Metric $1$ -median selection: Query complexity vs. approximation ratio

Ching-Lueh Chang ¹¹1Department of Computer Science and Engineering, Yuan Ze University, Taoyuan, Taiwan. Email: clchang@saturn.yzu.edu.tw ²²2Innovation Center for Big Data and Digital Convergence, Yuan Ze University, Taoyuan, Taiwan.

Abstract

Consider the problem of finding a point in a metric space $(\{1,2,\ldots,n\},d)$ with the minimum average distance to other points. We show that this problem has no deterministic $o(n^{1+1/(h-1)})$ -query $(2h-\Omega(1))$ -approximation algorithms for any constant $h\in\mathbb{Z}^{+}\setminus\{1\}$ .

1 Introduction

The metric $1$ -median problem asks for a point in an $n$ -point metric space with the minimum average distance to other points. It has a Monte-Carlo $O(n/\epsilon^{2})$ -time $(1+\epsilon)$ -approximation algorithm for all $\epsilon>0$ [6, 7]. In $\mathbb{R}^{D}$ , Kumar et al. [8] give a Monte-Carlo $O(2^{\text{poly}(1/\epsilon)}D)$ -time $(1+\epsilon)$ -approximation algorithm for $1$ -median selection and another algorithm for $k$ -median selection, where $D\geq 1$ and $\epsilon>0$ . Guha et al. [5] give streaming approximation algorithms for $k$ -median selection in metric spaces.

Chang [3], Wu [11] and Chang [1] show that metric $1$ -median has a deterministic nonadaptive $O(n^{1+1/h})$ -time $(2h)$ -approximation algorithm for all constants $h\in\mathbb{Z}^{+}\setminus\{1\}$ . Furthermore, Chang [4] shows the nonexistence of deterministic $o(n^{2})$ -time $(4-\Omega(1))$ -approximation algorithms for metric $1$ -median. This paper generalizes his result to show that metric $1$ -median has no deterministic $o(n^{1+1/(h-1)})$ -query $(2h-\Omega(1))$ -approximation algorithms for any constant $h\in\mathbb{Z}^{+}\setminus\{1\}$ . Combining our result with an existing upper bound [11, 1],

			$\displaystyle\min\left\{c\geq 1\mid\text{{\sc metric $1$-median} has a deterministic $O(n^{1+\epsilon})$-query $c$-approx.\ alg.}\right\}$
		$\displaystyle=$	$\displaystyle\min\left\{c\geq 1\mid\text{{\sc metric $1$-median} has a deterministic $O(n^{1+\epsilon})$-time $c$-approx. alg.}\right\}$
		$\displaystyle=$	$\displaystyle 2\left\lceil\frac{1}{\epsilon}\right\rceil$

for all constants $\epsilon\in(0,1)$ . That is, we determine the best approximation ratio of deterministic $O(n^{1+\epsilon})$ -query (resp., $O(n^{1+\epsilon})$ -time) algorithms for all $\epsilon\in(0,1)$ .

As in the previous lower bounds for deterministic algorithms [4, 2], we use an adversarial method. Roughly speaking, our proof proceeds as follows:

(i)
Design an adversary Adv for answering the distance queries of any deterministic algorithm $A$ with query complexity $q(n)=o(n^{1+1/(h-1)})$ .
(ii)
Show that $A$ ’s output has a large average distance to other points, according to Adv’s answers to $A$ .
(iii)
Construct a distance function with respect to which a certain point $\hat{\alpha}$ has a small average distance to other points.
(iv)
Construct the final distance function $d(\cdot,\cdot)$ similar to that in item (iii).
(v)
Show that $d$ is a metric.
(vi)
Show the consistency of $d(\cdot,\cdot)$ with Adv’s answers.
(vii)
Compare $\hat{\alpha}$ in item (iii) with $A$ ’s output to establish our lower bound on $A$ ’s approximation ratio.

Central to our constructions are two graph sequences, $\{H^{(i)}\}_{i=0}^{q(n)}$ and $\{G^{(i)}\}_{i=0}^{q(n)}$ in Sec. 3, that are unseen in previous lower bounds [9, 2, 4]. Like in [4], we need a small set $S$ of points whose distances to other points are answered as large values during $A$ ’s execution, and yet we assign a small value to the distances from a certain point $\hat{\alpha}\in S$ to many other points in item (iii).

This paper is organized as follows. Sec. 2 introduces the terminologies. Sec. 3 proves our main theorem that metric $1$ -median has no deterministic $o(n^{1+1/(h-1)})$ -query $(2h-\Omega(1))$ -approximation algorithms for any constant $h\in\mathbb{Z}^{+}\setminus\{1\}$ . In particular, Secs. 3.1, 3.2, 3.3 and 3.4 correspond to items (ii), (iii), (iv)–(vi) and (vii) above, respectively.

2 Definitions

A finite metric space $(M,d)$ is a finite set $M$ endowed with a function $d\colon M^{2}\to[0,\infty)$ such that

•
$d(x,x)=0$ ,
•
$d(x,y)>0$ if $x\neq y$ ,
•
$d(x,y)=d(y,x)$ , and
•
$d(x,y)+d(y,z)\geq d(x,z)$

for all $x$ , $y$ , $z\in M$ [10]. For all $c\geq 1$ , a point $z\in M$ is said to be a $c$ -approximate $1$ -median of $(M,d)$ if

\sum_{x\in M}\,d\left(z,x\right)\leq c\cdot\sum_{x\in M}\,d\left(y,x\right)

for all $y\in M$ . For convenience, $[n]\stackrel{{\scriptstyle\text{def.}}}{{=}}\{1,2,\ldots,n\}$ .

For deterministic algorithms $A$ and ${\cal O}\colon\{1,2,\ldots,n\}^{2}\to\mathbb{R}$ , denote by $A^{\cal O}(1^{n})$ the execution of $A$ with oracle access to $\cal O$ and with input $1^{n}$ , where $n\in\mathbb{N}$ . As the input to $A$ will be $1^{n}$ throughout this paper, abbreviate $A^{\cal O}(1^{n})$ as $A^{\cal O}$ . If $A^{d}$ outputs a $c$ -approximate $1$ -median of $([n],d)$ for each finite metric space $([n],d)$ , then $A$ is said to be $c$ -approximate for metric $1$ -median, where $c\geq 1$ .

Fact 1 ([3, 1, 11]).

For each constant $h\in\mathbb{Z}^{+}\setminus\{1\}$ , metric $1$ -median has a deterministic nonadaptive $O(n^{1+1/h})$ -time $(2h)$ -approximation algorithm.

A weighted undirected graph $G=(V,E,w)$ has a finite vertex set $V$ , an edge set $E$ and a weight function $w\colon E\to(0,\infty)$ , where each edge is an unordered pair of distinct vertices in $V$ . If $w\colon Y\to(0,\infty)$ for a superset $Y$ of $E$ , interpret $(V,E,w)$ simply as $(V,E,w|_{E})$ , where $w|_{E}$ denotes the restriction of $w$ on $E$ . For all $v\in V$ , let

N_{G}(v)\stackrel{{\scriptstyle\text{def.}}}{{=}}\left\{u\in V\mid(u,v)\in E\right\}

and $\text{\rm deg}_{G}(v)\stackrel{{\scriptstyle\text{def.}}}{{=}}|N_{G}(v)|$ . For all $S\subseteq V$ , $N_{G}(S)\stackrel{{\scriptstyle\text{def.}}}{{=}}\bigcup_{v\in S}\,N_{G}(v)$ . For all $s$ , $t\in V$ , an $s$ - $t$ path $P$ in $G$ is a sequence $\{v_{i}\in V\}_{i=0}^{k}$ satisfying $k\in\mathbb{N}$ , $v_{0}=s$ , $v_{k}=t$ and $(v_{i},v_{i+1})\in E$ for all $i\in\{0,1,\ldots,k-1\}$ . Its weight (or length) is $w(P)\stackrel{{\scriptstyle\text{def.}}}{{=}}\sum_{i=0}^{k-1}\,w(v_{i},v_{i+1})$ .³³3 $w(P)$ is a common and convenient abuse of notation. The shortest $s$ - $t$ distance in $G$ is

d_{G}(s,t)=\inf\left\{w(P)\mid\text{$P$ is an $s$-$t$ path in $G$}\right\},

where $s$ , $t\in V$ . So $d_{G}(s,t)=\infty$ if $G$ has no $s$ - $t$ paths. Note that we allow only positive weights, i.e., $\mathop{\mathrm{Im}}(w)\subseteq(0,\infty)$ . So a shortest $s$ - $t$ path must be simple, i.e., it does not repeat vertices. If $w\equiv 1$ , abbreviate $(V,E,w)$ as $(V,E)$ and call it an unweighted graph.

The following fact is well-known.

Fact 2.

For each undirected graph $G=(V,E)$ ,

\sum_{v\in V}\,\text{\rm deg}_{G}(v)=2\cdot|E|.

For a predicate $P$ , let $\chi[P]=1$ if $P$ is true and $\chi[P]=0$ otherwise. The following fact about geometric series is not hard to see.

Fact 3.

For all $r\geq 2$ and $m\in\mathbb{N}$ ,

\sum_{k=0}^{m}\,r^{k}\leq 2r^{m}.

3 Query complexity vs. approximation ratio

Throughout this section,

•
$n\in\mathbb{Z}^{+}$ ,
•
$\delta\in(0,1)$ and $h\in\mathbb{Z}^{+}\setminus\{1\}$ are constants (i.e., they are independent of $n$ ),
•
$A$ is a deterministic $o(n^{1+1/(h-1)})$ -query algorithm for metric $1$ -median, and
•
$S=[\lfloor\delta n\rfloor]\subseteq[n]$ .

All pairs in $[n]^{2}$ are assumed to be unordered in this section. So, e.g., $(1,2)\in\{2\}\times[n]$ . By padding at most $n-1$ dummy queries, assume without loss of generality that $A$ will have queried for the distances between its output and all other points when halting. Denote $A$ ’s query complexity by

q(n)=o\left(n^{1+1/(h-1)}\right).

Without loss of generality, forbid making the same query twice or querying for the distance from a point to itself, where the queries for $d(x,y)$ and $d(y,x)$ are considered to be the same for $x$ , $y\in[n]$ . Furthermore, let $n$ be sufficiently large to satisfy

$\displaystyle q(n)$	$\displaystyle\leq$	$\displaystyle\delta n^{1+1/(h-1)},$	(1)
$\displaystyle\delta n^{1/(h-1)}$	$\displaystyle>$	$\displaystyle 3,$	(2)
$\displaystyle\frac{2q(n)}{\|S\|-1}$	$\displaystyle\leq$	$\displaystyle\delta n^{1/(h-1)}.$	(3)

Define two unweighted undirected graphs $G^{(0)}$ and $H^{(0)}$ by

$\displaystyle E_{G}^{(0)}$	$\displaystyle\stackrel{{\scriptstyle\text{def.}}}{{=}}$	$\displaystyle\left\{\left(u,v\right)\mid\left(u,v\in[n]\setminus S\right)\land\left(u\neq v\right)\right\},$	(4)
$\displaystyle G^{(0)}$	$\displaystyle\stackrel{{\scriptstyle\text{def.}}}{{=}}$	$\displaystyle\left([n],E_{G}^{(0)}\right),$	(5)
$\displaystyle E_{H}^{(0)}$	$\displaystyle\stackrel{{\scriptstyle\text{def.}}}{{=}}$	$\displaystyle\emptyset,$	(6)
$\displaystyle H^{(0)}$	$\displaystyle\stackrel{{\scriptstyle\text{def.}}}{{=}}$	$\displaystyle\left([n],E_{H}^{(0)}\right).$	(7)

1: Let

E_{G}^{(0)}

G^{(0)}

E_{H}^{(0)}

and

H^{(0)}

be as in equations (4)–(7);

2: for

i=1

2

\ldots

q(n)

3: Receive the

i

th query of

A

, denoted

(a_{i},b_{i})

;

4: if

d_{G^{(i-1)}}(a_{i},b_{i})\leq h

then

5: Find a shortest

a_{i}

b_{i}

path

P_{i}

G^{(i-1)}

;

E_{H}^{(i)}\leftarrow E_{H}^{(i-1)}\cup\{e\mid\text{$e$ is an edge on $P_{i}$}\}

;

H^{(i)}\leftarrow([n],E_{H}^{(i)})

;

E_{G}^{(i)}\leftarrow E_{G}^{(i-1)}\setminus\{(u,v)\in E_{G}^{(i-1)}\setminus E_{H}^{(i)}\mid(\text{deg}_{H^{(i)}}(u)\geq\delta n^{1/(h-1)}-2)\lor(\text{deg}_{H^{(i)}}(v)\geq\delta n^{1/(h-1)}-2)\}

;

G^{(i)}\leftarrow([n],E_{G}^{(i)})

;

10: else

11:

E_{H}^{(i)}\leftarrow E_{H}^{(i-1)}

;

12:

H^{(i)}\leftarrow([n],E_{H}^{(i)})

;

13:

E_{G}^{(i)}\leftarrow E_{G}^{(i-1)}

;

14:

G^{(i)}\leftarrow([n],E_{G}^{(i)})

;

15: end if

16:

Q^{(i)}\leftarrow([n],\{(a_{j},b_{j})\mid j\in[i]\})

;

17: Output

\min\{d_{H^{(i)}}(a_{i},b_{i}),h-(1/2)\cdot\chi[\exists v\in\{a_{i},b_{i}\},\,(v\in S)\land(\text{deg}_{Q^{(i)}}(v)\leq\delta n^{1/(h-1)})]\}

as the answer to the

i

th query of

A

;

18: end for

Figure 1: Algorithm Adv for answering

A

’s queries

Algorithm Adv in Fig. 1 answers $A$ ’s queries. In particular, for all $i\in[q(n)]$ , the $i$ th iteration of the loop of Adv answers the $i$ th query of $A$ , denoted $(a_{i},b_{i})\in[n]^{2}$ . It constructs three unweighted undirected graphs, $G^{(i)}=([n],E_{G}^{(i)})$ , $H^{(i)}=([n],E_{H}^{(i)})$ and $Q^{(i)}$ . As $G^{(i-1)}$ is unweighted for all $i\in[q(n)]$ , $P_{i}$ in line 5 of Adv is an $a_{i}$ - $b_{i}$ path in $G^{(i-1)}$ with the minimum number of edges. By line 16 of Adv, the edges of $Q^{(i)}$ are precisely the first $i$ queries of $A$ .

Lemma 4.

E_{H}^{(0)}\subseteq E_{H}^{(1)}\subseteq\ldots\subseteq E_{H}^{(q(n))}\subseteq E_{G}^{(q(n))}\subseteq E_{G}^{(q(n)-1)}\subseteq\ldots\subseteq E_{G}^{(0)}.

Proof.

By lines 6 and 11 of Adv in Fig. 1, $E_{H}^{(i-1)}\subseteq E_{H}^{(i)}$ for all $i\in[q(n)]$ . By lines 8 and 13, $E_{G}^{(i)}\subseteq E_{G}^{(i-1)}$ for all $i\in[q(n)]$ .

To show that $E_{H}^{(q(n))}\subseteq E_{G}^{(q(n))}$ , we shall prove the stronger statement that $E_{H}^{(i)}\subseteq E_{G}^{(i)}$ for all $i\in\{0,1,\ldots,q(n)\}$ by mathematical induction. By equation (6), $E_{H}^{(0)}\subseteq E_{G}^{(0)}$ . Assume as the induction hypothesis that $E_{H}^{(i-1)}\subseteq E_{G}^{(i-1)}$ . The following shows that $E_{H}^{(i)}\subseteq E_{G}^{(i-1)}$ by examining each $e\in E_{H}^{(i)}$ :

Case 1:
$e\in E_{H}^{(i-1)}$ . By the induction hypothesis, $e\in E_{G}^{(i-1)}$ .
Case 2:
$e\notin E_{H}^{(i-1)}$ . As $e\in E_{H}^{(i)}\setminus E_{H}^{(i-1)}$ , lines 6 and 11 show that $e$ is on $P_{i}$ (and that the $i$ th iteration of the loop of Adv runs line 6 rather than line 11). By line 5, each edge on $P_{i}$ is in $E_{G}^{(i-1)}$ . In particular, $e\in E_{G}^{(i-1)}$ .

Having shown that $E_{H}^{(i)}\subseteq E_{G}^{(i-1)}$ , lines 8 and 13 will both result in $E_{H}^{(i)}\subseteq E_{G}^{(i)}$ , completing the induction step. ∎

Lemma 5.

For all $i\in[q(n)]$ with $d_{G^{(i-1)}}(a_{i},b_{i})\leq h$ ,

d_{H^{(i)}}\left(a_{i},b_{i}\right)=d_{H^{(q(n))}}\left(a_{i},b_{i}\right)=d_{G^{(q(n))}}\left(a_{i},b_{i}\right)=d_{G^{(i-1)}}\left(a_{i},b_{i}\right).

Proof.

By line 4 of Adv, the $i$ th iteration of the loop runs lines 5–9. Lines 5–7 put (the edges of) a shortest $a_{i}$ - $b_{i}$ path in $G^{(i-1)}$ into $H^{(i)}$ ; hence

d_{H^{(i)}}\left(a_{i},b_{i}\right)\leq d_{G^{(i-1)}}\left(a_{i},b_{i}\right).

This and Lemma 4 complete the proof. ∎

Below is an easy consequence of Lemma 4.

Lemma 6.

For all $i\in[q(n)]$ with $d_{G^{(i-1)}}(a_{i},b_{i})>h$ ,

d_{G^{(q(n))}}(a_{i},b_{i})>h.

3.1 The average distance from $A$ ’s output to other points

This subsection shows that the output of $A^{\sf Adv}$ has a large average distance to other points, according to the answers of Adv.

Lemma 7.

For all $i\in[q(n)]$ and $v\in[n]$ ,

\text{\rm deg}_{H^{(i)}}(v)\leq\text{\rm deg}_{H^{(i-1)}}(v)+2.

Proof.

If the $i$ th iteration of the loop of Adv runs lines 11–14 but not 5–9, then $H^{(i)}=H^{(i-1)}$ , proving the lemma. So assume otherwise. Being shortest, $P_{i}$ in line 5 does not repeat vertices. Therefore, $v$ is incident to at most two edges on $P_{i}$ , which together with lines 6–7 complete the proof. ∎

Lemma 8.

For all $v\in[n]$ ,

\text{\rm deg}_{H^{(q(n))}}(v)<\delta n^{1/(h-1)}.

Proof.

Assume

\displaystyle\text{\rm deg}_{H^{(q(n))}}(v)\geq\delta n^{1/(h-1)}-2

(8)

for, otherwise, there is nothing to prove. Clearly,

\displaystyle\text{\rm deg}_{H^{(0)}}(v)\stackrel{{\scriptstyle\text{(\ref{initiallymarkededgeset})--(\ref{initiallymarkedgraph})}}}{{=}}0\stackrel{{\scriptstyle\text{(\ref{tediouscondition2})}}}{{<}}\delta n^{1/(h-1)}-2.

(9)

By inequalities (8)–(9), there exists $i\in[q(n)]$ satisfying

	$\displaystyle\text{\rm deg}_{H^{(i-1)}}(v)<\delta n^{1/(h-1)}-2,$		(10)
	$\displaystyle\text{\rm deg}_{H^{(i)}}(v)\geq\delta n^{1/(h-1)}-2.$		(11)

Clearly,

\displaystyle N_{G^{(i)}}(v)=\left\{u\in[n]\mid\left(u,v\right)\in E_{G}^{(i)}\right\}.

(12)

As $H^{(i-1)}\neq H^{(i)}$ by inequalities (10)–(11), the $i$ th iteration of the loop of Adv runs lines 5–9 but not 11–14. By inequality (11) and line 8 of Adv,

\displaystyle\left\{u\in[n]\mid\left(u,v\right)\in E_{G}^{(i)}\right\}=\left\{u\in[n]\mid\left(u,v\right)\in E_{G}^{(i-1)}\setminus\left(E_{G}^{(i-1)}\setminus E_{H}^{(i)}\right)\right\}.

(13)

Equations (12)–(13) and Lemma 4 give

\displaystyle N_{G^{(i)}}(v)=\left\{u\in[n]\mid\left(u,v\right)\in E_{H}^{(i)}\right\}.

(14)

By inequality (10) and Lemma 7,

\displaystyle\text{\rm deg}_{H^{(i)}}(v)<\delta n^{1/(h-1)}.

This and equation (14) imply $\text{\rm deg}_{G^{(i)}}(v)<\delta n^{1/(h-1)}$ , which together with Lemma 4 completes the proof. ∎

Lemma 9.

For all $v\in[n]$ ,

\left|\left\{u\in[n]\mid d_{H^{(q(n))}}\left(v,u\right)<h\right\}\right|\leq 2\delta^{h-1}n.

Proof.

By Lemma 8,

\left|\left\{u\in[n]\mid\text{$\exists$ $v$-$u$ path in $H^{(q(n))}$ with exactly $k$ edges}\right\}\right|\leq\left(\delta n^{1/(h-1)}\right)^{k}

for all $k\in\mathbb{N}$ . Consequently,

	$\displaystyle\left\|\left\{u\in[n]\mid\text{$\exists$ $v$-$u$ path in $H^{(q(n))}$ with at most $h-1$ edges}\right\}\right\|$	$\displaystyle\leq$	$\displaystyle\sum_{k=0}^{h-1}\,\left(\delta n^{1/(h-1)}\right)^{k}$
		$\displaystyle\stackrel{{\scriptstyle\text{(\ref{tediouscondition2})~{}and~{}Fact~{}\ref{geometricseriesbound}}}}{{\leq}}$	$\displaystyle 2\delta^{h-1}n.$

Finally, recall that $H^{(q(n))}$ is unweighted. ∎

Denote the output of $A^{\text{\sf Adv}}$ by $z$ . Furthermore,

\displaystyle I\stackrel{{\scriptstyle\text{def.}}}{{=}}\left\{j\in\left[q(n)\right]\mid z\in\left\{a_{j},b_{j}\right\}\right\}.

(15)

The following lemma analyzes the sum of the distances, as answered by line 17 of Adv, from $z$ to other points.

Lemma 10.

			$\displaystyle\sum_{i\in I}\,\min\left\{d_{H^{(i)}}\left(a_{i},b_{i}\right),h-\frac{1}{2}\cdot\chi\left[\exists v\in\left\{a_{i},b_{i}\right\},\,\left(v\in S\right)\land\left(\text{\rm deg}_{Q^{(i)}}(v)\leq\delta n^{1/(h-1)}\right)\right]\right\}$
		$\displaystyle\geq$	$\displaystyle n\cdot\left(h-2h\delta^{h-1}-o(1)-\delta\right).$

Proof.

By Lemma 4,

			$\displaystyle\sum_{i\in I}\,\min\left\{d_{H^{(i)}}\left(a_{i},b_{i}\right),h-\frac{1}{2}\cdot\chi\left[\exists v\in\left\{a_{i},b_{i}\right\},\,\left(v\in S\right)\land\left(\text{\rm deg}_{Q^{(i)}}(v)\leq\delta n^{1/(h-1)}\right)\right]\right\}$
		$\displaystyle\geq$	$\displaystyle\sum_{i\in I}\,\min\left\{d_{H^{(q(n))}}\left(a_{i},b_{i}\right),h-\frac{1}{2}\cdot\chi\left[\exists v\in\left\{a_{i},b_{i}\right\},\,\left(v\in S\right)\land\left(\text{\rm deg}_{Q^{(i)}}(v)\leq\delta n^{1/(h-1)}\right)\right]\right\}\,\,\,\,\,\,\,\,\,\,$
		$\displaystyle\geq$	$\displaystyle\sum_{i\in I}\,\min\left\{d_{H^{(q(n))}}\left(a_{i},b_{i}\right),h\right\}$
		$\displaystyle-$	$\displaystyle\sum_{i\in I}\,\frac{1}{2}\cdot\chi\left[\exists v\in\left\{a_{i},b_{i}\right\},\,\left(v\in S\right)\land\left(\text{\rm deg}_{Q^{(i)}}(v)\leq\delta n^{1/(h-1)}\right)\right].$

For all $i\in I$ , there exists $c_{i}\in[n]$ with $\{z,c_{i}\}=\{a_{i},b_{i}\}$ by equation (15). Therefore,

\sum_{i\in I}\,\min\left\{d_{H^{(q(n))}}\left(a_{i},b_{i}\right),h\right\}=\sum_{i\in I}\,\min\left\{d_{H^{(q(n))}}\left(z,c_{i}\right),h\right\}.

As we forbid repeated queries, $\{c_{i}\}_{i\in I}$ is a sequence of distinct points. So by Lemma 9,

\sum_{i\in I}\,\min\left\{d_{H^{(q(n))}}\left(z,c_{i}\right),h\right\}\geq h\cdot\left(|I|-2\delta^{h-1}n\right).

Recall that $A^{\sf Adv}$ will have queried for the distances between its output (which is $z$ ) and all other points when halting. So

|I|\geq n-1

by equation (15).⁴⁴4Because we forbid repeated queries and queries for the distance from a point to itself, we also have $|I|\leq n-1$ .

Clearly,

			$\displaystyle\sum_{i\in I}\,\chi\left[\exists v\in\left\{a_{i},b_{i}\right\},\,\left(v\in S\right)\land\left(\text{\rm deg}_{Q^{(i)}}(v)\leq\delta n^{1/(h-1)}\right)\right]$
		$\displaystyle=$	$\displaystyle\sum_{i\in I}\,\chi\left[\exists v\in\left\{z,c_{i}\right\},\,\left(v\in S\right)\land\left(\text{\rm deg}_{Q^{(i)}}(v)\leq\delta n^{1/(h-1)}\right)\right]$
		$\displaystyle\leq$	$\displaystyle\sum_{i\in I}\,\chi\left[\left(z\in S\right)\land\left(\text{\rm deg}_{Q^{(i)}}\left(z\right)\leq\delta n^{1/(h-1)}\right)\right]$
		$\displaystyle+$	$\displaystyle\sum_{i\in I}\,\chi\left[\left(c_{i}\in S\right)\land\left(\text{\rm deg}_{Q^{(i)}}\left(c_{i}\right)\leq\delta n^{1/(h-1)}\right)\right].$

By line 16 of Adv and equation (15),

\displaystyle\text{deg}_{Q^{(i)}}\left(z\right)=\left|\left\{j\in I\mid j\leq i\right\}\right|.

Therefore,

	$\displaystyle\sum_{i\in I}\,\chi\left[\left(z\in S\right)\land\left(\text{\rm deg}_{Q^{(i)}}\left(z\right)\leq\delta n^{1/(h-1)}\right)\right]$	$\displaystyle\leq$	$\displaystyle\sum_{i\in I}\,\chi\left[\left\|\left\{j\in I\mid j\leq i\right\}\right\|\leq\delta n^{1/(h-1)}\right]$
		$\displaystyle\leq$	$\displaystyle\delta n^{1/(h-1)},$

where the last inequality follows because $|\{j\in I\mid j\leq i\}|=k$ when $i$ is the $k$ th smallest element of $I$ , for all $k\in[|I|]$ . Recall the distinctness of the points in $\{c_{i}\}_{i\in I}$ . Therefore,

\displaystyle\sum_{i\in I}\,\chi\left[\left(c_{i}\in S\right)\land\left(\text{\rm deg}_{Q^{(i)}}\left(c_{i}\right)\leq\delta n^{1/(h-1)}\right)\right]\leq\sum_{i\in I}\,\chi\left[c_{i}\in S\right]\leq|S|=\left\lfloor\delta n\right\rfloor.

(17)

Inequalities (3.1)–(17) complete the proof. ∎

3.2 Planting a point with a small average distance to other points

This subsection constructs a distance function with respect to which a certain point has an average distance of approximately $1/2$ to other points.

Lemma 11.

\left|E_{H}^{(q(n))}\right|\leq h\cdot q(n).

Proof.

Consider the $i$ th iteration of the loop of Adv, where $i\in[q(n)]$ .

•
Running lines 4–5 results in $P_{i}$ having at most $h$ edges. Consequently,
$\displaystyle\left|E_{H}^{(i)}\right|\leq\left|E_{H}^{(i-1)}\right|+h$ (18)
by line 6.
•
Running line 11 yields $|E_{H}^{(i)}|=|E_{H}^{(i-1)}|$ , implying inequality (18) as well.

Now,

\left|E_{H}^{(q(n))}\right|-\left|E_{H}^{(0)}\right|=\sum_{i=1}^{q(n)}\,\left(\left|E_{H}^{(i)}\right|-\left|E_{H}^{(i-1)}\right|\right)\stackrel{{\scriptstyle\text{(\ref{theincreaseofthenumberofmarkededges})}}}{{\leq}}h\cdot q(n).

Finally, $|E_{H}^{(0)}|=0$ by equation (6). ∎

Lemma 12.

\left|\left\{u\in[n]\mid\text{\rm deg}_{H^{(q(n))}}(u)\geq\delta n^{1/(h-1)}-2\right\}\right|=\frac{h}{\delta}\cdot o(n).

⁵⁵5We explicitly write down the constants

h

and

\delta

on the right-hand side for clarity, although they can be absorbed within

o(\cdot)

Proof.

By Fact 2, the average degree in $H^{(q(n))}$ is

\frac{1}{n}\cdot 2\cdot\left|E_{H}^{(q(n))}\right|.

So by the averaging argument (that any finite nonempty sequence of nonnegative numbers with average $\bar{a}$ has at most an $\bar{a}/t$ fraction of numbers that are greater than or equal to $t>0$ ),

\frac{1}{n}\cdot\left|\left\{u\in[n]\mid\text{\rm deg}_{H^{(q(n))}}(u)\geq\delta n^{1/(h-1)}-2\right\}\right|\leq\frac{1}{n}\cdot 2\cdot\left|E_{H}^{(q(n))}\right|\cdot\frac{1}{\delta n^{1/(h-1)}-2},

where the rightmost denominator is positive and is $\Theta(\delta n^{1/(h-1)})$ by equation (2). This and Lemma 11 complete the proof. ∎

By inequality (2), $S\setminus\{z\}\neq\emptyset$ . Let

\displaystyle\hat{\alpha}\stackrel{{\scriptstyle\text{def.}}}{{=}}\mathop{\mathrm{argmin}}_{\alpha\in S\setminus\{z\}}\,\text{deg}_{Q^{(q(n))}}(\alpha),

(19)

breaking ties arbitrarily.

Lemma 13.

For all $i\in[q(n)]$ ,

\displaystyle\text{\rm deg}_{Q^{(i)}}\left(\hat{\alpha}\right)\leq\delta n^{1/(h-1)}.

Proof.

By line 16 of Adv,

\displaystyle\text{deg}_{Q^{(i)}}\left(\hat{\alpha}\right)\leq\text{deg}_{Q^{(q(n))}}\left(\hat{\alpha}\right).

(20)

By equation (19) and the averaging argument,

\displaystyle\text{deg}_{Q^{(q(n))}}(\hat{\alpha})\leq\frac{1}{|S\setminus\{z\}|}\cdot\sum_{\alpha\in S\setminus\{z\}}\,\text{deg}_{Q^{(q(n))}}(\alpha).

Furthermore,

\displaystyle\sum_{\alpha\in S\setminus\{z\}}\,\text{deg}_{Q^{(q(n))}}(\alpha)\leq\sum_{\alpha\in[n]}\,\text{deg}_{Q^{(q(n))}}(\alpha)=2q(n),\,\,\,

(21)

where the equality follows from Fact 2, line 16 of Adv and the non-repeating of queries. Finally,

\displaystyle\text{deg}_{Q^{(i)}}(\hat{\alpha})\stackrel{{\scriptstyle\text{(\ref{trivialbecausethequerygraphgrows})--(\ref{sumofdegreesinthequerygraph})}}}{{\leq}}\frac{2q(n)}{|S|-1}\stackrel{{\scriptstyle\text{(\ref{tediouscondition3})}}}{{\leq}}\delta n^{1/(h-1)}.

∎

Inductively, let

$\displaystyle V_{0}$	$\displaystyle\stackrel{{\scriptstyle\text{def.}}}{{=}}$	$\displaystyle\left\{\hat{\alpha}\right\},$	(22)
$\displaystyle V_{1}$	$\displaystyle\stackrel{{\scriptstyle\text{def.}}}{{=}}$	$\displaystyle N_{Q^{(q(n))}}\left(\hat{\alpha}\right)\setminus V_{0},$	(23)
$\displaystyle V_{j+1}$	$\displaystyle\stackrel{{\scriptstyle\text{def.}}}{{=}}$	$\displaystyle N_{H^{(q(n))}}\left(V_{j}\right)\setminus\left(\bigcup_{i=0}^{j}\,V_{i}\right)$	(24)

for all $j\in[h-2]$ . Furthermore,

\displaystyle V_{h}\stackrel{{\scriptstyle\text{def.}}}{{=}}[n]\setminus\left(\bigcup_{i=0}^{h-1}\,V_{i}\right).

(25)

The following lemma is not hard to see from equations (22)–(25).

Lemma 14.

$(V_{0},V_{1},\ldots,V_{h})$ is a partition of $[n]$ , i.e., $\bigcup_{k=0}^{h}\,V_{k}=[n]$ and $V_{i}\cap V_{j}=\emptyset$ for all distinct $i$ , $j\in\{0,1,\ldots,h\}$ .

Let

	$\displaystyle B$	$\displaystyle=$	$\displaystyle\left\{u\in[n]\mid\text{\rm deg}_{H^{(q(n))}}(u)\geq\delta n^{1/(h-1)}-2\right\},$		(26)
	$\displaystyle{\cal E}$	$\displaystyle\stackrel{{\scriptstyle\text{def.}}}{{=}}$	$\displaystyle\left[E_{G}^{(q(n))}\setminus\left(\bigcup_{i,j\in\{0,1,\ldots,h\},\,\|i-j\|\geq 2}\,V_{i}\times V_{j}\right)\right]\cup\left(\left\{\hat{\alpha}\right\}\times\left(V_{h}\setminus\left(B\cup S\right)\right)\right).\,\,\,\,\,\,\,\,\,\,\,$		(27)

By equation (19), $\hat{\alpha}\notin V_{h}\setminus(B\cup S)$ , which together with equation (4) and Lemma 4 forbids any edge in $\cal E$ from being a self-loop. For all distinct $u$ , $v\in[n]$ ,

\displaystyle w\left(u,v\right)\stackrel{{\scriptstyle\text{def.}}}{{=}}\left\{\begin{array}[]{ll}1/2,&\text{if one of $u$ and $v$ is $\hat{\alpha}$ and the other is in $V_{h}\setminus(B\cup S)$,}\\ 1,&\text{otherwise.}\end{array}\right.

(30)

Furthermore, let

\displaystyle{\cal G}

\displaystyle\stackrel{{\scriptstyle\text{def.}}}{{=}}

\displaystyle\left([n],{\cal E},w\right)

(31)

be a weighted undirected graph.

Lemma 15.

\sum_{j=1}^{h-1}\,\left|V_{j}\right|\leq 2\delta^{h-1}n.

Proof.

By Lemma 8 and equation (24),

\left|V_{j+1}\right|\leq\left|V_{j}\right|\cdot\delta n^{1/(h-1)}

for all $j\in[h-2]$ . Therefore, $\sum_{j=1}^{h-1}\,|V_{j}|$ is bounded from above by the $(h-1)$ -term geometric series with the common ratio of $\delta n^{1/(h-1)}$ and the initial value of $|V_{1}|$ . Consequently,

\displaystyle\sum_{j=1}^{h-1}\,\left|V_{j}\right|\stackrel{{\scriptstyle\text{(\ref{tediouscondition2})~{}and~{}Fact~{}\ref{geometricseriesbound}}}}{{\leq}}2\cdot\left|V_{1}\right|\cdot\delta^{h-2}n^{(h-2)/(h-1)}.

(32)

By Lemma 13, $|N_{Q^{(q(n))}}(\hat{\alpha})|\leq\delta n^{1/(h-1)}$ . So by equation (23), we have $|V_{1}|\leq\delta n^{1/(h-1)}$ , which together with inequality (32) completes the proof. ∎

Lemma 16.

\left|V_{h}\setminus\left(B\cup S\right)\right|\geq n\left(1-2\delta^{h-1}-\frac{h}{\delta}\cdot o(1)-\delta\right).

Proof.

By Lemma 12 and equation (26), $|B|=(h/\delta)\cdot o(n)$ . By construction, $|S|=\lfloor\delta n\rfloor$ . Finally,

\displaystyle\left|V_{h}\right|\stackrel{{\scriptstyle\text{Lemmas~{}\ref{disjointnessoflayers}--\ref{sizeofthenonlastlayers}}}}{{\geq}}n-2\delta^{h-1}n-\left|V_{0}\right|\stackrel{{\scriptstyle\text{(\ref{layer0})}}}{{=}}n-2\delta^{h-1}n-1.

∎

The following lemma says that $\hat{\alpha}$ has an average distance of approximately $1/2$ to other points w.r.t. the distance function $\min\{d_{\cal G}(\cdot,\cdot),h\}$ .

Lemma 17.

\displaystyle\sum_{v\in[n]}\,\min\left\{d_{\cal G}\left(\hat{\alpha},v\right),h\right\}\leq n\cdot\left(\frac{1}{2}+2h\delta^{h-1}+\frac{h^{2}}{\delta}\cdot o(1)+h\delta\right).

Proof.

By equations (27)–(31), $d_{\cal G}(\hat{\alpha},v)\leq 1/2$ for all $v\in V_{h}\setminus(B\cup S)$ . This and Lemma 16 complete the proof. ∎

3.3 A metric consistent with Adv’s answers

This subsection constructs a metric $d\colon[n]^{2}\to[0,\infty)$ consistent with Adv’s answers in line 17. So Lemma 10 will require $z$ , which is the output of $A^{\sf Adv}$ , to have an average distance (w.r.t. $d$ ) of at least approximately $h$ to other points. Although $d(\cdot,\cdot)$ will not be exactly $\min\{d_{\cal G}(\cdot,\cdot),h\}$ , Lemma 17 will forbid $\sum_{v\in[n]}\,d(\hat{\alpha},v)/n$ from exceeding $1/2$ by too much. Details follow.

Recall that $H^{(i)}$ and $G^{(i)}$ are unweighted for all $i\in\{0,1,\ldots,q(n)\}$ . They can be treated as having the weight function $w$ while preserving $d_{H^{(i)}}(\cdot,\cdot)$ and $d_{G^{(i)}}(\cdot,\cdot)$ , as shown by the lemma below.

Lemma 18.

For all $i\in\{0,1,\ldots,q(n)\}$ , each path $P$ in $H^{(i)}$ or $G^{(i)}$ has exactly $w(P)$ edges.

Proof.

As $\hat{\alpha}\in S$ by equation (19), equation (30) implies $w(u,v)=1$ for all distinct $u$ , $v\in[n]\setminus S$ . This and equation (4) imply that all edges in $E_{G}^{(0)}$ have weight $1$ w.r.t. $w$ . So by Lemma 4, the edges in $E_{H}^{(i)}\cup E_{G}^{(i)}$ have weight $1$ w.r.t. $w$ . Finally, recall that $H^{(i)}=([n],E_{H}^{(i)})$ and $G^{(i)}=([n],E_{G}^{(i)})$ . ∎

We now show that $H^{(q(n))}$ has an edge in $V_{i}\times V_{j}$ only if $|i-j|\leq 1$ .

Lemma 19.

E_{H}^{(q(n))}\cap\left(\bigcup_{i,j\in\{0,1,\ldots,h\},\,|i-j|\geq 2}\,V_{i}\times V_{j}\right)=\emptyset.

Proof.

Suppose for contradiction that there exists $e\in E_{H}^{(q(n))}$ with an endpoint in $V_{k}$ and the other in $V_{\ell}$ , where $k$ , $\ell\in\{0,1,\ldots,h\}$ and $\ell\geq k+2$ . Then $N_{H^{(q(n))}}(V_{k})\cap V_{\ell}\neq\emptyset$ , which together with Lemma 14 and $\ell\geq k+2$ implies

\displaystyle N_{H^{(q(n))}}\left(V_{k}\right)\not\subseteq\bigcup_{j=0}^{k+1}\,V_{j}.

(33)

As $\ell\geq k+2$ and $k$ , $\ell\in\{0,1,\ldots,h\}$ , we have $0\leq k\leq h-2$ .

Case 1:
$k=0$ . By equations (19) and (22), $V_{0}\subseteq S$ . So $N_{G^{(0)}}(V_{0})=\emptyset$ by equations (4)–(5). Consequently, $N_{H^{(q(n))}}(V_{0})=\emptyset$ by Lemma 4, contradicting relation (33).
Case 2:
$k\in[h-2]$ . Relation (33) contradicts equation (24) (with $j\leftarrow k$ ).

A contradiction occurs in either case. ∎

Lemma 20.

$E_{H}^{(q(n))}\subseteq{\cal E}$ .

Proof.

By Lemma 19 and equation (27), $E_{G}^{(q(n))}\cap E_{H}^{(q(n))}\subseteq{\cal E}$ . This and Lemma 4 complete the proof. ∎

Lemma 21.

Let $P$ be a path in $\cal G$ that visits no edges in $\{\hat{\alpha}\}\times(V_{h}\setminus(B\cup S))$ . If the first and the last vertices of $P$ are in $V_{h}$ and $V_{1}$ , respectively, then $w(P)\geq h-1$ .

Proof.

By Lemma 14, $\bigcup_{k=0}^{h}\,V_{k}=[n]$ , $V_{i+1}\cap V_{i}=\emptyset$ and $(V_{i+1}\times V_{i})\cap(V_{j+1}\times V_{j})=\emptyset$ for all distinct $i$ , $j\in[h-1]$ . Because $P$ is a path in $\cal G$ visiting no edges in $\{\hat{\alpha}\}\times(V_{h}\setminus(B\cup S))$ , no edges on $P$ are in $V_{i}\times V_{j}$ for any $i$ , $j\in\{0,1,\ldots,h\}$ with $|i-j|\geq 2$ by equations (27) and (31). This forces $P$ , which is a $V_{h}$ - $V_{1}$ path, to visit at least one edge in $V_{i+1}\times V_{i}$ for each $i\in[h-1]$ (for a total of at least $h-1$ edges). As $\hat{\alpha}\notin\bigcup_{i=1}^{h}\,V_{i}$ by equations (22)–(25), equation (30) gives $w(u,v)=1$ for all $(u,v)\in\bigcup_{i=1}^{h-1}\,V_{i+1}\times V_{i}$ . We have shown that $P$ has at least $h-1$ edges of weight (w.r.t. $w$ ) $1$ . ∎

We proceed to analyze shortest $a_{i}$ - $b_{i}$ paths in ${\cal G}$ , where $i\in[q(n)]$ . Clearly, such paths must be simple.

Lemma 22.

Let $P$ be a shortest $a_{i}$ - $b_{i}$ path in ${\cal G}$ , where $i\in[q(n)]$ . If $P$ visits exactly one edge in $\{\hat{\alpha}\}\times(V_{h}\setminus(B\cup S))$ and $\hat{\alpha}\in\{a_{i},b_{i}\}$ , then $w(P)\geq h-1/2$ .

Proof.

Being shortest, $P$ must be simple. Assume $\hat{\alpha}=a_{i}$ for now. Because $P$ is a simple $\hat{\alpha}$ - $b_{i}$ path in ${\cal G}$ visiting exactly one edge in $\{\hat{\alpha}\}\times(V_{h}\setminus(B\cup S))$ , it can be decomposed into an edge $(\hat{\alpha},v)$ , where $v\in V_{h}\setminus(B\cup S)$ , and a $v$ - $b_{i}$ path $\tilde{P}$ in ${\cal G}$ that visits no edges in $\{\hat{\alpha}\}\times(V_{h}\setminus(B\cup S))$ .⁶⁶6If the first edge on $P$ is not in $\{\hat{\alpha}\}\times(V_{h}\setminus(B\cup S))$ , then $P$ ’s later visit of an edge in $\{\hat{\alpha}\}\times(V_{h}\setminus(B\cup S))$ must make $P$ non-simple, a contradiction. As $\hat{\alpha}=a_{i}$ , we have $b_{i}\in N_{Q^{(q(n))}}(\hat{\alpha})$ by line 16 of Adv. So by equations (22)–(23), $b_{i}\in V_{1}\cup\{\hat{\alpha}\}$ , implying $b_{i}\in V_{1}$ because querying for the distance from a point to itself is forbidden and $\hat{\alpha}=a_{i}$ . In summary, $\tilde{P}$ is a path in $\cal G$ , from $v\in V_{h}\setminus(B\cup S)$ to $b_{i}\in V_{1}$ , that visits no edges in $\{\hat{\alpha}\}\times(V_{h}\setminus(B\cup S))$ . So by Lemma 21 (with $P\leftarrow\tilde{P}$ ),

\displaystyle w\left(\tilde{P}\right)\geq h-1.

(34)

As $v\in V_{h}$ , we have $\hat{\alpha}\neq v$ by equations (22) and (25). By the construction of $\tilde{P}$ ,

\displaystyle w(P)=w\left(\hat{\alpha},v\right)+w\left(\tilde{P}\right)\stackrel{{\scriptstyle\text{(\ref{newedgeweightfunction})}}}{{\geq}}\frac{1}{2}+w\left(\tilde{P}\right).

(35)

Inequalities (34)–(35) show that $w(P)\geq h-1/2$ . The case of $\hat{\alpha}=b_{i}$ is symmetric: Reverse $P$ and exchange all the above occurrences of “ $a_{i}$ ” with “ $b_{i}$ .” ∎

Lemma 23.

For all $i\in[q(n)]$ with $\hat{\alpha}\in\{a_{i},b_{i}\}$ ,

\chi\left[\exists v\in\left\{a_{i},b_{i}\right\},\,\left(v\in S\right)\land\left(\text{\rm deg}_{Q^{(i)}}(v)\leq\delta n^{1/(h-1)}\right)\right]=1.

Proof.

By equation (19), $\hat{\alpha}\in S$ . This and Lemma 13 complete the proof. ∎

Lemma 24.

For all distinct $u$ , $v\in[n]\setminus(B\cup S)$ , we have $(u,v)\in E_{G}^{(q(n))}$ .

Proof.

As $u$ , $v\in[n]\setminus B$ , equation (26) implies

	$\displaystyle\text{\rm deg}_{H^{(i)}}(u)$	$\displaystyle<$	$\displaystyle\delta n^{1/(h-1)}-2,$		(36)
	$\displaystyle\text{\rm deg}_{H^{(i)}}(v)$	$\displaystyle<$	$\displaystyle\delta n^{1/(h-1)}-2$		(37)

when $i=q(n)$ . So by Lemma 4, inequalities (36)–(37) hold for all $i\in[q(n)]$ .

As $u$ , $v\in[n]\setminus S$ and $u\neq v$ , we have $(u,v)\in E_{G}^{(0)}$ by equation (4). By lines 8 and 13 of Adv,

\displaystyle E_{G}^{(i-1)}\setminus\left\{\left(x,y\right)\in[n]^{2}\mid\left(\text{deg}_{H^{(i)}}(x)\geq\delta n^{1/(h-1)}-2\right)\lor\left(\text{deg}_{H^{(i)}}(y)\geq\delta n^{1/(h-1)}-2\right)\right\}\subseteq E_{G}^{(i)}

(38)

for all $i\in[q(n)]$ . By inequalities (36)–(37) and relation (38), $(u,v)\in E_{G}^{(i)}$ if $(u,v)\in E_{G}^{(i-1)}$ , for all $i\in[q(n)]$ . The proof is complete by mathematical induction. ∎

Lemma 25.

Let $P$ be a shortest $a_{i}$ - $b_{i}$ path in ${\cal G}$ , where $i\in[q(n)]$ . If $P$ visits exactly two edges in $\{\hat{\alpha}\}\times(V_{h}\setminus(B\cup S))$ , then $G^{(q(n))}$ has an $a_{i}$ - $b_{i}$ path with exactly $w(P)$ edges.

Proof.

Being shortest, $P$ must be simple. Therefore, the two edges of $P$ in $\{\hat{\alpha}\}\times(V_{h}\setminus(B\cup S))$ , denoted $(u,\hat{\alpha})$ and $(\hat{\alpha},v)$ , are consecutive on $P$ . Clearly, $u\neq v$ . Replace the subpath $(u,\hat{\alpha},v)$ of $P$ by the edge $(u,v)$ to yield an $a_{i}$ - $b_{i}$ path $\tilde{P}$ . Except for the two edges of $P$ in $\{\tilde{\alpha}\}\times(V_{h}\setminus(B\cup S))$ (which are $(u,\hat{\alpha})$ and $(\hat{\alpha},v)$ ), all edges of $P$ are in $E_{G}^{(q(n))}$ by equation (27) and $P$ ’s being a path in ${\cal G}=([n],{\cal E},w)$ . As $u$ , $v\in V_{h}\setminus(B\cup S)$ and $u\neq v$ , $(u,v)\in E_{G}^{(q(n))}$ by Lemma 24. In summary, all the edges of $\tilde{P}$ (including $(u,v)$ and the edges of $P$ not in $\{\hat{\alpha}\}\times(V_{h}\setminus(B\cup S))$ ) are in $E_{G}^{(q(n))}$ . Consequently, $\tilde{P}$ is an $a_{i}$ - $b_{i}$ path in $G^{(q(n))}=([n],E_{G}^{(q(n))})$ . So we are left only to prove that $\tilde{P}$ has exactly $w(P)$ edges, which, by Lemma 18 (with $P\leftarrow\tilde{P}$ and $i\leftarrow q(n)$ ), is equivalent to proving $w(\tilde{P})=w(P)$ .

Note that $\hat{\alpha}\notin V_{h}\setminus(B\cup S)$ by equation (19). By the construction of $\tilde{P}$ and recalling that $u$ , $v\in V_{h}\setminus(B\cup S)$ and $u\neq v$ ,

w\left(\tilde{P}\right)=w(P)-w\left(u,\hat{\alpha}\right)-w\left(\hat{\alpha},v\right)+w\left(u,v\right)\stackrel{{\scriptstyle\text{(\ref{newedgeweightfunction})}}}{{=}}w(P)-\frac{1}{2}-\frac{1}{2}+1=w(P).

∎

Lemma 26.

Every simple path in $\cal G$ visiting exactly one edge in $\{\hat{\alpha}\}\times(V_{h}\setminus(B\cup S))$ either starts or ends at $\hat{\alpha}$ .

Proof.

By equation (19), $\hat{\alpha}\in S$ . So by equation (4) and Lemma 4, $\hat{\alpha}$ is incident to no edges in $E_{G}^{(q(n))}$ . Consequently, the set of all edges of $\cal G$ incident to $\hat{\alpha}$ is $\{\hat{\alpha}\}\times(V_{h}\setminus(B\cup S))$ by equation (27). The lemma is now easy to see. ∎

Lemma 27.

For all $i\in[q(n)]$ ,

			$\displaystyle\min\left\{d_{H^{(i)}}\left(a_{i},b_{i}\right),h-\frac{1}{2}\cdot\chi\left[\exists v\in\left\{a_{i},b_{i}\right\},\,\left(v\in S\right)\land\left(\text{\rm deg}_{Q^{(i)}}(v)\leq\delta n^{1/(h-1)}\right)\right]\right\}$		(39)
		$\displaystyle\leq$	$\displaystyle\min\left\{d_{\cal G}\left(a_{i},b_{i}\right),h-\frac{1}{2}\cdot\chi\left[\exists v\in\left\{a_{i},b_{i}\right\},\,\left(v\in S\right)\land\left(\text{\rm deg}_{Q^{(i)}}(v)\leq\delta n^{1/(h-1)}\right)\right]\right\}.\,\,\,\,\,\,\,\,\,\,\,\,\,$		(39)

Proof.

Assume the existence of an $a_{i}$ - $b_{i}$ path in ${\cal G}$ for, otherwise, $d_{\cal G}(a_{i},b_{i})=\infty$ and inequality (39) trivially holds. Pick any shortest $a_{i}$ - $b_{i}$ path $P$ in ${\cal G}=([n],{\cal E},w)$ . Clearly,

\displaystyle w(P)=d_{\cal G}\left(a_{i},b_{i}\right).

(40)

Being shortest, $P$ must be simple.

We establish inequality (39) in the following exhaustive cases:

Case 1:

$P$ visits no edges in $\{\hat{\alpha}\}\times(V_{h}\setminus(B\cup S))$ . By equation (27), all edges of $P$ are in $E_{G}^{(q(n))}$ , i.e., $P$ is a path in $G^{(q(n))}$ . So by Lemma 18 (with $i\leftarrow q(n)$ ), $w(P)$ equals the length of $P$ in the unweighted graph $G^{(q(n))}$ . Therefore,

\displaystyle d_{G^{(q(n))}}\left(a_{i},b_{i}\right)\leq w(P).

(41)

If $d_{G^{(i-1)}}(a_{i},b_{i})\leq h$ , then

d_{H^{(i)}}\left(a_{i},b_{i}\right)=d_{G^{(q(n))}}\left(a_{i},b_{i}\right)

by Lemma 5. Otherwise, $d_{G^{(q(n))}}(a_{i},b_{i})>h$ by Lemma 6. In either case, equations (40)–(41) imply inequality (39).

Case 2:
$P$ visits exactly one edge in $\{\hat{\alpha}\}\times(V_{h}\setminus(B\cup S))$ and $\hat{\alpha}\in\{a_{i},b_{i}\}$ . By Lemma 22 and equation (40), $d_{\cal G}(a_{i},b_{i})\geq h-1/2$ . This and Lemma 23 force the right-hand side of inequality (39) to equal $h-1/2$ . By Lemma 23, the left-hand side of inequality (39) is less than or equal to $h-1/2$ . We have verified inequality (39).
Case 3:
$P$ visits exactly one edge in $\{\hat{\alpha}\}\times(V_{h}\setminus(B\cup S))$ and $\hat{\alpha}\notin\{a_{i},b_{i}\}$ . A contradiction to Lemma 26 occurs.
Case 4:
$P$ visits exactly two edges in $\{\hat{\alpha}\}\times(V_{h}\setminus(B\cup S))$ . Lemma 25 and that $G^{(q(n))}$ is unweighted imply inequality (41). Proceeding as in Case 1, equations (40)–(41) and Lemmas 5–6 imply inequality (39) no matter $d_{G^{(i-1)}}(a_{i},b_{i})\leq h$ or otherwise.
Case 5:
$P$ visits at least three edges in $\{\hat{\alpha}\}\times(V_{h}\setminus(B\cup S))$ . Clearly, $P$ is non-simple, a contradiction.

∎

Define $d\colon[n]^{2}\to[0,\infty)$ by

	$\displaystyle d\left(a_{i},b_{i}\right)=d\left(b_{i},a_{i}\right)$
$\displaystyle\stackrel{{\scriptstyle\text{def.}}}{{=}}$	$\displaystyle\min\left\{d_{\cal G}\left(a_{i},b_{i}\right),h-\frac{1}{2}\cdot\chi\left[\exists v\in\left\{a_{i},b_{i}\right\},\,\left(v\in S\right)\land\left(\text{deg}_{Q^{(i)}}(v)\leq\delta n^{1/(h-1)}\right)\right]\right\},\,\,\,\,\,\,\,\,\,\,$
	$\displaystyle d\left(u,v\right)$
$\displaystyle\stackrel{{\scriptstyle\text{def.}}}{{=}}$	$\displaystyle\min\left\{d_{\cal G}\left(u,v\right),h\right\}$	(43)

for all $i\in[q(n)]$ and $(u,v)\in[n]^{2}\setminus\{(a_{j},b_{j})\mid j\in[q(n)]\}$ . Because all pairs in $[n]^{2}$ are unordered in this section, $(b_{i},a_{i})\notin[n]^{2}\setminus\{(a_{j},b_{j})\mid j\in[q(n)]\}$ for all $i\in[q(n)]$ . Consequently, equation (43) does not redefine $d(b_{i},a_{i})$ . Because ${\cal G}$ is undirected, the right-hand side of equation (43) remains intact with $u$ and $v$ interchanged. As $A$ does not repeat queries, equation (3.3) defines $d(a_{i},b_{i})$ and $d(b_{i},a_{i})$ only once for each $i\in[q(n)]$ (note that forbidding repeated queries implies the nonexistence of distinct $i$ , $j\in[q(n)]$ satisfying (1) $a_{i}=a_{j}$ and $b_{i}=b_{j}$ or (2) $a_{i}=b_{j}$ and $b_{i}=a_{j}$ ). It is now clear that $d(\cdot,\cdot)$ is a well-defined function on $[n]^{2}$ , a set of unordered pairs.⁷⁷7Even if we considered each pair in $[n]^{2}$ to be ordered, our arguments would still have shown that $d(\cdot,\cdot)$ is well-defined and symmetric. So we have the following lemma.

Lemma 28.

For all $x$ , $y\in[n]$ , $d(x,y)=d(y,x)$ .

Lemma 29.

For all distinct $x$ , $y\in[n]$ , $d(x,x)=0$ and $d(x,y)\geq 1/2$ .

Proof.

Recall that ${\cal G}=([n],{\cal E},w)$ . As $\mathop{\mathrm{Im}}(w)\subseteq[1/2,\infty)$ by equation (30), we have $d_{\cal G}(x,y)$ , $d_{\cal G}(y,x)\geq 1/2$ . So by equations (3.3)–(43) and $h\in\mathbb{Z}^{+}\setminus\{1\}$ , $d(x,y)\geq 1/2$ . Because we forbid queries for the distance from a point to itself, $d(x,x)$ is not defined by equations (3.3). By equation (43), $d(x,x)=0$ . ∎

Lemma 30.

$([n],d)$ is a metric space.

Proof.

By Lemmas 28–29, we only need to show that

\displaystyle d\left(x,y\right)+d\left(y,z\right)\geq d\left(x,z\right)

(44)

for all $x$ , $y$ , $z\in[n]$ . It is well-known that a positively-weighted undirected graph induces a distance function obeying the triangle inequality; hence

\displaystyle d_{\cal G}\left(x,y\right)+d_{\cal G}\left(y,z\right)\geq d_{\cal G}\left(x,z\right).

(45)

Because $\cal G$ is undirected, $d_{\cal G}(\cdot,\cdot)$ is symmetric. So by equations (3.3)–(43),

\displaystyle d\left(x,y\right)\in\left\{\min\left\{d_{\cal G}\left(x,y\right),h\right\},\min\left\{d_{\cal G}\left(x,y\right),h-\frac{1}{2}\right\}\right\}

(46)

for all $x$ , $y\in[n]$ . Now verify inequality (44) in the following exhaustive (but not mutually exclusive) cases:

Case 1:
$x=y$ , $y=z$ or $x=z$ . Lemma 29 implies inequality (44).
Case 2:
$d_{\cal G}(x,y)\geq h-1/2$ and $y\neq z$ . By relation (46), $d(x,y)\geq h-1/2$ . As $y\neq z$ , $d(y,z)\geq 1/2$ by Lemma 29. By relation (46), $d(x,z)\leq h$ . Summarizing the above proves inequality (44).
Case 3:
$d_{\cal G}(y,z)\geq h-1/2$ and $x\neq y$ . Replace “ $(x,y)$ ,” “ $(y,z)$ ” and “ $y\neq z$ ” in the analysis of Case 2 by “ $(y,z)$ ,” “ $(x,y)$ ” and “ $x\neq y$ ,” respectively.
Case 4:
$d_{\cal G}(x,y)<h-1/2$ and $d_{\cal G}(y,z)<h-1/2$ . By relation (46), $d(x,y)=d_{\cal G}(x,y)$ and $d(y,z)=d_{\cal G}(y,z)$ . So inequalities (44)–(45) share a common left-hand side. To deduce inequality (44) from inequality (45), therefore, it suffices to show that $d_{\cal G}(x,z)\geq d(x,z)$ , which follows from relation (46).

∎

Lemma 31.

For all $i\in[q(n)]$ ,

d_{H^{(i)}}\left(a_{i},b_{i}\right)\geq d_{\cal G}\left(a_{i},b_{i}\right).

Proof.

Assume the existence of an $a_{i}$ - $b_{i}$ path in $H^{(i)}$ for, otherwise, $d_{H^{(i)}}(a_{i},b_{i})=\infty$ and there is nothing to prove. Take a shortest $a_{i}$ - $b_{i}$ path $P$ in the unweighted graph $H^{(i)}=([n],E_{H}^{(i)})$ . So $d_{H^{(i)}}(a_{i},b_{i})$ is the number of $P$ ’s edges. By Lemma 18, $P$ ’s number of edges equals $w(P)$ . By Lemma 4, $P$ ’s edges are in $E_{H}^{(q(n))}$ . So by Lemma 20, $P$ is a path in ${\cal G}=([n],{\cal E},w)$ , implying $d_{\cal G}(a_{i},b_{i})\leq w(P)$ . Summarizing the above proves the lemma. ∎

The following lemma says that line 17 of Adv answers queries consistently with $d(\cdot,\cdot)$ .

Lemma 32.

For all $i\in[q(n)]$ ,

			$\displaystyle\min\left\{d_{H^{(i)}}\left(a_{i},b_{i}\right),h-\frac{1}{2}\cdot\chi\left[\exists v\in\left\{a_{i},b_{i}\right\},\,\left(v\in S\right)\land\left(\text{\rm deg}_{Q^{(i)}}(v)\leq\delta n^{1/(h-1)}\right)\right]\right\}$		(47)
		$\displaystyle=$	$\displaystyle d\left(a_{i},b_{i}\right).$		(47)

Proof.

Lemma 27 and equation (3.3) prove the “ $\leq$ ” part of equation (47). On the other hand, Lemma 31 and equation (3.3) imply the “ $\geq$ ” part of equation (47). ∎

3.4 Putting things together

We now arrive at our main result.

Theorem 33.

Metric $1$ -median has no deterministic $o(n^{1+1/(h-1)})$ -query $(2h-\epsilon)$ -approximation algorithms for any constants $h\in\mathbb{Z}^{+}\setminus\{1\}$ and $\epsilon>0$ .

Proof.

By Lemma 32 and line 17 of Adv, Adv answers $A$ ’s queries consistently with $d(\cdot,\cdot)$ . This implies that $A^{\text{\sf Adv}}$ and $A^{d}$ have the same output.⁸⁸8See, e.g., [2, Lemma 8]. That is, $A^{d}$ outputs $z$ . By Lemma 30, $([n],d)$ is a metric space.

By relation (46), $d(x,y)\leq\min\{d_{\cal G}(x,y),h\}$ for all $x$ , $y\in[n]$ . Therefore,

\displaystyle\sum_{v\in[n]}\,d\left(\hat{\alpha},v\right)\leq n\cdot\left(\frac{1}{2}+2h\delta^{h-1}+\frac{h^{2}}{\delta}\cdot o(1)+h\delta\right)

(48)

by Lemma 17.

Recall that $A$ does not repeat queries. So by equation (15) and Lemmas 28–29,

\displaystyle\sum_{v\in[n]}\,d\left(z,v\right)\geq\sum_{i\in I}\,d\left(a_{i},b_{i}\right).

⁹⁹9In fact, this is an equality because

A^{\sf Adv}

will have queried for the distances between its output and all other points when halting.

By Lemmas 10 and 32,

\displaystyle\sum_{i\in I}\,d\left(a_{i},b_{i}\right)\geq n\cdot\left(h-2h\delta^{h-1}-o(1)-\delta\right).

(49)

By inequalities (48)–(49),

\displaystyle\frac{\sum_{v\in[n]}\,d\left(z,v\right)}{\sum_{v\in[n]}\,d\left(\hat{\alpha},v\right)}\geq\frac{h-2h\delta^{h-1}-o(1)-\delta}{1/2+2h\delta^{h-1}+(h^{2}/\delta)\cdot o(1)+h\delta}.

(50)

Note that all the derivations so far have been valid for all constants $h\in\mathbb{Z}^{+}\setminus\{1\}$ and $\delta\in(0,1)$ . Take $\delta=\delta(h,\epsilon)>0$ to be sufficiently small and $n$ to be sufficiently large so that the right-hand side of inequality (50) is greater than $2h-\epsilon$ .¹⁰¹⁰10Alternatively, we may take $\delta=\delta(n)=\left(\frac{\max\{q(n),n\}}{n^{1+1/(h-1)}}\right)^{1/3}$ from the beginning of this section. Then, as $q(n)=o(n^{1+1/(h-1)})$ , the right-hand side of inequality (50) is $2h-o(1)$ , and inequalities (1)–(3) remain true for all sufficiently large $n$ . Then inequality (50) forbids $z$ , which is the common output of $A^{\sf Adv}$ and $A^{d}$ , from being a $(2h-\epsilon)$ -approximate $1$ -median of $([n],d)$ . Note that $A$ can be any deterministic $o(n^{1+1/(h-1)})$ -query algorithm from the beginning of this section. ∎

Next, we use Theorem 33 and Fact 1 to determine the minimum value of $c\geq 1$ such that metric $1$ -median has a deterministic $O(n^{1+\epsilon})$ -query (resp., $O(n^{1+\epsilon})$ -time) $c$ -approximation algorithm, for each constant $\epsilon\in(0,1)$ .

Theorem 34.

For each constant $\epsilon\in(0,1)$ ,

			$\displaystyle\min\left\{c\geq 1\mid\text{{\sc metric $1$-median} has a deterministic $O(n^{1+\epsilon})$-query $c$-approx.\ alg.}\right\}$
		$\displaystyle=$	$\displaystyle\min\left\{c\geq 1\mid\text{{\sc metric $1$-median} has a deterministic $O(n^{1+\epsilon})$-time $c$-approx.\ alg.}\right\}$
		$\displaystyle=$	$\displaystyle 2\left\lceil\frac{1}{\epsilon}\right\rceil.$

Proof.

Take $h=\lceil 1/\epsilon\rceil$ ; hence $h\in\mathbb{Z}^{+}\setminus\{1\}$ . It is easy to verify that $n^{1+\epsilon}=o(n^{1+1/(h-1)})$ . So by Theorem 33, metric $1$ -median does not have a deterministic $O(n^{1+\epsilon})$ -query $(2\lceil 1/\epsilon\rceil-\epsilon^{\prime})$ -approximation algorithm for any constant $\epsilon^{\prime}>0$ .

Clearly, $n^{1+1/h}=O(n^{1+\epsilon})$ . So by Fact 1, metric $1$ -median has a deterministic $O(n^{1+\epsilon})$ -time $(2\lceil 1/\epsilon\rceil)$ -approximation algorithm.

The above analyses remain valid with “query” and “time” exchanged because every $O(n^{1+\epsilon})$ -time algorithm makes $O(n^{1+\epsilon})$ queries. Consequently, deterministic $O(n^{1+\epsilon})$ -query (resp., $O(n^{1+\epsilon})$ -time) algorithms can be $(2\lceil 1/\epsilon\rceil)$ -approximate but not $(2\lceil 1/\epsilon\rceil-\epsilon^{\prime})$ -approximate for any constant $\epsilon^{\prime}>0$ . ∎

The brute-force exact algorithm for metric $1$ -median is well-known to run in $O(n^{2})$ time. Therefore, there is no need to extend Theorem 34 to the case of $\epsilon\geq 1$ . On the other hand, the following corollary deals with the case of $\epsilon=0$ .

Corollary 35.

Metric $1$ -median does not have a deterministic $O(n^{1+o(1)})$ -query (resp., $O(n^{1+o(1)})$ -time) $O(1)$ -approximation algorithm.

Proof.

Take $h\to\infty$ in Theorem 33. ∎

Acknowledgments

The author is supported in part by the Ministry of Science and Technology of Taiwan under grant 103-2221-E-155-026-MY2.

Appendix A Optimizing the hidden factors in Theorem 33

This appendix discusses how the bound of $o(n^{1+1/(h-1)})$ in Theorem 33 hides factors dependent on $h$ . For all $i\in[q(n)]$ ,

\displaystyle B_{i-1}\stackrel{{\scriptstyle\text{def.}}}{{=}}\left\{v\in[n]\mid\text{deg}_{H^{(i-1)}}(v)\geq\delta n^{1/(h-1)}-2\right\}.

(51)

Lemma 36.

For all $i\in[q(n)]$ and distinct $u$ , $v\in[n]\setminus(B_{i-1}\cup S)$ , we have $(u,v)\in E_{G}^{(i-1)}$ .

Proof.

As $u$ , $v\in[n]\setminus B_{i-1}$ ,

	$\displaystyle\text{deg}_{H^{(j)}}(u)$	$\displaystyle<$	$\displaystyle\delta n^{1/(h-1)}-2,$
	$\displaystyle\text{deg}_{H^{(j)}}(v)$	$\displaystyle<$	$\displaystyle\delta n^{1/(h-1)}-2$

for all $j\in\{0,1,\ldots,i-1\}$ by equation (51) and Lemma 4. So by lines 8 and 13 of Adv, $(u,v)\in E_{G}^{(j)}$ if $(u,v)\in E_{G}^{(j-1)}$ , for all $j\in[i-1]$ . By equation (4), $(u,v)\in E_{G}^{(0)}$ . The proof is complete by mathematical induction. ∎

Lemma 37.

For each $i\in[q(n)]$ such that the $i$ th iteration of the loop of Adv runs lines 5–9, $P_{i}$ in line 5 does not have two non-consecutive vertices in $[n]\setminus(B_{i-1}\cup S)$ .

Proof.

By line 5 of Adv, two non-consecutive vertices on $P_{i}$ are not connected by an edge in $E_{G}^{(i-1)}$ . This and Lemma 36 complete the proof. ∎

Lemma 38.

For all $i\in[q(n)]$ and $v\in B_{i-1}$ ,

N_{G^{(i-1)}}(v)\subseteq N_{H^{(i-1)}}(v).

Proof.

By equation (51),

\text{deg}_{H^{(i-1)}}(v)\geq\delta n^{1/(h-1)}-2.

Clearly,

\text{\rm deg}_{H^{(0)}}(v)\stackrel{{\scriptstyle\text{(\ref{initiallymarkededgeset})}}}{{=}}0\stackrel{{\scriptstyle\text{(\ref{tediouscondition2})}}}{{<}}\delta n^{1/(h-1)}-2.

So there exists $j\in[i-1]$ satisfying

	$\displaystyle\text{deg}_{H^{(j-1)}}(v)$	$\displaystyle<$	$\displaystyle\delta n^{1/(h-1)}-2,$		(52)
	$\displaystyle\text{deg}_{H^{(j)}}(v)$	$\displaystyle\geq$	$\displaystyle\delta n^{1/(h-1)}-2.$		(53)

Clearly,

\displaystyle N_{G^{(j)}}(v)=\left\{u\in[n]\mid(u,v)\in E_{G}^{(j)}\right\}.

(54)

As $H^{(j-1)}\neq H^{(j)}$ by inequalities (52)–(53), the $j$ th iteration of the loop of Adv runs lines 5–9 but not 11–14. By inequality (53) and line 8 of Adv,

\displaystyle\left\{u\in[n]\mid(u,v)\in E_{G}^{(j)}\right\}=\left\{u\in[n]\mid(u,v)\in E_{G}^{(j-1)}\setminus\left(E_{G}^{(j-1)}\setminus E_{H}^{(j)}\right)\right\}.

(55)

Equations (54)–(55) and Lemma 4 give

N_{G^{(j)}}(v)=N_{H^{(j)}}(v).

This and Lemma 4 complete the proof. ∎

Lemma 39.

For all $i\in[q(n)]$ ,

\displaystyle\left|E_{H}^{(i)}\right|\leq\left|E_{H}^{(i-1)}\right|+1.

Proof.

Clearly, we may assume that the $i$ th iteration of the loop of Adv runs lines 5–9 but not 11–14. By line 6, we only need to show that

\displaystyle\left|\left\{e\mid\left(\text{$e$ is an edge on $P_{i}$}\right)\land\left(e\notin E_{H}^{(i-1)}\right)\right\}\right|\leq 1.

(56)

By Lemma 37, $P_{i}$ in line 5 has at most one edge in $([n]\setminus(B_{i-1}\cup S))^{2}$ . So, to prove inequality (56), it suffices to show that each edge $(u,v)$ on $P_{i}$ with $(u,v)\notin([n]\setminus(B_{i-1}\cup S))^{2}$ satisfies $(u,v)\in E_{H}^{(i-1)}$ , as done below:

Case 1:
$\{u,v\}\cap S\neq\emptyset$ . By equation (4) and Lemma 4, $(u,v)\notin E_{G}^{(i-1)}$ . Consequently, $P_{i}$ has an edge not in $E_{G}^{(i-1)}$ , contradicting line 5.
Case 2:
$\{u,v\}\cap B_{i-1}\neq\emptyset$ . By symmetry, assume $v\in B_{i-1}$ . So by Lemma 38, $N_{G^{(i-1)}}(v)\subseteq N_{H^{(i-1)}}(v)$ . Because $P_{i}$ is a path in $G^{(i-1)}$ by line 5 and $(u,v)$ is on $P_{i}$ , $u\in N_{G^{(i-1)}}(v)$ . In summary, $u\in N_{H^{(i-1)}}(v)$ . I.e., $(u,v)\in E_{H}^{(i-1)}$ .

∎

The following improvement over Lemma 11 is immediate from equation (6) and Lemma 39.

Lemma 40.

\left|E_{H}^{(q(n))}\right|\leq q(n).

Assuming $100\leq h=o(n^{1/(h-1)})$ , the following modifications to this paper show that the bound of $o(n^{1+1/(h-1)})$ in Theorem 33 depends on $h$ as $o(n^{1+1/(h-1)}/h)$ :

(1)

Take

$\displaystyle q(n)$	$\displaystyle=$	$\displaystyle o\left(\frac{n^{1+1/(h-1)}}{h}\right),$
$\displaystyle\delta$	$\displaystyle=$	$\displaystyle h\cdot\frac{\max\{q(n),n\}}{n^{1+1/(h-1)}},$
$\displaystyle\lambda$	$\displaystyle=$	$\displaystyle\delta^{h/8},$
$\displaystyle S$	$\displaystyle=$	$\displaystyle[\lfloor\lambda n\rfloor].$

(2)
Replace “ $\delta$ ” by “ $\sqrt{\delta}$ ” in inequality (2).
(3)
Replace “ $\delta$ ” by $1/\delta^{h/4}$ in inequality (3).
(4)
Replace the two occurrences of “ $\delta$ ” by “ $\sqrt{\delta}$ ” in line 8 of Adv.
(5)
Replace “ $\delta$ ” by “ $1/\delta^{h/4}$ ” in line 17 of Adv.
(6)
Replace all occurrences of “ $\delta$ ” by “ $\sqrt{\delta}$ ” in Lemma 8 and its proof.
(7)
Replace all occurrences of “ $\delta$ ” by “ $\sqrt{\delta}$ ” in Lemma 9 and its proof.
(8)
Replace “ $\delta n^{1/(h-1)}$ ” and “ $h-2h\delta^{h-1}-o(1)-\delta$ ” by “ $n^{1/(h-1)}/\delta^{h/4}$ ” and “ $h-2h{\sqrt{\delta}}^{h-1}-o(1)-\lambda/2-1/(2\delta^{h/4}n^{1-1/(h-1)})$ ,” respectively, in the statement of Lemma 10.
(9)
Replace all occurrences of “ $\delta n^{1/(h-1)}$ ,” “ $2\delta^{h-1}n$ ” and “ $\lfloor\delta n\rfloor$ ” by “ $n^{1/(h-1)}/\delta^{h/4}$ ,” “ $2{\sqrt{\delta}}^{h-1}n$ ” and “ $\lfloor\lambda n\rfloor$ ,” respectively, in the proof of Lemma 10.
(10)
Replace all occurrences of “ $\delta n^{1/(h-1)}$ ,” “ $(h/\delta)\cdot o(n)$ ” and “Lemma 11” by “ $\sqrt{\delta}\,n^{1/(h-1)}$ ,” “ $(1/\sqrt{\delta})\cdot O(q(n)/n^{1/(h-1)})$ ” and “Lemma 40,” respectively, in Lemma 12 and its proof.
(11)
That $\hat{\alpha}$ is well-defined in equation (19) follows from $|S|\geq 2$ , which holds for all sufficiently large $n$ by item (1) and $h\geq 100$ .
(12)
Replace all occurrences of “ $\delta$ ” by “ $1/\delta^{h/4}$ ” in Lemma 13 and its proof.
(13)
Replace “ $\delta$ ” by “ $\sqrt{\delta}$ ” in equation (26).
(14)
Replace “ $\delta^{h-1}$ ” by “ $\delta^{h/4-1}$ ” in the statement of Lemma 15.
(15)
Replace all occurrences of “ $\delta$ ” by “ $\sqrt{\delta}$ ” and “ $1/\delta^{h/4}$ ,” respectively, in the first and the second paragraphs of the proof of Lemma 15.
(16)
Replace “ $1-2\delta^{h-1}-(h/\delta)\cdot o(1)-\delta$ ” by “ $1-2\delta^{h/4-1}-(1/\sqrt{\delta})\cdot O(q(n)/n^{1+1/(h-1)})-\lambda$ ” in the statement of Lemma 16.
(17)
Replace all occurrences of “ $(h/\delta)\cdot o(n)$ ,” “ $\lfloor\delta n\rfloor$ ” and “ $\delta^{h-1}$ ” by “ $(1/\sqrt{\delta})\cdot O(q(n)/n^{1/(h-1)})$ ,” “ $\lfloor\lambda n\rfloor$ ” and “ $\delta^{h/4-1}$ ,” respectively, in the proof of Lemma 16.
(18)
Replace “ $\delta^{h-1}$ ,” “ $(h^{2}/\delta)\cdot o(1)$ ” and “ $h\delta$ ” by “ $\delta^{h/4-1}$ ,” “ $(h/\sqrt{\delta})\cdot O(q(n)/n^{1+1/(h-1)})$ ” and “ $h\lambda$ ,” respectively, in the statement of Lemma 17.
(19)
Replace “ $\delta$ ” by “ $1/\delta^{h/4}$ ” in the statement of Lemma 23.
(20)
Replace all occurrences of “ $\delta$ ” by “ $\sqrt{\delta}$ ” in the proof of Lemma 24.
(21)
Replace the two occurrences of “ $\delta$ ” by “ $1/\delta^{h/4}$ ” in the statement of Lemma 27.
(22)
Replace “ $\delta$ ” by “ $1/\delta^{h/4}$ ” in equation (3.3).
(23)
Replace “ $\delta$ ” by “ $1/\delta^{h/4}$ ” in the statement of Lemma 32.
(24)
Replace “ $\delta^{h-1}$ ,” “ $(h^{2}/\delta)\cdot o(1)$ ” and “ $h\delta$ ” by “ $\delta^{h/4-1}$ ,” “ $(h/\sqrt{\delta})\cdot O(q(n)/n^{1+1/(h-1)})$ ” and “ $h\lambda$ ,” respectively, in inequality (48).
(25)
Replace “ $h-2h\delta^{h-1}-o(1)-\delta$ ” by “ $h-2h{\sqrt{\delta}}^{h-1}-o(1)-\lambda/2-1/(2\delta^{h/4}n^{1-1/(h-1)})$ ” in the right-hand side of inequality (49).
(26)
Replace the numerator and the denominator on the right-hand side of inequality (50) by “ $h-2h{\sqrt{\delta}}^{h-1}-o(1)-\lambda/2-1/(2\delta^{h/4}n^{1-1/(h-1)})$ ” and “ $1/2+2h\delta^{h/4-1}+(h/\sqrt{\delta})\cdot O(q(n)/n^{1+1/(h-1)})+h\lambda$ ,” respectively.
(27)
Verify that the right-hand side of inequality (50) is $2h-o(1)$ . To see this, use item (1) and $100\leq h=o(n^{1/(h-1)})$ to verify that $\delta=o(1)$ , $\max_{x\geq 1}\,x\cdot\delta^{x/8}=O(\delta)=o(1)$ (which requires elementary calculus and reveals that $h\sqrt{\delta}^{h-1}=o(1)$ , $h\delta^{h/4-1}=o(1)$ and $h\lambda=h\delta^{h/8}=o(1)$ ), $\lambda=o(1)$ , $\delta^{h/4}\geq 1/n^{h/(4(h-1))}$ , $\delta^{h/4}\cdot n^{1-1/(h-1)}=n^{\Omega(1)}$ , $\sqrt{\delta}\geq\sqrt{h\cdot q(n)/n^{1+1/(h-1)}}$ and $\sqrt{h\cdot q(n)/n^{1+1/(h-1)}}=o(1)$ .
(28)
Replace all occurrences of “ $\delta$ ” by “ $\sqrt{\delta}$ ” in equation (51) as well as in the proofs of Lemmas 36 and 38.

References

[1] C.-L. Chang. A deterministic sublinear-time nonadaptive algorithm for metric $1$ -median selection. To appear in Theoretical Computer Science.
[2] C.-L. Chang. Some results on approximate $1$ -median selection in metric spaces. Theoretical Computer Science, 426:1–12, 2012.
[3] C.-L. Chang. Deterministic sublinear-time approximations for metric $1$ -median selection. Information Processing Letters, 113(8):288–292, 2013.
[4] C.-L. Chang. A lower bound for metric $1$ -median selection. Technical Report arXiv: 1401.2195, 2014.
[5] S. Guha, A. Meyerson, N. Mishra, R. Motwani, and L. O’Callaghan. Clustering data streams: Theory and practice. IEEE Transactions on Knowledge and Data Engineering, 15(3):515–528, 2003.
[6] P. Indyk. Sublinear time algorithms for metric space problems. In Proceedings of the 31st Annual ACM Symposium on Theory of Computing, pages 428–434, 1999.
[7] P. Indyk. High-Dimensional Computational Geometry. PhD thesis, Stanford University, 2000.
[8] A. Kumar, Y. Sabharwal, and S. Sen. Linear-time approximation schemes for clustering problems in any dimensions. Journal of the ACM, 57(2):5, 2010.
[9] R. R. Mettu and C. G. Plaxton. Optimal time bounds for approximate clustering. Machine Learning, 56(1–3):35–60, 2004.
[10] W. Rudin. Principles of Mathematical Analysis. McGraw-Hill, 3rd edition, 1976.
[11] B.-Y. Wu. On approximating metric $1$ -median in sublinear time. Information Processing Letters, 114(4):163–166, 2014.

Metric 111-median selection: Query complexity vs. approximation ratio

Abstract

1 Introduction

2 Definitions

Fact 1 ([3, 1, 11]).

Fact 2.

Fact 3.

3 Query complexity vs. approximation ratio

Lemma 4.

Proof.

Lemma 5.

Proof.

Lemma 6.

3.1 The average distance from A𝐴A’s output to other points

Lemma 7.

Proof.

Lemma 8.

Proof.

Lemma 9.

Proof.

Lemma 10.

Proof.

3.2 Planting a point with a small average distance to other points

Lemma 11.

Proof.

Lemma 12.

Proof.

Lemma 13.

Proof.

Lemma 14.

Lemma 15.

Proof.

Lemma 16.

Proof.

Lemma 17.

Proof.

3.3 A metric consistent with Adv’s answers

Lemma 18.

Proof.

Lemma 19.

Proof.

Lemma 20.

Proof.

Lemma 21.

Proof.

Lemma 22.

Proof.

Lemma 23.

Proof.

Lemma 24.

Proof.

Lemma 25.

Proof.

Lemma 26.

Proof.

Lemma 27.

Proof.

Lemma 28.

Lemma 29.

Proof.

Lemma 30.

Proof.

Lemma 31.

Proof.

Lemma 32.

Proof.

3.4 Putting things together

Theorem 33.

Proof.

Theorem 34.

Proof.

Corollary 35.

Proof.

Acknowledgments

Appendix A Optimizing the hidden factors in Theorem 33

Lemma 36.

Proof.

Lemma 37.

Proof.

Lemma 38.

Metric $1$ -median selection: Query complexity vs. approximation ratio

3.1 The average distance from $A$ ’s output to other points