Persistent Homology Observables

• Persistent homology observables are multiscale summaries that capture how topological features emerge, persist, and vanish through filtrations like barcodes and persistence diagrams.
• They derive key quantities such as Betti numbers, persistence lifetimes, entropy measures, and landscapes, providing a quantitative encoding of shape across scales.
• These observables enable effective noise filtering, feature extraction, and statistical learning in diverse applications from physics and biology to high-dimensional data analysis.

Persistent homology observables are multiscale summaries extracted from a filtration of spaces, simplicial complexes, or other categorical objects, and are used to record how topological features appear, persist, and disappear as a scale parameter varies. In the standard setting, the basic observables are Betti numbers, persistence intervals or barcodes, persistence diagrams, birth and death times, and lifetimes $\ell_i=d_i-b_i$; in more specialized settings they also include rank-type invariants, persistence landscapes, persistent Betti numbers, entropy-like functionals, and application-specific statistics built from merge trees, filtrations on graphs, hypergraphs, wavefunction profiles, or Fock-space landscapes (Kemme et al., 10 May 2025). Across the literature, these observables are treated not merely as counts of holes, but as quantitative encodings of shape across scales, with the choice of filtration determining which geometric, combinatorial, or physical structures become visible (He et al., 2022).

1. Foundational observables and the filtration framework

The standard mathematical setup begins with a filtration

$K_0 \subseteq K_1 \subseteq \cdots \subseteq K_n,$

where each $K_i$ is a simplicial complex at scale $i$. For each dimension $k$, one has a chain group $C_k(K)$, boundary operators

$$\partial_k : C_k(K) \to C_{k-1}(K), \qquad \partial_{k-1}\circ \partial_k = 0,$$

and homology groups

$$H_k(K)=\ker(\partial_k)/\operatorname{im}(\partial_{k+1}), \qquad \beta_k=\operatorname{rank}(H_k).$$

Persistent homology tracks the induced maps

$H_k(K_i)\to H_k(K_j), \qquad i\le j,$

so that a class is represented by an interval $[b_i,d_i)$ or by the point $K_0 \subseteq K_1 \subseteq \cdots \subseteq K_n,$0 in a persistence diagram (Kemme et al., 10 May 2025).

These observables have a standard geometric interpretation. $K_0 \subseteq K_1 \subseteq \cdots \subseteq K_n,$1 counts connected components, $K_0 \subseteq K_1 \subseteq \cdots \subseteq K_n,$2 counts one-dimensional cycles or loops, and $K_0 \subseteq K_1 \subseteq \cdots \subseteq K_n,$3 counts voids or cavities. Features that persist longer lie farther from the diagonal $K_0 \subseteq K_1 \subseteq \cdots \subseteq K_n,$4 in the persistence diagram, and long persistence is interpreted as topologically robust, while short persistence is often treated as noise or fine-scale irregularity (Kemme et al., 10 May 2025). A major refinement of this interpretation is that short intervals need not be mere noise: persistent homology can detect the curvature of disks from which points have been sampled, and in that setting the birth and death scales of short-lived cycles depend on the ambient geometry (Bubenik et al., 2019).

The same foundational observables reappear in specialized constructions. In quantum persistent homology, the central quantity is the persistent homology group

$K_0 \subseteq K_1 \subseteq \cdots \subseteq K_n,$5

with persistent Betti number

$K_0 \subseteq K_1 \subseteq \cdots \subseteq K_n,$6

which counts the $K_0 \subseteq K_1 \subseteq \cdots \subseteq K_n,$7-dimensional topological features that are born by scale $K_0 \subseteq K_1 \subseteq \cdots \subseteq K_n,$8 and still present at scale $K_0 \subseteq K_1 \subseteq \cdots \subseteq K_n,$9 (Ameneyro et al., 2022). In directed persistent homology, the observables are modified to detect directed cycles in odd dimensions, while in persistent intersection homology the same outputs—Betti numbers, persistent homology groups, persistence diagrams, and barcodes—are computed on a restricted chain complex adapted to stratified spaces with singularities (Méndez et al., 2020); (Rieck et al., 2019).

2. Derived summaries from barcodes and diagrams

A large part of the literature treats barcodes and persistence diagrams as intermediate objects from which more compact observables are derived. The most basic derived quantity is the persistence lifetime

$K_i$0

In one influential usage, persistent entropy is defined as the Shannon entropy of normalized lifetimes: $K_i$1 This observable measures the non-uniformity of topological feature lifetimes: if one feature dominates, $K_i$2 is low, and if lifetimes are broadly and evenly distributed, $K_i$3 is higher (He et al., 2022).

The same work defines a family of lifetime norms

$K_i$4

with $K_i$5 the sum of lifetimes, $K_i$6 the root mean squared lifetime, and $K_i$7 the lifetime of the longest-lived feature. For wavefunction intensity profiles, $K_i$8 and persistent entropy are constructed from the barcode or persistence diagram obtained by a sublevel-set filtration of the one-dimensional probability density, and they summarize the number of prominent peaks, relative prominence of peaks, hierarchy of local maxima and minima, and multiscale organization of the density (He et al., 2022).

Another major family of summaries is built from persistence landscapes. For a filtered complex, the persistent Betti number is

$K_i$9

and the persistence landscape is

$i$0

The average persistence landscape is then the expectation $i$1, viewed as an element of the Hilbert space $i$2. This turns barcodes into a continuous observable suitable for inverse problems and statistical learning (Bubenik et al., 2019).

Several papers define summary statistics directly on persistence diagrams. In nuclear-collision analysis, for a persistence diagram in homological dimension $i$3,

$i$4

is used to define a homological fractal dimension. Betti curves

$i$5

are treated as cluster distribution functions across scales, and cluster entropy is defined as

$i$6

the Shannon entropy of the cluster-size distribution (Hamilton et al., 2022). In many-body localization, persistent-homology-based observables include arithmetic mean, geometric mean, standard deviation of birth, death, and lifetime distributions; $i$7-norms of lifetime vectors; persistent entropy from normalized lifetimes; the connectivity threshold

$i$8

and the maximum Betti number $i$9 (Hamilton et al., 2023).

3. Observables beyond the classical one-parameter barcode

In one-parameter persistence, the barcode is a complete description because finitely generated modules over a principal ideal domain decompose into interval-like pieces. Several cited works emphasize that this picture breaks down in multiparameter or more general indexed settings. For multiparameter persistent homology, a persistence module is an $k$0-graded module over

$k$1

and there is generally no decomposition into intervals (Harrington et al., 2017).

The proposed observables in this setting are algebraic rather than barcode-like. The multigraded Hilbert function

$k$2

records the dimension at each multidegree, and the Hilbert series

$k$3

provides a compact encoding of graded dimensions. Associated primes $k$4 stratify the support of the module into coordinate directions of persistence, while local cohomology $k$5 measures the size of the pieces supported on a chosen stratum (Harrington et al., 2017). These observables generalize the free/torsion dichotomy of the one-parameter case into fully persistent, partially persistent, and transient behavior.

A related development appears in the theory of change action derivatives in persistent homology. There, the classical pair group

$k$6

is interpreted as the group of $k$7-cycles whose lifespan is exactly the interval $k$8, and its rank counts how many barcode intervals of dimension $k$9 equal $C_k(K)$0. The paper generalizes this counting philosophy to tame filtrations indexed by finite posets, using a homological lifespan functor

$C_k(K)$1

so that the observable becomes a kind of finite-difference derivative of a rank-type memory functor (Salja, 24 Nov 2025). This gives a functorial observable for “new homology born here and killed here” in settings where no complete decomposition theorem exists.

More general indexing categories produce further observables. Persistent homology over directed acyclic graphs defines, for a connected subgraph $C_k(K)$2, the $C_k(K)$3-persistent homology group

$C_k(K)$4

where persistence is the image of a canonical map from the limit to the colimit of the diagram. In the single-source single-sink case this reduces to the image

$C_k(K)$5

recovering ordinary persistence, while in lattice-indexed cases it recovers the rank invariant (Chambers et al., 2014). Persistent homology of partially ordered spaces similarly treats natural homology as a persistence object indexed by the trace poset, reconstructed as a colimit of one-dimensional persistent homologies along traces (Calk et al., 2023).

4. Structural observables beyond counts of classes

Several works argue that persistent homology contains combinatorial or geometric information not captured by Betti numbers or interval multiplicities alone. One such refinement is the cophenetic matroid. For a filtered simplicial complex $C_k(K)$6, the cophenetic rank function is

$C_k(K)$7

for finite $C_k(K)$8. This defines a matroid, and the resulting filtered matroids, irreducible sets, ramification trees, rooted forests, and cophenetic ultrametric record how collections of homology classes become linearly dependent as scale changes (Güzel et al., 2022).

The same paper defines a cophenetic distance on homology classes: $C_k(K)$9 which measures the first additional scale at which two classes become linearly dependent. This is an observable of genealogical dependence rather than mere lifespan (Güzel et al., 2022). A plausible implication is that two filtrations with the same barcode can still differ at the level of dependency evolution.

Other generalized observables replace homology classes by arbitrary categorical features. Steady persistence and ranging persistence begin with a feature $$\partial_k : C_k(K) \to C_{k-1}(K), \qquad \partial_{k-1}\circ \partial_k = 0,$$0 and define counting functions

$$\partial_k : C_k(K) \to C_{k-1}(K), \qquad \partial_{k-1}\circ \partial_k = 0,$$1

where the first counts features present continuously throughout the interval and the second counts features appearing before the interval and reappearing after it (Gazull, 9 Jun 2025). The paper proves the equivalence

$$\partial_k : C_k(K) \to C_{k-1}(K), \qquad \partial_{k-1}\circ \partial_k = 0,$$2

so stability is tied to a structural property of the feature itself (Gazull, 9 Jun 2025).

Directed persistent homology provides another structural modification. For a directed simplicial complex, directed $$\partial_k : C_k(K) \to C_{k-1}(K), \qquad \partial_{k-1}\circ \partial_k = 0,$$3-cycles are cycles with only non-negative coefficients, and the directed homology module is the submodule generated by such cycles. The resulting directed barcode is a refinement of the undirected barcode: each directed bar is contained in a matched undirected bar, with the same death time and possibly later birth, while some undirected bars remain unmatched (Méndez et al., 2020). In even degrees above $$\partial_k : C_k(K) \to C_{k-1}(K), \qquad \partial_{k-1}\circ \partial_k = 0,$$4, directed homology vanishes, whereas in dimension $$\partial_k : C_k(K) \to C_{k-1}(K), \qquad \partial_{k-1}\circ \partial_k = 0,$$5 it detects a genuinely directed polygon only when the edges are consistently oriented around the cycle (Méndez et al., 2020).

5. Domain-specific observables in physics, biology, and networked data

A notable feature of the recent literature is the construction of persistent homology observables tailored to specific scientific domains. In the generalized Aubry-André-Harper model, the observables are built from the topology of the intensity profile $$\partial_k : C_k(K) \to C_{k-1}(K), \qquad \partial_{k-1}\circ \partial_k = 0,$$6 or $$\partial_k : C_k(K) \to C_{k-1}(K), \qquad \partial_{k-1}\circ \partial_k = 0,$$7 using a sublevel-set filtration. Persistent entropy and the root mean squared lifetime behave similarly to Shannon entropy and inverse participation ratio, distinguish localized, extended, and critical phases, and can be applied to both eigenstates and wavepacket propagation dynamics (He et al., 2022). The same study emphasizes an additional capability: persistent entropy shows a pronounced dip on the ordered line

$$\partial_k : C_k(K) \to C_{k-1}(K), \qquad \partial_{k-1}\circ \partial_k = 0,$$8

thereby distinguishing ordered from disordered regimes of the model (He et al., 2022).

In many-body localization, the $$\partial_k : C_k(K) \to C_{k-1}(K), \qquad \partial_{k-1}\circ \partial_k = 0,$$9 localization landscape

$$H_k(K)=\ker(\partial_k)/\operatorname{im}(\partial_{k+1}), \qquad \beta_k=\operatorname{rank}(H_k).$$0

is treated as a scalar field on the Fock-space graph, and persistent homology is applied to the superlevel sets

$$H_k(K)=\ker(\partial_k)/\operatorname{im}(\partial_{k+1}), \qquad \beta_k=\operatorname{rank}(H_k).$$1

The resulting observables—$$H_k(K)=\ker(\partial_k)/\operatorname{im}(\partial_{k+1}), \qquad \beta_k=\operatorname{rank}(H_k).$$2, $$H_k(K)=\ker(\partial_k)/\operatorname{im}(\partial_{k+1}), \qquad \beta_k=\operatorname{rank}(H_k).$$3, persistent entropy, $$H_k(K)=\ker(\partial_k)/\operatorname{im}(\partial_{k+1}), \qquad \beta_k=\operatorname{rank}(H_k).$$4, and $$H_k(K)=\ker(\partial_k)/\operatorname{im}(\partial_{k+1}), \qquad \beta_k=\operatorname{rank}(H_k).$$5—are interpreted as morphological diagnostics of fragmentation, clustering, and connectivity in Fock space near the many-body localization transition (Hamilton et al., 2023).

In nuclear collisions, each event is treated as a point cloud in $$H_k(K)=\ker(\partial_k)/\operatorname{im}(\partial_{k+1}), \qquad \beta_k=\operatorname{rank}(H_k).$$6, and a Delaunay triangulation with Delaunay Triangulation Field Estimation is filtered by an inverse-density-like quantity

$$H_k(K)=\ker(\partial_k)/\operatorname{im}(\partial_{k+1}), \qquad \beta_k=\operatorname{rank}(H_k).$$7

The resulting observables include fractal dimension from persistence lifetimes, Betti curves, cluster entropy, local clustering statistics via dendrogram leaf $$H_k(K)=\ker(\partial_k)/\operatorname{im}(\partial_{k+1}), \qquad \beta_k=\operatorname{rank}(H_k).$$8-norms, and a cophenetic distance correlation function. These are used to extract clustering signatures and elliptic-anisotropy-like behavior from final-state hadron distributions (Hamilton et al., 2022).

In high-dimensional data analysis, the key claim is that persistent homology itself need not change, but the observable used to build the filtration should change. Rather than using raw Euclidean distance, the paper recommends spectral distances on the symmetric $$H_k(K)=\ker(\partial_k)/\operatorname{im}(\partial_{k+1}), \qquad \beta_k=\operatorname{rank}(H_k).$$9-nearest-neighbor graph, especially diffusion distance and corrected effective resistance: $H_k(K_i)\to H_k(K_j), \qquad i\le j,$0 together with a spectral embedding formula for effective resistance (Damrich et al., 2023). This suggests that in noisy high-dimensional settings the “observable” is often the metric feeding the Vietoris–Rips filtration, not only the barcode extracted afterward.

Biological and network examples illustrate the breadth of the term. A pedagogical introduction uses Betti numbers, barcodes, persistence diagrams, and lifetimes to analyze a 3-1 supercoiled DNA structure (Kemme et al., 10 May 2025). Hypergraph classification defines three filtration observables—Simplicial Complex Closure, Restricted Barycentric Subdivision, and Relative Barycentric Subdivision—and then converts 0-dimensional and 1-dimensional barcodes into fixed-length numerical features via the number of bars and four algebraic summary statistics (Aktas et al., 2023). Persistent intersection homology adapts barcodes and diagrams to singular or mixed-dimensional data, such as a circle with a whisker, wedges, or pinched spaces (Rieck et al., 2019).

6. Interpretation, stability, and recurrent misconceptions

A recurrent interpretive theme is that persistent homology observables summarize shape across scales rather than topology at a single scale. The filtration lets one see when clusters merge, when loops appear, and when cavities are filled (Kemme et al., 10 May 2025). In application-driven work, this often translates into observables of clustering, ordering, morphology, or transport rather than purely topological classification.

One common misconception is that long bars are signal and short bars are noise. The curvature-detection work explicitly disputes this thesis, showing that short intervals encode geometric information and can recover Gaussian curvature $H_k(K_i)\to H_k(K_j), \qquad i\le j,$1 from random point samples in disks of constant curvature (Bubenik et al., 2019). A second misconception is that standard scalar summaries are always sufficient. In the generalized Aubry-André-Harper model, inverse participation ratio and Shannon entropy are permutation-invariant over sites and cannot distinguish one broad peak from several narrower peaks if the occupation statistics are similar, whereas persistent homology tracks local maxima, minima, and their lifetimes (He et al., 2022). In many-body localization, persistent homology is presented as revealing the cluster structure, extrema, and connectivity of eigenstate-support landscapes, which are unobtainable by traditional means (Hamilton et al., 2023).

Stability is another central criterion for calling a quantity an observable. Average persistence landscapes define a continuous map from metric measure spaces to a Hilbert space, continuous with respect to the Gromov–Wasserstein metric (Bubenik et al., 2019). Directed persistence diagrams satisfy bottleneck-distance bounds with respect to the correspondence distortion distance $H_k(K_i)\to H_k(K_j), \qquad i\le j,$2 (Méndez et al., 2020). Steady and ranging persistence formulate stability through interleavings and balancedness, proving that for tame filtrations and categories satisfying the triangle condition, balancedness and stability coincide (Gazull, 9 Jun 2025). These results do not identify a single universal observable; rather, they show that different summaries can be stable when matched to appropriate categories and filtrations.

Computationally, observables are constrained by the structures from which they are extracted. Relative persistent homology for a pair $H_k(K_i)\to H_k(K_j), \qquad i\le j,$3 can be computed via a relative Delaunay Čech complex, avoiding the full relative Čech filtration in low-dimensional Euclidean space (Blaser et al., 2019). Quantum persistent homology uses a persistent Dirac operator and the persistent combinatorial Laplacian so that persistent Betti numbers become quantum-measurable observables (Ameneyro et al., 2022). In contrast, directed persistence leads to polyhedral computations of non-negative cycles, and multiparameter persistence lacks a complete barcode-like classification (Méndez et al., 2020); (Harrington et al., 2017).

Taken together, these developments show that “persistent homology observables” is best understood as a layered term. At its core are Betti numbers, persistence intervals, diagrams, birth and death times, and lifetimes. Around this core are derived summaries such as entropy, landscapes, rank-type quantities, matroidal or categorical refinements, and domain-specific statistics constructed from merge trees, cycle dependencies, or graph-based filtrations. This suggests that the observable is not fixed once and for all: it is the component of persistent homology that is chosen to retain the multiscale structure relevant to a given problem.