Presentation is loading. Please wait.

Presentation is loading. Please wait.

Uncertain Compression

Similar presentations


Presentation on theme: "Uncertain Compression"β€” Presentation transcript:

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 π‘šβˆˆ [𝑁] chosen from distribution 𝑃 Sends compression 𝑦 = 𝐸 𝑃 π‘š ∈ 0,1 βˆ— to Bob. Bob computes π‘š = 𝐷 𝑃 𝑦 (Alice and Bob both know 𝑃). Require π‘š = π‘š (whp). Goal: min 𝐸 𝑃 , 𝐷 𝑃 𝔼xp π‘šβˆΌπ‘ƒ | 𝐸 𝑃 π‘š | Further (technical) requirement: β€œPrefix-free” βˆ€π‘šβ‰ π‘šβ€², 𝐸 𝑃 (π‘š) not prefix of 𝐸 𝑃 π‘š β€² 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 𝑖 with length β„“ 𝑖 possible iff 𝑖 2 βˆ’ β„“ 𝑖 ≀1 [Shannon]: Assign length βˆ’ log 2 𝑃(𝑖) with message 𝑖 Expected length 𝔼 𝑖 βˆ’ log 𝑃(𝑖) ≀𝐻 𝑃 +1 Motivates 𝐻 𝑃 ≝ 𝔼 𝑖 βˆ’ log 𝑃 𝑖 [Huffman]: Constructive, explicit, optimal August 4, 2017 IITB: Uncertain Compression & Coloring

4 IITB: Uncertain Compression & Coloring
Fundamental problem: gzip, Lempel-Ziv … Leads to entropy: Fundamental measure. Fundamental role in β€œlearning” Learning ≑ 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 don’t 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 (𝐸/𝐷) so that Sender has distribution 𝑃 on [𝑁] Receiver has distribution 𝑄 on [𝑁] Sender gets π‘šβˆˆ[𝑁] Sends 𝐸(𝑃,π‘š) to receiver. Receiver receives 𝑦 = 𝐸(𝑃,π‘š) Decodes to π‘š =𝐷(𝑄,𝑦) Want: π‘š= π‘š (provided 𝑃,𝑄 close), While minimizing 𝔼 xp π‘šβˆΌπ‘ƒ |𝐸(𝑃,π‘š)| 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” 𝑃,𝑄 = max π‘šβˆˆ[𝑁] max log 𝑃 π‘š 𝑄 π‘š , log 𝑄 π‘š 𝑃 π‘š β€œSymmetrized worst-case KL divergence” KL Divergence: 𝐷 𝑃,𝑄 =𝔼 xp π‘šβˆΌπ‘ƒ log 𝑃 π‘š 𝑄(π‘š) (So trivially: 𝐷 𝑃,𝑄 ≀Δ(𝑃,𝑄) ) Question: Can message be compressed to within 𝑓 𝐻 𝑃 ,Ξ” ? Or is it 𝑓(𝐻 𝑃 ,Ξ”,𝑁)? August 4, 2017 IITB: Uncertain Compression & Coloring

9 Solution 1: Assuming Randomness
Assume sender+receiver share π‘ŸβˆΌ Unif( 0,1 𝑑 ) In particular π‘Ÿ independent of 𝑃,𝑄,π‘š Compression scheme: Let π‘Ÿ=( π‘Ÿ 1 ,…, π‘Ÿ 𝑁 ) with π‘Ÿ 𝑖 ∈ 0,1 𝑑 𝑁 Sender sends prefix 𝑧 of π‘Ÿ π‘š ; long enough so that βˆ€π‘šβ€² s.t. 𝑧 is a prefix of π‘Ÿ π‘š β€² : 𝑃 π‘š β€² < 𝑃 π‘š 4 Ξ” 𝔼 xp π‘Ÿ π‘Ÿ π‘š β‰ˆβˆ’ log 𝑃 π‘š +2Ξ” β‡’ Thm: Expected compression length ≀𝐻 𝑃 +2Ξ” Deterministic? August 4, 2017 IITB: Uncertain Compression & Coloring

10 Combinatorial Reinterpretation
Can fix 𝑃(π‘š) (adds βˆ’ log 𝑃 π‘š to compression). Define: 𝐴 0 = π‘š , 𝐴 1 = π‘š β€² |𝑃 π‘š β€² β‰₯ 4 βˆ’Ξ” ⋅𝑃(π‘š) , 𝐴 𝑖 = π‘š β€² |𝑃 π‘š β€² β‰₯ 4 βˆ’π‘–β‹…Ξ” ⋅𝑃(π‘š) Similarly 𝐡 1 = π‘š β€² | 𝑄 π‘š β€² β‰₯ 2 βˆ’Ξ” ⋅𝑃(π‘š) , 𝐡 𝑖 = π‘š β€² | 𝑄 π‘š β€² β‰₯ 2 βˆ’ 2π‘–βˆ’1 β‹…Ξ” ⋅𝑃(π‘š) Nesting: 𝐴 0 βŠ† 𝐡 1 βŠ† 𝐴 1 βŠ† 𝐡 2 βŠ† 𝐴 2 β‹― Sizes: 𝐴 𝑖 ≀𝐾⋅ 𝐢 2𝑖 , 𝐡 𝑖 ≀𝐾⋅ 𝐢 2π‘–βˆ’1 for 𝐾= 1 𝑃 π‘š ; 𝐢= 2 Ξ” Question: Given 𝐾,𝐢 can π‘š be distinguished from π‘š β€² ∈ 𝐡 1 with 𝑂 𝐾,𝐢 1 bits? August 4, 2017 IITB: Uncertain Compression & Coloring

11 Compression ≑ Coloring
Weak Uncertainty graph π‘Š 𝑁,𝐾,𝐢 Vertices = 𝐴 0 , 𝐴 1 ,…, 𝐴 β„“ : Nested, 𝐴 0 =1, 𝐴 𝑖 ≀𝐾⋅ 𝐢 2𝑖 , 𝐴 β„“ =[𝑁] Edges: 𝐴 0 , 𝐴 1 ,…, 𝐴 β„“ ↔ 𝐴 0 β€² , 𝐴 1 β€² ,…, 𝐴 β„“ β€² β€² iff 𝐴 0 β‰  𝐴 0 β€² 𝐴 𝑖 βŠ† 𝐴 𝑖+1 β€² ; 𝐴 𝑖 β€² βŠ† 𝐴 𝑖+1 Claim: Compression length =𝑓 𝐻 𝑃 ,Ξ” iff βˆ€πΎ,𝐢, 𝑁 πœ’ π‘Š 𝑁,𝐾,𝐢 = 𝑂 𝐾,𝐢 (1) πœ’ π‘Š 𝑁,𝐾,𝐢 = open! Did we reduce to a harder problem?  August 4, 2017 IITB: Uncertain Compression & Coloring

12 Bounding chromatic number
Upper bounds: Easy! Just give the coloring! … not always. E.g., β€œShift Graph” 𝑆 𝑛,π‘˜ Vertices: 𝑛 π‘˜ ≝ Sequences of π‘˜ distinct elements of [𝑛] Edges 𝑖 1 , 𝑖 2 ,…, 𝑖 π‘˜ ↔( 𝑖 2 , 𝑖 3 ,…, 𝑖 π‘˜ , 𝑖 π‘˜+1 ) Thm: [Cole-Vishkin, Linial] If π‘˜β‰₯ log βˆ— 𝑛 , then πœ’ 𝑆 𝑛,π‘˜ =3. More generally πœ’ 𝑆 𝑛,π‘˜ =max 3, log π‘˜+Θ 1 𝑛 Lower bounds … much more challenging! August 4, 2017 IITB: Uncertain Compression & Coloring

13 IITB: Uncertain Compression & Coloring
Some Results [Haramaty+S.] πœ’ π‘Š 𝑁,𝐾,𝐢 ≀ exp 𝐾. 𝐢 β„“ β‹… log β„“ 𝑁 βˆ€β„“ [Golowich] πœ’ π‘Š 𝑁,𝐾,𝐢 ≀ exp exp 𝐾⋅ 𝐢 β„“ β‹… log 2β„“ 𝑁 βˆ€β„“ [Trivial] πœ’ π‘Š 𝑁,𝐾,𝐢 β‰₯𝐾⋅ 𝐢 2 For further understanding define π‘Š 𝑁,𝐾,𝐢 β„“ Like π‘Š 𝑁,𝐾,𝐢 but without size restriction on 𝐴 β„“ . (so wlog 𝐴 β„“ =[𝑁]) Upper bounds hold even for π‘Š 𝑁,𝐾,𝐢 β„“ [Haramaty+S]: πœ’ π‘Š 𝑁,𝐾,𝐢 β„“ = log Ξ© β„“ 𝑁 Need slow growth for long to get 𝑁-independence August 4, 2017 IITB: Uncertain Compression & Coloring

14 Upper Bounds – 1 [Cole-Vishkin/Linial]
Coloring Shift graph: Given coloring πœ’: 𝑆 𝑛,π‘˜βˆ’1 β†’ 0,1 𝑐 , construct coloring πœ’ β€² : 𝑆 𝑛,π‘˜ β†’ 𝑐 Γ—{0,1} as follows: To color 𝑖 1 ,…, 𝑖 π‘˜ : Let (π‘Ž 1 ,…, π‘Ž 𝑐 )=πœ’( 𝑖 1 ,…, 𝑖 π‘˜βˆ’1 ) And 𝑏 1 ,…, 𝑏 𝑐 =πœ’ 𝑖 2 ,…, 𝑖 π‘˜ Let 𝑗 be least index s.t. π‘Ž 𝑗 β‰  𝑏 𝑗 (exists!) πœ’ 𝑖 1 ,…, 𝑖 π‘˜ = 𝑗, π‘Ž 𝑗 Valid? πœ’ 𝑖 2 , …, 𝑖 π‘˜+1 = 𝑗, 𝑏 𝑗 or 𝑗 β€² ,π‘₯ for 𝑗 β€² ≠𝑗 ! August 4, 2017 IITB: Uncertain Compression & Coloring

15 Generalizing: Homorphisms
𝐺 homomorphic to 𝐻 (𝐺→𝐻) if βˆƒ πœ™:𝑉 𝐺 →𝑉 𝐻 s.t. 𝑒 ↔ 𝐺 π‘£β‡’πœ™ 𝑒 ↔ 𝐻 πœ™ 𝑣 Homorphisms? 𝐺 is π‘˜-colorable ⇔ 𝐺→ 𝐾 π‘˜ 𝐺→𝐻 and 𝐻→𝐿⇒𝐺→𝐿 Homomorphisms and Shift/Uncertainty graphs. 𝑆 𝑛,π‘˜ β†’ 𝑆 𝑛,π‘˜βˆ’1 β†’ 𝑆 𝑛,π‘˜βˆ’2 β†’β‹― π‘Š 𝑁,𝐾,𝐢 = π‘Š 𝑁,𝐾,𝐢 𝑁 β†’ π‘Š 𝑁,𝐾,𝐢 π‘βˆ’1 β†’β‹―β†’ π‘Š 𝑁,𝐾,𝐢 β„“ β†’β‹― Suffices to upper bound πœ’ π‘Š 𝑁,𝐾,𝐢 β„“ August 4, 2017 IITB: Uncertain Compression & Coloring

16 Degree of Homomorphisms
Say πœ™:𝐺→𝐻 𝐺 𝑑 πœ™ 𝑒 ≝ πœ™ 𝑣 𝑣 ↔ 𝐺 𝑒 | 𝑑 πœ™ ≝max 𝑒 { 𝑑 πœ™ 𝑒 } 𝐻 Lemma [HS]: πœ’ 𝐺 ≀𝑂( 𝑑 πœ™ 2 log πœ’ 𝐻 ) Lemma [Golowich]: πœ’ 𝐺 ≀𝑂( exp 𝑑 πœ™ log log πœ’ 𝐻 ) For πœ™: π‘Š 𝑁,𝐾,𝐢 β„“ β†’ π‘Š 𝑁,𝐾,𝐢 β„“βˆ’1 𝑑 πœ™ = exp 𝐾⋅ 𝐢 β„“ August 4, 2017 IITB: Uncertain Compression & Coloring

17 Proof: (of πœ’ 𝐺 ≀𝑂( 𝑑 πœ™ 2 log πœ’ 𝐻 ) )
Denote πœ’ 𝐻 =𝑐; 𝑑 πœ™ =𝑑 Let 𝑀=𝑂 𝑑⋅ log 𝑐 ; 𝑑=2𝑑 Claim: βˆƒ β„Ž 1 ,…, β„Ž 𝑀 , β„Ž 𝑖 : 𝑐 β†’ 𝑑 s.t. βˆ€π‘–βˆ‰π‘†βŠ† 𝑐 , 𝑆 ≀𝑑, βˆƒπ‘— 𝑠.𝑑. β„Ž 𝑗 𝑖 βˆ‰ β„Ž 𝑗 𝑆 Proof: Pick β„Ž 𝑗 ’s at random Claim: βˆƒπ‘‘β‹…π‘€ coloring of 𝐺 Proof: Given πœ’:𝐻→[𝑐], let πœ’ β€² :𝐺→ 𝑀 Γ—[𝑑] be: Let 𝑆 𝑒 = πœ’ πœ™ 𝑣 ) 𝑣 ↔ 𝐺 𝑒 ; 𝑖=πœ’ πœ™ 𝑒 Let 𝑗 be s.t. β„Ž 𝑗 𝑖 βˆ‰ β„Ž 𝑗 𝑆 πœ’ β€² 𝑒 = 𝑗, β„Ž 𝑗 𝑖 August 4, 2017 IITB: Uncertain Compression & Coloring

18 Lower bounds: πœ’ π‘Š 𝑁,𝐾,𝐢 β„“ β‰₯ log (2β„“) 𝑁
Claim: 𝑆 𝑁,2β„“ subgraph of π‘Š 𝑁,𝐾,𝐢 β„“ Proof: by inspection … Linial’s Proof: πœ’ 𝑆 𝑁,β„“ β‰₯ log πœ’ 𝑆 𝑁,β„“βˆ’1 β‡πœ’ 𝑆 𝑁,β„“βˆ’1 ≀ 2 πœ’ 𝑆 𝑁,β„“ (lower bounds by upper bounds!) Given πœ’: 𝑆 𝑁,β„“ β†’[𝑐], let πœ’ β€² : 𝑆 𝑁,β„“βˆ’1 β†’ 2 [𝑐] be: πœ’ β€² 𝑖 1 ,…, 𝑖 β„“βˆ’1 = πœ’ 𝑖 1 ,… 𝑖 β„“ | 𝑖 β„“ Claim: πœ’ β€² 𝑖 1 ,…, 𝑖 β„“βˆ’1 β‰  πœ’ β€² 𝑖 2 ,…, 𝑖 β„“ Proof: πœ’ 𝑖 1 ,…, 𝑖 β„“ ∈ πœ’ β€² 𝑖 1 ,…, 𝑖 β„“ βˆ’1 But πœ’ 𝑖 1 ,…, 𝑖 β„“ β‰ πœ’ 𝑖 2 ,…, 𝑖 β„“+1 β‡’πœ’ 𝑖 1 ,…, 𝑖 β„“ βˆ‰ πœ’ β€² 𝑖 2 ,…, 𝑖 β„“ August 4, 2017 IITB: Uncertain Compression & Coloring

19 IITB: Uncertain Compression & Coloring
Conclusion Compression (Uncertain) ≑ Graph Coloring Unfortunately – latter is hard! (not only to solve optimally, but also to understand analytically) Intriguing Special Case: π‘Š 𝑁 βˆ— : 𝐴 𝑖 ≀3𝑖 (linear, not exponential, growth) Is πœ’ π‘Š 𝑁 βˆ— =𝑂 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


Download ppt "Uncertain Compression"

Similar presentations


Ads by Google