《An Integer Projected Fixed Point Method for GraphMatching and MAP Inference》论文阅读
Step 1:
确定迭代初始值 \(x^*=x_0\),记录初值 \(S_0=x_0^TMx_0\),\(k=0\)
Step 2.1:
确定 \(b_{k+1}=P_d(Mx_k)\)
\[b_{k+1}=y^*=argmax(x_k^Tmy)
\]
\[s.t.\ Ay=1
\]
\[\ \ \ \ \ \ y>0
\]
对于此式,问题退化成了 \(Linear program\),所以可以得到全局最优解(离散的)
Step 2.2 && Step 3:
\(C=x_k^TM(b_{k+1}-x_k),\ \ D=(b_{k+1}-x_k)^TM(b_{k+1}-x_k), \ \ S_k=x_k^TMx_k\),
现在来确定下一次迭代的点 \(x\):
\(x=x_k+t(b_{k+1}-x_k)\),其中 \(0 \leq t \leq 1\),而由于 \(C=x_k^TMb_{k+1}-x_k^TMx_k \geq 0\),则
\[f(t)=x^TMx=(x_k+t(b_{k+1}-x_k))^TM(x_k+t(b_{k+1}-x_k))=S_k+tx^TM(b_{k+1}-x_k)+t(b_{k+1}-x_k)^TMx_k+t^2D
\]
\[(b_{k+1}-x_k)^TMx_k=((b_{k+1}-x_k)^TMx_k)^T=x^TM(b_{k+1}-x_k)
\]
所以,下一次迭代点需要 \(f(t)>x_k^TMx_k\)
所以只需要关注 \(D\) 的大小,
如果\(D \geq 0\),则\(f(t)>x_k^TMx_k\),\(t\in [0,1]\) 取任意值即可,那么 \(x_{k+1}=b_{k+1}\),
如果\(D<0\),则\(g(t)=f(t)-S_k=2tC+t^2D\)是否大于取决于 \(t\),若要使其\(g(t)>0\)
\[2tC+tD>0 \Longrightarrow t(2C+tD)>0 \Longrightarrow t=min\{ \frac{-C}{D}, 1\}
\]
如\(t=1 \Longrightarrow \frac{-C}{D}>1_{D<0} \Longrightarrow \mid C \mid > \mid D \mid \Longrightarrow 2C+D>0\),那么\(x_{k+1}=x_k+t(b_{k+1}-x_k)\)
============
当我沉默时,我觉得充实。当我开口说话时,却觉得空虚
===========