Research article

Bounds of a unified integral operator for (s,m)-convex functions and their consequences

  • Received: 25 January 2020 Accepted: 18 June 2020 Published: 28 June 2020
  • MSC : 26D10, 31A10, 26A33

  • The unified integral operator presented in Definition 4 produces several kinds of known fractional and conformable integral operators. The goal of this paper is to obtain bounds of this unified integral operator by using the definition of (s, m)-convexity. The resulting inequalities in specific cases represent the bounds of many known fractional and conformable fractional integral operators in a compact form.

    Citation: Zitong He, Xiaolin Ma, Ghulam Farid, Absar Ul Haq, Kahkashan Mahreen. Bounds of a unified integral operator for (s,m)-convex functions and their consequences[J]. AIMS Mathematics, 2020, 5(6): 5510-5520. doi: 10.3934/math.2020353

    Related Papers:

    [1] Hui Zhou, Bo Lu, Huaguang Gu, Xianjun Wang, Yifan Liu . Complex nonlinear dynamics of bursting of thalamic neurons related to Parkinson's disease. Electronic Research Archive, 2024, 32(1): 109-133. doi: 10.3934/era.2024006
    [2] Zilu Cao, Lin Du, Honghui Zhang, Lianghui Qu, Luyao Yan, Zichen Deng . Firing activities and magnetic stimulation effects in a Cortico-basal ganglia-thalamus neural network. Electronic Research Archive, 2022, 30(6): 2054-2074. doi: 10.3934/era.2022104
    [3] Yapeng Zhang, Yu Chen, Quanbao Ji . Dynamical analysis of spontaneous Ca2+ oscillations in astrocytes. Electronic Research Archive, 2024, 32(1): 405-417. doi: 10.3934/era.2024020
    [4] Xia Shi, Ziheng Zhang . Multiple-site deep brain stimulation with delayed rectangular waveforms for Parkinson's disease. Electronic Research Archive, 2021, 29(5): 3471-3487. doi: 10.3934/era.2021048
    [5] Feibiao Zhan, Jian Song, Shenquan Liu . The influence of synaptic strength and noise on the robustness of central pattern generator. Electronic Research Archive, 2024, 32(1): 686-706. doi: 10.3934/era.2024033
    [6] Chengtian Ouyang, Huichuang Wu, Jiaying Shen, Yangyang Zheng, Rui Li, Yilin Yao, Lin Zhang . IEDO-net: Optimized Resnet50 for the classification of COVID-19. Electronic Research Archive, 2023, 31(12): 7578-7601. doi: 10.3934/era.2023383
    [7] Honghui Zhang, Yuzhi Zhao, Zhuan Shen, Fangyue Chen, Zilu Cao, Wenxuan Shan . Control analysis of optogenetics and deep brain stimulation targeting basal ganglia for Parkinson's disease. Electronic Research Archive, 2022, 30(6): 2263-2282. doi: 10.3934/era.2022115
    [8] Ping Wang, Changgui Gu, Huijiu Yang, Haiying Wang . Identify the characteristic in the evolution of the causality between the gold and dollar. Electronic Research Archive, 2022, 30(10): 3660-3678. doi: 10.3934/era.2022187
    [9] Ariel Leslie, Jianzhong Su . Modeling and simulation of a network of neurons regarding Glucose Transporter Deficiency induced epileptic seizures. Electronic Research Archive, 2022, 30(5): 1813-1835. doi: 10.3934/era.2022092
    [10] Yi Chen, Benhuan Nie, Zhehao Huang, Changhong Zhang . Spatial relevancy of digital finance in the urban agglomeration of Pearl River Delta and the influence factors. Electronic Research Archive, 2023, 31(8): 4378-4405. doi: 10.3934/era.2023224
  • The unified integral operator presented in Definition 4 produces several kinds of known fractional and conformable integral operators. The goal of this paper is to obtain bounds of this unified integral operator by using the definition of (s, m)-convexity. The resulting inequalities in specific cases represent the bounds of many known fractional and conformable fractional integral operators in a compact form.


    Throughout this article, let Cm×n stand for the set of all m×n matrices over the field of complex numbers; A stands for the conjugate transpose of ACm×n; r(A) stands for the rank of ACm×n; Im stands for the identity matrix of order m and [A,B] stands for a columnwise partitioned matrix consisting of two submatrices A and B. We introduce the concept of generalized inverses of a matrix. For an ACm×n, the Moore–Penrose generalized inverse of A is defined to be the unique matrix XCn×m that satisfies the four Penrose equations:

    (1) AXA=A,  (2) XAX=X,  (3) (AX)=AX,  (4) (XA)=XA. (1.1)

    A matrix XCn×m is called an {i,,j}-generalized inverse of ACm×n, denoted by A(i,,j), if it satisfies the ith, ,jth matrix equations in (1.1). Notice that solutions of each of the four equations in (1.1) are not necessarily unique. Hence, the set of all {i,,j}-generalized inverses of the matrix A is usually denoted by the symbol {A(i,,j)}. There are altogether 15 sorts of {i,,j}-generalized inverses of A by the above definition, but the following

    A,  A(1,3,4),  A(1,2,4),  A(1,2,3),  A(1,4),  A(1,3),  A(1,2),  A(1) (1.2)

    are usually called the eight commonly-used types of generalized inverses of A in the literature. Admittedly, the Moore–Penrose generalized inverse of a matrix is a natural extension of the standard inverse of an invertible matrix, which has become one of the basic and classic concepts in matrix theory and its applications due to its usefulness in dealing with singular matrices. For more basic results and facts concerning generalized inverses of matrices and their properties, we refer the reader to the three references [1,2,3].

    Let us consider the following two general matrix equalities:

    A1B(i1,,j1)1C1+A2B(i2,,j2)2C2++AkB(ik,,jk)kCk=D, (1.3)
    A1B(i1,,j1)1A2B(i2,,j2)2AkB(ik,,jk)kAk+1=A, (1.4)

    where it is assumed that A1, A2, , Ak+1, B1, B2, , Bk, C1, C2, , Ck and A and D are matrices of appropriate sizes. Clearly, (1.3) and (1.4) are generalizations of various concrete matrix equalities that involve mixed operations of ordinary sums and products of several matrices and their generalized inverses. For the sake of exposition, let us take the following two well-known examples:

    B(i,,j)=B(i1,,j1)1+B(i2,,j2)2++B(ik,,jk)k, (1.5)
    (BkB2B1)(i,,j)=B(i1,,j1)1B(i2,,j2)2B(ik,,jk)k, (1.6)

    which are called the parallel sum of matrices and the reverse-order law for generalized inverses of multiple matrix products, respectively. As demonstrated in Lemma 2.2 below, generalized inverses of a matrix can be written as certain linear and bilinear matrix expressions involving two arbitrary matrices. Hence, the two matrix equalities in (1.3) and (1.4) can be equivalently represented as two kinds of linear and nonlinear matrix equations involving multiple arbitrary matrices (variable matrices).

    As we know, in algebra theory, or a given algebraic expression or equality that involves variable entries, algebraists can propose many kinds of questions relevant to its properties and performances, but of particular interest regarding the expression or equality is to determine explicit conditions under which the expression is invariant with respect to the variable entries or the equality holds for some/all values of the variable entries included. In view of this consideration, it is natural to ask the questions related to determining necessary and sufficient conditions for the matrix equalities in the context of (1.3) and (1.4) to hold respectively for some/all of the generalized inverses of the matrices in them. In fact, this sort of research problems were proposed and investigated by lots of authors focused generalized inverses of matrices and its various applications [4,5,6,7]. Prompted by this kind of work, the present author and his collaborators considered a selection of linear and nonlinear matrix expressions or matrix equalities that involve one or more arbitrary matrices, established a fundamental theory related to the invariance properties of different types of linear and nonlinear matrix expressions and presented their essential utilizations in the characterizations of many kinds of matrix equalities that involve algebraic operations of generalized inverses of matrices [8,9,10,11,12,13,14]. Particularly in two recent papers [15,16], the present author considered (1.4) for {1}-generalized inverses of matrices of B1, B2, , Bk with k=2,3, deriving different types of identifying conditions under which the given matrix equalities always hold with respect to all possible {1}-generalized inverses of the given matrices by using a diversity of precise rank equalities associated with certain specified block matrices that are constructed from the given matrices in (1.4). Obviously, characterization problems of (1.3) and (1.4) involve lots of difficult and divisive issues because they are quite inclusive for various possible choices of {i,,j}-generalized inverses of the given matrices, thus, there exist many concrete cases of (1.3) and (1.4) that we need to give further deep explorations.

    As an ongoing study of the above algebraic equality characterization problems on matrices, the main objective of this article is to consider several cases of the matrix equalities in (1.3) and (1.4) for the mixed combinations of {1}- and {1,2}-generalized inverses of matrices with k=2,3, and to give identifying conditions under which these algebraic matrix equalities always hold by using a series of ordinary but skillful calculations of ranks of matrices. The rest of this paper is organized in five sections. In Section 2, the present author introduces a sequence of well-known or existing basic formulas, facts and results related to ranks, ranges and generalized inverses of matrices that we shall utilize in the derivations of our main results regarding (1.3) and (1.4). In Section 3, the author derives necessary and sufficient conditions for the matrix equalities in (1.3) to hold with mixed combinations of {1}- and {1,2}-generalized inverses of matrices for k=2,3 by means of the elementary but the powerful block matrix representation method (BMRM) and the matrix rank method (MRM). In Section 4, the author gives identifying conditions for the matrix equalities in (1.4) to always hold with mixed combinations of {1}- and {1,2}-generalized inverses of matrices for k=2,3. In Section 5, the author presents a variety of parallel conclusions related to several kinds of algebraic matrix equalities that involve different mixed algebraic operations of given matrices and their weighted {1}- and {1,2}-generalized inverses. An application is presented in Section 6 for a matrix equality involving generalized inverses of a 2×2 block matrix. Concluding remarks are given in Section 7.

    In this section, the author presents some well-known or existing results and conclusions regarding a two-sided linear matrix equation, the general expressions of generalized inverses of a matrix, as well as some equalities and inequalities associated with ranks of block matrices, which we shall adopt as fundamental and feasible matrix analysis tools in the examinations of the algebraic matrix equality problems described in Section 1.

    Lemma 2.1 ([17]). Let

    AXB=C (2.1)

    be a two-sided linear matrix equation, where ACm×n, BCp×q and CCm×q are given. Then, the following four statements are equivalent:

    (a) Equation (2.1) is solvable for X.

    (b) R(A)R(C) and R(B)R(C).

    (c) r[A,C]=r(A) and r[BC]=r(B).

    (d) AACBB=C.

    In this case, the general solution of (2.1) can be written in the following parametric expression

    X=ACB+FAU+VEB, (2.2)

    where FA=InAA, EB=IpBB and U,VCn×p are arbitrary matrices. Particularly, (2.1) holds for all matrices XCn×p if and only if

    [A,C]=0  or  [BC]=0. (2.3)

    Lemma 2.2 ([17]). Let ACm×n. Then, the general expressions of A(1) and A(1,2) can be written as

    A(1)=A+FAU+VEA, (2.4)
    A(1,2)=(A+FAU)A(A+VEA), (2.5)

    where EA=ImAA, FA=InAA and U,VCn×m are arbitrary.

    Recall a simple fact in linear algebra that a matrix is null if and only if its rank is zero. As a direct consequence of this fact, we see that two matrices A and B of the same size are equal if and only if the matrix rank equality r(AB)=0 holds. In other words, these obvious two equivalent statements connect the matrix equality A=B alternatively to the matrix rank equality r(AB)=0. In view of this assertion, we can see that if certain nontrivial analytical formulas for calculating the rank of AB are obtained, then we can directly employ them to reveal essential connections between the two matrices. Of course, the usefulness of this proposed methodology is based on the fact that we can precisely calculate ranks of matrix expressions by using elementary operations of matrices. As a matter of fact, algebraists realized that numerous matrix rank formulas can be established in reasonable ways, while these matrix rank formulas have been recognized as powerful tools to construct, characterize and understand hidden structures and properties of various algebraic matrix expressions and matrix equalities.

    Lemma 2.3 ([18]). Let ACm×n,BCm×k and CCl×n. Then,

    r[A,B]r(A), (2.6)
    r[ABC0]r(B)+r(C). (2.7)

    Lemma 2.4 ([19,20]). Let ACm×n, BCm×k, CCl×n and DCl×k be given. Then,

    maxA(1)r(DCA(1)B)=min{r[C,D],  r[BD],  r[ABCD]r(A)}, (2.8)
    maxA(1,2)r(DCA(1,2)B)=min{r(A)+r(D),r[C,D],r[BD], r[ABCD]r(A)}. (2.9)

    Therefore,

    CA(1)B=D holds for all A^{(1)}[C,D]=0, or [BD]=0, or r[ABCD]=r(A); (2.10)
    CA(1,2)B=D holds for all A^{(1,2)} A=0 and D=0, or [C,D]=0, or [BD]=0, or r[ABCD]=r(A). (2.11)

    The cases of the matrix equalities in (1.3) for k=2 and {1}- and {1,2}-generalized inverses of B1 and B2 are given by

    A1B(1)1C1+A2B(1)2C2=D, (3.1)
    A1B(1)1C1+A2B(1,2)2C2=D, (3.2)
    A1B(1,2)1C1+A2B(1)2C2=D, (3.3)
    A1B(1,2)1C1+A2B(1,2)2C2=D, (3.4)

    where A1Cm×p1, B1Cq1×p1, C1Cq1×n, A2Cm×p2, B2Cq2×p2, C2Cq2×n and DCm×n. In this section, we first give conditions under which the four matrix equalities in (3.1)–(3.4) always hold for all {1}- and {1,2}-generalized inverses of B1 and B2.

    Theorem 3.1. Let A1, B1, C1, A2, B2, C2 and D be as given in (3.1). Then, the following three statements are equivalent:

    (a) Equation (3.1) holds for all B(1)1 and B(1)2.

    (b) A1B(1)1C1+A2B(1)2C2 is invariant with respect to all B(1)1 and B(1)2 and A1B1C1+A2B2C2=D holds.

    (c) Any one of the following 9 conditions holds:

    (i) A1=0, A2=0 and D=0.

    (ii) A1=0, C2=0 and D=0.

    (iii) C1=0, A2=0 and D=0.

    (iv) C1=0, C2=0 and D=0.

    (v) A1=0 and r[B2C2A2D]=r(B2).

    (vi) C1=0 and r[B2C2A2D]=r(B2).

    (vii) A2=0 and r[B1C1A1D]=r(B1).

    (viii) C2=0 and r[B1C1A1D]=r(B1).

    (ix) r[B10C10B2C2A1A2D]=r(B1)+r(B2).

    Proof. By (2.10), we first obtain that (2.10) holds for all generalized inverses B(1)1 if and only if

    [A1,DA2B(1)2C2]=0 or [C1DA2B(1)2C2]=0 or r[B1C1A1DA2B(1)2C2]=r(B1). (3.5)

    Subsequently, we see from (2.10) that the first equality in (3.5) holds for all generalized inverses B(1)2 if and only if

    A1=0 and any one of [A2,D]=0,  [C2D]=0,  r[B2A2C2D]=r(B2), (3.6)

    as required for (i), (ii) and (v) in (c). Meanwhile by (2.10), the second equality in (3.5) holds for all generalized inverses B(1)2 if and only if

    C1=0 and any one of the three conditions [A2,D]=0,  [C2D]=0,  r[B2A2C2D]=r(B2), (3.7)

    as required for (iii), (iv) and (vi) in (c). Further, by (2.8),

    maxB(1)2r[B1C1A1DA2B(1)2C2]=maxB(1)2r([B1C1A1D][0A2]B(1)2[0,C2])=min{r[B1C10A10A2], r[B1C1A100C2], r[B10C10B2C2A1A2D]r(B2)}. (3.8)

    Hence, the third equality in (3.5) holds for all generalized inverses B(1)2 if and only if

    r[B1C10A10A2]=r(B1)  or  r[B1C1A100C2]=r(B1)  or  r[B10C10B2C2A1A2D]=r(B1)+r(B2). (3.9)

    Also by (2.6) and (2.7), we see that

    r[B1C10A10A2]r[B1,C1]+r(A2)r(B1)  and  r[B1C1A100C2]r[B1A1]+r(C2)r(B1).

    Combining the rank inequalities with the first two equalities in (3.9) yields (vii) and (viii) in (c). The third rank inequality in (3.9) is just (ix) in (c).

    It can be seen from the nine conditions in Theorem 3.1(c) that if the matrices Ai0 and Ci0 in (3.1) for i=1,2, then (3.1) holds for all B(1)1 and B(1)2 if and only if Condition (ix) in Theorem 3.1(c) holds. This special fact brings convenience of using the assertions in Theorem 3.1.

    By a similar approach, we can show that the following two theorems hold, and we omit the details of their proofs.

    Theorem 3.2. Let A1, B1, C1, A2, B2, C2 and D be as given in (3.3). Then, the following three statements are equivalent:

    (a) Equation (3.2) holds for all B(1,2)1 and B(1)2.

    (b) A1B(1,2)1C1+A2B(1)2C2 is invariant with respect to all B(1,2)1 and B(1)2 and A1B1C1+A2B2C2=D holds.

    (c) Any one of the following 12 conditions holds:

    (i) A1=0, A2=0 and D=0.

    (ii) A1=0, C2=0 and D=0.

    (iii) B1=0, A2=0 and D=0.

    (iv) B1=0, C2=0 and D=0.

    (v) C1=0, A2=0 and D=0.

    (vi) C1=0, C2=0 and D=0.

    (vii) A1=0 and r[B2C2A2D]=r(B2).

    (viii) B1=0 and r[B2C2A2D]=r(B2).

    (ix) C1=0 and r[B2C2A2D]=r(B2).

    (x) A2=0 and r[B1C1A1D]=r(B1).

    (xi) C2=0 and r[B1C1A1D]=r(B1).

    (xii) r[B10C10B2C2A1A2D]=r(B1)+r(B2).

    In particular, assume that the matrices Ai0 and Ci0 in (3.4) for i=1,2. Then (3.3) holds for all B(1,2)1 and B(1)2 if and only if Condition (xii) in (c) holds.

    Theorem 3.3. Let A1, B1, C1, A2, B2, C2 and D be as given in (3.4). Then, the following three statements are equivalent:

    (a) Equation (3.4) holds for all B(1,2)1 and B(1,2)2.

    (b) A1B(1,2)1C1+A2B(1,2)2C2 is invariant with respect to all B(1,2)1 and B(1,2)2 and A1B1C1+A2B2C2=D holds.

    (c) Any one of the following 16 conditions holds:

    (i) A1=0, A2=0 and D=0.

    (ii) A1=0, B2=0 and D=0.

    (iii) A1=0, C2=0 and D=0.

    (vi) B1=0, A2=0 and D=0.

    (v) B1=0, B2=0 and D=0.

    (vi) B1=0, C2=0 and D=0.

    (vii) C1=0, A2=0 and D=0.

    (viii) C1=0, B2=0 and D=0.

    (ix) C1=0, C2=0 and D=0.

    (x) A1=0 and r[B2C2A2D]=r(B2).

    (xi) B1=0 and r[B2C2A2D]=r(B2).

    (xii) C1=0 and r[B2C2A2D]=r(B2).

    (xiii) A2=0 and r[B1C1A1D]=r(B1).

    (xiv) B2=0 and r[B1C1A1D]=r(B1).

    (xv) C2=0 and r[B1C1A1D]=r(B1).

    (xvi) r[B10C10B2C2A1A2D]=r(B1)+r(B2).

    In particular, assume that the matrices Ai0 and Ci0 in (3.4) for i=1,2. Then (3.4) holds for all B(1,2)1 and B(1,2)2 if and only if Condition (xvi) in (c) holds.

    The case of (1.3) for k=3 is given by

    A1B(i1,,j1)1C1+A2B(i2,,j2)2C2+A3B(i3,,j3)3C3=D, (3.10)

    where A1Cm×p1, B1Cq1×p1, C1Cq1×n, A2Cm×p2, B2Cq2×p2, C2Cq2×n, A3Cm×p3, B3Cq3×p3, C3Cq3×n and DCm×n. Clearly, there are eight possible expressions in (3.10) for {1}- and {1,2}-generalized inverses of B1, B2 and B3. In the remaining part of this section, we only give two groups of conditions under which (3.10) always holds for all {1}- and {1,2}-generalized inverses of B1, B2 and B3. It is easy to see that, if any one Ai and Ci in (3.10) is zero for i=1,2,3, (3.10) reduces to two-term cases on the left-hand side. So, we assume that Ai0 and Ci0 in (3.10) for i=1,2,3 without loss of generality.

    Theorem 3.4. Let Ai, Bi, Ci and D be as given in (3.10) and assume that Ai0 and Ci0 for i=1,2,3. Then, the following three statements are equivalent:

    (a) A1B(1)1C1+A2B(1)2C2+A3B(1)3C3=D holds for all B(1)1, B(1)2 and B(1)3.

    (b) A1B(1)1C1+A2B(1)2C2+A3B(1)3C3 is invariant with respect to all B(1)1, B(1)2 and B(1)3 and A1B1C1+A2B2C2+A3B3C3=D holds.

    (c) r[B100C10B20C200B3C3A1A2A3D]=r(B1)+r(B2)+r(B3).

    Proof. Rewriting A1B(1)1C1+A2B(1)2C2+A3B(1)3C3=D as A1B(1)1C1+A2B(1)2C2=DA3B(1)3C3 and applying Theorem 3.1(c), we obtain that it holds for all B(1)1 and B(1)2 if and only if r[B10C10B2C2A1A2DA3B(1)3C3]=r(B1)+r(B2). Consequently, applying (2.8) and (2.10) to the matrix equality yields (c).

    By a similar approach, we can show the following result, so we omit the details of its proof.

    Theorem 3.5. Let Ai, Bi, Ci and D be as given in (3.10) and assume that Ai0, Bi0 and Ci0 for i=1,2,3. Then, the following three statements are equivalent:

    (a) A1B(1,2)1C1+A2B(1,2)2C2+A3B(1,2)3C3=D holds for all B(1,2)1, B(1,2)2 and B(1,2)3.

    (b) A1B(1,2)1C1+A2B(1,2)2C2+A3B(1,2)3C3 is invariant with respect to all B(1,2)1, B(1,2)2 and B(1,2)3 and A1B1C1+A2B2C2+A3B3C3=D holds.

    (c) r[B100C10B20C200B3C3A1A2A3D]=r(B1)+r(B2)+r(B3).

    By induction, we are able to derive the following result.

    Theorem 3.6. Assume that Ai0, Bi0 and Ci0 in (1.3) are for i=1,2,,k. Then, (1.3) holds for all mixed combinations of {1}- and {1,2}-generalized inverses of B1,B2,,Bk if and only if

    r[B1000C10B200C200BkCkA1A2AkD]=r(B1)+r(B2)++r(Bk).

    The cases of the matrix equalities in (1.4) for k=2 and {1}- and {1,2}-generalized inverses of B1 and B2 are given by

    A1B(1)1A2B(1)2A3=A, (4.1)
    A1B(1)1A2B(1,2)2A3=A, (4.2)
    A1B(1,2)1A2B(1)2A3=A, (4.3)
    A1B(1,2)1A2B(1,2)2A3=A, (4.4)

    where A1Cm1×m2, B1Cm3×m2, A2Cm3×m4, B2Cm5×m4, A3Cm5×m6 and ACm1×m6. Particularly, the case of (4.1) for {1}-generalized inverses of B1 and B2 was studied in [11,15], and the following results were obtained.

    Theorem 4.1 ([11,15]). Let A1, B1, A2, B2, A3 and A be as given in (4.1). Then, the following three statements are equivalent:

    (a) Equation (4.1) holds for all B(1)1 and B(1)2.

    (b) A1B(1)1A2B(1)2A3 is invariant with respect to B(1)1 and B(1)2 and A1B1A2B2A3=A holds.

    (c) Any one of the following 6 conditions holds:

    (i) A=0 and A1=0.

    (ii) A=0 and A2=0.

    (iii) A=0 and A3=0.

    (iv) A=0 and r[0A1A2B1]=r(B1).

    (v) A=0 and r[0A2A3B2]=r(B2).

    (vi) r[A0A10A2B1A3B20]=r[A2B1B20]=r(B1)+r(B2).

    In what follows, we give the results on the other three cases for {1}- and {1,2}-generalized inverses of B1 and B2 in (4.1).

    Theorem 4.2. Let A1, B1, A2, B2, A3 and A be as given in (4.2). Then, the following three statements are equivalent:

    (a) Equation (4.2) holds for all B(1)1 and B(1,2)2.

    (b) A1B(1)1A2B(1,2)2A3 is invariant with respect to all B(1)1 and B(1,2)2 and A1B1A2B2A3=A holds.

    (c) Any one of the following 7 conditions holds:

    (i) A=0 and A1=0.

    (ii) A=0 and A2=0.

    (iii) A=0 and B2=0.

    (iv) A=0 and A3=0.

    (v) A=0 and r[0A1A2B1]=r(B1).

    (vi) A=0 and r[0A2A3B2]=r(B2).

    (vii) r[A0A10A2B1A3B20]=r[A2B1B20]=r(B1)+r(B2).

    Proof. The equivalence of (a) and (b) follows from the definition of the Moore–Penrose inverse of a matrix and its uniqueness.

    By Lemma 4.1(c), we first see that the product A1B(1)1A2B(1,2)2A3=A holds for all generalized inverses B(1)1 if and only if any one of the following three conditions holds:

    (i) A=0 and A1=0.

    (ii) A=0 and A2B(1,2)2A3=0.

    (iii) r[AA1A2B(1,2)2A3B1]=r(B1).

    Consequently, we derive the conditions under which (ii) and (iii) hold, respectively. By (2.11), the equality A2B(1,2)2A3=0 in (ii) of the above three conditions holds for all generalized inverses B(1,2)2 if and only if

    A2=0  or  B2=0  or  A3=0  or  r[0A2A3B2]=r(B2) (4.5)

    holds. So, the above condition (ii) is equivalent to (ii), (iii), (iv) and (vi) in (c). By (2.9),

    maxB(1,2)2r[AA1A2B(1,2)2A3B1]=maxB(1,2)2r([AA10B1]+[0A2]B(1,2)2[A3,0])=min{r[AA10B1]+r(B2),  r[AA100B1A2],  r[AA10B1A30],  r[A0A10A2B1A3B20]r(B2)}.

    Hence, the rank equality in (iii) of the above three conditions holds for all generalized inverses B(1,2)2 if and only if

    r[AA10B1]+r(B2)=r(B1)  or  r[AA100B1A2]=r(B1)  or  r[AA10B1A30]=r(B1) (4.6)
    or  r[A0A10A2B1A3B20]=r(B1)+r(B2). (4.7)

    By (2.7),

    r[AA10B1]+r(B2)r(A)+r(B1)+r(B2)r(B1)

    holds. Combining it with the first equality in (4.6) leads to r(A)=r(B2)=0; thus, the first equality in (4.6) is equivalent to (iii) in (c). By (2.6) and (2.7),

    r[AA100B1A2]r(A)+r[B1,A2]r(B1)

    holds. Combining it with the second equality in (4.6) leads to A=0 and r[0A1A2B1]=r(B1); thus, the second equality in (4.6) is equivalent to (v) in (c). By (2.7),

    r[AA10B1A30]r[AA3]+r(B1)r(B1)

    holds. Combining it with the third equality in (4.6) leads to A=0 and A3=0; thus the third equality in (4.6) is involved in (iv) in (c). The rank equality in (4.7) is equivalent to (vii) in (c).

    We also declare that the following theorems can be shown by a similar approach, so the details of the proofs are omitted.

    Theorem 4.3. Let A1, B1, A2, B2, A3 and A be as given in (4.4). Then, the following three statements are equivalent:

    (a) Equation (4.4) holds for all B(1,2)1 and B(1,2)2.

    (b) A1B(1,2)1A2B(1,2)2A3 is invariant with respect to all B(1,2)1 and B(1,2)2 and A1B1A2B2A3=A holds.

    (c) Any one of the following 8 conditions holds:

    (i) A=0 and A1=0.

    (ii) A=0 and B1=0.

    (iii) A=0 and A2=0.

    (iv) A=0 and B2=0.

    (v) A=0 and A3=0.

    (vi) A=0 and r[0A1A2B1]=r(B1).

    (vii) A=0 and r[0A2A3B2]=r(B2).

    (viii) r[A0A10A2B1A3B20]=r[A2B1B20]=r(B1)+r(B2).

    The case of (1.4) for k=3 is given by

    A1B(i1,,j1)1A2B(i2,,j2)2A3B(i3,,j3)3A4=A, (4.8)

    where A1Cm1×m2, B1Cm3×m2, A2Cm3×m4, B2Cm5×m4, A3Cm5×m6, B3Cm7×m6, A4Cm7×m8 and ACm1×m8. Particularly, the case of (4.8) for {1}-generalized inverses of B1, B2 and B3 was studied in [16], and the following results were obtained.

    Theorem 4.4 ([16]). Let A1, B1, A2, B2, A3, B3, A4 and A be as given in (4.8). Then, the following three statements are equivalent:

    (a) A1B(1)1A2B(1)2A3B(1)3A4=A holds for all B(1)1, B(1)2 and B(1)3.

    (b) A1B(1)1A2B(1)2A3B(1)3A4 is invariant with respect to all B(1)1, B(1)2 and B(1)3 and A1B1A2B2A3B3A4=A holds.

    (c) Any one of the following 10 conditions holds:

    (i) A=0 and A1=0.

    (ii) A=0 and A2=0.

    (iii) A=0 and A3=0.

    (iv) A=0 and A4=0.

    (v) A=0 and r[0A1A2B1]=r(B1).

    (vi) A=0 and r[0A2A3B2]=r(B2).

    (vii) A=0 and r[0A3A4B3]=r(B3).

    (viii) A=0 and r[00A10A2B1A3B20]=r[A2B1B20]=r(B1)+r(B2).

    (ix) A=0 and r[00A20A3B2A4B30]=r[A3B2B30]=r(B2)+r(B3).

    (x) r[A00A100A2B10A3B20A4B300]=r[0A2B1A3B20B300]=r(B1)+r(B2)+r(B3).

    For the other seven matrix equalities in (4.8) with mixed combinations of {1}- and {1,2}-generalized inverses of matrices, we have the following results.

    Theorem 4.5. Let A1, B1, A2, B2, A3, B3, A4 and A be as given in (4.8). Then, the following three statements are equivalent:

    (a) A1B(1)1A2B(1)2A3B(1,2)3A4=A holds for all B(1)1, B(1)2 and B(1,2)3.

    (b) A1B(1)1A2B(1)2A3B(1,2)3A4 is invariant with respect to all B(1)1, B(1)2 and B(1,2)3 and A1B1A2B2A3B3A4=A holds.

    (c) Any one of the following 11 conditions holds:

    (i) A=0 and A1=0.

    (ii) A=0 and A2=0.

    (iii) A=0 and A3=0.

    (iii) A=0 and B3=0.

    (v) A=0 and A4=0.

    (vi) A=0 and r[0A1A2B1]=r(B1).

    (vii) A=0 and r[0A2A3B2]=r(B2).

    (viii) A=0 and r[0A3A4B3]=r(B3).

    (ix) A=0 and r[00A10A2B1A3B20]=r[A2B1B20]=r(B1)+r(B2).

    (x) A=0 and r[00A20A3B2A4B30]=r[A3B2B30]=r(B2)+r(B3).

    (xi) r[A00A100A2B10A3B20A4B300]=r[0A2B1A3B20B300]=r(B1)+r(B2)+r(B3).

    Proof. The equivalence of (a) and (b) follows from the definition of the Moore–Penrose inverse of a matrix and its uniqueness.

    By Lemma 4.1(c), we first see that the matrix equality A1B(1)1A2B(1)2A3B(1,2)3A4=A holds for all generalized inverses B(1)1 and B(1)2 if and only if any one of the following 6 conditions holds:

    (i) A=0 and A1=0.

    (ii) A=0 and A2=0.

    (iii) A=0 and A3B(1,2)3A4=0.

    (iv) A=0 and r[0A1A2B1]=r(B1).

    (v) A=0 and r[0A2A3B(1,2)3A4B2]=r(B2).

    (vi) r[A0A10A2B1A3B(1,2)3A4B20]=r[A2B1B20]=r(B1)+r(B2).

    We next derive the conditions under which (iii), (v) and (vi) hold, respectively. By (2.10), the equality A3B(1,2)3A4=0 in (iii) of the above 6 conditions holds for all generalized inverses B(1,2)3 if and only if

    A3=0  or  B3=0  or  A4=0  or  r[0A3A4B3]=r(B3) (4.9)

    holds. By (2.8),

    maxB(1,2)3r[0A2A3B(1,2)3A4B2]=maxB(1,2)3r([0A20B2]+[0A3]B(1,2)3[A4,0])=min{r[A2B2]+r(B3),  r[0A2A3B2],  r[0A20B2A40],  r[00A20A3B2A4B30]r(B3)}=min{r[A2B2]+r(B3),  r[0A2A3B2],  r[A2B2]+r(A4),  r[00A20A3B2A4B30]r(B3)}.

    So, the rank equality r[0A2A3B(1,2)3A4B2]=r(B2) in (v) of the above 6 conditions holds for all B(1,2)3 if and only if

    r[A2B2]+r(B3)=r(B2)  or  r[0A2A3B2]=r(B2)  or  r[A2B2]+r(A4)=r(B2) (4.10)

    or

    r[00A20A3B2A4B30]=r(B2)+r(B3), (4.11)

    where the first and third rank equalities in (4.10) mean that B3=0 and A4=0 by noting that r[A2B2]r(B2). By (2.9),

    maxB(1,2)3r[A0A10A2B1A3B(1,2)3A4B20]=maxB(1,2)3r([A0A10A2B10B20]+[00A3]B(1,2)3[A4,0,0])=min{r[A0A10A2B10B20]+r(B3), r[A00A100A2B10A3B20], r[A0A10A2B10B20A400], r[A00A100A2B10A3B20A4B300]r(B3)}=min{r[A0A10A2B10B20]+r(B3), r[A00A100A2B10A3B20], r[A0A10A2B10B20A400], r[A00A100A2B10A3B20A4B300]r(B3)}.

    Hence, the rank equality in (vi) of the above 6 conditions holds for all B(1,2)3 if and only if

    r[A0A10A2B10B20]+r(B3)=r(B1)+r(B2)  or  r[A00A100A2B10A3B20]=r(B1)+r(B2) (4.12)
    or r[A0A10A2B10B20A400]=r(B1)+r(B2)  or  r[A00A100A2B10A3B20A4B300]=r(B1)+r(B2)+r(B3). (4.13)

    Note from (2.6) and (2.7) that the following four groups of matrix rank inequalities

    r[A0A10A2B10B20]+r(B3)r[A0A10A2B10B20]r(A)+r[A2B1B20]r(B1)+r(B2), (4.14)
    r[A00A100A2B10A3B20]r(A)+r[0A2B1A3B20]r[A2B1B20]r(B1)+r(B2), (4.15)
    r[A0A10A2B10B20A400]r[A0A10A2B10B20]r(A)+r[A2B1B20]r(B1)+r(B2), (4.16)
    r[A0A10A2B10B20A400]r[0A2B10B20A400]=r(A4)+r[A2B1B20]r(B1)+r(B2) (4.17)

    hold. Combining (4.14) with the first equality in (4.12) leads to

    A=0,  B3=0  and  r[00A10A2B10B20]=r[A2B1B20]=r(B1)+r(B2). (4.18)

    This condition, however, is included in (4.9). Combining (4.15) with the second equality in (4.12) leads to

    A=0  and  r[00A10A2B1A3B20]=r(B1)+r(B2). (4.19)

    Combining (4.16) and (4.17) with the first equality in (4.13) leads to

    A=0 and A4=0. (4.20)

    Combining the conditions in (i), (ii) and (iv) of the proof with (4.12)–(4.20) leads to the 11 statements in (c).

    The following 6 theorems can be established by a similar approach; therefore, the details of the proofs are omitted.

    Theorem 4.6. Let A1, B1, A2, B2, A3, B3, A4 and A be as given in (4.8). Then, the following three statements are equivalent:

    (a) A1B(1,2)1A2B(1,2)2A3B(1)3A4=A holds for all B(1,2)1, B(1,2)2 and B(1)3.

    (b) A1B(1,2)1A2B(1,2)2A3B(1)3A4 is invariant with respect to all B(1,2)1, B(1,2)2 and B(1)3 and A1B1A2B2A3B3A4=A holds.

    (c) Any one of the following 12 conditions holds:

    (i) A=0 and A1=0.

    (ii) A=0 and B1=0.

    (iii) A=0 and A2=0.

    (iv) A=0 and B2=0.

    (v) A=0 and A3=0.

    (vi) A=0 and A4=0.

    (vii) A=0 and r[0A1A2B1]=r(B1).

    (viii) A=0 and r[0A2A3B2]=r(B2).

    (ix) A=0 and r[0A3A4B3]=r(B3).

    (x) A=0 and r[00A10A2B1A3B20]=r[A2B1B20]=r(B1)+r(B2).

    (xi) A=0 and r[00A20A3B2A4B30]=r[A3B2B30]=r(B2)+r(B3).

    (xii) r[A00A100A2B10A3B20A4B300]=r[0A2B1A3B20B300]=r(B1)+r(B2)+r(B3).

    Theorem 4.7. Let A1, B1, A2, B2, A3, B3, A4 and A be as given in (4.8). Then, the following three statements are equivalent:

    (a) A1B(1,2)1A2B(1,2)2A3B(1,2)3A4=A holds for all B(1,2)1, B(1,2)2 and B(1,2)3.

    (b)A1B(1,2)1A2B(1,2)2A3B(1,2)3A4 is invariant with respect to all B(1,2)1, B(1,2)2 and B(1,2)3 and A1B1A2B2A3B3A4=A holds.

    (c) Any one of the following 13 conditions holds:

    (i) A=0 and A1=0.

    (ii) A=0 and B1=0.

    (iii) A=0 and A2=0.

    (iv) A=0 and B2=0.

    (v) A=0 and A3=0.

    (vi) A=0 and B3=0.

    (vii) A=0 and A4=0.

    (viii) A=0 and r[0A1A2B1]=r(B1).

    (ix) A=0 and r[0A2A3B2]=r(B2).

    (x) A=0 and r[0A3A4B3]=r(B3).

    (xi) A=0 and r[00A10A2B1A3B20]=r[A2B1B20]=r(B1)+r(B2).

    (xii) A=0 and r[00A20A3B2A4B30]=r[A3B2B30]=r(B2)+r(B3).

    (xiii) r[A00A100A2B10A3B20A4B300]=r[0A2B1A3B20B300]=r(B1)+r(B2)+r(B3).

    In addition to the classical definition in (1.1), some other weighted generalized inverses of a matrix can be defined from common solutions of given matrix equations derived from theory and applied points of view, alternatively there are many approaches and discussions concerning different kinds of weighted generalized inverses of matrices in the literature. Here, we mention such an example of weighted generalized inverses: given three matrices ACm×n, PCp×m and QCn×q, A is said to be weighted Moore–Penrose invertible with respect to the two matrices (P,Q) if and only if there exists a matrix XCq×p that satisfies the following four matrix equations

    (1) AQXPA=A,  (2) XPAQX=X,  (3) (PAQX)=PAQX,  (4) (XPAQ)=XPAQ. (5.1)

    In this case, the matrix X is denoted by X=A(,P,Q). In addition, a matrix X is said to be a {i,,j,P,Q}-generalized inverse of A, denoted by A(i,,j,P,Q), if it satisfies the ith, ,jth matrix equations in (5.1), while the set of all {i,,j,P,Q}-generalized inverses of the matrix A is, as usual, denoted by {A(i,,j,P,Q)}.

    Concerning the existence of X satisfying (5.1) and the corresponding expressions of X, we have the following results.

    Lemma 5.1. Let ACm×n, PCp×m and QCn×q. Then, A is weighted Moore–Penrose invertible with respect to the two matrices (P,Q) if and only if the two range equalities R((PA))=R(A) and R(AQ)=R(A) hold, or equivalently, the matrix rank equality r(PAQ)=r(A) holds. In such a case, the matrix equality

    A(i,,j,P,Q)=(PAQ)(i,,j) (5.2)

    always holds. In particular, the general expressions of A(1,P,Q) and A(1,2,P,Q) are given by

    A(1,P,Q)=(PAQ)(1)=(PAQ)+(Iq(AQ)AQ)G+H(IpPA(PA)), (5.3)
    A(1,2,P,Q)=(PAQ)(1,2)=((PAQ)+(Iq(AQ)AQ)G)PAQ((PAQ)+H(IpPA(PA))), (5.4)

    and the following matrix equalities

    AQA(1,P,Q)=AQA(1,2,P,Q)=A(PA)+AQH(IpPA(PA)), (5.5)
    A(1,P,Q)PA=A(1,2,P,Q)PA=(AQ)A+(Iq(AQ)AQ)GPA, (5.6)
    PAQA(1,P,Q)=PAQA(1,2,P,Q)=PA(PA)+PAQH(IpPA(PA)), (5.7)
    A(1,P,Q)PAQ=A(1,2,P,Q)PAQ=(AQ)AQ+(Iq(AQ)AQ)GPAQ (5.8)

    hold, where G,HCq×p are two arbitrary matrices.

    Proof. By Lemma 2.1, the first equation in (5.1) is solvable for the matrix X if and only if R(AQ)R(A) and R((PA))R(A) hold, both of which are equivalent to the matrix range equalities R((PA))=R(A) and R(AQ)=R(A), or equivalently, the matrix rank equality r(PAQ)=r(A). In this case, the first equation in (5.1) is equivalent to PAQXPAQ=PAQ. In this case, we obtain (5.3) and (5.4) from Lemma 2.2.

    Analogous to (4.8), it is possible to formulate four additive matrix equalities that involve weighted {1}- and {1,2}-generalized inverses as follows:

    A1B(1,P1,Q1)1C1+A2B(1,P2,Q2)2C2=D, (5.9)
    A1B(1,2,P1,Q1)1C1+A2B(1,P2,Q2)2C2=D, (5.10)
    A1B(1,P1,Q1)1C1+A2B(1,2,P2,Q2)2C2=D, (5.11)
    A1B(1,2,P1,Q1)1C1+A2B(1,2,P2,Q2)2C2=D, (5.12)

    where A1Cm×q1, P1Cp1×s1, B1Cs1×t1, Q1Ct1×q1, C1Cp1×n, A2Cm×q2, P2Cp2×s2, B2Cs2×t2, Q2Ct2×q2, C2Cp2×n and DCm×n are given.

    Theorem 5.2. Assume that B(1,P1,Q1)1 and B(1,P2,Q2)2 exist. Then, the following three statements are equivalent:

    (a) Equation (5.9) holds for all B(1,P1,Q1)1 and B(1,P2,Q2)2.

    (b) A1B(1,P1,Q1)1C1+A2B(1,P2,Q2)2C2 is invariant with respect to all B(1,P1,Q1)1 and B(1,P2,Q2)2 and A1(P1B1Q1)C1+A2(P2B2Q2)C2=D holds.

    (c) Any one of the following 9 conditions holds:

    (i) A1=0, A2=0 and D=0.

    (ii) A1=0, C2=0 and D=0.

    (iii) C1=0, A2=0 and D=0.

    (iv) C1=0, C2=0 and D=0.

    (v) A1=0 and r[P2B2Q2C2A2D]=r(B2).

    (vi) C1=0 and r[P2B2Q2C2A2D]=r(B2).

    (vii) A2=0 and r[P1B1Q1C1A1D]=r(B1).

    (viii) C2=0 and r[P1B1Q1C1A1D]=r(B1).

    (ix) r[P1B1Q10C10P2B2Q2C2A1A2D]=r(B1)+r(B2).

    Proof. By (5.3), the matrix equality in (5.9) is equivalent to

    A1(P1B1Q1)(1)C1+A2(P2B2Q2)(1)C2=D.

    In this case, applying Theorem 3.1 to the matrix equality yields the results in (a), (b) and (c).

    The following theorems can be shown by a similar approach.

    Theorem 5.3. Assume that B(1,P1,Q1)1 and B(1,P2,Q2)2 exist. Then, the following three statements are equivalent:

    (a) Equation (5.10) holds for all B(1,2,P1,Q1)1 and B(1,P1,Q1)2.

    (b) A1B(1,2,P1,Q1)1C1+A2B(1,P2,Q2)2C2 is invariant with respect to all B(1,2,P1,Q1)1 and B(1,P2,Q2)2 and A1(P1B1Q1)C1+A2(P2B2Q2)C2=D holds.

    (c) Any one of the following 12 conditions holds:

    (i) A1=0, A2=0 and D=0.

    (ii) A1=0, C2=0 and D=0.

    (iii) B1=0, A2=0 and D=0.

    (iv) B1=0, C2=0 and D=0.

    (v) C1=0, A2=0 and D=0.

    (vi) C1=0, C2=0 and D=0.

    (vii) A1=0 and r[P2B2Q2C2A2D]=r(B2).

    (viii) B1=0 and r[P2B2Q2C2A2D]=r(B2).

    (ix) C1=0 and r[P2B2Q2C2A2D]=r(B2).

    (x) A2=0 and r[P1B1Q1C1A1D]=r(B1).

    (xi) C2=0 and r[P1B1Q1C1A1D]=r(B1).

    (xii) r[P1B1Q10C10P2B2Q2C2A1A2D]=r(B1)+r(B2).

    Theorem 5.4. Assume that B(1,P1,Q1)1 and B(1,P2,Q2)2 exist. Then, the following three statements are equivalent:

    (a) Equation (5.12) holds for all B(1,2,P1,Q1)1 and B(1,2,P1,Q1)2.

    (b) A1B(1,2,P1,Q1)1C1+A2B(1,2,P2,Q2)2C2 is invariant with respect to all B(1,2,P1,Q1)1 and B(1,2,P2,Q2)2 and A1(P1B1Q1)C1+A2(P2B2Q2)C2=D holds.

    (b) Any one of the following 16 conditions holds:

    (i) A1=0, A2=0 and D=0.

    (ii) A1=0, B2=0 and D=0.

    (iii) A1=0, C2=0 and D=0.

    (vi) B1=0, A2=0 and D=0.

    (v) B1=0, B2=0 and D=0.

    (vi) B1=0, C2=0 and D=0.

    (vii) C1=0, A2=0 and D=0.

    (viii) C1=0, B2=0 and D=0.

    (ix) C1=0, C2=0 and D=0.

    (x) A1=0 and r[P2B2Q2C2A2D]=r(B2).

    (xi) B1=0 and r[P2B2Q2C2A2D]=r(B2).

    (xii) C1=0 and r[P2B2Q2C2A2D]=r(B2).

    (xiii) A2=0 and r[P1B1Q1C1A1D]=r(B1).

    (xiv) B2=0 and r[P1B1Q1C1A1D]=r(B1).

    (xv) C2=0 and r[P1B1Q1C1A1D]=r(B1).

    (xvi) r[P1B1Q10C10P2B2Q2C2A1A2D]=r(B1)+r(B2).

    Analogous to (1.4), it is possible to formulate some matrix equalities that involve alternative products of matrices and their weighted {1}- and {1,2}-generalized inverses. In what follows, we present a sequence of extensions of the results in the preceding sections to matrix equalities that involve mixed products of two or three weighted {1}- and {1,2}-inverses. In particular, we are able to construct four matrix equalities that involves weighted {1}- and {1,2}-generalized inverses as follows:

    A1B(1,P1,Q1)1A2B(1,P2,Q2)2A3=A, (5.13)
    A1B(1,2,P1,Q1)1A2B(1,P2,Q2)2A3=A, (5.14)
    A1B(1,P1,Q1)1A2B(1,2,P2,Q2)2A3=A, (5.15)
    A1B(1,2,P1,Q1)1A2B(1,2,P2,Q2)2A3=A, (5.16)

    where A1Cm1×m2, P1Cm3×q1, B1Cq1×p1, Q1Cp1×m2, A2Cm3×m4, P2Cm5×q2, B2Cq2×p2, Q2Cp2×m4, A3Cm5×m6 and ACm1×m6.

    Theorem 5.5. Assume that B(1,P1,Q1)1 and B(1,P2,Q2)2 exist. Then, the following three statements are equivalent:

    (a) Equation (5.13) holds for all B(1,P1,Q1)1 and B(1,P2,Q2)2.

    (b) A1B(1,P1,Q1)1A2B(1,P2,Q2)2A3 is invariant with respect to all B(1,P1,Q1)1 and B(1,P2,Q2)2 and A1(P1B1Q1)A2(P2B2Q2)A3=A holds.

    (c) Any one of the following 6 conditions holds:

    (i) A=0 and A1=0.

    (ii) A=0 and A2=0.

    (iii) A=0 and A3=0.

    (iv) A=0 and r[0A1A2P1B1Q1]=r(B1).

    (v) A=0 and r[0A2A3P2B2Q2]=r(B2).

    (vi) r[A0A10A2P1B1Q1A3P2B2Q20]=r[A2P1B1Q1P2B2Q20]=r(B1)+r(B2).

    Proof. Under the given assumptions, we obtain from (5.3) that A1B(1,P1,Q1)1A2B(1,P2,Q2)2A3=A is equivalent to

    A1B(1,P1,Q1)1A2B(1,P2,Q2)2A3=A1(P1B1Q1)(1)A2(P2B2Q2)(1)A3=A.

    In this case, applying Theorem 4.1 to the matrix equality yields the results in (a), (b) and (c).

    By a similar approach, we obtain the following three groups of results from Theorems 4.2 and 4.3.

    Theorem 5.6. Assume that B(1,P1,Q1)1 and B(1,P2,Q2)2 exist. Then, the following three statements are equivalent:

    (a) Equation (5.14) holds for all B(1,2,P2,Q2)1 and B(1,P2,Q2)2.

    (b) A1B(1,2,P2,Q2)1A2B(1,P2,Q2)2A3 is invariant with respect to all B(1,2,P2,Q2)1 and B(1,P2,Q2)2 and A1(P1B1Q1)A2(P2B2Q2)A3=A holds.

    (c) Any one of the following 7 conditions holds:

    (i) A=0 and A1=0.

    (ii) A=0 and B1=0.

    (iii) A=0 and A2=0.

    (iv) A=0 and A3=0.

    (v) A=0 and r[0A1A2P1B1Q1]=r(B1).

    (vi) A=0 and r[0A2A3P2B2Q2]=r(B2).

    (vii) r[A0A10A2P1B1Q1A3P2B2Q20]=r[A2P1B1Q1P2B2Q20]=r(B1)+r(B2).

    Theorem 5.7. Assume that B(1,P1,Q1)1 and B(1,P2,Q2)2 exist. Then, the following three statements are equivalent:

    (a) Equation (5.16) holds for all B(1,2,P1,Q1)1 and B(1,2,P2,Q2)2.

    (b) A1B(1,2,P1,Q1)1A2B(1,2,P2,Q2)2A3 is invariant with respect to all B(1,2,P1,Q1)1 and B(1,2,P2,Q2)2 and A1(P1B1Q1)A2(P2B2Q2)A3=A holds.

    (c) Any one of the following 8 conditions holds:

    (i) A=0 and A1=0.

    (ii) A=0 and B1=0.

    (iii) A=0 and A2=0.

    (iv) A=0 and B2=0.

    (v) A=0 and A3=0.

    (vi) A=0 and r[0A1A2P1B1Q1]=r(B1).

    (vii) A=0 and r[0A2A3P2B2Q2]=r(B2).

    (viii) r[A0A10A2P1B1Q1A3P2B2Q20]=r[A2P1B1Q1P2B2Q20]=r(B1)+r(B2).

    Finally, we present extensions of Theorems 4.4 and 4.7 to matrix equalities that involve mixed products of three weighted {1}- and {1,2}-inverses. Let

    A1Cm1×m2, P1Cm3×q1, B1Cq1×p1, Q1Cp1×m2, A2Cm3×m4, (5.17)
    P2Cm5×q2, B2Cq2×p2, Q2Cp2×m4, A3Cm5×m6, P3Cm7×q3, (5.18)
    B3Cq3×p3, Q3Cp3×m6, A4Cm7×m8. (5.19)

    Then, the matrix product A1B(1,P1,Q1)1A2B(1,P2,Q2)2A3B(1,P3,Q3)3A4 is defined accordingly, and we have the following results.

    Theorem 5.8. Given the matrices in (5.17)–(5.19) and ACm1×m8, and assume that B(1,P1,Q1)1, B(1,P2,Q2)2 and B(1,P3,Q3)3 exist. Then, the following three statements are equivalent:

    (a) A1B(1,P1,Q1)1A2B(1,P2,Q2)2A3B(1,P3,Q3)3A4=A holds for all B(1,P1,Q1)1, B(1,P2,Q2)2 and B(1,P3,Q3)3.

    (b) A1B(1,P1,Q1)1A2B(1,P2,Q2)2A3B(1,P3,Q3)3A4 is invariant with respect to all B(1,P1,Q1)1, B(1,P2,Q2)2 and B(1,P3,Q3)3 and A1(P1B1Q1)A2(P2B2Q2)A3(P3B3Q3)A4=A holds.

    (c) Any one of the following 10 conditions holds:

    (i) A=0 and A1=0.

    (ii) A=0 and A2=0.

    (iii) A=0 and A3=0.

    (iv) A=0 and A4=0.

    (v) A=0 and r[0A1A2P1B1Q1]=r(B1).

    (vi) A=0 and r[0A2A3P2B2Q2]=r(B2).

    (vii) A=0 and r[0A3A4P3B3Q3]=r(B3).

    (viii) A=0 and r[00A10A2P1B1Q1A3P2B2Q20]=r[A2P1B1Q1P2B2Q20]=r(B1)+r(B2).

    (ix) A=0 and r[00A20A3P2B2Q2A4P3B3Q30]=r[A3P2B2Q2P3B3Q30]=r(B2)+r(B3).

    (x) r[A00A100A2P1B1Q10A3P2B2Q20A4P3B3Q300]=r[0A2P1B1Q1A3P2B2Q20P3B3Q300]=r(B1)+r(B2)+r(B3).

    Theorem 5.9. Given the assumptions in (5.17)–(5.19) and ACm1×m8, assume that B(1,P1,Q1)1, B(1,P2,Q2)2 and B(1,P3,Q3)3 exist. Then, the following three statements are equivalent:

    (a) A1B(1,2,P1,Q1)1A2B(1,2,P2,Q2)2A3B(1,2,P3,Q3)3A4=A holds for all B(1,2,P1,Q1)1, B(1,2,P2,Q2)2 and B(1,2,P3,Q3)3.

    (b) A1B(1,2,P1,Q1)1A2B(1,2,P2,Q2)2A3B(1,2,P3,Q3)3A4 is invariant with respect to all B(1,2,P1,Q1)1, B(1,2,P2,Q2)2 and B(1,2,P3,Q3)3 and A1(P1B1Q1)A2(P2B2Q2)A3(P3B3Q3)A4=A holds.

    (c) Any one of the following 13 conditions holds:

    (i) A=0 and A1=0.

    (ii) A=0 and B1=0.

    (iii) A=0 and A2=0.

    (iv) A=0 and B2=0.

    (v) A=0 and A3=0.

    (vi) A=0 and B3=0.

    (vii) A=0 and A4=0.

    (viii) A=0 and r[0A1A2P1B1Q1]=r(B1).

    (ix) A=0 and r[0A2A3P2B2Q2]=r(B2).

    (x) A=0 and r[0A3A4P3B3Q3]=r(B3).

    (xi) A=0 and r[00A10A2P1B1Q1A3P2B2Q20]=r[A2P1B1Q1P2B2Q20]=r(B1)+r(B2).

    (xii) A=0 and r[00A20A3P2B2Q2A4P3B3Q30]=r[A3P2B2Q2P3B3Q30]=r(B2)+r(B3).

    (xiii) r[A00A100A2P1B1Q10A3P2B2Q20A4P3B3Q300]=r[0A2P1B1Q1A3P2B2Q20P3B3Q300]=r(B1)+r(B2)+r(B3).

    There exist many concrete matrix equalities, as described in (1.5) and (1.6), that can be represented in the forms of (1.3) and (1.4). As an application, we consider in this section the following partitioned matrix:

    M=[ABCD],

    where ACm1×n1, BCm1×n2, CCm2×n1 and DCm2×n2. Given such a partitioned matrix, we are able to construct a special matrix equality

    [ABCD](i,,j)=[A(i1,,j1)C(i2,,j2)B(i3,,j3)D(i4,,j4)] (6.1)

    for the generalized inverses. Because generalized inverses of matrices are not necessarily unique, (6.1) can be divided into the following four cases:

    {[ABCD](i,,j)}{[A(i1,,j1)C(i2,,j2)B(i3,,j3)D(i4,,j4)]},{[ABCD](i,,j)}{[A(i1,,j1)C(i2,,j2)B(i3,,j3)D(i4,,j4)]}, (6.2)
    {[ABCD](i,,j)}{[A(i1,,j1)C(i2,,j2)B(i3,,j3)D(i4,,j4)]},{[ABCD](i,,j)}={[A(i1,,j1)C(i2,,j2)B(i3,,j3)D(i4,,j4)]}. (6.3)

    It is a tedious work to characterize these four matrix set relations for different choices of the generalized inverses. In what follows, we only consider a special case of the first matrix set inclusion in (6.3). By definition, the first matrix set inclusion in (6.3) holds for {1}-inverses of the matrices if and only if

    [ABCD][A(1)C(1)B(1)D(1)][ABCD]=[ABCD] (6.4)

    holds for all A(1), B(1), C(1) and D(1), which in turn can be represented in the form

    P1A(1)Q1+P2B(1)Q1+P1C(1)Q2+P2D(1)Q2=M, (6.5)

    where P1=[AC], P2=[BD], Q1=[A,B] and Q2=[C,D]. In this case, we are able to find from Theorem 3.6 the following result.

    Theorem 6.1. Let Pi and Qi be as given in (6.5) and assume that Pi0 and Qi0 for i=1,2. Then, (6.4) holds for all A(1), B(1), C(1) and D(1) if and only if

    r[A000Q10B00Q100C0Q2000DQ2P1P2P1P2M]=r(A)+r(B)+r(C)+r(D),

    or equivalently,

    r[ABA0C0CDAB0B0DCD]=r(A)+r(B)+r(C)+r(D)r(M).

    We investigated the two sorts of general matrix equalities in (1.3) and (1.4) that involve ordinary sum and product operations of matrices and their generalized inverses, and we obtained a variety of necessary and sufficient conditions for them to hold. It can be seen that the corresponding results and facts are fairly easy to understand and utilize within the common discipline of matrix theory and generalized inverses of matrices since they were explicitly derived through various common algebraic calculations of matrices and their generalized inverses, as well as the skillful employments of the elementary but powerful BMRM and MRM.

    In addition to the work related to {1}- and {1,2}-generalized inverses of matrices, it would be of interest to consider other {i,,j}-generalized inverses of the matrices in the two matrix equalities in (1.3) and (1.4); yet, it is not difficult to see that we can easily extend them by induction to general matrix equalities through certain special block matrices that are constructed from the given matrices in the equalities. Finally, we conclude that this study will reveal numerous novel and intrinsic properties and performances of matrix expressions and matrix equalities that are encountered in the theory of generalized inverses. We also believe that these results and facts will provide enough supports to approach various concrete matrix equality problems from theoretical and applied points of view.

    The author declares that he has not used artificial intelligence tools in the creation of this article.

    The author would like to express his sincere thanks to the handling editor and three anonymous reviewers for their helpful comments and suggestions.

    The author declares that there is no conflict of interest.



    [1] A. A. Kilbas, H. M. Srivastava, J. J Trujillo, Theory and applications of fractional differential equations, North-Holland Mathematics Studies, 204, Elsevier, New York-London, 2006.
    [2] Y. C. Kwun, G. Farid, W. Nazeer, et al. Generalized Riemann-Liouville k-fractional integrals associated with Ostrowski type inequalities and error bounds of Hadamard inequalities, IEEE Access, 6 (2018), 64946-64953. doi: 10.1109/ACCESS.2018.2878266
    [3] S. Mubeen, A. Rehman, A note on k-Gamma function and Pochhammer k-symbol, J. Inform. Math. Sci., 6 (2014), 93-107.
    [4] M. Andrić, G. Farid, J. Pečarić, A further extension of Mittag-Leffler function, Fract. Calc. Appl. Anal., 21 (2018), 1377-1395. doi: 10.1515/fca-2018-0072
    [5] S. G. Farid, A unified integral operator and its consequences, Open J. Math. Anal., 4 (2020), 1-7. doi: 10.30538/psrp-oma2020.0047
    [6] H. Chen, U. N. Katugampola, Hermite-Hadamard and Hermite-Hadamard-Fejér type inequalities for generalized fractional integrals, J. Math. Anal. Appl., 446 (2017), 1274-1291. doi: 10.1016/j.jmaa.2016.09.018
    [7] S. S. Dragomir, Inequalities of Jensens type for generalized k-g-fractional integrals of functions for which the composite fg-1 is convex, Fract. Differ. Calc., 1 (2018), 127-150.
    [8] S. Habib, S. Mubeen, M. N. Naeem, Chebyshev type integral inequalities for generalized kfractional conformable integrals, J. Inequal. Spec. Funct., 9 (2018), 53-65.
    [9] F. Jarad, E. Ugurlu, T. Abdeljawad, et al. On a new class of fractional operators, Adv. Differ. Equ., 2017 (2017), 247.
    [10] T. U. Khan, M. A. Khan, Generalized conformable fractional operators, J. Comput. Appl. Math., 346 (2019), 378-389. doi: 10.1016/j.cam.2018.07.018
    [11] S. Mubeen, G. M. Habibullah, k-fractional integrals and applications, Int. J. Contemp. Math., 7 (2012), 89-94.
    [12] T. R. Parbhakar, A singular integral equation with a generalized Mittag-Leffler function in the kernel, Yokohama Math. J., 19 (1971), 7-15.
    [13] G. Rahman, D. Baleanu, M. A. Qurashi, et al. The extended Mittag-Leffler function via fractional calculus, J. Nonlinear Sci. Appl., 10 (2017), 4244-4253. doi: 10.22436/jnsa.010.08.19
    [14] T. O. Salim, A. W. Faraj, A generalization of Mittag-Leffler function and integral operator associated with integral calculus, J. Fract. Calc. Appl., 3 (2012), 1-13. doi: 10.1142/9789814355216_0001
    [15] M. Z. Sarikaya, M. Dahmani, M. E. Kiris, et al. (k, s)-Riemann-Liouville fractional integral and applications, Hacet. J. Math. Stat., 45 (2016), 77-89.
    [16] H. M. Srivastava, Z. Tomovski, Fractional calculus with an integral operator containing generalized Mittag-Leffler function in the kernel, Appl. Math. Comput., 211 (2009), 198-210.
    [17] T. Tunc, H. Budak, F. Usta, et al. On new generalized fractional integral operators and related fractional inequalities, ResearchGate, 2017. Available from: https://www.researchgate.net/publication/313650587.
    [18] Y. C. Kwun, G. Farid, S. Ullah, et al. Inequalities for a unified integral operator and associated results in fractional calculus, IEEE Access, 7 (2019), 126283-126292. doi: 10.1109/ACCESS.2019.2939166
    [19] N. Eftekhari, Some remarks on (s, m)-convexity in the second sense, J. Math. Inequal., 8 (2014), 489-495.
    [20] Y. C. Kwun, G. Farid, S. M. Kang, et al. Derivation of bounds of several kinds ofoperators via (s, m)-convexity, Adv. Differ. Equ., 2020 (2020), 1-14. doi: 10.1186/s13662-019-2438-0
    [21] G. Farid, W. Nazeer, M. S. Saleem, et al. Bounds of Riemann-Liouville fractional integrals in general form via convex functions and their applications, Mathematics, 6 (2018), 248.
    [22] L. Chen, G. Farid, S. I. Butt, et al. Boundedness of fractional integral operators containing MittagLeffler functions, Turkish J. Inequal., 4 (2020), 14-24.
    [23] G. Farid, Estimation of Riemann-Liouville k-fractional integrals via convex functions, Acta et Commentat. Univ. Tartuensis de Math., 23 (2019), 71-78.
    [24] G. Farid, Some Riemann-Liouville fractional integral for inequalities for convex functions, J. Anal., 27 (2019), 1095-1102. doi: 10.1007/s41478-018-0079-4
  • Reader Comments
  • © 2020 the Author(s), licensee AIMS Press. This is an open access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0)
通讯作者: 陈斌, bchen63@163.com
  • 1. 

    沈阳化工大学材料科学与工程学院 沈阳 110142

  1. 本站搜索
  2. 百度学术搜索
  3. 万方数据库搜索
  4. CNKI搜索

Metrics

Article views(4370) PDF downloads(248) Cited by(6)

/

DownLoad:  Full-Size Img  PowerPoint
Return
Return

Catalog