{
  "id": "topological-data-analysis",
  "title": "Topological Data Analysis",
  "lang": "en",
  "source_lang": "en",
  "summary": "Topological data analysis uses multiscale mathematical descriptions of connectivity and holes to investigate the structure of data.",
  "category": "mathematics",
  "subcategory": "geometry",
  "type": "concept",
  "aliases": [
    "TDA",
    "topological methods in data analysis"
  ],
  "generated_at": "2026-11-04",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/topological-data-analysis/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/topological-data-analysis/",
      "json": "https://aiwiki.page/zh/topological-data-analysis.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "Topological data analysis (TDA) is a collection of mathematical and computational methods for studying the shape of data. Drawing on topology, geometry, and statistics, it describes features such as connected components, loops, cavities, and branching structures. Its inputs include point clouds, images, networks, sampled functions, and time-dependent measurements. Rather than relying exclusively on coordinates or a single choice of scale, TDA often follows how structural features appear and disappear as a resolution parameter changes.\n\nThe best-known method is persistent homology, which records the lifetimes of topological features across a sequence of spaces. Another widely used construction, Mapper, produces a network summarizing the organization of a data set relative to selected measurement functions. These approaches can reveal structure that is not adequately represented by averages, linear projections, or a single clustering.\n\nTDA is not a universal method for recovering an intrinsic “true shape.” Its conclusions depend on how observations are represented, which distances or functions are used, and how sampling and noise are modeled. Its distinctive contribution is a framework for separating geometric construction, topological description, computational approximation, and statistical interpretation."
    },
    {
      "heading": "From observations to geometric spaces",
      "text": "A finite collection of observations has little useful topology if each observation is treated simply as an isolated point. TDA therefore begins by constructing a space that expresses relationships among observations. For a point cloud \\(X=\\{x_1,\\ldots,x_n\\}\\), these relationships commonly come from a distance \\(d\\), making \\(X\\) a metric space. Measurements might lie in Euclidean space, but they can also be compared through distances between sequences, graphs, shapes, or probability distributions. The choice of distance is substantive: standardizing variables, changing feature weights, or replacing Euclidean distance with a domain-specific metric can change the topology subsequently detected.\n\nA basic construction replaces each point by a ball of radius \\(r\\) and studies their union,\n\n\\[U_r(X)=\\bigcup_{i=1}^{n}B(x_i,r).\\]\n\nAt small radii, the union consists of many separate pieces. As the radius increases, pieces join, loops may form, and holes eventually fill. A finite sample distributed around a circle, for example, initially produces isolated components, then a connected ring, and finally a filled region. The parameter \\(r\\) expresses a scale of observation rather than an additional measured variable.\n\nComputations usually replace a continuous union of balls by a simplicial complex, assembled from vertices, edges, triangles, tetrahedra, and their higher-dimensional analogues. In a Čech complex at radius \\(r\\), a collection of vertices forms a simplex when their corresponding radius-\\(r\\) balls have a common intersection. The nerve theorem explains why this construction is useful: under suitable conditions on the covering sets and their intersections, the complex and the union have the same homotopy type. Convex balls in Euclidean space provide an important setting in which these conditions hold.\n\nThe Vietoris–Rips complex is easier to construct from pairwise distances. Under an edge-threshold convention, a set of vertices forms a simplex at parameter \\(t\\) whenever every pair has distance at most \\(t\\). Pairwise intersection does not generally imply common intersection, so Rips and Čech complexes are not identical. Their scale conventions must also be distinguished: radius-\\(r\\) balls overlap when their centers are at most \\(2r\\) apart. Comparisons between software packages and mathematical results can otherwise introduce a factor-of-two discrepancy.\n\nOther representations suit other data types. Alpha complexes use the geometry of Voronoi cells and Delaunay triangulations to provide relatively economical models of Euclidean unions of balls. Cubical complexes are natural for pixel and voxel arrays. A grayscale image can be analyzed by adding pixels in order of intensity, while a terrain can be analyzed through increasing elevation thresholds. Thus, the geometric object studied by TDA need not originate from a point cloud or an explicit distance matrix."
    },
    {
      "heading": "Homology and the meaning of topological features",
      "text": "The principal algebraic descriptor in TDA is homology. Informally, homology identifies cycles that do not bound higher-dimensional pieces within a space. Degree zero records connected components; degree one records independent loops; degree two records enclosed cavities. A circle has one connected component and one independent one-dimensional hole. A sphere's surface has one connected component and one two-dimensional cavity, but no one-dimensional hole. The solid ball bounded by that sphere has no corresponding two-dimensional homology class because its interior fills the cavity.\n\nThe computation begins with chain groups. For a simplicial complex \\(K\\), the vector space \\(C_k(K;\\mathbb F)\\) consists of formal combinations of its \\(k\\)-dimensional simplices, with coefficients in a chosen field \\(\\mathbb F\\). The boundary operator\n\n\\[\\partial_k:C_k(K;\\mathbb F)\\longrightarrow C_{k-1}(K;\\mathbb F)\\]\n\nmaps a simplex to a signed combination of its faces. Its defining identity is\n\n\\[\\partial_{k-1}\\partial_k=0,\\]\n\nwhich states that every boundary has zero boundary. A triangle's three edges, for example, form a cycle because their endpoint contributions cancel.\n\nThe cycles are the elements of \\(\\ker\\partial_k\\), and boundaries are the elements of \\(\\operatorname{im}\\partial_{k+1}\\). Homology is the quotient vector space\n\n\\[H_k(K;\\mathbb F)= \\ker\\partial_k/\\operatorname{im}\\partial_{k+1}.\\]\n\nThis quotient identifies cycles that differ by a boundary. The boundary of a filled triangle therefore represents zero in first homology, whereas a triangular loop with no filling triangle represents a nonzero class. The Betti number \\(\\beta_k=\\dim H_k(K;\\mathbb F)\\) counts independent classes in degree \\(k\\).\n\nCoefficients matter. Computation often uses the two-element field \\(\\mathbb F_2\\), where signs disappear and addition becomes exclusive-or. This simplifies implementations but can hide distinctions visible with other coefficients. For example, the real projective plane has nonzero first homology over \\(\\mathbb F_2\\) but zero first homology over the rational numbers. Ordinary Betti numbers over a field also do not retain all the torsion information available in integer homology. The coefficient field is consequently part of the specification of an analysis, not merely an implementation detail.\n\nHomology captures selected structural properties rather than a complete description of a space. Different spaces can have identical Betti numbers, and identical homology need not imply homeomorphism or even homotopy equivalence. A homology class also usually has many cycle representatives. A computed loop may therefore be algebraically valid without being the shortest, smoothest, or most interpretable representative. Extracting a representative that corresponds closely to a scientific mechanism is an additional geometric or optimization problem."
    },
    {
      "heading": "Filtrations, persistence modules, and barcodes",
      "text": "Persistent homology studies a filtration: a nested family of spaces\n\n\\[K_s\\subseteq K_t\\qquad\\text{whenever }s\\leq t.\\]\n\nIncreasing the distance threshold in a Rips complex gives one example. Another is the sublevel-set filtration of a function \\(f\\),\n\n\\[K_t=f^{-1}((-\\infty,t]).\\]\n\nFor a terrain, connected components begin near local minima and merge as the water level rises. For an image, the same construction tracks structures as an intensity threshold changes. A superlevel-set filtration reverses which values enter first.\n\nApplying homology to every space produces vector spaces \\(H_k(K_t;\\mathbb F)\\). Inclusions between spaces induce linear maps between their homology groups. This entire system is a persistence module. Retaining these maps is essential: a list of Betti numbers at separate scales does not specify which feature at one scale continues to which feature at another. A loop may disappear while a different loop appears, leaving the Betti number unchanged despite a structural transition.\n\nFor finite filtrations over a field, the module decomposes into interval modules. Each interval \\([b,d)\\) represents a class born at parameter \\(b\\) that remains present until parameter \\(d\\). These intervals form a persistence barcode. Equivalently, finite intervals can be displayed as points \\((b,d)\\) in a persistence diagram. The lifetime is \\(d-b\\); points near the diagonal have short lifetimes. Classes that remain through the filtration may have infinite death time, depending on the construction and the reporting convention.\n\nIn degree zero, this description closely resembles hierarchical clustering. Vertices begin as separate components, and edges merge them. A common bookkeeping convention retains the older component when two components merge and records the death of the younger one. For a distance filtration, the resulting merger information is closely connected to single-linkage clustering and minimum spanning trees. Higher-dimensional persistence adds information about loops and cavities that cannot be reconstructed from the component hierarchy alone.\n\nThe order in which simplices enter must respect their faces: an edge cannot precede its endpoints, and a triangle cannot precede its boundary edges. Multiple simplices may have the same filtration value, requiring a compatible ordering during computation. Such choices can affect zero-length intervals and selected representatives without changing the positive-length persistence information. Ordinary and reduced homology also differ in their treatment of the final connected component.\n\nLong intervals often indicate structures that survive a broad range of scales, but lifetime is not a universal measure of scientific importance. A short interval can encode a meaningful small feature, while a long interval can arise from an outlier, a sampling gap, or an unsuitable metric. Persistence provides a structured description whose interpretation depends on the data-generating process."
    },
    {
      "heading": "Stability, approximation, and reconstruction",
      "text": "A central mathematical justification for persistent homology is that small perturbations can produce only bounded changes in suitable persistence diagrams. The standard comparison is the bottleneck distance. It matches points in two diagrams, allowing unmatched points to be paired with the diagonal, and minimizes the largest matching cost measured in the maximum norm. Matching a point \\((b,d)\\) to the diagonal costs \\((d-b)/2\\), which makes short-lived features inexpensive to remove.\n\nFor suitably tame real-valued functions \\(f\\) and \\(g\\) on the same space, a fundamental stability theorem states\n\n\\[d_B\\bigl(\\operatorname{Dgm}_k(f), \\operatorname{Dgm}_k(g)\\bigr) \\leq \\|f-g\\|_\\infty.\\]\n\nThus, a uniformly bounded change in function values gives an equally bounded change in the persistence diagram. The hypotheses ensure that the homological information admits an appropriate diagram. This result concerns a precisely defined perturbation model; it does not imply robustness to arbitrary changes in preprocessing or sampling.\n\nA related geometric result uses the Hausdorff distance. For compact subsets \\(X\\) and \\(Y\\) of an ambient metric space, the corresponding distance-to-set functions satisfy\n\n\\[\\|d(\\,\\cdot\\,,X)-d(\\,\\cdot\\,,Y)\\|_\\infty \\leq d_H(X,Y).\\]\n\nCombined with stability, this controls the persistence of their offset filtrations. For Rips filtrations on abstract metric spaces, analogous bounds use Gromov–Hausdorff distance, with constants depending on parameter conventions. These statements explain why sufficiently accurate geometric approximations can preserve multiscale topological information.\n\nStability is distinct from reconstruction. Stability compares two mathematical inputs; reconstruction asks whether observed data recover properties of an unobserved object. A sample from a smooth manifold may support correct recovery when it is sufficiently dense and the noise is sufficiently small relative to geometric feature sizes. One relevant quantity is reach, the distance within which points have unique nearest projections onto the underlying set. Features that are thin, sharply curved, or closely spaced require finer sampling than broad, well-separated features.\n\nThe distinction is particularly important for outliers. A single observation far from a compact cloud can greatly increase Hausdorff distance and create a long-lived component. Robust alternatives therefore replace ordinary distance-to-set functions with density-sensitive constructions. The distance-to-a-measure function, for example, aggregates distances needed to capture a specified amount of probability mass, reducing the influence of isolated observations. Its robustness depends on both the chosen mass parameter and the metric used to compare distributions.\n\nStability also has limits as a scientific guarantee. Rescaling one coordinate can substantially alter pairwise distances without constituting a small perturbation under the relevant theorem. Adding observations from a previously unsampled region can legitimately change the inferred topology. Moreover, a summary can be stable while discarding information needed for a particular task. The appropriate interpretation is that persistence offers controlled continuity under specified transformations, not invariance under every plausible alteration of the data."
    },
    {
      "heading": "Statistical inference and topological summaries",
      "text": "Statistical TDA treats observations as random and asks how an estimated topological summary relates to a population-level target. Possible targets include the support of a probability distribution, a density superlevel set, a regression function, or the distribution of persistence diagrams across experimental units. These are different inferential problems. The topology of a distribution's support, for example, can differ from the topology of its high-density regions, especially when low-density bridges join otherwise distinct clusters.\n\nSampling assumptions determine what can be estimated. Independent observations from a compact support with adequate probability mass in small neighborhoods can yield convergence of the empirical support in Hausdorff distance. Stability can then transfer a geometric estimation bound into a persistence bound. Sparse sampling, heavy-tailed distributions, measurement error, or dependent observations require different arguments. In particular, treating every point in a recorded trajectory as an independent replicate can substantially overstate the amount of information available.\n\nA confidence interval for a numerical parameter has a topological analogue in a confidence region around an estimated diagram. If a procedure supplies a valid bottleneck error bound \\(\\varepsilon\\), features farther than \\(\\varepsilon\\) from the diagonal cannot all be explained by matching them to the diagonal within that bound. This produces a principled threshold, but its interpretation remains conditional on the target, assumptions, and coverage statement. It does not automatically assign a probability of truth to each individual loop.\n\nResampling provides another route to uncertainty assessment. Bootstrap methods, subsampling, and permutation tests can evaluate variation in topological summaries or differences between groups. Their validity is not automatic because diagram-valued statistics can be nonsmooth, and some support-estimation problems do not behave like ordinary sample means. The resampling unit must also match the experimental design: subjects, specimens, or independent trajectories may be the appropriate units rather than individual measurements within them. Multiple dimensions, scales, and alternative preprocessing choices introduce multiple-testing concerns.\n\nFor statistical learning, diagrams are often transformed into vectors or functions. A persistence landscape replaces each interval by a triangular function and orders these functions pointwise. Persistence images transform and weight diagram points, smooth them, and integrate the result over a grid. Betti curves record the number of active classes at each parameter. These summaries enable averages, functional comparisons, kernels, and conventional classifiers, although each retains and discards different information.\n\nPredictive assessment requires the same separation of training and evaluation used elsewhere in machine learning. Choices of scaling, filtration, homology degree, image resolution, and model parameters can all influence performance. When these choices are learned from data, they belong within the training process and, where appropriate, within cross-validation. A favorable classification result does not by itself establish that the associated holes or branches are causal features of the underlying system."
    },
    {
      "heading": "Mapper and topological network representations",
      "text": "Mapper is a method for constructing a compact network representation of data relative to a selected function, commonly called a filter or lens. Let \\(f:X\\to Z\\) map observations into a low-dimensional space. Cover the image \\(f(X)\\) by overlapping sets \\(\\{U_i\\}\\), then cluster each pullback \\(f^{-1}(U_i)\\). Each resulting cluster becomes a vertex. Two vertices are joined when their clusters share at least one observation; higher-order intersections can define higher-dimensional simplices, although visualization usually emphasizes the graph.\n\nThe overlap is essential. Without overlapping cover elements, clusters from adjacent regions would not share observations, and the construction would lose the connections that tie the summary together. For a real-valued lens, a common cover consists of intervals with specified lengths and overlaps. The lens might be a measured quantity, a density estimate, an eccentricity function, or a coordinate obtained from principal component analysis. More than one lens coordinate can be used, but the cover becomes more expensive as its dimension increases.\n\nMapper is related to the Reeb graph, which identifies points lying in the same connected component of a function's level set. For example, the Reeb graph of the height function on a torus can split and rejoin as horizontal sections change their connectivity. Mapper replaces exact level sets and continuous connected components with overlapping bands and finite-data clustering. Under appropriate assumptions and controlled refinements, mathematical results relate Mapper constructions to Reeb-type objects. An arbitrary empirical Mapper graph, however, is not automatically a faithful reconstruction of a unique latent graph.\n\nIts output depends on several interacting choices: the lens, cover resolution, overlap, metric, and clustering algorithm. A branch can appear because a population genuinely separates, because local density changes, or because a clustering threshold fragments a region. Fine covers may reveal local detail but become sensitive to sparse sampling; coarse covers may merge distinct structures. Changing overlap can alter graph connectivity even when the lens remains fixed. Stability and multiscale extensions address parts of this sensitivity, but ordinary Mapper remains a parameter-dependent exploratory construction.\n\nMapper also differs from a conventional network in the meaning of its vertices and edges. A vertex represents a cluster within one cover element, not necessarily a single observation or a disjoint population. An observation can belong to several vertices, and an edge represents shared membership rather than correlation, physical contact, or causal influence. Vertex counts therefore cannot simply be added to obtain a sample size.\n\nThe method is useful for displaying branching trajectories, heterogeneous populations, and relationships among covariates without forcing all observations into a flat partition. Its visual clarity can nevertheless encourage excessive interpretation. A graph layout is an additional representation step: distances and angles in a force-directed drawing need not equal distances in the original data. Scientific interpretation consequently depends on examining memberships, annotations, parameter sensitivity, and independent evidence rather than the visual prominence of a branch alone."
    },
    {
      "heading": "Algorithms, complexity, and software",
      "text": "The classical persistent-homology algorithm represents boundary operators as a matrix whose rows and columns follow filtration order. Columns are reduced by adding earlier columns so that nonzero reduced columns have distinct lowest nonzero entries. A resulting pivot identifies a birth–death pairing. This procedure resembles elimination in linear algebra, but it preserves the ordering information needed to track persistence. With coefficients in \\(\\mathbb F_2\\), arithmetic can be implemented efficiently using bit operations or sparse representations.\n\nIf a filtered complex contains \\(N\\) simplices, straightforward matrix reduction has cubic worst-case running time and potentially quadratic storage. Practical performance can be much better because boundary matrices are initially sparse and many columns can be simplified or avoided. Clearing, compression, specialized ordering strategies, and persistent cohomology reduce costs in many applications. Persistent cohomology uses dual algebraic information and can recover the same interval data over a field, while supporting implementations with favorable memory behavior.\n\nThe number of simplices is often the more serious bottleneck. A Rips complex on \\(n\\) vertices can contain every subset of vertices as a simplex. Even when only dimensions up to a fixed degree are needed, the number of candidate simplices can grow rapidly. Computing homology through degree \\(k\\) generally requires information about \\((k+1)\\)-simplices, because these determine which \\(k\\)-cycles become boundaries. Restricting an analysis to loops therefore does not eliminate the need to account for triangles.\n\nSeveral approaches reduce this burden. Sparse Rips constructions approximate the full filtration with explicit approximation guarantees under appropriate conditions. Landmark and witness constructions use a subset of points to define a smaller complex, though their guarantees depend on the particular construction. Alpha complexes can be economical in low ambient dimension, but Delaunay-based methods become expensive in high dimensions. Cubical algorithms exploit the regular structure of images, and implicit representations avoid explicitly storing every simplex whenever possible.\n\nWidely used software includes GUDHI, Ripser, Dionysus, PHAT, and giotto-tda. Their scopes differ: some emphasize particular persistence algorithms, while others provide geometric constructions, statistical representations, or machine-learning interfaces. Ripser is especially associated with efficient Vietoris–Rips persistence, and GUDHI provides a broader collection of complexes and topological tools. The availability of an implementation does not remove the need to specify coefficient fields, filtration conventions, truncation thresholds, and whether reported homology is reduced or unreduced.\n\nReproducibility also requires attention to numerical details. Floating-point arithmetic can change the ordering of nearly equal filtration values, while sparse or approximate nearest-neighbor graphs can omit edges required by an intended construction. Computing on such a graph may be entirely appropriate, but it defines a different input unless an approximation theorem applies. Recording the metric, preprocessing, maximum homology degree, parameter range, software version, and randomization settings is therefore part of documenting the mathematical object that was actually analyzed."
    },
    {
      "heading": "Applications and domain-specific interpretation",
      "text": "In materials science, persistent homology describes multiscale arrangements of atoms, particles, pores, and phases. A point cloud of atomic positions can produce loops associated with ring-like arrangements and cavities associated with larger voids. Filtrations based on unequal atomic radii can incorporate chemical differences that an unweighted point cloud ignores. Studies of glasses and other disordered materials have used persistence diagrams to investigate medium-range organization that may be difficult to express through pair-correlation functions alone. A detected homology class, however, is not automatically a unique physical ring or pore.\n\nMolecular applications combine geometric descriptors with chemical information. Coordinates from protein structures can be partitioned by atom type or residue class before constructing filtrations. Such element-specific constructions distinguish interactions that ordinary geometry would treat alike. Topological summaries have been used as features in predictive models of molecular properties and interactions. These are task-dependent statistical uses: persistence does not directly calculate binding free energy, establish a biochemical mechanism, or replace physical modeling.\n\nIn neuroscience and dynamical systems, topology can reveal latent organization in high-dimensional activity. Neural population measurements may occupy a space with circular or toroidal structure when they encode periodic variables. A 2022 study of grid-cell population activity reported toroidal topology in the joint activity of cells within individual modules. Such findings concern the organization of population states rather than the physical shape of the brain. The inference depends on preprocessing, sampling coverage, and evidence distinguishing a torus from other possible structures.\n\nFor time series, delay-coordinate constructions convert successive measurements into points such as\n\n\\[x_t=(s_t,s_{t-\\tau},\\ldots,s_{t-(m-1)\\tau}),\\]\n\nwhere \\(\\tau\\) is a lag and \\(m\\) an embedding dimension. Periodic behavior can produce loop-like geometry, while more complicated dynamical systems can produce higher-dimensional structures. Sliding-window persistence has therefore been used to quantify periodicity and changes in dynamical organization. Trends, noise, uneven sampling, and the selected delay window can all affect the resulting topology.\n\nIn imaging, sublevel or superlevel filtrations summarize connected regions, tunnels, and cavities across thresholds. Applications include porous media, microscopy, and biomedical image analysis. A three-dimensional image of a porous material may contain many connected voids but few independent tunnels, or the reverse; these distinctions affect how topology describes its structure. Measurements such as the Euler characteristic provide computationally economical summaries, although they combine Betti numbers and can conceal offsetting changes in different dimensions.\n\nNetworks admit another family of applications. A weighted graph can be filtered by edge weight, and its clique complex adds a simplex whenever all required pairwise edges exist. Loops then describe missing higher-order fillings in this constructed complex, not necessarily literal pathways in the original system. Correlation-based networks require particular caution: thresholding correlations and filling cliques are modeling decisions, and the resulting topology does not establish causation. Across domains, the strongest interpretations connect a precisely defined topological feature to independently supported scientific structure."
    },
    {
      "heading": "Historical development and broader directions",
      "text": "TDA developed from several traditions rather than a single invention. Classical algebraic topology supplied homology and related invariants; computational geometry supplied finite representations of shapes; and statistical learning supplied questions about noisy, high-dimensional observations. In the early 1990s, Patrizio Frosini and other researchers developed size-theoretic descriptions of shape, while Vanessa Robins investigated persistent topological information in sampled sets later in that decade. These developments helped establish the idea that features should be followed across a changing parameter rather than measured only once.\n\nA major algorithmic milestone was the 2002 paper on topological persistence by Herbert Edelsbrunner, David Letscher, and Afra Zomorodian. Work by Zomorodian and Gunnar Carlsson subsequently clarified the algebraic structure and computation of persistence over fields. David Cohen-Steiner, Edelsbrunner, and John Harer established a fundamental persistence-diagram stability theorem in work published in 2007. These results connected a computable summary to a mathematical account of decomposition and perturbation.\n\nMapper was introduced by Gurjeet Singh, Facundo Mémoli, and Carlsson in 2007. Carlsson's 2009 article “Topology and data” helped articulate a broader program connecting topological methods with data analysis. During the following decade, statistical foundations, efficient implementations, and interfaces with machine learning expanded substantially. The field increasingly distinguished exploratory visualization, geometric reconstruction, hypothesis testing, and predictive feature engineering as related but separate objectives.\n\nOne important extension is multiparameter persistence. Ordinary persistence varies one parameter, but applications may require both distance and density, or spatial scale and time. A bifiltration \\(K_{r,s}\\) is nested in each coordinate, giving a persistence module indexed by a partially ordered plane. Unlike the finite one-parameter case, general multiparameter modules do not admit a complete classification by a simple collection of intervals. Researchers therefore use summaries such as rank invariants, Hilbert functions, presentations, and one-dimensional restrictions, each exposing different parts of the structure.\n\nZigzag persistence allows maps in alternating directions, accommodating sequences in which objects are both added and removed. This is useful for evolving networks and moving observation windows, where ordinary inclusion-only filtrations may be unnatural. Extended persistence combines ordinary homological information with relative constructions to pair features that would otherwise remain essential. Persistent cohomology also supports circular coordinates: suitable one-dimensional classes can be converted into circle-valued functions that describe cyclic organization.\n\nOther directions connect topology with learning and optimization. Differentiable persistence-based objectives allow selected topological properties to influence learned representations, image segmentation, or geometric reconstruction. Their derivatives are generally piecewise defined and can change when persistence pairings change. Meanwhile, research on computational limits, robust estimation, representative cycles, and uncertainty seeks to distinguish what an algorithm can calculate from what a data set can justify. Multiparameter models make this distinction especially explicit: richer descriptions can preserve more structure while also demanding more computation and more careful choices of interpretable summaries."
    }
  ],
  "links": [
    "topology",
    "geometry",
    "statistics",
    "persistent-homology",
    "metric-space",
    "euclidean-space",
    "field-mathematics",
    "quotient-vector-space",
    "finite-field",
    "homeomorphism",
    "linear-map",
    "cluster-analysis",
    "manifold",
    "probability-distribution",
    "confidence-interval",
    "bootstrap-sampling",
    "multiple-testing",
    "machine-learning",
    "cross-validation",
    "principal-component-analysis",
    "matrix",
    "linear-algebra",
    "floating-point-arithmetic",
    "protein",
    "time-series",
    "dynamical-system",
    "euler-characteristic",
    "correlation",
    "causation"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}