Distributed Algorithms Below the Greedy Regime
Yannic Maus
Coloring
This project has received funding from the European Unionโs Horizon 2020 Research and Innovation Programme under grant agreement no. 755839.
Mohsen Ghaffari , Juho Hirvonen, Fabian Kuhn, Jara Uitto
Communication Network = Problem Instance:
3
LOCAL Model [Linial; FOCS โ87]
5
15
6
11
1
21
33
26
9
278
2
7
Discrete synchronous rounds:โข local computationsโข exchange messages with all neighbors
๐ฎ = ๐ฝ, ๐ฌ ,๐ = ๐ฝ
(computations unbounded, message sizes are unbounded)
time complexity = number of rounds
I am green
5
15
6
11
1
21
33
26
9
2
7
827
5
15
6
11
1
21
33
26
9
2
7
827
Communication Network = Problem Instance:
4
CONGEST MODEL
5
15
6
11
1
21
33
26
9
278
2
7
Discrete synchronous rounds:โข local computationsโข exchange messages with all neighbors
(computations unbounded, message sizes are unbounded) )
time complexity = number of rounds
๐ถ(๐ฅ๐จ๐ ๐) bits
๐ฎ = ๐ฝ, ๐ฌ ,๐ = ๐ฝ
5
15
6
11
1
21
33
26
9
2
7
827
5
Classic Big Four (Greedy Regime)
Maximal Ind. Set (MIS) ๐ซ + ๐ -Vertex Coloring
(๐๐ซ โ ๐)-Edge ColoringMaximal Matching
(ฮ: maximum degree of ๐บ)
6
In the LOCAL Model โฆGreedy Below Greedy
Maximal IS 2๐ log ๐ 2๐ log ๐ Maximum IS, ๐ โ ๐ -approx.
vertex cover, 2-approx.
poly log ๐ 2๐ log ๐ vertex cover, ๐ + ๐ -approx.
min. dominating set,(๐ + ๐) ๐ฅ๐จ๐ ๐ซ-approx.
2๐ log ๐ 2๐ log ๐ min dominating set,๐ + ๐ -approx.
hypergraph vertex cover, rank-approx.
2๐ log ๐ 2๐ log ๐ hypergraph vertex cover,๐ + ๐ -approx.
๐ซ + ๐ -vertex coloring 2๐ log ๐ 2๐ log ๐ ๐ซ-vertex coloring
๐๐ซ โ ๐ -edge coloring poly log ๐ poly log ๐ ๐ + ๐ ๐ซ-edge coloring
maximal matching poly log ๐ poly log ๐ Maximum Matching,๐ + ๐ -approx.
CONGEST
7
In the LOCAL Model โฆGreedy Below Greedy
Maximal IS 2๐ log ๐ 2๐ log ๐ Maximum IS, ๐ โ ๐ -approx.
๐ซ + ๐ -vertex coloring 2๐ log ๐ 2๐ log ๐ ๐ซ-vertex coloring
๐๐ซ โ ๐ -edge coloring poly log ๐ poly log ๐ ๐ + ๐ ๐ซ-edge coloring
maximal matching poly log ๐ poly log ๐ Maximum Matching,๐ + ๐ -approx.
CONGEST
โProblems that do not have easy sequential greedy algorithms.โ
โProblems that do haveeasy sequential greedy algorithms.โ
8
OutlineThis Talk: How do we use LOCAL? Below Greedy
Maximum IS7/8-approx.
ฮฉ(๐2) 2๐ log ๐ Maximum IS, ๐ โ ๐ -approx.
vertex cover, exact
ฮฉ(๐2) 2๐ log ๐ vertex cover, ๐ + ๐ -approx.
min. dominating set,exact
ฮฉ(๐2) 2๐ log ๐ min dominating set,๐ + ๐ -approx.
hypergraph vertex cover, exact
ฮฉ(๐2) 2๐ log ๐ hypergraph vertex cover,๐ + ๐ -approx.
๐(๐ฎ)-vertex coloring ฮฉ(๐2) 2๐ log ๐ ๐ซ-vertex coloring
edge coloring ? poly log ๐ ๐ + ๐ ๐ซ-edge coloring
Maximum matchingexact
? poly log ๐ Maximum Matching,๐ โ ๐ -approx.
Technique 1: Ball growing
Technique 2: Local filling
Technique 3: Aug. paths
CONGEST
[Ghaffari, Kuhn, Maus; STOC โ17]
[Ghaffari, Hirvonen, Kuhn, Maus; PODC โ18]
[Ghaffari, Kuhn, Maus, Uitto; STOC โ18]
9
Technique 1: Ball GrowingThis Talk: How do we use LOCAL? Below Greedy
Maximum IS7/8-approx.
ฮฉ(๐2) 2๐ log ๐ Maximum IS, ๐ โ ๐ -approx.
vertex cover, exact
ฮฉ(๐2) 2๐ log ๐ vertex cover, ๐ + ๐ -approx.
min. dominating set,exact
ฮฉ(๐2) 2๐ log ๐ min dominating set,๐ + ๐ -approx.
hypergraph vertex cover, exact
ฮฉ(๐2) 2๐ log ๐ hypergraph vertex cover,๐ + ๐ -approx.
๐(๐ฎ)-vertex coloring ฮฉ(๐2) 2๐ log ๐ ๐ซ-vertex coloring
edge coloring ? poly log ๐ ๐ + ๐ ๐ซ-edge coloring
Maximum matchingexact
? poly log ๐ Maximum Matching,๐ โ ๐ -approx.
Technique 1: Ball growing
Technique 2: Local filling
Technique 3: Aug. paths
10
Sequential Ball Growing (MaxIS)
๐
Safe Ball ๐ฉ๐(๐): |๐๐๐ฅ๐ผ๐(๐ต๐+1)| < (1 + ๐) โ ๐๐๐ฅ๐ผ๐(๐ต๐)
Terminates with small radius ๐ = ๐(๐โ1 log ๐).
Find safe ball: Set ๐ = 0 and increase ๐ until ball ๐ต๐ is safe.
11
Sequential Ball Growing (MaxIS)
Safe Ball ๐ฉ๐(๐): |๐๐๐ฅ๐ผ๐(๐ต๐+1)| < (1 + ๐) โ ๐๐๐ฅ๐ผ๐(๐ต๐)
Terminates with small radius ๐ = ๐(๐โ1 log ๐).
Find safe ball: Set ๐ = 0 and increase ๐ until ball ๐ต๐ is safe.
12
Sequential Ball Growing (MaxIS)
Safe Ball ๐ฉ๐(๐): |๐๐๐ฅ๐ผ๐(๐ต๐+1)| < (1 + ๐) โ ๐๐๐ฅ๐ผ๐(๐ต๐)
Terminates with small radius ๐ = ๐(๐โ1 log ๐).
Find safe ball: Set ๐ = 0 and increase ๐ until ball ๐ต๐ is safe.
13
Sequential Ball Growing (MaxIS)
Safe Ball ๐ฉ๐(๐): |๐๐๐ฅ๐ผ๐(๐ต๐+1)| < (1 + ๐) โ ๐๐๐ฅ๐ผ๐(๐ต๐)
Terminates with small radius ๐ = ๐(๐โ1 log ๐).
Find safe ball: Set ๐ = 0 and increase ๐ until ball ๐ต๐ is safe.
14
Sequential Ball Growing (MaxIS)
Safe Ball ๐ฉ๐(๐): |๐๐๐ฅ๐ผ๐(๐ต๐+1)| < (1 + ๐) โ ๐๐๐ฅ๐ผ๐(๐ต๐)
Terminates with small radius ๐ = ๐(๐โ1 log ๐).
Find safe ball: Set ๐ = 0 and increase ๐ until ball ๐ต๐ is safe.
15
Sequential Ball Growing (MaxIS)
Safe Ball ๐ฉ๐(๐): |๐๐๐ฅ๐ผ๐(๐ต๐+1)| < (1 + ๐) โ ๐๐๐ฅ๐ผ๐(๐ต๐)
Terminates with small radius ๐ = ๐(๐โ1 log ๐).
Find safe ball: Set ๐ = 0 and increase ๐ until ball ๐ต๐ is safe.
Sequentially computes a 1 + ๐ โ1-approximation for MaxIS.(using unbounded computation)
16
Parallel Ball Growing
TheoremUsing (๐ฉ๐จ๐ฅ๐ฒ ๐ฅ๐จ๐ ๐ , ๐ฉ๐จ๐ฅ๐ฒ ๐ฅ๐จ๐ ๐)-network decompositions โsequentially ball growingโ can be โdone in parallelโ in LOCAL.
Corollary
There are ๐ฉ๐จ๐ฅ๐ฒ ๐ฅ๐จ๐ ๐ randomized and ๐๐ถ ๐ฅ๐จ๐ ๐ deterministic ๐ + ๐ -approximation algorithms for covering and packing
integer linear programs.
This includes maximum independent set, minimum dominating set, vertex cover, โฆ .
[STOC โ17, Ghaffari, Kuhn, Maus]
[STOC โ17, Ghaffari, Kuhn, Maus]
17
Technique 2: Local FillingThis Talk: How do we use LOCAL? Below Greedy
Maximum IS7/8-approx.
ฮฉ(๐2) 2๐ log ๐ Maximum IS, ๐ โ ๐ -approx.
vertex cover, exact
ฮฉ(๐2) 2๐ log ๐ vertex cover, ๐ + ๐ -approx.
min. dominating set,exact
ฮฉ(๐2) 2๐ log ๐ min dominating set,๐ + ๐ -approx.
hypergraph vertex cover, exact
ฮฉ(๐2) 2๐ log ๐ hypergraph vertex cover,๐ + ๐ -approx.
๐(๐ฎ)-vertex coloring ฮฉ(๐2) 2๐ log ๐ ๐ซ-vertex coloring
edge coloring ? poly log ๐ ๐ + ๐ ๐ซ-edge coloring
Maximum matchingexact
? poly log ๐ Maximum Matching,๐ โ ๐ -approx.
Technique 1: Ball growing
Technique 2: Local filling
Technique 3: Aug. paths
18
ฮ-Coloring
Definition: An induced subgraph ๐ป โ ๐บ is called an easy component if any ๐ฅ๐บ-coloring of ๐บ โ ๐ป can be extended to a ๐ฅ๐บ-coloring of ๐บ without changing the coloring on ๐บ โ ๐ป.
[PODC โ18; Ghaffari, Hirvonen, Kuhn, Maus]
Theorem: Let ๐บ be a graph (โ clique) with max. degree ฮ โฅ 3. Every node of ๐บ has a small diameter easy component in distance at most O(log n).
Well studied under the name degree chosable components.
โ โ[Erdลs et al. โ79, Vizing โ76]
Previous Work: [Panconesi, Srinivasan; STOC โ93]
19
Find an MIS ๐ด of small diameter easy componentsDefine ๐ log๐ Layers: ๐ณ๐ = ๐ฃ ๐ฃ in distance ๐ to some component in ๐ด}For ๐ = ๐(log ๐) to 1
color nodes in ๐ฟ๐ through solving a (deg+1)-list coloringColor easy components in M
20
Find an MIS ๐ด of small diameter easy componentsDefine ๐ log๐ Layers: ๐ณ๐ = ๐ฃ ๐ฃ in distance ๐ to some component in ๐ด}For ๐ = ๐(log ๐) to 1
color nodes in ๐ฟ๐ through solving a (deg+1)-list coloringColor easy components in M
21
Find an MIS ๐ด of small diameter easy componentsDefine ๐ log๐ Layers: ๐ณ๐ = ๐ฃ ๐ฃ in distance ๐ to some component in ๐ด}For ๐ = ๐(log ๐) to 1
color nodes in ๐ฟ๐ through solving a (deg+1)-list coloringColor easy components in M
Greedy regime
22
Find an MIS ๐ด of small diameter easy componentsDefine ๐ log๐ Layers: ๐ณ๐ = ๐ฃ ๐ฃ in distance ๐ to some component in ๐ด}For ๐ = ๐(log ๐) to 1
color nodes in ๐ฟ๐ through solving a (deg+1)-list coloringColor easy components in M
Greedy regime
23
Find an MIS ๐ด of small diameter easy componentsDefine ๐ log๐ Layers: ๐ณ๐ = ๐ฃ ๐ฃ in distance ๐ to some component in ๐ด}For ๐ = ๐(log ๐) to 1
color nodes in ๐ฟ๐ through solving a (deg+1)-list coloringColor easy components in M
Greedy regime
24
Find an MIS ๐ด of small diameter easy componentsDefine ๐ log๐ Layers: ๐ณ๐ = ๐ฃ ๐ฃ in distance ๐ to some component in ๐ด}For ๐ = ๐(log ๐) to 1
color nodes in ๐ฟ๐ through solving a (deg+1)-list coloringColor easy components in M
Greedy regime
25
Technique 3: Augmenting PathsThis Talk: How do we use LOCAL? Below Greedy
Maximum IS7/8-approx.
ฮฉ(๐2) 2๐ log ๐ Maximum IS, ๐ โ ๐ -approx.
vertex cover, exact
ฮฉ(๐2) 2๐ log ๐ vertex cover, ๐ + ๐ -approx.
min. dominating set,exact
ฮฉ(๐2) 2๐ log ๐ min dominating set,๐ + ๐ -approx.
hypergraph vertex cover, exact
ฮฉ(๐2) 2๐ log ๐ hypergraph vertex cover,๐ + ๐ -approx.
๐(๐ฎ)-vertex coloring ฮฉ(๐2) 2๐ log ๐ ๐ซ-vertex coloring
edge coloring ? poly log ๐ ๐ + ๐ ๐ซ-edge coloring
Maximum matchingexact
? poly log ๐ Maximum Matching,๐ โ ๐ -approx.
Technique 1: Ball growing
Technique 2: Local filling
Technique 3: Aug. paths
26
1 + ๐ ฮ-Edge Coloring
Theorem1 + ๐ ฮ-edge coloring can be efficiently reduced to the
computation of weighted maximum matching approximations .
The reduction can be executed in the CONGEST model.
[Ghaffari, Kuhn, Maus, Uitto; STOC โ18]
For ๐ = 1 to 2ฮ โ 1compute a maximal matching ๐ of ๐บ ๐good
color edges of ๐ with color ๐ ๐good
remove ๐ from ๐บ๐good
Next
Well known: 2ฮ โ 1 iterations suffice to color all edges.
27
1 + ๐ ฮ-Edge Coloring
๐ + ๐ ๐ซ-edge coloring through good matchings:Reduce the max degree (amortized) at a rate of (1 โ ๐).
For ๐ = 1 to 1 + ๐ ฮcompute a good matching ๐good of ๐บ
color edges of ๐good with color ๐
remove ๐good from ๐บ
Next
Theorem1 + ๐ ฮ-edge coloring can be efficiently reduced to the
computation of weighted maximum matching approximations .
The reduction can be executed in the CONGEST model.
[Ghaffari, Kuhn, Maus, Uitto; STOC โ18]
Technique 1: Sequential ball growingProblems: Approx. for MaxIS, MinDS, MinVC and many more โฆHow do we (ab)use LOCAL?
Compute optimal solutions in small diameter graphs
28
Summary Techniques (LOCAL)
Technique 2: Local fillingProblems: ฮ-Coloring, ?How do we (ab)use LOCAL?
โExistenceโ + small diameter is enough to obtain a solution
Technique 3: Augmenting pathsProblems: 1 + ๐ ฮ Edge Coloring, Maximum Matching Approx.How do we (ab)use LOCAL?
Finding a maximal set of augmenting paths
CONGEST
29
Lower Bounds in CONGESTLower Bounds in CONGEST Below Greedy
Maximum IS7/8-approx.
ฮฉ(๐2) 2๐ log ๐ Maximum IS, ๐ โ ๐ -approx.
vertex cover, exact
ฮฉ(๐2) 2๐ log ๐ vertex cover, ๐ + ๐ -approx.
min. dominating set,exact
ฮฉ(๐2) 2๐ log ๐ min dominating set,๐ + ๐ -approx.
hypergraph vertex cover, exact
ฮฉ(๐2) 2๐ log ๐ hypergraph vertex cover,๐ + ๐ -approx.
๐(๐ฎ)-vertex coloring ฮฉ(๐2) 2๐ log ๐ ๐ซ-vertex coloring
?-edge coloring ? poly log ๐ ๐ + ๐ ๐ซ-edge coloring
Maximum matchingalmost exact
ฮฉ( ๐ poly log ๐ Maximum Matching,๐ โ ๐ -approx.
CONGEST[BCDELP โ19], [AKO โ18], [ACK โ16]
30
Letโs Discuss โฆ
A lot was spared in this talk (randomized!)
โข (1 โ ๐) max cut approximation: [Zelke โ09]Similar to technique 1: Subsampling + solving optimally
โข spanners, e.g., [Censor-Hillel, Dory; PODC โ18]
โข randomized edge coloring below the greedy regime, e.g., [Elkin, Pettie, Su; SODA โ15], [Chang, He, Li, Pettie, Uitto; SODA โ18], [Su, Vu; STOC โ19]
โข MPC: Maximum Matching approx. in time ๐(log log ๐)[Behnezhad, Hajiaghayi, Harris; FOCS โ19]
โข lots more โฆ
What can or cannot be done below the greedy regime in distributed models with limited communication (CONGEST, CONGESTED CLIQUE, MPC, โฆ)?
Thank you