Download presentation
Presentation is loading. Please wait.
1
General Strong Polarization
Madhu Sudan Harvard University Joint work with Jaroslaw Blasiok (Harvard), Venkatesan Guruswami (CMU), Preetum Nakkiran (Harvard) and Atri Rudra (Buffalo) Oct. 8, 2018 Berkeley: General Strong Polarization
2
Part I: Motivation Achieving Shannon Capacity
Oct. 8, 2018 Berkeley: General Strong Polarization
3
Shannon and Channel Capacity
Shannon [β48]: For every (noisy) channel of communication, β capacity πΆ s.t. communication at rate π
<πΆ is possible and π
>πΆ is not. formalization β omitted. Example: Acts independently on bits Capacity =1 β π»(π) ; π»(π) = binary entropy! π» π =πβ
log 1 π + 1βπ β
log 1 1βπ Capacity? β πΈ: π½ 2 π π β π½ 2 π π·: π½ 2 π β π½ 2 π π w. Pr π· π΅ππΆ πΈ π β π =π(1) lim πββ π π π β₯π
(Rate) πβ π½ 2 π w.p. 1βπ 1βπ w.p. π π΅ππΆ(π) Oct. 8, 2018 Berkeley: General Strong Polarization
4
βAchievingβ Shannon Capacity
Commonly used phrase. Many mathematical interpretations. Some possibilities: How small can π be? Get π
>πΆβπ with polytime algorithms? [Luby et al.β95] running time poly π π ? (or π=poly 1 π ?) Open till 2008 Arikanβ08: Invented βPolar Codesβ β¦ Final resolution: Guruswami+Xiaβ13, Hassani+Alishahi+Urbankeβ13 β Strong analysis Shannon β48: π=π 1 π 2 ; πβπΆβπ
Forney β66 :time =ππππ¦ π, π 2 Oct. 8, 2018 Berkeley: General Strong Polarization
5
Polar Codes and Polarization
Arikan: Defined Polar Codes, one for every π‘ Associated Martingale π 0 ,β¦, π π‘ ,β¦ π π‘ β[0,1] π‘th code: Length π= 2 π‘ If Pr π π‘ β π π‘ ,1β πΏ π‘ β€ π π‘ then π‘th code is π π‘ + πΏ π‘ -close to capacity, and Pr Decode BSC Encode π β π β€πβ
π π‘ Need π π‘ =π 1 π (strong polarization βπ=πππ(π)) Need π π‘ =1/ π Ξ©(1) (β strong polarization) Arikan et al. π=πππ π ;π=π 1 ; [GX13,HAU13] β Oct. 8, 2018 Berkeley: General Strong Polarization
6
Part II: Martingales, Polarization, Strong & Local
Oct. 8, 2018 Berkeley: General Strong Polarization
7
Martingales and Polarization
[0,1]-Martingale: Sequence of r.v.s π 0 , π 1 , π 2 ,β¦ π π‘ β 0,1 , πΌ π π‘+1 | π 0 ,β¦, π π‘ = π π‘ Well-studied in algorithms: [MotwaniRaghavan, Β§4.4], [MitzenmacherUpfal, Β§12] E.g., π π‘ = Expected approx. factor of algorithm given first π‘ random choices. Usually study βconcentrationβ E.g. β¦ whp π final β[ π 0 Β±π] Today: Polarization: β¦ whp π final β π,1βπ β lim π‘ββ π π‘ = π(π 0 , π 2 ) β lim π‘ββ π π‘ = Bern π 0 Oct. 8, 2018 Berkeley: General Strong Polarization
8
Berkeley: General Strong Polarization
Some examples π π‘+1 = π π‘ + 4 βπ‘ w.p π π‘ β 4 βπ‘ w.p π π‘+1 = π π‘ + 2 βπ‘ w.p π π‘ β 2 βπ‘ w.p π π‘+1 = π π‘ w.p if π π‘ β€ π π‘ w.p if π π‘ β€ 1 2 π π‘+1 = π π‘ 2 w.p if π π‘ β€ π π‘ β π π‘ 2 w.p if π π‘ β€ 1 2 Converges! Uniform on [0,1] Polarizes (weakly)! Polarizes (strongly)! Oct. 8, 2018 Berkeley: General Strong Polarization
9
Main Result: Definition and Theorem
Strong Polarization: (informally) Pr π π‘ β(π,1βπ) β€π if π= 2 βπ(π‘) and π= 2 βπ(π‘) formally βπΎ>0 βπ½<1,π π .π‘. βπ‘ Pr π π‘ β( πΎ π‘ ,1β πΎ π‘ ) β€πβ
π½ π‘ Local Polarization: Variance in the middle: π π‘ β(π,1βπ) βπ>0 βπ>0 π .π‘. βπ‘, π π‘ β π,1βπ βπππ π π‘+1 π π‘ β₯π Suction at the ends: π π‘ β π,1βπ βπ>0, βπ<β, βπ>0 π .π‘. π π‘ <πβ Pr π π‘+1 < π π‘ π β₯π Theorem: Local Polarization β Strong Polarization. Both definitions qualitative! βlow endβ condition. Similar condition for high end Oct. 8, 2018 Berkeley: General Strong Polarization
10
Berkeley: General Strong Polarization
Proof (Idea): Step 1: The potential Ξ¦ π‘ β min π π‘ , 1β π π‘ decreases by constant factor in expectation in each step. βπΌ Ξ¦ π = exp βπ β Pr π π β₯ exp βπ/2 β€ exp βπ/2 Step 2: Next π time steps, π π‘ plummets whp 2.1: Say, If π π‘ β€π then Pr π π‘+1 β€ π π‘ β₯1/2. 2.2: Pr βπ‘β π,2π π .π‘. π π‘ >π β€ π π /π [Doob] 2.3: If above doesnβt happen π 2π < 5 βπ whp QED Oct. 8, 2018 Berkeley: General Strong Polarization
11
Part III: Polar Codes Encoding, Decoding, Martingale, Polarization
Oct. 8, 2018 Berkeley: General Strong Polarization
12
Lesson 0: Compression β Coding
Defn: Linear Compression Scheme: π,π· form compression scheme for bern π π if Linear map π: π½ 2 π β π½ 2 π Pr πβΌbern π π π· πβ
π β π =π 1 Hopefully π π β€π» π +π Compression β Coding Let π β₯ be such that πβ
π β₯ =0; Encoder: πβ¦ π β₯ β
π Error-Corrector: π= π β₯ β
π+πβ¦πβπ· πβ
π =πβπ· πβ
π β₯ β
π+π.π = π€.π. 1βπ(1) π β₯ β
π π π β₯ Oct. 8, 2018 Berkeley: General Strong Polarization
13
Question: How to compress?
Arikanβs key idea: Start with 2Γ2 βPolarization Transformβ: π,π β π+π,π Invertible β so does nothing? If π,π independent, then π+π βmore randomβ than either π | π+π βless randomβ than either Iterate (ignoring conditioning) End with bits that are almost random, or almost determined (by others). Output βtotally random partβ to get compression! Oct. 8, 2018 Berkeley: General Strong Polarization
14
Iteration elaborated:
Martingale: π π‘ β conditional entropy of random output after π‘ iterations B A+B A C+D A+B+C+D D C+D C D B+D H G+H G F E+F E F+H E+F+G+H Oct. 8, 2018 Berkeley: General Strong Polarization
15
The Polarization Butterfly
π π π,π = π π 2 π+π , π π 2 π Polarization ββπβ π π» π π βπ π π β0 π β€ π» π +π β
π π 1 π 2 π π 2 π π 2 +1 π π 2 +2 π π π 1 + π π 2 +1 π 2 + π π 2 +2 π π 2 + π π π π 2 +1 π π 2 +2 π π π 1 π 2 π π 2 π π 2 +1 π π 2 +2 π π π Compression πΈ π = π π π π Encoding time =π π log π (given π) π To be shown: 1. Decoding = π π log π 2. Polarization! Oct. 8, 2018 Berkeley: General Strong Polarization
16
The Polarization Butterfly: Decoding
Decoding idea: Given π, π π = π π π,π Compute π+πβΌBern 2πβ2 π 2 Determine πβΌ ? π 1 π 2 π π 2 π π 2 +1 π π 2 +2 π π π 1 + π π 2 +1 π 1 π 2 π π 2 π π 2 +1 π π 2 +2 π π π 2 + π π 2 +2 π|π+πβΌ π Bern( π π ) Key idea: Stronger induction! - non-identical product distribution π π 2 + π π π π 2 +1 Decoding Algorithm: Given π, π π = π π π,π , π 1 ,β¦, π π Compute π 1 ,.. π π 2 β π 1 β¦ π π Determine π+π=π· π π + , π 1 β¦ π π 2 Compute π 1 β¦ π π 2 β π 1 β¦ π π ;π+π Determine π=π· π π β , π 1 β¦ π π 2 π π 2 +2 π π Oct. 8, 2018 Berkeley: General Strong Polarization
17
Polarization Martingale
π 1 π 2 π π 2 π π 2 +1 π π 2 +2 π π π 1 + π π 2 +1 π 1 π 2 π π 2 π π 2 +1 π π 2 +2 π π Idea: Follow π π‘ = π» π π,π‘ π <π,π‘ for random π π 2 + π π 2 +2 π π,0 π π,1 β¦ Z π,π‘ β¦ π π, log π π π 2 + π π π π 2 +1 π π‘ Martingale! π π 2 +2 Polarizes? Strongly? β Locally? π π Oct. 8, 2018 Berkeley: General Strong Polarization
18
Local Polarization of Arikan Martingale
Variance in the Middle: Roughly: π» π ,π» π β(π» 2πβ2 π 2 , 2π» π βπ» 2πβ2 π 2 ) πβ π,1βπ β2πβ2 π 2 far from π + continuity of π» β
β π» 2πβ2 π 2 far from π»(π) Suction: High end:π» 1 2 βπΎ βπ» 1 2 β πΎ 2 π» 1 2 βπΎ =1 βΞ πΎ 2 β1β πΎ 2 β1 βΞ πΎ 4 Low end: π» π βπ log 1 π ;2π» π β2π log 1 π ; π» 2πβ2 π 2 βπ» 2π β2π log 1 2π β2π» π β2πβ 2β 1 log 1 π π» π Dealing with conditioning β more work (lots of Markov) Oct. 8, 2018 Berkeley: General Strong Polarization
19
Berkeley: General Strong Polarization
βNew contributionsβ So far: Reproduced old work (simpler proofs?) Main new technical contributions: Strong polarization of general transforms E.g. π,π,π β(π+π,π,π)? Exponentially strong polarization [BGSβ18] Suction at low end is very strong! Random πΓπ matrix yields π π‘+1 β π π‘ π whp Strong polarization of Markovian sources [GNSβ19?] Separation of compression of known sources from unknown ones. Oct. 8, 2018 Berkeley: General Strong Polarization
20
Berkeley: General Strong Polarization
Conclusions Polarization: Important phenomenon associated with bounded random variables. Not always bad ο ο Needs better understanding! Recently also used in CS: [Lovett Meka] β implicit [Chattopadhyay-Hatami-Hosseini-Lovett] - explicit Oct. 8, 2018 Berkeley: General Strong Polarization
21
Berkeley: General Strong Polarization
Thank You! Oct. 8, 2018 Berkeley: General Strong Polarization
22
Berkeley: General Strong Polarization
Rest of the talk Background/Motivation: The Shannon Capacity Challenge Resolution by [Arikanβ08], [Guruswami-Xiaβ13], [Hassani,Alishahi,Urbankeβ13]: Polar Codes! Key ideas: Polar Codes & Arikan-Martingale. Implications of Weak/Strong Polarization. Our Contributions: Simplification & Generalization Local Polarization β Strong Polarization Arikan-Martingale Locally Polarizes (generally) Oct. 8, 2018 Berkeley: General Strong Polarization
23
Context for todayβs talk
[Arikanβ08]: Martingale associated to general class of codes. Martingale polarizes (weakly) implies code achieves capacity (weakly) Martingale polarizes weakly for entire class. [GXβ13, HAUβ13]: Strong polarization for one specific code. Why so specific? Bottleneck: Strong Polarization Rest of talk: Polar codes, Martingale and Strong Polarization Oct. 8, 2018 Berkeley: General Strong Polarization
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.