The purpose of approximation is finding optimal point \(x^*\) i.e. \(\nabla F(x^*) = 0\). We need a step/search direction \(\Delta x\) and step size \(t\). Taylor approximation has polynomial arguments that is a step and parameters of derivatives at the start point. The first degree of Taylor approximation has one adding term from start point \((x_0, F(x_0))\). The adding term \(\nabla F(x) \Delta x\) is consistent with a parameter (gradient \(\nabla F(x)\)) and a argument (step \(\Delta x\)). The Taylor approximation does approximate \(F(x + \Delta x)\) for any search direction \(\Delta x\). We want to choose \(\Delta x\) for the direction to the optimal point.

The adding term of Taylor approximation \(\nabla F(x) \Delta x\) have level curve (level line). The smallest Euclidean norm of the level curve is achieved at the tangent. The gradient descent set the step to the gradient \(\nabla F(x)\). This makes the adding term biggest with Euclidean norm \(\Vert \nabla F(x) \Vert^2\) i.e. dual norm \(\Vert \nabla F(x) \Vert_*\).

Newton’s method is second degree of Taylor approximation \(F(x_0+\Delta x) \approx F(x_0) + \nabla F(x) \Delta x + 1/2\Delta x^T H \Delta x\). We want to find \(\Delta x\) to minimize the second degree of Taylor approximation. In this case, the minimizing step is tangent of first adding term \(\nabla F(x) \Delta x\) and second adding term \(\Delta x^T H \Delta x\) i.e. Steepest descent in H norm \(\Vert \cdot \Vert _H\). The newton’s method can be thought as approximation of gradient \(\nabla F(x)\). \(\nabla F(x_0 + \Delta x) \approx \nabla F(x_0) + H \Delta x = 0,\ \Delta x = -H^{-1} \nabla F(x_0)\). This is also the derivative of second degree of Taylor approximation with respect to \(\Delta x\).

But the Taylor approximation is local. In addition to a step, a step size is needed. A step size determines how far the step taken. Backtracking line search has two constant parameters 0 < \(\alpha\) < 0.5, 0 < \(\beta\) < 1. The approximation is below the convex function. \(\alpha\) tilts the slope i.e. gradient upside and the tilted approximation meets the convex function. \(\beta\) is the update rate of the step size until the the amount of the step is less than the point that tilted approximation meeets the convex function.

Jun Kang
Clinical Assistant Professor of Hospital Pathology

My research interests include pathology, oncology and statistics.