\documentclass[11pt]{article}
\usepackage{fullpage}
\usepackage{amsmath, amsfonts}
\usepackage{clrscode}
\begin{document}
\begin{center} \Large
{\scshape 6.897 Advanced Data Structures (Spring'05)} \\[1ex]
{Prof.~Erik Demaine \quad\quad TA: Mihai P\v{a}tra\c{s}cu} \\[2ex]
\framebox{Problem 8 -- Solution}
\end{center}
\paragraph{From RMQ to 3-sided queries.}
The relation is immediate. For each $y \in \{1, \dots, n\}$, let $A[y]
= \min \{ x \mid (x,y) \in S \}$. When we want to query the range
$[0,b] \times [c,d]$, we use an RMQ query to find $t = \min \{ A[y]
\mid y \in [c,d] \}$. If $t > b$, the range is empty; otherwise, it
contains at least one point.
\paragraph{From 3-sided to general queries.}
Without loss of generality, $n$ is a power of two. Consider a perfect
binary tree over the $x$-coordinate. Each node represents a vertical
strip of space; say this is $[a,b] \times [1,n]$. For each such strip
(except the root, which gives the entire space), we build a structure
for 3-sided queries. If the node is the right child of its parent,
this structure assumes the left side of the rectangle is on the $a$
abscissa. If the node is a left child, it assume the right side in
fixed to $b$. Even though we have described 3-sided queries as fixing
one side to 0, we can actually fix it to anything, by horizontal
translation, and possibly a reflection.
For every 3-sided structure that we build, we must also include a
predecessor structure. This is needed because the 3-sided structure
has less than $n$ points. Thus, we must convert from the original rank
space to this new (sparser) rank space. This can be done by building a
predecessor structure on the set of $y$-coordinates of the points in
the 3-sided structure. We then run two predecessor queries to convert
the $y$ boundaries of the rectangle to the new rank space. We
implement these predecessor structures using $y$-fast trees. Note that
the universe is $\{1, \dots, n\}$, so a query takes $O(\lg\lg n)$
time.
The space for each of the 3-sided structures is $\sigma$ times the
number of points in the structure. Every point appear in exactly $\lg
n$ structures (one for each ancestor in the tree of its $x$
coordinate). Thus, the total space is $O(n\lg n \cdot \sigma)$. The
predecessor structures take space linear in the number of points of
the $3$-sided structures, so they form an additional constant factor.
Now assume we want to query the range $[a,b] \times [c,d]$. We find
the lowest common ancestor (call it $v$) of $a$ and $b$. This can be
done in constant time: we take the xor of $a$ and $b$ and find the
most significant set bit (which was in our standard set of bit
tricks). Now, the left child of $v$ contains $a$; let $m$ be the
rightmost abscissa under this node. The right child of $v$ contains
$b$; the leftmost abscissa under it is $m+1$. We have broken our range
query into two queries: $[a,m] \times [c,d]$ and $[m+1,b] \times
[c,d]$. Since we are doing existential queries, we take the or of the
two answers. Both of these queries are 3-sided queries, in the left
and right children of $v$, respectively. We use the predecessor
structures to convert $c$ and $d$ into the rank space of the 3-sided
structures, and then run the 3-sided queries. Thus, our running time
is $2\tau + O(\lg\lg n)$.
\end{document}