A Diagonal BFGS Update Algorithm with Inertia Acceleration Technology for Minimizations

Address the unrestricted optimization dilemma of reducing a continuously differentiable function $f:\mathbb{R}^{n}\rightarrow\mathbb{R}$, represented as:

Quasi-Newton methods belong to a category of iterative optimization algorithms that employ quasi-Newton approximations of the Hessian matrix to address the problem that defined in (1.1). The underlying principle involves forming a sequence of iterates $\{x_{k}\}$ designed to converge towards a fixed point of $f$, achieved through the iterative refinement of an approximate Hessian matrix $B_{k}$ to expedite optimization procedures with $f$ at each iteration. The foundational principles of the quasi-Newton approach originated with Davidon [2], and subsequent enhancements and applications have been explored by numerous scholars [3, 4, 5]. We start with $x_{0}\in\mathbb{R}^{n}$ and $B_{0}\in\mathbb{R}^{n\times n}$. The iterative process is updated by generating a sequence $\{x_{k}\}$ to iteratively refine the estimation of the Hessian matrix:

In $\mathbb{R}^{n}$, the determination of the search direction is governed by the line search vector $d_{k}$, and $\alpha_{k}>0$ represents the determination of the $\alpha_{k}$ is conducted via line search procedure. The meaning of $g_{k}$ represents partial gradient of $f$ at $x_{k}$, represented as a vector of the function we are looking for:

$s_{k}$ represents the difference vector between consecutive iterates, computed as $x_{k+1}-x_{k}$, and the variable $y_{k}$ denotes the discrepancy vector between the gradients observed at consecutive iterations, formulated as $g_{k+1}-g_{k}$. In algorithm for approximating the Hessian matrices, the vectors $s_{k}$ and $y_{k}$ are crucial for the enhancement of the algorithm, which is realized by an update process. BFGS updating is the most widely used among those formulations, expressed as:

Computational evidence has consistently demonstrated the superior effectiveness to other quasi-Newton methods. Considerable scholarly investigations have been dedicated to exploring the convergence characteristics of the technique , particularly in a realm of convex minimization, as evidenced by studies such as [6, 7, 8, 3, 4, 5]. These formulations have demonstrated high effectiveness in practical scenarios across a diverse range of optimization problems. By incorporating the insights garnered from these studies, we refine our existing formula $B_{k}$ in (1.4), leading to enhanced convergence performance.

Drawing inspiration from Neculai Andrei’s work on diagonal Hessian matrix from a function $f$[9], we leverage the efficacy of the update formula (1.4) and its variants. The approach discerns the components positioned along the principal diagonal of the approximative matrix which, adhering to the principles outlined in the theory from Dennis and Wolkowicz’s [10] secant updating method with minimal changes, is a key aspect to be addressed. Let us now delve into this consideration:

$b_{k}^{i}>0$ holds for every $i=1,\ldots,n$, an estimate of components of the matrix $B_{k}$ for $f$ at iteration $k$. Incorporated within the algorithm, a computation of the search direction which is expressed as $d_{k}=-B_{k}^{-1}g_{k}$, specifically, $d_{k}^{i}=-{g_{k}^{i}}/{b_{k}^{i}}$ for $i=1,\ldots,n$. Notably, as $B_{k}$ is both diagonal and positive definite, every element undergoes multiplication through distinct positive factors, determined by the diagonal factors $B_{k}$, in this particular algorithm. To facilitate the use of the minimum change secant update strategy, let us consider the assumption that the diagonal matrix $B_{k}$ possesses positive definiteness. The updated matrix $B_{k+1}$, defined as an outcome of the update to $B_{k}$, can be expressed as:

This observation motivates us to combine these two derivations of $B_{k}$ and obtain a new improved $B_{k}$ formula, which simplifies the proof of convergence. In utilizing the iterative update approach, let us make the assumption that the matrix $B_{k}$ should be positive definite. So $B_{k+1}$ derived from the update $B_{k}$ which is defined by $B_{k+1}s_{k}=y_{k}^{*}$, expressing $y^{*}$ as a formula to approach $\nabla^{2}f(x)s$, recognizing $y$ and $Y_{k}^{(3)}$ are vectors. Often, in practice, $\alpha_{k}$ from (1.2) is defined from WWP step rules [11, 12] with the selection of positive coefficients $\sigma$ and $\rho$ ensuring $1>\sigma>\rho>0$.

Section 1 of this paper introduces the research problem and recent advancements in related studies. Section 2 explains the motivation behind this research and proposes the improved diagonal quasi-Newton algorithm, DMBFGS3, as well as the WDMBFGS3 algorithm, which incorporates extrapolation techniques. Section 3 proves the global linear convergence of both algorithms. Section 4 validates the rapid convergence and stability of the WDMBFGS3 algorithm through numerical experiments. Section 5 wraps up the findings discussed in this paper and proposes avenues for further investigation.

As previously mentioned, the matrix $\Delta_{k}$ from (1.6) is ascertained as the result of the following questions:

Then we can further modify the definitions of $B_{k}$ and $y_{k}$,obtain the following formula:

The objective function of problem (2.2) constitutes a linear combination of minimizing the difference between $B_{k}$ and $B_{k+1}$ according to the principle of variation, along with minimizing the sum of the diagonal elements of $B_{k+1}$. The incorporation of $\text{tr}(B_{k}+\Delta_{k})$ within the target function from (2.1) is motivated by the aim to derive a correction matrix $\Delta_{k}$,

where$\Delta_{k}=\operatorname{diag}(\delta_{k}^{1},\dots,\delta_{k}^{n})$, $s_{k}=\begin{pmatrix}s_{k}^{1},\dots,s_{k}^{n}\end{pmatrix}^{T}$, and $y_{k}=\nabla^{2}f(x_{k})s_{k}.$ The Lagrangian corresponding to problem (2.3) can be said to be:

that $\lambda$ represents the lagrangian coefficient.Solutions pursued to this problem (2.3) correspond to a fixed point of the Lagrangian. Consequently, we take the partial derivative of (2.4) for $\delta_{k}^{i}$, so we can obtain:

By transferring items, we can obtain:

To determine $\lambda$, we use the constraint in (2.3) and substitute (2.6) into it, which gives:

where $A_{k}=\operatorname{diag}((s_{k}^{1})^{2},\dots,(s_{k}^{n})^{2})$. By utilizing (2.7) in (2.6), we can express the diagonal elements of $\Delta_{k}$ as follows:

where $B_{k}$ is the serves as the Hessian matrix. This diagonal correction matrix $\Delta_{k}$ can be articulated as

Then we have the following equation:

i.e., on components,

The vector is then ubsequently performed as:

where $g_{k+1}$ is the gradient at the new point, and $d_{k+1}$ is the search direction. It is essential to note that, in order to prevent the update matrix from degenerating or becoming an indefinite matrix $B_{k+1}$, it’s necessary that $b_{k+1}^{i}=b_{k}^{i}+\delta_{k}^{i}>0$ holds for all $i=1,\ldots,n$. If $b_{k+1}^{i}\geq\epsilon$, then $d_{k+1}=-\frac{g_{k+1}}{b_{k+1}^{i}}$; otherwise, $d_{k+1}=-g_{k+1}$, where $\epsilon_{b}$ denotes a minor positive constant.

In practical scenarios, the pursuit of iterative algorithms with rapid convergence rates remains a consistent focus of research [13, 14, 15, 16, 17, 18]. Among these, the inertial extrapolation method has gained significant attention as an acceleration approach [19, 20]. These methods rely on iterative schemes where each subsequent term is derived from the preceding two terms. This concept originated from Polyak’s seminal work [21] and is rooted when dealing with a dissipative dynamical system that is second-order in time, we employ an implicit discretization technique.This arrangement is delineated by the subsequent differential equation:

Let $\eta>0$ and $f:\mathbb{R}^{n}\rightarrow\mathbb{R}^{n}$ be a differentiable function. The discretization of arrangement (2.13) enables the determination of the subsequent term $x_{k+1}$ given the previous terms $x_{k-1}$ and $x_{k}$ using an iterative scheme. This discretization process is crucial in the development of inertial extrapolation methods, which aim to enhance the convergence rate of iterative algorithms. By leveraging information from the two preceding terms, these methods efficiently approach the desired solution. The specific form of the discretization and the iterative scheme employed depend on the particular inertial extrapolation method in use. However, the underlying principle remains the same: harnessing the momentum generated by previous iterations to expedite the convergence process.
Arrangement (2.13) is discretized in such a manner that, possessing the terms $x_{k-1}$ and $x_{k}$, the subsequent term $x_{k+1}$ can be ascertained utilizing

where $q$ is the pace magnitude. Equation (2.14) begets the ensuing iterative plan:

where $\rho=1-\eta q$, $\alpha=q^{2}$, and $\rho(x_{k}-x_{k-1})$ is denoted as the inertial extrapolation term aimed at accelerating the convergence of the series produced from Equation 2.15.

Recently, a plethora of algorithms rooted in the inertial acceleration technique, initially proposed by Alvarez [22], have emerged to tackle a diverse array of optimization and engineering challenges. Notably, the fusion of inertial acceleration with finely crafted search direction strategies, as advocated by Sun and Liu [23], has garnered significant attention.Building upon this groundwork, [24] has made significant contributions for addressing nonlinear monotone equations. The convergence being confined and meeting certain descent conditions:

The inertial extrapolation step magnitude is selected as:

In this case, where $\alpha\in[0,1)$, the stipulation in (2.16) is satisfied automatically.As is commonly understood, the aforementioned condition is stronger than the one typically used:

There are some algorithms obtained through inertial extrapolation techniques [25, 20, 26, 19], it becomes apparent that by setting

Considering these two algorithms, we can observe that with increasing iterations, the discrepancy between $x_{k-1}$ and $x_{k}$ becomes smaller and smaller. Therefore, the step size determined by (2.16) will achieve better convergence compared to (2.18).

From the preceding process, we proposed the following algorithm:

Based on extrapolation techniques, we have accelerated the convergence speed of the DMBFGS3 algorithm using extrapolation techniques. The resulting algorithm is as follows:

Remark

• •

  We introduce a novel approach based on the theory of minimal alterations in secant updates proposed by Dennis and Wolkowicz [10]. In our method, we enforce the condition that the updating matrices are diagonal, which serves as the approximation to the Hessian. We have improved the search strategy for $y_{k}$ and achieved linear convergence for the entire algorithm.

• •

  Our experimental results demonstrate that the algorithm proposed by combining these two methods is more efficient and robust compared to other benchmark algorithms.

The subsequent assumptions are necessary for this article:

• •

  The function $f$ exhibits continuous differentiability over the set $\Omega$, with $\|g(m)-g(n)\|\leq L\|m-n\|$ for all $m,n\in\Omega$, where $L$ is a non-negative constant.

• •

  The set of level points, denoted as $X=\{t\mid f(t)\leq f(t_{0})\}$, is enclosed within the set $\Omega$.

• •

  Indicating that $v_{1},v_{2}>0,v_{1},v_{2}\in\mathbb{R}^{n}$, and the function $f$ is convex, such that $v_{1}\|t\|^{2}\leq t^{T}G(x)t\leq v_{2}\|t\|^{2}$ for all $t\in\mathbb{R}^{n}$.

• •

  There exists $x^{*}\in\Omega$ which represents the solution to the minimization problem stated in (1.1).