Chapter 5: Beyond the black-box model
- Section5.1: propose a first order method with a 1/t^2 convergence rate, despite the non-smoothness
- Section5.2: the function is the maximum of smooth functions
- Section5.3: a concise description of Interior Point Methods
Section 5.1: Sum of a smooth and a simple non-smooth term
the problem:
f is convex and beta-smooth, and g is convex.ISTA(Iterative Shrinkage-Thresholding Algorthm)
- 根據Gradient Descent on the smooth function f
xt+1的表達式 - 結合這個問題minimize f+g得到xt+1的表達式
這個就是ISTA算法的迭代式。 - 查閱論文可得到證明此算法的收斂率(函數convex並且smooth,收斂率爲1/t, 函數只convex時,收斂率爲1/根號t)
- 這個算法需要假設g is simple, 因爲計算xt+1本身是一個凸優化問題,當假設g is simple時,計算xt+1可以通過解決n個在一維空間上的凸優化問題。
- 根據Gradient Descent on the smooth function f
FISTA(Fast ISTA)
- 結合Nesterov’s Accelerated Gradient Descent得到對應的
- 收斂率(證明查詢相關論文)
- 結合Nesterov’s Accelerated Gradient Descent得到對應的
CMD and RDA
- ISTA and FISTA assume smoothness in the Euclidean metric.
- CMD and RDA use these ideas in a non-Euclidean metric
總結
- 當函數可以分解爲sum of f and g(f is convex and smooth, g is convex)時,這個凸優化問題的收斂率比只知道函數爲凸時的收斂率小。
未完待續