KdTree

Next Assignment due 3/26 (Tuesday after Spring Break)

This Thursday, precept will cover the assignment in detail using a great
worksheet (thanks Maia!).
Due two days after Spring break ends.
Second hardest assignment (after Collinear).
– Not bad as long as you fully understand the methods.
Before starting, fully comprehend:
– The precept worksheet.
– The 2d tree part of this lecture.
Thursday 3/14 - Sunday 3/24 no lab TAs
– They come back on Monday 3/25

1
Algorithms R OBERT S EDGEWICK | K EVIN W AYNE

G EOMETRIC A PPLICATIONS OF BST S

‣ 1d range search
‣ kd trees
‣ line segment intersection
Algorithms
‣ interval search trees
F O U R T H E D I T I O N

R OBERT S EDGEWICK | K EVIN W AYNE

‣ rectangle intersection
http://algs4.cs.princeton.edu
Overview

This lecture. Intersections among geometric objects.

2d orthogonal range search orthogonal rectangle intersection

Applications. CAD, games, movies, virtual reality, databases, GIS, .…

Efficient solutions. Binary search trees (and extensions).

3
 G EOMETRIC A PPLICATIONS OF BST S

‣ 1d range search
‣ kd trees
‣ interval search trees
Algorithms
‣ line segment intersection

R OBERT S EDGEWICK | K EVIN W AYNE

‣ rectangle intersection
http://algs4.cs.princeton.edu
1d range search

Extension of ordered symbol table.

Insert key-value pair.
Search for key k.
Delete key k.
Range search: find all keys between k1 and k2.
Range count: number of keys between k1 and k2.

Application. Database queries.

insert B B
insert D B D
insert A A B D
Geometric interpretation. insert I A B D I
Keys are point on a line. insert H A B D H I

Find/count points in a given 1d interval. insert F A B D F H I

insert P A B D F H I P
count G to K 2
search G to K H I

5
1d range search: elementary implementations

Unordered list. Fast insert, slow range search.

Ordered array. Slow insert, binary search for k1 and k2 to do range search.

order of growth of running time for 1d range search

data structure insert range count range search

unordered list 1 N N

ordered array N log N R + log N

goal log N log N R + log N

N = number of keys
R = number of keys that match

6
1d range count: BST implementation

1d range count. How many keys between lo and hi ? number of keys < E

rank(E) = 2
rank 6
rank(F) = 3
2 S 7 rank(S) = 6
1 E 5 X rank(T) = 7
A 0 3 R size(E, S) = 5
C H 4 size(E, T) = 5
M size(F, T) = 4

public int size(Key lo, Key hi)

{
if (contains(hi)) return rank(hi) - rank(lo) + 1;
else return rank(hi) - rank(lo);
} number of keys < hi

Proposition. Running time proportional to log N.

Pf. Nodes examined = search path to lo + search path to hi.
7
1d range search: BST implementation

1d range search. Find all keys between lo and hi.

Recursively find all keys in left subtree (if any could fall in range).
Check key in current node.
Recursively find all keys in right subtree (if any could fall in range).

searching in the range [F..T]

red keys are used in compares

but are not in the range
S
E X
A R
C H
M
L P black keys are
in the range

Range search in a BST

Proposition. Running time proportional to R + log N.
Pf. Nodes examined = search path to lo + search path to hi + matches.
8
 G EOMETRIC A PPLICATIONS OF BST S

‣ 1d range search
‣ kd trees
‣ interval search trees
Algorithms
‣ line segment intersection

R OBERT S EDGEWICK | K EVIN W AYNE

‣ rectangle intersection
http://algs4.cs.princeton.edu
2-d orthogonal range search

Extension of ordered symbol-table to 2d keys.

Insert a 2d key.
Delete a 2d key.
Search for a 2d key.
Range search: find all keys that lie in a 2d range.
Range count: number of keys that lie in a 2d range.

Applications. Networking, circuit design, databases, ...

Geometric interpretation.
Keys are point in the plane.
Find/count points in a given h-v rectangle

rectangle is axis-aligned

10
Binary search tree?

Tree construction
What order to store points?
– X coordinate determines order?
– Y coordinate determines order?

E R

11
2d orthogonal range search: grid implementation

Grid implementation.
Divide space into M-by-M grid of squares.
Create list of points contained in each square.
Use 2d array to directly index relevant square.
Insert: add (x, y) to list for corresponding square.
Range search: examine only squares that intersect 2d range query.

Old Algorithm: Look at all points, order of growth N

RT
New algorithm: Look at all points, order of growth ?

LB

12
2d orthogonal range search: grid implementation analysis

Space-time tradeoff.
Space: M 2 + N.
Time: 1 + N / M 2 per square examined, on average.

Choose grid square size to tune performance.

Too small: wastes space.
Too large: too many points per square.
Rule of thumb: √N-by-√N grid.

Running time. [if points are evenly distributed] RT

Initialize data structure: N.

choose M ~ √N
Insert point: 1.
Range search: 1 per point in range.

LB

13
Clustering

Grid implementation. Fast, simple solution for evenly-distributed points.

Problem. Clustering a well-known phenomenon in geometric data.

Lists are too long, even though average length is short.
Need data structure that adapts gracefully to data.

14
Clustering

Grid implementation. Fast, simple solution for evenly-distributed points.

Problem. Clustering a well-known phenomenon in geometric data.

Ex. USA map data.

13,000 points, 1000 grid squares

half the squares are empty half the points are

in 10% of the squares 15
Space-partitioning trees

Use a tree to represent a recursive subdivision of 2d space.

Grid. Divide space uniformly into squares.

2d tree. Recursively divide space into two halfplanes.
Quadtree. Recursively divide space into four quadrants.
BSP tree. Recursively divide space into two regions.

Grid 2d tree Quadtree BSP tree

16
Space-partitioning trees: applications

Applications.
Ray tracing.
2d range search.
Flight simulators.
N-body simulation.
Collision detection.
Astronomical databases.
Nearest neighbor search.
Adaptive mesh generation.
Accelerate rendering in Doom.
Hidden surface removal and shadow casting.

Grid 2d tree Quadtree BSP tree

17
2d tree construction

Recursively partition plane into two halfplanes.

8
1

6 3 2
9

4 6 7 8

1
5 5 10 9
2

10

18
Your turn

Draw the KdTree and matching graph when the following points are inserted:
(2,3) (4, 2) (4, 4) (3, 3) (1, 5)

(2, 3)

19
2d tree implementation

Data structure. BST, but alternate using x- and y-coordinates as key.

Search gives rectangle containing point.
Insert further subdivides the plane.

q
p

q
p
points points points points
left of p right of p below q above q
even levels odd levels

6 1
9
3 3 2

5 1
2
4 6 7 8
7
5 10 9
10

20
2d tree construction

Recursively partition plane into two halfplanes.

8
1

6 3 2
9

4 6 7 8

1
5 5 10 9
2

10

10

21
Range search in a 2d tree demo

Goal. Find all points in a query axis-aligned rectangle.

Check if point in node lies in given rectangle.
Recursively search left/bottom (if any could fall in rectangle).
Recursively search right/top (if any could fall in rectangle).

8
1

6 3 2
9

4 6 7 8

1
5 5 10 9
2

10

22
Range search in a 2d tree demo

Goal. Find all points in a query axis-aligned rectangle.

Check if point in node lies in given rectangle.
Recursively search left/bottom (if any could fall in rectangle).
Recursively search right/top (if any could fall in rectangle).

8
1

6 3 2
9

4 6 7 8

1
5 5 10 9
2

10

4 done

23
Range search in a 2d tree analysis

Typical case. R + log N.

Worst case (assuming tree is balanced). R + √N.

8
1

6 3 2
9

4 6 7 8

1
5 5 10 9
2

10

24
Nearest neighbor search in a 2d tree demo

Goal. Find closest point to query point.

8
1

6 3 2
9
query point
3

4 6 7 8

1
5 5 10 9
2

10

25
Nearest neighbor search in a 2d tree demo

Check distance from point in node to query point.

Recursively search left/bottom (if it could contain a closer point).
Recursively search right/top (if it could contain a closer point).
Organize method so that it begins by searching for query point.

8
1

6 3 2
9

4 6 7 8
3
1
5 5 10 9
2

10

4 nearest neighbor = 5

26
Nearest neighbor search in a 2d tree analysis

Typical case. log N.

Worst case (even if tree is balanced). N.

8
1

6 3 2
9

4 6 7 8
3
1
5 5 10 9
2

10

4 nearest neighbor = 5

27
30
Flocking birds

Q. What "natural algorithm" do starlings, migrating geese, starlings,

cranes, bait balls of fish, and flashing fireflies use to flock?

http://www.youtube.com/watch?v=XH-groCeKbE
28
Flocking boids [Craig Reynolds, 1986]

Boids. Three simple rules lead to complex emergent flocking behavior:

Collision avoidance: point away from k nearest boids.
Flock centering: point towards the center of mass of k nearest boids.
Velocity matching: update velocity to the average of k nearest boids.

29
Kd tree

Kd tree. Recursively partition k-dimensional space into 2 halfspaces.

Implementation. BST, but cycle through dimensions ala 2d trees.

level ≡ i (mod k)
points points
whose ith whose ith
coordinate coordinate
is less than p’s is greater than p’s

Efficient, simple data structure for processing k-dimensional data.

Widely used.
Adapts well to high-dimensional and clustered data.
Discovered by an undergrad in an algorithms class! Jon Bentley
30
N-body simulation

Goal. Simulate the motion of N particles, mutually affected by gravity.

G m1 m2
Brute force. For each pair of particles, compute force: F =
r2
Running time. Time per step is N 2.

http://www.youtube.com/watch?v=ua7YlN4eL_w

31
Appel's algorithm for N-body simulation

Key idea. Suppose particle is far, far away from cluster of particles.
Treat cluster of particles as a single aggregate particle.
Compute force between particle and center of mass of aggregate.

32
Appel's algorithm for N-body simulation

Build 3d-tree with N particles as nodes.

Store center-of-mass of subtree in each node.
To compute total force acting on a particle, traverse tree, but stop
as soon as distance from particle to subdivision is sufficiently large.

SIAM J. ScI. STAT. COMPUT. 1985 Society for Industrial and Applied Mathematics
Vol. 6, No. 1, January 1985 O08

AN EFFICIENT PROGRAM FOR MANY-BODY SIMULATION*

ANDREW W. APPEL

Abstract. The simulation of N particles interacting in a gravitational force field is useful in astrophysics,
but such simulations become costly for large N. Representing the universe as a tree structure with the
particles at the leaves and internal nodes labeled with the centers of mass of their descendants allows several
simultaneous attacks on the computation time required by the problem. These approaches range from
algorithmic changes (replacing an O(N ’) algorithm with an algorithm whose time-complexity is believed
to be O(N log N)) to data structure modifications, code-tuning, and hardware modifications. The changes
reduced the running time of a large problem (N 10,000) by a factor of four hundred. This paper describes
both the particular program and the methodology underlying such speedups.

1. Introduction. Isaac Newton calculated the behavior of two particles interacting

through the force of gravity, but he was unable to solve the equations for three particles.
In this he was not alone [7, p. 634], and systems of three or more particles can be
solved only numerically. Iterative methods are usually used, computing at each discrete
time interval the force on each particle, and then computing the new velocities and
Impact. Runningpositions
timefor per step is N log N ⇒ enables new research.
each particle.
A naive implementation of an iterative many-body simulator is computationally
very expensive for large numbers of particles, where "expensive" means days of Cray-1 33
time or a year of VAX time. This paper describes the development of an efficient
program in which several aspects of the computation were made faster. The initial
 G EOMETRIC A PPLICATIONS OF BST S

‣ 1d range search
‣ kd trees
‣ interval search trees
Algorithms
‣ line segment intersection

R OBERT S EDGEWICK | K EVIN W AYNE

‣ rectangle intersection
http://algs4.cs.princeton.edu
1d interval search

1d interval search. Data structure to hold set of (overlapping) intervals.

Insert an interval ( lo, hi ).
Search for an interval ( lo, hi ).
Delete an interval ( lo, hi ).
Interval intersection query: given an interval ( lo, hi ), find all intervals
(or one interval) in data structure that intersects ( lo, hi ).

Q. Which intervals intersect ( 9, 16 ) ?

A. ( 7, 10 ) and ( 15, 18 ).
(9, 16)

(7, 10) (21, 24)

(5, 8) (17, 19)

(4, 8) (15, 18)

35
1d interval search API

public class IntervalST<Key extends Comparable<Key>, Value>

IntervalST() create interval search tree

void put(Key lo, Key hi, Value val) put interval-value pair into ST

Value get(Key lo, Key hi) value paired with given interval

void delete(Key lo, Key hi) delete the given interval

all intervals that intersect

Iterable<Value> intersects(Key lo, Key hi)
the given interval

Nondegeneracy assumption. No two intervals have the same left endpoint.

36
Interval search trees

Create BST, where each node stores an interval ( lo, hi ).

Use left endpoint as BST key.
Store max endpoint in subtree rooted at node.

(17, 19) 24

(5, 8) 18 (21, 24) 24

binary search tree
(left endpoint is key)

(4, 8) 8 (15, 18) 18

max endpoint in
subtree rooted at node

(7, 10) 10

37
 Interval search tree demo

To insert an interval ( lo, hi ) :

Insert into BST, using lo as the key.
Update max in each node on search path.

insert interval (16, 22)

(17, 19) 24

(5, 8) 18 (21, 24) 24

(4, 8) 8 (15, 18) 18

(7, 10) 10

38
Interval search tree demo

To search for any one interval that intersects query interval ( lo, hi ) :
If interval in node intersects query interval, return it.
Else if left subtree is null, go right.
Else if max endpoint in left subtree is less than lo, go right.
Else go left.

interval intersection (17, 19) 24

search for (21, 23)

(5, 8) 22 (21, 24) 24

(4, 8) 8 (15, 18) 22

compare (21, 23) to (16, 22)
(intersection!)

(7, 10) 10 (16, 22) 22 (21, 23)

39
Search for an intersecting interval implementation

To search for any one interval that intersects query interval ( lo, hi ) :
If interval in node intersects query interval, return it.
Else if left subtree is null, go right.
Else if max endpoint in left subtree is less than lo, go right.
Else go left.

Node x = root;
while (x != null)
{
if (x.interval.intersects(lo, hi)) return x.interval;
else if (x.left == null) x = x.right;
else if (x.left.max < lo) x = x.right;
else x = x.left;
}
return null;

40
Search for an intersecting interval analysis

To search for any one interval that intersects query interval ( lo, hi ) :
If interval in node intersects query interval, return it.
Else if left subtree is null, go right.
Else if max endpoint in left subtree is less than lo, go right.
Else go left.

Case 1. If search goes right, then no intersection in left.

Pf. Suppose search goes right and left subtree is non empty.
Since went right, we have max < lo.
For any interval (a, b) in left subtree of x,
we have b ≤ max < lo. max

(c, max)
definition of max reason for going right
(a, b) (lo, hi)
Thus, (a, b) will not intersect ( lo, hi ).

left subtree of x right subtree of x

41
 Search for an intersecting interval analysis

To search for any one interval that intersects query interval ( lo, hi ) :
If interval in node intersects query interval, return it.
Else if left subtree is null, go right.
Else if max endpoint in left subtree is less than lo, go right.
Else go left.

Case 2. If search goes left, then there is either an intersection in left

subtree or no intersections in either.

Pf. Suppose no intersection in left.

Since went left, we have lo ≤ max.
Then for any interval (a, b) in right subtree of x, max
hi < c ≤ a ⇒ no intersection in right.
(c, max)

(lo, hi) (a, b)

no intersections intervals sorted
in left subtree by left endpoint

left subtree of x right subtree of x

42
Interval search tree: analysis

Implementation. Use a red-black BST to guarantee performance.

easy to maintain auxiliary information

using log N extra work per op

interval best
operation brute
search tree in theory

insert interval 1 log N log N

find interval N log N log N

delete interval N log N log N

find any one interval

N log N log N
that intersects (lo, hi)

find all intervals

N R log N R + log N
that intersects (lo, hi)

order of growth of running time for N intervals

43
 G EOMETRIC A PPLICATIONS OF BST S

‣ 1d range search
‣ kd trees
‣ interval search trees
Algorithms
‣ line segment intersection

R OBERT S EDGEWICK | K EVIN W AYNE

‣ rectangle intersection
http://algs4.cs.princeton.edu
Orthogonal line segment intersection

Given N horizontal and vertical line segments, find all intersections.

Quadratic algorithm. Check all pairs of line segments for intersection.

Nondegeneracy assumption. All x- and y-coordinates are distinct.

45
Orthogonal line segment intersection: sweep-line algorithm

Sweep vertical line from left to right.

x-coordinates define events.
h-segment (left endpoint): insert y-coordinate into BST.

3 3

2 2

1 1

0 0

y-coordinates 46
Orthogonal line segment intersection: sweep-line algorithm

Sweep vertical line from left to right.

x-coordinates define events.
h-segment (left endpoint): insert y-coordinate into BST.
h-segment (right endpoint): remove y-coordinate from BST.

3 3

1 1

0 0

y-coordinates 47
Orthogonal line segment intersection: sweep-line algorithm

Sweep vertical line from left to right.

x-coordinates define events.
h-segment (left endpoint): insert y-coordinate into BST.
h-segment (right endpoint): remove y-coordinate from BST.
v-segment: range search for interval of y-endpoints.

3 3

1d range
4
search
2

1 1

0 0

y-coordinates 48
Orthogonal line segment intersection: sweep-line analysis

Proposition. The sweep-line algorithm takes time proportional to N log N + R

to find all R intersections among N orthogonal line segments.

Pf.
Put x-coordinates on a PQ (or sort). N log N

Insert y-coordinates into BST. N log N

Delete y-coordinates from BST. N log N

Range searches in BST. N log N + R

Bottom line. Sweep line reduces 2d orthogonal line segment intersection

search to 1d range search.
49
 G EOMETRIC A PPLICATIONS OF BST S

‣ 1d range search
‣ kd trees
‣ interval search trees
Algorithms
‣ line segment intersection

R OBERT S EDGEWICK | K EVIN W AYNE

‣ rectangle intersection
http://algs4.cs.princeton.edu
Orthogonal rectangle intersection

Goal. Find all intersections among a set of N orthogonal rectangles.

Quadratic algorithm. Check all pairs of rectangles for intersection.

Non-degeneracy assumption. All x- and y-coordinates are distinct.

51
Microprocessors and geometry

Early 1970s. microprocessor design became a geometric problem.

Very Large Scale Integration (VLSI).
Computer-Aided Design (CAD).

Design-rule checking.
Certain wires cannot intersect.
Certain spacing needed between different types of wires.
Debugging = orthogonal rectangle intersection search.

COS / ELE 462

52
Algorithms and Moore's law

"Moore’s law." Processing power doubles every 18 months.

197x: check N rectangles.
197(x+1.5): check 2 N rectangles on a 2x-faster computer.

Gordon Moore

Bootstrapping. We get to use the faster computer for bigger circuits.

But bootstrapping is not enough if using a quadratic algorithm:

197x: takes M days.
197(x+1.5): takes (4 M) / 2 = 2 M days. (!)

quadratic 2x-faster
algorithm computer

Bottom line. Linearithmic algorithm is necessary to sustain Moore’s Law.

53
Orthogonal rectangle intersection: sweep-line algorithm

Sweep vertical line from left to right.

x-coordinates of left and right endpoints define events.
Maintain set of rectangles that intersect the sweep line in an interval
search tree (using y-intervals of rectangle).
Left endpoint: interval search for y-interval of rectangle; insert y-interval.
Right endpoint: remove y-interval.

2 3 2

0 0

1 1

y-coordinates
54
Orthogonal rectangle intersection: sweep-line analysis

Proposition. Sweep line algorithm takes time proportional to N log N + R log N

to find R intersections among a set of N rectangles.

Pf.
Put x-coordinates on a PQ (or sort). N log N

Insert y-intervals into ST. N log N

Delete y-intervals from ST. N log N

Interval searches for y-intervals. N log N + R log N

Bottom line. Sweep line reduces 2d orthogonal rectangle intersection

search to 1d interval search.
55
Geometric applications of BSTs

problem example solution

1d range search BST

2d orthogonal line sweep line reduces to

segment intersection 1d range search

kd range search kd tree

1d interval search interval search tree

2d orthogonal sweep line reduces to

rectangle intersection 1d interval search

56