Optimization¶

Weeks¶

Week 33¶

Spine: Boyd & Vandenberghe, Convex Optimization

Reading: Ch. 2 Convex sets, pp. 21-66; Ch. 3 convex functions, pp. 67-126

Know: Recognize convex structure; prove sets/functions convex; use epigraph and composition rules.

Reconstruct: Derive Jensen's inequality intuition and first-order convexity condition.

Do: Reformulate a resource-allocation problem until its convex/nonconvex pieces are explicit.

Defend: Why does convexity change what a local optimum means?

Gate: classify 12 problems/functions/sets with proof or counterexample.

Source: source

Week 34¶

Spine: Boyd & Vandenberghe

Reading: Ch. 4 Convex optimization problems, pp. 127-214; Ch. 5 Duality, pp. 215-288

Know: Formulate standard convex problems; understand Lagrangian, dual function, KKT and sensitivity.

Reconstruct: Derive Lagrange dual for a simple constrained problem and interpret dual variables as marginal values.

Do: Optimize energy/storage scheduling and interpret shadow prices under changing constraints.

Defend: When is a constraint's dual variable more informative than the primal solution?

Gate: derive KKT conditions on an unseen small problem and explain each term operationally.

Source: source

Week 35¶

Spine: Boyd & Vandenberghe

Reading: Ch. 6 approximation/fitting, pp. 291-350; Ch. 7 statistical estimation, pp. 351-396

Know: Connect optimization to fitting, regularization, estimation and design tradeoffs.

Reconstruct: Derive ridge regression objective and closed-form solution; explain regularization geometrically.

Do: Fit an inverse problem under noise with L1/L2 penalties; compare sparsity, bias and robustness.

Defend: How does a regularizer encode a prior belief or design preference?

Gate: choose objective/regularizer for a new problem and defend against two alternatives.

Source: source

Week 36¶

Spine: Boyd & Vandenberghe

Reading: Ch. 9-11, pp. 457-630: unconstrained, equality-constrained, interior-point methods

Know: Understand descent/Newton methods, line search, equality constraints and barrier/interior-point ideas.

Reconstruct: Derive Newton step from quadratic local model and equality-constrained KKT linear system.

Do: Implement gradient descent vs Newton on an ill-conditioned objective and visualize convergence geometry.

Defend: Why can a mathematically superior method be practically worse?

Gate: Module gate: formulate, solve, stress-test and objective-audit one real design problem.

Source: source

Exit gate¶

Closed-book: 120 min: convexity proof/counterexample; KKT derivation; dual interpretation; Newton step; regularization choice.

Novel problem: Formulate an unfamiliar engineering design as variables/objective/constraints, then expose nonconvexities.

Artifact: Solve with two methods; perturb constraints and interpret sensitivity/dual values.

Defend: Defend objective function, constraints, regularizer, solver and what the optimizer is blind to.

Pass criterion: Pass if no proxy objective is treated as the real-world goal without explicit justification.

Transfer problems¶

Try these before consulting solutions or asking for the complete answer.

1. Convex sets: Prove or disprove convexity of five engineering feasible regions.

2. Convex functions: Classify functions using definition/known composition rules.

3. Formulation: Turn a verbal allocation problem into variables/objective/constraints with units.

4. Duality: Derive dual of a small constrained quadratic/linear problem and interpret prices.

5. KKT: Solve a constrained problem from KKT conditions and identify active constraints.

6. Sensitivity: Perturb a resource constraint and compare objective change to dual prediction.

7. Regularization: Compare L1/L2 regularization on noisy inverse problem.

8. Newton: Compare gradient descent/Newton under ill-conditioning and line search choices.

9. Multiobjective: Construct a Pareto frontier for performance vs energy/cost.

10. Objective audit: Show a metric that can be gamed while the real-world goal worsens; redesign objective/constraints.

Textbooks¶

See the five-book resource page.