Recently, the following problems have been encountered when using the CVX toolbox :
Only quadratic scalar forms can be specified in CVX
The objective function contains X * X ', where X is a complex vector of 10 * 1, how to solve it?
That’s a non-convex bilinear expression. If there is any hope of using CVX, you’d have to provide context to see whether there is a reformulation which works for it. But first carefully read
including the part about proving your problem is convex.
Thank you very much for your reply，Mark.The context as following：
To tell the truth，I don’t sure if it can be solved by CVX.But the paper says it can be solved by CVX toolbox.So I hope to get your help.Thank you most sincerely
Use the fact that
U*x*x'*U' = square_pos(norm(U*x))
I have tried, but it is wrong : the function ’ square _ poa ’ corresponding to the input parameter of type ’ cvx ’ is not defined. What can I do to solve it?
You can wait for me to correct the typo, which I just did.
Ok, waiting for you.
Ok, I see what you modified, but there is still an error : the operand must be a real number. Is that because the variable X is plural ? Or there will be other reasons？
Please provide reproducible code, complete with all input data, which exhibits the error. You probably have some inputs or other code which don’
t make sense.
clear
%% 初始化
NT=10; %发射天线数
NR=10; %接收天线数
K=5; %单天线通信用户总数
L=1; %传播路径数
P0=30; %功率分配
vk=1; %莱斯分布系数
m_max=10; %最大迭代次数，假设为10
epsilong=1;
alpha0=sqrt(5)+sqrt(5)*1i;
theta=[0 -50 -20 20 50]; %目标位置和杂波源位置
I=length(theta); %杂波源总数
alpha1=sqrt(10)+sqrt(10)*1i;
[alpha2,alpha3,alpha4]=deal(alpha1);
Gamma_T=-1; %信噪比阈值
Gamma_k=15; %信噪比阈值
sigma_R=1;
sigma_Ck=1;
ommiga_k1=[0 -50 -20 20 50];
ommiga_k0=[-90 -45 0 45 90]; %假设出发角均匀分布
%ommiga_k1=[-90 -45 0 45 90]; %假设出发角均匀分布
z=normrnd(0,sigma_R,1,10); %噪声 生成均值为零，标准差为sigma_R的1*10随机数矩阵
nk=normrnd(0,sigma_Ck); %噪声 生成均值为零，标准差为sigma_Ck的正态分布随机变量
A=[exp((pi/4)*1j),exp((3*pi/4)*1j),exp((5*pi/4)*1j),exp((7*pi/4)*1j)]; %诱导符号集
%% 雷达信号模型
at_theta=zeros(NT,K);
ar_theta=zeros(NR,K);
b=((abs(alpha1))^2)/sigma_R;
miu=((abs(alpha0))^2)/sigma_R;
for k=1:K
at_theta(:,k)=(1/sqrt(NT))*[1 exp(-1j*pi*sind(theta(k))) exp(-1j*2*pi*sind(theta(k))) exp(-1j*3*pi*sind(theta(k))) exp(-1j*4*pi*sind(theta(k))) exp(-1j*5*pi*sind(theta(k))) exp(-1j*6*pi*sind(theta(k))) exp(-1j*7*pi*sind(theta(k))) exp(-1j*8*pi*sind(theta(k))) exp(-1j*9*pi*sind(theta(k)))]';
ar_theta(:,k)=(1/sqrt(NR))*[1 exp(-1j*pi*sind(theta(k))) exp(-1j*2*pi*sind(theta(k))) exp(-1j*3*pi*sind(theta(k))) exp(-1j*4*pi*sind(theta(k))) exp(-1j*5*pi*sind(theta(k))) exp(-1j*6*pi*sind(theta(k))) exp(-1j*7*pi*sind(theta(k))) exp(-1j*8*pi*sind(theta(k))) exp(-1j*9*pi*sind(theta(k)))]';
U{k}=ar_theta(:,k)*at_theta(:,k)'; %注意：U{1}表示U0
end
%% 通信信号模型
at_ommiga=zeros(NT,K);
ommiga=zeros(K,5);
at_ommigakl=zeros(NT,K);
h_LOS=zeros(NT,K);
h_NLOS=zeros(NT,K);
h=zeros(NT,K);
c=zeros(K,1);
for k=1:K
at_ommiga(:,k)=(1/sqrt(NT))*[1 exp(-1j*pi*sind(ommiga_k0(k))) exp(-1j*2*pi*sind(ommiga_k0(k))) exp(-1j*3*pi*sind(ommiga_k0(k))) exp(-1j*4*pi*sind(ommiga_k0(k))) exp(-1j*5*pi*sind(ommiga_k0(k))) exp(-1j*6*pi*sind(ommiga_k0(k))) exp(-1j*7*pi*sind(ommiga_k0(k))) exp(-1j*8*pi*sind(ommiga_k0(k))) exp(-1j*9*pi*sind(ommiga_k0(k)))].';
at_ommigakl(:,k)=(1/sqrt(NT))*[1 exp(-1j*pi*sind(ommiga_k1(k))) exp(-1j*2*pi*sind(ommiga_k1(k))) exp(-1j*3*pi*sind(ommiga_k1(k))) exp(-1j*4*pi*sind(ommiga_k1(k))) exp(-1j*5*pi*sind(ommiga_k1(k))) exp(-1j*6*pi*sind(ommiga_k1(k))) exp(-1j*7*pi*sind(ommiga_k1(k))) exp(-1j*8*pi*sind(ommiga_k1(k))) exp(-1j*9*pi*sind(ommiga_k1(k)))].';
c(k,1)=sqrt(1/2)*(randn(1,1)+1j*randn(1,1)); %生成均值为0 方差为1的复高斯分布
h_LOS(:,k)=sqrt(NT)*at_ommiga(:,k); %视距传输
h_NLOS(:,k)=sqrt(NT/L)*c(k,1)*at_ommigakl(:,k); %非视距传输
h(:,k)=sqrt(vk/(1+vk))*h_LOS(:,k)+sqrt(1/(1+vk))*h_NLOS(:,k); %信道矢量
end
%% 优化目标函数
x_he=zeros(10,10);
m=1;
x0=[1-1i;1+1i;1-1i;1-1i;1-1i;1-1i;1-1i;1+1i;1-1i;1-1i].'; %随机x0
x(:,1)=x0;
AI=eye(10); %10*10单位矩阵
while m<=m_max
for i=1:4
x_he=x_he+b*U{i+1}*x(:,m)*x(:,m)'*U{i+1}';
end
U1=U{2}+U{3}+U{4}+U{5};
U2=U{2}'+U{3}'+U{4}'+U{5}';
cvx_begin
variable X(10) complex %定义复数向量X
w(:,m)=(inv(x_he+AI)*U{1}*x(:,m))/(x(:,m)'*U{1}'*inv(x_he+AI)*U{1}*x(:,m)); %更新w
u=(miu*((abs(w(:,m)'*U{1}*x(:,m)))^2))/(w(:,m)'*(x_he+AI)*w(:,m)); %更新u
% f1=(2*(x(:,m)'*U{1}'*w(:,m))*U{1}'*w(:,m))';
% f2=miu*real((2*(x(:,m)'*U{1}'*w(:,m))*U{1}'*w(:,m))'*x(:,m));
% f3=(u*w(:,m)'*(b*U1*X*X1*U2)*w(:,m));
% f4=(u*w(:,m)'*AI*w(:,m));
maximize(miu*(real((2*(x(:,m)'*U{1}'*w(:,m))*U{1}'*w(:,m))'*(X-x(:,m))))-(u*w(:,m)'*((b*square_pos(norm(U1*X))+AI)*w(:,m)))) %目标函数
subject to
square_pos(norm(X,2))<=P0; %约束条件 式（12b）
for k=1:K
i=unidrnd(4); %产生1到4之间的随机正整数
s(k)=A(1);
abs(imag(h(:,k)'*X*conj(s(k))))<=((real(h(:,k)'*X*conj(s(k))))-sqrt(sigma_Ck*Gamma_k))*tand(-45); %约束条件 式（12c）
end
cvx_end
SINR_rad(m)=cvx_optval;
m=m+1;
x(:,m)=X;
end
figure(1)
plot(SINR_rad,'-+b')
xlabel('迭代次数m');
ylabel('SINR_r_a_d(dB)');
title('收敛性分析');
grid onw(:,m) is complex, and it is multiplying the convex expression, square_pos(norm(U1*X)) which is not allowed by CVX because it can’t determine its convexity.
Your code basically has w(:,m)'*convex_expression*w(:,m)
A solution is to change it to convex_expression*(w(:,m)'*w(:,m)) or (w(:,m)'*w(:,m))*convex_expression. That way, there will be a nonnegative real scalar number multiplying the convex_expression, which will make CVX happy.
U1 is a matrix of 10*10，and X is a vector of 10 * 1, so the result of U1 * X * X ’ * U1 ’ should be a matrix of 10 * 10 , but the result of square _ pos ( norm ( U1 * X ) ) is a scalar, which is not true. Is there any other expression that can solve this problem?
Use square_pos(norm(X'*U1'*w)) in place of w'*U*X*X'*U1'*w)
There is still error :Cannot perform the operation: {complex affine} .* {convex}.Is there any solution？
Where is the error occurring? Please show the complete current version of your code.
This transformation seems to cause a change in the dimensions of the matrix, for example, if convex_expression is N x N and w is N x M, then w ’ * convex-expression * w is M x M, but convex-expression * (w (:, m)’ * w (:, m)) does not satisfy matrix multiplication
If convex expression is not scalar, I suggested using square_pos(norm(X'*U1'*w)) in place of w'*U*X*X'*U1'*w