Periodic Decomposition Block (PDB)

• PDB is a module that decomposes functions, tensors, or data by separating them into periodic and antiperiodic components using explicit projection operators.
• In tensor analysis, PDB underlies the t-product and t-SVD by exploiting block-circulant structures that are efficiently diagonalized via the FFT.
• Applied in time series forecasting, PDB transforms 1D signals into 2D representations, extracting dominant periodic patterns for robust multi-scale analysis.

A Periodic Decomposition Block (PDB) is a structural or algorithmic module designed to decompose signals, functions, or data tensors according to underlying periodicities. PDBs appear in diverse mathematical and machine learning contexts: (1) in harmonic analysis as operators that split periodic function spaces into direct sums of subspaces with smaller period or antiperiodic structure (Yadeta, 2022); (2) in tensor analysis as the algebraic backbone of the tensor t-product and t-SVD, where block-circulant structure underlies efficient multi-linear decompositions (Molavi et al., 2023); and (3) as an architectural block in neural forecasting pipelines, enabling extraction of periodic patterns by transforming time series into 2D tensor representations in the frequency domain (Nematirad et al., 31 Mar 2025). The unifying thread is the explicit separation and manipulation of periodic and antiperiodic (or, more generally, periodic in lower dimension) components in functional or data spaces, enabling both theoretical analysis and practical algorithmic improvement.

1. Functional Decomposition via the PDB in Periodic Function Spaces

Let $F$ denote the vector space of all real-valued functions on $\mathbb{R}$ with the shift operator $E^h f(x) = f(x+h)$. For any period $p > 0$, several relevant subspaces are defined:

• $P_p = \{ f \in F: E^p f = f \}$ is the subspace of all $p$-periodic functions.
• $P_{p/2} = \{ f \in F: E^{p/2} f = f \}$, the $p/2$-periodic subspace within $P_p$.
• $AP_{p/2} = \{ g \in F: E^{p/2} g = -g \}$, the subspace of $p$-periodic but $p/2$-antiperiodic functions.

The core result is the direct sum decomposition:

$P_p = P_{p/2} \oplus AP_{p/2}$

with explicit projections $P$ and $Q$ given by

$$P f = \frac{1}{2}(f + E^{p/2} f), \quad Q f = \frac{1}{2}(f - E^{p/2} f)$$

ensuring $f = Pf + Qf$ uniquely with orthogonality of subspaces ($P_{p/2} \cap AP_{p/2} = \{0\}$). This process can be iteratively applied to $P_{p/2}$, generating a recursive cascade of periodic and antiperiodic subspaces. For arbitrary $k \geq 1$,

$$F^{(k)} = \frac{1}{2}\big(F^{(k-1)} + E^{p/2^k} F^{(k-1)}\big) \in P_{p/2^k}$$

$$G^{(k)} = \frac{1}{2}\big(F^{(k-1)} - E^{p/2^k} F^{(k-1)}\big) \in AP_{p/2^k}$$

yielding the decomposition

$$f = F^{(n)} + \sum_{k=1}^n G^{(k)}, \qquad F^{(n)} \in P_{p/2^n}$$

and, under $\lim_{n\to\infty}\|F^{(n)}\|_\infty=0$, the convergent antiperiodic series

$f(x) = \sum_{k=1}^\infty G^{(k)}(x)$

This structural splitting is visualized via a lattice called a “periodicity diagram,” organizing subspaces according to period and antiperiod structure (as for $p=12$ in (Yadeta, 2022)).

2. PDB as the Algebraic Foundation of Tensor t-Product and t-SVD

In tensor analysis, PDB refers to the periodic block structure underlying the tensor-tensor product (t-product) and the t-SVD:

Given a third-order tensor $$\mathcal{A} \in \mathbb{R}^{n_1 \times \ell \times n_3}$$ and $$\mathcal{X} \in \mathbb{R}^{\ell \times n_2 \times n_3}$$:

• The “unfold” operation stacks frontal slices of $\mathcal{A}$ into a column block matrix.
• The block-circulant embedding forms $\mathrm{circ}(\mathrm{unfold}(\mathcal{A}))$ by cycling blocks according to periodic boundary conditions.
• The t-product $\star_t$ is then realized as

$$\mathcal{Y} = \mathcal{A} \star_t \mathcal{X} = \mathrm{fold}(\mathrm{circ}(\mathrm{unfold}(\mathcal{A})) \cdot \mathrm{unfold}(\mathcal{X}))$$

This algebra is efficiently diagonalized by the Discrete Fourier Transform (DFT) on the third mode, translating block convolution into independent matrix multiplications per frequency slice. The t-SVD follows: for any $\mathcal{A}$, there exist $\mathcal{U}, \mathcal{S}, \mathcal{V}$ (orthogonal and f-diagonal in the t-product sense) such that

$$\mathcal{A} = \mathcal{U} \star_t \mathcal{S} \star_t \mathcal{V}^\top$$

This periodic block structure forms the “Periodic Decomposition Block” in tensor algebra, unifying SVD-like decompositions for higher-order tensors (Molavi et al., 2023).

3. PDB in Machine Time Series Models: Times2D and Frequency-based 2D Lifting

The Times2D model employs a PDB as its front-end transform, explicitly extracting periodicities from 1D temporal signals by transposing selected dominant frequencies into a 2D “period vs. cycle” representation (Nematirad et al., 31 Mar 2025).

Given input $X_{1D} \in \mathbb{R}^{B \times S \times N}$ (batch, time, channels):

1. FFT is applied along the time axis, yielding amplitude spectra $A(b,f,n)$.
2. Top $k$ frequency indices $f_1,...,f_k$ are selected based on magnitude averaging.
3. For each frequency $f_i$, period $p_i = S / f_i$ is computed; the signal is reshaped into

$$X_{2D}^{(i)} \in \mathbb{R}^{B \times p_i \times f_i \times N}$$

mapping time series into a grid where each row/column corresponds to within-period/cross-period structure.

Subsequent processing—2D convolution, positional encoding, multi-head self-attention, flattening, and projection—yields multi-scale summary representations. The outputs from all $k$ period decompositions are summed to provide the final PDB output. This architecture allows simultaneous extraction of temporally local and global periodic dependencies, a property that empirical benchmarks demonstrate is essential for strong time series forecasting accuracy (Nematirad et al., 31 Mar 2025).

4. Structural and Algorithmic Properties

The PDB decomposition, whether in function space, tensor algebra, or neural pipeline, provides:

• Direct sum (orthogonal) splitting: The decomposition into smaller-period/antiperiodic subspaces in function analysis is unique, with explicit projection operators (Yadeta, 2022).
• Hierarchical refinement: Iterative splitting (the “PDB-cascade”) produces finer-scale periodic/antiperiodic components, giving rise to convergent antiperiodic series under mild norm decay.
• Block-circulant embeddings: The tensor algebraic PDB (t-product) acts as a periodic multi-linear convolution, equating to FFT-diagonalizable block-circulant multiplication (Molavi et al., 2023).
• Fast computation: In tensors, the t-SVD leverages the FFT, leading to computational cost proportional to FFT and batch SVDs across frequency bands.
• Expressive transformation: In learning tasks, PDB enables a lossless or near-lossless lift from 1D signals to 2D tensors encoding periodicity, allowing complex feature extractors (e.g., Conv2D or attention) to operate across both within-period and cross-period axes (Nematirad et al., 31 Mar 2025).

5. Comparative and Contextual Remarks

In tensor decomposition, the PDB (periodic/FFT) and its alternatives (e.g., reflective/Discrete Cosine Transform, DCT) are distinguished by:

Product	Boundary Condition	Block Structure	Diagonalization
$\star_t$	periodic	block-circulant	FFT (complex)
$\star_c$	reflective	block–Toeplitz+Hankel	DCT (real)

• $\star_t$ admits straightforward implementation and inherits SVD-like conditioning, but introduces complex values via the DFT; $\star_c$ achieves better edge handling in signals and real-number arithmetic.
• Both algebraic settings formalize tensor-tensor products as convolutional block-operations, with the PDB uniquely formalizing the periodic (FFT) variant (Molavi et al., 2023).

In functional analysis, the PDB’s recursive splitting is visualized via periodicity diagrams, revealing a lattice structure of periods and antiperiods for all divisors of a fundamental period; this has combinatorial and operator-theoretic significance (Yadeta, 2022).

In time series learning applications, the PDB (as formalized in Times2D) provides a learnable but interpretable lifting that empirically enables robust forecasting under both short-term and long-term horizon constraints, especially when combined with complementary blocks capturing non-periodic transient features (e.g., derivative heatmaps) (Nematirad et al., 31 Mar 2025).

6. Applications and Empirical Performance

PDBs are foundational for:

• Decomposition and reconstruction formulas in periodic function analysis, providing convergence guarantees for antiperiodic expansions under decay conditions (Yadeta, 2022).
• Efficient, scalable tensor decompositions (t-SVD, truncated t-SVD) in high-dimensional image analysis, classification, and clustering; as a baseline, PDB/t-SVD matches or exceeds quality and efficiency, with further improvements possible via boundary condition specialization (e.g., DCT for reflective BC) (Molavi et al., 2023).
• Time series forecasting, where explicit exploitation of periodic structure enables state-of-the-art results by flexibly combining periodic-feature maps with alternative representations of sharp change (Nematirad et al., 31 Mar 2025).

A plausible implication is that further generalizations of PDB architecture, including extensions to irregular time series, multiresolution adaptive boundary handling, and combinations with derivative- and event-driven primitives, may become increasingly relevant in both mathematical and applied contexts.