Toeplitz-Block Toeplitz Matrices

• Toeplitz-block Toeplitz matrices are defined as block matrices with Toeplitz structure at both the block level and within each block, generalizing classical Toeplitz systems.
• They enable efficient inversion and minimal data recovery using operator identities and low-rank block encodings, forming the backbone of fast algorithms in multidimensional settings.
• Their rich spectral properties and positivity criteria offer practical insights for numerical analysis, multivariate convolution, and quantum information applications.

A Toeplitz-block Toeplitz (TBT) matrix is a structured matrix of size $mn \times mn$ exhibiting Toeplitz structure both at the block level and within each block. These matrices are of central interest in multivariate convolution, operator theory, random matrix theory, and numerical linear algebra, owing to their rich algebraic and spectral properties, as well as their role in fast algorithms and multidimensional systems. This article presents a detailed, mathematically precise overview of TBT matrices, their algebraic and operator-theoretic structure, inversion theory, minimal data recovery, positivity, and spectral properties, referencing foundational and recent results.

1. Definition and Matrix Representation

Let $m, n \in \mathbb{N}$. A Toeplitz-block Toeplitz matrix $T \in \mathbb{C}^{mn \times mn}$ is defined such that $T$ is block-Toeplitz, with each block itself Toeplitz: $T = [T_{i-j}]_{i,j=1}^n,$ where, for each $k = -(n-1), \ldots, n-1$, $T_k = [t_{p-q}^{(k)}]_{p,q=1}^m$ is an $m \times m$ Toeplitz matrix. Thus, for $1 \leq i,j \leq n$ and $1 \leq p,q \leq m$,

$T_{(i,p),(j,q)} = t_{p-q}^{(i-j)}.$

This double-Toeplitz structure generalizes classical Toeplitz matrices (the $m=1$ case) and block Toeplitz matrices, encoding separable dependencies with respect to two (or more) spatial or temporal variables.

2. Operator-Theoretic Realization and Tensor Structures

TBT matrices correspond naturally to operator systems and tensor products of Toeplitz operator algebras. For $n \times n$ and $m \times m$ Toeplitz operator systems $T_n$, $T_m$, their algebraic tensor product $T_n \otimes T_m$ is the linear span of $s_n^k \otimes S_k$, where $s_n$ is the shift and $S_k \in T_m$ are Toeplitz. This realizes TBT matrices as elements of $T_n \otimes T_m$, i.e., as block Toeplitz matrices whose blocks are themselves Toeplitz. Isomorphic identifications hold at the level of operator systems, with dualities to truncated trigonometric polynomial spaces, as established in the Connes–van Suijlekom isomorphism and subsequent work (Farenick, 2021).

3. Inversion Theory and Minimal Data Recovery

Comprehensive analysis of the inversion of TBT matrices shows that their inverses are again TBT, with explicit operator identities and minimal information for reconstruction. Let $T$ be invertible. The main result (Sakhnovich, 2017) is that $T^{-1}$ can be uniquely reconstructed from a finite set of low-rank data:

• There exist $2m \times 2n$ (resp. $2n \times 2m$) matrices $g_{12}$ or $g_{21}$ encoding $T^{-1} Q_p T^{-1}$, where $Q_p$ derive from Toeplitz commutator identities with block integration matrices $A_1, A_2$.
• The fundamental operator identities for $T$ are $A_p T - T A_p^* = Q_p$ with $Q_p$ rank at most $2$ times the block size.
• Introducing a $2 \times 2$ block matrix function $G(\mu_1, \mu_2)$ (the "symbol"), $T$ is invertible if and only if $G$ is invertible everywhere on a suitable (e.g., toroidal) domain, $\Delta(\mu_1,\mu_2) = \det G(\mu_1,\mu_2)$ does not vanish, and then $T^{-1}$ is computable from $G$ and the $g_{12}$ or $g_{21}$ data.
• Explicit formulas produce, via contour or Fourier expansions, the unique coefficients of $T^{-1}$ expressed in an appropriate basis.

The minimal data required for recovery is a single such $g_{12}$ or $g_{21}$ block (size $4mn$ scalar entries), making the reconstruction efficient and non-redundant (Sakhnovich, 2017).

4. Structural and Algebraic Properties of the Inverse

The inverse $T^{-1}$ of a TBT matrix preserves the TBT structure. Theorem 3.3 in (Sakhnovich, 2017) asserts:

• $T^{-1}$ satisfies the same commutation relations with the block-integration matrices: $A_p T^{-1} - T^{-1} A_p^* = i\,[I_{mn}]\,\initializer_p$.
• Each $m \times m$ sub-block of $T^{-1}$ is Toeplitz in the block indices, and the entries within each block depend only on the difference of inner (block) indices.
• Fourier or complex-analytic representations imply $T^{-1}$ exhibits off-diagonal decay corresponding to the "almost banded" nature familiar in multidimensional Toeplitz systems.

These properties link the double Toeplitz structure to multidimensional convolution operators and their invertibility in analysis and signal processing (Sakhnovich, 2017).

5. Positivity, Tensor Products, and Operator System Theory

In operator system theory, positivity and tensor norm distinctions for TBT matrices encode intricate convex geometric and functional-analytic behavior. Key results include (Farenick, 2021):

• TBT matrices as $T_n \otimes T_m$ inherit two extremal operator system tensor products, $\otimes_\mathrm{min}$ (spatial) and $\otimes_\mathrm{max}$ (maximal).
• Min and max positivity are genuinely distinct: there exist explicit TBT matrices that are positive in the maximal but not in the minimal sense if $n, m \geq 2$.
• Criterion: A block Toeplitz matrix $X = [S_{i-j}]_{i,j=1}^m$ with each $S_k \in T_n$ is positive (in the spatial sense) iff, for every positive trigonometric polynomial matrix $F(z)$ of suitable degree, $\sum_{k=-n+1}^{n-1} S_k \otimes A_k$ is positive definite.
• The maximally entangled element $E_n = \sum_{k=-(n-1)}^{n-1} s_n^k \otimes s_n^{-k}$ is an extremal, entangled, and pure element in the corresponding positive cone, highlighting nontrivial entanglement phenomena and extremal ray structure for the set of positive TBT matrix-valued functions.

This operator-system perspective is closely connected to functional analysis, quantum information (due to entanglement structure), and multivariate moment problems.

6. Spectral and Asymptotic Analysis

Random TBT matrices and multilevel block Toeplitz matrices with (possibly) matrix-valued symbols exhibit universal and symbol-dependent spectral laws:

• For large-dimensional TBT matrices with i.i.d. or structured random blocks, the limiting spectral distribution (LSD), under various asymptotic regimes, converges to the Wigner semicircular law (when blocks are i.i.d.) or to the convolution of Toeplitz LSDs, depending on block structure (Basu et al., 2011).
• In multilevel block Toeplitz settings, spectra of preconditioned matrices $T_n^{-1}(g) T_n(f)$ cluster according to the symbol $g^{-1} f$ under suitable regularity and topological assumptions on the essential range of $g^{-1} f$. Provided $g$ is sectorial (spectrum bounded away from zero), and the essential range of $g^{-1} f$ does not disconnect $\mathbb{C}$ and has empty interior, eigenvalues globally distribute as prescribed by $g^{-1} f$ (Donatelli et al., 2014).
• This spectral clustering enables the design of optimal preconditioners, ensuring uniform GMRES iteration counts.

These spectral results have significant implications for numerical analysis, statistical mechanics, and the study of high-dimensional stochastic systems.

7. Generalizations and Applications

The algebraic properties and inversion theory for TBT matrices extend to block TBT matrices (i.e., higher-order multilevel Toeplitz structures) and triple-structured Toeplitz matrices. Construction of matrix-valued Weyl functions and reflection coefficients generalizes the recovery of the inverse from minimal interface data (Roitberg et al., 2020). Applications span multivariate convolution, time-frequency analysis, multidimensional system theory, and large-scale numerical simulations, with explicit operator-identity frameworks enabling efficient reconstruction and algorithmic implementation.

A key structural fact is that operator-identity techniques, low-rank block encodings, and symbol-based representations provide a unified approach to both the theory and computation with multidimensional Toeplitz structures, underlining deep connections with harmonic analysis and algebraic systems theory (Sakhnovich, 2017, Roitberg et al., 2020, Donatelli et al., 2014).

---

References:

(Sakhnovich, 2017): Inversion of the Toeplitz-block Toeplitz matrices and the structure of the corresponding inverse matrices (Farenick, 2021): The operator system of Toeplitz matrices (Basu et al., 2011): Limiting Spectral Distribution of Block Matrices with Toeplitz Block Structure (Donatelli et al., 2014): Spectral behavior of preconditioned non-Hermitian multilevel block Toeplitz matrices with matrix-valued symbol (Roitberg et al., 2020): On the inversion of the block double-structured and of the triple-structured Toeplitz matrices and on the corresponding reflection coefficients