Stanford AA203 Optimal and Learning-Based Control | Spring 2026 | Lecture 2: Optimization Theory

Optimization techniques rely on necessary conditions—such as the gradient equaling zero—to identify local minima, which can be extended via Lagrange multipliers to handle constraints, or simplified by identifying convex functions where local minima act as global minima.

• NOC theorem — For unconstrained local minima, the gradient must be zero; if the function is twice differentiable, the matrix of second-order derivatives must be positive semi-definite [13:01].

• Sufficient conditions — A point is a guaranteed local minimum if the gradient is zero and the second-order derivative matrix is positive definite [17:55].

• Convex functions — These shapes simplify computation because they ensure any found local minimum acts as the global minimum [22:28].

• Iterative descent — Optimization algorithms update guesses by moving in directions where the gradient product is negative, ensuring consistent value reduction [39:04].

• Step size — The parameter alpha controls jump distance; large values risk overshooting, while diminishing values ensure fine-tuned convergence [33:32].

• Curvature info — Incorporating second-order derivatives adjusts the descent direction for elongated level sets, improving efficiency despite higher computation costs [45:02].

• Lagrange multipliers — This method augments the objective function with constraints scaled by additional variables to convert constrained problems into unconstrained ones [1:17:21].

• Geometric interpretation — Optimal points occur at the tangency between the objective function’s level sets and the constraint boundary [1:09:00].

• How do gradient descent methods differentiate between local and global minima?

• What determines the computational cost of using the Hessian matrix in an optimization algorithm?