Skip to main content

00. Optimization::

// PROBLEM_TO_ALGORITHM — a map for rereading an optimization problem from its definition through algorithm selection

Optimization is the process of expressing "what can be changed, what counts as good, and what must be respected" in equations, then exploiting that structure to find a solution. This section connects problem formulation, gradient descent, Newton and quasi-Newton methods, conjugate gradient, and surrogate-based optimization in one sequence.

If you are reading for the first time, align the language with the six Background articles before reading the seven Core articles. If you already know calculus and linear algebra, you can begin with Problem Formulation.

Background (6 articles)

Core optimization (7 articles on problems and algorithms)

Coverage checklist

Every requested topic appears explicitly at the locations below.

AreaIncluded topicsLocation
Problem formulationdecision variables, objective function, constraints, feasible region07
Mathematical foundationsgradient/Jacobian/Hessian, Taylor approximation, convexity, norm/inner product, orthogonality, matrix decomposition08
Optimality conditionsfirst-order, second-order, positive definiteness09
Optimization algorithmsgradient descent, line search, momentum, Nesterov, Newton, DFP, BFGS, coordinate descent, conjugate gradient, preconditioning10–12
Expensive optimizationDOE, LHS, surrogate modeling, surrogate-based optimization13
Backgroundnotation, calculus, linear algebra, numerical analysis, probability/statistics, engineering modeling01–06

Reading perspective

  • Formulation is half the problem — if variables and constraints are chosen poorly, even a good algorithm only solves the wrong problem quickly.
  • Derivatives are local information — the gradient tells you which direction to move, and the Hessian tells you about curvature.
  • Linear algebra is geometry — norms, inner products, orthogonality, and eigenvalues form the language for explaining algorithmic convergence.
  • Numerical analysis is practicality — without floating point, condition numbers, and convergence, a method that works on paper breaks in code.
  • Expensive optimization is a budget problem — when one objective evaluation is costly, samples must be chosen more intelligently and a surrogate must be built.