dana moshkovitz mit joint work with subhash khot, nyu 1

Post on 14-Dec-2015

222 Views

Category:

Documents

4 Downloads

Preview:

Click to see full reader

TRANSCRIPT

Hardness of Approximately Solving Linear Equations Over Reals

Dana MoshkovitzMIT

Joint work with Subhash Khot, NYU1

Best Approximations Known For 2CSP vs. 3CSP

2 3OR 0.931.. [FG95] 7/8

XOR 0.878... [GW92] 1/2AND 0.859.. [FG95] 1/2

any better

approximation:

NP-hard [Håstad97]

2

MAX-CUT: If exists cut with (1-±) fraction of edges, efficiently cut at least (1-c±)

fraction of edges.

1

2(n)

nO(1)

exponential

polynomial

running-time

Exponential hardness: time 2n1-o(1)

Sharp threshold:Drop at ½+o(1)

Assuming it takes 2(n) time to solve 3SAT exactly on size n inputs.

1/2

The Complexity of Approximating 3XOR [=3LIN(GF(2))][AroraSafra92, AroraLundMotwaniSudanSzegedy92, Raz94, BellareGoldreichSudan95, Håstad97,

MRaz08]

Approx. factor

3

dmoshkov

2 3OR 0.931.. [FG95] 7/8

XOR 0.878... [GW92] 1/2AND 0.859.. [FG95] 1/2

Best Approximations Known For 2CSP vs. 3CSP

any better

approximation:

NP-hrd [Håstad97]

Any better approximation: NP-hard assuming the

Unique Games Conjecture [K02,KKMO04,R08]

4

Hard time showing easy problems are hard!

Proving Hardness of Constraint Satisfaction Problems

5

The Bellare-Goldreich-Sudan-Håstad paradigm

2CSP(GF(q))desired problem

composition with long-code/dictator code

[Khot02]: Start with 2LIN(GF(q)).

The Unique Games Conjecture: 2LIN(GF(q)) is extremely hard to approximate.

6

The Unique Games Conjecture (UGC) - The Hardness of 2LIN Over Large Alphabets

Unique Games Conjecture [Khot02, formulation by KKMO04]:For all ,δ>0, for sufficiently large q, given a set of linear equations of the form

xi – xj = c (mod q)

where 1-δ fraction of the equations can be satisfied, it is NP-hard to find an assignment that satisfies fraction of the equations.

[Raghavendra08]: Assuming the Unique Games Conjecture, the best approximation for all CSPs is obtained by rounding a natural SDP.

Alphabet/Hardness Rules of Thumb: As alphabet gets larger, problem gets harder.

Important Exception: problem can (but not necessarily) become easy if alphabet is R: linear/semidefinite programming can be used.

Typically,Easy, if allow fractional solutions Hard, if encode large discrete alphabets (e.g.,

exact 3LIN(R) [GR07])

7

• A natural optimization problem: find a balanced assignment for real homogeneous linear equations:

• SDP-based algorithm: Best approximation known obtained by rounding natural SDP (when 1- equations

exactly satisfiable, get margin O() for (1) of equations),– Unlike GF(q): Same for 2 or 3 variables per equation!

• We show NP-hardness for 3 vars per equation!• Approach to proving the Unique Games Conj.

8

x15 - x231 + x37 = 0

x1 - 2x3 + x89 = 0. . .

Work with Subhash Khot, 2010:

Margin of equation ax+by+cz = 0 is |ax+by+cz|

Main Theorem

Theorem [KM10]: For any ±>0, it is NP-hard, given a system of linear equations over reals where (1-±) fraction can be exactly satisfied by a balanced assignment,

to find an assignment where 0.99 fraction of the equations have margin at most 0.0001¢±.

• Tight: Efficient algorithm gives margin O() for (1) of equations.

• Blow up: Reduction from SAT has blow-up nnpoly(1/), matching the recent result of Arora, Barak, Steurer.

9

Approach to Proving The Unique Games Conjecture

1. Show that approximate 3LIN over the reals is NP-hard.

2. Show that approximate 2LIN over the reals is NP-hard, possibly by reduction from 3LIN.

3. Deduce the Unique Games Conjecture using parallel repetition.

10

Approximation Algorithm

Min E[| ax + by + cz |2] s.t. assignment is balanced

Analysis & Rounding. Assume (1-) equations exactly satisfiable.Then the SDP minimum is at most O().

Solve the SDP to get vector assignment. Pick random Gaussian ³, get real assignment xi=<vi,³> with similar value. The margin is O() for constant fraction of equations.

SDP

11

Hardness of Approximately Solving Real Linear Equations

with Three Variables Per Equation

12

Proving Hardness of Approximate Real Equations

1.Dictatorship test over reals.2.Adapting the paradigm

13

The Bellare-Goldreich-Sudan-Håstad paradigm

2CSP(GF(q))desired problem

composition with long-code/dictator code

dictatorship test for desired problem

Real Functions

•Variables correspond to points in Rn. •Assignments correspond to functions RnR. • We consider the Gaussian distribution over Rn,

where each coordinate has mean 0 and variance 1.

Fact: If x,y are independent Gaussians, then px+qy is Gaussian where p2+q2 = 1.

14

F

Dictator Testing

Tester

F is a dictator, i.e., F(x1,…,xn)=xi

No linear F’ junta (poly(1/) coord) with

|F-F’|2·²|F|2

• Function F:RnR. Two possibilities for F.• A tester queries three positions x,y,zRn, tests

aF(x)+bF(y)+cF(z) = 0 ?

• If dictator, equation satisfied exactly with probability 1-.

• If not approximated by linear junta, there is a margin 0.001 with probability 0.01.

15

Noise Sensitivity Approach to Dictator Testing

• Pick Gaussian x2Rn.• Perturb x by re-sampling each coordinate with

probability ±. Obtain x’.• Check F(x)=F(x’).

16

Problem with Noise Sensitivity Approach

• For F half-space,– With probability 1-c±,

equality holds.– With probability c±,

constant margin.• We want: with probability

¸ 0.01, margin ¸ 0.001±.

17

Linear Non-Juntas Are Noise-SensitiveObservation: If F = i2I aixi for |I|À 1/±2, then

(F(x) – F(x’)) is Gaussian with variance ¸ ±.Hence |F(x)-F(x’)| is ¸ 0.001± with prob¸0.01.

Basic Idea: test linearity and noise sensitivity.18

Hermite Analysis – Fourier Anlysis for Real Functions

Fact: For any F:RnR with bounded |f|2, can write

F = ci1,…,in¢Hi1

(x1)¢ …¢ Hin(xn),

where Hd is a polynomial (“Hermite polynomial”) of degree d.

Notation: F·1 is the linear part of F, and F>1(x) is its non-linear part.

19

Linearity Testing

• Pick Gaussian x,y2Rn, • Pick p2[0,1]; set q=(1-p2).• Check F(px+qy)=pF(x)+qF(y).

Lemma: |F-F·1|22 · E[Margin2].

Proof: Via Hermite analysis.

20

Equation depends on three variables!

Decompose into linear and non-linear part:

Ex|F(x)-F(x’)|2 = Ex|F·1(x)-F·1(x’)|2 + Ex|F>1(x)-F>1(x’)|2

Non junta ) with prob 0.1 margin¸ 0.1

Linearity test ) |F>1(x)|2 ·0.001

Cancellation problem: On average |F>1(x)-F>1(x’)|·0.001, but it can be much larger when |F·1(x)-F·1(x’)| is large and cancel |F·1(x)-F·1(x’)| !

21

Cancelations Don’t Arise!• coordinate-wise perturbation x»cx’: x’ is obtained

from x by re-sampling with prob ±.• random perturbation x»rx’: x’ is obtained from x as

px+qy for p=(1-±), p2+q2=1 • Cancelation:

Px »c x’[|F>1(x)-F>1(x’)|2 ¸ 0.1 ] ¸ 0.1

• By linearity testing:Px »r x’[|F>1(x)-F>1(x’)|2 ·0.01 ] ¸ 0.99

• We’ll show that they cannot co-exist!22

Cancelations Don’t Arise!• coordinate-wise perturbation x»cx’: x’ is obtained

from x by re-sampling with prob ±.• random perturbation x»rx’: x’ is obtained from x as

px+qy for p=(1-±), p2+q2=1 • Cancelation:

Px »c x’[|F>1(x)-F>1(x’)|2 ¸ 0.1 ] ¸ 0.1

• By linearity testing:Px »r x’[|F>1(x)-F>1(x’)|2 ·0.01 ] ¸ 0.99

• We’ll show that they cannot co-exist!23

Claim: For any G:RnR,Ex »r x’ |G(x)-G(x’)|2 ¸ Ex »c x’|G(x)-G(x’)|2 .

Main Lemma: From Small Difference to Constant Difference

Main Lemma: If for H:RnR,Px »c x’ [|H(x)-H(x’)|2 ¸ A¢T] ¸ 10®,

Px »r x’ [|H(x)-H(x’)|2 · T] ¸ 1-®,

Then, there is Boolean B:Rn {0,1} where Px »c x’ [|B(x)-B(x’)|2 = 1] ¸ 0.01,

Px »r x’ [|B(x)-B(x’)|2 = 0] ¸ 1-1/A,

24

Proof: Combinatorial, by considering “perturbation graphs”, and constructing an appropriate cut.

Thank you!

25

Proof of Main LemmaPerturbation graphs:

• Gc = (Rn,Ec)

• (x,x’)2Ec if:

– |H(x)|,|H(x’)|·1/2,– |H(x)-H(x’)|¸ 0.1.• Edge weight is prob x »c x’.

• Gr = (Rn,Er)

• (x,x’)2Er if:

– |H(x)|,|H(x’)|·1/2,– |H(x)-H(x’)| · 0.01.• Edge weight is prob x »r x’.

Our task: Find a cut in CµRn that cuts:• Ec edges of weight ¸ 0.09.

• Er edges of weight · 0.02.26

Cutting The Gaussian Space

Claim: There is a distribution D over cuts s.t.:• Every edge e2Ec is cut with prob ¸ 0.1.

• Every edge e2Er is cut with prob · 0.01.

-1/2 1/2H(x) H(x’)

27

From Small Cuts To Large Cuts

• In previous lemma, cut only ¼ weight. • Want to cut constant weight.

28

Larger Cut

• Let m=11/. Sample from D cuts C1,…,Cm.

• Pick random Iµ[m]. Let C = i2I Ci.

Claim: Edge e2Ec is cut with prob ¸ 0.5¢(1-1/e2).

• Edge e2Er is cut with prob 0.01 ¢10/ = 0.1.

Corr: A cut with 0.2¢0.1 weight of Ec and 0.8¢0.99 weight of Er.

29

top related