Multiscale Topological Differentiation

• Multiscale Topological Differentiation is a methodological pattern that computes scale-dependent invariants using filtrations, bifiltrations, and spectral methods.
• It integrates topological summaries such as persistence diagrams, Betti curves, and Euler characteristic surfaces to distinguish structures in varied data modalities.
• Applications span materials discovery, spatial transcriptomics, and network analysis by separating meaningful signal from noise and parameter-induced artifacts.

Searching arXiv for papers mentioning “multiscale topological differentiation” and closely related formulations. Multiscale topological differentiation is a family of methods in which topology is computed across varying scales and then used as a discriminative signal for structure, dynamics, or function. In the cited literature, the relevant scales may be metric thresholds, cover overlaps, filtration radii, time, codensity, mark-weighted distance, or pairs of starting and ending resolutions; the topological summaries may be homology groups, Betti curves, persistence diagrams, persistence landscapes, Euler characteristic surfaces, Hilbert functions, persistent Laplacians, or localized divergence scores (Xia et al., 2015, Chen et al., 2019, Roy et al., 2024, Du et al., 2023). The term therefore does not denote a single canonical algorithm. Rather, it denotes a methodological pattern: one constructs a scale-indexed family of complexes or related objects, extracts scale-dependent invariants, and then uses their evolution to classify, screen, rank, compare, or localize salient phenomena.

1. Conceptual scope and development

Related formulations appeared early in chromatin analysis, where a single scale parameter $\gamma$ controlled domain size and a consensus procedure extracted persistent domains across resolutions (Filippova et al., 2013). A closely aligned development was multiresolution persistent homology, which introduced a resolution parameter $\eta$ through a rigidity density and described the resulting procedure as a “topological microscope” for large data sets (Xia et al., 2015). In dynamical-systems settings, persistent homology was then used to build “multiscale topological portraits” of coordination dynamics, while the evolutionary de Rham–Hodge framework tracked zero and nonzero Hodge spectra along filtrations of evolving manifolds (Zhang et al., 2019, Chen et al., 2019).

Subsequent work expanded the same idiom across several data modalities. Persistent local homology and Euclidicity were used for multiscale singularity detection (Rohrscheidt et al., 2022). TopACT and multiparameter persistent homology landscapes connected subcellular, cellular, and multicellular scales in spatial transcriptomics (Benjamin et al., 2022). Persistent Laplacians were used for node-level differentiation in interactomic networks (Du et al., 2023). Localized multiscale descriptors were developed for knots via the multiscale Gauss link integral (Shen et al., 2023), for scalar fields via Scalar Function Topology Divergence (Trofimov et al., 2024), for time series via Euler Characteristic Surfaces (Roy et al., 2024), for superionic conductor discovery via multiscale topological learning (Chen et al., 2024), for heterogeneous cell data via persistence-weighted death simplices and tuned vectorizations (Torras-Pérez et al., 5 May 2025), for soft-matter images via the $\Psi$-function (Orlova et al., 28 Jul 2025), for multiscale Mapper via dimension-1 persistence (Fritze, 26 Sep 2025), for nonhierarchical sequences of partitions via the Multiscale Clustering Bifiltration (Schindler et al., 16 Oct 2025), and for marked point processes via Euler Characteristic envelopes and local Z-scores (Eckardt et al., 14 May 2026).

This distribution across domains suggests that the unifying content of multiscale topological differentiation is formal rather than disciplinary. It is the coupling of topology with scale variation, not the choice of application area, that defines the paradigm.

2. Mathematical architecture

The most common mathematical substrate is a filtration of simplicial complexes. In persistent-homology formulations, one starts from a point cloud or metric space, builds a scale-indexed family such as the Vietoris–Rips complexes

$$R_\varepsilon(X)=\{\sigma\subset X\mid \mathrm{diam}(\sigma)\le \varepsilon\},$$

and computes homology

$$H_k\bigl(C(\varepsilon)\bigr)=\ker(\partial_k)\big/\mathrm{im}(\partial_{k+1}), \qquad \beta_k(\varepsilon)=\mathrm{rank}\,H_k\bigl(C(\varepsilon)\bigr),$$

with persistence diagrams recording the birth and death of scale-dependent features (Zhang et al., 2019). The same general pattern appears in Vietoris–Rips, alpha, witness, sublevel-set, and mark-weighted Rips constructions (Xia et al., 2015, Torras-Pérez et al., 5 May 2025, Eckardt et al., 14 May 2026).

A second recurrent architecture is multiparameter. In spatial transcriptomics, the bifiltration

$$X_{r,t}=\mathrm{Rips}\bigl(\{p\in P:\rho_k(p)\le t\},r\bigr)$$

combines Rips radius and codensity (Benjamin et al., 2022). In cluster-sequence analysis, the Multiscale Clustering Bifiltration defines

$$K^{s,t} = \bigcup_{r\in[s,t]}\;\bigcup_{C\in \theta(r)} \Delta(C),$$

and studies the corresponding two-parameter persistence module through Hilbert functions

$\mathrm{HF}_k(s,t)=\dim_\Bbbk H_k(K^{s,t}),$

with $\mathrm{HF}_0$ measuring refinement failures and $\mathrm{HF}_1$ measuring higher-order inconsistencies (Schindler et al., 16 Oct 2025). Persistent local homology introduces a tri-parameter filtration over annulus radii and Rips scale (Rohrscheidt et al., 2022). Euler Characteristic Surfaces use a spatiotemporal parameter pair and define

$\eta$0

thereby encoding both scale and time in a single surface (Roy et al., 2024).

A third architecture is spectral. In interactomic networks, the persistent $\eta$1-Laplacian

$\eta$2

combines Betti information, via the nullity, with nonzero eigenvalues that capture finer geometric evolution (Du et al., 2023). In the evolutionary de Rham–Hodge framework, the zero eigenvalues of the evolving Hodge Laplacian recover persistent Betti numbers, while nonzero spectra quantify geometric progression during manifold evolution (Chen et al., 2019).

A fourth architecture is localization. Persistence-weighted death simplices place the simplex responsible for a death event back in the original embedding (Torras-Pérez et al., 5 May 2025). Scalar Function Topology Divergence constructs an F-Cross-Barcode on a doubled graph so that topological mismatches are localized to the same region of the domain (Trofimov et al., 2024). In marked point processes, local connectivity scores and Z-scores are computed at the critical filtration scale to identify hubs and barriers (Eckardt et al., 14 May 2026). These formulations are significant because they weaken the common assumption that multiscale topology is only a global summary.

3. Operational modes of differentiation

One operational mode is threshold-based screening. In lithium superionic conductor discovery, the multiscale topological learning framework separates each crystal into Li-only and Li-free substructures, computes persistent homology on both, and derives the minimum connectivity distance

$\eta$3

and the cycle density

$\eta$4

Compounds with $\eta$5 Å are filtered out, and among the remainder those with $\eta$6 are removed, reducing a library of $\eta$7 Li-containing materials to $\eta$8 for deeper analysis. Affinity Propagation on 22 normalized topological features then yields 32 clusters, 8 of which contain the majority of known LSICs; chemical screening reduces the pool to 44 candidates; AIMD validation identifies 14 LSICs, four of which were independently validated in recent experiments (Chen et al., 2024). Here “differentiation” is explicitly a downselection mechanism.

A second operational mode is persistence-guided parameter choice. In Multiscale 2-Mapper, one varies cover parameters, aligns successive 2-Mapper complexes into a filtration, computes dimension-1 persistence, and interprets long $\eta$9 bars as genuine loops while short bars are treated as noise. In the Klein bottle experiment, a single persistent class appears roughly in the range $\Psi$0; for $\Psi$1 the cover is too fine, and for $\Psi$2 it is too coarse (Fritze, 26 Sep 2025). In chromatin, varying the scale parameter $\Psi$3 changes the penalty on larger intervals, and persistence across a set of $\Psi$4 values is converted into a weighted interval-scheduling problem to obtain consensus domains (Filippova et al., 2013). In multiresolution persistent homology, changing the resolution parameter $\Psi$5 reveals different structural bands, as in the three-scale hexagonal fractal (Xia et al., 2015). In these cases, differentiation means separating stable topological signal from parameter-induced artifacts.

A third mode is perturbative ranking. In interactomic networks, each node is deleted in turn, persistent-Laplacian feature vectors are recomputed, and node importance is scored by the Euclidean perturbation

$\Psi$6

then aggregated across STRING thresholds by

$\Psi$7

The intersection of the top-25 genes across thresholds $\Psi$8 yields robust key genes (Du et al., 2023). In this setting, multiscale topological differentiation is neither classification nor clustering but sensitivity analysis on a scale-dependent spectral fingerprint.

A fourth mode is local anomaly or manifoldness assessment. Persistent local homology defines a persistent intrinsic dimension

$\Psi$9

and Euclidicity compares local barcodes to a Euclidean model annulus via bottleneck distances averaged over a grid of annulus parameters (Rohrscheidt et al., 2022). High Euclidicity indicates singular or highly curved regions; low Euclidicity indicates neighborhoods that behave like flat $$R_\varepsilon(X)=\{\sigma\subset X\mid \mathrm{diam}(\sigma)\le \varepsilon\},$$0.

A fifth mode is direct topology comparison. SFTD computes the $$R_\varepsilon(X)=\{\sigma\subset X\mid \mathrm{diam}(\sigma)\le \varepsilon\},$$1-norm of F-Cross-Barcode interval lengths so that topological discrepancies between scalar functions are measured across scales and forced to be spatially matched (Trofimov et al., 2024). Euler Characteristic Surfaces differentiate in scale and time by examining spikes or finite differences in $$R_\varepsilon(X)=\{\sigma\subset X\mid \mathrm{diam}(\sigma)\le \varepsilon\},$$2 and $$R_\varepsilon(X)=\{\sigma\subset X\mid \mathrm{diam}(\sigma)\le \varepsilon\},$$3 (Roy et al., 2024). In marked point processes, global envelopes of Euler characteristic curves test random labelling, while the critical scale

$$R_\varepsilon(X)=\{\sigma\subset X\mid \mathrm{diam}(\sigma)\le \varepsilon\},$$4

is used to localize the strongest mark–space coupling (Eckardt et al., 14 May 2026). These formulations treat differentiation as comparison against a reference, null, or perturbed system.

4. Domain-specific realizations

Domain	Representative formulation	Differentiation target
Solid-state materials	MTL on Li-only and Li-free substructures (Chen et al., 2024)	Li-network connectivity, framework cycle density, LSIC screening
Spatial biology	TopACT + MPH landscapes; PWDS, Betti curves, persistence images (Benjamin et al., 2022, Torras-Pérez et al., 5 May 2025)	Cell-type recovery, immune rings, cavities, disease-associated tissue architecture
Dynamics and soft matter	Topological portraits; $$R_\varepsilon(X)=\{\sigma\subset X\mid \mathrm{diam}(\sigma)\le \varepsilon\},$$5-function (Zhang et al., 2019, Orlova et al., 28 Jul 2025)	Collective transitions, breathing, fingering, clustering regimes
Networks and multiscale clusterings	Persistent Laplacians; MCbiF (Du et al., 2023, Schindler et al., 16 Oct 2025)	Key-gene ranking, refinement failures, higher-order assignment loops
Curves, scalar fields, and time series	mGLI; SFTD; ECS (Shen et al., 2023, Trofimov et al., 2024, Roy et al., 2024)	Local entanglement, topology divergence, critical spatiotemporal changes
Singular and marked spatial data	PLH/Euclidicity; Euler Characteristic envelopes (Rohrscheidt et al., 2022, Eckardt et al., 14 May 2026)	Singularity detection, manifoldness, mark–space dependence, hubs and barriers

These realizations differ substantially in data model and target variable. Some start from point clouds, others from graphs, scalar fields, partitions, images, or evolving manifolds. Some are inherently one-parameter, others genuinely multi-parameter. Some emphasize localization, others emphasize invariance or statistical stability. Yet the same technical motif recurs: a scale-indexed topological object is built, a summary of its evolution is extracted, and that summary is used to separate favorable from unfavorable, stable from noisy, Euclidean from singular, or hierarchical from nonhierarchical regimes.

The domain dependence also affects what counts as the relevant topological feature. In LSIC discovery, connectedness of the Li sublattice and loop density of the Li-free framework are functionally interpretable proxies for diffusion pathways (Chen et al., 2024). In tissue imaging, degree-1 features may correspond to cavities, alveoli, follicles, or peripheral immune rings (Benjamin et al., 2022, Torras-Pérez et al., 5 May 2025). In cluster bifiltrations, nonzero $$R_\varepsilon(X)=\{\sigma\subset X\mid \mathrm{diam}(\sigma)\le \varepsilon\},$$6 is not a geometric loop in Euclidean space but a higher-order inconsistency in assignments across scales (Schindler et al., 16 Oct 2025). In knot data analysis, scale-banded Gauss link scores encode local entanglement rather than Betti numbers (Shen et al., 2023). The differentiation step is therefore application-specific even when the formal machinery is homologous.

5. Validation, interpretability, and computational trade-offs

Validation strategies in this literature are unusually heterogeneous. In materials discovery, topological screening is followed by chemical filtering and then by AIMD, with final criteria on $$R_\varepsilon(X)=\{\sigma\subset X\mid \mathrm{diam}(\sigma)\le \varepsilon\},$$7, $$R_\varepsilon(X)=\{\sigma\subset X\mid \mathrm{diam}(\sigma)\le \varepsilon\},$$8, and stability window; 14 candidates satisfy the declared LSIC conditions, and four were independently validated experimentally (Chen et al., 2024). In subcellular spatial transcriptomics, TopACT predictions were benchmarked on synthetic data, podocyte localization was checked against independent Bin-20 Seurat calls, and immune-ring findings were validated by CD45 immunofluorescence, with reported increases in glomerular immune counts at $$R_\varepsilon(X)=\{\sigma\subset X\mid \mathrm{diam}(\sigma)\le \varepsilon\},$$9 and fluorescence-ratio evidence at $$H_k\bigl(C(\varepsilon)\bigr)=\ker(\partial_k)\big/\mathrm{im}(\partial_{k+1}), \qquad \beta_k(\varepsilon)=\mathrm{rank}\,H_k\bigl(C(\varepsilon)\bigr),$$0 (Benjamin et al., 2022). In chromatin, multiscale consensus boundaries were reported to be strongly enriched for CTCF binding and histone marks, and Table 1 gives an IMR90 CTCF boundary-overlap fraction of 44% versus 20% for the comparison method (Filippova et al., 2013). In scalar-field comparison, SFTD used as an additional loss in 3D reconstruction reduced IoU-error from 0.487 to 0.449, compared with 0.467 for the Wasserstein-based alternative, and localized mistaken voxels in segmentation despite identical ordinary persistence barcodes (Trofimov et al., 2024).

Interpretability is often obtained by explicit localization. PWDS overlays death simplices on tissue images and colors them by birth-scale partition and persistence weight (Torras-Pérez et al., 5 May 2025). SFTD associates F-Cross-Barcode intervals with specific vertices or cliques in the original domain (Trofimov et al., 2024). Marked point-process analysis maps local Z-scores at the critical scale, so that points with $$H_k\bigl(C(\varepsilon)\bigr)=\ker(\partial_k)\big/\mathrm{im}(\partial_{k+1}), \qquad \beta_k(\varepsilon)=\mathrm{rank}\,H_k\bigl(C(\varepsilon)\bigr),$$1 act as connectivity hubs and points with $$H_k\bigl(C(\varepsilon)\bigr)=\ker(\partial_k)\big/\mathrm{im}(\partial_{k+1}), \qquad \beta_k(\varepsilon)=\mathrm{rank}\,H_k\bigl(C(\varepsilon)\bigr),$$2 act as topological barriers (Eckardt et al., 14 May 2026). This localization is methodologically important because classical persistence diagrams, by themselves, do not always indicate where in the data a discrepancy is located.

Computational cost remains a limiting factor and explains many of the methodological variants. Persistent local homology with Vietoris–Rips complexes inherits exponential growth in the neighborhood size parameter $$H_k\bigl(C(\varepsilon)\bigr)=\ker(\partial_k)\big/\mathrm{im}(\partial_{k+1}), \qquad \beta_k(\varepsilon)=\mathrm{rank}\,H_k\bigl(C(\varepsilon)\bigr),$$3 (Rohrscheidt et al., 2022). Standard Vietoris–Rips on $$H_k\bigl(C(\varepsilon)\bigr)=\ker(\partial_k)\big/\mathrm{im}(\partial_{k+1}), \qquad \beta_k(\varepsilon)=\mathrm{rank}\,H_k\bigl(C(\varepsilon)\bigr),$$4 points in $$H_k\bigl(C(\varepsilon)\bigr)=\ker(\partial_k)\big/\mathrm{im}(\partial_{k+1}), \qquad \beta_k(\varepsilon)=\mathrm{rank}\,H_k\bigl(C(\varepsilon)\bigr),$$5 can require $$H_k\bigl(C(\varepsilon)\bigr)=\ker(\partial_k)\big/\mathrm{im}(\partial_{k+1}), \qquad \beta_k(\varepsilon)=\mathrm{rank}\,H_k\bigl(C(\varepsilon)\bigr),$$6 time/memory to list all simplices of dimension $$H_k\bigl(C(\varepsilon)\bigr)=\ker(\partial_k)\big/\mathrm{im}(\partial_{k+1}), \qquad \beta_k(\varepsilon)=\mathrm{rank}\,H_k\bigl(C(\varepsilon)\bigr),$$7, which motivated multiresolution PH via rigidity densities and cubical filtrations with overall cost about $$H_k\bigl(C(\varepsilon)\bigr)=\ker(\partial_k)\big/\mathrm{im}(\partial_{k+1}), \qquad \beta_k(\varepsilon)=\mathrm{rank}\,H_k\bigl(C(\varepsilon)\bigr),$$8 (Xia et al., 2015). Euler Characteristic Surfaces were introduced in part because they are substantially computationally cheaper than available alternate methods such as persistent homology (Roy et al., 2024). SFTD localizes discrepancies but requires a doubled graph or extended cubical lattice, roughly doubling the problem size on large domains (Trofimov et al., 2024). The literature therefore repeatedly trades full generality for tractable summaries, sparse surrogates, or lower-dimensional projections.

6. Recurrent misconceptions and conceptual limits

A common misconception is that multiscale topological differentiation is synonymous with ordinary persistent homology. The literature does not support that identification. Persistent homology is central, but persistent Laplacians, evolutionary Hodge spectra, Mapper towers, Euler characteristic surfaces, F-Cross-Barcodes, and multiscale Gauss link integrals all instantiate the same broader pattern while departing from standard barcode-only analysis (Du et al., 2023, Chen et al., 2019, Fritze, 26 Sep 2025, Roy et al., 2024, Shen et al., 2023).

A second misconception is that “multiscale” always means a single filtration radius. In practice, the scale variables may be cover overlap, resolution, time, codensity, annulus radii, mark-weighted distance, or a pair $$H_k\bigl(C(\varepsilon)\bigr)=\ker(\partial_k)\big/\mathrm{im}(\partial_{k+1}), \qquad \beta_k(\varepsilon)=\mathrm{rank}\,H_k\bigl(C(\varepsilon)\bigr),$$9 indexing a bifiltration (Benjamin et al., 2022, Rohrscheidt et al., 2022, Schindler et al., 16 Oct 2025, Eckardt et al., 14 May 2026). Some methods are explicitly multi-parameter rather than single-parameter. This matters because the target phenomenon may itself be two-scale or nonhierarchical.

A third misconception is that “differentiation” implies a literal derivative. In some papers it does refer to temporal or scale derivatives of an Euler characteristic surface (Roy et al., 2024). In many others it means ranking by perturbation (Du et al., 2023), thresholding by physically motivated metrics (Chen et al., 2024), separating long-lived topological classes from noise (Fritze, 26 Sep 2025), measuring deviation from a Euclidean local model (Rohrscheidt et al., 2022), or testing against a random-labelling null via envelopes (Eckardt et al., 14 May 2026). The word is therefore operational rather than uniquely mathematical.

Finally, the literature repeatedly notes parameter sensitivity and representational limits. Mapper requires parameter choice, and the 2-Mapper work develops tools guided by persistent $$X_{r,t}=\mathrm{Rips}\bigl(\{p\in P:\rho_k(p)\le t\},r\bigr)$$0 information precisely because this is a common issue (Fritze, 26 Sep 2025). Euclidicity depends on a correct choice of local dimension $$X_{r,t}=\mathrm{Rips}\bigl(\{p\in P:\rho_k(p)\le t\},r\bigr)$$1 (Rohrscheidt et al., 2022). Very large $$X_{r,t}=\mathrm{Rips}\bigl(\{p\in P:\rho_k(p)\le t\},r\bigr)$$2 in multiresolution PH over-smooths the data and can eventually kill all topology (Xia et al., 2015). In PWDS, the death simplex may not be uniquely determined if multiple simplices die simultaneously (Torras-Pérez et al., 5 May 2025). These points do not undermine the framework; they delimit its regime of reliability and indicate that multiscale topological differentiation is best understood as an inferential toolkit whose outputs are strongest when paired with domain validation, model checking, and scale-aware interpretation.