Toeplitz-block Toeplitz Matrices

• Toeplitz-block Toeplitz matrices are defined by a block structure with inner Toeplitz patterns, capturing nested shift invariance.
• They utilize bivariate generating symbols to describe spectral properties and enable explicit inversion formulas.
• Applied in PDE discretizations and signal processing, these matrices support fast, FFT-based algorithms and reduced computational complexity.

A Toeplitz-block Toeplitz structure refers to matrices possessing two or more nested levels of Toeplitz symmetry, typically arising in block-structured linear algebra and operator theory. Fundamentally, such matrices combine the features of block Toeplitz matrices (constant block-diagonal patterns) with each block itself exhibiting Toeplitz structure (entries constant along sub-diagonals). Toeplitz-block Toeplitz matrices capture the algebraic and spectral complexity characteristic of multi-dimensional discretizations, array processing, structured system theory, and stochastic modeling.

1. Formal Definition and Symbolic Structure

Let $m,n \geq 1$ be integers. A Toeplitz-block Toeplitz (TBT) matrix $T$ is an $(mn) \times (mn)$ array such that $T$ is partitioned into $n \times n$ blocks, each of size $m \times m$. The $(j,k)$ block $T_{j-k}$ is itself Toeplitz:

• $T = \left[ T_{j-k} \right]_{j,k=1}^n$, where $T_\ell = \left[a^{(\ell)}_{r-s}\right]_{r,s=1}^m$
• Thus, $T$0

This induces two shift invariances: block-diagonals fixed by $T$1 and sub-diagonals within each block dictated by $T$2 (Sakhnovich, 2017). Beyond two levels, multilevel or multidimensional Toeplitz-block Toeplitz matrices generalize the concept to higher tensor product structures (Hon et al., 2024), with generating symbols $T$3 encoding both macro and micro Toeplitz arrangements (Furci et al., 2024).

2. Symbol, Spectral, and Asymptotic Properties

Matrices of TBT type admit symbolic representations via bivariate Laurent polynomials or matrix-valued generating functions. For TBT matrices of size $T$4, the symbol is

$T$5

The spectral distribution of large TBT matrices $T$6 is governed—in the Weyl sense—by the symbol: for eigenvalues (if Hermitian) or singular values (in general)

$T$7

for continuous functions $T$8 (Furci et al., 2024, Barakitis et al., 24 Jan 2025). This canonical GLT-based framework extends spectral analysis to arbitrary block-wise and multilevel Toeplitz sequences (Hon et al., 2024).

3. Inversion and Minimal Data

TBT matrices admit explicit inversion formulas generalizing those of 1-D Toeplitz. A key result is the minimality of inversion data: for a $T$9 TBT matrix $(mn) \times (mn)$0, the entire inverse $(mn) \times (mn)$1 can be recovered from a small $(mn) \times (mn)$2 block $(mn) \times (mn)$3 constructed via displacement identities (Sakhnovich, 2017). The procedure involves assembling a low-rank polynomial matrix $(mn) \times (mn)$4 whose determinant encodes invertibility and whose generating function yields all entries of $(mn) \times (mn)$5 as block-Toeplitz Toeplitz matrices,

$(mn) \times (mn)$6

Via Cauchy-type contour integrals on the symbol, one recovers entries and spectral characteristics (Roitberg et al., 2020). For multilevel structures, the extension involves recursive application of these identities, entwined with the algebraic structure induced by shift operators.

4. Maximum Rank and Structural Theory

Under “structural” conditions—where the pattern of nonzeros is fixed but each possible nonzero is a unique free parameter—the maximum (generic) rank of a block lower-triangular Toeplitz-block matrix equals its term rank (maximum matching size in the associated bipartite graph) (Reißig, 2013): $(mn) \times (mn)$7 This algebraic-combinatorial duality establishes that key rank properties are dictated by matching algorithms, with implications for structural controllability and observer design in systems theory.

5. Fast Algorithms and Practical Applications

Toeplitz-block Toeplitz structure substantially reduces computational and storage complexity in linear algebraic systems and array processing. For $(mn) \times (mn)$8-level TBT matrices, standard FFT-based multiplication scales poorly due to padding in each dimension. The split FFT (lazy embedding, eager projection) approach processes only minimally required branches, reducing computation and memory by asymptotic factors (Siron et al., 2024):

• Computational cost ratio $(mn) \times (mn)$9
• Peak memory usage factor $T$0

In expedited solution of PDEs or boundary control problems, block-Levinson and multigrid algorithms exploit TBT structure for $T$1 or $T$2 inversion (Belishev et al., 2021, Donatelli et al., 2019). Symbol-based $T$3 preconditioners constructed from the absolute value of the symbol ensure mesh-independent iterative convergence for large multilevel block Toeplitz systems (Hon et al., 2024, Barakitis et al., 24 Jan 2025).

6. Maximal Algebras, Spectral Bands, and Random Ensembles

Maximal commutative subalgebras of block Toeplitz (and hence TBT) matrices are classified via cocycle-type relations linking blocks through polynomial generators (the “generalized circulants”) (Khan, 2018). In the random matrix regime, the limiting spectral distribution of symmetric random TBT ensembles is determined by their block and entry links. In the joint large-size limit, blocks with i.i.d. entries approach the semicircle law, while Toeplitz blocks yield the “square” of Toeplitz spectral volumes (Basu et al., 2011).

7. Extensions, Open Questions, and Future Directions

Contemporary research generalizes TBT analysis to higher dimensions, general block structures, and PDE discretizations involving variable coefficients. Open problems include:

• Robust spectral analysis for irrational block size ratios and low-rank Hankel corrections (Furci et al., 2024)
• Construction of optimal preconditioners and FFT-based approximants with preserved symbol distribution
• Characterization of joint eigenvalue/singular value distributions for multilevel block modifications

Ongoing numerical and analytical research continues to refine theoretical understanding and algorithmic exploitation of Toeplitz-block Toeplitz structures in both deterministic and stochastic domains.

---

Select References

• Rei{\ss}ig, "On the maximum rank of Toeplitz block matrices of blocks of a given pattern" (Reißig, 2013)
• Sakhnovich, "Inversion of the Toeplitz-block Toeplitz matrices ..." (Sakhnovich, 2017)
• Basu et al., "Limiting Spectral Distribution of Block Matrices with Toeplitz Block Structure" (Basu et al., 2011)
• Furci et al., "Block structured matrix-sequences and their spectral and singular value canonical distributions: a general theory" (Furci et al., 2024)
• Siron & Molesky, "A Split Fast Fourier Transform Algorithm for Block Toeplitz Matrix-Vector Multiplication" (Siron et al., 2024)
• Khan, "A Family of Maximal Algebras of Block Toeplitz matrices" (Khan, 2018)
• Ferrari, Furci, Serra-Capizzano, "Symbol-based multilevel block τ preconditioners ..." (Hon et al., 2024)
• Belishev & Karazeeva, "Toeplitz matrices in the Boundary Control method" (Belishev et al., 2021)
• Bini et al., "Computing the Exponential of Large Block-Triangular Block-Toeplitz Matrices ..." (Bini et al., 2015)
• Donatelli et al., "Multigrid methods for block-Toeplitz linear systems ..." (Donatelli et al., 2019)
• Roitberg & Sakhnovich, "On the inversion of the block double-structured ..." (Roitberg et al., 2020)
• Furci, Adriani, Serra-Capizzano, "Blocking structures, approximation, and preconditioning" (Barakitis et al., 24 Jan 2025)
• Kammerer et al., "An Array Decomposition Method for Finite Arrays with Electrically Connected Elements ..." (Åkerstedt et al., 5 Jun 2025)