Manifold Reconstruction Using Tangential Delaunay Complexes
 620 Downloads
 11 Citations
Abstract
We give a provably correct algorithm to reconstruct a kdimensional smooth manifold embedded in ddimensional Euclidean space. The input to our algorithm is a point sample coming from an unknown manifold. Our approach is based on two main ideas: the notion of tangential Delaunay complex defined in Boissonnat and Flötotto (Comput. Aided Des. 36:161–174, 2004), Flötotto (A coordinate system associated to a point cloud issued from a manifold: definition, properties and applications. Ph.D. thesis, 2003), Freedman (IEEE Trans. Pattern Anal. Mach. Intell. 24(10), 2002), and the technique of sliver removal by weighting the sample points (Cheng et al. in J. ACM 47:883–904, 2000). Differently from previous methods, we do not construct any subdivision of the ddimensional ambient space. As a result, the running time of our algorithm depends only linearly on the extrinsic dimension d while it depends quadratically on the size of the input sample, and exponentially on the intrinsic dimension k. To the best of our knowledge, this is the first certified algorithm for manifold reconstruction whose complexity depends linearly on the ambient dimension. We also prove that for a dense enough sample the output of our algorithm is isotopic to the manifold and a close geometric approximation of the manifold.
Keywords
Tangential Delaunay complex Weighted Delaunay triangulation Manifold reconstruction Manifold learning Sampling conditions Sliver exudation1 Introduction
Manifold reconstruction consists of computing a piecewise linear approximation of an unknown manifold \(\mathcal{M}\subset \mathbb{R}^{d}\) from a finite sample of unorganized points \(\mathcal {P}\) lying on \(\mathcal{M}\) or close to \(\mathcal{M}\). When the manifold is a twodimensional surface embedded in \(\mathbb{R}^{3}\), the problem is known as the surface reconstruction problem. Surface reconstruction is a problem of major practical interest which has been extensively studied in the fields of Computational Geometry, Computer Graphics and Computer Vision. In the last decade, solid foundations have been established and the problem is now pretty well understood. Refer to Dey’s book [25], and the survey by Cazals and Giesen in [16] for recent results. The output of those methods is a triangulated surface that approximates \(\mathcal{M}\). This triangulated surface is usually extracted from a threedimensional subdivision of the ambient space (typically a grid or a triangulation). Although rather inoffensive in threedimensional space, such data structures depend exponentially on the dimension of the ambient space, and all attempts to extend those geometric approaches to more general manifolds have led to algorithms whose complexities depend exponentially on d [7, 15, 19, 40].
The problem in higher dimensions is also of great practical interest in data analysis and machine learning. In those fields, the general assumption is that, even if the data are represented as points in a very highdimensional space \(\mathbb{R}^{d}\), they in fact live on a manifold of much smaller intrinsic dimension [44]. If the manifold is linear, wellknown global techniques like principal component analysis (PCA) or multidimensional scaling (MDS) can be efficiently applied. When the manifold is highly nonlinear, several more local techniques have attracted much attention in visual perception and many other areas of science. Among the prominent algorithms are Isomap [45], LLE [42], Laplacian eigenmaps [8], Hessian eigenmaps [26], diffusion maps [37, 39], principal manifolds [48]. Most of those methods reduce to computing an eigendecomposition of some connection matrix. In all cases, the output is a mapping of the original data points into \(\mathbb{R}^{k}\) where k is the estimated intrinsic dimension of \(\mathcal{M}\). Those methods come with no or very limited guarantees. For example, Isomap provides a correct embedding only if \(\mathcal{M}\) is isometric to a convex open set of \(\mathbb{R}^{k}\) and LLE can only reconstruct topological balls. To be able to better approximate the sampled manifold, another route is to extend the work on surface reconstruction and to construct a piecewise linear approximation of \(\mathcal{M}\) from the sample in such a way that, under appropriate sampling conditions, the quality of the approximation can be guaranteed. First breakthrough, along this line, was the work of Cheng, Dey and Ramos [15] and this was followed by the work of Boissonnat, Guibas and Oudot [7]. In both cases, however, the complexity of the algorithms is exponential in the ambient dimension d, which highly reduces their practical relevance.
In this paper, we extend the geometric techniques developed in small dimensions and propose an algorithm that can reconstruct smooth manifolds of arbitrary topology while avoiding the computation of data structures in the ambient space. We assume that \(\mathcal{M}\) is a smooth manifold of known dimension k and that we can compute the tangent space to \(\mathcal{M}\) at any sample point. Under those conditions, we propose a provably correct algorithm that construct a simplicial complex of dimension k that approximates \(\mathcal{M}\). The complexity of the algorithm is linear in d, quadratic in the size n of the sample, and exponential in k. Our work builds on [7] and [15] but dramatically reduces the dependence on d. To the best of our knowledge, this is the first certified algorithm for manifold reconstruction whose complexity depends only linearly on the ambient dimension. In the same spirit, Chazal and Oudot [20] have devised an algorithm of intrinsic complexity to solve the easier problem of computing the homology of a manifold from a sample.
Our approach is based on two main ideas: the notion of tangential Delaunay complex introduced in [6, 29, 31], and the technique of sliver removal by weighting the sample points [14]. The tangential complex is obtained by gluing local (Delaunay) triangulations around each sample point. The tangential complex is a subcomplex of the ddimensional Delaunay triangulation of the sample points but it can be computed using mostly operations in the kdimensional tangent spaces at the sample points. Hence the dependence on k rather than d in the complexity. However, due to the presence of socalled inconsistencies, the local triangulations may not form a triangulated manifold. Although this problem has already been reported [31], no solution was known except for the case of curves (k=1) [29]. The idea of removing inconsistencies among local triangulations that have been computed independently has already been used for maintaining dynamic meshes [43] and generating anisotropic meshes [12]. Our approach is close in spirit to the one in [12]. We show that, under appropriate sample conditions, we can remove inconsistencies by weighting the sample points. We can then prove that the approximation returned by our algorithm is ambient isotopic to \(\mathcal{M}\), and a close geometric approximation of \(\mathcal{M}\).
Our algorithm can be seen as a local version of the cocone algorithm of Cheng et al. [15]. By local, we mean that we do not compute any ddimensional data structure like a grid or a triangulation of the ambient space. Still, the tangential complex is a subcomplex of the weighted ddimensional Delaunay triangulation of the (weighted) data points and therefore implicitly relies on a global partition of the ambient space. This is a key to our analysis and distinguishes our method from the other local algorithms that have been proposed in the surface reconstruction literature [22, 34].
Organization of the Paper
In Sect. 2, we introduce the basic concepts used in this paper. We recall the notion of weighted Voronoi (or power) diagrams and Delaunay triangulations in Sect. 2.1 and define sampling conditions in Sect. 2.2. We introduce various quantities to measure the shape of simplices in Sect. 2.3 and, in particular, the central notion of fatness. In Sect. 2.4, we define the two main notions of this paper: the tangential complex and inconsistent configurations.
The algorithmic part of the paper is given in Sect. 3.
The main structural results are given in Sect. 4. Under some sampling condition, we bound the shape measure of the simplices of the tangential complex in Sect. 4.2 and of inconsistent configurations in Sect. 4.3. A crucial fact is that inconsistent configurations cannot be fat. We also bound the number of simplices and inconsistent configurations that can be incident on a point in Sect. 4.4. In Sects. 4.5 and 4.6, we prove the correctness of the algorithm, and space and time complexity, respectively. In Sect. 5, we prove that the simplicial complex output by the algorithm is indeed a good approximation of the sampled manifold. Finally, in Sect. 6, we conclude with some possible extensions.
The list of main notations has been added at the end of the paper as a reference for the readers.
2 Definitions and Preliminaries
We will denote the topological boundary and interior of a set \(X \subseteq \mathbb{R}^{d}\) by ∂X and \({\rm int} X\), respectively.
For a map f:X→Y and X _{1}⊆X, \(f\mid_{X_{1}}\) denotes the restriction of map f to the subset X _{1}.
A jsimplex is the convex hull of j+1 affinely independent points. If τ is a jsimples with vertices {x _{0},…,x _{ j }}, we will also denote τ as [x _{0},…,x _{ j }]. For convenience, we often identify a simplex and the set of its vertices. Hence, if τ is a simplex, p∈τ means that p is a vertex of τ. If τ is a jsimplex, \({\rm aff}(\tau)\) denotes the jdimensional affine hull of τ and N _{ τ } denotes the (d−j)dimensional normal space of \({\rm aff}(\tau)\).
For all x,y in \(\mathbb{R}^{d}\), [xy] denotes the line segment joining the points x and y.
In this paper, \(\mathcal{M}\) denotes a differentiable manifold of dimension k embedded in \(\mathbb{R}^{d}\) and \(\mathcal {P}= \{ p_{1},\ldots,p_{n}\}\) a finite sample of points from \(\mathcal{M}\). We will further assume that \(\mathcal{M}\) has positive reach (see Sect. 2.2). We denote by T _{ p } and N _{ p } the kdimensional tangent space and (d−k)dimensional normal space at point \(p\in \mathcal{M}\), respectively.
For a given \(p \in \mathbb{R}^{d}\) and r≥0, B(p,r) (\(\bar{B}(p,r)\)) denotes the ddimensional Euclidean open (close) ball centered at p of radius r, and \(B_{\mathcal{M}}(p, r)\) (\(\bar{B}_{\mathcal{M}}(p,r)\)) denotes \(B(p, r) \cap \mathcal{M}\) (\(\bar{B}(p,r) \cap \mathcal{M}\)).
The following lemma is a direct consequence of the above definition. We have included the proof for completeness.
Lemma 2.1
 1.
If U ^{⊥} and V ^{⊥} are the orthogonal complements of U and V in \(\mathbb{R}^{d}\), then ∠(U,V)=∠(V ^{⊥},U ^{⊥}).
 2.
If dim(U)=dim(V) then ∠(U,V)=∠(V,U).
Proof
Without loss of generality we assume that the affine spaces U,V,U ^{⊥} and V ^{⊥} are vector subspaces of \(\mathbb{R}^{d}\), i.e. they all passes through the origin.
Since dimV ^{⊥}≤dimU ^{⊥}, and the orthogonal complement is a symmetric relation on subspaces, the same argument yields the reverse inequality.
2. Let ∠(U,V)=α, and let P:U→V denotes the projection map of the vector space U on V.
Case a. α≠π/2. Since α≠π/2 and dim(U)=dim(V), the projection map P is an isomorphism between vector spaces U and V. Therefore, for any unit vector v∈V there exists a vector u∈U such that P(u)=v. From the definition of angle between affine space and the linear map P, we have ∠(u,v) (=∠(v,u)). This implies, ∠(V,U)≤∠(U,V)=α. Using the same arguments, we can show that ∠(U,V)≤∠(V,U) hence ∠(U,V)=∠(V,U).
Case b. α=π/2. We have ∠(V,U)=π/2, otherwise using the same arguments as in Case 1 we can show that α=∠(U,V)≤∠(V,U)<π/2. □
2.1 Weighted Delaunay Triangulation
2.1.1 Weighted Points
A weighted point is a pair consisting of a point p of \(\mathbb{R}^{d}\), called the center of the weighted point, and a nonnegative real number ω(p), called the weight of the weighted point. It might be convenient to identify a weighted point (p,ω(p)) and the hypersphere (we will simply say sphere in the sequel) centered at p of radius ω(p).
Two weighted points (or spheres) (p,ω(p)) and (q,ω(q)) are called orthogonal when ∥p−q∥^{2}=ω(p)^{2}+ω(q)^{2}, further than orthogonal when ∥p−q∥^{2}>ω(p)^{2}+ω(q)^{2}, and closer than orthogonal when ∥p−q∥^{2}<ω(p)^{2}+ω(q)^{2}.
Given a point set \(\mathcal {P}= \{ p_{1},\ldots ,p_{n}\} \subseteq \mathbb{R}^{d}\), a weight function on \(\mathcal {P}\) is a function ω that assigns to each point \(p_{i}\in \mathcal {P}\) a nonnegative real weight ω(p _{ i }): \(\omega(\mathcal {P}) = (\omega(p_{1}), \ldots, \omega(p_{n}))\). We write \(p_{i}^{\omega}= (p_{i},\omega(p_{i}))\) and \(\mathcal {P}^{\omega} = \{ p^{\omega}_{1},\ldots,p^{\omega}_{n}\}\).
In the paper, we will assume that \(\tilde{\omega}\leq\omega_{0}\), for some constant ω _{0}∈[0,1/2) (see Hypothesis 3.2). Under this hypothesis, all the balls bounded by weighted spheres are disjoint.
Given a subset τ of d+1 weighted points whose centers are affinely independent, there exists a unique sphere orthogonal to the weighted points of τ. The sphere is called the orthosphere of τ and its center o _{ τ } and radius Φ _{ τ } are called the orthocenter and the orthoradius of τ. If the weights of the vertices of τ are 0 (or all equal), then the orthosphere is simply the circumscribing sphere of τ whose center and radius are, respectively, called circumcenter and circumradius. If τ is a jsimplex, j<d, the orthosphere of τ is the smallest sphere that is orthogonal to the (weighted) vertices of τ. Its center o _{ τ } lies in \({\rm aff}(\tau)\). Note that a simplex τ may have an orthoradius which is imaginary, i.e. \(\varPhi_{\tau}^{2} < 0\). This situation, however, cannot happen if the relative amplitude of the weight function is <1/2.
A finite set of weighted points \(\mathcal {P}^{\omega}\) is said to be in general position if there exists no sphere orthogonal to d+2 weighted points of \(\mathcal {P}^{\omega}\).
2.1.2 Weighted Voronoi Diagram and Delaunay Triangulation
Alternatively, a dsimplex τ is in \({\rm Del}^{\omega}(\mathcal {P})\) if the orthosphere o _{ τ } of τ is further than orthogonal from all weighted points in \(\mathcal {P}^{\omega}\setminus\tau\).
Observe that the definition of weighted Voronoi diagrams makes sense if, for some \(p\in \mathcal {P}\), ω(p)^{2}<0, i.e. some of the weights are imaginary. In fact, since adding a same positive quantity to all ω(p)^{2} does not change the diagram, handling imaginary weights is as easy as handling real weights. In the sequel, we will only consider positive weight functions with relative amplitude <1/2.
The weighted Delaunay triangulation of a set of weighted points can be computed efficiently in small dimensions and has found many applications, see e.g. [4]. In this paper, we use weighted Delaunay triangulations for two main reasons. The first one is that the restriction of a ddimensional weighted Voronoi diagram to an affine space of dimension k is a kdimensional weighted Voronoi diagram that can be computed without computing the ddimensional diagram (see Lemma 2.2). The other main reason is that some flat simplices named slivers can be removed from a Delaunay triangulation by weighting the vertices (see [7, 14, 15] and Sect. 3).
Lemma 2.2
Proof
2.2 Sampling Conditions
2.2.1 Local Feature Size
2.2.2 Sampling Parameters
The point sample \(\mathcal {P}\) is said to be an (ε,δ)sample (where 0<δ<ε<1) if (1) for any point \(x \in \mathcal{M}\), there exists a point \(p \in \mathcal {P}\) such that \(\xp\\leq \varepsilon \,{\rm lfs}(x)\), and (2) for any two distinct points \(p,q \in \mathcal {P}\), \(\pq\\geq\delta \,{\rm lfs}(p)\). The parameter ε is called the sampling rate, δ the sparsity, and ε/δ the sampling ratio of the sample \(\mathcal {P}\).
The following lemma, proved in [35], states basic properties of manifold samples. As before, we write \({\rm n\! n}(p)\) for the distance between \(p \in \mathcal {P}\) and its nearest neighbor in \(\mathcal {P}\setminus\{p\}\).
Lemma 2.3
 1.For all \(p \in \mathcal {P}\), we have$$\delta \,{\rm lfs}(p) \leq {\rm n\! n}(p) \leq\frac{2\varepsilon }{1\varepsilon }\,{\rm lfs}(p) . $$
 2.For any two points p,q ∈ \(\mathcal{M}\) such that \(\pq\ = t \,{\rm lfs}(p)\), 0<t<1,$$\sin\angle(pq, T_p) \leq t/2. $$
 3.Let p be a point in \(\mathcal{M}\). Let x be a point in T _{ p } such that \(\px\ \leq t \,{\rm lfs}(p)\) for some 0<t≤1/4. Let x′ be the point on \(\mathcal{M}\) closest to x. Then$$\bigl\ xx'\bigr\ \leq2t^2 \,{\rm lfs}(p). $$
2.3 Fat Simplices
Consider a jsimplex τ, where 1≤j≤k+1. We denote by R _{ τ },Δ _{ τ },L _{ τ },V _{ τ } and Γ _{ τ }=Δ _{ τ }/L _{ τ } the circumradius, the longest edge length, the shortestedge length, the jdimensional volume, and spread i.e. the longest edge to shortestedge ratio of τ, respectively.
The following important lemma is due to Whitney [46, Chap. II].
Lemma 2.4
We deduce from the above lemma the following corollary. See also Lemma 1 in [32] and Lemma 16 in [15].
Corollary 2.5
(Tangent space approximation)
Proof
A sliver is a special type of flat simplex. The property of being a sliver is defined in terms of a parameter Θ _{0}, to be fixed later in Sect. 4.
The following definition is a variant of a definition given in [36].
Definition 2.6
(Θ _{0}fat simplices and Θ _{0}slivers)
Given a positive parameter Θ _{0}, a simplex τ is said to be Θ _{0}fat if the fatness of τ and of all its subsimplices is at least Θ _{0}.
A simplex of dimension at least 2 which is not Θ _{0}fat but whose subsimplices are all Θ _{0}fat is called a Θ _{0}sliver.
2.4 Tangential Delaunay Complex and Inconsistent Configurations
We denote by tangential Delaunay complex or tangential complex for short, the simplicial complex \(\{ \tau : \tau\in {\rm st}(u), u\in \mathcal {P}\}\). We denote it by \({\rm Del}^{\omega}_{T\mathcal{M}}(\mathcal {P})\). By our assumption above, \({\rm Del}^{\omega}_{T\mathcal{M}}(\mathcal {P})\) is a kdimensional subcomplex of \({\rm Del}^{\omega}(\mathcal {P})\). Note that \({\rm st}(u) = {\rm st}(u, {\rm Del}^{\omega}_{u}(\mathcal {P}))\) and \({\rm st}(u, {\rm Del}^{\omega}_{T\mathcal{M}}(\mathcal {P}))\), the star of u in the complex \({\rm Del}_{T\mathcal{M}}^{\omega}(\mathcal {P})\), are in general different.
By duality, computing \({\rm st}(u)\) is equivalent to computing the restriction to T _{ u } of the (weighted) Voronoi cell of u, which, by Lemma 2.2, reduces to computing a cell in a kdimensional weighted Voronoi diagram embedded in T _{ u }. To compute such a cell, we need to compute the intersection of \(\mathcal {P}1\) halfspaces of T _{ u } where \(\mathcal {P}\) is the cardinality of \(\mathcal {P}\). Each halfspace is bounded by the bisector consisting of the points of T _{ u } that are at equal weighted distance from u ^{ ω } and some other point in \(\mathcal {P}^{\omega}\). This can be done in optimal time [17, 21]. It follows that the tangential complex can be computed without constructing any data structure of dimension higher than k, the intrinsic dimension of \(\mathcal{M}\).
We now formally define an inconsistent configuration as follows.
Definition 2.7
(Inconsistent configuration)

The ksimplex τ=ϕ∖{w} is in \({\rm st}(u)\) but not in \({\rm st}(v)\).

\({\rm Vor}^{\omega}(w)\) is one of the first weighted Voronoi cells of \({\rm Vor}^{\omega}(\mathcal {P})\), other than the weighted Voronoi cells of the vertices of τ, that is intersected by the line segment [m _{ u } m _{ v }] oriented from m _{ u } to m _{ v }. Here \(m_{u} = T_{u} \cap {\rm Vor}^{\omega}(\tau)\) and \(m_{v} = T_{v}\cap {\rm aff}({\rm Vor}^{\omega}(\tau ))\). Let i _{ ϕ } denote the point where the oriented segment [m _{ u } m _{ v }] first intersects \({\rm Vor}^{\omega}(w)\).

τ is a Θ _{0}fat simplex.
Note that i _{ ϕ } is the center of a sphere that is orthogonal to the weighted vertices of τ and also to w ^{ ω }, and further than orthogonal from all the other weighted points of \(\mathcal {P}^{\omega}\). Equivalently, i _{ ϕ } is the point on [m _{ u } m _{ v }] that belongs to \({\rm Vor}^{\omega}(\phi)\).
An inconsistent configuration is therefore a (k+1)simplex of \({\rm Del}^{\omega} (\mathcal {P})\). However, an inconsistent configuration does not belong to \({\rm Del}_{T\mathcal{M}}^{\omega}(\mathcal {P})\) since \({\rm Del}_{T\mathcal{M}}^{\omega}(\mathcal {P})\) has no (k+1)simplex under our general position assumption. Moreover, the lower dimensional faces of an inconsistent configuration do not necessarily belong to \({\rm Del}_{T\mathcal{M}}^{\omega}(\mathcal {P})\).
Since the inconsistent configurations are k+1dimensional simplices, we will use the same notations for inconsistent configurations as for simplices, e.g. R _{ ϕ } and c _{ ϕ } for the circumradius and the circumcenter of ϕ, ρ _{ ϕ } and Θ _{ ϕ } for its radiusedge ratio and fatness, respectively.
An important observation, stated as Lemma 4.9 in Sect. 4.3, is that, if ε is sufficiently small with respect to Θ _{0}, then the fatness of ϕ is less than Θ _{0}. Hence, if the subfaces of ϕ are Θ _{0}fat simplices, ϕ will be a Θ _{0}sliver. This observation is at the core of our reconstruction algorithm.
3 Manifold Reconstruction
The algorithm removes all Θ _{0}slivers from \(C^{\omega}(\mathcal {P})\) by weighting the points of \(\mathcal {P}\). By the observation mentioned above, all inconsistencies in the tangential complex will then also be removed. All simplices being consistent, the resulting weighted tangential Delaunay complex \(\hat{\mathcal{M}}\) output by the algorithm will be a simplicial kmanifold that approximates \(\mathcal{M}\) well, as will be shown in Sect. 5.
In this section, we describe the algorithm. Its analysis is deferred to Sect. 4.
We will make the two following hypotheses.
Hypothesis 3.1
\(\mathcal{M}\) is a compact smooth submanifold of \(\mathbb{R}^{d}\) without boundary, and \(\mathcal {P}\) is an (ε,δ)sample of \(\mathcal{M}\) of sampling ratio ε/δ≤η _{0} for some positive constant η _{0}.
As shown in [23, 35], we can estimate the tangent space T _{ p } at each sample point p and also the dimension k of the manifold from \(\mathcal {P}\) and η _{0}. We assume now that T _{ p }, for any point \(p\in \mathcal {P}\), k and η _{0} are known.
The algorithm fixes a bound ω _{0} on the relative amplitude of the weight assignment:
Hypothesis 3.2
\(\tilde{\omega}\leq\omega_{0}\), for some constant ω _{0}∈[0,1/2).
Observe that, under this hypothesis, all the balls bounded by weighted spheres are disjoint.
The algorithm also fixes Θ _{0} to a constant defined in Theorem 4.15, which depends on k, ω _{0} and η _{0}.
Outline of the Algorithm
The function build_star(p) calculates the weighted Voronoi cell of p, which reduces to computing the intersection of the halfspaces of T _{ p } bounded by the (weighted) bisectors between p and other points in LN(p).
The function build_inconsistent_configurations(u,τ) adds to \(C^{\omega}(\mathcal {P})\) all the inconsistent configurations of the form ϕ=τ∪{w} where τ is an inconsistent simplex of \({\rm st}(u)\). More precisely, for each vertex v≠u of τ such that \(\tau \not\in {\rm st}(v)\), we calculate the points w∈LN(p), such that (u,v,w) witnesses the inconsistent configuration ϕ=τ∪{w}. Specifically, we compute the restriction of the Voronoi diagram of the points in LN(u) to the line segment [m _{ u } m _{ v }], where \(m_{u} = T_{u} \cap {\rm aff}({\rm Vor}^{\omega}(\tau))\) and \(m_{v} = T_{v} \cap {\rm aff}({\rm Vor}^{\omega}(\tau))\). According to the definition of an inconsistent configuration, w is one of the sites whose (restricted) Voronoi cell is the first to be intersected by the line segment [m _{ u } m _{ v }], oriented from m _{ u } to m _{ v }. We add inconsistent configuration ϕ=τ∪{w} to the completed complex.
Let τ be a candidate simplex of p that is a Θ _{0}sliver. We associate to τ a forbidden interval W(p,τ) that consists of all squared weights ω(p)^{2} for which τ appears as a simplex in \(C^{\omega}(\mathcal {P})\) (the weights of the other points remaining fixed).
The function candidate_slivers(p,ω) varies the weight of p and computes all the candidate slivers of p and their corresponding weight intervals W(p,τ). More precisely, this function follows the following steps.
1. We first detect all candidate jsimplices for all 2≤j≤k+1. This is done in the following way. We vary the weight of p from 0 to \(\omega_{0} {\rm n\! n}(p)\), keeping the weights of the other points fixed. For each new weight assignment to p, we modify the stars and inconsistent configurations of the points in LN(p) and detect the new jsimplices incident to p that have not been detected so far. The weight of point p changes only in a finite number of instances \(0= P_{0} < P_{1} < \dots< P_{n1} < P_{n} = \omega _{0} {\rm n\! n}(p)\).
2. We determine the next weight assignment of p in the following way. For each new simplex τ currently incident to p, we keep it in a priority queue ordered by the weight of p at which τ will disappear for the first time. Hence the minimum weight in the priority queue gives the next weight assignment for p. Since the number of points in LN(p) is bounded, the number of simplices incident to p is also bounded, as well as the number of times we have to change the weight of p.
3. For each candidate sliver τ of p which is detected, we compute W(p,τ) on the fly. The candidate slivers are stored in a list with their corresponding forbidden intervals.
4 Analysis of the Algorithm
The analysis of the algorithm relies on structural results that will be proved in Sects. 4.2, 4.3 and 4.4. We will then prove that the algorithm is correct and analyze its complexity in Sects. 4.5 and 4.6. In Sect. 5, we will show that the output \(\hat{\mathcal{M}}\) of the reconstruction algorithm is a good approximation of \(\mathcal{M}\).
The bounds to be given in the lemmas of this section will depend on the dimension k of \(\mathcal{M}\), the bound η _{0} on the sampling ratio, and on a positive scalar Θ _{0} that bounds the fatness and will be used to define slivers, fat simplices and inconsistent configurations.
4.1 First Lemmas
4.1.1 Circumradius and Orthoradius
The following lemma states some basic facts about weighted Voronoi diagrams when the relative amplitude of the weighting function is bounded. Similar results were proved in [14].
Lemma 4.1
 1.
\(\forall z \in {\rm aff}({\rm Vor}^{\omega}(\tau))\), \(\qz\ \leq\frac{\pz\}{\sqrt{14\omega_{0}^{2}}}\).
 2.
\(\forall z \in {\rm aff}({\rm Vor}^{\omega}(\tau))\), \(\sqrt{\zp\^{2}\omega^{2}(p)}\) ≥Φ _{ τ }.
 3.
∀σ⊆τ, Φ _{ σ }≤Φ _{ τ }.
Proof
3. The result directly follows from part 2 and the fact that \({\rm aff}({\rm Vor}^{\omega}(\tau)) \subseteq {\rm aff}({\rm Vor}^{\omega}(\sigma))\) (since σ⊆τ). □
4.1.2 Altitude and Fatness
If p∈τ, we define τ _{ p }=τ∖{p} to be the (j−1)face of τ opposite to p. We also write D _{ τ }(p) for the distance from p to the affine hull of τ _{ p }. D _{ τ }(p) will be called the altitude of p in τ.
From the definition of fatness, we easily derive the following lemma.
Lemma 4.2
 1.
\(\varTheta^{j}_{\tau} \leq\frac{1}{j!}\).
 2.
\(j! \varTheta^{j}_{\tau} \leq\frac{D_{\tau}(p)}{\varDelta_{\tau}} \leq j \varGamma^{j1}_{\tau} \times \frac{\varTheta^{j}_{\tau}}{\varTheta^{j1}_{\tau_{p}}}\).
Proof
4.1.3 Eccentricity
Let τ be a simplex and p be a vertex of τ. We define the eccentricity H _{ τ }(p,ω(p)) of τ with respect to p as the signed distance from o _{ τ } to \({\rm aff}(\tau_{p})\). Hence, H _{ τ }(p,ω(p)) is positive if o _{ τ } and p lie on the same side of \({\rm aff}(\tau_{p})\) and negative if they lie on different sides of \({\rm aff}(\tau_{p})\).
The following lemma is a generalization of Claim 13 from [14]. It bounds the eccentricity of a simplex τ as a function of ω(p) where p is a vertex of τ. The proof is included for completeness even though it is exactly the same as the one given in [14].
Lemma 4.3
Proof
4.2 Properties of the Tangential Delaunay Complex
The following two lemmas are slight variants of results of [15]. The first lemma states that the restriction of a (weighted) Voronoi cell to a tangent space is small.
Lemma 4.4
Assume that Hypotheses 3.1 and 3.2 are satisfied. There exists a positive constant C _{1} such that for all \(T_{p} \cap {\rm Vor}^{\omega}(p)\), \(\xp\ \leq C_{1} \varepsilon \,{\rm lfs}(p)\).
Proof
Let q be a point on the line segment [px] such that \(\pq\= C_{1}\varepsilon \,{\rm lfs}(p)/2\). Let q′ be the point closest to q on \(\mathcal{M}\). From Lemma 2.3, we have \(\qq'\\leq C_{1}^{2}\varepsilon ^{2} \,{\rm lfs}(p)/2\).
The following lemma states that, under Hypotheses 3.2 and 3.1, the simplices of \({\rm Del}_{p}^{\omega}(\mathcal {P})\) are small, have a good radiusedge ratio and a small eccentricity.
Lemma 4.5
 1.
If pq is an edge of \({\rm Del}_{T\mathcal{M}}^{\omega}(\mathcal {P})\), then \(\pq\ < C_{2}\varepsilon \,{\rm lfs}(p)\).
 2.
If τ is a simplex of \({\rm Del}_{T\mathcal{M}}^{\omega}(\mathcal {P})\), then Φ _{ τ }≤C _{3} L _{ τ } and Γ _{ τ }=Δ _{ τ }/L _{ τ }≤C _{3}.
 3.
If τ is a simplex of \({\rm Del}_{T\mathcal{M}}^{\omega}(\mathcal {P})\) and p a vertex of τ, the eccentricity H _{ τ }(p,ω(p)) is at most \(C_{4}\varepsilon \,{\rm lfs}(p)\).
Proof
 (a)Consider first the case where pq is an edge of \({\rm Del}_{p}^{\omega}(\mathcal {P})\). Then \(T_{p} \cap {\rm Vor}^{\omega}(pq) \neq\emptyset\). Let \(x\in T_{p} \cap {\rm Vor}^{\omega}(pq)\). From Lemma 4.4, we have \(\px\ \leq C_{1}\varepsilon \,{\rm lfs}(p)\). By Lemma 4.1,Hence, \(\ pq\ \leq C_{1}' \varepsilon \,{\rm lfs}(p)\) where \(C_{1}'\stackrel{{\rm def}}{=}C_{1}(1+1/\sqrt{14\omega_{0}^{2}})\).$$\qx\ \leq\frac{\px\}{\sqrt{14\omega_0^2}} \leq \frac{C_1\varepsilon \,{\rm lfs}(p)}{\sqrt{14\omega_0^2}}. $$
 (b)From the definition of \({\rm Del}_{T\mathcal{M}}^{\omega}(\mathcal {P})\), there exists a vertex r of τ such that \([pq] \in {\rm st}(r)\). From (a), ∥r−p∥ and ∥r−q∥ are at most \(C_{1}'\varepsilon \,{\rm lfs}(r)\). Using the fact that \({\rm lfs}\) is 1Lipschitz, \({\rm lfs}(p) \geq {\rm lfs}(r)  \pr\ \geq(1C_{1}' \varepsilon ) \,{\rm lfs}(r)\) (from part 1 (a)), which yields \({\rm lfs}(r) \leq\frac{{\rm lfs}(p)}{1C'_{1}\varepsilon }\). It follows thatThe first part of the lemma is proved by taking \(C_{2}\stackrel{{\rm def}}{=} \frac{5C_{1}' }{2}\) and using 2C _{2} ε<1.$$\ pq\ \leq\ pr\ + \ rq\ \leq\frac{2C_1'\varepsilon \,{\rm lfs}(p)}{1C_1'\varepsilon }. $$
4.3 Properties of Inconsistent Configurations
We now give lemmas on inconsistent configurations which are central to the proof of correctness of the reconstruction algorithm given later in the paper. The first lemma is the analog of Lemma 4.5 applied to inconsistent configurations. Differently from Lemma 4.5, we need to use Corollary 2.5 to control the orientation of the facets of \({\rm Del}_{T\mathcal{M}}^{\omega}(\mathcal {P})\) and require the following additional hypothesis relating the sampling rate ε and the fatness bound Θ _{0}.
Hypothesis 4.6
2Aε<1 where \(A \stackrel{{\rm def}}{=} 2C_{2} C_{3}/\varTheta^{k}_{0}\), and C _{2} and C _{3} are the constants defined in Lemma 4.5.
Lemma 4.7
(Tangent approximation)
Let τ be a ksimplex of \({\rm Del}_{T\mathcal{M}}^{\omega}(\mathcal {P})\) that is Θ _{0}fat. For all vertices p of τ, we have \(\sin\angle({\rm aff}(\tau), T_{p}) = \sin\angle(T_{p}, {\rm aff}(\tau)) \leq A\varepsilon \). The constant A, defined in Hypothesis 4.6, depends on k, ω _{0}, η _{0} and Θ _{0}.
Proof
Lemma 4.8
 1.
\(\pi_{\phi}\ \leq\frac{C_{2}'}{2} \varepsilon \,{\rm lfs}(p)\) for all vertices p of ϕ.
 2.
If pq is an edge of ϕ then \(\pq\ \leq C_{2}'\varepsilon \,{\rm lfs}(p)\).
 3.
If σ⊆ϕ, then \(\varPhi_{\sigma} \leq C_{3}' L_{\sigma}\) and \(\varGamma_{\sigma} = {\varDelta_{\sigma}} / {L_{\sigma}} \leq C_{3}'\).
 4.
If σ⊆ϕ and p is any vertex of σ, H _{ σ }(p,ω(p)) of σ is at most \(C_{4}'\varepsilon \,{\rm lfs}(p)\).
Proof
The next crucial lemma bounds the fatness of inconsistent configurations.
Lemma 4.9
Proof
A consequence of the lemma is that, if the subfaces of ϕ are Θ _{0}fat simplices and if
Hypothesis 4.10
\(\frac{C_{2}' \varepsilon }{(k+1)!} ( 1+\frac{2C_{3}}{\varTheta^{k}_{0}} ) < \varTheta_{0}^{k+1}\)
is satisfied, then ϕ is a Θ _{0}sliver. Hence, techniques to remove slivers can be used to remove inconsistent configurations.
In the above lemmas, we assumed that ε is small enough. Specifically in addition to Hypotheses 3.1, 3.2 and 4.6, we assumed that 2C _{2} ε<1 in Lemma 4.5, \(C_{2}'\varepsilon <1\) in Lemma 4.8 and \(2C_{2}'\varepsilon < 1\) in Lemma 4.9. We will make another hypothesis that subsumes these two previous conditions.
Hypothesis 4.11
\(C_{2}'(1+C_{2}' \eta_{0}) \varepsilon <1/2\).
Observe that this hypothesis implies \(C_{2}'(1+C_{2}') \varepsilon <1/2\) since η _{0}>1. Observe also that Hypotheses 4.6, 4.10 and 4.11 are satisfied for fixed ω _{0}, η _{0}, Θ _{0} and a sufficiently small ε.
4.4 Number of Local Neighbors
We will use the result from this section for the analysis of the algorithm, and also for calculating its time and space complexity.
Let \(N \stackrel{{\rm def}}{=} (4C'_{2}\eta_{0}+6)^{k}\), where the constant \(C'_{2}\) is defined in Lemma 4.8.
Lemma 4.12
Proof
Lemmas 4.5 and 4.8 show that, in order to construct \({\rm st}(p)\) and search for inconsistencies involving p, it is enough to consider the points of \(\mathcal {P}\) that lie in ball \(B_{p}= B(p, C_{2}'\varepsilon \,{\rm lfs}(p))\). Therefore it is enough to count the number of points in \(B_{p} \cap \mathcal {P}\).
4.5 Correctness of the Algorithm
Definition 4.13
(Sliverity range)
Let ω be a weight assignment satisfying Hypothesis 3.2. The weight of all the points in \(\mathcal {P}\setminus\{p\}\) are fixed and the weight ω(p) of p is varying. The sliverity range Σ(p) of a point \(p\in \mathcal {P}\) is the measure of the set of all squared weights ω(p)^{2} for which p is a vertex of a Θ _{0}sliver in \(C^{\omega}(\mathcal {P})\).
Lemma 4.14
Proof
Theorem 4.15
Take \(\varTheta_{0} =\frac{\omega^{2}_{0}}{2 N^{k+1} C_{5}}\) and assume that Hypotheses 3.1, 3.2, 4.6, 4.10 and 4.11 are satisfied. Then, the simplicial complex \(\hat{\mathcal{M}}\) output by Algorithm 1 has no inconsistencies and its simplices are all Θ _{0}fat.
Proof
The sliverity range Σ(p) of p is at most \(2N^{k+1}C_{5} \varTheta_{0} {\rm n\! n}(p)^{2}\) from Lemma 4.14. Since \(\varTheta_{0} = \frac{\omega_{0}^{2}}{2N^{k+1} C_{5}}\), Σ(p) is less than the total range of possible squared weights \(\omega_{0}^{2} {\rm n\! n}(p)^{2}\). Hence, function weight (p,ω) will always find a weight for any point \(p\in \mathcal {P}\) and any weight assignment of relative amplitude at most ω _{0} for the points of \(\mathcal {P}\setminus \{p\}\).
Since the algorithm removes all the simplices of \(C^{\omega}(\mathcal {P})\) that are not Θ _{0}fat, all the simplices of \(\hat{\mathcal{M}}\) are Θ _{0}fat.
By Lemma 4.9 and Hypothesis 4.10, all the inconsistent configurations in \(C^{\omega}(\mathcal {P})\) are either Θ _{0}slivers or contain subfaces which are Θ _{0}slivers. It follows that \(\hat{\mathcal{M}}\) has no inconsistency since, when the algorithm terminates, all simplices of \(C^{\omega}(\mathcal {P})\) are Θ _{0}fat. □
4.6 Time and Space Complexity
Theorem 4.16
Proof
Space Complexity: For each point \(p \in \mathcal {P}\) we maintain LN(p). The total space complexity for storing LN(p) for each point \(p \in \mathcal {P}\) is thus \(O(N\mathcal {P})\) by definition of LN(p).
By Lemma 2.2, each \({\rm st}(p)\), \(p \in \mathcal {P}\), has the same combinatorial complexity as a Voronoi cell in the kdimensional flat T _{ p }. Since the sites needed to compute this Voronoi cell all belong to LN(p), there number is at most N by Lemma 4.12. From the Upper Bound Theorem of convex geometry, see e.g. [13], the combinatorial complexity of each star is therefore O(N ^{⌊k/2⌋}). Hence the total space complexity of the tangential Delaunay complex is \(O(kN^{\lfloor k/2 \rfloor}) \mathcal {P}\).
For a given inconsistent Θ _{0}fat ksimplex in \({\rm st}(p)\), we can have from Lemmas 4.8 (2) and 4.12, at most kLN(p)≤kN different inconsistent configurations. Hence, the number of inconsistent configurations to be stored in the completed complex \(C^{\omega}(\mathcal {P})\) is at most \(O(k^{2}N^{\lfloor k/2 \rfloor+1}) \mathcal {P}\).
The main loop of the algorithm processes each point \(p \in \mathcal {P}\) in turn, and finds a weight assignment ω(p) using the function weight( , ). In the function weight, we maintain a list of candidate simplices of p with their corresponding forbidden intervals and a priority queue of candidate simplices of p. By Lemmas 4.5 (1), 4.8 (2) and 4.12, the space complexities of both these data structures are bounded by O(kN ^{ k+1}).
The time complexity to build all the inconsistent configurations corresponding to \({\rm st}(u)\) is O(dkN ^{⌊k/2⌋+1}) since the number of ksimplices in the star of a point p is O(N ^{⌊k/2⌋}).
Observe that, since \(\mathcal {P}\) is an (ε,δ)sample of \(\mathcal{M}\) with ε/δ≤η _{0}, \(\mathcal {P} = O(\varepsilon ^{k})\).
5 Topological and Geometric Guarantees
In this section we give conditions under which \({\rm Del}^{\omega}_{T\mathcal{M}}(\mathcal {P})\) is isotopic to and a close geometric approximation of the manifold \(\mathcal{M}\). The results in this section will be used to prove the theoretical guarantees on the quality of the simplicial complex \(\hat{\mathcal{M}}\) output by Algorithm 1 in Sect. 3.
5.1 General Position Assumption
We assume that the weighted points in \(\mathcal {P}^{\omega}\) are in general position in \(\mathbb{R}^{d}\) and this will imply \({\rm Del}^{\omega}(\mathcal {P})\) is a simplicial complex, see Sect. 2.1.2. Therefore \({\rm Del}^{\omega}_{T\mathcal{M}}(\mathcal {P})\) is a simplicial complex.
We know from Lemma 2.2 that \({\rm Vor}^{\omega}(\mathcal {P})\cap T_{p}\) is identical to the weighted Voronoi diagram \({\rm Vor}^{\xi_{p}}(\mathcal {P}_{p})\) of \(\mathcal {P}^{\xi_{p}}\) in T _{ p }, where \(\mathcal {P}_{p}\) is obtained by projecting \(\mathcal {P}\) orthogonally onto T _{ p } and ξ _{ p } is the weight assignment defined in Lemma 2.2.
Under Hypotheses 3.1 and 3.2, \({\rm Vor}^{\xi_{p}}(p)\) is bounded (Lemma 4.4) and therefore \(\dim {\rm aff}(\mathcal {P}_{p}) = k\). In addition to the general position assumption of \(\mathcal {P}^{\omega}\), we will also assume that the weighted points \(\mathcal {P}^{\xi_{p}}_{p}\) are in general position on T _{ p }, meaning that there exists no orthosphere centered on T _{ p } that is orthogonal to k+2 weighted points of \(\mathcal {P}_{p}^{\xi_{p}}\). See Sects. 2.1.1 and 2.1.2. As we have already mentioned in Sect. 2.4, the general position assumption can be fulfilled by applying an infinitesimal perturbation to the point sample or to the weight assignment ω.
Under the general position assumption above, \({\rm Del}^{\xi_{p}}(\mathcal {P}_{p})\) is a triangulation of \({\rm conv}(\mathcal {P}_{p})\) [3, 4, 5]. Moreover, since \({\rm Vor}^{\xi_{p}}(p)\) is bounded (from Lemma 4.4), p is an interior vertex of \({\rm Del}^{\xi_{p}}(\mathcal {P}_{p})\). This implies the following result.
Lemma 5.1
For all \(p \in \mathcal {P}\), \({\rm st}(p)\) is isomorphic to the star of an interior vertex of a compact triangulated kdimensional polytope.
5.2 Medial Axis and the Projection Map
Lemma 5.2
 1.
The map π is a C ^{1}function.
 2.
For all \(x \in \mathbb{R}^{d} \setminus\mathcal{O}\), the kernel of the linear map \(d\pi(x) : \mathbb{R}^{d} \rightarrow T_{\pi(x)}\), where dπ(x) denotes the derivative of π at x, is parallel to N _{ π(x)} and has dimension d−k.
5.3 Main Result
The main result of this section is the following theorem.
Theorem 5.3
 C1.

\({\rm Del}^{\omega}_{T\mathcal{M}}(\mathcal {P})\) has no kdimensional inconsistent simplices.
 C2.

All the simplices in \({\rm Del}^{\omega}_{T\mathcal{M}}(\mathcal {P})\) are Θ _{0}fat.
 P1.

Tangent space approximation: Let τ be a ksimplex in \(\hat{\mathcal{M}}\). For all vertices p of τ, we have \(\sin\angle(T_{p}, {\rm aff}(\tau)) = \sin\angle({\rm aff}(\tau), T_{p}) = O(\varepsilon )\). Note that this property directly follows from Lemma 4.7.
 P2.

PL kmanifold without boundary: \(\hat{\mathcal{M}}\) is a piecewise linear kmanifold without boundary;
 P3.

Homeomorphism: The map \(\pi_{\hat{\mathcal{M}}}\) provides a homeomorphism between \(\hat{\mathcal{M}}\) and \(\mathcal{M}\);
 P4.

Pointwise approximation: \(\forall x \in \mathcal{M}\), \({\rm dist}(x, \pi_{\hat{\mathcal{M}}}^{1}(x)) = O(\varepsilon ^{2}\,{\rm lfs}(x))\);
 P5.

Isotopy: There exists an isotopy \(F : \mathbb{R}^{d} \times[0, 1] \rightarrow \mathbb{R}^{d}\) such that the map F(⋅,0) restricted to \(\hat{\mathcal{M}}\) is the identity map on \(\hat{\mathcal{M}}\) and \(F(\hat{\mathcal{M}}, 1) = \mathcal{M}\).
Since Conditions C1 and C2 are satisfied when the conditions of Theorem 4.15 are satisfied, it follows that the simplicial complex \(\hat{\mathcal{M}}\) output by Algorithm 1 satisfies the above Properties P1–P5 provided that the additional condition on ε in the above theorem is satisfied.
The rest of the section will be devoted to the proof of Theorem 5.3.
5.4 Piecewise Linear kManifold
The following lemma is a direct consequence of Condition C1.
Lemma 5.4
Under Condition C1 of Theorem 5.3 and for a sufficiently small ε, \({\rm st}(p) = {\rm st}(p, \hat{\mathcal{M}})\) for all \(p \in \mathcal {P}\).
We can now prove that \(\hat{\mathcal{M}}\) is a PL kmanifold without boundary (Property P2).
Lemma 5.5
(PL kmanifold without boundary)
Assume Condition C1 of Theorem 5.3. If ε is sufficiently small, then \(\hat{\mathcal{M}}\) is a PL kmanifold without boundary.
Proof
Since \({\rm Del}^{\xi_{p}}(\mathcal {P}_{p})\) is a triangulation of a kdimensional polytope, \({\rm Del}^{\xi_{p}}(\mathcal {P}_{p})\) is a PL kmanifold [47, Lem. 8 of Chap. 3] and, since p is an interior vertex of \({\rm Del}^{\xi_{p}}(\mathcal {P}_{p})\), \({\rm lk}(p, {\rm Del}^{\xi_{p}}(\mathcal {P}_{p}))\) is a PL (k−1)sphere [47, Proof of Lem. 9 of Chap. 3].
From Lemma 5.4, we have \({\rm st}(p) = {\rm st}(p, \hat{\mathcal{M}})\). By Lemma 5.1, \({\rm st}(p, \hat{\mathcal{M}})\) is isomorphic to the star of an interior vertex of a kdimensional triangulated convex domain. Hence, \({\rm lk}(p, \hat{\mathcal{M}})\) is a PL (k−1)sphere. This directly implies that \(\hat{\mathcal{M}}\) is a PL kmanifold with no boundaries. □
Property P2 follows since the output \(\hat{\mathcal{M}}\) of the Algorithm 1 has no inconsistency and therefore satisfies Condition C1.
5.5 Homeomorphism
Following lemma establishes Property P3 of Theorem 5.3, and the rest of the section is devoted to its proof.
Lemma 5.6
(Homeomorphism)
Assume Conditions C1 and C2 of Theorem 5.3. For ε sufficiently small, the map π restricted to \(\hat{\mathcal{M}}\) gives a homeomorphism between \(\hat{\mathcal{M}}\) and \(\mathcal{M}\).
In the broad outline, the proof of Lemma 5.6 is similar to the proof of homeomorphism given in [2] but the technical details are quite different in most places. The difficulties that arose in getting results analogous to the ones in [2] were handled using ideas from [46, Chap. II].
In the rest of this paper, we assume that ω _{0},η _{0},Θ _{0} are fixed constants, and that ε is small enough. We use the asymptotic notations O() and Ω() hiding the constants.
Remark 5.7
For the details of the missing proofs of Lemmas 5.8, 5.9, 5.10, and 5.11 refer to [33, Chap. 4].
5.5.1 Basic Lemmas
The following lemma is a generalization of Proposition 6.2 in [40] which bounds the variation of the angle between tangent spaces between two points on the manifold \(\mathcal{M}\).
Lemma 5.8
(Tangent variation)
Using Lemma 5.8, we can show that the derivative of the restriction \(\widetilde{\pi}_{p}\) of π _{ p } to a neighborhood of p is nonsingular and that \(\widetilde{\pi}_{p}\) is injective. Using the Inverse function theorem this implies that \(\widetilde{\pi}_{p}\) is a diffeomorphism.
Lemma 5.9
(\(\widetilde{\pi}_{p}\) is a C ^{1}diffeomorphism)
There exist absolute constants r _{1} and r _{2} such that \(\widetilde{\pi}_{p}\) restricted to \(\widetilde{T}_{p} \stackrel{\rm def}{=} B(p, r_{1}\,{\rm lfs}(p))\cap T_{p}\) is a diffeomorphism onto \(\widetilde{\pi}_{p}(\widetilde{T}_{p})\) and \(\widetilde{\pi}_{p}(\widetilde{T}_{p}) \supseteq\widetilde{\mathcal{M}}_{p} \stackrel{{\rm def}}{=} B_{\mathcal{M}}(p, r_{2}\,{\rm lfs}(p))\).
Note that \(\widetilde{\mathcal{M}}_{p} \subseteq\widetilde{\pi }_{p}(\widetilde{T}_{p})\) follows from Lemma 2.3(3).
The next lemma follows directly from Lemmas 2.3, 4.7, 5.8 and 5.9.
Lemma 5.10
We will need the following technical lemma bounding the fatness of a perturbed simplex. A similar inequality can be found in the proof of Lemma 14c in [46, Chap. 4].
Lemma 5.11
We will also need the following standard lemma from convex geometry which bounds the distance between an interior point and a point on the boundary of a simplex. The proof can be found in [46, Lem. 14b of Chap. 4].
Lemma 5.12
We will use the standard notion of oriented simplices, positively oriented simplices and oriented PL manifolds (see [41, 47] or [46, Appendix II]).
Recall that the restriction of π _{ p } to \({\rm st}(p)\) provides a simplicial isomorphism between \({\rm st}(p)\) and \({\rm st}(p, {\rm Del}^{\xi_{p}}(\mathcal {P}_{p}))\) (Lemma 2.2), and that \({\rm st}(p) = {\rm st}(p, \hat{\mathcal{M}})\) (by Lemma 5.4). We orient T _{ p } and orient positively the kdimensional simplices of \({\rm st}(p, {\rm Del}^{\xi_{p}}(\mathcal {P}_{p}) )\). We then orient the kdimensional simplices of \({\rm st}(p, \hat{\mathcal{M}})\) isomorphically, using the orientation of the corresponding kdimensional simplices of \({\rm st}(p, {\rm Del}^{\xi_{p}}(\mathcal {P}_{p}))\).
Since \({\rm lk}(p, \hat{\mathcal{M}})\) is a PL (j−1)sphere (see the proof of Lemma 5.5), we immediately see that \({\rm st}(p, \hat{\mathcal{M}})\) is an oriented PL kball.
Definition 5.13
(Simplexwise positive map)
Let σ be a positively oriented isimplex of \(\mathbb{R}^{i}\), and let \(f: \sigma\rightarrow \mathbb{R}^{i}\) be a C ^{1}function. The map f is called simplexwise positive if det(J(f))>0 for all x∈σ, where J(f) and det(J(f)) denote the Jacobian and the determinant of the Jacobian of the map f, respectively.
The following result is a special case of a standard result from PL topology (see, e.g., Appendix II of [46, Lemma 15a]).
Lemma 5.14
Let \(\mathcal{K}\) be a jdimensional PL manifold whose jsimplices are oriented, and let the continuous map \(f: \mathcal{K}\rightarrow \mathbb{R}^{j}\) be a simplexwise positive map. Then, for any connected open subset W of \(\mathbb{R}^{j}\setminus f(\partial\mathcal{K})\), any two points of W not in \(f(\mathcal{K}^{1})\) are covered the same number of times. If this number is 1, then f, restricted to the open subset W′=f ^{−1}(W) of \(\mathcal{K}\), is injective.
5.5.2 Properties of \(\pi^{*}_{p}\)
Let σ=[p,p _{1},…,p _{ k }] be a ksimplex of \({\rm st}(p, \hat{\mathcal{M}})\). Now, for all points q∈σ, let π _{ p,t }(q)=(1−t)q+tπ _{ p }(q), and σ _{ t }=π _{ p,t }(σ). Since π _{ p } is affine on each simplex, π _{ p,t } is also affine. Therefore σ _{ t } is a simplex with vertices p,p _{1t },…,p _{ kt } with p _{ jt }=π _{ p,t }(p _{ j }), j∈{1,…,k}.
Let τ be an isimplex, and let \(f : \tau\rightarrow \mathcal{M}\) be a C ^{1}function. The simplex τ is said to be C ^{1} embedded by f in \(\mathcal{M}\) if f is an injective mapping and for all x∈τ, the rank of the linear map \(d f(x) : \mathbb{R}^{i} \rightarrow T_{f(x)}\) is i, where T _{ f(x)} is the tangent space to \(\mathcal{M}\) at f(x).
Lemma 5.15
(π C ^{1}embeds σ _{ t })
Assume Condition C1 and C2 of Theorem 5.3. If ε is sufficiently small, then the map π C ^{1}embeds σ _{ t } in \(\mathcal{M}\), for all t∈[0,1].
Proof
By Lemma 4.5 (1) and the definition of σ _{ t }, we have \(\sigma_{t} \subset B(p, O(\varepsilon ) \,{\rm lfs}(p))\). Therefore, for a sufficiently small ε, σ _{ t } cannot intersect the medial axis of \(\mathcal{M}\), which implies, from Lemma 5.2 (1), that the restriction of the map π to σ _{ t } is a C ^{1}function. The rest of the proof is devoted to showing that the restriction of π to σ _{ t } is injective and that dπ is nonsingular.
Equation (25) implies that π restricted to σ _{ t } is injective for ε sufficiently small. Indeed otherwise, there would exist x _{1},x _{2}∈σ _{ t } such that π(x _{1})=π(x _{2}), and the line segment in \({\rm aff}(\sigma_{t})\) joining the points x _{1} and x _{2} would be parallel to \(N_{\pi(x_{1})}\), contradicting Eq. (25).
We deduce also from Eq. (25) and Lemma 5.2 (2) that the derivative dπ is nonsingular when π is restricted to σ _{ t }. This completes the proof of the lemma. □
Lemma 5.16
(π ^{∗} C ^{1}embeds σ _{ t })
For ε sufficiently small, \(\pi^{*}_{p}\) C ^{1}embeds σ _{ t } in T _{ p } for all t∈[0,1].
Proof
Then, by Lemma 5.9, \(\widetilde{\pi }^{1}_{p}\) is a diffeomorphism from \(\widetilde{\pi}_{p}(\widetilde{T}_{p}) \supseteq\widetilde{\mathcal{M}}_{p}\) to \(\widetilde{T}_{p}\), and, by Lemma 5.15, π C ^{1}embeds σ _{ t } into \(\mathcal{M}\). We conclude that \(\pi^{*}_{p} _{\sigma_{t}} = \widetilde{\pi}^{1}_{p} \circ\pi\) C ^{1}embeds σ _{ t } into T _{ p }. □
The proof of the following lemma is similar to Lemma 23a in Chap. 4 from [46] and it will be used for proving that \(\pi ^{*}_{p}\) restricted to open star \(\mathring{{\rm st}}(p, \hat{\mathcal{M}})\) is injective in Lemma 5.18. We give a proof for completeness.
The proof of the following lemma is similar to Lemma 23a in Chapter 4 from [46]. We give a proof for completeness.
Lemma 5.17
Assume Conditions C1 and C2 of Theorem 5.3. For ε sufficiently small, \(\pi^{*}_{p}\) is a simplexwise positive mapping of \({\rm st}(p, \hat {\mathcal{M}})\) into T _{ p }.
Proof
We use the same notation σ _{ t } as in the beginning of this section. We first observe that, since the simplex σ _{1} is in T _{ p }, \(\pi^{*}_{p}\) is the identity on σ _{1}. Therefore, \(\det(J(\pi^{*}_{p}))> 0\) on σ _{1}. Moreover, since from Lemma 5.16, \(\det (J(\pi^{*}_{p})) \neq0\) in σ _{ t } for all t∈[0,1], we also have \(\det(J(\pi^{*}_{p})) > 0\) in σ _{0}. Which proves the lemma. □
We are now ready to prove that \(\pi^{*}_{p}\) restricted to the open star \(\mathring{{\rm st}}(p, \hat{\mathcal{M}})\) is injective.
Lemma 5.18
(Injectivity of \(\pi^{*}_{p}\) restricted to \(\mathring {{\rm st}}(p, \hat{\mathcal{M}})\))
Assume conditions C1 and C2 of Theorem 5.3 and let ε be sufficiently small. For each point \(p \in \mathcal {P}\), the map \(\pi^{*}_{p}\) is injective on the open star \(\mathring{{\rm st}}(p, \hat{\mathcal{M}})\).
Proof
Note that, since \(\hat{\mathcal{M}}\) is a PL manifold without boundary (Lemma 5.5), \({\rm lk}(p, \hat{\mathcal{M}})\) is a PL (k−1)sphere, \({\rm st}(p, \hat{\mathcal{M}})\) is a PL kball, and \(\partial {\rm st}(p, \hat{\mathcal{M}}) = {\rm lk}(p, \hat{\mathcal{M}})\).
For convenience, rewrite \(f=\pi^{*}_{p}\) and \(S={\rm st}(p, \hat{\mathcal{M}})\) and orient S as indicated in Sect. 5.5.1. S is thus an oriented PL kball, and, by Lemma 5.17, \(f=\pi^{*}_{p}\) is a simplexwise positive mapping from S to T _{ p }. Let f(S ^{ k−1}) be the image by f of the (k−1)skeleton of S (i.e. the set of faces of S of dimension at most k−1) and let W be any connected open subset of T _{ p }∖f(∂S). From Lemma 5.14, any two points of W∖f(S ^{ k−1}) are covered the same number of times. If this number is 1, then f, restricted to the open subset f ^{−1}(W) of \({\rm st}(p)\), is injective onto W.
Lemma 5.16 together with Eq. (29) shows that \(\pi^{*}_{p}(x) \neq\pi^{*}_{p}(g)\) for all x∈S∖{g}. Hence, g is not covered by any other point of S, and the lemma follows. □
5.5.3 Properties of π
Lemma 5.19
(Injectivity of the restriction of π to \(\mathring {{\rm st}}(p, \hat{\mathcal{M}})\))
Assume Conditions C1 and C2 of Theorem 5.3. Let ε be sufficiently small. For all p in \(\mathcal {P}\), the map π restricted to \(\mathring{{\rm st}}(p, \hat{\mathcal{M}})\) is injective.
Proof
To reach a contradiction, assume that there exist x _{1},x _{2}(x _{1}≠x _{2}) in \({\rm st}(p) \setminus\partial {\rm st}(p)\) such that π(x _{1})=π(x _{2}). Then \(\pi^{*}_{p}(x_{1}) = \pi^{*}_{p}(x_{2}) = N_{\pi(x_{1})}\cap T_{p}\). Which contradicts the fact that \(\pi^{*}_{p}\) is injective when restricted to \(\mathring{{\rm st}}(p, \hat{\mathcal{M}})\) (Lemma 5.18). □
Lemma 5.20
(Surjectivity of \(\pi_{\hat{\mathcal{M}}}\))
Assume Conditions C1 and C2 of Theorem 5.3. If ε is sufficiently small, then \(\pi_{\hat{\mathcal{M}}}\) is surjective on \(\mathcal{M}\).
Proof
Since \(\hat{\mathcal{M}}\) and \(\mathcal{M}\) are topological kmanifolds without boundaries, and \(\mathring{{\rm st}}(p, \hat{\mathcal{M}})\), which an open set in \(\hat{\mathcal{M}}\) with \(\hat{\mathcal{M}} = \bigcup_{p \in \mathcal {P}} \mathring{{\rm st}}(p, \hat{\mathcal{M}})\), is mapped injectively by π into \(\mathcal{M}\), we get from the Generalized Invariance of Domain Theorem, π is an open map and \(\pi(\hat{\mathcal{M}})\) is an open set in \(\mathcal{M}\). Also, note that since \(\hat{\mathcal{M}}\) is compact and π is a continuous function, \(\pi(\mathcal{M})\) is a compact subset of \(\mathcal{M}\) [10, Thm. 7.6 of Chap. 1]. As \(\mathcal{M}\) is a Hausdorff space, this implies that \(\pi(\hat{\mathcal{M}})\) is a closed set of \(\mathcal{M}\) [10, Thm. 7.5 of Chap. 1].
Now consider a connected component \(\mathcal{M}_{\alpha}\) of \(\mathcal{M}\) with \({\mathcal{V}}_{\alpha} = \pi(\hat{\mathcal{M}}) \cap \mathcal{M}_{\alpha} \neq \emptyset\). From the discussion in the previous paragraph we have \({\mathcal{V}}_{\alpha}\) is both open and closed in \(\mathcal{M}_{\alpha}\). Since \(\mathcal{M}_{\alpha}\) is connected, this implies \({\mathcal{V}}_{\alpha} = \mathcal{M}_{\alpha}\) [10, Def. 4.1 of Chap. 1]. To complete the proof of surjectivity, we only need to show that \(\mathcal {P}\) has nonempty intersection with each connected components of \(\mathcal{M}\), since π(p)=p for all \(\mathcal {P}\). By using the proof technique of Giesen and Wagner [35, Lem. 3], we can show that if \(\mathcal {P}\) is an (ε,δ)sample of \(\mathcal{M}\) with ε<1, then \(\mathcal {P}\) intersects all the connected components of \(\mathcal{M}\). See, [33, Chap. 4] for details. This completes the proof. □
As discussed earlier, π is an open map \(\pi_{\mathring{{\rm st}}(p, \hat{\mathcal{M}})}\) is injective (Lemma 5.19) and \(\mathring{{\rm st}}(p, \hat{\mathcal{M}})\) is an open set in \(\hat{\mathcal{M}}\). The restriction of π to \(\mathring{{\rm st}}(p, \hat{\mathcal{M}})\) is thus a homeomorphism onto its image. In addition, we know from Lemma 5.20 that \(\pi (\hat{\mathcal{M}}) = \mathcal{M}\). Therefore, for all \(x \in \mathcal{M}\), there exist open neighborhoods U _{ x } of x in \(\mathcal{M}\) such that π ^{−1}(U _{ x }) is a disjoint union of open sets, each of which is contained in an open star of a vertex in \(\hat{\mathcal{M}}\) and mapped homeomorphically onto U _{ x } by π. In other words, \((\hat{\mathcal{M}}, \pi)\) is a covering space of \(\mathcal{M}\).
A standard result in topology [38, Lem. 3.4 of Chap. 5] asserts that π covers all the points of a connected component of its image \(\mathcal{M}\) the same number of times. Let us show that this number is exactly one. It cannot be zero by definition. Neither can it be more than one since, otherwise, some sample point p would be covered more than once, which would contradict Lemma A.1 from the Appendix. We have therefore proved that the restriction of π to \(\hat{\mathcal{M}}\) is injective.
Lemma 5.21
(Injectivity of \(\pi_{\hat{\mathcal{M}}}\))
Under the same conditions as in Lemma 5.21, π restricted to \(\hat{\mathcal{M}}\) is injective.
Since \(\hat{\mathcal{M}}\) is compact and \(\mathcal{M}\) is a Hausdorff space, Lemmas 5.20 and 5.21, and a standard result from topology [10, Thm. 7.8 of Chap. 1] implies that \(\pi: \hat{\mathcal{M}} \rightarrow \mathcal{M}\) is a homeomorphism.
This completes the proof of Lemma 5.6.
5.6 Pointwise Approximation
Following lemma is a direct consequence is Lemmas 5.10 and 5.6, and the fact that \({\rm lfs}\) is 1Lipschitz function. For the details of the proof, refer to [33, Chap. 4].
Lemma 5.22
(Pointwise approximation)
Under Conditions C1 and C2 of Theorem 5.3, and ε sufficiently, we have \({\rm dist}(x, \pi_{\hat{\mathcal{M}}}^{1}(x)) = O(\varepsilon ^{2})\,{\rm lfs}(x)\).
5.7 Isotopy
We need the following lemma to show the isotopy Property P5 of Theorem 5.3. The proof of the lemma follows directly from [35, Lem. 5] and one can check the details in [33, Chap. 4].
Lemma 5.23
Using the facts that \(\hat{\mathcal{M}}\) is compact, M _{ t } is a Hausdorff space (as M _{ t } is a subspace with the subspace topology of \(\mathbb{R}^{d}\)), and F _{ t } is continuous, we get [10, Thm. 7.8 of Chap. 1], the following result:
Lemma 5.24
Assume Conditions C1 and C2 of Theorem 5.3. If ε is sufficiently small, then for all t∈[0,1], \(F_{t} : \hat{\mathcal{M}} \rightarrow M_{t}\) is a homeomorphism.
This implies:
Lemma 5.25
(Isotopy)
6 Conclusion
We have given the first algorithm that is able to reconstruct a smooth closed manifold in a time that depends only linearly on the dimension of the ambient space. We believe that our algorithm is of practical interest when the dimension of the manifold is small, even if it is embedded in a space of high dimension. This situation is quite common in practical applications in machine learning. Unlike most surface reconstruction algorithms in \(\mathbb{R}^{3}\), our algorithm does not need to orient normals (a critical issue in practical applications) and, in fact, works for nonorientable manifolds.
The algorithm is simple. The basic ingredients we need are data structures for constructing weighted Delaunay triangulations in \(\mathbb{R}^{k}\). We have assumed that the dimension of \(\mathcal{M}\) is known. If not, we can use algorithms given in [23, 35] to estimate the dimension of \(\mathcal{M}\) and the tangent space at each sample point. Moreover, our algorithm is easy to parallelize. One interesting feature of our approach is that it is robust and still works if we only have approximate tangent spaces at the sample points. We will report on experimental results in a forthcoming paper.
We have assumed that we know an upper bound on the sampling ratio η _{0} of the input sample \(\mathcal {P}\). Ideas from [7, 30] may be useful to convert a sample to a subsample with a bounded sampling ratio.
We foresee other applications of the tangential complex and of our construction each time computations in the tangent space of a manifold are required, e.g. for dimensionality reduction and approximating the Laplace Beltrami operator [11]. It easily follows from [9] that our reconstruction algorithm can also be used in Bregman spaces where the Euclidean distance is replaced by any Bregman divergence, e.g. Kullback–Leibler divergence. This is of particular interest when considering statistical manifolds like, for example, spaces of images [18].
7 Main Notations
General Notations (Sect. 2)
 B(c,r)

is \(\{ x \in \mathbb{R}^{d}  \cx\ < r \}\)
 \(\bar{B}(c, r)\)

is \(\{ x \in \mathbb{R}^{d}  \cx\ \leq r \}\)
 dim(U)

dimension of the affine space U
 \({\rm dist}(p,X)\)

is min_{ x∈X }∥p−x∥
 ∂X

topological boundary of X
 \({\rm int}\, X\)

topological interior of X
 df(x)

derivative of the function f at x
 \(f\mid_{X_{1}}\)

map f restricted to the subset X _{1}
 [xy]

line segment connecting points x and y
 ∥x−y∥

standard Euclidean distance between the points x and y
 ∠(u,v)

angle between the vectors u and v
 ∠(U,V)

angle between the affine spaces U and V
Submanifold (Sect. 2)
 \(B_{\mathcal{M}}(p,r)\)

is \(B(p, r) \cap \mathcal{M}\)
 \(\bar{B}_{\mathcal{M}}(p,r)\)

is \(\bar{B}(p, r) \cap \mathcal{M}\)
 d

dimension of the ambient space
 k

intrinsic dimension of \(\mathcal{M}\)
 \(\mathcal{M}\)

manifold
 \({\mathcal{O}}\)

medial axis of \(\mathcal{M}\)
 \(\mathcal {P}\)

sample
 \(\mathcal {P}\)

cardinality of the sample \(\mathcal {P}\) (Theorem 4.16)
 T _{ x }

tangent space at x
 N _{ x }

normal space at x
Weights (Sect. 2.1.1)
 p ^{ ω }

weighted point (p,ω(p))
 ω

weight assignment
 \(\tilde{\omega}\)

relative amplitude
 ω _{0}

bound on the relative amplitude
Sampling (Sect. 2.2)
Shape measure of simplex τ (Sect. 2.3)
 c _{ τ }

circumcenter
 D _{ τ }(p)

altitude (Sect. 4.1)
 \({\rm aff}(\tau)\)

affine hull of τ
 m _{ u }(τ) or m _{ u }

see Sect. 2.4
 i _{ ϕ }

see Sect. 2.4
 N _{ τ }

affine space orthogonal to \({\rm aff}(\tau)\)
 Δ _{ τ }

longest edge length (diameter)
 H _{ τ }(p,ω(p))

eccentricity (Sect. 4.1)
 L _{ τ }

shortestedge length
 o _{ τ }

orthocenter
 Φ _{ τ }

orthoradius
 R _{ τ }

circumradius
 \(\varGamma_{\tau} = \frac{\varDelta_{\tau}}{L_{\tau}}\)

longest edge to shortestedge ratio
 τ _{ p }=τ∖{p}

face of τ opposite to p (Sect. 4.1)
 Θ _{ τ }

fatness
 Θ _{0}

fatness bound
 V _{ τ }

volume
Complexes (Sect. 2.4)
 \({\mathcal{K}}\)

simplicial complex (Sect. 2)
 \({\rm st}(\sigma, {\mathcal{K}})\)

\(\{ \tau\in{\mathcal{K}} : \mbox{for some}~\tau_{1} \in \mathcal {K}, \sigma, \tau\subseteq \tau_{1} \}\) (Sect. 2)
 \({\rm lk}(\sigma, {\mathcal{K}})\)

\(\{ \tau\in{\rm st}(\sigma , \mathcal {K}) : \tau\cap\sigma= \emptyset\}\) (Sect. 2)
 \(\mathring{{\rm st}}(\sigma, {\mathcal{K}})\)

\({\rm st}(\sigma , {\mathcal{K}})\setminus {\rm lk}(\sigma, {\mathcal{K}})\) (Sect. 2)
 \({\rm st}(p)\)

\({\rm st}(p, {{\rm Del}^{\omega}(\mathcal {P}_{p})})\)
 \(C^{\omega}(\mathcal {P})\)

Complete complex
 \({\rm Del}^{\omega}(\mathcal {P})\)

Weighted Delaunay triangulation (Sect. 2.1.2)
 \({\rm Del}^{\omega}_{p} (\mathcal {P})\)

restriction of \({\rm Del}^{\omega}(\mathcal {P})\) to T _{ p }
 \({\rm Del}^{\omega}_{T\mathcal{M}}(\mathcal {P})\)

Tangential Delaunay complex
 \({\rm Inc}^{\omega}(\mathcal {P})\)

contains Θ _{0}inconsistent configurations of \({\rm Del}^{\omega}_{T\mathcal{M}}(\mathcal {P})\)
Footnotes
Notes
Acknowledgements
The authors thank Mariette Yvinec and Ramsay Dyer for their careful reading of earlier drafts and insightful comments. The authors also gratefully acknowledge the work of the reviewers and thank them for their detailed comments that helped to improve the manuscript.
This research has been partially supported by the Agence Nationale de la Recherche (project GAIA 07BLAN032804) and by the Future and Emerging Technologies (FET) programme within the Seventh Framework Programme for Research of the European Commission, under FETOpen grant number: 255827 (Project CGLearning).
References
 1.Amenta, N., Bern, M.: Surface reconstruction by Voronoi filtering. Discrete Comput. Geom. 22, 481–504 (1999) CrossRefzbMATHMathSciNetGoogle Scholar
 2.Amenta, N., Choi, S., Dey, T.K., Leekha, N.: A simple algorithm for homeomorphic surface reconstruction. Int. J. Comput. Geom. Appl. 12, 125–141 (2002) CrossRefzbMATHMathSciNetGoogle Scholar
 3.Aurenhammer, F., Edelsbrunner, H.: An optimal algorithm for constructing the weighted Voronoi diagram in the plane. Pattern Recognit. 17(2), 251–257 (1984) CrossRefzbMATHMathSciNetGoogle Scholar
 4.Aurenhammer, F.: Power diagrams: properties, algorithms and applications. SIAM J. Comput. 16(1), 78–96 (1987) CrossRefzbMATHMathSciNetGoogle Scholar
 5.Boissonnat, J.D., Dyer, R., Ghosh, A.: The stability of Delaunay triangulations. Int. J. Comput. Geom. Appl. (special issue for SoCG 2012) (to appear) Google Scholar
 6.Boissonnat, J.D., Flötotto, J.: A coordinate system associated with points scattered on a surface. Comput. Aided Des. 36, 161–174 (2004) CrossRefGoogle Scholar
 7.Boissonnat, J.D., Guibas, L.J., Oudot, S.Y.: Manifold reconstruction in arbitary dimensions using witness complexes. Discrete Comput. Geom. 42(1), 37–70 (2009) CrossRefzbMATHMathSciNetGoogle Scholar
 8.Belkin, M., Niyogi, P.: Laplacian eigenmaps and spectral techniques for embedding and clustering. In: Advances in Neural Information Processing Systems, vol. 14, pp. 585–591 (2002) Google Scholar
 9.Boissonnat, J.D., Nielsen, F., Nock, R.: Bregman Voronoi Diagrams. Discrete Comput. Geom. 44(2), 281–307 (2010) CrossRefzbMATHMathSciNetGoogle Scholar
 10.Bredon, G.E.: Topology and Geometry. Graduate Text in Mathematics. Springer, Berlin (1994) Google Scholar
 11.Belkin, M., Sun, J., Wang, Y.: Discrete Laplace operator on meshed surfaces. In: Proc. ACM Symp. on Computational Geometry, pp. 278–287 (2008) Google Scholar
 12.Boissonnat, J.D., Wormser, C., Yvinec, M.: Locally uniform anisotropic meshing. In: Proc. ACM Symp. on Computational Geometry, pp. 270–277 (2008) Google Scholar
 13.Boissonnat, J.D., Yvinec, M.: Algorithmic Geometry. Cambridge University Press, Cambridge (1998) CrossRefzbMATHGoogle Scholar
 14.Cheng, S.W., Dey, T.K., Edelsbrunner, H., Facello, M.A., Teng, S.H.: Sliver Exudation. J. ACM 47, 883–904 (2000) CrossRefMathSciNetGoogle Scholar
 15.Cheng, S.W., Dey, T.K., Ramos, E.A.: Manifold Reconstruction from Point Samples. In: Proc. ACMSIAM Symp. Discrete Algorithms, pp. 1018–1027 (2005) Google Scholar
 16.Cazals, F., Giesen, J.: Delaunay triangulation based surface reconstruction. In: Boissonnat, J.D., Teillaud, M. (eds.) Effective Computational Geometry for Curve and Surfaces. Springer, Berlin (2006) Google Scholar
 17.Chazelle, B.: An optimal convex hull algorithm in any fixed dimension. Discrete Comput. Geom. 10, 377–409 (1993) CrossRefzbMATHMathSciNetGoogle Scholar
 18.Carlsson, G., Ishkhanov, T., de Silva, V., Zomorodian, A.: On the local behavior of spaces of natural images. Int. J. Comput. Vis. 76(1), 1–12 (2008) CrossRefGoogle Scholar
 19.Chazal, F., Lieutier, A.: Smooth manifold reconstruction from noisy and nonuniform approximation with guarantees. Comput. Geom. 40, 156–170 (2008) CrossRefzbMATHMathSciNetGoogle Scholar
 20.Chazal, F., Oudot, S.Y.: Towards persistencebased reconstruction in Euclidean spaces. In: Proc. ACM Symp. on Computational Geometry, pp. 232–241 (2008) Google Scholar
 21.Clarkson, K.L., Shor, P.W.: Applications of random sampling in computational geometry. II. Discrete Comput. Geom. 4, 387–421 (1989) CrossRefzbMATHMathSciNetGoogle Scholar
 22.CohenSteiner, D., Da, T.K.F.: A greedy Delaunay based surface reconstruction algorithm. Vis. Comput. 20, 4–16 (2004) CrossRefGoogle Scholar
 23.Cheng, S.W., Wang, Y., Wu, Z.: Provable dimension detection using principle component analysis. Int. J. Comput. Geom. Appl. 18(5), 415–440 (2008) CrossRefzbMATHMathSciNetGoogle Scholar
 24.Delaunay, B.: Sur la sphère vide. Izv. Akad. Nauk SSSR, Otdelenie Matematicheskii i Estestvennyka Nauk 7, 793–800 (1934) Google Scholar
 25.Dey, T.K.: Curve and Surface Reconstruction: Algorithms with Mathematical Analysis. Cambridge University Press, Cambridge (2006) CrossRefGoogle Scholar
 26.Donohu, D.L., Grimes, C.: Hessian eigenmaps: new locally linear embedding techniques for high dimensional data. Proc. Acad. Nat. Sci. 100, 5591–5596 (2003) CrossRefGoogle Scholar
 27.Federer, H.: Curvature measures. Trans. Am. Math. Soc. 93(3), 418–491 (1959) CrossRefzbMATHMathSciNetGoogle Scholar
 28.Federer, H.: Geometric Measure Theory. Springer, New York (1969) zbMATHGoogle Scholar
 29.Flötotto, J.: A coordinate system associated to a point cloud issued from a manifold: definition, properties and applications. Ph.D. thesis, Université of Nice SophiaAntipolis (2003) Google Scholar
 30.Funke, S., Ramos, E.: Smoothsurface reconstruction in nearlinear time. In: Proc. ACMSIAM Symp. Discrete Algorithms, pp. 781–790 (2002) Google Scholar
 31.Freedman, D.: Efficient simplicial reconstructions of manifolds from their samples. IEEE Trans. Pattern Anal. Mach. Intell. 24(10) (2002) Google Scholar
 32.Fu, J.H.G.: Convergence of curvature in secant approximations. J. Differ. Geom. 37, 117–190 (1993) Google Scholar
 33.Ghosh, A.: Piecewise linear reconstruction and meshing of submanifolds of Euclidean space. Ph.D. thesis, INRIA Sophia Antipolis & Université of Nice SophiaAntipolis (2012) Google Scholar
 34.Gopi, M., Khrisnan, S., Silva, C.T.: Surface reconstruction based on lower dimensional localized Delaunay triangulation. In: Proc. Eurographics, pp. 363–371 (2000) Google Scholar
 35.Giesen, J., Wagner, U.: Shape dimension and intrinsic metric from samples of manifolds. Discrete Comput. Geom. 32(2), 245–267 (2004) CrossRefzbMATHMathSciNetGoogle Scholar
 36.Li, X.Y.: Generating wellshaped ddimensional Delaunay meshes. Theor. Comput. Sci. 296(1), 145–165 (2003) CrossRefzbMATHGoogle Scholar
 37.Lafon, S., Lee, A.B.: Diffusion maps and coarsegraining: a unified framework for dimensionality reduction, graph partitioning, and data set parameterization. IEEE Trans. Pattern Anal. Mach. Intell. 28, 1393–1403 (2006) CrossRefGoogle Scholar
 38.Massey, W.S.: Algebraic Topology: An Introduction. Graduate Texts in Mathematics, vol. 56. Springer, Berlin (1967) zbMATHGoogle Scholar
 39.Nadler, B., Lafon, S., Coifman, R.R., Kevrekidis, I.G.: Diffusion maps, spectral clustering and eigenfunctions of FokkerPlanck operators. In: Advances in Neural Information Processing Systems, vol. 18 (2005) Google Scholar
 40.Niyogi, P., Smale, S., Weinberger, S.: Finding the homology of submanifolds with high confidence from random samples. Discrete Comput. Geom. 39(1), 419–441 (2008) CrossRefzbMATHMathSciNetGoogle Scholar
 41.Rourke, C.P., Sanderson, B.J.: Introduction to PiecewiseLinear Topology. Springer, Berlin (1972) CrossRefzbMATHGoogle Scholar
 42.Roweis, S.T., Saul, L.K.: Nonlinear dimensionality reduction by locally linear embedding. Science 290, 2323–2326 (2000) CrossRefGoogle Scholar
 43.Shewchuk, J.: Star splaying: an algorithm for repairing Delaunay triangulations and convex hulls. In: Proc. ACM Symp. on Computational Geometry, pp. 237–246 (2005) Google Scholar
 44.Seung, H.S., Lee, D.D.: The manifold ways of perception. Science 290, 2268–2269 (2000) CrossRefGoogle Scholar
 45.Tenenbaum, J.B., de Silva, V., Langford, J.C.: A global geometric framework for nonlinear dimensionality reduction. Science 290, 2319–2323 (2000) CrossRefGoogle Scholar
 46.Whitney, H.: Geometric Integration Theory. Princeton University Press, Princeton (1957) zbMATHGoogle Scholar
 47.Zeeman, E.C.: In: Seminar on Combinatorial Topology. Institut des Hautes Études Scientifiques (Paris) and University of Warwick (Coventry), Notes, pp. 1963–1966 Google Scholar
 48.Zhang, Z., Zha, H.: Principal manifolds and nonlinear dimension reduction via local tangent space alignment. SIAM J. Sci. Comput. 26(1), 313–338 (2004) CrossRefzbMATHMathSciNetGoogle Scholar