\`x^2+y_1+z_12^34\`
Advanced Search
Article Contents
Article Contents

Capturing dynamics of time-varying data via topology

  • * Corresponding author: Lori Ziegelmeier

    * Corresponding author: Lori Ziegelmeier

L.X. was funded by Macalester College through a grant to L.Z. H.A. was supported by NSF grant 1934725, DELTA: Descriptors of Energy Landscapes by Topological Analysis. C.M.T. was supported by NSF grant DMS-1813752, Variational and Topological Approaches to Complex Systems. L.Z. was supported by NSF grant CDS & E-MSS-1854703, Exact Homological Algebra for Computational Topology (ExHACT)

Abstract Full Text(HTML) Figure(17) / Table(4) Related Papers Cited by
  • One approach to understanding complex data is to study its shape through the lens of algebraic topology. While the early development of topological data analysis focused primarily on static data, in recent years, theoretical and applied studies have turned to data that varies in time. A time-varying collection of metric spaces as formed, for example, by a moving school of fish or flock of birds, can contain a vast amount of information. There is often a need to simplify or summarize the dynamic behavior. We provide an introduction to topological summaries of time-varying metric spaces including vineyards [19], crocker plots [55], and multiparameter rank functions [37]. We then introduce a new tool to summarize time-varying metric spaces: a crocker stack. Crocker stacks are convenient for visualization, amenable to machine learning, and satisfy a desirable continuity property which we prove. We demonstrate the utility of crocker stacks for a parameter identification task involving an influential model of biological aggregations [57]. Altogether, we aim to bring the broader applied mathematics community up-to-date on topological summaries of time-varying metric spaces.

    Mathematics Subject Classification: 37N99, 55N31, 62R40, 92B99.

    Citation:

    \begin{equation} \\ \end{equation}
  • 加载中
  • Figure 1.  A filled-in disk (left) has Betti numbers $ (\beta_0, \beta_1, \beta_2, \beta_3, \ldots) = (1, 0, 0, 0, \ldots) $. A hollow square (center) has Betti numbers $ (1, 1, 0, 0, \ldots) $. A hollow two-torus (right) has Betti numbers $ (1, 2, 1, 0, \ldots) $. If we consider the union of these three shapes, the Betti numbers are $ (\beta_0, \beta_1, \beta_2, \beta_3, \ldots) = (3, 3, 1, 0, \ldots) $. Image of the torus taken from Wikimedia Commons https://commons.wikimedia.org/wiki/File:Torus.svg, available for reuse under CCA BY-SA 3.0

    Figure 2.  An illustrative example of a crocker stack for $ H_0 $ computed from a simulation from the Viscek model with noise parameter $ \eta = 0.02 $; see Section 5. The variable $ \alpha $ is a smoothing parameter which captures the persistence of topological features

    Figure 3.  Persistence diagram (Left) and corresponding persistence barcode (Right). Let the persistence barcode $ V $ consist of the intervals $ [1, 7] $, $ [2, 9] $, $ [3, 11] $, $ [5, 10] $, $ [5, 9] $, and let $ 4 = i \le j = 8 $. Then $ \mathrm{rank}(V(4) \to V(8)) $ is two since there are two intervals in the persistence diagram that contain the interval $ [i, j] = [4, 8] $

    Figure 4.  Round points (in red) represent points in persistence diagram $ A $. Triangular points (in blue) represent points in persistence diagram $ B $. The bottleneck distance between persistence diagrams $ A $ and $ B $ is computed by taking the largest $ L_\infty $ distance between matched round and triangular points of a bijection between $ A $ and $ B $, and then taking the infimum over all bijections. The three boxes (in green) show the optimal matching of three pairs of round and triangular points. The unboxed points match to the closest points on the diagonal

    Figure 5.  Vines and vineyard. Each point (in red) represents a point on a persistence diagram. Each dashed curve (in blue) is a vine traced out by a persistent point on time-varying persistence diagrams. The horizontal direction denotes time

    Figure 6.  The effect of $ \alpha $-smoothing. (Top) A persistence diagram, the corresponding persistence intervals (drawn vertically), and one column of a crocker plot matrix. If we had points moving in time, then we would get a time-varying persistence diagram, a time-varying persistence barcode, and a complete crocker plot matrix (swept out from left to right as time increases). (Bottom) A persistence diagram with the thick line (in red) reflecting the choice of $ \alpha $-smoothing, along with the corresponding $ \alpha $-smoothed persistence intervals, and one column of an $ \alpha $-smoothed crocker plot matrix. The $ y $-intercept of the diagonal thick red line is $ 2\alpha $. All persistence diagram points under the thick line are ignored under $ \alpha $-smoothing

    Figure 7.  To update the heading of an agent (centered, round blue point) according to the Vicsek model, we first find the nearby neighbors within a radius $ R $ (denoted by the dashed circle in red) and then take the average of its neighbors' headings, plus some noise

    Figure 8.  A plot of order parameters for three simulations of the Vicsek model with different noise parameters $ \eta $. For smaller values of $ \eta $, particles become more aligned, i.e. move in the same direction, over time

    Figure 9.  An example $ H_0 $ crocker plot of a simulation from the Viscek model with noise parameter $ \eta = 0.02 $. This is the same as an $ \alpha $-cross section of a crocker stack when $ \alpha $ = 0. In the region below the lowest curve (purple) where $ \beta_0 \geq 5 $, there can be many connected components, which we interpret as noise and is thus not displayed

    Figure 10.  An example $ H_0 $ and $ H_1 $ crocker stack for a simulation from the Viscek model with noise parameter $ \eta = 0.02 $. This figure shows the shifts of Betti curves in $ H_0 $ and $ H_1 $ as smoothing parameter $ \alpha $ increases from $ 0 $ to $ 0.01 $ and $ 0.03 $

    Figure 11.  The color scale corresponds to values in the distance matrix; denser color (red) means larger distances, and lighter color (yellow) means smaller distances. The 100 simulations of each noise parameter $ \eta = 0.01, 0.5, 1, 1.5, 2 $ are listed in order and annotated in the left matrix. The $ H_{0, 1} $ crocker distance matrix is more structured (Left) than the order parameter distance matrix (Right)

    Figure 12.  The $ H_{0, 1} $ crocker distance matrix is more structured (Left) than the order parameter distance matrix (Right)

    Figure 13.  The $ H_{0, 1} $ crocker distance matrix is more structured (Left) than the order parameter distance matrix (Right)

    Figure 14.  The $ H_{0, 1} $ crocker distance matrix is more structured (Left) than the order parameter distance matrix (Right)

    Figure 15.  Let $ Z = \mathbb{R}^2 $ with the Euclidean metric. The thicker solid curve (in red) represents subset $ X $ of $ Z $, and the thinner solid curve (in blue) represents subset $ Y $ of $ Z $. To compute the Hausdorff distance between $ X $ and $ Y $, we first take the supremum over all points in $ Y $ of the distance to the closest point in $ X $. In this figure, the distance is $ a = \sup_{y \in Y} \inf_{x \in X} d(x, y) $. Then, we do the same for the supremum over all points in $ X $ of the distance to the closest point in $ Y $, as shown by $ b = \sup_{x\in X} \inf_{y \in Y} d(x, y) $. Finally, we take the maximum of the two suprema, $ d_H^Z(X, Y) = \max\{a, b\} = a $

    Figure 16.  Consider the persistence diagrams $ \mathrm{Dgm}_k(V) = \{ (3, 6), (2, 8) \} $, denoted with red circles, and $ \mathrm{Dgm}_k(W) = \{ (1, 7), (3, 7.5) \} $, denoted with blue triangles. The bottleneck distance between the two persistence diagrams is 1.5, while the erosion distance is 1

    Figure 17.  Suppose $ A $ and $ B $ are dynamic metric spaces. The distance between any two of the four round points (in red) in $ A $ is 1 for all times $ t $. The distance between any two of the four triangular points (in blue) in $ B $ is $ 1+\varepsilon $ for all times $ t $. At an arbitrary time step $ t $ and scale parameter $ 1+\frac{\varepsilon}{2} $, we have $ \beta_0^A = 1 $ and $ \beta_0^B = 4 $

    Table 1.  Summary of the clustering accuracy on four different experiments (abbreviated Exp.) with three different feature vectors: order parameters, crocker plots, and crocker stacks. For crocker plots and crocker stacks, we distinguish different homological dimensions: $ H_{0, 1} $, $ H_0 $, and $ H_1 $. The top accuracy scores of each column are bolded. This table summarizes results with time step 1 for order parameters and time step 10 for crocker representations. Results with other time steps are discussed in Section 5.3.1. Clustering results with feature vectors that have been reduced to 3 dimensionsby PCA are shown in italics, while full feature vectors are not italicized

    Exp. 1 Exp. 2 Exp. 3 Exp. 4
    Order Parameters 0.63 0.51 0.61 0.59 0.35 0.35 0.21 0.17
    Crocker Plots, $ H_{0, 1} $ 1.00 1.00 0.67 0.67 0.44 0.43 0.42 0.43
    Crocker Plots, $ H_0 $ 1.00 1.00 0.67 0.77 0.45 0.43 0.39 0.43
    Crocker Plots, $ H_1 $ 0.98 0.99 0.71 0.67 0.36 0.35 0.37 0.33
    Crocker Stacks, $ H_{0, 1} $ 1.00 0.98 0.67 0.67 0.47 0.38 0.41 0.35
    Crocker Stacks, $ H_0 $ 1.00 1.00 0.67 0.67 0.49 0.46 0.41 0.41
    Crocker Stacks, $ H_1 $ 0.96 0.98 0.63 0.67 0.34 0.37 0.32 0.35
     | Show Table
    DownLoad: CSV

    Table 2.  Confusion matrix using $ K $-medoids to cluster the $ H_{0, 1} $ crocker plots corresponding to simulations of Experiment 3 ($ \eta $ = 0.01, 0.02, 0.19, 0.2, 1.99, 2). The rows represent the actual simulations corresponding to each noise parameter $ \eta $ while the columns represent the parameter of the cluster medoid to which a simulation is assigned. Even though $ K = 6 $ clusters were formed, the six medoids correspond to simulations from only three distinct noise parameters $ \eta = 0.02, 0.2, 2 $

    $ K $-medoids Clusters
    $ \eta $ = 0.02 $ \eta $ = 0.20 $ \eta $ = 2.00
    $ \eta $ = 0.01 69 31 0
    $ \eta $ = 0.02 64 36 0
    $ \eta $ = 0.19 4 96 99
    $ \eta $ = 0.2 1 99 0
    $ \eta $ = 1.99 0 0 100
    $ \eta $ = 2.00 0 0 100
     | Show Table
    DownLoad: CSV

    Table 3.  Summary of the clustering accuracy for the four experiments based on single $ \alpha $-smoothed crocker plots in $ H_{0, 1} $ as input feature vectors to $ K $-medoids, and the accuracy based on the crocker stack (with 18 $ \alpha $ values combined). The top accuracy scores of each column are bolded. Recall that when $ \alpha = 0 $, the $ \alpha $-smoothed crocker plot is equivalent to the standard crocker plot of [55]

    Exp. 1 Exp. 2 Exp. 3 Exp. 4
    stack 1.00 0.67 0.47 0.41
    $ \alpha $ = 0.00 1.00 0.67 0.44 0.42
    $ \alpha $ = 0.01 1.00 0.67 0.46 0.40
    $ \alpha $ = 0.03 1.00 0.67 0.49 0.40
    $ \alpha $ = 0.05 1.00 0.58 0.44 0.38
    $ \alpha $ = 0.08 0.99 0.67 0.39 0.36
    $ \alpha $ = 0.11 0.97 0.67 0.33 0.35
    $ \alpha $ = 0.13 0.92 0.73 0.41 0.28
    $ \alpha $ = 0.17 0.73 0.66 0.40 0.21
     | Show Table
    DownLoad: CSV

    Table 4.  Comparison of the $ K $-medoids clustering accuracy of Experiment 2 of the Euclidean distance on order parameters, crocker plots, and crocker stacks, as well as the bottleneck distance on the stacked set of persistence diagrams. All topological representations compute homology in dimensions 0 and 1, denoted $ H_0 $ and $ H_1 $. The top accuracy scores of each column are bolded. The parentheses on the order parameter row indicate that the same computation is performed in both columns since order parameters do not incorporate homology dimensions

    $ H_0 $ $ H_1 $
    Order Parameters (0.61) (0.61)
    Crocker Plots 0.67 0.71
    Crocker Stacks 0.67 0.63
    Stacked Persistence Diagrams 0.67 0.49
     | Show Table
    DownLoad: CSV
  • [1] H. Adams and G. Carlsson, Evasion paths in mobile sensor networks, International Journal of Robotics Research, 34 (2015), 90-104. 
    [2] H. Adams, T. Emerson, M. Kirby, R. Neville, C. Peterson, P. Shipman, S. Chepushtanova, E. Hanson, F. Motta and L. Ziegelmeier, Persistence images: A stable vector representation of persistent homology, J. Mach. Learn. Res., 18 (2017), Paper No. 8, 35 pp. http://jmlr.org/papers/v18/16-337.html.
    [3] H. Adams, D. Ghosh, C. Mask, W. Ott and K. Williams, Efficient evader detection in mobile sensor networks, arXiv preprint, arXiv: 2101.09813.
    [4] P. AroraD. Deepali and S. Varshney, Analysis of K-means and K-medoids algorithm for big data, Procedia Computer Science, 78 (2016), 507-512.  doi: 10.1016/j.procs.2016.02.095.
    [5] A. Banman and L. Ziegelmeier, Mind the gap: A study in global development through persistent homology, in Research in Computational Topology, Springer, 2018,125–144. doi: 10.1007/978-3-319-89593-2_8.
    [6] D. Bhaskar, A. Manhart, J. Milzman, J. T. Nardini, K. M. Storey, C. M. Topaz and L. Ziegelmeier, Analyzing collective motion with machine learning and topology, Chaos: An Interdisciplinary Journal of Nonlinear Science, 29 (2019), 123125, 12 pp. doi: 10.1063/1.5125493.
    [7] P. Bubenik, Statistical topological data analysis using persistence landscapes, J. Mach. Learn. Res., 16 (2015), 77-102. 
    [8] D. Burago, Y. Burago and S. Ivanov, A course in Metric Geometry, vol. 33, American Mathematical Society, Providence, 2001. doi: 10.1090/gsm/033.
    [9] G. Carlsson, Topology and data, Bull. Amer. Math. Soc. (N.S.), 46 (2009), 255-308.  doi: 10.1090/S0273-0979-09-01249-X.
    [10] G. Carlsson and V. de Silva, Zigzag persistence, Found. Comput. Math., 10 (2010), 367-405.  doi: 10.1007/s10208-010-9066-0.
    [11] G. CarlssonV. de SilvaS. Kališnik and D. Morozov, Parametrized homology via zigzag persistence, Algebr. Geom. Topol., 19 (2019), 657-700.  doi: 10.2140/agt.2019.19.657.
    [12] G. Carlsson, V. de Silva and D. Morozov, Zigzag persistent homology and real-valued functions, in Proceedings of the Twenty-Fifth Annual Symposium on Computational Geometry, ACM, 2009,247–256. doi: 10.1145/1542362.1542408.
    [13] G. Carlsson, G. Singh and A. Zomorodian, Computing multidimensional persistence, Algorithms and computation, 730–739, Lecture Notes in Comput. Sci., 5878, Springer, Berlin, 2009. doi: 10.1007/978-3-642-10631-6_74.
    [14] G. Carlsson and A. Zomorodian, The theory of multidimensional persistence, Discrete Comput. Geom., 42 (2009), 71-93.  doi: 10.1007/s00454-009-9176-0.
    [15] A. CerriB. D. FabioM. FerriP. Frosini and C. Landi, Betti numbers in multidimensional persistent homology are stable functions, Math. Methods Appl. Sci., 36 (2013), 1543-1557.  doi: 10.1002/mma.2704.
    [16] W. ChachólskiM. Scolamiero and F. Vaccarino, Combinatorial presentation of multidimensional persistent homology, J. Pure Appl. Algebra, 221 (2017), 1055-1075.  doi: 10.1016/j.jpaa.2016.09.001.
    [17] F. ChazalV. de Silva and S. Oudot, Persistence stability for geometric complexes, Geometriae Dedicata, 174 (2014), 193-214.  doi: 10.1007/s10711-013-9937-z.
    [18] D. Cohen-SteinerH. Edelsbrunner and J. Harer, Stability of persistence diagrams, Discrete Comput. Geom., 37 (2007), 103-120.  doi: 10.1007/s00454-006-1276-5.
    [19] D. Cohen-Steiner, H. Edelsbrunner and D. Morozov, Vines and vineyards by updating persistence in linear time, in Computational Geometry (SCG'06), ACM, 2006,119–126. doi: 10.1145/1137856.1137877.
    [20] P. Corcoran and C. B. Jones, Modelling topological features of swarm behaviour in space and time with persistence landscapes, IEEE Access, 5 (2017), 18534-18544.  doi: 10.1109/ACCESS.2017.2749319.
    [21] D. B. Damiano and M. R. McGuirl, A topological analysis of targeted in-111 uptake in SPECT images of murine tumors, J. Math. Biol., 76 (2018), 1559-1587.  doi: 10.1007/s00285-017-1184-8.
    [22] V. de Silva and R. Ghrist, Coordinate-free coverage in sensor networks with controlled boundaries via homology, The International Journal of Robotics Research, 25 (2006), 1205-1222.  doi: 10.1177/0278364906072252.
    [23] V. de Silva and R. Ghrist, Coverage in sensor networks via persistent homology, Algebr. Geom. Topol., 7 (2007), 339-358.  doi: 10.2140/agt.2007.7.339.
    [24] T. K. Dey and C. Xin, Computing bottleneck distance for 2-d interval decomposable modules, arXiv preprint, arXiv: 1803.02869.
    [25] M. R. D'OrsognaY. L. ChuangA. L. Bertozzi and L. S. Chayes, Self-propelled particles with soft-core interactions: Patterns, stability, and collapse, Phys. Rev. Lett., 96 (2006), 104302.  doi: 10.1103/PhysRevLett.96.104302.
    [26] H. Edelsbrunner and J. L. Harer, Computational Topology: An Introduction, American Mathematical Society, Providence, 2010. doi: 10.1090/mbk/069.
    [27] H. Edelsbrunner, D. Morozov and A. Patel, The stability of the apparent contour of an orientable 2-manifold, Topological Methods in Data Analysis and Visualization. Mathematics and Visualization., 27–41, Math. Vis., Springer, Heidelberg, 2011. doi: 10.1007/978-3-642-15014-2_3.
    [28] B. T. Fasy, J. Kim, F. Lecci, C. Maria, D. L. Millman and V. Rouvreau, Tda: Statistical tools for topological data analysis, https://cran.r-project.org/web/packages/TDA/index.html.
    [29] M. Feng and M. A. Porter, Persistent homology of geospatial data: A case study with voting, SIAM Rev., 63 (2021), 67-99.  doi: 10.1137/19M1241519.
    [30] M. Feng and M. A. Porter, Spatial applications of topological data analysis: Cities, snowflakes, random structures, and spiders spinning under the influence, Phys. Rev. Research, 2 (2020), 033426.  doi: 10.1103/PhysRevResearch.2.033426.
    [31] R. Ghrist, Barcodes: The persistent topology of data, ull. Amer. Math. Soc. (N.S.), 45 (2008), 61-75.  doi: 10.1090/S0273-0979-07-01191-3.
    [32] C. GiustiL. PapadopoulosE. T. OwensK. E. Daniels and D. S. Bassett, Topological and geometric measurements of force-chain structure, Physical Review E, 94 (2016), 032909.  doi: 10.1103/PhysRevE.94.032909.
    [33] I. T. Jolliffe, Principal Component Analysis, Springer Verlag, 1986. doi: 10.1007/978-1-4757-1904-8.
    [34] T. Kaczynski, K. Mischaikow and M. Mrozek, Computational Homology, vol. 157, pringer-Verlag, New York, 2004. doi: 10.1007/b97315.
    [35] L. Kaufman and P. Rousseeuw, Clustering by Means of Medoids, North-Holland, 1987.
    [36] W. Kim and F. Mémoli, Stable signatures for dynamic metric spaces via zigzag persistent homology, arXiv preprint, arXiv: 1712.04064.
    [37] W. Kim and F. Mémoli, Spatiotemporal persistent homology for dynamic metric spaces, Discrete Comput. Geom., 66 (2021), 831-875.  doi: 10.1007/s00454-019-00168-w.
    [38] M. Lesnick, The theory of the interleaving distance on multidimensional persistence modules, Found. Comput. Math., 15 (2015), 613-650.  doi: 10.1007/s10208-015-9255-y.
    [39] M. Maechler, Finding groups in data: Cluster analysis extended rousseeuw et al, https://cran.r-project.org/web/packages/cluster/cluster.pdf.
    [40] A. McCleary and A. Patel, Bottleneck stability for generalized persistence diagrams, Proc. Amer. Math. Soc., 148 (2020), 3149-3161.  doi: 10.1090/proc/14929.
    [41] A. McCleary and A. Patel, Edit distance and persistence diagrams over lattices, arXiv preprint, arXiv: 2010.07337.
    [42] E. Miller, Data structures for real multiparameter persistence modules, arXiv preprint, arXiv: 1709.08155.
    [43] N. Milosavljević, D. Morozov and P. Škraba, Zigzag persistent homology in matrix multiplication time, in Computational geometry (SCG'11), 2011,216–225. doi: 10.1145/1998196.1998229.
    [44] D. Morozov, Personal communication.
    [45] D. Morozov, Dionysus, http://www.mrzv.org/software/dionysus/.
    [46] J. R. Munkres, Topology, Prentice-Hall Englewood Cliffs, NJ, 1975.
    [47] C. Nilsen, J. Paige, O. Warner, B. Mayhew, R. Sutley, M. Lam, A. J. Bernoff and C. M. Topaz, Social aggregation in pea aphids: Experiment and random walk modeling, PLoS ONE, 8 (2013), e83343. doi: 10.1371/journal.pone.0083343.
    [48] N. OtterM. A. PorterU. TillmannP. Grindrod and H. A. Harrington, A roadmap for the computation of persistent homology, EPJ Data Science, 6 (2017), 17. 
    [49] S. Y. Oudot, Persistence Theory: From Quiver Representations to Data Analysis, vol. 209, American Mathematical Society Providence, RI, 2015. doi: 10.1090/surv/209.
    [50] H.-S. Park and C.-H. Jun, A simple and fast algorithm for $k$-medoids clustering, Expert Systems with Applications, 36 (2009), 3336-3341.  doi: 10.1016/j.eswa.2008.01.039.
    [51] A. Patel, Generalized persistence diagrams, J. Appl. Comput. Topol., 1 (2018), 397-419.  doi: 10.1007/s41468-018-0012-6.
    [52] V. Puuska, Erosion distance for generalized persistence modules, Homology Homotopy Appl., 22 (2020), 233-254.  doi: 10.4310/HHA.2020.v22.n1.a14.
    [53] M. ScolamieroW. ChachólskiA. LundmanR. Ramanujam and S. Öberg, Multidimensional persistence and noise, Found. Comput. Math., 17 (2017), 1367-1406.  doi: 10.1007/s10208-016-9323-y.
    [54] B. J. Stolz, H. A. Harrington and M. A. Porter, Persistent homology of time-dependent functional networks constructed from coupled time series, Chaos, 27 (2017), 047410, 17 pp. doi: 10.1063/1.4978997.
    [55] C. M. Topaz, L. Ziegelmeier and T. Halverson, Topological data analysis of biological aggregation models, PloS One, 10 (2015), e0126383. doi: 10.1371/journal.pone.0126383.
    [56] M. Ulmer, L. Ziegelmeier and C. M. Topaz, A topological approach to selecting models of biological experiments, PloS One, 14 (2019), e0213679. doi: 10.1371/journal.pone.0213679.
    [57] T. VicsekA. CzirókE. Ben-JacobI. Cohen and O. Shochet, Novel type of phase transition in a system of self-driven particles, Phys. Rev. Lett., 75 (1995), 1226-1229.  doi: 10.1103/PhysRevLett.75.1226.
    [58] T. Vicsek and A. Zafeiris, Collective motion, Physics Reports, 517 (2012), 71-140.  doi: 10.1016/j.physrep.2012.03.004.
    [59] X. Zhu, Persistent homology: An introduction and a new text representation for natural language processing, in Twenty-Third International Joint Conference on Artificial Intelligence, 2013.
    [60] A. Zomorodian and G. Carlsson, Computing persistent homology, Discrete Comput. Geom., 33 (2005), 249-274.  doi: 10.1007/s00454-004-1146-y.
  • 加载中

Figures(17)

Tables(4)

SHARE

Article Metrics

HTML views(946) PDF downloads(436) Cited by(0)

Access History

Other Articles By Authors

Catalog

    /

    DownLoad:  Full-Size Img  PowerPoint
    Return
    Return