arXiv subjects

Robert Krauthgamer

Publications and source records attributed to Robert Krauthgamer.

At least 37 records · Page 2Linked to original sources

arXiv2024

Streaming Algorithms for Geometric Steiner Forest

We consider an important generalization of the Steiner tree problem, the \emph{Steiner forest problem}, in the Euclidean plane: the input is a multiset $X \subseteq \mathbb{R}^2$, partitioned into $k$ color classes $C_1, C_2, \ldots, C_k \subseteq X$. The goal is to find a minimum-cost Euclidean graph $G$ such that every color class $C_i$ is connected in $G$. We study this Steiner forest problem in the streaming setting, where the stream consists of insertions and deletions of points to $X$. Each input point $x\in X$ arrives with its color $\textsf{color}(x) \in [k]$, and as usual for dynamic geometric streams, the input points are restricted to the discrete grid $\{0, \ldots, Δ\}^2$. We design a single-pass streaming algorithm that uses $\mathrm{poly}(k \cdot \logΔ)$ space and time, and estimates the cost of an optimal Steiner forest solution within ratio arbitrarily close to the famous Euclidean Steiner ratio $α_2$ (currently $1.1547 \le α_2 \le 1.214$). This approximation guarantee matches the state-of-the-art bound for streaming Steiner tree, i.e., when $k=1$, and it is a major open question to improve the ratio to $1 + ε$ even for this special case. Our approach relies on a novel combination of streaming techniques, like sampling and linear sketching, with the classical Arora-style dynamic-programming framework for geometric optimization problems, which usually requires large memory and has so far not been applied in the streaming setting. We complement our streaming algorithm for the Steiner forest problem with simple arguments showing that any finite approximation requires $Ω(k)$ bits of space.

cs.DS↗

arXiv2024

Coresets for Kernel Clustering

We devise coresets for kernel $k$-Means with a general kernel, and use them to obtain new, more efficient, algorithms. Kernel $k$-Means has superior clustering capability compared to classical $k$-Means, particularly when clusters are non-linearly separable, but it also introduces significant computational challenges. We address this computational issue by constructing a coreset, which is a reduced dataset that accurately preserves the clustering costs. Our main result is a coreset for kernel $k$-Means that works for a general kernel and has size $\mathrm{poly}(kε^{-1})$. Our new coreset both generalizes and greatly improves all previous results; moreover, it can be constructed in time near-linear in $n$. This result immediately implies new algorithms for kernel $k$-Means, such as a $(1+ε)$-approximation in time near-linear in $n$, and a streaming algorithm using space and update time $\mathrm{poly}(k ε^{-1} \log n)$. We validate our coreset on various datasets with different kernels. Our coreset performs consistently well, achieving small errors while using very few points. We show that our coresets can speed up kernel $k$-Means++ (the kernelized version of the widely used $k$-Means++ algorithm), and we further use this faster kernel $k$-Means++ for spectral clustering. In both applications, we achieve significant speedup and a better asymptotic growth while the error is comparable to baselines that do not use coresets.

cs.DS↗

arXiv2024

Cut Sparsification and Succinct Representation of Submodular Hypergraphs

In cut sparsification, all cuts of a hypergraph $H=(V,E,w)$ are approximated within $1\pmε$ factor by a small hypergraph $H'$. This widely applied method was generalized recently to a setting where the cost of cutting each hyperedge $e$ is provided by a splitting function $g_e: 2^e\to\mathbb{R}_+$. This generalization is called a submodular hypergraph when the functions $\{g_e\}_{e\in E}$ are submodular, and it arises in machine learning, combinatorial optimization, and algorithmic game theory. Previous work studied the setting where $H'$ is a reweighted sub-hypergraph of $H$, and measured the size of $H'$ by the number of hyperedges in it. In this setting, we present two results: (i) all submodular hypergraphs admit sparsifiers of size polynomial in $n=|V|$ and $ε^{-1}$; (ii) we propose a new parameter, called spread, and use it to obtain smaller sparsifiers in some cases. We also show that for a natural family of splitting functions, relaxing the requirement that $H'$ be a reweighted sub-hypergraph of $H$ yields a substantially smaller encoding of the cuts of $H$ (almost a factor $n$ in the number of bits). This is in contrast to graphs, where the most succinct representation is attained by reweighted subgraphs. A new tool in our construction of succinct representation is the notion of deformation, where a splitting function $g_e$ is decomposed into a sum of functions of small description, and we provide upper and lower bounds for deformation of common splitting functions.

cs.DS↗

arXiv2023

Recovery Guarantees for Distributed-OMP

We study distributed schemes for high-dimensional sparse linear regression, based on orthogonal matching pursuit (OMP). Such schemes are particularly suited for settings where a central fusion center is connected to end machines, that have both computation and communication limitations. We prove that under suitable assumptions, distributed-OMP schemes recover the support of the regression vector with communication per machine linear in its sparsity and logarithmic in the dimension. Remarkably, this holds even at low signal-to-noise-ratios, where individual machines are unable to detect the support. Our simulations show that distributed-OMP schemes are competitive with more computationally intensive methods, and in some cases even outperform them.

stat.ML↗

arXiv2023

Labelings vs. Embeddings: On Distributed Representations of Distances

We investigate for which metric spaces the performance of distance labeling and of $\ell_\infty$-embeddings differ, and how significant can this difference be. Recall that a distance labeling is a distributed representation of distances in a metric space $(X,d)$, where each point $x\in X$ is assigned a succinct label, such that the distance between any two points $x,y \in X$ can be approximated given only their labels. A highly structured special case is an embedding into $\ell_\infty$, where each point $x\in X$ is assigned a vector $f(x)$ such that $\|f(x)-f(y)\|_\infty$ is approximately $d(x,y)$. The performance of a distance labeling or an $\ell_\infty$-embedding is measured via its distortion and its label-size/dimension. We also study the analogous question for the prioritized versions of these two measures. Here, a priority order $π=(x_1,\dots,x_n)$ of the point set $X$ is given, and higher-priority points should have shorter labels. Formally, a distance labeling has prioritized label-size $α(\cdot)$ if every $x_j$ has label size at most $α(j)$. Similarly, an embedding $f: X \to \ell_\infty$ has prioritized dimension $α(\cdot)$ if $f(x_j)$ is non-zero only in the first $α(j)$ coordinates. In addition, we compare these prioritized measures to their classical (worst-case) versions. We answer these questions in several scenarios, uncovering a surprisingly diverse range of behaviors. First, in some cases labelings and embeddings have very similar worst-case performance, but in other cases there is a huge disparity. However in the prioritized setting, we most often find a strict separation between the performance of labelings and embeddings. And finally, when comparing the classical and prioritized settings, we find that the worst-case bound for label size often "translates" to a prioritized one, but also find a surprising exception to this rule.

cs.DS↗

arXiv2023

Comparison of Matrix Norm Sparsification

A well-known approach in the design of efficient algorithms, called matrix sparsification, approximates a matrix $A$ with a sparse matrix $A'$. Achlioptas and McSherry [2007] initiated a long line of work on spectral-norm sparsification, which aims to guarantee that $\|A'-A\|\leq ε\|A\|$ for error parameter $ε>0$. Various forms of matrix approximation motivate considering this problem with a guarantee according to the Schatten $p$-norm for general $p$, which includes the spectral norm as the special case $p=\infty$. We investigate the relation between fixed but different $p\neq q$, that is, whether sparsification in the Schatten $p$-norm implies (existentially and/or algorithmically) sparsification in the Schatten $q\text{-norm}$ with similar sparsity. An affirmative answer could be tremendously useful, as it will identify which value of $p$ to focus on. Our main finding is a surprising contrast between this question and the analogous case of $\ell_p$-norm sparsification for vectors: For vectors, the answer is affirmative for $p q$, but for matrices we answer negatively for almost all sufficiently distinct $p\neq q$. In addition, our explicit constructions may be of independent interest.

math.NA↗

arXiv2023

Lower Bounds for Pseudo-Deterministic Counting in a Stream

Many streaming algorithms provide only a high-probability relative approximation. These two relaxations, of allowing approximation and randomization, seem necessary -- for many streaming problems, both relaxations must be employed simultaneously, to avoid an exponentially larger (and often trivial) space complexity. A common drawback of these randomized approximate algorithms is that independent executions on the same input have different outputs, that depend on their random coins. Pseudo-deterministic algorithms combat this issue, and for every input, they output with high probability the same ``canonical'' solution. We consider perhaps the most basic problem in data streams, of counting the number of items in a stream of length at most $n$. Morris's counter [CACM, 1978] is a randomized approximation algorithm for this problem that uses $O(\log\log n)$ bits of space, for every fixed approximation factor (greater than $1$). Goldwasser, Grossman, Mohanty and Woodruff [ITCS 2020] asked whether pseudo-deterministic approximation algorithms can match this space complexity. Our main result answers their question negatively, and shows that such algorithms must use $Ω(\sqrt{\log n / \log\log n})$ bits of space. Our approach is based on a problem that we call Shift Finding, and may be of independent interest. In this problem, one has query access to a shifted version of a known string $F\in\{0,1\}^{3n}$, which is guaranteed to start with $n$ zeros and end with $n$ ones, and the goal is to find the unknown shift using a small number of queries. We provide for this problem an algorithm that uses $O(\sqrt{n})$ queries. It remains open whether $poly(\log n)$ queries suffice; if true, then our techniques immediately imply a nearly-tight $Ω(\log n/\log\log n)$ space bound for pseudo-deterministic approximate counting.

cs.DS↗

arXiv2023

Streaming Euclidean Max-Cut: Dimension vs Data Reduction

Max-Cut is a fundamental problem that has been studied extensively in various settings. We design an algorithm for Euclidean Max-Cut, where the input is a set of points in $\mathbb{R}^d$, in the model of dynamic geometric streams, where the input $X\subseteq [Δ]^d$ is presented as a sequence of point insertions and deletions. Previously, Frahling and Sohler [STOC 2005] designed a $(1+ε)$-approximation algorithm for the low-dimensional regime, i.e., it uses space $\exp(d)$. To tackle this problem in the high-dimensional regime, which is of growing interest, one must improve the dependence on the dimension $d$, ideally to space complexity $\mathrm{poly}(ε^{-1} d \logΔ)$. Lammersen, Sidiropoulos, and Sohler [WADS 2009] proved that Euclidean Max-Cut admits dimension reduction with target dimension $d' = \mathrm{poly}(ε^{-1})$. Combining this with the aforementioned algorithm that uses space $\exp(d')$, they obtain an algorithm whose overall space complexity is indeed polynomial in $d$, but unfortunately exponential in $ε^{-1}$. We devise an alternative approach of \emph{data reduction}, based on importance sampling, and achieve space bound $\mathrm{poly}(ε^{-1} d \logΔ)$, which is exponentially better (in $ε$) than the dimension-reduction approach. To implement this scheme in the streaming model, we employ a randomly-shifted quadtree to construct a tree embedding. While this is a well-known method, a key feature of our algorithm is that the embedding's distortion $O(d\logΔ)$ affects only the space complexity, and the approximation ratio remains $1+ε$.

cs.DS↗

arXiv2023

Streaming Facility Location in High Dimension via Geometric Hashing

In Euclidean Uniform Facility Location (UFL), the input is a set of clients in $\mathbb{R}^d$ and the goal is to place facilities to serve them, so as to minimize the total cost of opening facilities plus connecting the clients. We study the setting of dynamic geometric streams, where the clients are presented as a sequence of insertions and deletions of points in the grid $\{1,\ldots,Δ\}^d$, and we focus on the \emph{high-dimensional regime}, where the algorithm must use space polynomial in $d\cdot\logΔ$. We present a new algorithmic framework, based on importance sampling, for $O(1)$-approximation of UFL using only $\mathrm{poly}(d\cdot\logΔ)$ space. This framework is easy to implement in two passes, one for sampling points and the other for estimating their contribution. Over random-order streams, we can extend this to one pass by using the two halves of the stream separately. Our main result, for arbitrary-order streams, computes $O(d / \log d)$-approximation in one pass by combining the two passes differently. This improves upon previous algorithms that either need space $\exp(d)$ or only guarantee $O(d\cdot\log^2Δ)$-approximation, and therefore our algorithms for high dimension are the first to avoid the $O(\logΔ)$-factor in approximation that is inherent to the widely-used quadtree decomposition. Our improvement is achieved by employing a geometric hashing scheme that maps points in $\mathbb{R}^d$ into buckets of bounded diameter, with the key property that every point set of small-enough diameter is hashed into few buckets. By applying an alternative bound for this hashing, we also obtain an $O(1 / ε)$-approximation in one pass, using larger but still sublinear space $O(n^ε)$ where $n$ is the number of clients. We complement our results by showing $1.085$-approximation requires space exponential in $\mathrm{poly}(d\cdot\logΔ)$.

cs.DS↗

arXiv2022

Coresets for Clustering in Graphs of Bounded Treewidth

We initiate the study of coresets for clustering in graph metrics, i.e., the shortest-path metric of edge-weighted graphs. Such clustering problems are essential to data analysis and used for example in road networks and data visualization. A coreset is a compact summary of the data that approximately preserves the clustering objective for every possible center set, and it offers significant efficiency improvements in terms of running time, storage, and communication, including in streaming and distributed settings. Our main result is a near-linear time construction of a coreset for k-Median in a general graph $G$, with size $O_{ε, k}(\mathrm{tw}(G))$ where $\mathrm{tw}(G)$ is the treewidth of $G$, and we complement the construction with a nearly-tight size lower bound. The construction is based on the framework of Feldman and Langberg [STOC 2011], and our main technical contribution, as required by this framework, is a uniform bound of $O(\mathrm{tw}(G))$ on the shattering dimension under any point weights. We validate our coreset on real-world road networks, and our scalable algorithm constructs tiny coresets with high accuracy, which translates to a massive speedup of existing approximation algorithms such as local search for graph k-Median.

cs.DS↗

arXiv2022

Clustering Permutations: New Techniques with Streaming Applications

We study the classical metric $k$-median clustering problem over a set of input rankings (i.e., permutations), which has myriad applications, from social-choice theory to web search and databases. A folklore algorithm provides a $2$-approximate solution in polynomial time for all $k=O(1)$, and works irrespective of the underlying distance measure, so long it is a metric; however, going below the $2$-factor is a notorious challenge. We consider the Ulam distance, a variant of the well-known edit-distance metric, where strings are restricted to be permutations. For this metric, Chakraborty, Das, and Krauthgamer [SODA, 2021] provided a $(2-δ)$-approximation algorithm for $k=1$, where $δ\approx 2^{-40}$. Our primary contribution is a new algorithmic framework for clustering a set of permutations. Our first result is a $1.999$-approximation algorithm for the metric $k$-median problem under the Ulam metric, that runs in time $(k \log (nd))^{O(k)}n d^3$ for an input consisting of $n$ permutations over $[d]$. In fact, our framework is powerful enough to extend this result to the streaming model (where the $n$ input permutations arrive one by one) using only polylogarithmic (in $n$) space. Additionally, we show that similar results can be obtained even in the presence of outliers, which is presumably a more difficult problem.

cs.DS↗

arXiv2022

An Algorithmic Bridge Between Hamming and Levenshtein Distances

The edit distance between strings classically assigns unit cost to every character insertion, deletion, and substitution, whereas the Hamming distance only allows substitutions. In many real-life scenarios, insertions and deletions (abbreviated indels) appear frequently but significantly less so than substitutions. To model this, we consider substitutions being cheaper than indels, with cost $1/a$ for a parameter $a\ge 1$. This basic variant, denoted $ED_a$, bridges classical edit distance ($a=1$) with Hamming distance ($a\to\infty$), leading to interesting algorithmic challenges: Does the time complexity of computing $ED_a$ interpolate between that of Hamming distance (linear time) and edit distance (quadratic time)? What about approximating $ED_a$? We first present a simple deterministic exact algorithm for $ED_a$ and further prove that it is near-optimal assuming the Orthogonal Vectors Conjecture. Our main result is a randomized algorithm computing a $(1+ε)$-approximation of $ED_a(X,Y)$, given strings $X,Y$ of total length $n$ and a bound $k\ge ED_a(X,Y)$. For simplicity, let us focus on $k\ge 1$ and a constant $ε> 0$; then, our algorithm takes $\tilde{O}(n/a + ak^3)$ time. Unless $a=\tilde{O}(1)$ and for small enough $k$, this running time is sublinear in $n$. We also consider a very natural version that asks to find a $(k_I, k_S)$-alignment -- an alignment with at most $k_I$ indels and $k_S$ substitutions. In this setting, we give an exact algorithm and, more importantly, an $\tilde{O}(nk_I/k_S + k_S\cdot k_I^3)$-time $(1,1+ε)$-bicriteria approximation algorithm. The latter solution is based on the techniques we develop for $ED_a$ for $a=Θ(k_S / k_I)$. These bounds are in stark contrast to unit-cost edit distance, where state-of-the-art algorithms are far from achieving $(1+ε)$-approximation in sublinear time, even for a favorable choice of $k$.

cs.DS↗

arXiv2022

Conditional Lower Bounds for All-Pairs Max-Flow

We provide evidence that computing the maximum flow value between every pair of nodes in a directed graph on $n$ nodes, $m$ edges,and capacities in the range $[1..n]$, which we call the All-Pairs Max-Flow problem, cannot be solved in time that is significantly faster (i.e., by a polynomial factor) than $O(n^3)$ even for sparse graphs. Since a single maximum $st$-flow can be solved in time $\tilde{O}(m\sqrt{n})$ [Lee and Sidford, FOCS 2014], we conclude that the all-pairs version might require time equivalent to $\tildeΩ(n^{3/2})$ computations of maximum $st$-flow,which strongly separates the directed case from the undirected one. Moreover, if maximum $st$-flow can be solved in time $\tilde{O}(m)$,then the runtime of $\tildeΩ(n^2)$ computations is needed. The latter settles a conjecture of Lacki, Nussbaum, Sankowski, and Wulf-Nilsen [FOCS 2012] negatively. Specifically, we show that in sparse graphs $G=(V,E,w)$, if one can compute the maximum $st$-flow from every $s$ in an input set of sources $S\subseteq V$ to every $t$ in an input set of sinks $T\subseteq V$ in time $O((|S| |T| m)^{1-ε})$,for some $|S|$, $|T|$, and a constant $ε>0$,then MAX-CNF-SAT with $n'$ variables and $m'$ clauses can be solved in time ${m'}^{O(1)}2^{(1-δ)n'}$ for a constant $δ(ε)>0$,a problem for which not even $2^{n'}/poly(n')$ algorithms are known. Such runtime for MAX-CNF-SAT would in particular refute the Strong Exponential Time Hypothesis (SETH). Hence, we improve the lower bound of Abboud, Vassilevska-Williams, and Yu [STOC 2015], who showed that for every fixed $ε>0$ and $|S|=|T|=O(\sqrt{n})$, if the above problem can be solved in time $O(n^{3/2-ε})$, then some incomparable conjecture is false. Furthermore, a larger lower bound than ours implies strictly super-linear time for maximum $st$-flow problem, which would be an amazing breakthrough.

cs.DS↗

arXiv2022

Gap Edit Distance via Non-Adaptive Queries: Simple and Optimal

We study the problem of approximating edit distance in sublinear time. This is formalized as the $(k,k^c)$-Gap Edit Distance problem, where the input is a pair of strings $X,Y$ and parameters $k,c>1$, and the goal is to return YES if $ED(X,Y)\leq k$, NO if $ED(X,Y)> k^c$, and an arbitrary answer when $k < ED(X,Y) \le k^c$. Recent years have witnessed significant interest in designing sublinear-time algorithms for Gap Edit Distance. In this work, we resolve the non-adaptive query complexity of Gap Edit Distance for the entire range of parameters, improving over a sequence of previous results. Specifically, we design a non-adaptive algorithm with query complexity $\tilde{O}(n/k^{c-0.5})$, and we further prove that this bound is optimal up to polylogarithmic factors. Our algorithm also achieves optimal time complexity $\tilde{O}(n/k^{c-0.5})$ whenever $c\geq 1.5$. For $1<c<1.5$, the running time of our algorithm is $\tilde{O}(n/k^{2c-1})$. In the restricted case of $k^c=Ω(n)$, this matches a known result [Batu, Ergün, Kilian, Magen, Raskhodnikova, Rubinfeld, and Sami; STOC 2003], and in all other (nontrivial) cases, our running time is strictly better than all previous algorithms, including the adaptive ones. However, an independent work of Bringmann, Cassis, Fischer, and Nakos [STOC 2022] provides an adaptive algorithm that bypasses the non-adaptive lower bound, but only for small enough $k$ and $c$.

cs.DS↗

arXiv2022

The Power of Uniform Sampling for Coresets

Motivated by practical generalizations of the classic $k$-median and $k$-means objectives, such as clustering with size constraints, fair clustering, and Wasserstein barycenter, we introduce a meta-theorem for designing coresets for constrained-clustering problems. The meta-theorem reduces the task of coreset construction to one on a bounded number of ring instances with a much-relaxed additive error. This reduction enables us to construct coresets using uniform sampling, in contrast to the widely-used importance sampling, and consequently we can easily handle constrained objectives. Notably and perhaps surprisingly, this simpler sampling scheme can yield coresets whose size is independent of $n$, the number of input points. Our technique yields smaller coresets, and sometimes the first coresets, for a large number of constrained clustering problems, including capacitated clustering, fair clustering, Euclidean Wasserstein barycenter, clustering in minor-excluded graph, and polygon clustering under Fréchet and Hausdorff distance. Finally, our technique yields also smaller coresets for $1$-median in low-dimensional Euclidean spaces, specifically of size $\tilde{O}(\varepsilon^{-1.5})$ in $\mathbb{R}^2$ and $\tilde{O}(\varepsilon^{-1.6})$ in $\mathbb{R}^3$.

cs.DS↗

arXiv2022

Breaking the Cubic Barrier for All-Pairs Max-Flow: Gomory-Hu Tree in Nearly Quadratic Time

In 1961, Gomory and Hu showed that the All-Pairs Max-Flow problem of computing the max-flow between all $n\choose 2$ pairs of vertices in an undirected graph can be solved using only $n-1$ calls to any (single-pair) max-flow algorithm. Even assuming a linear-time max-flow algorithm, this yields a running time of $O(mn)$, which is $O(n^3)$ when $m = Θ(n^2)$. While subsequent work has improved this bound for various special graph classes, no subcubic-time algorithm has been obtained in the last 60 years for general graphs. We break this longstanding barrier by giving an $\tilde{O}(n^{2})$-time algorithm on general, weighted graphs. Combined with a popular complexity assumption, we establish a counter-intuitive separation: all-pairs max-flows are strictly easier to compute than all-pairs shortest-paths. Our algorithm produces a cut-equivalent tree, known as the Gomory-Hu tree, from which the max-flow value for any pair can be retrieved in near-constant time. For unweighted graphs, we refine our techniques further to produce a Gomory-Hu tree in the time of a poly-logarithmic number of calls to any max-flow algorithm. This shows an equivalence between the all-pairs and single-pair max-flow problems, and is optimal up to poly-logarithmic factors. Using the recently announced $m^{1+o(1)}$-time max-flow algorithm (Chen et al., March 2022), our Gomory-Hu tree algorithm for unweighted graphs also runs in $m^{1+o(1)}$-time.

cs.DS↗

arXiv2022

Exact Flow Sparsification Requires Unbounded Size

Given a large edge-capacitated network $G$ and a subset of $k$ vertices called terminals, an (exact) flow sparsifier is a small network $G'$ that preserves (exactly) all multicommodity flows that can be routed between the terminals. Flow sparsifiers were introduced by Leighton and Moitra [STOC 2010], and have been studied and used in many algorithmic contexts. A fundamental question that remained open for over a decade, asks whether every $k$-terminal network admits an exact flow sparsifier whose size is bounded by some function $f(k)$ (regardless of the size of $G$ or its capacities). We resolve this question in the negative by proving that there exist $6$-terminal networks $G$ whose flow sparsifiers $G'$ must have arbitrarily large size. This unboundedness is perhaps surprising, since the analogous sparsification that preserves all terminal cuts (called exact cut sparsifier or mimicking network) admits sparsifiers of size $f_0(k)\leq 2^{2^k}$ [Hagerup, Katajainen, Nishimura, and Ragde, JCSS 1998]. We prove our results by analyzing the set of all feasible demands in the network, known as the demand polytope. We identify an invariant of this polytope, essentially the slope of certain facets, that can be made arbitrarily large even for $k=6$, and implies an explicit lower bound on the size of the network. We further use this technique to answer, again in the negative, an open question of Seymour [JCTB 2015] regarding flow-sparsification that uses only contractions and preserves the infeasibility of one demand vector.

cs.DS↗

arXiv2022

Optimal Vertex-Cut Sparsification of Quasi-Bipartite Graphs

In vertex-cut sparsification, given a graph $G=(V,E)$ with a terminal set $T\subseteq V$, we wish to construct a graph $G'=(V',E')$ with $T\subseteq V'$, such that for every two sets of terminals $A,B\subseteq T$, the size of a minimum $(A,B)$-vertex-cut in $G'$ is the same as in $G$. In the most basic setting, $G$ is unweighted and undirected, and we wish to bound the size of $G'$ by a function of $k=|T|$. Kratsch and Wahlström [JACM 2020] proved that every graph $G$ (possibly directed), admits a vertex-cut sparsifier $G'$ with $O(k^3)$ vertices, which can in fact be constructed in randomized polynomial time. We study (possibly directed) graphs $G$ that are quasi-bipartite, i.e., every edge has at least one endpoint in $T$, and prove that they admit a vertex-cut sparsifier with $O(k^2)$ edges and vertices, which can in fact be constructed in deterministic polynomial time. In fact, this bound naturally extends to all graphs with a small separator into bounded-size sets. Finally, we prove information-theoretically a nearly-matching lower bound, i.e., that $\tildeΩ(k^2)$ edges are required to sparsify quasi-bipartite undirected graphs.

cs.DS↗