What problem does this paper attempt to address?

The problem that this paper attempts to solve is to accelerate the convergence of the inertial algorithm in a dynamical system with Hessian - driven damping and three constant coefficients by introducing a velocity restart scheme. Specifically, the author studied an acceleration restart scheme with Hessian - driven damping and established the linear convergence rate of the function value along the restart trajectory without assuming the strong convexity of the target function. ### Main problem decomposition: 1. **Improvement of the dynamical system**: The paper explored how to reduce the oscillatory behavior in the inertial system by introducing the Hessian - driven damping term. For example, the Hessian - driven damping term was added to the Damped Inertial Newton - like system (DIN), thus alleviating the typical oscillation caused by Asymptotic Vanishing Damping (AVD). 2. **Velocity restart strategy**: The author proposed a velocity restart strategy based on velocity change, aiming to further accelerate convergence. The core idea of this strategy is to restart the system when the velocity stops increasing, so as to avoid unnecessary oscillation and improve the convergence speed. 3. **Theoretical analysis and experimental verification**: The paper not only provided theoretical guarantees, proving that under certain conditions (such as when the Polyak - / Łojasiewicz inequality holds), the function value can converge to the optimal value at a linear rate; it also demonstrated the effectiveness of this method through numerical experiments, especially in high - dimensional optimization problems. ### Formula summary: - **DIN - AVD system**: \[ \ddot{x}(t)+\frac{\alpha}{t}\dot{x}(t)+\beta\nabla^{2}f(x(t))\dot{x}(t)+\nabla f(x(t)) = 0,\quad\alpha,\beta>0 \] - **WIN system**: \[ \ddot{x}(t)+\alpha\dot{x}(t)+\beta\nabla^{2}f(x(t))\dot{x}(t)+\gamma\nabla f(x(t)) = 0,\quad\alpha>0,\beta\geq0,\gamma>0 \] - **Velocity restart time**: \[ T(z)=\inf\left\{t > 0:\frac{d}{dt}\|\dot{x}_{z}(t)\|^{2}\leq0\right\} \] Through these improvements, the paper provides a new acceleration mechanism for optimization algorithms, especially when dealing with non - strongly convex target functions.

A Speed Restart Scheme for a Dynamical System with Hessian-Driven Damping and Three Constant Coefficients

Inertial Methods with Viscous and Hessian driven Damping for Non-Convex Optimization

An SDE Perspective on Stochastic Inertial Gradient Dynamics with Time-Dependent Viscosity and Geometric Damping

Solving convex optimization problems via a second order dynamical system with implicit Hessian damping and Tikhonov regularization

Restart of accelerated first order methods with linear convergence under a quadratic functional growth condition

Inertial primal-dual dynamics with Hessian-driven damping and Tikhonov regularization for convex-concave bilinear saddle point problems

Optimal Convergence Rates for Damped Inertial Gradient Dynamics with Flat Geometries

Accelerated Gradient Methods Combining Tikhonov Regularization with Geometric Damping Driven by the Hessian

Fast Convergence of Inertial Multiobjective Gradient-like Systems with Asymptotic Vanishing Damping

Optimal convergence rate of inertial gradient system with flat geometries and perturbations

Fast convergence rates and trajectory convergence of a Tikhonov regularized inertial primal\mbox{-}dual dynamical system with time scaling and vanishing damping

A fast continuous time approach with time scaling for nonsmooth convex optimization

Convergence and Speed Estimates for Semilinear Wave Systems with Nonautonomous Damping

Fast convex optimization via closed-loop time scaling of gradient dynamics

Fast second-order dynamics with slow vanishing damping approaching the zeros of a monotone and continuous operator

Accelerated Gradient Methods with Gradient Restart: Global Linear Convergence

Stochastic Inertial Dynamics Via Time Scaling and Averaging

Whiplash Gradient Descent Dynamics

Inertial dynamics with vanishing Tikhonov regularization for multobjective optimization

Convergence of Inertial Dynamics Driven by Sums of Potential and Nonpotential Operators with Implicit Newton-Like Damping