[LA] Lipschitz continuous gradient
2016-01-19 07:15
260 查看
Definitions
Summary
(1) ∥∇f(x)−∇f(y)∥2≤L∥x−y∥2(note that this does not assume convexity of f(x))
(2) L2xTx−f(x) is convex ( if dom(f) is convex )
(3) ∇2f(x)≤L⋅I ( if f(x) is twice differentiable )
(4) f(y)≤f(x)+∇f(x)⋅(y−x)+L2∥y−x∥22 ( if f(x) is convex)
(一) (1) to (2): From the equivalent definition of convexity, that
f(x) is convex iff (∇f(x)−∇f(y))T(x−y)≥0 and dom(f) is convex.
We just need to prove that
[(L⋅x−∇f(x))−(L⋅y−∇f(y))]T(x−y)≥0
[(L⋅x−∇f(x))−(L⋅y−∇f(y))]T(x−y)=L⋅∥x−y∥22−(∇f(x)−∇f(y))T(x−y)≥L⋅∥x−y∥22−∥∇f(x)−∇f(y)∥∥(x−y)∥≥L⋅∥x−y∥22−L⋅∥x−y∥22=0
(二)(2) to (3):
From the equivalent definition of convexity that
f(x) is convex iff ∇2f(x)≥0 if it is twice differentiable.
So
∇2(L2xTx−f(x))=L⋅I−∇2f(x)≥0
So
∇2f(x)≤L⋅I
(三)(3) to (4) ( f(x) does not need to be convex):
From (3), L⋅I−∇2f(x) is positive definite matrix. Then for any ω, we have
ωT(L⋅I−∇2f(x))ω≥0
i.e.
ω∇2f(x)ω≤L∥ω∥22
Then from the taylor decomposition, we have
f(y)=f(x)+∇f(x)⋅(y−x)+(y−x)T∇2f(z)(y−x)≤f(x)+∇f(x)⋅(y−x)+L∥y−x∥22
(四)(4) to (1):( f(x) needs to be convex)
Let y←x+t(∇f(y)−∇f(x)) plug in the equality:
f(x+t(∇f(y)−∇f(x)))≤f(x)+t∇f(x)⋅(∇f(y)−∇f(x))+Lt22∥∇f(y)−∇f(x)∥22
from the convexity of f(x), we have
f(x+t(∇f(y)−∇f(x)))≥f(y)+∇f(y)⋅(x−y+t(∇f(y)−∇f(x)))
combining these two equalities together,
f(y)−f(x)+∇f(y)⋅(x−y)+t∥∇f(y)−∇f(x)∥22≤Lt22∥∇f(y)−∇f(x)∥22
adding
f(x)−f(y)−∇f(y)⋅(x−y)≤L2∥y−x∥22
on both sides,
t∥∇f(y)−∇f(x)∥22≤Lt22∥∇f(y)−∇f(x)∥22+L2∥y−x∥22
Letting t=1L
⇒∥∇f(x)−∇f(y)∥22≤L2∥x−y∥22
If f(x) is convex and has Lipschitz derivative, which is equivalent to all these four conditions.
Summary
1. Definitions
∇f(x) is L- Lipschitz continuous, then we have(1) ∥∇f(x)−∇f(y)∥2≤L∥x−y∥2(note that this does not assume convexity of f(x))
(2) L2xTx−f(x) is convex ( if dom(f) is convex )
(3) ∇2f(x)≤L⋅I ( if f(x) is twice differentiable )
(4) f(y)≤f(x)+∇f(x)⋅(y−x)+L2∥y−x∥22 ( if f(x) is convex)
(一) (1) to (2): From the equivalent definition of convexity, that
f(x) is convex iff (∇f(x)−∇f(y))T(x−y)≥0 and dom(f) is convex.
We just need to prove that
[(L⋅x−∇f(x))−(L⋅y−∇f(y))]T(x−y)≥0
[(L⋅x−∇f(x))−(L⋅y−∇f(y))]T(x−y)=L⋅∥x−y∥22−(∇f(x)−∇f(y))T(x−y)≥L⋅∥x−y∥22−∥∇f(x)−∇f(y)∥∥(x−y)∥≥L⋅∥x−y∥22−L⋅∥x−y∥22=0
(二)(2) to (3):
From the equivalent definition of convexity that
f(x) is convex iff ∇2f(x)≥0 if it is twice differentiable.
So
∇2(L2xTx−f(x))=L⋅I−∇2f(x)≥0
So
∇2f(x)≤L⋅I
(三)(3) to (4) ( f(x) does not need to be convex):
From (3), L⋅I−∇2f(x) is positive definite matrix. Then for any ω, we have
ωT(L⋅I−∇2f(x))ω≥0
i.e.
ω∇2f(x)ω≤L∥ω∥22
Then from the taylor decomposition, we have
f(y)=f(x)+∇f(x)⋅(y−x)+(y−x)T∇2f(z)(y−x)≤f(x)+∇f(x)⋅(y−x)+L∥y−x∥22
(四)(4) to (1):( f(x) needs to be convex)
Let y←x+t(∇f(y)−∇f(x)) plug in the equality:
f(x+t(∇f(y)−∇f(x)))≤f(x)+t∇f(x)⋅(∇f(y)−∇f(x))+Lt22∥∇f(y)−∇f(x)∥22
from the convexity of f(x), we have
f(x+t(∇f(y)−∇f(x)))≥f(y)+∇f(y)⋅(x−y+t(∇f(y)−∇f(x)))
combining these two equalities together,
f(y)−f(x)+∇f(y)⋅(x−y)+t∥∇f(y)−∇f(x)∥22≤Lt22∥∇f(y)−∇f(x)∥22
adding
f(x)−f(y)−∇f(y)⋅(x−y)≤L2∥y−x∥22
on both sides,
t∥∇f(y)−∇f(x)∥22≤Lt22∥∇f(y)−∇f(x)∥22+L2∥y−x∥22
Letting t=1L
⇒∥∇f(x)−∇f(y)∥22≤L2∥x−y∥22
2. Summary
From last section, we know thatIf f(x) is convex and has Lipschitz derivative, which is equivalent to all these four conditions.
相关文章推荐
- 用java开发微信公众号:公众号接入和access_token管理(二)
- 重载 struct hash<class T>
- 用java开发微信公众号:测试公众号与本地测试环境搭建(一)
- 4007: [JLOI2015]战争调度 搜索
- 用百度语音识别添加自选股
- Android内存管理原理
- 学习笔记-post的传值随手记
- android pmem 和ashmem 介绍及实例分析
- Gulp: Getting Started
- 【tornado】静态文件
- 大数据(十) - Pig
- 大数据(九) - Hive
- win7 mac 共享文件
- FLTK 1.3.3 VS 2010 Configuration 配置
- 手机号码格式校验
- linux 内核 编译 makefile menuconfig过程讲解
- Java 多态特点
- [First order method] Gradient descent
- C# basic
- 慢牛系列五:用百度语音识别添加自选股