Christoph Lenzen and Roger Wattenhoferwebcourse.cs.technion.ac.il/236825/Spring2015/ho...Β Β· -...
Transcript of Christoph Lenzen and Roger Wattenhoferwebcourse.cs.technion.ac.il/236825/Spring2015/ho...Β Β· -...
Leveraging Linialβs Locality limit
Christoph Lenzen and Roger Wattenhofer
Presented by Yam Chemel
Linial (1992)
3-coloring and MIS
on a ring (π π)
takes at least logβ(π) rounds
Reminder
Prove logβ(π) Lower bound for MaxIS algorithms in rings
Our Goal:
How?
MaxIS π(logβ π )
3-coloring π logβ π
Contradicts Linial!
MaxIS vs. MIS - Example
Independent Set MIS MaxIS
(general graphs)
MIS βΈ MaxIS
MaxIS in rings
1 2
3 4
1 2
3
4
5
Q: What if each even node returns 1, and each odd node returns 0?
1 2
3 4
π 4 MaxIS
1 2
3
4
5 π 5
MaxIS
1 3
2 4
MaxIS
Identifiers are not necessarily in order!
1 3
2 4
MaxIS vs. MIS β in rings
MIS βΈ MaxIS
Algorithm π΄ MaxIS
π(logβ π )
Algorithm π΄ MIS
π logβ π
Contradicts Linial!
f-approximation algorithm for MaxIS
π -MaxIS solution πΌ β algorithm A solution
πΌ = π
If π΄ is π(π)-approximation: πΌ β€ π
and also:
πΌ β π(π) β₯ |π|
π΄ π π = π π
A is 2-approximation for MaxIS
πΌ =π
4
π
4β 2 β₯ π =
π
2
Leveraging Linialβs Locality limit
MaxIS approximation π(πππβ π )
3-coloring π πππβ π
Leveraging Linialβs Locality limit
MaxIS approximation π π
3-coloring π(π)
Leveraging Linialβs Locality limit
π(π)-alternating π(π)
MaxIS approximation π π
3-coloring π(π)
π(π)-alternating algorithm - Definition:
Suppose A is an algorithm operating on π π which assigns each node π£π β ππ a value π(π£π) β {0, 1}. We call A π(n)-alternating, if the length of any monochromatic sequence
π π£π = π π£π+1 = β― = π π£π+π is smaller than π(n).
< π(π)
π΄ π π = π π
π(π)-alternating algorithm - examples
n is even
Longest monochromatic sequence = 1
n is odd
Longest monochromatic sequence = 2
A is 3-alternating (also 4-alternating, 5-alternatingβ¦)
(Not necessarily MaxIS algorithm)
π΄ π π = π π
A is (n+1)-alternating
(Not necessarily MaxIS algorithm)
π(π)-alternating algorithm - examples
Lemma - Modified MaxIS approximation Suppose an f-approximation algorithm A for the MaxIS problem on the ring π π running in at most g(n) β₯ 1 rounds is given, where we have π(π)π(π) β π(logβ(π)). Then an π(logβ(π))-alternating algorithm Aβ² requiring π logβ π communication rounds exists.
A
MaxIS, π(π)-approximation,
π(π) rounds, π π π π β π πππβ π
Aβ π πππβ π -alternating,
π πππβ π rounds
Modified MaxIS approximation - Proof
General idea
ππ
π π
π πβ² π
ππβ²
Observation: For each node π£π, π(π£π) is a function of:
π£π
π(π) π(π)
π π = #πππ’πππ
β’ The IDs of its π(π) neighbors on each side
β’ π β’ Its ID
π(π) π(π)
Modified MaxIS approximation - Proof
β’ πππ‘ π π β 10π(π)π(π)
ππ =
π£1, β¦ , π£π π , π£π π +1, β¦, π£π π +π π , π£π π +1+π π , β¦ , π£2π π +π π£
| βπ β π£π π +1, β¦, π£π π +π π π π£π = 0
0 0
?
?
0
? ?
? ?
π(π) π(π) π(π)
Modified MaxIS approximation - Proof
β’ Define exactly the set of sequences preventing that A is π(π)-alternating
π(π) π(π)
? ? ? 0 0 0 ? ? ?
Id=2 Id=11 Id=8 Id=9 Id=14 Id=21 Id=7 Id=6 Id=3
π(π)
π = 2, 11, 8, 9, 14, 21, 7, 6, 3 β ππ
Take 2π π + π(π) consecutive nodes in π π
Assign them identifiers
Run A on the sequence / on π π containing the sequence.
If the π(π) center nodes compute 0, add s to ππ
Id=2 Id=11 Id=8 Id=9 Id=14 Id=21 Id=7 Id=6 Id=3
Building ππ
2π π + π(π)
Construct a sequence of identifiers for π π:
π 2
π 1 π β π
π π π π
Modified MaxIS approximation - Proof
1. Choose an arbitrary sequence s from ππ and assign the identifiers to π£1, β¦ , π£ π .
2. Assuming we already assigned labels to the nodes π£1, β¦ , π£π:
While there exists a sequence π β ππ that can be appended to π£1, β¦ , π£π without duplicating an identifier, we do so.
3. If no further sequence fits, we add π β π arbitrary identifiers not yet present in π£1, β¦ , π£π to
complete the labeling (π£1, β¦ , π£π) of π π.
Assume for contradiction, that for arbitrarily large n it is possible to
label π π as described, with at least π βπ
5π π identifiers
stemming from sequences out of ππ.
# πππππ π‘βππ‘ πππππ’π‘π 1 ππ π π β€ π βπ π β π
π π + 2π πβ
π
5π π
β₯ πππππ π‘βππ‘ πππππ’π‘π 0 ππ π β ππ
πππππ ππ π β ππβ π β #(πππππ πππ‘ ππ π β ππ)
Modified MaxIS approximation - Proof
# πππππ π‘βππ‘ πππππ’π‘π 0 ππ π π β₯
=π π
π π + 2π(π)β π β
π
5π π
# πππππ π‘βππ‘ πππππ’π‘π 1 ππ π π β€ π βπ π β π
π π + 2π πβ
π
5π π
= π 1 βπ π
π π + 2π π+
1
5π π
π π = 10π π π(π)
Modified MaxIS approximation - Proof
= π 1 β10π π π π
10π π π π + 2π π+
1
5π π
= π 1 β5π π
5π π + 1+
1
5π π
= π 1 β 1 β1
5π π + 1+
1
5π π
= π1
5π π + 1+
1
5π π
β€ π1
5π π + 0+
1
5π π
=2π
5π π
πΌ = # πππππ π‘βππ‘ πππππ’π‘π 1 ππ π π β€2π
5π π
However, according to f-approximation definition: π π πΌ β₯ π
(M -an arbitrary MaxIS of G)
π 6
Contradicts the assumption! (βFor arbitrarily large n it is possible to label π π as described, with at least
π βπ
5π π identifiers stemming from sequences out of ππ.β)
Modified MaxIS approximation - Proof
π(π) πΌ β€2
5π
Choosing every other node in π π is a MaxIS solution:
π =π
2
Combining the equations:
π π πΌ β₯π
2
At least π
ππ π identifiers remain which cannot form a further
sequence from ππ.
Set πβ² β max π0, 5π π β π
In π πβ²
#πππππππππ πππππ‘ππππππ
β₯πβ²
5π πβ²
For a large n (π > π0) it is possible to label π π as described, with
at most π βπ
5π π identifiers stemming from sequences out of πΊπ
Modified MaxIS approximation - Proof
π 2
π 1 π β π
π π π π
β₯5π π β π
5π π= π
π π π πβ²
ππ
Algorithm Aβ : ππ’π βΆ π΄(πβ², π πππππ‘ππππππ ππππ π πβ² )
Modified MaxIS approximation - Proof
π 4 π 11
1
2
3
4
3
8
1
4
π4
Algorithm Aβ : π π’π π΄(11, 3, 8, 1, 4 )
Example:
Modified MaxIS approximation - Proof
Aβ is π πβ² -alternating
No π(πβ²) consecutive 0βs No π(πβ²) consecutive 1βs
π΄ computes an independent setβ no 2 neighbors in independent set β
no 2 consecutive 1βs
No π β ππβ² in remaining n nodes of π πβ² β No π πβ² consecutive nodes compute 0 in
Aβ run
Maximum monochromatic sequence : π πβ² β 1 β π π πβ² π πβ² β π logβ πβ² = π logβ π
Aβ running time -π(πππβ(π)) rounds π π β€ π π π π β π(πππβ π)
π πβ² β π(πππβ π)
Aβ is π πππβ π -alternating
Leveraging Linialβs Locality limit
π(π)-alternating π(π)
MaxIS approximation π π
3-coloring π(π)
Leveraging Linialβs Locality limit
π(π)-alternating π(π)
MaxIS approximation π π
3-coloring π(π)
2. Proof : Lemma β Given a π(π)-alternating algorithm A running in π(π(π)) rounds, a 3-coloring of the ring can be computed in π(π(π)) rounds.
1. Run A. Let π(π£) β {0, 1} denote the result of this run.
2. Find a pair of neighboring nodes {π€1, π€2} with π π€1 β π π€2 which is closest to v.
0 1 π π€2
1 0 π π€2
If π£ β {π€1, π€2}: if π(π£) = 0: set π π£ β π otherwise: set π(π£) β π
0 1 π π€2
1 0 π π€2
1 0 1 1
π€2 π€1 π π€
1 0 1 1
Else: denote by πΏ the distance to the closer node in {π€1, π€2}, w.l.o.g. π€1. if πΏ β 2β: set π π£ β π(π€1) else: set π π£ β π(π€2)
1 0 1 1
π€2 π€1 π π€ (πΏ β 2β)
πΏ
3 8 π π€
8 3 π π€
3. If v has a neighbor w with π(π£) = π(π€) and v > w, set π(π£) β π.
3 8 π π€
8 3 π π€
3 8 π π€
1 2 3 8 π π€
1 2 (3)
4. If v has a neighbor w with π(π£) = π(π€) = π and v > w, set c(v) to the color none of the neighbors of v has.
3 8 π π€
1 2 (4)
2. Proof : Lemma β Given a π(π)-alternating algorithm A running in π(π(π)) rounds, a 3-coloring of the ring can be computed in π(π(π)) rounds.
Running time: πΆ π π
Step 1: Running A- π π π
Step 2: Finding a pair of neighbors with different d - π π .
No more than π(π£) consecutive nodes take the same decision d(v)
since A is π(π)-alternating.
Valid 3-coloring of πΉπ
Step 2: Each node π£ chooses different from one of its neighbors,
1 1 0 1 1 1 1 0 1
so at most one of the neighbors of π£ may take the same choice.
Step 3: From each pair of neighbors with the same color one chooses g.
Step 4: If that same color was g, v chooses the color non of its neighbors has.
v v
v
1 1 0
v
1 1 1 1 0 1
v v
Leveraging Linialβs Locality limit
π(π)-alternating π(π)
MaxIS approximation π π
3-coloring π(π)
Therefore, there isnβt a MaxIS approximation algorithms running on a ring
in less than log*(n)
Proof - Summary
Assume by contradiction that there exists a MaxIS approximation algorithm A running in less than πππβ(π).
Construct a π πππβ π -alternating algorithm running in π(πππβ(π)) using algorithm A.
By lemma 2, a 3-coloring of the ring can be computed in π(πππβ(π)) rounds.
This contradicts Linialβs 3-coloring lower bound.
- Discuss MDS lower bound - Compare MDS and MaxIS difficulty
Our (new) Goal:
No! Weβll show a case where MIS is easier than MDS
MDS on rings O(1)
Is MDS always easier than MIS?
How?
MDS β Minimum Dominating Set
DS MDS
MDS in rings
Taking every third node gives a minimum dominating set
π =π
3
Taking every node gives a 3-approximation MDS in 1 round β π logβ π
π π
π π
There is no πππβ(π) bound MDS approximation in rings! MDS approximation in rings takes πΆ(π) rounds
MDS approximation in rings
Can we compare MaxIS and MDS difficulty?
MDS
MaxIS
We saw that in π π MDS can be computed in 1 round, but MaxIS requires at least πππβ(π) round.
Is it always easier to compute MDS than MaxIS?
MaxIS graph family
πΊ π£ π€
Any graph that can be constructed this way
Lemma (Local computation of a MaxIS on MaxIS Graphs): The set {π£ β π | |π1
+(π£)|πππ 2 = 1} is a MaxIS for any MaxIS Graph.
Proof β part 1:
- For π£π β ππ π β {3,4}, π£π has 2|π1+ π£ | neighbors.
- π1+ π£π = 2 π1
+ π£ + 1 β π1+ π£π ππ πππ
- For π£π β ππ π β {1,2}, π1+ π£π = 4 π1
+ π£ β π1+ π£π ππ ππ£ππ
- {π£ β π | |π1+(π£)|πππ 2 = 1} = π3 βͺ π4
Proof β part 2:
- π3 βͺ π4 is an Independent set β according to the construction of
MaxIS graph
- (π£1, π£3, π£2, π£4) forms a cycle, so for each 4 nodes as such, only 2 can be in the IS.
- MaxIS canβt be larger than π
2
- π3 βͺ π4 =π
2
β π3 βͺ π4 ππ π πππ₯πΌπ
Lemma (Local computation of a MaxIS on MaxIS Graphs): The set {π£ β π | |π1
+(π£)|πππ 2 = 1} is a MaxIS for any MaxIS Graph.
Conclusion β MaxIS on a MaxIS graph can be determined locally, without communication (in π(1) rounds).
MDS on MaxIS graphs
We can prove that MDS on MaxIS graphs is as efficient as in general graphs, meaning:
Ξ©log π
log log π
MaxIS
MDS
Ring graphs - π π MaxIS graphs
MaxIS
MDS π(1)
Ξ©log π
log log π
MDS
MaxIS π(1)
Ξ© logβ π
MaxIS and MDS are not comparable in general graphs!