1.

4 Analysis of Algorithms

observations mathematical models order-of-growth classifications dependencies on inputs memory

Algorithms, 4th Edition

Robert Sedgewick and Kevin Wayne

Copyright 20022010

September 20, 2010 8:12:50 PM

Cast of characters

Programmer needs to develop a working solution. Student might play any or all of these roles someday.

Client wants to solve problem efficiently.

Theoretician wants to understand.

Basic blocking and tackling is sometimes necessary. [this lecture]

2

Running time

As soon as an Analytic Engine exists, it will necessarily guide the future course of the science. Whenever any result is sought by its aid, the question will ariseBy what course of calculation can these results be arrived at by the machine in the shortest time? Charles Babbage (1864)

how many times do you have to turn the crank?

Analytic Engine
3

Reasons to analyze algorithms Predict performance. Compare algorithms. Provide guarantees. Understand theoretical basis.
theory of algorithms (COS 423) this course (COS 226)

Primary practical reason: avoid performance bugs.

client gets poor performance because programmer did not understand performance characteristics

Some algorithmic successes Discrete Fourier transform. Break down waveform of N samples into periodic components. Applications: DVD, JPEG, MRI, astrophysics, . Brute force: N 2 steps. FFT algorithm: N log N steps, enables new technology.
time
64T

Friedrich Gauss 1805

quadratic

32T

16T 8T

linearithmic linear
1K 2K 4K 8K
5

size

Some algorithmic successes N-body simulation. Simulate gravitational interactions among N bodies. Brute force: N 2 steps. Barnes-Hut algorithm: N log N steps, enables new research.
Andrew Appel PU '81

time
64T

quadratic

32T

16T 8T

linearithmic linear
1K 2K 4K 8K
6

size

The challenge Q. Will my program be able to solve a large practical input?

Why is my program so slow ?

Why does it run out of memory ?

Key insight. [Knuth 1970s] Use scientific method to understand performance.

Scientific method applied to analysis of algorithms A framework for predicting performance and comparing algorithms. Scientific method. Observe some feature of the natural world. Hypothesize a model that is consistent with the observations. Predict events using the hypothesis. Verify the predictions by making further observations. Validate by repeating until the hypothesis and observations agree.

Principles. Experiments must be reproducible. Hypotheses must be falsifiable.

Feature of the natural world = computer itself.

8

observations mathematical models order-of-growth classifications dependencies on inputs memory

Example: 3-sum 3-sum. Given N distinct integers, how many triples sum to exactly zero?

% more 8ints.txt 8 30 -40 -20 -10 40 0 10 5 % java ThreeSum < 8ints.txt 4

a[i] 1 2 3 4 30 30 -40 -10

a[j] -40 -20 40 0

a[k] 10 -10 0 10

sum 0 0 0 0

Context. Deeply related to problems in computational geometry.

10

3-sum: brute-force algorithm

public class ThreeSum { public static int count(int[] a) { int N = a.length; int count = 0; for (int i = 0; i < N; i++) for (int j = i+1; j < N; j++) for (int k = j+1; k < N; k++) if (a[i] + a[j] + a[k] == 0) count++; return count; } public static void main(String[] args) { int[] a = StdArrayIO.readInt1D(); StdOut.println(count(a)); } }

check each triple we ignore any integer overflow

11

Measuring the running time Q. How to time a program? A. Manual.

70 % java ThreeSum < 2Kints.txt
tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick

% java ThreeSum < 1Kints.txt

tick tick tick

528 % java ThreeSum < 4Kints.txt

tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick tick

4039

Observing the running time of a program

12

Measuring the running time Q. How to time a program? A. Automatic.

public class Stopwatch Stopwatch() double elapsedTime()

create a new stopwatch time since creation (in seconds)

public static void main(String[] args) { int[] a = StdArrayIO.readInt1D(); Stopwatch stopwatch = new Stopwatch(); StdOut.println(ThreeSum.count(a)); double time = stopwatch.elapsedTime(); } client code

13

Measuring the running time Q. How to time a program? A. Automatic.

public class Stopwatch Stopwatch() double elapsedTime()

create a new stopwatch time since creation (in seconds)

public class Stopwatch { private final long start = System.currentTimeMillis(); public double elapsedTime() { long now = System.currentTimeMillis(); return (now - start) / 1000.0; } }
implementation (part of stdlib.jar)
14

Empirical analysis Run the program for various input sizes and measure running time.

N 250 500 1,000 2,000 4,000 8,000 16,000

time (seconds) 0.0 0.0 0.1 0.8 6.4 51.1 ?

15

Data analysis Standard plot. Plot running time T (N) vs. input size N.

standard plot

50

log-log plot

51.2 25.6

stra of

40

12.8

running time T(N )

30

lg(T(N ))
1K 2K 4K 8K

6.4 3.2 1.6 .8 .4

20

10

.2 .1

1K

problem size N Analysis of experimental data (the running time of ThreeSum)

16

Data analysis Log-log plot. Plot running time vs. input size N using log-log scale.
log-log plot

51.2 25.6 12.8

straight line of slope 3

lg(T(N ))

6.4 3.2 1.6 .8 .4 .2 .1

lg(T (N)) = b lg N + c b = 2.999 c = -33.2103 T (N) = a N b, where a = 2 c

8K

1K

2K

4K

8K

lg N
power law slope

mental data (the running time of ThreeSum)

Regression. Fit straight line through data points: a N b.

Hypothesis. The running time is about 1.006 10 10 N 2.999 seconds.

17

Prediction and validation Hypothesis. The running time is about 1.006 10 10 N 2.999 seconds.

Predictions. 51.0 seconds for N = 8,000. 408.1 seconds for N = 16,000.

Observations.

N 8,000 8,000 8,000 16,000

time (seconds) 51.1 51.0 51.1 410.8

validates hypothesis!

18

Doubling hypothesis Doubling hypothesis. Quick way to estimate b in a power-law hypothesis. Run program, doubling the size of the input.
N 250 500 1,000 2,000 4,000 8,000 time (seconds) 0.0 0.0 0.1 0.8 6.4 51.1 4.8 6.9 7.7 8.0 8.0

ratio

lg ratio 2.3 2.8 2.9 3.0 3.0

seems to converge to a constant b 3

Hypothesis. Running time is about a N b with b = lg ratio. Caveat. Cannot identify logarithmic factors with doubling hypothesis.
19

Doubling hypothesis Doubling hypothesis. Quick way to estimate b in a power-law hypothesis. Q. How to estimate a ? A. Run the program!

N 8,000 8,000 8,000

time (seconds) 51.1 51.0 51.1

51.1 = a 80003 a = 9.98 10 11

Hypothesis. Running time is about 9.98 10 11 N 3 seconds.

almost identical hypothesis to one obtained via linear regression

20

Experimental algorithmics System independent effects. Algorithm. Input data.

determines exponent b in power law

System dependent effects. Hardware: CPU, memory, cache, Software: compiler, interpreter, garbage collector, System: operating system, network, other applications,
helps determines constant a in power law

Bad news. Difficult to get precise measurements. Good news. Much easier and cheaper than other sciences.
e.g., can run huge number of experiments
21

War story (from COS 126) Q. How long does this program take as a function of N ?

String s = StdIn.readString(); int N = s.length(); ... for (int i = 0; i < N; i++) for (int j = 0; j < N; j++) distance[i][j] = ... ...

N 1,000 2,000 4,000 8,000

time 0.11 0.35 1.6 6.5

N 250 500 1,000 2,000

time 0.5 1.1 1.9 3.9

Jenny ~ c1 N2 seconds

Kenny ~ c2 N seconds
22

observations mathematical models order-of-growth classifications dependencies on inputs memory

23

Mathematical models for running time Total running time: sum of cost frequency for all operations. Need to analyze program to determine set of operations. Cost depends on machine, compiler. Frequency depends on algorithm, input data.

Donald Knuth 1974 Turing Award

In principle, accurate mathematical models are available.

24

Cost of basic operations

operation integer add integer multiply integer divide floating-point add floating-point multiply floating-point divide sine arctangent ...

example a + b a * b a / b a + b a * b a / b Math.sin(theta) Math.atan2(y, x) ...

nanoseconds 2.1 2.4 5.4 4.6 4.2 13.5 91.3 129.0 ...

Running OS X on Macbook Pro 2.2GHz with 2GB RAM

25

Cost of basic operations

operation variable declaration assignment statement integer compare array element access array length 1D array allocation 2D array allocation string length substring extraction string concatenation

example int a a = b a < b a[i] a.length new int[N] new int[N][N] s.length() s.substring(N/2, N) s + t

nanoseconds c1 c2 c3 c4 c5 c6 N c7 N c8 c9 c10 N
2

Novice mistake. Abusive string concatenation.

26

Example: 1-sum Q. How many instructions as a function of input size N ?

int count = 0; for (int i = 0; i < N; i++) if (a[i] == 0) count++;

operation variable declaration assignment statement less than compare equal to compare array access increment

frequency

2 2 N+1 N N N to 2 N
27

Example: 2-sum Q. How many instructions as a function of input size N ?

int count = 0; for (int i = 0; i < N; i++) for (int j = i+1; j < N; j++) if (a[i] + a[j] == 0) count++;
0 + 1 + 2 + . . . + (N 1) = 1 N (N 1) 2 N = 2

operation variable declaration assignment statement less than compare equal to compare array access increment

frequency

N+2 N+2 (N + 1) (N + 2) N (N 1)
tedious to count exactly

N (N 1) N to 2 N
28

Simplification 1: cost model Cost model. Use some basic operation as a proxy for running time.
int count = 0; for (int i = 0; i < N; i++) for (int j = i+1; j < N; j++) if (a[i] + a[j] == 0) count++;
0 + 1 + 2 + . . . + (N 1) = 1 N (N 1) 2 N = 2

operation variable declaration assignment statement less than compare equal to compare array access increment

frequency

N+2 N+2 (N + 1) (N + 2) N (N 1) N (N 1) N to 2 N
29

cost model = array accesses

Simplification 2: tilde notation

Estimate running time (or memory) as a function of input size N. Ignore lower order terms.

- when N is large, terms are negligible - when N is small, we don't care

N 3/6

Ex 1. Ex 2. Ex 3.

N 3 + 20 N + 16
N 3 + 100 N 4/3 + 56
N3 - N
2

~ N3 ~ N3 ~ N3
N 1,000 Leading-term approximation 166,666,667 N 3/6 N 2/2 + N /3 166,167,000

+ N

discard lower-order terms (e.g., N = 1000: 500 thousand vs. 166 million)

Technical definition. f(N) ~ g(N) means

lim

f (N ) = 1 g( N )
30

Simplification 2: tilde notation

Estimate running time (or memory) as a function of input size N. Ignore lower order terms.

- when N is large, terms are negligible - when N is small, we don't care

operation variable declaration assignment statement less than compare equal to compare array access increment

frequency

tilde notation

N+2 N+2 (N + 1) (N + 2) N (N 1) N (N 1) N to 2 N

~N ~N ~ N2 ~ N2 ~ N2 ~ N to ~ 2 N
31

Example: 2-sum Q. Approximately how many array accesses as a function of input size N ?

int count = 0; for (int i = 0; i < N; i++) for (int j = i+1; j < N; j++) if (a[i] + a[j] == 0) count++;

"inner loop"

A. ~ N 2 array accesses.

0 + 1 + 2 + . . . + (N 1) =

1 N (N 1) 2 N = 2

Bottom line. Use cost model and tilde notation to simplify frequency counts.
32

Example: 3-sum Q. Approximately how many array accesses as a function of input size N ?

int count = 0; for (int i = 0; i < N; i++) for (int j = i+1; j < N; j++) for (int k = j+1; k < N; k++) if (a[i] + a[j] + a[k] == 0) count++;

"inner loop"

A. ~ N 3 array accesses.

N = 3

N (N 1)(N 2) 3! 1 3 N 6

Bottom line. Use cost model and tilde notation to simplify frequency counts.
33

Estimating a discrete sum Q. How to estimate a discrete sum? A1. Take COS 340. A2. Replace the sum with an integral, and use calculus!

Ex 1. 1 + 2 + + N.

N i=1

x=1

x dx

1 2 N 2

Ex 2. 1 + 1/2 + 1/3 + + 1/N.

N 1 i=1

x=1

1 dx = ln N x

Ex 3. 3-sum triple loop.

N N N i=1 j =i k=j

x=1

y =x

z =y

dz dy dx

1 3 N 6

34

Mathematical models for running time In principle, accurate mathematical models are available. In practice, Formulas can be complicated. Advanced mathematics might be required. Exact models best left for experts.
costs (depend on machine, compiler)

TN = c1 A + c2 B + c3 C + c4 D + c5 E
A = array access B = integer add C = integer compare D = increment E = variable assignment
frequencies (depend on algorithm, input)

Bottom line. We use approximate models in this course: T(N) ~ c N 3.

35

observations mathematical models order-of-growth classifications dependencies on inputs memory

36

Common order-of-growth classificiations

200T

Good news. the small set of functions

1, log N, N, N log N, N 2, N 3, and 2N

100K 200K

time

100T

logarithmic constant
500K

suffices to describe order-of-growth of typical algorithms.

size log-log plot

exponential

tic

qua dra

ea lin
size

64T

time

8T 4T 2T T

logarithmic constant
1K 2K 4K 8K 512K

Typical orders of growth

37

lin

hm i ea c r

cubi

512T

r it

Common order-of-growth classificiations

growth rate

name

typical code framework

description

example

T(2N) / T(N)

constant

a = b + c;

statement

add two numbers

log N

logarithmic

while (N > 1) N = N / 2; ...

divide in half

binary search

~1

linear

for (int i = 0; i < N; i++) { ... }

loop

find the maximum

N log N

linearithmic

[see mergesort lecture]

divide and conquer

mergesort

~2

N2

quadratic

for (int i = 0; i < N; i++) for (int j = 0; j < N; j++) { ... } for (int i = 0; i < N; i++) for (int j = 0; j < N; j++) for (int k = 0; k < N; k++) { ... }

double loop

check all pairs

N3

cubic

triple loop

check all triples

2N

exponential

[see combinatorial search lecture]

exhaustive search

check all subsets

T(N)
38

Practical implications of order-of-growth

problem size solvable in minutes growth rate 1970s 1980s 1990s 2000s

any

any

any

any

log N

any

any tens of millions millions

any hundreds of millions millions

any

millions hundreds of thousands hundreds

billions hundreds of millions tens of thousands thousands

N log N

N2

thousand

thousands

N3

hundred

hundreds

thousand

2N

20

20s

20s

30

39

Binary search Goal. Given a sorted array and a key, find index of the key in the array? Successful search. Binary search for 33.

6
0

13
1

14
2

25
3

33
4

43
5

51
6

53
7

64
8

72
9

84
10

93
11

95
12

96
13

97
14

lo

mid

hi

40

Binary search Goal. Given a sorted array and a key, find index of the key in the array? Successful search. Binary search for 33.

6
0

13
1

14
2

25
3

33
4

43
5

51
6

53
7

64
8

72
9

84
10

93
11

95
12

96
13

97
14

lo

mid

hi

41

Binary search Goal. Given a sorted array and a key, find index of the key in the array? Successful search. Binary search for 33.

6
0

13
1

14
2

25
3

33
4

43
5

51
6

53
7

64
8

72
9

84
10

93
11

95
12

96
13

97
14

lo

mid

hi

42

Binary search Goal. Given a sorted array and a key, find index of the key in the array? Successful search. Binary search for 33.

lo = hi

6
0

13
1

14
2

25
3

33
4

43
5

51
6

53
7

64
8

72
9

84
10

93
11

95
12

96
13

97
14

mid return 4

43

Binary search: Java implementation Trivial to implement? First binary search published in 1946; first bug-free one published in 1962. Java bug in Arrays.binarySearch() not fixed until 2006.

public static int binarySearch(int[] a, int key) { int lo = 0, hi = a.length-1; while (lo <= hi) { int mid = lo + (hi - lo) / 2; if (key < a[mid]) hi = mid - 1; else if (key > a[mid]) lo = mid + 1; else return mid; } return -1; }

Invariant. If key appears in the array a[], then a[lo] key a[hi].
44

Binary search: mathematical analysis Proposition. Binary search uses at most 1 + lg N compares to search in a sorted array of size N. Def. T (N) # compares to binary search in a sorted subarray of size N. Binary search recurrence. T (N) T (N / 2) + 1 for N > 1, with T (1) = 1. Pf sketch.
T (N) T (N / 2) + 1 T (N / 4) + 1 + 1 T (N / 8) + 1 + 1 + 1 ... T (N / N) + 1 + 1 + + 1 = 1 + lg N
45

left or right half

given apply recurrence to first term apply recurrence to first term

stop applying, T(1) = 1

An N2 log N algorithm for 3-sum Step 1. Sort the N numbers. Step 2. For each pair of numbers a[i] and a[j], binary search for -(a[i] + a[j]).
input

30 -40 -20 -10 40

sort

0 10

-40 -20 -10

binary search

5 10 30 40

Analysis. Order of growth is N 2 log N. Step 1: N2 with insertion sort. Step 2: N 2 log N with binary search.

(-40, -20) (-40, -10) (-40, 0) (-40, 5) (-40, 10) (-40, 40) (-10, 0) (-20, 10) ( 10, 30) ( 10, 40) ( 30, 40)

60 30 40 35 30 0 10 10 -40 -50 -70

46

only count if a[i] < a[j] < a[k] to avoid double counting

Comparing programs Hypothesis. The N 2 log N three-sum algorithm is significantly faster in practice than the brute-force N 3 one.

N 1,000 2,000 4,000 8,000

time (seconds) 0.1 0.8 6.4 51.1

N 1,000 2,000 4,000 8,000 16,000 32,000 64,000

time (seconds) 0.14 0.18 0.34 0.96 3.67 14.88 59.16

ThreeSum.java

ThreeSumDeluxe.java

Bottom line. Typically, better order of growth faster in practice.

47

observations mathematical models order-of-growth classifications dependencies on inputs memory

48

Types of analyses Best case. Lower bound on cost. Determined by easiest input. Provides a goal for all inputs.

Worst case. Upper bound on cost. Determined by most difficult input. Provides a guarantee for all inputs.

Average case. Expected cost for random input. Need a model for random input. Provides a way to predict performance.
Ex 1. Array accesses for brute-force 3 sum. Best: ~ N3 Average: ~ N 3 Worst: ~ N3 Ex 2. Compares for binary search. Best: ~ 1 Average: ~ lg N Worst: ~ lg N
49

Types of analyses Best case. Lower bound on cost. Worst case. Upper bound on cost. Average case. Expected cost.

Actual data might not match input model? Need to understand input to effectively process it. Approach 1: design for the worst case. Approach 2: randomize, depend on probabilistic guarantee.

50

Commonly-used notations

notation

provides

example

shorthand for

used to

Tilde

leading term

~ 10 N2

10 N2 10 N2 + 22 N log N 10 N2 + 2 N + 37 N2

provide approximate model

Big Theta

asymptotic growth rate

(N2)

9000 N2 5 N2 + 22 N log N + 3N N2

classify algorithms

Big Oh

(N2) and smaller

O(N2)

100 N 22 N log N + 3 N 9000 N2

develop upper bounds

Big Omega

(N2) and larger

(N2)

N5 N3 + 22 N log N + 3 N

develop lower bounds

Common mistake. Interpreting big-Oh as an approximate model.

51

Tilde notation vs. big-Oh notation We use tilde notation whenever possible. Big-Oh notation suppresses leading constant. Big-Oh notation only provides upper bound (not lower bound).

time/memory f(N) values represented by O(f(N))

time/memory c f(N) values represented by ~ c f(N)

input size

input size
52

observations mathematical models order-of-growth classifications dependencies on inputs memory

53

Typical memory requirements for primitive types in Java Bit. 0 or 1. Byte. 8 bits. Megabyte (MB). 1 million bytes. Gigabyte (GB). 1 billion bytes.

type boolean byte char int float long double

bytes 1 1 2 4 4 8 8

for primitive types

54

Typical memory requirements for arrays in Java Array overhead. 16 bytes.

type char[] int[] double[]

bytes 2N + 16 4N + 16 8N + 16

type char[][] int[][] double[][]

bytes ~2MN ~4MN ~8MN

for one-dimensional arrays

for two-dimensional arrays

Ex. An N-by-N array of doubles consumes ~ 8N 2 bytes of memory.

55

Typical memory requirements for objects in Java Object overhead. 8 bytes. Reference. 4 bytes. Ex 1. A Complex object consumes 24 bytes of memory.

public class Complex { private double re; private double im; ... }

8 bytes (object overhead)

8 bytes (double) 8 bytes (double)

24 bytes

bject

c class Complex

24 bytes
object overhead real imag

ivate double real; ivate double imag;

double

value

56

Typical memory requirements for objects in Java Object overhead. 8 bytes. Reference. 4 bytes. Ex 2. A virgin String of length N consumes ~ 2N bytes of memory.

public class String { private int offset; private int count; private int hash; private char[] value; ... }
24 bytes
object overhead value offset count hash

8 bytes (object overhead) 4 bytes (int) 4 bytes (int) 4 bytes (int) 4 bytes (reference to array) 2N + 16 bytes (char[] array)

2N + 40 bytes

ct (Java library)
char[] value; int offset; int count; int hash;

lass String

te te te te

reference values
57

int

Example Q. How much memory does WeightedQuickUnionUF use as a function of N ?

public class WeightedQuickUnionUF { private int[] id; private int[] sz; public WeightedQuickUnionUF(int N) { id = new int[N]; sz = new int[N]; for (int i = 0; i < N; i++) id[i] = i; for (int i = 0; i < N; i++) sz[i] = 1; } public boolean find(int p, int q) { ... } public void union(int p, int q) { ... } }

58

Turning the crank: summary Empirical analysis. Execute program to perform experiments. Assume power law and formulate a hypothesis for running time. Model enables us to make predictions.

Mathematical analysis. Analyze algorithm to count frequency of operations. Use tilde notation to simplify analysis. Model enables us to explain behavior.

Scientific method. Mathematical model is independent of a particular system; applies to machines not yet built. Empirical analysis is necessary to validate mathematical models and to make predictions.

59