Abstract
In this paper, the DMP and CMP inverses of tensors via Einstein product are defined. Some characterizations, representations, and properties for these generalized inverses are investigated. The perturbation bounds related to the DMP and CMP inverses are also developed.
Similar content being viewed by others
Avoid common mistakes on your manuscript.
1 Introduction
Baksalary and Trenkler (2010) introduced a new pseudoinverse of a matrix named as the core inverse. Malik and Thome (2014) extended this definition and defined a new generalized inverse of a square matrix of an arbitrary index. They used the Drazin inverse (D) and the Moore–Penrose (MP) inverse, and therefore, this new generalized inverse is called the DMP inverse. The DMP inverse is analyzed from both algebraic as well as geometrical approaches establishing the equivalence between them. DMP inverse extends the notion of core inverse. Recently, Mehdipour and Salemi (2018) introduced another new inverse of a square matrix A, named after CMP inverse. The Drazin inverse, Moore–Penrose inverse, the weighted Moore–Penrose inverse, core and core-EP inverse, and outer inverse via Einstein product can be found in Behera and Mishra (2017), Behera et al. (2019), Ji and Wei (2017), Ji and Wei (2018), Jin et al. (2017), Liang and Zheng (2019), Liang et al. (2019), Ma (2018), Ma et al. (2019), Sahoo et al. (2019), Stanimirović et al. (2018), and Sun et al. (2016). DMP inverse and CMP inverse via Einstein product of tensors provide a new class of generalized inverses of tensors. Recently, Miao et al. (2019a, b) investigated the tensor functions via the tensor singular value decomposition and tensor Jordan canonical form based on the T-product for the tensor Moore–Penrose inverse and the tensor Drazin inverse, respectively. The monographs on the theory and computation of tensors and generalized inverses can be found in Ding and Wei (2016), Wei et al. (2018).
For a positive integer N, let \({\mathbf {I}}_1,\ldots ,{\mathbf {I}}_N\) be positive integers. An order N tensor \({\mathcal {A}}=({\mathcal {A}}_{i_{1},i_{2},\ldots ,i_{{N}}})_{1\le i_{j}\le {\mathbf {I}}_{j}}\), \((j = 1,\ldots , {N})\) is a multidimensional array with \({\mathfrak {I}}={\mathbf {I}}_{1}{\mathbf {I}}_{2}\cdots {\mathbf {I}}_{N}\) entries, where \({\mathbf {I}}_1,\ldots ,{\mathbf {I}}_N\) are positive integers. Let \({\mathbb {C}}^{{\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{N}}\) (resp. \({\mathbb {R}}^{{\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{N}}\)) be the set of the order N tensors of dimension \({\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{N}\) over complex numbers \({\mathbb {C}}\) (resp. real numbers \({\mathbb {R}}\)).
For a tensor \({\mathcal {A}} = ({\mathcal {A}}_{i_{1},\ldots ,i_{N},i_{1},\ldots ,i_N}) \in {\mathbb {C}}^{{\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{N} \times {\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{N}}\), if there exists a tensor \({\mathcal {X}}\), such that \({\mathcal {A}} *_{N} {\mathcal {X}} = {\mathcal {X}} *_{N} {\mathcal {A}} = {\mathcal {I}}\), then tensor \({\mathcal {A}}\) is invertible. In this case, \({\mathcal {X}}\) is called the inverse of \({\mathcal {A}}\) and denoted by \({\mathcal {A}}^{-1}\). For a tensor \({\mathcal {A}} = ({\mathcal {A}}_{i_{1},\ldots ,i_{M},j_{1},\ldots ,j_N}) \in {\mathbb {C}}^{{\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{M} \times {\mathbf {J}}_{1} \times \cdots \times {\mathbf {J}}_{N}}\), the tensor \({\mathcal {A}}^{\mathrm {T}}=({\mathcal {A}})_{i_{1},\ldots ,i_{M},j_{1},\ldots ,j_N} \in {\mathbb {C}}^{{\mathbf {J}}_{1} \times \cdots \times {\mathbf {J}}_{N} \times {\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{M}}\) is the transpose of \({\mathcal {A}}\). The conjugate transpose of a tensor \({\mathcal {A}}\) is denoted by \({\mathcal {A}}^{*}\) and elementwise defined as \(({\mathcal {A}}^{*})_{j_{1},\ldots ,j_{N},i_{1},\ldots ,i_M}=(\overline{{\mathcal {A}}})_{i_{1},\ldots ,i_{M},j_{1},\ldots ,j_N} \in {\mathbb {C}}^{{\mathbf {J}}_{1} \times \cdots \times {\mathbf {J}}_{N} \times {\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{M}}\) where the overline means the conjugate operator.
The Einstein product of tensors is defined in Einstein (2007) by the operation \(*_{N}\) via:
where \({\mathcal {A}} \in {\mathbb {C}}^{{\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{{N}}\times {\mathbf {K}}_{1} \times \cdots \times {\mathbf {K}}_{{N}}}\), \({\mathcal {B}} \in {\mathbb {C}}^{{\mathbf {K}}_{1}\times \cdots \times {\mathbf {K}}_{{N}}\times {\mathbf {J}}_{1}\times \cdots \times {\mathbf {J}}_{{M}}}\) and \({\mathcal {A}}*_{N}{\mathcal {B}} \in {\mathbb {C}}^{{\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{{N}} \times {\mathbf {J}}_{1} \times \cdots \times {\mathbf {J}}_{{M}}}\).
The associative law of this tensor product holds. In the above formula, when \({\mathcal {B}} \in {\mathbb {C}}^{{\mathbf {K}}_{1} \times \cdots \times {\mathbf {K}}_{{N}}}\), then
where \({\mathcal {A}}*_{N}{\mathcal {B}} \in {\mathbb {C}}^{{\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{{N}}}\).
Definition 1.1
(Sun et al. 2016) Let \({\mathcal {A}} \in {\mathbb {C}}^{{\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{{N}} \times {\mathbf {K}}_{1} \times \cdots \times {\mathbf {K}}_{{N}}}\). The tensor \({\mathcal {X}} \in {\mathbb {C}}^{{\mathbf {K}}_{1} \times \cdots \times {\mathbf {K}}_{{N}} \times {\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{{N}}}\) which satisfies:
is called the Moore–Penrose inverse of \({\mathcal {A}}\), abbreviated by MP inverse, denoted by \({\mathcal {A}}^\dagger \). If the equation (i) of the above Eqs. (1)–(4) holds, \({\mathcal {X}}\) is called an \((i)-\)inverse of \({\mathcal {A}}\), denoted by \({\mathcal {A}}^{(i)}\).
Definition 1.2
(Ji and Wei 2017) For \({\mathcal {A}} \in {\mathbb {C}}^{{\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{{N}} \times {\mathbf {K}}_{1} \times \cdots \times {\mathbf {K}}_{{N}}}\), the range \({\mathcal {R}}({\mathcal {A}})\) and the null space \({\mathcal {N}}({\mathcal {A}})\) of \({\mathcal {A}}\) are defined by:
where \({\mathcal {O}}\) is an appropriate zero tensor.
Definition 1.3
(Ji and Wei 2018, Lemma 2.1) Let \({\mathcal {X}} \in {\mathbb {C}}^{{\mathbf {N}}_{1} \times \cdots \times {\mathbf {N}}_{{K}}}\). The spectral norm \(\Vert {\mathcal {X}} \Vert _{2}\) is defined as:
where \(\lambda _{\max } ({\mathcal {X}}^* *_N {\mathcal {X}})\) denotes the largest eigenvalue of \({\mathcal {X}}^* *_N {\mathcal {X}}\).
Lemma 1.1
(Ma et al. 2019) Let \({\mathcal {E}} \in {\mathbb {C}}^{{\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{{N}} \times {\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{{N}}}\). Suppose that \(\Vert {\mathcal {E}}\Vert _{2} < 1\). Then, \({\mathcal {I}}+{\mathcal {E}}\) is nonsingular and
Lemma 1.2
(Ma et al. 2019) Let \({\mathcal {E}}\in {\mathbb {C}}^{{\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{{K}} \times {\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{{K}}}\). If \(\Vert {\mathcal {E}}\Vert _{2}<1\), then
and
Definition 1.4
(Behera et al. 2019; Ji and Wei 2018) Assume that \({\mathcal {A}} \in {\mathbb {C}}^{{\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{{N}} \times {\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{{N}}}\). Define
It is easy to see that
and
The smallest non-negative integer p, such that \({\mathcal {R}}({\mathcal {A}}^{p+1}) = {\mathcal {R}}({\mathcal {A}}^{p})\) (or \({\mathcal {N}}({\mathcal {A}}^{p+1}) = {\mathcal {N}}({\mathcal {A}}^{p})\)), denoted by Ind\(({\mathcal {A}})\), is called the index of \({\mathcal {A}}\).
Definition 1.5
(Behera et al. 2019; Ji and Wei 2018) Let \({\mathcal {A}} \in {\mathbb {C}}^{{\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{{N}} \times {\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{{N}}}\). The tensor \({\mathcal {X}} \in {\mathbb {C}}^{{\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{{N}} \times {\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{{N}}}\) which satisfies:
is called the Drazin inverse of \({\mathcal {A}}\), denoted by \({\mathcal {A}}^{d}\). Especially, if Ind\(({\mathcal {A}})=1\), \({\mathcal {X}}\) is called the group inverse of \({\mathcal {A}}\), denoted by \({\mathcal {A}}_{g}\).
For a tensor \({\mathcal {A}} \in {\mathbb {C}}^{{\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{{N}} \times {\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{{N}}}\), the singular value decomposition (SVD) Brazell et al. (2013), Sun et al. (2016) of \({\mathcal {A}}\) has the form:
where \({\mathcal {U}} \in {\mathbb {C}}^{{\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{{N}} \times {\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{{N}}}\) and \({\mathcal {V}} \in {\mathbb {C}}^{{\mathbf {K}}_{1} \times \cdots \times {\mathbf {K}}_{{N}} \times {\mathbf {K}}_{1} \times \cdots \times {\mathbf {K}}_{{N}}}\) are unitary tensors, and the tensor \({\mathcal {D}} \in {\mathbb {C}}^{{\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{{N}} \times {\mathbf {K}}_{1} \times \cdots \times {\mathbf {K}}_{{N}}}\) is a diagonal tensor satisfying:
where \(\mu _{i_{1} \ldots i_{{N}}}\) are the singular values of \({\mathcal {A}}\).
The diagonal tensor \({\mathcal {D}}\) can be written as: \({\mathcal {D}}= \begin{pmatrix} \Sigma &{}{\mathcal {O}} \\ {\mathcal {O}}&{}{\mathcal {O}} \end{pmatrix}\), where \(\Sigma \in {\mathbb {C}}^{{\mathbf {R}}_{1} \times \cdots \times {\mathbf {R}}_{{N}} \times {\mathbf {R}}_{1} \times \cdots \times {\mathbf {R}}_{{N}}}\) is a diagonal tensor of singular values of \({\mathcal {A}}\). Then, the singular value decomposition of \({\mathcal {A}}\) can be written as follows (Brazell et al. 2013):
Multiplying (1.5) by \({\mathcal {U}}*_{{N}}{\mathcal {U}}^{*}(={\mathcal {I}})\) on the right-hand side, and assuming that unitary \({\mathcal {V}}^{*}*_{{N}}{\mathcal {U}}\) is partitioned according to:
where \({\mathcal {K}}\in {\mathbb {C}}^{{\mathbf {R}}_{1} \times \cdots \times {\mathbf {R}}_{{N}} \times {\mathbf {R}}_{1} \times \cdots \times {\mathbf {R}}_{{N}}}\).
Hartwig and Spindelböck decomposition (Hartwig and Spindelböck 1983) of tensor arrived at the following result.
Lemma 1.3
Let \({\mathcal {A}} \in {\mathbb {C}}^{{\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{{N}} \times {\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{{N}}}\). Then, there exist unitary \({\mathcal {U}}\in {\mathbb {C}}^{{\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{{N}} \times {\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{{N}}}\), such that
where \(\Sigma \in {\mathbb {C}}^{{\mathbf {R}}_{1} \times \cdots \times {\mathbf {R}}_{{N}} \times {\mathbf {R}}_{1} \times \cdots \times {\mathbf {R}}_{{N}}}\) is a diagonal tensor of singular values of \({\mathcal {A}}\), and the tensors \({\mathcal {K}}\in {\mathbb {C}}^{{\mathbf {R}}_{1} \times \cdots \times {\mathbf {R}}_{{N}} \times {\mathbf {R}}_{1} \times \cdots \times {\mathbf {R}}_{{N}}}\), \({\mathcal {L}}\in {\mathbb {C}}^{{\mathbf {R}}_{1} \times \cdots \times {\mathbf {R}}_{{N}} \times ({\mathbf {I}}_{1}-{\mathbf {R}}_{1}) \times \cdots \times ({\mathbf {I}}_{N}-{\mathbf {R}}_{N})}\) satisfy:
From (1.6), the Drazin inverse and the Moore–Penrose inverse of \({\mathcal {A}}\) are presented as follows:
Theorem 1.1
(Behera et al. 2019) Let \({\mathcal {A}} \in {\mathbb {C}}^{{\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{{N}} \times {\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{{N}}}\). Then, \({\mathcal {A}}\) can be written as the sum of two tensors \({\mathcal {C}}_{{\mathcal {A}}}\) and \({\mathcal {N}}_{{\mathcal {A}}}\), i.e., \({\mathcal {A}}={\mathcal {C}}_{{\mathcal {A}}}+{\mathcal {N}}_{{\mathcal {A}}}\), where Ind\(({\mathcal {C}}_{{\mathcal {A}}})\le 1\), \({\mathcal {N}}_{{\mathcal {A}}}\) is nilpotent and \({\mathcal {C}}_{{\mathcal {A}}} *_{N}{\mathcal {N}}_{{\mathcal {A}}}= {\mathcal {N}}_{{\mathcal {A}}}*_{N}{\mathcal {C}}_{{\mathcal {A}}}={\mathcal {O}}\).
The tensors \({\mathcal {C}}_{{\mathcal {A}}}\) and \({\mathcal {N}}_{{\mathcal {A}}}\) called the core and nilpotent parts of the tensor \({\mathcal {A}}\), respectively. We know that if Ind\(({\mathcal {A}})\le 1\), then \({\mathcal {A}}={\mathcal {C}}_{{\mathcal {A}}}\). Also, it is valid that \({\mathcal {C}}_{{\mathcal {A}}}= {\mathcal {A}}*_{N}{\mathcal {A}}^{d}*_{N}{\mathcal {A}}\).
Theorem 1.2
(Schur decomposition) (Liang et al. 2019) Let \({\mathcal {A}} \in {\mathbb {C}}^{{\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{{N}} \times {\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{{N}}}\) be a tensor of index k. Then, it can be factorized as the Schur form of \({\mathcal {A}}\):
where \({\mathcal {U}} \in {\mathbb {C}}^{{\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{{N}} \times {\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{{N}}}\) is unitary, \({\mathcal {T}}_{11}\) is a nonsingular upper triangular tensor, and \({\mathcal {T}}_{22}\) is a nilpotent tensor with index k.
2 Preliminary results
In this section, we define the DMP and CMP inverses of tensor via Einstein product. Furthermore, we collect a few properties of DMP and CMP inverses.
Let \({\mathcal {A}} \in {\mathbb {C}}^{{\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{{N}} \times {\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{{N}}}\) have index k and consider the system of equations:
Theorem 2.1
If system (2.1) has a solution, then it is unique.
Proof
Assume that \({\mathcal {X}}_{1}\) and \({\mathcal {X}}_{2}\) satisfy (2.1). Then, using that \({\mathcal {A}}*_{N}{\mathcal {A}}^{d}={\mathcal {A}}^{d}*_{N}{\mathcal {A}}\), we get
\(\square \)
Theorem 2.2
The system of (2.1) is consistent and has a unique solution: \({\mathcal {X}}= {\mathcal {A}}^{d} *_{N} {\mathcal {A}}*_{N}{\mathcal {A}}^{\dagger }\).
Proof
It is easy to see that \({\mathcal {A}}^{d}*_{N}{\mathcal {A}}*_{N}{\mathcal {A}}^{\dagger }\) satisfies the three equations in system (2.1). Now, Theorem 2.1 gives the uniqueness. \(\square \)
Thus, for a given tensor \({\mathcal {A}} \in {\mathbb {C}}^{{\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{{N}} \times {\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{{N}}}\), the tensor \({\mathcal {A}}^{d}*_{N}{\mathcal {A}}*_{N}{\mathcal {A}}^{\dagger }\) is the unique tensor satisfying system of (2.1).
Definition 2.1
Let \({\mathcal {A}} \in {\mathbb {C}}^{{\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{{N}} \times {\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{{N}}}\) be a tensor of index k. The DMP inverse of \({\mathcal {A}}\), denoted by \({\mathcal {A}}^{d,\dagger }\), is defined to be the tensor:
DMP inverse has the following several important properties (Malik and Thome 2014).
From (1.8), the DMP inverse of \({\mathcal {A}}\) is given by:
Theorem 2.3
Let \({\mathcal {A}} \in {\mathbb {C}}^{{\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{{N}} \times {\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{{N}}}\) be of the form (1.6). Then:
Lemma 2.1
Let \({\mathcal {A}} \in {\mathbb {C}}^{{\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{{N}} \times {\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{{N}}}\) be a tensor of index k has the form (1.6). Then, Ind\((\Sigma *_{N} {\mathcal {K}}) = k-1\).
Proof
Since
with \({\mathcal {Y}}\), \({\mathcal {Z}}\) of adequate sizes, such that the tensor \(\begin{pmatrix} \Sigma *_{N} {\mathcal {K}}&{}\Sigma *_{N} {\mathcal {L}}\\ {\mathcal {Y}}&{} {\mathcal {Z}}\end{pmatrix}\) is nonsingular, we have \({\mathcal {N}}({\mathcal {A}}^{k}) = {\mathcal {N}}((\Sigma *_{N} {\mathcal {K}})^{k-1})\). And then, \({\mathcal {N}}({\mathcal {A}}^{k+1}) = {\mathcal {N}}((\Sigma *_{N} {\mathcal {K}})^{k})\). Since Ind\(({\mathcal {A}}) = k\), we can obtain that \(k-1\) is the smallest non-negative integer satisfying \({\mathcal {N}}((\Sigma *_{N} {\mathcal {K}})^{k-1}) = {\mathcal {N}}((\Sigma *_{N} {\mathcal {K}})^{k})\), that is Ind\((\Sigma *_{N} {\mathcal {K}}) = k-1\). \(\square \)
Theorem 2.4
The DMP inverse \({\mathcal {X}} = {\mathcal {A}}^{d,\dagger }\) of a tensor \({\mathcal {A}} \in {\mathbb {C}}^{{\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{{N}} \times {\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{{N}}}\) satisfies the equations:
where \({}^{d,\dagger }{\mathcal {C}}_{{\mathcal {A}}} = {\mathcal {A}} *_{N} {\mathcal {A}}^{d,\dagger } *_{N} {\mathcal {A}}\) denotes the DMP core part of \({\mathcal {A}}\).
Proof
(1) Using the Hartwig–Spindelbock decomposition of tensor \({\mathcal {A}}\) (Lemma 1.3), from \({\mathcal {C}}_{{\mathcal {A}}} = {\mathcal {A}} *_{N} {\mathcal {A}}^{d} *_{N} {\mathcal {A}}\) and \({}^{d,\dagger }{\mathcal {C}}_{{\mathcal {A}}} = {\mathcal {A}} *_{N} {\mathcal {A}}^{d,\dagger } *_{N} {\mathcal {A}}\), we have:
and
Thus, \({\mathcal {C}}_{{\mathcal {A}}} = {}^{d,\dagger }{\mathcal {C}}_{{\mathcal {A}}}\). The core part of \({\mathcal {A}}\) is its DMP core part. \({\mathcal {A}}^{d,\dagger }\) is a solution of \({\mathcal {A}} *_{N} {\mathcal {X}} *_{N} {\mathcal {A}} = {\mathcal {C}}_{{\mathcal {A}}}\).
(2) From \({}^{d,\dagger }{\mathcal {C}}_{{\mathcal {A}}} *_{N} {\mathcal {A}}^{\dagger }\), we have
The proof is completed. \(\square \)
Theorem 2.5
If \({\mathcal {A}} \in {\mathbb {C}}^{{\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{{N}} \times {\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{{N}}}\) has index k, then the following statements hold:
- (1)
\({\mathcal {A}} *_{N} {\mathcal {A}}^{d,\dagger }\) is a projector onto \({\mathcal {R}}({}^{d,\dag }{\mathcal {C}}_{{\mathcal {A}}})\) along \({\mathcal {N}}({\mathcal {A}}^{d} *_{N} {\mathcal {A}}^{\dagger })\);
- (2)
\({\mathcal {A}}^{d,\dagger } *_{N} {\mathcal {A}} = {\mathcal {A}}^{d} *_{N} {\mathcal {A}}\) is a projector onto \({\mathcal {R}}({\mathcal {A}}^{k})\) along \({\mathcal {N}}({\mathcal {A}}^{k})\).
Proof
(1) From (2.1), \({\mathcal {A}} *_{N} {\mathcal {A}}^{d,\dagger }\) is a projection. It is obvious that
and
(2) Since \({\mathcal {A}}^{d,\dagger } *_{N} {\mathcal {A}} = {\mathcal {A}}^{d} *_{N} {\mathcal {A}}\) and \({\mathcal {A}}^{d} *_{N} {\mathcal {A}}\) is a projection of \({\mathcal {A}}\), we have:
and
\(\square \)
Theorem 2.6
If \({\mathcal {A}} \in {\mathbb {C}}^{{\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{{N}} \times {\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{{N}}}\) has index k, then \({\mathcal {A}}^{d,\dagger }\) is the unique tensor \({\mathcal {X}}\) that satisfies:
Proof
We know that \({\mathcal {A}} *_{N} {\mathcal {A}}^{d,\dagger }\) is idempotent from Theorem 2.5. Moreover:
Assume that \({\mathcal {X}}_{1}\), \({\mathcal {X}}_{2}\) satisfy (2.5). Then, \({\mathcal {A}} *_{N} {\mathcal {X}}_{1} = {\mathcal {A}} *_{N} {\mathcal {X}}_{2} = {\mathcal {P}}_{{\mathcal {R}} ({}^{d}{\mathcal {C}}_{{\mathcal {A}}}), {\mathcal {N}}({\mathcal {A}}^{d} *_{N} {\mathcal {A}}^{\dagger })}\), \({\mathcal {R}}({\mathcal {X}}_{1}) \subseteq {\mathcal {R}}({\mathcal {A}}^{k})\) and \({\mathcal {R}}({\mathcal {X}}_{2}) \subseteq {\mathcal {R}}({\mathcal {A}}^{k})\). Since \({\mathcal {A}}({\mathcal {X}}_{1} - {\mathcal {X}}_{2}) = 0\), we get \({\mathcal {R}}({\mathcal {X}}_{1} - {\mathcal {X}}_{2}) \subseteq {\mathcal {N}}({\mathcal {A}})\). From \({\mathcal {R}}({\mathcal {X}}_{1}) \subseteq {\mathcal {R}}({\mathcal {A}}^{k})\) and \({\mathcal {R}}({\mathcal {X}}_{2}) \subseteq {\mathcal {R}}({\mathcal {A}}^{k})\), we get \({\mathcal {R}}({\mathcal {X}}_{1} - {\mathcal {X}}_{2}) \subseteq {\mathcal {R}}({\mathcal {A}}^{k})\); that is \({\mathcal {R}}({\mathcal {X}}_{1} - {\mathcal {X}}_{2}) \subseteq {\mathcal {N}}({\mathcal {A}}^{k}) \cap {\mathcal {R}}({\mathcal {A}}^{k}) = \{0\}\), since \({\mathcal {A}}\) has index k. Thus, there is only one \({\mathcal {X}}\) satisfying conditions. \(\square \)
Proposition 2.1
Let \({\mathcal {A}} \in {\mathbb {C}}^{{\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{{N}} \times {\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{{N}}}\) be a tensor of index k. Then:
- (a)
\({\mathcal {A}}^{d,\dagger } = {\mathcal {A}}^{d} *_{N} {\mathcal {A}} *_{N} {\mathcal {A}}^{\dagger }\).
- (b)
\({\mathcal {A}}^{d,\dagger }\) is an outer inverse of \({\mathcal {A}}\).
- (c)
\(({\mathcal {A}}^{d,\dagger })^{n}= {\left\{ \begin{array}{ll} ({\mathcal {A}}^{d} *_{N} {\mathcal {A}}^{\dagger })^{\frac{n}{2}}, &{} if\ n\ is\ even, \\ {\mathcal {A}}*_{N}({\mathcal {A}}^{d} *_{N} {\mathcal {A}}^{\dagger })^{\frac{n+1}{2}}, &{} if\ n\ is\ odd. \end{array}\right. }\)
- (d)
\(({\mathcal {A}}^{d,\dagger })^{\dagger } = (({\mathcal {A}} *_{N} {\mathcal {A}} *_{N} {\mathcal {A}}^{\dagger })^{\dagger }\).
- (e)
\((({\mathcal {A}}^{d,\dagger })^{d})^{d} = {\mathcal {A}}^{d,\dagger }\).
- (f)
\({\mathcal {A}}^{d,\dagger } = {\mathcal {O}}\) if and only if \({\mathcal {A}}\) is nilpotent or \({\mathcal {A}} = {\mathcal {O}}\).
Proof
(a) and (b) We can obtain (a) and (b) from definition and properties of the Moore–Penrose and Drazin inverses.
(c) We calculate \(({\mathcal {A}}^{d,\dagger })^{1} = {\mathcal {A}}^{d} *_{N} {\mathcal {A}} *_{N} {\mathcal {A}}^{\dagger } = {\mathcal {A}} *_{N} {\mathcal {A}}^{d} *_{N} {\mathcal {A}}^{\dagger }\) and \(({\mathcal {A}}^{d,\dagger })^{2} = {\mathcal {A}}^{d} *_{N} {\mathcal {A}} *_{N} {\mathcal {A}}^{\dagger } *_{N} {\mathcal {A}}^{d} *_{N} {\mathcal {A}} *_{N} {\mathcal {A}}^{\dagger } = {\mathcal {A}}^{d} *_{N} ({\mathcal {A}} *_{N} {\mathcal {A}}^{\dagger } *_{N} {\mathcal {A}}) *_{N} {\mathcal {A}}^{d} *_{N} {\mathcal {A}}^{\dagger } = {\mathcal {A}}^{d} *_{N} {\mathcal {A}} *_{N} {\mathcal {A}}^{d} *_{N} {\mathcal {A}}^{\dagger } = {\mathcal {A}}^{d} *_{N} {\mathcal {A}}^{\dagger }\). Then, we can obtain the formula.
(d) and (e) We can proof them through (2.4).
(f) Suppose that \({\mathcal {A}} \ne {\mathcal {O}}\) and \({\mathcal {A}}^{d,\dagger } = {\mathcal {O}}\). We can obtain two cases if \({\mathcal {A}}\) has the form (1.6) and \({\mathcal {A}}^{d,\dagger }\) has the form (2.4).
(i) \(\Sigma *_{N}{\mathcal {K}} \ne {\mathcal {O}}\). In this case, according to \({\mathcal {A}}^{d,\dagger } = {\mathcal {O}}\), we obtain \((\Sigma *_{N}{\mathcal {K}})^{d} = {\mathcal {O}}\). Therefore, \(\Sigma *_{N}{\mathcal {K}}\) is nilpotent. Hence, \({\mathcal {A}}\) must be nilpotent.
(ii) \(\Sigma *_{N}{\mathcal {K}} = {\mathcal {O}}\). In this case, the tensor \({\mathcal {A}}={\mathcal {U}}*_{N} \begin{pmatrix} {\mathcal {O}}&{}\Sigma *_{N} {\mathcal {L}}\\ {\mathcal {O}}&{} {\mathcal {O}} \end{pmatrix}*_{N}{\mathcal {U}}^{*}\) is clearly nilpotent. The converse is evident because in both \({\mathcal {A}} = {\mathcal {O}}\) and \({\mathcal {A}}\) nilpotent cases its Drazin inverse is the null tensor. \(\square \)
By \({\mathcal {C}}_{{\mathcal {A}}}= {\mathcal {A}}*_{{N}} {\mathcal {A}}^{d}*_{{N}}{\mathcal {A}}\), we obtain the following:
Theorem 2.7
Let \({\mathcal {A}} \in {\mathbb {C}}^{{\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{{N}} \times {\mathbf {K}}_{1} \times \cdots \times {\mathbf {K}}_{{N}}}\). The tensor \({\mathcal {X}}\in {\mathbb {C}}^{{\mathbf {K}}_{1} \times \cdots \times {\mathbf {K}}_{{N}} \times {\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{{N}}}\) is the unique tensor that satisfies the following system of equations:
Proof
Easy computation shows that the tensor \({\mathcal {A}}^{c,\dagger }= {\mathcal {A}}^{\dagger }*_{N}{\mathcal {C}}_{{\mathcal {A}}}*_{N} {\mathcal {A}}^{\dagger }\) is a solution of this system.
Let \({\mathcal {X}}_{1}\) and \({\mathcal {X}}_{2}\) be two tensors satisfying (2.7). Then:
Thus:
The result holds. \(\square \)
CMP inverse has several important properties (Mehdipour and Salemi 2018).
Proposition 2.2
Let \({\mathcal {A}}\in {\mathbb {C}}^{{\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{{N}} \times {\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{{N}}}\) with core-nilpotent decomposition \({\mathcal {A}}={\mathcal {C}}_{{\mathcal {A}}}+{\mathcal {N}}_{{\mathcal {A}}}\). Then, the following holds:
(1) \({\mathcal {A}}^{c,\dagger }={\mathcal {Q}}_{{\mathcal {A}}} *_{N}{\mathcal {A}}^{d}*_{N}{\mathcal {P}}_{{\mathcal {A}}}\), where \({\mathcal {P}}_{{\mathcal {A}}}={\mathcal {A}}*_{N} {\mathcal {A}}^{\dagger }\) and \({\mathcal {Q}}_{{\mathcal {A}}}={\mathcal {A}}^{\dagger } *_{N}{\mathcal {A}}\) are orthogonal projections onto \({\mathcal {R}}({\mathcal {A}})\) and \({\mathcal {R}}({\mathcal {A}}^{*})\), respectively;
(2) \({\mathcal {A}}^{c,\dagger }*_{N}{\mathcal {C}}_{{\mathcal {A}}} *_{N}{\mathcal {A}}^{c,\dagger } ={\mathcal {A}}^{c,\dagger }\) and \({\mathcal {C}}_{{\mathcal {A}}}*_{N}{\mathcal {A}}^{c,\dagger } *_{N}{\mathcal {C}}_{{\mathcal {A}}} ={\mathcal {C}}_{{\mathcal {A}}}\);
(3) \({\mathcal {C}}_{{\mathcal {A}}}*_{N}{\mathcal {A}}^{c,\dagger }={\mathcal {A}}*_{N} {\mathcal {A}}^{c,\dagger }\) and \({\mathcal {A}}^{c,\dagger }*_{N}{\mathcal {C}}_{{\mathcal {A}}} ={\mathcal {A}}^{c,\dagger }*_{N}{\mathcal {A}}\).
Proof
(1) This part follows from \({\mathcal {A}}^{c,\dagger }={\mathcal {A}}^{\dagger }*_{N} {\mathcal {C}}_{{\mathcal {A}}}*_{N}{\mathcal {A}}^{\dagger } ={\mathcal {A}}^{\dagger }*_{N}{\mathcal {A}}*_{N} {\mathcal {A}}^{d}*_{N}{\mathcal {A}}*_{N}{\mathcal {A}}^{\dagger }= {\mathcal {Q}}_{{\mathcal {A}}}*_{N}{\mathcal {A}}^{d}*_{N} {\mathcal {P}}_{{\mathcal {A}}}\).
(2) The proof follows from the representations give in (1) and \({\mathcal {C}}_{{\mathcal {A}}}={\mathcal {A}}*_{N}{\mathcal {A}}^{d}*_{N}{\mathcal {A}}\):
Since
we obtain
Also, the same method as above shows that \({\mathcal {C}}_{{\mathcal {A}}}*_{N} {\mathcal {A}}^{c,\dagger } *_{N}{\mathcal {C}}_{{\mathcal {A}}} = {\mathcal {C}}_{{\mathcal {A}}}\). Hence, the statement is proved.
(3) This part can also be demonstrated by combining \({\mathcal {C}}_{{\mathcal {A}}}= {\mathcal {A}}*_{N}{\mathcal {A}}^{d}*_{N}{\mathcal {A}}\) and \({\mathcal {A}}^{c,\dagger }= {\mathcal {Q}}_{{\mathcal {A}}}*_{N}{\mathcal {A}}^{d}*_{N} {\mathcal {P}}_{{\mathcal {A}}}\).
Similarly:
\(\square \)
Theorem 2.8
Suppose that \({\mathcal {A}}\in {\mathbb {C}}^{{\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{{N}} \times {\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{{N}}}\) is a tensor with Ind\(({\mathcal {A}})=k\). Then
- (1)
\({\mathcal {A}}^{k+1}*_{N}{\mathcal {A}}^{c,\dagger }= {\mathcal {A}}^{k+1}*_{N}{\mathcal {A}}^{\dagger }\);
- (2)
\({\mathcal {A}}^{c,\dagger }*_{N}{\mathcal {A}}^{k}= {\mathcal {A}}^{\dagger }*_{N}{\mathcal {A}}^{k}\).
Proof
(1) Since \({\mathcal {A}}^{m}*_{N}{\mathcal {Q}}_{{\mathcal {A}}}={\mathcal {A}}^{m}\) for every positive integer m. By Proposition 2.2:
(2) As \({\mathcal {P}}_{{\mathcal {A}}} *_{N}{\mathcal {A}}^{m}= {\mathcal {A}}^{m}\) for every positive integer m. By Proposition 2.2:
\(\square \)
Theorem 2.9
Suppose that \({\mathcal {A}}\in {\mathbb {C}}^{{\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{{N}} \times {\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{{N}}}\). Then, \({\mathcal {A}}^{c,\dagger }={\mathcal {A}}^{\dagger }\) if and only if Ind\(({\mathcal {A}})\le 1\).
Proof
If \({\mathcal {A}}^{c,\dagger }={\mathcal {A}}^{\dagger }\). By (1.8) and (2.6), we get \({\mathcal {K}}^{*}*_{N}\Sigma ^{-1}= {\mathcal {K}}^{*}*_{N}{\mathcal {K}}*_{N}(\Sigma *_{N}{\mathcal {K}})^{d}\) and \({\mathcal {L}}^{*}*_{N}\Sigma ^{-1}= {\mathcal {L}}^{*}*_{N}{\mathcal {K}}*_{N}(\Sigma *_{N}{\mathcal {K}})^{d}\). Multiplying both of these equalities by \({\mathcal {K}}\) and \({\mathcal {L}}\) on the left-hand side, respectively, and using (1.7), we obtain \({\mathcal {K}}*_{N}(\Sigma *_{N}{\mathcal {K}})^{d}=\Sigma ^{-1}\). Hence, \(\Sigma *_{N}{\mathcal {K}}\) is nonsingular. Moreover, by (1.6):
Obviously, \({\mathcal {U}}*_{N} \begin{pmatrix} \Sigma *_{N}{\mathcal {K}}&{}{\mathcal {O}}\\ {\mathcal {O}}&{}{\mathcal {I}} \end{pmatrix}*_{N}{\mathcal {U}}^{*}\) is invertible. Therefore, \({\mathcal {N}}({\mathcal {A}})= {\mathcal {N}}({\mathcal {A}}^{2})\), and hence, Ind\(({\mathcal {A}})\le 1\). Conversely, let \({\mathcal {A}}={\mathcal {C}}_{{\mathcal {A}}}+ {\mathcal {N}}_{{\mathcal {A}}}\) be the core-nilpotent decomposition of \({\mathcal {A}}\). If Ind\(({\mathcal {A}})\le 1\), then \({\mathcal {A}}={\mathcal {C}}_{{\mathcal {A}}}\), and hence, \({\mathcal {A}}^{c,\dagger }= {\mathcal {A}}^{\dagger }*_{N} {\mathcal {C}}_{{\mathcal {A}}} *_{N}{\mathcal {A}}^{\dagger }= {\mathcal {C}}_{{\mathcal {A}}}^{\dagger }*_{N}{\mathcal {C}}_{{\mathcal {A}}} *_{N}{\mathcal {C}}_{{\mathcal {A}}}^{\dagger }= {\mathcal {C}}_{{\mathcal {A}}}^{\dagger }={\mathcal {A}}^{\dagger }\). This completes the proof. \(\square \)
Theorem 2.10
Let \({\mathcal {A}}\in {\mathbb {C}}^{{\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{{N}} \times {\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{{N}}}\). Then, the following conditions are equivalent:
- (1)
\({\mathcal {A}}*_{N}{\mathcal {A}}^{c,\dagger }={\mathcal {A}}*_{N}{\mathcal {A}}^{\dagger }\);
- (2)
\({\mathcal {A}}^{c,\dagger }*_{N}{\mathcal {A}}={\mathcal {A}}^{\dagger }*_{N}{\mathcal {A}}\);
- (3)
\({\mathcal {A}}^{c,\dagger }= {\mathcal {O}} \) if and only if \({\mathcal {A}}\) is nilpotent.
Proof
(1) We know that \({\mathcal {A}}^{c,\dagger }*_{N}{\mathcal {A}}*_{N}{\mathcal {A}}^{c,\dagger } ={\mathcal {A}}^{c,\dagger }\). If \({\mathcal {A}}*_{N}{\mathcal {A}}^{c,\dagger }={\mathcal {A}}*_{N}{\mathcal {A}}^{\dagger }\), pre-multiplying by \({\mathcal {A}}^{\dagger }\), we get \({\mathcal {A}}^{c,\dagger }={\mathcal {A}}^{\dagger }\). Conversely, if \({\mathcal {A}}^{c,\dagger }={\mathcal {A}}^{\dagger }\), we multiplying \({\mathcal {A}}^{c,\dagger }={\mathcal {A}}^{\dagger }\) by \({\mathcal {A}}\) on the left-hand side, we obtain that \({\mathcal {A}}*_{N}{\mathcal {A}}^{c,\dagger }={\mathcal {A}}*_{N}{\mathcal {A}}^{\dagger }\).
(2) Also, by the same method as in (1), the result is obvious.
(3) If \({\mathcal {A}}^{c,\dagger }= {\mathcal {O}}\), that is \({\mathcal {K}}^{*}*_{N}{\mathcal {K}}*_{N}(\Sigma *_{N}{\mathcal {K}})^{d}={\mathcal {O}}\) and \({\mathcal {L}}^{*}*_{N}{\mathcal {K}}*_{N}(\Sigma *_{N}{\mathcal {K}})^{d}={\mathcal {O}}\) by (2.6). Multiplying both of these equalities by \({\mathcal {K}}\) and \({\mathcal {L}}\) on the left-hand side, respectively, and using:
we obtain \({\mathcal {K}}*_{N}(\Sigma *_{N}{\mathcal {K}})^{d}={\mathcal {O}}\). Then \((\Sigma *_{N}{\mathcal {K}})^{d}*_{N}\Sigma *_{N}{\mathcal {K}}*_{N} (\Sigma *_{N}{\mathcal {K}})^{d}=(\Sigma *_{N}{\mathcal {K}})^{d}= {\mathcal {O}}\). Moreover, \((\Sigma *_{N}{\mathcal {K}})^{k}=(\Sigma *_{N}{\mathcal {K}})^{k+1}*_{N}(\Sigma *_{N}{\mathcal {K}})^{d}={\mathcal {O}}\). Therefore, \(\Sigma *_{N}{\mathcal {K}}\) is nilpotent, and hence, \({\mathcal {A}}\) is nilpotent.
Conversely, if \({\mathcal {A}}\) is nilpotent, let Ind\(({\mathcal {A}})=k\), then \({\mathcal {A}}^{d}=({\mathcal {A}}^{l+1})_{g}*_{N}{\mathcal {A}}^{l}={\mathcal {O}}\), where \(l\ge k\), and hence \({\mathcal {A}}^{c,\dagger }={\mathcal {A}}^{\dagger }*_{N}{\mathcal {A}}*_{N} {\mathcal {A}}^{d}*_{N}{\mathcal {A}}*_{N}{\mathcal {A}}^{\dagger }={\mathcal {O}}\). \(\square \)
3 Main results
In this section, we investigate the perturbations for DMP and CMP inverses. First, we extend the recent results on the DMP inverse from the linear operator (Yu and Deng 2016) to the tensor.
Theorem 3.1
Let \({\mathcal {A}} \in {\mathbb {C}}^{{\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{{N}} \times {\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{{N}}}\) be a tensor of index k. There is a Schur form (1.9) of \({\mathcal {A}}\). Then, the Moore–Penrose inverse can be expressed by:
where \(\Delta = [{\mathcal {T}}_{11} *_{N} {\mathcal {T}}_{11}^{*} + {\mathcal {T}}_{12} *_{N} ({\mathcal {I}} - {\mathcal {T}}_{22}^{\dagger } *_{N} {\mathcal {T}}_{22}) *_{N} {\mathcal {T}}_{12}^{*}]^{-1}\).
Proof
Since \({\mathcal {A}}\) has the Schur form (1.9) and
we have
In view of Definition 1.1, it is easy to compute the first equation that
Furthermore, the second equation follows from:
The third equation is verified as:
and the fourth equation can be verified by:
where \({\mathcal {H}} = {\mathcal {T}}_{22}^{\dagger } *_{N} {\mathcal {T}}_{22}+({\mathcal {I}} - {\mathcal {T}}_{22}^{\dagger } *_{N} {\mathcal {T}}_{22}) *_{N} {\mathcal {T}}_{12}^{*} *_{N} \Delta *_{N} {\mathcal {T}}_{12} *_{N} ({\mathcal {I}} - {\mathcal {T}}_{22}^{\dagger } *_{N} {\mathcal {T}}_{22})\).
The tensor \({\mathcal {X}}\) satisfies four equations. Let \({\mathcal {X}}\) and \({\mathcal {Y}}\) satisfy four equations. To prove the uniqueness:
We know that the conclusion hold. \(\square \)
Let \({\mathcal {A}} \in {\mathbb {C}}^{{\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{{N}} \times {\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{{N}}}\) be a tensor of index k. Let us denote the following condition by \(({\mathcal {W}})\):
Theorem 3.2
(Ji and Wei (2018)) Suppose that condition \(({\mathcal {W}})\) holds and Ind\(({\mathcal {B}}) = k\). Then
and
Moreover
From (1.9) and (3.1), we have:
Now, we develop the perturbation bounds for the DMP inverse of the tensor.
Theorem 3.3
Let \({\mathcal {A}} \in {\mathbb {C}}^{{\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{{N}} \times {\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{{N}}}\) be of the form (1.9) and of index k, \({\mathcal {B}} = {\mathcal {A}} + {\mathcal {E}}\). If the perturbation \({\mathcal {E}}\) satisfies \({\mathcal {A}}^{d,\dagger } *_{N} {\mathcal {A}} *_{N} {\mathcal {E}} = {\mathcal {E}} *_{N} {\mathcal {A}}^{d,\dagger } *_{N} {\mathcal {A}} = {\mathcal {E}}\) and \(\Vert {\mathcal {A}}^{d,\dagger } *_{N} {\mathcal {E}}\Vert _{2} < 1\), then
and
Furthermore:
Proof
Assume that the perturbation \({\mathcal {E}} =\begin{pmatrix} {\mathcal {E}}_{11}&{}{\mathcal {E}}_{12} \\ {\mathcal {E}}_{21}&{}{\mathcal {E}}_{22} \end{pmatrix}\) satisfies:
then we have:
Since \({\mathcal {B}}= {\mathcal {A}} + {\mathcal {E}}\), we can obtain:
From \({\mathcal {A}}^{d,\dagger } *_{N} {\mathcal {A}} = {\mathcal {A}}^{d} *_{N} {\mathcal {A}} *_{N} {\mathcal {A}}^{\dagger } *_{N} {\mathcal {A}} = {\mathcal {A}}^{d} *_{N} {\mathcal {A}} = {\mathcal {A}} *_{N} {\mathcal {A}}^{d}\), we verify the conditions:
if and only if
and
Therefore, the \(({\mathcal {W}})\) condition holds. Then
and
Since \(\Vert {\mathcal {A}}^{d,\dagger } *_{N} {\mathcal {E}}\Vert _{2} < 1\), from Lemma 1.1, we have \({\mathcal {I}}+{\mathcal {A}}^{d,\dagger } *_{N} {\mathcal {E}}\) is invertible. And \({\mathcal {T}}_{22}\) is nilpotent tensor with index k. From Theorem 3.1, we can obtain:
where \({\mathcal {G}} = {\mathcal {T}}_{22}^{\dagger } -({\mathcal {I}} - {\mathcal {T}}_{22}^{\dagger } *_{N} {\mathcal {T}}_{22}) *_{N} ({\mathcal {T}}_{12}+{\mathcal {E}}_{12})^{*} *_{N} \Delta *_{N} ({\mathcal {T}}_{12}+{\mathcal {E}}_{12}) *_{N} {\mathcal {T}}_{22}^{\dagger }\). Through calculations, we can obtain:
It is obvious that
We can obtain
Thus
Similarly, we can prove that:
and
Moreover, from (3.3), taking norms of both sides, we obtain:
The proof is complete. \(\square \)
Now, we present the perturbation of Moore–Penrose inverse under the two-sided conditions.
Lemma 3.1
Let \({\mathcal {A}}, {\mathcal {E}}\in {\mathbb {C}}^{{\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{{N}} \times {\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{{N}}}\), \({\mathcal {B}}={\mathcal {A}}+{\mathcal {E}}\). If the perturbation \({\mathcal {E}}\) satisfies \({\mathcal {E}}= {\mathcal {A}}*_{N} {\mathcal {A}}^{\dagger } *_{N}{\mathcal {E}}={\mathcal {E}}*_{N}{\mathcal {A}}^{\dagger }*_{N}{\mathcal {A}}\) and \(\Vert {\mathcal {A}}^{\dagger }*_{N}{\mathcal {E}}\Vert _{2}<1\), then
Proof
Since \({\mathcal {E}}={\mathcal {A}}*_{N}{\mathcal {A}}^{\dagger } *_{N}{\mathcal {E}}={\mathcal {E}}*_{N}{\mathcal {A}}^{\dagger }*_{N}{\mathcal {A}}\),
Since \(\Vert {\mathcal {A}}^{\dagger } *_{N} {\mathcal {E}}\Vert _{2}<1\), by using (1.2) of Lemma 1.2, we know that:
In view of Definition 1.1, it is easy to compute the first equation that:
Furthermore, the second equation follows from:
The third equation is verified as:
and the fourth equation can be verified by:
Therefore, the result holds. \(\square \)
We estimate the perturbation bounds for the CMP inverse of the tensor.
Theorem 3.4
Let \({\mathcal {A}}, {\mathcal {E}}\in {\mathbb {C}}^{{\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{{N}} \times {\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{{N}}}\), Ind\(({\mathcal {A}})=k\) and \({\mathcal {B}}={\mathcal {A}}+{\mathcal {E}}\) be such that \({\mathcal {E}}={\mathcal {A}}*_{N}{\mathcal {A}}^{c,\dagger }*_{N}{\mathcal {E}} ={\mathcal {E}}*_{N} {\mathcal {A}}^{c,\dagger } *_{N} {\mathcal {A}} \) and \(\Vert {\mathcal {A}}^{c,\dagger }*_{N}{\mathcal {E}}\Vert _{2}<1\). Denote
satisfies
i.e., \(X={\mathcal {B}}^{c,\dagger }\).
Proof
Since \({\mathcal {E}}={\mathcal {A}}*_{N}{\mathcal {A}}^{c,\dagger }*_{N}{\mathcal {E}}\),
By Lemma 1.2, \({\mathcal {I}}+{\mathcal {A}}^{c,\dagger }*_{N}{\mathcal {E}}\) is invertible and
Since \({\mathcal {A}}*_{N}{\mathcal {A}}^{c,\dagger } *_{N}{\mathcal {E}}= {\mathcal {C}}_{{\mathcal {A}}} *_{N}{\mathcal {A}}^{\dagger }*_{N}{\mathcal {E}}={\mathcal {E}}\), we obtain:
Moreover:
Similarly:
we obtain
and
which are also true.
Since \(\Vert {\mathcal {A}}^{d}*_{N}{\mathcal {E}}\Vert _{2}= \Vert {\mathcal {A}}^{d}*_{N}{\mathcal {A}} *_{N}{\mathcal {A}}^{c,\dagger }*_{N}{\mathcal {E}}\Vert _{2} \le \Vert {\mathcal {A}}^{d}*_{N}{\mathcal {A}}\Vert _{2}\Vert {\mathcal {A}}^{c,\dagger }*_{N}{\mathcal {E}}\Vert _{2}<1\), by Ji and Wei (2018), we know that:
By using the same method, we also get \(\Vert {\mathcal {A}}^{\dagger }*_{N}{\mathcal {E}}\Vert _{2}<1\), so
And
Using \({\mathcal {B}}*_{N}{\mathcal {X}}={\mathcal {A}}*_{N}{\mathcal {A}}^{c,\dagger }\) of (3.9), we have
Finally, Since \({\mathcal {E}}={\mathcal {E}}*_{N} {\mathcal {A}}^{c,\dagger } *_{N}{\mathcal {A}}*_{N}\),
we have:
Thus, the proof of the theorem is complete. \(\square \)
Theorem 3.5
Let \({\mathcal {A}}, {\mathcal {E}}\in {\mathbb {C}}^{{\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{{N}} \times {\mathbf {I}}_{1} \times \cdots \times {\mathbf {I}}_{{N}}}\), Ind\(({\mathcal {A}})=k\) and \({\mathcal {B}}={\mathcal {A}}+{\mathcal {E}}\) be such that \({\mathcal {E}}={\mathcal {A}}*_{N}{\mathcal {A}}^{c,\dagger } *_{N}{\mathcal {E}}={\mathcal {E}}*_{N}{\mathcal {A}}^{c,\dagger } *_{N}{\mathcal {A}}\). If \(\Vert {\mathcal {A}}^{c,\dagger }*_{N}{\mathcal {E}}\Vert _{2}<1\), then
Moreover:
and
References
Baksalary O, Trenkler G (2010) Core inverse of matrices. Linear Multilinear Algebra 58:681–697
Behera R, Mishra D (2017) Further results on generalized inverses of tensors via the Einstein product. Linear Multilinear Algebra 65:1662–1682
Behera R, Nandi A, Sahoo J (2019) Further results on the Drazin inverse of even-order tensors, arXiv:1904.10783
Brazell M, Li N, Navasca C, Tamon C (2013) Solving multilinear systems via tensor inversion. SIAM J Matrix Anal Appl 34:542–570
Ding W, Wei Y (2016) Theory and computation of tensors: multi-dimensional arrays. Academic Press, New York
Einstein A (2007) The foundation of the general theory of relativity. In: Kox A, Klein M, Schulmann R (eds) The collected papers of Albert Einstein, vol 6. Princeton University Press, Princeton, pp 146–200
Hartwig RE, Spindelböck K (1983) Matrices for which \(A^{\ast } \) and \(A^{\dagger }\) commute. Linear Multilinear Algebra 14:241–256
Ji J, Wei Y (2017) Weighted Moore–Penrose inverses and the fundamental theorem of even-order tensors with Einstein product. Front Math China 12:1317–1337
Ji J, Wei Y (2018) The Drazin inverse of an even-order tensor and its application to singular tensor equations. Comput Math Appl 75:3402–3413
Jin H, Bai M, Benítez J, Liu X (2017) The generalized inverses of tensors and an application to linear models. Comput Math Appl 74:385–397
Liang M, Zheng B (2019) Further results on Moore–Penrose inverses of tensors with application to tensor nearness problems. Comput Math Appl 77:1282–1293
Liang M, Zheng B, Zhao R (2019) Tensor inversion and its application to the tensor equations with Einstein product. Linear Multilinear Algebra 67:843–870
Ma H (2018) Optimal perturbation bounds for the core inverse. Appl Math Comput 336:176–181
Ma H, Li N, Stanimirović PS, Katsikis VN (2019) Perturbation theory for Moore–Penrose inverse of tensor via Einstein product. Comput Appl Math 38:111. https://doi.org/10.1007/s40314-019-0893-6
Malik S, Thome N (2014) On a new generalized inverse for matrices of an arbitrary index. Appl Math Comput 226:575–580
Mehdipour M, Salemi A (2018) On a new generalized inverse of matrices. Linear Multilinear Algebra 66:1046–1053
Miao Y, Qi L, Wei Y (2019a) Generalized tensor function via the tensor singular value decomposition based on the T-product, arXiv:1901.04255v3
Miao Y, Qi L, Wei Y (2019b) T-Jordan canonical form and T-Drazin inverse based on the T-product, arXiv:1902.07024
Sahoo J, Behera R, Stanimirović PS, Katsikis VN, Ma H (2019) Core and Core-EP inverses of tensors. Comput Appl Math (to appear); arXiv:1905.07874
Stanimirović PS, Cirić M, Katsikis VN, Li C, Ma H (2018) Outer and \((b, c)\) inverses of tensors. Linear Multilinear Algebra. https://doi.org/10.1080/03081087.2018.1521783
Sun L, Zheng B, Bu C, Wei Y (2016) Moore–Penrose inverse of tensors via Einstein product. Linear Multilinear Algebra 64:686–698
Wei Y, Stanimirović P, Petković M (2018) Numerical and symbolic computations of generalized inverses. World Scientific Publishing Co. Pte. Ltd., Hackensack
Yu A, Deng C (2016) Characterizations of DMP inverse in a Hilbert space. Calcolo 53:331–341
Author information
Authors and Affiliations
Corresponding author
Additional information
Communicated by Jinyun Yuan.
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Haifeng Ma: Supported by the bilateral project between China and Poland (No. 37-18) and the National Natural Science Foundation of China under Grant 11971136.
Rights and permissions
About this article
Cite this article
Wang, B., Du, H. & Ma, H. Perturbation bounds for DMP and CMP inverses of tensors via Einstein product. Comp. Appl. Math. 39, 28 (2020). https://doi.org/10.1007/s40314-019-1007-1
Received:
Revised:
Accepted:
Published:
DOI: https://doi.org/10.1007/s40314-019-1007-1