vijay v. vazirani georgia tech

80
Vijay V. Vazirani Georgia Tech A Postmortem of the Last Decade and Some Directions for the Future

Upload: arnie

Post on 12-Jan-2016

29 views

Category:

Documents


1 download

DESCRIPTION

A Postmortem of the Last Decade and Some Directions for the Future. Vijay V. Vazirani Georgia Tech. Although this may seem a paradox, all exact science is dominated by the idea of approximation. Bertrand Russell (1872-1970). Exact algorithms have been studied - PowerPoint PPT Presentation

TRANSCRIPT

Page 1: Vijay V. Vazirani Georgia Tech

Vijay V. Vazirani

Georgia Tech

A Postmortem of the Last Decade

and

Some Directions for the Future

Page 2: Vijay V. Vazirani Georgia Tech

Although this may seem a paradox,

all exact science is dominated by

the idea of approximation.

Bertrand Russell (1872-1970)

Page 3: Vijay V. Vazirani Georgia Tech

Exact algorithms have been studied

intensively for over four decades,

and yet basic insights are still being obtained.

Since polynomial time solvability is the exception

rather than the rule, it is only reasonable

to expect the theory of approximation algorithms

to grow considerably over the years.

Page 4: Vijay V. Vazirani Georgia Tech
Page 5: Vijay V. Vazirani Georgia Tech
Page 6: Vijay V. Vazirani Georgia Tech
Page 7: Vijay V. Vazirani Georgia Tech
Page 8: Vijay V. Vazirani Georgia Tech
Page 9: Vijay V. Vazirani Georgia Tech

Beyond the list …

Unique Games Conjecture

Simpler proof of PCP Theorem

Online algorithms for AdWords problem

Page 10: Vijay V. Vazirani Georgia Tech

Beyond the list …

Unique Games Conjecture

Simpler proof of PCP Theorem

Online algorithms for AdWords problem

Integrality gaps vs approximability

Page 11: Vijay V. Vazirani Georgia Tech
Page 12: Vijay V. Vazirani Georgia Tech
Page 13: Vijay V. Vazirani Georgia Tech

Raghevendra, 2008: Assuming UGC,

for every constrained satisfaction problem:

Can achieve approximation factor

= integrality gap of “standard SDP”

NP-hard to approximate better.

Page 14: Vijay V. Vazirani Georgia Tech

Future Directions

Status of UGC

Raghavendra-type results for LP-relaxations

Randomized dual growth in

primal-dual algorithms

Page 15: Vijay V. Vazirani Georgia Tech

Approximability: sharp thresholds

For a natural problem:

Can achieve approximation factor in P.

If we can achieve in P

=> complexity-theoretic disaster

α(n)

α(n) − ∈(n)

Page 16: Vijay V. Vazirani Georgia Tech

Conjecture

There is a natural problem

having sharp thresholds

w.r.t. time classes

α1(n) > α 2 (n) > ... > α k (n)

P=T1(n) ⊂ T2 (n) ⊂ ... Tk(n)

Page 17: Vijay V. Vazirani Georgia Tech

Group Steiner Tree Problem

Chekuri & Pal, 2005:

Halperin & Krauthgamer, 2003:

log2−∈n factor algorithm in time

2^ (2^ ( log nO(∈)))

time = 2^(2^( log no(∈)))⇒ subexponential algorithm for 3SAT

Page 18: Vijay V. Vazirani Georgia Tech
Page 19: Vijay V. Vazirani Georgia Tech
Page 20: Vijay V. Vazirani Georgia Tech

What lies at the core of

approximation algorithms?

Page 21: Vijay V. Vazirani Georgia Tech

What lies at the core of

approximation algorithms?

Combinatorial optimization!

Page 22: Vijay V. Vazirani Georgia Tech

Combinatorial optimization

Central problems have LP-relaxations

that always have integer optimal solutions!

ILP: Integral LP

Page 23: Vijay V. Vazirani Georgia Tech

Combinatorial optimization

Central problems have LP-relaxations

that always have integer optimal solutions!

ILP: Integral LP

i.e., it “behaves” like an IP!

Page 24: Vijay V. Vazirani Georgia Tech

Massive accident!

Page 25: Vijay V. Vazirani Georgia Tech

Cornerstone problems in P

Matching (general graph) Network flow Shortest paths Minimum spanning tree Minimum branching

Page 26: Vijay V. Vazirani Georgia Tech

Is combinatorial optimizationrelevant today?

Why design combinatorial algorithms,

especially today that LP-solvers are so fast?

Page 27: Vijay V. Vazirani Georgia Tech
Page 28: Vijay V. Vazirani Georgia Tech
Page 29: Vijay V. Vazirani Georgia Tech
Page 30: Vijay V. Vazirani Georgia Tech

Combinatorial algorithms

Very rich theory

Gave field of algorithms some of its formative

and fundamental notions, e.g. P

Preferable in applications, since efficient

and malleable.

Page 31: Vijay V. Vazirani Georgia Tech

Helped spawn off algorithmic areas,

e.g., approximation algorithms and

parallel algorithms.

Page 32: Vijay V. Vazirani Georgia Tech

Problemsadmitting ILPs

Combinatorial optimization studied:

Page 33: Vijay V. Vazirani Georgia Tech

Problems admittingLP-relaxations with

bounded integrality gaps

Approximation algorithms studied:

Page 34: Vijay V. Vazirani Georgia Tech

Problems admittingLP-relaxations with

bounded integrality gaps

Problemsadmitting ILPs

Page 35: Vijay V. Vazirani Georgia Tech

Rational convex program

A nonlinear convex program that

always has a rational solution (if feasible),

using polynomially many bits,

if all parameters are rational.

Page 36: Vijay V. Vazirani Georgia Tech

Rational convex program

Always has a rational solution (if feasible)

using polynomially many bits,

if all parameters are rational.

i.e., it “behaves” like an LP!

Page 37: Vijay V. Vazirani Georgia Tech

Rational convex program

Always has a rational solution (if feasible)

using polynomially many bits,

if all parameters are rational.

i.e., it “behaves” like an LP!

Do they exist??

Page 38: Vijay V. Vazirani Georgia Tech
Page 39: Vijay V. Vazirani Georgia Tech
Page 40: Vijay V. Vazirani Georgia Tech

KKT optimality conditions

−∇ f0 (x) = yi fii

∑ '(x) + z jj

∑ a j

yi ≥ 0 for 1≤ i ≤ m

yi > 0 ⇒ fi (x) = 0 for 1≤ i ≤ m

fi (x) ≤ 0 for 1≤ i ≤ m

a jT x ≤ b j for 1≤ j ≤ p

Page 41: Vijay V. Vazirani Georgia Tech

Possible RCPs

Pick fi 's linear, and

f0 quadratic or logarithmic.

Page 42: Vijay V. Vazirani Georgia Tech

Quadratic RCPs

fo (x) = xT P x+ qT x

convexity requires:

∇2 f0 f =0, i.e., P f = 0

Page 43: Vijay V. Vazirani Georgia Tech

Two opportunities for RCPs:

Program A: Combinatorial, polynomial time

(strongly poly.) algorithm

Program B: Polynomial time (strongly poly.)

algorithm, given LP-oracle.

Page 44: Vijay V. Vazirani Georgia Tech

Helgason, Kennington & Lall, 1980Single constraint

Minoux, 1984Minimum quadratic cost flow

Frank & Karzanov, 1992Closest point from origin to bipartite perfect

matching polytope.

Karzanov & McCormick, 1997Any totally unimodular matrix.

Combinatorial Algorithms

Page 45: Vijay V. Vazirani Georgia Tech

Ben-Tal & Nemirovski, 1999

Polyhedral approximation of second-order cone

Main technique: Solves any quadratic RCP

in polynomial time, given an LP-oracle.

Page 46: Vijay V. Vazirani Georgia Tech

Ben-Tal & Nemirovski, 1999

Polyhedral approximation of second-order cone

Main technique: Solves any quadratic RCP

in polynomial time, given an LP-oracle.

Strongly polynomial algorithm?

Page 47: Vijay V. Vazirani Georgia Tech

Logarithmic RCPs

f0 (x) = − mii∑ log(ui (x))

where mi > 0 and ui (x) is linear in x.

Page 48: Vijay V. Vazirani Georgia Tech

Logarithmic RCPs

Rationality is the exception to the rule,

and needs to be established piece-meal.

f0 (x) = − mii∑ log(ui (x))

where mi > 0 and ui (x) is linear in x.

Page 49: Vijay V. Vazirani Georgia Tech

Logarithmic RCPs

Optimal solutions to such RCPs capture

equilibria for various market models!

f0 (x) = − mii∑ log(ui (x))

where mi > 0 and ui (x) is linear in x.

Page 50: Vijay V. Vazirani Georgia Tech

Arrow-Debreu Theorem, 1954

Celebrated theorem in Mathematical Economics

Established existence of market equilibrium under very general conditions using a deep theorem from topology - Kakutani fixed point theorem.

Page 51: Vijay V. Vazirani Georgia Tech

Arrow-Debreu Theorem, 1954

Celebrated theorem in Mathematical Economics

Established existence of market equilibrium under very general conditions using a theorem from topology - Kakutani fixed point theorem.

Highly non-constructive!

Page 52: Vijay V. Vazirani Georgia Tech

Arrow-Debreu Theorem, 1954

Celebrated theorem in Mathematical Economics

Established existence of market equilibrium under very general conditions using a theorem from topology - Kakutani fixed point theorem.

Continuous, quasiconcave,

satisfying non-satiation.

Page 53: Vijay V. Vazirani Georgia Tech

Complexity-theoretic question

For “reasonable” utility fns.,

can market equilibrium be computed in P?

If not, what is its complexity?

Page 54: Vijay V. Vazirani Georgia Tech

Short summary

So far, all markets

whose equilibria can be computed efficiently

admit convex or quasiconvex programs,

many of which are RCPs!

Page 55: Vijay V. Vazirani Georgia Tech

Combinatorial Algorithm for Linear Case of Fisher’s Model

Devanur, Papadimitriou, Saberi & V., 2002

By extending primal-dual paradigm to setting of convex programs & KKT conditions

Page 56: Vijay V. Vazirani Georgia Tech

Eisenberg-Gale Program, 1959

max log

. .

:

: 1

: 0

i ii

i ij ijj

iji

ij

m v

s t

i v

j

ij

u xx

x

Page 57: Vijay V. Vazirani Georgia Tech

Eisenberg-Gale Program, 1959

max log

. .

:

: 1

: 0

i ii

i ij ijj

iji

ij

m v

s t

i v

j

ij

u xx

x

prices pj

Page 58: Vijay V. Vazirani Georgia Tech

KKT conditions

1). ∀j : pj ≥0

2). ∀j : pj > 0 ⇒ xij =1i∑

3). ∀i, j :uij

pj

≤vi

m(i)

4). ∀i, j : xij > 0 ⇒uij

pj

=vi

m(i)=

uijxijj∑m(i)

Page 59: Vijay V. Vazirani Georgia Tech

Proof of rationality

Guess positive allocation variables (say k).

Substitute 1/pj by a new variable.

LP with (k + g) equations and

non-negativity constraint for each variable.

Page 60: Vijay V. Vazirani Georgia Tech

Auction for Google’s TV ads

N. Nisan et. al, 2009:

Used market equilibrium based approach.

Combinatorial algorithms for linear case

provided “inspiration”.

Page 61: Vijay V. Vazirani Georgia Tech

Long-standing open problem

Complexity of finding an equilibrium for

Fisher and Arrow-Debreu models under

separable, piecewise-linear, concave utilities?

Page 62: Vijay V. Vazirani Georgia Tech

utility

Piecewise linear, concave

amount of j

Additively separable over goods

Page 63: Vijay V. Vazirani Georgia Tech

Long-standing open problem

Complexity of finding an equilibrium for

Fisher and Arrow-Debreu models under

separable, piecewise-linear, concave utilities?

Equilibrium is rational!

Page 64: Vijay V. Vazirani Georgia Tech

Markets with separable, plc utilitiesare PPAD-complete

Chen, Dai, Du, Teng, 2009

Chen & Teng, 2009

V. & Yannakakis, 2009

Page 65: Vijay V. Vazirani Georgia Tech

Markets with separable, plc utilitiesare PPAD-complete

Chen, Dai, Du, Teng, 2009

Chen & Teng, 2009

V. & Yannakakis, 2009

(Building on combinatorial insights from DPSV)

Page 66: Vijay V. Vazirani Georgia Tech

Theorem (V., 2002): Generalized linear Fisher market to Spending constraint utilities. Polynomial time algorithm for computing equilibrium.

Page 67: Vijay V. Vazirani Georgia Tech

Is there a convex program for this model?

“We believe the answer to this question should be ‘yes’. In our experience, non-trivial polynomial time algorithms for problems are rare and happen for a good reason – a deep mathematical structure intimately connected to the problem.”

Page 68: Vijay V. Vazirani Georgia Tech

EG convex program = Devanur’s program

Price disc. Market

Goel & V.

Spending constraint marketV., 2005

Nash BargainingV., 2008

Eisenberg-Gale MarketsJain & V., 2007

EG[2] MarketsChakrabarty, Devanur & V.

2008

Page 69: Vijay V. Vazirani Georgia Tech

V., 2010: Assuming perfect price

discrimination, can handle:

Continuously differentiable, quasiconcave

(non-separable) utilities, satisfying non-satiation.

Page 70: Vijay V. Vazirani Georgia Tech

V., 2010:

Continuously differentiable, quasiconcave

(non-separable) utilities, satisfying non-satiation.

Compare with Arrow-Debreu utilities!!

continuous, quasiconcave, satisfying non-satiation.

Page 71: Vijay V. Vazirani Georgia Tech

A new development

Orlin, 2009: Strongly polynomial algorithm

for Fisher’s linear case, using scaling.

Open: For rest

Page 72: Vijay V. Vazirani Georgia Tech

Are there other classes of RCPs?

Page 73: Vijay V. Vazirani Georgia Tech

Sturmfels & Uhler, 2009:

S f =0 n×n, sample covariance matrix

G=([n],E) chordal graph

Then the following is an RCP:min log det Σ

s.t. Σij =Sij ∀(i, j)∈E or i = j

Page 74: Vijay V. Vazirani Georgia Tech
Page 75: Vijay V. Vazirani Georgia Tech

EG convex program = Devanur’s program

Price disc. Market

Goel & V.

Spending constraint marketV., 2005

Nash BargainingV., 2008

Eisenberg-Gale MarketsJain & V., 2007

EG[2] MarketsChakrabarty, Devanur & V.

2008

Page 76: Vijay V. Vazirani Georgia Tech

Building on

Karzanov & McCormick, 1997:

Combinatorial algorithm for min cost flow

under concave cost functions on edges.

Page 77: Vijay V. Vazirani Georgia Tech

EG convex program = Devanur’s program

Price disc. Market

Goel & V.

Spending constraint marketV., 2005

Nash BargainingV., 2008

Eisenberg-Gale MarketsJain & V., 2007

EG[2] MarketsChakrabarty, Devanur & V.

2008

Page 78: Vijay V. Vazirani Georgia Tech

Rational (combinatorial)approximations

to convex programs

Problemsadmitting RCPs

Page 79: Vijay V. Vazirani Georgia Tech

EG convex program = Devanur’s program

Price disc. Market

Goel & V.

Spending constraint marketV., 2005

Nash BargainingV., 2008

Eisenberg-Gale MarketsJain & V., 2007

EG[2] MarketsChakrabarty, Devanur & V.

2008

Page 80: Vijay V. Vazirani Georgia Tech

Rational (combinatorial) approximations

to convex programs

Problemsadmitting RCPs