1: %%%%%%%%%%%%% July 2005 %%%%%%%%%%%
2: %
3: %% Version 1, May 2005.
4: %
5: %% Version 2, July 2005. New proof of Lemma 3.9, and a few typos fixed.
6: %
7: %% Version 3, January 2006. Cleaned up, references to recent developments.
8:
9: \documentclass[epsbox,12pt,reqno]{amsart}
10: % reqno: equations numbers on the right for amsart.
11:
12: \usepackage{amsmath,amssymb,latexsym,amscd,amsthm} % ams auxiliaries.
13: \usepackage{graphicx}
14:
15: % \usepackage{times} % Causes problems with accents & spacing in PDF.
16:
17: %%% Custom page size for draft versions.
18: \setlength{\textwidth}{6.5in}
19: \setlength{\textheight}{8.45in}
20: \setlength{\oddsidemargin}{0in}
21: \setlength{\evensidemargin}{0in}
22: \setlength{\voffset}{-0.3in}
23: \setlength{\topmargin}{-0.1in} % Needed this for the ArXiv version
24: \setlength{\headsep}{0.2in}
25:
26: %%% Theorem environments.
27: \newtheorem{theorem}{Theorem}[section] % If subsection, get 3.1.2, etc.
28: \newtheorem{corollary}[theorem]{Corollary}
29: \newtheorem{proposition}[theorem]{Proposition}
30: \newtheorem{lemma}[theorem]{Lemma}
31: \theoremstyle{definition}
32: \newtheorem{definition}[theorem]{Definition}
33: \newtheorem{example}[theorem]{Example}
34: \newtheorem{question}[theorem]{Question}
35: \newtheorem{conjecture}[theorem]{Conjecture}
36: \theoremstyle{remark}
37: \newtheorem{remark}[theorem]{Remark}
38: % \numberwithin{equation}{section}
39:
40: %%% Special commands used.
41: \def\sd{\operatorname{depth}}
42: \def\csd{\mathbf{depth}}
43: \def\S{\mathbf{S}}
44: \def\setS{\text{S}}
45: \def\T{\mathbf{S}''}
46: \def\U{\mathbf{U}}
47: \def\V{\mathbf{V}}
48: \def\f{\mathbf{f}}
49: \def\g{\mathbf{g}}
50: \def\core{\operatorname{core}}
51: \def\cone{\operatorname{cone}}
52: \def\conv{\operatorname{conv}}
53: \def\aff{\operatorname{Aff}}
54: \def\south{\operatorname{south}}
55: \def\north{\operatorname{north}}
56: \def\interior{\operatorname{int}}
57: \def\bara{B\'ar\'any}
58: \def\cara{Carath\'eodory}
59: \def\R{\mathbb{R}}
60: \def\E{\text{E}}
61: \def\X{\text{X}}
62: \def\P{\text{H}}
63: \def\Sph{\mathbb{S}}
64: \def\B{\mathbb{B}}
65: \def\m{\operatorname{m}}
66: \def\zero{{\bf 0}}
67: \def\Cap{\operatorname{Cap}}
68: \def\Can{\operatorname{Can}}
69: \def\Pole{\operatorname{Pole}}
70: \def\Capbar{\overline{\text{Cap}}}
71:
72:
73: %%% HEADERS
74: \title{Colourful Simplicial Depth}
75:
76: \author{Antoine Deza, Sui Huang, Tamon Stephen and Tam{\' a}s Terlaky}
77:
78: \address{Advanced Optimization Laboratory,
79: Department of Computing and Software,
80: Faculty of Engineering,
81: 1280 Main St.~West,
82: McMaster University,
83: Hamilton, Ontario,
84: Canada
85: L8S 4K1.
86: }
87:
88: \email{\{deza,huangs3\}@mcmaster.ca, tamon@optlab.mcmaster.ca, terlaky@mcmaster.ca}
89:
90: \thanks{Research supported
91: by NSERC Discovery grants for the four authors,
92: by the Canada Research Chair program for the first and last authors
93: and by a MITACS grant for the second and third authors.}
94:
95: \subjclass[2000]{52C45, 52A35}
96:
97:
98: %
99: %%% BEGIN HERE %%%
100: %
101: \begin{document}
102:
103:
104: \def\minus{-}
105: \setcounter{page}{1}
106:
107: \maketitle
108:
109: \begin{abstract}
110: Inspired by \bara's colourful \cara~theorem \cite{Bar82}, we introduce
111: a colourful generalization of Liu's simplicial depth \cite{Liu90}.
112: We prove a parity property and conjecture that the minimum
113: colourful simplicial depth of any core point in any $d$-dimensional
114: configuration is $d^2+1$ and that the maximum is $d^{d+1}+1$.
115: We exhibit configurations attaining each of these depths,
116: % with core points of depth $d^2+1$ and $d^{d+1}+1$,
117: and apply our results to the problem of bounding
118: monochrome (non-colourful) simplicial depth.
119: \end{abstract}
120:
121: %
122: % INTRODUCTION
123: %
124: \section{Introduction}\label{se:intro}
125:
126: In statistics there are several measures of the depth of a
127: point $p$ in ${\R}^d$ relative to a fixed set $S$ of sample points.
128: Two recent surveys on data depth are \cite{Alo05} and \cite{FR05},
129: see references therein.
130: % The depth is a measure of how representative $p$ is of the sample.
131: The depth measure we are interested in is the {\it simplicial depth}
132: of $p$, which is the number of simplices generated by points
133: in $S$ that contain $p$.
134: A point of maximum simplicial depth can be viewed as a type of
135: $d$-dimensional median. We would like to obtain a lower bound for
136: the depth of simplicial medians.
137:
138: To do this, we consider a generalized
139: problem where the sample points are colourful. That is, in
140: dimension $d$ we consider sample points given in each of at least
141: $(d+1)$ colours. Then we define the {\it colourful simplicial depth} of a
142: point $p$ relative to this sample to be the number of {\it colourful
143: simplices} (i.e.~simplices with one vertex of each colour) that
144: contain $p$. We focus on the situation where the point $p$ is
145: in the intersection of the convex hulls of the individual colours,
146: which is called the {\it core} of the configuration.
147:
148: If $p$ is a core point we would typically
149: expect the simplicial depth of $p$ to be more than exponential
150: in $d$. However, we exhibit configurations where $p$ is a core point
151: but is contained in only $d^2+1$ colourful simplices. We
152: conjecture that any core point $p$ of any $d$-dimensional
153: colourful configuration is contained in at least $d^2+1$
154: colourful simplices.
155: Along the way, we notice that both in the colourful and
156: monochrome cases the simplicial depth of points in general position
157: (relative to the sample set) sometimes has pleasant parity properties.
158: We conclude by mentioning some other natural % computational geometry
159: problems relating to the colourful
160: and monochrome simplicial depth.
161:
162: %
163: % BACKGROUND
164: %
165: \section{Definitions and Background}\label{se:defs}
166:
167: \subsection{Simplicial Depth}\label{se:depth}
168: The (closed) simplicial depth of a point $p$ relative to a set
169: $S$ of $n=|S|$ points in ${\R}^d$ is the number of (closed) simplices
170: generated by sets of $(d+1)$ points from $S$ containing $p$
171: in their convex hull.
172: This was introduced by Liu \cite{Liu90} as a measure of how
173: representative $p$ is of the points in $S$.
174: Denote the simplicial depth of $p$ relative to $S$ as $\sd_S(p)$.
175: The simplicial depth of $p$ can be interpreted
176: as the probability that $p$ lies in a random simplex of $S$
177: times a constant factor of $n^{d+1}$ if we sample points from $S$
178: uniformly with replacement, or times $\binom{n}{d+1}$ if we
179: sample without replacement.
180:
181: We are most interested in the case when $S \cup \{p\}$ is in
182: {\it general position}, that is for all $k < d$ there are
183: no $k$-dimensional affine
184: subspaces contain $k+2$ points from $S \cup \{p\}$.
185: With this assumption, $p$ will always be in the
186: interior of any simplices that contain it, so the notions of closed
187: and open simplicial depth coincide.
188: Without this assumption the closed simplicial depth will be larger.
189:
190: For a set of points $S$, define $f(S)$ to be the maximum simplicial
191: depth of a point $p$ relative to $S$, that is:
192: \begin{equation}\label{eq:maxf}
193: f(S) = \max_{p \in R^d} \sd_S(p)
194: \end{equation}
195: A point $p$ maximizing $f(S)$ can be understood as a
196: higher dimensional median point. We will call any such $p$
197: a {\it simplicial median}. Indeed for $d=1$, this is the
198: usual definition of a median ${\R}$. In higher dimensions,
199: this definition retains many
200: desirable properties of the median, such as affine
201: invariance and a high breakdown point (see e.g.~\cite{Alo05},
202: \cite{FR05}, \cite{GSW92}, \cite{Liu90}).
203: However, this maximum will not be attained at a point in general position.
204:
205: We will consider a similar quantity, the maximum simplicial
206: depth of a point $p$ that maintains $S \cup \{p\}$
207: in general position:
208: \begin{equation}\label{eq:maxg}
209: g(S) = \max_{S \cup \{p\} \text{ in general position}} \sd_S(p)
210: \end{equation}
211: Equivalently, $g$ is the maximum open simplicial depth of a point
212: ${\R}^d$. In this way the definition of $g$ can be extended to
213: the case when $S$ is not in general position.
214: While the maximum in (\ref{eq:maxf}) will be
215: attained on a discrete set of points in ${\R}^d$, the
216: maximum in (\ref{eq:maxg}) will be attained on an open set.
217: For non-empty $S$, we will have $g(S)<f(S)$.
218:
219: \subsection{Colourful Simplicial Depth}\label{se:csd}
220: Now consider a situation where points are given in each of $r \ge d+1$
221: colours. Then the sample consists of colourful sets
222: $S_1, S_2, \ldots, S_r$ which define a colourful configuration $\S$.
223: In the following we use a bold font for colourful objects.
224: A {\it colourful simplex} from these sets
225: is any simplex whose vertices are chosen from distinct sets.
226: We define $\csd_\S(p)$ the {\it colourful simplicial depth} of $p$
227: relative to the configuration $\S$ as the number of colourful simplices
228: containing $p$. As with monochrome simplicial depth, colourful
229: depth can be interpreted probabilistically. In the case where
230: $r = d+1$, colourful depth corresponds to specifying separate
231: distributions for each vertex of the simplex.
232: Dividing the depth by $|S_1| \cdot |S_2| \cdot \ldots \cdot|S_{d+1}|$
233: gives the probability that $p$ lies in a random colourful simplex
234: (sampled uniformly).
235:
236: A choice of sets $S_1, \ldots, S_r$ specifies a {\it colourful
237: configuration} $\S$ of points. We call the intersection of the
238: convex hulls of the $S_i$'s in a configuration
239: the {\it core} of $\S$. \bara~proved
240: that core points are contained in some colourful simplex;
241: this is known at the Colourful \cara~ Theorem \cite{Bar82}.
242: In the remainder of the paper,
243: except where noted, we assume that all configurations and $p$ are in general
244: position and have a non-empty (hence full-dimensional) core.
245: We remark that our results hold under weaker conditions, such as
246: $p$ not lying on any hyperplanes generated by points from the
247: configuration.
248:
249: \subsection{Background}\label{se:background}
250: Even before the notion of simplicial depth was studied in statistics,
251: the question of computing bounds for $f(S)$ and $g(S)$ given $n$ and $d$
252: was studied in
253: the combinatorics and computational geometry communities.
254: The two-dimensional question dates back at least to K{\'a}rteszi \cite{Kar55}
255: who showed that for $n$ points in the plane, $g(S)$ is at most
256: $\frac{1}{24}(n^3-n)$ for odd $n$ and at most $\frac{1}{24}(n^3-4n)$
257: for even $n$,
258: and showed that these bounds were attained when $S$ is the set of
259: vertices of a regular $n$-gon.
260: In the early 1980's, Boros and F{\"u}redi \cite{BF84}
261: showed $g(S)$ is at least $\frac{n^3}{27} + O(n^2)$, and gave configurations
262: achieving this bound.
263:
264: In a beautiful paper, \bara~\cite{Bar82} gave bounds for the
265: monochrome simplicial depth in dimension $d$ as an application
266: of his Colourful \cara~Theorem. He obtained a lower bound by
267: showing that after colouring the points, some point $p$ must be
268: contained in many colourful simplices.
269: A key point of \bara's proof is that a core point $p$ of
270: a colourful configuration must lie in at least {\it one} colourful
271: simplex. Using this fact, for a set $S$ of $n$ points in general
272: position in ${\R}^d$ \bara~obtains a lower bound of:
273: \begin{equation}\label{eq:lb}
274: g(S) \ge \frac{1}{(d+1)^{d+1}}\binom{n}{d+1}+O(n^d)
275: \end{equation}
276: This result is asymptotically sharp up to a constant factor as function
277: of $n$ (for fixed $d$). However, as \bara~remarks, the constant is
278: probably quite far from the truth. Indeed, he gives a sharp upper
279: bound of:
280: \begin{equation}\label{eq:ub}
281: g(S) \le \frac{1}{2^d (d+1)!} n^{d+1} + O(n^d)
282: \end{equation}
283: We speculate that the true lower bound is not much less than the
284: upper bound.
285:
286: One way to improve (\ref{eq:lb}) would be to show that a core point $p$
287: must lie in more than one colorful simplex. In \bara's original
288: paper, he notes that $p$ must in fact lie in at least $(d+1)$ colourful
289: simplices, thereby improving (\ref{eq:lb}) to:
290: \begin{equation}\label{eq:lb2}
291: g(S) \ge \frac{1}{(d+1)^{d}}\binom{n}{d+1}+O(n^d)
292: \end{equation}
293: More generally, if we could show that any core point $p$ of a
294: $d$-dimensional configuration is contained
295: in at least $\mu(d)$ simplices, then we can improve the constant
296: in equation (\ref{eq:lb}) by a factor of $\mu(d)$.
297:
298: %
299: % MAIN
300: %
301: \section{Colourfully Covering the Core}\label{se:covering}
302: This leads us to ask: What is the minimum number $\mu(d)$ of simplices that
303: can contain a core point $p$ in a colourful configuration?
304: Given a colourful configuration $\S$ with colourful
305: sets $S_1, \ldots, S_r$ we can define:
306: \begin{equation}\label{eq:m}
307: \m(\S)= \min_{p \in \core(\S)} \csd_\S(p)
308: \end{equation}
309: We remark that if $\core(\S)$ has a non-empty interior, the
310: minimum in (\ref{eq:m}) will be attained on an open set of
311: points that are in general position relative to $\S$.
312:
313: In this notation, our objective is to find the minimum value of $\m(\S)$
314: over all configurations $\S$ with full-dimensional core in dimension $d$.
315: For a fixed $d$, it is clear that some configuration with
316: $(d+1)$ points in each of $(d+1)$ colours attains this minimum,
317: which depends only on the dimension.
318: Hence we can define:
319: \begin{equation}\label{eq:mu}
320: \mu(d)= \min_{\text{$d$ configurations } \S, ~ p \in \core(\S)} \csd_\S(p)
321: \end{equation}
322:
323: One might suppose that $\m(\S)$ is often large.
324: As a thought experiment, consider choosing a configuration
325: at random. If we take $(d+1)$ points in
326: ${\R}^d$ from a distribution that is ``nice'' and centrally
327: symmetric about the origin \zero, the probability that $\zero$ is
328: contained in their convex hull is $\frac{1}{2^d}$ (see e.g.~\cite{WW01}).
329: This suggests that for random $\S$, a typical value for
330: $\csd_\S(\zero)$ would be $\frac{1}{2^d}(d+1)^{d+1}$.
331: For a set $S$ of $(d+1)^2$
332: points in the plane, plugging this value into \bara's analysis
333: gives us an estimate of $g(S)$ very close to \bara's
334: upper bound (\ref{eq:ub}). However, it is not immediately clear
335: if we should expect $\m(\S)$ to be much smaller than $\csd_\S(\zero)$.
336:
337: If we take a configuration $\S^\triangle$ with
338: $S^\triangle_1$ given by $(d+1)$ points
339: in general position and
340: $S^\triangle_1 = S^\triangle_2 = \ldots = S^\triangle_{d+1}$
341: we get $\m(\S^\triangle) = (d+1)!$. In Section~\ref{se:cons} we
342: exhibit a configuration $\S^\minus$ with $\m(\S^\minus) \le d^2+1$.
343:
344: In the remainder of the paper, except where noted, we consider
345: configurations with $(d+1)$ points in each of $(d+1)$ colours.
346:
347: \subsection{Preliminaries}\label{se:prelim}
348: In \cite{BO97}, \bara~and Onn consider the problem of
349: {\it colourful linear programming}. This is the
350: algorithmic version of the colourful \cara~ problem:
351: Given a core point $p$, how can we {\it find} a
352: colourful simplex containing $p$?
353: They begin with some preprocessing which is also helpful here.
354:
355: Take a colourful configuration $\S$ of $(d+1)$ colourful sets
356: in $\R^d$, $\S = \{S_1, \ldots, S_{d+1}\}$.
357: Take $p \in \interior(\core(\S))$.
358: Without loss of generality we assume that the core point $p=\zero$.
359: Given any finite set of points $T \subseteq \R^d$, scaling the points of
360: $T$ does not affect whether $\zero$ lies in the convex hull
361: of $T$ since the coefficients in a convex combination can
362: themselves be rescaled.
363: This allows us to normalize
364: $\S$ by rescaling its points to unit vectors.
365:
366: Let $\conv(T)$ be the convex hull of the points in $T$ and
367: $\cone(T)$ be the set of non-negative linear combinations
368: of points of $T$. A cone is {\it simplicial} if it
369: can be generated by a set of $d$ linearly independent points
370: in $\R^d$.
371: If $T \subseteq \R^d$ is a set of points, $\zero \notin T$, but
372: $\zero \in \conv(T)$, then $\cone(T)$ must contain a non-trivial
373: linear subspace of $\R^d$. A closed, convex cone is called
374: {\it pointed} if it does not contain such a subspace, so we
375: summarize this as:
376: \begin{lemma}\label{le:cones}
377: Given any finite set of non-zero points $T \subseteq \R^d$,
378: $\zero$ is in $\conv(T)$ if and
379: only if $\cone(T)$ is not pointed.
380: \end{lemma}
381: When $T$ is a finite set of points on the unit $d$-sphere
382: $\Sph^d \subseteq \R^d$,
383: Lemma~\ref{le:cones} is equivalent to saying that $\zero \in \conv(T)$
384: if and only if $T$ is not contained in any open
385: hemisphere of $\Sph^d$. One direction is proved by building a hemisphere
386: from a hyperplane through $\zero$ whose normal lies in the interior of
387: $\cone(T)$ when this cone is pointed. The other direction is
388: proved by observing that an open hemisphere never contains both
389: a point $p$ and its antipode $-p$.
390:
391: We would like to put Lemma~\ref{le:cones} in a form that is convenient
392: for counting how many simplices generated from $T$ contain \zero.
393: To do this, we find it helpful to think about the
394: antipode of one of the points.
395: \begin{lemma}\label{le:antipodes}
396: If $T=\{p_1, p_2, \ldots, p_{d+1}\}$ is a set of non-zero affinely
397: independent points in $\R^d$,
398: $\zero$ is in $\conv(T)$ if and only if the antipode
399: $-p_{d+1}$ is in $\cone(p_1, p_2, \ldots, p_d)$.
400: \end{lemma}
401: \begin{proof}
402: Let $K = \cone(p_1, p_2, \ldots, p_d)$.
403: Since $K$ is a cone generated by $d$ linearly independent points in
404: $\R^d$, $K$ is simplicial and hence pointed.
405: If $-p_{d+1} \in K$, then we can write it as a conic combination
406: of the remaining $p_i$, that is:
407: ${-p_{d+1} = \sum_{i=1}^d a_i p_i}$
408: for some $a_1, \ldots, a_d \ge 0$. Moving the $p_{d+1}$ term
409: to the right hand side of the equation and dividing by $1+\sum_{i=1}^d a_i$
410: gives $\zero$ as a convex combination of the $p_i$'s.
411: If $-p_{d+1}$ is not in $K$, then we can
412: strictly separate $-p_{d+1}$ from $K$ with a hyperplane $H$ through \zero.
413: Then both $K$ and $p_{d+1}$ lie strictly on the same side of $H$, and the
414: cone generated by $T$ must be pointed.
415: \end{proof}
416:
417: \subsection{A Variational Approach}\label{se:var}
418: Take a point $p$ from a finite set $S \in \R^d$.
419: Call a simplex generated by points in $S$ a {\it $p$-simplex} if
420: $p$ is one of the points used to generate the simplex, and
421: call a $p$-simplex {\it zero-containing} if it contains $\zero$ in its interior.
422: Define $z_S(p)$ to be the number of zero-containing $p$-simplices for a
423: given $S$.
424:
425: Lemma~\ref{le:antipodes} tells us that $z_S(p)$ is the number of
426: simplicial cones generated by $S \setminus \{p\}$ that contain $-p$.
427: We find it useful to think about what happens to $z_S(p)$
428: if we move $p$ while fixing the remaining points of $S$.
429: This is particularly illustrative if we confine $p$ to the surface
430: of the unit sphere $\Sph^{d}$ centred at \zero.
431:
432: Let $U = S \setminus \{p\}$ with $|U| = u$.
433: Initially $z_S(p)$ will be the number of simplicial cones generated
434: by sets of $d$ points from $U$ that contain $-p$. Now consider
435: what happens as $p$ (and hence $-p$) move. The value of $z_S(p)$ will
436: stay fixed until $-p$ crosses the boundary of some simplicial cone
437: from $U$. These boundaries are defined by the hyperplanes
438: generated by $\zero$ and sets of $(d-1)$ points from $U$.
439: Taking all $(d-1)$ sets from $U$, we can generate all such boundaries.
440: They divide the surface of $\Sph^d$ into open {\it cells} that are
441: $(d-1)$-dimensional open sets. We can define the {\it depth} of a cell
442: of $S$ to be the number of simplicial cones generated by $S$ containing
443: any given point in the interior of the cell.
444:
445: Consider moving $p$ along the surface of $\Sph^d$ to a new point
446: $p'$. If $-p$ and $-p'$ are in the same cell, we will have
447: $z_S(p)=z_S(p')$. Now suppose $-p$ is in a cell $C$ adjacent to
448: the cell $C'$ containing $-p'$.
449: Then as we move from $-p$ to $-p'$ we cross a single hyperplane $H$
450: defined by a set $U^0$ of $(d-1)$ points from $U$ belonging to $H$.
451: Let's say that
452: $-p$ is on the left of $H$ and $-p'$ is on the right.
453: For the moment we will assume that only $(d-1)$ points of $U$
454: lie on $H$.
455: Let $U^-$ be the
456: set of $k$ points from $U$ on the left of $H$, and $U^+$ be the
457: $u-k-(d-1)$ points from $U$ on the right. Since $-p$ is in
458: a cell bordered by $H$, it lies in the cone defined by the points
459: from $U^0$ and $x$ for any point $x \in U^-$. On the other hand,
460: $-p$ is separated by $H$ from the cones formed by $U^0$ and $y$
461: for any $y \in U^+$. Hence $-p$ is contained in exactly $k$
462: simplicial cones from $S$ generated by $U^0$ and a single other
463: point. Similarly, $-p'$ is contained in exactly $u-k-(d-1)$ such
464: cones. Simplicial cones that do not contain $U^0$ in their generating
465: set will not have $H$ as a facet, so they will contain $-p$ if and only
466: if they contain $-p'$. Suppose $-p$ is in $l$ such cones.
467: Then $z_S(p)=l+k$, while $z_S(p')=l+u-k-(d-1)$.
468:
469: We conclude that given the value of $z_S(p)$ at some point $p$,
470: we can in principle compute $z_S(p')$ for any other point $p'$
471: by tracing a path from $-p$ to $-p'$, and seeing how each hyperplane
472: generated from points in $U = S \setminus \{p\}$ divides the points
473: of $U$. To do this formally, we need a topological lemma that
474: says we can always draw a path between two points on $\Sph^d$ that
475: crosses only hyperplanes from $U$ (as opposed to passing through
476: cones generated by fewer than $(d-1)$ points).
477: This reduces to the following fact which can be proved using
478: algebraic topology, see for example \cite{Mun84}:
479: \begin{lemma}\label{le:top}
480: The sphere $\Sph^d$, a $(d-1)$ dimensional manifold, remains path
481: connected after
482: removing finitely many $(d-3)$-dimensional manifolds. % from it.
483: \end{lemma}
484:
485: \subsection{Parity}\label{se:parity}
486: The variational approach to computing $z_S(p)$ explains the
487: following parity phenomenon:
488: \begin{proposition}\label{pr:mincfg}
489: For any colourful configuration $\S$ of $(d+1)$ points in each of $(d+1)$
490: colours in {\rm odd} dimension $d$ and any point $p$ with $\S$ and $p$
491: in general position, the colourful simplicial depth of $p$
492: with respect to $\S$ is {\rm even}.
493: \end{proposition}
494: The authors were surprised by this fact while experimenting with
495: configurations. However, it is easy to explain this
496: via a colourful version of the method described in
497: Section~\ref{se:var}. Suppose we begin with a % (invalid)
498: configuration $\S^0$ with $(d+1)$ points in each of $(d+1)$
499: colours clustered near the North Pole of $\Sph^d$. (We remarked
500: in Section~\ref{se:prelim} that it is sufficient to consider
501: configurations on the surface of $\Sph^d)$.
502: Then we can move one point at a time from its initial position
503: in $\S^0$ to its final position in $\S$ generating a sequence of
504: configurations $\S^0, \S^1, \S^2, \ldots, \S^{(d+1)^2}=\S$.
505: Clearly $\csd_{\S^0}(\zero)=0$. As we move a given point $p_i$ of colour
506: $j$ from its
507: initial position in $\S^0$ (and $\S^i$) to its final position
508: in $\S$ (and $\S^{i+1}$), we need only to know what happens
509: when the antipode $-p_i$ crosses colourful hyperplanes defined by
510: a set of $(d-1)$ points of $(d-1)$ colours, and not of colour $j$.
511: Such a colourful hyperplane $H$ will miss only one other colour, $j'$.
512: There will be $k$ points of colour $j'$ on one side of $H$,
513: and $(d+1-k)$ on the other side.
514: Here we are assuming that the points from $\S$ are in general position,
515: but we can argue by continuity that this assumption is not necessary.
516: As $-p_i$ crosses
517: $H$ the number of simplicial cones containing $-p_i$ generated by points
518: from $H$ and a point of colour $j'$ changes from $k$ to $(d+1-k)$.
519: As long as $(d+1)$ is even, the parity doesn't change.
520:
521: Examining this proof, we can see that Proposition~\ref{pr:mincfg} can
522: be generalized:
523: \begin{theorem}\label{th:cparity}
524: If $\:\S=\{S_1, S_2, \ldots, S_r\}$ is a $d$-dimensional colourful
525: configuration of points
526: and for each $i=1,2, \ldots, r$ we have $|S_i|$ even, and $p$ is
527: and any point $p$ with $\S$ and $p$ in general position, then the colourful
528: simplicial depth of $p$ with respect
529: to $\S$ is also even.
530: \end{theorem}
531:
532: For monochrome depth, as we move point $p$ around $\Sph^d$
533: we need to consider all possible hyperplanes formed from
534: $S \setminus \{p\}$. Using the same reasoning as
535: Theorem~\ref{th:cparity} we get:
536: \begin{theorem}\label{th:mparity}
537: If $S$ is a set of $n$ points in $\R^d$, and $n-d$ is even,
538: and $p$ is a point such that $S \cup \{p\}$ is in general position,
539: then the simplicial depth of $p$ with respect to $S$ is even.
540: \end{theorem}
541:
542: \begin{remark}
543: The variational approach suggested in Section~\ref{se:var}
544: has appeared in various guises in discussions of monochrome
545: simplicial depth. In particular, it underlies many of the
546: algorithms suggested for computing monochrome simplicial depth.
547: Several such algorithms have been proposed recently, see for
548: example the discussion in \cite{Alo05}.
549: Many of these focus on the 2-dimensional problem,
550: \cite{GSW92} and \cite{CO01} use variational ideas in
551: 3- and 4-dimensional algorithms.
552:
553: For this reason,
554: we believed that Theorem~\ref{th:mparity} existed as
555: folklore for some time.
556: Baker remarks on the two-dimensional version in a
557: recreational mathematics note \cite{Bak78}, but this fact,
558: which impressed the authors with its simple elegance,
559: seems curiously neglected in the literature.
560: We speculate that one reason for this
561: is that in statistics the focus has been on computing the depth of the
562: sample points themselves, which are not in general position and
563: do not retain nice parity conditions.
564: \end{remark}
565:
566: \subsection{Configurations with Small Minimal Colourful Depth}\label{se:cons}
567: We now describe how to build a colourful configuration $\S^\minus$
568: that contains $\zero$ in its core, but where only $d^2+1$ colourful
569: simplices contain \zero. Our strategy is to fix the first $d$
570: colourful sets $S_1^\minus, S_2^\minus, \ldots, S_d^\minus$ and then consider possible
571: placements of $(d+1)$ points $p_1, p_2, \ldots, p_{d+1}$ to form
572: $S_{d+1}^\minus$. We will place the points from
573: $S_1^\minus \cup S_2^\minus \cup \ldots \cup S_d^\minus$ on
574: the sphere $\Sph^d$ in such a way that some regions of $\Sph^d$
575: are sparsely covered by simplices from
576: $S_1^\minus \cup S_2^\minus \cup \ldots \cup S_d^\minus$.
577:
578: We begin by fixing $\epsilon = \frac{1}{100d}$.
579: We will place the points from $\S^\minus$ in three locations on $\Sph^d$.
580: The first on the {\it Tropic of Capricorn}, which we define to be
581: the set of points on $\Sph^d$ whose $d$th coordinate is $-2\epsilon$.
582: The second is on the {\it Tropic of Cancer}, whose $d$th coordinate
583: is $\epsilon$. % (** see picture?)
584: The two tropics are topologically copies of $\Sph^{d-1}$, but
585: unlike their namesakes they are not equally spaced from the equator.
586: The final region is the {\it polar region}
587: of points in $\Sph^d$ which are within $\epsilon$ of the North Pole
588: $p_{\north} = (0,0,\ldots,0,1)$.
589: \begin{figure}[h!bt]
590: \begin{center}
591: \includegraphics[height=5.4cm]{regions}
592: \caption{3-dimensional illustration of the regions used in constructing
593: $\S^\minus$.}\label{fig:regions}
594: \end{center}
595: \end{figure}
596:
597: Now let's fix the positions of the points
598: $\{x_1, x_2, \ldots, x_{d+1}\} \in S_1^\minus$.
599: Take:
600: $$x_1 = (\sqrt{1-4\epsilon^2}, 0, 0, \ldots, 0, -2\epsilon)
601: \qquad \text{ and } \qquad
602: x_2 = (-\sqrt{1-\epsilon^2}, 0, 0, \ldots, 0, \epsilon)$$
603: Note that the line segment between $x_1$ and $x_2$ passes just
604: below the origin in the sense that it contains a point whose
605: first $(d-1)$ coordinates are 0, and whose $d$th coordinate is
606: negative (and small).
607: We now place the remaining points $x_3, \ldots, x_{d+1}$ in the
608: polar region in such a way as to ensure that
609: $\zero \in \interior(\conv(S_1^\minus))$.
610: For $d=2$ we can do this by placing $x_3$ at the North Pole.
611: For $d \ge 3$ we can place the points on the section of the
612: Arctic Circle (points with distance $\epsilon$ to the North Pole)
613: with zero initial coordinate. Topologically the Arctic Circle is a copy of
614: $\Sph^{d-2}$; we can take $x_3, \ldots, x_{d+1}$ to be the vertices
615: of a regular simplex inscribed on this sphere.
616:
617: The points of colours $2,3, \ldots, d$ are chosen similarly.
618: The first points from each of the $d$ colours are arranged in a regular
619: simplex on Capricorn. The remaining points in the
620: same relative position to the first point, so that each $S_i^\minus$
621: is a rotation of $S_1^\minus$ around the $d$th coordinate axis.
622: In particular, for each $i=1,2, \ldots, d$, the second point
623: of $S_i^\minus$ will lie on Cancer and the final $(d-1)$ points will
624: lie in the polar region.
625:
626: We finish our construction by considering possible placements
627: of the points $p_1, \ldots, p_{d+1}$ of $S_{d+1}^\minus$.
628: We want to place the $p_i$'s in such a way that their antipodes
629: (the $-p_i$'s) are contained in few colourful simplicial cones generated
630: from $\S^\minus$.
631:
632: Consider the cell $C_{\south}$ defined by colours $1,\ldots,d$ of
633: $\S^\minus$ on $\Sph^d$ which contains the South Pole
634: $p_{\south}=(0,0, \ldots, 0,-1)$.
635: We claim this is exactly the intersection of $\Sph^d$ with
636: the single colourful simplicial cone $K_{\Cap}$ defined by the $d$
637: colourful points on Capricorn.
638: This follows since any other colourful cone
639: is generated by a set of $d$ coloured points
640: chosen from Capricorn, Cancer and the northern polar region.
641: Fix such a cone and call these sets ${G_{\Cap}}$, $G_{\Can}$ and
642: $G_{\Pole}$ and let $K_G = \cone(G_{\Cap} \cup G_{\Can} \cup G_{\Pole})$.
643: We assume that we have $|G_{\Cap}| < d$.
644: We need to show that
645: $\interior(K_{\Cap}) \cap \interior(K_G) = \emptyset$.
646: To do this, we find a hyperplane separating $K_{\Cap}$ and $K_G$.
647: If $G_{\Cap}=\emptyset$ the hyperplane through the Equator will do.
648: Otherwise, take the colours from $G_{\Cap}$ and consider any
649: facet $F$ of $K_{\Cap}$ containing generators of each of these colours.
650: Then $F$ separates $K_{\Cap}$ from all the polar points and all the
651: Cancer points of colours from $\{1,2,\ldots,n\} \setminus G_{\Cap}$.
652: (To be absolutely
653: proper, in higher dimension we would have to move Capricorn up
654: towards the equator to ensure the separation of the Cancer points,
655: i.e. we would have to reduce the constant $2\epsilon$ to
656: $(1+\delta)\epsilon$ for some $\delta>0$.)
657: This completes the proof. We conclude that the cell $C_{\south}$
658: is covered only by the colourful cone $K_{\Cap}$ and
659: closely approximates the spherical cap bounded by Capricorn.
660:
661: It is a good strategy to place the antipodes $-p_i$ in $C_{\south}$.
662: If we do this for all of $S_{d+1}^\minus$, however,
663: the resulting configuration will not have $\zero \in \conv(S_{d+1}^\minus)$
664: ($S_{d+1}^\minus$ would certainly be contained in an open hemisphere).
665: So we must have at least one antipode, say $-p_1$ above Capricorn.
666: Indeed, if we place the remaining $-p_i$ below Capricorn, we
667: would need to have $-p_1$ above the ring of the antipodes
668: of Capricorn. More precisely, this is the set of points on $\Sph^d$
669: with final coordinate value exactly $2\epsilon$. In particular,
670: it is above Cancer.
671:
672: Let $A=\{a_1, a_2, \ldots, a_d\}$ be the points from
673: $S_1^\minus, S_2^\minus, \ldots, S_d^\minus$
674: on Capricorn. Similarly, let $B=\{b_1, b_2, \ldots, b_d\}$ be
675: the points on Cancer. Let's count how many simplicial cones
676: from $\S^\minus$ must contain $-p_1$ if we place $-p_1$ above Cancer.
677: To do this, we start with $-p_1$ in $C_{\south}$ and then move it
678: above Cancer noting which cell boundaries it crosses as suggested
679: in Section~\ref{se:var}. This structure of the cell boundaries
680: is a topological question, so we find it convenient to remove the
681: $p_{\south}$ and equate $\Sph^d$ with $\R^{d-1}$.
682:
683: With the exception of the single colourful cone that contains
684: $C_{\south}$, the colourful simplicial cones generated by $\S^\minus$
685: correspond to colourful simplices in $\R^{d-1}$.
686: The polar points on $\Sph^d$ will be clustered near the origin
687: in $\R^{d-1}$. Let $A'=\{a_1', \ldots, a_d'\}$ and
688: $B'=\{b_1', \ldots, b_d'\}$ be the projections of $A$ and $B$
689: respectively in $\R^{d-1}$.
690: Then $\conv(A')$ and $\conv(B')$ form nested simplices which contain
691: the projection of the polar region. The boundaries of the colourful
692: simplicial cones on $\Sph^d$ map to facets of simplices in $\R^{d-1}$;
693: both are defined by sets of $(d-1)$ colourful points.
694: Moving $-p_1$ from below Capricorn to above
695: Cancer corresponds to moving $-p_1'$ from outside $\conv(A')$ to
696: inside $\conv(B')$.
697:
698: Let us now see what simplicial facets $-p_1'$ must cross to do this.
699: If we keep $-p_1'$ far away from the $a_i'$ and $b_i'$'s themselves,
700: we can avoid any facets involving the polar points: These facets
701: involve at most $(d-2)$ generators from $A'$ and $B'$, and hence have
702: ends that are at most $(d-3)$ dimensional manifolds in
703: $\conv(A') \setminus \interior(\conv(B'))$. The ends can be avoided by
704: Lemma~\ref{le:top}.
705:
706: This still leaves $d 2^{d-1}$ colourful facets defined by choosing
707: $(d-1)$ colourful points from $A'$ and $B'$. We can enumerate them by
708: first choosing an index (colour) to omit and then representing the
709: choices of $a_i'$'s and $b_i'$'s by a 0-1 vector of length $(d-1)$.
710: Letting 0 represent the choice of an $a_i'$, $\conv(A')$ is bounded by
711: the facets defined by $d$ index choices and a vector of 0's, while
712: $\conv(B')$ is bounded by the facets defined by $d$ index choices
713: and a vector of 1's. In fact there are $2^d$ colourful simplices
714: generated by $A'$ and $B'$, and they are enumerated by 0-1 vectors of
715: length $d$. Their facets are enumerated by choosing an
716: index to drop from the enumerating sequence. Therefore the sums
717: of the 0-1 vectors enumerating the facets of a given simplex
718: can differ by at most 1.
719:
720: Now start with $-p_1'$ outside $\conv(A')$.
721: To bring $-p_1'$ inside $\conv(B')$, we must start by bringing it
722: into $\conv(A')$. This involves crossing some boundary face of
723: $\conv(A')$, say the one defined by $a_1, \ldots, a_{d-1}$.
724: This is enumerated as $(d,0,0,\ldots,0,0)$. We can proceed through
725: facets $(d-1,0,0,\ldots,0,1)$, $(d-2,0,0,\ldots,0,1,1)$
726: until finally we cross $(1,1,1, \ldots, 1)$ into a cell of $\conv(B')$.
727: This involves $d$ facet crossings, which is minimal since at
728: each crossing we can only add a single 1 to the 0-1 part
729: of the enumerating vector.
730:
731: We claim that as $-p_1'$ crosses each facet, it makes a net
732: gain of $d-1$ containing simplices.
733: At the first facet, $(d,0,0,\ldots,0,0)$, $-p_1'$
734: leaves the single exterior simplex defined by the points $A'$
735: projected from Capricorn and enters the $d$ simplices defined
736: by $a_1', \ldots, a_{d-1}'$ and
737: the $d$ points of colour $d$ other than $a_d'$.
738: At subsequent facet crossings, the same thing happens for the remaining
739: colours: $-p_1'$ leaves the simplex defined by the crossing
740: facet and $a_i'$. As $-p_1'$ leaves, it enters the simplices
741: defined by this facet and the $d$ remaining points of colour $i$.
742: Hence the number of simplices containing $-p_1'$ immediately after
743: crossing into $\conv(B')$ is exactly $1+d(d-1)$.
744:
745: We will now return our attention to $\Sph^d$.
746: Denote by $C_p$ the cell containing $-p_1$ whose projection lies
747: inside $\conv(B')$. From our construction, $C_p$ is a cell above
748: Cancer. We want to claim that in fact it contains some point above
749: the set of antipodes of Capricorn, that is, a point whose antipode
750: is in $C_{\south}$.
751: This is a complicated geometric calculation. However, we
752: observe that nothing in our topological argument above changes if we
753: change the constant $2\epsilon$ in our definition of Capricorn to
754: $c \: \epsilon$ for any $c \ge 0$. In particular, the cell $C_p$
755: does not degenerate if we move the antipodes of Capricorn
756: towards Cancer by decreasing $c$ to 1.
757: Therefore for some $c>1$ (this condition maintains
758: $\zero \in \interior(\conv(S_i^\minus))$ for $i=1,\ldots,d$),
759: $C_p$ includes some point
760: above the antipodes of Capricorn. Any such $c$ and
761: point in $C_p$ would be sufficient for our construction.
762: We have used $c=2$ for concreteness and take it as an article
763: of faith that this is a small enough for our choice of $\epsilon$.
764: \begin{figure}[h!bt]
765: \begin{center}
766: \includegraphics[height=5.4cm]{placement}
767: \caption{Placement of points of colours
768: $1,2,3$ and antipodes of colour $4$ in
769: the 3-dimensional $\S^\minus$.}\label{fig:placement}
770: \end{center}
771: \end{figure}
772:
773: The construction can now be completed.
774: Take $-p_2$ to be the midpoint of shortest spherical segment between
775: Capricorn and $p_1$ (which lies below Capricorn).
776: Let $z < -2\epsilon$ be the final coordinate of $-p_2$ and
777: arrange the remaining points so that $-p_2, -p_3, \ldots, -p_{d+1}$
778: form a regular simplex on $\Sph^d \cap \{x \in \R^d | x_d=z\}$.
779: Then $\zero$ is in the convex hull of the $-p_i$ (and hence $S_{d+1}^\minus$).
780: Finally we can calculate $\csd_{\S^\minus}(\zero)$
781: from the location of the $-p_i$: $\zero$ lies in $1+d(d-1)$ colourful
782: simplices generated with $p_1$ and one simplex each including
783: $p_2, p_3, \ldots, p_{d+1}$. Hence:
784: $$\csd_{\S^\minus}(\zero)=1+d(d-1)+d = d^2+1$$
785:
786: \begin{remark}\label{re:alt}
787: There are other nice configurations with
788: $\csd_{\S^\minus}(\zero)=d^2+1$.
789: Consider a configuration $\S'$ similar to $\S^\minus$ but with the tropics
790: pushed to the north, taking Cancer's final coordinate to $3\epsilon$
791: and Capricorn's to $-\epsilon$. We can then move each of
792: $-p_1, \ldots, -p_d$ across Capricorn and the equator through a
793: single boundary facet. Finally place $-p_{d+1}$ at the South Pole.
794: Using the same analysis as above, we have $p_1, \ldots, p_d$ points
795: forming $1+(d-1)$ simplices containing \zero, and $p_{d+1}$
796: forming one such simplex for a total of $d^2+1$.
797:
798: Both $\S^\minus$ and $\S'$ have symmetry for the first $d$ colours,
799: but not the last one. We can also propose a configuration $\T$
800: with symmetry between all the colours. Follow the recipe for
801: $\S^\minus$ but place one point of each colour on Cancer and Capricorn
802: and place the remaining points in the polar region.
803: This brings a number of technical
804: difficulties, however. The points will not be in general
805: position, since the tropical hyperplanes include $(d+1)$ points.
806: It is also a bit less natural to evenly space $(d+1)$ points
807: on copies of $\Sph^{d-1}$, indeed for $d=2$ this construction
808: does not make sense. When there is a nice way to do this for $d \ge 3$
809: this (e.g.~4 points on $\Sph^2$) we may end up with some points
810: being antipodes. This would cause $\zero$ to be on the faces of
811: some simplices and increase its colourful simplicial depth.
812: Most of these problems can be fixed by perturbing $\T$, but even
813: so $\T$ is not well-suited to our proof technique.
814: One might also consider configurations that are not confined
815: to the sphere.
816: \end{remark}
817:
818: \subsection{Evaluating $\mu(d)$}\label{se:min}
819: The configuration $\S^\minus$ of Section~\ref{se:cons} satisfies
820: $\m(\S^\minus) \le d^2+1$ where $\m(\S)$ is the minimum colourful simplicial
821: depth of core point defined in (\ref{eq:m}).
822: We would like to prove that $\m(\S^\minus)=d^2+1$ and in fact
823: that for any colourful configuration $\S$ we will have
824: $\m(\S) \ge d^2+1$, or equivalently $\mu(d) \ge d^2+1$.
825: The second half of this proposition clearly
826: implies the first. We suggest it is also more approachable since
827: we can move the core point of minimum depth to $\zero$ during preprocessing,
828: whereas a direct attack on $\m(\S^\minus)$ requires understanding the
829: shape of the core of $\S^\minus$.
830:
831: \bara's original Colourful \cara~ theorem is exactly
832: that $\mu(d) \ge 1$. He further shows that
833: for any $\S$ any colourful point from $\S$ is part of some
834: generating set for a colourful simplex containing \zero. This
835: immediately yields $\mu(d) \ge d+1$. In $\S^\minus$
836: we see that $p_2, p_3, \ldots, p_{d+1}$ all generate a
837: {\it unique} colourful simplex containing $\zero$. Thus
838: the minimum number of colourful simplices containing
839: $\zero$ generated by an arbitrary point in a configuration is 1.
840: To get a stronger lower bound than $\mu(d) \ge d+1$ we need to
841: understand some global information about configurations.
842:
843: \begin{lemma}\label{le:depthd}
844: Fix the sets $S_1, \ldots, S_d$ from a colourful configuration $\S$
845: with $\zero$ in its core, and consider the cells created
846: on $\Sph^d$ by the colourful simplicial cones from these sets.
847: Then every cell has depth at least 1, and if there is a cell of
848: depth 1 it is unique and all other cells have depth at least $d$.
849: \end{lemma}
850: \begin{proof}
851: The fact that every cell has depth at least 1 is equivalent to the
852: fact that every colourful point generates some colourful simplex
853: that contains \zero, proved in \cite{Bar82}.
854: Suppose now that there is a cell $C$ of depth 1.
855: Any point exiting $C$ through a bounding hyperplane will be
856: exiting some colourful simplex. Since the depth of $C$ is 1,
857: this will always be the same simplex.
858: Thus the extreme points of $C$ must be a colourful set
859: $A=\{a_1, \ldots, a_d\}$ with $a_i \in S_i$ generating this simplex.
860: We can puncture $\Sph^d$ at $p \in C$ and project $\Sph^d \setminus \{p\}$
861: into $\R^{d-1}$. The $a_i$'s project to a set
862: $A'=\{a_1', \ldots, a_d'\}$ that forms a $(d-1)$-simplex
863: in $\R^{d-1}$.
864: The remaining colourful points project to points in $\conv(A')$.
865:
866: Take a point $q$ inside $\conv(A')$.
867: We want to show that $q$ is contained in at least $d$
868: colourful simplices in addition to $\conv(A')$ after projection.
869: To do this, it is sufficient to show that if we take any colourful
870: set $B'=\{b_1', \ldots, b_d'\}$ of projected points with $b'_i$ of
871: colour $i$ and $A' \cap B' = \emptyset$, then $q$ is in some colourful
872: simplex generated from points of $A' \cup B'$ with some generators
873: from $B'$. Equivalently, we want to show that $\conv(A')$ is
874: covered by colourful simplices generated from $A' \cup B'$
875: (excluding $\conv(A')$ itself from the covering).
876: Then by partitioning the projections of the colourful
877: points into $(d+1)$ colourful sets $A', B_1', B_2', \ldots, B_d'$
878: we get Lemma~\ref{le:depthd}.
879:
880: Consider the collection $\X$ of colourful $(d-1)$-simplices
881: generated by $A'$ and $B'$ in $\R^{d-1}$ and let
882: $\tilde{X}$ be the set of points
883: contained in the colourful simplices of $X$ other than $\conv(A')$.
884: The elements of $\X$ are all the simplices
885: formed by taking for each colour $i=1,2,\ldots,d$ either
886: $a_i'$ or $b_i'$ as a generating vertex.
887: This construction resembles the $d$-dimensional
888: cross-polytope $\beta_d$ (the dual of the $d$-cube), a regular polytope
889: in $\R^d$ with $2d$ vertices and $2^d$ facets.
890: The cross-polytope $\beta_d$ is generated by taking as vertices the
891: standard unit vectors $E^+ = \{e_1, \ldots, e_d\}$
892: and their negatives $E^- = \{-e_1, \ldots, -e_d\}$.
893: The facets of $\beta_d$ are the convex hulls
894: generated by choosing for each $i=1,\ldots,d$ either $e_i$ or $-e_i$.
895:
896: We can see that $\X$ is obtained from $\beta_d$ as follows:
897: We have $A' \cup B' \subset \R^{d-1}$.
898: Embed $\R^{d-1}$ as an affine subspace $\aff(A')$ in $\R^d$.
899: Take $\P$ to be an affine hyperplane in $\R^d$ parallel to $\aff(A')$.
900: For $i=1, \ldots, d$ let $p_i$ be the intersection
901: point of $\P$ with the line through $b_i'$ perpendicular to $\R^{d-1}$.
902: Let $P =\{ p_1, \ldots p_d\}$ and generate a set $Q$ of
903: $(d-1)$-simplices
904: by taking for each $i = 1, \ldots, d$ either $a_i'$ or $p_i$.
905: By construction $X$ is the projection of $Q$ into $\aff(A')$.
906: Now we claim that $Q$ is a continuous image of the facets of $\beta_d$.
907: We can exhibit such a map by first finding an affine transformation $T_1$
908: with $T_1(e_i)=a_i'$ for $i=1,\ldots,d$ and $T_1(\aff(E^-))=\P$.
909: Note that $T_1(\beta_d)$ is a polytope.
910: Then applying a further affine transformation $t_2$ to $\P$ with
911: $t_2(-e_i)=p_i$ for $i=1, \ldots, d$ and extending this to
912: $T_2$ on $\R^d$ so that $T_2$ fixes $\aff(A')$,
913: we see that the composition $T_2 \circ T_1$ is the required map.
914:
915: We proceed by contradiction. Assume that $\tilde{X}$
916: does not cover $\conv(A')$. Then we can find a retraction of
917: $\tilde{X}$ to its boundary $\partial(\conv(A'))$.
918: By composing $T_2$, the projection taking $Q$ onto $X$ and the
919: retraction of $\tilde{X}$, we get
920: a retraction of $T_1(\beta_d) \setminus \conv(A')$
921: onto $\partial(\conv(A'))$.
922: However, $T_1(\partial(\beta_d))$ is a $d$-dimensional
923: polytope topologically equivalent to $\Sph^d$ and hence
924: $T_1(\partial(\beta_d)) \setminus \conv(A')$ is topologically equivalent to a
925: $(d-1)$-dimensional disk $\B^{d-1}$. But $\partial(\conv(A'))$ is
926: topologically equivalent to $\Sph^{d-1}$ and a well-known theorem
927: of algebraic topology says that there does not exist a retraction
928: of $\B^{d-1}$ to $\Sph^{d-1}$ (see for example section 21 of \cite{Mun84}).
929: This is the required contradiction, hence the colourful simplices
930: of $\X \setminus \conv(A')$ cover $\conv(A')$.
931: \end{proof}
932:
933: \begin{corollary}\label{co:muge2d}
934: The minimum colourful simplicial depth of any core point in any
935: colourful configuration is at least $2d$.
936: That is, we have $\mu(d) \ge 2d$.
937: \end{corollary}
938: \begin{proof}
939: It suffices to prove this for a configuration $\S$ with $(d+1)$
940: in $(d+1)$ colours.
941: Observe that
942: if we have no cell of depth 1 then each of the $(d+1)$ points
943: of $S_{d+1}$ will generate at least two colourful simplices
944: containing \zero, and if we do have such a cell $C$, we must place
945: at least one point, say $p_1 \in S_{d+1}$ outside of $C$ to
946: get $\zero \in \conv(S_{d+1})$. Then $p_1$ is generates at least $d$ simplices
947: containing $\zero$ in addition to the $d$ required of the remaining
948: points in $S_{d+1}$.
949: \end{proof}
950:
951:
952: \subsection{The Two-dimensional Case}\label{se:2d}
953: We will briefly illustrate our methods by describing how core
954: points can be contained in configurations in $\R^2$.
955: Consider such a configuration $\S=\{X,Y,Z\}$ with core point $p$.
956: We assume general position, and as discussed in Section~\ref{se:prelim},
957: we may without loss of generality take the core point
958: $p=\zero$ and place the points of $\S$ on the unit circle $\Sph^2$.
959:
960: Then the points of $X$ and $Y$ divide $\Sph^2$ into six segments.
961: Let $X=\{x_1,x_2,x_3\}$, $Y=\{y_1,y_2,y_3\}$.
962: These points generate 9 simplicial cones and divide $\Sph^2$
963: into 6 segments. The boundaries between cones are simply the
964: rays through the $x_i$'s and $y_i$'s.
965: Because no three points of $X$ or $Y$ lie in the same half-circle,
966: each hyperplane through $\zero$ and $x_i$ divides
967: the $y_i$'s 2 to 1 and vice-versa.
968: Then as the antipode of a point from $Z$ crosses $x_i$ or $y_i$
969: the number of containing simplicial cones changes by exactly one.
970:
971: To get a configuration $\S^\minus$ where only 5 simplices contain 0,
972: we take $x_1=(-\sqrt{1-4\epsilon^2}, -2\epsilon)$,
973: $x_2=(\sqrt{1-\epsilon^2},\epsilon)$,
974: $x_3=(-\epsilon,\sqrt{1-\epsilon^2})$,
975: $y_1=(\sqrt{1-4\epsilon^2},-2\epsilon)$,
976: $y_2=(-\sqrt{1-\epsilon^2},\epsilon)$,
977: and
978: $y_3=(\epsilon,\sqrt{1-\epsilon^2})$.
979: Observe there is a large cell of depth 1 between $x_1$ and $y_1$.
980: The reader can verify that the sequence of colourful cell
981: depths is: 1,2,3,4,3,2.
982: \begin{figure}[h!bt]
983: \begin{center}
984: \includegraphics[height=1.2cm]{circle}
985: \caption{Covering depths for a circle with a depth 1 cell.}\label{fig:circle}
986: \end{center}
987: \end{figure}
988:
989: Let $Z=\{z_1, z_2, z_3\}$.
990: Place $z_2=(-\sqrt{1-9\epsilon^2},3\epsilon)$ and
991: $z_3=(\sqrt{1-9\epsilon^2},3\epsilon)$ so that their
992: antipodes lie between $x_1$ and $y_1$. They each generate one
993: simplex containing \zero. Finally, to ensure that $\zero \in \conv(Z)$,
994: we see that $-z_1$ must lie above $y_2$ and $x_2$. Take
995: $z_1=(-\sqrt{1-16\epsilon^2}, -4\epsilon)$.
996: Then $-z_1$ is contained in 3 colourful
997: simplicial cones generated by $X$ and $Y$. This configuration
998: has $\zero$ in the interior of its core and $\zero$ lies in $1+1+3=5$
999: colourful simplices.
1000: \begin{figure}[h!bt]
1001: \begin{center}
1002: \includegraphics[height=5cm]{min2d}
1003: \caption{Valid configuration $\S^\minus$ in dimension 2
1004: with $\csd_{\S^\minus}(\zero)=5$.}\label{fig:min2d}
1005: \end{center}
1006: \end{figure}
1007:
1008: Using the analysis in Section~\ref{se:min} we see that the
1009: cells generated by $X$ and $Y$ have colourful covering depth at
1010: least 1. If no cell attains this, then our configuration must
1011: yield at least 6 colourful simplices containing \zero.
1012: If some cell has depth 1, we can place at most two of the
1013: $-z_i$'s in this cell. The remaining $z_i$ must then have
1014: depth at least 2, for a minimum of 4. In fact, we can strengthen
1015: this to show that our configuration is minimal
1016: by observing that we cannot place all of $Z$ in two adjacent cells.
1017: We conclude that $\mu(2)=5$.
1018: A similar observation in three dimensions shows that
1019: $\mu(3) \ge 8$. Given the construction of Section~\ref{se:cons}
1020: and Proposition~\ref{pr:mincfg} we know that $\mu(3)$ is
1021: either 8 or 10.
1022: \bara~and Matou{\v s}ek \cite{BM06} have shown that
1023: $\mu(3) \ge 9$ which combined with Proposition~\ref{pr:mincfg}
1024: implies that
1025: $\mu(3)=10$.
1026:
1027: %
1028: % CONCLUSIONS
1029: %
1030: \section{Conclusions}\label{se:dis}
1031: Let us return to our original goals. Using the bound $\mu(d) \ge 2d$
1032: from Section~\ref{se:min}, we see that we can improve \bara's
1033: lower bound (\ref{eq:lb2}) for
1034: the depth of the monochrome simplicial median to:
1035: \begin{equation}\label{eq:lb3}
1036: g(S) \ge \frac{2d}{(d+1)^{d+1}}\binom{n}{d+1} + O(n^d)
1037: \end{equation}
1038: This is a modest improvement. Unfortunately, the construction
1039: in Section~\ref{se:cons} shows that simply bounding $\mu(d)$
1040: cannot give a stronger bound than:
1041: \begin{equation}\label{eq:lbc}
1042: g(S) \ge \frac{d^2+1}{(d+1)^{d+1}}\binom{n}{d+1} + O(n^d)
1043: \end{equation}
1044:
1045: Quite recently, Wagner proved exactly the bound (\ref{eq:lbc}) in
1046: his thesis \cite{Wag03} as a special case of his
1047: First Selection Lemma.
1048: This is, to our knowledge, the first improvement of
1049: (\ref{eq:lb2}) since \bara's original paper \cite{Bar82}.
1050: Wagner's result uses a continuous version of the
1051: Upper Bound Theorem for polytopes and other techniques from
1052: probability without any reference to colouring.
1053: We find the appearance of the constant $d^2+1$, which for
1054: us arrives from colourful combinatorics, quite remarkable.
1055:
1056: \subsection{Bounds for Core Point Depth}
1057: Recalling that $\m(\S)$ is the minimum value of a
1058: core point in a configuration $\S$ and that $\mu(d)$ is the
1059: minimum value of $\m(\S)$ over all $d$-dimensional
1060: colourful configurations $\S$, our main result is:
1061:
1062: \begin{theorem}\label{th:bounds}
1063: The minimal colourful simplicial depth of any interior core point in
1064: any colourful configuration is between $2d$ and $d^2+1$.
1065: That is, we have: $2d \le \mu(d) \le d^2+1$.
1066: \end{theorem}
1067:
1068: \begin{conjecture}\label{co:mu}
1069: The minimum colourful simplicial depth of any interior core point in any
1070: colourful configuration is $d^2+1$. That is,
1071: we have $\mu(d) = d^2+1$.
1072: \end{conjecture}
1073: This conjecture implies that the configuration $\S^\minus$
1074: minimizes $\m(\S)$ for $d$-dimensional colourful
1075: configurations. It would also give an elementary proof of (\ref{eq:lbc}).
1076: It is easy to see that this holds for $d=1$. As we noted
1077: in Section~\ref{se:2d}, Conjecture~\ref{co:mu} holds for $d=2$ and $d=3$.
1078: The non-uniqueness of configurations attaining $\m(\S)=d^2+1$
1079: suggests that any such proof cannot be completely trivial
1080: but it may be possible to do this through improved bookkeeping.
1081: The authors generated random low-dimensional configurations
1082: by computer and did not find any counterexamples to
1083: Conjecture~\ref{co:mu}.
1084:
1085: \begin{remark}\label{re:improved}
1086: The lower bound for $\mu(d)$ was improved very recently
1087: independently by
1088: \bara~and Matou{\v s}ek \cite{BM06} and
1089: Stephen and Thomas \cite{ST06}
1090: to
1091: $\displaystyle \max\left(3d, \frac{1}{5}d(d+1)\right)$ for $d>2$
1092: and $\displaystyle \left\lfloor \frac{(d+2)^2}{4} \right\rfloor$
1093: respectively. We know that $\mu(1)=2$, $\mu(2)=5$ and $\mu(3)=10$.
1094: Combining the improved bounds with the parity conditions of
1095: Proposition~\ref{pr:mincfg}
1096: we have the following bounds on $\mu(d)$ for $d > 3$:
1097: $$12 \le \mu(4)\le 17, ~
1098: 16 \le \mu(5) \le 26, ~ 18 \le \mu(6) \le 37, ~ 22 \le \mu(7) \le 50,$$
1099: and for $d > 7$:
1100: $$
1101: \displaystyle \left\lfloor \frac{(d+2)^2}{4} \right\rfloor
1102: \le \mu(d) \le d^2+1.$$
1103: \end{remark}
1104: \vspace{1mm}
1105:
1106: It is also natural to ask what type of colourful configuration
1107: % of $(d+1)$ points in each of $(d+1)$ colours
1108: has a core point of {\it maximum} colourful simplicial depth.
1109: For this question to be interesting, we must fix the number and
1110: size of the colourful sets. Hence we restrict our attention to
1111: $d$-configurations with $(d+1)$ points in each of $(d+1)$ colours.
1112: We also require $p$
1113: to lie in the interior of the core since moving to the boundary
1114: of a simplex increases the depth.
1115: We define:
1116: \begin{equation}\label{eq:nu}
1117: \nu(d)= \max_{d \text{ configurations }\S,
1118: ~ p \in \interior(\core(\S))} \csd_\S(p)
1119: \end{equation}
1120:
1121: Our method is well suited to analyzing $\nu(d)$ simply
1122: by changing our objective to creating deep cells and placing
1123: antipodes in them.
1124: We remark that $\nu(1)=2$.
1125: An analysis similar to that of Section~\ref{se:2d} shows that
1126: $\nu(2)=9$. The key observation is after placing two sets of
1127: three colourful points on the circle, the sequence of cell depths
1128: that we obtain is either 1,2,3,4,3,2
1129: or 3,2,3,2,3,2. In the first case we also need to argue that
1130: the cells of depth at least 3 cover less than half the
1131: circle and that opposite every point of depth 4 is a point of
1132: depth 1.
1133:
1134: The minimal core depth configuration $\S^\minus$
1135: used to prove $\mu(2)=5$ is topologically unique, so it is
1136: interesting to observe that, up to topology, there are two distinct
1137: configurations that contain $\zero$ in 9
1138: colourful simplices. The first corresponds to the sequence of
1139: cell depths 1,2,3,4,3,2 and contains a point $z_3$
1140: that generates a unique \zero-containing colourful simplex.
1141: The second corresponds to the sequence 3,2,3,2,3,2 and is a
1142: combinatorially symmetric configuration where each colourful
1143: point is in exactly three \zero-containing colourful simplices.
1144: The configurations are illustrated in Figure~\ref{fig:max2d}.
1145: \begin{figure}[h!tb]
1146: \begin{center}
1147: \includegraphics[height=5cm]{max2d}
1148: \caption{The two configurations in dimension 2
1149: with $\csd_{\S}(\zero)=9$.}\label{fig:max2d}
1150: \end{center}
1151: \end{figure}
1152:
1153: We can build a configuration $\S^+$ with
1154: $\csd_{\S^+}(\zero) =d^{d+1}+1$ by following the strategy for
1155: $\S^\minus$ but building a deep cell rather than a shallow one.
1156: To do this, we place the polar region points of colour $i$ close
1157: to the geodesic between $p_{\north}$ and the point of colour $i$
1158: on Cancer. Then $p_{\north}$ is contained in every colourful
1159: cone generated by points from Cancer and the polar region
1160: (in fact these are all the colourful cones containing $p_{\north}$).
1161: Hence the cell $C_{\north}$ containing $p_{\north}$ has depth $d^d$.
1162: By placing the points of $S^+_{d+1}$ so that $d$ of their antipodes
1163: are in $C_{\north}$ and the final antipode is at $p_{\south}$,
1164: we get $\S^+$ with $\csd_{\S^+}(\zero) = d \cdot d^d+1$.
1165: The two-dimensional $\S^+$ appears as the left element of
1166: Figure~\ref{fig:max2d}.
1167:
1168: A more symmetric (but similar) construction places one point of
1169: each colour at the vertices of a regular simplex, and the remaining
1170: points surround the antipode of the same colour.
1171:
1172: It follows that $\nu(d) \ge d^{d+1}+1$.
1173: We conjecture that this bound is tight.
1174: As with Conjecture~\ref{co:mu} a computer search did not turn up
1175: any counterexamples.
1176: \begin{conjecture}\label{co:nu}
1177: The maximum colourful simplicial depth of any point in the
1178: interior of the core of any colourful configuration of
1179: $(d+1)$ points in each of $(d+1)$ colours is $d^{d+1}+1$.
1180: That is, we have $\nu(d)=d^{d+1}+1$.
1181: \end{conjecture}
1182:
1183: \begin{remark}\label{re:32}
1184: For any $d$, there exists a colourful configuration $\S$
1185: which contains $\zero$ in at least $32\%$ of its colourful simplices.
1186: \end{remark}
1187: A configuration of $(d+1)$ points in each of $(d+1)$ colours
1188: generates $(d+1)^{d+1}$ colourful
1189: simplices, so Remark~\ref{re:32} follows immediately from
1190: the construction of $\S^+$. The minimum fraction of colourful
1191: simplices containing $\zero$ from an $\S^+$ configuration is $82/256$
1192: attained when $d=3$.
1193:
1194: %
1195: % QUESTIONS
1196: %
1197: \section{Open Questions}
1198: We would like to conclude by mentioning that there are many more
1199: natural questions relating to colourful and monochrome simplicial depth.
1200: The first is:
1201: \begin{question}\label{q:mavg}
1202: What is a typical value of $\m(\S)$ for a random
1203: configuration $\S$ of $(d+1)$ points in each of $(d+1)$ colours?
1204: \end{question}
1205: In Section~\ref{se:covering}, we remarked that such random
1206: configurations could be expected to have a simplicial depth
1207: on the order of $\frac{1}{2^d}(d+1)^{d+1}$ at the origin.
1208: We also gave a colourful configuration
1209: $\S^\triangle$ that has $\m(\S^\triangle)=(d+1)!$.
1210: However, $\S^\triangle$ is not in general position.
1211: Our construction $\S^\minus$ from Section~\ref{se:cons} is
1212: in general position and has a low value of $\m(\S^\minus)$.
1213: It is not clear if this behaviour is typical,
1214: i.e.~if most configurations have some point $p$ near the edge of
1215: the core that drags down $\m(\S)$, or if our configuration is
1216: statistically unlikely. Indeed we can consider the
1217: possibility that {\it all} configurations in general position
1218: have such a point near the edge of the core.
1219: \begin{question}\label{q:mmax}
1220: What is the maximum value of $\m(\S)$ for a colourful configuration $\S$
1221: of $(d+1)$ points in each of $(d+1)$ colours? What if $\S$ is not
1222: assumed to be in general position?
1223: \end{question}
1224:
1225: We observe that in fact our construction of a colourful
1226: configuration $\S^\minus$ with $\m(\S^\minus)=d^2+1$ contains points of
1227: high colourful simplicial depth, but away from \zero.
1228: This leads us to consider the colourful analogues of the
1229: functions $f(S)$ and $g(S)$ of Section~\ref{se:depth}.
1230: For a colourful configuration $\S$, define:
1231: \begin{equation}\label{eq:fg}
1232: \f(\S)= \max_{p \in \R^d} \csd_\S(p) \qquad \text{ and } \qquad
1233: \g(\S)= \max_{p \text{ in general position}} \csd_\S(p)
1234: \end{equation}
1235: We focus on the case where we have $(d+1)$ colours.
1236: It is clear that given the sizes of the colourful sets
1237: $S_1,\ldots,S_{d+1}$ comprising $\S$ that the maximum of
1238: $\f(\S)$ and $\g(\S)$ is $|S_1| \cdot \ldots \cdot |S_{d+1}|$
1239: and is attained by placing the points of each colour at
1240: (or near) the vertices of a simplex.
1241: If we restrict $\S$ to be a configuration of $(d+1)$ points in
1242: each of $(d+1)$ colours and take the maximum over the interior of the core, we
1243: get exactly the question of finding $\nu(d)$
1244: (Conjecture~\ref{co:nu}).
1245: We are also interested in lower bounds for $\f(\S)$ and $\g(\S)$.
1246: \begin{question}\label{q:minmax}
1247: For $d$-dimensional configurations consisting of $n$ points in each of
1248: $(d+1)$ colours, find lower bounds for $\f(\S)$ and $\g(\S)$.
1249: \end{question}
1250:
1251: In a survey paper on the colourful \cara~ theorem,
1252: \bara~and Onn \cite{BO97b}
1253: mention that the results of \cite{ABFK92} can be applied to give
1254: a lower bound for $\g(\S)$ when $n$ is large of the form:
1255: \begin{equation}\label{eq:colourlb}
1256: \g(\S) \ge c_d \binom{n}{d+1}
1257: \end{equation}
1258: Unfortunately, the constant $c_d$ is doubly exponential in $d$
1259: so the bound is only non-trivial if $n \gg e^{4d^2}$.
1260: In particular, it sheds no light on the $n=d+1$ case.
1261:
1262: One can also get a lower bound for $\g(\S)$ directly from the
1263: Colourful Tverberg Theorem \cite{ZV92}, which is used to derive the
1264: results in \cite{ABFK92}:
1265: \begin{equation}\label{eq:tverlb}
1266: \g(\S) \ge \frac{1}{4}\left(\frac{n}{d+1} +3\right)
1267: \end{equation}
1268: This still does not help for $n=d+1$, but
1269: for small $n$ the bound is stronger than (\ref{eq:colourlb})
1270: and comes with the additional guarantee that colourful
1271: simplices involved are disjoint! This suggests that there is
1272: much room for improvement.
1273:
1274: \subsection{Monochrome Questions}
1275: The authors would also like to mention that they do not know
1276: the answers to some fairly basic questions about monochrome
1277: simplicial depth.
1278: Recall the maximum closed and open depth functions $f(S)$ and $g(S)$
1279: for a set of points $S$ in $\R^d$ defined in Section~\ref{se:depth}.
1280: \begin{question}\label{q:limits}
1281: Are the points $p$ attaining the maximum $f(S)$ in (\ref{eq:maxf})
1282: always limit points of the set of maxima attaining $g(S)$
1283: in (\ref{eq:maxg})?
1284: \end{question}
1285: We feel that a positive answer to this question would provide
1286: a further natural justification for studying $g(S)$ in place of $f(S)$
1287: when the former is more tractable. Similarly, it would be interesting
1288: to get conditions on $S$ such that $f(S)$ is not much larger
1289: than $g(S)$.
1290:
1291: We are also curious about the expected values of $f(S)$ and $g(S)$:
1292: \begin{question}\label{q:mono}
1293: Given $n$ points in $\R^d$ distributed independently and symmetrically
1294: about \zero, what is the expected deepest simplicial depth of the
1295: resulting configuration? That is, what is the expected
1296: depth of the simplicial median of the points?
1297: \end{question}
1298: Wagner and Welzl \cite{WW01} give an expression for the
1299: expected depth of \zero, but $\zero$ will not always
1300: be the deepest point.
1301: Indeed if $n=d+1$ the expected simplicial depth of $\zero$ will be
1302: $\frac{1}{2^d}$ while the simplicial median always has depth 1.
1303: For fixed $d$ the expected depth of $\zero$ is
1304: $\frac{1}{2^d}\binom{n}{d+1}$ which has the same asymptotic behaviour as
1305: \bara's sharp upper bound (\ref{eq:ub}) for $g(S)$.
1306: However, when $n$ is not much larger than $(d+1)$, the gap between
1307: the expected depth of $\zero$ and \bara's upper bound is substantial and it is
1308: not clear to us where the expected depth of the simplicial median lies.
1309:
1310: \bara's method of proving
1311: (\ref{eq:lb}) combined with a solution to Question~\ref{q:mavg}
1312: might lend some insight into Question~\ref{q:mono}, but a direct
1313: approach would be better.
1314:
1315: %
1316: % ACKNOWLEDGMENTS
1317: %
1318: \section{Acknowledgments}
1319: We would like to thank {\sc Imre B\'ar\'any} for discussions
1320: which triggered this work and the anonymous referees for
1321: suggestions which improved the presentation of this paper.
1322:
1323:
1324: % Setup for bibtex
1325: % \nocite{*} % bib entries to include that are not actually cited.
1326: \bibliographystyle{hacked}
1327: % End setup for bibtex.
1328:
1329: \bibliography{refs}
1330:
1331: \end{document}
1332:
1333: