--- title: Discrete Euclidean k-Center Lower Bound url: https://www.emergentmind.com/topics/center-lower-bound type: topic --- # Discrete Euclidean k-Center Lower Bound Searching arXiv for the primary paper and closely related center lower-bound work. “Center Lower Bound” most commonly refers, in geometric algorithms and fine-grained complexity, to lower bounds for the computational complexity of center-selection problems, especially the **discrete \(k\)-Center** problem in fixed-dimensional Euclidean space. In that setting, the central result is that for every fixed dimension \(d\ge 2\), both exact solution and \((1+\epsilon)\)-approximation for discrete Euclidean \(k\)-Center have an intrinsic complexity governed by the exponent \(k^{1-1/d}\), under the **Exponential Time Hypothesis (ETH)** [2203.08328]. The phrase should be distinguished from other “center” lower-bound notions, such as the **Lower Bounded Center** clustering problem, where the number of centers is unrestricted but each chosen center must serve at least \(\lambda\) points [1304.7318]. ## 1. Discrete Euclidean \(k\)-Center and the lower-bound question In the discrete \(k\)-Center problem, one is given a metric space \((P,\mathrm{dist})\) with \(|P|=n\) and an integer \(k\le n\), and seeks a set of centers \[ C\subseteq P,\qquad |C|=k, \] minimizing \[ \max_{p\in P}\min_{c\in C}\mathrm{dist}(p,c). \] Equivalently, the task is to find the minimum radius \(r\) such that \(k\) closed balls of radius \(r\), centered at points of \(P\), cover all of \(P\) [2203.08328]. The lower-bound results concern the geometric regime \(P\subseteq \mathbb R^d\) with Euclidean distance \(\ell_2\), for fixed \(d\ge 2\). The focus is explicitly on **discrete centers**: the chosen centers must belong to the input point set \(P\). This restriction is essential, because the lower-bound reduction does **not** extend to the continuous version in which centers may be arbitrary points of \(\mathbb R^d\) [2203.08328]. The question addressed by the main lower-bound theory is not merely whether discrete Euclidean \(k\)-Center is hard in a classical NP-hardness sense, but whether the known dependence on \(k\), \(n\), and \(\epsilon\) in the best algorithms can be asymptotically improved. The answer given is negative under ETH: the characteristic exponent \(k^{1-1/d}\) cannot be eliminated from either exact or approximation algorithms in fixed dimension [2203.08328]. ## 2. ETH-based lower bounds and their exact form The main theorem states that for every fixed \(d\ge 2\), under ETH, discrete \(k\)-Center in \(d\)-dimensional Euclidean space does not admit a \((1+\epsilon)\)-approximation in time \[ f(k)\cdot \left(\frac{1}{\epsilon}\right)^{o\left(k^{1-1/d}\right)}\cdot n^{o\left(k^{1-1/d}\right)} \] for any computable function \(f\), and also cannot be solved exactly in time \[ f(k)\cdot n^{o\left(k^{1-1/d}\right)} \] for any computable function \(f\) [2203.08328]. These bounds apply to every fixed dimension \(d\ge 2\). For \(d=2\), the exponent becomes \[ k^{1-1/2}=\sqrt{k}, \] recovering the classical square-root behavior. More generally, the exponent \[ k^{1-1/d} \] is the higher-dimensional analogue of that planar phenomenon [2203.08328]. The paper also proves a concrete gap theorem. Given a \(d\)-dimensional geometric CSP instance \(\mathcal I=(V,D,C)\), it constructs a discrete \(k\)-Center instance \(\mathcal U\subseteq \mathbb R^d\) with \[ k=|V| \] such that, with \[ r:=\frac14,\qquad \epsilon^*:=\frac{r^2}{(d-1)\delta^2}=\frac{1}{16(d-1)\delta^2}, \] where \(D=[\delta]^d\), one has: \[ \mathcal I \text{ satisfiable } \Longrightarrow \mathrm{OPT}(\mathcal U)<2r, \] and \[ \mathcal I \text{ unsatisfiable } \Longrightarrow \mathrm{OPT}(\mathcal U)\ge 2r(1+\epsilon^*). \] Thus the construction creates a multiplicative gap of \(1+\epsilon^*\) between YES and NO instances [2203.08328]. A common misunderstanding is to read this as a lower bound for arbitrary metric \(k\)-Center or for the continuous geometric problem. The result is narrower and more precise: it is a lower bound for **discrete Euclidean \(k\)-Center in fixed dimension**, and the paper explicitly states that its reduction does not extend to the continuous variant [2203.08328]. ## 3. Tightness relative to classical upper bounds The lower bounds are formulated as matching results against algorithms of Agarwal and Procopiuc. For approximation, Agarwal–Procopiuc gave a \((1+\epsilon)\)-approximation algorithm for \(d\)-dimensional Euclidean \(k\)-Center running in \[ O(dn\log k)+\left(\frac{k}{\epsilon}\right)^{O\left(k^{1-1/d}\right)}\cdot n^{O(1)}. \] For exact solution, they gave an algorithm running in \[ n^{O\left(d\cdot k^{1-1/d}\right)}. \] The lower-bound paper interprets these algorithms as essentially optimal and asymptotically optimal, respectively, because ETH rules out improving away the exponent \(k^{1-1/d}\) up to constant-factor losses in the exponent and polynomial factors [2203.08328]. For approximation, the point is subtle. The lower bound excludes running times of the form \[ f(k)\cdot \left(\frac{1}{\epsilon}\right)^{o\left(k^{1-1/d}\right)}\cdot n^{o\left(k^{1-1/d}\right)}, \] while the upper bound is polynomial in \(n\). This suggests that one cannot asymptotically beat the exponent \(k^{1-1/d}\) simultaneously in the dependence on \(1/\epsilon\) and in the dependence on \(n\), even if arbitrary computable dependence on \(k\) is allowed [2203.08328]. For exact algorithms, the comparison is cleaner. Since \(d\) is fixed, the upper bound \[ n^{O\left(d\cdot k^{1-1/d}\right)} \] differs from the forbidden \[ f(k)\cdot n^{o\left(k^{1-1/d}\right)} \] only by a constant-factor loss in the exponent. This is why the exact algorithm is described as asymptotically optimal [2203.08328]. The broader significance is that the lower bound is not merely a hardness statement; it identifies the correct asymptotic exponent for discrete Euclidean \(k\)-Center in fixed dimension. ## 4. Source of the exponent: reduction from geometric CSP The exponent \(k^{1-1/d}\) comes from a reduction from a \(d\)-dimensional geometric CSP of Marx and Sidiropoulos. In that source problem, a binary CSP instance \(\mathcal I=(V,D,C)\) has \[ V\subseteq [N]^d, \] viewed as vertices of the \(d\)-dimensional grid, domain \[ D=[\delta]^d, \] and a constraint graph that is an induced subgraph of the grid. Binary constraints occur only between neighboring grid vertices \(a'=a\oplus e_i\), and have the form \[ R_{a,a'}=\{(x,y)\in R_a\times R_{a'}\mid x[i]\ge y[i]\}. \] The paper notes that Marx–Sidiropoulos used \(\le\)-constraints, but \(\ge\)-constraints are equivalent by a simple value-complement transformation [2203.08328]. The crucial source theorem states that for fixed \(d\ge 2\), if such a geometric CSP instance \(\mathcal I\) can be solved in time \[ f(|V|)\cdot |\mathcal I|^{o\left(|V|^{1-1/d}\right)} \] for some computable \(f\), then ETH fails. This is the mechanism that transfers the \(d\)-dimensional grid phenomenon into the \(k\)-Center exponent \(k^{1-1/d}\) [2203.08328]. At a high level, the reduction assigns one variable gadget per CSP variable and sets \[ k=|V|. \] A low-radius solution must effectively choose one representative center per variable gadget. These selected centers encode a CSP assignment. Completeness shows that a satisfying assignment yields radius \(<2r\); soundness shows that any radius \(<2r(1+\epsilon)\) solution induces an assignment, and a violated constraint forces at least one secondary point to remain uncovered at that radius [2203.08328]. This suggests that the lower bound is driven not by an ad hoc geometric obstruction, but by the combinatorics of grid-like constraint graphs in dimension \(d\). ## 5. Geometric construction and distance encoding The reduction is explicit. Fix \(d\ge 2\), let \(\mathcal I=(V,D,C)\) with \(D=[\delta]^d\), and define \[ r:=\frac14,\qquad \epsilon:=\frac{r^2}{(d-1)\delta^2}=\frac{1}{16(d-1)\delta^2}. \] The construction repeatedly uses \[ 0<\epsilon\le \epsilon\delta\le \epsilon\delta^2\le \epsilon\delta^2(d-1)=r^2=\frac1{16}. \] The \(k\)-Center instance is a point set \(\mathcal U\subseteq \mathbb R^d\) built from several gadget families [2203.08328]. For each variable \(a\in V\), there are **border points** \[ B_a^{+i}=a\oplus e_i\cdot r(1-\epsilon)\oplus (1^d-e_i)\cdot 2\epsilon\delta, \] \[ B_a^{-i}=a\ominus e_i\cdot r(1-\epsilon)\ominus (1^d-e_i)\cdot 2\epsilon\delta, \] and for each allowed value \(x\in R_a\subseteq [\delta]^d\), a **core point** \[ C_a^x=a\oplus \epsilon x. \] For each CSP edge \((a,a')\) with \(a'=a\oplus e_i\), and each \(\ell\in[\delta]\), there is a **secondary point** \[ S_{\{a,a'\}^\ell = a\oplus e_i\cdot\big((1-\epsilon)2r+\epsilon\ell\big). \] The total number of points satisfies \[ n\le |V|\cdot 2d + |C| + |V|^2\cdot \delta = |\mathcal I|^{O(1)}. \] Thus the reduction remains polynomial in the CSP input size [2203.08328]. Several distance lemmas force the combinatorics of feasible center sets. For each variable \(a\in V\) and coordinate \(i\in[d]\), \[ \mathrm{dist}(B_a^{+i},B_a^{-i})\ge 2r(1+\epsilon), \] and in fact \[ \mathrm{dist}(B_a^{+i},B_a^{-i})^2 = (2r(1-\epsilon))^2+(d-1)(4\epsilon\delta)^2 = (2r(1+\epsilon))^2. \] This separation is what prevents a single small-radius center from serving incompatible roles inside the same gadget [2203.08328]. At the same time, every pair of core points in one variable gadget is at distance \(x[i] \Longrightarrow \mathrm{dist}(C_a^x,S_{\{a,a'\}^\ell)\ge 2r(1+\epsilon), \] \[ \ell>y[i] \Longrightarrow \mathrm{dist}(C_{a'}^y,S_{\{a,a'\}^\ell)<2r, \] \[ \ell\le y[i] \Longrightarrow \mathrm{dist}(C_{a'}^y,S_{\{a,a'\}^\ell)\ge 2r(1+\epsilon). \] These four inequalities are the core of the gap construction. A secondary point is coverable from the left gadget exactly when \(x[i]\ge \ell\), and from the right gadget exactly when \(y[i]<\ell\). Because the CSP constraint is \(x[i]\ge y[i]\), every threshold \(\ell\) is coverable from one side or the other exactly in the satisfiable case [2203.08328]. ## 6. Variants, special cases, and related center lower bounds A central distinction is between **discrete** and **continuous** \(k\)-Center. The lower bounds discussed here apply only when \[ C\subseteq P. \] The paper explicitly notes that while Agarwal–Procopiuc’s upper bounds also apply to the continuous version, the lower-bound reduction does not extend there [2203.08328]. Another distinction is between these ETH-based bounds for parameterized approximation and exact computation, and lower bounds for **small fixed \(k\)** under fine-grained hypotheses. For example, later work proved that Euclidean discrete \(2\)-center in \(\mathbb R^{13}\) has a conditional lower bound \[ \Omega(n^{2-\delta}) \] under the Hyperclique Hypothesis, and Euclidean discrete \(k\)-center in \(\mathbb R^{7k}\) has a conditional lower bound \[ \Omega(n^{k-\delta}) \] for fixed \(k\ge 3\) [2305.01892]. Those results address a different regime: exact algorithms for constant \(k\), rather than ETH-based dependence on parameter \(k\) and dimension \(d\). The phrase “center lower bound” can also refer to a different clustering problem entirely, the **Lower Bounded Center** problem. There, one is given a set \(P\) and a lower bound \(\lambda\), and must choose a set of centers \(C\subseteq P\) and assign every point to a center so that each center gets at least \(\lambda\) assigned points, minimizing the maximum assignment distance [1304.7318]. That problem has a near-linear-time \((4+\varepsilon)\)-approximation in fixed-dimensional Euclidean space and a hardness threshold of \(\frac{\sqrt{13}}{2}\approx 1.80\) in the plane [1304.7318]. It is therefore a distinct “lower-bounded center” problem, not the same object as the ETH lower bounds for discrete Euclidean \(k\)-Center. This terminological distinction is important because “center lower bound” may denote either a **lower bound on computational complexity for \(k\)-Center** or a **lower bound on cluster size in Lower Bounded Center**. The two problems are related by subject matter but differ in objective, constraints, and proof techniques. ## 7. Conceptual significance The principal message of the discrete Euclidean \(k\)-Center lower-bound theory is that the exponent \[ k^{1-1/d} \] is the correct asymptotic measure of difficulty for fixed-dimensional discrete Euclidean \(k\)-Center, both for exact computation and for PTAS-style \((1+\epsilon)\)-approximation [2203.08328]. The lower bounds are “tight” in the sense that they match, up to constant factors in the exponent and polynomial factors, the classical Agarwal–Procopiuc upper bounds. The reduction also clarifies why the exponent depends on dimension in precisely this way. The source of hardness is a \(d\)-dimensional geometric CSP whose grid structure produces the same \(k^{1-1/d}\) phenomenon. In \(d=2\), this is the square-root regime; in higher dimensions it becomes the natural \(d\)-dimensional generalization [2203.08328]. A plausible implication is that future improvements for discrete Euclidean \(k\)-Center are unlikely to come from asymptotically removing the \(k^{1-1/d}\) dependence. More promising directions would have to exploit restrictions outside the lower-bound framework, such as different center models, different metrics, or structurally special instances. The paper itself already isolates one such boundary: its lower bounds do not cover the continuous version of Euclidean \(k\)-Center [2203.08328]. Source: https://www.emergentmind.com/topics/center-lower-bound