1
$\begingroup$

Given diagonal $A\in\mathbb{R}^{n\times n}$, symmetric $B\in\mathbb{R}^{n\times n}$, and $\mathrm{vec}(Y)=(A\otimes A-I_n)^{-1}\mathrm{vec}(B)$, find the following:

\begin{align} \operatorname{trace}(BY^{-1}). \end{align}

$A-I>0,$ $B$ is nonsingular.


My attempt:

For special case $A=aI_n$:

\begin{align} \mathrm{vec}(Y)=(A\otimes A-I_n)^{-1}\mathrm{vec}(B)=(aI_n\otimes aI_n-I_n)^{-1}\mathrm{vec}(B)=(a^2-1)^{-1}\mathrm{vec}(B). \end{align}

Thus

\begin{align} Y^{-1}=(a^2-1)B^{-1}, \end{align}

and \begin{align} \operatorname{trace}(BY^{-1})=\operatorname{trace}((a^2-1)I_n)=n(a^2-1). \end{align}

For special case $A=\begin{bmatrix} a_1 & & \\ & a_2 & \\ &&a_2 \end{bmatrix}:$

\begin{align} \operatorname{trace}(BY^{-1})=a_1^2a_2^2 + a_2^2 - 2. \end{align}


A=[3 0 0; 0 2 0; 0 0 2];
a=[3 2 2]';
T=zeros(1000,1);
for i=1:1000
    B=rand(3,2);
    [X,K,L] = idare(A,B,zeros(3,3));
    T(i)=a'*X.*inv(X)*a;
end
plot(T)
$\endgroup$

1 Answer 1

1
$\begingroup$

Define the vectors $$\eqalign{ a = {\rm diag}(A),\quad b = {\rm vec}(B),\quad y = {\rm vec}(Y) \\ }$$ and the matrices $$\eqalign{ A &= {\rm Diag}(a) \\ M &= A\otimes A - I_n\otimes I_n }$$ Use these to rearrange the given relationship into a matrix equation.
$$\eqalign{ y &= M^{-1}b \\ b &= My = \left(A\otimes A - I_n\otimes I_n\right)y &= {\rm vec}(AYA - Y) \\ B &= AYA - Y \\ BY^{-1} &= AYAY^{-1} - I_n \\ }$$ The function in question is the trace of that last expression. $$\eqalign{ \phi &= {\rm Tr}(BY^{-1}) \\&= {\rm Tr}(AYAY^{-1}) - {\rm Tr}(I_n) \\ &= YAY^{-1}:A - n \\ &= {\rm diag}(YAY^{-1}):{\rm diag}(A) - n \\ &= \left(Y^{-T}\odot Y\right)a:a - n \\ &= R_Ya:a - n \\ &= a^TR_Ya - n \\ }$$ The matrix $R_Y$ is known as the Relative Gain Array of $Y$ and has some interesting properties.
For example, denoting the all-ones vector by ${\tt1},\,$ it demonstrates Markov-like behavior
$$R_Y{\tt1}={\tt1},\qquad {\tt1}^TR_Y{\tt1}={\tt1}^T{\tt1}=n$$ Re-examining the special case of $\;\Big(A=\alpha I\iff a=\alpha{\tt1}\Big)\;$ yields $$\eqalign{ \phi &= (\alpha{\tt1})^TR_Y(\alpha{\tt1}) - n \\ &= \alpha^2n - n \\ \\ }$$


In the above, the symbol $(\odot)$ denotes the elementwise/Hadamard product
while a colon $(:)$ represents the trace/Frobenius product, i.e. $$\eqalign{ A:B = {\rm Tr}(A^TB) }$$ Finally, the diag() function creates a column vector from the main diagonal of the input matrix, while Diag() creates a diagonal matrix from a vector argument.


Update

I don't think you're interpreting your simulation correctly.

Here is a short snippet Julia/Matlab code which demonstrates that changing $B$ (while holding $A$ constant) definitely affects the value of $\phi$ and $R_Y$
a  = 100*rand(3,1);  A = Diag_(a)
   77.2603   0.0      0.0   
    0.0     83.9703   0.0   
    0.0      0.0     58.4881
     
B  = rand(3,3);  B += B';
y  = inv(kron(A,A)-I)*vec(B);  Y = reshape(y,size(B));
Ry = inv(Y').*Y
   0.564956   2.99807   -2.56303
   2.99807   -0.343684  -1.65439
  -2.56303   -1.65439    5.21742

a'*Ry*a
   18283.48744321049

B  = rand(3,3);  B += B';
y  = inv(kron(A,A)-I)*vec(B);  Y = reshape(y,size(B));
Ry = inv(Y').*Y
3×3 Array{Float64,2}:
   21.8927   17.0631   -37.9558
   17.0631    1.88734  -17.9505
  -37.9558  -17.9505    56.9063

a'*Ry*a
   40704.04750730478
$\endgroup$
10
  • $\begingroup$ thanks for detailed answer $\endgroup$
    – Lee
    Jun 15, 2020 at 16:38
  • $\begingroup$ Hi greg, in simulation $R_Y$ doesn't depend on $B$ when $A$ is diagonal. Is it possible to express $R_Y$ in terms of diagonal elements of $A$? For my second special case I found that $R_Y(1,1)=\frac{(a_1a_2-1)^2}{(a_1-a_2)^2}$, but for general case I couldn't figure out $\endgroup$
    – Lee
    Jun 21, 2020 at 10:10
  • $\begingroup$ What does your simulation look like? I tried some code of my own (above) and don't see the effect that you claim. $\endgroup$
    – greg
    Jun 21, 2020 at 15:03
  • $\begingroup$ I fixed diagonal $A$ and for any random $B$ we will get same $a^TR_Ya$, and elements of $A$ are larger than 1. And also $Y$ is solution to DARE, i use the functuon idare(A,B,Q=0). I will include it in my question soon $\endgroup$
    – Lee
    Jun 21, 2020 at 15:32
  • $\begingroup$ I wrote the code from my memory, because my laptop is not with me, but I think should be no bugs. $\endgroup$
    – Lee
    Jun 21, 2020 at 16:00

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.