Constrained Dynamics for Searching Saddle Points on General Riemannian Manifolds
Yukuan Hu, Laura Grazioli · 2026
Read Constrained Dynamics for Searching Saddle Points on General Riemannian Manifolds by Yukuan Hu, Laura Grazioli (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2601-03931.txt.
Introduction
Yukuan Hu and Laura Grazioli develop dynamics for finding saddle points on curved surfaces set in plain Euclidean space. Saddle points of energy landscapes matter in physics and chemistry, where they mark transition states between stable configurations. Existing methods need clear formulas describing the surface. Those may be missing when the surface is defined in roundabout ways through projectors, factorizations, or rank constraints.
Their constrained saddle dynamics is built only from geometric building blocks of the embedded surface, avoiding clear surface formulas entirely. The setup lives compactly on the Grassmann bundle of the tangent bundle, and its geometry yields proofs of local stability and local linear convergence. The authors note this is the first iterate convergence result for discrete algorithms seeking saddle points of prescribed index in this setting, a guarantee numerical chemists can build on.
Catalog record
| CREATOR | Yukuan Hu, Laura Grazioli |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 121 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2601-03931
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Constrained Dynamics for Searching Saddle Points on General Riemannian Manifolds Yukuan Hu, Laura Grazioli Abstract Finding constrained saddle points on Riemannian manifolds is significant for analyzing energy landscapes arising in physics and chemistry. Existing works have been limited to special manifolds that admit global regular level-set representations, excluding applications such as electronic excited-state calculations. In this paper, we develop a constrained saddle dynamics applicable to smooth functions on general Riemannian manifolds. Our dynamics is formulated compactly over the Grassmann bundle of the tangent bundle. By analyzing the Grassmann bundle geometry, we achieve universality via incorporating the second fundamental form, which captures variations of tangent spaces along the trajectory. We rigorously establish the local linear stability of the dynamics and the local linear convergence of the resulting algorithms. Remarkably, our analysis provides the first convergence guarantees for discretized saddle-search algorithms in manifold settings. Moreover, by respecting the intrinsic quotient structure, we remove unnecessary nondegeneracy assumptions on the eigenvalues of the Riemannian Hessian that are present in existing works. We also point out that locating saddle points can be more ill-conditioning than finding local minimizers, and requires using nonredundant parametrizations. Finally, numerical experiments on linear eigenvalue problems and electronic excited-state calculations showcase the effectiveness of the proposed algorithms and corroborate the established local theory. 1 Introduction Finding the saddle points (SPs) of potential energy functionals is a fundamental task in various scientific and engineering applications, particularly those involving energy landscape analysis. For example, the transition states between two (meta)stable states, which are crucial in physics [ 31 ] , chemistry [ 82 ] , and biology [ 76 ] , can be identified as index-1 SPs [ 2 ] . Higher-index SPs are useful for collective, multi-mode, or concerted transitions and play a key role in the construction of solution landscape [ 39 , 88 ] . Here, an unconstrained index- k SP is a critical point where the Euclidean Hessian has exactly k negative eigenvalues. Additionally, Riemannian manifolds can naturally arise as constraint sets due to the incorporation of physical laws, such as in the Thomson problem [ 64 , 81 ] and Bose-Einstein condensation [ 7 , 12 , 28 ] . In such cases, index- k constrained SPs can be analogously defined by using the Riemannian gradient and Hessian. In comparison with finding local or global minimizers, locating SPs exhibits two distinct and major challenges: (1) there are always unknown descent directions at SPs, rendering off-the-shelf optimization methods unstable; (2) in general, it is impossible to construct a global merit function which is variationally minimized (or maximized) at SPs, posing significant difficulties for designing globally convergent numerical methods [ 50 ] . There have been numerous algorithmic developments concerning saddle search in the unconstrained settings, together with rigorous theoretical analyses. On the contrary, the exploration in the constrained settings has been confined to special manifolds that admit global regular level-set representations (see Eq. ( 4 ) later for definition). This limitation rules out applications such as electronic excited-state calculations [ 27 , 69 , 84 ] , where the underlying quotient structures substantially complicate manifold representations. In some cases, omitting the quotient structures allows applying the existing methods directly. But unlike searching for local or global minimizers, this simplification can be detrimental in our context; see Example 2 and Section 5.1 later for an example involving the Stiefel and Grassmann manifolds. In this work, we develop a constrained saddle dynamics on general Riemannian manifolds and analyze the theoretical properties of the continuous dynamics and the resulting discretized algorithms. We further demonstrate the effectiveness of algorithms through electronic excited-state calculations on standard benchmark molecular systems. 1.1 Literature review In the following, we review the existing works for finding SPs in both unconstrained and constrained settings. Most of them primarily focus on developing locally convergent methods, with only few exceptions [ 49 , 79 ] . Our discussions are confined to cases where objective values, first-order derivatives, and (approximate) Hessian-vector products are available, excluding direct applications of Newton-type methods [ 4 , 6 , 20 , 78 ] . Unconstrained settings. In these cases, the numerical methods for finding the index-1 SPs can be mainly categorized into two classes: single-ended and double-ended methods. The double-ended methods (also known as the path-finding or chain-of-states methods) [ 41 , 45 , 55 ] are fed with two candidates of local minimizers and target at finding the minimum energy path (which passes through an index-1 SP under certain conditions by the mountain pass theorem [ 2 ] ). The single-ended ones (also known as the surface walking or eigenvector-following methods) [ 68 ] start with a single initial point without a priori knowledge about the final state. In this work, we focus on the class of single-ended methods 1 1 1 In fact, the double-ended methods are unsuitable for locating higher-index SPs by their nature and cannot be easily generalized. , covering the activation-relaxation technique ( nouveau ) [ 9 , 10 , 18 , 22 , 25 , 60 , 62 ] , gentlest ascent dynamics [ 26 , 71 ] , (shrinking) dimer methods [ 42 , 44 , 70 , 91 , 92 ] , among others [ 53 , 65 ] . All the mentioned three methods can be set in the flow \displaystyle\frac{\,\mathrm{d}\bm{x}}{\,\mathrm{d}t}(t) \displaystyle=-\mathrm{Proj}_{\bm{v}(t)^{\perp}}\big(\mathrm{grad}\,f(\bm{x}(t))\big)+\left\langle\bm{v}(t),\mathrm{grad}\,f(\bm{x}(t))\right\rangle\bm{v}(t) (1) \displaystyle=-R_{\bm{v}(t)}\big(\mathrm{grad}\,f(\bm{x}(t))\big), where f:\mathcal{E}\to\mathbb{R} is the objective (or energy) functional, \left\langle\bullet,\bullet\right\rangle:\mathcal{E}\times\mathcal{E}\to\mathbb{R} is the inner product of the ambient Euclidean space, \bm{v}(t)\in\mathcal{E} is an additional direction variable, \mathrm{Proj}_{\bm{v}(t)^{\perp}} denotes the orthogonal projection operator onto \mathrm{span}\{\bm{v}(t)\}^{\perp} , defined as \mathrm{Proj}_{\bm{v}(t)^{\perp}}(\bm{u}):=\bm{u}-\left\langle\bm{v}(t),\bm{u}\right\rangle\bm{v}(t),\quad\forall~\bm{u}\in\mathcal{E}, and R_{\bm{v}(t)} represents the (Householder) reflection operator defined using \bm{v}(t) : R_{\bm{v}(t)}(\bm{u}):=\bm{u}-2\left\langle\bm{v}(t),\bm{u}\right\rangle\bm{v}(t)=\mathrm{Proj}_{\bm{v}(t)^{\perp}}(\bm{u})-\left\langle\bm{v}(t),\bm{u}\right\rangle\bm{v}(t),\quad\forall~\bm{u}\in\mathcal{E}. Briefly speaking, instead of following the gradient flow, which leads the trajectory to a local or global minimizer, the dynamics ( 1 ) increases the objective value by climbing up along \pm\bm{v}(t) (whose sign is determined by the angle between \bm{v}(t) and \mathrm{grad}\,f(\bm{x}(t)) ), while decreases the value in all the directions perpendicular to \bm{v}(t) . An effective candidate for \bm{v}(t) is the normalized eigenvector corresponding to the lowest eigenvalue of the Hessian \mathrm{Hess}\,f(\bm{x}(t)) . In a neighborhood of a nondegenerate index-1 unconstrained SP, the dynamics ( 1 ) can then be viewed as the gradient flow of a local strongly convex merit function [ 30 , 33 ] , thereby providing a local stabilization for the index-1 SP. This underlies the activation-relaxation technique, where \bm{v}(t) is the solution of \min_{\bm{u}}~\left\langle\bm{u},\mathrm{Hess}\,f(\bm{x}(t))[\bm{u}]\right\rangle,\quad\mathrm{s.~t.}~\|\bm{u}\|=1, (2) possibly solved approximately by Krylov subspace methods such as the Lanczos algorithm [ 46 ] . Instead of solving the \bm{v} -subproblem ( 2 ) directly, the gentlest ascent dynamics includes a direction dynamics to track the lowest eigenvector, \frac{\,\mathrm{d}\bm{v}}{\,\mathrm{d}t}(t)=-\mathrm{Proj}_{\bm{v}(t)^{\perp}}\big(\mathrm{Hess}\,f(\bm{x}(t))[\bm{v}(t)]\big), which follows from the Euler-Lagrange equation of problem ( 2 ). The (shrinking) dimer methods further approximate the Hessian-vector product through a central finite-difference scheme, \mathrm{Hess}\,f(\bm{x}(t))[\bm{v}(t)]\approx\frac{1}{2\ell(t)}\left(\mathrm{grad}\,f\big(\bm{x}(t)+\ell(t)\bm{v}(t)\big)-\mathrm{grad}\,f\big(\bm{x}(t)-\ell(t)\bm{v}(t)\big)\right), with 2\ell(t)>0 the so-called dimer length, which is sometimes driven to 0^{+} as t\to+\infty [ 91 ] . In recent years, there have been extensions for locating index- k unconstrained SPs [ 21 , 26 , 30 , 58 , 59 , 71 , 88 , 89 ] . For instance, the gentlest ascent dynamics can be generalized by incorporating k direction dynamics [ 89 ] ,