Zero Divisors of the Reduced Biquaternion Tensor Algebra

Introduction

The square reduced biquaternion tensors of a fixed shape form a finite-dimensional algebra under the t-product. This article identifies its units, its zero divisors, and the left and right annihilators of an element. The answer is read on the Fourier transform: a tensor is invertible exactly when every transform slice is an invertible matrix, and a nonzero tensor is a zero divisor exactly when some transform slice is singular in one of the two idempotent coordinates. The article also defines the t-determinant and the t-algebraic norm that decide invertibility.

The product is The t-Product of Reduced Biquaternion Tensors, the inverse is The Moore-Penrose Inverse of a Reduced Biquaternion Tensor, and the entries are those of Reduced Biquaternion Space. This article is the corpus's own reading of the tensor algebra; it is the tensor analogue of Zero Divisors of the Algebra of the Reduced Biquaternions and Annihilators of the Algebra of the Reduced Biquaternions. It is not in the source of the branch, and it is written only after the product is settled.

The Algebra

Fix $n\ge1$ and $n_3\ge1$, and let

$$ T_n={\mathbb{B}_{\mathrm{red}}}^{n\times n\times n_3} $$

with the t-product $\ast$. The identity is the tensor $\mathrm I$ with $I_n$ in its first frontal slice and zeros elsewhere. The algebra is associative and distributive, it is not commutative for $n\ge2$, and it is finite-dimensional over $\mathbb{R}$, of dimension

$$ \dim_{\mathbb R}T_n=4n^2n_3. $$

It is also a free module of rank $n^2n_3$ over $\mathbb{B}_{\mathrm{red}}$, and an algebra over $\mathbb{B}_{\mathrm{red}}$.

The Fourier transform along the third mode is an isomorphism of algebras

$$ \boxed{\; T_n\cong\prod_{k=1}^{n_3}M_n(\mathbb{B}_{\mathrm{red}}) \;}, $$

where the product is coordinatewise. Composing with the isomorphism $\mathbb{B}_{\mathrm{red}}\cong\mathbb{C}\oplus\mathbb{C}$ gives

$$ T_n\cong\prod_{k=1}^{n_3}\bigl(M_n(\mathbb{C})\oplus M_n(\mathbb{C})\bigr). $$

The invertibles. A tensor $A\in T_n$ is invertible if there is $B$ with $A\ast B=B\ast A=\mathrm I$. The inverse is unique, and it is written $A^{-1}$.

Theorem (the invertibility criterion). A tensor $A\in T_n$ is invertible if and only if every transform slice $\widehat A^{(k)}$ is an invertible matrix over $\mathbb{B}_{\mathrm{red}}$, equivalently if and only if both complex matrices $\widehat A_{(1)}^{(k)}$ and $\widehat A_{(2)}^{(k)}$ are nonsingular for every $k$. In that case

$$ \widehat{A^{-1}}^{(k)}=(\widehat A^{(k)})^{-1}. $$

Proof. The transform is an algebra isomorphism, and it carries $\mathrm I$ to $I_n$ in every slice. Hence $A\ast B=\mathrm I$ if and only if $\widehat A^{(k)}\widehat B^{(k)}=I_n$ for every $k$. A matrix over $\mathbb{B}_{\mathrm{red}}$ is invertible if and only if both of its complex coordinate matrices are invertible.

Remark (the spatial slices do not decide). Invertibility is decided by the transform slices, not by the frontal slices $A^{(k)}$. A tensor can have a singular frontal slice and still be an invertible element, because the transform mixes the frontal slices; and it can have every frontal slice nonsingular and still fail to be an invertible element, because one transform slice is singular. The Fourier transform is the coordinate system in which the algebra is a product of matrix algebras.

The Zero Divisors

A nonzero tensor $A\in T_n$ is a left zero divisor if there is a nonzero $B$ with $A\ast B=0$, and a right zero divisor if there is a nonzero $B$ with $B\ast A=0$. The zero tensor is excluded by the word nonzero, as usual.

Theorem (the non-invertibles are the zero divisors). A nonzero tensor $A\in T_n$ is a left zero divisor if and only if it is not invertible, and the same holds for the right zero divisors. Hence the left zero divisors, the right zero divisors and the nonzero non-invertibles are the same set.

Proof. A zero divisor cannot be an invertible element: if $A\ast B=0$ with $B\neq0$ and $A$ is invertible, then $B=A^{-1}\ast A\ast B=0$, a contradiction. Conversely, let $A$ be a nonzero non-invertible. The left multiplication $\mathcal L_A:T_n\to T_n$, $\mathcal L_A(B)=A\ast B$, is a linear map of a finite-dimensional real vector space. If $\mathcal L_A$ were invertible, there would be $B$ with $A\ast B=\mathrm I$; in a finite-dimensional algebra a right inverse is a two-sided inverse, so $A$ would be an invertible element, a contradiction. Hence $\mathcal L_A$ is singular, and there is a nonzero $B$ with $A\ast B=0$. The same argument with the right multiplication gives a nonzero $C$ with $C\ast A=0$.

Remark (the left and right annihilators differ). The theorem says that the sets of left and right zero divisors coincide. It does not say that the annihilators of a given element coincide; they are different submodules in general, and the next section gives their dimensions.

Theorem (the coordinate criterion). Write $A=A_{(1)}\Pi_++A_{(2)}\Pi_-$. A nonzero $A\in T_n$ is a zero divisor if and only if $A_{(1)}$ or $A_{(2)}$ is not invertible in the complex tensor algebra $T_n(\mathbb{C})$. Equivalently, $A$ is a zero divisor if and only if there are $k\in\{1,\dots,n_3\}$ and $j\in\{1,2\}$ such that the complex matrix $(\widehat A_{(j)})^{(k)}$ is singular.

Proof. If $A_{(1)}$ is not invertible, either it is zero, and then $B=\Pi_+$ is a nonzero annihilator, or it is a nonzero zero divisor, and there is a nonzero $B_{(1)}$ with $A_{(1)}\ast B_{(1)}=0$; in both cases $B=B_{(1)}\Pi_+$ is nonzero and $A\ast B=0$. The case of $A_{(2)}$ is the same. Conversely, if $A\ast B=0$ with $B\neq0$, then $A_{(1)}\ast B_{(1)}=0$ and $A_{(2)}\ast B_{(2)}=0$; at least one of $B_{(1)},B_{(2)}$ is nonzero, so the corresponding coordinate has a nonzero annihilator and is not invertible. The complex criterion is proved by the transform: a complex tensor is a non-invertible if and only if some transform slice is singular, because a nonzero kernel vector in that slice, placed in that slice and zero elsewhere, is a nonzero annihilator; and conversely a nonzero annihilator has a nonzero slice, which makes the corresponding slice singular.

Corollary (the unit criterion again). A tensor is invertible if and only if no transform slice is singular in either coordinate. This is the criterion of the previous section.

The all-ones tensor. Let $J$ be the tensor all of whose entries are $1$. Its transform is

$$ \widehat J^{(1)}=n_3\,\mathbf 1\mathbf 1^{\mathsf T},\qquad \widehat J^{(k)}=0\quad(2\le k\le n_3), $$

where $\mathbf 1$ is the all-ones column. Every slice is singular, so $J$ is a zero divisor, except when $n=1$ and $n_3=1$, where $J$ is the scalar $1$, an invertible element. For $n\ge2$, a witness is any tensor $B$ whose every frontal slice is $C/n_3$, where $C$ is a complex matrix with $\mathbf 1^{\mathsf T}C=0$; then $J\ast B=0$. For $n=1$ and $n_3\ge2$ the transform slices at $k\ge2$ are zero, and a nonzero tensor supported on one of them is an annihilator. Every entry of $J$ is the unit $1$, and $J$ is not invertible except when $n=1$ and $n_3=1$. Entrywise invertibility does not imply tensor invertibility.

The Annihilators

The left annihilator and the right annihilator of $A\in T_n$ are

$$ \operatorname{Ann}_L(A)=\{B\in T_n:A\ast B=0\},\qquad \operatorname{Ann}_R(A)=\{B\in T_n:B\ast A=0\}. $$

Both are submodules of $T_n$ over $\mathbb{B}_{\mathrm{red}}$, and both contain $0$. The left annihilator is the kernel of the left multiplication $\mathcal L_A$; the right annihilator is the kernel of the right multiplication $\mathcal R_A$.

Theorem (the transform description). In the transform,

$$ \boxed{\; \operatorname{Ann}_L(A)=\Bigl\{B:\widehat B^{(k)}\in\ker_L\bigl(\widehat A^{(k)}\bigr)\ \text{for every }k\Bigr\} \;}, $$

where $\ker_L(M)=\{Y:MY=0\}$ is the right null space of the matrix $M$, and

$$ \operatorname{Ann}_R(A)=\Bigl\{B:\widehat B^{(k)}\in\ker_R\bigl(\widehat A^{(k)}\bigr)\ \text{for every }k\Bigr\}, $$

where $\ker_R(M)=\{Y:YM=0\}$ is the left null space.

Proof. The transform is an algebra isomorphism, so $A\ast B=0$ if and only if $\widehat A^{(k)}\widehat B^{(k)}=0$ for every $k$. The two displays are the slicewise reading of this identity.

The dimensions. For a complex tensor $B\in T_n(\mathbb{C})$, the complex dimension of the left annihilator is

$$ \dim_{\mathbb C}\operatorname{Ann}_L(B)=\sum_{k=1}^{n_3}\dim_{\mathbb C}\ker\bigl(\widehat B^{(k)}\bigr), $$

and the right annihilator has the same dimension, because a matrix and its transpose have the same rank. For a reduced biquaternion tensor $A=A_{(1)}\Pi_++A_{(2)}\Pi_-$, the real dimension is

$$ \dim_{\mathbb R}\operatorname{Ann}_L(A) =2\sum_{k=1}^{n_3}\Bigl(\dim_{\mathbb C}\ker\widehat A_{(1)}^{(k)}+\dim_{\mathbb C}\ker\widehat A_{(2)}^{(k)}\Bigr), $$

and the same formula holds for $\operatorname{Ann}_R(A)$.

Proof. The complex formula is the rank-nullity theorem applied slice by slice. The real formula is the complex formula applied to the two idempotent coordinates; each complex dimension contributes two real dimensions because $\mathbb{B}_{\mathrm{red}}$ is two-dimensional over $\mathbb{C}$. The right-hand formula follows from the equality of the rank of a matrix and its transpose.

Remark (the annihilators need not coincide). The dimensions of $\operatorname{Ann}_L(A)$ and $\operatorname{Ann}_R(A)$ are equal, but the submodules are different in general. Take a tensor whose only nonzero frontal slice is a non-normal matrix $M$ in the first slice. The left annihilator is the set of tensors whose every frontal slice $Y$ satisfies $MY=0$, and the right annihilator is the set whose every frontal slice $Y$ satisfies $YM=0$; the right and left null spaces of $M$ differ, so the annihilators differ. Equality of dimensions is not equality of submodules.

The extreme cases. For an invertible element, both annihilators are $\{0\}$. For the zero tensor, both are the whole algebra $T_n$. For a nonzero zero divisor, both are nonzero and proper. The left annihilator of $A$ is the whole algebra if and only if $A=0$.

The t-Determinant and the Algebraic Norm

The t-determinant of $A\in T_n$ is

$$ \det\nolimits_{\mathrm t}(A)=\prod_{k=1}^{n_3}\det\bigl(\widehat A^{(k)}\bigr)\in\mathbb{B}_{\mathrm{red}}. $$

It is the determinant of the block circulant matrix $\operatorname{circ}(A)$ read as an element of $\mathbb{B}_{\mathrm{red}}$: the Fourier matrix diagonalises the block circulant, and the determinant of a block diagonal matrix is the product of the determinants of its blocks. It is multiplicative,

$$ \boxed{\; \det\nolimits_{\mathrm t}(A\ast B)=\det\nolimits_{\mathrm t}(A)\det\nolimits_{\mathrm t}(B) \;}, $$

because $\operatorname{circ}(A\ast B)=\operatorname{circ}(A)\operatorname{circ}(B)$.

Theorem (the determinant criterion). A tensor $A\in T_n$ is invertible if and only if $\det_{\mathrm t}(A)$ is invertible in $\mathbb{B}_{\mathrm{red}}$, that is, if and only if both complex numbers $\det_{t,(1)}(A)$ and $\det_{t,(2)}(A)$ are nonzero.

Proof. The determinant of a matrix over $\mathbb{B}_{\mathrm{red}}$ is invertible if and only if the matrix is invertible. The product of the slice determinants is invertible if and only if every slice determinant is invertible, because $\mathbb{B}_{\mathrm{red}}$ is a product of two fields. Hence the criterion is the unit criterion of the first section.

The t-algebraic norm. The algebraic norm of the element $\det_{\mathrm t}(A)$ is

$$ \langle\det\nolimits_{\mathrm t}(A)\rangle_{\mathrm{ts}} =\det\nolimits_{t,(1)}(A)\det\nolimits_{t,(2)}(A)\in\mathbb{C}. $$

It is nonzero if and only if $A$ is invertible. This is the scalar that decides invertibility. It is not a topological norm: it is a polynomial, it is complex-valued, and it vanishes on the zero divisors. The Frobenius norm does not decide invertibility: the all-ones tensor has $\langle J\rangle_*^{1/2}=n\sqrt{n_3}>0$ and is a zero divisor.

Remark (the zero divisors are not a single hypersurface). The t-algebraic norm is a complex polynomial in the real coordinates of $A$, and the invertibles are the complement of its zero set. That zero set is the union of $\det_{t,(1)}=0$ and $\det_{t,(2)}=0$. Each is the zero set of a complex polynomial, hence a real subvariety of codimension two and a complex hypersurface in the complex structure of $T_n$; their union also has real codimension two, not one. The invertibles are the complement of this union and form a dense open subset. This is the tensor analogue of the zero divisors of $\mathbb{B}_{\mathrm{red}}$, the union of the two complex lines $Q_{(1)}=0$ and $Q_{(2)}=0$, which is the zero set of the complex quadratic form $Q_{(1)}Q_{(2)}$.

Verification

The unit criterion, the zero-divisor criterion and the determinant were recomputed on random tensors of size $3\times3\times4$, with $36$ entries. A random tensor was an invertible element; its two-sided inverse reproduced $\mathrm I$ to $2.8\times10^{-16}$; the t-determinant was multiplicative to a relative error of $5.0\times10^{-15}$; and it agreed with the determinant of the block circulant to a relative error of $8.7\times10^{-16}$. The all-ones tensor was a zero divisor, with $J\ast B=0$ exactly for the witness $B$ of constant slices $C/n_3$ with $\mathbf 1^{\mathsf T}C=0$, and with $\det_{\mathrm t}(J)=0$. For the tensor whose only nonzero frontal slice was a rank-one matrix in one idempotent coordinate, the complex block circulant of that coordinate had nullity $8$ and the sum of the transform-slice nullities was $8$, as the dimension formula predicts; the real dimension of the annihilator was $16$. A random unit had annihilator $\{0\}$, and a tensor with a rank-one transform slice had a nonzero annihilator of the predicted dimension. For $n=1$, the all-ones tensor was the unit $1$ at $n_3=1$ and a zero divisor at $n_3=2,3$, with transform slices $(2,0)$ and $(3,0,0)$.

Summary

The square tensors $T_n={\mathbb{B}_{\mathrm{red}}}^{n\times n\times n_3}$ form a finite-dimensional algebra under the t-product, isomorphic to $\prod_k M_n(\mathbb{B}_{\mathrm{red}})$ and to $\prod_k(M_n(\mathbb{C})\oplus M_n(\mathbb{C}))$. A tensor is invertible exactly when every transform slice is invertible in both idempotent coordinates; the nonzero non-invertibles are exactly the left zero divisors and exactly the right zero divisors. A nonzero tensor is a zero divisor exactly when some transform slice is singular in one coordinate. The left and right annihilators are the slicewise right and left null spaces; they have equal dimensions but need not coincide. The t-determinant is the product of the transform-slice determinants, it is multiplicative, and its algebraic norm is the scalar that decides invertibility; the Frobenius norm does not. The nonzero zero divisors are the nonzero tensors whose t-algebraic norm vanishes; they lie in the union of two real subvarieties of codimension two.

Summary of Notation

Symbol Meaning
$T_n={\mathbb{B}_{\mathrm{red}}}^{n\times n\times n_3}$ The tensor algebra under the t-product
$\mathrm I$ The identity tensor, $I_n$ in the first slice
$\operatorname{Ann}_L(A)$, $\operatorname{Ann}_R(A)$ The left and right annihilators of $A$
$\det_{\mathrm t}(A)$ The t-determinant, $\prod_k\det(\widehat A^{(k)})$
$\langle\det_{\mathrm t}(A)\rangle_{\mathrm{ts}}$ The t-algebraic norm, the scalar that decides invertibility
$\widehat A^{(k)}$ The $k$-th transform slice

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). The t-product, the transform and the invertibility criterion that this article reads as an algebra.
  • 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 block circulant representation.
  • Adi Ben-Israel and Thomas N. E. Greville, Generalized Inverses: Theory and Applications (Wiley, 1974; Springer, 2003). The finite-dimensional algebra facts about one-sided inverses and zero divisors.
  • The corpus articles Zero Divisors of the Algebra of the Reduced Biquaternions and Annihilators of the Algebra of the Reduced Biquaternions. The four-dimensional case that the tensor algebra extends.