Introduction

In the study of optimization and control systems, the problem of finding the minimum of a function $F(x)$ is often approached through gradient-based methods. Following the foundational work in \cite{1, 2, 3}, we consider the minimization of a function $F(x) = F(x_1, \dots, x_n)$ where the optimal point is denoted as $x^$, such that $F(x^) = \min F(x)$. A common approach is to utilize a continuous descent process defined by the differential equation:
$$\dot{x} = -A \nabla F(x)$$
where $A$ is a positive definite matrix. Under appropriate conditions, the trajectory $x(t)$ converges to $x^*$ as $t \to \infty$.

However, traditional gradient methods often suffer from slow convergence or sensitivity to the choice of step size $h$ when discretized. In discrete form, the iteration is typically expressed as:
$$x((k+1)h) = x(kh) - A \nabla F(x(kh)) \cdot h$$
As $h \to 0$, this approximation approaches the continuous trajectory, but for finite $h$, the error is $O(h)$. To improve the convergence properties and ensure finite-time stability, we introduce a modified approach based on sliding mode control principles \cite{4, 5}.

Sliding Mode Optimization

Let us define a switching surface $s(x) = 0$ that guides the system toward the optimum. We consider a Lyapunov function candidate $V(x) = F(x) - F(x^)$, where $F(x) > F(x^)$ for all $x \neq x^*$. To ensure that the system reaches the surface $s(x) = 0$ and remains there, we require that the derivative of the switching function satisfies:
$$\frac{ds(x)}{dt} = \langle \nabla s(x), \dot{x} \rangle < 0$$
when $s(x) > 0$, and conversely for $s(x) < 0$.

We propose a control law of the form:
$$\dot{x} = -K \nabla F(x) + b(x) u(x)$$
where $K$ is a constant gain, $b(x)$ is a vector function, and $u(x)$ is a switching control signal. Specifically, we can define:
$$u(x) = -\text{sgn } s(x)$$
$$b(x) = M \nabla s(x)$$
This leads to the following dynamics:
$$\dot{x} = -K \nabla F(x) - M \nabla s(x) \cdot \text{sgn } s(x)$$
where $M$ is chosen such that the condition for the existence of a sliding mode is satisfied:
$$M > \frac{|K \langle \nabla s(x), \nabla F(x) \rangle|}{|\nabla s(x)|^2}$$

Stability and Convergence

To analyze the stability of the equilibrium point $x^*$, we examine the behavior of the system on the sliding surface $s(x) = 0$. According to the method of equivalent control \cite{6, 7}, the motion on the surface is governed by:
$$\dot{x} = -K \nabla F(x) + K \frac{\langle \nabla s(x), \nabla F(x) \rangle}{|\nabla s(x)|^2} \nabla s(x)$$
This expression represents the projection of the gradient vector onto the tangent plane of the switching surface. By choosing $s(x)$ such that it relates to the gradient of the objective function, for example, $s(x) = \langle c, \nabla F(x) \rangle$, we can ensure that the trajectory moves directly toward the minimum.

The stability of this process can be verified using the Lyapunov function $V(x) = F(x) - F(x^)$. Calculating the time derivative along the trajectories of the system, we obtain:
$$\dot{V}(x) = \langle \nabla F(x), \dot{x} \rangle = -K |\nabla F(x)|^2 + K \frac{\langle \nabla s(x), \nabla F(x) \rangle^2}{|\nabla s(x)|^2}$$
By the Cauchy-Schwarz inequality, $\langle \nabla s(x), \nabla F(x) \rangle^2 \leq |\nabla s(x)|^2 |\nabla F(x)|^2$, which implies $\dot{V}(x) \leq 0$. The equality holds only when $\nabla F(x)$ is collinear with $\nabla s(x)$ or when $\nabla F(x) = 0$, corresponding to the optimum $x^$.

Numerical Results and Applications

The proposed method was tested on several benchmark optimization problems, including constrained linear and non-linear programming.

[TABLE:1]

The results in Table 1 demonstrate the convergence of the coordinates $x_i$ to their optimal values. For instance, in a test case involving four variables, the system reached the vicinity of the optimum within a small number of iterations, maintaining high precision even in the presence of constraints.

For constrained optimization problems of the form $\min F(x)$ subject to $L_i(x) \leq 0$, we utilize a penalty function approach or incorporate the constraints directly into the switching surface $s(x)$. As shown in the experiments, the sliding mode approach effectively handles these constraints, providing a robust path to the solution $x^* = (-0.16, -0.08, 4.04)$ for the tested non-linear system, which is consistent with results reported in \cite{10, 11, 12}.

Conclusion

The integration of sliding mode control into gradient-based optimization provides a powerful framework for solving complex mathematical programming problems. By defining appropriate switching surfaces and control laws, we ensure rapid convergence and robustness against numerical instabilities. This method is particularly suitable for real-time optimization where finite-time convergence to the sliding surface is a critical requirement.