Operator与优化

Relation

关系这个词跟映射有点相似，对于一个关系(R)，其是((x, y))的一个集合集合。其中( ext{dom }R={x|(x,y)in R})，(R(x)={yvert (x,y)in R})，其零集合是({x| (x,y)in R, y=0})。

Operations on Relation

• inverse. (R^{-1}={(y, x)vert (x,y)in R})
• composition. (RS={(x, y)vert (x,z)in R, (z,y)in S})
• scalar multiplication. (alpha R={(x, alpha y)vert (x,y)in R})
• addition. (R+S={(x, y+z)vert (x,y)in R, (x,z)in S})
• resolvent operator. (S=(I+lambda R)^{-1})

通过以上的运算可以看出，relation有点类似于凸函数中epigraph的那种集合定义。

Monotone Operations

对于一个单调的relation (F)，其定义为

[(u-v)^T(x-y)geq 0 ]

对于任意的((x, y), (u,v)in R). 一个最大单调(F)的定义为，没有其他单调relation包含(F)。

(F)是最大单调当且仅当(F)是一个连接的曲线，其斜率不存在负值。

Case: Subgradient (F=partial f(x))

Nonexpansive and contractive operator

对于一个(L-)Lipschitz连续的operator (F)，其nonexpansive和contraction的定义分别为(L=1)和(L<1)。

Characters:

Resolvent operation and Cayley operator

对于一个relation (F)，当(F)是单调且nonexpansive时，(R) operator是contractive的。(F)的cayley operator定义为

[C=2R-I=2(I+lambda F)^{-1}-I ]

同样当F是单调的时候，其cayley operator (C)是nonexpansive。

Proof:

Case:

1. Proximal

2. Indicator

Fixed point of operators & zero set of (F)

这里有个很重要的定理就是Cayley和resolvent的Fixed point等价于(F) relation的zero set。也就是

[F(x)in 0 Leftrightarrow C(x)=x Leftrightarrow R(x)=x ]

Theorem: Banach fixed point theorem

当(F)是contraction，dom (F=R^n)，那么(F(x))会收敛到一个唯一的fixed point。

Damped iteration of a nonexpansive operator

相对于

[x^{k+1}=F(x^k) ]

Damped iteration为一个(x^k)和(F(x^k))的组合

[x^{k+1} = heta^k x^k+(1- heta^k)F(x^k) ]

Proof:

Case:

Operator Splitting

这里要解决的问题是一个relation (F=A+B)，单独队(F)进行求解可能比较麻烦而分开对(A)和(B)求解更简单。

Theorem: 如果A和B是maximal monotone，那么

[0in A(x)+B(x) Leftrightarrow C_AC_B(z)=z ]

其中(x=R_B(z))

Proof:

证明也是比较简单，使用定义就可以得到。

Peaceman-Rachford & Douglas-Rachfold Splitting

[egin{align} & ext{Peaceman-Rachford}:qquad z^{k+1}=C_AC_B(z^k)\ & ext{Douglas-Rachfold}:qquad z^{k+1}=frac 1 2(I+C_AC_B)(z^k)\ end{align} ]

1. Douglas-Rachfold updating

The last equation:

Case: Alternating direction method of multipliers

Case: Constrained optimization

2. Peaceman-Rachford updating

[egin{align}x^{k+frac{1}{2}}&= ext{prox}_{alpha f}(z^k)\z^{k+frac{1}{2}}&=2x^{k+frac{1}{2}}-z^k\x^{k+1}&= ext{prox}_{alpha g}(z^{k+frac{1}{2}})\z^{k+1}&=2x^{k+1}-z^{k+frac{1}{2}}end{align} ]

Case: FedSplit, a consensus problem

对于loss函数(F)，以及consensus constrain，利用一阶方法求解最小值等价于

[0in abla F(x)+mathcal{N}^{ot} ]

其中(mathcal{N}^{ot})为其consensus的normal corn。

上图为其论文中的算法流程，这里的(A) operator为(mathcal{N}^{ot})，(B) operator为( abla F)而且由于(x=ar{z})在最后执行所以整个顺序都提前，并且算法中的第一步(a)直接整合了PR的中间两步。

Consensus Optimization

贴一下Boyd课程的代码吧（注释掉的是我修改的，更新就和公式一样了）

% Solves the QP
%       mininimze   (1/2)||Ax - b||_2^2
%       subject to  Fx <= g
% using D-R consensus. Note that the code has not been optimized for
% runtime and is only presened to give an idea of D-R consensu. For better
% performance, the inner loop should be run in parallel and should use a
% fast QP solver for small problems (e.g., CVXGEN).
%
% EE364b Convex Optimization II, S. Boyd
% Written by Eric Chu, 04/25/11
% 

close all; clear all
randn('state', 0); rand('state', 0);

%%% Generate problem instance
m = 1000;
n = 100;
k = 50;

xtrue = randn(n,1);
A = randn(m,n);
b = A*xtrue + randn(m,1);

F = randn(k,n);
g = F*xtrue;

%%% Use CVX to find solution
cvx_begin
    variable x(n)
    minimize ((1/2)*sum_square(A*x - b))
    subject to
        F*x <= g
cvx_end
xcvx = x;
fstar = cvx_optval; 
  
%%% Douglas-Rachford consensus splitting
N           = 10;      % number of subproblems
MAX_ITERS   = 50;
rho         = 200;

z           = zeros(n,N); 
xbar        = zeros(n,1);

for j = 1:MAX_ITERS,
    
    % x = prox_f(z), could be done in parallel
    for i = 1:N,
        Ai = A(m/N*(i-1) + 1:i*m/N,:);
        bi = b(m/N*(i-1) + 1:i*m/N);
        
        Fi = F(k/N*(i-1) + 1:i*k/N,:);
        gi = g(k/N*(i-1) + 1:i*k/N);
        
        % use CVX to solve prox operator
        zi = z(:,i);
        cvx_solver sdpt3
        cvx_begin quiet
            variable xi(n)
            minimize ( (1/2)*sum_square(Ai*xi - bi) + (rho/2)*sum_square(xi - zi) )
            subject to
                Fi*xi <= gi
        cvx_end
        x(:,i) = xi;
    end
    
    %% standard 
    %z_midterm = 2*x-z;
    %xbar_prev = xbar;
    %xbar = mean(z_midterm,2);
    
    %infeas(j) = sum(pos(F*xbar - g));
    %f(j) = (1/2)*sum_square(A*xbar - b);
    %z = z + (xbar*ones(1,N) - x);
    
    %% Boyd
    
    xbar_prev = xbar;
    xbar = mean(x,2);
    
    % record infeasibilities
    infeas(j) = sum(pos(F*xbar - g));
    
    % record objective value
    f(j) = (1/2)*sum_square(A*xbar - b);
    
    % update
    z = z + (xbar*ones(1,N) - x) + (xbar - xbar_prev)*ones(1,N);
end

%%% Make plots
subplot(2,1,1)
semilogy(1:MAX_ITERS, infeas);
ylabel('infeas'); set(gca, 'FontSize', 18); axis([1 MAX_ITERS 10^-2 10^2])
subplot(2,1,2)
plot(1:MAX_ITERS, f, [1 MAX_ITERS], [fstar fstar], 'k--');
xlabel('k'); ylabel('f'); axis([1 MAX_ITERS 300 2000]); set(gca, 'FontSize', 18);
print -depsc dr_consensus_qp.eps

左边是我修改的，右边是Boyd的代码。看下来效果好像差不多，但是我还没搞懂他的代码为啥这样写。

参考资料

• Monotone Operators - Boyd
• Monotone Operator Splitting Methods - Boyd