Multi-Scale Factorization Approaches

• Multi-scale factorization is a method for decomposing high-dimensional data into hierarchical structures that capture both local and global dependencies.
• Techniques such as block partitioning, sparse rotations, and reinforcement learning optimize factor extraction and error control across multiple scales.
• This approach is applied in fields like brain imaging, network analysis, and generative modeling to enhance data compression, interpretability, and computational efficiency.

Multi-scale factorization refers to a set of methodologies for decomposing high-dimensional data, operators, or matrices according to multiple hierarchical or scale-separated latent structures. These factorizations generalize the classical low-rank or flat factor models to forms that explicitly reflect modularity, hierarchy, or resolution, and frequently arise in fields such as network analysis, brain imaging, matrix computations, generative modeling, and partial differential equations.

1. Theoretical Frameworks and Model Classes

At the core of multi-scale factorization are three principal classes of models:

1. Hierarchical Multi-scale Factor Analysis (MSFA): Models partition variables (e.g., nodes in a network) into regional clusters, each modeled via localized factor analysis. Higher-level structures aggregate these clusters, often using additional factors or blockwise correlation structures, capturing both local and global dependencies simultaneously (Ting et al., 2017).
2. Multiresolution Matrix Factorization (MMF): Applied primarily to symmetric or structured matrices, MMF represents a matrix as a product of sparse orthogonal rotations (hierarchically organized) and a small core-diagonal matrix. The rotations at each level address the fine-to-coarse organization of the matrix structure, yielding localized and scale-separated components (Kondor et al., 2015, Ithapu et al., 2017, Hy et al., 2021, Mudrakarta et al., 2019).
3. Multi-level/Multiscale Generative or Functional Models: These include deep matrix factorizations for hierarchical feature extraction (e.g., levels of networks in fMRI (Li et al., 2018)), multilevel matrix factor models with nested global/local factors (Zhang et al., 2023), and flow-based models that iteratively factorize and gaussianize portions of the latent space based on learned or data-driven importance (Das et al., 2019).

Other frameworks—for instance, multi-scale decompositions in PDE inverse problems (Zangerl et al., 2020) or Bayesian models that encode multiscale priors via algebraic constructions such as block-constant orthogonal factors (Xu et al., 2020)—further illustrate the breadth of multi-scale approaches.

2. Mathematical Structures and Factorization Forms

Multi-scale factorizations universally exploit hierarchical representation:

• Block partitioning: Data or matrices are recursively split into clusters, blocks, or scales (e.g., clusters $\to$ sub-networks $\to$ global network in MSFA (Ting et al., 2017)).
• Sparse Transformation Sequences: In MMF, one constructs

$$A \approx Q_1^\top Q_2^\top \cdots Q_L^\top H Q_L \cdots Q_2 Q_1,$$

where each $Q_\ell$ is a $k$-point sparse rotation (e.g., Givens), and $H$ is a core-diagonal matrix remaining after elimination of high-frequency or fine-scale structure (Kondor et al., 2015).

• Composite Loading Hierarchies: Deep semi-nonnegative matrix factorization encodes hierarchical relationships through products of nonnegative loading matrices at each scale, so that the effective loadings at a coarse scale are compounded from products of finer-scale loadings (Li et al., 2018).
• Additive/Hybrid Structures: Several models (e.g., Asymmetric MMF) deploy additive splitting of a matrix into symmetric and skew-symmetric or low-rank and multiscale components, with each component factorized according to its structural properties (Mudrakarta et al., 2019).

Multiscale frames in convolutional PDE settings employ spatial and temporal frames such that each filtered component corresponds to a distinct frequency scale, leading to exact factorization (Zangerl et al., 2020).

3. Computational Algorithms and Estimation Techniques

The estimation of multi-scale factorizations involves a combination of local analyses, greedy or global optimization, and—in recent work—learning-based control:

• Blockwise Principal Component Analysis (PCA): In MSFA, local factors are estimated per cluster via PCA minimizing local Frobenius loss, while regional and global connections are constructed by assembling these factors and their covariances (Ting et al., 2017).
• Greedy Hierarchical Rotations: Incremental and batch MMF employ greedy search over $k$-tuples to determine which variables to rotate or eliminate at each level, taking advantage of recursive error decomposition (Ithapu et al., 2017, Kondor et al., 2015).
• Reinforcement Learning and Manifold Optimization: Learnable MMF versions frame the selection of split coordinates as a Markov Decision Process (MDP) solved via reinforcement learning, while the rotation matrices are optimized on the Stiefel manifold using Riemannian gradient descent and geodesic flows, ensuring global orthogonality and well-conditioned bases (Hy et al., 2021).
• Data-dependent Gaussianization in Flows: In multi-scale generative flow architectures, early factorization of dimensions is determined by per-dimension likelihood contributions, leading to data-dependent, staged gaussianization and improved sample quality (Das et al., 2019).
• Bayesian Posterior Inference: In Bayesian multiscale models, partition/tree priors are replaced with generative algebraic constructions (e.g., binary-valued random matrices and Cholesky whitening), and parameter uncertainty is quantified via posterior sampling (e.g., Hamiltonian Monte Carlo) (Xu et al., 2020).
• Parallelization and Scalability: pMMF organizes hierarchical rotations in parallelizable blockwise clusters, enabling application to very large sparse matrices (Kondor et al., 2015).

4. Theoretical Guarantees and Model Selection

Rigorous error control and identifiability properties are associated with multi-scale factorization approaches:

• Consistency and Asymptotic Normality: Under high-dimensional conditions (e.g., $p\to\infty, n\to\infty, p/n\to 0$), regional factor loadings and factors possess root-$n$ convergence rates, and covariance estimators concentrate at rates explicit in cluster size and sample size (Ting et al., 2017). Multilevel matrix factor models formalize consistency for both global and local loading estimators (Zhang et al., 2023).
• Greedy Error Decomposition: In MMF, the overall factorization error decomposes as a sum of per-level residuals, with the greedy one-step-at-a-time selection yielding optimal local reduction (Ithapu et al., 2017).
• Model Selection: The number of factors at any scale can be chosen via explained variance thresholds or information criteria (e.g., Bai–Ng BIC for local factor numbers) (Ting et al., 2017), or via eigen-ratio statistics exploiting spectral gap properties (Zhang et al., 2023).
• Identifiability: Multi-level matrix factor models separate global and local factors via orthogonality and uncorrelatedness, and identifiability is ensured under orthogonality and non-degeneracy of covariance eigenvalues (Zhang et al., 2023).

5. Applications and Empirical Results

Multi-scale factorizations have demonstrated efficacy and interpretability in several domains:

• Brain Connectivity Networks: MSFA, deep semi-NMF, and Bayesian multiscale models disentangle modular and hierarchical organization in resting-state and task-based fMRI, revealing interpretable structures and improving prediction of functional activations (Ting et al., 2017, Li et al., 2018, Xu et al., 2020).
• Matrix Compression and Graph Signal Processing: pMMF and learnable MMF outperform classic low-rank methods (SVD, Nyström, CUR) in compressing large matrices, producing sparse and localized wavelet bases for graphs, and serving as efficient preconditioners for linear systems (Kondor et al., 2015, Hy et al., 2021, Mudrakarta et al., 2019).
• Generative Modeling: Data-driven multi-scale architectures in flows achieve superior likelihoods and higher-fidelity sample quality by allocating transformation capacity according to learned importance at multiple scales (Das et al., 2019).
• Inverse Problems in Imaging: Multiscale factorization of the wave equation allows for more accurate and compressible reconstructions in photoacoustic tomography, by transferring temporal sparsity into spatial sparsity at multiple scales (Zangerl et al., 2020).
• Finance and Econometrics: Multilevel matrix factor models provide a two-way dimension reduction for panel time series in asset returns or macroeconomic indicators, reducing parameter count and enhancing interpretability versus flat models (Zhang et al., 2023).

6. Extensions, Limitations, and Comparative Analysis

Several key limitations and possible extensions are recognized:

• Beyond Binary Trees: Recent Bayesian constructions avoid the combinatorial complexity of explicit hierarchical trees via algebraic generation of multiscale orthogonal factors, enabling full posterior uncertainty quantification across scales (Xu et al., 2020).
• Hybrid Approaches: Hybrid schemes (e.g., CUR + MMF) leverage complementary strengths of low-rank and multiscale representations, achieving lower error and better compression for matrices with both global and local structure (Mudrakarta et al., 2019).
• Non-symmetric and Tensor Data: Asymmetric generalizations of MMF and multilevel tensor factor models extend the reach of multi-scale factorization to directed networks and higher-order data arrays (Mudrakarta et al., 2019, Zhang et al., 2023).
• Computational Bottlenecks: While local block decomposition and parallelizable approaches alleviate scalability issues, large-scale high-order model selection and optimization remain computationally expensive, especially for dense data or when manifold optimization is required (Hy et al., 2021, Ithapu et al., 2017).
• Adaptivity vs. Design-Fixity: Some models fix the factorization mask during training (e.g., certain multi-scale flow architectures), while others adaptively learn the multi-scale partition, with or without reinforcement (Das et al., 2019, Hy et al., 2021).
• Statistical Assumptions: Strong factor pervasiveness or mixing assumptions may be required for consistency, and identification can break down for weak or highly correlated factor structures (Zhang et al., 2023, Ting et al., 2017).

7. Comparative Table of Select Multi-Scale Factorization Approaches

Approach	Data/Domain	Core Mechanism
MSFA (Ting et al., 2017)	High-dimensional networks	Regional PCA + global factor
MMF/pMMF (Kondor et al., 2015)	Symmetric matrices/graphs	Hierarchical sparse rotations
Deep semi-NMF (Li et al., 2018)	fMRI/factor networks	Layered NMF + group sparsity
Learnable MMF (Hy et al., 2021)	Graph Laplacian/matrices	RL index selection + manifold
Multilevel MF (Zhang et al., 2023)	Panel matrix time series	Two-level (global + local)
Multi-scale flows (Das et al., 2019)	Generative models/images	Data-driven latent splitting
Bayesian multiscale (Xu et al., 2020)	Brain connectivity	Tree-free block-orthogonal
AMMF (Mudrakarta et al., 2019)	Asymmetric matrices	Bi-orthogonal rotations/sp.

Each approach addresses distinct classes of structure and data regimes, and the optimal method is typically domain- and data-dependent.

---

In sum, multi-scale factorization provides a principled approach for extracting, representing, and interpreting hierarchical or modular structures in high-dimensional data, offering improvements in compression, interpretability, and predictive utility across a wide spectrum of applications (Ting et al., 2017, Kondor et al., 2015, Hy et al., 2021, Li et al., 2018, Zhang et al., 2023, Xu et al., 2020, Das et al., 2019, Zangerl et al., 2020, Mudrakarta et al., 2019).