You already know: gradient descent, curvature and learning rates (Optimization, Jacobians and Taylor); maximum likelihood (Probability); joint distributions and the Gaussian (previous lesson).
Why this matters for Atlas. Convexity is the property that makes optimization trustworthy: for a convex function every local minimum is the global minimum, so linear and logistic regression, SVMs and many classical models have guaranteed solutions. Neural networks aren't convex, which is why their training needs the heuristics of D4. Lagrange multipliers turn constrained problems into unconstrained ones, and they show up in surprising places: the softmax and maximum-entropy distributions, the SVM's support vectors, and the KL-constrained objectives behind RLHF and DPO (D6).
1. Convex sets and functions
The question. When can you trust that the minimum you found is the minimum?
Start from nothing. A function is convex if the line segment between any two points on its graph lies on or above the graph: f(tx + (1−t)y) ≤ t f(x) + (1−t) f(y). Equivalently (for twice-differentiable f) its Hessian is positive semi-definite everywhere: curvature is never negative in any direction. Consequences: any local minimum is global, gradient descent with a small enough step converges to it, and sums of convex functions (plus regularizers like ‖w‖²) stay convex. Squared error of a linear model, logistic loss, and hinge loss are convex in the weights; a neural network's loss isn't, because composing nonlinear layers bends the landscape.