Download presentation
Presentation is loading. Please wait.
1
Uncertain Compression
& Graph Coloring Madhu Sudan Harvard Based on joint works with: (1) Adam Kalai (MSR), Sanjeev Khanna (U.Penn), Brendan Juba (WUStL) (2) Elad Haramaty (Harvard) (3) Badih Ghazi (MIT), Elad Haramaty (Harvard), Pritish Kamath (MIT) August 4, 2017 IITB: Uncertain Compression & Coloring
2
Classical Compression
The Shannon setting Alice gets πβ [π] chosen from distribution π Sends compression π¦ = πΈ π π β 0,1 β to Bob. Bob computes π = π· π π¦ (Alice and Bob both know π). Require π = π (whp). Goal: min πΈ π , π· π πΌxp πβΌπ | πΈ π π | Further (technical) requirement: βPrefix-freeβ βπβ πβ², πΈ π (π) not prefix of πΈ π π β² Ensures we can encode sequence of messages August 4, 2017 IITB: Uncertain Compression & Coloring
3
IITB: Uncertain Compression & Coloring
Shannon+Huffman [Kraft]: Prefix-free compression: Encoding message π with length β π possible iff π 2 β β π β€1 [Shannon]: Assign length β log 2 π(π) with message π Expected length πΌ π β log π(π) β€π» π +1 Motivates π» π β πΌ π β log π π [Huffman]: Constructive, explicit, optimal August 4, 2017 IITB: Uncertain Compression & Coloring
4
IITB: Uncertain Compression & Coloring
Fundamental problem: gzip, Lempel-Ziv β¦ Leads to entropy: Fundamental measure. Fundamental role in βlearningβ Learning β‘ Compression A goal in language too! Language evolves Words introduced, others fade. Why? Intuitive explanation: distributions on messages evolve! August 4, 2017 IITB: Uncertain Compression & Coloring
5
Compression as proxy for language
Understanding languages: Complex, not all forces well understood. Others hard to analyze. Compression: Clean mathematical problem. Faces similar issues as language. (Sometimes) easier to analyze. But to model issues associated with natural language, need to incorporate uncertainty! People donβt have same priors on messages. Need to estimate/bound each others priors Can compression work with uncertain priors? August 4, 2017 IITB: Uncertain Compression & Coloring
6
IITB: Uncertain Compression & Coloring
Outline Part 1: Motivation Part 2: Formalism Part 3: Randomized Solution Part 4: Issues with Randomized Solution Part 5: Deterministic Issues. August 4, 2017 IITB: Uncertain Compression & Coloring
7
Uncertain Compression
Design encoding/decoding schemes (πΈ/π·) so that Sender has distribution π on [π] Receiver has distribution π on [π] Sender gets πβ[π] Sends πΈ(π,π) to receiver. Receiver receives π¦ = πΈ(π,π) Decodes to π =π·(π,π¦) Want: π= π (provided π,π close), While minimizing πΌ xp πβΌπ |πΈ(π,π)| August 4, 2017 IITB: Uncertain Compression & Coloring
8
Proximity of Distributions
Many alternatives. Our goal: Find anything non-trivial that allows compression. Eventual choice: Ξ π,π = max πβ[π] max log π π π π , log π π π π βSymmetrized worst-case KL divergenceβ KL Divergence: π· π,π =πΌ xp πβΌπ log π π π(π) (So trivially: π· π,π β€Ξ(π,π) ) Question: Can message be compressed to within π π» π ,Ξ ? Or is it π(π» π ,Ξ,π)? August 4, 2017 IITB: Uncertain Compression & Coloring
9
Solution 1: Assuming Randomness
Assume sender+receiver share πβΌ Unif( 0,1 π‘ ) In particular π independent of π,π,π Compression scheme: Let π=( π 1 ,β¦, π π ) with π π β 0,1 π‘ π Sender sends prefix π§ of π π ; long enough so that βπβ² s.t. π§ is a prefix of π π β² : π π β² < π π 4 Ξ πΌ xp π π π ββ log π π +2Ξ β Thm: Expected compression length β€π» π +2Ξ Deterministic? August 4, 2017 IITB: Uncertain Compression & Coloring
10
Combinatorial Reinterpretation
Can fix π(π) (adds β log π π to compression). Define: π΄ 0 = π , π΄ 1 = π β² |π π β² β₯ 4 βΞ β
π(π) , π΄ π = π β² |π π β² β₯ 4 βπβ
Ξ β
π(π) Similarly π΅ 1 = π β² | π π β² β₯ 2 βΞ β
π(π) , π΅ π = π β² | π π β² β₯ 2 β 2πβ1 β
Ξ β
π(π) Nesting: π΄ 0 β π΅ 1 β π΄ 1 β π΅ 2 β π΄ 2 β― Sizes: π΄ π β€πΎβ
πΆ 2π , π΅ π β€πΎβ
πΆ 2πβ1 for πΎ= 1 π π ; πΆ= 2 Ξ Question: Given πΎ,πΆ can π be distinguished from π β² β π΅ 1 with π πΎ,πΆ 1 bits? August 4, 2017 IITB: Uncertain Compression & Coloring
11
Compression β‘ Coloring
Weak Uncertainty graph π π,πΎ,πΆ Vertices = π΄ 0 , π΄ 1 ,β¦, π΄ β : Nested, π΄ 0 =1, π΄ π β€πΎβ
πΆ 2π , π΄ β =[π] Edges: π΄ 0 , π΄ 1 ,β¦, π΄ β β π΄ 0 β² , π΄ 1 β² ,β¦, π΄ β β² β² iff π΄ 0 β π΄ 0 β² π΄ π β π΄ π+1 β² ; π΄ π β² β π΄ π+1 Claim: Compression length =π π» π ,Ξ iff βπΎ,πΆ, π π π π,πΎ,πΆ = π πΎ,πΆ (1) π π π,πΎ,πΆ = open! Did we reduce to a harder problem? ο August 4, 2017 IITB: Uncertain Compression & Coloring
12
Bounding chromatic number
Upper bounds: Easy! Just give the coloring! β¦ not always. E.g., βShift Graphβ π π,π Vertices: π π β Sequences of π distinct elements of [π] Edges π 1 , π 2 ,β¦, π π β( π 2 , π 3 ,β¦, π π , π π+1 ) Thm: [Cole-Vishkin, Linial] If πβ₯ log β π , then π π π,π =3. More generally π π π,π =max 3, log π+Ξ 1 π Lower bounds β¦ much more challenging! August 4, 2017 IITB: Uncertain Compression & Coloring
13
IITB: Uncertain Compression & Coloring
Some Results [Haramaty+S.] π π π,πΎ,πΆ β€ exp πΎ. πΆ β β
log β π ββ [Golowich] π π π,πΎ,πΆ β€ exp exp πΎβ
πΆ β β
log 2β π ββ [Trivial] π π π,πΎ,πΆ β₯πΎβ
πΆ 2 For further understanding define π π,πΎ,πΆ β Like π π,πΎ,πΆ but without size restriction on π΄ β . (so wlog π΄ β =[π]) Upper bounds hold even for π π,πΎ,πΆ β [Haramaty+S]: π π π,πΎ,πΆ β = log Ξ© β π Need slow growth for long to get π-independence August 4, 2017 IITB: Uncertain Compression & Coloring
14
Upper Bounds β 1 [Cole-Vishkin/Linial]
Coloring Shift graph: Given coloring π: π π,πβ1 β 0,1 π , construct coloring π β² : π π,π β π Γ{0,1} as follows: To color π 1 ,β¦, π π : Let (π 1 ,β¦, π π )=π( π 1 ,β¦, π πβ1 ) And π 1 ,β¦, π π =π π 2 ,β¦, π π Let π be least index s.t. π π β π π (exists!) π π 1 ,β¦, π π = π, π π Valid? π π 2 , β¦, π π+1 = π, π π or π β² ,π₯ for π β² β π ! August 4, 2017 IITB: Uncertain Compression & Coloring
15
Generalizing: Homorphisms
πΊ homomorphic to π» (πΊβπ») if β π:π πΊ βπ π» s.t. π’ β πΊ π£βπ π’ β π» π π£ Homorphisms? πΊ is π-colorable β πΊβ πΎ π πΊβπ» and π»βπΏβπΊβπΏ Homomorphisms and Shift/Uncertainty graphs. π π,π β π π,πβ1 β π π,πβ2 ββ― π π,πΎ,πΆ = π π,πΎ,πΆ π β π π,πΎ,πΆ πβ1 ββ―β π π,πΎ,πΆ β ββ― Suffices to upper bound π π π,πΎ,πΆ β August 4, 2017 IITB: Uncertain Compression & Coloring
16
Degree of Homomorphisms
Say π:πΊβπ» πΊ π π π’ β π π£ π£ β πΊ π’ | π π βmax π’ { π π π’ } π» Lemma [HS]: π πΊ β€π( π π 2 log π π» ) Lemma [Golowich]: π πΊ β€π( exp π π log log π π» ) For π: π π,πΎ,πΆ β β π π,πΎ,πΆ ββ1 π π = exp πΎβ
πΆ β August 4, 2017 IITB: Uncertain Compression & Coloring
17
Proof: (of π πΊ β€π( π π 2 log π π» ) )
Denote π π» =π; π π =π Let π=π πβ
log π ; π‘=2π Claim: β β 1 ,β¦, β π , β π : π β π‘ s.t. βπβπβ π , π β€π, βπ π .π‘. β π π β β π π Proof: Pick β π βs at random Claim: βπ‘β
π coloring of πΊ Proof: Given π:π»β[π], let π β² :πΊβ π Γ[π‘] be: Let π π’ = π π π£ ) π£ β πΊ π’ ; π=π π π’ Let π be s.t. β π π β β π π π β² π’ = π, β π π August 4, 2017 IITB: Uncertain Compression & Coloring
18
Lower bounds: π π π,πΎ,πΆ β β₯ log (2β) π
Claim: π π,2β subgraph of π π,πΎ,πΆ β Proof: by inspection β¦ Linialβs Proof: π π π,β β₯ log π π π,ββ1 βπ π π,ββ1 β€ 2 π π π,β (lower bounds by upper bounds!) Given π: π π,β β[π], let π β² : π π,ββ1 β 2 [π] be: π β² π 1 ,β¦, π ββ1 = π π 1 ,β¦ π β | π β Claim: π β² π 1 ,β¦, π ββ1 β π β² π 2 ,β¦, π β Proof: π π 1 ,β¦, π β β π β² π 1 ,β¦, π β β1 But π π 1 ,β¦, π β β π π 2 ,β¦, π β+1 βπ π 1 ,β¦, π β β π β² π 2 ,β¦, π β August 4, 2017 IITB: Uncertain Compression & Coloring
19
IITB: Uncertain Compression & Coloring
Conclusion Compression (Uncertain) β‘ Graph Coloring Unfortunately β latter is hard! (not only to solve optimally, but also to understand analytically) Intriguing Special Case: π π β : π΄ π β€3π (linear, not exponential, growth) Is π π π β =π 1 ? Fundamental underlying question: Is entropy the correct measure of natural compressibility August 4, 2017 IITB: Uncertain Compression & Coloring
20
IITB: Uncertain Compression & Coloring
Thank You August 4, 2017 IITB: Uncertain Compression & Coloring
Similar presentations
© 2024 SlidePlayer.com. Inc.
All rights reserved.