2-Mapper: Multi-Interpretation in TDA

• 2-Mapper is a term in TDA for constructions employing a two-dimensional filter or two-skeleton of clustered covers, each with distinct technical nuances.
• The approach extends traditional Mapper by capturing multidirectional features and refining one-dimensional topology via β1, improving the interpretation of cycles.
• Its sensitivity to parameter settings—such as cover resolution, overlap, and clustering methods—necessitates careful tuning for robust and stable topological inference.

2-Mapper is used in several distinct technical senses. Within the Mapper literature, the most relevant interpretation is usually “Mapper with a 2-dimensional filter,” namely a multivariate Mapper in which the lens $f$ maps into $\mathbb{R}^2$ (Oulhaj et al., 2024). A more specific recent usage defines 2-Mapper as the two-skeleton of the nerve of the clustered pullback cover, introduced in order to “observe two-dimensional features, and capture one-dimensional topology” through $\beta_1$ (Fritze, 26 Sep 2025). The same label also appears in unrelated settings, including “2-Mapper (MFM-Mapper)” for video-to-audio generation (Chen et al., 5 Sep 2025), the hierarchical quantum mapper Squash 2 (Dousti et al., 2015), and Two-Tier Mapper (TTMap) for gene expression analysis (Jeitziner et al., 2017). A common misconception is therefore that 2-Mapper has a single universally accepted definition; the published record instead shows a family of related and unrelated usages.

1. Terminological scope

In topological data analysis, “2-Mapper” most often refers either to Mapper with a two-dimensional lens or to a Mapper complex enriched up to dimension two. The former interpretation is explicit in the differentiable Mapper literature: “The most relevant interpretation within the Mapper literature is ‘Mapper with a 2-dimensional filter,’ i.e., a multivariate Mapper where the lens $f$ maps into $\mathbb{R}^2$” (Oulhaj et al., 2024). The latter interpretation is explicit in the multiscale construction: “The 2-Mapper proposed in this paper is the two-skeleton of the nerve of the clustered pullback cover” (Fritze, 26 Sep 2025).

Usage	Definition	Reference
2D-filter Mapper	Mapper with $f:X\to\mathbb{R}^2$	(Oulhaj et al., 2024)
Two-skeleton 2-Mapper	$N^2(f^*(\mathcal U))$ of the clustered pullback cover	(Fritze, 26 Sep 2025)
Other non-TDA usages	MFM-Mapper, Squash 2, TTMap	(Chen et al., 5 Sep 2025, Dousti et al., 2015, Jeitziner et al., 2017)

This terminological plurality matters because different papers attach different mathematical objects to the same name. In one line of work, 2-Mapper remains fundamentally a cover-and-cluster construction whose added complexity comes from a multivariate codomain. In another, it is specifically the passage from the Mapper graph to the full two-skeleton, so that 2-simplices encode triple overlaps and directly affect $\beta_1$. A plausible implication is that any technical discussion of 2-Mapper must specify whether “2” refers to the codomain dimension, the simplicial dimension retained, or an application-specific system name.

2. Core construction in the Mapper literature

The classical Mapper pipeline is common to both 1- and 2-dimensional variants. Let $X$ be a finite dataset, let $f:X\to\mathbb{R}^m$, choose a finite cover $\mathbb{R}^2$0 of $\mathbb{R}^2$1, pull back the cover to $\mathbb{R}^2$2, cluster each pullback set as $\mathbb{R}^2$3, and define

$\mathbb{R}^2$4

Equivalently, in the pullback-cover formulation,

$\mathbb{R}^2$5

with

$\mathbb{R}^2$6

Here the nerve $\mathbb{R}^2$7 has one vertex per cluster and includes a simplex whenever the corresponding clusters have non-empty intersection (Deb et al., 2019, Alvarado et al., 2024).

For 2-Mapper in the multivariate sense, the codomain is $\mathbb{R}^2$8 rather than $\mathbb{R}^2$9. In the one-dimensional exposition, the cover consists of overlapping intervals; “for $\beta_1$0 (2-Mapper), replace intervals by overlapping sets in $\beta_1$1 (e.g., rectangles/disks in a grid), with resolution and gain generalized to 2D” (Oulhaj et al., 2024). A standard usage is therefore a rectangular grid cover induced by a lens $\beta_1$2, with pullback bins $\beta_1$3 and clustering performed separately inside each pullback region (Belchí et al., 2019).

In the cubical-cover formulation, with $\beta_1$4 intervals per coordinate and overlap fraction $\beta_1$5, the boxes are given by

$\beta_1$6

and

$\beta_1$7

This makes explicit that 2-Mapper inherits all of Mapper’s dependence on cover resolution, overlap, and clustering choices (Fritze, 26 Sep 2025).

The technical motivation for moving from $\beta_1$8 to $\beta_1$9 is not merely visual. Multi-dimensional filters can “capture multi-directional phenomena, disentangle structures not visible in a single lens, and reduce projection-induced artifacts,” while the trade-off is that “cover design and parameter tuning are more complex,” the “combinatorial size of pullback sets increases,” and “more samples may be needed to stably populate overlapping 2D cover elements” (Oulhaj et al., 2024). This suggests that 2-Mapper is best understood as an extension of Mapper’s expressivity rather than a cosmetic generalization.

3. 2-Mapper as the two-skeleton of the clustered pullback nerve

A distinct recent definition sets 2-Mapper equal to the two-skeleton of the nerve of the clustered pullback cover:

$f$0

Under this definition, vertices correspond to clusters $f$1, edges are included when $f$2, and a 2-simplex is included exactly when a triple intersection is nonempty:

$f$3

The construction is motivated by the observation that “most tools only render the 1-skeleton (Mapper graph),” which omits the 2-simplices that determine whether graph cycles should be filled (Fritze, 26 Sep 2025).

The central topological claim is that “two-dimensional features (2-simplices) are exactly the combinatorial certificates of triple overlaps.” Adding them “suppresses spurious cycles and encodes local clustering density,” because triangles fill 1-cycles that would otherwise remain visible in the 1-skeleton. Conversely, if cover overlap or clustering thresholds are too sparse, cycles remain unfilled and $f$4 is inflated. In this formulation, 2-Mapper is designed to capture one-dimensional topology more faithfully through the first Betti number $f$5 (Fritze, 26 Sep 2025).

The multiscale version constructs a tower of covers $f$6, builds a family of complexes $f$7, and studies the persistence module

$f$8

through the induced maps $f$9. The resulting $\mathbb{R}^2$0 barcode is used for parameter selection: one chooses scales where the barcode shows long-lived one-dimensional classes and the 2-Mapper visualization remains interpretable. The paper’s explicit recommendation from the Klein bottle experiment is an overlap range $\mathbb{R}^2$1 (Fritze, 26 Sep 2025).

The same work also states stability results for multiscale 2-Mapper. For the cubical tower with $\mathbb{R}^2$2 and overlap $\mathbb{R}^2$3, the cover tower is $\mathbb{R}^2$4-good, and the induced tower of cluster covers under DBSCAN with MinPts $\mathbb{R}^2$5 and no free border points is $\mathbb{R}^2$6-good with $\mathbb{R}^2$7. By the stated Dey–Mémoli–Wang framework, the resulting persistence diagrams inherit the same type of stability under reasonable perturbations of the cover and lens (Fritze, 26 Sep 2025). This places 2-Mapper within the broader multiscale Mapper program while giving it a specifically homological target.

4. Instability, parameter sensitivity, and practical tuning

Mapper-type algorithms are notoriously parameter-sensitive. The instability framework of Belchi et al. is formulated for arbitrary covers and therefore applies directly to 2-Mapper, including the case of a 2D lens with a rectangular grid cover. In the empirical setting, for Mapper functions $\mathbb{R}^2$8 on a sample $\mathbb{R}^2$9, the Mapper distance is

$f:X\to\mathbb{R}^2$0

and Mapper instability is the expectation of this mismatch across subsamples:

$f:X\to\mathbb{R}^2$1

The theoretical upper bound

$f:X\to\mathbb{R}^2$2

shows that instability is controlled by tube mass near decision boundaries, deviations of empirical clusterings from population clusterings, and empty pullback bins (Belchí et al., 2019).

For 2-Mapper, the parameterization is naturally

$f:X\to\mathbb{R}^2$3

where $f:X\to\mathbb{R}^2$4 are resolutions and $f:X\to\mathbb{R}^2$5 are overlaps in the two filter directions. The same paper recommends estimating instability by subsampling, computing pairwise $f:X\to\mathbb{R}^2$6 on intersections, averaging over subsamples, and identifying “local minima” of instability as reliable parameter settings. It also reports that structural transitions in Mapper graphs coincide with ridges or spikes in instability, while local minima correspond to reliable configurations (Belchí et al., 2019).

The practical implications are direct. Increasing the number of bins tends to increase instability because per-bin masses $f:X\to\mathbb{R}^2$7 shrink. Low overlap can produce fragmentation, while high boundary complexity and many near-minimizers of the clustering objective increase sensitivity. The paper therefore recommends moderate overlaps, avoiding too many tiny bins, and preferring clustering methods whose quality functions have distinctive optima (Belchí et al., 2019). For 2-Mapper specifically, these considerations explain why codomain dimensionality increases flexibility and instability at the same time.

5. Differentiable, adaptive, and inverse formulations

Differentiable Mapper extends the 2-Mapper viewpoint by making the cover assignment probabilistic and smooth. The construction introduces a latent cover assignment matrix $f:X\to\mathbb{R}^2$8, a Mapper-complex generating function

$f:X\to\mathbb{R}^2$9

and a smooth assignment scheme $N^2(f^*(\mathcal U))$0 whose Bernoulli parameters are given by soft cover membership functions $N^2(f^*(\mathcal U))$1. The optimization objective is

$N^2(f^*(\mathcal U))$2

minimized by stochastic gradient descent. Crucially for 2-Mapper, the paper remarks that “the same relaxed construction can be made for a multi-dimensional Mapper, i.e., for filter functions taking values in $N^2(f^*(\mathcal U))$3, by making slight adjustments to the definition of $N^2(f^*(\mathcal U))$4” (Oulhaj et al., 2024). This makes 2D filter optimization part of the differentiable-topological pipeline rather than a separate preprocessing task.

A separate adaptive line of work replaces a single global scale with local scale selection. Multimapper defines

$N^2(f^*(\mathcal U))$5

thereby gluing together region-specific local Mappers. The motivation is density sensitivity: denser regions can be refined more aggressively while sparser regions are kept coarser so as to avoid shattering. In experiments using a 2D t-SNE lens, Multimapper magnifies selected regions while preserving global connectivity, and the proposed brick cover keeps intersection degree at most $N^2(f^*(\mathcal U))$6 in 2D for overlaps below $N^2(f^*(\mathcal U))$7, so the nerve has at most 2-simplices (Deb et al., 2019).

The inverse problem provides a different cautionary perspective. “Any Graph is a Mapper Graph” proves that if $N^2(f^*(\mathcal U))$8 has edge set $N^2(f^*(\mathcal U))$9 and isolated vertex set $\beta_1$0, and $\beta_1$1, then one can choose a cover and a function $\beta_1$2 so that Mapper with trivial clustering is isomorphic to $\beta_1$3. The paper also distinguishes “1-Mapper” and “2-Mapper,” with the latter meaning a two-dimensional filter $\beta_1$4 (Alvarado et al., 2024). It follows that the visible graph structure of a Mapper output, by itself, is not a certificate of intrinsic topology. The same work further notes that arbitrary graphs cannot all be realized by convex covers in $\beta_1$5; by Wegner’s theorem, arbitrary graph representability by convex sets requires dimension $\beta_1$6 (Alvarado et al., 2024). A plausible implication is that 2-Mapper interpretation must always be tied to the chosen lens, cover geometry, and clustering model.

6. Other technical uses of the name

Outside classical TDA, “2-Mapper” has been used for several unrelated systems. In video-to-audio generation, “2-Mapper (MFM-Mapper)” denotes “a dual-foundation-model mapper for efficient Video-to-Audio generation.” Its architecture freezes CAVP, TimeChat, AudioLDM-2, AudioMAE, and FLAN-T5, and fine-tunes only a GPT-2 mapper that predicts pooled AudioMAE embeddings from fused visual features. The reported result is better semantic fidelity and synchronization than a prior mapper baseline while using only $\beta_1$7 of that baseline’s training scale (Chen et al., 5 Sep 2025). Here “mapper” means a learned cross-modal connector, not a topological nerve construction.

In quantum compilation, Squash 2 is described as a hierarchical and scalable quantum mapper for the Requp multi-core reconfigurable quantum processor. It decomposes a quantum circuit into modules, partitions each module across $\beta_1$8 cores, and schedules execution while sharing physical and logical ancilla qubits. The emphasis is on fault-tolerant resource management, not on topological data analysis (Dousti et al., 2015).

In computational biology, Two-Tier Mapper (TTMap, “2-Mapper”) is a topology-based, user-independent clustering workflow for global gene expression analysis. It combines Hyperrectangle Deviation Assessment with a Global-to-Local Mapper using a two-tier cover

$\beta_1$9

a mismatch distance

$X$0

and a data-driven choice of the closeness parameter $X$1 via a Chen–Stein Poisson approximation (Jeitziner et al., 2017). In this setting, “2” refers to a two-tier network organization rather than to a two-dimensional filter or a two-skeleton.

Taken together, these usages show that “2-Mapper” is not a single artifact but a family of names centered on the idea of a mapper or mapping mechanism. In contemporary TDA, however, the dominant technical meanings remain the 2D-filter variant of Mapper and the two-skeleton variant designed to capture $X$2 more faithfully.