Gentlest Ascent Dynamics: Saddle Search Framework

• Gentlest Ascent Dynamics is a continuous-time saddle search method that stabilizes index-1 saddle points by reversing the force along the unstable direction.
• It applies to both gradient and non-gradient systems using Hessian min-modes or Jacobian eigenvectors to identify the unique unstable direction.
• GAD extends to high-index, constrained, and manifold formulations, offering practical insights for rare-event simulations, multiscale dynamics, and PDE discretizations.

Gentlest Ascent Dynamics (GAD) is a continuous-time dynamical framework for escaping the basins of attraction of stable invariant sets and converging to index-1 saddle points. Its central construction is to reverse the original force only along a distinguished unstable direction while retaining descent or relaxation in all orthogonal directions, thereby turning the target saddle into a stable equilibrium of an extended dynamics. In gradient systems this direction is the min-mode of the Hessian; in non-gradient systems it is represented by appropriate right and left Jacobian eigenvectors. From its original formulation, GAD has developed into a broad family of saddle-search dynamics encompassing non-gradient flows, constrained and high-index settings, manifold formulations, multiscale systems, and PDE discretizations (E et al., 2010).

1. Core formulation and geometric mechanism

For a gradient system with energy $V$, GAD augments the state variable $x\in\mathbb{R}^d$ by a direction variable $v\in\mathbb{R}^d$ and evolves

$$\dot{x}(t) = -\nabla V(x) + 2 \frac{\nabla V(x)\cdot v}{v\cdot v}\, v,\qquad \gamma \dot{v}(t) = -\nabla^2V(x)\,v + (v\cdot \nabla^2V(x)v)\,v.$$

The $x$-equation flips the gradient component along $v$, so the dynamics climbs uphill in that direction while descending in all orthogonal directions. The $v$-equation is gradient descent on the Rayleigh quotient

$\min_{\|u\|=1} u^T H(x)u,\qquad H(x)=\nabla^2V(x),$

and therefore drives $v$ toward the smallest-eigenvalue eigenvector, i.e. the min-mode of the Hessian (Gu et al., 2018).

An equivalent geometric description uses the Householder reflection $H(v)$, defined by

$x\in\mathbb{R}^d$0

so that the GAD state equation can be written as

$x\in\mathbb{R}^d$1

In this form, the force is reflected across the hyperplane orthogonal to $x\in\mathbb{R}^d$2. The associated idealized saddle dynamics (ISD) assumes that the unstable direction is already known and evolves only the reflected $x\in\mathbb{R}^d$3-dynamics. On a Riemannian manifold $x\in\mathbb{R}^d$4, the intrinsic ISD becomes

$x\in\mathbb{R}^d$5

where $x\in\mathbb{R}^d$6 and $x\in\mathbb{R}^d$7 is the unit eigenvector field associated with the smallest eigenvalue of the Riemannian Hessian (Bello-Rivas et al., 2023).

For non-gradient systems $x\in\mathbb{R}^d$8, the original GAD introduces right and left direction variables $x\in\mathbb{R}^d$9 and $v\in\mathbb{R}^d$0, since the Jacobian $v\in\mathbb{R}^d$1 is generally asymmetric: $v\in\mathbb{R}^d$2 The normalization $v\in\mathbb{R}^d$3 and $v\in\mathbb{R}^d$4 is preserved by suitable $v\in\mathbb{R}^d$5. In the gradient case, symmetry collapses the two directions into a single min-mode, recovering the simpler formulation (Gu et al., 2018).

2. Fixed-point structure, stability, and convergence limits

The defining theoretical property of GAD is that stable fixed points of the extended dynamics correspond to index-1 saddles of the original system. In the gradient setting, if $v\in\mathbb{R}^d$6 is a stationary point of $v\in\mathbb{R}^d$7, then $v\in\mathbb{R}^d$8 is an equilibrium of GAD for every Hessian eigenvector $v\in\mathbb{R}^d$9, and exactly one of these equilibria is linearly stable if and only if $$\dot{x}(t) = -\nabla V(x) + 2 \frac{\nabla V(x)\cdot v}{v\cdot v}\, v,\qquad \gamma \dot{v}(t) = -\nabla^2V(x)\,v + (v\cdot \nabla^2V(x)v)\,v.$$0 is an index-1 saddle point. In the non-gradient setting, if $$\dot{x}(t) = -\nabla V(x) + 2 \frac{\nabla V(x)\cdot v}{v\cdot v}\, v,\qquad \gamma \dot{v}(t) = -\nabla^2V(x)\,v + (v\cdot \nabla^2V(x)v)\,v.$$1 is a fixed point of $$\dot{x}(t) = -\nabla V(x) + 2 \frac{\nabla V(x)\cdot v}{v\cdot v}\, v,\qquad \gamma \dot{v}(t) = -\nabla^2V(x)\,v + (v\cdot \nabla^2V(x)v)\,v.$$2 and $$\dot{x}(t) = -\nabla V(x) + 2 \frac{\nabla V(x)\cdot v}{v\cdot v}\, v,\qquad \gamma \dot{v}(t) = -\nabla^2V(x)\,v + (v\cdot \nabla^2V(x)v)\,v.$$3 has distinct real eigenvalues, then each $$\dot{x}(t) = -\nabla V(x) + 2 \frac{\nabla V(x)\cdot v}{v\cdot v}\, v,\qquad \gamma \dot{v}(t) = -\nabla^2V(x)\,v + (v\cdot \nabla^2V(x)v)\,v.$$4 is a GAD fixed point, but linear stability selects precisely the one associated with the unique unstable eigenvalue [(Gao et al., 2014); (E et al., 2010)].

The linearization mechanism is explicit. At a fixed point associated with eigenvalue $$\dot{x}(t) = -\nabla V(x) + 2 \frac{\nabla V(x)\cdot v}{v\cdot v}\, v,\qquad \gamma \dot{v}(t) = -\nabla^2V(x)\,v + (v\cdot \nabla^2V(x)v)\,v.$$5, the GAD Jacobian has eigenvalues

$$\dot{x}(t) = -\nabla V(x) + 2 \frac{\nabla V(x)\cdot v}{v\cdot v}\, v,\qquad \gamma \dot{v}(t) = -\nabla^2V(x)\,v + (v\cdot \nabla^2V(x)v)\,v.$$6

Linear stability therefore requires $$\dot{x}(t) = -\nabla V(x) + 2 \frac{\nabla V(x)\cdot v}{v\cdot v}\, v,\qquad \gamma \dot{v}(t) = -\nabla^2V(x)\,v + (v\cdot \nabla^2V(x)v)\,v.$$7 and $$\dot{x}(t) = -\nabla V(x) + 2 \frac{\nabla V(x)\cdot v}{v\cdot v}\, v,\qquad \gamma \dot{v}(t) = -\nabla^2V(x)\,v + (v\cdot \nabla^2V(x)v)\,v.$$8 for all $$\dot{x}(t) = -\nabla V(x) + 2 \frac{\nabla V(x)\cdot v}{v\cdot v}\, v,\qquad \gamma \dot{v}(t) = -\nabla^2V(x)\,v + (v\cdot \nabla^2V(x)v)\,v.$$9, which is exactly the index-1 saddle condition for a general vector field. For energy landscapes, this corresponds to a stationary point with exactly one negative Hessian eigenvalue (E et al., 2010).

This stability transformation explains why GAD is often described as the saddle analogue of steepest descent. Steepest descent makes minima stable; GAD makes index-1 saddles stable. However, the convergence properties are not unrestricted. GAD has a clear dynamical interpretation and a large attraction domain compared with root-finding methods, but in configuration space it is only linearly convergent, in direct analogy with steepest descent dynamics (Gao et al., 2014). Moreover, a complete global convergence theory is not available in general. In a two-dimensional gradient example, the effective “GAD potential” is piecewise-defined and discontinuous across switching curves where the selected eigenvector changes, and instabilities can occur (E et al., 2010). A common misconception is therefore that GAD is a globally convergent saddle solver; the available results instead establish a local stability mechanism and, in several extensions, more structured convergence under additional assumptions.

3. Non-gradient, high-index, and constrained extensions

For non-gradient systems, the asymmetry of $x$0 motivates variants that reduce the cost of direction tracking. A simplified GAD replaces the original $x$1 system in $x$2 by an $x$3 or $x$4 system in $x$5: $x$6 or, alternatively,

$x$7

This preserves the same saddle-search property as the original GAD, reduces computational cost for directions by half, and can avoid explicit computation of $x$8 when the $x$9-version is used (Gu et al., 2018).

High-index generalizations reverse the flow on an unstable subspace rather than along a single direction. The original GAD paper discusses index-2 constructions for both the case of a complex conjugate unstable pair and the case of two positive real unstable eigenvalues, using unstable-plane identification or deflation-based recovery of successive unstable modes (E et al., 2010). A more systematic constrained high-index development is given by constrained GAD (CGAD), formulated on a constraint manifold

$v$0

With tangent projection $v$1, projected gradient

$v$2

and projected Hessian

$v$3

the index-$v$4 CGAD evolves

$v$5

Its linearly stable steady states are exactly nondegenerate constrained saddle points of Morse index $v$6, with $v$7 spanning the tangent negative spectral subspace. The corresponding idealized CGAD, in which the $v$8 are assumed to be the exact lowest tangent eigenmodes, satisfies a local exponential decay estimate

$v$9

near a nondegenerate constrained saddle (Liu et al., 2022).

4. Variational reformulations, multiscale dynamics, and PDE discretization

A major algorithmic reinterpretation replaces the continuous GAD flow by a local optimization map. In the iterative minimization formulation for index-1 saddle search, one first computes the min-mode

$v$0

then updates $v$1 by minimizing a modified local objective

$v$2

or, in the parameterized form,

$v$3

with $v$4. Near an index-1 saddle $v$5, the local minimizer map $v$6 satisfies $v$7, so the fixed-point iteration is quadratically convergent when the subproblem is solved exactly. If only a single steepest-descent step is taken on the local minimization problem, one recovers an Euler discretization of a GAD-like projected flow; if the direction subproblem is also solved by steepest descent, one recovers the original GAD (Gao et al., 2014).

For slow–fast stochastic systems, multiscale GAD (MsGAD) targets the effective averaged dynamics

$v$8

without assuming $v$9 is explicitly available. The key identity is

$\min_{\|u\|=1} u^T H(x)u,\qquad H(x)=\nabla^2V(x),$0

so the Jacobian of the averaged drift contains a covariance correction induced by the $\min_{\|u\|=1} u^T H(x)u,\qquad H(x)=\nabla^2V(x),$1-dependence of the fast invariant measure. The resulting MsGAD is solved numerically by HMM or SCM, with accelerations based on time averaging of the output, adaptive sample size $\min_{\|u\|=1} u^T H(x)u,\qquad H(x)=\nabla^2V(x),$2 in HMM, and adaptive boost parameter $\min_{\|u\|=1} u^T H(x)u,\qquad H(x)=\nabla^2V(x),$3 in SCM (Gu et al., 2016).

For semilinear elliptic PDEs, GAD can be lifted to a continuous-in-space parabolic system. For

$\min_{\|u\|=1} u^T H(x)u,\qquad H(x)=\nabla^2V(x),$4

the continuous-in-space index-1 saddle dynamics is

$\min_{\|u\|=1} u^T H(x)u,\qquad H(x)=\nabla^2V(x),$5

with $\min_{\|u\|=1} u^T H(x)u,\qquad H(x)=\nabla^2V(x),$6-inner product and $\min_{\|u\|=1} u^T H(x)u,\qquad H(x)=\nabla^2V(x),$7 preserved for all $\min_{\|u\|=1} u^T H(x)u,\qquad H(x)=\nabla^2V(x),$8. The stationary pairs $\min_{\|u\|=1} u^T H(x)u,\qquad H(x)=\nabla^2V(x),$9 correspond exactly to index-1 saddle points of the elliptic problem and the associated negative eigenfunction of $v$0. Standard $v$1 finite elements yield a semi-discrete error estimate

$v$2

while a backward Euler fully discrete scheme with normalization/retraction satisfies

$v$3

and the analysis establishes $v$4 stability and index preservation of the discrete saddle (Zhang et al., 3 Aug 2025).

5. Manifold formulations and data-driven intrinsic geometry

GAD has also been formulated intrinsically on manifolds that are not given analytically. In the point-cloud approach, the manifold is represented locally by a sampled cloud $v$5 near a current point $v$6, obtained either by perturbing $v$7 and flowing short-time trajectories or by integrating a stochastic differential equation such as

$v$8

Diffusion maps provide reduced coordinates $v$9, and Gaussian process regression is used to fit a smooth local chart $H(v)$0 and inverse $H(v)$1. The metric in local coordinates is then

$H(v)$2

the Levi-Civita connection is computed from the Christoffel symbols, and the intrinsic Hessian

$H(v)$3

provides the smallest-eigenvalue eigenvector field $H(v)$4 used in the intrinsic ISD update $H(v)$5 (Bello-Rivas et al., 2023).

This construction is explicitly intrinsic rather than extrinsic. It does not require explicit equality constraints defining the manifold, a global atlas, or precomputed reaction coordinates. Instead, it alternates between local point-cloud sampling, learned-chart construction, intrinsic geometric estimation, local ISD integration, and chart switching when the trajectory approaches the boundary of the reliable data support. The method requires the reactant, the ability to generate local samples around points on the manifold, and the ability to evaluate or simulate the local dynamics $H(v)$6. The numerical behavior reported in the paper indicates convergence in a finite number of outer iterations in the examples, while also emphasizing that accuracy depends on how well the point cloud approximates the manifold and that Hessian and eigenvector estimates become noisier when samples lie farther from the true manifold (Bello-Rivas et al., 2023).

6. Applications and related ascent–descent frameworks

In atomistic rare-event simulation, GAD was developed into deterministic, stochastic, and inertial finite-temperature forms for sampling index-1 saddles on high-dimensional PES. Deterministic GAD uses

$H(v)$7

while stochastic GAD adds white noise and MD-GAD evolves

$H(v)$8

A practical sampling protocol identifies an active region from the force pattern near a local minimum, initializes sparse random direction vectors on a small subset of atoms, and monitors the bond-stretching variable

$H(v)$9

For a surface vacancy and ad-atom pair on Cu(111), 661 initial direction vectors produced saddle events including ad-atom diffusion at $x\in\mathbb{R}^d$00 eV, concerted ad-atom/vacancy migration at $x\in\mathbb{R}^d$01 eV, vacancy diffusion at $x\in\mathbb{R}^d$02 eV, and bulk vacancy plus ad-atom formation at $x\in\mathbb{R}^d$03 eV. For a heptamer island on Cu(111), 843 initial direction vectors identified island sliding at $x\in\mathbb{R}^d$04 eV, island rotation at $x\in\mathbb{R}^d$05 eV, and subsurface vacancy plus ad-atom formation at $x\in\mathbb{R}^d$06 eV (Samanta et al., 2011).

In constrained quantum many-body models, CGAD has been used to compute excited states of Bose–Einstein condensates as constrained saddles of the Gross–Pitaevskii energy under the normalization constraint $x\in\mathbb{R}^d$07. In the linear case $x\in\mathbb{R}^d$08, the paper proves that an eigenfunction is an index-$x\in\mathbb{R}^d$09 excited state iff $x\in\mathbb{R}^d$10, so the ground state is the only constrained minimizer and higher excited states are ordered by Morse index. In nonlinear cases, the reported numerics show that higher Morse index generally correlates with higher energy and chemical potential, though not in a strict one-to-one way (Liu et al., 2022).

In power-system transient stability, GAD is used to compute 1-saddles on the DOA boundary of the post-fault stable equilibrium. Under assumptions A1–A5, the boundary satisfies

$x\in\mathbb{R}^d$11

so it can be reconstructed from the stable manifolds of index-1 critical elements. The numerical pipeline samples near the stable equilibrium, integrates GAD to locate 1-saddles, computes their stable manifolds, and augments them with periodic-orbit manifolds when necessary. Two-machine and three-machine benchmark systems are used to validate the reconstruction, including a case in which one 1-saddle covers only two faces of the boundary and the remaining four faces are supplied by index-1 periodic orbits located by an adjoint method and analyzed by perturbed GAD (Wu et al., 5 May 2026).

A related but distinct strand analyzes mixed ascent–descent dynamics through an energy/Lyapunov decomposition rather than explicit eigenvector following. For gradient descent-ascent,

$x\in\mathbb{R}^d$12

the dynamics can be rewritten as a Newtonian system with potential $x\in\mathbb{R}^d$13, antisymmetric “magnetic” term $x\in\mathbb{R}^d$14, and symmetric residual $x\in\mathbb{R}^d$15. When

$x\in\mathbb{R}^d$16

one can choose $x\in\mathbb{R}^d$17 so that the residual is purely dissipative and the Lyapunov function $x\in\mathbb{R}^d$18 decreases monotonically; bounded trajectories then converge to isolated steady states. This suggests a conceptual link to GAD: both frameworks separate stable and unstable directions in saddle-focused flows, but the former does so through a global curvature-ordering and dissipation argument rather than explicit min-mode dynamics (Seung, 2019).