Blue: f(x) (true convex function)Red: each tangent (linear under-estimators)Green: model m(x) = max(tangents)● Marker: suggested minimizer of the model
Click on the plot to add a tangent at that x. Or press Step to automatically add a tangent at the current model minimizer (Kelley iteration).
How to use / intuition
The algorithm keeps linear under-estimators (tangents). The model m(x) is the maximum of these lines.
Minimizing m(x) is easy (we numerically search). The minimizer suggests where the true function might be small.
Adding a tangent at that suggestion refines the model — repeated steps converge toward the true minimizer.