Low-Rank Structural Couplings

• Low-Rank Structural Couplings are models that integrate low-rank approximations of global dependencies with additional structural constraints to capture local variations.
• They employ methodologies such as alternating minimization, randomized solvers, and Riemannian optimization to efficiently address high-dimensional statistical estimation and simulation.
• These models have practical applications in covariance estimation, surrogate modeling, optimal transport, and medical imaging, balancing interpretability with computational scalability.

Low-rank structural couplings describe models, algorithms, and parameterizations in which a low-rank structure (often reflecting a small set of global factors or principal components) is integrally coupled with additional structural constraints or local corrections. Such couplings pervade modern approaches to high-dimensional statistical estimation, numerical simulation, surrogate modeling, inverse problems, and optimal transport. They offer interpretable inductive bias, computational scalability, and the ability to capture both global dependencies and essential local or domain-specific structure.

1. Mathematical Foundations of Low-Rank Structural Couplings

Low-rank structural couplings arise when a mathematical object—matrix, tensor, or function—must satisfy both a low-rank constraint and one or more structural constraints. The canonical example is the low-rank plus diagonal (LRPD) model for symmetric matrices: given $\Sigma\in\mathbb{R}^{n\times n}$, one seeks

$$\Sigma \approx LL^\top + D,\quad L\in\mathbb{R}^{n\times r},\ D=\operatorname{diag}(d_1,\ldots,d_n).$$

Here, the low-rank term ($LL^\top$) captures global correlation or shared factors, while the diagonal ($D$) corrects local variance. The LRPD decomposition generalizes to block-diagonal corrections and underpins a large class of factorial and spatiotemporal models (Yeon et al., 18 Dec 2025).

Structured low-rank matrix and tensor learning extends this coupling:

• For a matrix $X\in\mathbb{R}^{d\times T}$ subjected to a constraint $A(X)=b$, the optimization reads

$$\min_{X}\, L(X) + C\|X\|_*^2\quad\text{s.t. }A(X)=b,$$

where the low-rank and constraint terms are handled via a latent variable decomposition, separating the low-rank factor from the structural component (Jawanpuria et al., 2017).

• In tensors, a coupled factorization might produce

$$W = \sum_{k=1}^K \lambda_k (Z + A^*(s)) \times_k \Theta_k,$$

imposing low unfolding rank per mode and, for example, a Hankel or nonnegativity structure (Naram et al., 2023).

In the context of optimal transport, low-rank couplings refer to restricting the transport plan $P\in\mathbb{R}_+^{n\times m}$ to nonnegative rank (transport rank) at most $r$, via explicit parametrizations:

$$\Sigma \approx LL^\top + D,\quad L\in\mathbb{R}^{n\times r},\ D=\operatorname{diag}(d_1,\ldots,d_n).$$0

with affine marginal constraints and inner shared marginals $$\Sigma \approx LL^\top + D,\quad L\in\mathbb{R}^{n\times r},\ D=\operatorname{diag}(d_1,\ldots,d_n).$$1 (Scetbon et al., 2021, Halmos et al., 2024).

2. Algorithmic Approaches and Structural Optimization

Low-rank structural couplings frequently require specialized algorithms to optimize over the constrained feasible set. Typical strategies include:

• Alternating Minimization: For LRPD, the Alt algorithm alternates between extracting the best low-rank approximation (via spectral methods) and updating the diagonal to match the residual variances. Monotonic error decrease, local contraction, and convergence under mild spectral gap conditions are provable (Yeon et al., 18 Dec 2025).
• Randomized Solvers: For massive matrices where explicit storage or full spectral decomposition is infeasible, randomized Nyström sketching and stochastic estimation of the diagonal via Rademacher queries (Diag++) drastically reduce computational cost while preserving provable approximation guarantees (Yeon et al., 18 Dec 2025).
• Riemannian Optimization: For matrices or tensors under joint low-rank and structural constraints, reformulating the problem on a quotient or product manifold (e.g., spectrahedron, sphere products) enables efficient use of Riemannian gradient, trust-region, and conjugate-gradient methods. These approaches utilize explicit projections, retractions, and duality-gap monitoring to ensure correctness and stationarity (Jawanpuria et al., 2017, Naram et al., 2023, Yang et al., 23 Jan 2025).
• Latent and Pairwise Factorizations in High-Dimensional OT and Surrogate Modeling: In high-dimensional optimal transport and function surrogates, the "latent coupling" factorization decouples the transport or surrogate mapping into a sequence of smaller OT problems or into parameter-wise and pairwise low-rank interactions (e.g., PLRNet for EM simulation) (Halmos et al., 2024, Sun et al., 20 Mar 2026).

3. Domains of Application and Model Structures

Low-rank structural coupling is foundational in:

• Covariance Estimation: LRPD models allow simultaneous capture of global correlations and localized idiosyncratic variances or sector/block structure, crucial in finance, spatial statistics, and kernel learning (Yeon et al., 18 Dec 2025).
• Surrogate Modeling in Physics: In electromagnetic simulation and mechanics, high-dimensional response functions admit low-rank tensor decompositions (Tucker, TT, TR, and pairwise/plenary couplings). This dramatically reduces sample complexity, parameter count, and supports accurate surrogate model construction under severe sampling constraints (Sun et al., 20 Mar 2026, Karmakar et al., 24 Sep 2025).
• Contact Mechanics and Fluid-Structure Interaction: Reduced-order models for complex mechanics (including nonlinearity and contact constraints) exploit structural decomposition, but local phenomena may require over-complete dictionaries or non-linear manifold interpolants (e.g., dynamic time warping enrichment) when strict low-rankness fails (Kollepara, 2024, Benner et al., 2020).
• Optimal Transport and Gromov–Wasserstein Problems: Low-rank couplings in OT and GW enable both statistical regularization and computational scalability, supporting scalable domain alignment, batch correction, and clustering in genomics and imaging (Scetbon et al., 2021, Halmos et al., 2024, Scetbon et al., 2021).
• Structured Causal Inference: In settings with interference and unknown or complex dependencies, assuming a low-rank potential outcome structure supports robust, unbiased causal effect estimation, circumventing the infeasibility of balancing on all joint assignments (Sengupta et al., 15 Dec 2025).
• Medical Imaging and Shape Analysis: SLoRD demonstrates how low-rank contour descriptors lead to anatomically plausible, consistent segmentations with pronounced gains in segmentation accuracy for 3D structural data (You et al., 2024).

4. Structural Decomposition and Manifold Geometry

Solving optimization problems with coupled low-rank and structural constraints often entails working on embedded or quotient manifolds. Several key ideas recur:

• Space-Decoupling Product Manifolds: Introducing an auxiliary Grassmann variable or product coordinates (e.g., for encoding rank via $$\Sigma \approx LL^\top + D,\quad L\in\mathbb{R}^{n\times r},\ D=\operatorname{diag}(d_1,\ldots,d_n).$$2 with $$\Sigma \approx LL^\top + D,\quad L\in\mathbb{R}^{n\times r},\ D=\operatorname{diag}(d_1,\ldots,d_n).$$3) decouples low-rank and structural constraints, shifting the problem to a smooth manifold where explicit Riemannian geometry enables efficient algorithmic design (Yang et al., 23 Jan 2025).
• Partial Dualization and Decoupling: For trace-norm or nuclear-norm penalized models with constraints (linear, Hankel, nonnegativity), partial dualization separates the low-rank factor from the structural term in the Lagrangian or dual variable, clarifying feasible directions and providing efficient updates (Jawanpuria et al., 2017, Naram et al., 2023).
• Projection, Retraction, and Duality-Gap Monitoring: Essential geometric operations (ambient-to-tangent projections, manifold-specific retractions, and explicit Hessian actions) enable first- and second-order methods to approach critical points while maintaining feasibility and facilitating convergence checks via duality gap computations (Naram et al., 2023, Jawanpuria et al., 2017, Yang et al., 23 Jan 2025).

5. Theoretical Guarantees and Empirical Performance

Rigorous analysis underpins the efficacy of low-rank structural couplings:

• Monotonicity and Convergence: Alt-type alternating minimization in LRPD decompositions features monotone residual decrease and local contraction properties with only weak assumptions (eigenvalue gaps or non-axis-aligned structure) (Yeon et al., 18 Dec 2025).
• Stationarity and Global Optimality: Riemannian optimization in quotient and product manifold settings guarantees convergence to first- or second-order criticality; rank-deficiency at stationarity indicates global optimality in convex programs (Jawanpuria et al., 2017, Naram et al., 2023).
• Complexity Bounds: For randomized and block-coordinate methods, per-iteration complexity is reduced from cubic in ambient size to linear or near-linear in the target rank and sketch parameters, enabling application to previously intractable large-scale problems (Yeon et al., 18 Dec 2025, Halmos et al., 2024, Scetbon et al., 2021, Benner et al., 2020).
• Data Efficiency and Robustness: In multi-physics surrogates (e.g., xLRA for microstructural elasticity), physically informed low-rank CP-type decompositions achieve high accuracy ($$\Sigma \approx LL^\top + D,\quad L\in\mathbb{R}^{n\times r},\ D=\operatorname{diag}(d_1,\ldots,d_n).$$4) with as little as 4–5\% of the underlying dataset, six orders of magnitude fewer FLOPs compared to neural-operator baselines (Karmakar et al., 24 Sep 2025).
• Empirical Dominance: In EM surrogate tasks, the pairwise low-rank network (PLRNet) achieves uniformly lower mean relative errors with drastically lower parameter footprint compared to Tucker, TT, and dense MLP baselines (Sun et al., 20 Mar 2026). In batch correction and clustering, low-rank OT coupling yields lower misclassification and faster computation than entropic or plug-in analogues (Forrow et al., 2018, Scetbon et al., 2021, Halmos et al., 2024).

6. Limitations, Challenges, and Extensions

Low-rank structural coupling is not a panacea. In contact mechanics, sharp locality and moving interfaces cause a breakdown of linear separability, necessitating dictionaries or manifold-aware interpolation (e.g., dynamic time warping) when pure low-rank models underperform (Kollepara, 2024). Similar caveats apply in high-dimensional GW/OT, where over-aggressive rank truncation can miss essential structure.

Extensions include:

• Multi-marginal OT via further factorizations or block-coordinate schemes (Halmos et al., 2024).
• Unbalanced and semi-relaxed OT via relaxed marginal constraints (Halmos et al., 2024).
• Nonlinear or approximate low-rank surrogates via hybrid explicit-implicit or dictionary-based approaches (Kollepara, 2024).
• Model selection heuristics and data-driven diagnostics for rank determination, balancing feasibility, bias-variance tradeoff, and collinearity (Sengupta et al., 15 Dec 2025).

Low-rank structural couplings thus constitute a central principle for extracting, representing, and learning global–local structure in large-scale scientific computation and data analysis. They unify disparate algorithmic and statistical advances, underpinned by explicit geometric and variational structure, rigorous analysis, and demonstrated empirical superiority across domains (Yeon et al., 18 Dec 2025, Naram et al., 2023, Jawanpuria et al., 2017, Halmos et al., 2024, Scetbon et al., 2021, Yang et al., 23 Jan 2025, Sengupta et al., 15 Dec 2025, Benner et al., 2020, Kollepara, 2024, Sun et al., 20 Mar 2026, Karmakar et al., 24 Sep 2025, Forrow et al., 2018, Scetbon et al., 2021, You et al., 2024).