Source-linked AI summary
A Tensor Neural Network Method for High-Order Homogenization of Locally Periodic Elliptic Problems
Huaijia Zhang, Haochen Liu, Xia Ji, Hehu Xie
TL;DR
Locally periodic elliptic problems require high-order correctors that depend on the slow variable, complicating homogenization beyond the purely periodic case. The paper derives a computable recursive expansion and boundary-layer-free H1 estimate, then solves the corrector hierarchy with TNNs and deterministic quadrature. Experiments show accurate correctors and errors consistent with the proved or formally predicted high-order behavior.
Problem
Locally periodic coefficients make cell problems and high-order correctors depend on the slow variable, complicating high-order homogenization beyond the purely periodic case.
Method
The paper combines a recursive high-order two-scale hierarchy and boundary-layer-free convergence analysis with a TNN solver using deterministic tensorized quadrature for high-dimensional corrector computations.
Results
Experiments show accurate high-order correctors; H1 semi-norm errors are consistent with the proved estimate, while point-normalized L2 errors show systematic reduction as successive correctors are added.
Takeaways & Limitations
The framework provides a computable high-order treatment of locally periodic corrector hierarchies while retaining deterministic integration for the tensor-structured numerical realization.
Takeaways & Limitations
The convergence theorem is restricted to boundary-layer-free settings, and standard multidimensional Dirichlet problems require boundary-layer correctors.
Abstract
from arXiv · showhide
We develop a high-order tensor neural network (TNN) method for locally periodic elliptic multiscale problems of the form $-\nabla\cdot(A(x,x/\varepsilon)\nabla u_\varepsilon)=f$. Because the coefficient depends on both the slow variable $x$ and the fast periodic variable $y=x/\varepsilon$, the high-order cell problems and macroscopic corrector equations are more involved than in the classical case $A=A(y)$, and the correctors depend parametrically on $x$. We derive a computable high-order two-scale expansion and prove an $H^1$ convergence estimate for the partial expansion in boundary-layer-free settings, including periodic domains and ideal boundary-matching configurations. The proof uses the recursive compatibility structure of the corrector hierarchy and a zero-mean oscillation estimate in $H^{-1}$. We then construct a TNN framework for the high-dimensional corrector problems. Its tensor-product structure permits deterministic one-dimensional quadrature for the cell problems, homogenized coefficients, macroscopic source terms, and loss functions, avoiding Monte Carlo integration error. The numerical realization assumes that the coefficient entries and assembled data admit finite or controlled tensor-product representations; this computational assumption is separate from the general matrix-valued coefficient class used in the analysis. Experiments with scalar locally periodic coefficients show accurate high-order correctors. The $H^1$ semi-norm errors are consistent with the proved estimate, while point-normalized $L^2$ errors display the nominal high-order behavior predicted by the formal expansion.
1 Introduction
The paper addresses high-order homogenization for locally periodic elliptic problems, where slow-variable dependence makes correctors parameter-dependent and complicates the hierarchy. It derives a computable expansion, proves a boundary-layer-free estimate, and develops a tensor neural network solver with deterministic quadrature.
- Motivation: Locally periodic coefficients depend on both slow and fast variables, making cell functions, homogenized coefficients, and high-order correctors depend parametrically on the slow variable.This setting is more delicate than purely periodic homogenization.
- Contributions: The paper derives a recursive high-order asymptotic expansion consisting of parameterized cell problems, macroscopic corrector equations, and oscillatory correction terms.The formulation is designed to be directly computable.
- Contributions: The authors prove a high-order H1 convergence estimate for partial expansions in boundary-layer-free settings using recursive compatibility, a residual identity, and an ε-order gain in H−1.The supported settings include periodic domains and ideal boundary-matching configurations.
- Contributions: The TNN method uses tensor-structured coefficients and data with deterministic tensorized quadrature for losses, cell averages, homogenized coefficients, and high-order source terms.This avoids Monte Carlo sampling for the high-dimensional integrals.
- Numerical evidence: Numerical experiments verify accurate computed correctors, with H1 semi-norm tests compared against the proved estimate and point-normalized L2 results consistent with formal expansion orders.The experiments also show that high-accuracy integration helps preserve the asymptotic hierarchy.
2 High-Order Two-Scale Expansion
The paper constructs a recursive high-order two-scale expansion for locally periodic elliptic problems, combining parameterized cell problems with macroscopic corrector equations. Its convergence analysis applies in boundary-layer-free settings and supports the algorithmic construction of successive correctors.
- Model and assumptions: The coefficient is assumed Y-periodic in y, symmetric, and uniformly elliptic, with bounds λ|ξ|^2 ≤ ξ^T A(x,y)ξ ≤ Λ|ξ|^2.These assumptions define the admissible coefficient class for the expansion and convergence analysis.
- Recursive expansion: The two-scale expansion substitutes periodic correctors into the differential operator and produces a recursive cascade of cell and macroscopic equations.The construction separates fast-variable oscillatory components from slow-variable macroscopic corrections.
- Recursive expansion: The first-order cell functions χ_i(x,y) solve parameterized periodic problems, and their solvability yields the homogenized equation for u0.The correctors and homogenized coefficients depend on the slow variable x because the coefficient is locally periodic.
- Compatibility and normalization: The recursive cell problems are solvable because their right-hand sides have zero Y-average, while zero-mean normalization selects representatives modulo additive constants.Numerical point normalization is also allowed when shifts are consistently absorbed into macroscopic correctors.
- Convergence setting: The high-order H1 convergence theorem applies when no boundary layer is generated, including the flat torus and ideal boundary-matching configurations.Standard multidimensional Dirichlet problems generally require boundary-layer correctors, which are outside this analysis.
- Convergence setting: The full expansion also satisfies an H1 estimate under the theorem’s assumptions and an additional uniform H1 bound on the highest macroscopic corrector.The proof framework uses the recursive hierarchy, compatible error equations, and a Poincare inequality in the chosen energy space.
3 Tensor Neural Network Method
The method represents the locally periodic corrector hierarchy with tensor-product neural networks, hard constraints, sequential least-squares solves, and deterministic tensorized quadrature. Its computational realization requires tensor-product representations for coefficient entries and assembled source terms, while supporting multiple corrector representations and higher-order extensions.
- 3.1 Brief Review of TNN Architecture: TNN trial functions are tensor products of one-dimensional subnetworks for macroscopic or two-scale variables.Two-scale functions use separate subnetworks for slow and fast variables, with sine activations and automatic differentiation in the reported experiments.
- 3.1 Brief Review of TNN Architecture: Periodicity, gauges, and endpoint conditions are imposed through hard architectural transformations rather than penalty terms.Reference-point normalization preserves fast-variable periodicity, while zero-mean and endpoint constraints are imposed through explicit transformations.
- 3.2 Deterministic Tensorized Quadrature: Deterministic tensorized quadrature reduces high-dimensional inner products in losses, cell averages, homogenized coefficients, and source terms to one-dimensional contractions.The reported experiments divide each one-dimensional interval into 200 uniform subintervals with four Gauss points per subinterval.
- 3.3 TNN Method for High-Order Asymptotic Expansion: Each corrector equation is linear after assembling its right-hand side from previously computed terms, enabling sequential least-squares solves.Scaling coefficients can be eliminated analytically for positive-definite systems; constrained semidefinite systems use rank-revealing factorization or a Moore–Penrose pseudoinverse.
- 3.3 TNN Method for High-Order Asymptotic Expansion: Tensorized computation assumes finite or accurately truncated tensor-product forms for coefficient entries and assembled source terms.This is a computational assumption rather than an analytical requirement; arbitrary non-tensor data require preliminary tensor approximation, whose error must be added to other errors.
- 3.3 TNN Method for High-Order Asymptotic Expansion: The implementation represents homogenized and macroscopic functions with single-variable TNNs, while cell and oscillatory correctors use two-scale TNNs.An equivalent cell-equation formulation for δu2 provides a lower-rank representation and reduces subsequent assembly cost in the reported two-dimensional experiment.
4 Numerical Results
The TNN experiments reproduce the predicted high-order approximation hierarchy in one- and two-dimensional locally periodic problems, with H1 behavior aligned with the boundary-layer-free theory and L2 behavior treated as empirical. Non-product coefficients make the cell functions slow-variable dependent and harder to resolve.
- Error interpretation: H1 semi-norm errors are compared with the proved boundary-layer-free estimate, whereas point-normalized L2 slopes are empirical because the L2 norm is not gauge invariant.The distinction reflects the different treatment of additive representatives.
- One-dimensional problem: The one-dimensional Dirichlet test uses a special exact boundary-matching configuration with ε = 1/N, so every truncated expansion satisfies the boundary values without a boundary layer.The oscillatory correctors are endpoint-normalized and macroscopic components use homogeneous Dirichlet data.
- One-dimensional problem: Adding correctors reduces oscillations, with u0, u0+εu1, and u0+εu1+ε2u2 showing approximately first-, second-, and third-order L2 convergence, respectively.The third-order partial expansion further reduces the error.
- One-dimensional problem: The one-dimensional H1 fits give rates approximately 0.983 for U∗1 and 1.987 for U∗2, confirming the corrector hierarchy in this exact-matching setting.These results are not a boundary-layer-free validation for general multidimensional Dirichlet problems.
- Two-dimensional product-form problem: For the two-dimensional periodic product-form problem, L2 data show approximately first-, second-, and third-order convergence for u0, U1, and U2, while U∗3 further decreases the error.The H1 data show the expected improvement from the partial expansions.
- Two-dimensional non-product problem: The non-product problem has cell functions depending on both slow and fast variables despite low-rank input data, increasing the difficulty of resolving higher-order oscillatory correctors.The reported losses for ˆu2 and ˆu3 are noticeably higher than in the product-form case.
- Two-dimensional non-product problem: For the non-product problem, L2 rates are approximately 0.977 for u0, 2.02 for U1, 2.72 for U2, and 2.56 for U∗3, with U∗3 reducing error at every listed scale.The two higher-order curves remain below the nominal third-order rate on the reported scales.
- Two-dimensional non-product problem: In the non-product H1 results, U∗1 converges at approximately first order and U∗2 reduces error on all reported scales, although the finest scale deviates from ideal second-order behavior.A fit using the first four scales gives an approximate rate of 1.51 for U∗2, compared with 1.13 over all five scales.
5 Conclusions
The paper develops a high-order TNN framework with a recursive two-scale hierarchy, deterministic tensorized quadrature, and a boundary-layer-free H1 convergence estimate. Experiments show systematic L2 error reduction as correctors are added, while the three-dimensional visualization remains qualitative and boundary-layer corrections are left for future work.
- The framework combines recursive two-scale correctors, strong-form TNN least-squares problems, deterministic tensorized quadrature, and a common training protocol.The method targets locally periodic elliptic problems and implements the full hierarchy numerically.
- The one-dimensional and two-dimensional periodic experiments show systematic L2 error reduction as successive correctors are added.The non-product two-dimensional coefficient also tests cell problems with slow-variable dependence, although higher-order observed rates are below nominal orders on the reported scales.
- The three-dimensional example visualizes the learned expansion but is not a convergence test because no fine-scale reference solution is available.The cross-sections display the total approximation and first- and second-order coefficients across decreasing ε values.
- Boundary layer correctors for standard Dirichlet problems remain a future direction.The current convergence analysis is restricted to boundary-layer-free settings.