Abstract
In the study of energy functionals, finding transition states—mathematically defined as index-1 saddle points—is essential for understanding the pathways between stable states. While the convex splitting method is widely recognized for ensuring unconditional stability in gradient flow equations like Allen-Cahn and Cahn-Hilliard, its application to saddle point searches has remained relatively unexplored. This paper presents a novel framework that adapts the convex splitting idea to the Iterative Minimization Formulation (IMF) for saddle points. By introducing a general principle for constructing convex splitting forms for auxiliary functionals and utilizing a linearization technique to avoid nonlinear solvers, the proposed method enables the use of large time steps without compromising stability. The effectiveness of this approach is demonstrated through numerical experiments on 1D Ginzburg-Landau and 2D Landau-Brazovskii energy functionals.
Introduction
In fields ranging from materials science to quantum chemistry, researchers are often tasked with identifying the "mountain passes" of an energy landscape. These transition states represent the critical bottlenecks that a system must overcome to move from one local minimum (stable state) to another. However, calculating these index-1 saddle points is significantly more complex than finding local minima.
Traditional numerical methods, such as Gentlest Ascent Dynamics (GAD) or standard implementations of the Iterative Minimization Formulation (IMF), often suffer from stability issues. To prevent the simulation from diverging or "blowing up," researchers are frequently forced to use very small time steps, leading to high computational costs. This work addresses this limitation by applying the convex splitting method to the IMF framework, providing a path toward unconditional stability and increased efficiency.
Methodology
Iterative Minimization Formulation (IMF)
The IMF transforms the challenge of finding a saddle point into a sequence of minimization subproblems. At each cycle, an auxiliary functional is minimized to move the system closer to the transition state. The core difficulty lies in ensuring that these minimization steps are both stable and computationally inexpensive.
Convex Splitting and Linearization
The convex splitting method involves decomposing an energy functional $F$ into a convex part ($F_c$) and an expansive part ($F_e$), such that $F = F_c - F_e$. In a typical numerical scheme, the convex part is treated implicitly, while the expansive part is treated explicitly.
The authors' primary innovation is a general principle for constructing these splitting forms specifically for the auxiliary functionals used in the IMF. To further enhance efficiency, they introduced a "linearization trick." By strategically combining two different splitting forms of the original energy, they ensured that the resulting equations are linear. This allows the system to be solved using standard linear algebra, bypassing the need for expensive nonlinear solvers at every iteration.
Numerical Experiments and Results
The proposed method was validated using two primary physical models:
- 1D Ginzburg-Landau Functional: This model was used to simulate both non-conserved (Allen-Cahn) and conserved (Cahn-Hilliard) dynamics. In these tests, the Convex Splitting (CS) scheme demonstrated superior stability compared to non-convex (nCS) benchmarks. For the Allen-Cahn model, the CS scheme reached the target error in only 14 iterations using a large time step ($\Delta t = 10$), whereas standard schemes required hundreds of iterations at much smaller steps to maintain stability.
- 2D Landau-Brazovskii Functional: This complex model describes the phase transitions of diblock copolymers, such as the shift from lamellar (striped) to cylindrical (spotted) phases. The CS-based IMF successfully identified the transition states, accurately capturing the bottleneck configurations between these distinct physical patterns.
Discussion and Key Insights
The primary advantage of the convex splitting approach in this context is not necessarily the speed of a single iteration, but the ability to take massive time steps. While a standard semi-implicit scheme might diverge at $\Delta t = 0.2$, the CS scheme remains stable at $\Delta t = 10.0$ or higher. Furthermore, the authors demonstrated that if a convex splitting exists for a standard gradient flow, a stable splitting for the corresponding saddle point search can be derived automatically.
Conclusion
By integrating convex splitting into the IMF, this work provides a robust and efficient framework for calculating transition states in complex energy landscapes. The ability to maintain linearity and unconditional stability represents a significant practical advancement for high-dimensional physical simulations. Future work will explore applying this framework to even more challenging systems, such as the Kohn-Sham density functional in quantum chemistry, where numerical stability is a persistent hurdle.
Summary updated