Sie sind auf Seite 1von 67

Algorithms R OBERT S EDGEWICK | K EVIN W AYNE

2.3 Q UICKSORT
quicksort
selection
duplicate keys
Algorithms
system sorts
F O U R T H E D I T I O N

R OBERT S EDGEWICK | K EVIN W AYNE

http://algs4.cs.princeton.edu
Two classic sorting algorithms: mergesort and quicksort

Critical components in the worlds computational infrastructure.


Full scientific understanding of their properties has enabled us
to develop them into practical system sorts.
Quicksort honored as one of top 10 algorithms of 20th century
in science and engineering.

Mergesort. [last lecture]

...

Quicksort. [this lecture]

...

2
Quicksort t-shirt

3
2.3 Q UICKSORT
quicksort
selection
duplicate keys
Algorithms
system sorts

R OBERT S EDGEWICK | K EVIN W AYNE

http://algs4.cs.princeton.edu
Quicksort

Basic plan.
Shuffle the array.
Partition so that, for some j
entry a[j] is in place
no larger entry to the left of j
no smaller entry to the right of j
Sort each subarray recursively.

input Q U I C K S O R T E X A M P L E
shuffle K R A T E L E P U I M Q C X O S
partitioning item
partition E C A I E K L P U T M Q R X O S
not greater not less
sort left A C E E I K L P U T M Q R X O S
sort right A C E E I K L M O P Q R S T U X
result A C E E I K L M O P Q R S T U X
Quicksort overview
5
Tony Hoare
n u m b e r ) . 9.9 X 10 45 is u s e d to r e p r e s e n t i n f i n i t y . I m a g i n a r y
v a l u e s of x m a y n o t be n e g a t i v e a n d reM v a l u e s of x m a y n o t be
conunent I a n d J are o u t p u t v a r i a b l e s , a n d A is t h e a r r a y ( w i t h
s u b s c r i p t b o u n d s M : N ) w h i c h is o p e r a t e d u p o n b y t h i s p r o c e d u r e .
s m a l l e r t h a n 1. P a r t i t i o n t a k e s t h e v a l u e X of a r a n d o m e l e m e n t of t h e a r r a y A,
V a l u e s of Qd~'(x) m a y be c a l c u l a t e d e a s i l y by h y p e r g e o m e t r i c a n d r e a r r a n g e s t h e v a l u e s of t h e e l e m e n t s of t h e a r r a y in s u c h a
series if x is n o t t o o s m a l l n o r (n - m ) t o o large. Q~m(x) c a n be w a y t h a t t h e r e e x i s t i n t e g e r s I a n d J w i t h t h e following p r o p e r t i e s :
c o m p u t e d f r o m a n a p p r o p r i a t e set of v a l u e s of Pnm(X) if X is n e a r M _-< J < I =< N p r o v i d e d M < N

Invented quicksort to translate Russian into English.


1.0 or ix is n e a r 0. L o s s of s i g n i f i c a n t d i g i t s o c c u r s for x as s m a l l as A[R] =< X f o r M =< R _-< J
1.1 if n is l a r g e r t h a n 10. L o s s of s i g n i f i c a n t d i g i t s is a m a j o r diffi- A[R] = X f o r J < R < I
c u l t y in u s i n g finite p o l y n o m i M r e p r e s e n t a t i o n s also if n is l a r g e r A[R] ~ X f o r I =< R ~ N

[ but couldn't explain his algorithm or implement it! ]


t h a n m . H o w e v e r , Q L E G h a s b e e n t e s t e d in r e g i o n s of x a n d n The procedure uses an integer procedure random (M,N) which
b o t h large a n d small; chooses equiprobably a random integer F between M and N, and
procedure Q L E G ( m , n m a x , x, ri, R, Q); v a l u e In, n m a x , x, ri; also a p r o c e d u r e e x c h a n g e , w h i c h e x c h a n g e s t h e v a l u e s of its t w o
r e a l In, m n a x , x, ri; r e a l a r r a y R , Q; parameters ;

Learned Algol 60 (and recursion).


begin r e a l t, i, n, q0, s;
n : = 20;
i f n m a x > 13 t h e n
begin real X; integer F;
F : = r a n d o m ( M , N ) ; X : = A[F];
I:=M; J:=N;

Implemented quicksort.
n := nmax +7; up: for I : = I s t e p 1 u n t i l N d o
i f ri = 0 t h e n i f X < A [I] t h e n g o to d o w n ;
begin ifm = 0then I:=N;
Q[0] : = 0.5 X 10g((x + 1 ) / ( x - 1)) down: f o r J : = J s t e p --1 u n t i l M d o
else if A[J]<X then go to change;
begin t : = - - 1 . 0 / s q r t ( x X x - - 1); J:=M;
q0 : = 0; c h a n g e : i f I < J t h e n b e g i n e x c h a n g e (A[IL A[J]); Tony Hoare
Q[O] : = t ; I := I+ 1;J:= J - 1;
for i : = 1 step 1 until m do g o to u p 1980 Turing Award
begin s := (x+x)X(i-1)Xt end 2: b e g i n
4
Q[0]+ (3i-ii-2)q0; else i f [ < F t h e n b e g i n e x c h a n g e (A[IL A[F]) i i f c >= 0 t h e n
q0 : = Q[0]; I:=I+l 3:begin
Q[0] : = s e n d e n d ; end e:= aXe;f := bXd;goto8
if x = 1 then else i f F < J t l l e n b e g i n e x c h a n g e (A[F], A[J]) ; end 3 ;
Q[0] : = 9.9 I" 45; J:=J-1 e:=bXc;
R [ n + 1] : = x - s q r t ( x X x - 1); end ; ifd ~ 0then
for i : = n s t e p --1 u n t i l 1 d o end partition 4: begin
R[i] : = (i + m ) / ( ( i + i + 1) X x f:=bXd; goto8
+(m-i- 1) X R [ i + l ] ) ; e n d 4;
ALGORITHM 61
g o to t h e e n d ; f:=aXd; goto8
if m = 0 then ALGORITHM
PROCEDURES 64F O R R A N G E ARITHMETIC 5: e n d 2;
b e g i n i f x < 0.5 t b e n QUICKSORT
ALLAN GIBB* ifb > 0 then
Q[0] : = a r c t a n ( x ) - 1.5707963 e l s e U n iA.
C. i t y HOARE
v e r sR. of A l b e r t a , C a l g a r y , A l b e r t a , C a n a d a 6: b e g i n
Q[0] : = - a r e t a n ( 1 / x ) e n d e l s e if d > 0 then
Elliott
b e g i n Brothers Ltd., Borehamwood, Hertfordshire, Eng.
begin t : = 1 / s q r t ( x X x + 1); begin
p r o c e d u r e R A N G E S U M (a, b, c, d, e, f);
q0 : = 0; procedure quicksort (A,M,N); value M,N; e : = M I N ( a X d, b X c);
real a , b , c , d , e , f ;
q[0] := t; a r r a y A; i n t e g e r M , N ; f : = M A X ( a X c , b X d); go t o 8
c o m m e n t The term "range n u m b e r " was used by P. S. Dwyer,
for i : = 2 step 1 until m do comment Q u i c k s o r t is a v e r y f a s t a n d c o n v e n i e n t m e t h o d of e n d 6;
b e g i n s : = (x + x) X (i -- 1) X t X Q[0I
Linear Computations (Wiley, 1951). Machine procedures for
s o r t i n g a n a r r a y in t h e r a n d o m - a c c e s s s t o r e of a c o m p u t e r . T h e e : = b X c; f : = a X c; go t o 8
range arithmetic were developed about 1958 by Ramon Moore,
+(3i+iX i -- 2) q0; e n t i r e c o n t e n t s of t h e s t o r e m a y be s o r t e d , since no e x t r a s p a c e is e n d 5;
"Automatic Error Analysis in Digital C o m p u t a t i o n , " LMSD
qO : = Q[0]; r e q u i r e d . T h e a v e r a g e n u m b e r of c o m p a r i s o n s m a d e is 2 ( M - - N ) In f:=aXc;
Report 48421, 28 Jan. 1959, Lockheed Missiles and Space Divi-
Q[0] := s e n d e n d ; ( N - - M ) , a n d t h e a v e r a g e n m n b e r of e x c h a n g e s is one s i x t h t h i s i f d _-<O t h e n
sion, Palo Alto, California, 59 pp. If a _< x -< b and c ~ y ~ d,
R[n + 1] : = x - s q r t ( x x + 1); a m o u n t . S u i t a b l e r e f i n e m e n t s of t h i s m e t h o d will be d e s i r a b l e for 7: b e g i n
then R A N G E S U M yields an interval [e, f] such t h a t e =< (x + y)
for i := n step - 1 until 1 do its i m p l e m e n t a t i o n on a n y a c t u a l c o m p u t e r ; e:=bXd; goto8
f. Because of machine operation (truncation or rounding) the
R[i] : = (i + m ) / ( ( i -- m + 1) R[i + 1] begin i n t e g e r 1,J ; end 7 ;
machine sums a -4- c and b -4- d may not provide safe end-points
--(i+i+ 1) X x); if M < N then begin partition (A,M,N,I,J); e:=aXd;
of the output interval. Thus R A N G E S U M requires a non-local
for i : = 1 step 2 until nmax do quicksort (A,M,J) ; 8: e : = A D J U S T P R O D (e, - 1 ) ;
real procedure A D J U S T S U M which will compensate for the
Rill : = -- Rill; q u i c k s o r t (A, I, N ) f := A D J U S T P R O D (f, 1)
machine arithmetic. The body of A D J U S T S U M will be de-
the: for i : = 1 step 1 until nnmx do end end RANGEMPY;
pendent upon the type of machine for which it is written and so
Q[i] : = Q[i - 1] X R[i] end quieksort p r o c e d u r e R A N G E D V D (a, b, c, d, e, f) ;
is not given here. (An example, however, appears below.) I t
end QLEG; real a, b, c, d, e, f;
is assumed t h a t A D J U S T S U M has as parameters real v and w,
c o m m e n t If the range divisor includes zero the program
* T h i s p r o c e d u r e w a s d e v e l o p e d in p a r t u n d e r t h e s p o n s o r s h i p and integer i, and is accompanied by a non-local real procedure
exists to a non-local label " z e r o d v s r " . R A N G E D V D assumes a
of t h e A i r F o r c e C a m b r i d g e R e s e a r c h C e n t e r .
Communications of the ACM (July 1961)
C O R R E C T I O N 65
ALGORITHM which gives an upper bound to the magnitude
of the error involved in the machine representation of a number.
non-local real procedure A D J U S T Q U O T which is analogous
FIND (possibly identical) to A D J U S T P R O D ;
The output A D J U S T S U M provides the left end-point of the 6
C.output A. R.intervalHOARE begin
of R A N G E S U M when A D J U S T S U M is called
Tony Hoare

Invented quicksort to translate Russian into English.


[ but couldn't explain his algorithm or implement it! ]
Learned Algol 60 (and recursion).
Implemented quicksort.
Tony Hoare
1980 Turing Award
There are two ways of constructing a software design: One way is
to make it so simple that there are obviously no deficiencies, and
the other way is to make it so complicated that there are no obvious
deficiencies. The first method is far more difficult.

I call it my billion-dollar mistake. It was the invention of the null


reference in 1965 This has led to innumerable errors,
vulnerabilities, and system crashes, which have probably caused
a billion dollars of pain and damage in the last forty years.

7
Bob Sedgewick

Refined and popularized quicksort.


Analyzed many versions of quicksort.

rithms, and we can learn from that experience to separate


good algorithms from bad ones. Third, if the tile fits into
Bob Sedgewick
the memory of the computer, there is one algorithm,
called Quicksort, which has been shown to perform well
Programming S. L. Graham, R. L. Rivest in a variety of situations. Not only is this algorithm
Techniques Editors simpler than many other sorting algorithms, but empir-

Implementing ical [2, ll, 13, 21] and analytic [9] studies show that
Quicksort can be expected to be up to twice as fast as its

Quicksort Programs nearest competitors. The method is simple enough to be


learned by programmersActa Informatica
9 by
who have no
Springer-Verlag
7, 327--355 (1977)
previous experi-
1977
ence with sorting, and those who do know other sorting
Robert Sedgewick methods should also find it profitable to learn about
Brown University Quicksort.
Because of its prominence, it is appropriate to study
The Analysis of Quicksort Programs*
how Quicksort might be improved. This subject has
This paper is a practical study of how to implement received considerable attention (see, for example, [1, 4,
the Quicksort sorting algorithm and its best variants on 11, 13, 14, 18, 20]), but few real improvements have been Robert Sedgewick
real computers, including how to apply various code suggested beyond those described by C.A.R. Hoare, the
optimization techniques. A detailed implementation inventor of Quicksort, in his original papers [5, 6].Received Hoare January 19, t976
combining the most effective improvements to also showed how to analyze Quicksort and predict its
running time. The analysis Summary. The been
has since Quicksort
extendedsorting
to algorithm and its best variants are presented
Quicksort is given, along with a discussion of how to and analyzed. Results are derived which make it possible to obtain exact formulas de-
implement it in assembly language. Analytic results the improvements that he suggested, and used to indicate
scribing the total expected running time of particular implementations on real com-
describing the performance of the programs are how they may best be implemented
puters of Quick,sort [9,and15,an17]. The
improvement called the median-of-three modification.
summarized. A variety of special situations are subject of the carefulDetailed analysis of the effect of has
implementation of Quicksort an implementation technique called loop unwrapping
considered from a practical standpoint to illustrate not been studied as widely as global
is presented. The paper improvements
is intended tonot only to present results of direct practical utility,
Quicksort's wide applicability as an internal sorting the algorithm, butbut thealso to illustrate
savings to be the intriguing
realized are asmathematics which arises in the complete analysis
of thisofimportant
significant. The history Quicksortalgorithm.
is quite complex,
method which requires negligible extra storage.
Key Words and Phrases: Quicksort, analysis of and [15] contains a full survey of the many variants
algorithms, code optimization, sorting which, have been proposed. 1. Introduction
CR Categories: 4.0, 4.6, 5.25, 5.31, 5.5 The purpose of this paper is to describe in detail how
Quicksort can best beInimplemented
t96t-62 C.A.R. Hoareactual
to handle presented a new algorithm called Quicksort [7, 8]
applications on real computers. A general description ofinto order by computer. This method combines
which is suitable for putting files
elegancebyand
the algorithm is followed efficiency,of and
descriptions it remains today the most useful general-purpose
the most
effective improvementssortingthat method
have for beencomputers.
proposed The (as practical utility of the algorithm has meant
demonstrated in [15]). Next, an implementation of to countless modifications (though few real
not only that it has been sfibjected 8
Quicksort partitioning demo

Repeat until i and j pointers cross.


Scan i from left to right so long as (a[i] < a[lo]).

Scan j from right to left so long as (a[j] > a[lo]).

Exchange a[i] with a[j].

K R A T E L E P U I M Q C X O S

lo i j

9
Quicksort partitioning demo

Repeat until i and j pointers cross.


Scan i from left to right so long as (a[i] < a[lo]).

Scan j from right to left so long as (a[j] > a[lo]).

Exchange a[i] with a[j].


When pointers cross.
Exchange a[lo] with a[j].

E C A I E K L P U T M Q R X O S

lo j hi

partitioned!
10
The music of quicksort partitioning (by Brad Lyon)

https://googledrive.com/host/0B2GQktu-wcTicjRaRjV1NmRFN1U/index.html

11
Quicksort: Java code for partitioning

private static int partition(Comparable[] a, int lo, int hi)


{
int i = lo, j = hi+1;
while (true)
{
while (less(a[++i], a[lo])) find item on left to swap
if (i == hi) break;

while (less(a[lo], a[--j])) find item on right to swap


if (j == lo) break;

if (i >= j) break; check if pointers cross


exch(a, i, j); swap
}

exch(a, lo, j); swap with partitioning item


before v
return j; return index of item now known to be in place
} lo hi

before v during v "v !v

lo hi i j

during after "v v !v


before v v "v !v

lo hi i j lo j hi

during v "v !v after "v v !v Quicksort partitioning overview


12
Quicksort quiz 1

Q. How many compares (in the worst case) to partition an array of length N ?

A. ~N

B. ~N

C. ~N

D. ~ N lg N

E. I don't know.

13
Quicksort: Java implementation

public class Quick


{
private static int partition(Comparable[] a, int lo, int hi)
{ /* see previous slide */ }

public static void sort(Comparable[] a)


{
shuffle needed for
StdRandom.shuffle(a);
performance guarantee
sort(a, 0, a.length - 1);
(stay tuned)
}

private static void sort(Comparable[] a, int lo, int hi)


{
if (hi <= lo) return;
int j = partition(a, lo, hi);
sort(a, lo, j-1);
sort(a, j+1, hi);
}
}

14
Quicksort trace

lo j hi 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
initial values Q U I C K S O R T E X A M P L E
random shuffle K R A T E L E P U I M Q C X O S
0 5 15 E C A I E K L P U T M Q R X O S
0 3 4 E C A E I K L P U T M Q R X O S
0 2 2 A C E E I K L P U T M Q R X O S
0 0 1 A C E E I K L P U T M Q R X O S
1 1 A C E E I K L P U T M Q R X O S
4 4 A C E E I K L P U T M Q R X O S
6 6 15 A C E E I K L P U T M Q R X O S
no partition 7 9 15 A C E E I K L M O P T Q R X U S
for subarrays
of size 1 7 7 8 A C E E I K L M O P T Q R X U S
8 8 A C E E I K L M O P T Q R X U S
10 13 15 A C E E I K L M O P S Q R T U X
10 12 12 A C E E I K L M O P R Q S T U X
10 11 11 A C E E I K L M O P Q R S T U X
10 10 A C E E I K L M O P Q R S T U X
14 14 15 A C E E I K L M O P Q R S T U X
15 15 A C E E I K L M O P Q R S T U X

result A C E E I K L M O P Q R S T U X

Quicksort trace (array contents after each partition)

15
Quicksort animation

50 random items

algorithm position
in order
current subarray
not in order
http://www.sorting-algorithms.com/quick-sort
16
Quicksort: implementation details

Partitioning in-place. Using an extra array makes partitioning easier


(and stable), but is not worth the cost.

Terminating the loop. Testing whether the pointers cross is trickier


than it might seem.

Equal keys. When duplicates are present, it is (counter-intuitively)


better to stop scans on keys equal to the partitioning item's key. stay tuned

Preserving randomness. Shuffling is needed for performance guarantee.


Equivalent alternative. Pick a random partitioning item in each subarray.

17
Quicksort: empirical analysis (1961)

Running time estimates:


Algol 60 implementation.
National-Elliott 405 computer.

Elliott 405 magnetic disc


sorting N 6-word items with 1-word keys
(16K words)

18
Quicksort: empirical analysis

Running time estimates:


Home PC executes 108 compares/second.
Supercomputer executes 1012 compares/second.

insertion sort (N2) mergesort (N log N) quicksort (N log N)

computer thousand million billion thousand million billion thousand million billion

home instant 2.8 hours 317 years instant 1 second 18 min instant 0.6 sec 12 min

super instant 1 second 1 week instant instant instant instant instant instant

Lesson 1. Good algorithms are better than supercomputers.


Lesson 2. Great algorithms are better than good ones.
19
Quicksort: best-case analysis

Best case. Number of compares is ~ N lg N.

initial values

random shuffle

20
Quicksort: worst-case analysis

Worst case. Number of compares is ~ N 2 .

initial values

random shuffle

21
Quicksort: average-case analysis

Proposition. The average number of compares CN to quicksort an array of


N distinct keys is ~ 2N ln N (and the number of exchanges is ~ N ln N ).

Pf. CN satisfies the recurrence C0 = C1 = 0 and for N 2:


left right
partitioning

C0 + CN 1 C1 + CN 2 CN 1+ C0
CN = (N + 1) + + + ... +
N N N

Multiply both sides by N and collect terms: partitioning probability

N CN = N (N + 1) + 2(C0 + C1 + . . . + CN 1)

Subtract from this equation the same equation for N - 1:


N CN (N 1)CN 1 = 2N + 2CN 1

Rearrange terms and divide by N (N + 1):


CN CN 1 2
= +
N +1 N N +1
22
Quicksort: average-case analysis

Repeatedly apply previous equation:


CCNN CC
NN 1 1 22
= ++
NN++11 NN NN++1 1
CN 2 2 2
= + + substitute previous equation
N 1 N N +1
CN 3 2 2 2
= + + +
N 2 N 1 N N +1
2 2 2 2
= + + + ... +
3 4 5 N +1

Approximate sum by an integral:



1 1 1 1
CN = 2(N + 1) + + + ...
3 4 5 N +1
Z N +1
1
2(N + 1) dx
3 x

Finally, the desired result:


CN 2(N + 1) ln N 1.39N lg N
23
Quicksort: summary of performance characteristics

Quicksort is a (Las Vegas) randomized algorithm.


Guaranteed to be correct.
Running time depends on random shuffle.
Average case. Expected number of compares is ~ 1.39 N lg N.
39% more compares than mergesort.
Faster than mergesort in practice because of less data movement.
Best case. Number of compares is ~ N lg N.
Worst case. Number of compares is ~ N 2.
[ but more likely that lightning bolt strikes computer during execution ]

24
Quicksort properties

Proposition. Quicksort is an in-place sorting algorithm.


Pf.
Partitioning: constant extra space.
Depth of recursion: logarithmic extra space (with high probability).
can guarantee logarithmic depth by recurring
on smaller subarray before larger subarray
(requires using an explicit stack)

Proposition. Quicksort is not stable.


Pf. [ by counterexample ]
i j 0 1 2 3

B1 C1 C2 A1

1 3 B1 C1 C2 A1

1 3 B1 A1 C2 C1

0 1 A1 B1 C2 C1

25
Quicksort: practical improvements

Insertion sort small subarrays.


Even quicksort has too much overhead for tiny subarrays.
Cutoff to insertion sort for 10 items.

private static void sort(Comparable[] a, int lo, int hi)


{
if (hi <= lo + CUTOFF - 1)
{
Insertion.sort(a, lo, hi);
return;
}
int j = partition(a, lo, hi);
sort(a, lo, j-1);
sort(a, j+1, hi);
}

26
Quicksort: practical improvements

Median of sample.
Best choice of pivot item = median.
Estimate true median by taking median of sample.
Median-of-3 (random) items.
~ 12/7 N ln N compares (14% fewer)
~ 12/35 N ln N exchanges (3% more)

private static void sort(Comparable[] a, int lo, int hi)


{
if (hi <= lo) return;

int median = medianOf3(a, lo, lo + (hi - lo)/2, hi);


swap(a, lo, median);

int j = partition(a, lo, hi);


sort(a, lo, j-1);
sort(a, j+1, hi);
}

27
2.3 Q UICKSORT
quicksort
selection
duplicate keys
Algorithms
system sorts

R OBERT S EDGEWICK | K EVIN W AYNE

http://algs4.cs.princeton.edu
Selection

Goal. Given an array of N items, find the kth smallest item.


Ex. Min (k = 0), max (k = N - 1), median (k = N / 2).

Applications.
Order statistics.
Find the "top k."
Use theory as a guide.
Easy N log N upper bound. How?
Easy N upper bound for k = 1, 2, 3. How?
Easy N lower bound. Why?
Which is true?
N log N lower bound? is selection as hard as sorting?

N upper bound? is there a linear-time algorithm?

29
Quick-select

Partition array so that:


Entry a[j] is in place.
No larger entry to the left of j.
No smaller entry to the right of j.
Repeat in one subarray, depending on j; finished when j equals k.

k) v
before
public static Comparable select(Comparable[] a, int
{
lo if a[k] is here if a[k] is here
hi
StdRandom.shuffle(a); set hi to j-1 set lo to j+1
int lo = 0, hi = a.length - 1; during v " v !v
while (hi > lo)
{ i j

int j = partition(a, lo, hi); after !v


"v v
if (j < k) lo = j + 1;
else if (j > k) hi = j - 1; lo j hi
else return a[k];
Quicksort partitioning overview
}
return a[k];
}

30
Quick-select: mathematical analysis

Proposition. Quick-select takes linear time on average.

Pf sketch.
Intuitively, each partitioning step splits array approximately in half:
N + N / 2 + N / 4 + + 1 ~ 2N compares.
Formal analysis similar to quicksort analysis yields:
CN = 2 N + 2 k ln (N / k) + 2 (N k) ln (N / (N k))

Ex: (2 + 2 ln 2) N 3.38 N compares to find median k = N / 2.

31
Theoretical context for selection

Proposition. [Blum, Floyd, Pratt, Rivest, Tarjan, 1973] Compare-based


selection algorithm whose worst-case running time is linear.

Time Bounds for Selection

bY .
Manuel Blum, Robert W. Floyd, Vaughan Watt,
Ronald L. Rive&, and Robert E. Tarjan

Abstract

L The number of comparisons required to select the i-th smallest of

n numbers is shown to be at most a linear function of n by analysis of


i
a new selection algorithm -- PICK. Specifically, no more than
i 5.4305 n comparisons are ever required. This bound is improved for
L
extreme values of i , and a new lower bound on the requisite number

L of comparisons is also proved.

Remark. Constants are high not used in practice.

Use theory as a guide.


L

Still worthwhile to seek practical linear-time (worst-case) algorithm.


Until one is discovered, use quick-select if you dont need a full sort.
32
2.3 Q UICKSORT
quicksort
selection
duplicate keys
Algorithms
system sorts

R OBERT S EDGEWICK | K EVIN W AYNE

http://algs4.cs.princeton.edu
Duplicate keys

Often, purpose of sort is to bring items with equal keys together.


Sort population by age.
Remove duplicates from mailing list.
Sort job applicants by college attended.
sorted by time sorted by city (unstable)
Chicago 09:00:00 Chicago 09:25:52 C
Typical characteristics of such applications.
Phoenix 09:00:03 Chicago 09:03:13 C
Houston 09:00:13 Chicago 09:21:05 C
Huge array. Chicago 09:00:59
Houston 09:01:10
Chicago 09:19:46
Chicago 09:19:32
C
C
Small number of key values. Chicago 09:03:13
Seattle 09:10:11
Chicago 09:00:00
Chicago 09:35:21
C
C
Seattle 09:10:25 Chicago 09:00:59 C
Phoenix 09:14:25 Houston 09:01:10 H
Chicago 09:19:32 Houston 09:00:13 NOT H
Chicago 09:19:46 Phoenix 09:37:44 sorted P
Chicago 09:21:05 Phoenix 09:00:03 P
Seattle 09:22:43 Phoenix 09:14:25 P
Seattle 09:22:54 Seattle 09:10:25 S
Chicago 09:25:52 Seattle 09:36:14 S
Chicago 09:35:21 Seattle 09:22:43 S
Seattle 09:36:14 Seattle 09:10:11 S
Phoenix 09:37:44 Seattle 09:22:54 S

Stability when sorting on a second key


key

34
Duplicate keys: stop on equal keys

Our partitioning subroutine stops both scans on equal keys.

scan until P scan until P

P G E P A Q B P Y C O U P Z S R

Q. Why not continue scans on equal keys?

scan until > P scan until < P

P G E P A Q B P Y C O U P Z S R

35
Quicksort quiz 2

What is the result of partitioning the following array (skip over equal keys)?

scan until > A scan until < A

A A A A A A A A A A A A A A A A
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15

A. A A A A A A A A A A A A A A A A

B. A A A A A A A A A A A A A A A A

C. A A A A A A A A A A A A A A A A

D. I don't know.
36
Quicksort quiz 3

What is the result of partitioning the following array (stop on equal keys)?

scan until A scan until A

A A A A A A A A A A A A A A A A
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15

A. A A A A A A A A A A A A A A A A

B. A A A A A A A A A A A A A A A A

C. A A A A A A A A A A A A A A A A

D. I don't know.
37
Partitioning an array with all equal keys

38
Duplicate keys: partitioning strategies

Bad. Don't stop scans on equal keys.


[ ~ N 2 compares when all keys equal ]

B A A B A B B B C C C A A A A A A A A A A A

Good. Stop scans on equal keys.


[ ~ N lg N compares when all keys equal ]

B A A B A B C C B C B A A A A A A A A A A A

Better. Put all equal keys in place. How?


[ ~ N compares when all keys equal ]

A A A B B B B B C C C A A A A A A A A A A A

39
DUTCH NATIONAL FLAG PROBLEM

Problem. [Edsger Dijkstra] Given an array of N buckets, each containing a


red, white, or blue pebble, sort them by color.

input

sorted

Operations allowed.
swap(i, j): swap the pebble in bucket i with the pebble in bucket j.
color(i): color of pebble in bucket i.
Requirements.
Exactly N calls to color().
At most N calls to swap().
Constant extra space.
40
3-way partitioning

Goal. Partition array into three parts so that:


Entries between lt and gt equal to the partition item.
No larger entries to left of lt.
No smaller entries to right of gt.
before v

lo hi
before v <v =v >v
during
lo hi
lt i gt
during <v =v >v
after <v =v >v
lt i gt
lo lt gt hi
after <v =v >v
3-way partitioning
lo lt gt hi

3-way partitioning
Dutch national flag problem. [Edsger Dijkstra]
Conventional wisdom until mid 1990s: not worth doing.
Now incorporated into C library qsort() and Java 6 system sort.
41
Dijkstra 3-way partitioning demo

Let v be partitioning item a[lo].


Scan i from left to right.
(a[i] < v): exchange a[lt] with a[i]; increment both lt and i
(a[i] > v): exchange a[gt] with a[i]; decrement gt
(a[i] == v): increment i

lt i gt

P A B X W P P V P D P C Y Z

lo hi

before v
invariant
lo hi

during <v =v >v

lt i gt

after <v =v >v 42


Dijkstra 3-way partitioning demo

Let v be partitioning item a[lo].


Scan i from left to right.
(a[i] < v): exchange a[lt] with a[i]; increment both lt and i
(a[i] > v): exchange a[gt] with a[i]; decrement gt
(a[i] == v): increment i

lt gt

A B C D P P P P P V W Y Z X

lo hi

before v
invariant
lo hi

during <v =v >v

lt i gt

after <v =v >v 43


Dijkstra's 3-way partitioning: trace

v a[]
lt i gt 0 1 2 3 4 5 6 7 8 9 10 11
0 0 11 R B W W R W B R R W B R
0 1 11 R B W W R W B R R W B R
1 2 11 B R W W R W B R R W B R
1 2 10 B R R W R W B R R W B W
1 3 10 B R R W R W B R R W B W
1 3 9 B R R B R W B R R W W W
2 4 9 B B R R R W B R R W W W
2 5 9 B B R R R W B R R W W W
2 5 8 B B R R R W B R R W W W
2 5 7 B B R R R R B R W W W W
2 6 7 B B R R R R B R W W W W
3 7 7 B B B R R R R R W W W W
3 8 7 B B B R R R R R W W W W
3 8 7 B B B R R R R R W W W W
3-way partitioning trace (array contents after each loop iteration)

44
3-way quicksort: Java implementation

private static void sort(Comparable[] a, int lo, int hi)


{
if (hi <= lo) return;
int lt = lo, gt = hi;
Comparable v = a[lo];
int i = lo;
while (i <= gt)
{
int cmp = a[i].compareTo(v);
if (cmp < 0) exch(a, lt++, i++);
else if (cmp > 0) exch(a, i, gt--);
else i++;
}
before v
sort(a, lo, lt - 1); lo hi
sort(a, gt + 1, hi); during <v =v >v
}
lt i gt

after <v =v >v

lo lt gt hi

3-way partitioning
45
3-way quicksort: visual trace

equal to partitioning element

Visual trace of quicksort with 3-way partitioning

46
Duplicate keys: lower bound

Sorting lower bound. If there are n distinct keys and the ith one occurs
xi times, any compare-based sorting algorithm must use at least
n

N! xi
lg xi lg N lg N when all distinct;
x1 ! x2 ! xn ! i=1
N linear when only a constant number of distinct keys

compares in the worst case.

Proposition. [Sedgewick-Bentley 1997] proportional to lower bound

Quicksort with 3-way partitioning is entropy-optimal.


Pf. [beyond scope of course]

Bottom line. Quicksort with 3-way partitioning reduces running time


from linearithmic to linear in broad class of applications.

47
Sorting summary

inplace? stable? best average worst remarks

selection N2 N2 N2 N exchanges

use for small N


insertion N N2 N2
or partially ordered

tight code;
shell N log3 N ? c N 3/2
subquadratic

N log N guarantee;
merge N lg N N lg N N lg N
stable

improves mergesort
timsort N N lg N N lg N
when preexisting order

N log N probabilistic guarantee;


quick N lg N 2 N ln N N2
fastest in practice

improves quicksort
3-way quick N 2 N ln N N2
when duplicate keys

? N N lg N N lg N holy sorting grail

48
2.3 Q UICKSORT
quicksort
selection
duplicate keys
Algorithms
system sorts

R OBERT S EDGEWICK | K EVIN W AYNE

http://algs4.cs.princeton.edu
Sorting applications

Sorting algorithms are essential in a broad variety of applications:


Sort a list of names.
Organize an MP3 library. obvious applications

Display Google PageRank results.


List RSS feed in reverse chronological order.
Find the median.
Identify statistical outliers. problems become easy once items
are in sorted order

Binary search in a database.


Find duplicates in a mailing list.
Data compression.
Computer graphics. non-obvious applications

Computational biology.
Load balancing on a parallel computer.
...
50
War story (system sort in C)

A beautiful bug report. [Allan Wilks and Rick Becker, 1991]

We found that qsort is unbearably slow on "organ-pipe" inputs like "01233210":

main (int argc, char**argv) {


int n = atoi(argv[1]), i, x[100000];
for (i = 0; i < n; i++)
x[i] = i;
for ( ; i < 2*n; i++)
x[i] = 2*n-i-1;
qsort(x, 2*n, sizeof(int), intcmp);
}

Here are the timings on our machine:


$ time a.out 2000
real 5.85s
$ time a.out 4000
real 21.64s
$time a.out 8000
real 85.11s

51
War story (system sort in C)

Bug. A qsort() call that should have taken seconds was taking minutes.

Why is qsort() so slow?

At the time, almost all qsort() implementations based on those in:


Version 7 Unix (1979): quadratic time to sort organ-pipe arrays.
BSD Unix (1983): quadratic time to sort random arrays of 0s and 1s.

52
Engineering a system sort (in 1993)

Bentley-McIlroy quicksort.
samples 9 items
Cutoff to insertion sort for small subarrays.
Partitioning item: median of 3 or Tukey's ninther.
Partitioning scheme: Bentley-McIlroy 3-way partitioning.
similar to Dijkstra 3-way partitioning
SOFTWAREPRACTICE AND EXPERIENCE, VOL. 23(11), 12491265 (NOVEMBER 1993)
(but fewer exchanges when not many equal keys)

Engineering a Sort Function

JON L. BENTLEY
M. DOUGLAS McILROY
AT&T Bell Laboratories, 600 Mountain Avenue, Murray Hill, NJ 07974, U.S.A.

SUMMARY
We recount the history of a new qsort function for a C library. Our function is clearer, faster and more
robust than existing sorts. It chooses partitioning elements by a new sampling scheme; it partitions by a
novel solution to Dijkstras Dutch National Flag problem; and it swaps efficiently. Its behavior was
assessed with timing and debugging testbeds, and with a program to certify performance. The design
techniques apply in domains beyond sorting.
KEY WORDS Quicksort Sorting algorithms Performance tuning Algorithm design and implementation Testing

INTRODUCTION
C libraries have long included a qsort function to sort an array, usually implemented by
Very widely used. C, C++, Java 6, .
Hoares Quicksort. 1 Because existing qsorts are flawed, we built a new one. This paper
summarizes its evolution.
Compared to existing library sorts, our new qsort is fastertypically about twice as
fastclearer, and more robust under nonrandom inputs. It uses some standard Quicksort 53
A beautiful mailing list post (Yaroslavskiy, September 2009)

Replacement of quicksort in java.util.Arrays with new dual-pivot quicksort

Hello All,

I'd like to share with you new Dual-Pivot Quicksort which is faster than the
known implementations (theoretically and experimental). I'd like to propose
to replace the JDK's Quicksort implementation by new one.

...

The new Dual-Pivot Quicksort uses *two* pivots elements in this manner:

1. Pick an elements P1, P2, called pivots from the array.


2. Assume that P1 <= P2, otherwise swap it.
3. Reorder the array into three parts: those less than the smaller pivot,
those larger than the larger pivot, and in between are those elements
between (or equal to) the two pivots.
4. Recursively sort the sub-arrays.

The invariant of the Dual-Pivot Quicksort is:

[ < P1 | P1 <= & <= P2 } > P2 ]

...

http://mail.openjdk.java.net/pipermail/core-libs-dev/2009-September/002630.html
54
A beautiful mailing list post (Yaroslavskiy-Bloch-Bentley, October 2009)

Replacement of quicksort in java.util.Arrays with new dual-pivot quicksort

Date: Thu, 29 Oct 2009 11:19:39 +0000


Subject: Replace quicksort in java.util.Arrays with dual-pivot implementation

Changeset: b05abb410c52
Author: alanb
Date: 2009-10-29 11:18 +0000
URL: http://hg.openjdk.java.net/jdk7/tl/jdk/rev/b05abb410c52

6880672: Replace quicksort in java.util.Arrays with dual-pivot implementation


Reviewed-by: jjb
Contributed-by: vladimir.yaroslavskiy at sun.com, joshua.bloch at google.com,
jbentley at avaya.com

! make/java/java/FILES_java.gmk
! src/share/classes/java/util/Arrays.java
+ src/share/classes/java/util/DualPivotQuicksort.java

http://mail.openjdk.java.net/pipermail/compiler-dev/2009-October.txt

55
Dual-pivot quicksort

Use two partitioning items p1 and p2 and partition into three subarrays:
Keys less than p . 1

Keys between p and p .


1 2

Keys greater than p . 2

< p1 p1 p1 and p2 p2 > p2

lo lt gt hi

Recursively sort three subarrays.

degenerates to Dijkstra's 3-way partitioning

Note. Skip middle subarray if p1 = p2.


56
Dual-pivot partitioning demo

Initialization.
Choose a[lo] and a[hi] as partitioning items.
Exchange if necessary to ensure a[lo] a[hi].

S E A Y R L F V Z Q T C M K

lo hi

exchange a[lo] and a[hi]


Dual-pivot partitioning demo

Main loop. Repeat until i and gt pointers cross.


If (a[i] < a[lo]), exchange a[i] with a[lt] and increment lt and i.

Else if (a[i] > a[hi]), exchange a[i] with a[gt] and decrement gt.
Else, increment i.

< p1 p1 p1 and p2 ? p2 > p2

lo lt i gt hi

K E A F R L M C Z Q T V Y S

lo lt i gt hi
Dual-pivot partitioning demo

Finalize.
Exchange a[lo] with a[--lt].
Exchange a[hi] with a[++gt].

< p1 p1 p1 and p2 p2 > p2

lo lt gt hi

C E A F K L M R Q S Z V Y T

lo lt gt hi

3-way partitioned
Dual-pivot quicksort

Use two partitioning items p1 and p2 and partition into three subarrays:
Keys less than p . 1

Keys between p and p .


1 2

Keys greater than p . 2

< p1 p1 p1 and p2 p2 > p2

lo lt gt hi

Now widely used. Java 7, Python unstable sort, Android,

60
Three-pivot quicksort

Use three partitioning items p1, p2, and p3 and partition into four subarrays:
Keys less than p . 1

Keys between p and p . 1 2

Keys between p and p . 2 3

Keys greater than p . 3

< p1 p1 p1 and p2 p2 p2 and p3 p3 > p3

lo a1 a2 a3 hi

Multi-Pivot Quicksort: Theory and Experiments


ight; see http://www.siam.org/journals/ojsa.php

Shrinu Kushagra Alejandro Lopez-Ortiz J. Ian Munro


skushagr@uwaterloo.ca alopez-o@uwaterloo.ca imunro@uwaterloo.ca
University of Waterloo University of Waterloo University of Waterloo
Aurick Qiao
a2qiao@uwaterloo.ca
University of Waterloo
November 7, 2013

Abstract In 2009, Vladimir Yaroslavskiy introduced a novel 61


Quicksort quiz 4

Why do 2-pivot (and 3-pivot) quicksort perform better than 1-pivot?

A. Fewer compares.

B. Fewer exchanges.

C. Fewer cache misses.

D. I don't know.

62
Quicksort quiz 4

Why do 2-pivot (and 3-pivot) quicksort perform better than 1-pivot?

A. Fewer compares.

B. Fewer exchanges. # entries scanned is a good proxy


for cache performance when
comparing quicksort variants
C. Fewer cache misses.

partitioning compares exchanges entries scanned

1-pivot 2 N ln N 0.333 N ln N 2 N ln N

median-of-3 1.714 N ln N 0.343 N ln N 1.714 N ln N

2-pivot 1.9 N ln N 0.6 N ln N 1.6 N ln N

3-pivot 1.846 N ln N 0.616 N ln N 1.385 N ln N

Reference: Analysis of Pivot Sampling in Dual-Pivot Quicksort by Wild-Nebel-Martnez

Bottom line. Caching can have a significant impact on performance.


beyond scope of this course
63
Which sorting algorithm to use?

Many sorting algorithms to choose from:

sorts algorithms

elementary sorts insertion sort, selection sort, bubblesort, shaker sort, ...

subquadratic sorts quicksort, mergesort, heapsort, shellsort, samplesort, ...

system sorts dual-pivot quicksort, timsort, introsort, ...

external sorts Poly-phase mergesort, cascade-merge, psort, .

radix sorts MSD, LSD, 3-way radix quicksort,

parallel sorts bitonic sort, odd-even sort, smooth sort, GPUsort, ...

64
Which sorting algorithm to use?

Applications have diverse attributes.


Stable?
Parallel?
In-place?
Deterministic?
Duplicate keys?
Multiple key types?
Linked list or arrays?
Large or small items? many more combinations of

Randomly-ordered array? attributes than algorithms

Guaranteed performance?

Q. Is the system sort good enough?


A. Usually.

65
System sort in Java 7

Arrays.sort().

Has method for objects that are Comparable.


Has overloaded method for each primitive type.
Has overloaded method for use with a Comparator.
Has overloaded methods for sorting subarrays.
Algorithms.
Dual-pivot quicksort for primitive types.
Timsort for reference types.

Q. Why use different algorithms for primitive and reference types?

66
Ineffective sorts

http://xkcd.com/1185 67

Das könnte Ihnen auch gefallen