Intrinsic Finite-Difference Schemes

• Intrinsic finite-difference schemes are numerical methods defined using the underlying geometric or fractal structure, avoiding reliance on external coordinate systems.
• They provide tailored derivative approximations that ensure stability and error control in applications like Riemannian optimization, PDE discretization on fractals, and spectral differentiation.
• Adaptive algorithmic workflows utilizing tangent space bases and retraction operations yield dimension-optimal convergence and efficient computational performance.

An intrinsic finite-difference scheme is a numerical method for approximating derivatives or gradient operators that is constructed directly with respect to the geometric or analytic structure of the underlying space rather than by embedding into or referencing an external coordinate system. This approach is especially significant for problems posed on Riemannian manifolds, fractal graphs, or periodic domains, offering accuracy, stability, and convergence properties tailored to the intrinsic structure. Such schemes have found applications in Riemannian optimization, analysis on fractals, and spectrally-accurate differentiation, distinct from traditional finite-difference approaches that rely on Euclidean coordinates or spectral expansions.

1. Mathematical Framework and Definitions

Intrinsic finite-difference schemes are defined on structures such as Riemannian manifolds or self-similar fractals rather than Euclidean grids. On a smooth Riemannian manifold $\mathcal{M}$ of dimension $d$, equipped with an inner product $\langle\cdot,\cdot\rangle_x$ and norm $\|\cdot\|_x$ on each tangent space $T_x\mathcal{M}$, along with a retraction $R\colon T\mathcal{M}\to\mathcal{M}$, the goal is typically to approximate the Riemannian gradient of a function $f:\mathcal{M}\rightarrow \mathbb{R}$ or to discretize evolution equations intrinsic to the manifold (Taminiau et al., 13 Jan 2026).

The intrinsic forward finite-difference gradient at accuracy $h>0$ is: $$g_h(x) = \sum_{l=1}^d \frac{f\bigl(R_x(h\,e_l(x))\bigr)-f(x)}{h}\,e_l(x),$$ where $\{e_l(x)\}_{l=1}^d$ is an orthonormal basis of $T_x\mathcal{M}$. Central-difference variants are also possible, utilizing only variations along tangent directions to ensure computations remain on the manifold.

On discrete self-similar sets such as Sierpiński simplices, the finite-difference scheme is constructed recursively via graph Laplacians tailored to the fractal's combinatorial structure, avoiding reliance on external spectral data (Riane et al., 2018).

2. Construction Techniques and Algorithmic Workflow

Intrinsic schemes decompose directional derivatives or Laplacians in terms of the manifold's tangent space or native combinatorial decompositions. In Riemannian optimization, the main ingredients are:

• Selection of an orthonormal basis in $T_x\mathcal{M}$.
• Use of a global retraction $R$ to move in tangent directions while staying on $\mathcal{M}$.
• Discrete gradient approximation evaluated only with function values at intrinsic points.

A key advancement is the use of adaptive control parameters: a conservative estimate $\tau_k$ for regulating finite-difference accuracy, and an optimistic step-length parameter $\sigma_k$ for Riemannian updates. The algorithm iterates as follows (Taminiau et al., 13 Jan 2026):

• Select $h_k=\frac{2\,\epsilon}{5\,\sqrt{d}\,\tau_k}$.
• Compute $g_{h_k}(x_k)$.
• Evaluate success/failure via magnitude and Armijo-type decrease conditions.
• Adaptively update $\tau_k$ and $\sigma_k$.

On self-similar fractals, the combinatorics of the graph Laplacian at refinement level $m$ defines the intrinsic scheme: $\Delta_m u(X) = \sum_{Y \sim_m X} (u(Y) - u(X)),$ with scaling and normalization such that, as $m\to\infty$,

$$\Delta u(X) \approx \frac{d}{2}(d+2)^{m}\Delta_m u(X) + \mathcal{O}(2^{-m\alpha}).$$

Time-evolution equations are discretized intrinsically by explicit schemes using only graph-theoretic and measure-theoretic quantities (Riane et al., 2018).

3. Error Analysis, Consistency, and Stability

Intrinsic finite-difference schemes are accompanied by rigorous error analysis reflecting the geometry of the underlying space:

• For Riemannian gradients, the finite-difference error satisfies

$$\|\nabla f(x) - g_h(x)\|_x \leq \frac{L_M \sqrt{d}}{2} h$$

under $(L_M,R)$–smoothness assumptions, permitting explicit error control and adaptive step selection (Taminiau et al., 13 Jan 2026).

• On self-similar graphs, Taylor expansions and spectral scaling arguments provide local truncation error estimates of the form

$\varepsilon^m_{k,i} = O(h) + O(2^{-m\alpha}),$

leading to consistency as $h\to 0, m\to \infty$ (Riane et al., 2018).

• Stability is typically governed by a CFL-type condition adapted to the geometry. For the Sierpiński simplex,

$h \leq \frac{d^2}{2} (d+2)^{-m},$

ensures that the eigenvalues of the update matrix remain in the stable region.

• The convergence theorem binds the global error in terms of time discretization and graph refinement, yielding

$$\max_{0\le k\le N}\|u(t_k)-u^m_h(t_k)\| = O(h) + O(2^{-m\alpha}),$$

under stability constraints.

4. Optimality, Complexity, and Practical Considerations

The intrinsic approach is dimension-optimal for gradient-based methods on manifolds:

• The overall complexity to find an $\epsilon$-critical point is $O(d\,\epsilon^{-2})$ function evaluations and retractions, which matches known lower bounds for such algorithms (Taminiau et al., 13 Jan 2026).
• On fractal domains, efficient recursive constructions of the Laplacian allow for intrinsic discretizations without spectral decimation or eigenvalue approximation, yielding both constructive error and convergence guarantees (Riane et al., 2018).

Key practical aspects:

• Intrinsic schemes require a mechanism for constructing tangent space bases and retractions; for many classical matrix manifolds, these are available in closed-form or via standard linear algebra routines.
• Parallelization is straightforward since function evaluations in different directions are independent.
• Adaptive schemes avoid reliance on a priori global smoothness constants, instead learning via control parameters during execution.

5. Distinction from Extrinsic and Spectral Approaches

Intrinsic finite-difference methods contrast sharply with extrinsic and spectral approaches:

• Extrinsic finite differences perturb coordinates in an embedding space, which may not respect manifold constraints and often require projection steps, whereas intrinsic schemes operate purely along tangent directions and retain manifold structure throughout (Taminiau et al., 13 Jan 2026).
• Spectral and eigenfunction-based methods often rely on the availability of eigenvalues or spectral expansions, which may be unattainable on general domains or combinatorially complex spaces. Intrinsic schemes (as on the Sierpiński gasket or simplex) eschew spectral information, relying entirely on internal graph-theoretic structure for difference operators (Riane et al., 2018).
• Distributional and spectrally-accurate schemes (e.g., those built from singular Cauchy integrals) offer another class of intrinsic finite-difference operators developed directly in the function space without recourse to polynomial or trigonometric interpolation. These are prominent for periodic problems where physically intrinsic convolutional stencils can achieve spectral accuracy, outperforming polynomial-based finite differences in convergence order for smooth or analytic functions (Nachbin, 2019).

6. Applications and Impact

The intrinsic finite-difference paradigm has been influential in several domains:

• Derivative-free Riemannian optimization, where manifold constraints preclude extrinsic perturbations and intrinsic structures facilitate efficient, provably convergent algorithms with optimal iteration complexity (Taminiau et al., 13 Jan 2026).
• Numerical PDEs on fractals, enabling time-evolution problems to be discretized and solved without relying on spectral decimation, thus broadening the class of fractal geometries amenable to analysis (Riane et al., 2018).
• Spectrally-accurate differentiation, especially for periodic problems, where intrinsic, distributional finite-difference schemes achieve exponentially fast convergence for analytic data, rivaling global transforms but operating locally and in physical space (Nachbin, 2019).

7. Theoretical and Algorithmic Advancements

The development of intrinsic finite-difference schemes has led to multiple theoretical and algorithmic innovations:

• Adaptive regularity control via dual estimates of smoothness constants, enhancing robustness and removing the need for global smoothness a priori (Taminiau et al., 13 Jan 2026).
• Fully constructive convergence proofs using graph-intrinsic constructions (e.g., on Sierpiński simplices), allowing for explicit error, stability, and consistency bounds derived directly from the combinatorics of the fractal (Riane et al., 2018).
• Distributional kernel and multi-resolution stencils, providing a new physical-space alternative to spectral differentiation, with explicit stencils yielding both algebraic and spectral rates depending on regularity (Nachbin, 2019).

These advances collectively broaden the reach of finite-difference methods, enabling numerical methods to exploit intrinsic geometric or analytic structures beyond the classical Euclidean setting.