The Moore-Penrose Inverse of a Reduced Biquaternion Tensor
Introduction
The Moore-Penrose inverse of a matrix is the unique matrix satisfying the four Penrose equations. This article defines the same four equations for a reduced biquaternion tensor under the t-product, proves that their solution exists and is unique, computes it from the t-SVD, and records its main properties. The inverse is written $A^\dagger$; the dagger here is the Moore-Penrose inverse of a tensor, not the adjoint of an operator.
The product is The t-Product of Reduced Biquaternion Tensors, the decomposition is The Singular Value Decomposition of a Reduced Biquaternion Tensor, and the conjugate transpose is the bar-transpose fixed there. The matrix counterpart is the Moore-Penrose inverse of Least Squares and Total Least Squares over the Reduced Biquaternions; the equations that use the inverse are Tensor Equations over the Reduced Biquaternions.
The source is Cui-E Yu, Xin Liu, Hui Luo and Yang Zhang, "Reduced biquaternion tensors and applications in color video processing", Section 4.
The Definition
Let $A\in{\mathbb{B}_{\mathrm{red}}}^{n_1\times n_2\times n_3}$. A tensor $X\in{\mathbb{B}_{\mathrm{red}}}^{n_2\times n_1\times n_3}$ is a Moore-Penrose inverse of $A$ if
$$ A\ast X\ast A=A,\qquad X\ast A\ast X=X,\qquad (A\ast X)^{\mathsf H}=A\ast X,\qquad (X\ast A)^{\mathsf H}=X\ast A. $$
These are the four Penrose equations, written with the t-product and the conjugate transpose of the branch. The inverse is written $A^\dagger$.
Theorem (existence and uniqueness). Every $A\in{\mathbb{B}_{\mathrm{red}}}^{n_1\times n_2\times n_3}$ has exactly one Moore-Penrose inverse.
Proof. Write $A=A_{(1)}\Pi_++A_{(2)}\Pi_-$. The product, the conjugate transpose and the idempotents act coordinatewise, so the four equations are equivalent to the two complex t-product Penrose systems
$$ A_{(i)}\ast X_{(i)}\ast A_{(i)}=A_{(i)},\quad X_{(i)}\ast A_{(i)}\ast X_{(i)}=X_{(i)},\quad (A_{(i)}\ast X_{(i)})^{\mathsf H}=A_{(i)}\ast X_{(i)},\quad (X_{(i)}\ast A_{(i)})^{\mathsf H}=X_{(i)}\ast A_{(i)}, $$
for $i=1,2$. The Fourier transform identifies each complex t-product with the slicewise matrix product; in each slice the matrix Moore-Penrose inverse exists and is unique. Hence each complex system has a unique solution, and the pair is the unique solution of the tensor system.
The proof also gives the computation: transform, take the matrix Moore-Penrose inverse of every transform slice in each coordinate, and invert the transform. The tensor inverse is not the entrywise inverse.
The SVD Formula
Let $A=U\ast S\ast V^{\mathsf H}$ be the t-SVD of The Singular Value Decomposition of a Reduced Biquaternion Tensor, with $U$ and $V$ $*$-unitary and $S$ f-diagonal. Let $S^\dagger$ be the f-diagonal tensor defined in the transform by
$$ \widehat{S^\dagger}^{(k)}=\operatorname{diag}\bigl((\widehat S^{(k)}(1,1))^\dagger,\dots,(\widehat S^{(k)}(r,r))^\dagger\bigr),\qquad 1\le k\le n_3, $$
with $r=\min(n_1,n_2)$ and the elementwise Moore-Penrose inverse
$$ s^\dagger=s_{(1)}^\dagger\Pi_++s_{(2)}^\dagger\Pi_-,\qquad z^\dagger=\begin{cases}1/z,&z\neq0,\\0,&z=0,\end{cases} $$
for $s\in\mathbb{B}_{\mathrm{red}}$ and $z\in\mathbb{C}$. This is the scalar Moore-Penrose inverse with respect to the bar conjugation; it satisfies $ss^\dagger s=s$, $s^\dagger ss^\dagger=s^\dagger$, and $ss^\dagger$ and $s^\dagger s$ are each $0$ or $1$ in each idempotent coordinate. Then
$$ \boxed{\; A^\dagger=V\ast S^\dagger\ast U^{\mathsf H} \;}. $$
Remark (the inversion is in the transform). The diagonal entry of $S$ at $(i,i,k)$ is an element of $\mathbb{B}_{\mathrm{red}}$, but it is not inverted entry by entry in the time domain. A diagonal tube of invertibles can have a vanishing Fourier coefficient: for $n_3=2$, the tube $(1,1)$ has transform $(2,0)$, and the entrywise inverse $(1,1)$ is not the Moore-Penrose inverse of the tube. The reciprocal-or-zero is taken on the transform slices, as in the display.
Proof of the formula. In the transform, $S$ is f-diagonal, so the four Penrose equations for $S$ and $S^\dagger$ reduce to the scalar equations on each diagonal entry of each slice; hence $S^\dagger$ is the Moore-Penrose inverse of $S$. For $*$-unitary $U,V$, the four equations for $A=U\ast S\ast V^{\mathsf H}$ and $X=V\ast S^\dagger\ast U^{\mathsf H}$ follow by cancellation:
$$ A\ast X\ast A=U\ast S\ast S^\dagger\ast S\ast V^{\mathsf H}=U\ast S\ast V^{\mathsf H}=A, $$
$$ X\ast A\ast X=V\ast S^\dagger\ast S\ast S^\dagger\ast U^{\mathsf H}=V\ast S^\dagger\ast U^{\mathsf H}=X, $$
and $A\ast X=U\ast S\ast S^\dagger\ast U^{\mathsf H}$ and $X\ast A=V\ast S^\dagger\ast S\ast V^{\mathsf H}$ are Hermitian because $S\ast S^\dagger$ and $S^\dagger\ast S$ are f-diagonal with transform slices diagonal and entries $0$ or $1$. Uniqueness is the theorem above.
Remark (zero divisors). If a transform diagonal entry $s$ is a nonzero zero divisor, one of its complex coordinates is zero and the other is not. The elementwise inverse $s^\dagger$ then has the reciprocal in the nonzero coordinate and zero in the other; it is not the algebra inverse of $s$, which does not exist. The Moore-Penrose inverse exists even when the algebra inverse does not, because the Penrose equations are solved coordinatewise.
The Coordinate Reduction
The theorem of the first section can be read as a statement about the two idempotent coordinates:
$$ \boxed{\; A^\dagger=A_{(1)}^\dagger\Pi_++A_{(2)}^\dagger\Pi_- \;}, $$
where $A_{(i)}^\dagger$ is the Moore-Penrose inverse of the complex tensor $A_{(i)}$ under the complex t-product. The two complex inverses are computed independently. In the Fourier domain, for each $k$,
$$ \widehat{A^\dagger}^{(k)}=(\widehat A^{(k)})^\dagger, $$
the matrix Moore-Penrose inverse of the $k$-th transform slice, and the two coordinates are the two complex matrix pseudoinverses.
The algorithm. Transform $A_{(1)}$ and $A_{(2)}$ along the third mode; for each $k$, compute the matrix Moore-Penrose inverse of $\widehat A_{(1)}^{(k)}$ and of $\widehat A_{(2)}^{(k)}$ by the complex SVD; assemble the two coordinates; invert the transform. The cost is the cost of the t-SVD,
$$ O\bigl(n_1n_2n_3\log n_3+n_3\min(n_1^2n_2,n_1n_2^2)\bigr). $$
The Properties
Proposition (reflexivity). $(A^\dagger)^\dagger=A$.
Proof. The four equations are symmetric in $A$ and $A^\dagger$: the first and second swap, and the third and fourth swap. Hence $A$ is a Moore-Penrose inverse of $A^\dagger$, and uniqueness gives the identity.
Proposition (the adjoint identity). $(A^\dagger)^{\mathsf H}=(A^{\mathsf H})^\dagger$.
Proof. Apply the conjugate transpose to the four equations for $A^\dagger$, using $(X\ast Y)^{\mathsf H}=Y^{\mathsf H}\ast X^{\mathsf H}$ and $(A^{\mathsf H})^{\mathsf H}=A$. The result is the four equations for $(A^{\mathsf H})^\dagger$ with the unknown $(A^\dagger)^{\mathsf H}$; uniqueness gives the identity.
Proposition ($*$-unitary invariance). If $P$ and $Q$ are $*$-unitary tensors of compatible sizes, then
$$ \boxed{\; (P\ast A\ast Q)^\dagger=Q^{\mathsf H}\ast A^\dagger\ast P^{\mathsf H} \;}. $$
Proof. Substitute $X=Q^{\mathsf H}\ast A^\dagger\ast P^{\mathsf H}$ into the four equations for $P\ast A\ast Q$ and cancel the $*$-unitary factors using $P^{\mathsf H}\ast P=Q\ast Q^{\mathsf H}=\mathrm I$.
Proposition (idempotency). The tensors $A\ast A^\dagger$ and $A^\dagger\ast A$ are Hermitian idempotents:
$$ (A\ast A^\dagger)^2=A\ast A^\dagger,\qquad (A^\dagger\ast A)^2=A^\dagger\ast A, $$
and both are Hermitian. They are the orthogonal projections onto the range of $A$ and onto the range of $A^\dagger$, respectively.
Proof. The Hermiticity is the third and fourth Penrose equations. The idempotency follows from $A\ast A^\dagger\ast A=A$:
$$ (A\ast A^\dagger)^2=A\ast A^\dagger\ast A\ast A^\dagger=A\ast A^\dagger, $$
and similarly for $A^\dagger\ast A$.
Proposition (Hermitian idempotents). If $A$ is Hermitian and idempotent, $A^{\mathsf H}=A$ and $A\ast A=A$, then $A^\dagger=A$.
Proof. The tensor $A$ itself satisfies the four equations: $A\ast A\ast A=A$, $A\ast A\ast A=A$, and $A\ast A=A$ is Hermitian. Uniqueness gives $A^\dagger=A$.
The projectors. Write $\mathrm I_n$ for the identity tensor of shape $n\times n\times n_3$, and define
$$ L_A=\mathrm I_{n_2}-A^\dagger\ast A,\qquad R_A=\mathrm I_{n_1}-A\ast A^\dagger. $$
Then
$$ A\ast L_A=0,\qquad R_A\ast A=0,\qquad L_A\ast A^\dagger=0,\qquad A^\dagger\ast R_A=0, $$
$$ L_A^2=L_A,\qquad R_A^2=R_A,\qquad L_A^{\mathsf H}=L_A,\qquad R_A^{\mathsf H}=R_A, $$
and
$$ L_{A^{\mathsf H}}=R_A,\qquad R_{A^{\mathsf H}}=L_A. $$
The tensor $L_A$ is the orthogonal projection of the domain onto the kernel of $A$; the tensor $R_A$ is the orthogonal projection of the codomain onto the orthogonal complement of the range of $A$. The two projectors are the ones used by Tensor Equations over the Reduced Biquaternions.
Proof. The annihilation identities are direct substitutions: $A(\mathrm I-A^\dagger A)=A-A A^\dagger A=0$, and similarly for the other three. The idempotency follows by expanding the square and using $A^\dagger A A^\dagger=A^\dagger$ and $A A^\dagger A=A$. The Hermiticity follows from the third and fourth Penrose equations. The exchange identities follow from $(A^{\mathsf H})^\dagger=(A^\dagger)^{\mathsf H}$ and $(A\ast A^\dagger)^{\mathsf H}=A\ast A^\dagger$.
The normal case. If $A$ is normal, $A\ast A^{\mathsf H}=A^{\mathsf H}\ast A$, then $A$ and $A^\dagger$ commute,
$$ A\ast A^\dagger=A^\dagger\ast A. $$
Proof. In the transform, every slice $\widehat A^{(k)}$ is a normal matrix, and a normal matrix commutes with its Moore-Penrose inverse. Hence $\widehat A^{(k)}\widehat{A^\dagger}^{(k)}=\widehat{A^\dagger}^{(k)}\widehat A^{(k)}$ for every $k$; inverting the transform gives $A\ast A^\dagger=A^\dagger\ast A$.
The f-diagonal case. If $A$ is f-diagonal, then $A^\dagger$ is f-diagonal, and its transform diagonal entries are the elementwise Moore-Penrose inverses of those of $A$. It equals $A^{-1}$ exactly when $A$ is invertible, equivalently when no transform diagonal entry vanishes. Time-domain entries that are invertible do not suffice: the diagonal tube $(1,1)$ has transform $(2,0)$ and is not invertible. The source states the f-diagonal case entrywise; the tube $(1,1)$ is the counterexample, and the statement is correct only in the transform. This is the transform computation of the SVD formula read on the diagonal.
Verification
The four Penrose equations were recomputed on random tensors of sizes $3\times2\times4$ and $3\times3\times4$, with at most $36$ entries. The residuals of the four equations were $2.3\times10^{-15}$, $8.3\times10^{-17}$, $2.5\times10^{-16}$ and $4.9\times10^{-16}$; reflexivity held to $2.7\times10^{-15}$; the adjoint identity held to $1.3\times10^{-16}$; the two idempotency identities held to $4.8\times10^{-16}$ and $2.8\times10^{-16}$; the projector identities held to $2.2\times10^{-15}$; $*$-unitary invariance held to $3.5\times10^{-16}$; and the Hermitian idempotent reproduced its own inverse exactly. The computation used the transform, the slicewise complex SVD and the reciprocal-or-zero rule on the transform diagonal. The entrywise inverse was checked to fail: for the diagonal tube $(1,1)$ with $n_3=2$, the entrywise inverse differs from $S^\dagger$ by $1.06$ in Frobenius norm, and the product $S\ast S^\dagger$ is not the identity.
Summary
Every reduced biquaternion tensor $A$ has a unique Moore-Penrose inverse $A^\dagger$, the solution of the four Penrose equations under the t-product. It is computed from the t-SVD by $A^\dagger=V\ast S^\dagger\ast U^{\mathsf H}$, where $S^\dagger$ is f-diagonal with the elementwise pseudoinverses of the transform diagonal entries; the scalar pseudoinverse is the reciprocal in each nonzero idempotent coordinate and zero in a vanishing coordinate. Equivalently, the inverse is the pair of the two complex t-product pseudoinverses, computed slicewise in the Fourier domain. It satisfies $(A^\dagger)^\dagger=A$, $(A^\dagger)^{\mathsf H}=(A^{\mathsf H})^\dagger$, $*$-unitary invariance, and the idempotency of $A\ast A^\dagger$ and $A^\dagger\ast A$; a Hermitian idempotent is its own inverse. The projectors $L_A=\mathrm I-A^\dagger\ast A$ and $R_A=\mathrm I-A\ast A^\dagger$ are the orthogonal projections onto the kernel and the cokernel. The inverse exists even when the algebra inverse does not, because the Penrose equations are solved coordinatewise.
Summary of Notation
| Symbol | Meaning |
|---|---|
| $A^\dagger$ | The Moore-Penrose inverse of $A$ under the t-product |
| $A\ast X\ast A=A$ etc. | The four Penrose equations |
| $S^\dagger$ | The f-diagonal pseudoinverse of the middle factor |
| $s^\dagger$ | The scalar pseudoinverse, reciprocal-or-zero in each coordinate |
| $L_A=\mathrm I-A^\dagger\ast A$ | The projector onto the kernel of $A$ |
| $R_A=\mathrm I-A\ast A^\dagger$ | The projector onto the cokernel of $A$ |
Further Reading
- Cui-E Yu, Xin Liu, Hui Luo and Yang Zhang, "Reduced biquaternion tensors and applications in color video processing", arXiv:2403.01690 (2024), Section 4. The definition, the SVD formula and the properties of the tensor inverse.
- Adi Ben-Israel and Thomas N. E. Greville, Generalized Inverses: Theory and Applications (Wiley, 1974; Springer, 2003). The four Penrose equations and the uniqueness of their solution.
- Misha E. Kilmer and Carla D. Martin, "Factorization strategies for third-order tensors", Linear Algebra and its Applications 435 (2011) 641–658. The t-product and the t-SVD that the formula uses.
- Soo-Chang Pei, Ja-Han Chang, Jian-Jiun Ding and Ming-Yang Chen, "Eigenvalues and singular value decompositions of reduced biquaternion matrices", IEEE Transactions on Circuits and Systems I 55 (2008) 2673–2685. The matrix pseudoinverse and the two idempotent coordinates.