Classifying unavoidable Tverberg partitions

For a positive integer $n$, denote $[n]\stackrel{{\scriptstyle\text{\tiny{def}}}}{{=}}\{1,2,\ldots,n\}$. Define $T(d,r)\stackrel{{\scriptstyle\text{\tiny{def}}}}{{=}}(r-1)(d+1)+1$. Tverberg’s theorem [Tve66] (see also [Mat02]) asserts that for every point sequence $P\stackrel{{\scriptstyle\text{\tiny{def}}}}{{=}}(p_{1},\ldots,p_{T(d,r)})$ in $\mathbb{R}^{d}$ there exists a partition $\mathcal{I}\stackrel{{\scriptstyle\text{\tiny{def}}}}{{=}}\{I_{1},\ldots,I_{r}\}$ of $[T(d,r)]$ into $r$ parts, such that the $r$ convex hulls $\conv\{p_{i}\mid i\in I_{j}\}$, $1\leq j\leq r$, intersect at a common point. Such a partition $\mathcal{I}$ is called a Tverberg partition for $P$. If the convex hulls intersect at a single point (which happens whenever $P$ is generic), then that point is called the Tverberg point of the partition. The special case $r=2$ of Tverberg’s theorem is Radon’s lemma [Rad21]; in that case we refer to the Tverberg point as the Radon point.

We call a partition of $[T(d,r)]$ into $r$ parts a Tverberg type.

Let $\tv_{\mathcal{I}}(P)$ be the $T(d,r)$-ary predicate stating that the Tverberg type $\mathcal{I}$ is a Tverberg partition for $P$.

Our main objective in this paper is to classify the Tverberg-type predicates according to the three possibilities mentioned above.

It is sometimes convenient to encode a Tverberg type as a string $\sigma\in[r]^{T(d,r)}$, indicating to which part each integer belongs. Since the order of the parts within the partition does not matter, there are $r!$ different ways of encoding each Tverberg type. For example, for $d=2$, $r=3$, the Tverberg type $\bigl\{\{1,3,6\},\{2,7\},\{4,5\}\bigr\}$ can be encoded as $1213312$.

In this representation, the colorful Tverberg types are those $\sigma$ for which $\{\sigma(i+1),\ldots,\sigma(i+r)\}=[r]$ for each $i=0,r-1,2(r-1),\ldots,d(r-1)$. An example of a colorful Tverberg type for $d=3$, $r=5$ is $\underline{1234\overline{5}}\overline{241\underline{3}}\underline{514\overline{2}}\overline{5134}$. The lines above and below the digits indicate the intervals in which all “colors” must show up.

It is easily seen that the number of colorful Tverberg types with parameters $d,r$ is $(r-1)!^{d}$. We will expound on the significance of this number below.

We call the colorful Tverberg type encoded by $12\cdots r\cdots 212\cdots r\cdots$ the zigzag type. For example, the zigzag type for $r=3$ and $d=4$ is $12321232123$.

The results of Theorem 1.4 regarding zigzag types for $d\leq 4$ were previously announced in [BM14].

Our proof strategy for proving that a given predicate $\Phi$ is unavoidable is as follows: Suppose for a contradiction that for every $n$ there exists an $n$-point sequence $P$ that avoids $\Phi$. Let $\Psi_{1},\ldots,\Psi_{k}$ be other predicates. By Ramsey’s theorem, by making $n$ large enough, we can guarantee the existence of arbitrarily large subsequences $P^{\prime}$ of $P$ that are homogeneous with respect to $\Psi_{1},\ldots,\Psi_{k}$.

Hence, in our search for a contradiction, we can assume without loss of generality that our sequence $P$ not only avoids $\Phi$, but is also homogeneous with respect to a fixed finite family $\overline{\Psi}\stackrel{{\scriptstyle\text{\tiny{def}}}}{{=}}(\Psi_{1},\ldots,\Psi_{k})$ of other predicates. Furthermore, if we previously showed that some of the $\Psi_{i}$’s are unavoidable, then we can assume that $P$ specifically avoids their negations $\neg\Psi_{i}$.

In our proofs, we will start with an orientation-homogeneous sequence (which corresponds to taking $\Psi_{1}=\orient$), and we will add additional predicates to $\overline{\Psi}$ “on-the-fly”; this is OK as long as we do it only a finite number of times.

We say that predicates $\Phi_{1}$ and $\Phi_{2}$ are equivalent if there is an unavoidable predicate $\Phi$ such that $\Phi\implies(\Phi_{1}\iff\Phi_{2})$. In our proofs, $\Phi$ will usually be the orientation predicate. We write $\Phi_{1}\equiv\Phi_{2}$ if $\Phi_{1}$ and $\Phi_{2}$ are equivalent.

Let $P$ be a sequence of $T(d,r)$ points in $\mathbb{R}^{d}$. Tverberg’s theorem asserts the existence of at least one Tverberg partition for $P$. However, usually there is more than one Tverberg partition. If $r=2$ (the case of Radon’s lemma), and the points of $P$ are in general position, then the Radon partition is indeed unique. However, it seems that for $r\geq 3$ the Tverberg partition is never unique.

Sierksma conjectured in 1979 ([Sie79]; cited by Reay in [Rea82, Problem 14]) that the number of Tverberg partitions is always at least $(r-1)!^{d}$. Sierksma pointed out that there exist $T(d,r)$-point sets that have exactly $(r-1)!^{d}$ Tverberg partitions: Choose $d+1$ affinely independent points $p_{1},\ldots,p_{d+1}$, and let $q$ be a point in the interior of the simplex $\conv\{p_{1},\ldots,p_{d+1}\}$. Replace each $p_{i}$ by a tiny cloud $\mathcal{P}_{i}$ of $r-1$ points. Then, the Tverberg partitions of this point set are exactly those that have $r-1$ parts containing exactly one point from each cloud, plus an $r$-th part containing only $q$. Hence, the number of Tverberg partitions here equals exactly $(r-1)!^{d}$. (However, for $d\geq 2$ none of these partitions are colorful, no matter how the points are linearly ordered, since for $d\geq 2$, colorful partitions never contain parts of size $1$.)

White [Whi15] recently found a more general family of $T(d,r)$-point sets that have exactly $(r-1)!^{d}$ Tverberg partitions. In fact, he constructs, for every partition $T(d,r)=n_{1}+\cdots+n_{r}$ of $T(d,r)$ into $r$ integers satisfying $1\leq n_{i}\leq d+1$, a $T(d,r)$-point set $P$ that has exactly $(r-1)!^{d}$ Tverberg partitions, all of which have parts of sizes $n_{1},\ldots,n_{r}$ and have the origin as their Tverberg point. Furthermore, in his construction, each point $p_{i}$ is only specified by a vector of signs $v_{i}\in\{+,0,-\}^{d}$, which indicates the sign of each coordinate of $p_{i}$; the magnitudes of the coordinates can be chosen arbitrarily.

Regarding lower bounds, Hell [Hel08], extending earlier work of Vučić and Zivaljević [VŽ93], showed that the number of Tverberg partitions is always at least $(r-d)!$, and that if $r=p^{k}$ is a prime power, then the number is at least

For large $d$ and $r$, this number is roughly the square root of Sierksma’s conjectured bound.

Very recently, and independently, Attila Pór announced [Pór16] that he has found a full proof of Conjecture 1.3. Pór uses a different approach from the one we use in this paper.

In Section 2 we briefly show that the predicates we consider in this paper are semialgebraic. In Section 3 we prove Theorem 1.4. We first show how solving Tverberg-type predicates reduces to solving hyperplane-side predicates. Then we solve many hyperplane-side predicates, first in the plane, then in $d=3$, and then in $d=4$. In Section 4 we prove Theorem 1.2 and our result regarding Sierksma’s conjecture. A key technical lemma that is obvious but whose proof is quite tedious is proven in Appendix A. Finally, Appendix B contains Mathematica code for some computer enumerations that we performed.

To avoid subscripts we shall use integers to denote points. So, for example, $7$ will denote the $7$th point of sequence $P$. Points after the $9$th are denoted by the letters $A,B,C,\ldots$. Similarly, we shall write $p_{1}<p_{2}$ if point $p_{1}$ precedes point $p_{2}$ in $P$.

The convex hull operation will be indicated by the concatenation of the corresponding integers. So, for example $27$ is the line segment going from the $2$nd point to the $7$th, and $245$ is the triangle with vertices $2$, $4$ and $5$.

We shall use two notations for intersection. First, we use the conventional $179\cap 028$ to denote the intersection of triangles $179$ and $028$. Secondly, we denote the same by placing the two objects to be intersected one above the other, like so: ${}^{179}_{028}$.

Next, we introduce notation for hyperplane-separation statements. If $p_{1},\ldots,p_{d}$, $q_{1}$, $q_{2}$, $\ldots$ , $r_{1}$, $r_{2}$, $\ldots$ are points in $\mathbb{R}^{d}$, then we write

to mean that the hyperplane spanned by $p_{1},\ldots,p_{d}$ separates $q_{1},q_{2},\ldots$, from $r_{1},r_{2},\ldots$. In other words, $\orient(p_{1},\ldots,p_{d},q_{i})$ has the same value for all $i$, which is the opposite of $\orient(p_{1},\ldots,p_{d},r_{i})$ for all $i$. For instance, $148(2:7_{258}^{\,37})$ means that $7$ and ${}_{258}^{\,37}$ are on one side of the hyperplane $148$, whereas $2$ is on the other. (Of course, in order for $1,4,8$ to span a hyperplane and for ${}_{258}^{\,37}$ to be a Radon point, we must be in $\mathbb{R}^{3}$.)

Given a hyperplane-separation statement $s\stackrel{{\scriptstyle\text{\tiny{def}}}}{{=}}H(p:q)$ involving $k$ distinct points, we denote by $\Pi[s]$ the corresponding $k$-ary geometric predicate. For example, in $\mathbb{R}^{2}$, consider the two statements $s_{1}\stackrel{{\scriptstyle\text{\tiny{def}}}}{{=}}25(1:{}^{14}_{36})$ and $s_{2}\stackrel{{\scriptstyle\text{\tiny{def}}}}{{=}}37(1:{}^{15}_{49})$. Then $\Pi[s_{1}]\equiv\Pi[s_{2}]$; both denote the $6$-ary semialgebraic predicate that asserts, given a length-$6$ orientation-homogeneous sequence $(p_{1},\ldots,p_{6})$, that the line through $p_{2}$ and $p_{5}$ separates $p_{1}$ from $p_{1}p_{4}\cap p_{3}p_{6}$; the latter intersection point exists if $p_{1},\ldots p_{6}$ are orientation-homogeneous by the Radon-partition lemma (Lemma 3.1). (If the input points are not orientation-homogeneous, then we do not care what the predicate asserts.)

Thus, for example, in $\mathbb{R}^{4}$ we have $1368(27:459)$.

For example, in $\mathbb{R}^{2}$ we have $14(5:{}^{13}_{25})$, since the segment $13$ is entirely on one side of the line through $14$, specifically, on the side opposite to $5$.

We now proceed to show that, for all planar non-colorful Tverberg types, their negation is unavoidable. For this, we first prove a lemma that will be very useful in higher dimensions as well:

We now present a technique for lifting results to higher dimensions. Let $H$ be a hyperplane in $\mathbb{R}^{d}$. To define an orientation predicate within $H$, fix a rigid motion $\rho$ in $R^{d}$ that takes $H$ to the hyperplane $H_{0}$ given by $x_{d}=0$. Then define $\orient_{H}$ by $\orient_{H}(p_{1},\ldots,p_{d})\stackrel{{\scriptstyle\text{\tiny{def}}}}{{=}}\orient(\rho(p_{1}),\ldots,\rho(p_{d}))$ (where in the last predicate we ignore the last coordinates of the points, which equal $0$).

Fix a point $q\in\mathbb{R}^{d}\setminus H$. Each choice of $\rho$ produces one of two possible predicates $\orient_{H}$, which are negations of one another, depending on whether $\rho$ sends $q$ above or below $H_{0}$. Let us fix a $\rho$ that sends $q$ above $H_{0}$. Call the resulting predicate $\orient_{H,q}$. It is not hard to see that

Given a point set $X\subset\mathbb{R}^{d}$ and a point $p\notin\conv X$, fix a hyperplane $H$ that separates $p$ from $X$. Then we define the central projection of $X$ from $p$ into $H$ by taking each point $q\in X$ to the intersection point $q^{\prime}\stackrel{{\scriptstyle\text{\tiny{def}}}}{{=}}pq\cap H$; and we define the orientation within $H$ by $\orient_{H,p}$.

Note that projection from an intermediate point $p_{i}$ does not produce an orientation-homogeneous sequence, since then the orientation of a $d$-tuple of projected points will depend on the parity of the number of points that appear after $p_{i}$.

Hence, we can lift up lower-dimensional results by adding a new point at the beginning or at the end. For example:

Note that there are many plane-side predicates in $\mathbb{R}^{3}$ involving a Radon point, which are not covered by the above result, nor are they trivially solved by the same-side rule; for example, $\Pi[345({}:{}^{168}_{\,27})]$ and $\Pi[368({}:{}^{147}_{\,25})]$. Below in Section 3.6 we will solve the latter one (and its mirror image).

We now give a useful way of reinterpreting Lemma 3.12 and Corollary 3.13.

Let $P\in(\mathbb{R}^{3})^{n}$ be orientation-homogeneous, and let $a_{1},\allowbreak a_{2},a_{3},\allowbreak a^{\prime}_{1},a^{\prime}_{2},\allowbreak a^{\prime}_{3},b_{1},\allowbreak b_{2}\in P$, such that $b_{1},b_{2}$ interlace both $a_{1},a_{2},a_{3}$ and $a_{1}^{\prime},a_{2}^{\prime},a_{3}^{\prime}$, meaning, $a_{1}<b_{1}<a_{2}<b_{2}<a_{3}$ and $a_{1}^{\prime}<b_{1}<a_{2}^{\prime}<b_{2}<a_{3}^{\prime}$. Define the Radon points

Hence, both $r$ and $r^{\prime}$ lie along the segment $b_{1}b_{2}$. The question is in which order.

If $a^{\prime}_{1}\leq a_{1}$ and $a^{\prime}_{2}\geq a_{2}$ and $a^{\prime}_{3}\leq a_{3}$, then, by the same-side rule (Observation 3.3), it is immediate that the order is $b_{1},r,r^{\prime},b_{2}$.

However, what happens if there are “conflicting” movements, e.g. if $a^{\prime}_{1}>a_{1}$ and $a^{\prime}_{2}>a_{2}$? In fact, in such cases both outcomes are possible. However, one of the outcomes is unavoidable. Specifically:

Unlike in the planar case, for $d=3$ we have incomplete results.

Every colorful Tverberg type with $d=3$ is either of the form

where $\pi_{1},\pi_{4}$ are permutations of $[r]\setminus\{1\}$ and $\pi_{2},\pi_{3}$ are permutations of $[r]\setminus\{1,2\}$; or of the form

where $\pi_{1}$ is a permutation of $[r]\setminus\{1\}$, $\pi_{2}$ is a permutation of $[r]\setminus\{1,2\}$, $\pi_{3}$ is a permutation of $[r]\setminus\{2,3\}$, and $\pi_{4}$ is a permutation of $[r]\setminus\{3\}$.

In the case (4), part $1$ has size $2$, part $2$ has size $3$, and all the other parts have size $4$. In the case (5), parts $1$, $2$, and $3$ have size $3$, and all the other parts have size $4$.