程序代写代做代考 C ER algorithm AI Notes on Better Master Theorems

Notes on Better Master Theorems
for
Divide􏰀and􏰀Conquer Recurrences
􏰂 Intro duction
Divide􏰀and􏰀conquer recurrences are ubiquitous known for solving recurrences such as􏰑
in
the
analysis of algorithms􏰈
Many metho ds are
Tom
Mathematics Laboratory for
Leighton
Department and
Computer Science Massachusetts Institute of Technology
Cambridge􏰁 Massachusetts 􏰅􏰆􏰂􏰇􏰄
Octob er 􏰄􏰁 􏰂􏰄􏰄􏰃
Abstract
Techniques
Computer Science
known as the Master Metho d 􏰉􏰆􏰊􏰈 Recently􏰁 Akra and Bazzi 􏰉􏰂􏰊 discovered a surprisingly elegant generalization of the Master Metho d that yields a very simple formula for solving most divide􏰀 and􏰀conquer recurrence s􏰈 In these notes􏰁 we provide a simple inductive pro of of the Akra􏰀Bazzi result and we extend the result to handle variations of divide􏰀and􏰀conquer recurrences that commonly arise in practice􏰈
for solving divide􏰀and􏰀conquer recurrences are routinely
students each year􏰈 The dominant approach to solving such recurrence s is
T􏰑n􏰒 􏰙
􏰂
􏰆T 􏰑dn􏰙􏰆e􏰒 􏰓 O􏰑n􏰒 if n 􏰞 􏰂􏰜
but p erhaps the most widely taught approach is the Master Metho d that is describ ed in the seminal algorithms text by Cormen􏰁 Leiserson and Rivest 􏰉􏰆􏰊􏰈
The Master Metho d is fairly p owerful and results in a closed form solution for divide􏰀and􏰀conquer recurrences with a sp ecial 􏰑but commonly􏰀o ccurring􏰒 form􏰈 Recently Akra and Bazzi 􏰉􏰂􏰊 discovered a far more general solution to divide􏰀and􏰀conquer recurrences􏰈 The Akra􏰀Bazzi analysis is based on a sp ecial functional transform that they call the 􏰚order transform􏰈􏰐
In these notes􏰁 we give a simple inductive pro of of the Akra􏰀Bazzi result that is suitable for use in an undergraduate algorithms or discrete math class􏰈 We also show that the Akra􏰀Bazzi result holds for a more general class of recurrences that commonly arise in practice and that are often considered to b e di􏰏cult to solve􏰈
􏰂
if n 􏰙 􏰂
taught to
thousands of

􏰆 The Akra􏰀Bazzi Solution
We b egin with of the form􏰘
where􏰂
a
simple
inductive pro of of the
k
i􏰙􏰂 aiT 􏰑bix􏰒 􏰓 g􏰑x􏰒
number for which i􏰙􏰂 aibi 􏰙 􏰂􏰈 Then 􏰡
􏰡 Z x
g􏰑u􏰒 􏰢􏰢
du 􏰘 up􏰓􏰂
T􏰑x􏰒􏰙 P
􏰑 􏰋􏰑􏰂􏰒
for 􏰂 􏰦 x 􏰦 x 􏰅
for x 􏰞 x􏰅
Akra􏰀Bazzi result􏰈 The result
holds
for recurrences
􏰑􏰂􏰒
􏰂􏰈 x􏰧􏰂isarealnumber􏰁
􏰆􏰈 x􏰅isaconstantsuchthatx􏰅􏰧􏰂􏰙biandx􏰅􏰧􏰂􏰙􏰑􏰂􏰁bi􏰒for􏰂􏰦i􏰦k􏰁
􏰇􏰈 ai􏰞􏰅isaconstantfor􏰂􏰦i􏰦k􏰁
􏰔􏰈 bi􏰆􏰑􏰅􏰜􏰂􏰒isaconstantfor􏰂􏰦i􏰦k􏰁
􏰕􏰈 k 􏰧 􏰂 is a constant􏰁 and
􏰃􏰈 g􏰑x􏰒 is a nonnegative function that satis􏰍es the polynomial􏰀growth condition speci􏰍ed below􏰈
De􏰍nition􏰈 We say that g 􏰑x􏰒 satis􏰍es the polynomial􏰀growth condition if there exist positive con􏰀 stants c􏰂􏰁 c􏰆 such that for all x 􏰧 􏰂􏰁 for all 􏰂 􏰦 i 􏰦 k􏰁 and for all u 􏰆 􏰉bix􏰜 x􏰊􏰁
c􏰂g 􏰑x􏰒 􏰦 g 􏰑u􏰒 􏰦 c􏰆g 􏰑x􏰒􏰘
Remark􏰈 If jg 􏰅 􏰑x􏰒j is upp er b ounded by a p olynomial in x􏰁 then g 􏰑x􏰒 satis􏰍es the p olynomial􏰀 growth condition􏰈 For example􏰁 g 􏰑x􏰒 􏰙 x􏰌 log 􏰍 x satis􏰍es the p olynomial􏰀growth condition for any constants 􏰌􏰜 􏰍 􏰆 R􏰈
Theorem 􏰂 􏰑􏰉􏰂􏰊􏰒􏰈PGiven a recurrence of the form speci􏰍ed in Equation 􏰂􏰁 let p be the unique real kp
p T􏰑x􏰒􏰙􏰋 x 􏰂􏰓
􏰂
Examples􏰈
􏰛 If
􏰛 If
􏰛 If
􏰛 If
T 􏰑x􏰒 􏰙 T 􏰑x􏰒 􏰙 T 􏰑x􏰒 􏰙 T 􏰑x􏰒 􏰙 T 􏰑x􏰒 􏰙
􏰆T 􏰑x􏰙􏰔􏰒 􏰓 􏰇T 􏰑x􏰙􏰃􏰒 􏰓 􏰋􏰑x log x􏰒􏰁 then
p 􏰙 􏰂
and T 􏰑x􏰒 􏰙 􏰋􏰑x log􏰆 x􏰒􏰈
􏰆T 􏰑x􏰙􏰆􏰒 􏰓 􏰗 T 􏰑􏰇x􏰙􏰔􏰒 􏰓 􏰋􏰑x􏰆􏰙 log x􏰒􏰁 then p 􏰙 􏰆 and T 􏰑x􏰒 􏰙 􏰋􏰑x􏰆􏰙 log log x􏰒􏰈 􏰄
T 􏰑x􏰙􏰆􏰒 􏰓 􏰋􏰑log x􏰒􏰁 then p 􏰙 􏰅 and T 􏰑x􏰒 􏰙 􏰋􏰑log􏰆 x􏰒􏰈
􏰂 T 􏰑x􏰙􏰆􏰒 􏰓 􏰋􏰑􏰂􏰙x􏰒􏰁 then p 􏰙 􏰁􏰂 and T 􏰑x􏰒 􏰙 􏰋􏰑􏰑log x􏰒􏰙x􏰒􏰈 􏰆
􏰛 If
These conditions are somewhat less restrictive than those of 􏰉􏰂􏰊􏰈
􏰔T 􏰑x􏰙􏰆􏰒 􏰓 􏰋􏰑x􏰒􏰁 then p 􏰙 􏰆 and T 􏰑x􏰒 􏰙 􏰋􏰑x􏰆􏰒􏰈
􏰂
􏰆

The pro of of Theorem 􏰂 makes use of the following simple lemma from calculus􏰈
Lemma
there are positive
Proof􏰈
From the
p olynomiaZl􏰀growth
bi x
condition we know that
􏰂􏰈 If g 􏰑x􏰒 is a nonnegative function that satis􏰍es the polynomial􏰀growth condition􏰁 then constants c􏰇 􏰁 c􏰔 such that Zfor 􏰂 􏰦 i 􏰦 k and al l x 􏰧 􏰂􏰁
xp
x g 􏰑u􏰒
up􏰓􏰂
du 􏰦 xp􏰑x 􏰁 b x􏰒 i
c􏰆 g 􏰑x􏰒
minf􏰑b x􏰒p􏰓􏰂 􏰜 xp􏰓􏰂 g i
x g 􏰑u􏰒
c g􏰑x􏰒 􏰦 xp du 􏰦 c g􏰑x􏰒􏰘 􏰇 up􏰓􏰂 􏰔
where we
de􏰍ne c􏰔
to b e
a constant
for
which
bi x
􏰑􏰂 􏰁 bi 􏰒c􏰆
􏰙
􏰦 c􏰔 g 􏰑x􏰒
g 􏰑x􏰒
min f􏰂􏰜 bp􏰓􏰂g i
c􏰔 􏰧 􏰑􏰂 􏰁 bi 􏰒c􏰆
minf􏰂􏰜 bp􏰓􏰂 g i
for 􏰂 􏰦 i 􏰦 k􏰈 Similarly􏰁
Z
bi x
where we
de􏰍ne c􏰇
constant
􏰧 c􏰇g 􏰑x􏰒
for which
xp
to b e a
du 􏰧 xp􏰑x 􏰁 b x􏰒 i
maxf􏰑b
x􏰒p􏰓􏰂􏰜 xp􏰓􏰂 g
x
g 􏰑u􏰒 up􏰓􏰂
c􏰂g 􏰑x􏰒
􏰙
􏰑􏰂 􏰁 bi 􏰒c􏰂
maxf􏰂􏰜 bp􏰓􏰂g i
g 􏰑x􏰒
i
c􏰇 􏰦 􏰑􏰂 􏰁 bi 􏰒c􏰆
maxf􏰂􏰜 bp􏰓􏰂g i
for 􏰂 􏰦 i 􏰦 k􏰈
We will use induction
into intervals I􏰅 􏰙 􏰉􏰂􏰜 x􏰅􏰊 and Ij 􏰙 􏰑x􏰅 􏰓 j 􏰁 􏰂􏰜 x􏰅 􏰓 j􏰊 for j 􏰧 􏰂􏰈
to
prove Theorem 􏰂􏰁 and so it will
b e
helpful
to partition the
domain
of x
By the de􏰍nition of x􏰅􏰁 we know that if x 􏰆 Ij for some j 􏰧 􏰂􏰁 then for 􏰂 􏰦 i 􏰦 k􏰁
bix 􏰆 Ij􏰅 for
some j􏰅 􏰝 j􏰈 This is because b x 􏰞 b 􏰑x 􏰓 j 􏰁 􏰂􏰒 􏰧 b x 􏰧 􏰂􏰁 and because b x 􏰦 b 􏰑x 􏰓 j􏰒 􏰦 ii􏰅i􏰅ii􏰅
x􏰅 􏰓 j 􏰁 􏰑􏰂 􏰁 bi􏰒x􏰅 􏰦 x􏰅 􏰓 j 􏰁 􏰂􏰈 As a consequence􏰁 we know that the value of T in any interval after 􏰉􏰂􏰜 x􏰅 􏰊 dep ends only on the values of T in prior intervals􏰈
􏰇

Proof of Theorem 􏰂􏰈 We 􏰍rst show that there is a p ositive constant
c􏰕 such that for all x 􏰞 x􏰅 􏰁
􏰡 Zxg􏰑u􏰒 􏰢 T􏰑x􏰒􏰧c xp 􏰂􏰓 du 􏰘
􏰂 up􏰓􏰂
on the interval Ij containing x􏰈 when x 􏰆 􏰉􏰂􏰜 x􏰅􏰊 􏰑provided that we
The inductive step is argued as follows􏰘 Xk
The
the fact that T 􏰑x􏰒 􏰙 􏰋􏰑􏰂􏰒
case when j 􏰙 􏰅 follows from small enough􏰒􏰈
pro of is by induction
The base cho ose c􏰕
T 􏰑x􏰒 􏰙
􏰧 􏰙
􏰧
􏰙
􏰙
􏰧cxp 􏰂􏰓 􏰕
i􏰙􏰂 Xk
i􏰙􏰂
a c 􏰑b x􏰒p i􏰕i
􏰡 Zbixg􏰑u􏰒 􏰢
ai T 􏰑bi x􏰒 􏰓 g 􏰑x􏰒
􏰂 􏰓
up􏰓􏰂 􏰂
du 􏰓 g􏰑x􏰒
􏰑by induction􏰒
􏰑by Lemma 􏰂􏰒
􏰕
c􏰕 xp
c􏰕 xp
cxp 􏰕
Xk 􏰡 Zxg􏰑u􏰒 Zxg􏰑u􏰒 􏰢 aibpi 􏰂 􏰓 p􏰓􏰂 du 􏰁 p􏰓􏰂 du
i􏰙􏰂 􏰂u bixu
Xk 􏰡 Zxg􏰑u􏰒 c 􏰢
􏰓 g􏰑x􏰒
aibp 􏰂 􏰓 i
up􏰓􏰂
du 􏰁 􏰔 g􏰑x􏰒 xp
􏰓 g􏰑x􏰒
􏰡Z􏰂􏰢
i􏰙􏰂
x g 􏰑u􏰒 c􏰔
du 􏰁 g􏰑x􏰒 􏰓 g􏰑x􏰒
􏰂 􏰓
􏰡 Z􏰂up􏰓􏰂 􏰢xp
that c􏰕 􏰦 􏰂􏰙c􏰔 􏰈
provided
The pro of that there is a
p ositive constant c􏰃
T􏰑x􏰒􏰦c xp 􏰃
such that
for all
􏰞 x􏰅 􏰁
x g 􏰑u􏰒
􏰕 up􏰓􏰂 􏰕􏰔
cxp
du 􏰓 g􏰑x􏰒 􏰁 c c
􏰂 􏰓 􏰡Z􏰂􏰢
g􏰑x􏰒
x g 􏰑u􏰒 up􏰓􏰂
du
􏰂
x
􏰡 Zxg􏰑u􏰒 􏰢
􏰂􏰓 du 􏰂 up􏰓􏰂
is nearly identical􏰈 We need only insure that c􏰃 is chosen large enough so that the base case is
satis􏰍ed and so
as claimed􏰈
Remark􏰈 If
do es not necessarily hold if g 􏰑x􏰒 do es not satisfy the p olynomial􏰀growth condition􏰈
that c􏰃 􏰧 􏰂􏰙c􏰇 􏰈 As a consequence􏰁 we can conclude that
􏰡 􏰡 Zxg􏰑u􏰒 􏰢􏰢
p
T􏰑x􏰒􏰙􏰋 x 􏰂􏰓 du 􏰜
􏰂 up􏰓􏰂
g 􏰑x􏰒 grows faster than any p olynomial in x􏰁 then T 􏰑x􏰒 􏰙 􏰋􏰑g 􏰑x􏰒􏰒􏰈 Hence􏰁 Theorem 􏰂
􏰔

􏰇 Variations
Although the class of recurrences analyized in Section 􏰆 is quite broad􏰁 recurrences that arise in practice often di􏰌er in small ways from the class sp eci􏰍ed in Equation 􏰂􏰈 For example􏰁 in algorithm design􏰁 recurrences of the form
are common􏰈
Generally sp eaking􏰁 the inclusion
Xk
T 􏰑x􏰒 􏰦 ai T 􏰑dbi xe􏰒 􏰓 g 􏰑x􏰒
i􏰙􏰂
of 􏰎o ors and ceilings in a recurrence do es not signi􏰍cantly change
for all
where
􏰂􏰈 x􏰁 x􏰅 􏰁
tend to b e fairly tedious and of variations 􏰑which includes ceilings􏰒 and we show that the variations in this class do not a􏰌ect the solution of the
the nature
sp ecialized
􏰎o ors and
recurrence 􏰑up to constant factors􏰒􏰈 In particular􏰁 we show that the solution of Theorem 􏰂 holds
of the solution 􏰑e􏰈g􏰈􏰁 see 􏰉􏰂􏰁 􏰆􏰊􏰒􏰁 but the pro ofs of this fact in nature􏰈 In what follows􏰁 we describ e a general class
recurrences
of the form􏰑􏰘 􏰋􏰑􏰂􏰒
ai 􏰁
bi 􏰁
k 􏰁 and g 􏰑x􏰒 all satisfy the conditions sp eci􏰍ed in Section 􏰆􏰁
T􏰑x􏰒􏰙 P k
for 􏰂 􏰦 x 􏰦 x
􏰅 􏰑􏰆􏰒
for x 􏰞 x􏰅
i􏰙􏰂 aiT 􏰑bix 􏰓 hi􏰑x􏰒􏰒 􏰓 g􏰑x􏰒
􏰆􏰈 there is some constant 􏰛 􏰞 􏰅 for which jh 􏰑x􏰒j 􏰦 x􏰙􏰑log􏰂􏰓􏰛 x􏰒 for 􏰂 􏰦 i 􏰦 k whenever x 􏰧 x 􏰁 i􏰅
􏰇􏰈 there exist p ositive constants c􏰂 and c􏰆 such that for all x 􏰧 􏰂􏰁 for all 􏰂 􏰦 i 􏰦 k 􏰁 and for all
u 􏰆 􏰉bix 􏰓 hi 􏰑x􏰒􏰜 x􏰊􏰁
c􏰂g 􏰑x􏰒 􏰦 g 􏰑u􏰒 􏰦 c􏰆 g 􏰑x􏰒􏰜
and
􏰔􏰈 x is chosen to be a large enough constant􏰆 so that for any i 􏰦 k and any x 􏰧 x 􏰁 􏰅􏰅􏰂􏰅
􏰡􏰢􏰥A 􏰑a􏰒􏰂􏰁􏰂p􏰂􏰓􏰟􏰂 􏰠􏰧􏰂􏰓􏰂􏰁
bi log􏰂􏰓􏰛 x 􏰅 log􏰛􏰙􏰆 b x 􏰓 x 􏰂 log􏰛􏰙􏰆 x i 􏰂􏰓􏰛
􏰡 􏰢 logx
p􏰥A
􏰑b􏰒 􏰂􏰓 􏰂 􏰂􏰁 􏰟 􏰂 􏰠 􏰦􏰂􏰁 􏰂 􏰁
bi log􏰂􏰓􏰛 x log􏰛􏰙􏰆 b x 􏰓 x log􏰛􏰙􏰆 x 􏰡 􏰢 ilog􏰂􏰓􏰛x
􏰂􏰂
􏰑c􏰒 􏰡􏰂􏰓 􏰛􏰙􏰆􏰢􏰦􏰂􏰁
􏰆 logx
􏰑d􏰒􏰆􏰂􏰁􏰂 􏰧􏰂􏰈 􏰛􏰙􏰆
􏰆
log x
Such a constant value of x􏰅 can b e shown to exist using standard Taylor series expansions and asymptotic analysis􏰈
􏰕

For example􏰁 we might cho ose hi so that
hi 􏰑x􏰒 􏰙 dbi xe 􏰁 bix􏰜
thereby extending Theorem 􏰂 to handle ceiling functions􏰈 In this case􏰁 jh 􏰑x􏰒j 􏰝 􏰂􏰈 We can also use pi
much larger functions􏰁 however􏰈 For example􏰁 we could set h 􏰑x􏰒 􏰙 􏰁 x or h 􏰑x􏰒 􏰙 x􏰙􏰑log􏰆 x􏰒 for ii
x 􏰞 􏰂􏰈
To analyze the more general recurrence􏰁 we will need the following analogue of Lemma 􏰂􏰈
There are positive constants cZ 􏰁 c such that for 􏰂 􏰦 i 􏰦 k and all x 􏰧 􏰂􏰁 􏰇􏰔
x
Lemma 􏰆􏰈
Proof􏰈 of the
TheoremP􏰆􏰈 Given a recurrence of the form speci􏰍ed in Equation 􏰆􏰁 let p be the unique real number kp
􏰡 􏰡 Zxg􏰑u􏰒 􏰢􏰢
p
T􏰑x􏰒􏰙􏰋 x 􏰂􏰓 du 􏰘
􏰂 up􏰓􏰂
Proof􏰈 The pro of is very similar to that of Theorem 􏰂􏰈 The main di􏰌erence is that we use a slightly stronger inductive hyp othesis􏰈 In particular􏰁 we b egin by showing that there is a p ositive constant c􏰕
c g􏰑x􏰒 􏰦 xp
􏰇 up􏰓􏰂􏰔
bi x􏰓hi 􏰑x􏰒
g 􏰑u􏰒
pro of is identical to that for Lemma 􏰂 except that we use constraint 􏰇 ab ove in place
The
p olynomial􏰀growth condition of Section 􏰆􏰈
for which i􏰙􏰂 ai bi 􏰙 􏰂􏰈 Then
du 􏰦 c g􏰑x􏰒􏰘
such that for all x 􏰞 x􏰅 􏰁
􏰡 􏰂􏰢􏰡Zxg􏰑u􏰒􏰢 T􏰑x􏰒􏰧c xp 􏰂􏰓 􏰂􏰓 du 􏰘
fact that T 􏰑x􏰒 􏰙 􏰋􏰑􏰂􏰒 when x 􏰆 􏰉􏰂􏰜 x􏰅 􏰊 􏰑provided that c􏰕
is chosen to b e a small enough constant􏰒􏰈
􏰕
log 􏰛􏰙􏰆 x 􏰂 up􏰓􏰂
The pro of is by induction on the interval Ij containing x􏰈 The base case when j 􏰙 􏰅 follows from the
􏰃

The inductive step is argued as follows􏰘 Xk
T 􏰑x􏰒 􏰙
􏰧
i􏰙􏰂
ai T 􏰑bix 􏰓 hi 􏰑x􏰒􏰒 􏰓 g 􏰑x􏰒 􏰣􏰤
Xk
i􏰕ii
􏰂
log 􏰛􏰙􏰆 􏰑bi x 􏰓 hi 􏰑x􏰒􏰒
a c 􏰑b x 􏰓 h 􏰑x􏰒􏰒p
􏰂 􏰓
i􏰙􏰂
􏰣Z􏰤 bi x􏰓hi 􏰑x􏰒 g 􏰑u􏰒
􏰋
􏰂 􏰓
p􏰓􏰂 du 􏰓 g 􏰑x􏰒
􏰑by induction􏰒
􏰂u􏰅 􏰂 Xk􏰡􏰢p􏰥 A
􏰧abpcxp􏰂􏰁􏰂 􏰂􏰓􏰟􏰂 􏰠 ii􏰕 􏰂􏰓􏰛 􏰛􏰙􏰆 x
i􏰙􏰂
bi log x log bix 􏰓 􏰂􏰓􏰛 log x
􏰣ZxZx􏰤
􏰂 􏰓 p􏰓􏰂 􏰂 u
bi x􏰓hi 􏰑x􏰒
u
g 􏰑u􏰒
g 􏰑u􏰒
du 􏰁
Xk 􏰡􏰂􏰢􏰡Zxg􏰑u􏰒c􏰢
􏰋
􏰧 aibpc􏰕xp 􏰂 􏰓
i􏰙􏰂
􏰡 􏰂 􏰢􏰡 Zxg􏰑u􏰒
􏰙 cxp􏰂􏰓 􏰂􏰓
􏰕 􏰛􏰙􏰆 up􏰓􏰂
p􏰓􏰂 du 􏰓 g􏰑x􏰒
􏰑by the
bounds on h􏰒
􏰧cxp 􏰂􏰓 􏰕
􏰂
log 􏰛􏰙􏰆 x
x g􏰑u􏰒
The pro of of the
case􏰁 we
show by
induction that
there is a
positive
constant
c􏰃 such
x 􏰞 x􏰅 􏰁
i log􏰛􏰙􏰆x
􏰂 􏰓 du 􏰁 􏰔 g􏰑x􏰒 􏰂 up􏰓􏰂 xp
􏰓 g􏰑x􏰒
􏰑by constraint 􏰔􏰑a􏰒 on
􏰡 􏰂 􏰢􏰡 Zxg􏰑u􏰒 c 􏰢
􏰙 c􏰕xp 􏰂􏰓 􏰂􏰓
log􏰛􏰙􏰆 x 􏰂 up􏰓􏰂
du 􏰁
􏰢
􏰔 g􏰑x􏰒 xp
􏰓 g􏰑x􏰒 􏰡
􏰂 􏰢
􏰑by upp er b ound
x􏰅 􏰒􏰈
􏰂􏰓
􏰂
du up􏰓􏰂
x􏰅 and Lemma 􏰆􏰒
du 􏰓g􏰑x􏰒􏰁c c
􏰂􏰓 g􏰑x􏰒
􏰡 log x􏰢􏰡 Z􏰂
􏰢
􏰕􏰔 􏰛􏰙􏰆 log x
provided that c􏰕 􏰦 􏰂􏰙􏰑􏰆c􏰔 􏰒
constraint 􏰔􏰑c􏰒 on
is quite similar􏰈 In this
that for all
T􏰑x􏰒􏰦c xp 􏰃
􏰡 􏰂􏰢􏰡Zxg􏰑u􏰒􏰢
􏰂􏰁 􏰂􏰓 du 􏰘 log 􏰛􏰙􏰆 x 􏰂 up􏰓􏰂
􏰖

The base case is as b efore􏰈 The inductive step is argued as
Xk
T 􏰑x􏰒 􏰙 ai T 􏰑bix 􏰓 hi 􏰑x􏰒􏰒 􏰓 g 􏰑x􏰒
follows􏰘
􏰦
i􏰙􏰂 􏰣 􏰤 Xk 􏰂
a c 􏰑b x 􏰓 h 􏰑x􏰒􏰒p 􏰂 􏰁 i􏰃ii
log 􏰛􏰙􏰆 􏰑bi x 􏰓 hi 􏰑x􏰒􏰒
i􏰙􏰂
􏰣Z􏰤 bi x􏰓hi 􏰑x􏰒 g 􏰑u􏰒
􏰋
􏰂 􏰓
p􏰓􏰂 du 􏰓 g 􏰑x􏰒
􏰑by induction􏰒
􏰂u􏰅 􏰂 Xk􏰡􏰢p􏰥 A
􏰦abpcxp􏰂􏰓􏰂 􏰂􏰁􏰟􏰂 􏰠 ii􏰃 􏰂􏰓􏰛
i􏰙􏰂
bi log x log􏰛􏰙􏰆 b x 􏰓 x
i 􏰂􏰓􏰛
􏰋
􏰦 aibpc􏰃xp 􏰂 􏰁
i􏰙􏰂
p􏰓􏰂 du 􏰓 g􏰑x􏰒
􏰑by the
bounds
on h􏰒
log x 􏰣ZxZx􏰤
􏰂 􏰓 p􏰓􏰂 􏰂 u
bi x􏰓hi 􏰑x􏰒
u
g 􏰑u􏰒
g 􏰑u􏰒
du 􏰁
Xk 􏰡􏰂􏰢􏰡Zxg􏰑u􏰒c􏰢
i log􏰛􏰙􏰆x
􏰂 􏰓 du 􏰁 􏰇 g􏰑x􏰒 􏰂 up􏰓􏰂 xp
􏰓 g􏰑x􏰒
􏰑by constraint 􏰔􏰑b􏰒 on
􏰡 􏰂 􏰢􏰡 Zxg􏰑u􏰒 c 􏰢
du 􏰁
􏰙 cxp􏰂􏰁 􏰂􏰓
􏰇 g􏰑x􏰒 xp
x􏰅 and Lemma 􏰆􏰒
􏰃
􏰓 g􏰑x􏰒 􏰡
􏰂􏰁 g􏰑x􏰒 􏰇􏰃 􏰛􏰙􏰆
􏰡 􏰡 Zxg􏰑u􏰒 􏰢􏰢
T 􏰑x􏰒 􏰙 􏰋 xp 􏰂 􏰓 du 􏰜 􏰂 up􏰓􏰂
log􏰛􏰙􏰆 x 􏰂 up􏰓􏰂
􏰢
du 􏰓g􏰑x􏰒􏰁c c
􏰂 􏰢
􏰡 􏰂 􏰢􏰡 Zxg􏰑u􏰒
􏰙 cxp􏰂􏰁 􏰂􏰓
􏰃 􏰛􏰙􏰆 up􏰓􏰂
log x
􏰡 log x􏰢􏰡 Z􏰂 􏰢 􏰂 x g􏰑u􏰒
􏰦cxp 􏰂􏰁 􏰂􏰓 du
log 􏰛􏰙􏰆 x 􏰂 up􏰓􏰂
provided that c􏰃 􏰧 􏰆􏰙c􏰇 Hence􏰁 we can conclude that
as desired􏰈
􏰃
􏰑by constraint 􏰔􏰑d􏰒 on x􏰅 􏰒􏰈
Remark􏰈 It is worth noting that the x􏰙 log􏰂􏰓􏰛 x limit on the size of jh 􏰑x􏰒j is nearly tight􏰁 since i
the solution
of the
recurrence 􏰑 T􏰑x􏰒 􏰙
􏰋􏰑􏰂􏰒 􏰟􏰠
􏰆T x 􏰓 x
􏰆 log x
for 􏰂 􏰦 x 􏰦 x􏰅 for x 􏰞 x􏰅
is T 􏰑x􏰒 􏰙 x log􏰋􏰑􏰂􏰒 x􏰁 which is di􏰌erent than the solution of 􏰋􏰑x􏰒 for the recurrence without the x􏰙 log x term􏰈
􏰗

References
􏰉􏰂􏰊 M􏰈 Akra and L􏰈 Bazzi􏰈 􏰚On the solution of linear recurrence equations􏰈􏰐 To app ear􏰁 􏰂􏰄􏰄􏰃􏰈
􏰉􏰆􏰊 Thomas H􏰈 Cormen􏰁 Charles E􏰈 Leierson􏰁 and Ronald L􏰈 Rivest􏰈 Introduction to Algorithms􏰈 The MIT Press􏰁 Cambridge􏰁 Massachusetts􏰁 􏰂􏰄􏰄􏰅􏰈
􏰄