Block Toeplitz with Toeplitz Blocks Matrices

• Block Toeplitz with Toeplitz Blocks matrices are multi-level structured arrays characterized by translation invariance both at the block and within-block levels, essential in signal processing and PDE discretization.
• Their spectral properties, derived from both random and deterministic frameworks, facilitate fast solvers and asymptotic analyses through techniques such as split FFT and circulant embeddings.
• Explicit inversion and preconditioning algorithms, which leverage operator identities and low-rank displacement structures, significantly reduce computational complexity in high-dimensional systems.

A Block Toeplitz with Toeplitz Blocks (BTTB) matrix is a multi-level patterned matrix whose structure arises naturally in a wide range of applications, especially in multidimensional signal processing, statistical modeling, and the discretization of partial differential operators on rectangular grids. BTTB matrices exhibit a nested form of translation-invariance: the entries depend only on the differences of their respective indices, first at the block level and then within each block. This structure allows for significant algebraic, spectral, and algorithmic properties directly relevant to fast solvers, spectral analysis, and operator system theory.

1. Rigorous Definition and Structural Properties

Formally, for positive integers $m$, $n$, a BTTB matrix $T$ of size $(mn)\times(mn)$ is defined by

$T = [ T_{i-j} ]_{i,j=1}^m,$

where each $T_k$ is itself an $n\times n$ Toeplitz matrix,

$T_k = [ t^{(k)}_{u-v} ]_{u,v=1}^n.$

Equivalently, $T$ is the block matrix: $$T = \begin{bmatrix} T_0 & T_{-1} & \dots & T_{-(m-1)} \ T_1 & T_0 & \dots & T_{-(m-2)} \ \vdots & \ddots & \ddots & \vdots \ T_{m-1} & \dots & T_1 & T_0 \end{bmatrix},$$ with each $n$0 Toeplitz, i.e., $n$1 [$n$2].

This concept generalizes naturally to higher dimensions and block sizes, yielding matrices structured as tensor products or multi-level extensions, with the general $n$3-level BTTB matrix being indexed by differences along each dimension: $n$4 [$n$5].

2. Spectral Theory, Asymptotics, and Limiting Distributions

The spectral properties of random and deterministic BTTB matrices are determined by their underlying translational invariance. For random BTTB matrices with symmetric block structure as in

$n$6

where $n$7 are $n$8 Toeplitz or random matrices, the limiting spectral distribution (LSD) after scaling by $n$9 can be established in multiple regimes:

• For fixed $T$0 and $T$1 or vice versa, the empirical spectral distribution converges almost surely to a deterministic limit.
• If both $T$2 then the ESD converges to the semicircular law if the blocks are i.i.d.
• When blocks themselves are Toeplitz, the limit is described by the "squared-Toeplitz" moment law (Bryc–Dembo–Jiang, Hammond–Miller): no closed-form density is known, but the even moments are

$T$3

where $T$4 is the weight for each pair-matched word $T$5 in a single symmetric Toeplitz ensemble [$T$6].

For deterministic BTTB matrix-sequences generated by matrix-valued symbols $T$7, under scaling assumptions on the blocks, the singular/eigenvalue distribution asymptotically coincides with the distribution of $T$8 as $T$9 ranges over $(mn)\times(mn)$0, per a Szegő-type theorem. When $(mn)\times(mn)$1 is Hermitian, the eigenvalue distribution of $(mn)\times(mn)$2 converges to that of the function $(mn)\times(mn)$3 [$(mn)\times(mn)$4].

3. Algebraic Structure and Operator Systems

The BTTB class forms a vector space isomorphic to a tensor product of Toeplitz spaces but is not an algebra under ordinary multiplication unless further restrictions apply. The operator system of $(mn)\times(mn)$5 Toeplitz matrices, $(mn)\times(mn)$6, admits a complete order isomorphism with the dual of the space of trigonometric polynomials of degree $(mn)\times(mn)$7. Block Toeplitz matrices with Toeplitz blocks arise naturally as tensor products in the operator system framework: $(mn)\times(mn)$8 is, under the minimal operator system tensor product, exactly the vector space of BTTB matrices with entries in an operator system $(mn)\times(mn)$9 [$T = [ T_{i-j} ]_{i,j=1}^m,$0].

Distinct notions of positivity (e.g., min- vs. max-positivity) emerge for BTTB matrices, with the min cone associated to positivity under commuting Toeplitz representations and the max cone tied to decompositions into positive elementary tensors. These cones differ when the blocks are themselves Toeplitz, exemplified by the maximally entangled Toeplitz matrix $T = [ T_{i-j} ]_{i,j=1}^m,$1 generating an extremal ray in the corresponding cone [$T = [ T_{i-j} ]_{i,j=1}^m,$2].

4. Inversion, Explicit Algorithms, and Computational Complexity

Explicit formulas for the inverse of BTTB matrices have been developed via operator-identity methods. These rely on multivariate generalizations of the Gohberg–Semencul/Levinson approach:

• The inverse is again BTTB.
• The entire matrix is determined by two low-rank displacement generators of size $T = [ T_{i-j} ]_{i,j=1}^m,$3 (or analogs in higher dimensions), which can be extracted algorithmically.
• The key step is the construction of a block-matrix-valued rational pencil $T = [ T_{i-j} ]_{i,j=1}^m,$4 whose inverse yields generating polynomials encoding $T = [ T_{i-j} ]_{i,j=1}^m,$5; this reduces inversion complexity from $T = [ T_{i-j} ]_{i,j=1}^m,$6 to $T = [ T_{i-j} ]_{i,j=1}^m,$7 plus lower order terms.
• The minimal boundary data required is significantly less than the full matrix: typically $T = [ T_{i-j} ]_{i,j=1}^m,$8 parameters vs $T = [ T_{i-j} ]_{i,j=1}^m,$9 entries [$T_k$0, $T_k$1].

For ARMA-type block Toeplitz structures, finite explicit formulas for $T_k$2 can be derived in terms of the Fourier coefficients of the outer phase function attached to the spectral density. In such cases, linear-time algorithms yielding the $T_k$3 block inverse are possible, bypassing cubic complexity [$T_k$4].

5. Fast Algorithms and Preconditioning

BTTB matrices can be exploited for algorithmic acceleration via Fourier techniques:

• Classical circulant embedding maps the $T_k$5-level BTTB structure to a $T_k$6-dimensional circulant, enabling $T_k$7-time matrix-vector products using FFTs. However, this increases memory and compute by a factor $T_k$8.
• The split FFT algorithm ("lazy embedding, eager projection") reduces the arithmetic and memory overhead. At each level, a recursive decomposition splits the action into branches indexed by parity, with phase-shifted FFTs; a merge step projects back onto the Toeplitz support, ensuring correctness while reducing storage.
• The computational cost is reduced by a factor $T_k$9, and the peak memory is similarly compressed.
• These algorithms are naturally parallelizable, with each branch or sub-tree of the split recursion assignable to separate devices or threads [$n\times n$0].

Preconditioning strategies based on simplifying the block structure (e.g., replacing Toeplitz blocks by circulant surrogates) yield preconditioners $n\times n$1 whose spectra cluster at $n\times n$2. For Hermitian positive-definite BTTB matrices, preconditioned Krylov solvers achieve $n\times n$3 or $n\times n$4 iteration counts, resulting in nearly linear complexity even for large-scale systems [$n\times n$5].

6. Maximal Commutative Subalgebras and Simultaneous Diagonalization

The classification of maximal commutative subalgebras within the space of BTTB matrices reveals intricate algebraic structures. For a maximal commutative subalgebra $n\times n$6, the space of block Toeplitz matrices with blocks in $n\times n$7 can be organized into algebras $n\times n$8, parameterized by pairs $n\times n$9 with $T_k = [ t^{(k)}_{u-v} ]_{u,v=1}^n.$0: $T_k = [ t^{(k)}_{u-v} ]_{u,v=1}^n.$1 Each $T_k = [ t^{(k)}_{u-v} ]_{u,v=1}^n.$2 admits explicit two-stage simultaneous diagonalization by first diagonalizing $T_k = [ t^{(k)}_{u-v} ]_{u,v=1}^n.$3, then applying the standard $T_k = [ t^{(k)}_{u-v} ]_{u,v=1}^n.$4 DFT to each block, resulting in a full 2D discrete Fourier-type transform. The dimension of such an algebra is $T_k = [ t^{(k)}_{u-v} ]_{u,v=1}^n.$5 [$T_k = [ t^{(k)}_{u-v} ]_{u,v=1}^n.$6].

7. Applications, Special Cases, and Theoretical Generalizations

BTTB matrices encompass several important special classes:

• Ordinary Toeplitz matrices ($T_k = [ t^{(k)}_{u-v} ]_{u,v=1}^n.$7) and block Toeplitz matrices ($T_k = [ t^{(k)}_{u-v} ]_{u,v=1}^n.$8) are both special cases.
• The theory extends directly to three or more levels, yielding multidimensional BTTB (e.g., $T_k = [ t^{(k)}_{u-v} ]_{u,v=1}^n.$9-level) structures for multidimensional convolution operators or covariance matrices.
• The BTTB framework naturally appears in the operator systems approach to Toeplitz matrices, facilitating duality principles, complete positivity, and criteria for separability in the context of signal processing and quantum information theory [$T$0].
• Applications include fast solvers for PDEs discretized on grids, spectral estimation for multidimensional random fields, preconditioning of linear systems, and explicit invertibility criteria for space-time models.

All results are conditioned on structural assumptions—block sizes, independence or moment conditions for random ensembles, symbol regularity for deterministic sequences—ensuring that the translation-invariance is preserved and that the key algebraic manipulations and limit theorems apply [$T$1, $T$2, $T$3, $T$4, $T$5, $T$6, $T$7, $T$8].