Combinatorial 2D Vector Field Topology Extraction and Simplification

Transcription

1 Combinatorial 2D Vector Field Topology Extraction and Simplification Jan Reininghaus, Ingrid Hotz 2 Zuse Institute Berlin (ZIB), Germany. reininghaus@zib.de 2 Zuse Institute Berlin (ZIB), Germany. hotz@zib.de Summary: This paper investigates a combinatorial approach to vector field topology. The theoretical basis is given by Robin Forman s work on a combinatorial Morse theory for dynamical systems defined on general simplicial complexes. We formulate Forman s theory in a graph theoretic setting and provide a simple algorithm for the construction and topological simplification of combinatorial vector fields on 2D manifolds. Given a combinatorial vector field we are able to extract its topological skeleton including all periodic orbits. Due to the solid theoretical foundation we know that the resulting structure is always topologically consistent. We explore the applicability and limitations of this combinatorial approach with several examples and determine its robustness with respect to noise. Introduction Topological methods have developed into an integral part in data analysis and visualization, for scalar as well as for vector valued data. While the results of scalar and vector field topology coincide when dealing with gradient vector fields, the respective mathematical approach originated in different fields. Vector field topology is usually motivated and defined as a clustering of streamlines of the flow generated by the vector field. This clustering has a theoretical sound foundation in dynamical systems theory. One first defines limit sets of the streamlines. Features of the vector field are then defined as intersections of these limit sets. The extracted structure of the vector field is therefore connected to the equivalence classes generated by the flow and is called topology of the vector field. The extraction of the topological skeleton builds on a continuously defined vector field using numerical methods. Scalar field topology on the other hand is usually defined through Morse theory. Given a non degenerate function f : M R this theory defines a structure M, called Morse complex. This structure consists of minima, maxima, saddle points and their connectivity. A main result of Morse theory is that M is isomorphic to the singular homology of the domain of the function M. As the singular homology is a rather fine grained topological invariant of the domain M, this structure is also called the topology of the scalar function f. In general the topology is extracted using a combinatorial approach building on the data given in discrete points.

2 In this paper we approach vector field topology from a Morse theoretic viewpoint. Forman describes a discrete Morse theory for general vector fields in [4], our theoretical foundation. Lewiner has successfully applied Forman s discrete Morse theory for gradient vector fields [5] in [] which motivated our investigation. Forman s definitions lead to a structure M that contains all critical points, their connectivity and all closed streamlines. If M contains no closed streamlines, then this structure is isomorphic to the singular homology of the domain M. Morse theoretic vector field topology therefore degenerates to scalar field topology when dealing with gradient vector fields. Should M contain closed streamlines, then there are still many topological invariants of M that are contained in M, e.g. there are the strong Morse inequalities which guarantee that the topological complexity of M is bounded by the complexity of M. The proof of these inequalities provides a certain inherent consistency to this vector field topology. For example, a vector field on a sphere always contains at least one critical point, a vector field on a torus with no critical points contains at least two periodic orbits. These kinds of topological constraints are always reflected in M. Note that this consistency is preserved in practice, as we can compute M exactly because of the combinatorial nature of the definitions. 2 Related Work The use of topological methods for scalar and vector field visualization developed almost independently over the last two decades. Both areas build on solid mathematical foundations. While methods for vector fields mostly refer back to Poincaré Index Theory, topological scalar field analysis is based on Morse Theory. For a basic overview over these theories and their relation we refer to [3]. Vector field topology was introduced to visualization by Helman and Hesselink [8]. They defined the concept of a topological skeleton consisting of critical points and connecting separatrices to segment the field into regions of topologically equivalent streamline behavior. An algorithm to extract periodic orbits, completing the topological structure, was first proposed by Wischgoll et al. [9] based on the analysis of cell cycles. Later Theisel et al. [5] suggested an algorithm to detect isolated closed streamlines in 2D vector fields by intersecting certain stream surfaces in 3D. Following this work the use of topological methods has been further advanced according to many aspects like topology tracking, extraction of boundary topology and extensions to 3D. In the following we only point out some specific publications dealing with topological simplification, which is also a central question in our work. For a more complete overview we refer to the survey paper [] and the references therein. A major obstacle in the application of vector field topology is the high feature density in complex data sets. Therefore much attention has been payed to scaling and simplification of the topological skeleton. Solutions have been proposed by

3 Tricoche et al. following two different strategies. A scaling approach determines groups of close singularities according to a given measure. These are replaced consistently by fewer structures consisting of higher order singularities [7]. The main target of this method is the removal of visual clutter - the structural influence of singularities is ignored allowing for possibly important flow features to be removed. The second approach focuses on pair annihilations leading to a progressive simplification of the topological graph. The pairs are sorted according to a relevance measure, e.g. the Euclidean distance of the critical points [6]. In both cases the vector field is changed to fit the simplified structure. Similar to the scaling approach Weinkauf et al. [8] cluster close critical points in 3D fields and replace them by one critical point of higher order represented by a specific icon. Klein et al. [9] apply a scale space technique to track critical points over multiple spatial scales in order to assess the importance of a critical point to the overall behavior of the underlying flow field using Gaussian filter kernels. All this work is based on numerical analysis of continuous vector fields resulting from interpolated data values given at single vertices. Scalar field topology developed mostly independently from vector field topology. The main application areas in visualization include segmentation, transfer function design, and ridge extraction. A variety of algorithms for its extraction have been introduced, some using the corresponding gradient vector field, others suggesting combinatorial approaches using Morse-Smale theory. The Morse- Smale complex is a topological structure partitioning the domain into regions of uniform gradient flow. We would like to point out a few publications here, which motivated our work. Edelsbrunner et al. [3] define Morse-Smale complexes for piecewise linear data resulting in a combinatorial algorithm for its extraction. To reduce the often very complex structure a controlled simplification is suggested based on the general concept of persistence introduced in [2]. These ideas have been applied for visualization of 3D data in [7]. Forman has chosen a slightly different approach by developing a discrete Morse theory [6] for scalar fields defined on cell complexes. Rather than choosing a suitable class of continuous functions on the space, he assigns a single number to each cell of the complex and all further steps are combinatorial. Lewiner et al. [2] has applied this theory successfully to scalar fields on triangulated manifolds. Motivated by the success of these combinatorial algorithms for scalar field topology, recently first steps have been taken towards a combinatorial vector field topology. Chen [] et al. have developed a combinatorial topology extraction method based on Conley Index Theory. Contrary to our approach they however make use of continuous numerical methods for the subsequent simplification.

4 Fig.. Basic definitions, from left to right: a) simplicial graph of a single triangle; b) combinatorial vector field (dashed); c) critical point of index 2 (red); d) attracting periodic orbit (red) 3 Basic Ideas and Definitions In this section we present Forman s combinatorial Morse theory for vector fields [4] in graph theoretic terms. For simplicity we restrict ourselves to triangulated 2D manifolds while Forman s theory is defined in a far more general setting. We refer to graph edges as links to avoid confusion with the edges of the triangulation. Given a triangulation of a manifold M we first define its simplicial graph (see Figure a). The nodes of the graph consist of the vertices, edges and triangles of the triangulation and each node α p is labeled with the dimension p of the geometric simplex it represents. The links of the graph encode the neighborhood relation of the triangulation: If the simplex represented by the node α p is in the boundary of the simplex represented by the node β p+ then {α p, β p+ } is a link in the graph. Note that only simplices whose dimension differs by one are linkable. A matching of a graph is defined as a subset of links such that no two links are adjacent. We are now ready for the central definition of this paper: A combinatorial vector field V is a matching of a simplicial graph (see Figure b). The nodes of the graph that are not covered by V are called critical points (see Figure c). If α p is a critical point of V we say that α p has index p. A critical point of index p is called sink (p = ), saddle (p = ) or source (p = 2). Because of the matching property non critical nodes always appear in pairs whose associated dimension p differs by exactly one. Therefore one can immediately conclude the combinatorial Poincaré-Hopf formula (see [4]) χ(m) = = 2 ( ) p {number of p-simplices} p= 2 ( ) p {number of critical points with index p}, p= () which already hints at the inherent topological consistency provided by Forman s combinatorial approach.

5 A combinatorial p-streamline is a path in the graph whose links alternate between V and the complement of V and the dimension of the nodes of the path alternates between p and p +. A p-streamline connecting two critical points is called a separatrix. If a p-streamline is closed we call it either an attracting periodic orbit (p = ) (see Figure d) or a repelling periodic orbit (p = ). When a combinatorial vector field contains no periodic orbits we call it a discrete gradient vector field (see Lewiner). As the relation of the critical points to the topology of the manifold is quite hard to state in general combinatorial vector fields, we restrict ourselves to combinatorial Morse-Smale vector fields, i.e. combinatorial vector fields whose periodic orbits are pairwise disjoint. Note that combinatorial vector fields of simplicial graphs of 2D manifolds are always of type Morse-Smale due to the simple graph structure: Each -simplex has exactly two -simplices in its boundary and is itself in the boundary of at most two 2-simplices. Therefore p-streamlines can merge but not split, which implies that periodic orbits are pairwise disjoint. Let c p denote the number of critical points with index p, A p the number of closed p-streamlines and b p the p-th Betti number of M. Forman has proven in [4] that for any k there holds the strong Morse inequality A k + c k c k +... ± c b k b k +... ± b. (2) This result shows that the critical points and periodic orbits carry the topology of the triangulated manifold. As the topological features (critical points, separatrices and periodic orbits) can be computed exactly in a combinatorial vector field due to the finite nature of the definitions we always get a skeleton which is consistent with the topology of the underlying manifold. Note that the separatrices do not take part in (2). However, there is an intricate topological invariant, called Reidemeister torsion, whose computation involves the separatrices (see [4]). Therefore we consider the separatrices to be part of the topological skeleton of a vector field. 4 Algorithm In this section we present the algorithm we developed to construct a combinatorial vector field and extract its topological skeleton. Furthermore we show how the same algorithm that is used for the construction of the field can be employed to consistently simplify its topological skeleton. The input of the algorithmic pipeline is assumed to consist of a triangulation with vectors given on the vertices (see Figure 2). We begin by constructing the simplicial graph G = (S, L) (see Section 3) of the given triangulation. We add all simplices (vertices, edges and triangles) to the graph and then encode the connectivity information by adding the corresponding links to the graph (see Figure a). To represent the given vector field data we compute link weights w : L R in the simplicial graph. Let c : S R 3 denote the coordinates of the midpoints of the simplices represented in the graph. To calculate the auxiliary data values

6 Fig.2. Algorithmic pipeline, from left to right: a) input triangulation with vectors on vertices; b) link weighted simplicial graph; c) computed combinatorial vector field (red) with a sink in the center (blue) f : S R 3 on the nodes of the graph we average the data values given on the vertices of the triangulation. The weight of a link e = {n p+, n p 2 } L is then given by w(e) = ( ) ( ) f(n p+ ) + f (n p 2 2 ) c(n p+ ) c(n p 2 ), (3) where the dot denotes the scalar product. This term corresponds to the tangential component of the given data value scaled with the geometric length of the link. Note that the link weighted graph contains enough information to reconstruct the tangential vector field induced by the given data. Let M denote the set of all matchings of the simplicial graph G and let w(m) = e M w(e) denote the weight of the matching M. Then a combinatorial vector field V representing the given data can be constructed by computing V = arg max M M w(m), (4) i.e. by finding the heaviest matching of the link weighted simplicial graph. Note that the graph G = (S, L) is bipartite, as the links only connect odd with even dimensional simplices. This greatly simplifies the algorithm needed to solve (4). The graph problem (4) has been thoroughly studied, usually called maximum weighted bipartite matching. For this work we chose an exact approach based on the Hungarian method with a computational complexity of O( S 2 log( S )). This approach iteratively calculates the sequence of matchings M k := arg max w(m). (5) M M, M =k As this sequence does not contain several isolated maxima we can stop the computation once w(m k+ ) < w(m k ) and return the combinatorial vector field V := M k satisfying (4). For more details, a correctness proof and the computational complexity analysis we refer to [4]. Let s(m k ) S denote the critical points of the combinatorial vector field M k. Then s(m k+ ) s(m k ) and s(m k ) s(m k+ ) = 2. The vector fields M k+l

7 can therefore be interpreted as topologically simplified versions of the vector field V = M k. Given a combinatorial vector field we can easily extract its topological skeleton, i.e. its critical points, separatrices and periodic orbits. To find the critical points of the vector field we only have to find the nodes of the graph that are not covered by the matching. For the classification into sinks, sources and saddles we just have to look at the label p. Overall all critical points can be found and classified in O( S ). To extract all periodic orbits we make use of a depth first search through the graph with the side constraints of a p-streamline. Because p-streamlines of simplicial graphs of 2D manifolds can merge but not split (see section 3) we can extract all periodic orbits in O( S ). The classification into attracting and repelling orbits is again given by the label p of the participating nodes. The separatrices can be found analogously with the constrained depth first search mentioned above with a computational complexity O( S ) for the extraction of all separatrices. 5 Results The goal of this section is to explore the properties of the combinatorial vector fields constructed in Section 4. The robustness of the algorithm when dealing with noisy data is analyzed in the first example. The second example verifies the topological consistency guaranteed in Section 3. Finally we present a real world data set to examine the potential of this method for multi-scale topological vector field analysis. Example The purpose of this data set is to show the connection of our simplification algorithm with the concept of persistency from scalar field topology. To do this, we analyzed the gradient field of the scalar function f(x, y) = sin(x)sin(y)exp 3(x2 +y 2 ) on the domain [, ] 2 (see Figure 3 top-left) discretized with a uniform triangulation consisting of 6, vertices. To determine the robustness of our algorithm with respect to noise we added random numbers in the range of [.,.] to the scalar values of the mesh before taking the gradient (see Figure 3 top-right). This results in a perturbed gradient vector field containing many critical points (see Figure 3 bottom-left). Our topological simplification behaves like a persistence based approach: the extrema of the gradient field are generally canceled in the order given by the absolute value of the difference of the corresponding scalar values (see Figure 3 bottom-right). The data set is designed such that this order is reflected in the distance to the origin. Our simplification algorithm recovers the topology of the original unperturbed field. The presence of noise is indicated by the graph shown in Figure 4. The x-axis represents the simplification threshold, i.e. we stop the simplification process when w(m k+l ) w(m k+l+ ) is larger than the given threshold.

8 Fig. 3. Top-left: height field of unperturbed data set; top-right: height field of perturbed data set with simplified topological skeleton; bottom-left: sources (red), sinks (blue), and saddles (yellow) of initial combinatorial vector field on planar LIC of the gradient vector field; bottom-right: critical points of simplified topological skeleton - the blue circles represent tiny attracting periodic orbits critical points 4 2 original data perturbed data simplification threshold Fig.4. Simplification graph of example. The sharp drop of the red graph at. indicates the presence of noise in the range of [.,.]

9 The y-axis represents the number of the critical points of the corresponding combinatorial vector field. The sharp drop in the red graph at a simplification threshold of. indicates the presence of noise in the range of [.,.]. The symmetry of the unperturbed data f can be seen in the blue graph as there are eight cancellations with the same simplification threshold of about.47. Note that the combinatorial vector fields computed in this example are not always discrete gradient vector fields, i.e. they may contain periodic orbits. The gradient field property is not representable by the point set of vectors that is given for the construction of the vector field. However, most of the time when we sample a gradient vector field on the nodes of a triangulation and compute the combinatorial vector field, we get a discrete gradient vector field containing no periodic orbits. Fig.5. Random vector field on a torus with three levels of topological simplification l sinks saddles sources repelling orbits attracting orbits Table. Topological signature of a sequence of combinatorial vector fields M k+l on the torus, see Example 2. The strong Morse inequalities (2) hold in each case. Example 2 This data set consists of a random vector field defined on a torus. We created a random vector field on a grid and then sampled the trilinear interpolant on the triangulation of a torus with 6, vertices (see Figure 5). As we know the Betti numbers of a torus (b =, b = 2, b 2 = ) we can verify that the initial and all simplified topological skeletons satisfy the strong Morse inequalities (2) as expected (see Table ).

10 Fig. 6. Top-left: topological skeleton of a vector field on a cerebral aneurysm given by a blood flow simulation; top-right: zoom in of an area exhibiting recurrent flow behavior indicated by attracting periodic orbits (red) and repelling periodic orbits (blue); bottom-left: simplified topological skeleton with saddle, 2 sinks and source; bottom-right: simplified topological skeleton with 2 sinks and repelling orbit Example 3 This real world example depicts the surface velocity field of a simulation of blood flow through a cerebral aneurysm done by the Biofluid Mechanics Lab of the Charité - Universitätsmedizin Berlin. The simplicial graph of the triangulation consists of 6, nodes. The computation time for the construction, simplification and topology extraction is 29 minutes on a 3 GHz CPU with our current implementation. Almost all time is spent in the maximum weighted bipartite matching code that solves (4). The critical points in this vector field are stagnation points and thus of interest for the flow analysis. While standard vector field topology is able to reliably extract these critical points too, our algorithm delivers a hierarchy of topological skeletons which captures the dominant nature of the flow (see Figure 6 bottomleft). The blood enters the aneurysm at the bottom, and leaves it horizontally. This behavior is found by our algorithm and the global separation on the surface is extracted. This reduced flow structure may serve as a basis when comparing different cerebral aneurysms.

11 6 Discussion and Future Work This first approach to combinatorial vector field topology based on Forman s discrete Morse theory looks promising. The extracted topological skeleton is always consistent, the multi scale applicability has been demonstrated, and our algorithm is able to reliably deal with noisy data (see Section 5). We have identified three areas where the algorithm could be further improved: The quadratic runtime stemming from the maximum weighted matching problem (4) reduces the applicability of our algorithm to real world data sets. A greedy approximation algorithm for (4) should alleviate this problem. Note that the consistency of the topological skeleton would not be affected by such an approximation as it is induced by the matching property. While continuous vector field topology is able to compute the position of the separatrices accurately, the separatrices computed by our combinatorial approach do not coincide with the exact separatrices of a given continuous vector field - regardless of the resolution of the mesh employed used to sample the continuous vector field (see Figure 6 top-right). The same problem is exhibited by periodic orbits. This behavior might be corrected by a more sophisticated definition of link weights of the graph (3). Consider a vector field on a torus with two attracting and two repelling orbits. In principle it is possible to simplify this topology. However, our algorithm is unable to do so. The simplification is based on matchings and always cancels pairs of critical points. If there are no critical points in the vector field, then the algorithm cannot simplify its topology. In conclusion, the presented algorithm delivers a completely combinatorial approach to vector field topology on triangulated manifolds that is always consistent. The formulation of the algorithm in standard graph terminology allows for a straight forward implementation using a graph library and the only parameter present is the level of topological detail we want to extract. Acknowledgements We would like to thank David Günther and Tino Weinkauf for many fruitful discussions on this topic and Christian Löwen for his great implementational efforts. This work was funded by the DFG Emmy-Noether research programm. All visualizations in this paper have been created using AMIRA - a system for advanced visual data analysis (see References. G. Chen, K. Mischaikow, R.S. Laramee, P. Pilarczyk, and E. Zhang. Vector field editing and periodic orbit extraction using morse decomposition. IEEE Transactions in Visualization and Computer Graphics, 3: , Herbert Edelsbrunner and J. Harer. Persistent homology a survey. In J. E. Goodman, J. Pach, and R. Pollack, editors, Surveys on Discrete and Computational Geometry: Twenty Years Later, volume 458, pages AMS Bookstore, 28.

12 3. Herbert Edelsbrunner, John Harer, Vijay Natarajan, and Valerio Pascucci. Morsesmale complexes for piecewise linear 3-manifolds. In SCG 3: Proceedings of the nineteenth annual symposium on Computational geometry, pages 36 37, New York, NY, USA, 23. ACM. 4. Robin Forman. Combinatorial vector fields and dynamical systems. Mathematische Zeitschrift, 228:629 68, Robin Forman. Morse theory for cell complexes. Advances in Mathematics, 34:9 45, Robin Forman. A user s guide to discrete morse theory. In Proceedings of the 2 Internat. Conf. on Formal Power Series and Algebraic Combinatorics, Advances in Applied Mathematics, Attila Gyulassy, Vijay Natarajan, Valerio Pascucci, Peer-Timo Bremer, and Bernd Hamann. A topological approach to simplification of three-dimensional scalar functions. IEEE Transactions on Visualization and Computer Graphics, 2(4): , J. Helman and L. Hesselink. Representation and display of vector field topology in fluid flow data sets. Computer, 22(8):27 36, August Thomas Klein and Thomas Ertl. Scale-space tracking of critical points in 3d vector fields. In Hans Hagen Helwig Hauser and Holger Theisel, editors, Topology-based Methods in Visualization, Mathematics and Visualization, pages Springer Berlin Heidelberg, May 27.. Robert S. Laramee, Helwig Hauser, Lingxiao Zhao, and Frits H. Post. Topologybased flow visualization, the state of the art. In Hans Hagen Helwig Hauser and Holger Theisel, editors, Topology-based Methods in Visualization, Mathematics and Visualization, pages 9. Springer Berlin Heidelberg, May 27.. Thomas Lewiner. Geometric discrete Morse complexes. PhD thesis, Department of Mathematics, PUC-Rio, 25. Advised by Hlio Lopes and Geovan Tavares. 2. Thomas Lewiner, Helio Lopes, and Geovan Tavares. Applications of forman s discrete morse theory to topology visualization and mesh compression. IEEE Transactions on Visualization and Computer Graphics, (5):499 58, J.J. Sanchez-Gabites. Dynamical systems and shapes. RACSAM: Geometry and Topology, 2:27 59, Alexander Schrijver. Combinatorial Optimization. Springer, Holger Theisel, Tino Weinkauf, Hans-Christian Hege, and Hans-Peter Seidel. Gridindependent detection of closed stream lines in 2d vector fields. In Proceedings of the VMV Conference 24, page 665, Stanford, USA, November Xavier Tricoche, Gerik Scheuermann, and Hans Hagen. Continuous topology simplification of planar vector fields. In VIS : Proceedings of the conference on Visualization, pages 59 66, Washington, DC, USA, 2. IEEE Computer Society. 7. Xavier Tricoche, Gerik Scheuermann, Hans Hagen, and Stefan Clauss. Vector and tensor field topology simplification on irregular grids. In D. Ebert, J. M. Favre, and R. Peikert, editors, VisSym : Proceedings of the symposium on Data Visualization 2, pages 7 6, Wien, Austria, May Springer-Verlag. 8. Tino Weinkauf, Holger Theisel, K. Shi, Hans-Christian Hege, and Hans-Peter Seidel. Extracting higher order critical points and topological simplification of 3D vector fields. In Proc. IEEE Visualization 25, pages , Minneapolis, U.S.A., October Thomas Wischgoll and Gerik Scheuermann. Detection and visualization of closed streamlines in planar flows. IEEE Transactions on Visualization and Computer Graphics, 7(2):65 72, 2.