Enhanced Matrix Power Function for Cryptographic Primitive Construction

: A new enhanced matrix power function (MPF) is presented for the construction of cryptographic primitives. According to the deﬁnition in previously published papers, an MPF is an action of two matrices powering some base matrix on the left and right. The MPF inversion equations, corresponding to the MPF problem, are derived and have some structural similarity with classical multivariate quadratic (MQ) problem equations. Unlike the MQ problem, the MPF problem seems to be more complicated, since its equations are not deﬁned over the ﬁeld, but are represented as left–right action of two matrices deﬁned over the inﬁnite near-semiring on the matrix deﬁned over the certain inﬁnite, additive, noncommuting semigroup. The main results are the following: (1) the proposition of inﬁnite, nonsymmetric, and noncommuting algebraic structures for the construction of the enhanced MPF, satisfying associativity conditions, which are necessary for cryptographic applications; (2) the proof that MPF inversion is polynomially equivalent to the solution of a certain kind of generalized multivariate quadratic (MQ) problem which can be reckoned as hard; (3) the estimation of the effectiveness of direct MPF value computation; and (4) the presentation of preliminary security analysis, the determination of the security parameter, and speciﬁcation of its secure value. These results allow us to make a conjecture that enhanced MPF can be a candidate one-way function (OWF), since the effective (polynomial-time) inversion algorithm for it is not yet known. An example of the application of the proposed MPF for the Key Agreement Protocol (KAP) is presented. Since the direct MPF value is computed effectively, the proposed MPF is suitable for the realization of cryptographic protocols in devices with restricted computation resources.


Introduction
Quantum computers and Internets of Things (IoTs) are the recent technologies influencing the development of cryptographic protocols. The resistance to quantum cryptanalysis became important after Peter W. Shor proposed polynomial-time quantum cryptanalysis [1] for the traditional cryptographic primitives such as Diffie-Hellman key exchange protocol, RSA and ElGamal cryptosystems, Digital signature algorithm (DSA), and Elliptic Curve cryptosystem (ECC). These primitives were based on so-called classical candidate one-way functions (OWFs), such as discrete exponent and the multiplication of large prime numbers.
The demand for the effective realization of cryptographic protocols is influenced by the Internet of Things (IoT) technology.
One of the perspective trends is the creation of OWFs, the security of which relies on the NP-hard problems [2]. So far, there are no known effective quantum cryptanalytic algorithms solving NP-hard problems; therefore, this cryptographic trend is a significant part of so-called post-quantum cryptography. Several trends to create cryptographic primitives that can resist quantum cryptanalysis (2) Prove that MPF inversion is polynomially equivalent to the solution of a certain kind of generalized multivariate quadratic (MQ) problem which can be reckoned as hard; (3) Estimate the effectiveness of direct MPF value computation; (4) Perform preliminary security analysis and determine security parameters and their secure values.
By completing these tasks, we show that proposed the MPF can be a candidate OWF for cryptographic applications, namely for the realization of a key agreement protocol (KAP) presented in Construction 1.
The rest of this paper is organized as follows: In Section 2, we present an abstract MPF construction without the specification of concrete algebraic structures. We also present the main notations and definitions as well as a KAP in its symbolic form.
Algebraic structures for the introduction of the platform semigroup are proposed in Section 3, namely a medial semigroup S M and a modified medial semigroup S. The normal forms of these semigroups are constructed. On the base of the exponent semiring R = N 0 of natural numbers with zero, the near-semiring NSR is introduced to define exponent functions on S M and S.
In Section 4, the enhanced MPF, based on platform semigroup S and with power matrices defined over NSR, is constructed. It is proved that this MPF is associative and that the MPF problem is polynomially equivalent to a certain kind of generalized MQ problem, which is reckoned as hard. The conjecture is made that the MPF problem is hard as well. The asymptotic estimates of the effective computation of the direct MPF value are given.
In Section 5, some comments and conclusions are presented.
In Section 6 provides an illustration of a Key Agreement Protocol (KAP) with artificially small matrix orders.

The Construction of the Abstract MPF
An MPF is the function that computes the matrix obtained by powering some given matrix by two numerical matrices-one on the left and the other on the right. It is somewhat similar to matrix multiplication by two matrices on the left and right, respectively. The matrix that is powered is named the base matrix and the matrices that are powering the base matrix are named power matrices. In general, we define the base matrix over the multiplicative (semi)group S and power matrices over some numerical (semi)ring R. We call semigroup S a platform (semi)group-which, according to the MPF definition, is multiplicative-and R an exponent (semi)ring. The appropriate matrix semigroups M S and M R contain base matrices and power matrices, respectively.
There is some analogy of MPF with the well-known Diffie-Hellman function in terms of the action of one algebraic structure on the other. Let us assume that we have the commutative multiplicative group of integers modulo p, Z * p = {1, 2, . . . , p − 1}, and commutative additive group of integers modulo p − 1, Z p−1 = {0, 1, . . . , p − 2}. Then for any ω ∈ Z * p , and any x ∈ Z p−1 , the following discrete exponent (power) function can be defined: In this case we can say that group Z p−1 is acting on the group. Since this action is defined in commutative numerical algebraic structures, the notation of left and right action y ω x has no real meaning and is equivalent to ω xy = ω yx .
Let for example, X, Y, and W be matrices where X, Y are defined over the set of integers Z = {0, ±1, ±2, . . .} in matrix set M Z , and matrix W is defined over the set of complex integers C = Z + iZ, where i 2 = −1 in matrix set M C . Then, since matrices are non-commuting, the notion of left and right action of M Z in M C can be introduced in the following way: where matrices X, W, and Y are multiplied using the convenient matrix multiplication rule. In a similar way, we can define the two-sided MPF (or simply MPF), using matrices defined over certain algebraic structures. This symbolic form of MPF is as follows: We first define one-sided MPFs in the following way: The left-sided MPF corresponding to matrix W powered by matrix X on the left with MPF value equal to matrix C = c ij has the following form: Definition 2. The right-sided MPF corresponding to matrix W powered by matrix Y on the right with MPF value equal to matrix D = d ij has the following form: Using Definitions 1 and 2, we can now define the two-sided MPF.
Definition 3. The two-sided MPF corresponds to matrix W powered by matrix X on the left and by matrix Y on the right with MPF value equal to matrix Q = q ij , and is expressed in the following way: The MPF definition is related to the following associativity identities.

Identity 1.
MPF is one-side associative, (left-side and right-side associative, respectively) if the following identities hold: Identity 2. MPF is two-side associative if the following identities hold: Definition 4. MPF is associative if it is one-side and two-side associative.
It follows from Definition 3 that, in general, an MPF is a function F : M R × M S × M R → M S . Further, to be short, we will use the notation MPF R S for the definition of an MPF with a base matrix defined over platform semigroup S and with power matrices defined over exponent semiring R. Furthermore, we present the following definitions: Definition 5. The direct MPF value computation is to find matrix Q in Equation (3), when matrices X, Y, and W are given. Definition 6. The inverse MPF value computation is to find matrices X and Y in Equation (3), when matrices W and Q are given.
So far, all matrices in the construction of the MPF were defined over certain commutative algebraic structures, namely, the base matrix was defined over the commutative numerical (semi)group S and power matrices over the commutative numerical ring R.

Lemma 1.
If R is a commutative numerical semiring and S is a commutative semigroup, then the MPF is associative.
The proof can be found in [19]. In [28], the categorical interpretation of MPF, taken from [22], is presented in the context of the construction of several key agreement protocols. We slightly reformulate the notions used in this interpretation by the following statement, which is more appropriate for our study.
This means that there exist bilinear, right, and left actions of matrix semiring M R on matrix semigroup M S . According to the definition of action, it must satisfy the associative law corresponding to Definition 4. Since matrix semigroup M S is multiplicative, then the M R -semibimodule M S is multiplicative in our case.
Previously, we used the MPF in our research to construct cryptographic primitives, namely, KAP, symmetric, and asymmetric encryption protocols. The suggested protocols are based on the problem, defined as follows: The MPF problem is the computation of the MPF inverse value.
Having in mind that the existence of one-way functions (OWFs) is not proven yet, we will follow the common practice for the proposition of new candidate OWFs for cryptographic application. Consequently, the two necessary (but not sufficient) conditions for MPF to be a candidate (OWF) are the following.

Definition 8.
(1) The computation of the MPF direct value is computationally easy; (2) The MPF problem is polynomially equivalent to a certain hard problem without a known polynomial time algorithm.
By MPF-based OWF security, we mean the complexity of computation of the inverse MPF value, which corresponds to the solution of the MPF problem. In some cases, e.g., when a discrete logarithm function can be applied to the MPF, the MPF problem can be transformed to the special system of MQ equations defined over the field or ring. In our earlier publications, we referred to this problem as the matrix MQ (MMQ) problem [20]. However, MQ and MMQ problems have significant differences. The classical MQ problem corresponds to the system of random generated MQ equations consisting of unknown quadratic and linear monomials, while the MMQ problem corresponds to the system of equations generated by random matrices and has only quadratic monomials with unknown variables.
The computation of the direct MPF value is effective and its asymptotic polynomial time estimation is presented below in Section 4.
On the other hand, referencing the complexity of certain generalized MQ problems related to the MPF R S problem, it is shown that there is no known polynomial-time algorithm for its solution. Thus, we make the following conjecture. Conjecture 1. The necessary conditions for the construction of cryptographic protocols based on MPF R S are the following: (1) It is associative; (2) Matrices X and Y are members of some subsets M R 1 and M R 2 of commuting matrices in M R respectively, i.e., for any U ∈ M R 1 and V ∈ M R 2 UX = XU, YV = VY (9) (3) MPF R S satisfies the clauses of Definition 8.
The intriguing idea was to extend the MPF construction to noncommutative algebraic structures, namely, S and R, thereby giving higher expected complexity of the MPF problem and higher potential security for the construction of cryptographic primitives. The main problem with this approach was the loss of associativity of MPF, which made its application in cryptography impossible.
A breakthrough in the construction of an associative MPF based on noncommuting algebraic structures occurred when we found out that the infinite, noncommutative medial semigroup [29] (denoted by S M ) can be used. This paper presents the development of this idea by introducing modified medial semigroup S as a platform semigroup and a new kind of exponent functions on S with exponents defined in the specially constructed exponent near-semiring (NSR). The notion of the near-semiring can be found in [30]. The generic properties of semigroup S directly induce the properties of the NSR. The semigroup S is constructed from the medial semigroup S M by introducing two extra relations. Semigroups S M and S are infinite, multiplicative, and noncommutative. The NSR is infinite and noncommutative with respect to the addition operation.
If matrix W is defined over the noncommuting platform semigroup S, then for the construction of cryptographic protocols it is required to introduce (use) the normal form in S. This normal form is introduced in the next section. If MPF R S satisfies Conjecture 1, then the construction of a key agreement protocol (KAP) in symbolic form can be realized in the following way. Construction 1. Let X, U ∈ M R 1 and Y, V ∈ M R 2 (i.e., UX = XU, YV = VY), and let the public parameters be matrix W ∈ M S and subsets M R 1 and M R 2 . Then the KAP consists of the following steps.
(1) Alice chooses two secret matrices, X ∈ M R 1 and Y ∈ M R 2 , at random, then computes the MPF value A = X W Y and sends it to Bob; (2) Bob chooses two secret matrices, U ∈ M R 1 and V ∈ M R 2 , at random, then computes the MPF value B = U W V and sends it to Alice; (3) Alice and Bob compute the same secret key K A = K B = K in the following way: The security analysis of KAP, constructed on the base of algebraic structures introduced in Section 3, is presented in Section 4.

The Definition of Algebraic Structures
To construct a platform semigroup for MPF, the class of multiplicative medial semigroups [25] is used. We consider medial semigroup S M when its presentation consists of two generators a, b and relations R M written in the following way: where ω 1 and ω 2 are arbitrary nonempty words in S M , written in terms of generators a and b.
The reason for the introduction of the medial semigroup is the existence of the following identity, based on the relation R M , valid for all words ω 1 , ω 2 ∈ S M and any exponent e ∈ N 0 , where N 0 is the semiring of natural numbers with zero: To construct platform semigroup S for the MPF in Equation (3), two extra relations R 1 and R 2 are added to S M : These relations can be generalized in other forms. In our manuscript we will deal only with these ones.
Thus, modified medial semigroup S has the following presentation: Notice that we define S as a multiplicative, noncommuting, noncancellative, and infinite semigroup which is a nonsymmetric algebraic structure.

Remark 1.
The sum of exponents of generators a and b equals 5 on the left side of R 1 and R 2 in Equation (13) and equals 2 on the right side.
Semigroups S M and S are made monoids by introducing an empty word as a multiplicatively neutral element, denoted by 1. Then, conveniently, the following identities hold for all ω ∈ S M : Definition 9. In any semigroup S, two words ω and ω are equivalent, i.e., ω ≡ ω , if ω is obtained from ω by applying any sequence of relations defined in S to ω and vice versa. Two words ω and ω are equal, i.e., ω = ω if they are written in the same form, e.g., ω = ab 2 a 3 bb 4 and ω = aa 2 b 3 ab 4 .

Definition 10.
Equivalence class E q ω of any ω ∈ S consists of all words equivalent to ω.
For the introduction of the normal form in any semigroup S, we will follow the convenient normal form definition for groups. Definition 11. The normal form in any semigroup S is defined if there exists a surjective function η : S → S η (when S η ⊂ S) based on the set of relations defined in S and satisfying the following condition: ω 1 ≡ ω 2 if, and only if, images of any ω 1 and ω 2 in S have equal images in S η , i.e., (16) Returning to the semigroup S in Equation (14), there are infinitely many equivalent words to a certain word ω ∈ S-e.g., let ω = bb 4 a 5 a; then ω 1 = ba 5 b 4 a, ω 2 = ba 2 ba, ω 3 = ba 8 b 7 a, etc., are equivalent words being in the equivalence class On the contrary, there are elements with equivalence classes consisting of the single element, e.g., elements a i and b j .
Relations in S can be used in the direction to reduce the value of exponents of generators a and b. There are also words in S for which the exponent of generators a and b cannot be reduced. We call these elements elementary and they are included in the set ε, where i, j ∈ N 0 : According to the convenient practice in noncommuting cryptography, the construction of cryptographic protocols requires the introduction of the normal form. The purpose of a normal form is the unique interpretation of operations performed in noncommuting structures. In our case, the unique interpretation of entries of matrix Q in Equation (3) is required.
It is easy to see that any ω ∈ S M can be uniquely transformed to the following equivalent form using relation R M : where g 1 is the left boundary generator and g 2 is the right boundary generator, i.e., g 1 , g 2 ∈ {a, b}. This representation is obtained by grouping together generators a with their exponents and moving them to the left (and, analogously, generators b to the right), using relation R M , while g 1 , g 2 remain unchanged. Then, the exponents of the same grouped generators are summed up. Since relation R M preserves the sum of exponents of generators a and b, such a representation of ω is unique.
Proof. We must prove that any word ω ∈ S M can be uniquely expressed in the form of Equation (19) and that the normal forms of two words ω 1 and ω 2 are equal if, and only if, ω 1 ≡ ω 2 . According to Equation (18), let ω be expressed uniquely in the form ω 0 = g 1 a i 0 b j 0 g 2 . Then ω 0 can be rewritten to ω n f by assigning to the exponents β and α maximal values, either 0 or 1, defined by boundary generators g 1 and g 2 . If g 1 = b, then β = 1; else, β = 0. If g 2 = a, then α = 1; else, α = 0. Since this representation is in one-to-one correspondence with g 1 and g 2 , it is unique.
To prove the second statement, the set of words in their normal forms b β a i b j a α is considered. In this set, the multiplication operation can be introduced by the transformation of the resulting word ω = ω 1 ω 2 to the form ω n f . We denote the set of words written in the form ω n f by S M,n f .

Lemma 2.
Let T be an additive noncommuting semigroup consisting of the tuples (β, i, j, α), where α, β ∈ {0, 1} ⊂ N 0 and i, j ∈ N 0 , with the following addition operation: Then there is an isomorphism ϕ : S M,n f → T , which can be expressed by the following relation for any word ω n f : To be concise, we omit the proof. Now we prove that if ω 1 ≡ ω 2 , then Equation (16) holds, and vice versa. Let n f (ω) = ω n f as defined by Equation (19). Normal forms of ω 1 and ω 2 are expressed in the following way: n f (ω 1 ) = ω 1,n f = b β 1 a i 1 b j 1 a α 1 and n f (ω 2 ) = ω 2,n f = b β 2 a i 2 b j 2 a α 2 . Referencing Lemma 2, ϕ(ω 1,n f ) = (β 1 , i 1 , j 1 , α 1 ) and ϕ(ω 2,n f ) = (β 2 , i 2 , j 2 , α 2 ). Since the sum of exponents is preserved under transformations using relation R M and according to the rule of assignment of values 0 and 1 to the exponents α 1 , β 1 and α 2 , β 2 , we have The proof in the opposite direction is the following: let ω 1,n f = ω 2,n f = ω n f ; then ω n f represents its equivalency class with respect to relation R M . The members of this equivalency class will be words ω 1 and ω 2 obtained by applying R M ; hence, ω 1 ≡ ω 2 .
The normal form in S is constructed on the base of the normal form in S M by using relations R M , R 1 , and R 2 for the subsequent minimization of exponents i a and j b in Equation (19). The following functions are required to be introduced: the floor function i/j for the ratio of natural numbers i, j and the minimum function min(i, j) for i, j ∈ N 0 . Taking into account Remark 1, we introduce the following variable based on these functions: Let us consider the word ω n f written in equivalent form ω 0 in Equation (18). If i 0 = j 0 = 5, then when using R M and any of relations R 1 or R 2 the following identity holds: This relation can be applied for the one-step minimization of exponents i 0 , j 0 ≥ 5. Taking in mind Remark 1, the exponents i 0 and j 0 can be expressed in the following unique way: Then, using variable µ defined in Equation (20), the word ω 0 can be rewritten in the following equivalent form: Taking into account relations R M , R 1 , and R 2 and Equation (21), we obtain the following word ω 1 equivalent to ω 0 and ω ∈ S: These considerations allow us to create the normal form in semigroup S.
Theorem 2. For the given word ω n f in normal form of S M , its normal form ω η in S is represented by the function η : S M → S and expressed by a finite n-step recurrent minimization procedure of exponents i a and j b in Equation (19) using relations R M , R 1 , and R 2 : Proof. We rewrite word ω n f to ω 0 according to Equation (18) and perform the minimization procedure for this word, obtaining word ω 1 : Let both i 0 , j 0 ≥ 5. Then, the first step to minimize exponents i 0 and j 0 is performed using Equation (18)- (21). As a result, the new equivalent word ω 1 The unique word ω 0 representation by ω 1 follows from the unique representation of i 0 and j 0 by Equation (22). If both i 1 , j 1 ≥ 5 in ω 1 , then the second step of minimization is performed analogously, and the unique word ω 2 is obtained with exponents i 2 = i 1 − 3µ 1 , j 2 = j 1 − 3µ 1 and This recurrence is continued up to the (n − 1) th step while relation i n−1 , j n−1 ≥ 5 does not hold. The unique word ω n−1 is obtained in the form: If ω n−1 is in the set of elementary words ε defined above in Equation (17), then the minimization procedure stops. The word ω n−1 is uniquely transformed to the normal form in S by the function n f (ω n−1 ) = ω η in Equation (19). Otherwise, the word ω n−1 is subsequently minimized using the last nth step minimization procedure according to the following two alternative conditions. If β n−1 = 1, i n−1 ≥ 5, and j n−1 = 4, then the final nth step of minimization is applied by using relation R 1 in Equation (11). If α n−1 = 1, i n−1 = 4, and j n−1 ≥ 5, then R 2 is applied to end the minimization process. Finally, the unique normal form ω n ∈ S in S is obtained, represented by Equation (25).
So far, we have considered exponent functions defined on S, determined by non-negative exponents in semiring R = N 0 . We generalize these functions by introducing certain "imaginary" exponents, yielding "complex" exponents and having some weak analogy with complex numbers in classical numerical algebra based on the imaginary unit i (i 2 = −1).
According to [26] and other related sources, the set of all mappings on an additive semigroup with identity zero is the standard example of a so-called near-semiring. It is a closed algebraic structure with two operations, namely, addition and multiplication (composition) of mappings.
A near-semiring is a nonempty set A with two binary operations "+" and "·", and a constant 0 such that (A; +; 0) is a monoid (not necessarily commutative) and (A;·) is a semigroup. These structures are related by one (right or left) distributive law, and, accordingly, the 0 is the one-side (right or left, respectively) absorbing element.
Following this general construction, we introduce a special kind of near-semiring (NSR), defining exponent functions on the modified medial semigroup S. In order to preserve the main properties of exponent functions (they are specified in Equation (39) below), we must replace the one-side distributive law and absorbing (zero) element law by two-sided ones, respectively.
Despite the difference between the convenient definition and ours, we will assume that NSR introduced below stands for the special-type near-semiring. The notion special-type is implied by the fact that medial semigroups are special-type semigroups and that exponent functions are special-type functions on these semigroups.

Definition 12
A near-semiring NSR is a nonempty set with two binary operations "+" and "·", such that (NSR; +; 0) is an additive monoid with neutral element 0 ∈ N 0 , and (NSR; ·; 1) is a multiplicative monoid with neutral element 1 ∈ N 0 , satisfying the following (two-sided) axioms for all x, y, z in NSR: Since exponent functions are mappings on the modified medial semigroup S, defined by exponent values in the NSR, we supply the monoid (NSR; +; 0) with the following extra relation R + of the medial semigroup type introduced in Equation (11). For all nonzero x, y, z 1 , z 2 in NSR, the following relation R + holds: In addition, we assume that multiplicative monoid (NSR; ·, 1) is commutative, i.e., for all x, y ∈ NSR, In analogy with complex numbers over the integers or reals, we introduce the "imaginary" unit denoted by ι and satisfying the following relations for all u ∈ N 0 : where the first relation is implied by Equation (33). The exponent ι of generators a and b in S is defined in the following way: We claim, that NSR consists of the union of the following sets: where the set N 0 + ιN 0 + N 0 defines the class of elements {t + ι · u + v} and the set ιN 0 Theorem 3. The set NSR is closed with respect to addition and multiplication operations and is a near-semiring according to Definition 12.

Proof:
The closure of NSR follows directly from the relation Equations (11)- (13). Indeed, we can consider only two classes of elements in NSR: {t + ι · u + v} defining the set N 0 + ιN 0 + N 0 and {ι · t + u + ι · v} defining the set ιN 0 + N 0 + ιN 0 , where t, u, v ∈ N 0 . The classes of elements {t + ι · u} = N 0 + ιN 0 and {ι · t + u} = ιN 0 + N 0 are the partial cases of corresponding sets N 0 + ιN 0 + N 0 and ιN 0 + N 0 + ιN 0 , respectively, when v = 0. After performing operations between the elements of classes {t + ι · u} and {ι · t + u} and grouping similar terms according to relation R + in Equation (32), we obtain an element either in the set ιN 0 + N 0 + ιN 0 or in N 0 + ιN 0 + N 0 . The operations with any other kind of pairs of elements does not yield any other elements except the elements in the sets N 0 + ιN 0 + N 0 and ιN 0 + N 0 + ιN 0 when similar terms are grouped using relation R + .
Referencing to Equations (30)- (35) and Theorem 3, the only two kinds of "complex" exponents x = t + ι · u + v and y = ι · t + u + ι · v, where t, u, v ∈ N 0 , can be defined for any generator a or b. For example, for generator a we claim, that a x = a t+ι·u+v = a t a ι·u a v = a t b u a v ; a y = a ι·t+u+ι·v = a ι·t a u a ι·v = b t a u b v .
Generalizing the last equations to any word ω in S, we obtain the following relations: where the word ω is obtained from ω by interchanging generators (a to b and b to a, respectively), according to Equations (37). It is evident that the exponent function in S M and S satisfies the following convenient identities for any ω, ω 1 , ω 2 ∈ S (S M ) and any x, y ∈ NSR: Theorem 5. Let ω be either in S M or in S and x ∈ NSR; then the normal form of ω x in S is expressed by the word b β a i b j a α defined in Theorem 2.

Proof:
Let ω ∈ S M and x ∈ NSR. Then, referencing Theorems 3 and 4 and using Equations (37)-(39), ω x can be transformed to the word ω ∈ S M with exponents of generators in N 0 . Then, by applying Theorem 1, word ω can be transformed to its normal form ω n f in S M , and by applying Theorem 2, to its normal form ω η in S. If ω ∈ S, the same procedure is applied.
Hence, we constructed near-semiring NSR, defining exponent functions on semigroups S M and S.

Enhanced MPF Construction and Its Security Investigation
According to notation introduced in Section 2, we are dealing with the problems denoted by are associative, i.e., they satisfy the identities given in Equations (7) and (8).
We start from the consideration of MPF N 0 S M and MPF N 0 S problems. We assume that, initially, the base matrix W entries ω ij in Equation (3) are in semigroup S and are expressed in normal forms according to Equation (25) (Theorem 2), i.e., Power matrices X = {x il } and Y = y jk are defined over the semiring N 0 , i.e., x il , y jk ∈ N 0 , where i, j, k, l ∈ I (m) = {1, 2, . . . , m}. Using Equation (40), the entry q ik of matrix Q in Equation (3) can be expressed in the following way: Referencing Equation (19) (Theorem 1), every q ik can be transformed to its normal form q ik,n f in S M , thus obtaining matrix Q n f : where exponents i a,ik and j b,ik are exponents in N 0 , α ik , β ik ∈ {0, 1}, and i, j, k, l ∈ I (m) .
The latter Equations (41) and (42) represent the system of equations corresponding to MPF N 0 S M in Equation (3). According to [31], it can be called a special type of multivariate quadratic power problem, when unknown monomials are presented in exponents. However, the results of NP-completeness presented in [31] cannot be applied to this problem since it is defined over different algebraic structures with additional commutation constraints (Equation (9)) to random generated matrices X, Y.
The MPF N 0 S M problem can be transformed to an MPF N 0 S problem, transforming entries q ik,n f to their normal forms q ik,η ∈ S. Then, according to Equation (25) (Theorem 2), the normal forms of q ik,η are found after the n-step recurrent minimization procedure: As a result, we obtain matrix Q η = q ik,η , which replaces matrix Q in Equation (3).
In the case of an MPF N 0 S M problem, the explicit relations Equations (41) and (42) can be used, relating the exponents of the entries of matrix Q n f = q ik,n f with the entries of power matrices X = {x il } and Y = y jk , and with the exponents of generators a and b in the entries of the base matrix W = ω ij . Since (in this case) x il and y jk are elements in N 0 , the sum of exponents of generators a and b in the word q ik are preserved and can be expressed by the following formulas: These equations can be rewritten in a matrix form by introducing the following matrix notations: Matrices E a , E b and H a , H b are obtained from matrices W and Q = Q n f in Equation (3), respectively.
Referencing Definition 9, we can redefine MPF N 0 S M as follows: Definition 13. The MPF N 0 S M problem is to find matrices X and Y in Equation (44), satisfying commutation conditions Equation (9), when matrices E a , E b and H a , H b are given. Equation (44) represents a special type of multivariate quadratic (MQ) problem, since it is generated by randomly generated matrices X and Y and defined over the semiring N 0 . In our previous publication, we called this kind of problem a matrix MQ (MMQ) problem [20]. MMQ equations do not have linear monomials and consist only of quadratic ones. The significant difference between MQ problems and MMQ problems, represented by Equation (44), is that in the latter case matrix equations are defined over the semiring N 0 rather than over the field or ring.
Structurally, the MPF N 0 S M problem is related with the known exact non-negative matrix factorization (Exact NMF) problem [32]. We denote the non-negative matrix H with m rows and n columns by H m×n .

Definition 14.
Exact NMF problem: The input is a matrix H m×n with non-negative entries whose rank is exactly k, for k ≥ 1 . The output is a pair of non-negative matrices X m×k , Y k×n , satisfying the equation If no such pair of matrices (X m×k , Y k×n ) exists, then the output is a statement of the nonexistence of the solution. The decisional version of the Exact NMF problem takes the same input and gives as output "YES" if such a (X m×k , Y k×n ) pair exists and outputs "NO" otherwise.
In [32] it is proved that the Exact NMF problem is NP-hard.
In [33] the exponential-or even super-exponential-time for a generalization of the Exact NMF problem solution was presented.
In [34], a polynomial-time algorithm for the Exact (and Approximate) NMF problem for every constant k was given. This result holds also for the instances when m = n = k, corresponding to instances of matrix Equation (44), assuming that matrices E a , E b are identity matrices.
However, this trivial (polynomial-time) reduction from Exact NMF to Equation (44) does not correspond to the MPF N 0 S M problem according to Definition 14, since Exact NMF do not include commutativity conditions on matrices X and Y in Equation (9). The Exact NMF problem complexity with commutativity constraints is not yet known and, therefore, this relation cannot tell anything about the MPF N 0 S M problem complexity. In the case of the MPF N 0 S problem, the entries of matrix Q in Equation (3) are transformed to the normal forms of semigroup S by the finite minimization procedure of exponents of generators a and b, according to Theorem 2, thus obtaining matrix Q η . Then, instead of Equation (3), we have the following equation: where X and Y are any matrices satisfying the commutation conditions in Equation (3). This equation will not have a solution with high probability, since, in general, the resulting exponents of generators in the entries of matrix Q η will be too small to satisfy Equation (46). For example, let us consider the left side MPF in Equation (2) and, for simplicity, let m = 2. Let, for example, c 11 = (aba) 2 (ba) 3 ; then, after powering words a 5 b 3 a and ba by exponents 2 and 3, respectively, and transforming the resulting word to the normal form of semigroup S M , we obtain the word a 15 b 12 . When this word is transformed to the normal form of S, we obtain c 11 = a 6 b 3 . As we see, there are no two exponents to obtain the word a 6 b 3 by exponentiating initial words a 5 b 3 a and ba without their minimization, i.e., transforming the exponentiation result to the normal form of S. The same is valid when Equation (46) is transformed to the system of matrix equations in the form of Equation (44).

Theorem 7.
If instances of the MPF N 0 S problem are generated in such a way that entries of matrix W are written in the normal form of S and the product of any two entries of matrix W is reducible by relations R M , R 1 , R 2 in Equations (11) and (13), then Equation (46) has solution if, and only if, matrix X has exactly one entry equal to 1 ∈ N 0 in each row and matrix Y has exactly one entry equal to 1 ∈ N 0 in each column, while all other entries in matrices X and Y are equal to 0 ∈ N 0 .
Proof. Sufficiency: Assume that matrix X has exactly one entry equal to 1 ∈ N 0 in each row and matrix Y has exactly one entry equal to 1 ∈ N 0 in each column, while all other entries in matrices X and Y are equal to 0 ∈ N 0 . Then matrix Q η entries in Equation (46) will be equal to certain entries of matrix W depending on the distribution of 1s in matrices X and Y. So, Equation (46) will have a solution.
Necessity: Assume that Equation (46) has a solution, then powers of generators a and b in entries of elements of matrix Q η are not reduced using R 1 , R 2 . Taking in mind that the product of any two entries of matrix W is reducible, we conclude that matrix X has exactly one entry equal to 1 ∈ N 0 in each row and matrix and Y has exactly one entry equal to 1 ∈ N 0 in each column, while all other entries in matrices X and Y are equal to 0 ∈ N 0 .
The probability is negligible that a random generated matrix X has one entry equal to 1 ∈ N 0 in each row, and Y has one entry equal to 1 ∈ N 0 in each column, with other entries in X and Y equal to 0 ∈ N 0 . For example, if entries of m × m matrices X and Y are randomly generated with uniform distribution in the set of natural numbers {0, 1, . . . , 2 q − 1}, then this probability is p = m 2m 2 −2qm 2 . For example, for m = 4 and q = 3, the probability is p = 2 −64 ; for m = 8 and q = 3, it is p = 2 −336 . For brevity, the following proposition is formulated without proof. Proposition 1. If instances of matrix W are generated at random and satisfy the conditions of Theorem 7, and matrices X and Y are generated at random with uniform distribution in the set of natural numbers {0, 1, . . . , 2 q − 1}, then Equation (46) has a solution with asymptotic, negligible, exponentially decreasing probability in parameters m and q.
The possible way to obtain an equivalent system in order to find any matrices X and Y, when matrices W and Q η = q ik,η are given, is to transform matrix Q η to equivalent matrix Q R by using relations R M , R 1 , R 2 in Equations (11) and (13), in the reverse direction than was done when matrix Q in Equation (3) was transformed to the normal form Q η . We denote the transformation of the word ω in the reverse direction in S by R −r ω , where r denotes the number of transformation steps. Then we denote the transformation of q ik,η in the reverse direction by r ik steps by R −r ik ik . To express the reverse transformation of matrix Q η we construct a transformation matrix R η = R −r ik ik , i, k ∈ I (m) = {1, 2, . . . , m}. Then, using the Hadamard matrix product , the matrix Q R = R η Q η is obtained; this can be expressed in the following way: As a result, the powers of generators a and b in Q R will be increased, expecting to satisfy Equation (47) with new matrix Q R instead of Q η in the right side. However, the transformation of words to their normal forms is a surjective mapping, denoted by η above. In general, a word normal form represents infinitely many equivalent words in S. Therefore, if the words in their normal forms are presented in an m × m matrix Q η , then it is not clear which equivalent matrix Q R to choose to guarantee the solution of Equation (47). If matrices X and Y are generated at random, then definitely, for every entry q ik,η of Q η , the different number of steps r ik will be required to search for matrix Q R satisfying Equation (47). If we are transforming matrix Q η in the reverse direction in at most one step, i.e., r < 2, then the exhausting set of possible equivalent matrices consists of 2 m 2 elements. If r > 1, then this set consists of (r − 1) m 2 elements and is super-exponentially growing in m. For every matrix Q R in this set, the analogous system of matrix equations like Equation (44) can be retrieved and this system will have a solution if, and only if, Equation (47) has a solution.

Definition 15
The MPF N 0 S problem is to find any X and Y satisfying Equation (47) and commutation conditions Equation (9), when matrices W and Q η are given, where Q R = R η Q η and R η is any reverse transformation matrix found by a random search procedure in an exponentially large set, providing solution existence of Equation (47).  (3) if matrices X and Y are defined over the NSR and are in M NSR , and matrix W is defined over the S and is in M S . We will show that this problem is much more complicated than the MPF N 0 S problem. It inherits the MPF N 0 S problem's difficulties, since the statements of Theorem 6 and Proposition 1 can be reformulated for the MPF NSR S problem as well. Therefore, the analogue of Equation (46) will not have a solution with an overwhelming probability and a search procedure for the suitable matrix Q R in the right side of Equation (46) must be performed in the same way as in the case of the MPF N 0 S problem. An Equation of (47) type can be found with the difference that matrices X and Y are defined over the NSR. That is what the MPF NSR S problem inherits from the MPF N 0 S , problem together with the same formal Definition 15. Therefore, for further consideration, the references to Equations (46) and (47) will be used to represent the MPF NSR S problem, as well.
Recall that, according to Lemma 2 in Theorem 1, T is an additive noncommuting semigroup consisting of the tuples (β, i, j, α) and there exists an isomorphism ϕ : S M,n f → T , where S M,n f is a semigroup of words in normal forms of S M . Recall that ϕ b β a i b j a α = (β, i, j, α). Since NSR possesses an imaginary variable, which we denoted by ι, the transformation of Equation (47) to a system of the same type as Equation (44) is not possible, since exponents of generators a and b are confused and cannot be equated. We denote the set of matrices over non-commuting additive semigroup T by M T , which is a semigroup with respect to the matrix sum, when matrix elements are summed according to Lemma 2 in Theorem 1. Proof. Referencing isomorphism ϕ in Theorem 1 and Equations (14) and (35), the action • of the "imaginary" unit ι in NSR to the tuple (β, i, j, α) in T for distinct α, β ∈ {0, 1} and i, j ∈ N 0 is expressed in the following way: (37) and (38), the action of "complex" exponents x = t + ι · u + v and y = ι · t + u + ι · v (where t, u, v ∈ N 0 ) to the elements of T can be found. For example, if α = β = 1, the following relation takes place for x:

Referencing Equations
For distinct α, β ∈ {0, 1} and two kinds of "complex" exponents x and y, we obtain eight equations of type Equation (48), which are omitted here for brevity.
Since T is isomorphic to the sub-semigroup of normal forms of the medial semigroup S M , we can claim that there exists a bilinear (right and left) action of the matrix near-semiring M NSR on matrix semigroup M T in the same way as bilinear action is defined for MPF in Equation (3). Since the set of words ω η in normal forms of S is a subset of S M,n f , then by applying isomorphism ϕ to matrix W and Q η in Equation (46) we obtain matrices E and P η in M T , respectively: Then, referencing Equations (30)-(39), the following relations take place: where • is an MPF-induced action operation of matrices in M NSR to matrices in M T and corresponds to formal matrix multiplication as in Equation , while multiplication between entries is performed according to Equation (48).
Since ϕ is an isomorphism and is computed in polynomial time, by combining Equations (49) and (50), Equation (48) can be rewritten in the following form: Analogously to Equation (48), the obtained system will not have a solution with overwhelming probability and, hence, the same search procedure of matrix P R = ϕ(Q R ) must be performed to solve the MPF NSR S problem. Then Equation (51) has to be replaced by the following equation: To prove the polynomial equivalence of Equations (47) and (52), we must prove that if X and Y is a solution of Equation (47), then it is a solution of Equation (52), and vice versa. This statement holds, since ϕ is the polynomial-time computable isomorphism and since identities Equations (30)-(39) hold.
Referencing Theorem 8, the following polynomially equivalent definition to Definition 15 can be formulated.

Definition 16. The MPF NSR
S problem is to find any X and Y over NSR satisfying Equation (52) and commutation conditions Equation (9), when matrices E and P η are given, where P R = R η P η , and R η is any reverse transformation matrix, found by a random search procedure in an exponentially large set, providing the solution existence of Equation (52).
Since isomorphism ϕ is computed effectively, an analogous dichotomy can be formulated in this case.

Dichotomy 2.
If the procedure of the matrix P R search can be performed in polynomial time, then the decisional MPF NSR S problem is in NP; otherwise, it is not in NP and not in Co-NP.
The MPF NSR S problem according to Definition 16 represents a new kind of MQ problem-namely a matrix MQ (MMQ) problem-which is not defined over the (finite) field or ring. It can be interpreted as the bilinear action of matrices X and Y over the NSR to matrix E over the infinite additive noncommuting semigroup T.
We are making a conjecture that this problem is hard, since it is defined over much more complicated algebraic structures than the classical MQ problem which is used for the creation of cryptographic primitives, e.g., in the HFE cryptosystem. The classical solution methods of the MQ problem, such as Grobner bases or Linearization, are not adequate in this case, since we are dealing with noncommuting and nonsymmetric algebraic systems. So far, there is no understanding on how to deal with system Equation (52). By way of analogy with the classical MQ problem, we can say that matrix Equation (52) corresponds to the system of m 2 equations and 2m 2 unknown variables in NSR, satisfying Equations (37) and (38). In terms of the MQ problem, this system is neither over-defined, nor under-defined; that increases its complexity.

Conjecture 2.
Since the MPF NSR S problem is polynomially equivalent to the special kind of generalized MQ problem, which is reckoned to be hard, it can be considered as a candidate one-way function (OWF) for the construction of cryptographic primitives.
The effectiveness of the computation of the direct MPF NSR S value is based on the computation of small exponents using multiplication and addition. The size of exponents can be up to 8 bits. After powering matrix W by matrices X and Y in Equation (3), the generators in entries of the obtained matrix are rearranged using Equation (11) and exponents of same generators are summed up, representing the product of the same generators by the one corresponding generator with a certain exponent. After that, the transformation to the normal form of the semigroup S is performed. The asymptotic computation complexity is estimated using the complexity of multiplications of three matrices presented in Equation (2)  The security parameter is the order m of matrices used in the MPF. Since there is no theory dealing with Equation (52) due to its definition over noncommuting and nonsymmetric algebraic structures, the secure value of security parameter m is determined by heuristic analogy with the known classical MQ problem. For example, there is no theory on finding a probability of solution existence for an MPF problem for random chosen matrix Q in Equation ((3) as it can be done easily for the system of linear equations. Even for the classical MQ problem and decades of its investigation, based on more or less developed theory, this problem is investigated using a modelling technique on toy examples [35]. Hence, it is impossible so far to obtain a rigorous lower bond for the security parameter m for the MPF NSR S problem, taking in mind Definition 16. Many authors have shown that known methods for the solution of a randomly generated MQ problem (e.g., based on Grobner bases and Linearization) rarely give results better than an exhaustive search method. Since, according to Equation (52), the MPF problem is isomorphic to some kind of generalized MQ problem, we accept the assumption that, at this moment, the best attack to the MPF problem could be the exhaustive search attack.
Referencing the analogy of a randomly generated MQ problem over the field GF(2)-when the solution of a system of equations with more than 80 equations and 80 variables, using known methods of solution (say, Grobner bases or Linearization), is intractable-the security parameter value for the MPF problem is determined. In this case, the exhaustive search of 80 variables in GF(2) runs in time 2 80 . For example, considering the hidden field equation (HFE) algorithm described in the extended version of [35], the attack presented in [36] is expected to run in time 2 152 , when the extended linearization (XL) method is used. According to [37], a possible improvement of this attack runs in time 2 82 , which is still worse than the 2 80 complexity of exhaustive search.
Taking an analogy of these considerations, we propose to choose m ≥ 10, hence determining the lower bound of the m value heuristically, which may seem to have extra in reserve. If m = 10, then the matrix Equation (52) corresponds to m 2 = 100 equations with 2m 2 = 200 unknown variables in NSR. We are making a conjecture that it will be currently enough to prevent any cryptanalysis attack (though unknown yet) on the key agreement protocol presented in Construction 1, including by an exhaustive search attack. In this consideration, the additional complexity to find matrix P R in Equation (52) was not taken into account, which contributes to a significant part of the MPF NSR S problem's complexity.
According to Construction 1, the public key (PuK) for the KAP construction is matrix W and the private key (PrK) is matrices X, Y (and U, V, respectively) of dimension m×m. Let entries w lj = b β lj a r lj b s lj a α lj ; a, b ∈ S; α lj , β lj ∈ {0, 1}; r lj , s lj ∈ N of matrix W be expressed in the normal form of S according to the Theorem 2. Assume, that w lj is expressed by 8 bits by assigning two bits to α lj , β lj and six bits to r lj , s lj . If the value of security parameter m = 10, then 8 × 100 = 800 bits will be required for the storage of matrix W.
Let entries x il , y jk of matrices X, Y be expressed in NSR in the form (36). Then, assigning the same 8 bit length for these entries, the storage of these requires 2 × 8 × 100 = 1600 bits.
With this connection, we think we can currently formulate the following conjecture.

Discussion and Conclusions
Despite the fact that the existence of one-way functions (OWFs) is not yet proved, many authors are trying to propose new so-called conjectured or candidate OWFs for cryptographic applications and showing the complexity of their inversion by associating them with other polynomially equivalent hard problems. This paper is just one more attempt to do so by introducing a new enhanced matrix power function (MPF) based on infinite, nonsymmetric, and noncommutative algebraic systems. It can be expected that the use of such algebraic systems can increase the complexity of the inversion of the corresponding candidate OWF and, as a consequence, the security of the constructed cryptographic primitives.
The main problem in making useful MPFs based on noncommuting algebraic systems for cryptography is the loss of associativity. This paper presents a solution to this problem.
The complexity of the MPF NSR S problem implies the complexity of the inversion of the corresponding candidate OWF and the security of the cryptographic primitives. It is shown that the MPF NSR S problem is polynomially equivalent to some generalized MQ problem, which is reckoned to be hard. It is the so-called matrix MQ (MMQ) problem, which is not defined over the (finite) field, but can be interpreted as bilinear left-right action of matrices X and Y over the NSR to matrix E over a certain (introduced here) infinite, nonsymmetric, additive, and noncommuting semigroup T.
It is not yet proven that the computational MPF NSR S problem is NP-hard, since the polynomial-time reduction from any NP-hard to MPF NSR S problem is not yet found, despite significant attempts. As is known, quantum cryptanalytic algorithms can easily cope with problems which have some periodicities or symmetries. Since MPF NSR S is defined over infinite, nonsymmetric, and noncommuting algebraic structures (having no periodicities or symmetries), it seems that current quantum cryptanalytic algorithms could face a problem dealing with this kind of problem. The actual complexity class of the MPF NSR S problem is not yet known. If the proof that the MPF NSR S problem is NP-hard is found, and if a certain quantum cryptanalysis algorithm solves any problem in the NP-hard complexity class in polynomial time, then it will solve the MPF NSR S problem in polynomial time as well.
The security parameter m is defined as corresponding to the order of matrices used in the MPF definition. Since there is no theory dealing with the system of MPF equations (due to its definition over the noncommuting and nonsymmetric infinite algebraic structures), the security parameter's value is determined by heuristic analogy with the known classical MQ problem.
In this connection, the secure value for m is proposed to be m ≥ 10. Then, in order to solve the MPF NSR S problem and to break the key agreement protocol (KAP) in Construction 1, the adversary must solve matrix Equation (52) corresponding to the system of m 2 = 100 equations with 2m 2 = 200 unknown variables in NSR.
The asymptotic time for the computation of the direct MPF NSR S value is effective and is performed in O m 3 log 2 m . The computation of the direct value of many traditional candidate OWFs is based on the modular exponent function in large groups. These operations are time-consuming and usually require the use of special co-processors to speed up computations. Instead of exponentiation of large integers (512-2048 bits long) in the case of traditional candidate OWFs, we can deal with m × m matrices and perform operations with their entries of a few bits in length.
The main practical result of this paper is the construction of a KAP, definition of security parameters, and calculation of secure parameter values. The public key (PuK) for KAP construction is matrix W and the private key (PrK) is matrices X, Y (and U, V, respectively) of dimension m × m. The security parameter is m and its secure value is proposed to be m ≥ 10. Then the storage requirement for the public and private keys are 800 and 1600 bits, respectively. The latter is based on heuristic analogy with the well-known MQ problem since there is no theory yet dealing with the problems involved (i.e., noncommuting and nonsymmetric algebraic systems). Deeper consideration of the MPF complexity presented here can be performed in subsequent studies, thereby influencing the key lengths and security parameter values.
Referencing the heuristic analogy with the MQ problem, when existing methods for the MQ problem solution rarely give results better than total scan (except for in some special cases), the proposed PrK length prevents total scan attack. Since the methods of solution of MPF NSR S are unknown yet and there is not even any theory on how to handle such a kind of system of equations, we hypothesize that the proposed key length in practice will provide sufficient security against a total scan attack and the attack presented in [27].
Taking in mind that the computation of the direct MPF NSR S value is effective and its inversion is polynomially equivalent to a certain type of generalized MQ problem, we are making a conjecture that MPF NSR S can be a candidate OWF. Overall, the novelty of results presented in the paper can be summarized as follows: a new type of candidate OWF based on infinite, noncommuting, and nonsymmetric algebraic structures is presented in the class of OWFs of noncommuting cryptography. According to our knowledge so far, only a few nonsymmetric OWFs have been proposed in this class. Hence, we can expect to achieve greater security even against quantum cryptanalysis since it is more effective in cryptosystems using symmetric structures with periodicities.

Example of KAP Realization with Artificially Small Matrix Orders
We consider matrices of order 3 × 3. The public matrix W with entries written in normal and in shortened form when a 0 = 1, b 0 = 1 is chosen as follows: According to Construction 1, we have matrices X, U ∈ M R 1 , Y, V ∈ M R 2 , where M R 1 and M R 2 are subsets of M S of commuting matrices. For effective protocol realization we propose to use M R 1 and M R 2 as subsets of circulant matrices [38]. According to Equation (36), we define subsets M R 1 and M R 2 as sets of circulant matrices defined either over the set N 0 + ιN 0 or the set ιN 0 + N 0 .
The following theorem is presented without proof.
Theorem 9. For any circulant matrix X ∈ M R 1 , and any circulant matrix U ∈ M R 2 (or vice versa), X and U are commuting.

Further Research
The main challenge to investigating the security of the proposed MPF is to continue attempts to prove the NP-completeness of the MPF problem. On the other side, a more consistent investigation of MPF security against possible attack scenarios is required. These objectives can be achieved by deeper theoretical investigation of MPF cryptanalytic equations, since there is no theory yet dealing with the problems involved.