Does a distance of zero always mean identical modules?

Not in the most general setting; additional restrictions matter.

Is epsilon necessarily a time delay?

No. It is a shift in the chosen scale parameter or parameters.

Does a small distance guarantee identical model predictions?

No. A downstream decision rule needs its own analysis or evaluation.

AI Glossary / Data science and analytics / Advanced

Interleaving distance

Definition

Interleaving distance measures how closely two persistence modules can be related after allowing a shift in their scale parameters. Persistence modules describe information, such as topological features, across a changing threshold together with the maps connecting those stages. The distance is the infimum of the nonnegative shifts for which compatible maps in both directions form an interleaving. It is a specialized concept in topological data analysis, not an ordinary distance between rows of a spreadsheet.

What it means

A topological summary can change as an analyst varies a distance threshold, density level, or another parameter. Comparing only one chosen threshold may miss how features develop across scales. An interleaving compares the organized sequence of information and its connections. A small value means the modules can approximate each other using a small allowed shift under the chosen convention. It does not mean the original datasets are identical or that every business-relevant property is similar.

How it works

For one real-valued parameter, an epsilon-interleaving maps the information in module M at scale t into module N at scale t plus epsilon, and also maps N back into M with the same shift. The maps must respect movement along the scale. Going across and back must agree with the module's own map from t to t plus twice epsilon. These consistency conditions distinguish interleaving from arbitrary matching of snapshots. The distance takes the infimum over admissible epsilon values. More general settings use specified parameter shifts, so the definition and units must be stated.

Three practical examples

Topological-method research: investigators compare a detailed persistence construction with an approximation and seek a bound on how much information changes across scales.
Data-pipeline assessment: a team studies whether a perturbation of the input produces a controlled change in its chosen persistent-homology representation.
Multiparameter analysis: researchers compare modules tracking more than one scale, while recognizing that exact computation and interpretation can be substantially harder.

Think of it this way

Think of two descriptions of a process that can be compared with a small allowance on their scale readings. Each description must translate consistently into the other, including after moving forward. The analogy is limited: the parameter need not be time, and the mathematical maps need not pair individual objects or make the descriptions literally identical.

Worked example

Consider a simple one-parameter interval module representing one feature present from scale 2 up to, but not including, scale 8. Inside that interval its information is one-dimensional and the internal maps are identities; outside it, the module is zero. Compare it with the zero module, which has no such feature at any scale.
Any maps to or from the zero module are zero. For an epsilon-interleaving to work, the interval module's own map across a shift of twice epsilon must therefore also be zero everywhere. The interval has length 6. If epsilon is less than 3, twice epsilon is less than 6, and there are scales t where both t and t plus twice epsilon lie inside the interval. The internal map there is an identity, so the required condition fails. At epsilon equal to 3, a shift of 6 cannot leave both endpoints inside the half-open interval. All those shifted maps are zero, and the condition holds. The interleaving distance to zero is consequently 3. This toy calculation explains why short-lived features can be close to absence under this distance.

Illustrative case study

In a fictional shape-analysis project, a team uses persistent homology to summarize scanned parts at different thresholds. It wants to replace a detailed computation with a faster approximation. A researcher derives an interleaving bound for a specified representation and parameter convention. The team records the bound separately from the downstream classifier's accuracy and tests the latter independently. A small topological discrepancy may support the approximation analysis, but it does not by itself certify that defect decisions remain unchanged. The case illustrates a possible evaluation workflow, not a reported industrial result.

Common misconceptions and mistakes

Interleaving distance is generally an extended pseudometric: infinity is possible, and in broad module classes distinct nonisomorphic modules can have distance zero. Stronger metric statements require appropriate restrictions or identifications. It is not simply the difference between feature counts at one threshold. In suitable one-parameter settings it agrees with bottleneck distance between persistence summaries, but this relationship should not be transferred without qualification to arbitrary multiparameter modules. An available upper bound is also not necessarily the exact distance.

Business use

This term is most relevant when a project already uses persistent homology or related topological methods. Ask which modules are compared, what the parameter measures, and whether the reported number is exact, approximate, or a bound. A rescaling of a parameter can change its numerical interpretation. For a business decision, connect mathematical stability to the actual task through separate validation. Do not assume that a theoretically small distance means negligible operational risk, or that computing it is practical for every representation.

Check your understanding

Why does the interval from 2 to 8 have interleaving distance 3 from zero in the example?

Answer: Interleaving requires the module's maps across twice the shift to vanish. A shift of 6 crosses the entire half-open interval, while any smaller shift leaves some nonzero internal map.

Related terms

Persistence module
persistent homology
filtration
bottleneck distance
barcode
stability theorem
multiparameter persistence
pseudometric

Learn next

Learn how a filtration produces persistent homology and how an interval represents a feature's lifespan. Then trace the two shifted maps in the definition of an interleaving. Study the assumptions of the one-parameter isometry theorem before comparing bottleneck and interleaving distances in more general settings.

Sources

Michael Lesnick, original interleaving-distance theory, section 2.2: https://arxiv.org/pdf/1106.5305 Bubenik and colleagues, topological spaces of persistence modules and extended pseudometric properties: https://people.clas.ufl.edu/peterbubenik/files/s41468-018-0022-4.pdf Håvard Bakke Bjerkevik, original research, definition 3.10 and one-parameter comparison: https://link.springer.com/article/10.1007/s10208-025-09695-w

Previous: Interleaving Next: Interlingua