1. Introduction to Algorithms
6.046J/18.401J LECTURE 3
Divide and Conquer
Binary search
Powering a number
Fibonacci numbers
Matrix multiplication
Strassen’s algorithm
VLSI tree layout
Prof. Erik D. Demaine
Copyright © Erik D. Demaine and Charles E. Leiserson
September 14, 2005
L2.1

2. The divide-and-conquer design paradigm
Divide the problem (instance) into subproblems.
Conquer the subproblems by solving them recursively.
Combine subproblem solutions.
September 14, 2005
Copyright © Erik D. Demaine and Charles E. Leiserson
L2.2

3. Merge sort Divide: Trivial. Conquer: Recursively sort 2 subarrays.
Combine: Linear-time merge.
September 14, 2005
Copyright © Erik D. Demaine and Charles E. Leiserson
L2.3

4. Merge sort Divide: Trivial. Conquer: Recursively sort 2 subarrays.
Combine: Linear-time merge.
T(n) = 2 T(n/2) + (n)
# subproblems
subproblem size
work dividing and combining
September 14, 2005
Copyright © Erik D. Demaine and Charles E. Leiserson
L2.4

5. Master theorem (reprise)
T(n) = a T(n/b) + f (n)
CASE 1: f (n) = O(nlogba – ), constant  > 0
 T(n) = (nlogba) .
CASE 2: f (n) = (nlogba lgkn), constant k  0
 T(n) = (nlogba lgk+1n) .
CASE 3: f (n) = (nlogba +  ), constant  > 0, and regularity condition
 T(n) = ( f (n)) .
September 14, 2005
Copyright © Erik D. Demaine and Charles E. Leiserson
L2.5

6. Master theorem (reprise)
T(n) = a T(n/b) + f (n)
CASE 1: f (n) = O(nlogba – ), constant  > 0
 T(n) = (nlogba) .
CASE 2: f (n) = (nlogba lgkn), constant k  0
 T(n) = (nlogba lgk+1n) .
CASE 3: f (n) = (nlogba +  ), constant  > 0, and regularity condition
 T(n) = ( f (n)) .
Merge sort: a = 2, b = 2  nlogba = nlog22 = n
 CASE 2 (k = 0)  T(n) = (n lg n) .
September 14, 2005 Copyright © Erik D. Demaine and Charles E. Leiserson L2.6

7. Binary search Find an element in a sorted array:
Divide: Check middle element.
Conquer: Recursively search 1 subarray.
Combine: Trivial.
September 14, 2005
Copyright © Erik D. Demaine and Charles E. Leiserson
L2.7

8. Binary search Find an element in a sorted array:
Divide: Check middle element.
Conquer: Recursively search 1 subarray.
Combine: Trivial.
Example: Find 9
September 14, 2005
Copyright © Erik D. Demaine and Charles E. Leiserson
L2.8

9. Binary search Find an element in a sorted array:
Divide: Check middle element.
Conquer: Recursively search 1 subarray.
Combine: Trivial.
Example: Find 9
September 14, 2005
Copyright © Erik D. Demaine and Charles E. Leiserson
L2.9

10. Binary search Find an element in a sorted array:
Divide: Check middle element.
Conquer: Recursively search 1 subarray.
Combine: Trivial.
Example: Find 9
September 14, 2005
Copyright © Erik D. Demaine and Charles E. Leiserson
L2.10

11. Binary search Find an element in a sorted array:
Divide: Check middle element.
Conquer: Recursively search 1 subarray.
Combine: Trivial.
Example: Find 9
September 14, 2005
Copyright © Erik D. Demaine and Charles E. Leiserson
L2.11

12. Binary search Find an element in a sorted array:
Divide: Check middle element.
Conquer: Recursively search 1 subarray.
Combine: Trivial.
Example: Find 9
September 14, 2005
Copyright © Erik D. Demaine and Charles E. Leiserson
L2.12

13. Binary search Find an element in a sorted array:
Divide: Check middle element.
Conquer: Recursively search 1 subarray.
Combine: Trivial.
Example: Find 9
September 14, 2005
Copyright © Erik D. Demaine and Charles E. Leiserson
L2.13

14. Recurrence for binary search
T(n) = 1 T(n/2) + (1)
# subproblems
subproblem size
work dividing and combining
September 14, 2005
Copyright © Erik D. Demaine and Charles E. Leiserson
L2.14

15. Recurrence for binary search
T(n) = 1 T(n/2) + (1)
# subproblems
subproblem size
work dividing and combining
nlogba = nlog21 = n0 = 1  CASE 2 (k = 0)
 T(n) = (lg n) .
September 14, 2005
Copyright © Erik D. Demaine and Charles E. Leiserson
L2.15

16. Powering a number Problem: Compute an, where n  N.
Naive algorithm: (n).
September 14, 2005
Copyright © Erik D. Demaine and Charles E. Leiserson
L2.16

17. a n/2  a n/2 a (n–1)/2  a (n–1)/2  a Powering a number
Problem: Compute an, where n  N.
Naive algorithm: (n).
Divide-and-conquer algorithm:
a n/2  a n/2
a (n–1)/2  a (n–1)/2  a
if n is even; if n is odd.
an =
September 14, 2005
Copyright © Erik D. Demaine and Charles E. Leiserson
L2.17

18. a n/2  a n/2 a (n–1)/2  a (n–1)/2  a Powering a number
Problem: Compute an, where n  N.
Naive algorithm: (n).
Divide-and-conquer algorithm:
a n/2  a n/2
a (n–1)/2  a (n–1)/2  a
if n is even; if n is odd.
an =
T(n) = T(n/2) + (1)  T(n) = (lg n) .
September 14, 2005
Copyright © Erik D. Demaine and Charles E. Leiserson
L2.18

19. Fibonacci numbers Recursive definition: if n = 0; if n = 1; Fn =
Fn–1 + Fn–2 if n  2. L
Fn =
September 14, 2005
Copyright © Erik D. Demaine and Charles E. Leiserson
L2.19

20. Fn–1 + Fn–2 Fibonacci numbers 5)/ 2 Recursive definition:
if n = 0; if n = 1;
if n  2. L 1
Fn–1 + Fn–2
Fn =
Naive recursive algorithm: ( n) (exponential time), where  = (1 is the golden ratio.
5)/ 2
September 14, 2005 Copyright © Erik D. Demaine and Charles E. Leiserson
L2.20

21. Computing Fibonacci numbers
Bottom-up:
Compute F0, F1, F2, …, Fn in order, forming each number by summing the two previous.
Running time: (n).
September 14, 2005
Copyright © Erik D. Demaine and Charles E. Leiserson
L2.21

22. Computing Fibonacci numbers
Bottom-up:
Compute F0, F1, F2, …, Fn in order, forming each number by summing the two previous.
Running time: (n).
Naive recursive squaring:
Fn =  n/ 5 rounded to the nearest integer.
Recursive squaring: (lg n) time.
This method is unreliable, since floating-point arithmetic is prone to round-off errors.
September 14, 2005 Copyright © Erik D. Demaine and Charles E. Leiserson L2.22

23. 1⎤n ⎡Fn1 ⎣⎢ F 0⎥⎦ . Recursive squaring Fn ⎤ ⎡1 Theorem: ⎥⎦  ⎢⎣1 F n
0⎥⎦ .
⎥⎦  ⎢⎣1 F n
n1
September 14, 2005
Copyright © Erik D. Demaine and Charles E. Leiserson
L2.23

24. 1⎤n ⎡Fn1 ⎣⎢ F 0⎥⎦ . Recursive squaring Fn ⎤ ⎡1 Theorem: ⎥⎦  ⎢⎣1 F
0⎥⎦ .
⎥⎦  ⎢⎣1 F
n n1
Algorithm: Recursive squaring.
Time = (lg n) .
September 14, 2005
Copyright © Erik D. Demaine and Charles E. Leiserson
L2.24

25. 1⎤n ⎡Fn1 ⎣⎢ F 0⎥⎦ . 1⎤ 1 ⎢⎣1 0⎥⎦ Recursive squaring Fn ⎤ ⎡1 Theorem:
0⎥⎦ .
⎥⎦  ⎢⎣1 F
n n1
Algorithm: Recursive squaring.
Time = (lg n) .
Proof of theorem. (Induction on n.)
1⎤ 1
F1 ⎤  ⎡1
Base (n = 1): ⎡F2 .
⎢⎣ F F ⎥⎦
⎢⎣1 0⎥⎦
1 0
September 14, 2005
Copyright © Erik D. Demaine and Charles E. Leiserson
L2.25

26. 1. ⎡F F ⎡ F F ⎣ Fn Fn1⎦ ⎣Fn1 Fn2 ⎦ ⎣1 1⎤n1 ⎡1 ⎡1 0⎥⎦ 1⎤n 0⎥⎦ ⎢⎣1
Recursive squaring
Inductive step (n  2): 1.
⎡F F ⎤
⎡ F F
⎤ ⎡1 ⎤
n1 n
⎥  ⎢ n
n1
⎥  ⎢ 0⎥⎦ ⎢
⎣ Fn
Fn1⎦
⎣Fn1
Fn2 ⎦ ⎣1
1⎤n1 ⎡1 ⎡1 1⎤
0⎥⎦ 1⎤n 0⎥⎦
 ⎢⎣1
⎢⎣1
0⎥. ⎦ ⎡1
 ⎢ ⎣1
September 14, 2005
Copyright © Erik D. Demaine and Charles E. Leiserson
L2.26

27. Matrix multiplication
Input: A = [aij], B = [bij].
Output: C = [cij] = A B.
i, j = 1, 2,… , n.
c12 L c1n ⎤ ⎡a11
c22
⎢ M M O M ⎥
a12 L a1n ⎤ ⎡b11 b12 L b1n ⎤
a22 b22
⎢ M M O M n2
⎡c11 ⎥
L b
⎢c21
L c2n ⎥ ⎢a21
L a2n ⎥ ⎢b21 ⎢
⎥  ⎢
⎥  ⎢ 2n ⎥
M ⎥ ⎢ M
O M ⎥ ⎢c
⎥ ⎢a
⎥ ⎢b ⎥
c L c
a L a
b L b
⎣ n1 n2
nn ⎦ ⎣ nn ⎦ ⎣ n1
nn ⎦
n1 n2 n
cij  aik  bkj
k 1
Copyright © Erik D. Demaine and Charles E. Leiserson
September 14, 2005
L2.27

28. Standard algorithm for i  1 to n do for j  1 to n do cij  0
for k  1 to n
do cij  cij + aik bkj
September 14, 2005
Copyright © Erik D. Demaine and Charles E. Leiserson
L2.28

29. Standard algorithm for i  1 to n do for j  1 to n do cij  0
for k  1 to n
do cij  cij + aik bkj
Running time = (n3)
September 14, 2005
Copyright © Erik D. Demaine and Charles E. Leiserson
L2.29

30. Divide-and-conquer algorithm
nn matrix = 22 matrix of (n/2)n/2) submatrices:
⎡r s ⎤  ⎡a b ⎤  ⎡ e f ⎥⎤
h ⎦ B
⎢⎣t u⎥⎦
C = A
⎣⎢c d ⎥⎦ ⎢⎣g 
r = ae + bg s = af + bh t = ce + dg u = cf + dh
September 14, 2005
8 mults of (n/2)n/2) submatrices 4 adds of (n/2)n/2) submatrices
Copyright © Erik D. Demaine and Charles E. Leiserson
L2.30

31. Divide-and-conquer algorithm
nn matrix = 22 matrix of (n/2)n/2) submatrices:
⎡r s ⎤  ⎡a b ⎤  ⎡ e f ⎥⎤
h ⎦ B
⎢⎣t u⎥⎦
C = A
⎣⎢c d ⎥⎦ ⎢⎣g 
r = ae + bg s = af + bh t = ce + dh u = cf + dg
September 14, 2005
recursive
8 mults of (n/2)n/2) submatrices ^
4 adds of (n/2)n/2) submatrices
Copyright © Erik D. Demaine and Charles E. Leiserson
L2.31

32. Analysis of D&C algorithm
T(n) = 8 T(n/2) + (n2)
# submatrices
submatrix size
work adding submatrices
September 14, 2005
Copyright © Erik D. Demaine and Charles E. Leiserson
L2.32

33. Analysis of D&C algorithm
T(n) = 8 T(n/2) + (n2)
# submatrices
submatrix size
work adding submatrices
nlogba = nlog28 = n3
 CASE 1  T(n) = (n3).
September 14, 2005
Copyright © Erik D. Demaine and Charles E. Leiserson
L2.33

34. Analysis of D&C algorithm
T(n) = 8 T(n/2) + (n2)
# submatrices
submatrix size
work adding submatrices
nlogba = nlog28 = n3
 CASE 1  T(n) = (n3).
No better than the ordinary algorithm.
September 14, 2005
Copyright © Erik D. Demaine and Charles E. Leiserson
L2.34

35. Strassen’s idea Multiply 22 matrices with only 7 recursive mults.
September 14, 2005
Copyright © Erik D. Demaine and Charles E. Leiserson
L2.35

36. Strassen’s idea Multiply 22 matrices with only 7 recursive mults.
P1 = a  ( f – h) P2 = (a + b)  h P3 = (c + d)  e P4 = d  (g – e)
P5 = (a + d)  (e + h) P6 = (b – d)  (g + h)
P7 = (a – c)  (e + f )
September 14, 2005
Copyright © Erik D. Demaine and Charles E. Leiserson
L2.36

37. Strassen’s idea Multiply 22 matrices with only 7 recursive mults.
P1 = a  ( f – h) P2 = (a + b)  h P3 = (c + d)  e P4 = d  (g – e)
P5 = (a + d)  (e + h) P6 = (b – d)  (g + h)
P7 = (a – c)  (e + f )
r = P5 + P4 – P2 + P6 s = P1 + P2
t = P3 + P4
u = P5 + P1 – P3 – P7
September 14, 2005
Copyright © Erik D. Demaine and Charles E. Leiserson
L2.37

38. Strassen’s idea Multiply 22 matrices with only 7 recursive mults.
P1 = a  ( f – h) P2 = (a + b)  h P3 = (c + d)  e P4 = d  (g – e)
P5 = (a + d)  (e + h) P6 = (b – d)  (g + h)
P7 = (a – c)  (e + f )
r = P5 + P4 – P2 + P6 s = P1 + P2
t = P3 + P4
u = P5 + P1 – P3 – P7
7 mults, 18 adds/subs. Note: No reliance on commutativity of mult!
7 mults, 18 adds/subs.
Note: No reliance on commutativity of mult!
September 14, 2005
Copyright © Erik D. Demaine and Charles E. Leiserson
L2.38

39. Strassen’s idea Multiply 22 matrices with only 7 recursive mults.
P1 = a  ( f – h) P2 = (a + b)  h P3 = (c + d)  e P4 = d  (g – e)
P5 = (a + d)  (e + h) P6 = (b – d)  (g + h)
P7 = (a – c)  (e + f )
r = P5 + P4 – P2 + P6
= (a + d) (e + h)
+ d (g – e) – (a + b) h
+ (b – d) (g + h)
= ae + ah + de + dh
+ dg –de – ah – bh
+ bg + bh – dg – dh
= ae + bg
September 14, 2005
Copyright © Erik D. Demaine and Charles E. Leiserson
L2.39

40. Strassen’s algorithm
Divide: Partition A and B into (n/2)(n/2) submatrices. Form terms to be multiplied using + and – .
Conquer: Perform 7 multiplications of (n/2)(n/2) submatrices recursively.
Combine: Form C using + and – on (n/2)(n/2) submatrices.
September 14, 2005
Copyright © Erik D. Demaine and Charles E. Leiserson
L2.40

41. Strassen’s algorithm
Divide: Partition A and B into (n/2)(n/2) submatrices. Form terms to be multiplied using + and – .
Conquer: Perform 7 multiplications of (n/2)(n/2) submatrices recursively.
Combine: Form C using + and – on (n/2)(n/2) submatrices.
T(n) = 7 T(n/2) + (n2)
September 14, 2005
Copyright © Erik D. Demaine and Charles E. Leiserson
L2.41

42. Analysis of Strassen T(n) = 7 T(n/2) + (n2) September 14, 2005
Copyright © Erik D. Demaine and Charles E. Leiserson
L2.42

43. nlogba = nlog27  n2.81 Analysis of Strassen T(n) = 7 T(n/2) + (n2)
 CASE 1  T(n) = (nlg 7).
September 14, 2005
Copyright © Erik D. Demaine and Charles E. Leiserson
L2.43

44. nlogba = nlog27  n2.81 Analysis of Strassen T(n) = 7 T(n/2) + (n2)
 CASE 1  T(n) = (nlg 7).
The number 2.81 may not seem much smaller than 3, but because the difference is in the exponent, the impact on running time is significant. In fact, Strassen’s algorithm beats the ordinary algorithm on today’s machines for n  32 or so.
September 14, 2005
Copyright © Erik D. Demaine and Charles E. Leiserson
L2.44

45. Best to date (of theoretical interest only): (n2.376L).
Analysis of Strassen
T(n) = 7 T(n/2) + (n2)
nlogba = nlog27  n2.81
 CASE 1  T(n) = (nlg 7).
The number 2.81 may not seem much smaller than 3, but because the difference is in the exponent, the impact on running time is significant. In fact, Strassen’s algorithm beats the ordinary algorithm on today’s machines for n  32 or so.
Best to date (of theoretical interest only): (n2.376L).
September 14, 2005 Copyright © Erik D. Demaine and Charles E. Leiserson L2.45

46. VLSI layout
Problem: Embed a complete binary tree with n leaves in a grid using minimal area.
September 14, 2005
Copyright © Erik D. Demaine and Charles E. Leiserson
L2.46

47. VLSI layout
Problem: Embed a complete binary tree with n leaves in a grid using minimal area.
W(n)
H(n)
September 14, 2005
Copyright © Erik D. Demaine and Charles E. Leiserson
L2.47

48. VLSI layout
Problem: Embed a complete binary tree with n leaves in a grid using minimal area.
W(n)
H(n)
H(n) = H(n/2) + (1)
= (lg n)
September 14, 2005
Copyright © Erik D. Demaine and Charles E. Leiserson
L2.48

49. VLSI layout
Problem: Embed a complete binary tree with n leaves in a grid using minimal area.
W(n)
H(n)
H(n) = H(n/2) + (1)
= (lg n)
W(n) = 2 W(n/2) + (1)
= (n)
September 14, 2005
Copyright © Erik D. Demaine and Charles E. Leiserson
L2.49

50. VLSI layout
Problem: Embed a complete binary tree with n leaves in a grid using minimal area.
W(n)
H(n)
H(n) = H(n/2) + (1)
W(n) = 2 W(n/2) + (1)
= (lg n)
= (n)
Area = (n lg n)
September 14, 2005
Copyright © Erik D. Demaine and Charles E. Leiserson
L2.50

51. H-tree embedding L(n) L(n) September 14, 2005
Copyright © Erik D. Demaine and Charles E. Leiserson
L2.51

52. H-tree embedding L(n) L(n) L(n/4) (1) L(n/4) September 14, 2005
Copyright © Erik D. Demaine and Charles E. Leiserson
L2.52

53. H-tree embedding L(n) L(n) = 2 L(n/4) + (1) = ( n ) Area = (n)
September 14, 2005
(1) L(n/4)
Copyright © Erik D. Demaine and Charles E. Leiserson
L2.53

54. Conclusion
Divide and conquer is just one of several powerful techniques for algorithm design.
Divide-and-conquer algorithms can be analyzed using recurrences and the master method (so practice this math).
The divide-and-conquer strategy often leads to efficient algorithms.
September 14, 2005
Copyright © Erik D. Demaine and Charles E. Leiserson
L2.54