Linear algebra¶

Weeks¶

Week 10¶

Spine: Boyd & Vandenberghe, Introduction to Applied Linear Algebra

Reading: Ch. 1-3, pp. 1-68: vectors; linear functions; norm/distance

Know: Represent state/data as vectors; use norms, inner products, angles and linear maps.

Reconstruct: Derive Cauchy-Schwarz geometrically; derive projection onto a vector.

Do: Represent a multi-sensor robot state and design three different norms; show how each changes what 'closest state' means.

Defend: Why is the choice of norm a modeling decision, not merely notation?

Gate: Closed-book vector identities + one modeling problem where units differ across coordinates.

Source: source

Week 11¶

Spine: Boyd & Vandenberghe, Introduction to Applied Linear Algebra

Reading: Ch. 5-6, approx. pp. 89-128: linear independence; matrices

Know: Diagnose redundancy, basis choice, rank and matrix-as-map thinking.

Reconstruct: Regenerate definitions of independence/span/basis; prove uniqueness of coordinates in a basis.

Do: Construct a sensor matrix with deliberate redundancy; identify which sensor combinations are informationally redundant.

Defend: What is the operational meaning of rank loss?

Gate: reduce an unseen matrix model to independent degrees of freedom and defend the basis.

Source: source

Week 12¶

Spine: Boyd & Vandenberghe, Introduction to Applied Linear Algebra

Reading: Ch. 8-11, pp. 147-224: linear equations; linear dynamical systems; multiplication; inverses

Know: Solve linear systems; reason about reachability over repeated matrix dynamics; interpret inverses and condition.

Reconstruct: Derive solution conditions for Ax=b in terms of span; derive x(t)=A^t x(0) for a discrete linear system.

Do: Model a 4-compartment resource flow system and explore stable, unstable, and conserved modes numerically.

Defend: Why can an inverse exist mathematically yet be useless numerically?

Gate: Closed-book solve + condition/sensitivity explanation on an almost-singular system.

Source: source

Week 13¶

Spine: Boyd & Vandenberghe, Introduction to Applied Linear Algebra

Reading: Ch. 12-13, pp. 225-284: least squares; data fitting

Know: Derive least squares as projection; understand residual orthogonality and regression geometry.

Reconstruct: Derive normal equations from minimizing ||Ax-b||² and derive QR-based solution conceptually.

Do: Fit a calibration curve with outliers; compare normal equations and QR numerically.

Defend: What exactly is being optimized in least squares, and what assumptions make that meaningful?

Gate: derive least squares two ways (geometry + calculus) and identify when both views fail.

Source: source

Week 14¶

Spine: Boyd & Vandenberghe, Introduction to Applied Linear Algebra + supplement

Reading: Ch. 16 constrained least squares, pp. 339-356; Ch. 18 nonlinear least squares, pp. 381-418; supplement eigen/SVD from FNC Ch. 7

Know: Handle constraints, nonlinear fitting, SVD/eigenstructure, low-rank approximation and ill-conditioning.

Reconstruct: Derive rank-1 SVD approximation intuition; derive Gauss-Newton linearization for nonlinear least squares.

Do: Compress a synthetic sensor dataset with SVD and quantify reconstruction/error tradeoff; then fit a nonlinear model.

Defend: Why are small singular values simultaneously useful information and a numerical warning?

Gate: diagnose rank, conditioning, identifiability and compression on a fresh dataset.

Source: source

Exit gate¶

Closed-book: 120 min: projection, least squares, rank/nullspace, basis, eigenmode and SVD questions.

Novel problem: Given an unknown sensor/actuator matrix, diagnose redundancy, identifiability and conditioning.

Artifact: Implement QR least squares + SVD diagnostic; compare against naive inverse/normal equations.

Defend: Explain rank loss, small singular values, basis dependence and coordinate-invariant statements.

Pass criterion: Pass if learner predicts numerical failure before running code and justifies representation choices.

Transfer problems¶

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

1. Projection: Project an observation onto a model subspace and interpret the residual physically.

2. Rank: Create a matrix whose columns look different but are linearly dependent; explain the hidden redundancy.

3. Nullspace: Find a nonzero actuator command that produces zero net output in a toy mechanism.

4. Conditioning: Construct two nearly collinear columns and measure solution sensitivity to 0.1% data perturbation.

5. Dynamics: Analyze repeated x_{k+1}=Ax_k for modes that decay, persist, or grow.

6. Least squares: Fit an overdetermined calibration system and verify residual orthogonality.

7. QR vs normal: Solve the same ill-conditioned least-squares problem both ways and compare.

8. SVD: Find the best rank-1 approximation to a small dataset and quantify lost variance/information.

9. Identifiability: Given y=Ax with fewer independent measurements than state dimensions, describe all indistinguishable states.

10. Constraints: Solve a constrained least-squares allocation problem and interpret active constraints.

Textbooks¶

See the five-book resource page.