Mapper Construction: Topology and HD Mapping

• Mapper Construction is a method that converts spaces into nerve complexes via filter functions and connected pullbacks, forming a discrete summary akin to Reeb graphs.
• It leverages categorical formalism and convergence bounds to ensure robust approximations, with outcomes sensitive to cover design and clustering strategies.
• In autonomous driving, mapper pipelines aggregate local lane or instance observations using temporal association and spatial fusion to build global HD maps.

Mapper construction most commonly denotes the topological data analysis procedure that converts a space or dataset, a filter function, and a cover of the filter image into a nerve complex built from connected pullbacks; in a distinct recent usage in autonomous driving, the same phrase denotes pipelines that aggregate local lane or instance observations into a global vectorized HD map (Munch et al., 2015, Cho et al., 14 Apr 2025). In the topological setting, Mapper is a discrete summary closely related to Reeb graphs and Reeb spaces, and its behavior is governed by the geometry of the cover, the clustering used on pullbacks, and the stability properties of the resulting nerve. In the mapping setting, the emphasis shifts to temporal association, spatial fusion, and large-scale vector aggregation (Zhu et al., 5 Mar 2025).

1. Classical topological construction

In its standard form, Mapper starts with a dataset or topological space $X$, a filter or lens $f:X\to Y$, and an open cover $\mathcal{U}=\{U_\alpha\}_{\alpha\in A}$ of $f(X)$. The pullback cover is obtained by taking the path-connected components of each preimage $f^{-1}(U_\alpha)$:

$$f^*(\mathcal{U})=\bigcup_{\alpha\in A}\pi_0\bigl(f^{-1}(U_\alpha)\bigr).$$

The Mapper is then the nerve of this pullback cover,

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

where simplices encode nonempty intersections among the connected pullback components (Munch et al., 2015).

This construction is often specialized to scalar filters $f:X\to\mathbb{R}$, where the cover is given by overlapping intervals. In that setting, the practical parameterization is usually expressed through the lens $f$, a bin diameter or cover scale $r$, and a bin overlap $f:X\to Y$0; in discrete point-cloud settings, path-connected components are approximated by clusters inside each pullback set (Deb et al., 2019). In the 1-dimensional case, the 1-skeleton of the nerve is the usual Mapper graph, and for suitable interval covers it can be regarded as a “relaxed” Reeb graph (Munch et al., 2015).

A technically refined 1-dimensional formulation uses a gomic, namely a cover of $f:X\to Y$1 that is generic, open, minimal, and whose elements are intervals. Because no more than two elements of a minimal open interval cover can intersect at a time, the corresponding Mapper is a graph rather than a higher-dimensional simplicial complex (Carrière et al., 2015). This scalar setting is the basis for many theoretical results on structure, convergence, and stability.

2. Reeb-space interpretation and categorical formalization

The Reeb space of a continuous map $f:X\to Y$2 identifies points that lie in the same path-connected component of the same fiber. Formally,

$f:X\to Y$3

The quotient $f:X\to Y$4 is the Reeb space; when $f:X\to Y$5, it is the Reeb graph (Munch et al., 2015).

A key development is the categorical representation of both Mapper and Reeb space. For a good cover $f:X\to Y$6 with nerve $f:X\to Y$7, categorical Mapper is encoded as a functor

$f:X\to Y$8

where $f:X\to Y$9. The Reeb space is encoded as the functor

$\mathcal{U}=\{U_\alpha\}_{\alpha\in A}$0

A pushforward functor $\mathcal{U}=\{U_\alpha\}_{\alpha\in A}$1 transfers categorical Mapper into the same target category as the categorical Reeb space, making them directly comparable (Munch et al., 2015).

The comparison is made by an interleaving distance on $\mathcal{U}=\{U_\alpha\}_{\alpha\in A}$2. If $\mathcal{U}=\{U_\alpha\}_{\alpha\in A}$3 denotes the supremum of diameters of cover elements, the main convergence theorem states

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

Hence, as cover resolution tends to zero, Mapper converges to the Reeb space in interleaving distance, and at fixed resolution the same bound quantifies approximation quality (Munch et al., 2015). In the scalar case, this categorical convergence can be promoted to geometric convergence of a reconstructed graph-valued Mapper to the Reeb graph.

3. Cover design, clustering, and learned variants

A central practical issue is that Mapper is highly sensitive to cover design. Standard constructions with a single global scale often require substantial parameter tuning, and a uniform cover can simultaneously be too coarse in dense regions and too fine in sparse regions (Deb et al., 2019). This has motivated variants that modify the construction at the level of the cover rather than only at the level of the filter.

Multimapper replaces a single global cover by a family of locally chosen covers. Given a collection of subsets $\mathcal{U}=\{U_\alpha\}_{\alpha\in A}$5 with $\mathcal{U}=\{U_\alpha\}_{\alpha\in A}$6 and covers $\mathcal{U}=\{U_\alpha\}_{\alpha\in A}$7, it defines

$\mathcal{U}=\{U_\alpha\}_{\alpha\in A}$8

This permits locally finer or coarser resolutions, with scale selection guided by data density and persistence-based diagnostics; for 2-dimensional lenses, the brick cover further limits simultaneous intersections so that the Mapper has at most 2-simplices (Deb et al., 2019).

A different line of work attempts to learn the cover directly. G-Mapper optimizes the cover by repeatedly splitting intervals according to an Anderson–Darling normality test, using a Gaussian mixture model to place overlapping child intervals in a data-adaptive way rather than fixing a uniform resolution in advance (Alvarado et al., 2023). A related probabilistic formulation introduces implicit intervals through a hidden assignment matrix induced by a Gaussian mixture model on the filter values, with posteriors

$\mathcal{U}=\{U_\alpha\}_{\alpha\in A}$9

and defines a Mapper graph mode as a point estimate, optimized by stochastic gradient descent with a likelihood-plus-topology loss (Tao et al., 2024).

Fuzzy generalizations extend cover construction from hard pullback clusters to thresholded membership functions. In particular, Gustafson–Kessel Mapper replaces spherical fuzzy covers with ellipsoidal covers induced by Gustafson–Kessel fuzzy clustering, making the construction geometry-adaptive for anisotropic and non-spherical data; the resulting graph is locally stable under small perturbations of the fuzzifier, and graph changes occur only at threshold-crossing critical events (Sen et al., 19 Jun 2026).

4. Stability, convergence, and pathologies

In the scalar Morse-type setting, Mapper structure can be read directly from the extended persistence diagram of the Reeb graph. For a cover $f(X)$0, ordinary, relative, and extended features survive or disappear according to whether their diagram points lie outside or inside staircase regions $f(X)$1, $f(X)$2, and $f(X)$3. If the granularity of $f(X)$4 is at most $f(X)$5, then every Reeb-graph feature with persistence $f(X)$6 is guaranteed to appear in the MultiNerve Mapper, and the MultiNerve Mapper converges to the Reeb graph in functional distortion distance with error bounded by $f(X)$7; for the ordinary Mapper, the corresponding bound is $f(X)$8 (Carrière et al., 2015).

A probabilistic refinement of this program introduces the enhanced mapper graph as the display locale of a mapper cosheaf. In that setting, the deterministic approximation bound becomes

$f(X)$9

and under sampling from a density concentrated near the underlying $f^{-1}(U_\alpha)$0-space, the random enhanced mapper graph converges with high probability to the true Reeb graph, again at error controlled by cover resolution (Brown et al., 2019).

These positive results coexist with strong negative and cautionary statements. One common misconception is that a Mapper graph is close to an intrinsic invariant of the dataset. In fact, the inverse problem has an affirmative answer: if $f^{-1}(U_\alpha)$1 is a graph with edge set $f^{-1}(U_\alpha)$2 and isolated vertex set $f^{-1}(U_\alpha)$3, and $f^{-1}(U_\alpha)$4, then there exists a cover and a function $f^{-1}(U_\alpha)$5 such that the Mapper graph with trivial clustering is isomorphic to $f^{-1}(U_\alpha)$6 (Alvarado et al., 2024). A second misconception is that clustering choice is a secondary implementation detail. Fixed-count clustering methods such as $f^{-1}(U_\alpha)$7-means are described as systematically over-producing clusters in some cover elements and under-producing them in others, thereby creating holes or quotienting distinct topological structures; the conclusion is that globally fixed cluster counts should be avoided in Mapper construction (Vejdemo-Johansson, 8 Jul 2025). These results indicate that the output depends strongly on lens choice, cover design, and the clustering model used on pullbacks.

5. Domain-specific and compositional constructions

Several specialized constructions adapt Mapper to settings where the ambient geometry or computational constraints differ substantially from the standard point-cloud model.

On simply connected image domains, Mapper can be computed directly from a continuous scalar height function on the pixel grid. The range is covered by overlapping intervals $f^{-1}(U_\alpha)$8, split into even and odd subcovers so that no two intervals in the same subcover overlap; connected regions are then found by BFS on the pixel adjacency graph, edges are recovered from overlap checks on candidate pixels, and valence-2 nodes can be removed without changing topological events. With suitable covers tied to critical values, this image-specific Mapper recovers contour, join, and split trees on simply connected domains (Robles et al., 2017).

For weighted undirected graphs, Mapper on Graphs replaces metric clustering in pullback sets by clustering on induced subgraphs. A topological lens $f^{-1}(U_\alpha)$9 is defined on graph vertices, the normalized filter range is covered by intervals, each pullback induces a subgraph $$f^*(\mathcal{U})=\bigcup_{\alpha\in A}\pi_0\bigl(f^{-1}(U_\alpha)\bigr).$$0, and graph clustering yields nodes in the Mapper-on-Graphs skeleton. Instead of classical overlap edges, the construction uses graph cuts between clusters, producing homology-preserving multi-scale skeletons of the input graph (Rosen et al., 2018).

Compositional variants analyze how topology changes when scalar filters are combined. For two univariate mappers $$f^*(\mathcal{U})=\bigcup_{\alpha\in A}\pi_0\bigl(f^{-1}(U_\alpha)\bigr).$$1 and $$f^*(\mathcal{U})=\bigcup_{\alpha\in A}\pi_0\bigl(f^{-1}(U_\alpha)\bigr).$$2, the stitching construction

$$f^*(\mathcal{U})=\bigcup_{\alpha\in A}\pi_0\bigl(f^{-1}(U_\alpha)\bigr).$$3

builds a bivariate mapper from intersections of pullback cover elements, and localized homological differences, local relative Euler characteristic, and localized entropy differences quantify the additional information gained during stitching (Zhou et al., 2021). A related enrichment is the Steinhaus Mapper filtration $$f^*(\mathcal{U})=\bigcup_{\alpha\in A}\pi_0\bigl(f^{-1}(U_\alpha)\bigr).$$4, which assigns Steinhaus weights to simplices of a single Mapper and supports stable-path computations on the 1-skeleton; this produces overlap-based filtrations and path explanations without requiring a tower of covers (Arendt et al., 2019).

Domain	Construction principle	Outcome
Images	Even/odd interval covers, BFS connected components, valence-2 simplification	Fast Mapper computation; contour/join/split tree recovery
Graphs	Lens on vertices, induced-subgraph clustering, graph-cut edges	Homology-preserving graph skeletonization
Composed filters	Stitching $$f^*(\mathcal{U})=\bigcup_{\alpha\in A}\pi_0\bigl(f^{-1}(U_\alpha)\bigr).$$5 and $$f^*(\mathcal{U})=\bigcup_{\alpha\in A}\pi_0\bigl(f^{-1}(U_\alpha)\bigr).$$6 into $$f^*(\mathcal{U})=\bigcup_{\alpha\in A}\pi_0\bigl(f^{-1}(U_\alpha)\bigr).$$7	Bivariate Mapper and topological gains
Single-cover enrichment	Steinhaus weights on Mapper simplices	Stable paths and overlap-based filtration

6. Lane-level and online HD map construction

In autonomous-driving research, “mapper construction” designates a different class of problems: building vectorized HD maps from visual or multimodal observations. SIO-Mapper is a lane-level HD map construction framework that uses only satellite imagery and OpenStreetMap data, without physical site visits. Its pipeline combines a dual-encoder network, SIO-Net, for lane extraction with a lane-integration stage that converts per-tile lane masks into a city-scale vector map. The integration module contains a clustering-based mapper $$f^*(\mathcal{U})=\bigcup_{\alpha\in A}\pi_0\bigl(f^{-1}(U_\alpha)\bigr).$$8, a graph-based mapper $$f^*(\mathcal{U})=\bigcup_{\alpha\in A}\pi_0\bigl(f^{-1}(U_\alpha)\bigr).$$9, and a map merger based on the Hungarian algorithm; the final map $$M(\mathcal{U},f):=\mathrm{Nrv}\bigl(f^*(\mathcal{U})\bigr),$$0 is evaluated with map coverage $$M(\mathcal{U},f):=\mathrm{Nrv}\bigl(f^*(\mathcal{U})\bigr),$$1, map accuracy $$M(\mathcal{U},f):=\mathrm{Nrv}\bigl(f^*(\mathcal{U})\bigr),$$2, and mean vertex distance $$M(\mathcal{U},f):=\mathrm{Nrv}\bigl(f^*(\mathcal{U})\bigr),$$3 (Cho et al., 14 Apr 2025).

IC-Mapper addresses online vectorized map construction from multi-camera video. It maintains a global map

$$M(\mathcal{U},f):=\mathrm{Nrv}\bigl(f^*(\mathcal{U})\bigr),$$4

of static road elements, where $$M(\mathcal{U},f):=\mathrm{Nrv}\bigl(f^*(\mathcal{U})\bigr),$$5 is a semantic class, $$M(\mathcal{U},f):=\mathrm{Nrv}\bigl(f^*(\mathcal{U})\bigr),$$6 is a BEV polyline or polygon, and $$M(\mathcal{U},f):=\mathrm{Nrv}\bigl(f^*(\mathcal{U})\bigr),$$7 is a persistent instance ID. The framework consists of an instance-centric temporal association module and an instance-centric spatial fusion module. Temporal association computes a similarity tensor

$$M(\mathcal{U},f):=\mathrm{Nrv}\bigl(f^*(\mathcal{U})\bigr),$$8

combining feature and geometric correspondence between current detections and tracked instances; spatial fusion samples points from the historical global map in the current patch and fuses them with current detections to update the global vector map in real time (Zhu et al., 5 Mar 2025).

The two usages are therefore distinct. In topological data analysis, Mapper construction studies nerves of pullback covers, Reeb-space approximation, and stability under cover refinement. In online HD mapping, mapper construction studies lane extraction, instance association, spatial fusion, and global vector-map aggregation at city or trajectory scale (Munch et al., 2015, Zhu et al., 5 Mar 2025).