Abstract
We completely characterize the unimodal category for functions \(f:{\mathbb {R}}\rightarrow [0,\infty )\) using a decomposition theorem obtained by generalizing the sweeping algorithm of Baryshnikov and Ghrist. We also give a characterization of the unimodal category for functions \(f:S^1\rightarrow [0,\infty )\) and provide an algorithm to compute the unimodal category of such a function in the case of finitely many critical points. We then turn to the monotonicity conjecture of Baryshnikov and Ghrist. We show that this conjecture is true for functions on \({\mathbb {R}}\) and \(S^1\) using the above characterizations and that it is false on certain graphs and on the Euclidean plane by providing explicit counterexamples. We also show that it holds for functions on the Euclidean plane whose Morse–Smale graph is a tree using a result of Hickok, Villatoro and Wang.
Similar content being viewed by others
Change history
13 October 2021
On page 28 and 33 the equations K & L was revised.
Notes
Note that they seem to be using the notation \(\mathrm{supp}f\) to denote \(f^{-1}(0,\infty )\), i.e. the set-theoretic support of f.
We use the standard convention that \(\inf \emptyset =\infty \) and \((a,b)=\emptyset \) if \(a\ge b\).
This actually allows us to compute \(\mathbf{ucat}(f)\) of \(f:X\rightarrow [0,\infty )\) for any set \(X \subseteq {\mathbb {R}}\): if \(f^{-1}(0,\infty )\) has infinitely many components, \(\mathbf{ucat}(f)\) is infinite, otherwise \(f^{-1}(0,\infty )\) is a finite union of intervals, each of which can be treated separately.
Note that \(u_n\) is automatically increasing if \(u_{n-1}(1)\ne 0\).
If these conditions are satisfied, we say that a majorizes b.
Here, xy is shorthand for \(\{x,y\}\).
Here, \(p_1p_2\ldots p_n\) refers to the union of segments \(p_1p_2,p_2p_3,\ldots ,p_{n-1}p_n\) and \(\overline{p_1p_2\ldots p_n}\) denotes the compact region bounded by \(p_1p_2\ldots p_n p_1\); see “Appendix A.1” for more details.
Here \(\partial Q\) is the topological boundary of Q and \(\mathrm{int}\,Q\) is the interior of Q.
The notation is not meant to imply any relation with a and b at this point.
Here by “Morse function” we mean a compactly supported function, which is Morse on its support.
References
Améndola, C., Engström, A., Haase, C.: Maximum number of modes of Gaussian mixtures. Inf. Inference 9(3), 587–600 (2020). https://doi.org/10.1093/imaiai/iaz013
Apostol, T.M.: Mathematical Analysis, 2nd edn. Addison-Wesley Publishing Co., Reading (1974)
Baryshnikov, Y., Ghrist, R.: Unimodal category and topological statistics. In: Proceedings of NOLTA (2011)
Behboodian, J.: On the modes of a mixture of two normal distributions. Technometrics 12(1), 131–139 (1970)
Bing, R.H.: The Geometric Topology of 3-Manifolds, vol. 40. American Mathematical Society Colloquium Publications, Providence (1983). https://doi.org/10.1090/coll/040
Cornea, O., Lupton, G., Oprea, J., Tanré, D.: Lusternik-Schnirelmann Category. Mathematical Surveys and Monographs, vol. 103. American Mathematical Society, Providence (2003). https://doi.org/10.1090/surv/103
Doob, J.L.: Measure Theory. Graduate Texts in Mathematics, vol. 143. Springer, New York (1994). https://doi.org/10.1007/978-1-4612-0877-8
Edelsbrunner, H., Fasy, B.T., Rote, G.: Add isotropic Gaussian kernels at own risk: more and more resilient modes in higher dimensions. Discret. Comput. Geom. 49(4), 797–822 (2013). https://doi.org/10.1007/s00454-013-9517-x
Eisenberger, I.: Genesis of bimodal distributions. Technometrics 6, 357–363 (1964). https://doi.org/10.2307/1266090
Ghrist, R.W.: Elementary Applied Topology, vol. 1. CreateSpace, Seattle (2014)
Govc, D.: Unimodal category. PhD thesis, University of Ljubljana (2017)
Hickok, L., Villatoro, J., Wang, X.: Unimodal category of 2-dimensional distributions (full), unpublished note. https://faculty.math.illinois.edu/~xwang105/unimodal.pdf (n.d.)
Huntsman, S.: Topological density estimation. In: Proceedings of NOLTA (2017)
Huntsman, S.: Topological mixture estimation. In: Dy, J., Krause, A. (eds.) Proceedings of the 35th International Conference on Machine Learning, PMLR, Stockholmsmässan, Stockholm Sweden, Proceedings of Machine Learning Research, vol. 80, pp. 2088–2097 (2018). http://proceedings.mlr.press/v80/huntsman18a.html
Kadelburg, Z., Đukić, D., Lukić, M., Matić, I.: Inequalities of Karamata, Schur and Muirhead, and some applications. Teach. Math. 8(1), 31–45 (2005)
Kakiuchi, I.: Unimodality conditions of the distribution of a mixture of two distributions. Math. Sem. Notes Kobe Univ. 9(2), 315–325 (1981)
Kemperman, J.H.B.: Mixtures with a limited number of modal intervals. Ann. Stat. 19(4), 2120–2144 (1991). https://doi.org/10.1214/aos/1176348389
Maehara, R.: The Jordan curve theorem via the Brouwer fixed point theorem. Am. Math. Mon. 91(10), 641–643 (1984). https://doi.org/10.2307/2323369
Milnor, J.: Morse Theory. Based on Lecture Notes by M. Spivak and R. Wells. Annals of Mathematics Studies, No. 51. Princeton University Press, Princeton (1963)
Nicolaescu, L.: An Invitation to Morse Theory, 2nd edn. Springer, New York (2011). https://doi.org/10.1007/978-1-4614-1105-5
Ray, S., Lindsay, B.G.: The topography of multivariate normal mixtures. Ann. Stat. 33(5), 2042–2065 (2005). https://doi.org/10.1214/009053605000000417
Robertson, C.A., Fryer, J.G.: Some descriptive properties of normal mixtures. Skand. Aktuarietidskr. 1969, 137–146 (1969). https://doi.org/10.1080/03461238.1969.10404590
Shirinkam, S., Alaeddini, A., Gross, E.: Identifying the number of components in Gaussian mixture models using numerical algebraic geometry. J. Algebra Appl. 19(11), 2050204 (2020). https://doi.org/10.1142/S0219498820502047
Acknowledgements
The author would like to thank his thesis advisors Dušan Repovš and Primož Škraba, the latter of whom suggested the problem. Thanks also to Jaka Smrekar for encouragement during the early stages of this research.
Author information
Authors and Affiliations
Corresponding author
Ethics declarations
Conflict of interest
The author states that there is no conflict of interest.
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Dejan Govc was supported by the Slovenian Research Agency Grant P1-0292-0101.
Appendices
A Direct descriptions of the counterexamples in \({\mathbb {R}}^2\)
1.1 A.1 First example
Piecewise linearity of the functions \(u_1,u_2,F:{\mathbb {R}}^2\rightarrow [0,\infty )\) appearing in Sect. 4.4 follows from the properties of the \(\infty \)-distance. For concreteness, we explicitly describe the decomposition of the plane with respect to which the functions are piecewise linear. The main advantage of \(u_1\) and F being piecewise linear is that other facts about these functions can be verified completely computationally.
Before we begin, we need some notation. Two points \(p_1,p_2\in {\mathbb {R}}^2\) determine a segment
We write \(p_1p_2\ldots p_n\) for the union of segments \(p_1p_2,p_2p_3,\ldots ,p_{n-1}p_n\). If \(p_1p_2\ldots p_np_1\) is a topological circle in \({\mathbb {R}}^2\), it is the boundary of a uniquely determined compact set in \({\mathbb {R}}^2\), which we denote by \(\overline{p_1p_2\ldots p_n}\).
We can describe \(u_1,u_2\) and F as piecewise linear functions defined by their values on the vertices of a polygonal decomposition of \(S=\mathrm{supp}f\) consisting of 44 vertices, 95 edges and 52 faces, namely triangles, trapezoids and two non-convex quadrilaterals (see Fig. 17). Note that we do not count the “face at infinity” and we consider parallelograms to be a special case of trapezoids. Some care must be taken as not every choice of values at the vertices of a quadrilateral can be extended to a linear function. First, we list the vertices (indexed lexicographically):
We omit listing the edges \(h_1,h_2,\ldots ,h_{95}\) (ordered lexicographically by the indices of the vertices) as they are simply the edges of the 52 polygons in the decomposition. Finally, we list the faces, using the notation defined above (ordered lexicographically by the corresponding sets of vertices):
We can now state the alternative descriptions of \(u_1,u_2\) and F. The function \(u_1\) can be defined on the vertices of the above decomposition:
Similarly, we have
Summing these, we obtain
The values of F (and the polygonal decomposition of \(S=\mathrm{supp}f\)) are pictured in Fig. 17.
1.2 A.2 Second example
For the same reason as our first example, \(f:{\mathbb {R}}^2\rightarrow [0,\infty )\) is actually a piecewise linear function, so it can be described by its function values at the vertices of a polygonal decomposition of its support \(S=\mathrm{supp}f\). This decomposition consists of 47 vertices, 94 edges and 46 faces, namely triangles, trapezoids and two non-convex quadrilaterals (see Fig. 18). Again note that not every choice of values at the vertices of a trapezoid can be extended to a linear function, but in our case such issues do not arise as the two function values on each of two parallel sides agree. We first list the vertices (indexed lexicographically):
We again omit listing the edges \(h_1,h_2,\ldots ,h_{94}\) and proceed to the faces:
An alternative definition of the piecewise linear function \(f:{\mathbb {R}}^2\rightarrow [0,\infty )\) is then given by specifying it on the vertices as follows (see Fig. 18):
Proposition 5.1
The function \(f:{\mathbb {R}}^2\rightarrow [0,\infty )\) has a unimodal \(\infty \)-decomposition of length 2.
Proof
Define two piecewise linear functions \(v_1,v_2:{\mathbb {R}}^2\rightarrow [0,\infty )\) on a polygonal decomposition of S by specifying their values at the vertices. Start with the decomposition of S defined above. Further subdivide it by adding the edges \(x_{15}x_{18}, x_{17}x_{23}, x_{30}x_{37}, x_{35}x_{39}\). The function \(v_1\) assumes the value 1 at the vertices \(x_3,x_4,x_{10},x_{15},x_{16},x_{17},x_{24},x_{25},x_{26},x_{27},x_{28}\); the value 3 at the vertex \(x_9\); and the value 0 at all the other vertices. The function \(v_2\) assumes the value 1 at the vertices \(x_3,x_4,x_{24},x_{25},x_{26},x_{27},x_{28},x_{37},x_{38},x_{39},x_{44}\); the value 3 at the vertex \(x_{45}\); and the value 0 at all the other vertices. It is clear that \(\max \{v_1,v_2\}=f\), however, these functions are not unimodal, as \(x_3x_{24}x_{28}x_4x_3\) yields a non-trivial cycle in some of the superlevel sets. However, we can modify \(v_1\) and \(v_2\) to obtain unimodal functions \(u_1,u_2:{\mathbb {R}}^2\rightarrow [0,\infty )\). To retain the property \(\max \{u_1,u_2\}=f\), we modify them on sets with disjoint interiors, namely, \(u_1\) is modified on the sets \(R_1=[-1,1]\times [2,3]\), \(R_2=[-1,1]\times [-2,-1]\) and \(R_3=[-7,-5]\times [-2,-1]\), whereas \(u_2\) is modified on the sets \(R_4=[-1,1]\times [1,2]\), \(R_5=[-1,1]\times [-3,-2]\) and \(R_6=[-7,-5]\times [1,2]\). They are modified in the same way on each of these. Namely, if R is of one of these rectangles, subdivide it into four triangles using the center point of R. Then define a piecewise linear function on R that takes the value 1 at the center point of R and value 0 at the vertices of the rectangle, and extend it by 0 to the whole plane to obtain a function \(\varphi _R:{\mathbb {R}}^2\rightarrow [0,\infty )\). Now, \(u_1\) and \(u_2\) are defined by putting \(u_1=v_1-\varphi _{R_1}-\varphi _{R_2}-\varphi _{R_3}\) and \(u_2=v_2-\varphi _{R_4}-\varphi _{R_5}-\varphi _{R_6}\). (These functions are piecewise linear w.r.t. a further subdivision of S that takes into account the six rectangles.) A straightforward verification shows that \(u_1\) and \(u_2\) are unimodal, but it is ultimately unenlightening, so instead of this, we provide the graphs of the two functions (Fig. 19). \(\square \)
Proposition 5.2
The function \(f:{\mathbb {R}}^2\rightarrow [0,\infty )\) has a unimodal decomposition of length 3.
Proof
An explicit unimodal decomposition of \(f:{\mathbb {R}}^2\rightarrow [0,\infty )\) can be described as follows: take the polygonal decomposition of S as defined in the beginning. Further subdivide it by adding the edges \(x_1x_4, x_4x_5, x_{15}x_{18}, x_{20}x_{25}, x_{21}x_{27}, x_{25}x_{32}, x_{27}x_{33}, x_{35} x_{39}\). Now, define piecewise linear functions \(u_1,u_2,u_3:{\mathbb {R}}^2\rightarrow [0,\infty )\) by specifying their values at the vertices of this decomposition. For the sake of brevity, we only specify the nonzero values. The function \(u_1\) is defined by taking the value 3 at the vertex \(x_9\) and the value 1 at the vertices \(x_4,x_{10},x_{15},x_{16},x_{17},x_{27},x_{28}\). The function \(u_2\) is defined by taking the value 3 at the vertex \(x_{45}\) and the value 1 at the vertices \(x_3,x_{24},x_{25}, x_{37},x_{38},x_{39},x_{44}\). Finally, \(u_3\) takes the value 1 at the vertex \(x_{26}\). It is straightforward to verify that these are all unimodal and that \(f=u_1+u_2+u_3\). (See the graphs in Fig. 20.) \(\square \)
B Algorithm for the circle
Theorem 3.7 provides a generalization of the sweeping algorithm computing \(\mathbf{ucat}(f)\) for \(f:{\mathbb {R}}\rightarrow [0,\infty )\) with finitely many critical points, given in Baryshnikov and Ghrist (2011). Here we describe Algorithm 1 which computes \(\mathbf{ucat}(f)\) for \(f:S^1\rightarrow [0,\infty )\) in the case of finitely many critical points. Although this is possible in principle, as the proofs of the theorems leading to the algorithm are constructive in nature, we do not compute the explicit decomposition.
Rights and permissions
About this article
Cite this article
Govc, D. Unimodal category and the monotonicity conjecture. J Appl. and Comput. Topology 5, 621–669 (2021). https://doi.org/10.1007/s41468-021-00077-z
Received:
Revised:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s41468-021-00077-z