randomizedlocal networkcomputing€¦ · coloring with one bad edge is ok, but not more. x...
TRANSCRIPT
![Page 1: RandomizedLocal NetworkComputing€¦ · Coloring with one bad edge is ok, but not more. X f-resilienttasksandBPLD Strategy: If the coloring is good, accept If the coloring is bad,](https://reader034.vdocuments.mx/reader034/viewer/2022050401/5f7f545909f91466a6735f56/html5/thumbnails/1.jpg)
Randomized LocalNetwork Computing
Laurent Feuilloley 1 · Pierre Fraigniaud 2
1ENS Cachan · Universite Paris Diderot · Aalto University2Universite Paris Diderot
SPAA 2015 · Portland · June 15
![Page 2: RandomizedLocal NetworkComputing€¦ · Coloring with one bad edge is ok, but not more. X f-resilienttasksandBPLD Strategy: If the coloring is good, accept If the coloring is bad,](https://reader034.vdocuments.mx/reader034/viewer/2022050401/5f7f545909f91466a6735f56/html5/thumbnails/2.jpg)
A derandomization theorem
Informally our theorem is :
In a distributed computing model
If a language can be checked locally with
randomization Then :
If it can be constructed locally with
randomization Then it can be constructed locally without
randomization
![Page 3: RandomizedLocal NetworkComputing€¦ · Coloring with one bad edge is ok, but not more. X f-resilienttasksandBPLD Strategy: If the coloring is good, accept If the coloring is bad,](https://reader034.vdocuments.mx/reader034/viewer/2022050401/5f7f545909f91466a6735f56/html5/thumbnails/3.jpg)
LOCAL model
A network of machines
Every vertex has a unique identifier
ID : 235Computation
Communication
![Page 4: RandomizedLocal NetworkComputing€¦ · Coloring with one bad edge is ok, but not more. X f-resilienttasksandBPLD Strategy: If the coloring is good, accept If the coloring is bad,](https://reader034.vdocuments.mx/reader034/viewer/2022050401/5f7f545909f91466a6735f56/html5/thumbnails/4.jpg)
LOCAL model
A network of machines
Every vertex has a unique identifier
ID : 235Computation
Computational power = ∞No failure
CommunicationBandwidth = ∞
![Page 5: RandomizedLocal NetworkComputing€¦ · Coloring with one bad edge is ok, but not more. X f-resilienttasksandBPLD Strategy: If the coloring is good, accept If the coloring is bad,](https://reader034.vdocuments.mx/reader034/viewer/2022050401/5f7f545909f91466a6735f56/html5/thumbnails/5.jpg)
LOCAL model
First point of view : minimize the number of rounds
max degree = ∆
![Page 6: RandomizedLocal NetworkComputing€¦ · Coloring with one bad edge is ok, but not more. X f-resilienttasksandBPLD Strategy: If the coloring is good, accept If the coloring is bad,](https://reader034.vdocuments.mx/reader034/viewer/2022050401/5f7f545909f91466a6735f56/html5/thumbnails/6.jpg)
LOCAL model
Second point of view : local computation
max degree = ∆
![Page 7: RandomizedLocal NetworkComputing€¦ · Coloring with one bad edge is ok, but not more. X f-resilienttasksandBPLD Strategy: If the coloring is good, accept If the coloring is bad,](https://reader034.vdocuments.mx/reader034/viewer/2022050401/5f7f545909f91466a6735f56/html5/thumbnails/7.jpg)
Theorem
In the LOCAL model
If a language can be checked locally withrandomization
Then :
If it can be constructed locally withrandomization
Then it can be constructed locally withoutrandomization
![Page 8: RandomizedLocal NetworkComputing€¦ · Coloring with one bad edge is ok, but not more. X f-resilienttasksandBPLD Strategy: If the coloring is good, accept If the coloring is bad,](https://reader034.vdocuments.mx/reader034/viewer/2022050401/5f7f545909f91466a6735f56/html5/thumbnails/8.jpg)
Construction vs Decision
Example : (∆ + 1)-coloring
Construction from a global perspective.
ID = 45
ID = 3
ID = 2
![Page 9: RandomizedLocal NetworkComputing€¦ · Coloring with one bad edge is ok, but not more. X f-resilienttasksandBPLD Strategy: If the coloring is good, accept If the coloring is bad,](https://reader034.vdocuments.mx/reader034/viewer/2022050401/5f7f545909f91466a6735f56/html5/thumbnails/9.jpg)
Construction vs Decision
Construction from a local perspective.
![Page 10: RandomizedLocal NetworkComputing€¦ · Coloring with one bad edge is ok, but not more. X f-resilienttasksandBPLD Strategy: If the coloring is good, accept If the coloring is bad,](https://reader034.vdocuments.mx/reader034/viewer/2022050401/5f7f545909f91466a6735f56/html5/thumbnails/10.jpg)
Construction vs Decision
Theorem (Linial’92) :
Constructing a (∆ + 1)-colouring requires Ω(log∗n)rounds.
![Page 11: RandomizedLocal NetworkComputing€¦ · Coloring with one bad edge is ok, but not more. X f-resilienttasksandBPLD Strategy: If the coloring is good, accept If the coloring is bad,](https://reader034.vdocuments.mx/reader034/viewer/2022050401/5f7f545909f91466a6735f56/html5/thumbnails/11.jpg)
Construction vs Decision
Decision from a local perspective
X
![Page 12: RandomizedLocal NetworkComputing€¦ · Coloring with one bad edge is ok, but not more. X f-resilienttasksandBPLD Strategy: If the coloring is good, accept If the coloring is bad,](https://reader034.vdocuments.mx/reader034/viewer/2022050401/5f7f545909f91466a6735f56/html5/thumbnails/12.jpg)
Construction vs Decision
Decision from a local perspective
![Page 13: RandomizedLocal NetworkComputing€¦ · Coloring with one bad edge is ok, but not more. X f-resilienttasksandBPLD Strategy: If the coloring is good, accept If the coloring is bad,](https://reader034.vdocuments.mx/reader034/viewer/2022050401/5f7f545909f91466a6735f56/html5/thumbnails/13.jpg)
Construction vs Decision
Decision from a global perspective
X
XX
XX
X
X
X
XX
X
X
![Page 14: RandomizedLocal NetworkComputing€¦ · Coloring with one bad edge is ok, but not more. X f-resilienttasksandBPLD Strategy: If the coloring is good, accept If the coloring is bad,](https://reader034.vdocuments.mx/reader034/viewer/2022050401/5f7f545909f91466a6735f56/html5/thumbnails/14.jpg)
Construction vs Decision
Decision from a global perspective
X
XX
X
X
X
XX
X
X
![Page 15: RandomizedLocal NetworkComputing€¦ · Coloring with one bad edge is ok, but not more. X f-resilienttasksandBPLD Strategy: If the coloring is good, accept If the coloring is bad,](https://reader034.vdocuments.mx/reader034/viewer/2022050401/5f7f545909f91466a6735f56/html5/thumbnails/15.jpg)
Theorem
In the LOCAL model
If a language can be checked locally withrandomization
Then :
If it can be constructed locally withrandomization
Then it can be constructed locally withoutrandomization
![Page 16: RandomizedLocal NetworkComputing€¦ · Coloring with one bad edge is ok, but not more. X f-resilienttasksandBPLD Strategy: If the coloring is good, accept If the coloring is bad,](https://reader034.vdocuments.mx/reader034/viewer/2022050401/5f7f545909f91466a6735f56/html5/thumbnails/16.jpg)
Locally ?
Here locally means constant number of rounds
Coloring verification can be done locally → 1 round,
but coloring construction cannot→ log∗n rounds.
![Page 17: RandomizedLocal NetworkComputing€¦ · Coloring with one bad edge is ok, but not more. X f-resilienttasksandBPLD Strategy: If the coloring is good, accept If the coloring is bad,](https://reader034.vdocuments.mx/reader034/viewer/2022050401/5f7f545909f91466a6735f56/html5/thumbnails/17.jpg)
Locally ?
What can be constructed locally ?
→ Weak coloring, fractional
coloring, and someapproximations
![Page 18: RandomizedLocal NetworkComputing€¦ · Coloring with one bad edge is ok, but not more. X f-resilienttasksandBPLD Strategy: If the coloring is good, accept If the coloring is bad,](https://reader034.vdocuments.mx/reader034/viewer/2022050401/5f7f545909f91466a6735f56/html5/thumbnails/18.jpg)
Theorem
In the LOCAL model
If a language can be checked in O(1) rounds,with randomization
Then :
If it can be constructed in O(1) rounds withrandomization
Then it can be constructed in O(1) roundswithout randomization
![Page 19: RandomizedLocal NetworkComputing€¦ · Coloring with one bad edge is ok, but not more. X f-resilienttasksandBPLD Strategy: If the coloring is good, accept If the coloring is bad,](https://reader034.vdocuments.mx/reader034/viewer/2022050401/5f7f545909f91466a6735f56/html5/thumbnails/19.jpg)
Languages and classes
a language : is a set(G , x) satisfying a property P
A class is a set of languages
→ LD = the languages thatcan be checked in constanttime deterministically.
X
![Page 20: RandomizedLocal NetworkComputing€¦ · Coloring with one bad edge is ok, but not more. X f-resilienttasksandBPLD Strategy: If the coloring is good, accept If the coloring is bad,](https://reader034.vdocuments.mx/reader034/viewer/2022050401/5f7f545909f91466a6735f56/html5/thumbnails/20.jpg)
Languages and classes
BPLD = the languages that can be checked in
constant time using randomization.
More precisely :
there exists a checker, and p ∈(
1
2, 1]
s.t. :
If (G , x) ∈ L, then Pr[all nodes accept] ≥ p
If (G , x) /∈ L then Pr[a node rejects] ≥ p
![Page 21: RandomizedLocal NetworkComputing€¦ · Coloring with one bad edge is ok, but not more. X f-resilienttasksandBPLD Strategy: If the coloring is good, accept If the coloring is bad,](https://reader034.vdocuments.mx/reader034/viewer/2022050401/5f7f545909f91466a6735f56/html5/thumbnails/21.jpg)
f -resilient tasks and BPLD
BPLD
LD
f -resilient tasks
Lf = L ∈ LD
+ errors accepted on f nodes.
![Page 22: RandomizedLocal NetworkComputing€¦ · Coloring with one bad edge is ok, but not more. X f-resilienttasksandBPLD Strategy: If the coloring is good, accept If the coloring is bad,](https://reader034.vdocuments.mx/reader034/viewer/2022050401/5f7f545909f91466a6735f56/html5/thumbnails/22.jpg)
f -resilient tasks and BPLD
Coloring with one bad edge is ok, but not more.
X
![Page 23: RandomizedLocal NetworkComputing€¦ · Coloring with one bad edge is ok, but not more. X f-resilienttasksandBPLD Strategy: If the coloring is good, accept If the coloring is bad,](https://reader034.vdocuments.mx/reader034/viewer/2022050401/5f7f545909f91466a6735f56/html5/thumbnails/23.jpg)
f -resilient tasks and BPLD
Strategy : If the coloring is good, accept
If the coloring is bad, reject withprobability q
Number of bad edges
Probability that all nodes accept1
1 2
b
bp
![Page 24: RandomizedLocal NetworkComputing€¦ · Coloring with one bad edge is ok, but not more. X f-resilienttasksandBPLD Strategy: If the coloring is good, accept If the coloring is bad,](https://reader034.vdocuments.mx/reader034/viewer/2022050401/5f7f545909f91466a6735f56/html5/thumbnails/24.jpg)
Back to the derandomization
theorem
More formally the theorem is :
In the LOCAL model
If L ∈ BPLD Then :
If L can be constructed in O(1) rounds withrandomization
Then it can be constructed in O(1) rounds
deterministically
![Page 25: RandomizedLocal NetworkComputing€¦ · Coloring with one bad edge is ok, but not more. X f-resilienttasksandBPLD Strategy: If the coloring is good, accept If the coloring is bad,](https://reader034.vdocuments.mx/reader034/viewer/2022050401/5f7f545909f91466a6735f56/html5/thumbnails/25.jpg)
A glimpse of the proof
Two main steps :
Using Ramsey theory to reduce to a special case(from Naor-Stockmeyer’93)
Proving that locality prevent weird correlations
![Page 26: RandomizedLocal NetworkComputing€¦ · Coloring with one bad edge is ok, but not more. X f-resilienttasksandBPLD Strategy: If the coloring is good, accept If the coloring is bad,](https://reader034.vdocuments.mx/reader034/viewer/2022050401/5f7f545909f91466a6735f56/html5/thumbnails/26.jpg)
Further works
LD → BPLD → ?
Get a better understanding of randomization innetwork distributed computing.
![Page 27: RandomizedLocal NetworkComputing€¦ · Coloring with one bad edge is ok, but not more. X f-resilienttasksandBPLD Strategy: If the coloring is good, accept If the coloring is bad,](https://reader034.vdocuments.mx/reader034/viewer/2022050401/5f7f545909f91466a6735f56/html5/thumbnails/27.jpg)
Further works
LD → BPLD → ?
Get a better understanding of randomization in
network distributed computing.
Thank you !