RMT（二阶等价）

随机矩阵系列

随机矩阵的二阶等价结果

---

---

前言

本文仿真了一阶等价和二阶等价结果

---

一、主要目的

$X=C^{\frac{1}{2}}Z$ $Z\in N(0,I_p)$
求$Q(z)=(\frac{1}{n}X{X}^T-z{I}_p)$和$Q(z_1)MQ(z_2)$的确定性等价。

二、简化情况$C=I_p$

1.$\overline{Q}(z)$

RMT 有如下结果：$\overline{Q}(z)=-\frac{1}{z}(I_p+\tilde{m}(z)C{)}^{-1}$
${\tilde{m}}_p(z)=(-z+\frac{1}{n}tr(C(I_p+{\tilde{m}}_p(z)C)){)}^{-1}$

close all; clear; clc
coeffList = 1:10;
LoopNum = 50;
Res = zeros(length(coeffList),1);
temp1 = zeros(1,LoopNum);
for ii = 1:length(coeffList)
    disp(ii)
    for Loop_i = 1:LoopNum
        [~,~,temp1(Loop_i)]= GetfirstQ(coeffList(ii));
    end
    Res(ii,1) = mean(temp1);
end
plot(coeffList,abs(Res).')

function [Q_bar1,Q1,error1] = GetfirstQ(coeff)

p = 100*coeff;  
n = 1000*coeff;
eigs_C  = [ones(p,1)];
C = eye(p);
Z = randn(p,n);
X = sqrtm(C)*Z;
SCM = X*(X')/n;
clear i % make sure i stands for the imaginary unit
z1 = 2 + 1i;
[Q_bar1,Q1,error1] = SolveQbar(z1,eigs_C,n,p,C,SCM);
end

function [Q_bar,Q,error]= SolveQbar(z,eigs_C,n,p,C,SCM)
%UNTITLED2 此处提供此函数的摘要
%   此处提供详细说明
tilde_m = 0;
tilde_m_tmp=-1;
while abs(tilde_m-tilde_m_tmp)>1e-6
    tilde_m_tmp=tilde_m;
    tilde_m = 1/( -z + 1/n*sum(eigs_C./(1+tilde_m*eigs_C)) );
end

 Q_bar = -1/z * inv(eye(p) + tilde_m * C);
 Q = inv(SCM - z *eye(p));
 error = trace(Q_bar - Q)/p;
end

结果如下，一阶没有问题（横轴p n 增加，纵轴误差）

2. $Q(z_1)MQ(z_2)$

$E[Q(z_1)MQ(z_2)]=\overline{Q}(z_1)M\overline{Q}(z_2)+\overline{Q}(z_1)C\overline{Q}(z_2)f(z_1,z_2)$

$f(z_1,z_2)=\frac{\frac{\frac{1}{n}tr(C\overline{Q}(z_1)M\overline{Q}(z_2))}{1-\frac{\frac{1}{n}tr(C\overline{Q}(z_1)C\overline{Q}(z_2))}{(1+\frac{1}{n}tr(C\overline{Q}(z1)))(1+\frac{1}{n}tr(C\overline{Q}(z2)))}}}{(1+\frac{1}{n}tr(C\overline{Q}(z1)))(1+\frac{1}{n}tr(C\overline{Q}(z2)))}\in C$
不是简单的$E[Q(z_1)MQ(z_2)]=\overline{Q}(z_1)M\overline{Q}(z_2)$
下面仿真有无第二项的区别
代码如下（示例）：

coeffList = 1:10;
LoopNum = 100;
Res = zeros(length(coeffList),2);
temp1 = zeros(1,LoopNum);
temp2 = zeros(1,LoopNum);
for ii = 1:length(coeffList)
    disp(ii)
    for Loop_i = 1:LoopNum
        [temp1(Loop_i),temp2(Loop_i)]= Solveerror(coeffList(ii));
       
    end
    Res(ii,1) = mean(temp1);
    Res(ii,2) = mean(temp2);
end
plot(coeffList,abs(Res(:,1)).');
hold on;
plot(coeffList,abs(Res(:,2)).');
legend('Only first item','first plus second item')

结果如下：

上面是$z_1=z_2$的情况，若$z_1\ne z_2$,结果类似

---

三、简化情况$C\ne I_p$

一阶等价

二阶等价

所以实际上正确的确定性等价是很小的，第二项也是不可忽略的

总结

考虑的$C=I_p$和$C\ne I_p$情况