BWT-Based Compression Algorithms Haim Kaplan and Elad Verbin Tel-Aviv University Presented in CPM ’07, July 8, 2007.

BWT-Based Compression Algorithms

Haim Kaplan and Elad VerbinTel-Aviv University

Presented in CPM ’07, July 8 , 2007

Results

• Cannot show constant c<2 s.t.

• Similarly,• no c<1.26 for BWRL

• no c<1.3 for BWDC

• Probabilistic technique.

0. 0( ) ( ) ( )s BW s c nH s o n

Outline● Part I: Definitions● Part II: Results● Part III: Proofs● Part IV: Experimental Results

Part I: Definitions

BW0The Main Burrows-Wheeler Compression

Algorithm:

Compressed String S’

String S BWTBurrows-Wheeler Transfor

m

MTFMove-to-

front

Order-0 Encoding

Text with local uniformity

Text in English (similar contexts -> similar character)

Integer string with many small numbers

The BWT

● Invented by Burrows-and-Wheeler (‘94)● Analogous to Fourier Transform (smooth!)

string with context-regularity

BWT

string with spikes (close repetitions)

s

s

mississippi

ipssmpissii

[Fenwick]

p i#mississi pp pi#mississ is ippi#missi ss issippi#mi ss sippi#miss is sissippi#m i

i ssippi#mis s

m ississippi #i ssissippi# m

The BWTT = mississippi#

mississippi#ississippi#mssissippi#mi sissippi#mis

sippi#missisippi#mississppi#mississipi#mississipi#mississipp#mississippi

ssippi#missiissippi#miss Sort the rows

# mississipp ii #mississip pi ppi#missis s

F L=BWT(T)

T

BWT sorts the characters by their post-context

BWT Facts

1. permutes the text2. (≤n+1)-to-1 function

Move To Front

● By Bentley, Sleator, Tarjan and Wei (’86)

string with spikes (close repetitions)

ipssmpissii

integer string with small numbers

0,0,0,0,0,2,4,3,0,1,0

move-to-front

s

's

Move to Front

a,b,r,c,dabracadabra

Move to Front

a,b,r,c,d0abracadabraa,b,r,c,dabracadabra

Move to Front

b,a,r,c,d0,1abracadabraa,b,r,c,d0abracadabraa,b,r,c,dabracadabra

Move to Front

r,b,a,c,d0,1,2abracadabrab,a,r,c,d0,1abracadabraa,b,r,c,d0abracadabraa,b,r,c,dabracadabra

Move to Front

a,r,b,c,d0,1,2,2abracadabrar,b,a,c,d0,1,2abracadabrab,a,r,c,d0,1abracadabraa,b,r,c,d0abracadabraa,b,r,c,dabracadabra

Move to Front

c,a,r,b,d0,1,2,2,3abracadabraa,r,b,c,d0,1,2,2abracadabrar,b,a,c,d0,1,2abracadabrab,a,r,c,d0,1abracadabraa,b,r,c,d0abracadabraa,b,r,c,dabracadabra

Move to Front

a,c,r,b,d0,1,2,2,3,1abracadabrac,a,r,b,d0,1,2,2,3abracadabraa,r,b,c,d0,1,2,2abracadabrar,b,a,c,d0,1,2abracadabrab,a,r,c,d0,1abracadabraa,b,r,c,d0abracadabraa,b,r,c,dabracadabra

Move to Front

0,1,2,2,3,1,4,1,4,4,2abracadabraa,c,r,b,d0,1,2,2,3,1abracadabrac,a,r,b,d0,1,2,2,3abracadabraa,r,b,c,d0,1,2,2abracadabrar,b,a,c,d0,1,2abracadabrab,a,r,c,d0,1abracadabraa,b,r,c,d0abracadabraa,b,r,c,dabracadabra

After MTF● Now we have a string with small numbers:

lots of 0s, many 1s, …● Skewed frequencies: Run Arithmetic!

Character frequencies

BW0The Main Burrows-Wheeler Compression

Algorithm:

Compressed String S’

String S BWTBurrows-Wheeler Transfor

m

MTFMove-to-

front

Order-0 Encoding

Text with local uniformity

Text in English (similar contexts -> similar character)

Integer string with many small numbers

BWRL (e.g. bzip)

Compressed String S’

String S

BWTBurrows-Wheeler Transfor

m

MTFMove-to-

front

? RLE

Run-Length encodi

ng

Order-0 Encoding

Many more BWT-based algorithms

● BWDC: Encodes using distance coding instead of MTF

● BW with inversion frequencies coding● Booster-Based [Ferragina-Giancarlo-

Manzini-Sciortino]● Block-based compressor of Effros et al.

order-0 entropy

Lower bound for compression without context information

n

nnsnH log)(0

S=“ACABBA”

1/2 `A’s: Each represented by 1 bit

1/3 `B’s: Each represented by log(3) bits

1/6 `C’s: Each represented by log(6) bits

6*H0(S)=3*1+2*log(3)+1*log(6)

order-k entropy

= Lower bound for compression with order-k contexts

)(snH k

)()( 0 sw

sk wHwsnHk

order-k entropy

mississippi:Context for i: “mssp”Context for s: “isis”Context for p: “ip”

6)"("4 0 msspH4)"("4 0 isisH

2)"("2 0 ipH

1(" ") 12nH mississippi

0 (" ") 20.03nH mississippi

Part II: Results

Measuring against Hk

● When performing worst-case analysis of lossless text compressors, we usually measure against Hk

● The goal – a bound of the form:|A(s)|≤ c×nHk(s)+lower order term

● Optimal: |A(s)|≤ nHk(s)+lower order term

Bounds

lower Upper

BW0 2 [KaplanVerbin07] 3.33 [ManziniGagie07]

BWDC 1.3 [KaplanVerbin07] 1.7 [KaplanLandauVerbin06]

BWRL 1.26 [KaplanVerbin07] 5 [Manzini99]

gzip 1 1

PPM 1 1

Bounds

lower Upper

BW0 2 [KaplanVerbin07] 3.33 [ManziniGagie07]

BWDC 1.3 [KaplanVerbin07] 1.7 [KaplanLandauVerbin06]

BWRL 1.26 [KaplanVerbin07] 5 [Manzini99]

gzip 1 1

PPM 1 1

a

. 0( ) 2 ( ) (1)ks k BW s nH s o

Bounds

lower Upper

BW0 2 [KaplanVerbin07] 3.33 [ManziniGagie07]

BWDC 1.3 [KaplanVerbin07] 1.7 [KaplanLandauVerbin06]

BWRL 1.26 [KaplanVerbin07] 5 [Manzini99]

gzip 1 1

PPM 1 1

Bounds

lower

BW0 2 [KaplanVerbin07]

BWDC 1.3 [KaplanVerbin07]

BWRL 1.26 [KaplanVerbin07]

gzip 1

PPM 1

Surprising!! Since BWT-based compressors

work better than gzip in practice!

Possible Explanations

1. Asymptotics:

and real compressors cut into blocks, so

2. English Text is not Markovian!• Analyzing on different model might show BWT's

superiority

( ) ( ) ( / log )

( ) 1.7 ( ) (log )

k

DC k

gzip s nH s O n n

BW s nH s O n

n

Part III: Proofs

Lower bound● Wish to analyze BW0=BWT+MTF+Order0● Need to show s s.t.● Consider string s: 103 `a', 106 `b'

Entropy of s● BWT(s):

same frequencies MTF(BWT(s)) has: 2*103 `1', 106-103 `0‘ Compressed size: about

3 310 log(10 )

3 32 10 log(10 / 2)

need BWT(s) to have many isolated `a’s

00( ) 2 ( ) (1)BW s nH s o

many isolated `a’s● Goal: find s such that in BWT(s), most `a’s are

isolated● Solution: probabilistic.

BWT is (≤n+1)-to-1 function.● A random string s’ has ≥1/(n+1) chance of being a

BWT-image● A random string has ≥1-1/n2 chance of having “many”

isolated `a’s Therefore, such a string exists

General Calculation● s contains pn `a’s, (1-p)n `b’s.

Entropy of s:● MTF(BWT(s)) contains 2p(1-p)n `1’s, rest `0’s

compressed size of MTF(BWT(s)):

● Ratio:

1 1log (1 ) log

1p p

p p

12 (1 ) log

2 (1 )p p

p p

0

12 (1 ) log

2 (1 )2

1 1log (1 ) log

1

p

p pp p

p pp p

Lower bounds on BWDC, BWRL

● Similar technique. p infinitesimally small gives compressible string. So maximize ratio over p.

● Gives weird constants, but quite strong

Experimental Results

• Sanity Check: Picking texts from above Markov models really shows behavior in practice

• Picking text from “realistic” Markov sources also shows non-optimal behavior• (“realistic” = generated from actual texts)

• On long Markov text, gzip works better than BWT

Bottom Line● BWT compressors are not optimal

(vs. order-k entropy) ● We believe that they are good since English text is

not Markovian.● Find theoretical justification!

● also improve constants, find BWT algs with better ratios, ...

Thank You!

Additional Slides (taken out for lack of time)

BWT - Invertibility

● Go forward, one character at a time

Main Property: L F mapping● The ith occurrence of c in L

corresponds to the ith occurrence of c in F.

● This happens because the characters in L are sorted by their post-context, and the occurrences of character c in F are sorted by their post-context.

p i#mississi pp pi#mississ is ippi#missi ss issippi#mi ss sippi#miss is sissippi#m i

i ssippi#mis s

m ississippi #i ssissippi# m

# mississipp ii #mississip pi ppi#missis s

F Lunknown

BW0 vs. Lempel-Ziv

● BW0 dynamically takes advantage of context-regularity

● Robust, smooth, alternative for Lempel-Ziv

BW0 vs. Statistical Coding● Statistical Coding (e.g. PPM):

Builds a model for each context Prediction -> Compression

Exploits similarities between similar contexts

Optimally models each context

Explicit partitioning – produces a model for

each context

No explicit partitioning to contexts

PPMBW0

Compressed Text Indexing

● Application of BWT● Compressed representation of text, that

supports: fast pattern matching (without

decompression!) Partial decompression

● So, no need to ever decompress! space usage: |BW0(s)|+o(n)

● See more in [Ferragina-Manzini]

Musings

● On one hand: BWT based algorithms are not optimal, while Lempel-Ziv is.

● On the other hand: BWT compresses much better

● Reasons:

1. Results are Asymptotic. (EE reason)

2. English text was not generated by a Markov source (real reason?)

● Goal: Get a more honest way to analyze● Use a statistic different than Hk?