Geometry-Adaptive Harmonic Basis

• Geometry-adaptive harmonic bases are locally orthonormal functions defined through eigen-decomposition of geometry-sensitive operators, accurately capturing local geometric features.
• They enable adaptive shrinkage and sparse representation, yielding optimal approximations in PDE discretizations and efficient operator compression across high-dimensional applications.
• Their construction leverages methods such as Jacobian eigendecomposition, Laplace–Beltrami eigenfunctions, and localized SVD, making them essential in image denoising, machine learning, and computational geometry.

Geometry-adaptive harmonic bases are locally orthonormal sets of functions explicitly constructed to adapt to the underlying geometry of a domain, signal, or dataset. They arise as eigenfunctions of Laplacian-type operators or as singular vectors of locally defined restriction/extension operators. Such bases are essential for optimal approximation, operator compression, and adaptive regularization in high-dimensional signal processing, PDE discretizations, and machine learning. Their construction, spectral properties, and effectiveness are tightly linked to local geometry (level sets, inclusion boundaries, surface curvature), yielding sparse and nearly optimal representations across diverse application domains.

1. Mathematical Definition and Operator-Theoretic Characterization

A geometry-adaptive harmonic basis (GAHB) is a local orthonormal basis $\{e_k(y)\}$ for a vector space (e.g., image patches, finite element neighborhoods, surface parameterizations), determined by eigenvalue problems associated with geometric operators.

• In score-based diffusion models and MMSE denoisers, the GAHB is defined pointwise: given a (noisy) datum $y\in\mathbb{R}^d$ and denoiser $f(y)$, one forms the Jacobian $J(y)=\nabla f(y)$, diagonalizes $J(y)=\sum_{k=1}^d\lambda_k(y)e_k(y)e_k(y)^T$, with $\lambda_1\ge \cdots \ge \lambda_d \ge 0$, and takes $\{e_k(y)\}$ as the basis functions (Kadkhodaie et al., 2023).
• For elliptic PDEs, GAHB arises as harmonic characteristic functions $\chi_{D_m}$ solving $\Delta\chi_{D_m}=0$ in the background, with support and boundary conditions reflecting the geometry of inclusions (Poveda et al., 2014).
• On smooth surfaces, basis functions $\{\phi_i\}$ are defined as Laplace–Beltrami eigenfunctions $-\Delta_S\phi_i=\lambda_i\phi_i$; these are globally harmonic and adapt to the manifold geometry (Alsnayyan et al., 2021).
• For edges in multiscale finite element methods, edge basis functions are formed by SVD of the local restriction operator acting on $a$-harmonic functions in oversampling neighborhoods, yielding nearly exponential approximation rates and geometry-adaptive construction (Chen et al., 2020).

2. Spectral Shrinkage, Approximation, and Optimality

Central to the utility of GAHBs is their role in adaptive shrinkage and sparse expansion:

• In image denoising, any input $y$ can be represented $y=\sum_k c_k(y)e_k(y)$, with coefficients $c_k(y)=\langle y,e_k(y)\rangle$. The denoiser output is $f(y)=\sum_k \lambda_k(y)c_k(y)e_k(y)$, i.e., shrinkage along each eigenvector direction by geometry-adaptive factors $\lambda_k(y)\in[0,1]$ (Kadkhodaie et al., 2023). For the MMSE oracle, $$\lambda_k^{oracle}(x)=\frac{\langle x,e_k\rangle^2}{\langle x,e_k\rangle^2+\sigma^2}$$.
• In high-contrast elliptic problems, harmonic characteristic functions $\chi_{D_m}$ capture leading-order solution behavior, while localized truncation $\chi_{D_m}^\delta$ enable efficient approximation with rapidly decaying error in the truncation parameter $\delta$ (Poveda et al., 2014).
• For surface representations, Laplace–Beltrami eigenfunctions have frequency ordering: low-index functions represent global modes while high-index encode localized geometry (Alsnayyan et al., 2021).
• Edge-adaptive harmonic bases, constructed via SVD of restriction operators, yield exponential decay of singular values $\lambda_{e,m}$ and guarantee nearly exponential convergence of energy-norm error $$\|u^h-u_{H,m}\|_{H^1_a}\le C'\exp(-m^{1/(d+1)-\epsilon})\|f\|_{L^2}$$ (Chen et al., 2020).

3. Geometric Adaptivity and Empirical Patterns

Geometry-adaptive harmonic bases inherently track local geometric features:

• On images, leading eigenvectors $e_k(y)$ oscillate tangentially along contours and exhibit smooth, low-frequency structure in homogeneous regions. The spatial frequency grows with $k$, and shrinkage factors $\lambda_k(y)$ decay rapidly, ensuring sparsity (Kadkhodaie et al., 2023).
• For domains with inclusions, harmonic characteristics are supported in inclusion-centered neighborhoods, with truncation ensuring fast decay outside the inclusion (Poveda et al., 2014).
• On surfaces, Laplace–Beltrami harmonics adapt to curvature and global shape, forming an orthonormal “Fourier” basis whose support scales inversely with the harmonic index (Alsnayyan et al., 2021).
• In parameterization schemes, THB-spline basis construction leverages local refinement and truncation so that only active functions in regions of geometric interest (e.g., boundary protrusions) are included, and projection aids in preserving parametric quality (Hinz et al., 2020).

4. Construction Algorithms and Computational Strategies

Geometry-adaptive harmonic bases are constructed via localized or global eigendecomposition, operator SVD, or goal-oriented adaptive refinement:

Approach	Key Steps	Reference
DNN Jacobian Eigendecomposition	Form $J(y)=\nabla f(y)$, diagonalize, use $\{e_k(y)\}$	(Kadkhodaie et al., 2023)
Harmonic Characteristic Functions	Solve $\Delta\chi_{D_m}=0$ with inclusion-adapted BCs	(Poveda et al., 2014)
Surface Laplace–Beltrami	Assemble mass/stiffness, solve $S c_i=\lambda_i M c_i$	(Alsnayyan et al., 2021)
Edge SVD (MsFEM)	Discretize restriction operator, compute SVD, extend harmonically	(Chen et al., 2020)
THB-Spline Goal-Oriented Adaptivity	Mark and refine locally via weighted DWR indicators	(Hinz et al., 2020)

In all cases, localization (e.g., truncation, oversampling, refinement) is crucial for computational efficiency and for ensuring basis functions adapt to significant geometric structure.

5. Generalization, Inductive Bias, and Empirical Validation

Empirical studies and theoretical analysis demonstrate that GAHBs are closely linked to generalization and optimality:

• In score-based DNNs, networks trained on non-overlapping subsets produce near-identical score functions when trained on sufficient data ($N\approx10^5$ for $80\times80$ images), with low model variance and strong generalization (Kadkhodaie et al., 2023).
• For regular image classes (e.g., $C^\alpha$ “cartoon” images), the theoretical minimax PSNR slope $\alpha/(\alpha+1)$ (via bandlet theory) is matched by DNN denoisers employing GAHBs, and basis vectors resemble bandlet atoms (Kadkhodaie et al., 2023).
• In multiscale elliptic PDEs, edge-adaptive harmonic basis construction via SVD and harmonic extension achieves exponential accuracy, robust across rough and high-contrast coefficients (Chen et al., 2020).
• Truncated harmonic characteristic bases in high-contrast PDEs exhibit rapid error decay in truncation parameter $\delta$, with small-dimensional multiscale spaces accurately capturing leading-order solution components (Poveda et al., 2014).
• On manifold surfaces, Laplace–Beltrami manifold harmonics enable operator compression—halving the number of basis functions can yield relative errors $10^{-3}$–$10^{-4}$ in current and RCS metrics (Alsnayyan et al., 2021).

6. Applications across Disciplines

Geometry-adaptive harmonic bases are applied in:

• Score-based diffusion image generation and denoising (local basis shrinkage yields near-minimax performance, density consistency) (Kadkhodaie et al., 2023).
• Multiscale finite element methods for high-contrast, heterogeneous materials (localized harmonics for efficient solution representation) (Poveda et al., 2014).
• Computational geometry, surface PDEs, and electromagnetic simulation (operator compression via manifold harmonics) (Alsnayyan et al., 2021).
• Adaptive edge-basis methods for robust and accurate resolution of rough-coefficient PDEs (Chen et al., 2020).
• Planar and surface parameterization with localized THB-spline refinement and geometry-adaptive PDE-based mapping (Hinz et al., 2020).

These bases are central to achieving sparse representation, computational efficiency, and geometric fidelity in diverse scientific and engineering tasks.

7. Limitations, Contingencies, and Open Directions

GAHB constructions critically depend on the underlying geometry and the chosen operator:

• In scenarios where geometry is randomized or permuted (e.g., shuffled pixels in images), inductive bias toward local harmonic adaptation fails, and empirical performance degrades (Kadkhodaie et al., 2023).
• For low-dimensional manifolds where the optimal basis is simply tangent vector direction, GAHB modes capture only part of the optimal representation (slight suboptimality observed) (Kadkhodaie et al., 2023).
• The construction cost (e.g., local solves, SVD) is manageable for localized bases, but scaling to extreme problem sizes or extremely fine geometric features may require further algorithmic advances or hierarchical compression strategies (1410.02932106.11907).

A plausible implication is that future developments may combine data-driven learning of geometry-adaptive harmonic modes with goal-oriented adaptive refinement and nonlinear domain optimization for even more efficient high-dimensional representations.