StructuralDecompose: Decomposition Frameworks

• StructuralDecompose is a family of formal methods that decompose signals, matrices, graphs, and other structures into interpretable components.
• It underpins diverse applications, from multiphase powder X-ray diffraction and 3D shape modeling to algebraic, categorical, and time-series analyses.
• Each method enforces structure via consistency conditions, canonical factorization, or learned latent subspaces, enabling precise reconstruction and downstream identification.

StructuralDecompose denotes a family of decomposition formalisms rather than a single method. In the cited literature, the term is used for prior-free set decomposition of multiphase powder X-ray diffraction, linear factorization of whole-shape embeddings into semantic part subspaces, discrete dictionary-based decomposition of Tensor Product Representation components, grammar-based matrix model selection, canonical decomposition of non-manifold simplicial complexes, modular time-series decomposition, and categorical theories of structured decompositions and temporal narratives (Gao et al., 7 May 2026, Dubrovina et al., 2019, Park et al., 2024, Grosse et al., 2012, Morando, 2019, Sunny, 6 Oct 2025, Bumpus et al., 2022). Across these usages, the decomposed object may be a signal, a latent code, a graph- or poset-indexed object, a simplicial complex, a matrix polynomial, or a logical model; the output may be a set of source-resolved components, a family of subspaces, a decomposition tree, a collection of canonical factors, or a set of parameterized submodels.

1. Foundational formulations

A recurrent formulation treats decomposition as the recovery of structured components together with a consistency condition that reconstructs the original object. In multiphase PXRD, XDecomposer takes a mixed intensity profile $x \in \mathbb{R}_+^L$ and models it by the linear superposition

$$y(\theta) = \sum_{i=1}^{N} w_i y_i(\theta) + \epsilon(\theta),$$

with unknown phase number, unknown coefficients $w_i \ge 0$, and mixture consistency $x(\theta) \approx \sum_k \hat y_k(\theta)$; the output is an unordered set of phase-resolved components $\{\hat y_k\}$ with slot activity gating and a fixed maximum $K_{\max}$ (Gao et al., 7 May 2026). In latent 3D shape modeling, structural decomposition is the linear map from a whole-shape code $v \in \mathbb{R}^n$ to part codes $v_i = P_i v$ using learned projections satisfying

$$P_i^2 = P_i,\qquad P_i P_j = 0 \ (i \neq j),\qquad P_1+\cdots+P_K = I,$$

so that $V = V_1 \oplus \cdots \oplus V_K$ and composition becomes $$y(\theta) = \sum_{i=1}^{N} w_i y_i(\theta) + \epsilon(\theta),$$0 (Dubrovina et al., 2019). In TPR-based neuro-symbolic models, structural decomposition means recovering role vectors $$y(\theta) = \sum_{i=1}^{N} w_i y_i(\theta) + \epsilon(\theta),$$1, filler vectors $$y(\theta) = \sum_{i=1}^{N} w_i y_i(\theta) + \epsilon(\theta),$$2, and unbinding vectors from a superposed representation

$$y(\theta) = \sum_{i=1}^{N} w_i y_i(\theta) + \epsilon(\theta),$$3

with decoding via $$y(\theta) = \sum_{i=1}^{N} w_i y_i(\theta) + \epsilon(\theta),$$4 when $$y(\theta) = \sum_{i=1}^{N} w_i y_i(\theta) + \epsilon(\theta),$$5; the D3 layer replaces continuous component emitters with discrete key-value dictionaries assigned to roles, fillers, and unbinding operators (Park et al., 2024).

These formulations differ in domain, but each fixes a structured target space before learning or inference begins. In some settings that target space is a set with unknown cardinality; in others it is a direct-sum latent space, a tensor-product factorization, or a categorical diagram. A plausible implication is that “structural decomposition” is best understood not as one algorithmic primitive but as a design pattern in which the admissible output family is constrained strongly enough to support inversion, reconstruction, or downstream identification.

2. Physical and signal-domain realizations

In structural characterization, XDecomposer casts multiphase PXRD analysis as prior-free blind source separation with set prediction and physics-consistent reconstruction. Its architecture combines a hierarchical analyzer $$y(\theta) = \sum_{i=1}^{N} w_i y_i(\theta) + \epsilon(\theta),$$6, a Transformer encoder $$y(\theta) = \sum_{i=1}^{N} w_i y_i(\theta) + \epsilon(\theta),$$7 pretrained by masked reconstruction on single-phase patterns, and a phase-query-driven latent decomposition module. Learnable queries $$y(\theta) = \sum_{i=1}^{N} w_i y_i(\theta) + \epsilon(\theta),$$8 cross-attend to the global context, produce activity probabilities $$y(\theta) = \sum_{i=1}^{N} w_i y_i(\theta) + \epsilon(\theta),$$9, and modulate latent features by FiLM parameters before a decoder predicts masks $w_i \ge 0$0 with $w_i \ge 0$1. The formulation builds in non-negativity and boundedness, $w_i \ge 0$2, and enforces mixture consistency, geometric shape constraints, and PIT/Hungarian alignment under unknown phase number (Gao et al., 7 May 2026).

In periodic KS-DFT, structural decomposition appears as an atom-resolved partition of the one-particle density matrix and the total energy. The theory starts from Bloch sums of crystalline Gaussian-type orbitals,

$w_i \ge 0$3

constructs localized linear combinations in a supercell at $w_i \ge 0$4, and assigns localized orbitals to atoms through PM or IBO localization with IAO-based population measures. The resulting atomic $w_i \ge 0$5-RDMs $w_i \ge 0$6 decompose the periodic KS-DFT energy into atom-resolved kinetic, electron–nuclear, Hartree, exchange–correlation, and Ewald structural terms, with cohesive energy written as $w_i \ge 0$7 (Zamok et al., 2024).

In time-series analysis, the R package "StructuralDecompose" (Sunny, 6 Oct 2025) implements a modular pipeline that separates changepoint detection, anomaly detection, smoothing, and decomposition rather than treating decomposition as a monolithic procedure. The supported method choices are exposed explicitly through breakpoint = "strucchange"|"pelt"|"binseg"|"cusum", anomaly = "rolling_median"|"zscore"|"mad"|"rolling_stats", smoother = "loess"|"moving_average"|"spline", and decomposition = "stl_additive"|"stl_multiplicative". The result object persists intermediate artifacts such as breakpoints, anomaly scores, trend, seasonal component, and residual, so decomposition is inseparable from a staged preprocessing and segmentation semantics.

3. Learned latent and symbolic decomposition

The Decomposer–Composer architecture for semantic 3D shape modeling learns a factorized embedding space in which semantic structure is encoded by part-specific subspaces. A 3D CNN encoder maps a voxel occupancy grid $w_i \ge 0$8 to a whole-shape latent code $w_i \ge 0$9, the projections $x(\theta) \approx \sum_k \hat y_k(\theta)$0 produce canonical part embeddings $x(\theta) \approx \sum_k \hat y_k(\theta)$1, and a shared decoder reconstructs normalized part volumes $x(\theta) \approx \sum_k \hat y_k(\theta)$2. A 3D spatial transformer network then predicts affine parameters $x(\theta) \approx \sum_k \hat y_k(\theta)$3, applies trilinear resampling, and assembles a full output

$x(\theta) \approx \sum_k \hat y_k(\theta)$4

The training objective combines partition-of-identity regularization, part reconstruction BCE, transformation loss, and cycle consistency (Dubrovina et al., 2019). Here structural decomposition is not merely segmentation; it is a linear factorization of representation space with explicit geometric reassembly.

In D3, decomposition is imposed at the level of symbolic component generation. For each component $x(\theta) \approx \sum_k \hat y_k(\theta)$5, a query $x(\theta) \approx \sum_k \hat y_k(\theta)$6 is matched against a discrete dictionary $x(\theta) \approx \sum_k \hat y_k(\theta)$7, the top-$x(\theta) \approx \sum_k \hat y_k(\theta)$8 keys are selected, and the resulting code is

$x(\theta) \approx \sum_k \hat y_k(\theta)$9

followed by a residual projection and final linear map. Roles and unbinding operators can share a dictionary to enforce TPR duality constraints, and the layer acts as a drop-in replacement in FWM, TPR-RNN, and Linear Transformer backbones (Park et al., 2024). The paper’s central claim is not simply improved accuracy but improved decomposition of unseen combinatorial inputs into structured role–filler representations.

A third learned formulation, "Compositional Structures in Neural Embedding and Interaction Decompositions" (Trager et al., 2024), decomposes an embedding $\{\hat y_k\}$0 into pure interaction components

$\{\hat y_k\}$1

For softmax models

$\{\hat y_k\}$2

the paper proves that conditional independence $\{\hat y_k\}$3 holds if and only if $\{\hat y_k\}$4 whenever $\{\hat y_k\}$5 intersects both $\{\hat y_k\}$6 and $\{\hat y_k\}$7. Structural decomposition in this setting is thus a necessary-and-sufficient linear criterion for conditional independence, not merely a descriptive embedding analysis.

4. Categorical and graph-theoretic decomposition systems

A general categorical theory appears in "Structured Decompositions: Structural and Algorithmic Compositionality" (Bumpus et al., 2022). There a $\{\hat y_k\}$8-valued structured decomposition of shape $\{\hat y_k\}$9 is a functor

$K_{\max}$0

where $K_{\max}$1 is the Grothendieck construction of the shape graph. The associated category $K_{\max}$2 collects such decompositions, tame decompositions are those with only monomorphisms, and semantics are assigned by a spine functor $K_{\max}$3 followed by colimit. Width is then defined via completions into objects of subcategories $K_{\max}$4, recovering treewidth, complemented treewidth, layered treewidth, Carmesin’s graph decomposition width, and a fine-grained variant of $K_{\max}$5-treewidth.

"Decomposing time-varying data into simple pieces: structured decompositions of narratives" (Bumpus et al., 11 Jul 2026) lifts this static theory to persistent narratives. A persistent narrative is a presheaf $K_{\max}$6 on a discrete time category satisfying a pullback condition for interval decompositions, and under the paper’s hypotheses any static spined sd-category yields a temporal sd-category on $K_{\max}$7. This produces temporal analogues of ordinary tree-width, complemented tree-width, and tree-independence number, with widths defined pointwise over time and aggregated through the lifted spine.

In Petri-net theory, structural decomposition is realized by wiring decompositions in the algebra of nets with boundaries. A net $K_{\max}$8 carries left and right boundary ports, a contention relation, and a labeled transition system on markings with labels $K_{\max}$9 describing boundary activity. Sequential composition $v \in \mathbb{R}^n$0 synchronizes along a shared boundary, tensor product $v \in \mathbb{R}^n$1 forms parallel composition, and decomposition width bounds both leaf size and intermediate interface size in the wiring expression. The paper gives constant-width families such as clique nets $v \in \mathbb{R}^n$2 with decomposition width $v \in \mathbb{R}^n$3 and subset nets $v \in \mathbb{R}^n$4 with decomposition width $v \in \mathbb{R}^n$5, while conjecturing that grid nets require growing width (Rathke et al., 2013).

5. Topological and geometric decomposition theories

For non-manifold simplicial complexes, structural decomposition is defined combinatorially through the totally exploded complex $v \in \mathbb{R}^n$6 and gluing instructions. The canonical decomposition is

$v \in \mathbb{R}^n$7

where $v \in \mathbb{R}^n$8 glues exactly along manifold $v \in \mathbb{R}^n$9-faces. The result is unique up to isomorphism, computable in linear time, and yields connected components that are Initial-Quasi-Manifolds. The same work introduces the Extended Winged Data structure, whose first layer stores TV/TT data for each IQM component and whose second layer stores inter-component stitching information across singular loci (Morando, 2019). Structural decomposition here is canonical rather than learned.

In the theory of Białynicki–Birula decompositions, a smooth complete $v_i = P_i v$0-variety with finite fixed locus decomposes into attracting cells

$v_i = P_i v$1

and repelling cells $v_i = P_i v$2. The paper characterizes filterability through the orbit graph $v_i = P_i v$3: the positive decomposition is filterable if and only if $v_i = P_i v$4 has no directed cycles, and stratification is characterized equivalently by closure containment, strict increase of cell dimensions along directed edges, or the numerical condition $v_i = P_i v$5 whenever $v_i = P_i v$6. Both filterability and stratification are invariant under reversing the $v_i = P_i v$7-action (Gonzales et al., 30 Apr 2026). This is a structural decomposition in the classical algebro-geometric sense of cell stratification.

In flag-complex theory, decomposition becomes enumerative. A flag complex is the clique complex of its $v_i = P_i v$8-skeleton, and the paper studies Boolean decompositions of complexes $v_i = P_i v$9 satisfying $$P_i^2 = P_i,\qquad P_i P_j = 0 \ (i \neq j),\qquad P_1+\cdots+P_K = I,$$0 for CM or vertex decomposable flag complexes. Edge subdivision preserves generalized Boolean decomposition, Tchebyshev triangulation satisfies

$$P_i^2 = P_i,\qquad P_i P_j = 0 \ (i \neq j),\qquad P_1+\cdots+P_K = I,$$1

and the mirroring/Danzer complex obeys

$$P_i^2 = P_i,\qquad P_i P_j = 0 \ (i \neq j),\qquad P_1+\cdots+P_K = I,$$2

These formulas connect $$P_i^2 = P_i,\qquad P_i P_j = 0 \ (i \neq j),\qquad P_1+\cdots+P_K = I,$$3-, $$P_i^2 = P_i,\qquad P_i P_j = 0 \ (i \neq j),\qquad P_1+\cdots+P_K = I,$$4-, and $$P_i^2 = P_i,\qquad P_i P_j = 0 \ (i \neq j),\qquad P_1+\cdots+P_K = I,$$5-data to explicit geometric transformations and to positivity properties lying between unimodality and real-rootedness (Park, 2024).

6. Algebraic, logical, and homological decompositions

Several algebraic literatures use StructuralDecompose for recursive model reduction or canonical factorization. In grammar-based matrix decomposition, a real-valued data matrix $$P_i^2 = P_i,\qquad P_i P_j = 0 \ (i \neq j),\qquad P_1+\cdots+P_K = I,$$6 is explained by an algebraic expression generated from primitives $$P_i^2 = P_i,\qquad P_i P_j = 0 \ (i \neq j),\qquad P_1+\cdots+P_K = I,$$7, $$P_i^2 = P_i,\qquad P_i P_j = 0 \ (i \neq j),\qquad P_1+\cdots+P_K = I,$$8, $$P_i^2 = P_i,\qquad P_i P_j = 0 \ (i \neq j),\qquad P_1+\cdots+P_K = I,$$9, $V = V_1 \oplus \cdots \oplus V_K$0 and productions such as $V = V_1 \oplus \cdots \oplus V_K$1, $V = V_1 \oplus \cdots \oplus V_K$2, $V = V_1 \oplus \cdots \oplus V_K$3, and $V = V_1 \oplus \cdots \oplus V_K$4. Model selection proceeds by greedy search over the grammar, production-specific initialization, and predictive likelihood scoring, so decomposition is simultaneously generative and structural (Grosse et al., 2012).

In CSPs with global constraints, structural decomposition centers on hypergraph width only after augmenting it with sparse intersections and partial assignment checking. For a constraint $V = V_1 \oplus \cdots \oplus V_K$5, the intersection variables are $V = V_1 \oplus \cdots \oplus V_K$6, and the central result is that classes with sparse intersections over a PAC catalogue reduce in polynomial time to classic CSP instances with the same solutions. The same logic extends to subproblem decompositions and weighted CSPs (Thorstensen, 2015).

In linear SEM identification, recursive decomposition operates on c-components. A coefficient set $V = V_1 \oplus \cdots \oplus V_K$7 with head $V = V_1 \oplus \cdots \oplus V_K$8 is identifiable if a $V = V_1 \oplus \cdots \oplus V_K$9-HT-admissible witness set exists, yielding a linear system

$$y(\theta) = \sum_{i=1}^{N} w_i y_i(\theta) + \epsilon(\theta),$$00

The paper then recursively marginalizes descendant sets, decomposes into c-component submodels, and reuses newly identified edges to enlarge the allowed witness set, thereby increasing identification power and subsuming the nonparametric identification algorithm for direct effects (Chen, 2015).

In modular noetherian right $$y(\theta) = \sum_{i=1}^{N} w_i y_i(\theta) + \epsilon(\theta),$$01-groups with strong order unit, the strong order interval $$y(\theta) = \sum_{i=1}^{N} w_i y_i(\theta) + \epsilon(\theta),$$02 decomposes into directly indecomposable factors indexed by central dual atoms $$y(\theta) = \sum_{i=1}^{N} w_i y_i(\theta) + \epsilon(\theta),$$03, and the global group decomposes as

$$y(\theta) = \sum_{i=1}^{N} w_i y_i(\theta) + \epsilon(\theta),$$04

These beam factors are built from frozen powers $$y(\theta) = \sum_{i=1}^{N} w_i y_i(\theta) + \epsilon(\theta),$$05, and beams of dimension $$y(\theta) = \sum_{i=1}^{N} w_i y_i(\theta) + \epsilon(\theta),$$06 admit coordinatization by $$y(\theta) = \sum_{i=1}^{N} w_i y_i(\theta) + \epsilon(\theta),$$07-lattices in $$y(\theta) = \sum_{i=1}^{N} w_i y_i(\theta) + \epsilon(\theta),$$08 for a noncommutative discrete valuation field $$y(\theta) = \sum_{i=1}^{N} w_i y_i(\theta) + \epsilon(\theta),$$09 with valuation ring $$y(\theta) = \sum_{i=1}^{N} w_i y_i(\theta) + \epsilon(\theta),$$10 (Dietzel, 2023). In poset type decomposition, a distinguished $$y(\theta) = \sum_{i=1}^{N} w_i y_i(\theta) + \epsilon(\theta),$$11 plays the role of center, $$y(\theta) = \sum_{i=1}^{N} w_i y_i(\theta) + \epsilon(\theta),$$12-complete subsets define the type class, and the canonical splitter is either

$$y(\theta) = \sum_{i=1}^{N} w_i y_i(\theta) + \epsilon(\theta),$$13

or

$$y(\theta) = \sum_{i=1}^{N} w_i y_i(\theta) + \epsilon(\theta),$$14

with orthogonality relations characterizing the complementary type (Bice, 2014).

Two further algebraic examples are explicitly spectral and homological. For $$y(\theta) = \sum_{i=1}^{N} w_i y_i(\theta) + \epsilon(\theta),$$15-structured matrix polynomials

$$y(\theta) = \sum_{i=1}^{N} w_i y_i(\theta) + \epsilon(\theta),$$16

the paper gives a unified spectral decomposition in terms of a standard pair $$y(\theta) = \sum_{i=1}^{N} w_i y_i(\theta) + \epsilon(\theta),$$17 and a parameter matrix $$y(\theta) = \sum_{i=1}^{N} w_i y_i(\theta) + \epsilon(\theta),$$18 satisfying $$y(\theta) = \sum_{i=1}^{N} w_i y_i(\theta) + \epsilon(\theta),$$19 and $$y(\theta) = \sum_{i=1}^{N} w_i y_i(\theta) + \epsilon(\theta),$$20, together with coefficient recovery formulas such as

$$y(\theta) = \sum_{i=1}^{N} w_i y_i(\theta) + \epsilon(\theta),$$21

This machinery is then used for inverse eigenvalue construction and eigenvalue embedding with no spill-over (Zhao et al., 5 Jun 2026). In monomial resolutions, structural decomposition expresses the multigraded Betti numbers of a monomial ideal $$y(\theta) = \sum_{i=1}^{N} w_i y_i(\theta) + \epsilon(\theta),$$22 in terms of dominant and purely nondominant ideals. The first decomposition has the form

$$y(\theta) = \sum_{i=1}^{N} w_i y_i(\theta) + \epsilon(\theta),$$23

and iterating it yields a second decomposition into dominant and purely nondominant summands, with applications to minimal resolutions and projective dimension (Alesandroni, 2017).

Taken together, these usages show that StructuralDecompose functions as a domain-specific name for exact splitting rules, recursive factorization procedures, or width-controlled compositional schemes. The recurring outputs are not interchangeable, but they are uniformly structural: phase-resolved signals, part embeddings, symbolic components, canonical topological factors, graph-shaped diagrams, central algebraic parts, or recursively simplified submodels.