Search arXiv⌕ Search
Home/Archive/arXiv

arXiv · 2611.07692

Hybrid Newton-BFGS high-index saddle dynamics for accelerating saddle-point searches

Jianyuan Yin·Lei Zhang·Shixue Zhang
View original source ↗Read source document ↗Export metadata ↓

Abstract

We propose a hybrid Newton-BFGS high-index saddle dynamics (HiSD) method for accelerating saddle-point searches. The method constructs a Newton correction on the unstable subspace from the available negative eigenpairs, while BFGS updates based on projected secant information approximate the inverse Hessian on the stable subspace. By combining Newton and quasi-Newton corrections, the method incorporates curvature information in both subspaces while preserving the ascent-descent structure of HiSD. Accounting for the evolving stable subspaces, we establish local $Q$-superlinear convergence of the full-memory version near nondegenerate saddle points under suitable assumptions. We further develop a limited-memory version for large-scale problems and provide fixed-step and line search strategies. Numerical experiments show local convergence behavior consistent with the theory and demonstrate acceleration on ill-conditioned problems and degenerate saddle manifolds in neural-network loss landscapes. The method also supports efficient upward and downward saddle searches for a liquid crystal model, further illustrating its broad applicability to saddle-point computation.

Explore related subjects

math.NAcs.NA

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Jianyuan Yin, Lei Zhang, Shixue Zhang. 2026-11-05. Hybrid Newton-BFGS high-index saddle dynamics for accelerating saddle-point searches. https://arxiv.org/abs/2611.07692

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

Overcoming Spectral Bias via Cross-Attention

Neural networks often exhibit spectral bias, a phenomenon where high-frequency components may converge substantially more slowly than low-frequency ones. To address this imbalance, we propose a cross-attention-based network that combines a multiscale random Fourier feature bank with learnable amplitude weighting. The amplitude envelope modulates the Fourier features according to their frequencies, while the cross-attention residual structure provides an input-dependent mechanism to emphasize informative scales. As a result, the proposed design accelerates high-frequency convergence relative to comparison networks built on the same multiscale bank. Moreover, the feature bank can be incrementally enriched with dominant modes identified through discrete Fourier analysis of intermediate approximations, without modifying the hidden-layer architecture. We further extend the framework to PDE learning through a two-network representation in which one subnetwork approximates higher-frequency solution components and the other approximates lower-frequency components, with a learnable or analytically computed mixing coefficient balancing their contributions. Numerical experiments on function approximation, image reconstruction, and representative elliptic PDEs demonstrate the accuracy of the proposed method for problems involving high-frequency, multiscale, and discontinuous features.

math.NA↗
arXiv2026

Spherical Cap $L_2$ Discrepancy -- Blessing of Dimensionality and a Distance-Kernel Variant

We prove a dimension-explicit upper bound for the information-complexity (also called the inverse of the discrepancy) of the classical spherical cap $L_2$ discrepancy on the $d$-dimensional sphere $\mathbb{S}^d$. For every fixed $\varepsilon\in(0,1)$, this bound implies that a single point suffices in all sufficiently large dimensions, which shows a ``blessing of dimensionality'' for the associated numerical integration problem. We then introduce a one-parameter family of distance reproducing kernels for the Sobolev space $H^{(d+1)/2}(\mathbb{S}^d)$. For the balanced kernel $K(\boldsymbol{x},\boldsymbol{y})=1-\tfrac{1}{\sqrt{2}}\|\boldsymbol{x}-\boldsymbol{y}\|$, a lower bound together with a random-point upper bound shows that the normalized information-complexity is $Θ_\varepsilon(d)$ for every fixed $\varepsilon\in(0,1)$. Thus its dependence on $d$ is determined up to constants depending only on $\varepsilon$. Finally, we study the one-sided spherical cap discrepancy obtained by restricting the height parameter to $[0,1]$. We derive its kernel and Stolarsky's invariance principle explicitly. A sharp mean-distance inequality, attained by regular simplices, determines its normalized information-complexity exactly in all sufficiently large dimensions and proves strong polynomial tractability with optimal $\varepsilon$-exponent two.

math.NA↗
arXiv2026

Symplectic Hamiltonian Direct Discontinuous Galerkin Method for Wave Propagation

This paper presents a symplectic Hamiltonian direct discontinuous Galerkin (DDG) method for approximating wave propagation problems, including the linear and semilinear wave equations. Within an auxiliary-variable-free DG framework, we prove that the symmetry of the numerical flux bilinear form is equivalent to the existence of a discrete Hamiltonian structure. It follows that methods such as the symmetric interior penalty, BR2, and symmetric DDG (SDDG) methods admit a discrete Hamiltonian structure, whereas the Baumann--Oden and DDG methods do not possess this property. Exploiting this structure, we construct fully discrete symplectic schemes by combining the SDDG spatial discretization with symplectic time integrators. We also develop a unified treatment of periodic, Dirichlet, Neumann, and Robin boundary conditions that preserves the discrete Hamiltonian structure of the SDDG scheme. We further derive error estimates for the SDDG method applied to semilinear wave equations under these boundary conditions, showing the optimal convergence rate for the displacement and the suboptimal convergence rate for the velocity. Numerical experiments validate the theoretical convergence rates and demonstrate that the symplectic Hamiltonian DDG method achieves superior long-time energy conservation and accuracy.

math.NA↗