1

A 3-CNF-SAT descriptor algebra and the

solution of the P =NP conjecture
Prof. Marcel Rémon and Dr. Johan Barthélemy

Abstract
arXiv:1609.05709v1 [cs.CC] 25 Jul 2016

The relationship between the complexity classes P and NP is an unsolved question in the

field of theoretical computer science. In this paper, we investigate a descriptor approach

based on lattice properties. In a previous paper, we tried to prove that neither P 6= NP

nor P = NP was “unprovable” within the a-temporal framework of Mathematics. See [4].

A part of the proof about the impossibility to prove that P = NP turns to be inexact, and

yields the first author to investigate deeper into the possibility of P = NP .

This paper proposes a new way to decide the satisfiability of any 3-CNF-SAT problem. The

analysis of this exact [non heuristical] algorithm shows a strictly bounded exponential

complexity. The complexity of any 3-CNF-SAT solution is bounded by O(2490 ). This

[over-estimated] bound is reached by an algorithm working on the smallest description (via

descriptor functions) of the evolving set of solutions in function of the already considered

clauses, without exploring these solutions. Any remark about this paper is warmly welcome.

Index Terms

Algorithm Complexity, P − NP problem, 3-CNF-SAT problem

I. The 3-CNF-SAT problem

Boolean formulae are built in the usual way from propositional variables xi and the logical

connectives ∧, ∨ and ¬, which are interpreted as conjunction, disjunction, and negation,

respectively. A literal is a propositional variable or the negation of a propositional variable,

and a clause is a disjunction of literals. A Boolean formula is in conjunctive normal form if

and only if it is a conjunction of clauses.

M.Rémon, Department of Mathematics, Namur University, Belgium; [email protected]
J.Barthélemy, SMART Infrastructure Facilities, University of Wollongong, Australia; [email protected]

January 6, 2020—8 : 41 am DRAFT

A 3-CNF-SAT DESCRIPTOR ALGEBRA AND THE SOLUTION OF THE P =NP CONJECTURE 2

A 3-CNF formula ϕ is a Boolean formula in conjunctive normal form with exactly three

literals per clause, like ϕ := (x1 ∨ x2 ∨ ¬x3 ) ∧ (¬x2 ∨ x3 ∨ ¬x4 ) := ψ1 ∧ ψ2 . A 3-CNF formula

is composed of n propositional variables xi and m clauses ψj .

The 3-CNF-satisfiability or 3-CNF-SAT problem is to decide whether there exists or not

logical values for the propositional variables, so that ϕ can be true. Until now, we do not

know whether it is possible or not to check the satisfiability of any given 3-CNF formula ϕ

in a polynomial time with respect of n, as the 3-CNF-SAT problem is known to belong to

the class NP of problems. See [2] for details.

II. A matrix representation of a 3-CNF formula

A. Definitions

The size of a 3-CNF formula ϕ is defined as the size of the corresponding Boolean circuit,

i.e. the number of logical connectives in ϕ. Let us note the following property :

size(ϕ) = O(m) = O(α × n) (1)

where α = m/n is the ratio of clauses with respect to variables. It seems that α ≈ 4.258

gives the most difficult 3-CNF-SAT problems. See [3].

Let ϕ(x1 , x2 , · · · , xn ) be a 3-CNF formula. The set Sϕ of all satisfying solutions is

Sϕ = {(x1 , · · · , xn ) ∈ {0, 1}n | ϕ(x1 , · · · , xn ) = 1} (2)

Let Σϕ = # Sϕ and s̄1 , · · · , s̄Σϕ be the ordered elements of Sϕ . For 1 ≤ j ≤ Σϕ : s̄j =

(s1j , · · · , sij , · · · , snj ). We define the 3-CNF-matrix representation of ϕ as [ϕ] :

 
x1 xi xn
 1 
 s1 ··· sn1 
[ϕ] =  . (3)
 
 .. .. 
 sij . 

s1Σϕ ··· snΣϕ

January 6, 2020—8 : 41 am DRAFT

A 3-CNF-SAT DESCRIPTOR ALGEBRA AND THE SOLUTION OF THE P =NP CONJECTURE 3

B. Examples

Each clause ψi will be represented by a 7 × 3 matrix. For example,

   
x1 x2 x3 x2 x3 x4
   
 0 0 0   0 0 0 
   
   
 0 1 0   0 0 1 
  
   
 0 1 1   0 1 0 
[ψ1 ] = [(x1 ∨ x2 ∨ ¬x3 )] = 

 and [ψ2 ] = [(¬x2 ∨ x3 ∨ ¬x4 )] = 
 
 1 0 0  0 1 1 


   
 1 0 1   1 0 0 
  
   
 1 1 0   1 1 0 
  
1 1 1 1 1 1
The 3-CNF formula ϕ = ψ1 ∧ ψ2 will be represented by a 12 × 4 matrix :
 
x1 x2 x3 x4
 
 0 0 0 0 
 
 
 0 0 0 1 
 
 0 1 0 0 
 
 
 0 1 1 0 
 
 
 0 1 1 1 
 
[ϕ] = [(x1 ∨ x2 ∨ ¬x3 ) ∧ (¬x2 ∨ x3 ∨ ¬x4 )] =  1 0 0 0 
 
 
 1 0 0 1 
 
 
 1 0 1 0 
 
 
 1 0 1 1 
 
 1 1 0 0 
 
 
 1 1 1 0 
 
1 1 1 1
This paper defines an algebra on this type of matrices such that [ϕ] = [ψ1 ] ∧ [ψ2 ].

III. First definitions and properties for 3-CNF-matrices

A. Extension to new variables

Let A be such a matrix, A can be extended to new propositional variables by adding columns

filled with the neutral sign “.”, meaning that the corresponding variable can be set either to

0 or 1. This new matrix A is equivalent to A.

   
x1 x2 x4 x1 x2 x3 x4
   
 a1 a21 a41   a11 a21 .[01 ] a41 
A= 1 ≡ =A (4)
   
 a1 a2j a4j   a1j a2j . a4j 
 j   
a1Σϕ a2Σϕ a4Σϕ a1Σϕ a2Σϕ . 4
aΣϕ
January 6, 2020—8 : 41 am DRAFT
A 3-CNF-SAT DESCRIPTOR ALGEBRA AND THE SOLUTION OF THE P =NP CONJECTURE 4

B. Reduction of 3-CNF-matrices

The inverse operation, called reduction, replaces two same lines only differing by a 0 and a

1 for a variable, by a unique line with a neutral sign for this variable :
 
x1 x2 x3
 
 0 0 0 
  
x1 x2 x3
  
 
 0 0 1 
   x1 x2 x3
0 0 . 

    
 0 1 0  
≡
  0 . . 
A= 0 1 . ≡ (5)
  
  
0 1 1     1 0 . 


  
 1 0 . 

 
 1 0 0  1 1 0
1 1 0
 
 

 1 0 1 

1 1 0

C. Disjunction of 3-CNF-matrices

Let A and B be two matrices and {x1 , · · · , xn } the union of their support variables. Let A

and B be their extensions over {x1 , · · · , xn }. Then we define the disjunction of A and B by
 
x1 · · · xn
 
A∨B = 
 A 
 (6)
B

Of course, this new matrix should be reordered so that the lines are in a ascending order,

which can yield sometimes in replacing a line with a neutral sign by two lines with a one

and a zero.

D. Block decomposition of 3-CNF-matrices

Let A a matrix such that the reduction process yields to lines with neutral sign, then A can

be rewritten as the disjunction of smaller matrices. For example,

 
x1 x2 x3
 

 0 0 0 

 

 0 1 0 

  ! ! !
 0 1 1  x1 x1 x2 x1 x2 x3
[ψ1 ] =  = ∨ ∨
1 0 0  1 0 1 0 0 0
 

 

 1 0 1 

 

 1 1 0 

1 1 1
January 6, 2020—8 : 41 am DRAFT
A 3-CNF-SAT DESCRIPTOR ALGEBRA AND THE SOLUTION OF THE P =NP CONJECTURE 5

The block decomposition is not unique. For instance, there are 6 different block decompo-

sitions for a 3-variables clause.

E. Conjunction of 3-CNF-matrices

Let A and B be two matrices, A and B their extensions to the joint set of propositional

variables. Let Ak and B l be the one line matrices such that :

ΣA ΣB
_ _
A= Ak and B = Bl (7)
k=1 l=1

We define the conjunction of A and B as

   
ΣA ΣB ΣA ΣB ΣA ΣB
_ _ _ _
_ _
A∧B ≡ A∧B =  Ak  ∧  Bl = Ak ∧ B l = C k,l (8)
k=1 l=1 k=1 l=1 k=1 l=1

where

i
 ∅ if ∃ cm = “NaN”
! ! 

x1 xi xn x1 xi xn !
C k,l = ∧ = x1 xi xn (9)
a1k aik ank b1l bil bnl  otherwise
c1 ci cn


m m m

and



 aik if aik = bil

 ai if ai 6= bi and bi = “ · ”

k k l l
cim = (10)
i i i i


 b l if ak 6
= b l and a k = “·”


 “NaN” otherwise

F. The empty and full 3-CNF-matrices

Let us call ∅, the empty matrix, with no line at all. The empty matrix is neutral for the

disjunction operator ∨ and absorbing for the conjunction operator ∧.

Let us define Ω, the full matrix, as a one line matrix with only neutral signs in it. The full

matrix is neutral for ∧ and absorbing for ∨.

January 6, 2020—8 : 41 am DRAFT

A 3-CNF-SAT DESCRIPTOR ALGEBRA AND THE SOLUTION OF THE P =NP CONJECTURE 6

G. Example of operations

Let us consider the following block decompositions for [ψ1 ] and [ψ2 ] with x2 and (x2 x3 )

as common supports.
 
x x2 x3
 1 
 0 0 0 
 
 
 0 1 0 
  ! ! !
 
 0 1 1  x2 x2 x3 x1 x2 x3
[ψ1 ] =  = ∨ ∨
 1 0 0  1 0 0 1 0 1
 
 
 1 0 1 
 
 
 1 1 0 
 
1 1 1
and
 
x2 x3 x4
 

 0 0 0 

 

 0 0 1 

  ! ! !
 0 1 0  x2 x2 x3 x2 x3 x4
[ψ2 ] =  = ∨ ∨
0 1 1  0 1 1 1 0 0
 

 

 1 0 0 

 

 1 1 0 

1 1 1

[ψ1 ] ∧ [ψ2 ]
! ! ! !
x2 x3 x2 x3 x4 x2 x3 x1 x2 x3
= ∅∨ ∨ ∨ ∨∅∨∅∨ ∨∅∨∅
1 1 1 0 0 0 0 1 0 1
 
x1 x2 x3 x4
 
 . 0 0 . 
 
= . 1 0 0 
 
 
 . 1 1 . 
 
1 0 1 .

H. Lattice structure of 3-CNF-matrices

A semi-lattice (X, ∨) is a pair consisting of a set X and a binary operation ∨ which is

associative, commutative, and idempotent.

January 6, 2020—8 : 41 am DRAFT
A 3-CNF-SAT DESCRIPTOR ALGEBRA AND THE SOLUTION OF THE P =NP CONJECTURE 7

Let us note A the set of all the 3-CNF-matrices. Then (A, ∨) and (A, ∧) are both semi-

lattices, respectively called join and meet semi-lattices.

Let us define the two absorption laws as x = x ∨ (x ∧ y) and its dual x = x ∧ (x ∨ y). A

lattice is an algebra (X, ∨, ∧) satisfying equations expressing associativity, commutativity,

and idempotence of ∨ and ∧, and satisfying the two absorption equations.

(A, ∨, ∧) is a lattice over the set of 3-CNF-matrices with respect to the disjunction and

conjunction operators. Moreover, (A, ∨, ∧) is a distributive bounded lattice as ∧ is

distributive with respect to ∨ and A ∨ Ω = Ω & A ∧ ∅ = ∅ ∀A ∈ A. See [1] for more details

over lattices.

IV. Characterization theorems via functional descriptors

Theorem IV.1: Every non empty 3-CNF-matrix can be characterized by a one-

line parameterized 3-CNF-matrix, called its functional matrix description.

 
x1 xi xn
 1 
 s1 ··· sn1 
∀ [ϕ] =  6= ∅ , ∃ n functions fi : {0, 1}i → {0, 1} such that
 
 . ..
 .. sij . 
 
s1Σϕ ··· snΣϕ

!
_ x1 ··· xi ··· xn
[ϕ] = (11)
(α1 ,··· ,αn )∈{0,1}n f1 (α1 ) · · · fi (α1 , · · · , αi ) · · · fn (α1 , · · · , αn )

 
f1 (α1 )
notation  .. 
≡ 
 .

 (12)
fn (α1 , · · · , αn )

So, the knowledge of f1 (α1 ), · · · , fi (α1 , · · · , αi ), · · · , fn (α1 , · · · , αn ) characterizes fully [ϕ].

These modulo-2 functions are called the functional descriptors of ϕ.

January 6, 2020—8 : 41 am DRAFT

A 3-CNF-SAT DESCRIPTOR ALGEBRA AND THE SOLUTION OF THE P =NP CONJECTURE 8

Example :

[ϕ] = [(x1 ∨ x2 ∨ ¬x3 ) ∧ (¬x2 ∨ x3 ∨ ¬x4 )]

!
_ x1 x2 x3 x4
=
(α1 ,··· ,α4 )∈{0,1}4 α1 α2 (α1 + 1)(α2 + 1)α3 + α3 α2 (α3 + 1)α4 + α4
(mod 2)

Proof:

• The theorem is satisfied for n = 1 as

 
! ! ! ! x1 !
x1 x1 x1 x1   _ x1
= ; =  0 =
; 
1 f1 (α1 ) ≡ 1 0 f1 (α1 ) ≡ 0 α1 ∈{0,1} α1
1

• Let the theorem be true for n − 1 and [ϕ] be a 3-CNF-matrix of dimension n. There exist

two 3-CNF-matrices [ϕ1 ] and [ϕ2 ] of size n − 1 such that :

! !
_ x1 x2 · · · xn _ x1 x2 · · · xn
[ϕ] =
αi ∈{0,1} 0 f2 (α2 ) · · · fn (α2 , · · · , αn ) αi ∈{0,1} 1 g2 (α2 ) · · · gn (α2 , · · · , αn )

!
_ x1 · · · xn
Thus [ϕ] =
αi ∈{0,1} h1 (α1 ) · · · hn (α1 , · · · , αn )

where

h1 (α1 ) = α1

hi (α1 , · · · , αi ) = (α1 + 1)fi (α2 , · · · , αi ) + α1 gi (α2 , · · · , αi ) (mod 2) for i 6= 1

Corollary IV.1: The functional descriptors of ϕ are modulo-2 multi-linear com-

binations of αi .

Proof: This is a mere consequence of the recursive definition of hi (α1 , · · · , αi ).

X
So, hi (α1 , · · · , αi ) = ∆i (δ1 , · · · , δi ) α1δ1 · · · αiδi (mod 2) (13)
(δ1 ,··· ,δi )∈{0,1}i

where ∆i (δ1 , · · · , δi ) ∈ {0, 1}

∆i (δ1 , · · · , δi ) is called the signature of hi (α1 , · · · , αi ).

January 6, 2020—8 : 41 am DRAFT

A 3-CNF-SAT DESCRIPTOR ALGEBRA AND THE SOLUTION OF THE P =NP CONJECTURE 9

Example :

Consider a clause ψ ≡ [¬]xr ∨ [¬]xs ∨ [¬]xt where 1 ≤ r < s < t ≤ n. [ψ] is then characterized

by the following characterization functions :

hi (α1 , · · · , αi ) = αi ∀i<t



 (αr + 1)(αs + 1)(αt + 1) + αt if ψ = xr ∨ xs ∨ xt





 (αr + 1)(αs + 1) αt + αt if ψ = xr ∨ xs ∨ ¬xt

(αr + 1) αs (αt + 1) + αt if ψ = xr ∨ ¬xs ∨ xt





if ψ = xr ∨ ¬xs ∨ ¬xt

 (αr + 1) αs αt + αt
ht (αr , αs , αt ) = (14)


 αr (αs + 1)(αt + 1) + αt if ψ = ¬xr ∨ xs ∨ xt

if ψ = ¬xr ∨ xs ∨ ¬xt



 αr (αs + 1) αt + αt


αr αs (αt + 1) + αt if ψ = ¬xr ∨ ¬xs ∨ xt





if ψ = ¬xr ∨ ¬xs ∨ ¬xt

 αr αs αt + αt

Theorem IV.2: The conjunction operator ∧ between two sets of clauses can be

rewritten as the merging of their characterization functions : [hi (·)] = [fi (·)]∧[gi (·)].
   
f1 (α1 ) g1 (α1 )
 ..  and [ϕ0 ] ≡ 
  .. 
Let [ϕ] ≡ 
 .   .


fn (α1 , · · · , αn ) gn (α1 , · · · , αn )

Note : ϕ or ϕ0 should be extended if necessary in order to get the same support of propo-

sitional variables. Remember that all operations are modulo 2 : αi + αi = 0, αi2 = αi and

(αi + 1)αi = 0 for all αi .

 
h1 (α1 )
..
Then [ϕ] ∧ [ϕ0 ] ≡ 
 
.

 
hn (α1 , · · · , αn )

January 6, 2020—8 : 41 am DRAFT

A 3-CNF-SAT DESCRIPTOR ALGEBRA AND THE SOLUTION OF THE P =NP CONJECTURE 10

where for 1 ≤ t ≤ n :

ht (α1 , · · · , αt ) = (αt + 1) · { [ft (α1 , · · · , αt−1 , 0) + gt (α1 , · · · , αt−1 , 0)] (15)

· [ft (α1 , · · · , αt−1 , 1) · gt (α1 , · · · , αt−1 , 1)]

+ [ft (α1 , · · · , αt−1 , 0) · gt (α1 , · · · , αt−1 , 0)] }

+ αt · { [ft (α1 , · · · , αt−1 , 1) + gt (α1 , · · · , αt−1 , 1)]

· [ft (α1 , · · · , αt−1 , 0) + gt (α1 , · · · , αt−1 , 0)]

+ [ft (α1 , · · · , αt−1 , 1) + gt (α1 , · · · , αt−1 , 1)]

· [ft (α1 , · · · , αt−1 , 0) · gt (α1 , · · · , αt−1 , 0)]

+ [ft (α1 , · · · , αt−1 , 1) · gt (α1 , · · · , αt−1 , 1)]} (mod 2)

Moreover if there exists a (unique) j < t, related to the highest αj such that :

gj∗ (α1 , · · · , αj ) ≡ [ft (α1 , · · · , αt−1 , 0) + gt (α1 , · · · , αt−1 , 0)] ·

[ft (α1 , · · · , αt−1 , 1) + gt (α1 , · · · , αt−1 , 1)] 6= 0 (16)

[An additional constraint over αj induced by the conjunction operation]

⇒ call to a new merging fj (α1 , · · · , αj ) ∧ gj∗ (α1 , · · · , αj )

using a recursive call to definition (15).

(17)

Recursivity will end as soon as there is no longer such gj∗ (α1 , · · · , αj ) = 1 or when gj∗ (α1 , · · · , αj )

is no longer a function of αi but a constant always equal to 1.

Proof: Consider the possible values for ft (α1 , · · · , αt ) and gt (α1 , · · · , αt ) in equation (15)

when αt ∈ {0, 1} :

January 6, 2020—8 : 41 am DRAFT

A 3-CNF-SAT DESCRIPTOR ALGEBRA AND THE SOLUTION OF THE P =NP CONJECTURE 11

• ft (α1 , · · · , αt ) = gt (α1 , · · · , αt ) for αt ∈ {0, 1}

ht (α1 , · · · , αt ) = (αt + 1) · { [ft (·, 0) + gt (·, 0)] · [ft (·, 1) · gt (·, 1)] + [ft (·, 0) · gt (·, 0)] } +

αt · { [ft (·, 1) + gt (·, 1)] · [ft (·, 0) + gt (·, 0)] +

[ft (·, 1) + gt (·, 1)] · [ft (·, 0) · gt (·, 0)] + [ft (·, 1) · gt (·, 1)] }

= (αt + 1) · ft (·, 0) + αt · ft (·, 1) [as ft () + gt () = 0 and ft () · gt () = ft2 () = ft ()]

= ft (α1 , · · · , αt ) = ht (α1 , · · · , αt )

[ht () is thus the conjunction of ft () and gt ()]

• ft (α1 , · · · , 0) = gt (α1 , · · · , 0) but ft (α1 , · · · , 1) 6= gt (α1 , · · · , 1)

ht (α1 , · · · , αt ) = (αt + 1) · ft (·, 0) + αt · ft (·, 0) [as ft (·, 1) + gt (·, 1) = 1 and ft (·, 1) · gt (·, 1) = 0]

= ft (α1 , · · · , 0) = gt (α1 , · · · , 0)

[ht () sends αt to the value where ft () = gt ()]

• ft (α1 , · · · , 1) = gt (α1 , · · · , 1) but ft (α1 , · · · , 0) 6= gt (α1 , · · · , 0)

ht (α1 , · · · , αt ) = (αt + 1) · ft (·, 1) + αt · ft (·, 1) [as ft (·, 0) + gt (·, 0) = 1 ; ft (·, 0) · gt (·, 0) = 0]

= ft (α1 , · · · , 1) = gt (α1 , · · · , 1)

[ht () sends αt to the value where ft () = gt ()]

January 6, 2020—8 : 41 am DRAFT

A 3-CNF-SAT DESCRIPTOR ALGEBRA AND THE SOLUTION OF THE P =NP CONJECTURE 12

• ft (α1 , · · · , 1) 6= ht (α1 , · · · , 1) and ft (α1 , · · · , 0) 6= gt (α1 , · · · , 0)

⇓ [Impossibility to merge the two functions ft () and gt () ]

⇓ [No constraint over αt but a induced constraint over some αj , j < t]

ht (α1 , · · · , αt ) = αt [as ft () + gt () = 1 and ft () · gt () = 0]

but [ft (·, 0) + gt (·, 0)] · [ft (·, 1) + gt (·, 1)] = f onction(α1 , · · · , αj ) = 1

and gj∗ (α1 , · · · , αj ) := [ft (·, 0) + gt (·, 0)] · [ft (·, 1) + gt (·, 1)] + gj (α1 , · · · , αj ) = 1

[New additional constraint over αj , j < t, so that the impossibility

cannot occur anymore, as gj (·, αj ) → gj (·, αj ) + 1 when it appears.]

Example :

Consider the following sets of clauses :

• ϕ = (x1 ∨ x2 ∨ ¬x3 ) ∧ (¬x2 ∨ x3 ∨ ¬x4 ) ∧ (¬x1 ∨ x3 ∨ ¬x4 )

• ϕ0 = (¬x1 ∨ x2 ∨ ¬x3 ) ∧ (¬x2 ∨ x3 ∨ ¬x5 ) ∧ (¬x1 ∨ x3 ∨ ¬x5 )

Then
     
α1 α1 α1
     

 α2  
  α2  
  α2 

[ϕ] =  α1 α3 + α2 α3 + α1 α2 α3 ∧ α3 ∧ α3
     

     

 α4   α +α α +α α α
  4 2 4 2 3 4
  α +α α +α α α
  4 1 4 1 3 4


α5 α5 α5
 
α1
 

 α2 

= α1 α3 + α2 α3 + α1 α2 α3
 

 
 α +α α +α α +α α α +α α α +α α α +α α α α 
 4 1 4 2 4 1 2 4 1 3 4 2 3 4 1 2 3 4 
α5

January 6, 2020—8 : 41 am DRAFT

A 3-CNF-SAT DESCRIPTOR ALGEBRA AND THE SOLUTION OF THE P =NP CONJECTURE 13

     
α1 α1 α1
     

 α2  
  α2  
  α2 

0
[ϕ ] =  α3 + α1 α3 + α1 α2 α3  ∧  α3 ∧ α3
     

     

 α4 
 
 α4  
  α4 

α5 α5 + α2 α5 + α2 α3 α5 α5 + α1 α5 + α1 α3 α5
 
α1
 

 α2 

= α3 + α1 α3 + α1 α2 α3
 

 

 α4 

α5 + α1 α5 + α2 α5 + α1 α2 α5 + α2 α3 α5

And

 
α1
 

 α2 

0
[ϕ] ∧ [ϕ ] =  α2 α3
 

 
 α +α α +α α +α α α +α α α +α α α +α α α α 
 4 1 4 2 4 1 2 4 1 3 4 2 3 4 1 2 3 4 
α5 + α1 α5 + α2 α5 + α1 α2 α5 + α2 α3 α5

In this example, no recursive call is done. We get h3 (·) := (α1 α3 + α2 α3 + α1 α2 α3 ) ∧ (α3 +

(15)
α1 α3 +α1 α2 α3 ) ⇒ h3 (·) := (α3 +1)·{0·α2 +0}+α3 ·{(α2 +1)·0+(α2 +1)·0+α2 } = α2 α3 ,

as f3 (α1 , α2 , 0) = 0, f3 (α1 , α2 , 1) = α1 + α2 + α1 α2 , g3 (α1 , α2 , 0) = 0, g3 (α1 , α2 , 1) = 1 + α1 +

α1 α2 , f3 (α1 , α2 , 1) + g3 (α1 , α2 , 1) = α2 + 1 and f3 (α1 , α2 , 1) · g3 (α1 , α2 , 1) = α2 .

V. An approach to the 3-CNF-SAT problem via descriptors

A. Boolean descriptors

The usual presentation of a 3-CNF-SAT problem consists of a list on m 3-CNF clauses

defined over n propositional variables. These m clauses describe perfectly the set of solu-

tions for the 3-CNF-SAT problem and can be considered as Boolean descriptors of the

3-CNF-SAT problem. These Boolean descriptors are of linear complexity as they can be

represented by an array of dimension 3 × m.

January 6, 2020—8 : 41 am DRAFT

A 3-CNF-SAT DESCRIPTOR ALGEBRA AND THE SOLUTION OF THE P =NP CONJECTURE 14

The difficulty with such Boolean descriptors is that there is no simple or direct relation

between them and the set of solutions or the answer to the satisfiability question.

B. 3-CNF-matrix descriptors

This paper proposes in (3) a 3-CNF-matrix description of a 3-CNF-SAT problem. These

descriptors (each line in the 3-CNF-matrix) can be of exponential complexity as there are

as many descriptors as solutions. Even if one uses the reduction version of the 3-CNF-matrix

description as explained in (5), simulations show that the complexity remains exponential.

The interest of these descriptors is the direct link between them and the set of solutions

or the answer to the satisfiability question.

C. Functional descriptors

This paper proposes also in (11) a 3-CNF-matrix functional description for any 3-CNF-

SAT problem. These functional descriptors are of unknown complexity, at least at this

stage of the paper.

These functional descriptors are somehow in between both previous types of descriptors, as

they are in an exponential relation to the set of solutions and in an direct relation with

the satisfiability question. Indeed, given the functional descriptors, it is straightforward to

give the answer to the satisfiability question : no if the functional descriptors does not exist,

and yes otherwise. However, one needs to generate all possible values for αi to get the entire

set of solutions, and that can take an exponential time.

Conclusion : The approach of the 3-CNF-SAT problem via functional descriptors seems

to be promising as it does not consider the set of all solutions, but only focuses on the sole

question about satisfiability.

January 6, 2020—8 : 41 am DRAFT

A 3-CNF-SAT DESCRIPTOR ALGEBRA AND THE SOLUTION OF THE P =NP CONJECTURE 15

VI. Complexity analysis of the functional descriptor approach

A. A first measure of the complexity for functional descriptors

Theorem VI.1: The complexity of the functional descriptor approach for a 3-

CNF-SAT problem with m clauses and n propositional variables is

O(m n2 max max lenj (ht ) )

1≤t≤n 1≤j≤m

where lenj (ht ) is the number of terms in ht (·) when the j first clauses are considered :

X
Let len(ht ) ≡ ∆t (δ1 , · · · , δt ) [ see (13) for the definition of ∆t ] (18)
(δ1 ,··· ,δt )∈{0,1}t

So lenj (ht ) = len(ht ) at stage j of the computations.

Proof: Let us compute the complexity of ft (·)∧gt (·) in (15). First of all, one has to compute

the four functions in square brackets : [ft (·, 0) + gt (·, 0)], [ft (·, 1) + gt (·, 1)], [ft (·, 0) · gt (·, 0)]

and [ft (·, 1) · gt (·, 1)]. We have :

len(ft (·, 0)) ≤ len(ft (·, αt )) and len(ft (·, 1)) ≤ len(ft (·, αt )) [≡ len(ft )]

len(gt (·, 0)) ≤ len(gt (·, αt )) and len(gt (·, 1)) ≤ len(gt (·, αt )) [≡ len(gt )]

len(ft + gt ) ≤ len(ft ) + len(gt ) ≤ len(ft ) · len(gt ) when len(ft ) > 2 and len(gt ) > 2

The complexity for the four functions is then O(len(ft ) · len(gt )).

The complexity for computing ht (·) in (15) is :

O( 3 · [(len(ft ) · len(gt ))2 + (len(ft ) · len(gt ))] + 2 · [2(len(ft ) · len(gt ))2 + (len(ft ) · len(gt ))])

= O(7 · (len(ft ) · len(gt ))2 + 5 · (len(ft ) · len(gt )))

= O([len(ft ) · len(gt )]2 ) for large len(ft ) · len(gt ) (19)

Note : it needs three runs over the formula in the brackets to do the product with (αt + 1),

one to compute the formula, one to multiply it by αt and one to add both results. Similarly,

it takes two runs to compute the product with αt .

January 6, 2020—8 : 41 am DRAFT

A 3-CNF-SAT DESCRIPTOR ALGEBRA AND THE SOLUTION OF THE P =NP CONJECTURE 16

Using the same argumentation, we have :

len(ht ) = O([len(ft ) · len(gt )]2 ) for large len(ft ) · len(gt ) (20)

and for the recursive call with j < t [see (16)]

len(gj∗ ) = O([len(ft ) · len(gt )]2 ) for large len(ft ) · len(gt ) (21)

To solve the 3-CNF-SAT problem, one should compute all n functional descriptors ht (·),

each of them with at most n recursive calls, and this for each step of integration of the

m clauses. So, using the equivalence between (20) and (19), the overall complexity of the

functional approach to 3-CNF-SAT problem will be of order O(m n2 max max lenj (ht ) ).
1≤t≤n 1≤j≤m

B. Non uniformly distributed versus uniformly distributed literals in 3-CNF-SAT problems

Theorem VI.2: The most difficult 3-CNF-SAT problems are uniformly dis-

tributed ones.

Note : The invariance structure of 3-CNF-SAT problem is important with respect to the

complexity of the functional descriptor approach. It is then normal that problems with

uniformly distributed propositional variables are harder as no re-labeling of the variables

can be done to reduce the complexity. A simple example of the importance of re-labeling is

proposed just after the following proof.

Proof:

• Negative and positive literals

First, let us note that the computations involving negative literals are easier and faster than

for positive ones. This is a mere consequence of our definition for ht (·) in (14). So the

most difficult problems will be the balanced one with respect to the proportion of negative

and positive literals. Otherwise, we inverse some variables in order to get the maximum

of negative literals. Let us suppose from here that the proportion of positive and negative

literals is quasi equal for each variable.

• Some definitions

Let us divide now the clauses in two sets. The first set contains all the clauses with the
January 6, 2020—8 : 41 am DRAFT
A 3-CNF-SAT DESCRIPTOR ALGEBRA AND THE SOLUTION OF THE P =NP CONJECTURE 17

higher indexed literal being positive and the second with the negative ones :

[
Cl+ = Cl+ (xt )
t
[
= {ψi := [¬]xr ∨ [¬]xs ∨ xt with r < s < t, or any permutation of xr , xs , xt }
t
[
Cl− = Cl− (xt )
t
[
= {ψi := [¬]xr ∨ [¬]xs ∨ ¬xt with r < s < t, or any permutation of xr , xs , xt }
t

and V + (xt ) = {xi (i < t) | ∃ ψ ∈ Cl+ (xt ) : xi appears in ψ}

V − (xt ) = {xi (i < t) | ∃ ψ ∈ Cl− (xt ) : xi appears in ψ}

V (xt ) = V + (xt ) ∪ V − (xt )

By construction, there is at least one solution for each xt when considering clauses only in

Cl+ [xt = 1] or in Cl− [xt = 0]. Moreover, the computation of the functional descriptors will

not involved recursive calls [see (16)] as no impossibility exists for any xt .

As ht (·) is a multi-linear combination of the αi corresponding to the literals xi (i ≤ t) found

+
in clauses where xt has the highest index, there will be at most 2(#V (xt ) + 1)
[respectively
− +
2(#V (xt ) + 1)
] terms in ht (·) computed over Cl+ [resp. Cl− ]. So len(ht )|Cl+ ≤ 2(#V (xt ) + 1)

−
and len(ht )|Cl− ≤ 2(#V (xt ) + 1)
.

• The merging of Cl+ and Cl− .

One needs to compute ht (·) := ht (·)|Cl+ ∧ ht (·)|Cl− .

 For t = n and considering that hn (·) is a multi-linear combination of the αi appearing in

hn (·)|Cl+ or hn (·)|Cl− , we have that :

+
(xn ) ∪ V − (xn )) + 1
len(hn ) ≤ 2#(V

≤ 2#V (xn ) + 1 (22)

 For t = n − 1 also, hn−1 (·) is a multi-linear combination of the αi corresponding to

the variables xi (i < n − 1) found in common clauses with xn−1 as the highest variable :

len(hn−1 ) ≤ 2#V (xn−1 ) + 1 . But one has to add the potential αi involved in the recursive

January 6, 2020—8 : 41 am DRAFT

A 3-CNF-SAT DESCRIPTOR ALGEBRA AND THE SOLUTION OF THE P =NP CONJECTURE 18

∗
call gn−1 (·) from the previous computation of hn (·). [see (16)]

∗
From (16), gn−1 (α1 , · · · , αn−1 ) = [hn (·, 0)|Cl+ + hn (·, 0)|Cl− ] · [hn (·, 1)|Cl+ + hn (·, 1)|Cl− ].
∗
So gn−1 (·) will be a multi-linear combination of the same αi , except αn , as for hn (·).
∗
Therefore, hn−1 (·) ∧ gn−1 (·) will be a combination of the αi associated to the variables
Sn Sn
in i=n−1 [V + (xi ) ∪ V − (xi )] ⇒ len(hn−1 ∧ gn−1
∗
) ≤ 2#[ i=n−1 V (xi )] , as xn−1 should not be

counted in V (xn ).

Sn
 So, ∀ t : len(ht ∧ gt∗ ) ≤ 2#[ i=t V (xi )] + 1−(n−t)
. But as ht ∧ gt∗ is a combination of at most

t αi ’s, we have :

Sn
len(ht ∧ gt∗ ) ≤ min( 2#[ i=t V (xi )] + 1−(n−t)
, 2t ) (23)

• Uniform distribution of the literals

As V (xi ) is dependent of the ordering of the variables, it is possible to re-order the variables
Sn
so that 2#[ i=t V (xi )]
is minimal, except in the case of uniformly distributed literals. The

uniformly distributed case is then the most difficult problem, as no ordering can

reduce the maximum value in (23).

Example of non uniformly distributed 3-CNF-SAT problem :

• Consider the following 3-CNF-SAT problem with m clauses and 2m + 1 propositional

variables :

^
ϕ := (x2i−1 ∨ x2i ∨ x2m+1 )
1≤i≤m

Here we have :

V + (xt ) = V − (xt ) = ∅ ∀ t 6= 2m + 1

V + (x2m+1 ) = {x1 , · · · , x2m } and V − (x2m+1 ) = ∅

max max lenj (ht ) = 3m+1 − 3m + 1 = O(3m )

1≤t≤n 1≤j≤m

Note : The proof of this equality is more difficult than interesting, so we do not write it here.

January 6, 2020—8 : 41 am DRAFT

A 3-CNF-SAT DESCRIPTOR ALGEBRA AND THE SOLUTION OF THE P =NP CONJECTURE 19

• But the same 3-CNF-SAT problem can be formalized in terms of opposite literals

yi = ¬xi , ∀i ∈ {1, · · · , 2m + 1} :

^
ϕ := (¬y2i−1 ∨ ¬y2i ∨ ¬y2m+1 )
1≤i≤m

This time, we have :

V + (yt ) = V − (yt ) = ∅ ∀ t 6= 2m + 1

V + (y2m+1 ) = ∅ and V − (y2m+1 ) = {y1 , · · · , y2m }

max max lenj (ht ) = 2m = O(2m )

1≤t≤n 1≤j≤m

Indeed, ht (α1 , · · · , αt ) = αt ∀ t < 2m + 1

and h2m+1 (α1 , · · · , α2m+1 ) = [(α1 α2 α2m+1 + α2m+1 ) ∧ (α3 α4 α2m+1 + α2m+1 )] ∧ · · ·
see(15)
= [(α2m+1 + 1) · 0 + (α2m+1 ) · (α1 α2 + 1) · (α3 α4 + 1)] ∧ · · ·
m
Y
= (α2m+1 ) · (α2i−1 α2i + 1)
i=1
⇒ len2m+1 (h2m+1 ) = 2m

• Finally, the same 3-CNF-SAT problem can be formalized using re-ordered propositional

variables z1 = y2m+1 and zi = yi−1 , ∀i ∈ {2, · · · , 2m + 1} :

^
ϕ := (¬z2i ∨ ¬z2i+1 ∨ ¬z1 )
1≤i≤m

And, this time, we have :

V + (zt ) = ∅ ∀ t
(
− ∅ for t = 1 or t = 2i (1 ≤ i ≤ m)
V (zt ) =
{z1 , z2i } for t = 2i + 1 (1 ≤ i ≤ m)
max max lenj (ht ) = 2
1≤t≤n 1≤j≤m

So for this example, one can reach a linear complexity of O(2 · number of ht (·)) = O(2 m),

as only one ht (·) has to be computed at each step without any recursive call.

January 6, 2020—8 : 41 am DRAFT

A 3-CNF-SAT DESCRIPTOR ALGEBRA AND THE SOLUTION OF THE P =NP CONJECTURE 20

Example of uniformly distributed 3-CNF-SAT problem :

The smallest exact uniformly distributed and optimally re-ordered 3-CNF-SAT problem is :



 x1 ∨ ¬x2 ∨ ¬x3





 x1 ∨ x2 ∨ ¬x3

¬x1 ∨ ¬x2 ∨ ¬x3




8

¬x1 ∨ x2 ∨ ¬x3

^ ^
ϕ := ψi =
i=1


 x1 ∨ ¬x2 ∨ x3

x1 ∨ x2 ∨ x3






¬x1 ∨ ¬x2 ∨ x3





¬x1 ∨ x2 ∨ x3



No relabeling will reduced the 3-CNF-SAT complexity. This problem is “hard” in the sense

that each clause eliminates only one solution at a time. We have here :

V + (x1 ) = V − (x1 ) = ∅

V + (x2 ) = V − (x2 ) = ∅

V + (x3 ) = V − (x3 ) = {x1 , x2 }

and
 
Step h1 (·) h2 (·) h3 (·) maxt len(ht ) #Solutions
 

 ψ1 α1 α2 (α1 + 1)α2 α3 + α3 3 7 

ψ1 ∧ ψ2 α1 α2 α1 α3 1 6
 
 
 

 ∧3i=1 ψi α1 α2 α1 α2 α3 + α1 α3 2 5 

 

 ∧4i=1 ψi α1 α2 0 1 4 

∧5i=1 ψi
 

 α1 α1 α2 0 1 3 

∧6i=1 ψi 1 α2 0 1 2
 
 
 

 ∧7i=1 ψi 1 0 0 1 1 

∧8i=1 ψi @ @ @ 0 0

Conclusions :

The theorem about uniformity is important as it states : for any non uniformly distributed

3-CNF-SAT problem ϕ with m clauses and n variables, there exists an uniformly distributed

3-CNF-SAT problem ϕ0 with m clauses and n variables which is more difficult to solve, in

terms of functional descriptors.

January 6, 2020—8 : 41 am DRAFT

A 3-CNF-SAT DESCRIPTOR ALGEBRA AND THE SOLUTION OF THE P =NP CONJECTURE 21

C. A sorting algorithm to reduce complexity

We propose the following sorting algorithm of complexity O(m + n log(n) + m log(m)) :

 Relabel the propositional variables [xi → yj ] in order to get their occurrence [O(m)]

in a decreasing order : #{y1 } is maximal, · · · , #{yn } is minimal [O(n log(n))];

 Inverse the sign of the literals in order to get the maximum of negative literals;

 Sort the clauses to get a increasing order of the highest variable in the ordered

clauses [O(m log(m))];

 Within the set of clauses with the same highest variable, sort the clauses so that the

ones with negative highest variable appear before the ones with positive highest variable.

As we seldom have exact uniformly distributed 3-CNF-SAT problems, the complexity can

then be reduced drastically, as shown in Figure 1.

len(h_t)  for  20  variables  and  120  clauses   len(h_t)  for  20  variables  and  120  clauses  
30000   160   h_9  
h_10  
140  
25000   h_11  
120  
h_12  
20000  
h_1   100   h_13  

15000   h_2   80   h_14  

h_18   h_15  
60  
10000  
h_19   h_16  
40  
h_20   h_17  
5000  
20  
h_18  
0   0   h_19  
0   20   40   60   80   100   120   0   10   20   30   40   50   60   70   80   90   100  
3-­‐CNF-­‐Clauses   3-­‐CNF  Clauses   h_20  

Fig. 1 : Complexity for the same dataset before and after the sorting algorithm.

D. Exact uniformly distributed α-random 3-CNF-SAT problems

Definition : Let ϕ be a 3-CNF-SAT problem with n variables, each of them appearing

3α 3α
exactly 2 times as positive and 2 times as negative literal, for some α > 0. Let these 3 α n

literals be randomly distributed amongst the clauses. Such problem is called an exact

uniformly distributed α-random 3-CNF-SAT problem.

Remark : For exact uniformly distributed 3-CNF-SAT problem, the labeling part of the
3α
previous sorting algorithm has no effect, as the occurrence of each literal is 2 . Only the

re-ordering of the m clauses can reduce the complexity of the problem.

January 6, 2020—8 : 41 am DRAFT

A 3-CNF-SAT DESCRIPTOR ALGEBRA AND THE SOLUTION OF THE P =NP CONJECTURE 22

Theorem VI.3: For such exact uniformly distributed α-random 3-CNF-SAT

problems, the expected number of clauses where i (i > 2) is the highest index, is

noted mα (i) and given by :

(i − 1)(i − 2)
E[#{ψ = [¬]xr ∨ [¬]xs ∨ [¬]xt | max(r, s, t) = i}] = 3 α ≡ mα (i) (24)
(n − 1)(n − 2)

Proof: Let ψ be a clause with xi or ¬xi , there are C2i−1 · 3 α combinations with smaller

indices amongst C2n−1 · 3 α possibles combinations. So, the probability for xi to get the
(i−1)(i−2)
highest index is : (n−1)(n−2) . The expected value is obtained by multiplying the probability

by the number of occurrences of xi .

Figure 2 shows the theoretical density and cumulative distributions of mα (i) versus the

distributions for the observed values for #{ψ = [¬]xr ∨ [¬]xs ∨ [¬]xt | max(r, s, t) = i} in the

case of a 3-CNF-SAT problem with 175 variables and 753 random clauses.

Empirical  versus  theore1cal  distribu1on  of  m_α(i)   Empirical  versus  theore1cal  cumula1ve  distribu1on  
Total  number  of  variables  in  the    first  N  clauses    

14   200  
Number  of  clauses  with  i  as  maximal  index  

180  
12  
160  

10   140  

120  
8  
100  

6   80  

60  
4  
40  

2   20  

0  
0   0   100   200   300   400   500   600   700   800  
0   25   50   75   100   125   150   175   N  first  clauses    
Index  i  of  x_i  :  175  variables  and  753  random  clauses  (uuf175-­‐01.cnf)   (753  randon  clauses  from  uuf175-­‐01.cnf  hBp://www.satlib.org/ubcsat)  

Empirical  data  (175  variables  -­‐  alpha  =  4,30)   TheoreKcal  data  (175  var  -­‐  alpha  =  4,30)   Empirical  #  variables  in  N  clauses   TheoreKcal  #  variables  in  N  clauses  

Fig. 2 : Density and cumulative distributions of “sorted clauses” for n = 175 and α = 4, 30.

Theorem VI.4: For exact uniformly distributed α-random 3-CNF-SAT prob-

lems, the expected number of variables in V (xi ), for i > 2 and large n, is given

by :

(i − 1)(i − 2)
E[#V (xi )] = 2 mα (i) = 6α (25)
(n − 1)(n − 2)

Proof: There is C2i−1 possible triplets with xi being the highest indexed variable. The
i−2 2
probability for some xj (j < i) to appear in one of these triplets is C2i−1
= i−1 = p for any

j. The occurrence of xj follows a binomial model Bi(mα (i), p), as one can choose several
January 6, 2020—8 : 41 am DRAFT
A 3-CNF-SAT DESCRIPTOR ALGEBRA AND THE SOLUTION OF THE P =NP CONJECTURE 23

times the same triplet (given different clauses with respect to the negative or positive sign

of the included literals). The expected number of occurrence of xj in the mα (i) triplets is
6 α(i−2)
then mα (i) · p = (n−1)(n−2) < 1 for large n. So each variable is expected to appear at most

once in the mα (i) triplets-clauses. Therefore, the number of variables, different from xi ,

occurring in these mα (i) clauses is 2mα (i) as there are two variables distinct from xi in each

clause.

Theorem VI.5: For exact uniformly distributed α-random 3-CNF-SAT prob-

lems, the maximal expected complexity for the computation of ht (·) is bounded

by :

k
" #
X
min{2( mα (n(j) )) − (k − 1) , n(k) }
j=0

len(ht ) ≤ max 2
k≥0

k
X
where 2( mα (n(j) )) − (k − 1) is a concave quadratic function with respect to t or
j=0
k, as shown on figures 3 and 4.

Proof: From (25), we know that #V (xi ) is expected to be maximal for i = n, when

#V (xn ) = 2mα (n) = 6 α. So, from (22), we have that :

len(hn ) ≤ 2#V (xn ) + 1

≤ 22 mα (n)+1 = 2(6α+1) (26)

For the computation of the recursive call gj∗ (α1 , · · · , αj ) (see 16), the index j is the highest

index of the variables in V (xn ). Let us note it n(1) . We have thus # V (xn ) = 2 mα (n)

indexes uniformly chosen from {1, · · · , n − 1}. n(1) will be the expected maximal index from

an uniform distribution for 2 mα (n) iid variables ui ∼ U [1, · · · , n − 1] :

2 mα (n) 6α
n(1) = E[ max (ui )] = (n − 1) = (n − 1) (27)
1≤i≤2mα (n) 2 mα (n) + 1 6α + 1

January 6, 2020—8 : 41 am DRAFT

A 3-CNF-SAT DESCRIPTOR ALGEBRA AND THE SOLUTION OF THE P =NP CONJECTURE 24

So, for the recursive call, we will have to compute hn(1) (·) ∧ gn∗ (1) (·). We get :

(n(1) − 1)(n(1) − 2)
#V (xn(1) ) = 2 mα (n(1) ) = 6α from (25)
(n − 1)(n − 2)

len(hn(1) (·) ∧ gn∗ (1) (·)) ≤ 2#{V (xn ) ∪ V (xn(1) )} + 1−(2−1) from (23)
(1)
≤ 22(mα (n)+mα (n ))

And so on, for the next recursive calls. We get for the recursive call k (k > 1) :

n ≡ n(0)

ui ∼ U [1, · · · , n(k−1) − 1]

2 mα (n(k−1) )
n(k) = E[ max (ui )] = (n(k−1) − 1)
1≤i≤2mα (n(k−1) ) 2 mα (n(k−1) ) + 1

(n(k) − 1)(n(k) − 2)
#{V (xn(k) )} = 2 mα (n(k) ) = 6α
(n − 1)(n − 2)

k
 
[
(k)

 min{# {V (xn(j) )} − (k − 1) , n } 


len(hn(k) (·) ∧ gn∗ (k) (·)) ≤ 2 j=0

k
 
X
(j) (k)

 min{2( mα (n )) − (k − 1) , n } 


len(hn(k) (·) ∧ gn∗ (k) (·)) ≤ 2 j=0

 
min{Mα (n(k) ), n(k) }
≤2 (28)

This bound is only defined for the variables with n(k) as index. Note that n(k) are functions

of the starting index n(0) = n. We can compute similar bounds for other starting indexes

n(0) in [1, · · · , n − 1], so that Mα (·) can be defined for all t as shown in Figure 4.

Numerical computations show that, for large n, Mα (n(k) ) as well as Mα (t) are concave

quadratic functions with coefficients only depending on α. This can be easily explained

as a mere consequence of the iid randomness of the variables #{V (xn(k) )} and mα (n(j) ).

January 6, 2020—8 : 41 am DRAFT

A 3-CNF-SAT DESCRIPTOR ALGEBRA AND THE SOLUTION OF THE P =NP CONJECTURE 25

Mα(n(k))  
1400  

1200  

1000  

800   N=175  -­‐  alpha=4  

N=500  -­‐  alpha=4  

600   y  =  -­‐0,4729x2  +  29,386x  +  32,912   N=100.000  -­‐  alpha=4  

R²  =  0,99999  
N=5000  -­‐  alpha=8  
400   N=100.000  -­‐  alpha=8  

N=100.000  -­‐  alpha=5,19  

200  

0  
0   10   20   30   40   50   60  

-­‐200  
Recursive  call  index  k  

k
X
Fig. 3 : Complexity wrt k : Mα (n (k)
) = 2( mα (n(j) )) − (k − 1) where n(0) = n.
j=0

Indeed, the central limit theorem for the expectation of iid random variables predicts that
(k)
E[2Mα (n )
] follows a Normal distribution (censored by min). But X ∼ N (µ, σ 2 ) implies a
2
quadratic log-density : log(fX (x)) ∝ − (x−µ)
2σ . For each value of α, we can compute the cor-

responding µα , σα and the maximum value for Mα (n(k) ). Quadratic regression estimations

give maxk (Mα (n(k) )) ≈ 294 for α = 4, maxk (Mα (n(k) )) ≈ 490 for α = 5, 12 (see figure 3 and

below for the choice of such α) and maxt (Mα (t)) ≈ 1160 for α = 8.

Remark : It is now important to see whether different starting points n(0) yield not to

aggregating trajectories so that addition of bounds are to be considered. This situation can

be neglected as shown in the following theorem.

Theorem VI.6: The probability for a given variable xi to be in more than one

trajectory tends to zero for large n.

Proof: Let us consider separately the possible trajectories tr(xn(0) → xn(k) ) for n(0) = m ∈

{1, · · · , n} and k ∈ {1, · · · , n}. Let us note a given trajectory : tr(m, km ) with km such that

n(km ) > i. For each variable xi , there exists at most (n − i)(n − i − 1)/2 trajectories tr(m, km )

where xi could be the next highest indexed variable for n(km +1) : tr(n, 0), · · · , tr(n, kn ), · · · ,

January 6, 2020—8 : 41 am DRAFT

A 3-CNF-SAT DESCRIPTOR ALGEBRA AND THE SOLUTION OF THE P =NP CONJECTURE 26

tr(i + 1, 0). The probability for xi to get the highest index in a trajectory is :

P [i = max{l : xl ∈ tr(m, km )}]

l
j=k
[m
= P [i = max{l : xl ∈ V (xn(j) ) \ {xn(0) , · · · , xn(km ) }}]
l
j=0|n(0) =m

= P [i = max{l : xl ∈ tr(m, km )}|xi ∈ tr(m, km )] · P [xi ∈ tr(m, km )]

l

We have :

P [xi ∈ tr(m, km )] = #{clauses in tr(m, km )} · P [xi ∈ the clause and i is the maximum index]
j=k
Xm n(j) − 2
= mα (n(j) ) · (j) −1
[see (24)]
j=0|n(0) =m
C2n
j=k
Xm 2
= mα (n(j) ) ·
n(j) − 1
j=0|n(0) =m
j=k
Xm (n(j) − 1)(n(j) − 2)
= 6α
(n − 1)(n − 2)(n(j) − 1)
j=0|n(0) =m
j=k
Xm
(6α)
= (n(j) − 2)
(n − 1)(n − 2) j=0

For instance :

6α(n(0) − 2) 6α m − 2 6α
P [xi ∈ tr(m, 0)] = = ≤ as n(0) = m ≤ n
(n − 1)(n − 2) n−1 n−2 n−1

and

P [xi ∈ tr(m, 1)] ≤ P [xi ∈ tr(n, 1)]

6α
= (n(0) − 2) + (n(1) − 2)
(n − 1)(n − 2)
6α 6α 6α
= + ([ (n − 1)] − 2)
n − 1 (n − 1)(n − 2) 6α + 1
6α 6α
< + (n − 3)
n − 1 (n − 1)(n − 2)
6α n−3
< (1 + )
n−1 n−2

Finally,

6α n−3 n − (km + 2)
P [xi ∈ tr(m, km )] ≤ P [xi ∈ tr(n, km )] < (1 + + ··· + )
n−1 n−2 n−2
→ 0 for large n with respect to km and α.
January 6, 2020—8 : 41 am DRAFT
A 3-CNF-SAT DESCRIPTOR ALGEBRA AND THE SOLUTION OF THE P =NP CONJECTURE 27

Now, considering that the elements of tr(m, km ) are iid uniformly distributed random vari-

ables drawn from {1, · · · , m − 1}, we have :

P [i = max{l : xl ∈ tr(m, km )}|xi ∈ tr(m, km )]

l

= P [i = max{#tr(m, km )uniform random variables}]

[for large n, we use the expected value for #tr(m, km )]

j=k
Xm
= P[ 2mα (n(j) ) − 1 uniform iid variables ≤ i ]
j=0
Pj=km
j=0 2mα (n(j) )−1
Y i
=
(m − 1)
l=1
 Pj=km 2mα (n(j) )−1
i j=0
=
m−1

In conclusion,
 Pj=k m 2m (n(j) )−1
α j=k
Xm
i j=0
(6α)
P [i = max{l : xl ∈ tr(m, km )] = (n(j) − 2)
l m−1 (n − 1)(n − 2) j=0
j=k
Xm
(6α)
≤ (n(j) − 2) as i ≤ (m − 1)
(n − 1)(n − 2) j=0
→0 for large n with respect to k and α

Therefore, there is a negligible probability for a variable xi to be maximal in two or more

trajectories tr(m, km ), as we can see this event as the output of a binomial model with a very

small probability of success (“xi being maximal in some tr(m, km )”), over (n − i)(n − i − 1)/2

possible trajectories :

Let p = max P [i = max{l : xl ∈ tr(m, km )}]

m,km l

Then, P [Two or more successes]

= 1 − (P [0 success] + P [1 success])
(n−i)(n−i−1) (n − i)(n − i − 1) (n−i)(n−i−1)
−1
≤ 1 − ([(1 − p) 2 ]+[ p(1 − p) 2 ])
2
→0 for large n, as p → 0 for large n with respect to k and α.

The last thing to prove is that k is not O(n) as α is a given constant. Figure 3, which

is computed with the theoretical formula from (28), shows that the maximal value for k is
January 6, 2020—8 : 41 am DRAFT
A 3-CNF-SAT DESCRIPTOR ALGEBRA AND THE SOLUTION OF THE P =NP CONJECTURE 28

negligible with respect to n : k ≤ 30 for n = 100.000 and α = 4, and k ≤ 55 when n = 100.000

and α = 8.

Mα(t)  =  Mα(n(k))  with  t  =  n(k)            

300  

250  

200  
y  =  -­‐0,0037x2  +  1,8559x  +  19,51  
R²  =  0,99937  
150  
N=500  -­‐  alpha=4  (Theory)  

N=500  -­‐  alpha=4  (Example)  

100  

50  

0  
0   100   200   300   400   500   600  
Index  t  of  the  variables  

Fig. 4 : Complexity wrt variable index t : Mα (n(k) ) = Mα (t) for n(0) = n.

The example comes from the Dimacs generator at https://toughsat.appspot.com/

Remark : With some real generated 3-CNF-SAT problems, it is possible to observe a

“cluster” process, the size of one trajectory, i.e. the number of xi involved in that trajectory,

becoming more and more important so that this trajectory attracts all the variables. Then,

k is O(n), P [xi ∈ tr(n, k)] → 1 and the complexity becomes exponential. It is easy to solve

these cases. As the variables are uniformly distributed in random 3-CNF-SAT problems,

each variable xi being repeated approximatively 3α times, it is possible to permute joining

variables xj (belonging to two or more trajectories) with a smaller indexed variable, such as

xj−1 (or xj−2 if xj−1 is already in a previous trajectory, and so on). The two trajectories

will then be dissociated. We propose the following “permutation” algorithm :

 First, apply the sorting algorithm over the 3-CNF-SAT problem;

 Sort each clause [¬]xr ∨ [¬]xs ∨ [¬]xt so that r ≥ s ≥ t;

 Beginning with the last clause (with [¬]xn ), mark xj where j = max{i : xi ∈ V (xn )}

as already belonging in a trajectory and initialize W (xn ) := V (xn ) and W (xj ) := V (xn )

where W (xj ) ≡ ∪i V (xi ) for i such that xi ∈ tr(· · · , xj );

 Loop over k := 1 to k := n − 3 with clauses having [¬]xn−k as the highest in-

dexed variable; if xn−k is already marked as belonging to a trajectory, do W (xn−k ) :=

W (xn−k ) ∪ V (xn−k ) otherwise initialize W (xn−k ) := V (xn−k );

January 6, 2020—8 : 41 am DRAFT

A 3-CNF-SAT DESCRIPTOR ALGEBRA AND THE SOLUTION OF THE P =NP CONJECTURE 29

Consider xj where j = max{i : xi ∈ W (xn−k )}; do while (xj is already marked as

belonging in a trajectory and j ≥ 3α) relabel xj ↔ xj−1 and j := j − 1

[we do not consider j < 3α as merging of trajectories for small indexes is not a

problem because Mα (j) = j];

Initialize W (xj ) := W (xn−k ).

Figure 4 shows the result for a Dimacs generated 3-CNF-SAT problem with 500 variables

and α = 4. We apply the sorting and the permuting algorithms on the generated file to

eliminate joining trajectories.

If we have proved in this section that the complexity is bounded with respect to n, we still

have to show that complexity is not increasing with respect to α, which is not the case for

Mα (t).

VII. Complexity analysis with respect to α

It is easy to see that the complexity is an increasing function of α for exact uniformly

distributed α-random 3-CNF-SAT problems, at least for small α, as smaller α-random 3-

CNF-SAT problems can be viewed as subsets of larger α-random problems.

But there should be somewhere a threshold for α as large α-random problems are easy to

solve because unsatisfiability is often a consequence of a subset of the problem. Empirical

results from the literature suggest that this threshold for α is ≈ 4.258. See [3].

The analysis of complexity with respect to α will be done through Sϕ , the set of all satisfying

solutions for the 3-CNF-SAT problem ϕ. See definition (2).

Theorem VII.1: For exact uniformly distributed α-random 3-CNF-SAT prob-

lems ϕ = {ψj }1≤j≤m with n variables and m clauses, we get for large n and m the

following expected number of solutions :

7 7
E[#Sϕ ] = E[#{(x1 , · · · , xn ) ∈ {0, 1}n |ϕ(x1 , · · · , xn ) = 1}] = 7 ( )n−3 ( )(m−n+2) (29)
4 8

January 6, 2020—8 : 41 am DRAFT

A 3-CNF-SAT DESCRIPTOR ALGEBRA AND THE SOLUTION OF THE P =NP CONJECTURE 30

Proof:

• Let us re-order the m clauses ψj in such a way that each clause has only one new additional

variables with respect to the set of variables appearing in the previous clauses.

• Let Vk = {xi | ∃j , 1 ≤ j ≤ k : xi appears in ψj }. The re-ordering of the clauses yields

to embedded subsets V1 ⊆ V2 · · · ⊆ Vn−2 = {x1 , · · · , xn } with #V1 = 3, · · · , #Vk = k + 2, · · · ,

#Vn−2 = n and #Vk0 = n ∀ k 0 ≥ n − 2.

The cases where all clause ψk+1 introduces two or three new additional variables to Vk are to

be neglected, as this means that the 3-CNF-SAT problem can be split into two sub-problems

with one or zero common variable, which reduces drastically the complexity of the problem.

• Let us look at the expected effect of a clause ψj (1 ≤ j ≤ m) over the number of solutions :

1. Let us consider ψ1 .

The first clause yields to 7 · 2n−3 solutions. The matrix representation of ψ1 will be

a 7 × 3 matrix.

2. Let us consider ψ2 .

Let ψ2 introduces only one new additional variable xt , and let xr and xs be the two

common variables for ψ1 and ψ2 :

! !
xq xr xs xr xs xt
[ψ1 ] = and [ψ2 ] =
7 lines 7 lines

Depending on the sign of the literal xt in ψ2 , the result matrix for [ψ1 ∧ ψ2 ] will get

the 7 lines of [ψ1 ] with a zero in the column for xt if ψ2 = [¬]xr ∨ [¬]xs ∨ ¬xt or with

a one if ψ2 = [¬]xr ∨ [¬]xs ∨ xt . This corresponds to solutions where the literal [¬]xt

is satisfied.

On the contrary, when the value in the column for xt is opposite to the sign of

[¬]xt , the satisfiability of ψ2 should pass through the literals xr and xs . Among

the four possible values for (xr , xs ), only three will be accepted. One couple for

(xr , xs ) will be ruled out, as well in matrix [ψ1 ] as in [ψ2 ]. This corresponds to

one or two lines deleted in [ψ1 ], depending on the sign for xr and xs in ψ1 . The

expected number of lines deleted in [ψ1 ] will be : 1 · P [one deletion] + 2 · P [two

deletions] = 1 · 41 + 2 · 43 = 74 . Therefore, the expected number of lines in [ψ1 ∧ ψ2 ] will

be equal to 7 + (7− 74 ) = 7(1+ 43 ) = 7( 74 ) = 12, 25. And the boundaries for #[ψ1 ∧ψ2 ]
January 6, 2020—8 : 41 am DRAFT
A 3-CNF-SAT DESCRIPTOR ALGEBRA AND THE SOLUTION OF THE P =NP CONJECTURE 31

are [min2 , max2 ] = [12, 13].

3. Let us consider ψ3 .

Let ψ3 introduce a new additional variable. Using the same type of arguments as

for ψ2 , the expected number of deleted lines in [ψ1 ∧ ψ2 ] will be equal to :

E(# deletions in ψ1 ∧ ψ2 ) =
 2 
X13 X2
 d · P [d deletions |#[ψ1 ∧ ψ2 ] = k] · P [#[ψ1 ∧ ψ2 ] = k]
k=12 d=1

Let us consider here an example where #[ψ1 ∧ ψ2 ] = 13.

 
x1 x2 x3 x4
 

 0 0 0 0 

 

 0 1 0 0 

0 1 1 0 
 

 

 1 0 0 0 

 

 1 0 1 0 

 
 1 1 0 0 
For instance, [ϕ] = [(x1 ∨ x2 ∨ ¬x3 ) ∧ (x2 ∨ ¬x3 ∨ ¬x4 )] =  
1 1 1 0 
 

 

 0 0 0 1 

 

 0 1 0 1 

 
 0 1 1 1 
 
1 0 0 1 
 

 

 1 1 0 1 

1 1 1 1

Let us consider the couples (xi , xj ) and the number of deleted lines for each case :

xi xj # del. x1 x2 # del.
0 0 d1 0 0 2
0 1 d2 For the above example : 0 1 4
1 0 d3 1 0 3
1 1 d4 1 1 4
P4
We see that, whatever the value of #[ψ1 ∧ψ2 ], i=1 di = #[ψ1 ∧ψ2 ]. So, the expected

number of deleted clauses, independently from the case (xi , xj ), will be :

P
1 1 1 1 i di #[ψ1 ∧ ψ2 ]
d1 · + d2 · + d3 · + d4 · = =
4 4 4 4 4 4
January 6, 2020—8 : 41 am DRAFT
A 3-CNF-SAT DESCRIPTOR ALGEBRA AND THE SOLUTION OF THE P =NP CONJECTURE 32

V3
Therefore, the expected number of lines in [ i=1 ψi ] will be :

3
^ X
E[#[ ψi ]] = {#[ψ1 ∧ ψ2 ] + (#[ψ1 ∧ ψ2 ] − # deletions )} · P [#[ψ1 ∧ ψ2 ] = k]
i=1 k
13
X #[ψ1 ∧ ψ2 ]
= {#[ψ1 ∧ ψ2 ] + (#[ψ1 ∧ ψ2 ] −)} · P [#[ψ1 ∧ ψ2 ] = k]
4
k=12
   
1 3 1 1
= 12{1 + (1 − )} · + 13{1 + (1 − )} ·
4 4 4 4
2
^ 7
= E[#[ ψi ]] ·
i=1
4
7
= 7 · ( )2
4
= 21, 4375

V3
And #[ i=1 ψi ] ∈ [min3 , max3 ] = [12 + 12 − 4, 13 + 13 − 1] = [20, 25]

4. Let us now consider a given clause ψj (j ≤ n − 2).

We know that ψj introduces a new additional variable. Then, using the same type

of arguments as for ψ3 , the expected number of lines will be :

j j−1
^ ^ 7 7
E[#[ ψi ]] = E[#[ ψi ]] · ( ) = 7 · ( )j−1 .
i=1 i=1
4 4
j
^
and #[ ψi ] ∈ [2 minj−1 − 2j−1 , 2 maxj−1 − 1]
i=1
∈ [max{0, 2j−1 (8 − j)} , 6 · 2j−1 + 1]

5. So, for ψn−2 , we have (n ≥ 10) :

n−2
^ 7
E[#[ ψi ]] = 7 · ( )n−3 (30)
i=1
4
n−2
^
and #[ ψi ] ∈ [0 , 6 · 2n−3 + 1] (31)
i=1

6. For ψj where j > n − 2, no new variable will be added, and the number of

solution will only decrease.

Using the same previous argument, we can consider the six possible cases (xi , xj , xk )

and the corresponding di with 1 ≤ i ≤ 8. Here again, we get that :

8
X j−1
^
di = #[ ψi ]
i=1 i=1
January 6, 2020—8 : 41 am DRAFT
A 3-CNF-SAT DESCRIPTOR ALGEBRA AND THE SOLUTION OF THE P =NP CONJECTURE 33

#[ j−1
V
i=1 ψi ]
Vj−1
Thus, the expected number of deleted lines in [ i=1 ψi ] will be 8 .

There is no other operation to do for ψj . We only have to delete some lines in

Vj−1
[ i=1 ψi ]. So, computing the remaining lines, we get :
j j−1
^ ^ 7
E[#[ ψi ]] = E[#[ ψi ]] ·
i=1 i=1
8

And the boundaries will be :

j
^
#[ ψi ] ∈ [0, maxj−1 − 1]
i=1

7. Finally, for the last clause ψm , we get :

m
^ 7 7
E[#Sϕ ] = E[#[ ψi ]] = 7 ( )n−3 ( )(m−n+2)
i=1
4 8

m
^
and #[ ψi ] ∈ [0 , 6 · 2n−3 − m + n − 1]
i=1

Theorem VII.2: The most difficult exact uniformly distributed α-random 3-

m
CNF-SAT problems are the ones with a ratio α = n approximately equal to

5, 19.

Proof: Let us consider exact uniformly distributed α-random 3-CNF-SAT problems. The

most difficult problems are the ones where the decision between satisfiability and unsatisfia-

bility arises only when considering the last clause ψm . This is equivalent to have E[#Sϕ ] ≈ 1.

We get :

7 7
E[#Sϕ ] ≈ 1 ⇔ 7( )n−3 ( )m−n+2 ≈ 1
4 8
7
⇔ ( )m 2n ≈ 1
8
7
⇔ m log( ) + n log(2) ≈ 0
8
7
⇔ (α n) log( ) + n log(2) ≈ 0
8
7
⇔ α log( ) ≈ −log(2)
8
− log(2)
⇔α≈
log( 78 )
⇔ α ≈ 5, 19089307

January 6, 2020—8 : 41 am DRAFT

A 3-CNF-SAT DESCRIPTOR ALGEBRA AND THE SOLUTION OF THE P =NP CONJECTURE 34

Theorem VII.3: The relation between exact uniformly distributed α-random

3-CNF-SAT problems and usual α-random 3-CNF-SAT problems can be seen as

√
a reduction of the ratio α through the function : 3α − 1, 9098 α.

Proof:

When considering exact uniformly distributed α-random 3-CNF-SAT problems, each literal

occurs with exactly the same frequency in the m clauses, only the combination of the literals

in each clause being random. We have : #xi = 3 α.

But the usual uniform α-random 3-CNF-SAT problems are such that : E[#xi ] = 3 α, where

the variables are drawn randomly from a multinomial population with P [xi appears in a
3α
clause] = pi = m = n3 . For large n, the number of occurrence for each variable will asymptot-

ically follow a Normal distribution N (µ, σ 2 ) with µ = m·pi = 3 α and σ 2 = m·pi (1−pi ) ≈ 3 α.

If we consider, after sorting the clauses as explained in our descriptor approach, the sec-

ond half of the clauses (where Mα (t) ≥ t), we will get a folded normal distribution for

Di [usual α-random 3-CNF] = |#{xi } − E[#{xi }]| = |#{xi } − 3 α|. We have :

Di ∼ |N (0, 3 α)|
r
2
E[Di ] ≈ σ
π
r
6α p
≈ = 1, 9098 α
π
p
⇒ E[#{xi }] ≈ 3α − 1, 9098 α for the clauses where Mα (t) ≥ t

So, if we have #{xi } = α in the exact uniformly distributed α-random 3-CNF-SAT problems,
√
this corresponds to an “folded” expected occurency E[#{xi }] ≈ 3α − 1, 9098 α for usual

uniform α-random 3-CNF-SAT problems.

Corollary VII.1: The threshold α = 5, 19 found for exact uniformly distributed α-random

3-CNF-SAT problems corresponds approximatively to a reduced threshod

√
E[#{xi }] 1, 9 × 5, 19
α= = 5, 19 − = 4, 14135
3 3

for usual uniform α-random 3-CNF-SAT problems.

January 6, 2020—8 : 41 am DRAFT

A 3-CNF-SAT DESCRIPTOR ALGEBRA AND THE SOLUTION OF THE P =NP CONJECTURE 35

Log  (#  𝓢𝛗)  

18  

16  

14  

12  

10  

8  

6  

4  

2  

0  
0   50   100   150   200   250   300  
Ordered  clauses  for  3-­‐CNF-­‐SAT  problem  with  n=75  and  m=325  
uuf75-­‐01.cnf  from  hDp://www.satlib.org/ubcsat  

Computed  number  of  solu8ons   Expected  number  of  solu8ons  

Fig. 5 : Number of solutions with respect to the analyzed clauses

for a 3-CNF-SAT problem with n = 75 and m = 325

Note : This is still a theoretical value for the threshold. Indeed, for usual uniform α-random

generated 3-CNF-SAT problem, we detect a small difference between the observed and the
Vj
theoretical expected number of solutions with respect of the first j analyzed clauses i=1 ψi .
Vj
The theoretical expected number of solutions E[#[ i=1 ψi ]] is defined as 7 · ( 47 )s · ( 78 )t , where

s is the number of clauses in {ψ2 , · · · , ψj } introducing new additional variable and t the

number of remaining clauses. Figure 5 shows the situation for a 3-CNF-SAT problem with

n = 75 and m = 325, taken from http://www.satlib.org/ubcsat.

This difference shows that theoretical expected values are over-estimating the observed val-

ues. Let us note that we re-ordered the m clauses ψj in such a way that new additional

variables are appearing as lately as possible in the 3-CNF-SAT problem (in order not to

reach too large numbers for #Sϕ ).

VIII. Conclusions and future researches

Our researches were built on exact uniformly distributed α-random 3-CNF-SAT problems.

The complexity analysis was mainly done in terms of expected value for some characteristics.

We see that these expected values are over-estimating the real values. This means that our

conclusions about the most difficult value for α [= 5, 19], and therefore about the maximum

theoretical value for the complexity 2Mα (t) [= 2490 ] are over fitted. Future researches will try
January 6, 2020—8 : 41 am DRAFT
A 3-CNF-SAT DESCRIPTOR ALGEBRA AND THE SOLUTION OF THE P =NP CONJECTURE 36

to suppress this bias to be more accurate in our estimation of the complexity for the NP

problems.

We have seen that for α ≈ 5, 19, the maximum complexity for a α-random 3-CNF-SAT prob-

lem will be around 2490 whatever the number of variables. The NP problems are then

not exponential but bounded exponential problems. This makes them belong-

ing to P . But even with “yottaflops” computers (1024 instructions by second), this can

take about “10136575708788609985206606935908809922268405942697014391424964252461

88692463039064879247034987638184445605903560477” centuries to solve such problems. ^

¨

This is not exponential, only a huge constant upper bound.

Even if this paper is mostly theoretical, each theorem was validated by extensive numeri-

cal tests. Future researches will be to improve our different algorithms implementing the

descriptor approach for 3-CNF-SAT problems1 .

References

[1] Stanley Burris. A Course in Universal Algebra. Dover Pubns, City, 2012.

[2] Th. Cormen, Ch. Leiserson, R. Rivest, and Cl. Stein. Introduction to Algo-

ritmics. MIT Press, Cambridge, 2nd edition, 2001.

[3] James M. Crawford and Larry D. Auton. Experimental results on the

crossover point in random 3-sat. Artificial Intelligence, 81(12):31 – 57, 1996.

Frontiers in Problem Solving: Phase Transitions and Complexity.

[4] Rémon Marcel. About the impossibility to prove p != np or p

= np and the pseudo-randomness in np. Published in Arxiv :

http://arxiv.org/abs/0904.0698v2, January 2010.

[5] M. Sipser. The History and Status of the P versus NP Question. Proceed-

ings of the 24th Annual Meeting ACM, pages 603–618, 1992.

1
I would like to thank Dr. Johan Barthélemy for his help in terms of writing and testing these algorithms, as
well as the SMART department of the University of Wollongong for their welcome. Codes will be available on
www.github.com
January 6, 2020—8 : 41 am DRAFT