Why Care About Convexity? When a Local Answer Is the Best Answer

A ball that can only feel the slope under it rolls downhill and stops on flat ground, but is that the lowest point? This lesson shows why convexity turns that local clue into a global answer: the chord test, the proof that every local minimum of a convex function is global, the tangent-line bound that makes a zero gradient a certificate, and why gradient descent on a bowl reaches the bottom at a predictable speed. Then it builds real convex problems: least squares, the lasso and linear programs, and shows why neural networks are the famous exception. Background: derivatives and a little vector notation. Chapters 0:00 A ball that only feels the slope 0:15 Twenty balls, three valleys 0:36 Reshape the ground into a bowl 1:12 The chord test 1:53 Convex regions: the segment test 2:31 Local and global minima 2:48 Proof: every local minimum is global 3:49 Two warnings: no minimum, flat bottoms 4:23 Tangent lines stay below the graph 4:39 Proof: a zero slope is a certificate 5:47 Two variables: the gradient 6:23 A worked example and the Hessian 7:02 The saddle: flat proves nothing 7:21 Gradient descent on a bowl 7:50 How fast: the bend between 1 and 3 8:40 Constraints and a dented region 9:23 Building convex problems 10:15 Least squares and the lasso 11:02 Linear programs and neural networks 11:34 Why care about convexity? The dual Sources and further reading S. Boyd and L. Vandenberghe, Convex Optimization, Cambridge University Press (2004), ch. 2-5 and 9 Y. Nesterov, Lectures on Convex Optimization, Springer (2018), section 2.1 A. Beck, First-Order Methods in Optimization, SIAM (2017), Theorem 10.21 R. Tibshirani, Regression shrinkage and selection via the lasso, JRSS B 58 (1996) Related episode, Support Vector Machines:    • Support Vector Machine: Margins & Kernels ...   Narration uses a synthetic voice (Qwen3-TTS, cloned from the channel's usual voice). Soundtrack: original instrumental composition generated with ACE-Step 1.5.