Abstract

First of all we give some reasons that “natural proofs” built not a barrier to prove P 6= NP using Boolean complexity. Then we investigate the approximation method for its extension to prove superpolynomial lower bounds for the non-monotone complexity of suitable Boolean functions in NP or to understand why this is not possible. It is given some evidence that the approximation method alone cannot be used to prove a super-linear lower bound for any function f ∈Bn. Additionally, an overview on the methods for proving lower bounds of the non-monotone and the monotone complexity of Boolean functions is given. Finally, a personal opinion how to proceed the research on the P versus NP problem and also on proving a super-linear lower bound for the non-monotone complexity of a Boolean function in NP is given.

1 Introduction and Preliminaries

Understanding the power of negations is one of the most challenging problems in complexity theory. With respect to monotone Boolean functions, Razborov [37] was the first who could show that the gain, if using negations, can be super-polynomial in comparision to monotone Boolean networks. Tardos [42] has improved this to exponential. For the characteristic function of an NP-complete problem like the clique function, it is widely believed that negations cannot help enough to improve the Boolean complexity from exponential to polynomial. Since the computation of an one-tape Turing machine can be simulated by a non-monotone Boolean network of size at most the square of the number of steps [40, Chapter 3.9], a superpolynomial lower bound for the non-monotone network complexity of such a function would imply P 6= NP. For the monotone complexity of such a function, exponential lower bounds are known [4, 6, 36, 2, 25, 20, 22, 7, 3, 21]. But until now, no one could prove a super-linear lower bound for the nonmonotone complexity of any Boolean function in NP. An obvious attempt to get a super-polynomial lower bound for the non-monotone complexity of the clique function could be the extension of the method which has led to the proof of an exponential lower bound of its monotone complexity. This is the so-called “method of approximation” developed in 1984 independently by Andreev [4] and Razborov [36]. In 1989, at the 21st STOC, Razborov [38] has presented the sketch of a proof that his approximation method cannot be used to prove better than quadratic lower bounds for the non-monotone complexity of any Boolean function. But Razborov uses a very strong distance measure in his proof for the inability of the approximation method. As elaborated in [10], one can use the approximation method with a weaker distance measure to prove super-polynomial lower bounds. Our goal is the extension of the approximation method to non-monotone Boolean networks to prove a super-polynomial lower bound for the non-monotone complexity of a function in NP or to understand why this is not possible.

Firstly, we give some basic definitions. Bn := {f | f : {0,1}n → {0,1}} is the set of all n-ary Boolean functions. The ith variable is denoted by xi : {0,1}n → {0,1}, 1 ≤ i ≤ n. Let Vn := {xi | 1 ≤ i ≤ n} and

. Variables and negated variables are called

literals. A function m : {0,1}n → {0,1} which is the conjunction of some literals is called a monomial. If we delete some literals from a monomial m then we obtain a submonomial m′ of m and we write m′ ⊆ m. The empty monomial ε is the constant function 1. The disjunction of monomials is a formula in disjunctive normal form (DNF). The disjunction of some literals is called a clause. If we delete some literals from a clause d then we obtain a subclause of d. The empty clause is the constant function 0. The conjunction of clauses is a formula in conjunctive normal form (CNF). A monomial m is called an implicant of the function f if for all a ∈ {0,1}n, m(a) = 1 implies f(a) = 1. An implicant m is a prime implicant of f if no proper submonomial of m is an implicant of f. A clause d is called an f-clause if for all a ∈ {0,1}n, d(a) = 0 implies f(a) = 0. A prime clause d of f is an f-clause where no proper subclause of d is an f-clause. Let a := (a1,a2,...,an), b := (b1,b2,...,bn) ∈ {0,1}n. We write a ≤ b iff ai ≤ bi for 1 ≤ i ≤ n. A function f ∈ Bn is monotone iff a ≤ b implies f(a) ≤ f(b) for all a,b ∈ {0,1}n. Let Ω0 := {∧,∨,¬} and Ωm := {∧,∨}.

For Ω ∈ {Ω0,Ωm,B2}, an Ω-network β is a directed, acyclic graph such that each node has indegree at most two. The nodes g with indegree zero are input nodes and are labelled with op(g) ∈ Vn ∪ {0,1}. The nodes g with indegree larger than zero are the gates of β. Each gate g is labelled with an operator op(g) ∈ Ω where the indegree of g is equal the number of operands of op(g). A node with outdegree zero is an output node. For a node g in β let pred(g) := {h | h → g is an edge in β} be the set of its direct predecessors. With each node g, we associate a function resβ(g) : {0,1}n → {0,1} which

is defined by

op(g) g is an input node,

resβ(g) := ¬resβ(h1) op(g) = ¬, pred(g) = {h1},

resβ(h1) op(g) resβ(h2) otherwise, where pred(g) = {h1,h2}.

The functions resβ(g) with g is a node in β are computed by β. Let f ∈ Bn. The minimum number of gates in an Ω0-network which computes f where negations are not counted is the non-monotone complexity C(f) of f. Each operator in B2 can be realized by an Ω0-network using at most two gates. Hence, for proving a super-linear lower bound for the size of a B2-network realizing a Boolean function f we can restrict us to prove a superlinear lower bound for the non-monotone complexity of f. An Ωm-network is called a monotone network. Note that exactly the monotone Boolean functions can be computed by a monotone network. The minimum number of gates in a monotone network which computes the monotone function f is the monotone complexity Cm(f) of f.

Given any Ω0-network β, we can convert β to an equivalent Ω0-network β′ where all negations occur only at the input nodes. Moreover, the size of β is at most doubled. For doing this, we start at the output nodes and apply De Morgan rules for bringing the negations to the input nodes. Since gates can be simultaneously negated and non-negated, some gates have to be doubled. The resulting network is a so-called standard network where only input variables are negated. We consider a negated variable ¬xi as an input node g with op(g) = ¬xi. The standard complexity Cst(f) of a function f ∈ Bn is the size of a smallest standard network which computes f. Note that the standard and the non-monotone complexity of a function f differs at most by the factor two. Hence, for proving a super-linear lower bound for the non-monotone complexity of a Boolean function, we can restrict us to the consideration of standard networks.

Before the investigation of the approximation method with respect to its extension to standard networks, we have to deal with the paper of Razborov and Rudich [39]. In 1994, Razborov and Rudich have introduced the notion of a “natural proof”. They say that the known proofs of lower bounds on the complexity of explicit Boolean functions in non-monotone models fall within their definition of natural. They have shown that natural proofs cannot be used for separating P and NP unless hard pseudorandom generators do not exist. Since the existence of such generators is widely believed, natural proofs are widely accepted to be a barrier for proving P 6= NP using Boolean complexity. We discuss this in the subsequence.

Firstly, we give their definition of a “natural proof”. This is a proof which uses a natural combinatorial property. A combinatorial property is a subset {Cn ⊂ Bn | n ∈ N} of Boolean functions. Cn is called natural if there is Cn∗ ⊆ Cn which satisfies:

1. For all f ∈ Bn it can be decided in 2O(n) time if f ∈ Cn∗. (constructiveness)

2. |Cn∗| ≥ 2−O(n)|Bn|. (largeness)

The first property means that the characteristic function of can be computed in polynomial time in the size of the truth table of the input function f ∈ Bn. The second property says that a function randomly chosen from Bn is contained in with non-negligible probability. P/poly is the set of languages which are recognizable by a family of Boolean networks of polynomial size. Note that P ⊆ P/poly. A combinatorial property is useful against P/poly if the network complexity of any sequence f1,f2,...,fn,... where fn ∈ Cn is super-polynomial; i.e., for all k ∈ N there is nk ∈ N such that the network complexity of fn is larger than nk for all n > nk. A proof that a Boolean function does not have polynomial network complexity is natural against P/poly if the proof uses a natural combinatorial property Cn which is useful against P/poly.

Razborov and Rudich mention that “from experience it is plausible to say that we do not yet understand the mathematics of Cn outside exponential time (as a function of n) well enough to use them effectively in a combinatorial style proof.” This means that combinatorial properties used in a today’s lower bound proof are constuctive. With respect to the largeness property, they write: “In Section 5 we give some solid theoretical evidence for largeness, by showing that any Cn based on a formal complexity measure must be large.” We discuss now the “solid theoretical evidence for largeness” given by Razborov and Rudich.

A formal complexity measure is a function µ : Bn 7→ R+ such that

a) µ(f) ≤ 1 for f ∈ {x1,x2,...,xn,¬x1,¬x2,...,¬xn}, and

b) µ(f ∧ g) ≤ µ(f) + µ(g) and µ(f ∨ g) ≤ µ(f) + µ(g) for all f,g ∈ Bn.

A formula is a Boolean network where the underlying graph is a tree. The size of a formula β is the number of leaves in β. The formula size LΩ0(f) of a Boolean function f ∈ Bn is the size of a smallest Ω0-formula which computes f. Note that LΩ0 itself is a formal complexity measure. Moreover, by induction on the formula size, it can be shown that LΩ0(f) ≥ µ(f) for all f ∈ Bn and each formal complexity measure µ.

Razborov and Rudich show that “any formal complexity measure µ which takes a large value at a single function, must take large values almost everywhere.” This is formalized by the following theorem.

Theorem 1 Let µ be a formal complexity measure on Bn, and let µ(f) ≥ t for some f ∈ Bn. Then for at least 1/4 of all functions g ∈ Bn, µ(g) ≥ t/4.

Razborov and Rudich conclude that “every combinatorial property based on such a measure automatically satisfies the largeness condition in the definition of natural property.” But they do not formalize what they mean that a property is based on a formal complexity measure. The combinatorial property of f ∈ Bn cannot be “µ(f) ≥ t” for a certain bound t since the property is used to prove this lower bound. Let Cn ⊂ Bn be the combinatorial property used to prove µ(f) ≥ t; i.e., f ∈ Cn. Theorem 1 does not imply that a function g having the same measure as f has to be contained in Cn. Hence, Theorem 1 does not imply the largeness property for Cn.

Theorem 1 is not a surprise because of Shannon’s famous counting argument [41, 49] which shows that at least a fraction of (1 ) of the functions in Bn has non-monotone complexity at least . The theorem tells us that we can only use a formal complexity measure to prove a lower bound t for the non-monotone network complexity of a function f ∈ Bn which has the property that up to the constant factor , the measure of at least a quarter of the functions in Bn is at least as large as the measure of f. Using the counting argument again, we know that the formula size has this property. By Shannon’s counting argument, a combinatorial property which implies that a function having this property has complexity smaller than could never fulfill the largeness property. Hence, a proof which uses such a combinatorial property is not natural.

Altogether, it seems that “natural proofs” built not a barrier for proving P 6= NP using Boolean complexity. Therefore, it makes sence to investigate the approximation method with regard its expandability to standard networks. This is the aim of the paper.

The first problem which we have to investigate is the treatment of the negated variables. In [11], I have tried to treat the negated variables in a standard network which computes a given non-constant monotone Boolean function f at its output node gt in such a way that we can use the approximators developed for f with respect to monotone networks on standard networks. The proof of Theorem 6 in [11] is wrong. The mistake in the proof is explained in [12]. The motivation of the approach in [11] was the avoidance of the explicit consideration of the negated variables. The conclusion in [12] is that the negated variables have to be approximated as well. Therefore, we have to consider the negated variables explicitely. Before doing this, a lot of work has to be done.

We shall investigate the computation in standard networks in the next section. In Section 3, an overview on the methods for proving lower bounds of the non-monotone and of the monotone complexity of Boolean functions is given. We describe the general idea of the approximation method in Section 4. For monotone networks, two kinds of approximators are known, the CNF-DNF-approximators and the sunflower-approximators. Section 5 is devoted the description of these approximators with respect to monotone networks. The extension of these approximators to standard networks is treated in Section 6. It is given evidence that CNF-DNF- and also sunflowerapproximators alone cannot be used to prove a super-linear lower bound for the standard complexity of any function in Bn. In Section 7, a personal opinion how to proceed the research on the P versus NP problem and also on proving a super-linear lower bound for the non-monotone complexuty of a Boolean function in NP is given.

2 The Computation in Standard Networks

Let β be a standard network which computes a function f ∈ Bn at its output node gt. Let g be any node in β. The function resβ(g) can be written as a DNF-formula; i.e., res where each mj is a monomial.

We call this representation of resβ(g) the DNF-representation DNFβ(g) of resβ(g). The function resβ(g) can be written as a CNF-formula as well; i.e., res where each dj is a clause. We denote this formula the CNF-representation CNFβ(g) of resβ(g). In contrast to monotone networks, DNFβ(gt) must not contain the prime implicants of the function f as monomials. Furthermore, CNFβ(gt) must not contain the prime clauses of the function f as clauses. To extend the proof techniques developed for monotone networks to standard networks, it is useful to recognize the prime implicants in DNFβ(gt). Similarly, it is useful to recognize the prime clauses in CNFβ(gt). For doing this, let p1,p2,...,pk be the prime implicants and c1,c2,...,cs be the prime clauses of the function f. Each monomial mj in DNF is an implicant of the function f. Otherwise, it would exist an input (a1,a2,...,an) ∈ {0,1}n such that f(a1,a2,...,an) = 0 but resβ(gt)(a1,a2,...,an) = 1. This means that each monomial in DNFβ(gt) contains at least one prime implicant of the function f as a submonomial. If a monomial mj contains l > 1 prime implicants then we add l − 1 copies of mj to the DNF-representation. By separating in each monomial containing a prime implicant pj the prime implicant and the other literals, we can write

k lj

res ,

j=1 i=1

where DNFβ(gt) contains lj monomials including the prime implicant pj,

1 ≤ j ≤ k.

Similarly, we separate in each f-clause dj in CNFβ(gt) the contained prime clause from the other literals. If a clause dj contains l > 1 prime clauses then we add l − 1 copies of dj to the CNF-representation. By separating in each clause containing a prime clause cj the prime clause and the other literals, we can write

s lj

res ,

j=1 i=1

where CNFβ(gt) contains lj clauses including the prime clause cj, 1 ≤ j ≤ s.

Now we describe the DNF- and CNF-formulas constructed by a standard network β. Note that after a simplification of the network, no input node g with op(g) ∈ {0,1} exists. Assume that for all input nodes g, op(g) ∈

. Starting at the input nodes, the network β constructs the DNFformulas in the following way:

1. If g is an input node with op(g) = xi or op(g) = ¬xi then

DNFβ(g) := op(g).

2. If g is an ∨-gate with pred(g) = {h1,h2} then

DNFβ(g) := DNFβ(h1) ∨ DNFβ(h2).

3. If g is an ∧-gate with pred(g) = {h1,h2}, DNF and

DNF then

t1 t2

DNFβ(g) := __(mi ∧ m′j).

i=1 j=1

Each input a ∈ resβ(g)−1(1) satisfies a monomial mj of DNFβ(g). Each input b ∈ resβ(g)−1(0) does not satisfy any monomial in DNFβ(g). Hence, each monomial in DNFβ(g) contains a variable xi with bi = 0 or a negated variable ¬xj with bj = 1.

Starting at the input nodes, the network β constructs the CNF-formulas in the following way:

1. If g is an input node with op(g) = xi or op(g) = ¬xi then

CNFβ(g) := op(g).

2. If g is an ∧-gate with pred(g) = {h1,h2} then

CNFβ(g) := CNFβ(h1) ∧ CNFβ(h2).

3. If g is an ∨-gate with pred(g) = {h1,h2}, CNF and

CNF then

t1 t2

CNFβ(g) := ^^(di ∨ d′j).

i=1 j=1

Each input b ∈ resβ(g)−1(0) falsifies a clause dj of CNFβ(g). Each input a ∈ resβ(g)−1(1) does not falsify any clause in CNFβ(g). Hence, each clause in CNFβ(g) contains a variable xi with ai = 1 or a negated variable ¬xj with aj = 0.

The following theorem characterizes exactly the DNF-representation and the CNF-representation of resβ(gt) with respect to a standard network which computes a Boolean function f ∈ Bn at its output node gt.

Theorem 2 Let β be a standard network which computes a Boolean function f ∈ Bn at its output node gt. Then the following hold:

a) DNFβ(gt) contains only implicants of the function f. Furthermore, for each a ∈ f−1(1), DNFβ(gt) contains an implicant ma of f such that ma(a) = 1.

b) CNFβ(gt) contains only f-clauses. Furthermore, for each b ∈ f−1(0), CNFβ(gt) contains an f-clause db such that db(b) = 0.

Proof: Assume that DNFβ(gt) contains a monomial m which is not an implicant of f. Then, by the definition of an implicant of f, there exists b ∈ {0,1}n such that m(b) = 1 but f(b) = 0. This contradicts the assumption that β computes f at its output node gt. Hence, all monomials of DNFβ(gt) are implicants of f.

Assume that there is a ∈ f−1(1) such that m(a) = 0 for all implicants m in DNFβ(gt). Then resβ(gt)(a) = 0 but f(a) = 1. This contradicts the assumption that β computes f at its output node gt. Hence, for each a ∈ f−1(1), DNFβ(gt) contains an implicant ma of f such that ma(a) = 1.

This proves part a) of the theorem. Analogously, part b) of the theorem can be proved.

Every DNF-formula can be transformed into an equivalent CNF-formula.

To see this let be a DNF-formula which computes a Boolean function f ∈ Bn. To obtain an equivalent CNF-formula γ, we pick from each monomial mi, 1 ≤ i ≤ t0 one literal and perform the disjunction of all chosen literals. Then the conjunction of all clauses which can be constructed in this way is a CNF-formula which corresponds to the DNF-formula α. The following lemma shows that γ computes the function f.

Lemma 1 be a DNF-formula which computes a Boolean function f ∈ Bn. Let be the CNF-formula constructed from α as described above. Then γ computes f.

Proof: Consider a ∈ f−1(1). Then there is a monomial ml in α such that ml(a) = 1. Since each clause of γ contains a literal of ml, the input a satisfies all clauses in γ. Hence γ(a) = 1.

Let b ∈ f−1(0). Then each monomial in α contains a literal which is not satisfied by b. Consider a clause dl of γ which picks from each monomial a literal which is not satisfied by b. Obviously, dl(b) = 0. Hence, γ(b) = 0.

Altogether, we have shown that γ computes f.

We call such a transformation of a DNF-formula to an equivalent CNFformula a DNF/CNF-switch. A DNF/CNF-switch can be organized as the construction of a tree T in the following way:

1. Each edge in T is labelled by a literal. With each node w in T we associate the clause d(w) which is obtained by the disjunction of the literals on the unique path from the root of T to w. T is constructed while expanding the monomials m0,m1,m2,...,mt0 where m0 is the empty monomial.

2. While expanding m0, the root of T is created. The associated clause is the empty clause.

3. Suppose that w is a leaf that was created while expanding mi. Then the monomial mi+1 is expanded at the leaf w in the following way: The leaf w obtains for each literal in mi+1 a new son w′. The edge (w,w′) is labelled with the corresponding literal.

After the construction of the tree T, the clauses corresponding to the paths from the root of T to the leaves are the clauses contained in the CNF-formula γ obtained from by performing a DNF/CNF-switch.

Analogously, every CNF-formula can be transformed into an equivalent DNF-formula. We call such a transformation of a CNF-formula to an equivalent DNF-formula a CNF/DNF-switch.

3 On Proof Methods in Boolean Complexity

To get a lower bound for the size of a network β which computes a Boolean function f ∈ Bn, we have to count gates in β. The problem is that we have no knowledge about the structure of β; i.e., we can only use the fact that the network β computes the function f. Therefore, with respect to a complete basis B2 or Ω0, only small linear lower bounds for the network complexity of a function in NP could be proved. If the function f depends on each of the n input variables then each of the n input nodes has to be connected to the output node. For doing this, at least n − 1 gates with indegree two are needed. Hence, each Boolean network which computes a function depending on all n input variables contains at least n − 1 gates. Since functions like x1 ∧ x2 ∧ ... ∧ xn depend on all variables and can be realized with only n − 1 gates, without an additional argument, no larger lower bound can be proved. Slightly better lower bounds are obtained using the so-called gate-elimination method. The gate-elimination method uses induction. By an assignment of some variables with values from {0,1}, a specific small constant number of gates is eliminated in each step and the resulting function is of the same type as the function before the assignment. Over the years, the case analyses used in the proofs have become more and more complicated impoving the lower bounds only slightly. For an overview see [49, 10, 18]. I am convinced that the elimination method alone cannot be used to prove a super-linear lower bound for the network complexity of any function in NP.

What happen if we consider Boolean functions with many outputs as the Boolean matrix multiplication or the Boolean convolution? With respect to the convolution, each output depends on nearly all variables. Moreover, as shown by Valiant [45], the graph of any network computing the convolution is an n-superconcentrator. An n-superconcentrator is a directed graph with n input and n output nodes such that for each subset of the input nodes and each subset of the output nodes of the same size r there are r mutually nodedisjoint paths connecting the set of input nodes with the set of output nodes. Aho, Hopcroft and Ullman [1] have conjectured that an n-superconcentrator has at least nlogn edges. But Valiant [45] itself has shown that there exist superconcentrators of linear size destroying the hope to prove a super-linear lower bound using graph theoretical arguments only. Also for functions with many outputs, no super-linear lower bound for its network complexity is known.

The inability to prove super-linear lower bounds for the non-monotone complexity of explicit Boolean functions has led to the consideration of restricted models of Boolean networks like monotone or bounded-depth Boolean networks. For both restricted models, exponential lower bounds for the complexity of an explicit Boolean function in NP are known. We are interested in proving a super-linear lower bound of the non-monotone complexity of a Boolean function in NP. Bounding the depth of the network to be constant seems to be a much harder restriction than allowing only monotone networks. Some functions with linear network complexity are used to prove an exponential lower bound for the size of a constant depth network computing the function. Techniques for proving super-linear lower bounds for the monotone complexity of functions which are also candidates for proving a super-linear lower bound for its non-monotone complexty seems to be more suitable for their extension to get a super-linear lower bound for the non-monotone complexity of a function in NP. Therefore, we are interested in the methods developed for the proof of lower bounds for the monotone complexity of Boolean functions.

The core of each super-linear lower bound proof for the monotone complexity of a Boolean function is the successful application of certain replacement rules. In a monotone network β computing a Boolean function f, a replacement rule replaces a node u with resβ(u) = h by a node u′ which computes a function h′. In most cases, h′ depends on h. The first replacement rules used to prove some lower bounds have the additional property that the resulting monotone network β′ still computes the function f. Such a replacement rule is of Type 1. For a presentation of replacement rules of Type 1 see [28, 49].

Replacement rules of Type 1 in combination with the gate elimination method are used to prove lower bounds for the monotone complexity of Boolean sums [30, 33, 47, 29], Boolean matrix multiplication [35, 31, 28] and generalized Boolean matrix multiplication [46]. Replacement rules of Type 1 are used in different ways. With respect to Boolean matrix multiplication [31, 28], they are used for the characterization of an optimal monotone network. For the generalized matrix product [46] and for Boolean sums [47, 29] they are used explicitely; i.e., the gate u is replaced by a subnetwork which computes the function h′. This should be possible without additional cost. To get this, Wegener [46] has introduced a technique, very common in algebraic complexity, into Boolean complexity. Certain functions are given for free as inputs of the network. A lower bound for such a network implies the same lower bound for the monotone complexity of the considered function. Wegener [46] uses this technique in combination with the gate elimination method to prove an Ω(n2/log2 n) lower bound for the monotone complexity of the generalized Boolean matrix product. In [48], Wegner has introduced a further technique improving this lower bound to Ω(n2/logn). Instead using the gate elimination method, he has defined a suitable value function to estimate the contribution of each ∧-gate for the computation of the outputs. At each ∧-gate, the value function distributes at most the value 1 among the prime implicants. Then he has proved the necessity to give to each prime implicant at least the value obtaining a lower bound of half the number of prime implicants. The definition of the value function depends on the structure of the function computed at the ∧-gate under consideration. Important for the proof is that the function has many outputs and also the structure of the prime implicants of the functions computed at the output nodes.

The first super-linear lower bound for the monotone complexity of an explicit Boolean function has been proved by Neciporuk [30] in 1969 for a function in Mn,n, a set of so-called Boolean sums. A Boolean sum fi is the disjunction of a subset Fi ⊆ Vn of variables. He has considered the monotone complexity of sets of Boolean sums which have “nothing in common”. Nothing in common means that two distinct Boolean sums have at most one variable in common. We say then that the set of Boolean sums is (1,1)-disjoint. One can think that ∧-gates cannot reduce the monotone complexity of a set of Boolean sums in comparision to networks which use only ∨-gates. But Tarjan [43, 49] has given an example which shows that using ∧-gates can reduce the monotone complexity. For (1,1)-disjoint sets of Boolean sums, Neciporuk has proved that optimal monotone Boolean networks contain only ∨-gates. A well known construction of Ko˝v´ari, So´s and Tura´n [27] leads to an explicitely constructed (1,1)-disjoint set f = (f1,f2,...,fn) of Boolean sums with Ω(n3/2) prime implicants such that an Ω(n3/2) lower bound for the monotone complexity of this function has been proved. Some years later, Pippenger [33] and Mehlhorn [29] have generalized the approach of Neciporuk to sets of Boolean sums which are (h,k)-disjoint; i.e., any h + 1 different Boolean sums have at most k variables in common. Such a set of Boolean sums corresponds to a bipartite graph which does not contain a Kh+1,k+1 as a subgraph where Kh+1,k+1 denotes the complete bipartite graph with node sets of sizes h+1 and k+1, respectively. Bipartite graphs can be represented by Boolean matrices. A Boolean matrix A is (h,k)-free if it does not contain any (h+1)×(k+1) submatrix containing only ones. A bipartite graph contains no Kh+1,k+1 iff the corresponding Boolean matrix is (h,k)-free. The question about the maximal number of ones in a (h,k)-free (n×n)-matrix is the famous problem of Zarankievicz. Pippenger and Mehlhorn showed that using ∧-gates for the computation of a set of (h,k)-disjoint Boolean sums can save at most the factor max{h − 1,k − 1}. Using a construction of Brown [13], a (2,2)-disjoint set of Boolean sums with Ω(n5/3) prime implicants has been constructed such that an Ω(n5/3) lower bound for the monotone complexity of this Boolean function has been proved. In the subsequence, the explicit construction of further dense (h,k)free Boolean matrices [5, 26] has led to larger lower bounds.

Boolean sums, the Boolean matrix multiplication and the generalized Boolean matrix multiplication have some disjointness properties which the convolution does not have. Therefore, to prove a lower bound for the convolution, the situation becomes more difficult. The first approach for proving a lower bound for the monotone complexity of the convolution uses graphtheoretical properties of monotone networks realizing the convolution. Pippenger and Valiant [34] have studied shifting graphs and have proved that each monotone network for the convolution has to be a shifting graph obtaining an Ω(nlogn) lower bound for monotone complexity of the convolution. To prove a lower bound of size Ω(n4/3) for the number of ∧-gates needed in a monotone network which computes the convolution, the author [9] has introduced two further techniques into Boolean complexity. For the first time, a replacement rule changing the function computed at the output nodes of the network is used. An ∧-gate g such that the function computed at the output of the gate g has a certain property is replaced by 0. Therefore, the gate g is eliminated but, at the output nodes, the construction of some prime implicants could be destroyed. Because of the property of the function computed at the output of the gate g, the number of destroyed prime implicants is bounded. Only for inputs such that an output has to be one, a wrong value could be computed at the output node because of an application of the replacement rule. Such a replacement rule is of Type 2. To apply the replacement rule, we need that the monotone network has a certain structure. With respect to a given proof technique, we call a monotone network which allow the application of the proof technique a normal form network. Given any monotone network β computing a given monotone Boolean function, the network is transformed into normal form first and then, the corresponding proof technique is applied. Maybe, the transformation increases the size of the network such that this increase has to be taken into consideration to obtain a lower bound for the monotone complexity from the lower bound for a normal form network. Weiß [50] observed that on each path from the input node ai to an output node ck which depends on ai and has at least two prime implicants there has to be a first ∨-gate such that some ajbl with j 6= i is an implicant of the function. Using a replacement rule of Type 1, he showed that all these ∨-gates can be eliminated after setting ai to zero. Consider the assignment α which we obtain after setting with respect to each such an ∨-gate the variables of the prime implicant ajbl to one. Then all outputs of the network do not depend on ai. Therefore for each output function ck which depends on ai, the assignment α has to satisfy any prime implicant of ck. Since the Boolean convolution is semi-disjoint, each conjunction of a variable in A = {a0,a1,...,an−1} and a variable in B = {b0,b1,...,bn−1} is prime implicant of exactly one output function. Therefore, at most p2 prime implicants can be constructed if the assignment α is defined with respect to p first ∨-gates having the needed property. Since n output functions depend on ai, p2 has to be at least n such that an n3/2 lower bound for the number of ∨-gates needed in a monotone network which computes the convolution could be proved. Grinschuk and Sergeev [17] have constructed (h,k)-disjoint Boolean circulant matrices with many ones. The complexity of the corresponding set of Boolean sums is Ω(n2 log−6 n). Since circulant matrices are related to cyclic convolution and cyclic Boolean convolution can be reduced to Boolean convolution [17, 24], they obtain an Ω(n2 log−6 n) lower bound for the number of ∨-gates in a monotone network computing the Boolean convolution. Therefore, the used proof technique to prove the lower bound for the Boolean convolution was reduction.

Although since 1969 super-linear lower bounds for the monotone complexity of explicit functions in Mn,m where m = Θ(n) have been proved, before 1984, the largest lower bound for the monotone complexity of an explicit single output function was of size 4n [44]. All super-linear lower bound proofs for the monotone complexity of functions in Mn,m strongly depend on the property that a set of functions has to be computed. With respect to single output monotone Boolean functions, no technique for counting a super-linear number of gates has been developed before 1984. In 1984, Andreev [4, 6] and Razborov [36, 37] independently achieved the breakthrough. They have proved super-polynomial lower bounds for certain single output functions in NP. The functions resβ(g) computed at the gates g are replaced by a function which approximates resβ(g). The main point was the introduction of replacement rules which change the value of the function computed at the output node with respect to inputs in f−1(0) where f is the considered function. Such a replacement rule is of Type 3. The so-called approximation method was born. In the next section, we will describe the approximation method in detail.

4 The Approximation Method

Both, Andreev and Razborov have used set theoretical constructions to prove the lower bound. In a sence, this hides the effect of the approximation on the computation in the network. To understand this effect, we describe the approximation method for monotone networks directly on a monotone network which computes the function under consideration. For the extension of the approximation method to standard networks, this approach is more suitable than using a set theoretical construction as Razborov and Andreev.

To get a lower bound for the monotone complexity of a monotone function f ∈ Bn, we start with a monotone network β which computes f. We have no knowledge about the structure of β. In particular, we have no knowledge about the DNF-representations of the functions computed at the nodes of β. Let g1,g2,...,gt be the nodes of β numbered in any topological order. Starting with g1, the DNF-representations DNFβ(gi), 1 ≤ i ≤ t are treated in this order. The idea is to replace DNFβ(gi) by an approximation DNF’β(gi) such that we have the needed structural information. After the replacement, DNFβ(gj), j > i has to be updated such that for its construction DNF’β(gi) is used instead of DNFβ(gi). Therefore, not the function f but an approximation f′ of f is computed at the output node of β. Hence, there are inputs c ∈ {0,1}n such that f′(c) 6= f(c). Let gi be the last node for which DNFβ(gi) has been replaced by DNF’β(gi). In the subsequence, DNFβ(gj) denotes for j > i the DNF-representation of the current function computed at the node gj and for j ≤ i, DNFβ(gj) immediately before its replacement.

Let f1 (f2) denote the function computed at the output node gt after

the approximation of DNFβ(gi−1) (DNFβ(gi)) and before the approximation of DNFβ(gi) (DNFβ(gi+1)). We say that the approximator of the node gi introduces an error with respect to the input c ∈ {0,1}n if f1(c) = f(c) but f2(c) 6= f(c). Note that f′(c) 6= f(c) implies that there exists a node gi in β such that the approximator of gi introduces an error with respect to the input c. The approximators should be designed in a way such that the following is fulfilled:

1. After the replacement of DNFβ(gt), the number of inputs c ∈ {0,1}n with f′(c) 6= f(c) is “large”.

2. For all nodes gi, 1 ≤ i ≤ t, the number of inputs c ∈ {0,1}n where an error with respect to c is introduced by the replacement of DNFβ(gi) by DNF’β(gi) is “small”.

Note that these properties imply that a monotone network computing the function f has to contain “many” gates. How to approximate DNF-formulas such that these properties are fulfilled?

The general idea is to bound the size of the monomials in the DNFformulas constructed at the nodes in β. The size of a monomial can be its length; i.e., its number of distinct literals, or another measure. Let r be the upper bound for the size of a monomial in an approximator with respect to a node gi. An obvious way to bound the size of the monomials would be the following:

• For the construction of the approximator DNF’β(gi) construct the DNF-representation DNFβ(gi) of the current function computed at gi and remove each monomial of size larger than r.

The effect of the removal of monomials from DNFβ(gi) to the DNF-representation DNFβ(gt) of resβ(gt) is the removal of some monomials in DNDβ(gt). Hence, an error could be introduced only for inputs c ∈ f−1(1). Next we describe the construction of the approximators more in detail.

For input nodes, the approximator and the original DNF-representation are the same. For the comparision of the approximator DNF’β(gi) and the DNF-representation DNFβ(gi) of the current function computed at gi immediately before the approximation of DNFβ(gi) suppose that gi1 and gi2 are the direct predecessors of the gate gi. By construction, each monomial in DNF’ and in DNF’ has size at most r. Therefore, if gi is an ∨-gate, each monomial in the DNF-representation

DNFβ(gi) of the current function computed at gi has size at most r. Hence, we define DNF’β(gi) := DNFβ(gi). No error is introduced by the approximator DNF’β(gi). But the number of monomials in DNF’β(gi) could be the double of the number of monomials in DNF’β(gi1) or in DNF’β(gi2). If gi is an ∧-gate then

t1 t2

DNFβ(gi) = __(mj ∧ m′l).

j=1 l=1

To obtain DNF’β(gi), we remove from DNFβ(gi) all monomials mjm′l of size larger than r. To get a large lower bound for the function f ∈ Bn, the function f must have the following property:

F1 Only “few” inputs in f−1(1) fulfill a monomial of size larger than r.

If the number of monomials which are removed would be small enough then perhaps, we could prove an upper bound for the number of errors introduced by the approximation at an ∧-gate which is small enough. We need a mechanism which bounds the number of monomials removed during the construction of an approximator. Two such mechanisms are known, CNF-DNF-approximators which switch between CNF- and DNF-formulas and approximators which use the sunflower lemma discovered by Erd˝os and Rado [15]. We call such an approximator sunflower-approximator. Next, we review both approximators with respect to their use in monotone networks.

5 Approximators in Monotone Networks

Let f ∈ Bn be the monotone function for which we intend to prove a large lower bound of its monotone complexity. Let β = g1,g2,...,gt be a monotone network which computes f at its output node gt. To get a large lower bound, the approximation method has to take care that the following property is fulfilled:

A1 Only “few” monomials are removed to obtain DNF’β(gi) from DNFβ(gi).

The methods to obtain this property in CNF-DNF- and in sunflower-approximators are different. Since CNF-DNF-approximators are less difficult, we describe these approximators first.

5.1 CNF-DNF-Approximators

CNF-DNF-approximators are introduced implicitly by Haken [20] and explicitly by Jukna [22], Berg and Ulfberg [7] and Amano and Maruoka [3]. To understand the idea of CNF-DNF-approximators let us consider the organization of a CNF/DNF-switch as the construction of a tree T as described in Section 2. Obviously, the outdegree of an inner node w cannot be larger than the number of literals in the clause which is expanded at the leaf w during the construction of the tree T. Let m(w) denote the monomial associated with the path from the root to the node w. After the performance of the CNF/DNF-switch consider any path P from the root to a leaf w in T. Let v be any node on P. Obviously, the monomial m(v) is a prefix of the monomial of m(w). If we can ensure that on each edge starting from a node with outdegree at least two, the size of the corresponding monomial increases by one then the number of different prefixes of size exactly r could be bounded by l(k)r where l(k) is the maximal number of literals in a clause expanded during the CNF/DNF-switch. k will be an upper bound for the size of a clause used in the construction of an approximator. After the construction of T, all monomials of size larger than r are removed. Each such a monomial has a prefix of size r which has to be fulfilled by each input which fulfills the monomial. Therefore, an upper bound for the number of different such prefixes can be used in a lower bound proof. We call a CNF/DNF-switch followed by the elimination of all monomials of size larger than r an CNF/DNF-approximator switch. Analogously, if l(r) is the maximal number of literals in a monomial of size at most r expanded during a DNF/CNF-switch then we obtain an upper bound of l(r)k for the number of different prefixes of size exactly k. After the construction of T, all clauses of size larger than k are removed. Such a switch is called DNF/CNF-approximator switch.

This observation yields the idea to approximate with respect to each node gi in β DNFβ(gi) by a DNF-formula DNF’β(gi) which contains only monomials of size at most r and also CNFβ(gi) by a CNF-formula CNF’β(gi) which contains only clauses of size at most k. Note that by the removal of clauses from CNFβ(gi) only for inputs c ∈ f−1(0) an error could be introduced. To get a large lower bound for the monotone complexity of f using CNF-DNF-approximators, the function f must have the following additional property:

F2 Only “few” inputs in f−1(0) falsify a clause of size larger than k.

To get a large lower bound for the monotone complexity of f, the number of inputs for which CNF’β(gt) or DNF’β(gt) compute the wrong value has to be “large”. Therefore, the function f must have the following additional property:

F3 At least one of the following two properties is fulfilled:

1. If CNF’β(gt) is not the constant function one then “many” inputs in f−1(1) do not fulfill CNF’β(gt).

2. If DNF’β(gt) is not the constant function zero then “many” inputs in f−1(0) do not falsify DNF’β(gt).

Instead of using the whole sets f−1(1) and f−1(0), more appropriate subsets T1 ⊆ f−1(1) and T0 ⊆ f−1(0) could be used to prove the lower bound. Now we are prepared to give a precise description of CNF-DNFapproximators. We distinguish three cases.

Case 1: gi is an input node.

Then

CNF’β(gi) := CNFβ(gi) and DNF’β(gi) := DNFβ(gi).

Obviously, no error is introduced by both approximators.

Case 2: gi is an ∧-gate with direct predecessors gi1 and gi2.

Then

CNF’β(gi) := CNF’β(gi1) ∧ CNF’β(gi2).

Since the size of each clause in CNF’β(gi1) and in CNF’β(gi2) is at most k, each clause in CNF’β(gi) has also at most size k. Since CNF’β(gi) = CNFβ(gi), no error is introduced by the approximation.

DNF’β(gi) is obtained from CNF’β(gi) by performing a CNF/DNF-approximator switch.

Case 3: gi is an ∨-gate with direct predecessors gi1 and gi2.

Then

DNF’β(gi) := DNF’β(gi1) ∨ DNF’β(gi2).

Since the size of each monomial in DNF’β(gi1) and in DNF’β(gi2) is at most r, each monomial in DNF’β(gi) has also at most size r. Since DNF’β(gi) = DNFβ(gi), no error is introduced by the approximation.

CNF’β(gi) is obtained from DNF’β(gi) by performing a DNF/CNF-approximator switch.

For the application of CNF-DNF-approximators to a monotone function f ∈ Bn, to obtain a large lower bound for its monotone complexity, it has to exists large sets T1 ⊆ f−1(1) and T0 ⊆ f−1(0) of inputs such that the following properties are fulfilled:

1. There is a “small” upper bound n1 for the number of inputs in T1 which fulfill any monomial of size r + 1.

2. There is a “small” upper bound n0 for the number of inputs in T0 which falsify any clause of size k + 1.

3. At least one of the following two cases is fulfilled:

(a) There is a constant d1 < 1 such that a clause of size k is fulfilled by at most d1|T1| inputs in T1.

(b) There is a constant d0 < 1 such that a monomial of size r is falsified by at most d0|T0| inputs in T0.

To get a lower bound for the monotone complexity of f, the properties are used in the following way. Assume that the constant d1 exists. With respect to CNF’β(gt), two situations can arise. If CNF’β(gt) computes the constant function one then for no input c ∈ T0, the value f(c) is computed correctly. Therefore, for each input c ∈ T0 there is an ∨- gate gi such that the construction of CNF’β(gi) introduce an error with respect to the input c. Since at an ∨-gate, an error for at most l(r)kn0 inputs in T0 is introduced, we obtain the lower bound l(r|T)k0n| 0 for the monotone complexity of f. Otherwise, CNF’β(gt) contains a non-empty clause of size at most k. Therefore by the third property, CNF’β(gt) can be fulfilled by at most d1|T1| inputs in T1. Since at an ∧-gate, an error for at most l(k)rn1 inputs in T1 is introduced, we obtain the lower bound (1l−(kd)1r)n|T11| for the monotone complexity of f. The case that d0 exists can be discussed analogously.

5.2 Sunflower-Approximators

To bound the number of monomials which could be removed during the construction of an approximator, sunflower-approximators use the sunflower lemma or a modification of the sunflower lemma. Since CNF-DNF-approximators do not use such a combinatorial lemma, they seem to be simpler than sunflower-approximators. But the properties which a Boolean function must have to get a large lower bound for its monotone complexity using the methods are different. Both methods use the following property:

F1 Only “few” inputs in f−1(1) fulfill a monomial of size larger than r.

The properties F2 and F3 needed if we use CNF-DNF-approximators are replaced by two other properties. Therefore, the sets of functions for which the application of the methods results into a large lower bound for the monotone complexity may be different. Note that with respect to the perfect matching function, we know a proof of a super-polynomial lower bound which uses a sunflower-approximator [37] but no such a proof which uses a CNF-DNF-approximator. First of all, we will review the sunflower lemma and its use for proving a lower bound.

The sunflower lemma of Erd˝os and Rado [15] is the central combinatorial property to bound the number of monomials removed during the construction of an approximator. The basis of our description is the excellent presentation of Jukna in [23, 24].

A sunflower with p pedals and core T is a collection S1,S2,...,Sp of p sets such that Si ∩ Sj = T for 1 ≤ i < j ≤ p. Note that p pairwise disjoint sets is a sunflower with empty core. The following sunflower lemma means that each family of nonempty sets which is large enough must contain a sunflower with p pedals.

Lemma 2 Let F be a family of non-empty sets each of size at most r. If |F| > r!(p − 1)r then F contains a sunflower with p pedals.

If we relax the property that the core T lies entirely in all sets S1,...,Sp such that the differences Si \ T, 1 ≤ i ≤ p are non-empty and mutually disjoint then we obtain a lemma proved by Fu¨redi in 1978 [19]. The common part of p distinct finite sets S1,S2,...,Sp is the set T := Si6=j(Si ∩ Sj).

Lemma 3 Let F be a family of non-empty sets each of size at most r. If |F| > (p − 1)r then F contains p sets with common part of size less than r.

Razborov’s approximator [36, 37] are based on the sunflower lemma and on Fu¨redi’s lemma. Andreev [4, 6] uses his own modification of the sunflower lemma. No matter which modification of the sunflower lemma is used, the essential properties of the approximators are the same. Therefore, we only describe approximators which uses the sunflower lemma directly.

To use the sunflower lemma, a set S(m) of the same size as m is constructed for each monomial m. For example, if its length is the size of m, S(m) is the set of all variables in m. The set S(m) corresponds to the monomial m. The idea is to use an appropriate r as the upper bound for the size of a monomial and l := r!(p − 1)r as the upper bound for the number of monomials in the approximator DNF’β(gi), 1 ≤ i ≤ t. For the construction of the approximators, the nodes in β are considered in a topological order such that the approximators of the direct predecessors gi1 and gi2 are already constructed when DNF’β(gi) is constructed. Note that all monomials in DNF’β(gi1) and in DNF’β(gi2) have size at most r and the number of monomials in both approximators is at most l.

After the construction of DNFβ(gi) where gi is an ∨-gate, the number of monomials in DNFβ(gi) can exceed the upper bound l but is at most 2l. Then, by the sunflower lemma, there are p monomials m1,m2,...,mp such that the corresponding sets S(m1),S(m2),...,S(mp) form a sunflower. The core T of the sunflower corresponds to a monomial m(T). Then in DNFβ(gi), the monomials m1,m2,...,mp are replaced by the single monomial m(T) which is a submonomial of each monomial mj, 1 ≤ j ≤ p. This operation is called a plucking. The effect of a plucking to DNFβ(gt) is the replacement of some monomials by a proper submonomial. Hence, a plucking can only introduce an error for inputs in f−1(0). As long as possible pluckings are performed leading to the approximator DNF’β(gi). By the sunflower lemma, the number of monomials in DNF’β(gi) is at most l. Since at the beginning, the number of monomials is at most 2l, less than 2l pluckings are performed.

After the construction of DNFβ(gi) where gi is an ∧-gate, the size of some monomials can exceed the upper bound r. We remove all these monomials first. Since DNF’β(gi1) and also DNF’β(gi1) contain at most l monomials, at most l2 monomials are removed. Then, the plucking procedure is applied to the remaining monomials obtaining the approximator DNF’β(gi). Since the approximators of the direct predecessors of gi contain at most l monomials, at most l2 pluckings are performed.

For the application of sunflower-approximators to a monotone function f ∈ Bn, to obtain a large lower bound for its monotone complexity, it has to exist large sets T1 ⊆ f−1(1) and T0 ⊆ f−1(0) of inputs such that the following properties are fulfilled:

1. There is a “small” upper bound n1 for the number of inputs in T1 which fulfill any monomial of size larger than r.

2. There is a “small” upper bound n0 for the number of inputs in T0 for which an error is introduced because the performance of a plucking.

3. At least one of the following two cases is fulfilled:

(a) There is a constant d1 < 1 such that DNF’β(gt) computes the constant function one or DNF’β(gt) is satisfied by at most d1|T1| inputs in T1.

(b) There is a constant d0 < 1 such that DNF’β(gt) computes the constant function zero or DNF’β(gt) is falsified by at most d0|T0| inputs in T0.

To get a lower bound for the monotone complexity of the function f, sunflower approximators are used in the following way. Assume that the constant d1 exists. If DNF’β(gt) computes the constant function one then for no input c ∈ T0, the value f(c) is computed correctly. Only pluckings introduce an error for an input in T0. Since a single plucking introduces an error for at most n0 inputs in T0, at least |nT00| pluckings are performed. Since at most l2 pluckings are performed at a gate in β, we obtain the lower bound l|2Tn00| for the monotone complexity of f. Otherwise, there are (1 − d1)|T1| inputs in T1 which do not satisfy DNF’β(gt). Only the removal of a monomial at an ∧-gate can introduce an error for an input in T1. The removal of one monomial of size larger than r can introduce an error for at most n1 inputs in T1. Hence, at least (1−dn11)|T1| monomials are removed. Since at most l2 monomials are removed at an ∧-gate, we obtain the lower bound (1−ld21n)1|T1| for the monotone complexity of f. The case that d0 exists can be discussed analogously.

6 The Extension of the Approximation Method

Our goal is to extend the approximation method such that it can be used to prove a super-polynomial lower bound for the standard complexity of an appropriate Boolean function f ∈ Bn or to understand why this is not possible. Let β = g1,g2,...,gt be a standard network which computes a function f ∈ Bn at its output node gt. As in monotone networks our goal is to approximate the DNF-formulas constructed at the nodes in β; i.e., we replace DNFβ(gi) by an approximator DNF’β(gi). Exactly as in monotone networks, we define the notion that DNF’β(gi) intoduces an error with respect to the input c ∈ {0,1}n. Again, the general idea is to bound the size of the monomials in the DNF-formulas constructed at the nodes of β. In contrast to monotone networks, a monomial in DNFβ(gi) can contain both kinds of literals. Hence, with respect to the definition of the size of a monomial, we have two possibilities:

1. The size of a monomial depends on both kinds of literals.

2. The size of a monomial depends only on one kind of literals.

Before discussing both cases, let us review the needed properties of the function f such that a large lower bound could be proved for f using the approximation method. Since the approximation method obtains DNF’β(gi) by the removal of all monomials of size larger than a given bound r, the first property is that the number of inputs c ∈ T1 which fulfill a monomial of size larger than r is small enough. Since at the output node gt, the value of many inputs c ∈ T1 ∪ T0 has to be computed incorrectly, the second property is that the number of inputs c ∈ T1 ∪T0 with DNF’β(gt)(c) 6= f(c) is large enough. The approximation method takes care that the number of monomials which are removed from DNFβ(gi) to obtain DNF’β(gi) is small enough. This is done in the following way:

CNF-DNF-approximators:

At each gate gi, the CNF-formula CNFβ(gi) is approximated by a CNFformula CNF’β(gi) as well. CNF’β(gi) is obtained from CNFβ(gi) by the removal of all clauses of larger size than a given bound k. Therefore, the additional property that the number of inputs c ∈ T0 which falsify a clause of size larger than k is small enough is needed. In dependence of r and k, upper bounds hm(r) and hc(k) of the number of variables in a monomial of size at most r and in a clause of size at most k, respectively are derived. Then, upper bounds h1(k)r and h2(r)k for the number of different prefixes of size exactly r of the monomials removed at an ∧-gate gi and the number of different prefixes of size exactly k of the clauses removed at an ∨-gate gi are estimated. To get a large lower bound, p := h1(k) has to be non-constant.

Sunflower-approximators:

To bound the number of monomials removed from DNFβ(gi) to obtain DNF’β(gi), sunflower-approximators use parameters 2 ≤ r,p ≤ n where r is an given upper bound for the size of the monomials in DNF’β(gi) and p is the number of pedals with respect to the sunflower lemma. Using the sunflower lemma, the number of monomials in the approximators is bounded by l := r!(p−1)r. Hence, an upper bound of l2 for the number of monomials removed at an ∧-gate gi is obtained. To get a large lower bound, p has to be non-constant.

Now we will discuss the case that the size of a monomial depends on both kind of literals. We will give some evidence that in this case, the approximation method cannot be extended to prove a super-linear lower bound for the standard complexity of any Boolean function f ∈ Bn.

As described above, with respect to each known approximation method, the upper bound for the number of monomials removed at an ∧-gate gi for obtaining DNF’β(gi) from DNFβ(gi) is at least pr where r is the upper bound for the size of the monomials in DNF’β(gi) and p is non-constant. In the case that both kinds of literals are approximated, such an upper bound seems to be too large. We will explain this for sizes which depend on the number of negated and the number of non-negated variables in a monomial

m. Let r0 (r1) be the upper bound for the number of negated (non-negated) variables of a monomial m in DNF’β(gi); i.e., each monomial which does not fulfill both bounds is removed from DNFβ(gi). Let r := r0 + r1. Note that each monomial of length larger than r cannot fulfill both bounds. The following lemma shows that that 2r monomials of length r are sufficient such that each input c ∈ f−1(1) fulfills at least one of these monomials.

Lemma 4 Let f be any Boolean function in Bn. There are 2r monomials of length r such that each input c ∈ f−1(1) fulfills at least one of these monomials.

Proof: Fix any r variables xi1,xi2,...,xir. Consider any c ∈ f−1(1). Let

where for 1 ≤ j ≤ r

By construction, . There are 2r monomials which

use exactly the variables xi1,xi2,...,xir.

Note that each submonomial of is fulfilled by c as well. At most 2r monomials of length at most r could suffice such that their removal could introduce an error for each input c ∈ f−1(1). Since pr > 2r for non-constant p, the approximation of the DNF-formula with respect to one ∧-gate could destroy the correct computation of f(c) for all c ∈ f−1(1). This shows that the first property seems not be fulfilled if both kind of literals are approximated.

It remains the consideration of the case that the size of a monomial depends only on one kind of literals. We discuss CNF-DNF-approximators first.

6.1 Extended CNF-DNF-Approximators

We consider the subcase that the non-negated variables are approximated. The other subcase can be discussed in the same way. Let β = g1,g2,...,gt be a standard network which computes a Boolean function f ∈ Bn at its output node gt. The sizes of a monomial m or of a clause d depend only on its non-negated variables. This implies that any number of negated variables can be contained in each monomial m in DNF’β(gi) and also in each clause d in CNF’β(gi). Since β computes the function f at its output node gt, before any approximation, DNFβ(gt) contains for each c ∈ f−1(1) a monomial mc such that mc(c) = 1. Only for c = (1,1,... ,1), we can exclude that mc contains any negated variable. Furthermore, CNFβ(gt) contains for each c ∈ f−1(0) an f-clause dc such that dc(c) = 0. Before any approximation, each f-clause d in CNFβ(gt) contains a literal in mc for each c ∈ f−1(1).

We have to define large sets T1 ⊆ f−1(1) and T0 ⊆ f−1(0) of inputs such that the following properties are fulfilled:

1. There is a “small” upper bound n1 for the number of inputs in T1 which fulfill any monomial of size r + 1.

2. There is a “small” upper bound n0 for the number of inputs in T0 which falsify any clause of size k + 1.

3. At least one of the following two cases is fulfilled:

(a) There is a constant d1 < 1 such that a clause of size k is fulfilled by at most d1|T1| inputs in T1.

(b) There is a constant d0 < 1 such that a monomial of size r is falsified by at most d0|T0| inputs in T0.

But without any further information about the structure of the network it seems to be impossible to fulfill the third property without destroying the first or the second property. Since the third property divides into two cases, we have to consider the two situations that CNF’β(gt) is not the constant function one and that DNF’β(gt) is not the constant function zero.

If CNF’β(gt) is not the constant function one then CNF’β(gt) must contain at least one clause d. Note that d is a subclause of any f-clause d′. Let pc(d) be a prime clause which is a subclause of the f-clause d′. For each c ∈ f−1(1) such that c fulfills a literal contained in d, it cannot be excluded that f(c) is computed correctly by CNF’β(gt). Therefore, for each c ∈ f−1(1) such that d is not satisfied by c, cj = 1 if ¬xj is contained in the clause d. d can contain each ¬xj where the variable xj is not contained in pc(d). To get the property that a clause of size k could be fulfilled by at most d1|T1| inputs in T1, the set T1 should only contain inputs c ∈ f−1(1) such that there is a prime clause d(c) with cj = 1 for all xj not contained in d(c). But for such an input c ∈ T1, a monomial of size r+1 which is fulfilled by c must not be a submonomial of a prime implicant which is fullfilled by c. Therefore, a small upper bound n1 for the number of inputs in T1 which fulfill any monomial of size r + 1 cannot be proved in the usual way. Note that usually, the set T1 contains exactly those inputs which correspond to the prime implicants of the function. Because of the structure of the prime clauses of the considered functions, I have found no other way to prove a small upper bound.

If DNF’β(gt) is not the constant function zero then DNF’β(gt) must contain at least one monomial m. Note that m is a submonomial of an implicant m′ of f. Let pi(m) be a prime implicant which is a submonomial of the implicant m′. For each c ∈ f−1(0) such that c falsifies a literal contained in m, it cannot be excluded that f(c) is computed correctly by DNF’β(gt). Therefore, for each c ∈ f−1(0) such that m is satisfied by c, cj = 0 if ¬xj is contained in the monomial m. m can contain each ¬xj where the variable xj is not contained in pi(m). To get the property that a monomial of size r could be falsified by at most d0|T0| inputs in T0, the set T0 should only contain inputs c ∈ f−1(0) with the property that there is a prime implicant p(c) such that cj = 0 for all xj not contained in p(c). For such a set T0, a small upper bound n0 for the number of inputs in T0 which falsify any clause of size k + 1 cannot be proved in the usual way. Because of the structure of the prime implicants of the considered functions, I have found no other way to prove a small upper bound.

This gives evidence that with respect to CNF-DNF-approximators, the needed properties cannot be fulfilled if only one kind of literals is approximated and no other properties of the network are used.

Altogether, we have given some evidence that extended CNF-DNF-approximators alone cannot be used to prove a lower bound for the standard complexity of any Boolean function f ∈ Bn.

6.2 Extended Sunflower-Approximators

Now, we shall investigate the expandability of sunflower-approximators. Unlike CNF-DNF-approximators which have an obvious extension, we have to explain the extension of sunflower-approximators. Given any standard network β computing the considered function f ∈ Bn, we will use extended sunflower-approximators for the approximation of DNFβ(g) where g is a node in β. Hence, we separate in each monomial mj of DNFβ(g) the negated and the non-negated variables obtaining

mj = mj0mj1

where mj0 contains exactly the negated variables and mj1 contains exactly the non-negated variables of mj. If mj does not contain any negated (nonnegated) variable then mj0 = ε (mj1 = ε). After doing this, we write for DNF

s

DNFβ(g) = _ mj0mj1.

j=1

We describe the extension of sunflower-approximators for the case that the non-negated variables are approximated. The approximation of the negated variables can be done analogously. The construction is more complicated than in the monotone case. The difficulties come from the incorporation of the negated part of the monomials which we do not approximate. We consider the case that the size of the non-negated part m1 of a monomial m = m0m1 is its length; i.e., the number of different variables in m1. For the application of the sunflower lemma, we use for each monomial m = m0m1 the set V (m1) of variables contained in m1.

Let g1,g2,...,gt be the nodes of β numbered in any topological order. In dependence of the approximators corresponding to the direct predecessors of the considered node gi and the operation opi of gi, we define the approximator of DNF ) denotes the set of monomials consisting of only non-negated (negated variables) from Vn including the empty monomial. m denotes always a monomial in . First of all, we describe the general structure of the approximators. The approximator defined for the DNF-formula of a function computed at a node gi in β consists of

1. a set M(gi) ⊆ Mn of monomials each of size at most r,

2. for each monomial m ∈ M(gi), a set .

Then, the approximator DNF’β(gi) is defined by

DNF’β(gi) := _ i _i mm.

m∈M(g ) m∈D (m)

Consider i ≥ 1. Assume that the approximators with respect to gj, j < i are already defined. Now we define the approximator with respect to gi. In dependence of the kind of gi, we distinguish four cases.

Case 1: op(gi) = xj, j ∈ {1,2,... ,n}.

Let

M(gi) := {xj} and Di(xj) := {ε}.

Then

DNF’β(gi) := εxj.

Obviously, xj and DNF’β(gi) compute the same function.

Case 2: op(gi) = ¬xj, j ∈ {1,2,... ,n}.

Let

M(gi) := {ε} and Di(ε) := {¬xj}.

Then

DNF’β(gi) := ¬xjε.

Obviously, ¬xj and DNF’β(gi) compute the same function.

Case 3: gi is an ∨-gate with direct predecessors gi1 and gi2.

Before the performance of any plucking we have

M(gi) := M(gi1) ∪ M(gi2)

and for each m ∈ M(gi)

Di1(m) ∪ Di2(m) if m ∈ M(gi ) ∩ M(gi2),

,

Di2(m) if m 6∈ M(gi1).

Now we explain the effect of plucking with respect to these sets. Let m1,...,mp be p monomials where the corresponding sets V (m1),...,V (mp) form a sunflower with core T. A plucking replaces the monomials m1,...,mp by the monomial m(T) consisting of the variables in T. Hence, the following operations are performed:

,

and

M(gi) := M(gi) \ {m1,m2,...,mp} ∪ {m(T)}.

Case 4: gi is an ∧-gate with direct predecessors gi1 and gi2. Before the performance of any plucking, we have

M(gi) := {mm′ | m ∈ M(gi1),m′ ∈ M(gi2) and mm′ has size ≤ r}.

For each m ∈ M(gi) for all m1 ∈ M(gi1), m2 ∈ M(gi2) such that m = m1m2, we define

Di(m,m1,m2) := {mm′ | m ∈ Di1(m1) and m′ ∈ Di2(m2)} and

Di(m) := 1 i1 2[ i2 1 2 Di(m,m1,m2). m ∈M(g ),m ∈M(g ):m m =m

Then these sets are modified because of the performed pluckings as described in Case 3. This finishs the description of the extended approximators.

A sunflower-approximator for standard networks consists of two components. One component is the approximated part, the other component the non-approximated part. The combinatorial properties of the approximated part are the same as in sunflower-approximators for monotone Boolean networks. With respect to the non-approximated part, no combinatorial properties usable in a lower bound proof can be extracted without any further information about the structure of the network. Moreover, the nonapproximated part causes that arguments used in the lower bound proof for the monotone complexity do not work now.

The main problem is caused by the performance of pluckings. A necessary property is that the number of inputs in T0 for which a plucking introduces an error can be bounded. Let m1,m2,...,mp be p monomials where the corresponding sets V (m1),...,V (mp) form a sunflower with core T. With respect to monotone networks, an input c ∈ T0 for which an error is introduced because of the performance of the corresponding plucking cannot fulfill any monomial in {m1,m2,...,mp}. Otherwise, the error with respect to c would be already exist before the performance of the plucking. Exactly this fact is used to get the needed upper bound. But with respect to standard networks, an input c ∈ T0 for which an error is introduced because of the performance of the corresponding plucking can fulfill some monomials in {m1,m2,...,mp}. This comes from the non-approximated part of the monomials in the approximators. To see this consider c ∈ T0 for which an error is introduced. Then there is m ∈ Di(m(T)) such that c fulfills mm(T). By construction, there is l ∈ {1,2,... ,p} such that m ∈ Di(ml). Obviously, c cannot fulfill the monomial ml. Otherwise, c would fulfill mml such that the plucking does not introduce an error with respect to the input c. But with respect to each j ∈ {1,2,... ,p} with m ∈6 Di(mj) c could fulfill the monomial mj if each monomial in Di(mj) contains a literal not fulfilled by c. Therefore, we have to estimate an upper bound for the number of inputs in c ∈ T0 with

1. c satisfies a monomial mm(T) where m ∈ Di(m(T)) and

2. for all 1 ≤ j ≤ p, c does not fulfill mj or each monomial in Di(mj) contains a literal not fulfilled by c.

Without any knowledge about the structure of the monomials in the nonapproximated part of the approximators, I see no way to prove an upper bound for such inputs in T0 which is small enough. This gives evidence that sunflower-approximators alone cannot be used to prove a super-linear lower bound for the standard complexity of any Boolean function f ∈ Bn.

7 What should be done next?

To prove a super-linear lower bound for the standard complexity of any explicit Boolean function, we need more knowledge about the use of negations in a non-monotone Boolean network. Essential for the lower bound proofs for the monotone complexity of a Boolean function is the following property: In a monotone Boolean network each prime implicant of the function has to be constructed at the corresponding output node. A standard network computing a Boolean function must not have this property. Instead of constructing a prime implicant p at the output node, a set m1,m2,...,mr of monomials such that

1. p is a submonomial of each monomial; i.e, and

2.

could be constructed. To prove a lower bound for the standard complexity of a Boolean function, we have to prove a lower bound for the number of gates needed for the construction of such a DNF-representation of the function. The problem is that by a standard network, each algorithm for the computation of a solution could be realized. To clarify this by an example, let us consider the Boolean function f where the input variables encode an undirected graph on n nodes and f is one iff the input graph is a k-clique; i.e., the graph consists of a k-clique and n − k isolated nodes. Note that f is non-monotone and each prime implicant of f has a literal with respect to each possible edge. For a given graph G = (V,E), it is easy to decide if G is a k-clique. G is a k-clique iff G has exactly k nodes with degree k − 1 and in total k(k − 1) edges where each edge is counted with respect to each end node. f is the exact-clique function. For checking if exactly l of m variables are one, we can use a non-monotone network of size O(m) [49, Chapter 3.4]. Therefore, f can be computed by a standard network of linear size. If we relax the definition of f and we define that f is one iff the input graph G contains a k-clique, the situation changes dramatically. Now f is the NP-complete clique function and it is an open problem if for f a standard network of polynomial size exists. Each deterministic algorithm for the solution of the clique problem can be realized by a standard network such that the size of the network is polynomial in the time used by the algorithm.

To prove a lower bound for the size of a standard network which computes a given Boolean function f, we can only use the structure of the function f. We need an intuition which properties of the prime implicants or the prime clauses of a given function forces a standard network to use a certain amount of gates. How to get such an intuition? The structure of the prime implicants of the exact-clique function tells us directly how we can realize the function by a standard network of linear size. The knowledge about upper bounds may help us to get an intuition what make the function easy or difficult. Hence, before looking for properties which enable us to prove a lower bound, I would look for upper bounds for the function under consideration. Which functions should be chosen for the try to get the first proof of a super-linear lower bound for the non-monotone complexity of an explicit Boolean function?

As for monotone networks, I would first try to obtain a super-linear lower bound for a Boolean function with many outputs as the Boolean convolution, the Boolean matrix multiplication or (1,1)-disjoint Boolean sums. A nonconstant lower bound with respect to each output would result in a superlinear lower bound for the function. Firstly, upper bounds for the chosen function should be investigated to learn something about the usefulness of negations with respect to the considered function. Can we adapt any method developed for monotone networks to obtain a super-linear lower bound? The paper of Weiß [50] could be a good starting point.

How to proceed the work with respect to the P versus NP problem? Currently, I am convinced that we are far away to prove a super-polynomial lower bound for the non-monotone complexity of any explicit Boolean function. On the other hand, the strongest barrier towards proving P 6= NP could be that it holds P = NP. To ensure that the whole time spent for working on the P versus NP problem is not used to prove an impossible theorem, I would switch to the try to develop a polynomial algorithm for the solution of an NP-complete problem. Moreover, also in the case that P 6= NP, understanding why it is not possible to develop a polynomial algorithm for the solution of the considered NP-complete problem could be of help to prove a lower bound for the standard complexity of the corresponding Boolean function. What kind of NP-complete problem should be chosen? I think that a good candidate would be an NP-complete optimization problem for the following reasons.

A general approach for the solution of an optimization problem is the following: Start with a feasible solution of the optimization problem under consideration. As long as possible apply to the current feasible solution an improvement step. The improvement step replaces a part of the current solution by a part which is outside of the current solution such that the obtained solution is feasible and the value of the objective function is improved. To get a polynomial algorithm for the optimization problem, the following properties should be fulfilled:

1. A suboptimal feasible solution always allows the application of an improvement step,

2. after a polynomial number of improvement steps, an obtimal solutionis obtained, and

3. an improvement step can be performed in polynomial time.

A well known optimization problem where this approach has led to a polynomial time algorithm is the maximum matching problem. Let G = (V,E) be an undirected graph. M ⊆ E is a matching if no two edges in M have a common node. A matching M is maximal if there is no edge e ∈ E \ M such that M ∪{e} is a matching. A matching M is maximum if there exists no matching M′ ⊆ E of larger size. A maximal matching M is minimum if there is no maximal matching M′ of G such that |M′| < |M|. Given an undirected graph G = (V,E), the maximum matching problem is finding a maximum matching M ⊆ E. The minimum maximal matching problem is finding a minimum maximal matching M ⊆ E. Note that the minimum maximal matching problem is NP-complete [16]. Let M ⊆ E be a matching of G. A node v ∈ V is M-free iff v is not incident to an edge in M.

In 1891, Peterson introduced the technique of alternating paths. A path P = v0,v1,...,vk is M-alternating if it contains alternately edges in M and in E \ M. Let P = v0,v1,...,vk be a simple M-alternating path. P is M-augmenting if v0 and vk are M-free. M ⊕ P denotes the symmetric diffence of M and P; i.e., M ⊕ P = M \ P ∪ P \ M. If P is an Maugmenting path then M ⊕ P is a matching of G, and |M ⊕ P| = |M| + 1. In 1957, Berge proved that a matching M ⊆ E is maximum iff there is no M-augmenting path in G. Until 1963, for non-bipartite graphs only exponential time algorithms for the construction of M-augmenting paths has been known. Then in 1963, Edmonds [14] has shown how to construct an M-augmenting path in a non-bipartite graph in polynomial time if an M-augmenting path exists. This resulted in a polynomial algorithm for the maximum matching problem.

Berge’s characterization theorem has resulted in the construction of an improvement step. Edmonds has shown how this improvement step can be performed in polynomial time. If we try to apply an analogous approach to an NP-complete optimization problem, we need such a problem which allows the proof of a character

ization theorem which can be used for the construction of an improvement step. Then we can try to develop a polynomial implementation of the improvement step. I think that the minimum maximal matching problem could be a good candidate for such an NP-complete problem.