Radu I. Boţ,Enis Chenchene,Jalal M. Fadili
Abstract:We study accelerated Krasnosel'ki\uı-Mann-type methods with preconditioners in both continuous and discrete time. From a continuous time model, we derive a generalized fast Krasnosel'ki\uı-Mann method, providing a new yet simple proof of convergence that allows for unprecedented flexibility in parameter tuning. Our analysis unifies inertial and anchoring acceleration mechanisms and offers a broad range of parameter choices, which prove beneficial in practice. Additionally, we extend our analysis to the case where preconditioners are allowed to be degenerate, unifying the treatment of various splitting methods and establishing the weak convergence of shadow sequences.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is to develop and analyze an accelerated Krasnosel’ki˘ı–Mann method. This method combines pre - conditioning techniques and can effectively solve fixed - point problems in both continuous and discrete time. Specifically, the goals of the paper include:
1. **Introduce the generalized fast Krasnosel’ki˘ı–Mann method**: By introducing a more general parameter selection, namely \(b(t) := \eta \dot{\beta}(t)+\frac{\alpha \eta \beta(t)}{t}\), where \(\eta\in(0, 1)\) and \(\alpha_\eta:=\eta+(1 - \eta)(\alpha - 1)\), the existing Fast - OGDA system is generalized.
2. **Provide a new convergence proof**: Through a close connection with the continuous - time model, the author provides a simple convergence proof for the proposed discrete - time algorithm (Algorithm 1) that almost completely matches the continuous - time method. This enables them to cover cases not covered by previous methods, such as the case where \(\alpha = 2\).
3. **Handle degenerate pre - conditioners**: The paper also explores the case when the pre - conditioner is allowed to be degenerate, uniformly handles various splitting methods, and establishes the weak convergence of the so - called "shadow sequence".
4. **Verify by numerical experiments**: Numerical experiments show that the proposed method is superior to the baseline method in some cases, especially when \(\eta<\frac{1}{2}\), it shows better convergence performance.
### Main contributions
- **Generalize the Fast - OGDA system**: A more general parameter selection is introduced, making the method more flexible.
- **New convergence proof**: A simple convergence proof that is almost identical to the continuous - time method is provided, covering more parameter settings.
- **Handle degenerate pre - conditioners**: By introducing the degenerate Opial lemma, the situation of degenerate pre - conditioners is accommodated, ensuring the convergence of multiple splitting methods.
- **Numerical experiments**: Demonstrate the superior performance of the new method in practical applications.
### Mathematical formulas
The key formulas involved in the paper include:
- The dynamic equation of the continuous - time system:
\[
\ddot{x}(t)+\frac{\alpha}{t}\dot{x}(t)+\beta(t)\frac{d}{dt}Q(x(t)) + b(t)Q(x(t))=0
\]
where \(b(t)=\eta\dot{\beta}(t)+\frac{\alpha \eta \beta(t)}{t}\).
- The update rule of the discrete - time algorithm (Algorithm 1):
\[
x_{k + 1}=\left(1-\frac{\alpha_\eta}{k+\sigma}\right)x_k+\frac{\alpha_\eta}{k+\sigma}T(x_k)+\left(1-\frac{\alpha}{k+\sigma}\right)(T(x_k)-T(x_{k-1}))
\]
These formulas show how to adjust the parameters to accelerate the solution of the fixed - point problem and ensure the convergence and stability of the algorithm.