Tight Approximation Algorithms for Ordered

Covering

Jatin Batra, Syamantak Das and A. Jha

TIFR Mumbai, IIIT-Delhi and EPFL

2023
Sample frame title

This is some text in the first frame. This is some text in the first
frame. This is some text in the first frame.
Minimum Vertex Cover

Definition
Given a graph G = (V , E ), find the smallest cardinality subset
S ⊆ V such that for all edges e in E , e ∩ S ≠ ∅.

G = (V, E)

Vertex Cover
Figure: Vertex Cover in G
Alternate View (Scheduling)

1. One Machine.
2. The vertices are the jobs.

2
4
3 4 2 3 1 5 6
5
σ
M
1
6 G = (V, E)
Figure: Scheduling Vertices on a Machine

Objective Function?
Alternate View (Scheduling): Cover Times

Definition
Given a graph G = (V , E ) and a schedule 𝜎 : V → [n] of its
vertices, the cover time of an edge is the first moment a vertex in
the edge is scheduled.

Covσ (e) = 1
4 2 3 1 5 6 4
σ
6
Figure: Cover Time of an Edge

1. Vertex Cover: min 𝜎 maxe ∈E Cov 𝜎 (e).

Í
2. Min-Sum Vertex Cover: min 𝜎 e ∈E Cov 𝜎 (e) [?, ?, ?]
Can we unify?
Ordered Optimization [?]

Given a schedule 𝜎, arrange e in descending order as:

Cov 𝜎 (e1 ) ≥ Cov 𝜎 (e2 ) ≥ . . . Cov 𝜎 (em ).

m
minσ Covσ (e1 ) minσ j=1 Covσ (ej )

?
Figure: Generalized Objective Functions
Ordered Optimization [?]

Given a schedule 𝜎, arrange e in descending order as:

Cov 𝜎 (e1 ) ≥ Cov 𝜎 (e2 ) ≥ . . . Cov 𝜎 (em ).

m
minσ Covσ (e1 ) minσ Covσ (ej )
ℓ j=1

ℓ
minσ j=1 Covσ (ej )
Figure: Top-ℓ Objective Functions

Ordered Objective Functions:

Ím
min𝜎 j=1 wj Cov𝜎 (ej )
for w1 ≥ w2 ≥ . . . wm ≥ 0.
Our Contribution

Figure: One Ring to Rule Them All

Current State

Perhaps a slide to just mention that what the best approximation

ratios and hardness results are known as of now for completeness.
Our Results

Definition
Ordered Min-Sum Vertex Cover(OMSVC): Let G = (V , E ) be a
graph, and a w1 ≥ w2 ≥ . . . wm ≥ 0 be a sequence of non-negative
weights. Design a schedule 𝜎 of vertices to minimize
Í
j wj Cov 𝜎 (ej ) where ej is the ordering of edges in descending
order of cover times.

Theorem
There is a (2 + 𝜀)-approximation algorithm for OMSVC running in
time nO (1/𝜀) .
Our Results

Definition
Top-ℓ Set Cover: Let E be a set and I be a collection of subsets
Í
of E . Design a schedule 𝜎 of sets to minimize ℓj=1 Cov 𝜎 (ej ) for a
fixed ℓ ∈ [n] where the elements ej are ordered in descending order
of cover times.

Theorem
The natural greedy algorithm gives a (8 log2 (n/ℓ) + 16)-approx for
Top-ℓ Set Cover.

Theorem
It is hard to approximate the Top-ℓ Set Cover problem within a
factor better than max (Ω(1), Ω(log(n/ℓ)).
Overview of Techniques

Our algorithm works in two phases:

1. In the first phase, we reduce OMSVC to Discounted Min-Sum
Vertex Cover (Standard).
2. In the second phase, we gave an LP-rounding algorithm for
Discounted Min-Sum Vertex Cover.
Reduction

Cost Paid

Covσ (em ) ≤ ≤ Covσ (eℓ ) ≤ ≤ Covσ (e1 )

T 1 2 r k

Free of Cost
Figure: Top-ℓ Vertex Cover
Reduction

Problem? Do not know the

ordering before hand.
How to write LP?
Reduction

Shift by a Discount
Add the Discount back later λ = Covσ (eℓ )

(Covσ (em ) − λ)+ ≤ ≤ (Covσ (eℓ ) − λ)+ ≤ ≤ (Covσ (e1 ) − λ)+

ℓ
e∈E (Covσ (e) − λ)+ + λ · ℓ = j=1 Covσ (ej )

Oblivious to Ordering Fixed

Figure: Top-ℓ Vertex Cover

Reduction

Theorem
write this cleanly Let 𝜎 ∗ be the optimal schedule to an instance of
OMSVC. Then, a schedule 𝜎 satisfying:
∑︁ ∑︁
(Cov 𝜎 − 2𝜆)+ ≤ 2 (Cov 𝜎 ∗ − 𝜆)+
e ∈E e ∈E
Íℓ Íℓ
implies that j=1 Cov 𝜎 (ej ) ≤ j=1 Cov 𝜎 (ej ).
∗
DMSVC Linear Program

∑︁ ∑︁
Minimize ue,t , s.t.
t ≥𝜆 e ∈E
ue,t + xu,<t + xv ,<t ≥ 1, ∀ e = (u, v ), t = 1, 2, . . . (1)
∑︁
xv ,t ≤ 1 , ∀ t = 1, 2, . . . (2)
v ∈V
u, x ≥ 0 .
Intuition For Linear Program

Variables: ue,t , xv ,t

e for t′ + 1 ≤ t
v ue,t = 0
u
xu,t′′ = 1

v u

ue,t = 1
xv,t′ = 1
for 1 ≤ t ≤ t′
Figure: Intuition for variables and constraints
Intuition For Linear Program

Variables: ue,t , xv ,t
v e


u t≥λ ue,t

λ
Free
v
Paid

ue,t = 1
for 1 ≤ t ≤ t′
Figure: Intuition for variables and constraints
Bi-criteria algorithm for DMSVC
σ
v1 1
v2 2

2λ

vn n
V [n]
Figure: Assignment Graph
Older Approaches: FLT

Independently throw each vertex into the buckets.

v 2xv,n
2xv,1 2xv,2

T 1 2 n
Figure: [?]
Older Approaches: FLT

Break the buckets and schedule a uniformly random permutation

of vertices inside the bucket.

1 2 12 345
Figure: [?]
Problem?
Recall

Expectation

Reality
Figure: [?]