covers, so that the resulting topological summaries can be related by simplicial maps and analyzed by persistent homology. In its basic form, Mapper takes a dataset or topological space $X$, a map $f:X\to Z$, and a cover of $f(X)$, and returns the nerve of the pullback cover. Multiscale Mapper replaces that single cover with a tower of covers, producing a tower of simplicial complexes whose persistence diagrams summarize the evolution of topological features across scale (Dey et al., 2015). Subsequent work develops exact and combinatorial computation, stability for point-cloud data, density-sensitive local refinement, higher-skeleton constructions, and parallel execution (Deb et al., 2019, Hajij et al., 2017, Bungula et al., 2024, Fritze, 26 Sep 2025).

1. Formal construction and topological role

Mapper is defined by

$\mathrm{M}(\mathcal{U}, f) := N(f^*(\mathcal{U})),$

where $f^*(\mathcal{U})$ is obtained by pulling back each cover element and decomposing the preimage into path-connected components, and $N(\cdot)$ denotes the nerve simplicial complex (Dey et al., 2015). In the point-cloud setting summarized in later computational work, the same construction is typically implemented by clustering each pullback set and taking clusters as nerve vertices (Hajij et al., 2017, Bungula et al., 2024). Mapper has been described as approximating and generalizing the Reeb graph, the contour tree, split, and joint trees, and as a discrete approximation of the Reeb space in the $1$-dimensional case (Hajij et al., 2017, Deb et al., 2019).

Multiscale Mapper replaces a single cover by a tower of covers

$$\mathfrak{U}=\{\mathcal{U}_\varepsilon\}_{\varepsilon\ge r},$$

equipped with maps of covers $$w_{\varepsilon,\varepsilon'}:\mathcal{U}_\varepsilon\to \mathcal{U}_{\varepsilon'}$$ satisfying

$$w_{\varepsilon,\varepsilon}=\mathrm{id}, \qquad w_{\varepsilon',\varepsilon''}\circ w_{\varepsilon,\varepsilon'}=w_{\varepsilon,\varepsilon''}.$$

Applying pullback and nerve at each scale yields a tower of simplicial complexes, denoted

$f:X\to Z$0

and the induced simplicial maps define a persistence module after applying homology (Dey et al., 2015).

A central technical condition is the notion of a $f:X\to Z$1-good tower of covers. Such a tower has resolution $f:X\to Z$2, each $f:X\to Z$3 satisfies $f:X\to Z$4, and every subset $f:X\to Z$5 with $f:X\to Z$6 is contained in some $f:X\to Z$7 (Dey et al., 2015). This condition underlies the stability theory and excludes pathological cover systems.

2. Exact, combinatorial, and metric computation

For a simplicial-complex domain $f:X\to Z$8 and a real-valued piecewise-linear map $f:X\to Z$9, Multiscale Mapper admits an exact algorithm that computes the persistence diagram using only the $f(X)$0-skeleton of the input complex (Dey et al., 2015). The key condition stated in that analysis is that $f(X)$1 for all edges, where $f(X)$2 is determined by the cover tower. Under that condition, the multiscale construction on the full complex agrees with the one computed on the $f(X)$3-skeleton (Dey et al., 2015).

For general maps, the same paper gives a combinatorial algorithm that acts only on vertex sets connected by the $f(X)$4-skeleton graph. For each cover element $f(X)$5 at a given scale, one forms the induced subgraph on the vertices mapping into $f(X)$6, computes its graph connected components, and builds the nerve on the resulting vertex subsets (Dey et al., 2015). The resulting persistence diagrams approximate the exact multiscale mapper diagrams via the stated stability result.

The framework also admits a metric reinterpretation. Given $f(X)$7, the pullback pseudometric is

$f(X)$8

with $f(X)$9 (Dey et al., 2015). When $\mathrm{M}(\mathcal{U}, f) := N(f^*(\mathcal{U})),$0 is $\mathrm{M}(\mathcal{U}, f) := N(f^*(\mathcal{U})),$1-good, this pseudometric satisfies a relaxed triangle inequality, and the associated Čech filtration is, after log-reindexing, $\mathrm{M}(\mathcal{U}, f) := N(f^*(\mathcal{U})),$2-interleaved with Multiscale Mapper (Dey et al., 2015). This places Multiscale Mapper in direct computational relation with standard filtration-based TDA.

A common misconception is that computational multiscale analysis is simply a batch of unrelated Mapper runs at different resolutions. The tower-of-covers formalism is stricter: the scales are linked by maps of covers, these induce simplicial maps between nerves, and homological comparison is performed on the resulting persistence module rather than by informal visual comparison alone (Dey et al., 2015).

3. Point-cloud filtrations, DBSCAN, and stability

For point-cloud data, computational Multiscale Mapper depends not only on the cover parameter but also on the clustering algorithm and its parameters. In the DBSCAN setting, the relevant parameters are $\mathrm{M}(\mathcal{U}, f) := N(f^*(\mathcal{U})),$3 and $\mathrm{M}(\mathcal{U}, f) := N(f^*(\mathcal{U})),$4. If $\mathrm{M}(\mathcal{U}, f) := N(f^*(\mathcal{U})),$5, DBSCAN is equivalent to single linkage clustering with cutting height $\mathrm{M}(\mathcal{U}, f) := N(f^*(\mathcal{U})),$6; filtrations exist if cover size increases, $\mathrm{M}(\mathcal{U}, f) := N(f^*(\mathcal{U})),$7 increases, and/or $\mathrm{M}(\mathcal{U}, f) := N(f^*(\mathcal{U})),$8 decreases. If $\mathrm{M}(\mathcal{U}, f) := N(f^*(\mathcal{U})),$9, a filtration of mapper graphs may fail to exist except in the absence of free-border points (Bungula et al., 2024).

The point-cloud analysis in (Bungula et al., 2024) makes this dependence explicit by constructing a bi-filtration indexed by cover size and DBSCAN $f^*(\mathcal{U})$0. For each pair $f^*(\mathcal{U})$1, one obtains a Mapper nerve $f^*(\mathcal{U})$2, and these are linked across parameter increases by simplicial maps, yielding bi-filtrations of homology groups

$f^*(\mathcal{U})$3

(Bungula et al., 2024).

The same work shows that a $f^*(\mathcal{U})$4-dimensional filtration can be unstable under small perturbations of the input data, even when a filtration exists. Stability is recovered by increasing both the cover size and $f^*(\mathcal{U})$5 simultaneously. If each point is perturbed by at most $f^*(\mathcal{U})$6, then the resulting bi-filtrations of homology groups are $f^*(\mathcal{U})$7-interleaved (Bungula et al., 2024). In the language of persistence modules, there exist morphisms

$f^*(\mathcal{U})$8

whose compositions agree with the $f^*(\mathcal{U})$9-shift maps in the sense stated in the paper (Bungula et al., 2024).

The foundational multiscale stability theorem operates at the level of the map $N(\cdot)$0 and a $N(\cdot)$1-good tower of covers. If $N(\cdot)$2 satisfy $N(\cdot)$3, then after log-reindexing,

$N(\cdot)$4

where $N(\cdot)$5 is the bottleneck distance (Dey et al., 2015). Taken together, these results separate two issues that are often conflated: the abstract stability of Multiscale Mapper for good towers, and the additional clustering-induced complications that arise in point-cloud implementations.

4. Density-sensitive local refinement and Multimapper

Standard Mapper applies a globally chosen cover scale. The density-sensitive analysis in “Multimapper” identifies two opposite failure modes: in dense regions, a coarse scale can obscure local structure, while in sparse regions, a fine scale can fragment the visualization or “shatter” a single topological object (Deb et al., 2019). Multiscale Mapper improves on a single fixed scale by constructing a tower of covers at various global resolutions, but the critique made in that work is that each scale is still applied uniformly and therefore does not adapt to local density (Deb et al., 2019).

The proposed remedy is a locally adaptive construction. Regions of interest are selected, for example where a current Mapper appears too coarse or too fine. If $N(\cdot)$6 is a selected set of Mapper nodes with associated clusters $N(\cdot)$7, the restricted data are

$N(\cdot)$8

A finer or otherwise modified cover is then built only over $N(\cdot)$9, pulled back and clustered locally, while the rest of the data remain at the original scale. The refined and unmodified clusters are merged into a new family $1$0, and a new nerve is constructed on all intersections (Deb et al., 2019).

The paper formalizes this as

$1$1

where $1$2 is a cover of the dataset and $1$3 is the cover chosen for region $1$4 (Deb et al., 2019). The same work proposes a persistence-based diagnostic for deviations between Mapper and Reeb-space behavior: if for a simplex $1$5 in the Mapper nerve,

$1$6

persistently across scales, then the Mapper output is misrepresenting the topology (Deb et al., 2019). This diagnostic is phrased in terms of persistent homology on intersections and is intended to distinguish robust features from noise.

A further computational device introduced there is the brick cover, a brick-like offset-rectangle cover that reduces the maximum simplex order in Mapper nerve constructions to at most triangles in $1$7D (Deb et al., 2019). In practical terms, the Multimapper program extends multiscale analysis from globally indexed resolutions to region-specific refinement, while preserving global adjacencies through a single nerve construction.

5. Higher-skeleton Multiscale Mapper and parameter selection

The standard Mapper graph records only the $1$8-skeleton of the nerve. The higher-skeleton extension called $1$9-Mapper computes vertices, edges, and triangles, with the stated aim of observing two-dimensional features and capturing one-dimensional topology more faithfully (Fritze, 26 Sep 2025). Its basic definition is

$$\mathfrak{U}=\{\mathcal{U}_\varepsilon\}_{\varepsilon\ge r},$$0

where $$\mathfrak{U}=\{\mathcal{U}_\varepsilon\}_{\varepsilon\ge r},$$1 denotes the $$\mathfrak{U}=\{\mathcal{U}_\varepsilon\}_{\varepsilon\ge r},$$2-skeleton of the nerve (Fritze, 26 Sep 2025).

The multiscale version is built from a tower of cubical covers. For a bounding box $$\mathfrak{U}=\{\mathcal{U}_\varepsilon\}_{\varepsilon\ge r},$$3, each axis is subdivided into $$\mathfrak{U}=\{\mathcal{U}_\varepsilon\}_{\varepsilon\ge r},$$4 intervals with overlap fraction $$\mathfrak{U}=\{\mathcal{U}_\varepsilon\}_{\varepsilon\ge r},$$5, and each cover element is a box

$$\mathfrak{U}=\{\mathcal{U}_\varepsilon\}_{\varepsilon\ge r},$$6

Varying a scale parameter $$\mathfrak{U}=\{\mathcal{U}_\varepsilon\}_{\varepsilon\ge r},$$7, or equivalently the overlap fraction $$\mathfrak{U}=\{\mathcal{U}_\varepsilon\}_{\varepsilon\ge r},$$8, produces a tower of covers $$\mathfrak{U}=\{\mathcal{U}_\varepsilon\}_{\varepsilon\ge r},$$9 with canonical maps from coarser to finer covers (Fritze, 26 Sep 2025). In the reported implementation, DBSCAN is used with $$w_{\varepsilon,\varepsilon'}:\mathcal{U}_\varepsilon\to \mathcal{U}_{\varepsilon'}$$0, specifically to avoid free border points and to ensure filtrations behave well (Fritze, 26 Sep 2025).

A computational difficulty is the alignment of clusters across scales, since clusters may split or merge. The proposed solution defines a matching map by maximal Jaccard overlap: $$w_{\varepsilon,\varepsilon'}:\mathcal{U}_\varepsilon\to \mathcal{U}_{\varepsilon'}$$1 where $$w_{\varepsilon,\varepsilon'}:\mathcal{U}_\varepsilon\to \mathcal{U}_{\varepsilon'}$$2 is the data in cluster $$w_{\varepsilon,\varepsilon'}:\mathcal{U}_\varepsilon\to \mathcal{U}_{\varepsilon'}$$3 (Fritze, 26 Sep 2025). When multiple fine-scale clusters map to the same coarse-scale cluster, ghost vertices and additional simplices are added so that loops and homological features are preserved in the filtration. The implementation is described via a simplex tree, with GUDHI named explicitly, and the overall complexity is reported as $$w_{\varepsilon,\varepsilon'}:\mathcal{U}_\varepsilon\to \mathcal{U}_{\varepsilon'}$$4, where $$w_{\varepsilon,\varepsilon'}:\mathcal{U}_\varepsilon\to \mathcal{U}_{\varepsilon'}$$5 is the number of scales and $$w_{\varepsilon,\varepsilon'}:\mathcal{U}_\varepsilon\to \mathcal{U}_{\varepsilon'}$$6 the number of clusters per scale (Fritze, 26 Sep 2025).

This framework also turns persistent $$w_{\varepsilon,\varepsilon'}:\mathcal{U}_\varepsilon\to \mathcal{U}_{\varepsilon'}$$7 into a parameter-selection criterion. The paper proposes choosing parameter ranges for which $$w_{\varepsilon,\varepsilon'}:\mathcal{U}_\varepsilon\to \mathcal{U}_{\varepsilon'}$$8 is constant and matches the expected topology, rather than relying solely on visual inspection (Fritze, 26 Sep 2025). In experiments on $$w_{\varepsilon,\varepsilon'}:\mathcal{U}_\varepsilon\to \mathcal{U}_{\varepsilon'}$$9 points sampled from a Klein bottle in $$w_{\varepsilon,\varepsilon}=\mathrm{id}, \qquad w_{\varepsilon',\varepsilon''}\circ w_{\varepsilon,\varepsilon'}=w_{\varepsilon,\varepsilon''}.$$0 and projected to $$w_{\varepsilon,\varepsilon}=\mathrm{id}, \qquad w_{\varepsilon',\varepsilon''}\circ w_{\varepsilon,\varepsilon'}=w_{\varepsilon,\varepsilon''}.$$1, small overlap $$w_{\varepsilon,\varepsilon}=\mathrm{id}, \qquad w_{\varepsilon',\varepsilon''}\circ w_{\varepsilon,\varepsilon'}=w_{\varepsilon,\varepsilon''}.$$2 made the $$w_{\varepsilon,\varepsilon}=\mathrm{id}, \qquad w_{\varepsilon',\varepsilon''}\circ w_{\varepsilon,\varepsilon'}=w_{\varepsilon,\varepsilon''}.$$3-Mapper too sparse, large overlap $$w_{\varepsilon,\varepsilon}=\mathrm{id}, \qquad w_{\varepsilon',\varepsilon''}\circ w_{\varepsilon,\varepsilon'}=w_{\varepsilon,\varepsilon''}.$$4 filled in loops with $$w_{\varepsilon,\varepsilon}=\mathrm{id}, \qquad w_{\varepsilon',\varepsilon''}\circ w_{\varepsilon,\varepsilon'}=w_{\varepsilon,\varepsilon''}.$$5-simplices, and $$w_{\varepsilon,\varepsilon}=\mathrm{id}, \qquad w_{\varepsilon',\varepsilon''}\circ w_{\varepsilon,\varepsilon'}=w_{\varepsilon,\varepsilon''}.$$6 yielded a stable $$w_{\varepsilon,\varepsilon}=\mathrm{id}, \qquad w_{\varepsilon',\varepsilon''}\circ w_{\varepsilon,\varepsilon'}=w_{\varepsilon,\varepsilon''}.$$7 (Fritze, 26 Sep 2025). This use of persistent homology for parameter tuning directly addresses the long-standing ad hoc status of Mapper parameter choice.

6. Parallel execution and computational scaling

Parallel Mapper provides a divide-and-conquer strategy for the construction of Mapper on multiple processors. The codomain interval $$w_{\varepsilon,\varepsilon}=\mathrm{id}, \qquad w_{\varepsilon',\varepsilon''}\circ w_{\varepsilon,\varepsilon'}=w_{\varepsilon,\varepsilon''}.$$8 is decomposed into an $$w_{\varepsilon,\varepsilon}=\mathrm{id}, \qquad w_{\varepsilon',\varepsilon''}\circ w_{\varepsilon,\varepsilon'}=w_{\varepsilon,\varepsilon''}.$$9-chain cover $f:X\to Z$00 such that adjacent intervals overlap and non-adjacent intervals do not. For each $f:X\to Z$01, a local cover $f:X\to Z$02 is constructed so that the only nontrivial intersections across neighboring processors occur on the prescribed overlap interval $f:X\to Z$03 (Hajij et al., 2017).

Each processor computes Mapper independently on the restricted data $f:X\to Z$04 using $f:X\to Z$05, producing a local graph $f:X\to Z$06. The global output is recovered by quotient-based merging: clusters lying in $f:X\to Z$07 appear in both neighboring subproblems and are identified as single vertices in the merged graph (Hajij et al., 2017). The paper gives an inductive correctness proof showing that the parallel output matches the sequential Mapper computation for the same data and covers (Hajij et al., 2017).

The reported performance analysis uses Amdahl’s law,

$f:X\to Z$08

with the experimentally observed estimate $f:X\to Z$09 because clustering accounts for approximately $f:X\to Z$10 of runtime (Hajij et al., 2017). With $f:X\to Z$11 processes, this yields $f:X\to Z$12, and as $f:X\to Z$13 the speedup approaches $f:X\to Z$14 (Hajij et al., 2017). The experiments reported there indicate that speedup saturates beyond $f:X\to Z$15–$f:X\to Z$16 processes for typical datasets (Hajij et al., 2017).

Although this work is formulated for Mapper rather than the entire multiscale pipeline, its partition–cover–process–merge design is presented as naturally extendable to multiscale settings, where covers change across resolutions and local computations must still be merged coherently (Hajij et al., 2017). A plausible implication is that large-scale computational Multiscale Mapper inherits two distinct engineering problems: the cost of repeated clustering across scales, and the maintenance of consistent inter-scale or inter-region identifications during merging.