greedy merge degrading has optimal power law
play

Greedy-Merge Degrading has Optimal Power-Law Assaf Kartowsky and Ido - PowerPoint PPT Presentation

Greedy-Merge Degrading has Optimal Power-Law Assaf Kartowsky and Ido Tal Technion, Israel June 28, 2017 Assaf Kartowsky and Ido Tal Greedy-Merge Degrading has Optimal Power-Law June 28, 2017 1 / 17 Introduction Motivation Motivation X Y


  1. Greedy-Merge Degrading has Optimal Power-Law Assaf Kartowsky and Ido Tal Technion, Israel June 28, 2017 Assaf Kartowsky and Ido Tal Greedy-Merge Degrading has Optimal Power-Law June 28, 2017 1 / 17

  2. Introduction Motivation Motivation X Y W W : X → Y , P X , |Y| is very large Common problem in – Digital receiver design – Polar code construction ⇒ Quantize Y to L letters Assaf Kartowsky and Ido Tal Greedy-Merge Degrading has Optimal Power-Law June 28, 2017 2 / 17

  3. Introduction Motivation Motivation X Y Z W Φ Q Q : X → Z , |Z| = L ∆ I � I ( X ; Y ) − I ( X ; Z ) ≥ 0 Question Given |X| , what is ∆ I ∗ � min Q ∆ I = O (?) in terms of L ? Assaf Kartowsky and Ido Tal Greedy-Merge Degrading has Optimal Power-Law June 28, 2017 3 / 17

  4. Introduction Previous Results Previous Results Binary input, |X| = 2 O ( L − 1 . 5 log L ) Pedarsani et al. 2011 Finite |X| (constant) O ( L − 1 / ( |X|− 1) ) Gulcu, Ye, and Barg 2016 Ω( L − 2 / ( |X|− 1) ) Tal 2015 Related work Kurkoski and Yagi 2014 Nazer, Ordentlich, and Polyanskiy 2017 Assaf Kartowsky and Ido Tal Greedy-Merge Degrading has Optimal Power-Law June 28, 2017 4 / 17

  5. Main Result Main Result Theorem � |Y| > 2 |X| ∆ I ∗ = O ( L − 2 |X|− 1 ) = ⇒ L ≥ 2 |X| In particular, 2   |X|− 1 π |X| ( |X| − 1) 2 |X| ∆ I ∗ ≤ 2 · L − |X|− 1   � � �� � 2 1 + |X|− 1 1 Γ 2( |X|− 1) − 1 2 1 + 2 This bound is: Attained by “greedy-merge” algorithm Tight in power-law sense Assaf Kartowsky and Ido Tal Greedy-Merge Degrading has Optimal Power-Law June 28, 2017 5 / 17

  6. Proof Outline Main Ideas Proof - Main Ideas Greedy-merge algorithm Simple upper bounds on ∆ I “Sphere-packing” Assaf Kartowsky and Ido Tal Greedy-Merge Degrading has Optimal Power-Law June 28, 2017 6 / 17

  7. Proof Outline Notation Notation Channel, input and output probabilities: W ( y | x ) � P ( Y = y | X = x ) π x � P ( X = x ) W ( x | y ) � P ( X = x | Y = y ) π y � P ( Y = y ) Mutual information: � � I ( W , P X ) � I ( X ; Y ) = η ( π x ) − π y η ( W ( x | y )) x ∈X x ∈X , y ∈Y � − p log p p > 0 η ( p ) � 0 p = 0 Loss in mutual information: ∆ I = I ( W , P X ) − I ( Q , P X ) Assaf Kartowsky and Ido Tal Greedy-Merge Degrading has Optimal Power-Law June 28, 2017 7 / 17

  8. Proof Outline Greedy-Merge Algorithm Merging a Pair of Letters For y a , y b ∈ Y define: y a α x � W ( x | y a ) α � ( α x ) x ∈X π a � π y a y ab β x � W ( x | y b ) β � ( β x ) x ∈X π b � π y b y b Merging y a , y b to y ab : W ( x | y ab ) = π a α x + π b β x π y ab = π a + π b π a + π b Loss by a single merger: � π a α x + π b β x � ∆ I x � ( π a + π b ) η − π a η ( α x ) − π b η ( β x ) π a + π b � ∆ I = ∆ I x x ∈X Assaf Kartowsky and Ido Tal Greedy-Merge Degrading has Optimal Power-Law June 28, 2017 8 / 17

  9. Proof Outline Greedy-Merge Algorithm Greedy-Merge Algorithm Algorithm: – Merge y a , y b that minimize ∆ I – Repeat |Y| − L times If min ∆ I = O ( |Y| − |X| +1 |X|− 1 ) ⇒ proof is finished New Goal Prove existence of y a , y b ∈ Y s.t. ∆ I = O ( |Y| − |X| +1 |X|− 1 ) Assaf Kartowsky and Ido Tal Greedy-Merge Degrading has Optimal Power-Law June 28, 2017 9 / 17

  10. Proof Outline Simple upper bounds on ∆ I Simple upper bounds on ∆ I ∆ I is complicated Upper bound ∆ I : ∆ I x ≤ ( π a + π b ) ( α x − β x ) 2 ∆ I x ≤ ( π a + π b ) | α x − β x | min( α x ,β x ) � � | α x − β x | , ( α x − β x ) 2 ⇒ ∆ I ≤ ( π a + π b ) |X| · max x ∈X min min( α x , β x ) � �� � � d ( α , β ) Limit search to: � � |Y small | ≥ |Y| 2 Y small � y ∈ Y : π y ≤ |Y| 2 y a , y b ∈Y small ∆ I ≤ 4 |X| ⇒ min |Y| · y a , y b ∈Y small d ( α , β ) min Assaf Kartowsky and Ido Tal Greedy-Merge Degrading has Optimal Power-Law June 28, 2017 10 / 17

  11. Proof Outline Sphere-Packing Sphere-Packing Essentials y 4 y 6 y 2 1 2 r y 7 y 3 y 5 y 1 A metric d : M × M → R + 0 r / 2-radius volumed spheres � α , r � � ζ ∈ M : d ( α , ζ ) ≤ r � � B 2 2 Find r = r critical > 0 s.t.: � � α , r �� � � � B Vol = Vol whole space 2 α ∈ S ⇒ d ( α , β ) ≤ r for some α , β ∈ M Assaf Kartowsky and Ido Tal Greedy-Merge Degrading has Optimal Power-Law June 28, 2017 11 / 17

  12. Proof Outline Sphere-Packing Sphere-Packing Reasoning � � � � α , r β , r � = ∅ ⇒ d ( α , ζ ) , d ( β , ζ ) ≤ r B ∩ B 2 2 2 Triangle inequality: ⇒ d ( α , β ) ≤ d ( α , ζ ) + d ( ζ , β ) ≤ r B ( α i , r 2 ) d ≤ r α 3 α 2 α 5 α 1 α 4 Assaf Kartowsky and Ido Tal Greedy-Merge Degrading has Optimal Power-Law June 28, 2017 12 / 17

  13. Proof Outline Sphere-Packing Sphere-Packing Reasoning d is a semimetric Find Q ′ ( · , r ) s.t.: Q ′ ( α , r ) ∩ Q ′ ( β , r ) � = ∅ ⇒ d ( α , β ) ≤ r Q ′ ( α i , r ) r ≤ α 3 d α 2 α 5 α 1 α 4 Assaf Kartowsky and Ido Tal Greedy-Merge Degrading has Optimal Power-Law June 28, 2017 13 / 17

  14. Proof Outline Sphere-Packing Towards a “Sphere” Our d ( · , · ) is a semimetric B ( α , r ) is a box s 1 s 0 + s 1 = 1 B ( α , r ) ω ( α 1 , r ) α 1 ω ( α 1 , r ) s 0 α 0 ω ( α 0 , r ) ω ( α 0 , r ) �� � r 2 ω ( α x , r ) � max ( √ α x r , r ) 4 + α x r − r ω ( α x , r ) � max 2 , r � x ∈X α x = 1 ⇒ dimension reduction is preferable Assaf Kartowsky and Ido Tal Greedy-Merge Degrading has Optimal Power-Law June 28, 2017 14 / 17

  15. Proof Outline Sphere-Packing Towards a “Sphere” s 0 + s 1 = 1 s 1 B ( α , r ) ω ( α 1 , r ) Q ′ ( α , r ) B K ( α , r ) α 1 C ( α , r ) ω ( α 1 , r ) s 0 ω ( α 0 , r ) α 0 ω ( α 0 , r ) � ζ ∈ R |X| : � � B K ( α , r ) � B ( α , r ) ∩ x ∈X ζ x = 1 - complicated C ( α , r ) ⊆ B K ( α , r ) - a box in R |X|− 1 : C ( α , r ) ∩ C ( β , r ) � = ∅ d ( α , β ) ≤ r � Q ′ ( α , r ) ⊆ R |X|− 1 - suitable Assaf Kartowsky and Ido Tal Greedy-Merge Degrading has Optimal Power-Law June 28, 2017 15 / 17

  16. Proof Outline Sphere-Packing Weighted “Sphere”-Packing Variable volume “spheres” ϕ 1 2 √ s 1 Q ′ ( α , r ) s 1 ∼√ α 1 r α 1 |X| − 1 dimensional density: 1 � ϕ ( ζ ′ ) � 2 √ ζ x x ∈X ′ Volume → Weight Assaf Kartowsky and Ido Tal Greedy-Merge Degrading has Optimal Power-Law June 28, 2017 16 / 17

  17. Proof Outline Sphere-Packing Conclusion and Further Results Conclusion 2 ∆ I ∗ = O ( L − |X|− 1 ) Tight in power-law Attained by “greedy-merge” algorithm Further results (full paper) For the upgrading setting: 2 ∆ I ∗ = Ω( L − |X|− 1 ), same sequence of channels 2 ∆ I ∗ = O ( L − |X|− 1 ) for |X| = 2 Optimal upgrading algorithm for |X| = 2 Assaf Kartowsky and Ido Tal Greedy-Merge Degrading has Optimal Power-Law June 28, 2017 17 / 17

Download Presentation
Download Policy: The content available on the website is offered to you 'AS IS' for your personal information and use only. It cannot be commercialized, licensed, or distributed on other websites without prior consent from the author. To download a presentation, simply click this link. If you encounter any difficulties during the download process, it's possible that the publisher has removed the file from their server.

Recommend


More recommend