48
Ryan O’Donnell (CMU, IAS) joint work with Ankur Moitra (MIT)

Ryan O’Donnell (CMU, IAS) joint work with Ankur Moitra (MIT)

Embed Size (px)

Citation preview

Ryan O’Donnell (CMU, IAS)

joint work with

Ankur Moitra (MIT)

Binary optimization, linear objective

F ⊆ {0,1}n, “feasible solutions”

max

Binary optimization, linear objectives

F ⊆ {0,1}n, “feasible solutions”

“max”

Pareto optima

def: x ∈ F is Pareto optimal if not dominated on every objective by some y ∈ F.

Obj1

Obj2

xy

Pareto optima

Obj1

Obj2

xy

def: x ∈ F is Pareto optimal if not dominated on every objective by some y ∈ F.

Obj1

Obj2

thm:

If d+1 linear objectives are “semi-random”

(in the Smoothed Analysis sense)

then for any F ⊆ {0,1}n,

E[# Pareto optima in F] ≤ O(n2d).

Application: 0-1 Knapsack

[Nemhauser−Ullmann69]:

For i=1…n, compute Pareto optimal

weight,value pairs achievable by items 1…i.⟨ ⟩

[Beier−Vöcking03]:

With 2 semi-rand objs, E[# PO’s] ≤ O(n4).

⇒ O(n5)-time alg for 0-1 Knapsack in

Smoothed Analysis model!

Model [BV03]

F ⊆ {0,1}n arbitrary

d+1 linear objs

x

=

Obj(x)

ϕ ≥ 1 a parameter

Each is a r.v. on [0,1] with pdf bdd by ϕ.

All independent.

Model [BV03]

Each is a r.v. on [0,1] with pdf bdd by ϕ.

All independent.

1/ϕ

ϕ

Prior work

[BV03]: d = 1, E[# PO’s] ≤ O(ϕn4)

[B04]: Same for arbit. F. Conj: ≤ O(nf(d))

[BRV07]: d = 1, E[# PO’s] ≤ O(ϕn2)

[RT09]: E[# PO’s] ≤ roughly

[MO11]: E[# PO’s] ≤ 2(4ϕd)d(d+1)/2 n2d

Remark: All bounds here hold even assuming Obj0 adversarial, nonlinear, arbitrary.

for n ≥ exp(exp(O(d2 log d))

(F = {0,1}n)

[BV03]: Ω(n2) when Obj0 adversarial, ϕ = 1

[BR11]: Ω(ϕn2), d = 1; Ω(ϕn)d-logd roughly, d > 1

Warmup:

Alternate proof (sketch) for

[BRV07]’s d = 1 bound

F ⊆ {0,1}nObj0(x) arbitrary

Obj1(x) ϕ-semi-random E[# PO’s] ≤ O(ϕn2) ?

Obj0 01011

10100

01101

01111

01000

11011

00111

11110

Obj1

.8.4

Obj0 01011

10100

01101

01111

01000

11011

00111

11110

.2 .9

Obj0(x) arbitrary

Obj1(x) ϕ-semi-random

Obj1

F ⊆ {0,1}n

E[# PO’s] ≤ O(ϕn2) ?

Goal: For each ϵ-strip S,

E[# PO’s in S] ≈ Pr[∃ a PO in S] ≤ O(n) ϕ ϵ

Obj0 01011

10100

01101

01111

01000

11011

00111

11110

ϵ

ϵ2

Obj1

Union-bound over (n/ϵ) strips ⇒ E[# PO’s] ≤ O(ϕn2)

S

Obj0 01011

10100

01111

01000

11011

00111

11110

ϵ

Obj1

Given x, Pr[ Vx ∈ S ] ?

S

01101

Obj0 01011

10100

01111

01000

11011

00111

11110

ϵ

Obj1

Given x, Pr[ Vx ∈ S ] ?

S

Take any j s.t. xj = 1.

01101

Obj0 01011

10100

01101

01111

01000

11011

00111

11110

ϵ

Obj1

Given x, Pr[ Vx ∈ S ] ?

S

Take any j s.t. xj = 1.

.8.4.2 .9

= Pr[ Vj ∈ certain width-ϵ interval] ≤ ϕϵ.

Obj0 01011

10100

01101

01111

01000

11011

00111

11110

ϵ

Given x, Pr[ Vx ∈ S ] ?

S

Take any j s.t. xj = 1.

= Pr[ Vj ∈ certain width-ϵ interval] ≤ ϕϵ.

Boundedness Lemma: Pr[ Vx ∈ S] ≤ O(ϵ),

even just using the randomness of Vj.

Obj0 01011

10100

01101

01111

01000

11011

00111

11110

ϵ

Obj1

S

event Tx,S = “x is PO and Vx ∈ S”

.8.4.2 .9

Obj0 01011

10100

01101

01111

01000

11011

00111

11110

ϵ

Obj1

S

event Tx,S = “x is PO and Vx ∈ S”

.8.4.2 .9

Fantasy: Now a unique x s.t. Tx,S may yet occur

more complicated event:

Tx,j,S = “x is PO and xj = 1 and Vx ∈ S

and first PO to x’s left, call it y, has yj = 0”

Uniqueness Lemma:

Draw all Vi’s except Vj.

Then ∃ at most 1 x s.t. Tx,j,S may still occur.

Tx,j,S = “x is PO and xj = 1 and Vx ∈ S

and first PO to x’s left, call it y, has yj = 0”

Obj0 01011

10100

01101

01111

01000

11011

00111

11110

ϵ

Obj1

S

.8.4.2 .9

Remainder of the d=1 proof sketch

Tx,j,S = “x is PO and xj = 1 and Vx ∈ S and …”

Uniqueness Lemma: Draw all Vi’s except Vj.

Then ∃ at most 1 x s.t. Tx,j,S may still occur.Bddness Lemma: For that x, Pr[Vx ∈ S] ≤ ϕϵ.

Union-bound over all j, S:

E[ # PO x s.t. first PO to x’s left, y, has yj ≠ 1 = xj for some j ] ≤ n(n/ϵ)ϕϵ = n2ϕ.

For each PO x, ∃ j s.t. yj ≠ c = xj. Maybe c = 0…

Union-bound over c ∈ {0,1}… + another trick.

Our result: Larger d

Obj0 arbitrary; Obj1, Obj2 ϕ-semi-random

01011

10100

01101

01111

01000

11011

00111

11110

d=2:

Obj0

Obj1

Obj2

Obj0 01011

10100

01101

01111

01000

11011

00111

11110

Obj1

Obj2

V =

Obj0 01011

10100

01101

01111

01000

11011

00111

11110

Obj2

ϵS

ϵ

Obj1

V =

more complicated event:

Tx,j,S = “x is PO and xj = 1 and Vx ∈ S

and first PO to x’s left, call it y, has yj = 0”

more complicated event:

Tx,j,c,S = “x is PO and xj = c and Vx − ∈ S

and first PO to x’s left, call it y, has yj ≠ c”

more complicated event:

Tx,J,C,S = “x is PO and…”

still

J: list of d coordinates

C: d×d matrix of bits w/ certain pattern

S: d-dimensional strip, plus “nearby” d′-dim. strips for all d′ < d

more complicated event:

J: list of d coordinates

C: d×d matrix of bits w/ certain pattern

S: d-dimensional strip, plus “nearby” d’-dim. strips for all d’ < d

B: “blueprint” [AMR09]

+

=

Tx,B = “x is PO and conditions involving V, x, B.”

still

Given B, entries of V are partitioned into

two sets, V[fewB] and V[mostB].

V =

Tx,B = “x is PO and conditions involving V, x, B.”

Boundedness Lemma: For any V[mostB] outcome,

Pr [condits involving V, x, B] ≤ ϕd(d+1)/2 ϵd(d+1)/2.

Uniqueness Lemma: Draw V[mostB].

Then ∃ at most 1 x s.t. Tx,B may still

occur.

Tx,B = “x is PO and conditions involving V, x, B.”

V[fewB]

Counting Lemma: O(d/ϵ)d(d+1)/2 n2d possible B.

Union bound ⇒ E[# PO’s] ≤ O(dϕ)d(d+1)/2 n2d.

Given B, entries of V are partitioned into

two sets, V[fewB] and V[mostB].

Tx,j,c,S = “x is PO and xj = c and Vx − ∈ S

and first PO to x’s left, call it y, has yj ≠ c”

J: list of d coordinates

C: d×d matrix of bits w/ certain pattern

S: d-dimensional strip, plus “nearby” d′-dim. strips for all d′ < d

B: “blueprint”

+

=

Tx,B = “x is PO and conditions involving V, x, B.”

Tx,B = “x is PO and SKETCH(x,V) = B”

where SKETCH() is a certain

deterministic algorithm.

SKETCH(x,V):

Bddness Lemma: For any V[mostB] outcome,

Pr [SKETCH(x,V) = B] ≤ ϕd(d+1)/2 ϵd(d+1)/2.

Uniqueness Lemma: Draw V[mostB].

Then ∃ at most 1 x s.t. Tx,B may still occur.

V[fewB]

Tx,B = “x is PO and SKETCH(x,V) = B”

Uniqueness Lemma: Draw V[mostB].

Then ∃ at most 1 x s.t. Tx,B may still occur.

Tx,B = “x is PO and SKETCH(x,V) = B”

Equivalent Lemma:

Suppose x, V are such that x is PO.

Assume SKETCH(x,V) = B.

Then RECONSTRUCT(B,V[mostB]) = x.

RECONSTRUCT(B,V[most]):

Bddness Lemma: For any V[mostB] outcome,

Pr [SKETCH(x,V) = B] ≤ ϕd(d+1)/2 ϵd(d+1)/2.V[fewB]

Uniqueness Lemma:

Suppose x, V are such that x is PO.

Assume SKETCH(x,V) = B.

Then RECONSTRUCT(B,V[mostB]) = x.

Counting Lemma: O(d/ϵ)d(d+1)/2 n2d possible B.

A sketch of SKETCH(x,V) for d = 2

SKETCH(x,V):

1. Let F0 = F

2. Let D1 = { z ∈ F0 : V1z > V1x and V2z > V2x }

3. Let y1 = argmax{ Obj0(y) : y ∈ D1

}

4. Let j1 = least index s.t.

5. Let F1 = { z ∈ F0 : Obj0(z) > Obj0(y1) and }

6. Let D2 = { z ∈ F1 : V2z > V2x }

7. Let y2 = argmax{ Obj1(y) : y ∈ D2 }

8. Let j2 = least index s.t.

9. Let F2 = { z ∈ F1 : Obj1(z) > Obj1(y1) and }

// if Vx is PO then it must be argmax{ Obj2(z) : z ∈ F2 }

10. Output: J = (j1, j2), C =

S = ( strip of Obj1,2(x) − diag(Vj1,j2 C), strip of Obj2(x) − )

SKETCH(x,V):

• Close gap of nd vs. n2d.

• Lower bounds when all objs semirandom?

• F ⊆ {0,1,2,…,k}n ?

• Bounds on Var[ # POs ] ?

• More Smoothed Analysis via “blueprints”?

Future directions?