ES.1802, October 13, 2026

Overview

The ES.1802 class reviews multivariable critical-point classification, then introduces constrained optimization through level curves, boundary checks, and the Lagrange multiplier condition ∇f = λ∇g. It then previews an ES.1806 linear algebra connection: deriving the projection of a vector onto a subspace using the matrix P = A(AᵀA)⁻¹Aᵀ.

Key takeaways

For a symmetric Hessian at a critical point, all-positive eigenvalues indicate a local minimum, all-negative eigenvalues a local maximum, and mixed positive and negative eigenvalues a saddle; a zero eigenvalue makes the test inconclusive.
A constrained optimum occurs where the objective and constraint level curves are tangent, yielding the Lagrange multiplier condition ∇f = λ∇g.
To find the point on xy = 1 closest to the origin, minimize x² + y² subject to xy − 1 = 0; minimizing squared distance avoids the square root without changing the minimizer.
Constrained optimization over a bounded domain requires checking both interior candidates and boundary endpoints, as illustrated by y = 1 − x² in the first quadrant.
Projection onto a vector A is A(AᵀB)/(AᵀA); projection onto the column space of a full-column-rank matrix A generalizes this to A(AᵀA)⁻¹AᵀB.

Chapters

0:00 Hessian Eigenvalues Classify Multivariable Critical Points

A critical point in several variables satisfies ∇f = 0, the multivariable analogue of a zero first derivative.
The Hessian is the matrix of second partial derivatives; equality of mixed partials makes it symmetric.
Positive Hessian eigenvalues indicate a local minimum, negative eigenvalues a local maximum, and mixed signs a saddle.

6:00 Zero Hessian Eigenvalues and the Quadratic Approximation

If a Hessian eigenvalue is zero, the second-derivative test is inconclusive; other analysis is needed to classify the point.
Near a critical point, the quadratic approximation captures curvature that the tangent-plane approximation misses.
A positive-definite Hessian makes the quadratic form positive in every nonzero direction, producing local bowl-shaped behavior and a minimum.

11:10 Optimization Over Domains and Their Boundaries

A global maximum or minimum is defined relative to a specified domain, unlike a local extremum.
Constraints can define regions such as the unit sphere x² + y² + z² ≤ 1 or the first octant.
Optima may occur inside a domain or on its boundary, so boundary points must be included in the search.

15:15 Set Up the Closest-Point Problem on xy = 1

The example asks for the point on the hyperbola xy = 1 closest to the origin.
Distance is √(x² + y²), and minimizing it is equivalent to minimizing the simpler objective f(x,y) = x² + y².
The constraint is written as g(x,y) = xy − 1 = 0, keeping the objective and constraint separate rather than substituting one variable.

21:00 Tangency of Level Curves Leads to Lagrange Multipliers

Level curves of x² + y² are circles centered at the origin, while xy = 1 is the constraint curve.
At a constrained optimum, the objective’s level curve is tangent to the constraint, so their gradients are parallel.
The parallel-gradient condition is written ∇f = λ∇g, where λ is a scalar multiplier.

27:20 Solve Multiplier Equations and Check Domain Endpoints

For two variables, ∇f = λ∇g supplies two equations, which are solved together with the constraint equation.
Lagrange multiplier candidates should be evaluated and compared to determine which gives the desired extremum.
For a bounded curve segment—such as y = 1 − x² in the first quadrant—test endpoints as well as interior critical points.

33:00 Rewrite Projection onto a Vector with Dot Products

The projection of B onto a nonzero vector A is (A·B)/(A·A) times A.
The dot product A·A equals the squared magnitude of A, avoiding a square root in the projection formula.
Using matrix notation, the same vector projection can be written A(AᵀA)⁻¹AᵀB.

40:30 Project a Vector onto a Matrix Column Space

The column space of A consists of all linear combinations of its columns; multiplying A by a coefficient vector produces such a combination.
A vector B outside that space can be approximated by its projection P onto the column space.
For linearly independent columns, the projection is P = A(AᵀA)⁻¹AᵀB, extending the single-vector projection formula to a subspace.

47:00 Interpret the Projection Matrix P

The matrix A(AᵀA)⁻¹Aᵀ is the projection matrix, and multiplying it by B returns the component of B lying in A’s column space.
The projection formula generalizes the earlier vector projection and previews a useful ES.1806 linear algebra technique.
A’s columns span the set of outputs obtainable from A times coefficient vectors; projecting B finds the closest vector in that set.

Summary, takeaways, and chapters were generated by AI from the video's transcript and may contain errors. The video belongs to its creator, ChristianAltoSax.