On the Power of Unique 2-Prover 1-Round Games

Subhash Khot 
Princeton University, NJ-08544
khot@cs.princeton.edu

ABSTRACT We also show that a variation of the Unique Games Con-

A 2-prover game is alled unique if the answer of one prover je ture implies that for arbitrarily small onstant Æ > 0 it is
uniquely determines the answer of the se ond prover and hard to nd an independent set of size Æn in a graph that is
vi e versa (we impli itly assume games to be one round guaranteed to have an independent set of size
(n).
games). The value of a 2-prover game is the maximum a - The main idea in all the above results is to use the 2-prover
eptan e probability of the verier over all the prover strate- game given by the Unique Games Conje ture as an \outer
gies. We make the following onje ture regarding the power verier" and build new probabilisti ally he kable proof sys-
of unique 2-prover games, whi h we all the Unique Games tems (PCPs) on top of it. The uniqueness property plays a
Conje ture : ru ial role in the analysis of these PCPs.
In light of su h interesting onsequen es, we think it is an
The Unique Games Conje ture : For arbitrarily small important open problem to prove (or disprove) the Unique
onstants ; Æ > 0, there exists a onstant k = k(; Æ ) su h Games Conje ture. We also present a semi-denite pro-
that it is NP-hard to determine whether a unique 2-prover gramming based algorithm for nding reasonable prover strate-
game with answers from a domain of size k has value at least gies for a unique 2-prover game. Given a unique 2-prover
1  or at most Æ . game with value 1  and answers from a domain of size k,
We show that a positive resolution of this onje ture would this algorithm nds prover strategies q that make the verier
imply the following hardness results : a ept with probability 1 O(k2  1=5 log( 1 )). This result
1. For any 21 < t < 1, for all suÆ iently small on- shows that the domain size k = k(; Æ ) must be suÆ iently
stants  > 0, it is NP-hard to distinguish between large if the Unique Games Conje ture is true.
the instan es of the problem 2-Linear-Equations mod
2 where either there exists an assignment that satises
1  fra tion of equations or no assignment an satisfy 1. INTRODUCTION
more than 1 t fra tion of equations. As a orol- The dis overy of the PCP Theorem ([4℄, [3℄) and subse-
lary of the above result, it is NP-hard to approximate quent quantitative improvements in PCP onstru tions have
the Min-2CNF-deletion problem within any onstant led to (in many ases optimal) hardness of approximation re-
fa tor. sults for various optimization problems. For example Max-
2. For the onstraint satisfa tion problem where every Clique [14℄, Max-3-SAT [15℄ and Set Cover [9℄ to name a
onstraint is the predi ate Not-all-equal(a; b; ), a; b; 2 few.
GF (3) , it is NP-hard to distinguish between the in- However the PCP te hniques haven't been su essful in
stan es where either there exists an assignment that obtaining \good" hardness results for some problems like
satises 1  fra tion of the onstraints or no assign- Vertex Cover (see [7℄ for an ex iting new result), Min-2CNF-
ment satises more than 98 +  fra tion of the on- deletion and oloring of graphs and 3-uniform hypergraphs
straints for an arbitrarily small onstant  > 0. We with a small hromati number. In this paper we try to
also get a hardness result for a slight variation of ap- identify some promising new dire tions for atta king these
proximate oloring of 3-uniform hypergraphs. problems.
 This work was partly supported by Sanjeev Arora's David All PCP onstru tions today (with the possible ex eption
of [7℄) follow the basi paradigm of omposing a so alled
and Lu ile Pa kard Fellowship and NSF Grant CCR- \outer verier" with an \inner verier" (proof omposition
0098180 was rst introdu ed by Arora and Safra [4℄, but the kind of
omposition we are referring to was rst used by Bellare et
al [5℄). The fo us of most of the re ent resear h has been
Permission to make digital or hard copies of all or part of this work for on improving the quality of the inner verier. Many sophis-
personal or classroom use is granted without fee provided that copies are ti ated inner veriers have been onstru ted (see [14℄, [15℄,
not made or distributed for profit or commercial advantage and that copies [21℄, [13℄) based on the Long Codes introdu ed by Bellare
bear this notice and the full citation on the first page. To copy otherwise, to et al [5℄ and the Fourier Analysis te hniques developed by
republish, to post on servers or to redistribute to lists, requires prior specific Hastad ([14℄, [15℄). However the outer verier has remained
permission and/or a fee.
STOC’02, May 19-21, 2002, Montreal, Quebec, Canada. untou hed. All PCP onstru tions use the same outer ver-
Copyright 2002 ACM 1-58113-495-9/02/0005 ...$5.00. ier, namely the one obtained by parallel repetition of a
2-prover proto ol for Gap-3SAT. The soundness property mod 2, on an instan e with optimum
p 1  produ es
required of the outer verier is given by the Raz's Parallel a solution with value 1 O( ).
Repetition Theorem [20℄ and we hen eforth all this verier
the Raz Verier. 2. A simple redu tion from 2-Linear-Equations to 2-SAT
In this paper, we point out that one promising route for gives a similar result, i.e. a (1 ; 1 t ) gap for 2-SAT
getting good hardness results for problems for whi h PCP for any 21 < t < 1. As a orollary, it is NP-hard to ap-
te hniques have failed so far, is to onstru t an outer verier proximate Min-2SAT-deletion (also alled Min-2CNF-
with \better properties". The Raz Verier is basi ally a 2- deletion) within any onstant fa tor. On the algorith-
prover game with the following ru ial properties : mi side, Zwi k's algorithm [23℄, on a 2-SAT instan e
with optimum 1  produ es an assignment with value
1. For arbitrarily small Æ > 0, it is NP-hard to determine 1 O(1=3 ). Klein et al [19℄ give O(log n log log n) ap-
whether the value of the game is 1 or at most Æ . proximation for Min-2CNF-deletion.
2. The answers of the provers are from a domain of size 3. Guruswami et al [13℄ (also see [18℄) show that the on-
k where k is a onstant depending on Æ . straint satisfa tion problem asso iated with the pred-
i ate Not-all-equal(a; b; ; d) where a; b; ; d are binary
3. The answer of the se ond prover uniquely determines variables, is hard to approximate better than a random
the answer of the rst prover. assignment. They use this fa t to derive hardness re-
sults for 4-uniform hypergraph oloring. However their
One might expe t the property (3) to be even stronger, te hniques do not work for 3-uniform hypergraphs and
i.e. the answer of the se ond prover uniquely determines one of the reasons is that for the predi ate Not-all-
the answer of the rst prover and vi e versa. In fa t su h equal(a; b; ) over binary variables, there does exist an
games have been onsidered in literature before ([10℄, [8℄) algorithm that does better than a random assignment
and they are alled \unique games". However, to the best of [22℄.
our knowledge, the question whether unique 2-prover games However we show that the Unique Games Conje ture
(with (1 , Æ ) gap in their value) are powerful enough to implies that the predi ate Not-all-equal(a; b; ) over
apture NP hasn't been onsidered before. This question is ternary variables is hard to approximate better than
pre isely the fo us of this paper and we make the following a random assignment. We also derive hardness re-
(rather bold) onje ture : sult for a variation of 3-uniform hypergraph oloring
The Unique Games Conje ture : For arbitrarily small whi h we all \semi- oloring". In this problem we
onstants ; Æ > 0, there exists a onstant k = k(; Æ ) su h are given a 3-uniform hypergraph and the goal is to
that it is NP-hard to determine whether a unique 2-prover olor the verti es so that 1  fra tion of the edges
game with answers from a domain of size k has value at least are non-mono hromati (as opposed to all edges non-
1  or at most Æ . hromati ) where  is a given parameter. (the need for
onsidering this version of oloring is due to the inher-
An important point here is that one an trivially deter- ent loss of perfe t ompleteness in the Unique Games
mine whether a unique 2-prover game has value 1. Therefore Conje ture). We show that it is NP-hard to semi-
the gap in the above onje ture is (1 ; Æ ) as opposed to olor a 3-semi- olorable 3-uniform hypergraph with
the gap (1; Æ ) in the Raz Verier. In other words, NP-hard onstantly many olors.
unique games must lose perfe t ompleteness.
We show that a positive resolution of this onje ture would 4. Frieze and Jerrum [11℄ give an algorithm for Max-k-
have many interesting onsequen es. We use the 2-prover ut that a hieves a fa tor roughly 1 k1 + 2 kln2 k . There
game given by the Unique Games Conje ture as an outer is an almost-mat hing hardness result by Kann et al
PCP verier and build appropriate inner veriers to prove [17℄ who show a hardness fa tor of 1 341k for this
the following results : problem. However in their redu tion the value of the
maximum k- ut in the ompleteness ase is 1
( k1 )
1. For every 21 < t < 1, for all suÆ iently small  > 0, it is whi h is bounded away from 1. It is an interesting open
NP-hard to distinguish between instan es of 2-Linear- problem whether a similar hardness result holds with
Equations mod 2, where either there exists an assign- perfe t ompleteness or near-perfe t ompleteness.
ment satisfying at least 1  fra tion of equations
or no assignment satises more than 1 t fra tion of We show that for any t > 12 , for all suÆ iently large
equations. onstants k, it is NP-hard to distinguish between the
instan es of Max-k- ut where the optimum value of a
This result is essentially due to H astad [16℄. He pro- k- ut is either 1  or at most 1 k(log1 k)t where
posed a so alled \ odeword test" for testing Long  > 0 is an arbitrarily small onstant.
Codes and analyzed it using Bourgain's theorem [6℄
on Fourier spe trum of boolean fun tions. However he We also onsider the following relaxation of the uniqueness
wasn't able to give a \ onsisten y test" whi h would property. We say that a 2-prover game has \d-to-1 property"
work for the Raz Verier. The (minor) ontribution of if the answer of the se ond prover uniquely determines the
this paper is to show that if one uses the outer verier answer of the rst prover and for every answer of the rst
given by the Unique Games Conje ture, it is indeed prover, there are at most d answers for the se ond prover for
possible to onstru t and analyze a onsisten y test whi h the verier would a ept. We assume d to be a xed
implying the above hardness result. integer and d  2. Consider the following onje ture :
This hardness result is tight sin e the algorithm of d-to-1 Conje ture : For arbitrarily small onstant Æ > 0,
Goemans and Williamson [12℄ for 2-Linear-Equations there exists a onstant k = k(Æ ) su h that it is NP-hard
to determine whether a 2-prover game with d-to-1 property 2. PRELIMINARIES
and answers from a domain of size at most k has value 1 or This se tion gives a preliminary ba kground on PCPs, 2-
at most Æ . prover games, Long Codes and the basi paradigm of PCP
Note that in ontrast with the Unique Games Conje ture, onstru tions.
we an hope for perfe t ompleteness in the d-to-1 Conje -
ture (sin e d  2). We use some of the te hniques from 2.1 Probabilistically Checkable Proofs
Dinur and Safra's paper [7℄ to show that the d-to-1 Conje - A language L is said to have a probabilisti he kable
ture implies the following results : proof system with parameters (r; q; ; s) if there exists a
probabilisti polynomial time verier whi h on input x of
1. For arbitrarily small ; Æ > 0, it is hard to nd an size n and a proof ,
independent set of size Æn in a graph whi h is guaran-
teed to have an independent set of size (1 211=d )n.  Uses r = r(n) random bits and queries q = q(n) bits
(see [1℄ for an algorithmi result). Note that Dinur from the proof .
and Safra's result [7℄ does not imply su h a result for  Depending on the bits read from the proof it a epts
independent sets. Su h a result is equivalent to the or reje ts.
existen e of a PCP with zero free bits, ompleteness

(1) and arbitrarily low soundness, whi h is an open  It has the following two properties :
problem.
(Completeness) : If x 2 L, there exists a proof  whi h
2. From the above result it follows
p that if 2-to-1 Conje - the verier a epts with probability  .
ture is true, it would imply 2  hardness for Vertex (Soundness) : If x 62 L, the verier a epts any proof
Cover whi h is better than the fa tor 1:3606 by Dinur with probability at most s.
and Safra. In fa t, Dinur and Safra do use an analog
of 2-to-1 property. We do not elaborate on this due to The parameters 1  > s > 0 are alled ompleteness
spa e limitations. and soundness parameters respe tively. We re all the PCP
In light of su h interesting onsequen es of the Unique Theorem ([4℄, [3℄) that every language in NP has a PCP
Games Conje ture, we think it is an important open problem system with = 1; s = 21 and the verier uses O(log n)
to prove or disprove it. In this paper, we also present a semi- random bits and queries only a onstant number of bits from
denite programming based algorithm giving the following the proof.
theorem : 2.2 2-Prover 1-Round Games
Theorem 1. There exists a (poly-time) algorithm su h Consider the following game between 2 provers and a ver-
that given a unique 2-prover game with value 1  and an- ier. There is a set V of all possible \questions" that the
swers from a domain of size k, it nds prover strategies
q
that verier an ask the rst prover and a set of questions W
make the verier a ept with probability 1 O(k2 1=5 log( 1 )). that the verier an ask the se ond prover.
A \strategy" of the rst prover is a map LV : V ! N
Andersson et al [2℄ proved a similar result for the problem where N is a set of possible answers of the rst prover. On
2-Linear-Equations mod p, where the onstraints are lin- a question v 2 V , the prover returns an answer LV (v ) to the
ear equations mod p with every equation ontaining exa tly verier. Similarly the strategy of the se ond prover is a map
2 variables. Su h onstraints have the uniqueness property LW : W ! M where M is the set of his possible answers.
sin e the value to one variable in the equation uniquely de- The \a eptan e predi ate" of the verier is a map
termines the value to the se ond variable. Our algorithm is : V  N  W  M ! fT RUE; F ALSE g
simpler and more general than that of Andersson et al.
Theorem 1 shows that if at all the Unique Games Conje - The game works in the following way. The verier pi ks a
ture is true, the domain size required k = k(; Æ ) must be pair of questions (v; w), v 2 V; w 2 W with a ertain prob-
at least  1=110 . A trivial bound k  1Æ also holds, sin e the ability distribution on the set of all pairs. He asks question
provers an hoose their answers uniformly at random from v to the rst prover and the question w to the se ond prover
the domain of possible answers and satisfy the verier with who return answers LV (v ) and LW (w) respe tively. The
probability at least 1 .
k
verier a epts i
Overview of the paper : Se tion 2 provides the pre- (v; LV (v ); w; LW (w)) = T RUE
liminary ba kground. We prove the results for 2-Linear- The value of the game is dened as the maximum, over
Equations mod 2 and Min-2CNF-deletion in Se tion 3. We all possible prover strategies, of the a eptan e probability
prove the hardness of predi ate Not-all-equal(a; b; ) over of the verier.
ternary variables in Se tion 4. We prove Theorem 1 in Se - We will be interested in games where the answer of the
tion 5 and appendix A. Proofs of all the other results are se ond prover uniquely determines the answer of the rst
omitted from this extended abstra t sin e they are quite prover, i.e. for every question pair (v; w) asked by the verier
lengthy and involved. Se tion 6 on ludes with a few re- and every answer b 2 M of the se ond prover, there is a
marks as to why it would be diÆ ult to either prove or unique answer a 2 M su h that the verier a epts. In this
disprove the Unique Games Conje ture. ase, we an asso iate a fun tion vw : M ! N for every
pair (v; w) so that the verier a epts i
vw (LW (w)) = LV (v )
 A game is alled \unique" (see [10℄, [8℄) if M = N and 2.4 Constructing PCPs, Long Codes and Fourier
every fun tion vw is a bije tion, i.e. the answer of the Analysis
se ond prover uniquely determines the answer of the rst We brie y explain a basi paradigm for PCP onstru tions
prover and vi e versa. (see [5℄, [14℄, [15℄, [21℄, [13℄). The verier an be on eptu-
Remarks : (1) We assume throughtout this paper that
ally divided into an \outer" part and an \inner" part.
the sets of answers M and N are of onstant size. (2) The The verier redu es an arbitrary language in NP to a gap-
denition of the unique games diers slightly in ([10℄, [8℄). version of Label Cover instan e L as given by Theorem 2.
In their denition, for every answer of one prover, there is at This is alled the \outer" part of the verier.
most one possible answer of the other prover and vi e versa. The verier then expe ts the proof to ontain \Long Codes"
of the labels of verti es in the instan e L. The verier pi ks
2.3 The Label Cover Problem some edge(s) of the instan e L and performs some lo al
We dene a problem alled Label Cover whi h is equiv- he ks on the supposed long odes of the supposed labels
alent to 2-prover games with the property that the answer of the endpoints of these edge(s). This lo al he king is
of the se ond prover uniquely determines the answer of the alled the \inner" part of the verier.
rst prover. For the sake of onvenien e, we prefer to talk in For proving the soundness property of the verier, one
terms of the Label Cover problem instead of 2-prover games. shows that if the verier a epts the en oded proof with
\good" probability, then the proof an be \de oded" to de-
Denition 1. A Label Cover problem L onsists of a om- ne labels for the Label Cover instan e L with a \good"
plete bipartite graph G(V; W ), withPbipartition V; W . An value of OP T (L). This gives a ontradi tion provided we
edge (v; w) has a weight pvw with v;w pvw = 1. Every started with an instan e L with suÆ iently small value of
vertex in V is supposed to get a label from a set N and OP T (L). Theorem 2 guarantees that OP T (L) an be made
every vertex in W is supposed to get a label from a set M . arbitrarily small. The proof of the soundness of the verier
With every edge (v; w) there is asso iated a \proje tion" relies on the Fourier analysis of the Long Codes.
vw : M ! N . For an assignment of labels to the verti es of We dene the Long Codes in the following.
the graph, that is for fun tions LV : V ! N; LW : W ! M ,
an edge (v; w) is said to be satised if vw (LW (w)) = LV (v ). Denition 2. A binary Long Code on a set of labels M is
The goal is to nd an assignment of labels that maximizes indexed by all fun tions f : M ! f 1; 1g. The long ode A
the total weight of the satised edges. We dene OP T (L) of a label a 2 M is given by
to be the maximum weight of edges satised by any label- A(f ) = f (a) 8 f : M ! f 1; 1g
ing. A Label Cover problem is alled \unique" if M = N
and every proje tion v;w : M ! M is a bije tion (i.e. a A heating proof might ontain an arbitrary string/table
permutation). A instead of a orre t Long Code. Su h tables are handled
by their Fourier expansion (see [15℄ for a detailed exposition)
Clearly, a Label Cover problem is same as a 2-prover game A =
X
A^  (f ) where  (f ) =
Y
f (x)
where V; W are sets of questions the verier an ask the M x2
two provers and N; M are sets of answers by the provers
respe tively. The Fourier oeÆ ients Ab satisfy the Parseval's identity,
The following theorem is a onsequen e of the PCP The- b2
P

orem ([4℄, [3℄) and Raz's Parallel Repetition Theorem [20℄.  A = 1.

It an be found in any of the papers ([5℄, [15℄, [13℄). 3. HARDNESS OF 2-LINEAR-EQUATIONS
Theorem 2. For every onstant Æ > 0, there exists a MOD 2
onstant k = k(Æ ) su h that it is NP-hard to determine In this se tion we present a proof of the following theorem.
whether a Label Cover problem L with answers from sets
Theorem 3. The Unique Games Conje ture implies that
of size at most k (i.e. jM j; jN j  k) has OP T (L) = 1 or
OP T (L)  Æ . for every 21 < t < 1, for all suÆ iently small onstants
 > 0, it is NP-hard to distinguish between the instan es
Remark : It turns out that in the redu tion given by of 2-Linear-Equations mod 2, where the fra tion of satised
Theorem 2, we have jM j  jN j and the proje tions v;w : equations is at least 1  or at most 1 t .
M ! N are highly many-to-one (this many-to-one-ness in-
reases as Æ de reases). The PCP onstru tions in this paper This result is essentially due to Hastad [16℄. He proposed
do not work for su h proje tions. Our onstru tions need a test for he king a long ode and analyzed it using Bour-
a very stringent ondition that the proje tions be bije tions gain's re ent theorem [6℄ on Fourier spe trum of boolean
or d-to-1 for some xed d independent of Æ . fun tions, whi h itself was inspired by a question raised by
It is lear that the Unique Label Cover problem orre- Hastad.
sponds to a unique 2-prover game. Hen e the Unique Games The (minor) ontribution of this paper is to introdu e the
Conje ture an be restated as : Unique Games Conje ture and to show that Hastad's test
an be extended to test the onsisten y between two long
Unique Games Conje ture : For arbitrarily small on- odes, giving a PCP verier that makes a linear test on 2
stants ; Æ > 0, there exists a onstant k = k(; Æ ) su h that query bits, has ompleteness 1  and soundness 1 t .
it is NP-hard to determine whether a unique Label Cover Following the standard paradigm, the PCP verier takes
instan e with the label sets of size k (i.e. jM j = k) has the gap-version of the unique Label Cover problem L guar-
optimum at least 1  or at most Æ . anteed by the Unique Games Conje ture and expe ts the
proof to ontain, for every vertex v 2 V , the long ode of The probability of a eptan e of the verier is learly
the label LV (v ) and for every vertex w 2 W , the long ode
Pr[A ℄ = Ev;f; 1 + A(f )A(f) + Ew 1 + A(f )B (f Æ  )
  
of the label LW (w). These long odes are assumed to be 1
folded, i.e. A( f ) = A(f ) (see [15℄). 2 2 2
The verier pi ks some edges and he ks that the labels Using the Fourier expansion A =
P
 A  we get
b
along these edges satisfy the orresponding bije tions. There
is a te hni al issue of how the edges are pi ked. Let pv =
P Ef; [A(f )A(f)℄ = Ef; [
X
Ab1 Ab2 1 (f )2 (f )2 ()℄
w pvw . That is if an edge is pi ked with a probability equal 1 ;2
to its weight, pv is the probability that the left endpoint is
v . Let v : W ! [0; 1℄ be dened as v (w) = ppvw v . That is
Note that 1 ; 2 are subsets of M . We have
v (w) is the onditional probability that the right endpoint 1 (f )2 (f ) =
Y
f (x)
Y
f (x) =
Y
f (x)
of an edge is w given that the left endpoint is v . x21 x22 x2
2
A tion of the verier : where 1
2 is the symmetri dieren e between the sets
1. Pi k v 2 V with probability pv . Let A be the (sup- 1 and 2 . The expe tation over f is non-zero only if
posed) long ode of the (supposed) label of v . 1
2 = ;, i.e. 1 = 2 = . Also E [ ()℄ = (1 2)jj .
Hen e
2. Pi k a random fun tion f : M ! f 1; 1g and a \per- Ef; [A(f )A(f)℄ = Ab2 (1 2)jj
X

turbation fun tion"  : M ! f 1; 1g. For ea h x 2 

M , (x) = 1 with probability 1  and (x) = 1 P
with probability . Using the Fourier expansion B =  B  , we have
b

3. With probability 21 ea h, sele t one of the following

X
Ef; [A(f )B (f Æ  )℄ = Ef; [ Ab Bb  (f ) (f Æ  ) ()℄
a tions : ;
(1)
(a) (Codeword test) A ept i A(f ) = A(f) We have
(b) (Consisten y test) Pi k a vertex w 2 W with the Y Y

distribution v . Let B be the (supposed) long  (f Æ  ) = f ( (x)) = f (y ) = ( ) (f )

ode of the (supposed) label of w and  = vw : x2 y2( )
M ! M be the bije tion between v and w. A - Substituting this in (1) and taking expe tation over f we
ept i see that the expe tation is non-zero only if  =  ( ). Sin e
A(f ) = B (f Æ  )  is a bije tion,  =  1 (). Thus (1) an be written as
Ef; [A(f )B (f Æ  )℄ = Ab Bb 1 () (1 2)jj
X
where f Æ  denotes the omposition of fun tions.

Remark : Hastad proposed and analyzed the odeword Hen e the probability of a eptan e is
test. We propose the onsisten y test and show that Hastad's " #
analysis an be extended to he k onsisten y provided the
Pr[A ℄ = 1 + 1 Ev
h i
Ab2 (1 2)jj + Ab Ew Bb 1 ()
X X
Unique Games Conje ture is true. 2 4  
3.1 Completeness 1 1
= + E [R + Tv ℄
It is easy to see that the ompleteness of the test is 1  +2  2 4 v v
where the outer label over instan e has ompleteness 1  . If this probability is  1 81 t t where t is as in Theorem
The test may fail due to 2 reasons : (1) The edge (v; w) 4, we have Ev [Rv + Tv ℄  2 12 t t . This implies that over
pi ked by the verier may be an unsatised edge of the label
over instan e whi h happens with probability  . In this the hoi e of v , with probability at least 21 , Rv +Tv  2 t t .
ase, the onsisten y test fails. (2) In a orre t proof A is Fix any su h \good" v . We have Rv  1 t t and Tv 
a long ode of some a 2 M . The odeword test fails when 1 t t > 21 .
(a) = 1 whi h happens with probability .
Ab2 + e 2 Ab2
X X

The laim about the ompleteness follows. Note that by 1 t t  Rv 

the Unique Games Conje ture,  an be assumed to be ar-  : jj   1  : jj >  1
bitrarily small.
Ab2 < t t
X
=) (2)
3.2 Soundness Analysis  : jj >  1
We use the following (deep) theorem of Bourgain [6℄.
Taking k =  1 in Theorem 4, we get
Theorem 4. Let A be any boolean fun tion (for instan e
Ab2 < 1
X
a supposed long ode) and k > 0 an integer. Then for every (3)
1 < t < 1, there exists a onstant t > 0 su h that 100
2  : jAb j  10
1 4 k 2

Now we use the fa t that Tv > 21 . Call  \good" if   M

Ab2 < t k t then Ab2 < 1
X X
If
is nonempty, jj   1 and jAb j  101 4 k . We will show
2

 : jj>k
100
1 4
 : jAb j 10 k2
that the ontribution of bad 's to Tv is small. First of all,
sin e the tables are folded, Ab = 0 when jj is even (see Theorem 5. If the Unique Games Conje ture is true,

[15℄). In parti ular Ab = 0 when  is empty. Also then the following holds : for a onstraint satisfa tion prob-




lem with all onstraints of the form Not-all-equal(a; b; ) and
X the variables from a ternary alphabet, it is NP-hard to deter-

 

 Ab Ew [Bb 1 () ℄ 
 mine whether there exists an assignment that satises 1 
 : jj> 1 
fra tion of the onstraints or no assignment satises more
s
X
s
X 2 than 98 +  fra tion of the onstraints, where  > 0 is an
Ab2 
 Ew [Bb 1 () ℄  arbitrarily small onsatnt.
 : jj> 1 
s X p
We will onstru t a PCP that reads 3 symbols from a
Ab2 < t t proof over ternary alphabet, a epts i the 3 symbols are
 : jj> 1 not all equal, has ompleteness 1  and soundness 98 + .
We use Long Code over GF (3) on the set of labels M .
where we used (2). Similarly we use (3), and show that the Su h a ode is indexed by all fun tions f : M ! f1; !; ! 2 g
ontribution of 's su h that jAb j  101 4 k to Tv is at most where ! is the ube root of unity. The Long Code A of
2

1 . This implies that Tv when restri ted to good 's, still

10 a 2 M is dened as A(f ) = f (a). The Fourier expansion in
remains at least 14 . We have this setting is
f (x)(x)
2 3 X Y
" # A(f ) = Ab  (f ) where  (f ) =
1
Ab2 Bb 2 Ab2 Bb 2
X X
Ew  1 () jj   Ew 4  1 () 5  x2M
  good and  ranges over all fun tions  : M ! GF (3). The Fourier
(4) oeÆ ient Ab is given by
h i
 1 4 2k Ew
 100
2
b2
P
1
 good B 1 ()
X
Ab = jM j A(f ) (f )
  2  3 f :M !f1;!;!2 g
1 2
  100 4 Ew   good   1 () 
k 2  P
A
b B
b

 h i 2 Remark : In a orre t Long ode A, we will have A(!f ) =

  100 1 4 2k2 Ew P !A(f ) and we may want to for e this ondition on every
 good A B 1 () 
b b 

(supposed) Long ode in the proof. This is alled \folding"

 1 4 2k2 1
 100 in the PCP literature. However the spe i nature of the
16
The expression on the se ond-last line is just Tv restri ted predi ate Not-all-equal(a; b; ) forbids us from doing so and
to good 's whi h we showed to be at least 41 . Note that we this makes the analysis more diÆ ult. See [13℄ for a detailed
are assuming that v is good itself, whi h holds with proba- dis ussion on this issue.
bility 21 . The verier is given a unique Label Cover instan e L guar-
Now we dene a labeling for the Label Cover instan e as anteed by the Unique Games Conje ture. It expe ts as a
follows : For a good vertex v 2 V , pi k  with probability proof the Long odes of the labels of all the verti es in L.
Ab2 , pi k a random element of  and dene it to be the label The verier works as follows :
of v . For any vertex w 2 W , pi k  with probability Bb2 , 1. Pi k a vertex v 2 V with probability pv .
pi k a random element of  and dene it to be the label of
w. 2. Pi k 3 verti es w1 ; w2 ; w3 , ea h of them independently
It is easy to see that the weight of the edges satised by from the distribution v . Let A; B; C be the (sup-
this labeling equals the expression (4). Label of v will be posed) long odes of the (supposed) labels of the ver-
dened to be a random element x 2  and the label of w ti es w1 ; w2 ; w3 respe tively. Let  = vw1 ,  0 = vw2 ,
will be dened to be a random element y 2  1 (). With  00 = vw3 be the respe tive proje tions.
probability j1 j it holds that  (y ) = x and the edge (v; w) in
the Label Cover instan e is satised. 3. Pi k two random fun tions f; g : M ! f1; !; ! 2 g.
Sin e the expression (4) is at least
(4 2k ), we get a la-
2

4. Pi k a fun tion  : M ! f!; ! 2 g by dening for ea h

belling that satises edges of total weight
(4 2k ). How- x 2 M , (x) = ! with probability 21 and (x) = ! 2
2

ever this ontradi ts the fa t that OP T (L)  Æ if Æ was ho- with probability 21 .
sen suÆ iently small (see the Unique Games Conje ture).
This shows that the soundness is at most 1 81 t t where 5. A ept i
t > 12 is arbitrary, proving Theorem 3.
Not-all-equal(A(f Æ ); B (g Æ 0 ); C (((fg)Æ 00 )
))
Remark : A simple gadget (x  y = 0 7! x _ y; x _ y )
redu es 2-Linear-Equations to 2-SAT and implies a (1 The ompleteness is 1 3 where 1  is the ompleteness
; 1 t ) gap for 2-SAT for any t > 21 . of the outer label over instan e. The verier pi ks 3 edges
and ea h of them an be an unsatised edge of the Label
4. HARDNESS OF THE PREDICATE NOT- Cover instan e with probability  . If all the 3 edges are
ALL-EQUAL(A,B,C), A,B,C 2 GF (3) satised, A; B; C are the long odes of some a; b; 2 M
respe tively and  (a) =  0 (b) =  00 ( ) = d for some d 2 M .
In this se tion we will show hardness of the predi ate Thus
Not-all-equal(a,b, ) over GF (3). This predi ate is TRUE
i a; b; do not all have the same value. We will prove that A(f Æ ) = f ( (a)) = f (d); B (g Æ 0 ) = g ( 0 (b)) = g (d) and
 C ((fg) Æ  00
) = (fg)( 00 ( ))
( ) = f (d)g (d)( ) We will show that if the terms with 6= 0 are not small,
and not all three an be equal sin e () takes values only in one an extra t labels for the Label Cover instan e L giving
the set f!; ! 2 g. a \good" value of OP T (L). Lets assume
 
 
h
1 i
Ab Bb Cb ( )j j 
4.1 Soundness Analysis X
Æ  Ev;w1 ;w2 ;w3


The following lemma is easily proven. 
 ()=0 ( )=00 ( )6=0
2 
2
Lemma 1. Let x; y; z 2 f1; !; ! g. Then the expression
Applying Cau hy-S hwartz, this expression an be bounded
1 X
by
1 xr1 y r2 z r3
9 r1 ;r2 ;r3 2GF (3) hsX s

jBb j2 jCb j2 ( 14 )j j
X i
r1 +r2 +r3 =0 Ev;w1 ;w2 ;w3 jAb j2
equals 0 if x = y = z and 1 otherwise.  0 ( )=00 ( )6=0
From this lemma it is lear that the expression implying that
1 1 A(f Æ  )r1 B (g Æ  0 )r2 C ((fg ) Æ  00
)r3 jBb j2 jCb j2 j 1j
X h i
Æ 2  Ev;w2 ;w3
X

9 r1 ;r2 ;r3 2GF (3)

r1 +r2 +r3 =0 0 ( )=00 ( )6=0
equals 1 if the test a epts and 0 otherwise. Hen e the a ep- Now we an dene labels as follows. For a vertex w3 2 W ,
tan e probability of the verier is equal to the expe tation pi k with probability jCb j2 , pi k a random y 2 M with
of this expression over the hoi e of (v; w1 ; w2 ; w3 ; f; g; ). (y ) 6= 0 and dene it to be the label of w3 . For a vertex
Let us onsider this expe tation for a xed v . We divide v 2 V , pi k a random w2 2 v , pi k  with probability
the terms in the summation into 3 ases and onsider the jBb j2 , pi k a random x 2 M with  (x) 6= 0 and dene 0 (x)
expe tation of ea h term separately : (a) r1 = r2 = r3 = 0 to be the label of v . It is easy to see that this gives a labelling
(b) (r1 ; r2 ; r3 ) take values (0; 1; 1) in some order ( ) r1 = with OP T (L)  Æ 2 .
r2 = r3 = 1 (d) r1 = r2 = r3 = 1. Hen e we an hoose L su h that OP T (L) is suÆ iently
The ase (a) is trivial, the expe tation being 1 in this ase. small and ensure that the terms with 6= 0 in (5) are arbi-
In ase (b), lets say (r1 ; r2 ; r3 ) = (1; 1; 0), the other ases trarily small. The term with = 0 ontributes
being similar. The expe tation is
Ew ;w ;w [Ab0 Bb0 Cb0 ℄ = (Ew [Ab0 ℄)3 = 3
Ew1 ;w2 ;f;g [A(f Æ  )B (g Æ  0 )℄
1 2 3

The ase (d) is just the omplex onjugate of ase ( ) and

Sin e ;  0 are bije tions, f Æ  and g Æ  0 are distributed
identi ally as f and g respe tively. Hen e the expe tation is it ontributes 3 and terms whi h an be assumed to be arbi-
trarily small. Thus we an write the a eptan e probability
Ew1 ;w2 ;f;g [A(f )B (g )℄ = Ew1 ;f [A(f )℄
Ew2 ;g [B (g )℄ as
For a xed v , let  = Ew;f [A(f )℄ = Ew [Ab0 ℄. Sin e w1 ; w2 Pr[A ℄ = 1 1 6 Ev [jj2 ℄ 1 Ev [3 + 3 ℄
are identi ally distributed, the above expe tation is same as 9 9 9
Ew;f [A(f )℄
Ew;f [A(f )℄ = jEw;f [A(f )℄j2 = jj2 + (Terms with arbitrarily small magnitude )
Now onsider ase ( ). The expe tation is Sin e jj  1, this probability is maximized when  = 0
E [A(f Æ  )B (g Æ  0 )C ((fg ) Æ  00
)℄ (the reader familiar with this area will re ognize that this
means the proof better ontain folded tables. If tables are
Substituting Fourier expansions of A; B; C , we get not folded, it an only de rease the a eptan e probability).
h X i
Hen e a eptan e probability an be bounded by 98 +  for
E Ab Bb Cb
 (f Æ  )  (g Æ  0 )  ((fg) Æ  00
) arbitrarily small  > 0.
;;
Note that 4.2 Hardness of 3-uniform Hypergraph Semi-
f ( (x))(x) = f (y )( (y)) = () (f )
Y Y coloring
 (f Æ ) =
1

x2M y2M We show that the Unique Games Conje ture implies that
it is NP-hard to semi- olor a 3-uniform hypergraph with
where we dene, by an abuse of notation,  () to be the onstantly many olors when the hypergraph is given to be
fun tion  Æ  1 . The previous expression redu es to semi- olorable with 3 olors. This is proved by ombining
h X i
the te hniques in the previous se tion with the idea of ov-
E Ab Bb Cb
() 00 ( ) (f ) 0 ( ) 00 ( ) (g ) () ering omplexity of PCPs introdu ed by Guruswami et al
;; [13℄. We skip the proof.
Taking expe tation over f; g , we see that the terms in this
summation are zero unless  () =  0 ( ) =  00 ( ). Also it 5. PROOF OF THEOREM 1
is easy to he k that E [( ) ()℄ = ( 21 )j j where for a In this se tion we prove Theorem 1. Instead of unique 2-
fun tion : M ! GF (3) we dene j j to be the number of prover games, we work in a more general setting of onstraint
x 2 M su h that (x) 6= 0. Thus the expe tation redu es to satisfa tion problems with uniqueness property.
Ab Bb Cb ( 1 )j j
h X i
Ew1 ;w2 ;w3 (5) Problem : We are given a set X of n variables whi h
()=0 ( )=00 ( )
2 take values from the set [k℄ = f1; 2; : : : ; kg. For every pair
(u; v ) of variables, there is a \ onstraint" whi h is a bije -  By repla ing ~r by ~r if needed, assume that ~r
~s  0.
tion uv : [k℄ ! [k℄. This onstraint has a weight wuv with
P
(u;v) wuv = 1.  Constru t the following assignment A : for every vari-
For an assignment A : X ! [k℄ to the variables, a on- able u, let
straint on the pair (u; v ) is satised, if uv (A(u)) = A(v ). A(u) = i0 where ~r
~ui0 = 1max
The goal is to nd an assignment that maximizes the total ik(~r
~ui )
weight of satised onstraints.
We prove the following theorem in Appendix A whi h is
Algorithm : We use a semidenite program from Feige and suÆ ient to prove Theorem 1.
Lovasz's paper [10℄ and augment it with a suitable rounding
pro edure. Let us rst formulate the problem as a quadrati Theorem 6. If there exists an assignment that satises
integer program. For every variable u 2 X , let u1 ; u2 ; : : : ; uk onstraints with total weight 1 , then the above algorithm
be auxiliary variables taking 0-1 values. Pla e the following produ es an assignment thatq
satises onstraints with ex-
onstraints : pe ted weight 1 O(k2 1=5 log( 1 )).
u21 + u22 + : : : u2k = 1 8 u 2 X (6)
ui uj = 0 8 u 2 X and 8 i 6= j (7) 6. CONCLUSION
We intend that if an assignment assigns the value i0 2 [k℄ to It seems quite diÆ ult to prove (or disprove) the Unique
a variable u, then ui0 = 1 and ui = 0 8i 6= i0 . This would Games Conje ture.
satisfy the onstraints (6), (7). These onstraints imply that Proving the onje ture is equivalent to onstru ting a PCP
for every pair (u; v ) of variables that reads 2 symbols and a epts i these symbols satisfy a
bije tive onstraint. However the urrent tools appear quite
ui vj  0 8 i; j (8) weak for onstru ting PCPs that read 2 symbols. Paral-
X
ui vj = 1 (9) lel repetition of a unique game is a unique game and one
1i;jk might hope to amplify the soundness by parallel repetition.
However we do not have a hard instan e of a unique game
It is easy to see that the goal is to maximize the following to begin with. Theorem 1 shows that if the Unique Games
fun tion subje ted to the above onstraints. Conje ture is true, the domain size k(; Æ )   1=110 , thus the
X
wuv (u1 v(1) + u2 v(2) + : : : uk v(k) ) where  = uv (10) domain size would play a very ru ial role.
(u;v) On the other hand, disproving the onje ture may require
an algorithm that gives a theorem similar to Theorem 1 and
Now we onsider the semidenite programming relaxation whose performan e is independent of the domain size k.
of the problem. We allow the variables (u1 ; : : : ; uk ) to be A less ambitious goal (than proving the Unique Games
ve tors in a high dimensional spa e (in kn-dimensional spa e Conje ture) would be to show that the value of a unique
to be pre ise) and the onstraints (6)-(9) repla ed by the 2-prover game with domain size k is hard to approximate
onstraints : within fa tor f (k) where f (k) ! 1 as k ! 1. The only
known results are onstant fa tor hardness for 2-Linear-
~u1
~u1 + ~u2
~u2 + : : : + ~uk
~uk = 1 8 u 2 X (11) Equations mod 2 by Hastad [15℄ and for 2-Linear-Equations
~ui
~uj = 0 8 u 2 X 8 i 6= j (12) mod p by Andersson et al [2℄.
~ui
~vj  0 8 u; v 2 X 8 i; j (13)
X 7. ACKNOWLEDGEMENT
~ui
~vj = 1 8 u; v 2 X (14) I am greatful to Johan Hastad for showing me his analysis
1i;jk using Bourgain's theorem, whi h led me think about unique
The goal is to maximize the following fun tion subje ted to games. I thank Sanjeev Arora, Venkatesan Guruswami and
the above onstraints : Johan Hastad for many helpful dis ussions and their valu-
X able omments on an earlier version of this paper.
wuv (~u1
~v(1) + : : : + ~uk
~v(k) ) where  = uv (15)
(u;v)
8. REFERENCES
Observation : In any feasible solution of the SDP, for any
two variables u; v , we have from the onstraints (11), (12) [1℄ N. Alon and N. Kahale. Approximating the
and (14), independen e number via the -fun tion. Te hni al
Report, Tel Aviv University, 1995.
Xk X k Xk X k [2℄ G. Andersson, L. Engebretsen, and J. H astad. A new
k ~ui k = k ~vj k = 1 and ( ~ui )
( ~vj ) = 1 way of using semidenite programming with
i=1 j =1 i=1 j =1 appli ations to linear equations mod p. Journal of
Pk Pk
This implies that i=1 ~ui = j =1 ~vj . We denote ~s = Algorithms , 39(2):162{204, 2001.
[3℄ S. Arora, C. Lund, R. Motawani, M. Sudan, and
i=1 ~ui whi h is the same for all variables u and k~sk = 1.
Pk

M. Szegedy. Proof veri ation and the hardness of

We solve the semidenite program and onstru t an as- approximation problems. Journal of the ACM,
signment using the following rounding pro edure. 45(3):501{555, 1998.
 Choose a ve tor ~r from the normal distribution, i.e. [4℄ S. Arora and S. Safra. Probabilisti he king of proofs
hoose every oordinate of ~r from the distribution N (0; 1) : A new hara terization of np. Journal of the ACM,
independently. 45(1):70{122, 1998.
 [5℄ M. Bellare, O. Goldrei h, and M. Sudan. Free bits, Pro . of the 30th Annual ACM Symposium on Theory
p ps and non-approximability. Ele troni Colloquium of Computing, pages 551{560, 1998.
on Computational Complexity, Te hni al Report
TR95-024, 1995. APPENDIX
[6℄ J. Bourgain. On the distribution of the fourier
spe trum of boolean fun tions. manus ript. A. PROOF OF THEOREM 6
[7℄ I. Dinur and S. Safra. The importan e of being biased. Let uv = 1ik ~ui
~v(i) ;  = uv whi h is the part of
P
In Pro . of the 34thth Annual ACM Symposium on
Theory of Computing, 2002. the SDP obje tive fun tion (15) orresponding to the on-
[8℄ U. Feige. Error redu tion - the state of the art. straint on (u; v ). By the hypothesis, the SDP has a solution
Te hni al Report CS95-32, Weizmann Institute of
with value at least 1  implying that there exist ve tors
Te hnology, 1995.
(~ui )u2X;i2[k℄ satisfying
X
[9℄ U. Feige. A threshold of ln n for approximating set wuv uv  1 
over. Journal of the ACM, 45(4):634{652, 1998. (u;v)
[10℄ U. Feige and L. Lovasz. Two-prover one-round proof
systems, their power and their problems. In Pro . of wuv  1 21=5
X
=)
the 24th Annual ACM Symposium on Theory of
uv 1 =
1 4 5
Computing, pages 733{744, 1992. 2

[11℄ A. Frieze and M. Jerrum. Improved approximation Fix any (u; v ) with uv q1 21 4=5 . We will show that with
algorithms for max k- ut and max bise tion.
Algorihmi a, 18:67{81, 1997. probability 1 O(k2 1=5 log( 1 )), uv (A(u)) = A(v ). Let
[12℄ M. Goemans and D. Williamson. 0.878 approximation  = uv for simpli ity. The intuition behind the proof is
algorithms for max- ut and max-2sat. In Pro . of the simple : if uv = 1, the SDP onstraints (11-14) imply that
26th Annual ACM Symposium on Theory of ~ui = ~v(i) 8 i 2 [k℄ (this an be seen by substituting  = 0
Computing, pages 422{431, 1994. in Lemma 2). Thus for any ve tor ~r, if ~r
~ui is maximized
[13℄ V. Guruswami, J. Hastad, and M. Sudan. Hardness of for index i0 , then ~r
~vj is maximized at index  (i0 ). Hen e
approximate hypergraph oloring. In Pro . of the 41st the rounding pro edure will assign, A(u) = i0 ; and A(v ) =
IEEE Symposium on Foundations of Computer  (i0 ) satisfying the onstraint.
S ien e, pages 149{158, 2000. We however have uv  1 21 4=5 and it takes some eort
[14℄ J. Hastad. Clique is hard to approximate within n1  . to translate the intuition into a rigorous proof. We pro eed
In Pro . of the 37th Annual IEEE Symposium on to prove several simple lemmas.
Foundations of Computer S ien e, pages 627{636, Lemma 2. k~ ui ~v(i) k  2=5 8 i 2 [k℄.
1996.
[15℄ J. H. stad. Some optimal inapproximability results. In
Pro . of the 29th Annual ACM Symposium on Theory Proof.

of Computing, pages 1{10, 1997. 1 4=5 X X

1   ~ui
~v(i)  k~ui kk~v(i) k
[16℄ J. H
astad. On a proto ol possibly useful for min-2sat. 2 i i
unpublished manus ript.
[17℄ V. Kann, S. Khanna, J. Lagergren, and A. Pan onesi.

X k~ui k2 + k~v(i) k2 = 1
On the hardness of max k- ut and its dual. In Pro . of 2
the 5th Israel Symposium on Theory and Computing i
Systems, pages 61{67, 1996.
[18℄ S. Khot. Hardness rsults for approximate hypergraph =)
k~ui k2 + k~v(i) k2 ~u
~v  1 4=5 8 i
i (i)
oloring. In Pro . of the 34thth Annual ACM 2 2
Symposium on Theory of Computing, 2002. =) k~ui ~v(i) k2  4=5 8i
[19℄ P. Klein, S. Plotkin, S. Rao, and E. Tardos.
Approximation algorithms for steiner and dire ted
multi uts. Journal of Algorithms, 22(2):241{269, 1997. Lemma 3. If Y is distributed as N (0; 1),
[20℄ R. Raz. A parallel repetition theorem. SIAM J. of
Computing, 27(3):763{803, 1998. 2
Pr [jY j > ℄ e 2
[21℄ A. Samorodnitsky and L. Trevisan. A p p
hara terization of np with optimal amortized query
omplexity. In Pro . of the 32nd Annual ACM Proof. Standard inequality.
Symposium on Theory of Computing, pages 191{199, q

2000. Lemma 4. With probability 1 O(k2 1=5 log( 1 )), om-

[22℄ U. Zwi k. Approximation algorithms for onstraint ponents of ~r along the dire tions of ve tors
satisfa tion problems involving at most three variables f~ui gi2[k℄; f~ui ~uj gi6=j ; f~ui ~v(i) gi2[k℄
per onstraint. In Pro . of the 9th Annual ACM-SIAM
Symposium on Dis rete Algorithms, pages 201{210, have magnitude in the range
1998. r r

1=5 log( 1 ); 1 i
h
[23℄ U. Zwi k. Finding almost satisfying assignments. In log( )
 
 Proof. This follows from the fa t that ~ r is distributed Now we will show that
in a spheri ally symmetri manner and hen e its omponent
along any dire tion is distributed as N (0; 1). Hen e for any ~r
~v(i0 ) = 1max
j k
(~r
v~j ) (16)
unit ve tor ~t,
This would imply that the assignment A given by the round-
ing pro edure assigns A(u) = i0 ; A(v ) =  (i0 ) and the on-
" r # r

Pr j~r
~tj < 1=5 log( 1 ) < 21=5 log( 1 ) straint on the pair (u; v ) is satised.
Let j 6= i0 be any index. By Lemma 8 and Lemma 7,
" # r r
p ~r
~v(j )  ~r
~uj + 2=5 log( 1 )  ~r
~ui0 1
r

Pr j~r
~tj > 1

log( ) <  42=5 log( )
  
Also by Lemma (8) we have
where the rst inequality is trivial and the se ond follows
from Lemma 3. Now we take a union bound along the O(k2 )
r

dire tions spe ied in the statement of this lemma. ~r
~v(i0 )  ~r
~ui0 2=5 log( 1 )

q It follows that
Lemma 5. With probability 1 10k1=5 log( 1 ), the om-
q ~r
~v(i0 ) > ~r
~v(j ) 8 j 6= i0
ponent of ~r along ~s, that is j~r
~sj, is at least 5k1=5 log( 1 ).
nishing the proof of (16) and Theorem 6.

Proof. Trivial.
q
Thus ex ept with probability 1 O(k2 1=5 log( 1 )), we an
assume that ~r satises hypothesis of Lemma 4 and Lemma
5. Under this assumption, we prove the following 3 lemmas.
Let i0 2 [k℄ be su h that ~r
~ui0 = max1ik ~r
~ui .

Lemma 6. k~ui0 k  51=5 .

q
( ki=1 u~i )
~r = ~s
~r  5k1=5 log( 1 ) by Lemma
P
Proof.

5 andqi0 is the index that maximizes ~r
~ui . Hen e ~r
~ui0 

51=5 log( 1 ). But by Lemma 4, the omponent of ~r along
q
~ui0 has magnitude at most log( 1 ). This implies that
k~ui k  51=5 .
0

q
Lemma 7. 8 j 6= i0 ; ~r
~uj  ~r
~ui 0 52=5 log( 1 )

Proof.

~r
~ui0 ~r
~uj = j~r
~ui0 ~r
~uj j

= j~r
(~ui0 ~uj )j q
 k~ui0 ~uj kq1=5 log( 1 ) by Lemma 4
 k~ui0 kq1=5 log( 1 ) Sin e ~ui0 ? ~uj
 52=5 log( 1 ) by Lemma 6


q
Lemma 8. 8 i; j~r
~ui ~r
~v(i) j  2=5 log( 1 )

Proof.

j~r
~ui ~r
~v(i) j = jq~r
(~ui ~v(i) )j

 qlog( 1 ) k~ui ~v(i) k by Lemma 4
 log( 1 )2=5 by Lemma 2