Unification of algorithms for minimum mode optimization.

Area of Science:

• Computational chemistry
• Materials science
• Numerical analysis

Background:

• Minimum mode following algorithms are crucial for identifying saddle points in complex chemical and material systems.
• These algorithms require finding the minimum curvature mode of the Hessian matrix without direct computation, which is computationally infeasible for high-dimensional systems.
• Existing methods include Lanczos, dimer, Rayleigh-Ritz minimization, shifted power iteration, and locally optimal block preconditioned conjugate gradient.

Purpose of the Study:

• To unify various iterative minimum mode finding algorithms under a single theoretical framework.
• To theoretically compare the efficiency of different minimum mode finding methods.
• To establish the Lanczos method as a benchmark for minimum mode optimization.

Main Methods:

• Unification of iterative minimum mode finding algorithms using the Krylov subspace concept.
• Theoretical analysis demonstrating that smaller search subspaces used by other methods are contained within the Lanczos method's Krylov space.
• Numerical testing of the dimer method against the Lanczos method.

Main Results:

• All discussed iterative methods can be unified within the Krylov subspace framework.
• The theoretical efficiency of minimum mode finding methods is bounded by the Lanczos method.
• Numerical results show the dimer method with second-order optimizers approaches, but does not surpass, the efficiency of the Lanczos method.

Conclusions:

• The Krylov subspace provides a unified theoretical basis for minimum mode finding algorithms.
• The Lanczos method offers a theoretically optimal approach for minimum mode optimization.
• Practical methods like the dimer method are efficient but theoretically limited compared to Lanczos.