ASYMMETRIC NUMERAL SYSTEMS: ADDING FRACTIONAL BITS TO HUFFMAN CODER

Size: px

Start display at page:

Download "ASYMMETRIC NUMERAL SYSTEMS: ADDING FRACTIONAL BITS TO HUFFMAN CODER"

Meghan Kerrie Horn
5 years ago
Views:

1 ASYMMETRIC NUMERAL SYSTEMS: ADDING FRACTIONAL BITS TO HUFFMAN CODER Huffman coding Arithmetic coding fast, but operates on integer number of bits: approximates probabilities with powers of ½, getting inferior compression rate accurate, but many times slower (computationally more demanding) Asymmetric Numeral Systems (ANS) accurate and faster than Huffman we construct low state automaton optimized for given probability distribution Jarek Duda Kraków,

2 We need n bits of information to choose one of 2 n possibilities. For length n 0/1 sequences with pn of 1, how many bits we need to choose one? Entropy coder: encodes sequence with (p i ) i=1..m probability distribution using asymptotically at least H = m i=1 p i lg(1/p i ) bits/symbol (H lg(m)) Seen as weighted average: symbol of probability p contains lg(1/p) bits. Encoding (p i ) distribution with entropy coder optimal for (q i ) distribution costs ΔH = p i lg(1/q i ) p i lg(1/p i ) = p i lg ( p i ) 1 i i q i ln (4) i p i more bits/symbol - so called Kullback-Leibler distance (not symmetric). i (p i q i ) 2 2

Symbol frequencies from a version of the British National Corpus containing 67 million words ( http://www.yorku.ca/mack/uist2011.html#f1 ): Uniform: lg(27) 4.

3 Symbol frequencies from a version of the British National Corpus containing 67 million words ( ): Uniform: lg(27) bits/symbol (x 27x + s) Huffman uses: H = i p i r i bits/symbol Shannon: H = i p i lg(1/p i ) bits/symbol We can theoretically improve by 1% here ΔH bits/symbol order 1 Markov: ~3.3 bits/symbol order 2: ~3.1 bits/symbol, word: ~2.1 bits/symbol Currently the best text compression: durlica kingsize ( ) 10 9 bytes of text from Wikipedia (enwik9) into bytes: 1bit/symbol (lossy) video compression: ~ 1000 reduction 3

4 Huffman codes encode symbols as bit sequences Perfect if symbol probabilities are powers of 2 e.g. p(a) =1/2, p(b)=1/4, p(c)=1/4 A 0 0 B 1 1 C A 0 C 11 B 10 We can reduce ΔH by grouping m symbols together (alphabet size 2 m or 3 m ): Generally, the maximal depth (R = max r i ) grows proportionally to m here, i ΔH drops approximately proportionally to 1/m, ΔH 1/R (= 1/lg (L)) for ANS: ΔH 4 R (= 1/L 2 for L = 2 R is the number of states) 4

5 Fast Huffman decoding step for maximal depth R: let X contain last R bits (requires decodingtable of size proportional to L = 2 R ): t = decodingtable[x]; // X {0,, 2 R 1} is current state usesymbol(t. symbol); // use or store decoded symbol X = t. newx + readbits(t. nbbits); // state transition where t. newx for Huffman are unused bits of the state, shifted to oldest position: t. newx = (X nbbits) & (2 R 1) (mod(a, 2 R ) = a & (2 R 1)) tans: the same decoder, different t. newx: not only shifts the unused bits, but also modifies them accordingly to the remaining fractional bits (lg(1/p)): x = X + L {L,, 2L 1} buffer containing lg(x) [R, R + 1) bits: 5

$Operating on fractional number of bits restricting range L to length l subrange number x contains lg(l/l) bits contains lg(x) bits$

6 Operating on fractional number of bits restricting range L to length l subrange number x contains lg(l/l) bits contains lg(x) bits adding symbol of probability p - containing lg(1/p) bits lg(l/l) + lg(1/p) lg(l/l ) bits for l p l lg(x) + lg(1/p) = lg(x ) for x x/p 6

7 Asymmetric numeral systems redefine even/odd numbers: symbol distribution: s: N A (A alphabet, e.g. {0,1}) ( s(x) = mod(x, b) for base b numeral system: C(s, x) = bx + s) Should be still uniformly distributed but with density p s : # { 0 y < x: s(y) = s} xp s then x becomes x-th appearance of given symbol: D(x ) = (s(x ), # {0 y < x : s(y) = s(x )}) C(D(x )) = x D(C(s, x)) = (s, x) x x/p s example: range asymmetric binary systems (rabs) Pr(0) = 1/4 Pr(1) = 3/4 - take base 4 system and merge 3 digits, cyclic (0123) symbol distribution s becomes cyclic (0111): s(x) = 0 if mod(x, 4) = 0, else 1 to decode or encode 1, localize quadruple ( x/4 or x/3 ) if s(x) = 0, D(x) = (0, x/4 ) else D(x) = (1,3 x/4 + mod(x, 4) 1) C(0, x) = 4x C(1, x) = 4 x/3 + mod(x, 3) + 1 7

8 rans - range variant for large alphabet A = {0,.., m 1} assume Pr(s) = f s /2 n c s : = f 0 + f f s 1 start with base 2 n numeral system and merge length f s ranges for x {0,1,, 2 n 1}, s(x) = max{s: c s x} encoding: C(s, x) = x/f s n + mod(x, f s ) + c s decoding: s = s(x & (2 n 1)) (e.g. tabled, alias method) D(x) = (s, f s (x n) + ( x & (2 n 1)) c s ) Similar to Range Coding, but decoding has 1 multiplication (instead of 2), and state is 1 number (instead of 2), making it convenient for SIMD vectorization. ( ). Additionally, we can use alias method to store s for very precise probabilities. It is also convenient for dynamical updating. Alias method: rearrange probability distribution into m buckets: containing the primary symbol and eventually a single alias symbol 8

uabs - uniform binary variant (A = {0,1}) - extremely accurate Assume binary alphabet, p Pr (1), denote x s = {y < x: s(y) = s} xp s For uniform symbol distribution we can choose: x 1 = xp x 0 = x x

9 uabs - uniform binary variant (A = {0,1}) - extremely accurate Assume binary alphabet, p Pr (1), denote x s = {y < x: s(y) = s} xp s For uniform symbol distribution we can choose: x 1 = xp x 0 = x x 1 = x xp s(x) = 1 if there is jump on next position: s = s(x) = (x + 1)p xp decoding function: D(x) = (s, x s ) its inverse coding function: C(0, x) = x p C(1, x) = x p For p = Pr(1) = 1 Pr(0) = 0.3: 9

10 Stream version renormalization Currently: we encode using succeeding C functions into huge number x, then decode (in reverse direction!) using succeeding D. Like in arithmetic coding, we need renormalization to limit working precision - enforce x I = {L,, bl 1} by transferring base b youngest digits: ANS decoding step from state x (s, x) = D(x); usesymbol(s); while x < L, x = bx + readdigit(); encoding step for symbol s from state x while x > maxx[s] // = bl s 1 {writedigit(mod(x, b)); x = x/b }; x = C(s, x); For unique decoding, we need to ensure that there is a single way to perform above loops: I = {L,, bl 1}, I s = {L s,, bl s 1} where I s = {x: C(s, x) I} Fulfilled e.g. for - rabs/rans when p s = f s /2 n has 1/L accuracy: 2 n divides L, - uabs when p has 1/L accuracy: b Lp = blp, - in tabled variants (tabs/tans) it will be fulfilled by construction. 10

11 Single step of stream version: to get x I to I s = {L s,, bl s 1}, we need to transfer k digits: x s (C(s, x/b k ), mod(x, b k )) where k = log b (x/l s ) for given s, observe that k can obtain one of two values: k = k s or k = k s 1 for k s = log b (p s ) = log b (L s /L) e. g. : b = 2, k s = 3, L s = 13, L = 66, b k sx + 3 = 115, x = 14, p s = 13/66: 11

General picture: encoder prepares before consuming succeeding symbol decoder produces symbol, then consumes succeeding digits Decoding is in reversed direction: we have stack of symbols (LIFO)

12 General picture: encoder prepares before consuming succeeding symbol decoder produces symbol, then consumes succeeding digits Decoding is in reversed direction: we have stack of symbols (LIFO) - the final state has to be stored, but we can write information in initial state - encoding should be made in backward direction, but use context from perspective of future forward decoding. 12

13 In single step (I = {L,, bl 1}): lg(x) lg(x) + lg(1/p) Three sources of unpredictability/chaosity: 1) Asymmetry: behavior strongly dependent on chosen symbol small difference changes decoded symbol and so the entire behavior. 2) Ergodicity: usually log b (1/p) is irrational succeeding iterations cover entire range. 3) Diffusivity: C(s, x) is close but not exactly x/p s there is additional diffusion around expected value modulo lg(b) So lg(x) [lg(l), lg(l) + lg(b)) has nearly uniform distribution x has approximately: Pr(x) 1/x probability distribution contains lg (1/Pr (x)) lg (x) + const bits of information. 13

14 ΔH Redundancy: instead of p s we use q s = x/c(s, x) probability: ΔH 1 ln(4) (p s q s ) 2 ΔH 1 p s s s,x where inaccuracy: ε s (x) = p s x/c(s, x) drops like 1/L: 1 ln(4) L 2 (1 + 1 ) for uabs, ΔH m p 1 p ln(4) Pr(x) (ε s(x)) ln(4) L 2 1 p s p s 2 s for rans Denoting L = 2 R, ΔH 4 R and grows proportionally to (alphabet size) 2. for uabs: 14

15 Tabled variants: tans, tabs (choose L = 2 R ) I = {L,,2L 1}, I s = {L s,,2l s 1} where I s = {x: C(s, x) I} we will choose s symbol distribution (symbol spread): #{x I: s(x) = s} = L s approximating probabilities: p s L s /L Fast pseudorandom spread ( ): step = 5/8 L + 3; // step chosen to cover the whole range X = 0; // X = x L {0,, L 1} for table handling for s = 0 to m 1 do {symbol[x] = s; X = mod(x + step, L); } // symbol[x] = s(x + L) Then finding table for {t = decodingtable(x); usesymbol(t. symbol); X = t. newx + readbits(t. nbbits)} decoding: for s = 0 to m 1 do next[s] = L s ; // symbol appearance number for X = 0 to L 1 do // fill all positions {t. symbol = symbol[x]; // from symbol spread x = next[t. symbol] + +; // use symbol and shift appearance t. nbbits = R lg(x) ; // L = 2 R t. newx = (x t. nbbits) L; decodingtable[x] = t; } 15

16 Precise initialization (heuresis) N s = { 0.5+i p s : i = 0,, L s 1} are uniform we need to shift them to natural numbers. (priority queue with put, getminv) for s = 0 to n 1 do put((0.5/p s, s)); for X = 0 to L 1 do {(v, s) = getminv; put((v + 1/p s, s)); symbol[x] = s; } ΔH drops like 1/L 2 L alphabet size 16

tabs and tuning test all possible symbol distributions for binary alphabet store tables for quantized probabilities (p) e.g. 16 16 = 256 bytes for 16 state, ΔH 0.001 H.

17 tabs and tuning test all possible symbol distributions for binary alphabet store tables for quantized probabilities (p) e.g = 256 bytes for 16 state, ΔH H.264 M decoder (arith. coding) tabs t = decodingtable[p][x]; X = t. newx + readbits(t. nbbits); usesymbol(t. symbol); no branches, no bit-by-bit renormalization state is single number (e.g. SIMD) 17

Additional tans advantage simultaneous encryption we can use huge freedom while initialization: choosing symbol distribution slightly disturb s(x) using PRNG initialized with cryptographic key

18 Additional tans advantage simultaneous encryption we can use huge freedom while initialization: choosing symbol distribution slightly disturb s(x) using PRNG initialized with cryptographic key ADVANTAGES comparing to standard (symmetric) cryptography: standard, e.g. DES, AES ANS based cryptography (initialized) based on XOR, permutations highly nonlinear operations bit blocks fixed length pseudorandomly varying lengths brute force or QC attacks just start decoding to test cryptokey perform initialization first for new cryptokey, fixed to need e.g. 0.1s speed online calculation most calculations while initialization entropy operates on bits operates on any input distribution 18

19 Summary we can construct accurate and extremely fast entropy coders: - accurate (and faster) replacement for Huffman coding, - many times faster replacement for Range coding, - faster decoding in adaptive binary case (but more complex encoding), - can simultaneously encrypt the message. Perfect for: DCT/wavelet coefficients, lossless image compression Perfect with: Lempel-Ziv, Burrows-Wheeler Transform... and many others... Further research: - finding symbol distribution s(x): with minimal ΔH (and quickly), - tune accordingly to p s L s /L approximation (e.g. very low probable symbols should have single appearance at the end of table), - optimal compression of used probability distribution to reduce headers, - maybe finding other low state entropy coding family, like forward decoded, - applying cryptographic capabilities (also without entropy coding), -...? 19

Data Compression Techniques

Data Compression Techniques Part 1: Entropy Coding Lecture 4: Asymmetric Numeral Systems Juha Kärkkäinen 08.11.2017 1 / 19 Asymmetric Numeral Systems Asymmetric numeral systems (ANS) is a recent entropy