arXiv subjects
Publications and source records attributed to Tillmann Miltzow.
We investigate the computational complexity of computing the Hausdorff distance. Specifically, we show that the decision problem of whether the Hausdorff distance of two semi-algebraic sets is bounded by a given threshold is complete for the complexity class $\forall\exists_<\mathbb{R}$. This implies that the problem is NP-, co-NP-, $\exists\mathbb{R}$- and $\forall\mathbb{R}$-hard.
Given a neural network, training data, and a threshold, it was known that it is NP-hard to find weights for the neural network such that the total error is below the threshold. We determine the algorithmic complexity of this fundamental problem precisely, by showing that it is $\exists\mathbb R$-complete. This means that the problem is equivalent, up to polynomial-time reductions, to deciding whether a system of polynomial equations and inequalities with integer coefficients and real unknowns has a solution. If, as widely expected, $\exists\mathbb R$ is strictly larger than NP, our work implies that the problem of training neural networks is not even in NP. Neural networks are usually trained using some variation of backpropagation. The result of this paper offers an explanation why techniques commonly used to solve big instances of NP-complete problems seem not to be of use for this task. Examples of such techniques are SAT solvers, IP solvers, local search, dynamic programming, to name a few general ones.
We study algorithmic problems that belong to the complexity class of the existential theory of the reals (ER). A problem is ER-complete if it is as hard as the problem ETR and if it can be written as an ETR formula. Traditionally, these problems are studied in the real RAM, a model of computation that assumes that the storage and comparison of real-valued numbers can be done in constant space and time, with infinite precision. The complexity class ER is often called a real RAM analogue of NP, since the problem ETR can be viewed as the real-valued variant of SAT. In this paper we prove a real RAM analogue to the Cook-Levin theorem which shows that ER membership is equivalent to having a verification algorithm that runs in polynomial-time on a real RAM. This gives an easy proof of ER-membership, as verification algorithms on a real RAM are much more versatile than ETR-formulas. We use this result to construct a framework to study ER-complete problems under smoothed analysis. We show that for a wide class of ER-complete problems, its witness can be represented with logarithmic input-precision by using smoothed analysis on its real RAM verification algorithm. This shows in a formal way that the boundary between NP and ER (formed by inputs whose solution witness needs high input-precision) consists of contrived input. We apply our framework to well-studied ER-complete recognition problems which have the exponential bit phenomenon such as the recognition of realizable order types or the Steinitz problem in fixed dimension.
Exhibiting a deep connection between purely geometric problems and real algebra, the complexity class $\exists \mathbb{R}$ plays a crucial role in the study of geometric problems. Sometimes $\exists \mathbb{R}$ is referred to as the 'real analog' of NP. While NP is a class of computational problems that deals with existentially quantified boolean variables, $\exists \mathbb{R}$ deals with existentially quantified real variables. In analogy to $Π_2^p$ and $Σ_2^p$ in the famous polynomial hierarchy, we study the complexity classes $\forall \exists \mathbb{R}$ and $\exists \forall \mathbb{R}$ with real variables. Our main interest is the area-universality problem, where we are given a plane graph $G$, and ask if for each assignment of areas to the inner faces of $G$, there exists a straight-line drawing of $G$ realizing the assigned areas. We conjecture that area-universality is $\forall \exists \mathbb{R}$-complete and support this conjecture by proving $\exists \mathbb{R}$- and $\forall \exists \mathbb{R}$-completeness of two variants of area-universality. To this end, we introduce tools to prove $\forall \exists \mathbb{R}$-hardness and membership. Finally, we present geometric problems as candidates for $\forall \exists \mathbb{R}$-complete problems. These problems have connections to the concepts of imprecision, robustness, and extendability.
We show that the decision problem of determining whether a given (abstract simplicial) $k$-complex has a geometric embedding in $\mathbb R^d$ is complete for the Existential Theory of the Reals for all $d\geq 3$ and $k\in\{d-1,d\}$. This implies that the problem is polynomial time equivalent to determining whether a polynomial equation system has a real solution. Moreover, this implies NP-hardness and constitutes the first hardness results for the algorithmic problem of geometric embedding (abstract simplicial) complexes.
A family S of convex sets in the plane defines a hypergraph H = (S, E) as follows. Every subfamily S' of S defines a hyperedge of H if and only if there exists a halfspace h that fully contains S' , and no other set of S is fully contained in h. In this case, we say that h realizes S'. We say a set S is shattered, if all its subsets are realized. The VC-dimension of a hypergraph H is the size of the largest shattered set. We show that the VC-dimension for pairwise disjoint convex sets in the plane is bounded by 3, and this is tight. In contrast, we show the VC-dimension of convex sets in the plane (not necessarily disjoint) is unbounded. We provide a quadratic lower bound in the number of pairs of intersecting sets in a shattered family of convex sets in the plane. We also show that the VC-dimension is unbounded for pairwise disjoint convex sets in R^d , for d > 2. We focus on, possibly intersecting, segments in the plane and determine that the VC-dimension is always at most 5. And this is tight, as we construct a set of five segments that can be shattered. We give two exemplary applications. One for a geometric set cover problem and one for a range-query data structure problem, to motivate our findings.
We solve an open problem posed by Michael Biro at CCCG 2013 that was inspired by his and others' work on beacon-based routing. Consider a human and a puppy on a simple closed curve in the plane. The human can walk along the curve at bounded speed and change direction as desired. The puppy runs with unbounded speed along the curve as long as the Euclidean straight-line distance to the human is decreasing, so that it is always at a point on the curve where the distance is locally minimal. Assuming that the curve is smooth (with some mild genericity constraints) or a simple polygon, we prove that the human can always catch the puppy in finite time.
Given two shapes $A$ and $B$ in the plane with Hausdorff distance $1$, is there a shape $S$ with Hausdorff distance $1/2$ to and from $A$ and $B$? The answer is always yes, and depending on convexity of $A$ and/or $B$, $S$ may be convex, connected, or disconnected. We show that our result can be generalised to give an interpolated shape between $A$ and $B$ for any interpolation variable $α$ between $0$ and $1$, and prove that the resulting morph has a bounded rate of change with respect to $α$. Finally, we explore a generalization of the concept of a Hausdorff middle to more than two input sets. We show how to approximate or compute this middle shape, and that the properties relating to the connectedness of the Hausdorff middle extend from the case with two input sets. We also give bounds on the Hausdorff distance between the middle set and the input.
Many problems in Discrete and Computational Geometry deal with simple polygons or polygonal regions. Many algorithms and data-structures perform considerably faster, if the underlying polygonal region has low local complexity. One obstacle to make this intuition rigorous, is the lack of a formal definition of local complexity. Here, we give two possible definitions and show how they are related in a combinatorial sense. We say that a polygon $P$ has point visibility width $w=pvw$, if there is no point $q\in P$ that sees more than $w$ reflex vertices. We say that a polygon $P$ has chord visibility width $w=cvw $, if there is no chord $c=\textrm{seg}(a,b)\subset P$ that sees more than w reflex vertices. We show that \[ cvw \leq pvw ^{O( pvw )},\] for any simple polygon. Furthermore, we show that there exists a simple polygon with \[ cvw \geq 2^{Ω( pvw )}.\]
Given a simple polygon $\mathcal{P}$ on $n$ vertices, two points $x,y$ in $\mathcal{P}$ are said to be visible to each other if the line segment between $x$ and $y$ is contained in $\mathcal{P}$. The Point Guard Art Gallery problem asks for a minimum set $S$ such that every point in $\mathcal{P}$ is visible from a point in $S$. The Vertex Guard Art Gallery problem asks for such a set $S$ subset of the vertices of $\mathcal{P}$. A point in the set $S$ is referred to as a guard. For both variants, we rule out any $f(k)n^{o(k / \log k)}$ algorithm, where $k := |S|$ is the number of guards, for any computable function $f$, unless the Exponential Time Hypothesis fails. These lower bounds almost match the $n^{O(k)}$ algorithms that exist for both problems.
We study the complexity of Maximum Clique in intersection graphs of convex objects in the plane. On the algorithmic side, we extend the polynomial-time algorithm for unit disks [Clark '90, Raghavan and Spinrad '03] to translates of any fixed convex set. We also generalize the efficient polynomial-time approximation scheme (EPTAS) and subexponential algorithm for disks [Bonnet et al. '18, Bonamy et al. '18] to homothets of a fixed centrally symmetric convex set. The main open question on that topic is the complexity of Maximum Clique in disk graphs. It is not known whether this problem is NP-hard. We observe that, so far, all the hardness proofs for Maximum Clique in intersection graph classes $\mathcal I$ follow the same road. They show that, for every graph $G$ of a large-enough class $\mathcal C$, the complement of an even subdivision of $G$ belongs to the intersection class $\mathcal I$. Then they conclude invoking the hardness of Maximum Independent Set on the class $\mathcal C$, and the fact that the even subdivision preserves that hardness. However there is a strong evidence that this approach cannot work for disk graphs [Bonnet et al. '18]. We suggest a new approach, based on a problem that we dub Max Interval Permutation Avoidance, which we prove unlikely to have a subexponential-time approximation scheme. We transfer that hardness to Maximum Clique in intersection graphs of objects which can be either half-planes (or unit disks) or axis-parallel rectangles. That problem is not amenable to the previous approach. We hope that a scaled down (merely NP-hard) variant of Max Interval Permutation Avoidance could help making progress on the disk case, for instance by showing the NP-hardness for (convex) pseudo-disks.
Recently, various natural algorithmic problems have been shown to be $\exists \mathbb{R}$-complete. The reduction relied in many cases on the $\exists \mathbb{R}$-completeness of the problem ETR-INV, which served as a useful intermediate problem. Often some strengthening and modification of ETR-INV was required. This lead to a cluttered situation where no paper included all the previous details. Here, we give a streamlined exposition in a self-contained manner. We also explain and prove various universality results regarding ETR-INV. These notes should not be seen as a research paper with new results. However, we do describe some refinements of earlier results which might be useful for future research. We plan to extend and update this exposition as seems fit.
In a nutshell, we show that polynomials and nested polytopes are topological, algebraic and algorithmically equivalent. Given two polytops $A\subseteq B$ and a number $k$, the Nested Polytope Problem (NPP) asks, if there exists a polytope $X$ on $k$ vertices such that $A\subseteq X \subseteq B$. The polytope $A$ is given by a set of vertices and the polytope $B$ is given by the defining hyperplanes. We show a universality theorem for NPP. Given an instance $I$ of the NPP, we define the solutions set of $I$ as $$ V'(I) = \{(x_1,\ldots,x_k)\in \mathbb{R}^{k\cdot n} : A\subseteq \text{conv}(x_1,\ldots,x_k) \subseteq B\}.$$ As there are many symmetries, induced by permutations of the vertices, we will consider the \emph{normalized} solution space $V(I)$. Let $F$ be a finite set of polynomials, with bounded solution space. Then there is an instance $I$ of the NPP, which has a rationally-equivalent normalized solution space $V(I)$. Two sets $V$ and $W$ are rationally equivalent if there exists a homeomorphism $f : V \rightarrow W$ such that both $f$ and $f^{-1}$ are given by rational functions. A function $f:V\rightarrow W$ is a homeomorphism, if it is continuous, invertible and its inverse is continuous as well. As a corollary, we show that NPP is $\exists \mathbb{R}$-complete. This implies that unless $\exists \mathbb{R} =$ NP, the NPP is not contained in the complexity class NP. Note that those results already follow from a recent paper by Shitov. Our proof is geometric and arguably easier.
Consider an ordered point set $P = (p_1,\ldots,p_n)$, its order type (denoted by $χ_P$) is a map which assigns to every triple of points a value in $\{+,-,0\}$ based on whether the points are collinear(0), oriented clockwise(-) or counter-clockwise(+). An abstract order type is a map $χ: \left[\substack{n\\3}\right] \rightarrow \{+,-,0\}$ (where $\left[\substack{n\\3}\right]$ is the collection of all triples of a set of $n$ elements) that satisfies the following condition: for every set of five elements $S\subset [n]$ its induced order type $χ_{|S}$ is realizable by a point set. To be precise, a point set $P$ realizes an order type $χ$,if $χ_P(p_i,p_j,p_k) = χ(i,j,k)$, for all $i<j<k$. Planar point sets are among the most basic and natural geometric objects of study in Discrete and Computational Geometry. Properties of point sets are relevant in theory and practice alike. It is known, that deciding if an abstract order type is realizable is complete for the existential theory of the reals. Our results show that order type realizability is much easier for realistic instances than in the worst case. In particular, we can recognize instances in "expected \NP-time". This is one of the first $\exists\mathbb{R}$-complete problems analyzed under the lens of Smoothed Analysis.
In the Art Gallery Problem we are given a polygon $P\subset [0,L]^2$ on $n$ vertices and a number $k$. We want to find a guard set $G$ of size $k$, such that each point in $P$ is seen by a guard in $G$. Formally, a guard $g$ sees a point $p \in P$ if the line segment $pg$ is fully contained inside the polygon $P$. The history and practical findings indicate that irrational coordinates are a "very rare" phenomenon. We give a theoretical explanation. Next to worst case analysis, Smoothed Analysis gained popularity to explain the practical performance of algorithms, even if they perform badly in the worst case. The idea is to study the expected performance on small perturbations of the worst input. The performance is measured in terms of the magnitude $δ$ of the perturbation and the input size. We consider four different models of perturbation. We show that the expected number of bits to describe optimal guard positions per guard is logarithmic in the input and the magnitude of the perturbation. This shows from a theoretical perspective that rational guards with small bit-complexity are typical. Note that describing the guard position is the bottleneck to show NP-membership. The significance of our results is that algebraic methods are not needed to solve the Art Gallery Problem in typical instances. This is the first time an $\exists\mathbb{R}$-complete problem was analyzed by Smoothed Analysis.
We prove that the following problem is complete for the existential theory of the reals: Given a planar graph and a polygonal region, with some vertices of the graph assigned to points on the boundary of the region, place the remaining vertices to create a planar straight-line drawing of the graph inside the region. This strengthens an NP-hardness result by Patrignani on extending partial planar graph drawings. Our result is one of the first showing that a problem of drawing planar graphs with straight-line edges is hard for the existential theory of the reals. The complexity of the problem is open in the case of a simply connected region. We also show that, even for integer input coordinates, it is possible that drawing a graph in a polygonal region requires some vertices to be placed at irrational coordinates. By contrast, the coordinates are known to be bounded in the special case of a convex region, or for drawing a path in any polygonal region.
We prove that the art gallery problem is equivalent under polynomial time reductions to deciding whether a system of polynomial equations over the real numbers has a solution. The art gallery problem is a classical problem in computational geometry. Given a simple polygon $P$ and an integer $k$, the goal is to decide if there exists a set $G$ of $k$ guards within $P$ such that every point $p\in P$ is seen by at least one guard $g\in G$. Each guard corresponds to a point in the polygon $P$, and we say that a guard $g$ sees a point $p$ if the line segment $pg$ is contained in $P$. The art gallery problem has stimulated extensive research in geometry and in algorithms. However, the complexity status of the art gallery problem has not been resolved. It has long been known that the problem is $\text{NP}$-hard, but no one has been able to show that it lies in $\text{NP}$. Recently, the computational geometry community became more aware of the complexity class $\exists \mathbb{R}$. The class $\exists \mathbb{R}$ consists of problems that can be reduced in polynomial time to the problem of deciding whether a system of polynomial equations with integer coefficients and any number of real variables has a solution. It can be easily seen that $\text{NP}\subseteq \exists \mathbb{R}$. We prove that the art gallery problem is $\exists \mathbb{R}$-complete, implying that (1) any system of polynomial equations over the real numbers can be encoded as an instance of the art gallery problem, and (2) the art gallery problem is not in the complexity class $\text{NP}$ unless $\text{NP}=\exists \mathbb{R}$. As a corollary of our construction, we prove that for any real algebraic number $α$ there is an instance of the art gallery problem where one of the coordinates of the guards equals $α$ in any guard set of minimum cardinality. That rules out many geometric approaches to the problem.
In the Token Swapping problem we are given a graph with a token placed on each vertex. Each token has exactly one destination vertex, and we try to move all the tokens to their destinations, using the minimum number of swaps, i.e., operations of exchanging the tokens on two adjacent vertices. As the main result of this paper, we show that Token Swapping is $W[1]$-hard parameterized by the length $k$ of a shortest sequence of swaps. In fact, we prove that, for any computable function $f$, it cannot be solved in time $f(k)n^{o(k / \log k)}$ where $n$ is the number of vertices of the input graph, unless the ETH fails. This lower bound almost matches the trivial $n^{O(k)}$-time algorithm. We also consider two generalizations of the Token Swapping, namely Colored Token Swapping (where the tokens have different colors and tokens of the same color are indistinguishable), and Subset Token Swapping (where each token has a set of possible destinations). To complement the hardness result, we prove that even the most general variant, Subset Token Swapping, is FPT in nowhere-dense graph classes. Finally, we consider the complexities of all three problems in very restricted classes of graphs: graphs of bounded treewidth and diameter, stars, cliques, and paths, trying to identify the borderlines between polynomial and NP-hard cases.