Degree Complexity for a Modied Pigeonhole Principle

Maria Luisa Bonety Nicola Galesiz

Department de Llenguatges i Sistemes Informatics School of Mathematics
Universitat Politecnica de Catalunya Institute for Advanced Study
Jordi Girona Salgado 1-3 Barcelona Spain Princeton 08540
e-mail [email protected] New Jersey USA
e-mail: [email protected]

Abstract
We consider a modication of the pigeonhole principle, MPHP, introduced by Goerdt in
[7]. Using a technique of Razborov [9] and simplied by Impagliazzo, Pudlak and Sgall [8], we
prove that any Polynomial Calculus refutation of a set of polynomials encoding the MPHP,
requires degree
(log n). We also prove that the this lower bound is tight, giving Polynomial
Calculus refutations of MPHP of optimal degree.
Finally we prove a simple Lemma giving a simulation of Resolution by Polynomial Calculus.

1 Introduction
Razborov in [9] proved that any Polynomial Calculus (PC ) refutation of a polynomial encoding
of the pigeonhole principle (PHPnm ) requires degree at least
(n). Here the number of pigeons
m is any integer greater than n. While other techniques were developed to prove degree lower
bounds in PC [5, 2, 1] for other combinatorial principles, the technique introduced by Razborov
and simplied by [8] wasn't succesfully applied to other principles dierent from the PHP .
The core of Razborov's technique is to produce an explicit characterization of the vector space
of all polynomials derivable from the polynomial encoding of PHP , using low degree Polynomial
Calculus refutations. This way one can study what is the minimal degree d for which refutations
of PHP of degree d exist.
The main contribution of our result is extending Razborov's technique to a combinatorial princi-
ple somewhat dierent from the pigeonhole principle. We consider a modication of the pigeonhole
principle (MPHP ), introduced by Goerdt [7]. MPHP is dened over n pigeons and log n holes,
and diers from PHP since it allows many holes to hold more than one pigeon (see Denition 2.1
for further details). In fact, MPHP is minimally unsatisable. Note that the PC degree lower
bound of Razborov doesn't give lower bounds for this principle.
In Section 3 we introduce a polynomial formulation of the MPHP principle, Poly-MPHP ,
and we prove that any Polynomial Calculus refutation of this set of polynomials requires degree
 A prelimianry version appeared as part of the paper A Study of Proof Search Algorithms for Resolution and
Polynomial Calculus published in the Proceedings of the 40-th IEEE Conference on Foundations of Computer Science,
1999
y Partly supported by Project CICYT, TIC 98-0410-C02-01 and Promocion General del Conocimiento PB98-0937-
C04-03.
z This work was done while the author was at the Universitat Politecnica de Catalunya partly Supported by an
European Community grant under the TMR project.

1
O(log n) over any eld. Following Impagliazzo Pudlak and Sgall [8], we dene a new pigeon dance
tailored for the MPHP principle and we prove those properties that we need to dene an explicit
characterization of the vector space of all polynomials derivable from the Poly-MPHP using degree
d PC refutations. As a consequence we prove that the minimal degree for refuting Poly-MPHP
in PC over any eld is of the order of the number of holes in MPHP (i.e
(log n)).
In section 4 we introduce the Resolution system. Following [3], we consider the width (i.e. the
size of the largest clause used in a refutation) as a complexity measure and we show a Polyno-
mial Calculus simulation of Resolution. Under a xed standard translation of CNF formulas to
polynomials our simulation produces PC refutations of degree bounded by the width plus 1. The
consequences of this result are discussed in the last Section where we argue that our simulation
implies that the width based proof-search algorithm of [3] cannot have a better performance than
the Groebner Basis proof-search algorithm of [6].
In Section 5, we approach the problem of proving tight degree upper bounds for PC refutations
of Poly-MPHP . In fact using the simulation result of the previous section we give PC refutations
for Poly-MPHP of optimal degree (i.e. O(log n)).
In the nal section, we propose an alternative approach to obtain the upper bound of section
5 and we discuss some consequences of our results. Also we pose some problems arising from our
work that could be interesting to study.

2 Preliminaries
The Polynomial Calculus (PC) is a refutation system for formulas in CNF. We express a CNF
formula F as a sequence of polynomials p1 ; = 0; : : :; pm = 0 over some eld K . To force 0-1
solutions we always add among the initial polynomials the axioms x2 x = 0 for all variables x. A
PC refutation is a sequence of polynomials ending with 1 = 0 such that each line in the sequence
is either an initial polynomial or is obtained from two previous polynomials in the sequence by
the following rules: (1) ff +g g for ;  2 K ; and (2) f , for any variablex. The degree of
xf
a refutation is the maximal degree of a polynomial used in the proof. The complexity of a PC
refutation is given by its degree.
We dene a standard mapping tr from formulas in CNF to sets of polynomials in the following
way: (1) tr(x) = 1 x ; (2) tr(x) = x; (3) tr(x _ y ) = tr(x)
tr(y ). We denote by [n] the set
f1; 2; : : :; ng.
We will use a tautology encoding a modication of the pigeon hole principle dened in [7]. Let
n be a natural number of the form 2k , for some k, and let m = log2 n. For each j = 1; : : :; m, let
Part(j ) be the partition of [n] induced by j the following way:
Part(j ) := ffi; i + 1; : : :; i + (2j 1)g j i = 1; 1 + 2j ; 1 + 2
2j ; : : :; 1 + ( 2n 1)
2j g
If we consider [n] as a set of pigeons and [m] a set of holes, then for each hole j 2 [m], Part(j )
j

contains sets of pigeons, e.g.

Part(1) = ff1; 2g; f3; 4g; : : : fn 1; ngg
Part(2) = ff1; 2; 3; 4g; : : :; fn 3; n 2; n 1; ngg
..
.
Part(log2 n) = ff1; 2; : : :; ngg
Consider the following denition:
Denition 2.1 For all i; i0 2 [n], i and i0 are j -compatible if and only if they are in dierent
sets of Part(j ).

2
 We consider the following property for n pigeons and log2 n holes. If each pigeon is sitting in
some hole, then there must exist an hole j and two pigeons i and i0 that are not j -compatible sitting
in hole j . Our CNF formula MPHPn expresses the negation of the previous property with the
further restriction that each pigeon must sit in exactly one hole.
W
(1) mj=1 xi;j i 2 [n]
(2) xi;j _ xi0 ;j j 2 [m]; i 6= i0 2 [n] , not j -compatible
(3) xi;j _ xi;k i 2 [n]; j 6= k 2 [m]
Notice that the set of clauses dening our MPHPn subsumes the set of clauses dening the
MPHPn of [7]. First, we add clauses encoding the restriction that each pigeon must sit in ex-
actly one hole. Second, Goerdt considered the restricted, but more complicated, denition of not
compatibility between pigeons, allowing the clauses in (2) only for not compatible piegeons i and i0
lying in dierent half sets of Part(j ) (see [7]).

3 Degree Lower Bounds for the Modied PHP

In this section we show that any polynomial calculus refutation of the MPHPn requires degree

(log n). We will use the same technique of [9, 8]. Recall that m = log n, the denition of j -
compatible pigeons and the denition of the set Part(j ), for all j 2 [m] (see Section 2). Given
Qi := 1 Pj2[m] xi;j we adopt the following polynomial formulation of the MPHPn, that we call
Poly-MPHPn :
(1) Q = 0
i i 2 [n]
(2) x x = 0 i 2 [n]; j; k 2 [m]
i;j i;k

(3) x x = 0 j 2 [m]; i; k 2 [n] not j-compatible

i;j k;j

(4) x2 x = 0 i 2 [n]; j 2 [m]

i;j i;j

For a polynomial x which is a product of xi;j , let Pigeons(x; j ) be the set of i's such that xi;j
is a factor in x.
Denition 3.1 T is the set of the monomials x = xi1;j1 : : :xi ;j such that all ik are distinct and
for all jk 2 [m] and for all i and i0 in Pigeons(x; jk ) i and i0 are jk -compatible. Td is the set of
l l

monomials in T of degree at most d.

Using the identities (2), (3) and (4) any polynomial can be represented, without increasing its
degree, as a linear combination of monomials in T . Therefore any polynomial calculus refutation
carried on modulo the ideal I generated by the polynomials (2), (3) and (4), is in the vector space
Span(T ) generated from the monomials in T . From now on we assume that all the computations
are modulo the ideal I .
We want to build aQ basis BQd for the vector space Span(T ) such that the elements of Bd are
products of the form i;j xi;j i Qi . As in [8] (and [9]) the denition of Bd is obtained from a
process that maps partial assignments into partial assignments: the pigeon dance. We consider a
dummy hole 0, and we represent elements of Bd as partial assignments according to the following
denition:
Denition 3.2 A is the set of the partial mappings a from [n] to [m] [ f0g such that for all
i; i0 2 [n], i 6= i0, if a(i) = a(i0) = j 6= 0 then i and i0 are j -compatible.

3
Let Ad := fa 2 A : jaj  dg. For a 2 A with a = f(i1; j1); : : : (ik ; jk ); (i01; 0); : : :; (i0l; 0)g, let a^
denote the restriction Qf(i1; j1); : : : (ikQ; jk )g of a. Any element a 2 A denes a polynomial xa the
following way: xa = a(i)=j;j 6=0 xi;j a(i)=0 Qi . Therefore by denition of T any polynomial xa^
associated to a^ 2 Ad is in Td .
Our pigeon dance diers from that of [9, 8] since sometimes a pigeon can be sent to an occupied
hole. Consider the following denition:
Denition 3.3 Given a 2 A, we say that a hole j is Good for the pigeon i in a, and we write
j 2 Good(i; a), if j > a(i) and for all i0 2 a 1 (j ), i and i0 are j -compatible.
Given a 2 A, our pigeon dance works the following way: starting from the rst pigeon in dom(a)
we try to move all the pigeons i 2 dom(a) into a hole j which is good for i in a.
Denition 3.4 (Dance) Let a 2 A and consider dom(a). A pigeon dance on a is a sequence of
mappings a0 ; a1; : : :an in A with the same domain as a, dened the following way: a0 = a and for
all 0 < t  n, if a(t) is undened, then at = at 1 , otherwise
(
at(j ) = at 1 (j ) j 6= t
at(t) 2 Good(t; at 1)
Denition 3.5 (Minimal Dance) Let a 2 A be given and let t be a pigeon index in [n]. By
Dt(a) we denote a mapping b 2 A such that dom(b) = dom(a), and dened as follows:
b(i) = a(i) i 2 dom(a); i = 6 t
b(t) = minj2[m] [j 2 Good(t; a)]
If minj 2[m][j 2 Good(t; a)] does not exists, then Dt (a) is undened. The minimal pigeon dance
Dmin (a) on a is: Dmin (A) = Dn(Dn 1 (


(D1(a))


)
The minimal dance has two main properties. It can be always dened whenever a dance is
dened, and it denes a one-to-one mapping from partial assignments to partial assignments. We
show these properties in the following lemmas.
Lemma 3.1 If there exists a dance on a, then there always exists a minimal dance on a.
Proof. We prove by induction on t = 1; : : :; n that there is dance b = b0; b1; : : :; bn where b0 = a
such that its rst t steps correspond to the rst t steps of the minimal dance on a. The lemma
hence follows for t = n. Assume to have proved the claim for t 1, and let b = b0 ; b1; : : :; bn the
correct dance having the rst t 1 steps as in the minimal dance. We show how to build a new
correct dance c = c0; c1; : : :; cn having its rst t steps as in the minimal dance.
Let jmin = minj 2[m] [j 2 Good(t; bt 1)] and suppose j = bt(t). Observe that since b is a correct
dance, then jmin always exists and moreover jmin  j . If j = jmin , then b is making the right choice
at the t-th step. In this case we have no need to change b, so we dene ci = bi for all i = 0; : : :; n.
Assume instead that jmin < j . In this case we dene c = c0 ; c1; : : :; cn the following way: in the
rst t 1 steps c and b are the same, that is, for all i, i = 1; : : :; t 1, ci = bi ; at the t-th step, ct
is dened by: (
ct(i) = bjmin if i = t
(i) otherwise
t

4
 The denition of ci for for i > t is as follows:
ci(j ) = (ci 1 (j ) for j 6= i
ci(i) = bj (i) otherwise
if bi (i) = jmin and i and t are not jmin -compatible
i
We have to prove that c = c0 ; c1; : : :; cn is a properly dened dance and its rst t steps are
minimal. Observe that the rst t 1 steps of c are correct and minimal since they are the same of
b. The t-th step is correct and minimal by denition of jmin. Therefore it remains to prove that
the steps strictly greater than t dene a correct dance. By the denition of ci, for i > t, we have
to prove that for all i > t ci (i) 2 Good(i; ci 1).2

Claim 3.1 For all i > t, ci(i) 2 Good(i; ci 1).

Proof. (of Claim3.1)
By the denition of ci (i) for i > t, it is easy to see that we have to prove that for all i > t such
that bi (i) = jmin and i and t are not jmin -compatible, then j 2 Good(i; ci 1). We obtain the Claim
showing that: (1) there is at most one i such that bi (i) = jmin and i and t are not jmin -compatible;
and (2) for this i we have that j 2 Good(i; ci 1).
The rst property easily follows since if there exist two dierent pigeons i and i0 both not jmin -
compatible with t, then i; i0 and t are in the same set D 2 Part(jmin ). But this is not possible
since b is a correct dance and therefore i and i0 must be jmin -compatible.
For the second point, assume we have an i such that bi(i) = jmin and i and t are not jmin -
compatible. We prove that i is j -compatible with all elements in ci 11 (j ), which proves that j 2
Good(i; ci 1). If ci 11(j ) = ;, the result is trivial. Assume, for the sake of contradiction, that there is
a i0 2 ci 11 (j ), which is not j -compatible with i. Therefore i and i0 are in the same set B 2 Part(j ).
Since i is the only pigeon on which we have modied the dance b (except for t), then it follows that
i0 was already sent to j in b, that is bi0 (i0) = j . Observe that, since i and t are not jmin-compatible,
then i and t are in the same set C 2 Part(jmin ). But since jmin < j , then C  B and therefore
t 2 B. This is a contradiction since b is a correct dance and we cannot have that bt(t) = j and
b(i) = j , for two pigeons i and t not j -compatible. 2
Lemma 3.2 The minimal dance is a one-to-one mapping.
Proof. We show that for all t = 1; : : :; n, Dt(
) is a 1-1 mapping. The result then follows since the
minimal dance is a composition of the Dt mappings. We show that if Dt(a) = Dt (a0) then a = a0.
Suppose Dt (a) = Dt (a0). Then dom(a) = dom(a0) and a(i) = a0 (i) for all i 2 dom(a); i 6= t. It
remains to show that a(t) = a0 (t). We show that neither a(t) < a0(t) nor a0 (t) < a(t). Suppose the
former. We show the following two inequalities:
Dt(a)(t)  a0 (t) a0 (t) < Dt (a0)(t)
This leads to a contradiction since Dt (a0)(t) = Dt(a)(t) and by previous two inequalities we have
that Dt (a)(t) < Dt(a)(t). The second inequality just follows from the denition of minimal dance.
To obtain the rst inequality, observe that since for all i < t, a(i) = a0(i), then either Good(t; a) 
Good(t; a0) or Good(t; a0)  Good(t; a) and moreover Good(t; a0) and Good(t; a) share the same
minimal element Dt (a)(t) = Dt (a0)(t). Since a0 (t) 2 Good(t; a0), then a0 (t)  Dt(a)(t). The other
case a0 (t) < a(t) is completely symmetric. 2
5
 Consider the following fact:
Fact
If a pigeon dance ends successfully on an a 2 A, then the polynomial associated to the dance is
in T (this is because we are moving to always strictly greater holes and therefore at the end the
dummy hole 0 has disappeared).

Lemma 3.3 If d  log2 n and a 2 Ad , then there exists a dance on a if and only if there exists a
dance on ^a.
Proof. If there is a dance for a then obviously there is a dance for ^a, so one implication is easy. For
the other implication, assume that the number of Q factors in xa is dierent from 0, since otherwise
there is nothing to prove. Assume that a^ = f(i1; j1); : : :; (ik ; jk )g and that l is the number of Q's
factors in a, so that k + l < d = log2 n . Note that after excuting the dance on a^ we remain with
at least l holes unused. We will use these unused holes to dene a dance on the whole a. That
is, if the pigeon i is in dom(^a), then a(i) = a^(i). If the pigeon i 2 dom(a) dom(^a), then we
assign one of the unused l holes to move i in a(i). Since these are completely new holes and since
jdom(a)j jdom(^a)j  l, then the dance on a is well dened. 2
We can now proceed to the denition of the basis Bd .
Denition 3.6
Bd = fxa : a 2 Ad there is a dance on a^g
It is easy to prove that the following monotonicity properties hold for Bd : (1) Bd 1  Bd ; (2)
xa 2 Bd 1 if and only if for all i 62 dom(a), xaQi 2 Bd . In order to show that Bd is a basis for
Span(Td) we need to dene an order  on polynomials in Td . We will do it as in [8].
Denition 3.7 Let xa and xb be two polynomials in Td. then xa  xb if and only if deg(xa) <
deg (xb), or if deg (xa) = deg (xb), then for the largest pigeon i such that a(i) 6= b(i), we have that
a(i) < b(i).
Lemma 3.4 Bd is a basis for Span(Td) for any d  log2 n .
Proof. Under the hypothesis that the degree d is less than log2 n we show: (1) that jBdj  jTdj and
(2) that any xa 2 Td can be expressed as a linear combination of elements of Bd , from which the
Lemma follows. The rst property follows because the minimal dance denes a 1-1 into mapping
from Bd to Td . More precisely, if xa 2 Bd then we have a dance on a^ and since d  log2 n , then by
Lemma 3.3, there is dance on a and therefore by Lemma 3.1 there is a minimal dance on a that
by Lemma 3.2 is a 1-1 mapping. Finally the observation in the previous Fact proves the rst part.
For the second part we work by induction on . Assume that for all x0  xa, x0 2 Span(Bd ). We
show that xa 2 Span(Bd ). If there is a dance on a then xa is in Bd . Otherwise we show how to
express xa as a linear combination of the elements of Bd . Let PPt be the set of all possible correct
rst t steps of the dance on a. We prove that xa 2 Span(Bd ) i b2P xb 2 Span(Bd ) by induction
on t = 0; : : :; n. Since there is no dance on a, then Pn = ; and therefore the claim follows. The
t

base of the induction t = 0 follows since P0 = a. For the induction step observe that if t 62 dom(a)
then Pt = Pt 1 and so the claim follows by induction on t. Otherwise for any b 2 Pt 1 , xb is of the
form xt;j xc . We rewrite xt;j with respect to the relation Qt , so that xb can be rewritten as
X
(1) xc xc Qt xc xt;j 0
j 0 6=j

6
 equation 1 can be rewritten as:
X X X
xc xc Qt xc xt;j 0 xcxt;j0 xc xt;j 0
j 0 <j j 0 >j;j 0 2Good(t;b) j 0 >j;j 0 62Good(t;b)
Each monomial in the last term is equal 0, therefore in Span(Bd ). The rst three terms in the
above sum are in Span(Bb ) by induction on . The rst by the base case of the denition of .
The second by the induction case of  and by the monotonicity property of Bd . The third by (the
second case of the denition) . The fourth term corresponds exactly to all the possible correct
rst t steps of b. Therefore if we sum over all xb for b 2 Pt 1 we have that
X X
xb 2 Span(Bd ) i xb 2 Span(Bd )
b2P t b2P 1 t

This concludes the proof of the Lemma. 2

Theorem 3.1 Any polynomial calculus refutation of MPHPn has degree not less than log2 n .
Proof. The proof is as in [8]. That is we prove by induction on the length of the proof that each
polynomial derivable from the initial polynomials Qi with at most degree d is a linear combinations
of polynomials in Bd Td (i.e a combination of the elements of Bd that are multiples of some axioms
Qi ). Therefore since 1 2 Td and it has a unique representation in each basis, we cannot derive the
polynomial 1 with a proof of degree less than or equal to d.
Recall that we are considering refutation modulo the vector I . Therefore in the base case an
axiom is always of the form Qi for some i 2 [n], and the claim follows.
In the inductive step, if a line is inferred by the sum rule the result is immediate. For the case
of product, say we have x xx , with jaj  d 1. We want to prove that xi;j xa 2 Span(Bd Td ).
a

By induction, xa can be written as a sum of elements in Bd Td (i.e. multiples of Qi ). Therefore

a i;j

distributing xi;j along elements of this sum, we can write xa xi;j as a sum of multiples of Qi . By
the monotonicity properties of Bd , it is easy to see that this is a sum of scalar multiples of Qi and
therefore in Span(Bd Td). 2

4 Resolution lower bounds via degree lower bounds

Resolution is a refutation proof system for formulas in CNF form based on the following resolution
rule: C _Cx _Dx_D where if C and D have common literals, they appear only once in C _D. A resolution
proof of a CNF formula F is a derivation of the empty clause from the clauses dening F , using
the above inference rule. Following [3] the width w(F ) of a CNF formula F is dened to be the
the number of literals of the largest clauses in F . The width w(R) of a refutation R is dened as
the size of the greatest clause appearing in R. The width w(` F ) of refuting a formula F is dened
as min w(R)
R`F
We prove that degree lower bounds imply width lower bounds as long as the initial polynomials
of the PC proofs are obtained by the standard mapping tr of the initial clauses of the resolution
proofs.
Lemma 4.1 Given a set of unsatisable clauses F and a resolution refutation of F , there is a
polynomial calculus refutation of tr(F ) of degree less than or equal to w(` F )+1.

7
Proof. Observe that given two clauses A and B, it is easy to obtain a PC derivation of
tr(A) = 0 ` tr(A)tr(B) = 0
with degree w(A) + w(B ).
We show that for each clause A in the resolution proof we nd a PC derivation of tr(A) = 0
with degree at most the width of deriving A plus one. If A is an initial clause the result follows by
denition of tr. Now assume that at a resolution step we are in the following situation
A _ x x _ B
D
We will simulate the resolution rule by a few PC steps. Assume that A = A0 _ C and B = B 0 _ C ,
i.e. C is the clause formed by the literals that belong to both A and B . D = A0 _ B 0 _ C . By
induction we have derived
tr(A)(1 x) = tr(A0 )tr(C )(1 x) = 0 and tr(B)x = tr(B 0 )tr(C )x = 0
By the previous observation we can obtain the refutations
tr(A0)tr(C )tr(B 0 )(1 x) = 0 and tr(B 0 )tr(C )tr(A0)x = 0
An application of the sum rule gives tr(D) = 0. Note that the premises and conclusion of the
resolution rule get translated by polynomials of the same degree as the width of the clauses. The
steps added in the simulation can increment by 1 the degree respect to the width. 2
As a consequence of the previous lemma and the width-size trade-o [3](see also last Section),
a linear (in the number of variables) degree lower bound in polynomial calculus can give us an
exponential lower bound in resolution size.

5 Upper bounds for the Modied Pigeonhole Principle

In this section we prove that the lower bound obtained in Section 3 is tight giving degree O(log n)
PC refutations of Poly-MPHPn. We use the simulation Lemma 4.1 of the previous section.
First consider the following denition and Lemma from [5].
Denition 5.1 ([5]) Let P (~x) and Q(~y) be two sets of polynomials over a eld F . Then P is
(d1; d2)-reducible to Q if :
1. For ever yi , there is a degree d1 denition of yi in terms of ~x. That is for every i, there exist
a degree d1 polynomial ri sucht that yi can be viwed as dened by yi = ri (~x);
2. There exists a degree d2 PC refutation of the polynomials Q(~r(~x)) from the polynomials P (~x)
Lemma 5.1 ([5]) Suppose that P (~x) is (d1; d2)-reducible to Q(~y). Then if there is a degree d3 PC
refutation of Q(~y), then there is a degree max(d2; d3d1) PC refutation of P (~x).
We'll use the previous Lemma to obtain PC refutations of Poly-MPHPn from PC refutations
of the set of polynomials Poly-3MPHPn that we dene next.

8
 First we dene the formula 3MPHPn obtained from MPHPn as follows: for all i 2 [n], let
yi;0; : : :;Wyi;m be m + 1 new extension variables. Dene 3MPHP by changing for all i 2 [n], the
clauses mi=1 xi;j of MPHP by the set of clauses:
^
m
yi;0 ^ (yi;j 1 _ xi;j _ yi;j ) ^ yi;m
j =1
where yi;j means that i goes to one of the rst j holes.
Denition 5.2 Poly-3MPHPn is obtained applying the mapping tr to the clauses dening 3MPHPn.
Lemma 5.2 Poly-MPHPn is (1,3)-reducible to Poly-3MPHPn
Proof. Dene Xm
yi;j = 1 xi;k
k>j
We prove that the initial polynomials of Poly-3MPHPn (with the y substituted by the denitions
above) are derivable with a 3-degree polynomial calculus refutations from initial polynomials of
Poly-MPHPn .
Observe that yi;0 = Qi = 0 and yi;m = 1. So we can easily prove yi;0 = 0 and 1 yi;m = 0. A
generic initial polynomial of Poly-3MPHPn of the form
(1 yi;j 1 )(1 xi;j )(yi;j ) for 2 < j < n
is equivalent to
X
m X
m
( xi;k )(1 xi;j )( xi;k )
k>j 1 k>j
And using again Qi = 0, the previous equation can be rewritten as
jX1 X
m
(1 xi;k )(1 xi;j )(1 xi;k )
k=1 k>j
By simple calculations (using the initial axioms of Poly-MPHPn ) this is equal to
X
n jX1
(1 Qj ) + xi;k xi;j + (1 xi;j ) xi;k summk=j+1 xi;k
k=1;k6=j k=1
Using the initial axioms of Poly-MPHPn it is easy to see that each of the three terms of the above
polynomial is equal to 0. 2
In order to use lemma 4.1 to obtain O(log n) degree upper bounds for Poly-3MPHPn we need
to prove O(log n) width upper bounds (in resolution) for 3MPHPn . We adapt a proof sketched by
Goerdt from [7].
Lemma 5.3 There are resolution refutations of 3MPHPn of size nO(log n) and width O(log n).

9
Proof. First observe that there is an easy tree-like resolution refutation of size polynomial in n
and of width O(log n) to obtain the initial clauses of MPHPn from the initial clauses of 3MPHPn .
We show, following [7], that there is a resolution refutation of MPHPn of width O(log n) and size
nO(log n) .
The proof goes by induction on k = 4; : : :; log n. The case MPHP4 is easy and can be also
found in [7]. Assume to have, by induction, a refutation of MPHP2 , for k  4. We give a proof
k

of MPHP2 +1 .
k

Consider the intial clauses of MPHP2 +1 of the form

k

xi;1 _ : : : _ xi;k+1 (1)

for all i = 1; : : :; 2k+1 . Resolve all of these clauses in parallel with the clauses xi;k+1 _ xi0 ;k+1 for
all i0 6= i, such that i and i0 are not (k + 1)-compatibles. This leaves us with the following clauses
xi;1 _ : : : _ xi;k _ xi0 ;k+1
for all i = 1; : : :; 2k+1 and for all i0 6= i not (k +1)-compatibles. Applying to these clauses the proof
of MPHP2 we have by induction, we produce the singleton clauses xi0 ;k+1 . Now we use these
k

clauses and resolve with clauses in (1) to obtain for al i = 1; : : :; 2k the clauses
xi;1 _ : : : _ xi;k
With another application of the proof of MPHP2 applied to these clauses we obtain the empty
k

clauses.
It is straigthforward to see that in the refutation we never use clauses of width greater than
O(log n) and that the total number of clauses derived is at most quasipolynomial in n. 2
Theorem 5.1 There are O(log n) degree PC refutations of Poly-MPHPn.
Proof. Lemma 4.1 and the previous Lemma gives us O(log n) degree PC refutations of Poly-
3MPHPn . Then Lemma 5.2 implies the claim of the Theorem 2

6 Discussion and Open Problems

Resolution and Polynomial Calculus: width, size and degree
Consider the following two Theorems proved respectively in [6] and [7].
Theorem 6.1 ([6]) If a set of clauses F over n variables and of width at most k has a tree-like
resolution refutation of size S , then the set of polynomials tr(F ) has a PC refutation of degree
k + O(log S ).
Theorem 6.2 ([7]) There are tree-like resolution refutations of MPHPn of size nO(1) and width
O(n).
Following the strategy used in the previous section and the previos theorems we can obtain
the same degree O(log n) upper bound for PC refutations of Poly-MPHPn . The analysis of the
Theorem 6.1 shows that the number of poynomials produced by the simulation is roughly of the
order nO(log S ) . Therefore using this theorem we cannot improve the nO(log n) number of polynomials
10
that also our simulation produces, given that we are using quasipolynomial size resolution refutation
for MPHPn .
Consider the following resuming table for MPHPn and Poly-MPHPn :
System size width degree
tree-like Res nO(1) O(n) 
dag-like Res n O (log n) O(log n) 
PC n O (log n)  (log n)
Observe that there seems to be a trade-o between size and width for Resolution refutations
of MPHPn . To reduce the width from O(n) to O(log n) the size should increase considerably. We
don't know, for the case of MPHPn , if there exist refutations of O(log n) width and polynomial
size. We suspect that such proofs don't exists; i.e. that we cannot always obtain proofs of optimal
size and width at the same time. It could be interesting to study such a kind of questions also
for other tautologies or even in a general setting analyzing the relationships between size, width in
Resolution and degree in Polynomial Calculus.
Resolution: width optimality and proof searching
The previous question is also motivated by the problem of searching for proofs or refutations. Ben-
Sasson and Wigderson in [3], proposed an algorithm for searching for Resolution proofs based on
the width measure:
Algorithm BW
C:= Clauses of F
w := 0
While 2 62 C Do
w := w + 1
apply resolution rule to clauses in C to derive all possible clauses of width  w.
Add the clauses obtained to C
End

The running time of this algorithm is nO(w(`F )) when it searches for refutations of a formula F
over n variables. The relationship between size and width was studied by [3] in order to reduce the
problem of size lower bounds to that of width lower bounds. They proved the following theorem.
Theorem 6.3 ([3]) Let F be any unsatisable CNF formula over n variables and let S (` F ) be
the size of the shortest resolution refutation of F . Then:
!!
S (` F )  exp
(w (` F ) w (F ))2

n
We proved that this trade-o is optimal ([4]). It would be interesting to nd some unsatisable
formula F over O(n2) variables with polynomial size resolution refutations but requiring degree

(n) in Polynomial Calculus for some polynomial encoding. This result would show that there is a
tautology provable fast in Resolution, but requiring high degree in Polynomial Calculus. A possible
good candidate could be the formula GTn used by [4] to prove the optimality of the size-width trade
o.
11
Resolution and Polynomial Calculus: simulations
Recall the simulation Theorem of Clegg,Edmonds and Impagliazzo [6]:
Theorem 6.4 ([6]) If a set of clauses F over n variables and of width at most k, has a deag-like
p refutation of size S , then the set of polynomials tr(F ) has PC refutation of degree at
resolution
most 3 n loge S + k + 1.
The size-degree trade-o of the previous Theorem is optimal, since there is a trivial resolution
proof of size O(2n) for the PHPnn+1 , and Razborov in [9] showed that PHPnm require
(n) degree
PC refutations for all m. It could be interesting to prove the optimality of the previous simulation
also in the case when S is "small" (i.e. polynomial in the size of the formula). We suspect that
this result could be true. A way to obtain it is proving a
(n) degree lower bound for the GTn
principle.
Observe that our Lemma 4.1 is better (in the sense that it gives a smaller degree PC refutations)
than the corresponding simulation result (Theorem 6.4) of [6] in the case we have constant width
polynomial Resolution refutations of formulas having initial clauses of constant size. Moreover it
implies that under the tr translation the width based algorithm of [3] cannot be better than the
Grobner basis algorithm of [6].
Finally notice that there is no simulation of Polynomial Calculus by Resolution. Therefore it
would be also interesting to obtain the opposite direction of Lemma 4.1.

Acknowledgment
The authors would like to thank the anonymous referees of the PhD Thesis of N. Galesi for many
comments about the preliminary version of this paper. These comments contributed to improve
the quality of the presentation.

References
[1] M. Aleckhnovitch and A. Razborov. Lower bounds for the polynomial calculus: non binomial
case. Manuscript, September 2000.
[2] E. Ben-Sasson and R. Impagliazzo. Random CNF are hard for the Polynomial Calculus.
Proceedings of the IEEE Symposium on Foundations of Computer Science (FOCS), pages
412{421, 1999.
[3] E. Ben-Sasson and A. Wigderson. Short proofs are narrow - resolution made simple. To appear
in Journal of the ACM.
[4] M.L. Bonet and N. Galesi. Optimality of Size-Width Trade-O for Resolution In Proceedings
of the IEEE Symposium on Foundations of Computer Science (FOCS), pages 422{431, 1999.
Journal version submitted to Computational Complexity
[5] S. Buss, D. Grigoriev, R. Impagliazzo, and T. Pitassi. Linear gaps between degrees for the
polynomial calculus modulo distinct primes. In Proceedings of the ACM Symposium on Theory
of Computing (STOC), pages 547{556, 1999.

12
[6] Matthew Clegg, Jeery Edmonds, and Russell Impagliazzo. Using the Groebner basis algo-
rithm to nd proofs of unsatisability. In Proceedings of the Twenty-Eighth Annual ACM
Symposium on the Theory of Computing, pages 174{183, Philadelphia, Pennsylvania, 22{24
May 1996.
[7] A. Goerdt. Unrestricted resolution versus N-resolution. Theoretical Computer Science, 93:159{
167, 1992.
[8] R. Impagliazzo, P. Pudlak, and J. Sgall. Lower bounds for the polynomial calculus and the
groebner basis algorithm. Computational Complexity, 8(2):127 {144, 1999.
[9] A. Razborov. Lower bounds for the polynomial calculus. Computational Complexity, 7(4):291{
324, 1998.

13