Tensor Equations over the Reduced Biquaternions

Introduction

This article solves the tensor equation $A\ast X=B$ over the reduced biquaternions. It gives the solvability criterion, the general solution, the Hermitian solution, the least-squares solutions and the minimum-norm least-squares solution, and it treats the right-handed equation $X\ast A=B$. Every statement reduces to the pair of complex tensor equations carried by the two idempotent coordinates.

The product is The t-Product of Reduced Biquaternion Tensors, the inverse is The Moore-Penrose Inverse of a Reduced Biquaternion Tensor, and the norms are those of Reduced Biquaternion Tensors as a Vector Space. The matrix counterpart is Matrix Equations and Structured Solutions over the Reduced Biquaternions; the application is Color Video Compression and Deblurring with Reduced Biquaternion Tensors.

The source is Cui-E Yu, Xin Liu, Hui Luo and Yang Zhang, "Reduced biquaternion tensors and applications in color video processing", Section 5.

The Solvability Criterion

Let $A\in{\mathbb{B}_{\mathrm{red}}}^{n_1\times n_2\times n_3}$, $B\in{\mathbb{B}_{\mathrm{red}}}^{n_1\times n_4\times n_3}$ and $X\in{\mathbb{B}_{\mathrm{red}}}^{n_2\times n_4\times n_3}$. The equation is

$$ A\ast X=B. $$

The two projectors of The Moore-Penrose Inverse of a Reduced Biquaternion Tensor are

$$ L_A=\mathrm I_{n_2}-A^\dagger\ast A,\qquad R_A=\mathrm I_{n_1}-A\ast A^\dagger. $$

They satisfy $A\ast L_A=0$, $R_A\ast A=0$, $L_A\ast A^\dagger=0$ and $A^\dagger\ast R_A=0$.

Theorem (solvability and general solution). The equation $A\ast X=B$ has a solution if and only if

$$ \boxed{\; R_A\ast B=0 \;}. $$

When this holds, the general solution is

$$ X=A^\dagger\ast B+L_A\ast Y, $$

where $Y\in{\mathbb{B}_{\mathrm{red}}}^{n_2\times n_4\times n_3}$ is arbitrary.

Proof. If $X$ is a solution, then $R_A\ast B=R_A\ast A\ast X=0$ because $R_A\ast A=0$. Conversely, if $R_A\ast B=0$, then $A\ast A^\dagger\ast B=B$, so $X_0=A^\dagger\ast B$ is a solution. For an arbitrary solution $X$,

$$ X=A^\dagger\ast A\ast X+L_A\ast X=A^\dagger\ast B+L_A\ast X, $$

so every solution has the stated form. Conversely, every tensor of that form is a solution because $A\ast L_A=0$ and $A\ast A^\dagger\ast B=B$.

Corollary (the minimum-norm solution). When the equation is solvable, the solution of minimum Frobenius norm is $X=A^\dagger\ast B$.

Proof. The two terms of the general solution are orthogonal: $L_A\ast A^\dagger=0$, hence $(A^\dagger)^{\mathsf H}\ast L_A=0$, and

$$ \langle A^\dagger\ast B+L_A\ast Y\rangle_*=\langle A^\dagger\ast B\rangle_*+\langle L_A\ast Y\rangle_*. $$

The minimum is attained at $L_A\ast Y=0$.

The Least-Squares Solutions

When $R_A\ast B\neq0$ the equation has no solution, and one minimizes the residual instead.

Theorem (least squares). The least-squares solutions of $A\ast X=B$ are

$$ \boxed{\; X=A^\dagger\ast B+L_A\ast W \;}, $$

where $W$ is arbitrary, and the minimum-norm least-squares solution is $X=A^\dagger\ast B$. The least-squares solutions are exactly the solutions of the normal equation

$$ A^{\mathsf H}\ast(A\ast X-B)=0. $$

Proof. By the norm identity of The Singular Value Decomposition of a Reduced Biquaternion Tensor,

$$ \langle A\ast X-B\rangle_*^{1/2}=\frac1{\sqrt{n_3}}\langle \operatorname{diag}(\widehat A)\operatorname{diag}(\widehat X)-\operatorname{diag}(\widehat B)\rangle_*^{1/2}. $$

The minimum is therefore attained slice by slice, and the matrix least-squares solutions of $\widehat A^{(k)}\widehat X^{(k)}=\widehat B^{(k)}$ are

$$ \widehat X^{(k)}=(\widehat A^{(k)})^{\dagger}\widehat B^{(k)}+(\mathrm I-(\widehat A^{(k)})^{\dagger}\widehat A^{(k)})\widehat W^{(k)}. $$

Assembling the slices and inverting the transform gives the stated form. The normal equation is the derivative of the residual, and it is satisfied by the same expression because $A^{\mathsf H}(\mathrm I-A A^\dagger)=0$. The minimum-norm statement is the orthogonality argument of the previous corollary.

Remark (the two normal equations). The equation $A\ast X=B$ and its normal equation have the same least-squares solutions; the normal equation is solvable for every $B$, while the original equation is solvable only when $R_A\ast B=0$. The pseudoinverse solution $A^\dagger\ast B$ is the canonical one: it is the least-squares solution of minimum norm, and it is the exact solution whenever an exact solution exists.

The Hermitian Solution

Let $A,B\in{\mathbb{B}_{\mathrm{red}}}^{n\times n\times n_3}$ be square, and suppose that $X$ is required to be Hermitian, $X^{\mathsf H}=X$. The equation is $A\ast X=B$ with $B$ square.

Theorem (Hermitian solutions). The equation has a Hermitian solution if and only if

$$ \boxed{\; R_A\ast B=0,\qquad B\ast A^{\mathsf H}=A\ast B^{\mathsf H} \;}. $$

When these conditions hold, the general Hermitian solution is

$$ X=A^\dagger\ast B+(A^\dagger\ast B)^{\mathsf H} -A^\dagger\ast B\ast A^{\mathsf H}\ast(A^\dagger)^{\mathsf H} +L_A\ast U\ast L_A, $$

where $U\in{\mathbb{B}_{\mathrm{red}}}^{n\times n\times n_3}$ is an arbitrary Hermitian tensor.

Proof. If $X$ is Hermitian and $A\ast X=B$, then $R_A\ast B=R_A\ast A\ast X=0$, and $B^{\mathsf H}=X^{\mathsf H}\ast A^{\mathsf H}=X\ast A^{\mathsf H}$, so

$$ B\ast A^{\mathsf H}=A\ast X\ast A^{\mathsf H}=A\ast B^{\mathsf H}. $$

Conversely, assume the two conditions. The tensor $B\ast A^{\mathsf H}$ is Hermitian because its conjugate transpose is $A\ast B^{\mathsf H}$, which equals it by the second condition. The displayed $X$ is Hermitian: the first two terms are conjugate pairs, the third is Hermitian because $B\ast A^{\mathsf H}$ is, and the last is Hermitian because $L_A$ and $U$ are. Applying $A$ kills the last term, and the first term gives $B$ because $R_A\ast B=0$. The second term becomes $A\ast B^{\mathsf H}\ast(A^\dagger)^{\mathsf H}$; the third becomes $A\ast A^\dagger\ast B\ast A^{\mathsf H}\ast(A^\dagger)^{\mathsf H}=A\ast B^{\mathsf H}\ast(A^\dagger)^{\mathsf H}$, using $A\ast A^\dagger\ast A=A$ and $B\ast A^{\mathsf H}=A\ast B^{\mathsf H}$. The two cancel, so $A\ast X=B$.

For the general form, let $X_0$ be any Hermitian solution and set $U=X_0$. Put $P=A^\dagger\ast A$ and $B=A\ast X_0$. The first term of the display is $P\ast X_0$, the second is $X_0\ast P$, the third is

$$ A^\dagger\ast A\ast X_0\ast A^{\mathsf H}\ast(A^\dagger)^{\mathsf H} =P\ast X_0\ast A^{\mathsf H}\ast(A^\dagger)^{\mathsf H} =P\ast X_0\ast P, $$

because $X_0\ast A^{\mathsf H}=B^{\mathsf H}$ and $B^{\mathsf H}\ast(A^\dagger)^{\mathsf H}=(A^\dagger\ast B)^{\mathsf H}=(P\ast X_0)^{\mathsf H}=X_0\ast P$; the last term is $(\mathrm I-P)\ast X_0\ast(\mathrm I-P)$. The sum of the four is $X_0$, so the displayed family contains every Hermitian solution.

Remark (the source's condition). The source writes the condition as $A\ast B^{*}=B\ast A^{*}$, with its star the bar-transpose of this branch; this is $A\ast B^{\mathsf H}=B\ast A^{\mathsf H}$, the condition used here. The condition says that $A\ast B^{\mathsf H}$ is Hermitian, and it is what makes the cancellation in the proof work.

Remark (the rectangular case). When $A$ is $n_1\times n_2$ with $n_1\neq n_2$, the Hermitian unknown $X$ is $n_2\times n_2$ and $B$ is $n_1\times n_2$; the two sides of the condition $B\ast A^{\mathsf H}=A\ast B^{\mathsf H}$ have different shapes, so the condition is formable only in the square case. The rectangular Hermitian problem requires a different compatibility condition and is not treated here; the rectangular equation without the Hermitian constraint is covered by the solvability and least-squares theorems above.

The Right-Handed Equation

Let $A\in{\mathbb{B}_{\mathrm{red}}}^{n_2\times n_4\times n_3}$, $B\in{\mathbb{B}_{\mathrm{red}}}^{n_1\times n_4\times n_3}$ and $X\in{\mathbb{B}_{\mathrm{red}}}^{n_1\times n_2\times n_3}$. The equation is

$$ X\ast A=B. $$

Theorem (right-handed least squares). The least-squares solutions of $X\ast A=B$ are

$$ X=B\ast A^\dagger+W\ast R_A, $$

where $R_A=\mathrm I_{n_2}-A\ast A^\dagger$ and $W\in{\mathbb{B}_{\mathrm{red}}}^{n_1\times n_2\times n_3}$ is arbitrary. The minimum-norm least-squares solution is $X=B\ast A^\dagger$. The normal equation is

$$ (X\ast A-B)\ast A^{\mathsf H}=0. $$

Proof. Apply the left-handed theorem to the adjoint equation $A^{\mathsf H}\ast X^{\mathsf H}=B^{\mathsf H}$ and take the conjugate transpose. The general solution becomes $X=B\ast A^\dagger+W\ast R_A$ with $W$ arbitrary, and the minimum-norm statement is preserved by the conjugate transpose.

The Two Idempotent Coordinates

Write $A=A_{(1)}\Pi_++A_{(2)}\Pi_-$, $B=B_{(1)}\Pi_++B_{(2)}\Pi_-$ and $X=X_{(1)}\Pi_++X_{(2)}\Pi_-$. The equation $A\ast X=B$ is equivalent to the two complex tensor equations

$$ A_{(1)}\ast X_{(1)}=B_{(1)},\qquad A_{(2)}\ast X_{(2)}=B_{(2)}. $$

The projectors, the solvability criterion, the general solution, the least-squares solutions and the minimum-norm solutions are the pairs of their complex counterparts. The Hermitian condition is also coordinatewise: $X^{\mathsf H}=X$ is equivalent to $X_{(1)}^{\mathsf H}=X_{(1)}$ and $X_{(2)}^{\mathsf H}=X_{(2)}$, and the compatibility condition $B\ast A^{\mathsf H}=A\ast B^{\mathsf H}$ is equivalent to the two complex conditions

$$ B_{(1)}\ast A_{(1)}^{\mathsf H}=A_{(1)}\ast B_{(1)}^{\mathsf H},\qquad B_{(2)}\ast A_{(2)}^{\mathsf H}=A_{(2)}\ast B_{(2)}^{\mathsf H}. $$

The algorithm. Transform both coordinates, solve the matrix equation slice by slice by the matrix pseudoinverse, and invert the transform. The cost is the cost of the pseudoinverse,

$$ O\bigl(n_1n_2n_3\log n_3+n_3\min(n_1^2n_2,n_1n_2^2)\bigr), $$

plus the slicewise matrix multiplications. The two coordinates are independent and can be solved in parallel.

Verification

The statements were recomputed on random tensors of sizes $3\times2\times4$ and $3\times3\times4$, with at most $36$ entries. For a consistent system, the norms $\langle R_A\ast B\rangle_*^{1/2}$ and $\langle A\ast X_0-B\rangle_*^{1/2}$ were at most $1.6\times10^{-14}$, and the general solution reproduced $B$ to $2.4\times10^{-14}$. For an inconsistent system, the normal equation held to $1.5\times10^{-13}$, and the identity $\langle X\rangle_*=\langle A^\dagger\ast B\rangle_*+\langle L_A\ast W\rangle_*$ held to $1.8\times10^{-29}$. For the Hermitian problem, the residual was $3.1\times10^{-14}$, the symmetry $X^{\mathsf H}-X$ was $4.8\times10^{-15}$, and the completeness with $U=X_0$ held to $3.8\times10^{-15}$; a tensor in the range of $A$ whose product $B\ast A^{\mathsf H}$ was not Hermitian failed the condition by $84.07$. For the right-handed equation, the normal equation held to $1.7\times10^{-13}$, the minimum-norm identity to $2.7\times10^{-15}$, and the consistent solution to $1.9\times10^{-14}$.

Summary

The equation $A\ast X=B$ is solvable if and only if $R_A\ast B=0$, where $R_A=\mathrm I-A\ast A^\dagger$; its general solution is $X=A^\dagger\ast B+L_A\ast Y$, with $L_A=\mathrm I-A^\dagger\ast A$. The least-squares solutions are the same expression with $W$ in place of $Y$, and the minimum-norm least-squares solution is $A^\dagger\ast B$. For square $A$ and $B$, a Hermitian solution exists if and only if $R_A\ast B=0$ and $B\ast A^{\mathsf H}=A\ast B^{\mathsf H}$; its general form is $A^\dagger\ast B+(A^\dagger\ast B)^{\mathsf H}-A^\dagger\ast B\ast A^{\mathsf H}\ast(A^\dagger)^{\mathsf H}+L_A\ast U\ast L_A$ with $U$ Hermitian. The right-handed equation has the mirror solution $X=B\ast A^\dagger+W\ast R_A$. Every statement is the pair of its two complex idempotent-coordinate statements. The pseudoinverse is the canonical solution: exact when the equation is solvable, and of minimum norm among the least-squares solutions otherwise.

Summary of Notation

Symbol Meaning
$A\ast X=B$ The tensor equation
$A^\dagger$ The Moore-Penrose inverse of $A$
$L_A=\mathrm I-A^\dagger\ast A$ The projector onto the kernel of $A$
$R_A=\mathrm I-A\ast A^\dagger$ The projector onto the cokernel of $A$
$A_{(1)},A_{(2)}$ The two idempotent coordinates
$X\ast A=B$ The right-handed equation

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 5. The solvability, least-squares and Hermitian solutions of $A\ast_{Ht}X=B$.
  • Adi Ben-Israel and Thomas N. E. Greville, Generalized Inverses: Theory and Applications (Wiley, 1974; Springer, 2003). The matrix solvability criterion and the least-squares theory that the tensor statements extend.
  • Roger A. Horn and Charles R. Johnson, Matrix Analysis (Cambridge University Press, 2013). The Hermitian solutions of linear matrix equations and the projector identities.
  • 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 used in the proofs.