Bardia Sadri

How fast is the k -means method?

Abstract We present polynomial upper and lower bounds on the number of iterations performed by the k-means method (a.k.a. Lloyd’s method) for k-means clustering. Our upper bounds are polynomial in the number of points, number of clusters, and the spread of the point set. We also present a lower bound, showing that in the worst case the k-means heuristic needs to perform Ω(n) iterations, for n points on the real line and two centers. Surprisingly, the spread of the point set in this construction is polynomial. This is the first construction showing that the k-means heuristic requires more than a polylogarithmic number of iterations. Furthermore, we present two alternative algorithms, with guaranteed performance, which are simple variants of the k-means method. Results of our experimental studies on these algorithms are also presented.

Critical points of the distance to an epsilon-sampling of a surface and flow-complex-based surface reconstruction

The distance function to surfaces in three dimensions plays a key role in many geometric modeling applications such as medial axis approximations, surface reconstructions, offset computations, feature extractions and others. In most cases, the distance function induced by the surface is approximated by a discrete distance function induced by a discrete sample of the surface. The critical points of the distance function determine the topology of the set inducing the function. However, no earlier theoretical result has linked the critical points of the distance to a sampling of geometric structures to their topological properties. We provide this link by showing that the critical points of the distance function induced by a discrete sample of a surface either lie very close to the surface or near its medial axis and this closeness is quantified with the sampling density. Based on this result, we provide a new flow-complex-based surface reconstruction algorithm that, given a tight ε-sampling of a surface, approximates the surface geometrically, both in Hausdorff distance and normals, and captures its topology.

Medial axis approximation and unstable flow complex

The medial axis of a shape is known to carry a lot of information about it. In particular a recent result of Lieutier establishes that every bounded open subset of Rn has the same homotopy type as its medial axis. In this paper we provide an algorithm that, given a sufficiently dense but not necessarily uniform sample from the surface of a shape with smooth boundary, computes a core for its medial axis approximation, in form of a piecewise linear cell complex, that captures the topology of the medial axis of the shape. We also provide a natural method to freely augment this core in order to enhance it geometrically all the while maintaining its topological guarantees. The definition of the core and its extension method are based on the steepest ascent flow induced by the distance function to the sample. We also provide a geometric guarantee on the closeness of the core and the actual medial axis.

Flow-complex-based shape reconstruction from 3D curves

We address the problem of shape reconstruction from a sparse unorganized collection of 3D curves, typically generated by increasingly popular 3D curve sketching applications. Experimentally, we observe that human understanding of shape from connected 3D curves is largely consistent, and informed by both topological connectivity and geometry of the curves. We thus employ the flow complex, a structure that captures aspects of input topology and geometry, in a novel algorithm to produce an intersection-free 3D triangulated shape that interpolates the input 3D curves. Our approach is able to triangulate highly nonplanar and concave curve cycles, providing a robust 3D mesh and parametric embedding for challenging 3D curve input. Our evaluation is fourfold: we show our algorithm to match designer-selected curve cycles for surfacing; we produce user-acceptable shapes for a wide range of curve inputs; we show our approach to be predictable and robust to curve addition and deletion; we compare our results to prior art.

Lipschitz unimodal and isotonic regression on paths and trees

We describe algorithms for finding the regression of t, a sequence of values, to the closest sequence s by mean squared error, so that s is always increasing (isotonicity) and so the values of two consecutive points do not increase by too much (Lipschitz). The isotonicity constraint can be replaced with a unimodular constraint, for exactly one local maximum in s. These algorithm are generalized from sequences of values to trees of values. For each we describe near-linear time algorithms.

Critical Points of Distance to an Epsilon-sampling of a Surface and Flow-complex-based Surface Reconstruction

The distance function to surfaces in three dimensions plays a key role in many geometric modeling applications such as medial axis approximations, surface reconstructions, offset computations and feature extractions among others. In many cases, the distance function induced by the surface can be approximated by the distance function induced by a discrete sample of the surface. The critical points of the distance functions are known to be closely related to the topology of the sets inducing them. However, no earlier theoretical result has found a link between topological properties of a geometric object and critical points of the distance to a discrete sample of it. We provide this link by showing that the critical points of the distance function induced by a discrete sample of a surface fall into two disjoint classes: those that lie very close to the surface and those that are near its medial axis. This closeness is precisely quantified and is shown to depend on the sampling density. It turns out that critical points near the medial axis can be used to extract topological information about the sampled surface. Based on this, we provide a new flow-complex-based surface reconstruction algorithm that, given a tight e-sample of a surface, approximates the surface geometrically, both in distance and normals, and captures its topology. Furthermore, we show that the same algorithm can be used for curve reconstruction.

Manifold homotopy via the flow complex

It is known that the critical points of the distance function induced by a dense sample P of a submanifold Σ of ℝn are distributed into two groups, one lying close to Σ itself, called the shallow, and the other close to medial axis of Σ, called deep critical points. We prove that under (uniform) sampling assumption, the union of stable manifolds of the shallow critical points have the same homotopy type as Σ itself and the union of the stable manifolds of the deep critical points have the homotopy type of the complement of Σ. The separation of critical points under uniform sampling entails a separation in terms of distance of critical points to the sample. This means that if a given sample is dense enough with respect to two or more submanifolds of ℝn, the homotopy types of all such submanifolds together with those of their complements are captured as unions of stable manifolds of shallow versus those of deep critical points, in a filtration of the flow complex based on the distance of critical points to the sample. This results in an algorithm for homotopic manifold reconstruction when the target dimension is unknown.

MEDIAL AXIS APPROXIMATION AND UNSTABLE FLOW COMPLEX

The medial axis of a shape is known to carry a lot of information about the shape. In particular, a recent result of Lieutier establishes that every bounded open subset of ℝn has the same homotopy type as its medial axis. In this paper we provide an algorithm that computes a structure we call the core for the approximation of the medial axis of a shape with smooth boundary from a discrete sample of its boundary. The core is a piecewise linear cell complex that is guaranteed to capture the topology of the medial axis of the shape provided the sample of its boundary is sufficiently dense but not necessarily uniform. We also present a natural method for augmenting the core in order to extend it geometrically while maintaining the topological guarantees. The definition of the core and its extension are based on the steepest ascent flow map that results from the distance function induced by the sample point set. We also provide a geometric guarantee on the closeness of the core and the actual medial axis.