⚡ electrical engineering

Robust Constrained Optimization via Sliding Mode Control

This paper proposes a robust sliding mode control framework that reformulates Karush-Kuhn-Tucker conditions as a dynamical system to achieve finite-time convergence and exact constraint satisfaction for equality-constrained optimization, even in the presence of disturbances and without requiring objective convexity.

Original authors: Shyam Kamal, Baby Diana, Sunidhi Pandey, Sandip Ghosh, Thach Ngoc Dinh

Published 2026-05-01

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Shyam Kamal, Baby Diana, Sunidhi Pandey, Sandip Ghosh, Thach Ngoc Dinh

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to find the lowest point in a vast, foggy valley (this is your optimization problem). However, there are strict rules you must follow: you cannot cross a river, you must stay on a specific path, or you must avoid a wall (these are your constraints).

Most traditional methods for solving this are like a hiker who walks slowly downhill. They are careful and smooth, but they often get stuck in small dips (local minima) or, worse, they might wander too close to the riverbank and accidentally step into the water before slowly backing out. They take a very long time to realize they are off-track.

This paper proposes a new, much more aggressive strategy using a concept called Sliding Mode Control (SMC). Here is how it works, explained through simple analogies:

1. The "Magnetic Wall" Analogy (Enforcing Rules)

In traditional methods, the "rules" (constraints) are like soft fences. If you get too close, you feel a gentle push back. You might still wobble near the fence.

In this new method, the rules act like a magnetic wall or a slippery slide.

The Goal: The system is designed to slam the "violation" (how far you are from the rule) down to zero instantly (in finite time).
How it works: Imagine a robot trying to walk along a tightrope. If it starts to lean left, a powerful, sudden force yanks it back to the center immediately. It doesn't just "try" to stay on the line; it forces itself onto the line and stays there.
The Result: The system reaches the "safe zone" (where the constraints are perfectly met) in a guaranteed, short amount of time, regardless of whether the valley is smooth or full of bumps.

2. The "Two-Phase" Journey

The paper describes the process in two distinct phases, like a race car driver:

Phase 1: The Crash Course (Reaching Phase): The car drives aggressively toward the track (the constraint line). It might bounce a little, but a powerful controller forces it onto the track very quickly.
Phase 2: The Smooth Glide (Sliding Phase): Once the car is on the track, the controller switches modes. It stops fighting the track and simply guides the car along the track toward the finish line (the optimal solution). Because the car is now "locked" to the track, it can't drift off, even if the wind (disturbances) blows hard.

3. Handling the "Bumpy Road" (Robustness)

Real life is messy. There are wind gusts, uneven ground, and measurement errors (noise).

Old Methods: If a strong wind hits a traditional hiker, they might get blown off the path entirely and take a long time to recover.
This New Method: Because the "magnetic wall" is so strong, if a wind gust tries to push the system off the constraint line, the controller immediately slams it back. The paper proves mathematically that the system is immune to these "matched" disturbances. It's like a ship with a rudder so powerful that no wave can push it off its course.

4. Speeding Up the Finish (Finite-Time Convergence)

Standard methods often get slower and slower as they get closer to the solution, theoretically taking forever to actually arrive.

The Innovation: The authors introduce a special "turbo boost" (called Nonsingular Terminal Sliding Mode). This ensures that the system doesn't just get close to the solution; it actually arrives at the exact solution in a finite, predictable amount of time. It's the difference between a car that slows down to a crawl as it approaches a stop sign versus one that brakes hard and stops exactly at the line.

5. Real-World Tests (The "Proof")

The authors tested this "magnetic wall" strategy on several problems to show it works better than the old "soft fence" methods:

The Maze: In a robot navigation task, traditional methods got stuck in a "dead end" (a local minimum) or moved too slowly along the walls. The new method forced the robot through the narrow passage and to the goal quickly.
The Puzzle: They used it to solve a "Shidoku" puzzle (a 4x4 Sudoku). The system quickly adjusted the numbers to fit the rules, ignoring the fact that the puzzle had some redundant rules that usually confuse computers.
The Team: They tested a group of robots trying to agree on a single value (consensus). Even with noise and errors, they all snapped into agreement quickly and stayed there.

Summary

Think of this paper as a new set of instructions for a robot trying to solve a puzzle while following strict rules.

Old way: "Walk carefully, try to stay on the line, and hope you don't fall off." (Slow, fragile, often inaccurate).
New way: "If you are off the line, get snapped back to it immediately. Once you are on the line, slide straight to the finish. Ignore the wind." (Fast, unbreakable, and precise).

The paper claims this method is superior because it guarantees the rules are followed exactly and quickly, and it keeps working even when the environment is noisy or chaotic.

1. Problem Statement

The paper addresses the challenge of solving equality-constrained optimization problems in continuous time, specifically:
$\min_{x \in \mathbb{R}^n} \phi(x) \quad \text{subject to} \quad h(x) = 0$
where $\phi(x)$ is the objective function (potentially non-convex) and $h(x) = 0$ represents $m$ equality constraints.

Limitations of Existing Methods:

Asymptotic Convergence: Classical continuous-time methods (e.g., Primal-Dual Gradient Dynamics (PDGD), PI-Controlled Multipliers Optimization (PI-CMO)) typically guarantee only asymptotic stability. They approach the feasible set and optimal solution as $t \to \infty$ , which is slow and insufficient for real-time safety-critical applications.
Approximate Feasibility: Many methods only satisfy constraints approximately or require convexity assumptions to guarantee convergence.
Lack of Robustness: Existing approaches often suffer from instability or significant performance degradation in the presence of matched disturbances, structural uncertainties, or measurement noise.

2. Methodology

The authors propose a novel framework that reformulates the optimization problem as a control-theoretic system using Sliding Mode Control (SMC).

A. Control-Theoretic Reformulation

States: The optimization variables $x(t)$ are treated as system states.
Control Inputs: The Lagrange multipliers $\lambda(x(t))$ are treated as control inputs.
Outputs: The constraint violations $y(t) = h(x(t))$ are treated as regulated outputs.
Dynamics: The system is modeled as:
$\dot{x} = -\nabla \phi(x) - J_h(x)^\top \lambda$
where $J_h(x)$ is the Jacobian of the constraints. The goal is to design a feedback law $\lambda$ that drives $h(x)$ to zero in finite time.

B. Sliding Mode Control (SMC) Design

The core contribution is the design of a discontinuous control law for $\lambda$ to enforce the sliding manifold $S(x) = \{x \mid h(x) = 0\}$ .

Equivalent Control ( $\lambda_{eq}$ ): Derived to maintain the system on the manifold ( $\dot{h}=0$ ). It acts as a projected gradient flow:
$\lambda_{eq} = -(J_h J_h^\top)^{-1} J_h \nabla \phi$
Switching Control ( $\lambda_{sw}$ ): A discontinuous term designed to drive the system to the manifold in finite time:
$\lambda_{sw} = (J_h J_h^\top)^{-1} K \text{sgn}(h(x))$
where $K$ is a positive-definite gain matrix.
Resulting Dynamics: The closed-loop system exhibits two phases:
- Reaching Phase: The switching term drives $h(x) \to 0$ in finite time.
- Sliding Phase: Once on the manifold, the system evolves via projected gradient flow $\dot{x} = -P(x)\nabla \phi(x)$ , ensuring asymptotic convergence to the KKT point.

C. Finite-Time Optimality (NTSMC)

For strongly convex problems, the authors introduce a Nonsingular Terminal Sliding Mode (NTSMC) strategy.

They define a higher-order sliding variable $S = z_1 + \beta |z_2|^\gamma \text{sgn}(z_2)$ , where $z_1 = h(x)$ and $z_2 = \dot{h}(x)$ .
They replace the standard gradient flow with a normed gradient flow (dividing by the norm of the gradient).
This ensures that the system reaches the global optimum in finite time, not just the constraint manifold.

D. Robustness Analysis

The paper rigorously analyzes robustness using Lyapunov theory and differential inclusions (Filippov solutions):

Matched Disturbances: The SMC law rejects disturbances lying in the range of $J_h^\top$ (matched uncertainties) completely, maintaining finite-time convergence.
Structural Uncertainty: The method remains stable even if the constraint Jacobian $J_h$ has bounded time-varying perturbations, provided the rank condition is preserved.
Measurement Noise: Under $L_\infty$ -bounded noise, the system converges to a small boundary layer around the constraint manifold, rather than diverging.

3. Key Contributions

Finite-Time Constraint Enforcement: Unlike classical methods, the proposed SMC framework guarantees that equality constraints are satisfied exactly within a finite time $T$ , regardless of the convexity of the objective function.
Robustness: The method provides inherent robustness against matched disturbances, model uncertainties, and measurement noise, a feature often lacking in gradient-based continuous-time optimizers.
Finite-Time Global Optimality: For convex problems, the integration of NTSMC and normed gradient flows ensures convergence to the optimal solution in finite time.
Rigorous Stability Analysis: The paper provides a complete Lyapunov-based proof for finite-time reachability of the sliding manifold and asymptotic (or finite-time for convex) convergence to the KKT points.
Chattering Mitigation: Practical strategies are discussed, including the use of saturation functions ( $\text{sat}$ ) and the Super-Twisting Algorithm (STA) (in the Appendix) to smooth the control signal while preserving finite-time convergence properties.

4. Results and Validation

The authors validate the method through numerical simulations on diverse benchmark problems, comparing it against PDGD, PI-CMO, and Projected Gradient Flow (PGF).

Non-Convex Quadratic Optimization: Under external sinusoidal disturbances, PDGD and PI-CMO diverged or oscillated significantly. The proposed SMC method converged to the optimal point in 0.495 seconds, demonstrating superior disturbance rejection.
Obstacle Avoidance (Robotics): In a narrow passage with symmetric obstacles, the Artificial Potential Field (APF) method got trapped in a local minimum. The SMC method successfully navigated the robot to the goal in 0.41 seconds, enforcing safety constraints ( $h(x)=0$ ) in finite time.
Shidoku Puzzle (Combinatorial Optimization): The method solved a 4x4 Sudoku-like puzzle with redundant constraints (rank-deficient Jacobian). It achieved finite-time convergence to integer solutions, proving effectiveness even without the Linear Independence Constraint Qualification (LICQ).
Distributed Parameter Estimation: In a multi-agent consensus problem, the NTSMC approach ensured all agents reached the centralized optimal estimate in finite time ( $T_c \le 5.5$ s) despite matched disturbances, outperforming standard consensus algorithms.

5. Significance

This paper bridges the gap between control theory and optimization, offering a paradigm shift from asymptotic to finite-time constrained optimization.

Safety-Critical Applications: The ability to guarantee constraint satisfaction in finite time makes this approach ideal for robotics, autonomous systems, and power grids where safety violations cannot be tolerated even momentarily.
Robustness in Uncertain Environments: The method's resilience to noise and model errors makes it superior for real-world deployments where perfect models are unavailable.
Theoretical Advancement: It extends the applicability of SMC beyond traditional control systems to general optimization landscapes, including non-convex and rank-deficient scenarios, providing a new tool for solving complex constrained problems.

In conclusion, the proposed framework transforms constrained optimization into a robust control problem, achieving exact feasibility and optimality with guaranteed speed and resilience, overcoming the fundamental limitations of classical continuous-time optimization methods.