Dynamic Programming in Computer Science

Carnegie Mellon University
Research Showcase @ CMU

Computer Science Department School of Computer Science
1979
Dynamic programming in computer science

Kevin Q. Brown
Follow this and additional works at: htp://repository.cmu.edu/compsci
his Technical Report is brought to you for free and open access by the School of Computer Science at Research Showcase @ CMU. It has been
accepted for inclusion in Computer Science Department by an authorized administrator of Research Showcase @ CMU. For more information, please
contact research-showcase@andrew.cmu.edu.
NOTICE WARNING CONCERNING COPYRIGHT RESTRICTIONS:
The copyright law of the United States (title 17, U.S. Code) governs the making
o f photocopies or other reproductions of copyrighted material. Any copying of this
document without permission of its author may be prohibited by law.
CMU-CS-79-106
Dynamic Programming in Computer Science University Libraries

Pittsburgh PA 1 5 2 1 3 - 3 8 9 C
K e v i n Q. B r o w n
F e b r u a r y 1979
DEPARTMENT
of
COMPUTER SCIENCE
510.7808
C28R
79-106
Carnegie-Mel Ion University

PAGE I
T a b l e of C o n t e n t s
1. I n t r o d u c t i o n
2. A n Example
3 . B r i e f S u m m a r y of t h e T h e o r y
4 . S i n g l e S o u r c e S h o r t e s t Path Problems
5. Optimal Parenthesization Problems
5.1 O p t i m i z i n g t h e M u l t i p l i c a t i o n of N M a t r i c e s
5.2 C o n t e x t F r e e Language Recognition
5.3 O p t i m a l A l p h a b e t i c E n c o d i n g and Optimal S e a r c h T r e e s
5.4 O p t i m i z i n g t h e S e a r c h f o r N O r d e r Statistics
6. Optimal Partition Problems

7. Optimal Matching Problems
8. ' H a r d ' Combinatorial Problems
9. Conclusion
References
PAGE 1
1. I n t r o d u c t i o n
" D y n a m i c p r o g r a m m i n g is an important t e c h n i q u e f o r t h e s o l u t i o n o f p r o b l e m s
i n v o l v i n g t h e o p t i m i z a t i o n of a s e q u e n c e of d e c i s i o n s . T h e simple i d e a u n d e r l y i n g
t h i s t e c h n i q u e is to r e p r e s e n t the p r o b l e m b y a p r o c e s s w h i c h e v o l v e s f r o m s t a t e
t o s t a t e in r e s p o n s e to d e c i s i o n s . A t y p i c a l o p t i m i z a t i o n p r o b l e m t h e n b e c o m e s
o n e o f g u i d i n g t h e s y s t e m to a terminal s t a t e at minimum c o s t . W h e n t h e c o s t
s t r u c t u r e is a p p r o p r i a t e , t h e d e t e r m i n a t i o n of an optimal p o l i c y ( s e q u e n c e o f
d e c i s i o n s ) m a y b e r e d u c e d to the s o l u t i o n of a functional e q u a t i o n in w h i c h t h e
s t a t e a p p e a r s as an i n d e p e n d e n t v a r i a b l e . " Karp and Held, 1 9 6 7 .
Dynamic programming ( D P ) is o n e of s e v e r a l problem solving methods u s e d in computer

science (and operations research). O n e of t h e d i s t i n g u i s h i n g f e a t u r e s of a D P a l g o r i t h m is t h e
way it d e c o m p o s e s a p r o b l e m into s u b p r o b l e m s . A p r o b l e m of s i z e N ( t h a t is, N stages)
t y p i c a l l y d e c o m p o s e s t o s e v e r a l p r o b l e m s of s i z e N - 1, e a c h of w h i c h d e c o m p o s e s t o s e v e r a l
p r o b l e m s o f s i z e N - 2, e t c T o a v o i d r e c o m p u t a t i o n of common s u b p r o b l e m s , a DP a l g o r i t h m
typically records t h e s o l u t i o n s to all s u b p r o b l e m s it e n c o u n t e r s . T h u s , at t h e e x p e n s e of
e x t r a s t o r a g e , a D P a l g o r i t h m c a n r e d u c e the time r e q u i r e d to s o l v e t h e o r i g i n a l p r o b l e m . DP
is m o s t a p p r o p r i a t e l y a p p l i e d to p r o b l e m s f o r w h i c h this table of s u b p r o b l e m s o l u t i o n s helps
e l i m i n a t e a g r e a t n u m b e r of r e d u n d a n t c o m p u t a t i o n s . A s t h e q u o t e at t h e b e g i n n i n g o f this
c h a p t e r s u g g e s t s , D P is a p p l i c a b l e p r i m a r i l y to t h e t y p e s of p r o b l e m s t h a t c a n b e expressed
a s a s e q u e n c e o f d e c i s i o n s to b e made at e a c h of s e v e r a l s t a g e s .
For other t y p e s o f p r o b l e m s t h e r e are c o r r e s p o n d i n g problem solving methods. These

m e t h o d s i n c l u d e d i v i d e - a n d - c o n q u e r , linear and i n t e g e r p r o g r a m m i n g , and b r a n c h - a n d - b o u n d .
Divide-and-conquer is applicable to problems that can be divided into subproblems of
( a p p r o x i m a t e l y ) e q u a l s i z e , the r e s u l t s of w h i c h c a n b e e f f i c i e n t l y u s e d t o s o l v e t h e original
problem [2]. F o r e x a m p l e , a d i v i d e - a n d - c o n q u e r d e c o m p o s i t i o n of a p r o b l e m o f s i z e N m a y
produce two s u b p r o b l e m s of s i z e N/2. Linear and i n t e g e r p r o g r a m m i n g are applicable to
problems with linear optimization functions and linear constraints. [11]. (Gilmore [23]
describes some interesting interrelationships b e t w e e n linear and i n t e g e r programming and
dynamic programming.) Branch-and-bound is commonly used for pruning of large tree
s e a r c h e s , a n d is also c l o s e l y r e l a t e d to DP [ 3 9 , 4 9 ] .
Dynamic programming o r i g i n a t e d w i t h the w o r k of Bellman [ 5 ] and has b e e n applied to

problems in o p e r a t i o n s r e s e a r c h , economics, control t h e o r y , computer s c i e n c e , and several
o t h e r areas. T h e t h e o r y has e v o l v e d in s e v e r a l d i r e c t i o n s : discrete vs. continuous states,
finite vs. infinite (countable and uncountable) number of states, and deterministic vs.
stochastic systems. N o t s u r p r i s i n g l y , the l i t e r a t u r e o n d y n a m i c p r o g r a m m i n g is enormous.
The survey p a p e r b y T h o m a s [ 6 9 ] is a g o o d o v e r v i e w of t h e s e v e r a l b r a n c h e s o f D P . The
orientation of this paper is toward the uses of DP in c o m p u t e r science, in particular
combinatorial problems. T h e DP algorithms f o r t h e s e p r o b l e m s a r e t y p i c a l l y d i s c r e t e , f i n i t e ,
PAGE 2
and deterministic. T a r j a n ' s [ 6 8 ] r e c e n t s u r v e y article o n c o m b i n a t o r i a l a l g o r i t h m s i n c l u d e s a

b r i e f d e s c r i p t i o n o f t h e a p p l i c a t i o n s of DP t o c o m p u t e r s c i e n c e .
T h i s p a p e r is i n t e n d e d t o b e a r e f e r e n c e to the uses of d y n a m i c p r o g r a m m i n g in c o m p u t e r
science. M a n y p r o b l e m s a r e c o v e r e d , along w i t h the DP a l g o r i t h m s and r e f e r e n c e s t o r e l a t e d
p r o b l e m s and more advanced results. W e i n t r o d u c e dynami c p r o g r a m m i n g in t h e f o r m o f an
example in Chapter 2 below, and in C h a p t e r 3 present some of t h e t h e o r y of DP and
subsequently apply it t o t h e p r o b l e m of C h a p t e r 2. In the f o l l o w i n g four chapters we
p r e s e n t a n d a n a l y z e e x a m p l e s f r o m s e v e r a l s u b c l a s s e s of DP p r o b l e m s . These subclasses are
(for the most part) defined by the f o r m of t h e recursion for breaking a problem into
subproblems. C h a p t e r 4 c o v e r s Single S o u r c e S h o r t e s t Path P r o b l e m s , C h a p t e r 5 d e s c r i b e s
O p t i m a l P a r e n t h e s i z a t i o n P r o b l e m s , Optimal P a r t i t i o n P r o b l e m s a r e a n a l y z e d in C h a p t e r 6, i n
C h a p t e r 7 w e p r e s e n t Optimal M a t c h i n g P r o b l e m s , and ' H a r d ' C o m b i n a t o r i a l P r o b l e m s a r e t h e
t o p i c o f C h a p t e r 8. E a c h of t h e s e c h a p t e r s a r e i n d e p e n d e n t and c a n b e r e a d s e p a r a t e l y from
the others. ( I t is s u g g e s t e d , h o w e v e r , that C h a p t e r s 2 and 3 b e r e a d b e f o r e a t t a c k i n g a n y o f
t h e material in t h e later chapters.)
PAGE 3
2. A n E x a m p l e
B e f o r e d i v i n g i n t o t h e t h e o r y , it will b e helpful to g i v e some of t h e f l a v o r o f D P a l g o r i t h m s
by g i v i n g an example. T h e d i a g r a m in F i g u r e 2-1 is a g r a p h w i t h ( n o n - n e g a t i v e ) weighted
edges. These weights a r e i n t e r p r e t e d as the " d i s t a n c e s " along t h e e d g e s c o n n e c t i n g the
nodes. T h u s , t h e d i s t a n c e f r o m n o d e (1,1) to n o d e (2,1) is 3 and t h e d i s t a n c e f r o m n o d e ( 2 , 1 )
to n o d e ( 3 , 2 ) is 2. ( T h e d i s t a n c e f r o m n o d e (2,1) to n o d e ( 1 , 1 ) , o n t h e o t h e r h a n d , is not
d e f i n e d i n t h e g r a p h a n d c a n t h u s b e assumed to b e infinite.) T h e d i s t a n c e a l o n g p a t h ( 1 , 1 ) -
( 2 , 1 ) - ( 3 , 2 ) is 3 + 2 = 5. T h e p r o b l e m is to f i n d t h e s h o r t e s t p a t h f r o m n o d e ( 0 , 1 ) t o n o d e ( 4 ,
1).
F i g u r e 2 - 1 : F i n d t h e s h o r t e s t p a t h f r o m Node (0,1) t o N o d e ( 4 , 1 )
There a r e s e v e r a l a p p r o a c h e s that w e can take t o s o l v e this p r o b l e m . O n e a p p r o a c h is t o

e n u m e r a t e all o f t h e p a t h s f r o m n o d e (0,1) to n o d e (4,1) and c h o o s e t h e o n e w i t h minimum
distance. F o r t h e g r a p h o f F i g u r e 2-1 t h e r e a r e 3 3
= 27 d i f f e r e n t p a t h s f r o m n o d e ( 0 , 1 ) t o
node (4,1). Examing all t h e s e p a t h s is s o m e w h a t tedious, but definitely manageable. If,
h o w e v e r , i n s t e a d o f 3 c o l u m n s of 3 n o d e s t h e r e w e r e K columns of L n o d e s , t h e n t h e r e w o u l d
be L K
paths to examine! T h u s , t h e n u m b e r of p a t h s g r o w s v e r y r a p i d l y w i t h L a n d K a n d f o r
l a r g e p r o b l e m s o f this t y p e a n o t h e r m e t h o d will h a v e to b e u s e d .
A d y n a m i c p r o g r a m m i n g a p p r o a c h t u r n s o u t to b e j u s t w h a t w e n e e d . Let
- t h e ( n o n n e g a t i v e ) " d i s t a n c e " f r o m n o d e (i,j) t o n o d e ( i + l , k )
PAGE 4
a n d let
M j j = l e n g t h of the s h o r t e s t p a t h f r o m n o d e (0,1) to n o d e ( i , j ) .
T h e o p t i m i z a t i o n f o r t h e p r o b l e m in F i g u r e 2-1 p r o c e e d s in f o u r stages. In e a c h s t a g e i w e
f i n d t h e l e n g t h s o f t h e s h o r t e s t p a t h s f r o m n o d e (0,1) to t h e n o d e s (i,j), f o r ail j , w h i c h we
d e n o t e b y Mjj. S t a g e 1 is e a s y since t h e r e is o n l y o n e p a t h f r o m n o d e ( 0 , 1 ) t o e a c h o f the
n o d e s (1,1), (1,2), and (1,3). Thus, w e have
l l 011» 1 2 M
012' 1 3
= C
013- M = C M = C
S t a g e 2, h o w e v e r , r e q u i r e s some c o m p u t a t i o n since t h e r e a r e t h r e e p a t h s f r o m n o d e ( 0 , 1 ) t o
e a c h o f t h e n o d e s ( 2 , 1 ) , ( 2 , 2 ) , and (2,3). Thus,
M
21 = m i n
< 011
c + C
lll> C
012 + C
121> 0 1 3 C + C
131 *
M 2 2 = min ( C Q 1 1 + C n 2 , C 0 1 2 + 122' 013
c C + C
132 *
M 2 3 = min ( C 0 1 1 + C 1 1 3 , C Q 1 2 + C 1 2 3 , C Q 1 3 + C 1 3 3 )
I n s t a g e 3 w e f i n d t h e l e n g t h s of the s h o r t e s t path s f r o m n o d e (0,1) to n o d e s ( 3 , 1 ) , ( 3 , 2 ) , a n d
(3,3). H e r e is w h e r e t h e g r e a t a d v a n t a g e of dynamic p r o g r a m m i n g b e c o m e s a p p a r e n t . It is
n o t n e c e s s a r y t o e x a m i n e ail p a t h s f r o m n o d e (0,1) to n o d e s (3,1), (3,2), a n d ( 3 , 3 ) b e c a u s e we
c a n u s e t h e r e s u l t s f r o m s t a g e 2 as f o l l o w s :
M | = min ( M j ^ + C ^ | , M
3 2 2 2 2 * ^ 2 2 1 ' ^ 2 3 **" ^ 2 3 1 ^
M 3 2 = min ( M ^ + ^212* ^ 2 2 2 ^222* ^ 2 3 +
^232 ^
M 3 3 = min ( M j ^ + C j ^ , M 2 2 3 2 2 + ^223* ^ 2 3 ^233 ^
S i m i l a r l y , in s t a g e 4 w e c a n u s e the r e s u l t s f r o m s t a g e 3.
M 4 1 - min ( M 3 1 + C 3 1 1 , M 3 2 + C 3 2 1 , M 3 3 + C 3 3 1 )
T h e s o l u t i o n f o r F i g u r e 2 - 1 is p a t h (0,1) - (1,1) - (2,2) - (3,3) - (4,1) of l e n g t h 7. N o t e t h a t it
is s t r a i g h t f o r w a r d t o d e t e r m i n e that the l e n g t h M ^ j = 7, but r e c o v e r i n g t h e corresponding
path (or paths) is a slightly different problem. It is, h o w e v e r , s i m p l y s o l v e d by simply
recording a pointer Pjj w i t h e a c h d i s t a n c e Mjj w h i c h p o i n t s back to t h e n o d e ( i - l , k ) from
w h i c h t h e s h o r t e s t p a t h to n o d e (i,j) came.
Why c a n t h e r e s u l t s o f s t a g e 2 be u s e d to s o l v e s t a g e 3, and the r e s u l t s o f s t a g e 3 be

u s e d t o s o l v e s t a g e 4? T h i s is b e c a u s e of the Markov property of this t y p e o f p r o b l e m . The
o p t i m u m p a t h f r o m a n o d e (i,j) to n o d e (4,1) d o e s not d e p e n d o n t h e p a t h t a k e n f r o m n o d e ( 0 ,
1) t o n o d e (i,j). T h i s p r o p e r t y c a n be e x p r e s s e d as Bellman's P r i n c i p l e of O p t i m a l i t y :
" A n o p t i m a l p o l i c y has the p r o p e r t y that w h a t e v e r t h e initial s t a t e a n d initial
d e c i s i o n a r e , t h e r e m a i n i n g d e c i s i o n s must c o n s t i t u t e an optimal p o l i c y w i t h r e g a r d
to t h e s t a t e resulting from the first decision." Bellman[57], p.83.
PAGE 5
W e c a n u s e t h i s p r i n c i p l e f o r t h e g e n e r a l case of K columns of L n o d e s . The dynamic

programming approach will take only 0(K*L^) time, a significan t improvement over the
exhaustive approach of examining all L K
paths. B o t h of t h e s e methods require 0(K*L^)
storage to s t o r e the distances ( C j j ) . k T h e g e n e r a l formulas d e s c r i b i n g t h e o p t i m i z a t i o n for
t h i s t y p e o f p r o b l e m , k n o w n as t h e functional equations f o r this p r o b l e m , a r e
M
ij =
T ( M
<i-i),k + c
<i-n,k,j >
w h e r e M . i t = 0 , f o r all k. I n t h e n e x t c h a p t e r w e s h o w h o w to d e r i v e t h i s e q u a t i o n .
n
PAGE 6
3 . B r i e f S u m m a r y of the T h e o r y
B e l l m a n ' s P r i n c i p l e of O p t i m a l i t y , as s h o w n in the example of C h a p t e r 2, is an i m p o r t a n t a n d

p o w e r f u l p r i n c i p l e for dynamic programming. It is, h o w e v e r , lacking in at least t w o w a y s : (1)
I t is n o t s t a t e d v e r y r i g o r o u s l y , and ( 2 ) the p r o b l e m of o b t a i n i n g t h e f u n c t i o n a l e q u a t i o n s is
not e v e n addressed. T o o b t a i n the functional e q u a t i o n s , w e must r e l y o n m e t h o d s b e y o n d t h e
s c o p e of the P r i n c i p l e of Optimality.
By the mid 1960's several people had s o l v e d the first problem by showing that the
P r i n c i p l e o f O p t i m a l i t y d e r i v e s f r o m a m o n o t o n i c i t y p r o p e r t y of t h e c o s t f u n c t i o n s . (See Karp
a n d H e l d [ 4 4 ] f o r r e f e r e n c e s . ) K a r p and Held s u b s e q u e n t l y made c o n s i d e r a b l e p r o g r e s s o n t h e
second problem. In their f o r m u l a t i o n , a c o m b i n a t o r i a l d e c i s i o n p r o b l e m is a s s u m e d t o be
expressed in t h e f o r m o f a discrete decision process (DDP). T h i s f o r m is u s u a l l y a fairly
natural way to express the problem. The dynamic programming algorithm to solve the
p r o b l e m , o n t h e o t h e r h a n d , c o r r e s p o n d s to a sequential decision process ( S D P ) . Karp and
Held describe the t r a n s f o r m (if it e x i s t s ) f r o m a d i s c r e t e d e c i s i o n p r o c e s s t o a s e q u e n t i a l
decision process.
T h e r e s u l t o f K a r p a n d Held c a n be u s e d to d e r i v e the f u n c t i o n a l e q u a t i o n s f o r t h e p r o b l e m
of Figure 2-1. Their r e s u l t , h o w e v e r , will not be f o l l o w e d e x a c t l y b e c a u s e it is l o n g and
complicated to e x p l a i n (and thus discouraging for readers). T h u s , to d e r i v e t h e functional
e q u a t i o n s as p a i n l e s s l y as p o s s i b l e , a modified a p p r o a c h will be d e s c r i b e d w h i c h s h o u l d n ' t go
too far wrong. The f i r s t s t e p is to d e f i n e the d i s c r e t e d e c i s i o n p r o c e s s . The D D P is a
f o u r - t u p l e D « ( A , S, P, f ) w h e r e
A = { p r i m i t i v e d e c i s i o n s },
S = { f e a s i b l e p o l i c i e s } = { tuples of p r i m i t i v e d e c i s i o n s w h i c h t r a n s f o r m the
s t a r t s t a t e to a final s t a t e },
P = { s e t s o f d a t a u s e d b y the c o s t f u n c t i o n f }, and
f = c o s t f u n c t i o n , d e f i n e d o v e r all x £ A*,
w h e r e A * is t h e s e t o f all n - t u p l e s of elements of A f o r n > 0. T h e case n = 0 is t h e empty

s t r i n g a n d is d e n o t e d V \ F o r t h e p r o b l e m of F i g u r e 2-1 the DDP D = ( A , S, P, f ) is d e f i n e d
by
A = { aj, a 2 a 3 } w h e r e aj t r a n s f o r m s t h e s t a t e f r o m n o d e ( i - l , k ) t o n o d e ( i , j ) ,
S = { a a a ai j k 1 | i,j,k < { 1 , 2 , 3 } },
P = { matrices C | C j f k > 0, V i < {0,1,2,3}, j , k < {1,2,3} }, and
f is d e f i n e d r e c u r s i v e l y b y : f ( e , p ) = 0, f(aj,p) - C 0 1 j , f(xajaj,p) - f(xaj,p) + C n j j

w h e r e n = l e n g t h ( x ) + 1, and p * P.
PAGE 7
The c o n v e r s i o n f r o m D D P to SDP ( s e q u e n t i a l d e c i s i o n p r o c e s s ) is not e n t i r e l y automatic.

W e m u s t f i r s t d e f i n e t h e s t a t e s f o r the SDP (and thus the DP a l g o r i t h m ) o r d o s o m e t h i n g else
t h a t is e q u i v a l e n t . * Q n c e t h e s t a t e s a r e d e f i n e d , the o t h e r c o m p o n e n t s o f t h e S D P convert
fairly automatically from the DDP. The discrete decision process D is t r a n s f o r m e d to a
s e q u e n t i a l d e c i s i o n p r o c e s s W = ( Z, P, h, k ) w h e r e
Z is a f i n i t e s t a t e a u t o m a t o n Z = ( A , Q, qQ, F, \ ) w h e r e
A = a l p h a b e t f o r Z (same A as f o r the DDP a b o v e ) ,
Q = { s t a t e s },
q Q = initial s t a t e ,
F = { final s t a t e s }, / correspond to S in the DDP
X = transition function,
P = d a t a f o r t h e c o s t f u n c t i o n s h and k (same as P f o r t h e DDP),
h = c o s t f u n c t i o n d e f i n e d b y : h(r,q,a,p) = cost of r e a c h i n g s t a t e X ( q , a ) b y f i r s t
r e a c h i n g s t a t e q at a c o s t r and t h e n a p p l y i n g t h e s e q u e n c e of d e c i s i o n s a. T h e
f u n c t i o n h r e l a t e s to f b y : h(r,q,a,p) » f ( x a , p ) w h e r e f ( x , p ) = r and X ( q Q , x ) - q .
k ( p ) is t h e c o s t o f t h e null s e q u e n c e (of p r i m i t i v e d e c i s i o n s ) e.
T h e S D P f o r t h e p r o b l e m o f F i g u r e 2 - 1 is a f o u r - t u p l e W = ( Z, P, h, k ) w h e r e
Z is an f s a Z * ( A , Q, qQ, F, X ) w h e r e
A = { aj, a , a 2 3 },
Q 88
( 01>
s s
l 1' 1 2 > 13> 21> 22> 23> 31> 32> 33> 4 1 J»
s s s s s s s s s
^0 =
f s
01
F - { s 4 i },
X is d e f i n e d b y : X(sjj,a^) • s ^ + i ) ^
P = { matrices C | C j j k > 0, V i * {0,1,2,3}, j,k ^ {1,2,3} },
h is d e f i n e d b y : h ( r , S j j , a ^ , p ) = r + C j j ^ , and
k is d e f i n e d b y : k ( p ) = 0.
N o w t h a t w e h a v e d e f i n e d t h e S D P , if t h e m o n o t o n e p r o p e r t y ( d e s c r i b e d b e l o w ) is s a t i s f i e d
then the SDP is a monotone sequential decision process ( M S D P ) f r o m w h i c h t h e functional
e q u a t i o n s f o r the DP algorithm can be d e r i v e d .
1
Karp and Hold use a complicated argument that involves a (finite rank) equiresponse relation of an automaton
w h i c h the states are derived.
PAGE 8
Monotone property: If r^ < r 2 then h(rj,qQ,x,p) < h(r ,qQ,x,p), V x ( A * . 2
( F o r t h e p r o b l e m o f F i g u r e 2 - 1 this is e a s i l y s a t i s f i e d b e c a u s e o f t h e a d d i t i v i t y o f h ( r , q , a ,
p ) . ) T h e f u n c t i o n a l e q u a t i o n s d e f i n e a f u n c t i o n G s u c h that
G
< ^ > - { * | x C > - q } < <e>'V<>P>- h k
T h a t i s , G ( q , p ) is t h e minimum p o s s i b l e cost t o g e t t o state q , u n d e r t h e c o s t f u n c t i o n h. N o w it

is p o s s i b l e t o d e f i n e t h e f u n c t i o n a l e q u a t i o n s .
Functional Equations:
G ( q , p ) = m i n ( k(p),
0 { ( S ( a ) | x ^ ) = q o } h<G<s,p),q,a,p> )
G ( q
' p ) =
{<s,a)|X<T,a)-q} < < >P>.q»a>P> h G s f o r
<* * % •
F o r t h e p r o b l e m o f F i g u r e 2 - 1 this is
G(s ,p) = 0
0 1
G ( S
" P ) =
<W i > ( G(S
«-W P) + >
L e t t i n g M y = G ( s j j , p ) , this is
M Q 1 = 0
M
ii = t n ( M
(i-l),k + C
( -l).K,i '
i
) f 0 r (
^^ ( 0 1 )
-
T h i s e q u a t i o n is t h e same as t h e f u n c t i o n a l e q u a t i o n g i v e n f o r this p r o b l e m in C h a p t e r 2.
The results of Karp a n d Held p r o v i d e a p o w e r f u l a p p r o a c h f o r t h e c o n s t r u c t i o n o f D P

a l g o r i t h m s b u t s e v e r a l d e t a i l s a r e left u n s o l v e d . F o r e x a m p l e , K a r p a n d H e l d n o t e t h a t a g o o d
h e u r i s t i c f o r r e d u c i n g t h e time r e q u i r e d b y a DP algorithm is t o minimize t h e n u m b e r o f s t a t e s
in t h e S D P . T h e p r o b l e m o f minimizing the n u m b e r of s t a t e s , h o w e v e r , is n o t in g e n e r a l
s o l v a b l e i n a f i n i t e n u m b e r o f s t e p s [ 3 8 ] . A n o t h e r detail left u n s o l v e d b y K a r p a n d H e l d i s
t h e p r o b l e m o f t r a n s f o r m i n g t h e p r o b l e m statement to a d i s c r e t e d e c i s i o n p r o c e s s . Given a
p r o b l e m s t a t e m e n t , t h e r e may b e s e v e r a l w a y s t o r e p r e s e n t t h e c o r r e s p o n d i n g D D P .
I n t h e f o l l o w i n g c h a p t e r s w e will a p p l y t h e t e c h n i q u e s p r e s e n t e d in this c h a p t e r t o s e v e r a l
subclasses of DP problems. S i n c e t h e n o t a t i o n is s o m e w h a t c u m b e r s o m e a small c o m p r o m i s e
has b e e n made. T h e c o s t f u n c t i o n s f ( x , p ) and h(r,q,a,p) shall b e ( m o r e c o n v e n i e n t l y , a l t h o u g h
l e s s c o r r e c t l y ) d e n o t e d b y s i m p l y f ( x ) and h(r,q,a), r e s p e c t i v e l y .
PAGE 9
4. S i n g l e S o u r c e S h o r t e s t Path Problems
T h e p r o b l e m o f C h a p t e r 2 is a s p e c i a l case of the single s o u r c e s h o r t e s t p a t h p r o b l e m . But
b e f o r e d e s c r i b i n g t h e g e n e r a l p r o b l e m , w e s h o u l d f i r s t n o t e some o t h e r i n t e r e s t i n g problems
that fall into this special case. O n e of the b e t t e r known problems of this type is the
s t a g e c o a c h p r o b l e m ( [ 7 2 ] ] ) , w h i c h is commonly u s e d as an i n t r o d u c t i o n to D P in operations
research textbooks. A r a t h e r c u r i o u s e x a m p l e of this t y p e of a l g o r i t h m is t h e o r d i n a r y , g r a d e
school a l g o r i t h m f o r long division [8]. But of more i n t e r e s t to c o m p u t e r s c i e n t i s t s is the
V i t e r b i algorithm. T h e V i t e r b i a l g o r i t h m [71, 19] is f o r m a l l y a s t a g e c o a c h t y p e o f p r o b l e m , s o
it is n o t n e c e s s a r y t o d e s c r i b e it in detail h e r e . Its a p p l i c a t i o n s , h o w e v e r , i n c l u d e n o t only
e s t i m a t i o n p r o b l e m s in digital communications [ 9 ] b u t also t e x t r e c o g n i t i o n [ 1 9 ] a n d s y n t a c t i c
pattern recognition [61]. The s t a g e c o a c h p r o b l e m has also b e e n e n c o u n t e r e d in speech
r e c o g n i t i o n [ 3 ] a n d in image u n d e r s t a n d i n g s y s t e m s [ 1 6 ] .
We will now describe the general single source shortest path problem. Let G be a
( d i r e c t e d ) g r a p h o f N n o d e s and E e d g e s . T h e e d g e (if a n y ) f r o m n o d e Nj t o n o d e N j is e d g e
Ejj. T h e c o s t s t r u c t u r e is d e f i n e d b y the w e i g h t s C j j > 0 c o r r e s p o n d i n g to t h e e d g e s E j j . The
p r o b l e m is t o f i n d t h e s h o r t e s t p a t h f r o m node N j to node N^. F o r n o w w e w i l l a s s u m e t h a t
t h e g r a p h G is a c o m p l e t e g r a p h . ( F o r e a c h pair of n o d e s Nj and Nj t h e r e is an e d g e E j j . ) If a
particular g r a p h is n o t a l r e a d y c o m p l e t e , w e can e a s i l y make it so b y augmenting it with
e d g e s Ejj w i t h w e i g h t s C j j = oo.
Our goal now is t o o b t a i n the functional e q u a t i o n s f o r a DP a l g o r i t h m t h a t s o l v e s this

problem. T o d o t h i s , h o w e v e r , the p r o b l e m will h a v e to r e w o r d e d s l i g h t l y . Instead of d i r e c t l y
f i n d i n g t h e s h o r t e s t p a t h f r o m n o d e N j to n o d e N , the a l g o r i t h m will f i n d t h e length o f
N the
shortest path. F o r t h i s p r o b l e m o n l y a t r i v i a l modification of t h e a l g o r i t h m is t h e n necessary
to obtain the actual shortest path. F o r some p r o b l e m s in o t h e r c h a p t e r s , h o w e v e r , the
d i s t i n c t i o n is n o t s o t r i v i a l .
T o o b t a i n t h e f u n c t i o n a l e q u a t i o n s the first s t e p is to d e f i n e t h e d i s c r e t e d e c i s i o n p r o c e s s
( D D P ) D = ( A , S , P, f ). F o r t h e s i n g l e s o u r c e s h o r t e s t p a t h p r o b l e m this i s :
A = { a a2> }, where "af means "go to node L"
S = { A*a|vj }, = set of all feasible paths from node 1 to node N
P = { m a t r i c e s C | C j j > 0, Cjj = 0, i,j=l,2,...N }, and

f, t h e c o s t f u n c t i o n , is d e f i n e d r e c u r s i v e l y b y : f ( e ) = 0, f ( a j ) - C^, f(xajaj) =
f ( x a j ) + C j j , V x c A*.
T h e c o r r e s p o n d i n g S D P W - ( Z, P, h, k ) is
PAGE 10
Z = f s a ( A , Q , q Q , F, X ) w h e r e
A = s a m e as f o r t h e DDP,
Q = { Sj, s 2 > . . . S|\j }, where "s^" means "at node i"
qo =
{ s
l )>
I 7
= { s N },
X(Sj,aj) = Sj,
P = s a m e as f o r t h e DDP,
h(r,Sj,aj) = r + C j j , and
M p ) = 0.
T h e m o n o t o n i c i t y p r o p e r t y is s a t i s f i e d e a s i l y b y the a d d i t i v i t y of t h e c o s t f u n c t i o n . Thus, the

functional equations are:
G(s ) = 0{
G ( s
i> - h ( G ( s
iVi) " T < G s + C
Ji>
L e t t i n g M ( i ) = G ( s j ) t h i s is
M(1) = 0
M(i) = ^ ( M ( j ) + C,.)
W e w i l l n o w s h o w h o w to s o l v e t h e s e functional e q u a t i o n s .
It is n o t as i n t u i t i v e l y o b v i o u s h o w to s o l v e t h e s e e q u a t i o n s as it w a s f o r t h e stagecoach
problems. T h e s o l u t i o n is k n o w n as D i j k s t r a ' s algorithm [ 1 2 , 6, 6 0 ] and r e q u i r e s O ( N ^ ) time
and storage. ( F o r a g e n e r a l g r a p h this is the b e s t that o n e c a n d o b e c a u s e t h e r e a r e O(N^)
edges and any shortest p a t h a l g o r i t h m w i l l , in the w o r s t c a s e , h a v e t o l o o k at all o f the
edges.) The a l g o r i t h m r u n s in N - 1 s t a g e s and in e a c h s t a g e i f i n d s t h e i t h c l o s e s t n o d e to
node Nj.. But e a c h stage requires a comparison b e t w e e n 0(N) nodes, so the total time is
0(N ).
2
PAGE 11
A l g o r i t h m f o r S i n g l e S o u r c e S h o r t e s t Path
I n p u t : C o m p l e t e g r a p h G w i t h i n t e r n o d e distance s C ( I , J ) .
O u t p u t : G ( I ) = d i s t a n c e of s h o r t e s t p a t h f r o m Node 1 to Node I.
H(I> = last n o d e b e f o r e Node I in an optimal path f r o m N o d e 1 t o N o d e I.
Time: 0 ( N ) , Storage: 0 ( N )
2 2
/ Initialization;
For I « - 1 t h r u N d o
b e g i n G ( I ) <- C ( 1 , I ) ; H(I) <- 1; e n d ;
N O D E S <- { 1, 2, 3 , . . . N );
! S - set of nodes for which an optimal path has been found;

S <- { 1 }; N E W <- 1;
/ Optimization;
F o r I <- 2 t h r u N d o
begin
M I N C O S T <- co;
F o r J < N O D E S - S do
begin
N E W C O S T <- G ( i , N E W ) + C ( N E W , J ) ;
If N E W C O S T < G ( J ) t h e n
begin
G ( J ) <- N E W C O S T ; H(J) <- N E W ;
end;
If G ( J ) < M I N C O S T t h e n
begin
M I N C O S T <- G ( J ) ; M I N J 4 - J ;
end
end;
S 4 - S U { M I N J }; N E W <- M I N J ;
end;
This a l g o r i t h m c a n b e m o d i f i e d to r u n in 0 ( E l o g N ) time and 0 ( E ) s t o r a g e , w h e r e E is the

n u m b e r o f e d g e s in t h e g r a p h . ( A h o , H o p c r o f t , and Ullman [ 2 ] g i v e c r e d i t t o J o h n s o n [42].)
F o r g r a p h s w h e r e E < N / l o g N this is b e t t e r than 0 ( N ) time and s p a c e .
2 2
For example, any
p l a n a r g r a p h o f N n o d e s ( N > 2 ) has at most 3*N - 6 ( u n d i r e c t e d ) e d g e s ( o r 6*N - 12 d i r e c t e d
edges) [26]. T h u s , the single source shortest path problem for a planar g r a p h can b e s o l v e d
i n 0 ( N l o g N ) time a n d 0 ( N ) s t o r a g e . T h e s e and r e l a t e d g r a p h algorithms a r e d e s c r i b e d i n A h o ,
H o p c r o f t , a n d Ullman [ 2 ] . A n i n t e r e s t i n g a p p l i c a t i o n of D i j k s t r a ' s a l g o r i t h m s i n c e 1974 is
d e s c r i b e d in Duncan [14]. A r e c e n t s u r v e y of fast expected time a l g o r i t h m s f o r t h e s h o r t e s t
p a t h p r o b l e m ( a n d o t h e r s i m p l e p a t h p r o b l e m s ) is p r o v i d e d in P e r l [ 5 7 ] . Fredman [20] (and
r e f e r e n c e s ) d e s c r i b e t h e all points s h o r t e s t p a t h p r o b l e m .
PAGE 12
5. O p t i m a l Parenthesization Problems
In this chapter we present several problems that can be described as optimal

parenthesization problems. T h a t is, t h e y c a n all be p u t in the f o r m :
G i v e n a s t r i n g of N e l e m e n t s , find an optimal (in some s e n s e ) p a r e n t h e s i z a t i o n o f

the string.
T h e s e p r o b l e m s i n c l u d e o p t i m i z a t i o n of the multiplication of N m a t r i c e s , c o n t e x t f r e e l a n g u a g e
recognition, various k i n d s o f optimal s e a r c h t r e e s , and o p t i m i z i n g t h e s e a r c h f o r N order
statistics.
The obvious brute force method for solving an optimal p a r e n t h e s i z a t i o n problem is to

s i m p l y e x a m i n e all o f t h e p o s s i b l e c o m p l e t e p a r e n t h e s i z a t i o n s of N e l e m e n t s a n d c h o o s e the
best one. Unfortunately, this rapidly becomes unmanageable. The number of ways to
c o m p l e t e l y p a r e n t h e s i z e a s t r i n g of N elements is the Nth Catalan n u m b e r
It c a n b e s h o w n that X ( N ) > 2 ~ N 2
( [ 2 ] , p.73).
The DP a l g o r i t h m s f o r these problems o f f e r substantial improvement o v e r the b r u t e force

method. T h i s is b e c a u s e an o p t i m a l l y p a r e n t h e s i z e d s t r i n g must b e c o m p o s e d o f optimally
parenthesized substrings. (If a s u b s t r i n g c a n b e b e t t e r p a r e n t h e s i z e d , t h e n t h e e n t i r e string
c a n b e b e t t e r p a r e n t h e s i z e d . ) DP uses this p r o p e r t y of optimal p a r e n t h e s i z a t i o n p r o b l e m s to
a c h i e v e f u n c t i o n a l e q u a t i o n s of the f o r m
M., = m , n
ik i<j<k ijk
= f ( M
i j ' ( j * l ) , k >•
M
T h e o p t i m a l p a r e n t h e s i z a t i o n f o r t h e s u b s t r i n g (i,k) is d e t e r m i n e d b y e x a m i n i n g t h e k - i p a i r s
of Mjj and + ( w h i c h have already been computed). Since t h e r e a r e 0 ( N ) s u b s t r i n g s ( i ,
2
k ) , t h e t o t a l t i m e is 0 ( N ) , much less than 0 ( X ( N ) ) .

3
For two of the p r o b l e m s , optimal alphabetic encoding and optimal s e a r c h for N order
s t a t i s t i c s , t h e 0 ( N ) time c a n b e i m p r o v e d to 0 ( N log N) time.
3
Unfortunately, this speedup
r e l i e s o n p r o p e r t i e s t h a t a r e p e c u l i a r to t h e s e individual p r o b l e m s and c a n n o t b e a p p l i e d t o
all of the parenthesization problems. A nice decomposition into parallel computations,
h o w e v e r , is a v a i l a b l e f o r a n y optimal p a r e n t h e s i z a t i o n p r o b l e m . Guibas, Kung, and T h o m p s o n
[ 2 5 ] h a v e s h o w n h o w ( N + 1) ( N + 2) / 2 - 0 ( N ) mesh-connected parallel p r o c e s s o r s
2
can
solve an optimal parenthesization p r o b l e m in 0 ( N ) time. (In general, a K b y K triangular
m e s h - c o n n e c t e d p r o c e s s o r c a n i m p r o v e the time to 0 ( N / K ) . ) 3 2
PAGE 13
5*1 Optimizing the Multiplication of N Matrices

The first p r o b l e m that w e will d e s c r i b e in this c h a p t e r is t h e p r o b l e m o f o p t i m i z i n g the
multiplication of N matrices. Since matrix multiplication is a s s o c i a t i v e , t h e m u l t i p l i c a t i o n s can
b e g r o u p e d as d e s i r e d a n d ( m a t h e m a t i c a l l y ) the r e s u l t will still be t h e same. Computationally,
however, it may make a tremendous difference. For example, consider the following
multiplication problem:
M = M 1 x M 2 x M 3 x M 4 x M .
5
[5x10] [10x100] [100x2] [2x20] [20x50]
W e w i l l a s s u m e in t h i s d i s c u s s i o n that the multiplication of a p x q matrix a n d a q x r matrix

costs pqr operations. (Probert [62] generalizes this problem to allow Strassen matrix
m u l t i p l i c a t i o n as w e l l as t h e o r d i n a r y b l o c k multiplication.) If the m u l t i p l i c a t i o n s a r e g r o u p e d
. M = M 1 x ( M 2 x ( ( M 3 x M 4 ) x M 5 ) )
t h e n t h e m u l t i p l i c a t i o n w i l l c o s t 156,500 o p e r a t i o n s . If, h o w e v e r , t h e g r o u p i n g is
M = ( M 1 x ( M 2 x M 3 ) ) x ( M 4 x M 5 )
t h e n t h e multiplication will cost o n l y 4600 operations.
W e w i l l n o w d e s c r i b e t h e g e n e r a l p r o b l e m and the c o r r e s p o n d i n g O ( N ^ ) time D P a l g o r i t h m . *
Let
M = Mj x M 2 x . . . x Mjvj
where Mj is a c q b y c± m a t r i x , M 2 is a c j b y c 2 matrix, e t c . The discrete decision process
D - ( A , S , P, f ) is d e f i n e d b y
A = { a j j k | l < i < j < k < N } , where a^ means "multiply the result of M± x M^+j
x ... x Mj by the result of Mj+i % x x
S = { sequences of a ^ corresponding to complete p a r e n t h e s i z a t i o n of the N

. m a t r i c e s },
P = { c <E ( I )
+ N + 1
}, a n d
f ( e ) = 0 , f ( x a j j ) - f ( x ) + C j . j * Cj * c .
k k
W e m u s t f i r s t d e f i n e t h e s t a t e s Q b e f o r e c o n s t r u c t i n g the SDP. T h e most n a t u r a l s e t o f s t a t e s
for this problem is the set of all partial parenthesizations of N elements. The SDP
W = ( Z, P, h , k ) is t h e n d e f i n e d b y
^ h i n [ 1 0 ] describes a non-DP O(N) time approximation algorithm that is never w o r s e than a factor of 5/4 f r o m
optimal.
PAGE 14
Z = ( A , Q , q Q , F, \ ), ( T h e c o m p o n e n t s are o b v i o u s . )
P = s a m e as f o r t h e DDP,
h ( r , q , a j j ^ ) = r + c , _ i * Cj * c^, and
k ( p ) = 0.
M o n o t o n i c i t y o f t h e S D P is a s s u r e d b y the a d d i t i v i t y of h. T h u s t h e f u n c t i o n a l e q u a t i o n s a r e
G ( q ) » 0, and
Q
G
<«> = {(s,a)|Ms a)= } ( q
h ( G ( s )
' '
S a )
The o p t i m i z a t i o n p r o c e d u r e s u g g e s t e d b y t h e s e e q u a t i o n s will c e r t a i n l y p r o d u c e an o p t i m a l
parenthesization. T h e time s p e n t , h o w e v e r , will be p r o h i b i t i v e . T h i s is b e c a u s e t h e number
of states that will have t o b e e x a m i n e d is the n u m b e r of p a r t i a l p a r e n t h e s i z a t i o n s of N
e l e m e n t s , w h i c h is at l e a s t as g r e a t as the Nth C a t a l a n n u m b e r X ( N ) . This does not compare
w e l l w i t h t h e 0 ( N ) time that w a s p r o m i s e d a b o v e .
3
Obviously, the choice of states will h a v e
t o b e made m o r e c a r e f u l l y . T h e n a t u r a l choice is not a l w a y s the b e s t .
Let's re-examine the situation. T o p r o d u c e a f a s t e r DP a l g o r i t h m t h e n u m b e r o f states

must be reduced. T h e r e is no g e n e r a l p r o c e d u r e f o r this. In f a c t , I b a r a k i [ 3 8 ] h a s shown
that t h e p r o b l e m is in g e n e r a l u n s o l v a b l e . T h u s , to r e d u c e the n u m b e r o f s t a t e s w e must
exploit properties p e c u l i a r to this p a r t i c u l a r p r o b l e m . In this c a s e a c l u e is p r o v i d e d b y a
c l o s e r e x a m i n a t i o n o f t h e p r i m i t i v e o p e r a t o r s a ^ in t h e set A. If t h e s u b s t r i n g s (Mj,Mj j,...Mj) +
and (Mj |,Mj 2>—^K^

+ +
a r e a
' r e a c
* y ^ " y p a r e n t h e s i z e d , t h e n a p p l i c a t i o n of t h e o p e r a t o r a ^
u
will
m a k e ( M j , M j |,...Mj,...M^) a f u l l y p a r e n t h e s i z e d s u b s t r i n g .
+ T h i s implies that t h e s t a t e s should
b e o f t h e f o r m S j j , l < i < j < N w h e r e s y has the i n t e r p r e t a t i o n " s u b s t r i n g (Mj,Mj ^,...Mj) is f u l l y +
parenthesized."
U n f o r t u n a t e l y , t h e r e is n o w a y that a p a r e n t h e s i z a t i o n s u c h a s ( ( M x M ) x ( M x M ) ) can
b e p r o d u c e d b y the primitive o p e r a t o r s with only these states Sjj. Thus, another property
p e c u l i a r t o p a r e n t h e s i z a t i o n p r o b l e m s will h a v e to b e p u l l e d out of a hat. T h i s o n e is a k i n d
of decomposition property and it looks more like divide-and-conquer [2] than dynamic
programming. T h e p r o p e r t y is that the p a r e n t h e s i z a t i o n of a s u b s t r i n g (i,j) is i n d e p e n d e n t of
t h e p a r e n t h e s i z a t i o n o f a s u b s t r i n g (k,Z) iff i < j < k < L o r k < L < i < j . T h u s , t h e p r o b l e m o f
p r o d u c i n g a s t a t e , call it ( S j j u S g + j ^ ) , f o r w h i c h a j ^ f u l l y p a r e n t h e s i z e s s u b s t r i n g ( i , k ) ( a n d
produces s t a t e S j ^ ) c a n b e d e c o m p o s e d into the p r o b l e m of s o l v i n g Sjj a n d s o l v i n g S^j+i)^-
S i n c e t h e c o s t f u n c t i o n is a d d i t i v e ,
< < ij (j+l),k> > - G s u s G

< ij>
s + G
<S(j+i), >- k
T h u s , the f u n c t i o n a l equations become

PAGE 15
G ( q ) = 0, and
Q
G
 = (S, u 1 " , )
S
M k
G ( ( S
ii U S
<M>^ + C
-i * i* *
C C
- m
j n
G( S j j ) + G(S ( M ) > k ) + c M * c, * c . k
L e t t i n g Gjj - G ( S j j ) , this becomes
G H = 0, and
G., = m i n
G-. + G/- . + c , * c. * c . .
ik j u <j*l ),k i-l j n
k
w h i c h is t h e t y p e o f r e c u r r e n c e desired. H e r e is a p s e u d o - A l g o l a l g o r i t h m t h a t solves
functional equations:
A l g o r i t h m F o r C o m p u t i n g Optimal Multiplication of N Matrices
I n p u t : i n t e g e r N > 0 , array C [ 1 : N + 1 ] of p o s i t i v e i n t e g e r s
Output: A r r a y G w h e r e G ( i , j ) = minimum c o s t of p a r e n t h e s i z i n g s u b s t r i n g (i,j>.
A r r a y H w h e r e H(i,j) = s u b s c r i p t K w h e r e optimal p a r e n t h e s i s f o r s u b s t r i n g ( i , j ) i s
located.
Time: 0(N ), Storage: 0 ( N )
3 2
F o r L <- 1 t h r u N d o
G(L,L) 4- 0;
F o r L <- 2 t h r u N d o / L = length of substrings optimized
F o r I 4- 1 t h r u N + l - L do
begin
M I N C O S T <- oo;
J 4- I + L -1;
F o r K 4- I t h r u J - l do / Find optimal paren for substring (IJ)
begin
C O S T 4 - G(I,K) + G ( K + 1 , J ) + C ( I - l ) * C ( K ) * C ( J ) ;
If C O S T < M I N C O S T t h e n
begin
MINCOST 4- COST;
G(I,J) 4- C O S T ;
H(I,J) 4 - K
end
end
end;
PAGE 16
5.2 Context Free Language Recognition

C o n t e x t f r e e l a n g u a g e r e c o g n i t i o n is a w e l l s t u d i e d p r o b l e m . Younger [75] s h o w e d that a
Turing m a c h i n e c a n r e c o g n i z e a g e n e r a l c o n t e x t f r e e l a n g u a g e in 0 ( N ) t i m e , w h e r e 3
N is the
l e n g t h of the s t r i n g being recognized. Kasami and T o r i i [ 4 5 ] c o n s t r u c t e d an 0 ( N ) t i m e ( R A M ) 2
algorithm for unambiguous context free language recognition and E a r l y [15] produced an
algorithm that not only recognizes general context free languages in O ( N ^ ) t i m e , b u t also
u n a m b i g u o u s l a n g u a g e s in O ( N ^ ) time, and some f u r t h e r r e s t r i c t e d c l a s s e s of l a n g u a g e s (such
a s L R ( k ) ) in 0 ( N ) time. S i n c e t h e n Valiant [ 7 0 ] c o n s t r u c t e d an 0 ( N ^ 1 ) 1 time g e n e r a l context
f r e e r e c o g n i t i o n a l g o r i t h m ( b y using S t r a s s e n matrix multiplication), a n d G r a h a m , H a r r i s o n , a n d
Ruzzo [24] produced an 0(N^/logN) time on-line algorithm for general context free
recognition. I n t h i s p a p e r w e d e s c r i b e the simplest O ( N ^ ) time a l g o r i t h m f o r g e n e r a l context
f r e e language recognition.
The DP algorithm for general context free recognition that is p r e s e n t e d in t h i s paper

r e q u i r e s t h e c o n t e x t f r e e g r a m m a r to be put in C h o m s k y normal f o r m [ 3 2 ] . (Early [15] does
n o t r e q u i r e C h o m s k y n o r m a l f o r m , but his a l g o r i t h m is more c o m p l i c a t e d , t o o . ) T h e g r a m m a r is
a 4 - tuple
G = (Nonterminals, Terminals, Productions, Start symbol).

N o n t e r m i n a l s y m b o l s a r e i n d i c a t e d b y u p p e r case l e t t e r s ( A , B, C , . . .) a n d t e r m i n a l s y m b o l s
a r e i n d i c a t e d b y l o w e r c a s e l e t t e r s (a, b, c, . . .). T h e p r o d u c t i o n s a r e all o f t h e f o r m A - » B C
or A -» a ( C h o m s k y normal form). If m, n, and q a r e s t r i n g s of t e r m i n a l s a n d nonterminals,
t h e n m A n -> m q n iff A - » q is a p r o d u c t i o n . T h e l a n g u a g e g e n e r a t e d b y t h e g r a m m a r is t h e set
o f s t r i n g s o f t e r m i n a l s w s u c h that S -** w w h e r e - » * is the t r a n s i t i v e c l o s u r e o f
To construct t h e D P a l g o r i t h m w e must f i r s t d e f i n e t h e DDP. T h e DDP makes use of a

f u n c t i o n c a l l e d " P r e f i x " d e f i n e d as f o l l o w s :
L e t t h e i n p u t s t r i n g b e Z = z^Z2— |\j» * ' * ^ * ' S °f p r i m i t i v e o p e r a t o r s

z a n c e x e a s r n
aj^. P r e f i x ( x , a j j ) = t h e s t r i n g x minus all o p e r a t o r s to t h e r i g h t o f a ^ .

k If a ^
d o e s n o t o c c u r in x t h e n P r e f i x ( x , a j j ^ ) = e.
F o r e x a m p l e , if x « a
233 123 455 135>
a a a t h e n
Prefixfoa^g)- a
233 123 455*
a a T h e D D P i s
D - ( A , S , P, f ):
A = { a j f k | 1 * j + l j - f 2 - k }»
z z z
S - { A*a l j N },
N o w improved to 0 < N ' ) with tho new matrix multiplication algorithm of Pan [54].
2 7 9
PAGE 17
P = { p r o d u c t i o n s }, a n d
f(e) = e, f<xajj ) k = { B | B->CD w h e r e 3p C ( f(Prefix(x,a j)) j p and 3q D (
f(Prefix(x,a ( j + 1 ) > q > k ) ) }.
T h i s is a m o s t c u r i o u s w a y t o d e f i n e a DDP. The primitive o p e r a t o r s and, e s p e c i a l l y , t h e c o s t

f u n c t i o n r e t u r n sets r a t h e r t h a n s c a l a r s . T h e r e a s o n f o r r e t u r n i n g s e t s is t h a t i n t h i s problem
we a r e i n t e r e s t e d in f i n d i n g all of the n o n t e r m i n a l s that p r o d u c e t h e s u b s t r i n g s Z j Z j j . . . Z j . + If
t h e r e a r e t w o o r m o r e n o n t e r m i n a l s that can g e n e r a t e a s u b s t r i n g Z j Z j j . . . Z j it is n o t s u f f i c i e n t +
to keep only one n o n t e r m i n a l and t h r o w t h e o t h e r s a w a y b e c a u s e t h e p r o d u c t i o n s d o not

t r e a t the nonterminals equivalently. T h u s , if f ( x a j p ) = { B, D, E } a n d f ( x a j ) = { D, E }, t h e n
k q k
a
ipk w
'" c e r
^ ' 'y
a n
be p r e f e r r e d over a | q k . If, h o w e v e r , f ( x a j ) = { B, D, E } a n d f ( x a j ) = { C
p k r k
} then we cannot say w h i c h is b e t t e r because the t w o s e t s are not comparable. This

s i t u a t i o n is d e s c r i b e d as a partial order r e l a t i o n . It is d e f i n e d b y
f(xa j p k ) < f(xa j q k ) iff f ( x a j p k ) c f(xa j q k ).
F u r t h e r m o r e , t h e o b j e c t is t o maximize r a t h e r t h a n minimize f u n d e r t h e p a r t i a l o r d e r in t h i s
problem. ( A c c o r d i n g l y , w e w i l l r e p l a c e " m i n " w i t h " u n i o n " in the f u n c t i o n a l e q u a t i o n s . )
N o w it is time t o p r o d u c e the SDP. Let the s t a t e Sjj be i n t e r p r e t e d as " z j Z j j . . . Z j h a s b e e n +
parsed". To produce an e f f i c i e n t algorithm, w e use p s e u d o - s t a t e s (Sjj U S pk as in the

p r e v i o u s parenthesization problem. T h e SDP is W = ( Z, P, h, k ) w h e r e
Z = f s a ( A , Q , q Q , F, X ) w h e r e
A » s a m e as f o r t h e DDP,
Q =» { S j j , ( S j j U S k Z ) },
qo - ( Jt e
F = { S 1 N } ,
X( (Sjj U S j + ( 1 ) > k ), ajj k ) = s , ik
P = s a m e as f o r t h e D D P ,
h(r,(Sjj u S j ( + 1 ) ) k ),ajj ) = f(xak j j k ) w h e r e X ( q , x ) = (Sjj u S j

0 ( + 1 ) > k ) , and
k ( p ) = e.
B e f o r e w e d e r i v e t h e f u n c t i o n a l e q u a t i o n s w e must f i r s t s h o w that t h e m o n o t o n i c i t y property
is s a t i s f i e d . N o t e t h a t t h e c o s t f u n c t i o n is not s i m p l y a d d i t i v e as in t h e p r e v i o u s p r o b l e m s so
m o n o t o n i c i t y d o e s n o t c o m e q u i t e as automatically. T h u s , w e must s h o w t h a t if x , y « A * , t h e n
If X ( q , x ) - X ( q , y ) and f ( x ) > f ( y ) t h e n f ( x w ) > f ( y w ) , V w < A * .
0 0
T h i s is t r u e b e c a u s e if f ( x ) => f ( y ) t h e n c e r t a i n l y Vw<=A* f ( x w ) f(yw). We conclude that the
functional equations are

PAGE 18
G ( q ) = e, and
Q
G ( S
^ =
,o ,c U
,W G(S
ii (i*i),K>> <V <M), >> a , , p )
US S
k j k
<V (M>,k' ijk>

S
a
U f(x x a 1 2 j j k ) where M q ^ ) = S j j and X C q ^ ) = S ^
H e r e is a n a l g o r i t h m f o r s o l v i n g the f u n c t i o n a l e q u a t i o n s :
A l g o r i t h m F o r G e n e r a l C o n t e x t F r e e Language Recognition
Input: Nonterminals = {Aj,A2, . . . A ^ j } / = start symbol

T e r m i n a l s = { t i » t 2 » • • • *T^
F i n i t e s e t o f p r o d u c t i o n s of the f o r m : Aj->AjA^ o r A j - » t j
S t r i n g Z = Z2Z2».Z|\j w h e r e Zj ( T e r m i n a l s
O u t p u t : A r r a y M w h e r e M(I,K) = { A | A -»* z j Z j + 1 ...z k }
T h u s , Z is r e c o g n i z e d iff s t a r t s y m b o l « A j ( M(1,N).
Time: 0(N ), 3
Storage: 0(N ). 2
F o r I «- 1 t h r u N do
M(I,I) 4 - { A | " A - ^ a f < Productions }
F o r L *- 2 t h r u N d o / L = length of string;
F o r I <- 1 t h r u N + l - L do
begin
K « - I + L - 1;
M(I,K) «- e; / e = empty set
F o r J <- I t h r u K - l d o
M ( I , K ) <- M(I,K) u { A | 3 ^ M ( I , J ) , / S € M ( J + 1 , K ) ( A ^ / S ) }
end;
N o t e t h a t t h e i n n e r l o o p is n o t i m p l e m e n t e d t h e same h e r e as f o r t h e p r e v i o u s t w o algorithms.
T h i s is b e c a u s e t h e p a r t i a l o r d e r r e l a t i o n f o r c e s us to use t h e u n i o n o p e r a t i o n r a t h e r t h a n a
min. S i n c e t h e s e t o f n o n t e r m i n a l s is f i n i t e , the s i z e of e a c h of t h e s e t s M ( I , K ) is b o u n d e d by
a constant. T h u s , t h e e n t i r e a l g o r i t h m still r e q u i r e s o n l y O ( N ^ ) time.
5.3 Optimal Alphabetic Encoding and Optimal Search Trees

T h e r e a r e m a n y k i n d s o f optimal s e a r c h t r e e s and optimal a l p h a b e t i c e n c o d i n g corresponds
naturally to one of the simpler kinds. An optimal alphabetic encoding bears several
similarities to a Huffman code [34]. W e a r e g i v e n an a l p h a b e t A L P H of N c h a r a c t e r s a n d t h e
p r o b a b i l i t i e s p j t h a t a r a n d o m l y c h o s e n c h a r a c t e r is c h a r a c t e r i. E a c h of t h e c h a r a c t e r s must
PAGE 19
be represented by a prefix code c h o s e n f r o m a set of r s y m b o l s . * Furthermore, for an

a l p h a b e t i c e n c o d i n g , t h e c o d e s f o r t h e l e t t e r s of A L P H must be in i n c r e a s i n g n u m e r i c a l o r d e r .
( F o r a H u f f m a n e n c o d i n g t h e c o d e s a r e not r e s t r i c t e d to b e o r d e r e d . ) L e t l { « length of the
e n c o d i n g o f c h a r a c t e r i. T h e o b j e c t i v e is to minimize t h e e x p e c t e d l e n g t h o f a m e s s a g e , t h a t is
N
min T\ i
D
{alphabetic encodings} "\ i"

1=1
Gilbert a n d M o o r e [ 2 2 ] w e r e t h e f i r s t to use DP to s o l v e the p r o b l e m o f c o n s t r u c t i n g an
optimal alphabetic encoding. T h e i r a l g o r i t h m is v e r y similar t o t h e a l g o r i t h m s f o r t h e other
o p t i m a l p a r e n t h e s i z a t i o n p r o b l e m s , and it r e q u i r e s O(IM^) time and O ( N ^ ) s t o r a g e .
The tree corresponding t o an optimal a l p h a b e t i c e n c o d i n g is i l l u s t r a t e d in F i g u r e 5 - 1 . A

" 0 " i n t h e e n c o d i n g c o r r e s p o n d s to a left b r a n c h o f the t r e e and a " 1 " c o r r e s p o n d s t o a r i g h t
branch. E a c h l e a f o f t h e t r e e c o r r e s p o n d s t o o n e of t h e l e t t e r s of t h e a l p h a b e t . The internal
n o d e s d o not r e p r e s e n t l e t t e r s of the alphabet. A general binary search t r e e , on the other
h a n d , m a y h a v e e l e m e n t s at i n t e r n a l n o d e s as w e l l as t h e l e a v e s .
OPTIMAL
ALPHABETIC
OPTIMAL ALPHABETIC TREE
CHARACTER PROBABILITY CODE
A 0.15 000
B 0.10 001 ,
C 0.15 010
D 0.10 Oil
E 0.20 10
F 0.05 1100
0.10 1101
G
0.15 111
H
F i g u r e 5 - 1 : E x a m p l e o f Optimal A l p h a b e t i c E n c o d i n g and R e l a t e d D e c i s i o n T r e e
Knuth describes the c o n s t r u c t i o n of a g e n e r a l optimal s e a r c h t r e e in 0 ( N ) 2

time (and
storage) ( [ 4 6 ] , [47J,p.433-439, [2], p.l 19-123). Not o n l y does this search tree include
elements at internal nodes but it also is e x t e n d e d to allow (nonzero) probabilities that
For this presentation w e will assume that r - 2, thus making the codes binary codes.
l
PAGE 20
unsuccessful searches a r e made in the t r e e . T h a t is, the s e a r c h k e y m a y b e less t h a n A,

b e t w e e n A a n d B, b e t w e e n B and C, e t c . Hu and T u c k e r [ 3 6 ] p r o d u c e d a m u c h d i f f e r e n t (and
c o m p l i c a t e d ) a l g o r i t h m to c o n s t r u c t an optimal a l p h a b e t i c t r e e in 0 ( N ) time a n d 0 ( N )
2
storage
( s e e a l s o H u [ 3 3 ] ) , w h i c h K n u t h [ 4 7 ] i m p r o v e d to O ( N l o g N ) time a n d 0 ( N ) s t o r a g e . Since t h e n ,
I t a i [ 4 0 ] h a s p r o d u c e d DP s o l u t i o n s f o r s e v e r a l c o n s t r a i n e d v a r i a t i o n s of t h e p r o b l e m . Payne
a n d M e i s e l [ 5 6 ] d e s c r i b e an i n t e r e s t i n g DP a l g o r i t h m f o r an optimal m u l t i - d i m e n s i o n a l binary
tree. W e w i l l p r e s e n t t h e o p t i m a l a l p h a b e t i c e n c o d i n g a l g o r i t h m of G i l b e r t a n d M o o r e . (For
treatments o f t h e g e n e r a l o p t i m a l s e a r c h t r e e t h e r e a d e r is r e f e r r e d to t h e t e x t s o f Knuth
a n d A h o , H o p c r o f t , a n d Ullman c i t e d a b o v e . )
T h e f i r s t s t e p t o w a r d s o l v i n g the optimal a l p h a b e t i c e n c o d i n g p r o b l e m is t o d e f i n e t h e D D P .
L e t t h e a l p h a b e t A L P H = ( z j , z , . . . z ) . T h e DDP D = ( A , S, P, f ) is
2 N
A = { a ^ | 1 •
• • k> z
S = { s e q u e n c e s o f a ^ c o r r e s p o n d i n g to total p a r e n t h e s i z a t i o n s of N e l e m e n t s },
P = { (p ,p , . . .
x 2 p ) « (Reals )
N
+ N
}, and
f(e) = 0, f ( x a j j k ) = f(x) + X P . m
m=i,k
Note t h a t t h e c o s t f u n c t i o n f d o e s not seem to d e p e n d o n j at all. The dependence on j

comes from the constraint o n the legal p a r e n t h e s i z a t i o n s in S , as w i l l be shown in the
functional equations.
T o c o n s t r u c t t h e S D P w e must f i r s t d e f i n e t h e s t a t e s . Let t h e s t a t e Sjj b e i n t e r p r e t e d as

"an optimal alphabetic code has b e e n c o n s t r u c t e d f o r ZjZj j...Zj". T o
+ produce an efficient
algorithm, w e w i l l u s e p s e u d o - s t a t e s (Sjj U S | ) as in t h e p r e v i o u s o p t i m a l
k parenthesization
problems. T h e S D P is W = ( Z, P, h, k ) w h e r e
PAGE 21
Z = ( A , Q , q , F, X ) w h e r e
Q
Q - { S j j , ( S j j u S ) | l<i<j<k<l<N } U { e },
K I
= e
% '
F = { S 1 N } ,
X( (Sjj U S j ( + 1 ) > k ) , ajj k ) = s ,

ik
h( r, ( S . . u S ( > U k ) f a. ) = r +
jk S P , and
m
k ( p ) = 0.
The monotonicity property is s a t i s f i e d b y the a d d i t i v i t y of t h e c o s t f u n c t i o n and thus the
functional equations are
G ( e ) = 0, a n d
« V - (S US 8 "i" ),a, ) * « V W
k ik V W '*>
k
= 1" G
ii +
W +
2 P m .
m=i
L e t t i n g Gjj = G ( S j j ) this becomes
Gjj - 0, and
k
G
iK - T G
ii + G
(M).K +
2 P -
m
These f u n c t i o n a l e q u a t i o n s c a n b e s o l v e d in O ( N ^ ) time b e c a u s e t h e s u m m a t i o n f r o m m=i to
m=k of p m c a n b e c o m p u t e d in 0 ( 1 ) time a f t e r some p r e p r o c e s s i n g . The following algorithm
• s h o w s h o w t h a t is d o n e .
PAGE 22
A l g o r i t h m f o r O p t i m a l A l p h a b e t i c Encoding
I n p u t : i n t e g e r N > 0 , a r r a y P [ 1 : N ] of p o s i t i v e r e a l s
O u t p u t : G ( I , K ) = c o s t of o p t i m a l e n c o d i n g of 2jZj+^...z«.
H ( I , K ) = s u b s c r i p t J f o r optimal p a r t i t i o n of ZjZj ^...z^.
+
Time: 0 ( N ) , Storage: 0 ( N )
3 2
/ Initialization of G and C;
C ( 0 ) 4 - 0;
F o r I «- 1 t h r u N do
begin
G ( I , I ) «- 0; C ( I ) «- C ( I - l ) + P(I)
end;
/ Optimization over all pairs (I>K).;
F o r L *- 2 t h r u N d o / L = length of string;
F o r I «- L t h r u N + l - L do
begin
K <~ I + L - 1;
F o r J «- I t h r u K - l d o
begin
C O S T «- G ( I , J ) + G(J+1,K) + C ( K ) - C ( I - l ) ;
begin
M I N C O S T <- C O S T
G(I,K) <- C O S T ;
H(I,K) *- J ;
end
end
end;
5.4 Optimizing the Search for N Order Statistics

O p t i m i z i n g t h e s e a r c h f o r N o r d e r statistics is v e r y similar t o optimal a l p h a b e t i c encoding.
T h e p r o b l e m is d e s c r i b e d as f o l l o w s :
L e t X b e a s e t o f M r e a l n u m b e r s Xj. W e w a n t to find N ( N < M) o r d e r s t a t i s t i c s

Sj w h e r e Sj e q u a l s t h e aj'th smallest element of X. F o r e x a m p l e , if X = { 1 , 5 , 2 , 9 , - 2 } ,
M = 5 , N = 1, a n d a^ = 4 , t h e n s ^ = 5. T h e p r o b l e m is to f i n d t h e N o r d e r s t a t i s t i c s
Sj as q u i c k l y as p o s s i b l e .
One way t o f i n d t h e o r d e r s t a t i s t i c s is to s o r t t h e M e l e m e n t s of X in O ( M l o g M ) t i m e and

t h e n o b t a i n t h e N o r d e r s t a t i s t i c s in 0 ( N ) time ( b y using the aj as indices i n t o t h e s o r t e d set
X ) f o r a total time of O(MlogM); F o r large M we can do better. F o r the c a s e N = 1, t h e o r d e r
s t a t i s t i c s Sj c a n b e f o u n d in 0 ( M ) t i m e [ 7 , 6 3 ] . P o h l [ 5 9 ] d e s c r i b e s a s p e c i a l c a s e o f N « 2,.the
PAGE 23
m i n a n d t h e m a x , a n d s h o w s h o w t h e y can be f o u n d in | 3 M / 2 - 2 | comparisons, which is

b e t t e r t h a n t h e o b v i o u s 2 M - 2 c o m p a r i s o n algorithm. T h e DP a l g o r i t h m t h a t w e d e s c r i b e in
t h i s p a p e r m a k e s o p t i m a l u s e of a linear time s e l e c t i o n a l g o r i t h m . It d o e s n o t a t t e m p t t o do
a n y o p t i m i z a t i o n of the kind d e s c r i b e d b y Pohl.
To construct t h e D P a l g o r i t h m w e s t a r t , as a l w a y s , w i t h t h e d i s c r e t e decision process
D - < A , S , P, f ):
A = { aQ, a ^ , a2> . . . ajvj, ajvj+j }, / where a^ 1 < i < N finds the order statistic
s
i* 0
a
~ ^» anc
^ N+l
a =
M+i.
S = { x € A * | all aj a r e in x },
p = { a < M N
}, a n d
f ( e ) = 0 , f ( a j ) = M, f ( x a j ) = f ( x ) + a k - aj w h e r e j = max s u b s c r i p t o f a in x t h a t is
l e s s t h a n i, a n d k = min s u b s c r i p t of a in x that is g r e a t e r t h a n i.
When constructing t h e S D P w e e n c o u n t e r e x a c t l y the same p r o b l e m s w i t h t h e number of
states as with the problem of optimal multiplication of N matrices. Fortunately, the
d e c o m p o s i t i o n p r o p e r t y still a p p l i e s s o the trick of c o n s t r u c t i n g " s t a t e s " ( S j j U S p w i l l

k work.
S i n c e t h e d e t a i l s a r e b a s i c a l l y the same h e r e as f o r the p r e v i o u s p a r e n t h e s i z a t i o n problems,
w e w i l l p r e s e n t o n l y t h e r e c u r r e n c e r e l a t i o n and the a l g o r i t h m .
F u n c t i o n a l E q u a t i o n s f o r O p t i m i z i n g the S e a r c h f o r N O r d e r S t a t i s t i c s
G(i,i+1) = 0
G(i,K) = G(i,j) + G(j,k)

PAGE 24
A l g o r i t h m f o r O p t i m i z i n g t h e use of a Linear Selection A l g o r i t h m

t o F i n d N O r d e r Statistics
I n p u t : V e c t o r A s u c h that 0 = A ( 0 ) < A ( l ) < A(2) < . . . A(N) < A ( N + i ) = M + l

O u t p u t : A r r a y G w h e r e G ( I , K ) = minimum cost of finding all o r d e r s t a t i s t i c s A ( J ) ,
I < J < K, g i v e n that A ( I ) and A ( K ) h a v e a l r e a d y b e e n d e t e r m i n e d .
A r r a y H w h e r e H(I,K) = s u b s c r i p t J of optimal A ( J ) f o r s u b d i v i d i n g
the interval [A(I),A(K)].
Time: 0(N >, Storage: 0(N )
3 2
F o r L <- 0 t h r u N d o
G ( L , L + 1 ) <- 0;
F o r L <- 2 t h r u N + l d o / L =» Length of substrings optimized
F o r I <- 0 t h r u N + i - L d o
begin
K <- I + L;
/ Find optimal paren for substring (7,K;)
F o r J *- 1+1 t h r u K - l do
begin
C O S T <- G ( I , J ) + G(J,K) + A ( K ) - A ( I ) - 1;
begin
G ( I , K ) <- M I N C O S T f - C O S T ;
H(I,K) <- J
end
end
end;
W e c a n i m p r o v e t h i s a l g o r i t h m to O ( N l o g N ) time and 0 ( N ) s t o r a g e . T h i s is b e c a u s e it c a n b e
put it i n t h e f o r m o f an o p t i m a l a l p h a b e t i c e n c o d i n g p r o b l e m . In the f u n c t i o n a l equations
above, the quantity G(i,k) is i n t e r p r e t e d to equal the minimum c o s t of finding all order
s t a t i s t i c s a j , i < j < k, g i v e n t h a t aj and a k have already been determined. C h a n g e this to
G ( i , k ) = minimum c o s t o f finding all o r d e r statistics a j , i < j < k, g i v e n t h a t aj a n d

a
k + l h a v e already been determined.
A l s o r e w r i t e the quantity a ^ j - a as the sum ( f r o m m=i to m=k) of ( a + j - a ) .

} m m The new
f u n c t i o n a l equations are
PAGE 25
N = 4
M = 100 G 0 1 2 CO 4 5 H 0 1 2 3 4 5
0 0 34 103 177 229 0 1 2 2 2
ORDER
1 0 49 118 159 l 2 3 3
STATISTICS
2 0 54 95 2 3 3
20 3 0 30 3 4
35 4 0 4
70 5 5
F i g u r e 5 - 2 : E x a m p l e of f i n d i n g optimal u s e of a linear s e l e c t i o n a l g o r i t h m .
G(i,i) = 0, and
k
G(i,k> = G(i,j) + GCJ+1.K) • X (a m + 1 - a > - 1.

m
J m=i
The functional equations are now in the same f o r m as t h e f u n c t i o n a l e q u a t i o n s for an

optimal alphabetic encoding. T h u s , the O ( N l o g N ) time and 0 ( N ) s t o r a g e a l g o r i t h m f o l l o w s .
PAGE 26
6 . O p t i m a l Partition Problems
O p t i m a l p a r t i t i o n p r o b l e m s a r e not e n t i r e l y unlike optimal p a r e n t h e s i z a t i o n p r o b l e m s . They

b o t h i n v o l v e a p a r e n t h e s i z a t i o n of a s t r i n g of N e l e m e n t s . The partition problem, however,
does not p r o d u c e a complete p a r e n t h e s i z a t i o n of the N e l e m e n t s . I n s t e a d , it p a r t i t i o n s the
s t r i n g i n t o K s e t s o f c o n s e c u t i v e e l e m e n t s ( w i t h K s e t s of u n n e s t e d p a r e n t h e s e s ) f o r some
v a l u e o f K. T h e v a l u e o f K m a y b e a g i v e n c o n s t a n t o r it may b e left u n s p e c i f i e d .
P e t e r s o n , B i t n e r , and H o w a r d [ 5 8 ] describe a problem for which the number of p a r t i t i o n s K

is u n s p e c i f i e d . T h e p r o b l e m is to s e l e c t optimal tab s e t t i n g s to minimize t h e n u m b e r o f b l a n k s
in a document. N is t h e n u m b e r of columns and K is t h e n u m b e r of t a b s s e t . The functional
e q u a t i o n s , h o w e v e r , are of the form
G Q = 0, and
G. = m a x
G. + C
i l l ! )
w h i c h is v e r y s i m i l a r t o t h e s o l u t i o n i o r the single s o u r c e s h o r t e s t p a t h p r o b l e m d e s c r i b e d i n
Chapter 4. Because of this similarity this p r o b l e m will not b e d i s c u s s e d further. The
r e m a i n d e r o f t h i s c h a p t e r w i l l i n s t e a d d e s c r i b e an e x a m p l e w h e r e K is a g i v e n c o n s t a n t .
T h e a u t h o r k n o w s o f t w o n o n t r i v i a l p r o b l e m s w h i c h a r e of t h e optimal p a r t i t i o n t y p e (with
a given value for the required n u m b e r of p a r t i t i o n s K). One is the resource allocation
problem ( [ 1 3 ] , c h . 3 ) , b u t t h e b e s t example of this t y p e of p r o b l e m is t h e o n e s o l v e d by
Fisher's a l g o r i t h m [ 1 7 , 2 7 ] . * T o d e s c r i b e the p r o b l e m w e will f i r s t h a v e t o i n t r o d u c e some
notation. L e t t h e i n p u t s t r i n g of N elements be d e n o t e d x j , x , X 3 , . . . X|\j, w h e r e Xj is a r e a l .
2
A cluster ( i , j ) is s i m p l y t h e s u b s t r i n g (xj,Xj + j , . . . X j ) . C o r r e s p o n d i n g to e a c h c l u s t e r ( i , j ) is a
diameter D ( i , j ) > 0 t h a t is a f u n c t i o n of ( x j , X j j , . . . X j ) . T h e d i a m e t e r f u n c t i o n t h a t w i l l b e u s e d
+
i n t h i s d e s c r i p t i o n is
D(i,j) = jL ( x . - x ) , w h e r e x » ^2
k=i " k
" k-i (
J- , + 1 )
L e t a K - partition o f t h e N e l e m e n t s be r e p r e s e n t e d b y a K + l - t u p l e (Io*'l*^2* • " • W w

^ e r e
0 = Iq < I| < I2 < • • • < I « - l < I « = N. T h e i n t e r p r e t a t i o n is that t h e f i r s t c l u s t e r is (Iq+I,^),

t h e s e c o n d c l u s t e r is ( I j + 1 , I ) I 2
e t c
- T h e optimal p a r t i t i o n p r o b l e m ( w h i c h F i s h e r ' s algorithm
s o l v e s ) is
Vast
algorithms f o r several special cases of Fisher's clustering problem will be presented b y Shamos, B r o w n . Saxe
and Weide [66]. '
PAGE 27
k
m ,
^n
D ( I +1 I )
{K-partitions} ~ u
^j ' j.l -
+ i 1 ;
O n e m e t h o d o f s o l u t i o n is t o s i m p l y g e n e r a t e all p o s s i b l e K - p a r t i t i o n s a n d c h o o s e t h e one
w i t h minimum cost. But t h e r e are
( N - 1)!
(ID- ( N - K)! * (K - 1)!
possible partitions, which for large K and N will be prohibitive. Dynamic programming
p r o v i d e s considerable improvement. T o p r o d u c e the DP a l g o r i t h m w e f i r s t c o n s t r u c t t h e DDP
D - ( A , S , P, f ):
A = { a j j | 1 < i < j < N } w h e r e ajj d e f i n e s t h e s u b s t r i n g ( x j X j i...Xj) as o n e o f t h e +
K sets of the partition,
S = { s e q u e n c e s o f K a^'s that p r o d u c e legal p a r t i t i o n s },
P = { N - t u p l e s o f r e a l s }, and
f ( e ) = 0 , f ( x a j j ) = f ( x ) + D(i,j) f o r all legal s e q u e n c e s x a j j .
To define the S D P w e must f i r s t d e f i n e the s t a t e s . One w a y t o d o this is t o let S ^ j j be
interpreted " s u b s t r i n g (i,j) is p a r t i t i o n e d into h sets of c o n s e c u t i v e e l e m e n t s . " T h i s l e a d s t o a
D P a l g o r i t h m t h a t c o s t s 0 ( K N ) time and 0 ( K N ) s t o r a g e .
3 2
A n o t h e r c h o i c e o f s t a t e s is a v a i l a b l e ,
h o w e v e r , t h a t n a t u r a l l y l e a d s t o a DP algorithm that c o s t s o n l y O ( K N ^ ) time a n d 0 ( K N ) s t o r a g e .
The improvement is to c o n s i d e r o n l y s t a t e s S ^ j = S ^ i j , w h i c h a r e i n t e r p r e t e d as "substring
( l , j ) is p a r t i t i o n e d i n t o h s e t s of c o n s e c u t i v e e l e m e n t s . " T h u s , t h e S D P W = ( Z, P, h, k) is
Z = ( A , Q , q , F, X ) w h e r e Q
Q = { S h j | 1 < h < k, 1 < j < N } u { e },
qo = > e
F - { S K N },
x ( s
hi' (i+l),j
a ) = S
(h+l),j'
h(r,S ,a(
h i j + 1 ) j ) = r + D(i,j), and
k(p) = 0.
M o n o t o n i c i t y o f t h e S D P f o l l o w s f r o m a d d i t i v i t y of t h e c o s t f u n c t i o n D and t h e f a c t t h a t D ( i , j ) >
0. T h u s , t h e functional equations are

PAGE 28
G ( e ) = 0, a n d
G ( S
^ =
<Vui"(i.l>,i> h ( G (
Vl>,i>. W a
(i.D,i>
-1" G
<Vn,i> +
°< 'i>i+1
L e t t i n g G j = G(S| j) this becomes

h 1 >
Fisher's version of the algorithm for solving these equations first computes all of the
diameters D ( i , j ) a n d s t o r e s t h e m in an a r r a y , c o s t i n g O ( N ^ ) s t o r a g e . The algorithm below
c r e a t e s v e c t o r s S U M a n d S U M S Q , w h i c h e n a b l e the diameter to b e c o m p u t e d in 0 ( N ) s t o r a g e
at n o e x t r a e x p e n s e in time. ( T h i s trick d o e s not w o r k f o r ail d i a m e t e r f u n c t i o n s , t h o u g h . )
PAGE 29
A l g o r i t h m f o r C o n s t r u c t i n g an Optimal K - P a r t i t i o n w i t h the Sum of S q u a r e s

A b o u t the M e a n Diameter Function
I n p u t : X « N - t u p l e o f r e a l s , K = no. of s e t s in the p a r t i t i o n .
O u t p u t : A r r a y G s u c h t h a t G ( h , j ) = minimum cost f o r p a r t i t i o n i n g x j X 2 . » X j i n t o h sets
c o n s e c u t i v e elements. A r r a y H s u c h that H(h,j) = s u b s c r i p t k f o r t h e h'th
s e t o f t h e o p t i m a l p a r t i t i o n of x^x^y-y
T i m e : 0 ( K N ) , S t o r a g e : 0<KN)
2
/ Define statement function;

D(I,J) = (SUMSQ(J> - S U M S Q ( I - l ) ) - (SUM(J) - S U M ( I - l ) ) 2
/ ( J - I + 1)
/ Compute partial sums and sums of squares of X(I\)
S U M ( O ) <- S U M S Q ( O ) <- 0;
F o r J <- 1 t h r u N d o
begin
S U M ( J ) <- S U M ( J - l ) + X ( J ) ;
S U M S Q ( J ) <- S U M S Q ( J - l ) + X ( J ) * X ( J ) ;
G(1,J) «- D(1,J); H(1,J)«-J;
end;
/ Optimization;
F o r L <- 2 t h r u K d o / L = no. of sets in partition;
F o r J «- L t h r u N d o
F o r I <- L - l t h r u J - l d o
begin
C O S T <- G ( L - 1 , I ) + D ( I + 1 , J ) ;
begin
MINCOST 4- COST;
G ( L , J ) <- C O S T ;
H(L,J) 4 - I;
end
end;
W e c a n f u r t h e r r e d u c e t h e s t o r a g e r e q u i r e d b y this a l g o r i t h m to 0 ( N ) . T o f i n d t h e cost o f the

optimal partition in 0 ( N ) s t o r a g e and 0 ( K N ) time is f a i r l y s i m p l e , b u t t h e r e c o v e r y
2
of the
o p t i m a l p a r t i t i o n i t s e l f ( w i t h i n t h e same time and s p a c e b o u n d s ) is m o r e d i f f i c u l t . First, we
c o n s i d e r f i n d i n g t h e cost o f t h e optimal p a r t i t i o n in 0 ( N ) s t o r a g e a n d 0 ( K N ) t i m e .
2
The only trick is t o eliminate unnecessary storage that is b e i n g used. The diameter
f u n c t i o n is a l r e a d y c o m p u t e d in 0 ( N ) s t o r a g e . O n l y the a r r a y G c o s t s 0 ( K N ) s t o r a g e . (Forget
a b o u t H f o r n o w s i n c e o n l y t h e optimal cost is b e i n g c o m p u t e d . ) B u t in t h e i n n e r t w o l o o p s of
t h e a b o v e a l g o r i t h m o n l y t h e v e c t o r s G ( l - 1 , * ) and G(L,*) are r e f e r e n c e d . T h a t i s , at s t a g e L
o n l y t h e v e c t o r s G ( L - 1 , * ) a n d G(L,*) a r e n e e d e d . T h u s , the K x N a r r a y G c a n b e r e p l a c e d by
t w o v e c t o r s o f l e n g t h N a n d t h e c o s t of t h e optimal p a r t i t i o n c a n b e c o m p u t e d in O ( N ) s t o r a g e
PAGE 30
a n d 0 ( K I \ n time.
The recovery o f t h e o p t i m a l p a r t i t i o n itself, r a t h e r than j u s t t h e c o s t , in 0 ( N ) s t o r a g e and

0<KN ) 2
time can be accomplished by combining divide - and - conquer with dynamic
programming. The method is similar to the a l g o r i t h m of H i r s c h b e r g [29] that s o l v e s the
l o n g e s t c o m m o n s u b s e q u e n c e p r o b l e m ( t o be d e s c r i b e d later in t h i s p a p e r ) in 0 ( N ) storage
a n d 0 ( N ) ti me.
2
R e c a l l t h a t t h e s t a t e S^j is i n t e r p r e t e d to mean " x ^ X 2 » . X j a r e p a r t i t i o n e d i n t o h s e t s of

consecutive e l e m e n t s . " D e f i n e a n e w set of s t a t e s S'^j that a r e i n t e r p r e t e d "xjXj + j . . . xjnj are
p a r t i t i o n e d i n t o h s e t s of c o n s e c u t i v e e l e m e n t s . " Let G'(h,i) = the minimum c o s t t o r e a c h s t a t e
S'hj. Also, assume for convenience that K is a power of 2. Here is how the
d i v i d e - a n d - c o n q u e r algorithm works:
1. C o m p u t e G ( K / 2 , i ) f o r all i in 0 ( K / 2 * N ) time and c o m p u t e G'(K/2,j) f o r all j in

2
0 ( K / 2 * N ) time f o r 0 ( K N ) total time.

2 2
2 . F i n d t h e s u b s c r i p t i that minimizes G(K/2,i) + G'(K/2,i + i ) . T h i s s u b s c r i p t i e q u a l s

t h e I « / 2 ' t h e K + i - t u p l e (Iq,I|,I2,...I|<) that r e p r e s e n t s t h e optimal p a r t i t i o n o f
n
the N elements.
3. N o w f i n d a n d I3K/4 b y a p p l y i n g s t e p s 1. and 2. to the r e s p e c t i v e

subproblems. ( T h i s is the d i v i d e s t e p - s o l v i n g t w o s u b p r o b l e m s o f half t h e
size.)
4. R e p e a t t h e d i v i d e s t e p of s t e p 3. T h a t is, find I|</g, l3K/8> *5K/8> * *7K/8> * * a n c e c
A f t e r l o g K i t e r a t i o n s w e will h a v e d e t e r m i n e d the e n t i r e optimal p a r t i t i o n .
T h e t o t a l t i m e t o r e c o v e r t h e optimal p a r t i t i o n may at first a p p e a r to be 0 ( K N * l o g K ) . 2

With
careful a n a l y s i s , h o w e v e r , w e can s h o w it to be o n l y 0 ( K N ) . 2
Let U(K,N) be the time to
c o n s t r u c t t h e o p t i m a l K - p a r t i t i o n of N elements and let T ( K , N ) = 0 ( K N ) b e t h e time t o f i n d t h e
2
cost of the optimal K-partitions f o r all the c l u s t e r s (1,N), ( 1 , N - 1 ) , ( l , N - 2 ) , e t c . The above

d i v i d e - a n d - c o n q u e r a l g o r i t h m leads to the f o l l o w i n g e q u a t i o n s :
PAGE 31
T(I,J) = I J 2
logK 2 ' 1 1
U(K,N> = I I [ 2 T
( | . W ' " Q-W->) l +
°( iM2'-'
I
" ^j-Dk/a-)]
l o g K [" 2 j l
"J
i=l L J = l 2' J
logK
< S |"o<> N +
-^T N1- 2
0(KN )2
T h e time U(K,N) d e p e n d s s o m e w h a t o n t h e optimal p a r t i t i o n (Iq,Ij,...I )

k b u t in t h e w o r s t case
t h e t o t a l t i m e is still 0 < K N ) .
2
PAGE 32
7 . O p t i m a l Matching Problems
One of the major differences between optimal matching problems and optimal
p a r e n t h e s i z a t i o n o r p a r t i t i o n p r o b l e m s is that i n s t e a d of j u s t o n e s t r i n g of e l e m e n t s , t w o (or
m o r e ) are involved. T h e a u t h o r K n o w s of t w o p r o b l e m s in this class. The first, and most
i m p o r t a n t , is t h e l o n g e s t c o m m o n s u b s e q u e n c e ( L C S ) p r o b l e m ( a n d its v a r i a t i o n s ) , f o r which
DP produces a g o o d (if n o t o p t i m a l ) algorithm. T h e s e c o n d p r o b l e m is t h e g a m e o f N I M , f o r
w h i c h a m u c h f a s t e r a l g o r i t h m is k n o w n and thus will not be d e s c r i b e d f u r t h e r [ 2 1 ] . We will
n o w describe the LCS problem.
L e t Y = y 1Y2 • • • N y a n c
* ^ - z
i 2 • • •
z
be * w 0 s t r
' S
n s
°* l e n g t h N w h o s e e l e m e n t s y- f and
Zj a r e c h o s e n f r o m a finite alphabet ALPH. ( T h e s t r i n g s Y and Z a r e c h o s e n t o b e t h e s a m e
length only for c o n v e n i e n c e in p r e s e n t a t i o n . ) A s t r i n g B is a subsequence o f Y if B c a n be
o b t a i n e d b y d e l e t i n g c h a r a c t e r s f r o m Y. F o r e x a m p l e , if Y = "dyanamic p r o g r a m m i n g " , t h e n B =
" m a i n " is a s u b s e q u e n c e of Y. T h e l o n g e s t common s u b s e q u e n c e p r o b l e m ( f o r t h e t w o strings
Y a n d Z ) is t o f i n d t h e ( n o t n e c e s s a r i l y u n i q u e ) l o n g e s t s t r i n g B that is a s u b s e q u e n c e o f b o t h
Y a n d Z.
T h e L C S p r o b l e m c a n b e s o l v e d in 0 ( N ) time and s p a c e [ 2 9 ] .
2
The string editing p r o b l e m , a
g e n e r a l i z a t i o n o f t h e L C S p r o b l e m , is to d e t e r m i n e the minimum n u m b e r of c h a n g e s , i n s e r t i o n s ,
o r d e l e t i o n s t o t r a n s f o r m a s t r i n g Y to a s t r i n g Z. T h i s p r o b l e m c a n also b e s o l v e d in 0(N ) 2
time and space [73]. T h e r e h a v e b e e n s e v e r a l e f f o r t s to i m p r o v e t h e s e a l g o r i t h m s and/or

'-how lower bounds for the problems. Hirschberg [29] reduced the s p a c e for the LCS
problem from 0(N ) 2
to 0(N) while maintaining the b o u n d of 0 ( N ) 2
time. Paterson [55]
produced a n 0 ( N l o g l o g N / l o g N ) time algorithm f o r the LCS p r o b l e m [ 3 0 ] ) , w h i c h M a s e k
2
and
Paterson [52] have recently i m p r o v e d to 0 ( N / l o g N )
2
time w i t h t h e i r s o l u t i o n t o t h e string
editing problem. H u n t a n d S z y m a n s k i [ 3 5 ] h a v e c o n s t r u c t e d an a l g o r i t h m t h a t s p e n d s 0 ( ( R + N )
l o g N ) time w h e r e R is t h e total n u m b e r of pairs of c h a r a c t e r s at w h i c h t h e t w o sequences
match. T h u s , R m a y b e as l a r g e as N 2
but if it is small, t h e n t h e a l g o r i t h m t a k e s o n l y 0(N
l o g N ) time. W o n g a n d C h a n d r a [ 7 4 ] s h o w e d an fl(N )
2
l o w e r (time) b o u n d of t h e s t r i n g editing
problem for a model of computation allowing o n l y equal / u n e q u a l c o m p a r i s o n s between
e l e m e n t s of strings. A h o , H i r s c h b e r g , and Ullman [ 1 ] s h o w e d t h e ft(N ) 2
lower bound for the
LCS problem (with only equal / unequal comparisons). Hirschberg [ 3 1 ] has p r o d u c e d an
O(NlogN) lower bound for the LCS problem where less-than-or-equal comparisons are
allowed. Hirschberg [30] provides an o v e r v i e w of the r e s u l t s for the LCS and related
p r o b l e m s a n d S e l k o w [ 6 5 ] d e s c r i b e s an e x t e n s i o n of t h e s e s t r i n g p r o b l e m s t o t r e e s . Itakura
[41] gives an interesting a p p l i c a t i o n of the L C S p r o b l e m to s p e e c h r e c o g n i t i o n . Fitch and
M a r g o l i a s h [ 1 8 ] a p p l y t h e s t r i n g e d i t i n g p r o b l e m to d e t e r m i n a t i o n o f t h e m u t a t i o n d i s t a n c e of
t h e D N A of s e v e r a l species.
PAGE 33
W e w i l l n o w p r e s e n t t h e 0 ( N ) time and 0 ( N ) s p a c e a l g o r i t h m f o r t h e L C S p r o b l e m .
2 2
The
t w o s t r i n g s a r e , as b e f o r e , Y and Z, e a c h c o n s i s t i n g of N c h a r a c t e r s f r o m t h e a l p h a b e t ALPH.
T h e D D P D - ( A , S , P, f ) is d e f i n e d b y
A = { ajj } w h e r e aj j c o m p a r e s y j and Z j f o r e q u a l i t y ,
S = { e } U { a j | i,j < N } U { x a j j a j I x « S and i<k and ]<L }, /

} k The subscripts of
the a — must be strictly increasing.;
P = { 2 * N - t u p l e s o f e l e m e n t s in the a l p h a b e t A L P H }, and
f ( e ) = 0 , f ( a j j ) = C ( y j , Z j ) , f ( x a j j a ) = f(xajj> + C ( y j , Z j ) f o r x € A L P H * a n d i<k a n d j<Z,

w
w h e r e C ( y j , z j > = 1 if y j = Z j and C ( y j , Z j ) = 0 o t h e r w i s e .
Note that the cost function f is w r i t t e n s o that t h e o b j e c t ' is t o maximize f rather than
m i n i m i z e it. L e t t h e s t a t e s o f t h e SDP b e w r i t t e n Sjj and be i n t e r p r e t e d " ( y i , y 2 > — Y j ) h a s been
compared with (Z|Z2-..Zj) by a sequence ending in a j j . " T h u s , t h e S D P is W = ( V , P, h , k )
where
V = ( A , Q , q , F, X ) w h e r e
Q
A « s a m e as f o r t h e DDP,
Q = { Sjj | i,j = 1,2,...N } u { S 0 0 },
^ 0 oo> = s
F = { S N N },
X<Sjj,3 K Z > = S k / for k>i, Z>j,
h ( r , S j j , a ) = r + C ( y , z j ) f o r i<k, j<Z, and

k i k
k(p) = 0.
T h e S D P is m o n o t o n e b y a d d i t i v i t y of the cost f u n c t i o n . Thus, the functional equations are
G(S 0 0 ) = 0 , and
G ( S
ii> =
{ ( S ^ ) | K + C
^i' i>-
z
This gives an 0 ( N ) 4
time a l g o r i t h m . H o w e v e r , a monotone p r o p e r t y of the f u n c t i o n G c a n
minimize the search for max in the e q u a t i o n a b o v e . Since t h e s u b s c r i p t s of the ajj are
r e q u i r e d t o b e s t r i c t l y i n c r e a s i n g in e a c h legal s e q u e n c e of c o m p a r i s o n s , it f o l l o w s t h a t
If i < k and j < /, t h e n G ( S j j ) < G(S />. k
T h u s , t o f i n d G ( S j j ) it is s u f f i c i e n t to c o m p u t e
PAGE 34
G( S j j ) = max ( [™* * S K ( h l ) ) ] , [ ™ G(S , ( 1 ) > k )] ) + C(y.^.
T h e t i m e is n o w r e d u c e d t o O ( N ^ ) but f u r t h e r i m p r o v e m e n t is p o s s i b l e . T h e max(k<i) and the

max(k<j) can be propagated throughout the c o m p u t a t i o n of all G ( S j j ) s o t h a t o n l y constant
time per state Sjj is required. One way to do this is to c r e a t e an array MAXG(iJ) to
propagate these max's. Let G(i,j) = G(Sjj). When G(i,j) is d e f i n e d , p r o p a g a t e the max's
t h r o u g h the steps
G ( i , j ) <- M A X G ( i - l , j - l ) + C ( y j , z : )
M A X G ( i J ) <- G ( i , j )
M A X G ( i + l , j ) <- max( M A X G ( i + l , j ) , G(i,j) )
M A X G ( i , j + l ) <- max( M A X G ( i , j + l ) , G(i,j) ).
A n o t h e r w a y t o a c c o m p l i s h t h e p r o p a g a t i o n is to make a small c h a n g e in t h e d e f i n i t i o n o f G
so t h a t it a c c o m p l i s h e s t h e same task as M A X G , thus s a v i n g o n e a r r a y [ 2 9 ] . T h i s is w h a t is
d o n e in t h e f o l l o w i n g algorithm.
A l g o r i t h m f o r t h e L o n g e s t C o m m o n S u b s e q u e n c e of T w o S t r i n g s
I n p u t : i n t e g e r N > 0, a r r a y s Y [ 1 : N ] , Z[1:N]
O u t p u t : A r r a y G s u c h t h a t G ( I , J ) = l e n g t h of longest common s u b s e q u e n c e o f (yj,y2>—Yj)
and (zj,Z2,.»Zj). ^rray H s u c h that H(I,J) = p o i n t e r to e i t h e r ( I , J - 1 ) o r
( I - 1 , J ) , w h i c h e v e r has the l a r g e s t v a l u e of G.
Time: 0(N >, Storage: 0(N )
2 2
/ initialization;
F o r I <r- 0 t h r u N d o
F o r J 4- 0 t h r u N do
G ( I , J ) <- H(I,J) 4 - 0;
/ optimization;
F o r I <- 1 t h r u N do
F o r J «- 1 t h r u N do
begin
If Y ( I ) = Z ( J ) t h e n
G(I,J) G(I-1,J-1) + 1
else
G ( I , J ) <- max( G ( I , J - 1 ) , G ( I - 1 , J ) );
If G ( I , J - 1 ) > G ( I - 1 , J ) t h e n
H ( I , J ) <- ( I , J - 1 )
else
H ( I , J ) «- ( I - 1 . J )
end;
PAGE 35
G B A B B A H B A B B A
A 0 1 1 1 1 A 0,0 0,0 1,2 1,3 1,4
B A B B A B 1 1 2 2 2 B 0,0 2,1 2,2 2,3 2,4
\\\ I
A B A B A A 1 2 2 2 3 A 2,1 3,1 3,2 3,3 3,4
B 1 2 3 3 3 B 3,1 3,2 4,2 4,3 4,4
1 2 3 3 4 A 4,1 4,2 4,3 5,3 5,4

A
F i g u r e 7 - 1 : T h e m a t r i c e s G and H p r o d u c e d f o r the s t r i n g s A B A B A a n d B A B B A .
PAGE 36
8 . ' H a r d Combinatorial Problems

1
T h e p r o b l e m s d e s c r i b e d in t h e p r e v i o u s s e c t i o n s can all be s o l v e d in 0 ( N ) , O ( N ^ ) , o r
2
more
g e n e r a l l y t i m e t h a t is b o u n d e d a b o v e b y a p o l y n o m i a l in N. T h e p r o b l e m s in t h i s s e c t i o n h a v e
n e v e r b e e n s o l v e d in l e s s t h a n 0 ( 2 ^ ) time. M a n y of t h e s e p r o b l e m s a r e N P - C o m p l e t e [ 4 3 , 2 ] .
( T h e o t h e r p r o b l e m s a p p e a r to be at least this hard.) F o r p r o b l e m s w h e r e a n a i v e algorithm
c o s t s 0 ( N ! ) t i m e , D P c a n t y p i c a l l y b r i n g an i m p r o v e m e n t to 0 ( P ( N ) * 2 ) time w h e r e P ( N ) is a
N
p o l y n o m i a l in N. F o r e x a m p l e , the t r a v e l i n g s a l e s p e r s o n p r o b l e m ( T S P ) r u n s in t i m e 0(N^2'v
[28, 6]. An improved DP algorithm for the TSP and knapsack problems that utilizes
b r a n c h - a n d - b o u n d is d e s c r i b e d b y M o r i n [ 5 3 ] . C o n s t r u c t i n g optimal d e c i s i o n t r e e s f r o m a s e t
o f y e s / n o t e s t s is N P - c o m p l e t e [ 3 7 ] and the DP algorithm r u n s in 0 ( 3 ^ ) time a n d s p a c e [ 6 4 ] .
( L e w [ 5 0 ] d e s c r i b e s a s o l u t i o n to a g e n e r a l i z a t i o n of this p r o b l e m . ) B a y e s [ 4 ] d e s c r i b e s a d i s k
f i l e p l a c e m e n t p r o b l e m - a v a r i a t i o n of the module placement p r o b l e m in K a r p a n d H e l d [ 4 4 ] -
f o r w h i c h a D P a l g o r i t h m s p e n d s 0 ( 2 ) time and 0 ( C ( N , N / 2 ) ) s t o r a g e .
N
Kohler and Steiglitz [49]
and Kohler [48] describe DP ( a n d b r a n c h - a n d - b o u n d ) s o l u t i o n s of N P - C o m p l e t e scheduling
problems. G i l m o r e [ 2 3 ] s h o w s interconnections b e t w e e n cutting stock, linear programming,
k n a p s a c k i n g , d y n a m i c p r o g r a m m i n g , and i n t e g e r p r o g r a m m i n g . Shapiro [67] s h o w s h o w the
knapsack problem can be reduced to a (huge) shortest path problem, and Lloyd [51]
d e s c r i b e s a D P a l g o r i t h m f o r t h e minimum w e i g h t t r i a n g u l a t i o n p r o b l e m .
There a r e f a r t o o m a n y p r o b l e m s to d e s c r i b e them all in full d e t a i l . But to g i v e s o m e of

t h e f l a v o r o f w h a t a DP a l g o r i t h m f o r this t y p e of p r o b l e m c a n b e like, w e w i l l n o w describe
t h e t r a v e l i n g salesperson (TSP) problem.
The s a l e s p e r s o n must t r a v e l to e a c h of N cities, s t a r t i n g and e n d i n g at c i t y n u m b e r one.

The distances Djj between the cities are given. The p r o b l e m is to f i n d t h e route with
minimum total distance traveled. T h e n a i v e s o l u t i o n is to s i m p l y c o m p a r e all ( N - l ) ! possible
r o u t e s a n d c h o o s e t h e minimum. Use of a DP a l g o r i t h m , h o w e v e r , c a n r e d u c e t h i s t o CXN^^)
time.
The f i r s t s t e p in c o n s t r u c t i o n of the DP algorithm is to d e f i n e t h e D D P D = ( A , S , P, f )

where
A = { aj } w h e r e aj means "go to c i t y i",
S = { p e r m u t a t i o n s o f t h e aj e n d i n g in a^ },
P =* { m a t r i c e s Djj | Djj > 0, Vi,j }, and
f ( e ) = 0, f ( ) = D
a j u> f ( x a j a j ) = f ( x a j ) + Djj.
T o c o n s t r u c t t h e S D P w e must f i r s t d e f i n e the s t a t e s . It is t e m p t i n g to let t h e s t a t e s b e Sj

w h e r e Sj is i n t e r p r e t e d "at c i t y i". B u t since w e must t r a v e l to all o f the N c i t i e s , n o t j u s t t h e
PAGE 37
c l o s e s t o n e , this won't w o r k . I n s t e a d , w e must c o n s i d e r the cities a l r e a d y v i s i t e d b y t h e t i m e

the salesperson r e a c h e s c i t y i. Let C = { cities } = { 1, 2, 3 , . . . N }. If M is a s u b s e t o f C
r e p r e s e n t i n g t h e c i t i e s t h a t h a v e b e e n v i s i t e d b y the s a l e s p e r s o n a n d i £ M is t h e c i t y where
t h e s a l e s p e r s o n is c u r r e n t l y l o c a t e d , t h e n t h e s t a t e is Sj^jj. W i t h t h i s d e f i n i t i o n t h e S D P is
W - ( Z, P, h, k > w h e r e
Z = ( A , Q , q , F, \ ) w h e r e
Q
Q - { S M > i | M c C and i (• M },
^0 = S
{1},1'
F
- « s , i }, c
x ( s
M,i' j a ) = S
Mu{j},j'
h(r,S M > j , a j ) - r + D y , and
k(p) - 0.
M o n o t o n i c i t y o f t h e S D P f o l l o w s f r o m the a d d i t i v i t y of the cost f u n c t i o n . T h u s , the functional
equations are
G(S { 1 } ) » 0, and
G(S ,) = j
M
m n
G(S .
M { i } i J ) • D j f
H e r e is a n a l g o r i t h m t o s o l v e t h e functional e q u a t i o n s :
PAGE 38
A l g o r i t h m for T r a v e l i n g Salesperson Problem
I n p u t : " D i s t a n c e " m a t r i x D ( I , J ) , VI,J=1,...N

O u t p u t : ( L e t M b e a s u b s e t o f {1,2,...N}.) A r r a y G s u c h that G(M,I) = minimum c o s t
o f t r a v e l i n g t o t h e c i t i e s of M and ending at c i t y I. A r r a y H s u c h t h a t
H ( M , I ) = last c i t y r e a c h e d b e f o r e c i t y I in t h e optimal s u b t o u r c o r r e s p o n d i n g
to G(M,I).
T i m e : 0 ( N 2 ) , S t o r a g e : 0 ( N 2 ) ( T h e a n a l y s i s is in [ 2 8 ] . )
2 N N
/ MINIMIZE finds the best subtour for (M,ti)

p r o c e d u r e M I N I M I Z E * G , H, D, M, I )
begin
M I N C O S T « - oo;
F o r all J < M - { I } d o
begin
C O S T <- G ( M - { I } , J ) + D(J,I);
begin
G ( M , I ) <- M I N C O S T <- C O S T ;
H ( M , I ) *- J ;
end
end
end;
/ main routine;
! Initialization;
C « - { 2, 3 , 4, . . . N };
F o r I *- 1 t h r u N d o
G ( { I } , I ) «- D(1,I);
/ Optimization;
F o r K « - 2 t h r u N - l d o / K = size of subsets;
F o r all M c C s u c h that |M| = K do
F o r all I c M d o
MINIMIZE(G,H,DMI)j
MINIMIZE(G,HACu {!},!);
PAGE 39
9. C o n c l u s i o n
D y n a m i c P r o g r a m m i n g is a u s e f u l p r o b l e m s o l v i n g m e t h o d f o r many p r o b l e m s in c o m p u t e r
science. I n t h i s p a p e r w e h a v e d e s c r i b e d p r o b l e m s in s e v e r a l c a t e g o r i e s o f D P : Shortest
Path Problems, Optimal Parenthesization Problems, Optimal Partition Problems, Optimal
Matching P r o b l e m s , and ' H a r d ' Combinatorial Problems. T h e a p p r o a c h t a k e n in t h i s paper
t o w a r d s o l u t i o n o f t h e s e p r o b l e m s is a modification of the t r e a t m e n t in K a r p a n d H e l d [ 4 4 ] .
E a c h a n a l y s i s o f a p r o b l e m p r o c e e d s in the f o l l o w i n g s t e p s :
1. P r o d u c e t h e d i s c r e t e d e c i s i o n p r o c e s s f o r the p r o b l e m ,
2 . D e c i d e w h a t a r e t h e r e l e v a n t states and c o n s t r u c t the c o r r e s p o n d i n g sequential

decision process,
3. Check the monotonicity conditions and (if satisfied) produce the functional
equations, and
4. F r o m t h e f u n c t i o n a l e q u a t i o n s c o n s t r u c t a d y n a m i c p r o g r a m m i n g algorithm that
solves the problem.
In this paper we have also p r e s e n t e d n e w p r o b l e m s and r e s u l t s i n c l u d i n g optimizing the

s e a r c h f o r N o r d e r s t a t i s t i c s and a more s p a c e e f f i c i e n t v e r s i o n of F i s h e r ' s a l g o r i t h m f o r an
optimal partition.
Acknowledgments
Jon Bentley s u g g e s t e d d y n a m i c p r o g r a m m i n g as t h e topic f o r my a r e a q u a l i f i e r and has

g i v e n n u m e r o u s h e l p f u l c o m m e n t s that h a v e i m p r o v e d the q u a l i t y o f t h i s p a p e r .
PAGE 40
References
[I] A . V . A h o , D.S. H i r s c h b e r g , and J.D. Ullman, Bounds on the Complexity of the Longest
Common Subsequence Problem, J . A s s o c . C o m p . Mach., 23 ( 1 9 7 6 ) , p p . 1 - 1 2 .
[2] A . V . A h o , J . H o p c r o f t , a n d J.D. Ullman, The Design and Analysis of Computer

Algorithms, A d d i s o n - W e s l e y , Reading, Mass., 1974. C o n s t r u c t i n g o p t i m a l b i n a r y
s e a r c h t r e e , pp. 119-123..
[3] J . K . B a k e r , The DRAGON System - An Overview, I E E E T r a n s . A c o u s t . S p e e c h S i g n a l

P r o c e s s . , A S S P - 2 3 (1975), pp. 24-29.
[4] A . J . B a y e s , The Optimised Placing of Files on Disk Using Dynamic Programming,

O p t i m i z a t i o n , R.S. A n d e r s s e n , L S . J e n n i n g s , and D M R y a n , e d . , U n i v e r s i t y o f
Q u e e n s l a n d P r e s s , St. L u c i a , Q u e e n s l a n d , 1972, p p . 1 8 2 - 1 8 6 .
[5] R. B e l l m a n , Dynamic Programming, Princeton University Press, Princeton, N e w

J e r s e y , 1957.
[6] R. B e l l m a n , On a routing problem, Q u a r t . A p p l . Math., 16 ( 1 9 5 8 ) , p p . 8 7 - 9 0 .
[7] M. B l u m , R.W. F l o y d , V . R . P r a t t , R.L. R i v e s t , and R.E. T a r j a n , Time bounds for selection,

J . C o m p u t . S y s t e m Sci., 7 ( 1 9 7 2 ) , p p . 4 4 8 - 4 6 1 .
[8] D.W. B o y d , Long Division: An Example of Dynamic Programming, Amer. Inst. Indust.
E n g i n . T r a n s . , volume (1974), pp. 365-366.
[9] C . C a f f o r i o a n d F. R o c c a , Methods for Measuring Small Displacements of Television

Images, IEEE Trans. Information T h e o r y , I T - 2 2 (1976), pp. 5 7 3 - 5 7 9 .
[10] F.Y. C h i n , An 0(N) Algorithm for Determining a Near-Optimal Computation Order of

Matrix Chain Products, Comm. A C M , 21 ( 1 9 7 8 ) , p p . 5 4 4 - 5 4 9 .
[II] G . B . D a n t z i g , Linear Programming and Extensions, Princeton University Press,

P r i n c e t o n , N e w J e r s e y , 1963.
[12] E.W. D i j k s t r a , A Note on Two Problems in Connection With Graphs, Numer. Math., 1
(1959), pp. 269-271.
[13] S . E . D r e y f u s and A . M . L a w , The Art and Theory of Dynamic Programming, Academic

P r e s s , N e w Y o r k , 1977.
[14] G . T . D u n c a n , Optimal Diagnostic Questionnaires, O p e r a t i o n s Res., 23 ( 1 9 7 5 ) , p p .

22-32.
[15] J . E a r l y , An efficient context-free parsing algorithm, Comm. A C M , 13 ( 1 9 7 0 ) , p p .

94-102.
[16] M.A. F i s c h l e r and R.A. E l s c h l a g e r , The Representation and Matching of Pictorial

Structures, IEEE Trans. Computers, C - 2 2 (1973), pp. 67-92.
[17] W . D . F i s h e r , On grouping for maximum homogeneity, J . Amer. Statist. Assoc., 5 3

(1958), pp. 789-798.
[18] W . M . F i t c h a n d E. M a r g o l i a s h , Construction of Phylogenetic Trees, S c i e n c e , 155 ( 1 9 6 7 ) ,

pp. 279-284.
PAGE 41
[19] G . D . F o r n e y J r . , The Viterbi Algorithm, P r o c . I E E E , 61 ( 1 9 7 3 ) , p p . 2 6 8 - 2 7 8 .
[20] M.L. F r e d m a n , On the Decision Tree Complexity of the Shortest Path Problems, 1 6 t h
A n n u a l S y m p o s i u m o n F o u n d a t i o n s of C o m p u t e r S c i e n c e , I E E E , L o n g B e a c h , C A . , O c t . ,
• 1975, p p . 9 8 - 9 9 .
[21] M. G a r d n e r , Concerning the game of nim and its mathematical analysis* Mathematical
G a m e s s e c t i o n of S c i e n t i f i c A m e r i c a n , 198 ( 1 9 5 8 ) , p p . 1 0 4 - 1 1 1 .
[22] E.N. G i l b e r t a n d E.F. M o o r e , Variable length encodings, Bell S y s t e m T e c h . J . , 3 8 ( 1 9 5 9 ) ,

pp. 933-968.
[23] P. G i l m o r e , Catting stock, Linear Programming, knapsacking, Dynamic Programming*

Integer Programming: some interconnections, T e c h . Rep. R C 6 5 2 8 ( 2 8 1 8 8 ) , I B M
T J . W a t s o n R e s e a r c h C e n t e r , Y o r k t o w n Heights, N e w Y o r k , S e p t 1 9 7 7 . 3 3 p a g e s .
[24] S . L . G r a h a m , M.A. H a r r i s o n , and W.L. R u z z o , On Line Context Free Language

Recognition in Less Than Cubic Time, 8 t h A n n u a l A C M S y m p o s i u m o n T h e o r y o f
C o m p u t i n g , A s s o c . C o m p u t . Mach. Special I n t e r e s t G r o u p o n A u t o m a t a a n d C o m p u t i n g
T h e o r y , N e w Y o r k , M a y , 1976, p p . 1 1 2 - 1 2 0 .
[25] I I. G u i b a s , H.T. K u n g , and C D . T h o m p s o n , Direct VLSI Implementation of

Combinatorial Algorithms, C o n f e r e n c e o n V e r y L a r g e Scale I n t e g r a t i o n , I n d u s t r i a l
A s s o c i a t e s , C a l i f o r n i a I n s t i t u t e of T e c h n o l o g y , J a n . , 1979,
[26] F. H a r a r y , Graph Theory, A d d i s o n - W e s l e y Publishing C o m p a n y , R e a d i n g , M a s s . , 1 9 6 9 .
[27] J . A . H a r t i g a n , Clustering Algorithms, John Wiley & Sons, New York, 1975. Fisher
Algorithm on pages 130-140.
[28] M. H e l d a n d R.M. K a r p , A Dynamic Programming Approach to Sequencing Problems,

J . S I A M , 10 ( 1 9 6 2 ) , p p . 1 9 6 - 2 1 0 .
[29] D.S. H i r s c h b e r g , A Linear Space Algorithm for Computing Maximal Common

Subsequences, C o m m . A C M , 18 ( 1 9 7 5 ) , p p . 3 4 1 - 3 4 3 .
[30] D.S. H i r s c h b e r g , Complexity of Common Subsequence Problems, Springer-Verlag, New

Y o r k , 1977. L e c t u r e N o t e s in C o m p u t e r Science,No. 5 6 , F u n d a m e n t a l s o f C o m p u t a t i o n
Theory.
[31] D.S. H i r s c h b e r g , An information - theoretic lower bound for the longest common
subsequence problem, I n f o r m a t i o n P r o c e s s i n g Lett., 7 ( 1 9 7 8 ) , p p . 4 0 - 4 1 .
[32] J . E . H o p c r o f t a n d J . D . Ullman, Formal Languages and Their Relation to Automata,

A d d i s o n - W e s l e y , R e a d i n g , Mass., 1969.
[33] T . C . H u , A new proof of the T - C algorithm, S I A M J . A p p l . M a t h . , 25 ( 1 9 7 3 ) , p p .

83-94.
[34] D.A. H u f f m a n , A Method for the Construction of Minimum Redundancy Codes, Proc.
Inst. Radio Engin., 40 (1952), pp. 1098-1101.
[35] J . W . H u n t a n d T . G . S z y m a n s k i , A Fast Algorithm for Computing Longest Common

Subsequences, C o m m . A C M , 20 ( 1 9 7 7 ) , p p . 3 5 0 - 3 5 3 .
PAGE 42
[36] T . C . H u a n d A . C . T u c k e r , Optimum Binary Search Trees, S I A M J . A p p l . M a t h . , 21

(1971), pp. 514-532.
[37] L. H y a f i l a n d R.L. R i v e s t , Constructing Optimal Binary Decision Trees is NP-Complete,

I n f o r m a t i o n P r o c e s s i n g Lett., 5 (1976), pp. 15-17.
[38] T . I b a r a k i , Minimal Representations of Some Classes of Dynamic Programming,

I n f o r m a t i o n a n d C o n t r o l , 27 ( 1 9 7 5 ) , p p . 2 8 9 - 3 2 8 .
[39] T . I b a r a k i , Branch-and-bound procedure and state-space representation of

combinatorial optimization problems. I n f o r m a t i o n and C o n t r o l , 3 6 ( 1 9 7 8 ) , p p . 1 - 2 7 .
[40] A . I t a i , Optimal alphabetic trees, S I A M J . C o m p u t . , 5 ( 1 9 7 6 ) , p p . 9 - 1 8 .
[41] F. I t a k u r a , Minimum Prediction Residual Principle Applied to Speech Recognition,

I E E E T r a n s . A c o u s t . S p e e c h Signal P r o c e s s . , A S S P - 2 3 ( 1 9 7 5 ) , p p . 6 7 - 7 2 .
[42] D.B. J o h n s o n , Algorithms for shortest paths, Ph.D. T h e s i s , D e p t . of C o m p u t e r S c i e n c e ,

C o r n e l l U n i v e r s i t y , I t h a c a , N e w Y o r k , 1973. C i t e d in A h o , H o p c r o f t , a n d U l l m a n .
[43] R.M. K a r p , Reducibility among combinatorial problems, C o m p l e x i t y of C o m p u t e r

C o m p u t a t i o n s , M i l l e r a n d T h a t c h e r , e d . , Plenum P r e s s , N e w Y o r k , 1 9 7 2 , p p . 8 5 - 1 0 4 .
[44] R.M. K a r p a n d M. H e l d , Finite-State Processes and Dynamic Programming, SIAM

J . A p p l i e d M a t h , 15 ( 1 9 6 7 ) , p p . 6 9 3 - 7 1 8 .
[45] T . K a s a m i a n d K. T o r i i , A syntax-analysis procedure for unambiguous context-free

grammars, J . A s s o c . C o m p . Mach., 16 ( 1 9 6 9 ) , p p . 4 2 3 - 4 3 1 .
[46] D.E. K n u t h , Optimal Binary Search Trees, A c t a Informat., 1 ( 1 9 7 1 ) , p p . 1 4 - 2 5 .
[47] D.E. K n u t h , Sorting and Searching, A d d i s o n - W e s l e y , R e a d i n g , Mass., 1 9 7 3 . Algorithm

f o r b i n a r y s e a r c h t r e e in S e c t i o n 6.2.2.
[48] W . H . K o h l e r , Computational Experience with Efficient Exact and Approximate

Algorithms for an NP-Complete Scheduling Problem, P r o c e e d i n g s o f t h e N i n t h H a w a i i
I n t e r n a t i o n a l C o n f e r e n c e o n S y s t e m S c i e n c e s , U n i v e r s i t y of H a w a i i , M a n o a , 1 9 7 6 , p p .
11/- 1 1 o
[49] W . H . K o h l e r a n d K. S t e i g l i t z , Enumerative and Iterative Computational Approaches,

C o m p u t e r a n d J o b - S h o p S c h e d u l i n g T h e o r y , E.G. C o f f m a n , J r . , e d . , J o h n W i l e y & S o n s ,
N e w Y o r k , 1976, pp. 2 2 9 - 2 8 7 .
[50] A . L e w , Optimal Conversion of Extended-Entry Decision Tables with General Cost

Criteria, C o m m . A C M , 21 ( 1 9 7 8 ) , p p . 2 6 9 - 2 7 9 .
[51] E.L. L l o y d , On triangulations of a set of points in the plane, 18th A n n u a l S y m p o s i u m

o n F o u n d a t i o n s o f C o m p u t e r S c i e n c e , IEEE, Long B e a c h , C A , 0 c t . 3 1 - N o v . 2 , 1 9 7 7 , p p .
228-240. A l s o a v a i l a b l e as M I T T e c h n i c a l R e p o r t M I T / L C S / T M - 8 8 .
[52] W . J . M a s e k a n d M.S. P a t e r s o n , A Faster Algorithm Computing String Edit Distances,

T e c h . Rep. M I T / L C S / T M - 1 0 5 , MIT Laboratory for Computer Science, C a m b r i d g e , Mass.
[53] T . L . M o r i n , Branch and Bound Strategies for Dynamic Programming, Operations Res .
24 (1976), pp. 611-627.
PAGE 43
[54] V . Y . P a n , An introduction to the trilinear technique of aggregating, uniting and

cancelling and applications of the technique for constructing fast algorithms for
matrix operations,' P r o c e e d i n g s of the N i n e t e e n t h A n n u a l S y m p o s i u m o n t h e
F o u n d a t i o n s o f C o m p u t e r S c i e n c e , I E E E , Long B e a c h , C A , 1978,
[55] M.S. P a t e r s o n , U n i v e r s i t y of W a r w i c k , E n g l a n d .
[56] H J . P a y n e a n d W.S. M e i s e l , An Algorithm for Constructing Optimal Binary Decision

Trees, I E E E T r a n s . C o m p u t e r s , C - 2 6 ( 1 9 7 7 ) , p p . 9 0 5 - 9 1 6 .
[57] Y . P e r l , Average Analysis of Simple Path Algorithms, T e c h . R e p . U I U C D C S - R - 7 7 - 9 0 5

U I L U - E N G 7 7 1 7 5 9 , D e p t . of C o m p u t e r S c i e n c e , U n i v e r s i t y o f Illinois,
U r b a n a - C h a m p a i g n , N o v . 1977.
[58] J . L . P e t e r s o n , J.R. B i t n e r , and J.H. H o w a r d , The Selection of Optimal Tab Settings,

C o m m . A C M , 21 ( 1 9 7 8 ) , p p . 1 0 0 4 - 1 0 0 6 .
[59] I. P o h l , A sorting problem and its complexity, Comm. A C M , 15 ( 1 9 7 2 ) , p p . 4 6 2 - 4 6 4 .
[60] M. P o l l a c k a n d W. W i e b e n s o n , Solutions of the shortest route problem - A review,

O p e r a t i o n s Res., 8 ( 1 9 6 0 ) , p p . 2 2 4 - 2 3 0 .
[61] F.P. P r e p a r a t a and S.R. R a y , An approach to artificial symbolic cognition, Information

Sci., 4 (1972), pp. 6 5 - 8 6 .
[62] R.L. P r o b e r t , An Extension of Computational Duality To Sequences of Bilinear

Computations, S I A M J . Comput., 7 (1978), pp. 9 1 - 9 8 .
[63] A . S c h o n h a g e , M. P a t e r s o n , and N. P i p p e n g e r , Finding the median, J. Comput. System

S c i . , 13 ( 1 9 7 6 ) , p p . 1 8 4 - 1 8 9 .
[64] H. S c h u m a c h e r and K.C. S e v c i k , The Synthetic Approach to Decision Table Conversion,

C o m m . A C M , 19 ( 1 9 7 6 ) , p p . 3 4 3 - 3 5 1 . See also t h e T e c h n i c a l C o r r e s p o n d e n c e s e c t i o n
of the C . A C M 21(1978), pp. 179-180.
[65] S . M . S e l k o w , The tree-to-tree editing problem, I n f o r m a t i o n P r o c e s s i n g L e t t . , 6 ( 1 9 7 7 ) ,

pp. 184-186.
[66] M . I . S h a m o s , K.Q. B r o w n , J . S a x e , and B.W. W e i d e , C l u s t e r i n g in O n e D i m e n s i o n , I n

preparation.
[67] J . F . S h a p i r o , Shortest Route Methods for Finite State Space Deterministic Dynamic
Programming Problems, S I A M J . , 16 ( 1 9 6 8 ) , p p . 1 2 3 2 - 1 2 5 0 .
[68] R.E. T a r j a n , Complexity of Combinatorial Algorithms, S I A M R e v i e w , 2 0 ( 1 9 7 8 ) , p p .

457-491.
[69] M.E. T h o m a s , A Survey of the State of the Art in Dynamic Programming, Amer. Inst,
of Indust. Engin., 8 (1976), pp. 59-69.
[70] L . G . V a l i a n t , General Context-Free Recognition in Less than Cubic Time, J. Comput.

S y s t e m S c i . , 10 ( 1 9 7 5 ) , p p . 3 0 8 - 3 1 5 .
[71] A . J . V i t e r b i , Error bounds for convolutional codes and an asymptotically optimum

decoding algorithm, I E E E T r a n s . I n f o r m a t i o n T h e o r y , I T - 1 3 ( 1 9 6 7 ) , p p . 2 6 0 - 2 6 9 .
PAGE 44
[72] K M . W a g n e r , Principles of Operations Research, Prentice-Hall, E n g l e w o o d C l i f f s , N.J.,

1969.
[73] R.A. W a g n e r a n d M.J. F i s c h e r , The siring-to-string correction problem, J. Assoc. Comp.

M a c h . , 21 ( 1 9 7 4 ) , p p . 1 6 8 - 1 7 3 .
[74] C K . W o n g a n d A.K. C h a n d r a , Bounds for the String Editing Problem, J. Assoc. Comp.
Mach., 23 (1976), pp. 13-16.
[75] D.H. Y o u n g e r , Recognition and parsing of context-free languages, Infomation and

C o n t r o l , 10 ( 1 9 6 7 ) , p p . 1 8 9 - 2 0 8 .
UNCLASSIFIED llllSlffilllM ..... , 0 flM
SECURITY CLASSIFICATION O F T H I S P A G E (When Data Entered) I11 l|l 1 l|j ||i 1 If| || l| l|l 11|| 11| 1 If 1
1 11(I111 IfJ11 ||lfll f 1J11
REPORT DOCUMENTATION PAGE 3 fl4fl5 D ] i n pp ) m p l e t i n g N
f S
o r m
1. REPORT NUMBER 2. G O V T A C C E S S I O N ww. ^. . _ _ TALOG NUMBER
CMU-CS-79-106
5. TYPE OF REPORT & PERIOD COVERED
4. T I T L E (and Subtitle)
• Interim
DYNAMIC PROGRAMMING IN COMPUTER SCIENCE 6. PERFORMING ORG. REPORT NUMBER
8. CONTRACT OR GRANT NUMBERf*)

7. AUTHORS
Kevin Q. Brown N00014-76-C-0370
9. PERFORMING ORGANIZATION NAME AND ADDRESS 10. PROGRAM ELEMENT. PROJECT, TASK
A R E A 4 WORK U N I T N U M B E R S
Carnegie-Mellon University
Computer Science Department
Pittsburgh, PA 15213
11. CONTROLLING OFFICE NAME ANO ADORESS _ • 12. REPORT DATE
February 1979
Office of Naval Research
13. NUMBER OF PAGES
Arlington, VA 22217 •
49
14. MONITORING AGENCY N A M E & A D O R E S S f / / dltieront from Controlling Office) 15. SECURITY C L A S S , (of thla report)
UNCLASSIFIED
Same as above 15a. D E C L ASSI F l C A T I O N / DOWN G R A D I N G
SCHEDULE
16. DISTRIBUTION STATEMENT (of this Report)
Approved for.public release; distribution unlimited
17. DISTRIBUTION STATEMENT (of the abstract entered In Block 20, it different from Report)
18. SUPPLEMENTARY NOTES
19. KEY W O R D S (Continue on reverse aide if neceaeary and identify by block number)
20. A B S T R A C T (Continue on reverse aide it neceaeary and identify by block number)
DD i jan M
73 1473 E D I T I O N O F 1 N O V 6 5 IS O B S O L E T E UNCLASSIFIED
S/N 0102-014- 6601 | S E C U R I T Y C L A S S I F I C A T I O N O F T H I S P A G E (When Data Entered)

Dynamic Programming in Computer Science

Uploaded by

Dynamic Programming in Computer Science

Uploaded by

Carnegie Mellon University

Research Showcase @ CMU

Dynamic programming in computer science

Follow this and additional works at: htp://repository.cmu.edu/compsci

Dynamic Programming in Computer Science University Libraries

Carnegie-Mel Ion University

6. Optimal Partition Problems

Dynamic programming ( D P ) is o n e of s e v e r a l problem solving methods u s e d in computer

For other t y p e s o f p r o b l e m s t h e r e are c o r r e s p o n d i n g problem solving methods. These

Dynamic programming o r i g i n a t e d w i t h the w o r k of Bellman [ 5 ] and has b e e n applied to

and deterministic. T a r j a n ' s [ 6 8 ] r e c e n t s u r v e y article o n c o m b i n a t o r i a l a l g o r i t h m s i n c l u d e s a

There a r e s e v e r a l a p p r o a c h e s that w e can take t o s o l v e this p r o b l e m . O n e a p p r o a c h is t o

Why c a n t h e r e s u l t s o f s t a g e 2 be u s e d to s o l v e s t a g e 3, and the r e s u l t s o f s t a g e 3 be

W e c a n u s e t h i s p r i n c i p l e f o r t h e g e n e r a l case of K columns of L n o d e s . The dynamic

B e l l m a n ' s P r i n c i p l e of O p t i m a l i t y , as s h o w n in the example of C h a p t e r 2, is an i m p o r t a n t a n d

w h e r e A * is t h e s e t o f all n - t u p l e s of elements of A f o r n > 0. T h e case n = 0 is t h e empty

P = { matrices C | C j f k > 0, V i < {0,1,2,3}, j , k < {1,2,3} }, and

f is d e f i n e d r e c u r s i v e l y b y : f ( e , p ) = 0, f(aj,p) - C 0 1 j , f(xajaj,p) - f(xaj,p) + C n j j

The c o n v e r s i o n f r o m D D P to SDP ( s e q u e n t i a l d e c i s i o n p r o c e s s ) is not e n t i r e l y automatic.

A = a l p h a b e t f o r Z (same A as f o r the DDP a b o v e ) ,

F = { final s t a t e s }, / correspond to S in the DDP

P = d a t a f o r t h e c o s t f u n c t i o n s h and k (same as P f o r t h e DDP),

P = { matrices C | C j j k > 0, V i * {0,1,2,3}, j,k ^ {1,2,3} },

then the SDP is a monotone sequential decision process ( M S D P ) f r o m w h i c h t h e functional

e q u a t i o n s f o r the DP algorithm can be d e r i v e d .

Monotone property: If r^ < r 2 then h(rj,qQ,x,p) < h(r ,qQ,x,p), V x ( A * . 2

T h a t i s , G ( q , p ) is t h e minimum p o s s i b l e cost t o g e t t o state q , u n d e r t h e c o s t f u n c t i o n h. N o w it

The results of Karp a n d Held p r o v i d e a p o w e r f u l a p p r o a c h f o r t h e c o n s t r u c t i o n o f D P

Our goal now is t o o b t a i n the functional e q u a t i o n s f o r a DP a l g o r i t h m t h a t s o l v e s this

A = { a a2> }, where "af means "go to node L"

S = { A*a|vj }, = set of all feasible paths from node 1 to node N

P = { m a t r i c e s C | C j j > 0, Cjj = 0, i,j=l,2,...N }, and

Q = { Sj, s 2 > . . . S|\j }, where "s^" means "at node i"

T h e m o n o t o n i c i t y p r o p e r t y is s a t i s f i e d e a s i l y b y the a d d i t i v i t y of t h e c o s t f u n c t i o n . Thus, the

! S - set of nodes for which an optimal path has been found;

This a l g o r i t h m c a n b e m o d i f i e d to r u n in 0 ( E l o g N ) time and 0 ( E ) s t o r a g e , w h e r e E is the

In this chapter we present several problems that can be described as optimal

G i v e n a s t r i n g of N e l e m e n t s , find an optimal (in some s e n s e ) p a r e n t h e s i z a t i o n o f

The obvious brute force method for solving an optimal p a r e n t h e s i z a t i o n problem is to

The DP a l g o r i t h m s f o r these problems o f f e r substantial improvement o v e r the b r u t e force

k ) , t h e t o t a l t i m e is 0 ( N ) , much less than 0 ( X ( N ) ) .

5*1 Optimizing the Multiplication of N Matrices

[5x10] [10x100] [100x2] [2x20] [20x50]

W e w i l l a s s u m e in t h i s d i s c u s s i o n that the multiplication of a p x q matrix a n d a q x r matrix

t h e n t h e multiplication will cost o n l y 4600 operations.

W e w i l l n o w d e s c r i b e t h e g e n e r a l p r o b l e m and the c o r r e s p o n d i n g O ( N ^ ) time D P a l g o r i t h m . *

where Mj is a c q b y c± m a t r i x , M 2 is a c j b y c 2 matrix, e t c . The discrete decision process

S = { sequences of a ^ corresponding to complete p a r e n t h e s i z a t i o n of the N

W e m u s t f i r s t d e f i n e t h e s t a t e s Q b e f o r e c o n s t r u c t i n g the SDP. T h e most n a t u r a l s e t o f s t a t e s

Let's re-examine the situation. T o p r o d u c e a f a s t e r DP a l g o r i t h m t h e n u m b e r o f states

and (Mj |,Mj 2>—^K^

< < ij (j+l),k> > - G s u s G

T h u s , the f u n c t i o n a l equations become

L e t t i n g Gjj - G ( S j j ) , this becomes

A l g o r i t h m F o r C o m p u t i n g Optimal Multiplication of N Matrices

5.2 Context Free Language Recognition

The DP algorithm for general context free recognition that is p r e s e n t e d in t h i s paper

G = (Nonterminals, Terminals, Productions, Start symbol).

To construct t h e D P a l g o r i t h m w e must f i r s t d e f i n e t h e DDP. T h e DDP makes use of a

L e t t h e i n p u t s t r i n g b e Z = z^Z2— |\j» * ' * ^ * ' S °f p r i m i t i v e o p e r a t o r s

aj^. P r e f i x ( x , a j j ) = t h e s t r i n g x minus all o p e r a t o r s to t h e r i g h t o f a ^ .

T h i s is a m o s t c u r i o u s w a y t o d e f i n e a DDP. The primitive o p e r a t o r s and, e s p e c i a l l y , t h e c o s t

to keep only one n o n t e r m i n a l and t h r o w t h e o t h e r s a w a y b e c a u s e t h e p r o d u c t i o n s d o not

} then we cannot say w h i c h is b e t t e r because the t w o s e t s are not comparable. This

N o w it is time t o p r o d u c e the SDP. Let the s t a t e Sjj be i n t e r p r e t e d as " z j Z j j . . . Z j h a s b e e n +

parsed". To produce an e f f i c i e n t algorithm, w e use p s e u d o - s t a t e s (Sjj U S pk as in the

X( (Sjj U S j + ( 1 ) > k ), ajj k ) = s , ik

h(r,(Sjj u S j ( + 1 ) ) k ),ajj ) = f(xak j j k ) w h e r e X ( q , x ) = (Sjj u S j

T h i s is t r u e b e c a u s e if f ( x ) => f ( y ) t h e n c e r t a i n l y Vw<=A* f ( x w ) f(yw). We conclude that the

functional equations are

<V (M>,k' ijk>

U f(x x a 1 2 j j k ) where M q ^ ) = S j j and X C q ^ ) = S ^

Input: Nonterminals = {Aj,A2, . . . A ^ j } / = start symbol

5.3 Optimal Alphabetic Encoding and Optimal Search Trees

be represented by a prefix code c h o s e n f r o m a set of r s y m b o l s . * Furthermore, for an

{alphabetic encodings} "\ i"

L e t a K - partition o f t h e N e l e m e n t s be r e p r e s e n t e d b y a K + l - t u p l e (Io'l^2* • " • W w

4. R e p e a t t h e d i v i d e s t e p of s t e p 3. T h a t is, find I|</g, l3K/8> 5K/8> 7K/8> * a n c e c