Buss Arith

Bounded Arithmetic and
Propositional Proof Complexity

Samuel R. Buss∗
Departments of Mathematics and Computer Science
University of California, San Deigo
La Jolla, CA 92093-0112
Abstract
This is a survey of basic facts about bounded arithmetic and about
the relationships between bounded arithmetic and propositional proof
complexity. We introduce the theories S2i and T2i of bounded
arithmetic and characterize their proof theoretic strength and their
provably total functions in terms of the polynomial time hierarchy.
We discuss other axiomatizations of bounded arithmetic, such as
minimization axioms. It is shown that the bounded arithmetic
hierarchy collapses if and only if bounded arithmetic proves that the
polynomial hierarchy collapses.
We discuss Frege and extended Frege proof length, and the two
translations from bounded arithmetic proofs into propositional proofs.
We present some theorems on bounding the lengths of propositional
interpolants in terms of cut-free proof length and in terms of the
lengths of resolution refutations. We then define the Razborov-
Rudich notion of natural proofs of P 6= NP and discuss Razborov’s
theorem that certain fragments of bounded arithmetic cannot prove
superpolynomial lower bounds on circuit size, assuming a strong
cryptographic conjecture. Finally, a complete presentation of a proof
of the theorem of Razborov is given.
1 Review of Computational Complexity

1.1 Feasibility
This article will be concerned with various “feasible” forms of computability
and of provability. For something to be feasibly computable, it must be
computable in practice in the real world, not merely effectively computable
in the sense of being recursively computable.
∗ Supported in part by NSF grants DMS-9503247 and DMS-9205181.
The notion of effective computation is an intuitive (nonformal) concept:
a problem is effectively decidable if there is a definite procedure which, in
principle, can solve every instance of the problem. Church’s Thesis (for which
there is strong evidence) states that a problem is effectively computable if and
only if it can be recursively computed, or equivalently, if and only if it can be
decided by a Turing machine computation. However, effectively computable
problems may not be feasibly computable since the computational resources
such as time and space needed to solve an effectively computable problem
may be enormous. By “feasible” we mean “computable in practice” or
“computable in the real world”; or more to the point, a problem is feasibly
computable if for any reasonably sized instance of the problem, we are able
to solve that instance, albeit perhaps with a large (but not impossibly large)
commitment of computing resources.
Like the notion of “effective”, the notion of “feasible” is an intuitive
concept; so it is natural to look for a feasible analog of Church’s Thesis
that will give a formal, mathematical model for feasible computation. A
big advantage of having a mathematical model for feasible computability
is that this allows a mathematical investigation of the power and limits of
feasible computation. There is a widespread opinion that polynomial time
computability is the correct mathematical model of feasible computation.
This widespread belief is based primarily on two facts. First, the class
of polynomial time computable functions and predicates is a robust and
well-behaved class which has nice closure properties and is invariant for
many natural models of computation. Second, and more importantly, it is
an empirical observation about actual, real-world algorithms that it is nearly
always the case that decision problems which are known to have polynomial
time algorithms are also known to be feasibly decidable (see Cook [17] for a
discussion of this).
1.1.1 Feasibility Versus Infeasibility — An Example

For an example of feasibility and infeasibility, we will consider the problem
of integer multiplication. We’ll give two algorithms, the first one is effective
but not feasible, while the second one is effective and feasible.
For multiplication, we presume we are given two integers x and y in
binary notation. We also presume that we already have feasible algorithms
for addition which will be used as subroutines in our algorithms for
multiplication. Let the length of x, denoted |x|, be the number of bits
in the binary representation of x. For the purposes of computing runtimes,
we presume that our subroutine for adding x and y uses |x| + |y| steps
(where a “step” involves only a constant number of bit operations).
The infeasible algorithm: This is what we might call the first-grade
algorithm for multiplication, since in the U.S., students might learn this
algorithm in the first grade. The algorithm is very simple: given two positive
integers x and y , add x to itself repeatedly y − 1 times. The runtime of
2
this algorithm is (y − 1) · (|x| + |y|) steps.
The feasible algorithm: This algorithm can be called the fourth-grade
algorithm for multiplication (in California, this would be taught in the fourth
grade). This is the usual grade school algorithm for multiplication; e.g., to
multiply 6 and 5 in binary notation, one writes:
110 (= 6)
101 (= 5)
110
000
110
11110 (= 30)
The runtime of this algorithm is approximately |x| · |y| steps.

To see why the first algorithm is infeasible and the second one is feasible,
consider two 20 digit numbers x and y . To be as concrete as possible,
we estimate the runtimes in seconds, assuming that each “step” takes one
microsecond. Since x ≈ 1020 , we have |x| ≈ 70 . Likewise, |y| ≈ 70 . The
runtime of the infeasible algorithm is therefore
≈ 140 ∗ 1020 µs ≈ 1000 × (age of the universe).
On the other hand, the runtime of the feasible algorithm is

1
≈ 702 µs = 4900µs ≈ s
200
Clearly there is an impressive difference between the infeasible algorithm
and the feasible algorithm! In this example, it is hopeless to carry out
the infeasible algorithm to completion; whereas, the feasible algorithm is
extremely quick on modern computers.
Definition Let n denote the sums of the lengths of the inputs to an

algorithm. An algorithm is polynomial time if its runtime is bounded by nc
for some constant c . An algorithm is has exponential runtime if its runtime
c
is bounded by 2n for some constant c .
Clearly the infeasible algorithm has exponential runtime (but not poly-
nomial runtime); whereas the feasible algorithm is polynomial time.
Remarks: From our theoretical viewpoint, we are primarily interested in
the asymptotic behavior of the algorithms. We are also usually not interested
in constant factors. Thus, the fact that modern day computers have builtin
operations for multiplication of 32-bit or even 64-bit integers does not make a
significant difference, since it only speeds up the above estimates of runtime
by a factor of 32 or 64. Likewise, the fact that computers are currently
doubling in speed every couple years does not affect our interest in feasible
versus infeasible. (Actually, the doubling of speed is one of the primary
3
reasons for neglecting constant factors in runtime.)† The quadratic time,
O(n2 ), algorithm for multiplication is not asymptotically optimal: O(n log n)
runtime is possible using fast Fourier transform techniques.
1.2 P, NP, and the Polynomial-Time Hierarchy

We let N denote the set of non-negative integers.
Definition P is the set of polynomial time recognizable predicates on N.

Or equivalently, P is the set of polynomial time recognizable subsets of N
F P is the set of polynomial time computable functions.
We let |x| = dlog2 (x + 1)e denote the length of the binary representation
of x. We write |~x| =|x1 |, |x2 |, . . . , |xk |.
Remark: For us, functions and predicates are arithmetic: polynomial time
means in terms of the length |x| of the input x . Generally, computer
scientists use the convention that functions and predicates operate on strings
of symbols; but this is equivalent to operations on integers, if one identifies
integers with their binary representation.
Cobham [15] defined F P as the closure of some base functions under
composition and limited iteration on notation. As base functions, we can
take 0, S (successor), b 12 xc , 2 · x,
½
1 if x ≤ y
x≤y =
0 otherwise
½
y if x > 0
Choice(x, y, z) =
z otherwise
Definition Let q be a polynomial. f is defined from g and h by limited

iteration on notation with space bound q iff
f (~x, 0) = g(~x)
f (~x, y) = h(~x, y, f (~x, b 12 yc))
provided |f (~x, y)| ≤ q(|~x|, |y|) for all ~x, y .
Theorem 1 (Cobham [15]) F P is equal to the set of functions which can

be obtained from the above base functions by using composition and limited
iteration on notation.
† The rate of increase in computing performance presently shows little sign of abating;
however, if should not be expected to continue forever, since presumably physical limits
will eventually be met. It is interesting to note that if the rate of increase could be
sustained indefinitely, then exponential time algorithms would actually be feasible, since
one would merely wait a polynomial amount of time for computer power to increase
sufficiently to run the algorithm quickly.
4
A nondeterministic Turing machine makes “choices” or “guesses”: the
machine accepts its input iff there is at least one sequence of choices that
leads to an accepting state. On the other hand, a co-nondeterministic Turing
machine makes “universal choices”: it accepts iff every possible sequence of
choices leads to an accepting state. NP is defined to be the set of predicates
accepted by non-deterministic polynomial time Turing machines. The class
coNP is defined to equal the set of predicates accepted by co-nondeterministic
Turing machines, i.e., coNP is the set of complements of NP predicates. For
example, the set of (codes of) satisfiable propositional formulas is in NP .
And the set of (codes of) tautologies is in coNP .
We now give a second, equivalent, purely logical method of defining NP
and coNP .
Definition If Ψ is a set of predicates, PB∃(Ψ) is the set of predicates A

expressible as
~x ∈ A ⇔ (∃y ≤ 2p(|~x|) )B(~x, y)
for some polynomial p and some B ∈ Ψ .
PB∀(Ψ) is defined similarly with universal polynomially bounded quantifi-
cation.
Definition NP = PB∃(P ) and coNP = PB∀(P ) . In other words, NP is

the set of predicates A(~x) expressible in the form
A(~x) ⇔ (∃y ≤ 2p(|~x|) )B(~x, y)
for some B ∈ P and some polynomial p . The class coNP is similarly defined
with a universal polynomially bounded quantifier.
Definition If Ψ is a set of predicates, P Ψ (resp., F P Ψ ) is the set of

predicates (resp., functions) polynomial time recognizable with oracles for a
finite number of predicates in Ψ .
The notions of P , F P , NP and coNP are generalized by defining the

polynomial time hierarchy:
Definition The Polynomial Time “Hierarchy”:

p
1 = FP
∆p1 = P
Σpk = PB∃(∆pk )
Πpk = PB∀(∆pk )
p p
∆pk+1 = P Σk = P Πk
p p
p
k+1 = F P Σk = F P Πk
5
. .
. .
. .
∆p3 p
3
⊇
Πp2 Σp2
⊇
⊇
⊆
∆p2 p
2
⊆
⊇
coNP = Πp1 Σp1 = NP
⊇
⊇
P = ∆p1 p
1 = FP
It is an open question whether P = NP , NP = coNP and whether the

polynomial hierarchy is proper. To sum up the current state of knowledge,
all that is known is that if any two of the classes in the above hierarchy are
equal, then the hierarchy collapses to those classes.
Remark: The above definitions have given purely logical characteriza-
tions of the classes in the polynomial time hierarchy in terms of expressibility
in a formal language. In the next section we will give purely logical
characterizations of these classes in terms of derivability in a formal theory.
The most important such theory is S21 in which the provably recursive
functions are precisely the polynomial time computable functions. Thus
we can characterize the proof-theoretic strength of S21 as corresponding to
polynomial time computability.
2 Bounded Arithmetic
A constructive proof system is one in which proofs of existence contain, or
imply the existence of, algorithms for finding the object which is proved to
exist. For a feasibly constructive system, the algorithm will be feasible, not
merely effective. For instance, if ∀x∃yA(x, y) is provable then there should
be a feasible algorithm to find y as a function of x. In the next section, we
introduce feasible proof systems for number theory: more precisely, S21 will
be a feasible proof system, and other systems, S2i and T2i are systems that
have proof-theoretic strength corresponding to higher levels of the polynomial
time hierarchy.
2.1 The Language of Bounded Arithmetic

The theories of bounded arithmetic will be first-order theories for the natural
numbers N = {0, 1, 2, . . .}. The first-order language for bounded arithmetic
contains the predicates = and ≤ and contains function symbols 0 , S
6
(successor), +, ·, b 12 xc , |x| , # and relation symbol ≤ , where
x#y = 2|x|·|y|
It is easy to check that the # (pronounced “smash”) function allows us to

express 2q(|~a|) for q any polynomial with positive integer coefficients.
Definition A bounded quantifier is a quantifier of the form (Qx ≤ t) with

t a term not involving x. A sharply bounded quantifier is one of the form
(Qx ≤ |t|). (∀x) and (∃x) are unbounded quantifiers. A bounded formula is
one with no unbounded quantifiers.
A hierarchy of classes Σbk , Πbk of bounded formulas is defined by counting

alternations of bounded quantifiers, ignoring sharply bounded quantifiers.
(Analogously to defining the arithmetic hierarchy by counting unbounded
quantifiers, ignoring bounded quantifiers.)
Definition Σb0 = Πb0 is the set of formulas with only sharply bounded
quantifiers.
If A ∈ Σbk then (∀x ≤ |t|)A and (∃x ≤ t)A are in Σbk and (∀x ≤ t)A
is in Πbk+1 . Dually, if A ∈ Πbk then (∃x ≤ |t|)A and (∀x ≤ t)A are in Πbk
and (∃x ≤ t)A is in Σbk+1 . For formulas not in prenex form, we say that
a formula is in Σbi (resp., Πbi ) iff prenex operations can be used to put the
formula in to the prenex Σb1 (resp., Πbi ) form defined above.
One of the primary justifications for the definition of Σbi - and Πbi -formulas
is the following theorem.
Theorem 2 Fix k ≥ 1 . A predicate Q is in Σpk iff there is a Σbk formula

which defines it.
This theorem is essentially due to Stockmeyer [48] and Wrathall [51];

Kent and Hodgson [31] were the first to prove a full version of this.
Remarks: There are several reasons why the # function and sharply
bounded quantifiers are natural choices for inclusion in the language of
bounded arithmetic:
• The # function has the right growth rate for polynomial time
computation.
• The above theorem defines the Σ , Π classes of the polynomial hierarchy
syntactically (without use of computation),
• The presence of # in the language gives precisely the right growth rate
so that the following Quantifier Exchange Principle holds:
(∀x ≤ |a|)(∃y ≤ b)A(x, y) ↔
↔ (∃y ≤ (2a + 1)#(4(2b + 1)2 ))(∀x ≤ |a|)
[A(x, β(x + 1, y)) ∧ β(x + 1, y) ≤ b]
7
• The original use of the # function was by E. Nelson [38]; his primary
reason for its introduction was that the growth rate of # allows a
smooth treatment of sequence coding and of the metamathematics of
substitution. Wilkie and Paris [50] independently introduced the axiom
Ω1 (“xlog x is total”) for similar reasons, since it gives similar growth
rate.
The original use of polynomially bounded quantifiers was by Bennett [1];
they were first defined in form given above by [3, 4].
2.2 Induction Axioms for Bounded Arithmetic

The IND axioms are the usual induction axioms. The PIND and LIND
axioms are “polynomial” and “length” induction axioms that are intended
to be feasibly effective forms of induction.
Definition Let i ≥ 0. The following are axiom schemes often used for
theories of bounded arithmetic.
Σbk -IND: A(0) ∧ (∀x)(A(x) ⊃ A(x + 1)) ⊃ (∀x)A(x) for A ∈ Σbk .
Σbk -PIND: A(0) ∧ (∀x)(A(b 12 xc) ⊃ A(x)) ⊃ (∀x)A(x) for A ∈ Σbk .
Σbk -LIND: A(0) ∧ (∀x)(A(x) ⊃ A(x + 1)) ⊃ (∀x)A(|x|) for A ∈ Σbk .
The axiom schemes Σbk -LIND and Σbk -PIND typically are equivalent and are
(strictly?) weaker than Σbk -IND. Since exponentiation is not provably total
in Bounded Arithmetic, the |x| function is not provably surjective; therefore,
the LIND axioms do not appear to equal to the IND axioms in strength.
2.3 Theories of Bounded Arithmetic

Definition Let i ≥ 0. T2i is the first-order theory with language 0 , S , + ,
·, b 21 xc , |x| , # and ≤ and axioms:
(1) A finite set, BASIC, of (universal closures of) open axioms defining
simple properties of the function and relation symbols. BASIC properly
contains Robinson’s Q since it has to be used with weaker induction
axioms.
(2) The Σbi -IND axioms.
T2−1 has no induction axioms. T2 is the union of the T2i ’s.
T2 is equivalent to I∆0 + Ω1 (see Parikh [40] and Wilkie and Paris [50])
modulo differences in the nonlogical language.
Definition Let i ≥ 0. S2i is the first-order theory with language 0 , S , +,

·, b 21 xc , |x| , # and ≤ and axioms:
8
(1) The BASIC axioms, and
(2) The Σbi -PIND axioms.
S2−1 = T2−1 has no induction axioms. S2 is the union of the S2i ’s.
Remark: The theory S21 , which we will relate closely to polynomial
computability, is defined by PIND on NP properties (in light of Theorem 2).
The following, somewhat surprising, relationship holds between the

hierarchy of theories S2i and the hierarchy of theories T2i .
Theorem 3 (Buss [3, 4]). Let i ≥ 1 . T2i ` S2i and S2i ` T2i−1 . So S2 ≡ T2 .
Open Question: Are the following inclusions proper?
S21 ⊆ T21 ⊆ S22 ⊆ T22 ⊆ · · ·
2.4 Provably Recursive and Σbi -Definable Functions

We know come to one of the centrally important definitions for describing
the proof-theoretic complexity of bounded arithmetic:
Definition Let f : Nk 7→ N . The function f is Σbi -definable by a theory R
iff there is a formula A(~x, y) ∈ Σbi so that
(1) For all ~n ∈ Nk , A(~n, f (~n)) is true.
(2) R ` (∀~x)(∃y)A(~x, y)
(3) R ` (∀~x, y, z)(A(~x, y) ∧ A(~x, z) ⊃ y = z)
When a function is Σb1 -definable by a theory R , then we also say that the
function is provably recursive in R . By “provably recursive” we are intending
that there should be a Turing machine M which computes the function so
that R can prove the Turing machine always halts within polynomial time.
By a “bootstrapping procedure” it is possible to show that, for a given
Turing machine M and a given polynomial time bound p(n) , the statement
that M always halts within time p(|x|) can be expressed in the form
(∀x)(AM,p (x)) where AM,p is a Σb1 -formula. (Actually, this fact is immediate
from Theorem 2, so no further bootstrapping is needed.) On the other hand,
any Σb1 -definable function is certainly Turing computable; it can be computed
by a (non-polynomial time) brute-force search if nothing else. Therefore, it
makes sense to identify “provably recursive” and “ Σb1 -definable”.
The concept of Σbi -definability applies to functions only. The analogue
for predicates is the notion of ∆bi -definability:
Definition Let Q ⊆ N. Q is ∆bi -definable by a theory R iff there is
a Σbi -formula A and Πbi -formula B that define Q so that A and B are
provably equivalent in R . A formula is ∆bi with respect to R iff it is
provably equivalent to a Σbi - and to a Πbi -formula.
9
2.5 Bootstrapping Theorems
Theorem 4 [3, 4] Every polynomial time function is Σb1 -definable by S21
and every polynomial time predicate is ∆b1 -definable by S21 .
The above theorem shows that S21 , and the Σbi -definable functions and
∆b1 -definablepredicates are sufficiently strong to introduce polynomial time
properties. We will omit the proof of this theorem here: for details, the
reader can refer to Buss [3, 4], some improvements to the bootstrapping can
be found in Buss-Ignjatović [13] and a proof outline can be found in [7].
Alternative approaches to bootstrapping in different settings are given by
Wilkie-Paris [50] and in Hájek-Pudlák [25].
A large part of the importance of Σb1 -definable functions and ∆b1 -
predicates comes from the following theorem:
Theorem 5 [3, 4] Let i ≥ 1 . Any Σb1 -definable function or ∆b1 -definable

predicate of S2i may be introduced into the nonlogical language and used freely
in induction axioms. The same holds for T2i in place of S2i .
Proof (Sketch) Suppose f is Σb1 -defined by R , so that
R ` (∀x)(∃!y ≤ r(~x))Af (~x, y).
Then any atomic formula ϕ(f (~s)) is equivalent to both
(∃y ≤ r(~s))(Af (~s, y) ∧ ϕ(y))
and
(∀y ≤ r(~s))(Af (~s, y) ⊃ ϕ(y)).
Note that the first formula is in Σbi and the second is in Πbi . Thus, for
i odd, any Σbi -formula involving f is equivalent to one not involving f by
transforming atomic subformulas as above using the first equivalent formula
for positively occurring subformulas and the second equivalent formula for
negatively occurring subformulas. For i even, the roles of positive and
negative occurrences are reversed. (We have omitted some details from this
proof, since it is also necessary to remove f from terms in quantifier bounds;
this is possible because of the known bound r(~x) on the value of f (~x) .) 2
2.6 Main Theorems for S2i

The so-called “main theorems” for the theories S2i give an exact character-
ization of the Σbi -definable functions of S2i in terms of the computational
complexity.
10
Theorem 6 (Buss [3, 4]) Let i ≥ 1 . Let A be a Σbi -formula. Suppose
S2i ` (∀~x)(∃y)A(~x, y). Then there is a Σbi -formula B and a function f ∈ pi
and a term t so that
(1) S2i ` (∀~x, y)(B(~x, y) ⊃ A(~x, y)).

(2) S2i ` (∀~x)(∃!y)B(~x, y) .
(3) S2i ` (∀~x)(∃y ≤ t)B(~x, y) . (see Parikh [40])
(4) For all ~n , N |= B(~n, f (~n)).
p
Theorem 7 If f ∈ i then there is a formula B ∈ Σbi and a term t so that
(2), (3) and (4) hold.
Corollary 8 ( i ≥ 1) The Σbi -definable functions of S2i are precisely the

functions in pi .
The most interesting case of the above theorems and corollary is probably
the i = 1 case. For this, we have:
Corollary 9 The Σb1 -definable functions of S21 are precisely the polynomial
time functions.
It is this corollary that allows us to state that the proof-theoretic strength

of S21 corresponds to polynomial time computability.
The above theorems characterize the Σbi -definable functions of S2i . These
can be restated to characterize the ∆bi -definable predicates of S2i .
Definition A predicate Q(~x) is ∆bi -definable in a theory T provided it is

defined by a Σbi -formula A(~x) and a Πbi -formula B(~x) such that T proves
(∀~x)(A(~x) ↔ B(~x)).
Theorem 10 ( i ≥ 1). Suppose A(~x) ∈ Σbi and B(~x) ∈ Πbi and S2i ` A ↔
B . Then there is a predicate Q ∈ ∆pi so that, for all ~n ,
Q(~n) ⇔ N |= A(~n) ⇔ N |= B(~n)
Conversely, if Q ∈ ∆pi then there are A and B so that the above holds.
In other words, the ∆bi -definable predicates of S2i are precisely the ∆pi -
predicates.
The most interesting case of the last theorem is again probably the i = 1
case. For this, we have:
Corollary 11 If A is a formula which is S21 -provably in NP ∩ coNP then

A defines a polynomial time predicate (provably in S21 ). Being provably in
NP ∩ coNP means provably equivalent to a Σb1 - and to a Πb1 -formula.
11
We shall sketch a proof of the main theorem below; but first, we need to
take a diversion into the sequent calculus. The sequent calculus and the cut
elimination theorem will be one of the main tools in our proof of the main
theorems.
3 The Sequent Calculus

3.1 Gentzen’s Sequent Calculus
To prove the Main Theorem, we shall formalize S2i in Gentzen’s sequent
calculus.
We shall work in first-order logic with ∧ , ∨ , ¬, ⊃, ∀, ∃ the logical
symbols of our first language. In addition, there is one further symbol, → ,
which is the sequent connective. The sequent connective does not occur
in first-order formulas per se, but instead is used to mark the middle of a
sequent:
Definition A sequent is an expression of the form
A1 , A2 , . . . , An → B1 , B2 , . . . , Bk
where the Ai ’s and Bi ’s are formulas. Its intended meaning is
(A1 ∧ A2 ∧ · · · ∧ An ) ⊃ (B1 ∨ B2 ∨ · · · ∨ Bk )
We shall use Greek letters Γ, ∆, . . . to denote finite sequences of formulas

separated by commas. A sequent is thus denoted Γ → ∆ . The cedents Γ
and ∆ are called the antecedent and succedent of the sequent (respectively).
Gentzen’s sequent calculus is a proof system where lines in the proofs
are sequents. The sequent calculus is generally denoted LK , from “Logische
Kalkul”.
Definition An LK-proof is a tree of sequents: the leaves or initial sequents

must be of the form A → A ; the root, or endsequent, is what is proved; and
the valid inferences are:
Γ → ∆, A A, Γ → ∆
¬A, Γ → , ∆ Γ → ∆, ¬A
A, Γ → ∆ Γ → ∆, A Γ → ∆, B
A ∧ B, Γ → ∆ Γ → ∆, A ∧ B
B, Γ → ∆ Γ → ∆, A
A ∧ B, Γ → ∆ Γ → ∆, A ∨ B
A, Γ → ∆ B, Γ → ∆ Γ → ∆, B
A ∨ B, Γ → ∆ Γ → ∆, A ∨ B
12
Γ → ∆, A B, Γ → ∆ A, Γ → ∆, B
A ⊃ B, Γ → ∆ Γ → ∆, A ⊃ B
A(b), Γ → ∆ Γ → ∆, A(t)
(∃x)A(x), Γ → ∆ Γ → ∆, (∃x)A(x)
A(t), Γ → ∆ Γ → ∆, A(b)
(∀x)A(x), Γ → ∆ Γ → ∆, (∀x)A(x)
In the quantifier inferences, the free variable b is called the eigenvariable and
must not appear in the lower sequent.
Γ→ ∆ Γ→ ∆
A, Γ → ∆ Γ → ∆, A
Γ, A, B, Π → ∆ Γ → ∆, A, B, Λ
Γ, B, A, Π → ∆ Γ → ∆, B, A, Λ
A, A, Γ → ∆ Γ → ∆, A, A
A, Γ → ∆ Γ → ∆, A
Γ → ∆, A A, Π → Λ
Cut:
Γ, Π → ∆, Λ
The system LK , as defined above, gives a sound and complete proof
system for first-order logic. It is probably the most elegant way of
formulating first-order logic; and a primary factor in its elegance is the
following fundamental theorem:
Theorem 12 (Gentzen [23])
• LK is complete.
• LK without the Cut inference is complete.
In particular, if P is an LK-proof of Γ → ∆ then there is a cut-free proof
P ∗ of Γ → ∆. There is an effective (but not feasible) procedure to obtain
P ∗ from P .
Examination of the rules of inference for LK reveals that, with the
exception of the cut rule, every inference has the subformula property that
the formulas in the hypotheses of the inference are subformulas of formulas
in the conclusion (the lower sequent) of the inference. More precisely, every
formula in the upper sequents is obtained from a subformula of a formula
in the lower sequent, possibly after substitution of a term for a (formerly)
bound variable. In other words, every formula in an upper sequent is a
subformula in the wide sense of some formula in the lower sequent.
Thus, if there are no cuts, the logical complexity of formulas in the proof
will be at most the logical complexity of formulas in the endsequent Γ → ∆ .
We therefore say that cut-free sequent calculus proofs enjoy the subformula
property.
13
3.2 Sequent Calculus Formulations of Bounded
Arithmetic
We now wish to formulate a sequent calculus version of the theories of
bounded arithmetic in a way that preserves the essence of the this subformula
property. We there for enlarge LK as follows:
(1) Allow equality axioms and BASIC axioms as initial sequents. An initial
sequent will contain only atomic formulas.
(2) Add inferences for bounded quantifiers (the variable b occurs only as
indicated):
b ≤ s, A(b), Γ → ∆ Γ → ∆, A(t)
(∃x ≤ s)A(x), Γ → ∆ t ≤ s, Γ → ∆, (∃x ≤ s)A(x)
A(t), Γ → ∆ b ≤ s, Γ → ∆, A(b)
t ≤ s, (∀x ≤ s)A(x), Γ → ∆ Γ → ∆, (∀x ≤ s)A(x)
(3) Allow induction inferences: (for A ∈ Σbi )
A(b), Γ → ∆, A(b + 1)
Σbi -IND
A(0), Γ → ∆, A(t)
A(b 12 bc), Γ → ∆, A(b)

Σbi -PIND
A(0), Γ → ∆, A(t)
Definition S2i and T2i are formulated as sequent calculus systems with
BASIC axioms as initial sequents and with Σbi -PIND and Σbi -IND inference
rules, respectively. With side formulas, the induction inferences are
equivalent to the induction axioms.
3.3 Free-Cut Elimination

Definition A cut inference
Γ → ∆, A A, Π → Λ
Γ, Π → ∆, Λ
is free unless A is the direct descendant either of a formula in an initial

sequent or of a principal formula of an induction inference.
The next theorem is due to Gentzen and Takeuti (for fragments of Peano
arithmetic).
14
Theorem 13 Free-Cut Elimination Theorem If P is an S2i -proof (or
T2i -proof ) then there is a proof P ∗ in the same theory with the same endsequent
which contains no free cuts.
In a free-cut free proof, every formula will be a subformula (in the wide
sense) of an induction formula, of a formula in an axiom or of a formula in
the conclusion.
In S2i and T2i , cut formulas may be restricted to be Σbi -formulas.
Therefore, if Γ → ∆ is a sequent of Σbi -formulas which is a consequence
of S2i or T2i , then Γ → ∆ has an S2i -proof or T2i -proof (respectively) such
that every formula appearing in the proof is in Σbi .
3.4 Outline of Proof of the Main Theorem

There are two steps in the proof of Theorem 6
Step 1: By assumption, there is an S2i -proof P of
→ (∃y)A(~c, y).
Therefore, by free-cut elimination, there is an S2i proof P ∗ of
→ (∃y ≤ t)A(~c, y)
such that every formula in P ∗ is a Σbi -formula.
Step 2: Given the proof P ∗ we will extract an algorithm to compute
a function f (~c) such that A(~n, f (~n)) is true for all n . The function f
will be in pi and will be Σbi -defined by S2i . Furthermore, S2i will prove
(∀~x)A(~x, f (~x)). Thus, P ∗ can be thought of as a program plus a proof that
it is correct.
Step 1 is immediate from the free-cut elimination theorem. Before, we
can carry out Step 2, we need to introduce the Witness predicate.
3.5 The Witness Predicate

Definition Fix i ≥ 1. Let B(~a) be a Σbi -formula with all free variables
indicated. Then Witnessi,~
a
B (w, ~
a) is a formula defined inductively by:
(1) If B ∈ Σbi−1 ∪ Πbi−1 then Witnessi,~

a
a) ⇔ B(~a).
B (w, ~
(2) If B = C ∨ D then
Witnessi,~
a
a) ⇔ Witnessi,~
B (w, ~
a
a) ∨ Witnessi,~
C (β(1, w), ~
a
D (β(2, w), ~
a).
(3) If B = C ∧ D then
Witnessi,~
a
a) ⇔ Witnessi,~
B (w, ~
a
a) ∧ Witnessi,~
C (β(1, w), ~
a
D (β(2, w), ~
a).
15
(4) If B = (∃x ≤ t)C(~a, x) then
Witnessi,~
a
a) ⇔ β(1, w) ≤ t ∧ Witnessi,~
B (w, ~
a,b
C(~ a,b) (β(2, w), ~
a, β(1, w)).
(5) If B = (∀x ≤ |t|)C(~a, x) then
Witnessi,~
a
a) ⇔ (∀x ≤ |t|)Witnessi,~
B (w, ~
a,b
C(~ a,b) (β(x + 1, w), ~
a, x).
(6) If B = ¬C use prenex operations to push the negation sign inside.
Lemma 14 Fix i ≥ 1. Let B ∈ Σbi .

(1) For some term tB , S2i proves
B(~a) ↔ (∃w ≤ tB )Witnessi,~

a
B (w, ~
a).
(2) Witnessi,~
a p
B ∈ ∆i ( = P if i = 1).
(3) Witnessi,~
a b i
B is ∆i with respect to S2 .
Proof: This is easily proved by induction on the complexity of B .
3.6 The Main Lemma

Lemma 15 Suppose S2i ` Γ → ∆ where Γ and ∆ contain only Σbi -formulas.
Let ~c be the free variables in Γ and ∆ . Then, there is a function f such that
(a) f is Σbi -defined by S2i
(b) S2i ` Witnessi,~

VVc
c) ⊃ Witnessi,~
Γ (w, ~
WWc
∆ (f (w, ~
c), ~c)
p
(c) f ∈ i ( = F P if i = 1 )
Proof The proof is by induction on the number of inferences in a free-cut

free S2i -proof of Γ → ∆. As an example of one case of the proof of the main
lemma, suppose that P is a free-cut free proof and ends with the inference
B(b 12 ac) → B(a)

B(0) → B(t)
By the induction hypothesis, there is a function g so that

(1) g is Σbi -defined by S2i
p
(2) g is in i (= F P if i = 1 )
(3) S2i ` Witnessi,a,~c1 (w, a, ~c) ⊃ Witnessi,a,~

c
B(a) (g(w, a, ~
c), a, ~c).
B(b 2 ac)
16
(4) S2i ` (∀a, ~c)[g(w, a, ~c) ≤ tB (a, ~c)]
Now define f by limited iteration as
f (w, 0, ~c) = g(w, 0, ~c)

f (w, a, ~c) = g(f (w, b 12 ac, ~c), a, ~c)
so f (w, a, ~c) ≤ tB (a, ~c) and the following hold:

(1) f ∈ pi (= F P if i = 1 )
Pf: Since f is defined by limited iteration from g .
(2) S2i can Σbi -define f and prove that f satisfies the above conditions
(3) S2i ` Witnessi,a,~

c
B(0,~ c) ⊃ Witnessi,a,~
c) (w, a, ~
c
B(a,~ c) (f (w, a, ~
c), a, ~c).
Pf: Since Witnessi,b,~
c
B(a,~
b i b
c) is a Σi -formula, S2 can prove this by Σi -PIND
directly from the induction hypothesis.
Q.E.D. Main Lemma 15
3.7 Conclusion of Proof of Main Theorem

We can now prove the Main Theorem 6 for S2i from the above Lemma 15.
Proof Suppose S2i ` (∀~x)(∃y)A(~x, y) . By a theorem of Parikh, there is a

term t so that S2i proves → (∃y ≤ t)A(~c, y) . By the Main Lemma,
S2i ` Witnessi,~
c
(∃y≤t)A (g(~
c), ~c)
for some Σbi -defined function g . Define B(~c, y) to be the formula
y = β(1, g(~c)).
Since g is Σbi -defined by S2i , B is a Σbi -formula and by the properties of

Witness,
S2i ` (∀~x, y)(B(~x, y) ⊃ A(~x, y))
Finally, define f (~c) = β(1, g(~c)) .
Q.E.D. Main Theorem 6
4 Other Systems of Bounded Arithmetic

4.1 Alternative Axioms for Bounded Arithmetic
Let Ψ be a set of formulas. The axioms below are schemes where A ∈ Ψ :
17
Ψ-MIN: (Minimization)
(∃x)A(x) ⊃ (∃x)[A(x) ∧ (∀y < x)(¬A(y))]
Ψ-LMIN: (Length minimization)
(∃x)A(x) ⊃ A(0) ∨ (∃x)[A(x) ∧ (∀y ≤ b 21 xc)(¬A(y))]
Ψ-replacement:
(∀x ≤ |t|)(∃y ≤ s)A(x, y) ↔
↔ (∃w ≤ SqBd(t, s))(∀x ≤ |t|)(A(x, β(Sx, w)) ∧ β(Sx, w) ≤ s)
strong Ψ -replacement:
(∃w ≤ SqBd(t, s))(∀x ≤ |t|)
[(∃y ≤ s)A(x, y) ↔ A(x, β(Sx, w)) ∧ β(Sx, w) ≤ s]
The figure below shows the known relationships between fragments of

bounded arithmetic: (for i ≥ 1 , relative to S21 )
Σbi -IND ⇔ Πbi -IND ⇔ Σbi -MIN ⇔ ∆bi+1 -IND

w
w
Ä
Σbi -PIND ⇔ Πbi -PIND ⇔ Σbi -LIND ⇔ Πbi -LIND
~
w
Ä
Σbi -LMIN ⇔ strong Σbi -replacement
w ~
w w
Ä Ä
Σbi−1 -IND (Σbi+1 ∩ Πbi+1 )-PIND
Σbi+1 -MIN ⇔ Πbi -MIN
Σbi+1 -replacement ⇒ Σbi -PIND ⇒ Σbi -replacement
S2i+1 Â T2i
Σbi+1
S2i+1 Â T2i + Σbi+1 -replacement

B(Σbi+1 )
Due to space limitations, we shall only sketch proofs of two of the facts
pictured in the above figure.
Theorem 16 [3, 4] S21 + Σbi -P IN D ` ∆bi -IND . Hence S2i ⊃ T2i−1 .
18
Proof Suppose A is ∆bi w.r.t. S2i . Assume (∀x)(A(x) ⊃ A(x + 1)) and
argue inside S2i . Let B(x, z) be the formula
(∀w ≤ x)(∀y ≤ z + 1)(A(w −

.
y) ⊃ A(w)).
So B is equivalent to a Πbi -formula. Now by definition of B ,

(∀x, z)(B(x, b 12 zc) ⊃ B(x, z)) and hence by Πbi -PIND on B(x, z) w.r.t. z ,
(∀x)(B(x, 0) ⊃ B(x, x)).
By the assumption, (∀x)B(x, 0) ; hence (∀x)B(x, x) , from whence
(∀x)(A(0) ⊃ A(x)) 2
The second result is concerns the conservation results between T2i and
S2i+1 :
Theorem 17 [8] Fix i ≥ 1. S2i+1 is conservative over T2i with respect to

Σbi+1 -formulas, and hence with respect to ∀∃Σbi+1 -sentences.
This means that any ∀∃Σbi+1 -formula which is S2i+1 -provable is also
T2i -provable.
Proof (Idea). Fix i ≥ 1 and let Z be P V or T2i−1 as appropriate. First

show that every pi -function is definable in Z in an appropriate sense.
For i = 1, there is a function symbol for every polynomial time function;
for i > 1, we show that every pi -function can be “Qi -defined” — this is
stronger than “Σbi -defined”. Second, prove a stronger version of the Main
Lemma above; in essence, we partially formalize the Main Lemma in Z and
prove that the witnessing function f is defined appropriately in Z . Namely,
we prove that if S2i ` A with A ∈ Σbi then Z ` A .
We omit the rest of the proof details.
A nice consequence of the previous theorem and of the Main Theorem

for S2i is:
p
Theorem 18 [8] The Σbi+1 -definable theories of T2i are precisely the i+1 -
functions.
4.2 Witnessing Theorem for T21

The class Polynomial Local Search, PLS, was defined by Papadimitriou [39]
to capture a common kind of search problem. Two representative examples
of PLS problems are (1) linear programming and (2) simulated annealing.
Of course, it is known that there are polynomial time algorithms for linear
programming. On the other hand, there is no polynomial time algorithm
which is known to find even a local optimum solution based on simulated
annealing. So, it is open whether PLS is in P .
19
Definition A Polynomial Local Search (PLS), problem is specified by two
polynomial time functions N, c and a polynomial time predicate F which
satisfy the following conditions:
(1) c(s, x) is a cost function,
(2) N (s, x) is a neighborhood function, such that for all s such that F (s, x)
holds, we have
c(N (s, x), x) ≤ c(s, x) and F (N (s, x), x).
(3) {s : F (s, x)} is the solution space for input x. F (0, x) always holds;
and if F (s, x), then |s| < p(|x|) for p some polynomial.
A solution to the PLS problem is a (multivalued) function f , s.t., for all x,
c(N (f (x), s), s) = c(f (x), x) and F (f (x), x).
Theorem 19 [14] Suppose T21
proves (∀x)(∃y)A(x, y) where A ∈ Σb1 . Then
there is a PLS function f (x) = y and a polynomial time function π such that
T21 ` (∀x)A(x, π ◦ f (x)).
Furthermore, every PLS function (and every function π ◦ f ) is Σb1 -definable
by T21 .
Corollary 20 The same holds for S22 by conservativity of S22 over T21 .
Proof-idea: A free-cut free T21 -proof can be transformed into a PLS
problem. 2
4.3 Herbrand’s Theorem and the KPT Witnessing

Theorem
The following theorem is a version of Herbrand’s Theorem from Herbrand’s
dissertation [27, 30, 28]. A proof of (a strengthening of) this version of
Herbrand’s theorem can be found in [10].
Theorem 21 Let T be a theory axiomatized by purely universal axioms. Let
A(x, y, z) be quantifier-free. Suppose T proves
(∀x)(∃y)(∀z)A(x, y, z).
Then there is an integer k > 0 and there are terms t1 (x) , t2 (x, z1 ) ,
t3 (x, z1 , z2 ), . . . , tk (x, z1 , . . . , zk−1 ) so that:
T ` (∀x)[(∀z1 )[A(x, t1 (x), z1 )∨
(∀z2 )[A(x, t2 (x, z1 ), z2 ) ∨
(∀z3 )[A(x, t3 (x, z1 , z2 ), z3 ) ∨
..
.
(∀zk )[A(x, tk (x, z1 , . . . , zk−1 ), zk )] · · ·]]].
20
The KPT Witnessing Theorem (Theorem 22 below) applies this theorem
to the theories T2i . In order to do this, however, we must re-axiomatize T2i
to have purely universal axioms. In order to give a set of purely universal
axioms, we must enlarge the language of T2i by adding function symbols all
p
i+1 -functions. Of course, we want to have only a conservative extension
of T2i when we enlarge the language: By Theorem 18, T2i can already
Σbi+1 -define all bi+1 functions; therefore, we can add symbols for these
functions to the language of T2i and add the defining equations for these new
function symbols and still have a conservative extension of T2i .
Towards this end, we define:
Definition Let i ≥ 1. The theory P Vi = T2i ( pi+1 ) is defined to be

the theory T2i with language enlarged to include function symbols for all
p
i+1 -functions. In addition, P Vi includes the axioms that define these
p
i+1 -functions.
Using these new function symbols as Skolem functions, any Σbi -formula
of T2i is equivalent to a quantifier-free formula. The induction axioms
A(0) ∧ (∀x)(A(x) ⊃ A(x + 1)) ⊃ (∀x)A(x)
of T2i can be replaced by
A(0) ∧ ¬A(c) ⊃ A(fA (c)) ∧ ¬A(fA (c) + 1)

p
where fA is the i+1 -function such that
(
least x < c such that A(x) and ¬A(x + 1)
fA (c) =
0 if no such x < c exists
It therefore be shown that P Vi can be axiomatized by purely universal

formulas.
It is also possible to define P V1 as being the natural first-order, purely
universal theory which has function symbols from all polynomial time
functions (this is the same as the theory P V of Cook [16], except extended
to first-order logic).
Using P Vi in place of T2i , and applying Herbrand’s theorem to P Vi , one
obtains the following witnessing theorem.
Theorem 22 (Krajı́ček-Pudlák-Takeuti [35]) Suppose A ∈ Σbi+2 and T2i

proves (∀x)(∃y)(∀z)A(x, y, z). Then there are k > 0 and functions
fi (x, z1 , ..., zi−1 ) so that
(1) Each fi is Σbi+1 -defined by T2i .
21
(2) T2i proves
(∀x)[(∀z1 )[A(x, f1 (x), z1 )∨

(∀z2 )[A(x, f2 (x, z1 ), z2 ) ∨
(∀z3 )[A(x, f3 (x, z1 , z2 ), z3 ) ∨
..
.
(∀zk )[A(x, fk (x, z1 , . . . , zk−1 ), zk )] · · ·]]].
4.4 Collapse of the Polynomial Hierarchy

As mentioned earlier, it is open whether the hierarchy of theories of bounded
arithmetic is proper. Because of the close relationship between the fragments
of bounded arithmetic and the levels of the polynomial time hierarchy, it
is natural to try to find a connection between the questions of whether
these two hierarchies are proper. This is answered by the following theorem,
which shows that the hierarchy of theories of bounded arithmetic collapses
if and only if bounded arithmetic is able to prove that the polynomial time
hierarchy is proper.
Theorem 23 (Krajı́ček-Pudlák-Takeuti [35], Buss [11], Zambella [52])

If T2i = S2i+1 , then the polynomial time hierarchy collapses, provably in T2i .
In fact, in this case, T2i proves that every Σpi+3 predicate is (a) equivalent to
a Boolean combination of Σpi+2 -predicates and (b) is in Σpi+1 /poly .
Proof (Idea) For simplicity, assume i = 0 . Suppose T20 (P V ) = S21 . Let

ϕ represent a vector of Boolean formula ϕ = hϕ1 , . . . , ϕn i . Then T20 (P V )
proves
∀ϕ(∃` ≤ n)(∃hw1 , . . . , w` i)
[(∀j ≤ `)(wj satisfies ϕj ) ∧ “` = n or ϕ`+1 is unsatisfiable”]
The formula in [· · ·] is in Πb1 , so the KPT witnessing theorem can be applied

to get k > 0 and polynomial time functions f1 , . . . , fk so that T20 (P V )
proves (setting n = k ) that given ϕ1 , . . . , ϕk satisfied by w1 , . . . wk , that
one of fj (ϕ, w1 , . . . , wj−1 ) produces a witness to ϕj . [Note that fj has all
ϕi ’s as input.]
Let P reAdvice(a, hϕ`+1 , . . . , ϕk i) mean that for all ϕ1 , . . . , ϕ` < a
satisfied by w1 , . . . , w` , fj (ϕ, w1 , . . . , wj−1 ) satisfies ϕj for some j ≤ ` .
Let Advice(a, hϕ`+1 , . . . , ϕk i) mean that P reAdvice holds, and that ` is the
minimum possible value for which there is such PreAdvice.
Claim T20 (P V ) proves, that if ϕ` < a and if Advice(a, hϕ`+1 , . . . , ϕk i) ,

then ϕ` is satisfiable if and only if for all ϕ1 , . . . , ϕ` , satisfied by w1 , . . . , w` ,
there is j ≤ ` such that fj (ϕ, w1 , . . . , wj−1 ) satisfies ϕj .
22
Proof of claim: If the latter condition is true, then the only way for
hϕ` , . . . , ϕk i to not be “preadvice”, (which it isn’t, by definition of “advice”)
is for ϕ` to be satisfied by f` (ϕ, w) ~ for some ϕ1 , . . . , ϕ`−1 , w1 , . . . , w`−1 . 2
Note that this means that the NP complete property of satisfiability is

in coNP relative to the polynomial size advice, hϕ` , . . . , ϕk i .
The above shows that T20 (P V ) would prove NP ⊆ coNP/poly . From this,
Karp-Lipton style methods can show that T20 (P V ) proves the polynomial
time hierarchy collapses. In fact it can be shown that T20 (P V ) proves
that every polynomial time hierarchy predicate is equivalent to Boolean
combination of Σp2 predicates. The proof idea is that the property P reAdvice
is in coNP and therefore, property
PAlen(`) ≡ ∃hϕ`+1 , . . . , ϕk iP reAdvice(a, h~

ϕi)
is a Σp2 -property.
Similar methods work for i ≥ 1 . Q.E.D.
5 Lengths of Propositional Proofs

5.1 Frege and Extended Frege Systems
Definition Propositional Formulas are formed with logical connectives ∧ ,
∨ , ¬ and ⊃, variables p1 , p2 , . . . , and parentheses.
Cook’s Theorem: P = N P if and only if there is a polynomial time

algorithm for determining if a propositional formula is valid.
Frege systems are the usual “textbook” proof systems for propositional
logic:
Definition A Frege (F ) proof system is a usual proof system for

propositional logic with a finite set of axiom schemes and with only the
modus ponens rule. F is sound and complete.
Open Question: Does every tautology have a polynomial size F -proof?
If the answer to this question is “Yes”, then NP = coNP. This is since the
set of tautologies is coNP complete and having a polynomial size F -proof is
an NP property.
The size of Frege proofs is measured in terms of the number of symbols
occurring in the proof. Of course, the number of symbols may be large either
because there are a lot of steps in the proof, or because the formulas which
occur in the proof are large. In the latter case, there will generally be many
repeated subformulas; this motivates the definition of “extended Frege proof
systems” where repeated subformulas may be abbreviated by new symbols.
23
Definition The extended Frege (eF ) proof system is a Frege proof system
plus the extension rule:
Extension Rule: whenever q is a variable which has not been used in the
proof so far and does not appear in the final line of the proof or in ϕ then
we may infer
q ↔ ϕ.
This allows us to use q as abbreviation for ϕ. By iterating uses of extension
rule the extension rule can apparently make proofs logarithmically smaller
by reducing the formula size.
Tseı̌tin [49] first used the extension rule, for resolution proofs. See also
Statman [47] and Cook-Reckhow [18, 19]. Statman proved that the number
of symbols in an extended Frege proof can be polynomially bounded in terms
on the number of steps in the proof.
Theorem 24 Reckhow [45] The choice of axiom schemas or of logical

language does not affect the lengths of F - or eF -proofs by more than a
polynomial amount.
5.2 Abstract Proof Systems

The following definition of propositional proof system is due to Cook [16]:
Definition A propositional proof system is a polynomial time function f

with range equal to the set of all valid formulas.
An (extended) Frege proof system can be viewed as a propositional proof

system by letting f (w) equal the last line of w if w is a valid (e)F -proof.
Similarly, any theory (e.g. set theory) can be viewed as a propositional proof
system. For instance, ZF serves a proof system by letting f (w) be defined
so that if w codes a ZF -proof that ϕ is a tautology, then f (w) equals (the
code for) ϕ.
Theorem 25 (Cook [16]) NP = coNP if and only if there is a proof system f

for which tautologies have polynomial size proofs.
To prove this theorem, note that if f is such a proof system, then then
a formula is a tautology iff it has a proof of polynomial length. Therefore,
the coNP complete property of being a tautology would be in NP , so NP
would equal coNP . Conversely, an NP algorithm can be turned into a proof
system, by letting its accepting computations serve as proofs.
Definition A proof system f in which all tautologies have polynomial size

proof is called super.
It is open whether there exist super proof systems.
24
Definition Let S and T be proof systems (with the same propositional
language). S simulates T iff there is a polynomial p so that for any
T -proof of size n there is an S -proof of the same formula of size ≤ p(n) . S
p-simulates T iff the S -proof is obtainable as a polynomial time function of
the T -proof.
Open Question: Does F simulate eF ?
This open question is related to the question of whether Boolean circuits

have equivalent polynomial size formulas. By Ladner [36] and Buss [5] this
is a non-uniform version of the open question of whether P is equal to
alternating logarithmic time (ALOGTIME).
Open Question: Is there a maximal proof system which simulates all other
propositional proof systems?
Krajı́ček and Pudlák [33] have shown that if NEXP (non-deterministic

exponential time) is closed under complements then the answer is “Yes”.
5.3 The Propositional Pigeonhole Principle

We now introduce tautologies that express the pigeonhole principle.
Definition The propositional pigeonhole principle P HPn is the formula

^ _ _ _
pi,j ⊃ (pi,j ∧ pm,j )
0≤i≤n 0≤j<n 0≤i<m≤n 0≤j<n
states that n + 1 pigeons can’t fit singly into n holes. pi,j means “pigeon i
is in hole j ”.
Theorem 26 (Cook-Reckhow [18, 19]) There are polynomial size eF -proofs

of P HPn .
Theorem 27 ([6]) There are polynomial size F -proofs of P HPn .
Theorem 28 (Haken [26]) The shortest resolution proofs of P HPn are of

exponential size.
Cook and Reckhow had proposed P HPn as an example for showing that
F could not simulate eF . However, Theorem 27 implies this is not the case.
Presently there are no very good candidates of combinatorial principles that
might separate F from eF ; the paper [2] describes some attempts to find
such combinatorial principles.
25
5.4 P HPn has Polysize eF -Proofs
The pigeonhole principle provides a nice example of how extended Frege
(eF ) proofs can grow exponentially larger when translated straightforwardly
into Frege (F ) proofs. Although it is known (Theorem 27) that PHP
has polynomial size Frege proofs, the proofs are rather different than the
extended Frege proofs below.
We first give an intuitive, conceptual version of a proof of the pigeonhole
principle and then describe how it is formalized as an extended Frege proof.
Let [n] = {0, 1, 2, . . . , n}.
Conceptual Version: (by contradiction)

1−1
Given f : [n] 7−→ [n − 1]
1−1
define fk : [k] 7−→ [k − 1]
as fn (i) = f (i) ½,
fk+1 (i) if fk+1 (i) < k
fk (i) =
fk+1 (k + 1) otherwise.
1−1
For k = 1, f1 : [1] 7−→ [0] — contradiction.
k
The eF -proof: Uses qi,j for “fk (i) = j ”.
n
qi,j ↔ pi,j
k
qi,j ↔ qi,j
k+1
∨ (qi,k
k+1
∧ qk+1,j
k+1
)
Then prove for k = n, n − 1, . . . , 1 that qi,j
k
’s code a one-to-one function
from [k] to [k − 1]. For k = 1 , we have “f1 is total and one-to-one”:
1 1 1 1
q0,0 ∧ q1,0 ∧ ¬(q0,0 ∧ q1,0 )
which is impossible. 2
It a useful exercise to see what happens when the above polynomial
size extended Frege proofs are translated into Frege proofs by unwinding
n k
the definitions of the abbreviations qi,j . Since qi,j abbreviates a formula
k+1 1
involving three occurrences of variables qi,j , the variable qi,j abbreviates a
n−1
formula containing 3 occurrences of symbols.
6 Translations from Bounded Arithmetic into

Propositional Logic
There are two important translations of proofs in fragments of bounded
arithmetic into proofs in propositional logic. The first, due to Cook [16], is
a translation from PV-proofs (or more-or-less equivalently, from S21 -proofs)
into polynomial size extended Frege proofs. The second, due to Paris and
Wilkie [41], is a translation from proofs in the theory I∆0 or I∆0 + Ω − 1
or S2 = T2 into constant depth, polynomial size Frege proofs.
26
6.1 S21 and Polysize eF Proofs
P V is an equational theory of polynomial time functions; above, we
discussed P V1 = T20 ( p1 ) which is the conservative first-order extension
of P V . P V was first introduced by Cook [16], who showed that there is
an intimate translation between P V -proofs and polynomial size eF -proofs.
Namely, he showed that if A(x) is a polynomial time equation provable in
n
PV, then there is a family of tautologies A such that
n
(1) A is a polynomial size propositional formula,
n
(2) A says that A(x) is true whenever |x| ≤ n,
n
(3) A has polynomial size eF -proofs.
(Some generalizations of these results have been proved by Dowd [21] for
PSPACE and Krajı́ček and Pudlak [34] for various of the theories of bounded
arithmetic.)
In these notes, we shall prove the version of Cook’s theorem for S21
and Πb2 -formulas A. (Our proof below follows the version in [12]). This
version is completely analogous to Cook’s original theorem, in view of the
∀Σb1 -conservativity of S21 over P V1 .
Definition Let t(~a) be a term. The bounding polynomial of t is a polynomial

qt (~n) such that
(∀~x)(|t(~x)| ≤ qt (max{|~x|})).
The inductive definition is:
q0 (n) = 1
qa (n) = n for a a variable
qS(t) (n) = qt (n) + 1
qs+t (n) = qs (n) + qt (n)
qs·t (n) = qs (n) + qt (n)
qs#t (n) = qs (n) · qt (n) + 1
q|t| (n) = q 1 (n) = qt (n)
b 2 tc
Definition Let A(~a) be a bounded formula. The bounding polynomial

of A is a polynomial qA (~a) so that if |ai | ≤ n for all ai in ~a , then A(~a)
refers only to numbers of length ≤ qA (n) . The polynomial qA is inductively
defined by:
(1) qs≤t = qs=t = qs + qt
(2) q¬A = qA
(3) qA∧B = qA∨B = qA⊃B = qA + qB
(4) q(Qx≤t)A = qt (n) + qA (n + qt (n))
27
Next we define t m to be a vector of polynomial size formulas that
define (compute) the term t for values of length ≤ m . For this it is useful
to think of formulas as being circuits of fanout 1.
Definition Let + m be a polynomial size, fanout 1 circuit which accepts

2m binary inputs and outputs m binary signals; + m computes the bitwise
sum of two m -bit integers (and discards any overflow). Likewise define · m ,
# m , b 12 xc m , etc.
n
Definition Let t(~a) be a term and m ≥ qt (n) . t m is a vector of
m propositional formulas defining the lower m bits of the value of t(~a) when
|ai | ≤ n.
For b a free variable in t , a propositional variable vib represents the i-th bit
of b ’s value.
n
(1) 0 m is a sequence of m false formulas (for example p ∧ ¬p).
n
(2) For b a variable, b m
is a sequence of m − n false formulas followed by
b
vn−1 , . . . , v0b .
n n n
(3) s+t m is + m( s m, t m) (the formulas corresponding to the circuit
n n
for addition applied to the outputs of s m and t m ).
(4) And similarly for other cases.

n
Note that t m is a polynomial size formula (in m and n).
n
For A ∈ Πb2 , we define below a propositional formula A m for m ≥
qA (n). If B is a formula, we assign new ‘existential’ variables ²Bi and new
‘universal’ variables µB
i to B ( i ≥ 0 ). Different occurrences of B will
generally get assigned different such variables.
Definition EQm is a circuit for equality:
^
m−1
EQm (~
p, ~q) is (pk ↔ qk ).
k=0
p, ~q) is a circuit for ≤ :

LEm (~
Ã !
_ ^
p, ~q) ∨
EQm (~ qi ∧ ¬pi ∧ (qi ↔ pi ) .
0≤i<m i<j<m
Definition A is in negation-implication normal form (NINF) iff all negation

signs are applied to atomic subformulas and there are no implications in A .
28
Definition Assume A ∈ Πb2 and A is in NINF and m ≥ qA (n) . Define
n
A m inductively by:
n n n
(1) s=t m is EQm ( s m, t m)
n n n
(2) s≤t m
is LEm ( s m
, t m
)
n n
(3) ¬A m is ¬ A m for A atomic.
n n n
(4) A ∧ B m
is A m
∧ B m
n n n
(5) A∨B m is A m ∨ B m
n n
(6) (∃x ≤ t)A(x) m is x ≤ t ∧ A(x) m ({εi /vi }i=0 )
A x n−1
n n
(7) (∀x ≤ t)A(x) m
is ¬x ≤ t ∨ A(x) m i /vi }i=0 )
({µA x n−1
n V
m−1 n
(8) (∀x ≤ |t|)A(x) m
is ¬k ≤ |t| ∨ A(k) m
Note that |t| ≤ m (by our
k=0
assumption on m).
n W
m−1 n
(9) (∃x ≤ |t|)A(x) m is k ≤ |t| ∧ A(k) m
k=0
n
Proposition 29 The formula A m
is equivalent to A in that A(~a) is true
n
( |ai | ≤ n) iff for all truth assignments to the universal variables in A m
n
there is an assignment to the existential variables which satisfies A m.
We can extend the definition of A in the obvious way to formulas not

in NINF, by using prenex operations to transform A into NINF form.
n
Definition An eF -proof of A m is defined like an ordinary eF -proof
except now we additionally allow the existential variables (but not the other
n
variables) in A m to be defined by the extension rule (each existential
variable may be defined only once).
Theorem 30 (essentially Cook [16]). If A ∈ Πb2 and S21 ` (∀~x)A(~x) then

n
there are polynomial size (in n ) eF -proofs of A qA (n) . These eF -proofs
are obtainable in polynomial time.
Proof If Γ → ∆ is provable in S21 , we prove the theorem for
¬Γ ∨ ∆ .
By free-cut elimination it will suffice to do it for Γ ⊂ Σb1 and ∆ ⊂ Πb2 . We

proceed by induction on the number of inferences in a free-cut free proof.
29
Case (1): A logical axiom B → B . Obviously
¬B ∨ B = ¬B ∨ B
has a polynomial size eF -proof.

Case (2): A BASIC axiom. For example,
n
(x + y) + z = x + (y + z) 3n
has straightforward polynomial size F -proofs using techniques from [6] that
allow formalization of addition and multiplication in Frege proofs.
Case (3) The proof ends with a contraction:
Γ→∆, B, B
Γ→∆, B
Recall that all three B ’s are assigned different existential and universal
variables. The induction hypothesis says there are polynomial size eF -proofs
of
¬Γ ∨ ∆ ∨ B ∨ B .
Modify these proofs by (1) identifying the universal variables for different
B ’s; (2) at the end of the proof use extension to define
²00j ↔ ( B (~²) ∧ ²j ) ∨ (¬ B (~²) ∧ ²0j )
where ~²00 are the existential variables for the lower B and the others are the
existential variables for the upper B ’s; and (3) then extend to a proof of
¬Γ ∨ ∆ ∨ B (~²00 ).
Case (4) The proof ends with a Cut:
Γ → ∆, B B, Π → Λ
Γ, Π → ∆, Λ
By free cut elimination, B ∈ Σb1 ; so B has existential variables ~² and ¬B

has universal variables µ
~ . By induction hypothesis, there are polynomial
size eF -proofs of
¬Γ ∨ ∆ ∨ B (~²)
and
¬Π ∨ Λ ∨ ¬B (~
µ).
The polynomial size eF -proof of
¬Γ ∨ ¬Π ∨ ∆ ∨ Λ
30
consists of the first proof above followed by the second proof except with the
µ
~ ’s changed to ~² ’s followed by a (simulated) cut.
Case (5) For Σb1 -PIND inferences, iterate the construction for Cut and
contractions.
Case (6) If the proof ends with:
Γ→∆, A(t)
t ≤ s, Γ→∆, (∃x ≤ s)A(x)
Let ~² be the existential variables for (∃x ≤ s)A . The desired eF -proof
contains:
(a) Extension: ~² ↔ t .
(b) The proof from the induction hypothesis of
¬Γ ∨ ∆ ∨ A(t) .
(c) A further derivation of
¬t ≤ s ∨ ¬Γ ∨ ∆ ∨ (t ≤ s ∧ A(t)) .
(d) A derivation of
¬t ≤ s ∨ ¬Γ ∨ ∆ ∨ (∃x ≤ s)A(x)
by changing some t ’s to ² ’s. 2
6.2 Consequences of the S21 and eF Correspondence

Theorems 31-33 are due to Cook [16], although he stated them for P V
instead of for S21 .
Theorem 31 Let G ⊇ F be a propositional proof system. If S21 ` Con(G)

then eF p-simulates G.
Theorem 32 If S21 ` NP =coNP then eF is super.
Theorem 33 eF has polynomial size proofs of the propositional formulas

ConeF (n) which assert that there is no eF -proof of p ∧ ¬p of length ≤ n.
Theorem 34 [9]. F has polynomial size proofs of the self-consistency

formulas ConF (n).
31
Proof of Theorem 31 from Theorem 33: (Idea) Suppose there is a G proof
P of a tautology ϕ. A polynomial size eF proof of ϕ is constructed as
follows: Let p~ be the free variables in ϕ(~
p) . Reason inside eF . First show
that if ¬ϕ then there is an F -proof P1 of ¬ϕ(~ p) where p~ denotes a vector
of >’s and ⊥’s: the truth values of p~ . By substituting p~ for p~ in P and
combining this with P1 , we construct a G-proof P2 of a contradiction. This
proof has size polynomial in |P | since P1 has size polynomial in |ϕ| ≤ |P |.
By Theorem 33 there is a polynomial size eF -proof of ConG (|P2 |) so the
assumption that ¬ϕ is impossible; i.e., ϕ is true. 2
Proof of Theorem 33: S21 ` Con(eF) . 2
Proof of Theorem 34: The F -self-consistency proof is a “brute-force” proof

that truth is preserved by axioms and modus ponens using the fact that the
Boolean formula value problem is in ALOGTIME. 2
Definition A substitution Frege sF proof system is a Frege proof system

plus the substitution rule:
ψ(p)
ψ(ϕ)
for ψ , ϕ arbitrary formulas, all occurrences of p substituted for.
Theorem 35 (Cook-Reckhow [19], Dowd [22], Krajı́ček and Pudlák [33])

sF and eF p-simulate each other.
Proof (Idea) sF p-simulates eF is not hard to show directly. eF p-

simulates sF since S21 ` Con(sF) . 2
6.3 Constant Depth Frege Proofs

Let propositional formulas use connectives ∧ and ∨ with negations only
on variables. The depth of a formula is the maximum number of blocks
(alternations) of ∧ ’s and ∨ ’s on any branch of the formula, viewed as a tree.
The depth of a Frege proof is the maximum depth of formulas occurring in
the proof.
Theorem 36 Constant-depth Frege systems are complete (for constant depth

tautologies).
Proof By the cut-elimination theorem. 2
32
6.4 Translation from I∆0 /S2 into Constant Depth Frege
Paris and Wilkie [41] developed the following translation between provability
in I∆0 (or I∆0 + Ω1 ) and the lengths of constant depth Frege proofs. The
theory I∆0 + Ω1 is essentially identical to the theory S2 = T2 — the only
difference is the choice of first-order language: I∆0 uses the first-order
language containing the symbols 0, S, +, · and ≤. Actually, we shall work
with I∆0 (α, f ) or S2 (α, f ) where α and/or f are allowed to be new
predicate or function symbols (resp.) which may be used in induction
axioms.
Definition We translate closed (=variable-free) arithmetic formulas A into

propositional formulas APW : this is defined inductively as follows.
(1) (α(t))PW is the formula qi , where i is the numeric value of the variable-
free term t .
(2) (f (t) = s)PW is the formula pi,j , where i and j are the numeric values
of t and s. Without loss of generality, f occurs only in this context.
(3) For other atomic formulas, P (~t)PW is defined to be either the constant
> or the constant ⊥ .
(4) Boolean connectives are translated without any change. E.g., (A ∧ B)PW
is APW ∧ B PW .
PW
^
^
value(t)
PW
(5) [(∀x ≤ t)A(x)] is [A(i)] .
i=0
PW
_
_
value(t)
PW
(6) [(∃x ≤ t)A(x)] is [A(i)] .
i=0
Theorem 37 (Paris-Wilkie [41]) Suppose I∆0 (α, f ) proves (∀x)A(x).

Then the formulas {A(n)PW : n ≥ 0} are tautologies and have polynomial
size, constant-depth Frege proofs.
Proof (Idea) Given an I∆0 (α, f ) proof P (x) of A(x) and given n ≥ 0 ,
replace x everywhere with n , to get a proof P (n) of A(n) . W.l.o.g., P (x)
is free-cut free, so has only bounded formulas. Replace every formula B
in P (n) with its translation B PW . Thus every sequent Γ → ∆ in P (n)
becomes a propositional sequent ΓPW → ∆PW .
(a) Size of new formulas. A simple size analysis gives that there is a
constant c such that for every formula A ∈ P (n) , the formula APW as at
most nc many symbols. This is since every term t(n) is bounded by nc and
there are finitely many formulas A in P (n) .
(b) Size of propositional proofs of ΓPW → ∆PW is likewise bounded by nd
for some constant d . To prove this, consider how the propositional proof is
33
obtained from the proof P (n) : the general idea is to work from the bottom of
the proof upwards, always considering sequents in P (n) with values assigned
to all the free variables.
(b.i) A ∃ ≤:right inference in P (n) :
Γ → ∆, B(s)
s ≤ t, Γ → ∆, (∃x ≤ t)B(x)
If s ≤ t , the propositional translation of this is:
ΓPW → ∆PW , B(s)PW

∨:right’s
W
W
t
ΓPW → ∆PW , B(i)PW
i=0
W
W
t
>, ΓPW → ∆PW , B(i)PW
i=0
(b.ii) A ∀ ≤:right inference in P (n) :
a ≤ t, Γ → ∆, B(a)
Γ → ∆, (∀x ≤ t)B(x)
has propositional translation:
{>, ΓPW →∆ PW
, B(i)PW }i=0
t
∧:right’s
V
V
t
Γ PW
→∆ PW
, B(i) PW
i=0
(b.iii) A induction inference in P (n) :
Γ, B(a) → B(a + 1), ∆

Γ, B(0) → B(t), ∆
has propositional translation
{ΓPW , B(i)PW → B(i + 1) PW t−1

, ∆PW }i=0
Cuts
ΓPW , B(0)PW → B(t) , ∆ PW PW
Other inferences are handled similarly. Since the proof P (n) has constant
size, and since the values of terms are ≤ nα , for some constant α , the size
bound is proved. 2
34
When Ω1 is present as an axiom, then the function x 7→ xlog x is total. In
this case, an argument similar to the above establishes the following theorem:
Theorem 38 (Paris-Wilkie [41]) Suppose I∆0 (α, f ) + Ω1 proves (∀x)A(x).

Then the formulas {A(n)PW : n ≥ 0} are tautologies and have quasi-
polynomial size, constant-depth Frege proofs.
The above two theorems can be sharpened in the case of S2i - and
T2i -proofs. First, at the cost of adding a finite set polynomial time functions
such as the Gödel β function, we may assume that every formula in Σbi (α, f )
or Πbi (α, f ) consists of exactly i bounded quantifiers, then a sharply bounded
quantifier and then a Boolean combination of atomic formulas of the form
α(t) or f (t) = s or which do not use α or f . [Basically, because of the
quantifier exchange property and by contracting like quantifiers.] With this
convention, then if A ∈ Σbi or A ∈ Πbi then the translation APW is a depth
i + 1 propositional formula where the bottom depth has polylogarithmic
fanin. This gives the following theorem:
Theorem 39 Suppose T2i ` Γ → ∆ , sequent of Σbi ∪ Πbi formulas. Then

the sequents ΓPW → ∆PW have polynomial size propositional sequent calculus
proofs of depth i + 1 in which every formula has polylogarithmic fanin at the
bottom level.
Furthermore, there is a constant c such that every sequent in the
propositional proof has at most c formulas.
If every formula in the T2i -proof is in Πbi , then
V every formula in the
propositional proofs starts with a (topmost) block of ’s.
Proof Analogous to the above proof. We leave the details to the reader.
7 Interpolation Theorems for Propositional

Logic
7.1 Craig’s Theorem
The interpolation theorem is one of the fundamental theorems of mathemat-
ical logic; this was first proved in the setting of first-order logic by Craig [20].
For propositional logic, the interpolation is much simpler:
Theorem 40 Let A(~ p, ~q) and B(~

p, ~r) be propositional formulas involving
only the indicated variables. Suppose A(~ p, ~q) ⊃ B(~
p, ~r) is a tautology. Then
there is a propositional formula C(~p) using only the common variables, so
that A ⊃ C and C ⊃ B are tautologies.
Proof Since A(~ p, ~q) |= B(~

p, ~r) ; if we have already assigned truth values to
p~ = p1 , . . . , pk , then it is not possible to extend this to a truth assignment
on p~, ~q, ~r such that both A(~ p, ~q) and ¬B(~ p, ~r) hold.
35
Let τ1 , . . . , τn be the truth assignments to p1 , . . . , pk for which it is
possible to make A(~ p, ~q) true by further assignment of truth values to ~q .
Let C(~ p) say that one of τ1 , . . . , τn holds for p~ , i.e.,
n ³
_
_ ´
(i) (i) (i)
C = p 1 ∧ p2 ∧ . . . ∧ p k
i=1
where ½
(i) pj if τi (pj ) =True
pj =
¬pj otherwise
p, ~q) |= C(~
Then clearly, A(~ p) . Also, by the comment at the beginning of the
p) |= B(~
proof, C(~ p, ~r). 2
Note that C(~
p) may be exponentially larger than A(~
p, ~q) and B(~
p, ~r) .
Example: Let p1 , . . . , pk code the binary representation of a k -bit
integer P . Let A(~ p, ~q) be a formula which is satisfiable iff P is composite
(e.g. q codes two integers > 1 with product P ). Let B(~ p, ~r) be a formula
which is satisfiable iff P is prime (i.e., ~r codes a Pratt-primality witness).
P is prime ⇔ ∃~rB(~
p, ~r)
⇔ ¬∃~qA(~
p, ~q).
p, ~q) ⊃ ¬B(~
and A(~ p, ~r) is a tautology.
An interpolant C(~ p) must express “ p~ codes a composite”.
Generalizing the above example gives:
Theorem 41 (Mundici [37] If there is a polynomial upper bound on the
circuit size of interpolants in propositional logic, then
N P/poly ∩ coNP/poly = P/poly
Proof Let ∃~qA(~ p) and ∀~rB(~
p, ~q) express an N P/poly property R(~ p, ~r)
express R(~
p) in coNP/poly form. Then
∃~qA(~
p, ~q) |= ∀~rB(~
p, ~r),
which is equivalent to
p, ~q) ⊃ B(~
A(~ p, ~r)
being a tautology. Let C(~
p) be a polynomial size interpolant s.t.,
p, ~q) ⊃ C(~
A(~ p) and p) ⊃ B(~
C(~ p, ~r)
are tautologies. Thus
∃~qA(~
p, ~q) |= C(~
p) |= ∀~rB(~
p, ~r),
p) ⇔ C(~
I.e., R(~ p) and R(~
p) has a polynomial size circuit, so R(~
p) is in
P/poly . 2
36
7.2 Cut-Free Proofs and Interpolant Size
Definition Let P K be the propositional fragment of the Gentzen sequent
calculus. The size of a P K -proof, |P |, is the number of steps in P .
Normally, sequent calculus proofs are treelike; however, sometime we consider
non-treelike proofs, and in this case, we write |P |dag to denote the size of P .
V (A) denotes the set of free variables in A . For C a formula, |C| is the
number of ∧ ’s and ∨ ’s in C .
The next theorem states that tree-like cut-free proofs have interpolants
of polynomial formula size, and general cut-free proofs have interpolants of
polynomial circuit size.
Theorem 42 Let P be a cut-free P K proof of A → B , where V (A) ⊆ {~

p, ~q}
and V (B) ⊆ {~
p, ~q}. Then there is an interpolant C such that
(1) A ⊃ C and C ⊃ B are valid,
(2) V (C) ⊆ {~
p},
(3) |C| ≤ |P | and |C|dag ≤ |P |dag .
Remark: The theorem also holds for proofs which have cuts only on
formulas D such that V (D) ⊆ {~
p, ~r} or V (D) ⊆ {~
p, ~r}.
Proof We use induction on the number of inferences in P to prove a slightly

more general statement:
Claim: If P is a proof of Γ1 , Γ2 → ∆1 , ∆2 and if V (Γ1 , ∆1 ) ⊆ {~
p, ~q} and
V (Γ2 , ∆2 ) ⊆ {~
p, ~r}, then there is an interpolant C so that
(1) Γ1 → ∆1 , C and C, Γ2 → ∆2 are valid,
(2) V (C) ⊆ {~
p}, and
(3) The polynomial size bounds hold too.
Base Case: Initial sequent. If the initial sequent is of the form qi → qi , take
C to be ⊥ since
qi → qi , ⊥ and ⊥→
are valid. For an initial sequent of the form ri → ri , take C to be > . For
an initial sequent pi → pi , C will be either >, ⊥ , pi or (¬pi ) depending
on how the pi ’s are split into Γ1 , Γ2 , ∆1 , ∆2 .
Induction Step: There are a number of cases, depending on the type of the
last inference in the proof.
(1) For last inference an ∨ :right:
37
Γ → ∆, A, B
Γ → ∆, A ∨ B
the interpolant for the upper sequent still works for the lower sequent,
i.e., use C such that
Γ1 → ∆1 , A, B, C and C, Γ2 → ∆2 ,
or
Γ1 → ∆1 , C and C, Γ2 → ∆2 , A, B,
depending on whether A ∨ B is in ∆1 or ∆2 (respectively).
(2) For last inference an ∧ :right:
Γ → ∆, A Γ → ∆, B
Γ → ∆, A ∧ B
(2.a) If A ∧ B is in ∆1 , apply the induction hypothesis twice to have
interpolants CA and CB so that
Γ1 →∆−
1 , A, CA CA , Γ2 →∆2
Γ1 →∆−
1 , B, CB CB , Γ2 →∆2
are valid. Now the derivations
Γ1 → ∆−1 , A, CA Γ1 → ∆−
1 , B, CB
Γ1 → ∆−
1 , A, CA ∨ CB Γ1 → ∆−
1 , B, CA ∨ CB
Γ1 → ∆−
1 , A ∧ B, CA ∨ CB
and
CA , Γ2 → ∆2 CB , Γ2 → ∆2
CA ∨ CB , Γ2 → ∆2
show (CA ∨ CB ) is an interpolant.
(2b) If A ∧ B is in ∆2 applying the induction hypothesis twice gives
CA and CB so that
Γ1 →∆1 , CA CA , Γ2 →∆−
2 ,A
Γ1 →∆1 , CB CB , Γ2 →∆−
2 ,B
are valid. Now the two following derivations show (CA ∧ CB ) is
an interpolant:
CA , Γ2 → ∆−
2 ,A CB , Γ2 → ∆−
2 ,B
CA ∧ CB , Γ2 → ∆−
2 ,A CA ∧ CB , Γ2 → ∆−
2 ,B
CA ∧ CB , Γ2 → ∆−
2 ,A ∧ B
Γ1 → ∆1 , CA Γ1 → ∆1 , CB
Γ1 → ∆1 , CA ∧ CB
The other cases are similar and the size bounds on C are immediate. 2
38
7.3 Interpolation Theorems for Resolution
We start with a quick review of the well-known system of resolution for
propositional logic.
Definition A literal is a propositional variable p or a negated variable ¬p.

p is ¬p, and (¬p) is p. A clause is a set of literals; its intended meaning is
the disjunction of its members. A set of clauses represents the conjunction
of its members. Thus a set of clauses “is” a formula in conjunctive normal
form. A resolution inference is an inference of the form:
C ∪ {p} D ∪ {p}
C ∪D
For such resolution inferences, we assume w.l.o.g. that p, p 6∈ C and p, p 6∈ D .
A resolution refutation of a set Γ of clauses is a derivation of the empty
clause ∅ from Γ by resolution inferences.
Theorem 43 Resolution is refutation-complete (and sound).
Since resolution is a refutation, it does not prove implications A ⊃ B .

Therefore, to formulate the interpolation theorem for resolution, we work
with sets of clauses A ∪ B . An interpolant for the sets A and B is a
formula C such that A ⊃ C and C ⊃ ¬B . This gives the following form of
the interpolation theorem:
Theorem 44 Let {A1 (~ p, ~q)} and {B1 (~

p, ~q), . . . , Ak (~ p, ~r), . . . , B` (~
p, ~r)} be
sets of clauses, so that their union Γ is inconsistent. Then there is a
formula C(~p) such that for any truth assignment τ , domain(τ ) ⊇ {~ p, ~q, ~r} ,
p)) = False, then τ (Ai (~
(1) If τ (C(~ p, ~q)) = False, for some i .
p)) = True, then τ (Bj (~
(2) If τ (C(~ p, ~q)) = False, for some j .
Proof From Γ unsatisfiable, we have
A1 (~ p, ~q) → ¬B1 (~
p, ~q), . . . , Ak (~ p, ~r), . . . , ¬B` (~
p, ~r)
is valid. Thus there is an interpolant C(~

p) such that
A1 (~ p, ~q) → C(~
p, ~q), . . . , Ak (~ p)
and
p) → ¬B1 (~
C(~ p, ~r), . . . , ¬B` (~
p, ~r)
are valid. 2
The next theorem gives bounds on the size or computational complexity

of the interpolant, in terms of the number of inferences in the resolution
refutation.
39
Theorem 45 (Krajı́ček [32]) Let {Ai (~ p, ~q)}i ∪ {Bj (~
p, ~r)}j have a refutation
R of n resolution inferences. Then an interpolant, C(~ p) , can be chosen with
O(n) symbols in dag representation.
If R is tree-like, then C(~
p) is a formula with O(n) symbols.
Proof We view R as a dag or as a tree, each node corresponding to an

inference and labeled with the clause inferred at that inference. For each
clause E in R , define CE (~
p) as follows:
p, ~q), a hypothesis, set CE = ⊥ (False).

(1) For E = Ai (~
p, ~q), a hypothesis, set CE = > (True).
(2) For E = Bj (~
F ∪ {qi } G ∪ {q i }
(3) For an inference
F ∪G
set CF ∪G = CF ∪{qi } ∨ CG∪{qi } .
F ∪ {ri } G ∪ {ri }
F ∪G
set CF ∪G = CF ∪{ri } ∧ CG∪{ri } .
F ∪ {pi } G ∪ {pi }
F ∪G
set CF ∪G = (pi ∧ CF ∪{pi } ) ∨ (pi ∧ CG∪{pi } ) .
Lemma 46 For all clauses F ∈ R , CF (~ p) satisfies the following condition

that if τ is a truth assignment and τ (F ) = False, then
(a) if τ (CF ) = False, then τ (Ai (~
p, ~q)) = False for some i
(b) if τ (CF ) = True, then τ (Bj (~
p, ~r)) = False for some j
The proof of the lemma is by induction on the definition of CF .

Q.E.D. Lemma and Theorem.
7.4 Resolution with Limited Extension

By “extension” is meant the introduction of variables that represent complex
propositional formulas. We have already seen one system that uses extension,
namely the extended Frege proof system. “Limited extension” means that
extension variables may be introduced to abbreviate only formulas that
appear in the formula being proved.
When A is a formula, we let σA be the extension variable for A; in
particular, for p a variable, σp is just the same variable p , and for other
formulas A, σA is a new variable.
40
Definition When A is a formula, LE(A) is a set of clauses which define
the meanings of the extensions variables for all subformulas of A; to wit:
LE(p) = ∅
½ ¾
LE(¬A) = LE(A) ∪ {σ¬A , σA }, {σ¬A , σA }
| {z } | {z }
¬σA ⊃σ¬A σ¬A ⊃¬σA
½ ¾
LE(A ∧ B) = LE(A) ∪ LE(B) ∪ {σA∧B , σA }, {σA∧B , σB }, {σA∧B , σA , σB }
| {z } | {z } | {z }
σA∧B ⊃σA σA∧B ⊃σB σA ∧σB ⊃σA∧B
½ ¾
LE(A ∨ B) = LE(A) ∪ LE(B) ∪ {σA , σA∨B }, {σB , σA∨B }, {σA , σB , σA∨B }
| {z } | {z } | {z }
σA ⊃σA∨B σB ⊃σA∨B σA∨B ⊃σA ∨σB
Definition Let A be a set of formulas. Then LE(A) is ∪A∈A {LE(A)} .

Similarly,
p, ~q) = ∪{LE(A) : V (A) ⊆ {~

LE(~ p, ~q}}.
p, ~r) = ∪{LE(A) : V (A) ⊆ {~
LE(~ p, ~r}}.
Theorem 47 Let Γ be the set of clauses
p, ~q)}i ∪ {Bj (~
{Ai (~ p, ~r)}j ∪ LE(~
p, ~q) ∪ LE(~
p, ~r)
and suppose Γ has a refutation R of n resolution inferences. Then there is

an interpolant C(~ p, ~q)}i and {Bj (~
p) for the sets {Ai (~ p, ~r)}j of circuit size
O(n).
Proof Let C(~ p) be the interpolant for {Ai (~ p, ~q)}i ∪ LE(~

p, ~q) and
p, ~r)}j ∪ LE(~
{Bj (~ p, ~r) given by the earlier interpolation theorem.
Claim C(~
p) is the desired interpolant.
Proof of claim: Any truth assignment τ with domain {~ p, ~q} can be

uniquely extended to satisfy LE(~ p, ~q) . Suppose τ (C(~p)) = False. Extend
τ so as to satisfy LE(~ p, ~q) . By choice of C(~ p) , τ makes a clause from
p, ~q)}i ∪ LE(~
{Ai (~ p, ~q) false, hence makes one of the Ai ’s false.
A similar argument shows that if τ (C(~ p)) = True, then τ falsifies some
Bj (~
p, ~r).
Q.E.D. Claim and Theorem.
41
8 Natural Proofs, Interpolation and Bounded
Arithmetic
8.1 Natural Proofs
The notion of natural proofs was introduced by Razborov and Rudich [44].
In order for a proof to be a natural proof of P 6= NP , it must give a
suitably constructive way of proving that certain Boolean functions are not
in P (for example, the Boolean function Sat which recognized satisfiable
propositional formulas). More precisely, a proof is natural provided it is
possible to use the proof method to find a family Cn of Boolean functions
which satisfies the property of the next definition. We represent a Boolean
function fn (x1 , . . . , xn ) by its truth table (which of course has size N = 2n );
therefore, an n-ary Boolean functions is identified with a string of 2n 0 ’s
and 1’s.
Definition C = {Cn }n is quasipolynomial-time natural against P/poly if

and only if each Cn is a set of (strings representing) truth tables of n -ary
Boolean functions, and such that the following hold:
O(1)
Constructivity: The predicate “fn ∈ Cn ?” has circuits of size 2n , and
n
Largeness: |Cn | ≥ 2−cn · 22 for some c > 0 , and
Usefulness: If fn ∈ Cn for all n , then the family {fn }n is not in P/poly

(i.e., does not have polynomial size circuits).
The motivation for this definition of natural proofs is that “constructive”

proofs that N P 6⊂ P/poly ought to give (quasi)polynomial time property
which is natural against P/poly . Note that ‘quasipolynomial time’, is
measured as a function of the size of the truth table of fn .
8.2 Strong Pseudo-Random Number Generator Conjec-

ture
We now introduce the Strong Pseudo-Random Number Generator (SPRNG)
Conjecture: this conjecture implies the existence of pseudorandom number
generators which generate pseudorandom numbers that pass any feasible
test for randomness. This conjecture is closely related to the problem of the
existence of one-way functions, and is generally conjectured to be true by
researchers in cryptography. (But it is stronger that P 6= NP , and so far it
has not been proven.)
Definition Let Gn : {0, 1}n → {0, 1}2n be a pseudo-random number

generator. The hardness, H(Gn ) , of Gn is the least S > 0 such that, for
42
some circuit C of size S ,
¯ ¯
¯ ¯
¯ P rob [C(Gn (x))=1] − P rob [C(y)=1]¯ ≥ 1
¯x∈{0,1}n y∈{0,1} 2n ¯ S
SPRNG Conjecture: There are pseudorandom number generators Gn ,

²
computed by polynomial size circuits, with hardness H(Gn ) ≥ 2n , for some
² > 0.
One of the main results of Razborov and Rudich [44] is the following
theorem. We omit the proof from these notes.
Theorem 48 If the SPRNG conjecture is true, then there are no properties

which are quasipolynomial time/poly natural against P/poly .
The remainder of this article presents a theorem of Razborov that states,

in essence, that if S22 can prove certain kinds of superpolynomial lower
bounds on the sizes of circuits for polynomial hierarchy predicates (e.g., that
SAT is not in NP), then there is a natural proof that P 6= N P which,
by the previous theorem, would imply the falsity of the SPRNG conjecture.
Since the SPRNG conjecture is thought to be true, then we do not expect
to be able to obtain such a lower bound proof in S22 .
Let α be a new predicate symbol: we will use α to encode information
about a Boolean circuit. Let LB(t, S, α) be a formula which states that
the Boolean circuit encoded by α either does not correctly compute the
predicate S or has size greater than t . In other words, proving LB(t, S, α)
proves that t is a lower bound on the size of any circuit computing S . (We
give a more complete definition of LB() in the next section.)
Razborov [43] proved the following theorem:
Theorem 49 Let t denote any superpolynomial function and let S be any
bounded formula. If S22 (α) proves LB(t, S, α) , then the SPRNG conjecture
is false.
Razborov’s first proof of Theorem 49 used the conservativity of S22 (α)
over T21 (α) and used the resulting PLS algorithm to construct a kind of
interpolant which serves as a natural proof. He later gave a second, more
direct proof based on translating T21 (α) -proofs into propositional proofs [42].
We give below a variation of this second proof based on propositional
interpolation theorems, using a translation of T21 (α) proofs into limited
extension resolution refutations, generalizing a construction of Krajı́ček [32].
The remainder of this article is devoted to this proof.
8.3 Split Bounded Arithmetic Theories

Let α and β be new unary predicate symbols. We define theories S2i (α, β)
and T2i (α, β) of bounded arithmetic in the usual way by adjoining α and β
as new predicate symbols. In addition, these theories admit induction on
43
Σbi (α, β)-formulas (which are Σbi -formulas in which α and β may appear).
We let Σb∞ (α) denote all bounded formulas in the language of S2 plus α .
And we define the class of formulas SΣbi to equal
SΣbi = Σbi (Σb∞ (α), Σb∞ (β))
where Σb1 (X) indicates the closure of X under ∧ , ∨, sharply bounded
quantification and existential bounded quantification, where Πb1 (X) is
defined similarly and
Σbi+1 (X) = Σb1 (Πbi (X))
and Πbi+1 (X) is similarly defined.
The split versions of S2i and T2i are defined as follows:
Definition
SS2i = BASIC + SΣbi -PIND
ST2i = BASIC + SΣbi -IND
We now describe the way in which circuits are coded in the theories S2i (α)
and T2i (α), and the way in which superpolynomial lower bounds on circuit
size are expressed in these theories. Let N ≥ 0 and n = |N | ≈ log N ( where
n = |x| ). Also suppose t(n) = nω(1) (this will be the superpolynomial lower
bound) — note that the fact that t(n) = nω(1) need not be provable in S2 .
Finally, suppose that S(N, x) is a Σb∞ -formula.
Definition Let LB(t, S, α) be the statement
·
¬ α codes a circuit of size ≤ t(n) such that
¸
(∀x ∈ {0, 1} )(α(x) = 1 ↔ S(N, x))
n
where
(1) The free variables of LB(t, S, α) are N and α .
(2) By “α encodes a circuit” we mean that α encodes gate types and gate
connections in some straightforward manner, plus, α may encode the
full truth table description of the functions computed by every gate in
the circuit!
Theorem 50 If S22 (α) ` LB(t, S, α), then SS22 ` SLB(t, S, α, β) where
SLB(t, S, α, β) is
·
¬ α codes a circuit of size ≤ t(n)/2 − 1 and
β codes a circuit of size ≤ t(n)/2 − 1 such that ¸
∀x ∈ {0, 1}n ((α ⊕ β)(x) = 1 ↔ S(N, x))
44
Proof If ¬SLB(t, S, α, β), then the circuit α0





⊕ 




¡¡ @@ 



¡ @
¯L ¯L α0


¯ L ¯ L 

¯ L ¯ L 



¯ α L ¯ β L 


¯ L ¯ L 

satisfies ¬LB(t, S, α). 2
In order to rephrase SLB(t, S, α, β) , we let γ be a new predicate symbol.

Theories such as SS(α, β)22 and ST21 (α, β) may use the symbol γ in induction
formulas. If
SS22 ` SLB(t, S, α, β),
then
SS22 ` ¬CC(t/2 − 1, γ, α) ∨ ¬CC(t/2 − 1, S ⊕ γ, β)
where CC(t, T (x), α) states:
·
α codes a circuit of size ≤ t(n) such that
¸
∀x ∈ {0, 1}n (α(x) = 1 ↔ T (x))
or, in sequent form, SS22 proves
CC(t/2 − 1, γ, α), CC(t/2 − 1, S ⊕ γ, β) →
Since CC is a Σb1 formula, this sequent is also provable in ST21 by ∀Σb2 -

conservativity. (By the same proof that shows S22 is conservative over
T21 .)
Theorem 51 (Razborov [43]) If SS22 ` SLB(t, S, α, β) for some t = nω(1)

and S ∈ Σb∞ , then the SPRNG conjecture is false.
Corollary 52 If the SPRNG conjecture holds, then S22 does not prove
superpolynomial lower bounds on circuit size for any bounded formula (i.e.,
for any polynomial time hierarchy predicate).
The proof of Theorem 51 takes up the rest of this paper. We shall prove
that if
ST21 ` CC(t, γ, α), CC(t, S ⊕ γ, β) → ,
45
then there are quasipolynomial size circuits which are natural against P/poly .
The first step of the proof involves converting an ST21 -proof and the
sequent into a constant-depth propositional proof using the general method
of Theorem 38. We then convert this to a resolution refutation with limited
extension and finally apply the Interpolation Theorem 47. The interpolant
obtained in this way will serve as a natural proof against P/poly .
8.4 Conversion to Constant Depth Frege Proofs

In this section, we discuss how to apply the Paris-Wilkie translation of
Theorem 38 to T21 (α, β) proofs. The proof of Theorem 38 still mostly
applies, and the only new ingredient required for the translation is that
we must describe the translation of atomic formulas β(t) and γ(t) . This
is easily done by augmenting the definition of section 6.4 above to define
(β(t))PW to equal ri and (γ(t))PW to equal pi , where i is the numeric value
of the closed term t .
By expanding the language to include the sequence coding functions and
using SΠb1 -IND and applying free cut-elimination, we may assume that every
formula in the ST21 proof is of the form
(∀y ≤ r)(∃z ≤ |r0 |)(· · ·)
where (· · ·) is a Boolean combination of Σb∞ (α) formulas and Πb∞ (β)

formulas and of formulas γ(· · ·) .
When translated into propositional logic by the Paris-Wilkie translation,
this becomes
O(1)
2n^ ^ n_
O(1)
Ei,j
i=0 j=0
where each formula Ei,j either (1) is ±pi or (2) involves only p~, ~q or
(3) involves only p~, ~r .
Suppose we have a ST21 -proof of
CC(t, γ, α), CC(t, S ⊕ γ, β) → ,
Fixing N and f (x) = S(N, α) , the Paris-Wilkie translation gives a

propositional sequent calculus proof of
^
^ ^
^
Ai (~
p, ~q), p, ~r) →
Bj (~
i j
where
(1) {Ai (~
p, ~q)}i is a set of clauses stating that ~q codes a circuit of size t
computing the function γ with graph given by p~ .
46
(2) {Bj (~
p, ~q)}j is a set of clauses stating that ~r codes a circuit of size t
computing the function γ ⊕ f .
(3) f does not have a circuit of size 2t + 1

(4) Each formula in the proof is a conjunction of disjunctions of formulas
involving just p~, ~q or just p~, ~r .
(5) Each sequent has only c many formulas, where c is a constant

independent of N .
O(1)
(6) The proof has only 2n many symbols.
So far, we have obtained a polynomial size, tree-like sequent calculus

proof in which every formula consists of conjunctions of clauses. The
conjunctions may have polynomially many conjuncts; however, the clauses
will contain at most N O(1) many literals. In addition, there is a constant
upper bound, independent of N , on the number of formulas in each sequent
of the proof. We shall now replace this with a propositional sequent calculus
proof in which all formulas are merely clauses (i.e., disjunctions of atomic
and negated atomic formulas). The new proof will still be polynomial size
and tree-like; however, it will no longer have the property that there is a
constant upper bound on the number of formulas in each sequent.
To form this new proof, we apply the following translation to the
propositional proof obtained above:
(a) Given a sequent
^
^
p1 ^
^
pc0
^
^
q1 ^
^
qc00
E1,i , . . . , Ec0 ,i → F1,i , . . . , Fc00 ,i
i=1 i=1 i=1 i=1
replace it with the q1 · q2 · · · · · qc00 sequents
E1,1 , E1,2 , . . . , E1,p1 , E2,1 , . . . , Ec0 ,pc0 → F1,i1 , F1,i2 , . . . , Fc00 ,ic00
O(1) O(1))
Since each qi = 2n and c00 = O(1) , this still only 2n many
sequents.
(b) Build a new proof of all these sequents. The hardest case of making this
V
V
p
a valid new proof, is the case of a cut on Fi . For this, an inference
i=1
V
V
p V
V
p
Γ → ∆, Fi Fi , Γ → ∆
i=1 i=1
Γ→ ∆
47
is replaced by p cuts; i.e., by
Γ∗ → ∆∗ , F1 F1 , F2 , . . . , Fp , Γ∗ → ∆∗
Γ → ∆ , F2
∗ ∗
F2 , F3 , . . . , Fp , Γ∗ → ∆∗
Γ∗ → ∆∗ , F3 F3 , . . . , Fp , Γ∗ → ∆∗
·
·
·
Γ∗ → ∆∗
At the end of the second step, we have a treelike sequent calculus proof of
A1 (~
p, ~q), . . . , Ak (~
p, ~q), B1 (~ p, ~r) →
p, ~r), . . . B` (~
such that every formula in in the proof is a disjunction of formulas which
either involve just p~ and ~q or involve just p~ and ~r .
8.5 Translation to Resolution with Limited Extension

In the previous section, we have converted the ST21 (α, β) -proof into a
polynomial size, tree-like, propositional sequent calculus proof containing
only clauses as formulas. We next shall convert this proof into a resolution
refutation with limited extension.
Each sequent in the proof obtained in the second step has the form
_
p1 _
pu _
q1 _
qv
E1,i , . . . , Eu,i , → F1,i , . . . , Fv,i (A)
i=1 i=1 i=1 i=1
where each Ea,i , Fa,i involves only {~

p, ~q} or {~
p, ~r} .
Associate with sequent (A), the following set (B) of clauses:
{{E1,i }pi=1
1
, . . . , {Eu,i }pi=1
u
, {¬F1,1 , }, {¬F1,2 , }, . . . , {¬Fv,qv , }} (B)
Now (B) is not really a proper set of clauses, since clauses are supposed
to contain literals (not formulas). So instead of using (B), we introduce
extension variables to form the following set (C) of clauses:
© ª
{σE1,i }pi=1
1
, . . . , {σEu,i }pi=1
u
, {σ¬F1,1 }, {σ¬F1,2 }, . . . , {σ¬Fv,qv } (C)
If sequent (A) is Γ → ∆, then the set (C) of clauses is denoted (Γ → ∆)LE .
It is important for us that all the extension variables used in (C) are from
LE(~ p, ~q) and LE(~
p, ~r).
Lemma 53 If Γ → ∆ is derived in m lines of the sequent calculus proof
constructed in Step (2) above, then
(Γ → ∆)LE ∪ LE(~
p, ~q) ∪ LE(~
p, ~r)
has a resolution refutation (not necessarily tree-like) of O(m2 ) resolution
inferences.
48
Proof By induction on m. — splits into cases depending on the last
inference of the proof:
W
Case (1) Γ → ∆ is A → A . If A = Ai , then
{{σA1 , . . . , σAu }, {σ¬A1 }, . . . , {σ¬Au }} ∪ LE(A)
has a resolution refutation of O(u) inferences.

W
Case (2) Suppose A = i Ai involves only p
~, ~q . Then {σA } and
{σA1 , . . . , σAu } can be derived from each other (in the presence of
LE(A)). Therefore it is not important how we express formulas as
disjunctions when there is a choice.
Case (3) ∧ :left and ∨:right inferences involve only formulas that use just
p~, ~q or just p~, ~r ; these are therefore straightforward (the ∧:right is a
little harder than the ∨ :left case).
Case (4) An ∨ :left inference can be:
W W
i Ei , Γ → ∆ j Fj , Γ → ∆
∨{Ei , Fj }i,j , Γ → ∆
The induction hypotheses give refutations R1 and R2 :


(Γ→∆)LE 

{σEi }i 
R1
=⇒ ∅
LE(~p, ~q) 


LE(~p, ~r)
and

(Γ→∆)LE 

{σFj }j 
R
2
=⇒ ∅
LE(~p, ~q) 


LE(~p, ~r)
These can be combined as:

(Γ→∆)LE 

{σEi }i ∪ {σFj }j  R0 
1
=⇒ {σFj }j 

LE(~p, ~q)  

 (Γ→∆)LE R2
LE(~p, ~r) =⇒ ∅
LE(~p, ~q) 


LE(~p, ~r)
where R10 is like R1 but uses {σEi }i ∪ {σFj }j in place of {σEi }i .

(Remark: Note the refutation is not tree-like since {σFj }j may be used
multiple times in R2 .)
49
Case 5 Suppose the last inference is a cut:
W W
Γ → ∆, i Ai i Ai , Γ → ∆
Γ→ ∆
The induction hypotheses give refutations R1 and R2 :


(Γ→∆)LE 

{σ¬A1 }, . . . , {σ¬Au } 
R1
=⇒ ∅
LE(~p, ~q) 


LE(~p, ~r)
and

(Γ→∆)LE 

{σA1 , . . . , σAu } 
R2
=⇒ ∅
LE(~p, ~q) 


LE(~p, ~r)
Combine these as below, with R10 equal to R1 minus any uses of

{σ¬Ai }’s:

(Γ→∆)LE  0
R1 
LE(~p, ~q) =⇒ {σA1 , . . . , σAu } 

 
LE(~p, ~r) (Γ→∆)LE R2
=⇒ ∅
LE(~p, ~q) 


LE(~p, ~r)
Q.E.D. Lemma.
8.6 Conclusion of the Proof

From the previous Lemma 53 and from the Interpolation Theorem 47 for
O(1)
p) of size 2n
resolution with limited extension, there must be a circuit C(~
such that
p) = 0, then {Ai (~
(1) if C(~ p, ~q)}i is unsatisfiable, and
p) = 1, then {Bj (~
(2) if C(~ p, ~q)}j is unsatisfiable.
O(1)
Note the size of C(~ p) is 2n which is quasipolynomial in N = 2n . In
case (1), when C(~ p) = 0, the function γ(x) does not have a circuit of size
t = nω(1) . In case (2), when C(~ p) = 1 , the function (γ ⊕ f )(x) does not
have a circuit of size t = nω(1) . (Recall f (x) does not have a circuit of size
2t + 1.)
50
Definition Let
df
C ∗ (~ p)) ∨ C(~
p) = (¬C(~ p ⊕ f ),
where p~ ⊕ f is p0 ⊕ f (0), . . . pN −1 ⊕ f (N − 1) . (Each f (i) is 0 or 1, of
course.)
Claim Under the above assumptions, C ∗ (~

p) is a quasipolynomial time
property against P/poly .
Proof There are three things to show:

O(1)
(1) Constructivity: C ∗ has circuits of size 2n since C does.
(2) Largeness: For all γ , either C ∗ (γ) or C ∗ (γ ⊕ f ) holds (since either
¬C(γ) holds, or C((γ ⊕ f ) ⊕ f ) holds). Therefore, C ∗ (γ) holds for at
least half of the γ ’s.
(3) Usefulness: then γ does not have a polynomial size circuit.
(3.a) If ¬C(~ p), i.e., C(~
p) = 0 , then γ = p~ does not have a circuit of
size t , by choice of C .
(3.b) If C(~p ⊕ f ), i.e., C(~
p ⊕ f ) = 1 , then (~p ⊕ f ) ⊕ f = p~ (= γ)
likewise does not have a circuit of size t .
Q.E.D. Theorem 51
References
[1] J. Bennett, On Spectra, PhD thesis, Princeton University, 1962.
[2] M. L. Bonet, S. R. Buss, and T. Pitassi, Are there hard examples
for Frege systems?, in Feasible Mathematics II, P. Clote and J. Remmel,
eds., Boston, 1995, Birkhäuser, pp. 30–56.
[3] S. R. Buss, Bounded Arithmetic, PhD thesis, Princeton University,
1985.
[4] , Bounded Arithmetic, Bibliopolis, 1986. Revision of 1985 Princeton
University Ph.D. thesis.
[5] , The Boolean formula value problem is in ALOGTIME, in Proceed-
ings of the 19-th Annual ACM Symposium on Theory of Computing,
May 1987, pp. 123–131.
[6] , Polynomial size proofs of the propositional pigeonhole principle,
Journal of Symbolic Logic, 52 (1987), pp. 916–927.
[7] , An introduction to proof theory. Typeset manuscript, to appear
in S.R. Buss (ed.) Handbook of Proof Theory, North-Holland, 199?
51
[8] , Axiomatizations and conservation results for fragments of bounded
arithmetic, in Logic and Computation, proceedings of a Workshop held
Carnegie-Mellon University, 1987, vol. 106 of Contemporary Mathemat-
ics, American Mathematical Society, 1990, pp. 57–84.
[9] , Propositional consistency proofs, Annals of Pure and Applied
Logic, 52 (1991), pp. 3–29.
[10] , On Herbrand’s theorem, in Logic and Computational Complexity,
Lecture Notes in Computer Science #960, D. Leivant, ed., Berlin, 1995,
Springer Verlag, pp. 195–209.
[11] , Relating the bounded arithmetic and polynomial-time hierarchies,
Annals of Pure and Applied Logic, 75 (1995), pp. 67–77.
[12] S. R. Buss and et al., Weak formal systems and connections to
computational complexity. Student-written Lecture Notes for a Topics
Course at U.C. Berkeley, January–May 1988.
[13] S. R. Buss and A. Ignjatović, Unprovability of consistency statements
in fragments of bounded arithmetic, Annals of Pure and Applied Logic,
74 (1995), pp. 221–244.
[14] S. R. Buss and J. Krajı́ček, An application of Boolean complexity to
separation problems in bounded arithmetic, Proc. London Math. Society,
69 (1994), pp. 1–21.
[15] A. Cobham, The intrinsic computational difficulty of functions, in
Logic, Methodology and Philosophy of Science, Proceedings of the
Second International Congress, held in Jerusalem, 1964, Y. Bar-Hillel,
ed., Amsterdam, 1965, North-Holland.
[16] S. A. Cook, Feasibly constructive proofs and the propositional calculus,
in Proceedings of the Seventh Annual ACM Symposium on Theory of
Computing, 1975, pp. 83–97.
[17] , Computational complexity of higher type functions, in Proceedings
of the International Congress of Mathematicians, 1990, pp. 55–69.
[18] S. A. Cook and R. A. Reckhow, On the lengths of proofs in the
propositional calculus, preliminary version, in Proceedings of the Sixth
Annual ACM Symposium on the Theory of Computing, 1974, pp. 135–
148.
[19] , The relative efficiency of propositional proof systems, Journal of
Symbolic Logic, 44 (1979), pp. 36–50.
[20] W. Craig, Linear reasoning. A new form of the Herbrand-Gentzen
theorem, Journal of Symbolic Logic, 22 (1957), pp. 250–268.
52
[21] M. Dowd, Propositional representation of arithmetic proofs, in Pro-
ceedings of the 10th ACM Symposium on Theory of Computing, 1978,
pp. 246–252.
[22] , Model-theoretic aspects of P 6= N P . Typewritten manuscript,
1985.
[23] G. Gentzen, Die Widerspruchsfreiheit der reinen Zahlentheorie, Math-

ematische Annalen, 112 (1936), pp. 493–565. English translation in [24],
pp. 132-213.
[24] , Collected Papers of Gerhard Gentzen, North-Holland, 1969.
Editted by M. E. Szabo.
[25] P. Hájek and P. Pudlák, Metamathematics of First-order Arithmetic,

Springer-Verlag, Berlin, 1993.
[26] A. Haken, The intractability of resolution, Theoretical Computer
Science, 39 (1985), pp. 297–308.
[27] J. Herbrand, Recherches sur la théorie de la démonstration, PhD
thesis, University of Paris, 1930.
[28] , Investigations in proof theory: The properties of true propositions,
in From Frege to Gödel: A Source Book in Mathematical Logic, 1978-
1931, J. van Heijenoort, ed., Harvard University Press, Cambridge,
Massachusetts, 1967, pp. 525–581. Translation of chapter 5 of [27], with
commentary and notes, by J. van Heijenoort and B. Dreben.
[29] , Écrits logique, Presses Universitaires de France, Paris, 1968. Ed.

by J. van Heijenoort.
[30] , Logical Writings, D. Reidel, Dordrecht-Holland, 1971. Ed. by W.
Goldfarb, Translation of [29].
[31] C. F. Kent and B. R. Hodgson, An arithmetic characterization of
NP, Theoretical Comput. Sci., 21 (1982), pp. 255–267.
[32] J. Krajı́ček, Interpolation theorems, lower bounds for proof systems,

and independence results for bounded arithmetic. To appear in Journal
of Symbolic Logic.
[33] J. Krajı́ček and P. Pudlák, Propositional proof systems, the consis-

tency of first-order theories and the complexity of computations, Journal
of Symbolic Logic, 54 (1989), pp. 1063–1079.
[34] , Quantified propositional calculi and fragments of bounded arith-
metic, Zeitschrift für Mathematische Logik und Grundlagen der Math-
ematik, 36 (1990), pp. 29–46.
53
[35] J. Krajı́ček, P. Pudlák, and G. Takeuti, Bounded arithmetic and
the polynomial hierarchy, Annals of Pure and Applied Logic, 52 (1991),
pp. 143–153.
[36] R. E. Ladner, The circuit value problem is log space complete for P,
SIGACT News, 7 (1975), pp. 18–20.
[37] D. Mundici, Tautologies with a unique Craig interpolant, Annals of
Pure and Applied Logic, 27 (1984), pp. 265–273.
[38] E. Nelson, Predicative Arithmetic, Princeton University Press, 1986.
[39] C. H. Papadimitriou, On graph-theoretic lemmata and complexity
classes (extended abstract), in Proceedings of the 31st IEEE Symposium
on Foundations of Computer Science (Volume II), IEEE Computer
Society, 1990, pp. 794–801.
[40] R. J. Parikh, Existence and feasibility in arithmetic, Journal of
[41] J. B. Paris and A. J. Wilkie, ∆0 sets and induction, in Open Days
in Model Theory and Set Theory, W. Guzicki, W. Marek, A. Pelc, and
C. Rauszer, eds., 1981, pp. 237–248.
[42] A. A. Razborov, On provably disjoint NP-pairs, Tech. Rep. RS-94-
36, Basic Research in Computer Science Center, Aarhus, Denmark,
November 1994. http://www.brics.dk/index.html.
[43] , Unprovability of lower bounds on the circuit size in certain
fragments of bounded arithmetic, Izvestiya of the RAN, 59 (1995),
pp. 201–224.
[44] A. A. Razborov and S. Rudich, Natural proofs, in Proceedings of
the 26-th Annual ACM Symposium on Theory of Computing, 1994,
pp. 204–213.
[45] R. A. Reckhow, On the Lengths of Proofs in the Propositional Calculus,
PhD thesis, Department of Computer Science, University of Toronto,
1976. Technical Report #87.
[46] J. Siekmann and G. Wrightson, Automation of Reasoning, vol. 1&2,
Springer-Verlag, Berlin, 1983.
[47] R. Statman, Complexity of derivations from quantifier-free Horn
formulae, mechanical introduction of explicit definitions, and refinement
of completeness theorems, in Logic Colloquium ’76, Amsterdam, 1977,
North-Holland, pp. 505–517.
[48] L. J. Stockmeyer, The polynomial-time hierarchy, Theoretical Com-
put. Sci., 3 (1976), pp. 1–22.
54
[49] G. S. Tsejtin, On the complexity of derivation in propositional logic,
Studies in Constructive Mathematics and Mathematical Logic, 2 (1968),
pp. 115–125. Reprinted in: [46, vol 2].
[50] A. J. Wilkie and J. B. Paris, On the scheme of induction for bounded
arithmetic formulas, Annals of Pure and Applied Logic, 35 (1987),
pp. 261–302.
[51] C. Wrathall, Complete sets and the polynomial-time hierarchy,
Theoretical Comput. Sci., 3 (1976), pp. 23–33.
[52] D. Zambella, Notes on polynomially bounded arithmetic, Journal of

55

Buss Arith

Uploaded by

Copyright:

Available Formats

Buss Arith

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Buss Arith

Uploaded by

Copyright:

Available Formats

Bounded Arithmetic and

Propositional Proof Complexity

1 Review of Computational Complexity

1.1.1 Feasibility Versus Infeasibility — An Example

The runtime of this algorithm is approximately |x| · |y| steps.

≈ 140 ∗ 1020 µs ≈ 1000 × (age of the universe).

On the other hand, the runtime of the feasible algorithm is

Definition Let n denote the sums of the lengths of the inputs to an

1.2 P, NP, and the Polynomial-Time Hierarchy

Definition P is the set of polynomial time recognizable predicates on N.

Definition Let q be a polynomial. f is defined from g and h by limited

provided |f (~x, y)| ≤ q(|~x|, |y|) for all ~x, y .

Theorem 1 (Cobham [15]) F P is equal to the set of functions which can

Definition If Ψ is a set of predicates, PB∃(Ψ) is the set of predicates A

Definition NP = PB∃(P ) and coNP = PB∀(P ) . In other words, NP is

A(~x) ⇔ (∃y ≤ 2p(|~x|) )B(~x, y)

Definition If Ψ is a set of predicates, P Ψ (resp., F P Ψ ) is the set of

The notions of P , F P , NP and coNP are generalized by defining the

Definition The Polynomial Time “Hierarchy”:

It is an open question whether P = NP , NP = coNP and whether the

2.1 The Language of Bounded Arithmetic

It is easy to check that the # (pronounced “smash”) function allows us to

Definition A bounded quantifier is a quantifier of the form (Qx ≤ t) with

A hierarchy of classes Σbk , Πbk of bounded formulas is defined by counting

Theorem 2 Fix k ≥ 1 . A predicate Q is in Σpk iff there is a Σbk formula

This theorem is essentially due to Stockmeyer [48] and Wrathall [51];

2.2 Induction Axioms for Bounded Arithmetic

2.3 Theories of Bounded Arithmetic

Definition Let i ≥ 0. S2i is the first-order theory with language 0 , S , +,

The following, somewhat surprising, relationship holds between the

2.4 Provably Recursive and Σbi -Definable Functions

Theorem 5 [3, 4] Let i ≥ 1 . Any Σb1 -definable function or ∆b1 -definable

Proof (Sketch) Suppose f is Σb1 -defined by R , so that

R ` (∀x)(∃!y ≤ r(~x))Af (~x, y).

Then any atomic formula ϕ(f (~s)) is equivalent to both

(∃y ≤ r(~s))(Af (~s, y) ∧ ϕ(y))

2.6 Main Theorems for S2i

(1) S2i ` (∀~x, y)(B(~x, y) ⊃ A(~x, y)).

Corollary 8 ( i ≥ 1) The Σbi -definable functions of S2i are precisely the

It is this corollary that allows us to state that the proof-theoretic strength

Definition A predicate Q(~x) is ∆bi -definable in a theory T provided it is

Q(~n) ⇔ N |= A(~n) ⇔ N |= B(~n)

Corollary 11 If A is a formula which is S21 -provably in NP ∩ coNP then

3 The Sequent Calculus

Definition A sequent is an expression of the form

where the Ai ’s and Bi ’s are formulas. Its intended meaning is

We shall use Greek letters Γ, ∆, . . . to denote finite sequences of formulas

Definition An LK-proof is a tree of sequents: the leaves or initial sequents

(3) Allow induction inferences: (for A ∈ Σbi )

A(b 12 bc), Γ → ∆, A(b)

3.3 Free-Cut Elimination

is free unless A is the direct descendant either of a formula in an initial

3.4 Outline of Proof of the Main Theorem

3.5 The Witness Predicate

(1) If B ∈ Σbi−1 ∪ Πbi−1 then Witnessi,~

(5) If B = (∀x ≤ |t|)C(~a, x) then

(6) If B = ¬C use prenex operations to push the negation sign inside.

Lemma 14 Fix i ≥ 1. Let B ∈ Σbi .

B(~a) ↔ (∃w ≤ tB )Witnessi,~

Proof: This is easily proved by induction on the complexity of B .