From: babagamppis on
This is the html version of the file http://www.cs.princeton.edu/~arora/talks/pcptalk.ppt.

Executed and transcribed by M. M. Musatov and http://meami.org/

How NP got a new definition:


Probabilistically Checkable Proofs (PCPs) & Approximation Properties
of NP-hard problems


SANJEEV ARORA

PRINCETON UNIVERSITY






Recap of NP-completeness and its philosophical
importance.

Definition of approximation.

How to prove approximation is NP-complete
(new definition of NP; PCP Theorem)

Survey of approximation algorithms.


Talk Overview






A central theme in modern TCS: Computational Complexity


How much time (i.e., # of basic operations) are needed
to solve an instance of the problem?


Example: Traveling Salesperson
Problem on n cities


Number of all possible salesman tours = n!

(> # of atoms in the universe for n =49)


One key distinction: Polynomial time (n3, n7 etc.)

versus
Exponential time (2n, n!, etc.)


n =49






Is there an inherent difference between


being creative / brilliant


and


being able to appreciate creativity / brilliance?


Writing the Moonlight Sonata
Proving Fermat’s Last Theorem
Coming up with a low-cost
salesman tour


Appreciating/verifying
any of the above


When formulated as “computational effort”, just the P vs NP

Question.






P vs. NP


NP


P


NPC


“YES” answer has certificate of O(nc) size, verifiable in O(nc’)
time.


Solvable in O(nc) time.


NP-complete: Every NP problem is reducible to it in O(nc) time.
(“Hardest”)


e.g., 3SAT: Decide satisfiability of a boolean formula like






Pragmatic Researcher


Practical Importance of P vs NP: 1000s of optimization
problems are NP-complete/NP-hard. (Traveling Salesman,
CLIQUE, COLORING, Scheduling, etc.)


“Why the fuss? I am perfectly content with approximately
optimal solutions.” (e.g., cost within 10% of optimum)


Bad News: NP-hard for many problems.


Good news: Possible for quite a few problems.






Approximation Algorithms


MAX-3SAT: Given 3-CNF formula , find assignment
maximizing the number of satisfied clauses.


An -approximation algorithm is one that for every
formula, produces in polynomial time an assignment that
satisfies at least OPT/ clauses. ( ø 1).


Good News: [KZ’97] An 8/7-approximation algorithm exists.


Bad News: [Hastad’97] If P  NP then for every  > 0, an
(8/7 -)-approximation algorithm does not exist.






Observation (1960’s thru ~1990)


NP-hard problems differ with respect to approximability


[Johnson’74]: Provide explanation? Classification?


Last 15 years: Avalanche of Good and Bad news






Next few slides: How to rule out existence
of good approximation algorithms
(New definition of NP via PCP Theorems
and why it was needed)






Recall: “Reduction”


“If you give me a place to
stand, I will move the earth.”
– Archimedes (~ 250BC)







“If you give me a polynomial-time algorithm
for 3SAT, I will give you a polynomial-time
algorithm for every NP problem.”
--- Cook, Levin (1971)


“Every instance of an NP problem can be disguised as
an instance of 3SAT.”


a 1.01-approximation for MAX-3SAT


[A., Safra] [A., Lund, Motwani, Sudan, Szegedy] 1992


MAX-3SAT






Desired


Way to disguise instances of any NP problem as
instances of MAX-3SAT s.t.


“Yes” instances turn into satisfiable formulae

“No” instances turn into formulae in which < 0.99
fraction of clauses can be simultaneously satisfied


“Gap”






Cook-Levin reduction doesn’t produce instances
where approximation is hard.


Transcript of computation


?


Transcript is “correct” if it satisfies all “local” constraints.


Main point: Express
these as boolean formula


But, there always exists a transcript that satisfies almost all
local constraints! (No “Gap”)






New definition of NP….






Recall: Usual definition of NP


INPUT x


CERTIFICATE 


n


nc


M


x is a “YES” input

 there is a  s.t. M accepts
(x, )

x is a “NO” input

 M rejects (x, ) for every 






NP = PCP (log n, 1)
[AS’92][ALMSS’92]; inspired by [BFL’90], [BFLS’91][FGLSS’91]


x is a “YES” input

 there is a  s.t. M accepts (x, )

x is a “NO” input

for every , M rejects (x, )


INPUT x


CERTIFICATE 


n


nc


M


Reads Fixed number of bits
(chosen in randomized fashion)


Pr [ ] = 1


Pr [ ] > 1/2


Uses O(log n)
random bits


(Only 3 bits ! (Hastad 97))


Many other
“PCP Theorems”
known now.






Disguising an NP problem as MAX-3SAT


INPUT x


?


M


O(lg n) random bits


Note: 2O(lg n) = nO(1).

) M ≡ nO(1) constraints, each on O(1) bits

x is YES instance ) All are satisfiable

x is NO instance ) · ½ fraction satisfiable


“gap”






Of related interest….


Do you need to read a math proof completely to

check it?


Recall: Math can be axiomatized (e.g., Peano Arithmetic)


Proof = Formal sequence of derivations from axioms






Verification of math proofs


Theorem


Proof


M


M runs in poly(n) time


n bits


(spot-checking)


O(1) bits


PCP Theorem


Theorem correct  there is a proof that M accepts w. prob. 1
Theorem incorrect  M rejects every claimed proof w. prob 1/2







HITTING SET DOMINATING SET HYPERGRAPH - TRAVERSAL ...


[PY ’88]; OTHERS


[LY ’93]


[LY ’93, ABSS ’93]


Known Inapproximability Results
The tree of reductions [AL ‘96]


MAX-3SAT


MAX-3SAT(3)


CLIQUE


LABEL COVER


SET COVER


COLORING


[PY ’88]


[LY ’93]


[FGLSS ’91, BS ‘89]




Metric TSP Vertex Cover MAX-CUT STEINER...




NEAREST VECTOR MIN-UNSATISFY QUADRATIC -PROGRAMMING LONGEST
PATH

....




INDEPENDENT SET BICLIQUE COVER FRACTIONAL COLORING MAX-PLANAR
SUBGRAPH MAX-SET PACKING MAX-SATISFY


Class II O(lg n)


Class I 1+


Class III 2(lg n)1-


Class IV n






Proof of PCP Theorems
( Some ideas )






Need for “robust” representation


O(lg n) random bits


3 bits


1 0 0 1 0 1 1 1 0 1 0 0 0 1 0 1


 :


Randomly corrupt 1% of 


x


x


x


Correct proof still accepted with 0.97- probability!


Original proof of PCP Thm used polynomial representations,

Local “testing” algorithms for polynomials, etc. (~30-40 pages)






New Proof (Dinur’06); ~15-20 pages


Repeated applications of two operations on the clauses:


Globalize: Create new constraints using “walks” in the
adjacency graph of the old constraints.


Domain reduction: Change constraints so variables take values
in a smaller domain (e.g., 0,1)
(uses ideas from old proof)






Unique game conjecture and why it is useful


Problem: Given system of equations modulo p (p is prime).


7x2 + 2x4 = 6

5x1 + 3x5 = 2

 

7x5 + x2 = 21



2 variables per equation


UGC (Khot03): Computationally intractable to distinguish between the
cases:

0.99 fraction of equations are simultaneously satisfiable
no more than 0.001 fraction of equations are simultaneously
satisfiable.


Implies hardness of approximating vertex cover, max-cut, etc.

(K04), (KR05)(KKMO05)






Applications of PCP Techniques: Tour d’Horizon


Locally checkable / decodable codes.

List decoding of error-correcting codes.

Private Info Retrieval

Zero Knowledge arguments / CS proofs

Amplification of hardness / derandomization

Constructions of Extractors.

Property testing


[Sudan ’96, Guruswami-Sudan]


[Katz, Trevisan 2000]


[Kilian ‘94] [Micali]


[Lipton ‘88] [A., Sudan ’97] [Sudan, Trevisan, Vadhan]


[Safra, Ta-shma, Zuckermann]

[Shaltiel, Umans]


[Goldreich, Goldwasser, Ron ‘97]






Approximation algorithms: Some major ideas


Relax, solve, and round : Represent problem using a
linear or semidefinite program, solve to get fractional
solution, and round to get an integer solution. (e.g.,
MAX-CUT, MAX-3SAT, SPARSEST CUT)


Primal-dual: “Grow” a solution edge by edge; prove its
near optimality using LP duality. (Usually gives faster
algorithms.) e.g., NETWORK DESIGN, SET COVER


How can you prove that the solution you found has
cost at most 1.5 times (say) the optimum cost?


Show existence of “easy to find” near-optimal solutions:
e.g., Euclidean TSP and Steiner Tree





What is semidefinite programming?


Ans. Generalization of linear programming; variables are
vectors instead of fractions. “Nonlinear optimization.”


[Groetschel, Lovasz, Schrijver ’81]; first used in approximation
algorithms
by [Goemans-Williamson’94]


Next few slides: The semidefinite programming approach






G = (V,E)


n vertices


v1


v2


v3


vn


Rn


n vectors, d(vi,vj) satisfy some constraints.


Ex: 1.13 ratio for MAX-CUT, MAX-2SAT [GW ’93]

O(lg n) ratio for min-multicut, sparsest cut. [LLR ’94, AR ’94]

n1/4-coloring of 3-colorable graphs. [KMS ’94]

(lg n)O(1) ratio for min-bandwidth and related problems [F ’98, BKRV
’98]

8/7 ratio for MAX-3SAT [KZ ’97]

plog n-approximation for graph partitioning problems (ARV04)


Main Idea:


“Round”


How do you analyze these vector
programs?


Ans. Geometric arguments; sometimes very complicated






Ratio 1.13.. for MAX-CUT


[GW ’93]


G = (V,E) Find that maximizes capacity .


Quadratic Programming Formulation


Semidefinite Relaxation [DP ’91, GW ’93]






Randomized Rounding


[GW ’93]


v6


v2


v3


v5


Rn


v1


Form a cut by partitioning v1,v2,...,vn around a random hyperplane.


SDPOPT


vj


vi


ij


Old math rides to the rescue...






sparsest cut: edge expansion


Input: A graph G=(V,E).


S


E(S, S)


For a cut (S,S) let E(S,S) denote the edges

crossing the cut.


The sparsity of S is the value


The SPARSEST CUT problem is to find the cut which minimizes (S).


SDPs used to give plog n -approximation involves proving a
nontrivial fact about high-dimensional geometry [ARV04]






ARV structure theorem


Arora, Rao, and Vazirani showed how the SDP could be rounded

to obtain an approximation to Sparsest Cut (2004)


ARV structure theorem:

If the points xu 2 Rn are well-spread, e.g.

u,v (xu-xv)2 ø 0.1 and xu2 · 10 for u 2 V

and d(u,v) = (xu-xv)2 is a metric, then…


A


B


There exist two large, well-separated sets

A, B µ {x1, x2, …, xn}

with |A|,|B| ø 0.1 n and


After we have such A and B, it is easy to

extend them to a good sparse cut in G.






Unexpected progress in
other disciplines…


ARV structure theorem led to new understanding of
the interrelationship between l1 and l2 norms

(resolved open question in math)


l1 distances among n points can be realized as
l2 distances among some other set of n points, and
the distortion incurred is only plog n


[A., Lee, Naor’05], building upon [Chawla Gupta Raecke’05]






Theory of Approximability?


Desired Ingredients:

Definition of approximation-preserving reduction.
Use reductions and algorithms to show:


Approx. upto ()


factor ()


factor ()


.. . . . . . .


All interesting problems


Partial Progress

Max-SNP: Problems similar to MAX-3SAT. [PY ’88]

RMAX(2): Problems similar to CLIQUE. [PR ‘90]

F+2(1): Problems similar to SET COVER. [KT ’91]]

MAX-ONES CSP, MIN-CSP,etc. (KST97, KSM96)






Further Directions


Investigate alternatives to approximation
Average case analysis
Slightly subexponential algorithms (e.g. 2o(n) algorithm for
CLIQUE??)
Resolve the approximability of graph partitioning problems.
(BISECTION, SPARSEST CUT,
plog n vs loglog n??) and Graph Coloring


3. Complete the classification of problems w.r.t. approximability.

4. Simplify proofs of PCP Thms even further.

5. Resolve “unique games”conjecture.

6. Fast solutions to SDPs? Limitations of SDPs?






Attributions


Definition of PCP


Polynomial Encoding Method


Verifier Composition


PCP Hardness of Approx.


Fourier Transform Technique


[Fortnow, Rompel, Sipser ’88]

[Feige, Goldwasser, Lovįsz, Safra, Szegedy ’91]

[Arora, Safra ’92]


[Lund, Fortnow, Karloff, Nisan ’90]

[Shamir ’90]

[Babai, Fortnow ’90]

[Babai, Fortnow, Levin, Szegedy ’91]


[Arora, Safra ’92]


[FGLSS ’91]

[ALMSS ’92]


[Håstad ’96, ’97]






Constraint Satisfaction Problems


Let F = a finite family of boolean constraints.

An instance of CSP(F):


x1


x2


xn


g1


g2


gm


.. . . . . . . . . . . .


.. . . . . . . . . . . .


variables


functions from F


[Schaefer ’78]


Ex:


Dichotomy Thm:


P


NP Complete


{CSP(F) : F is finite}


Iff F is 0-valid, 1-valid, weakly positive or negative, affine, or
2CNF






MAX-CSP


MAX-ONES-CSP


MIN-ONES-CSP


P


MAX-SNP-hard (1+) ratio is NP-hard


Iff F is 0-valid, 1-valid, or 2-monotone


[Creignou ‘96]

[Khanna, Sudan, Williamson ‘97]


(Supercedes MAXSNP work)


Ex:


P


1+


n


Feasibilty NP-hard


Feasibility is undecidable


Ex:


[KSW ‘97]


[KST ‘97]


P


1+


n


Feasibilty NP-hard


NEAREST-CODEWORD-complete


MIN-HORN-DELETION-complete






Geometric Embeddings of Graphs


G = (V,E)


n vertices


v1


v2


v3


vn


Rn


n vectors, d(vi,vj) satisfy some constraints.


Ex: 1.13 ratio for MAX-CUT, MAX-2SAT [GW ’93]

O(lg n) ratio for min-multicut, sparsest cut. [LLR ’94, AR ’94]

n1/4-coloring of 3-colorable graphs. [KMS ’94]

(lg n)O(1) ratio for min-bandwidth and related problems [F ’98, BKRV
’98]

8/7 ratio for MAX-3SAT [KZ ’97]

plog n-approximation for graph partitioning problems (ARV04)






Example: Low Degree Test


F =GF(q)


f : F m ! F


Is f a

degree d polynomial ?


Easy: f is a degree d polynomial iff its restriction on every line is
a univariate degree d polynomial.

[Line ≡ 1 dimensional affine subspace] ≡ q points.


Does f agree with a

degree d polynomial

in 90% of the points?


Theorem: Iff on ~ 90% of lines, f has agreement ~90% with a univariate
degree d polynomial.

Weaker results: [Babai, Fortnow, Lund ‘90] [Rubinfeld Sudan
‘92] [Feige, Goldwasser, Lovįsz, Szegedy ‘91]

Stronger results: [A. Sudan ‘96]; [Raz, Safra ‘96]






Example: Low Degree Test


F =GF(q)


f : F m ! F


Is f a

degree d polynomial ?


Does f agree with a

degree d polynomial

in 90% of the points?


Theorem: Iff on ~ 90% of lines, f has agreement ~90% with a univariate
degree d polynomial.

Weaker results: [Babai, Fortnow, Lund ‘90] [Rubinfeld Sudan
‘92] [Feige, Goldwasser, Lovįsz, Szegedy ‘91]

Stronger results: [A. Sudan ‘96]; [Raz, Safra ‘96]






The results described in this paper indicate a possible classification
of optimization problems as to the behavior of their approximation
algorithms. Such a classification must remain tentative, at least
until the existence of polynomial-time algorithms for finding optimal
solutions has been proved or disproved. Are there indeed O(log n)
coloring algorithms? Are there any clique finding algorithms better
than O(ne) for all e>0? Where do other optimization problems fit into
the scheme of things? What is it that makes algorithms for different
problems behave the same way? Is there some stronger kind of
reducibility than the simple polynomial reducibility that will explain
these results, or are they due to some structural similarity between
te problems as we define them? And what other types of behavior and
ways of analyzing and measuring it are possible?


David Johnson, 1974






MAX-LIN(3): Given a linear system over GF(2) of the form


NP-hard Optimization Problems


MAX-3SAT: Given 3-CNF formula , find assignment maximizing the number
of satisfied clauses.


find its largest feasible subsystem.






Defn: An -approximation for MAX-LIN(3) is a polynomial-time algorithm
that computes, for each system, a feasible

subsystem of size ø . ( ø 1)


Approximation Algorithms


Easy Fact: 2-approximation exists.


Theorem : If P  NP, (2-)-approximation does not
exists.






Common Approx. Ratios






Early History


Graham’s algorithm for multiprocessor scheduling [approx. ratio =
2]
1971,72 NP-completeness

Sahni and Gonzalez: Approximating TSP is NP-hard
1975 FPTAS for Knapsack [IK]

1976 Christofides heuristic for metric TSP

1977 Karp’s probabilistic analysis of Euclidean TSP

1980 PTAS for Bin Packing [FL; KK]

1980-82 PTAS’s for planar graph problems [LT, B]







Subsequent Developments


1988 MAX-SNP: MAX-3SAT is complete problem [PY]

1990 IP=PSPACE, MIP=NEXPTIME

1991 First results on PCPs [BFLS, FGLSS]

1992 NP=PCP(log n,1) [AS,ALMSS]

1992-95 Better algorithms for scheduling, MAX- CUT [GW],
MAX-3SAT,...

1995-98 Tight Lowerbounds (H97); (1+ )- approximation for
Euclidean TSP, Steiner Tree...

1999-now Many new algorithms and hardness results.

2005 New simpler proof of NP=PCP(log n,1) (Dinur)






3SAT: Given a 3-CNF formula, like


decide if it has a satisfying assignment.



THEOREMS: Given decide

if T has a proof of length · n in Axiomatic Mathematics



Philosophical meaning of P vs NP: Is there an inherent difference
between
being creative / brilliant
and
being able to appreciate creativity / brilliance?


SOME NP-COMPLETE PROBLEMS






“Feasible” computations:

those that run in polynomial
(i.e.,O(nc)) time


(central tenet of theoretical computer science)


e.g., time is “infeasible”






NP=PCP(log n, 1)


[A., Safra ‘92]

[A., Lund, Motwani, Sudan, Szegedy ’92]


INPUT x


CERTIFICATE 


n


nc


M


O(1) bits


O(lg n) random bits


Accept / Reject


x is a “YES” input

 there is  s.t. M accepts

x is a “NO” input

 for every  M rejects



> 1 - 


> ½ + 


Håstad’s 3-bit PCP Theorem (1997)


Reads 3 bits; Computes sum mod 2


Pr[ ]


Pr[ ]






(2-)-approx. to MAX-LIN(3)
) P=NP


INPUT x


?


M


O(lg n) random bits


Note: 2O(lg n) = nO(1).

) M ≡ nO(1) linear constraints

x is YES instance ) > 1- fraction satisfiable

x is NO instance ) · ½+ fraction satisfiable


1- _ ½ +






Polynomial Encoding


Idea 1


[LFKN ‘90]

[BFL ’90]


Sequence of bits / numbers


2 4 5 7


Represent as m-variate degree d polynomial:


2x1x2 + 4x1(1-x2) + 5x2(1-x1) + 7(1-x1)(1-x2)


Evaluate at all points in GF(q)m


Note: 2 different polynomials differ in (1-d/q) fraction of points.






2nd Idea:
Many properties of polynomials are locally checkable.


Program Checking [Blum Kannan ‘88]

Program Testing / Correcting [Blum, Luby, Rubinfeld ‘90]

MIP = NEXPTIME [Babai, Fortnow, Lund ’90]


1st “PCP Theorem”


Dinur [05]’s proof uses random walks on expander graphs

instead of polynomials.






Håstad’s 3-bit Theorem (and “fourier method”)


NP = PCP(lg n, 1)


T1


T2


c bits


1 bit


YES instances ) 9 T1T2 Pr[Accept] = 1

NO instances ) 8 T1T2 Pr[Accept] < 1-


V0


Raz’s Thm


S1


S2


ck bits


k bits


Pr[Accept] = 1

vs. Pr[Accept] < 2-k/10


V1


22ck bits


22k bits


LONG CODING [BGS ’95]


Verifier Composition


V2


Suppose

Pr[Accept] > ½ + 


(A few pages of Fourier Analysis)


9 S1 S2 which V1 accepts w/ Prob ø 2-k/10

) x is a YES instance.






Sparsest Cut / Edge Expansion


S


S


G = (V, E)


c- balanced separator


Both NP-hard


(G) = min


S µ V


| E(S, Sc)|


|S|


|S| < |V|/2


c(G) = min


S µ V


| E(S, Sc)|


|S|


c |V| < |S| < |V|/2






c-balanced separator


c(G) = min


S µ V


| E(S, Sc)|


|S|


c |V| < |S| < |V|/2


S


S


Assign {+1, -1} to v1, v2, …, vn to minimize


(i, j) 2 E |vi –vj|2/4


Subject to i < j |vi –vj|2/4 ø c(1-c)n2


+1


-1


|vi –vj|2/4 =1


Semidefinite relaxation for


Find unit vectors


in <n


|vi –vj|2 + |vj –vk|2 ø |vi –vk|2 8 i, j, k


Triangle inequality


“cut semimetric”


|vi –vj|2 =0