solve matrix equation involving diag and weighted summation of PSD matrices

92 Views Asked by At

Some help for a non-expert, please?

Let $\mathbf{b}= [b_1\cdots b_m \cdots b_M]^T$ and $\mathbf{c}= [c_1\cdots c_m \cdots c_M]^T$.

Given the identity

$\mathbf{D}^T \, \mathrm{diag}(\mathbf{b}) \, \mathbf{D} = \sum_{m=1}^{M}b_mc_m\mathbf{A}_m,$

where $\mathbf{A}_m \in \Re^{N\times N}, \mathbf{A}_m \geq 0$.

Is there a way to solve for $\mathbf{D}$ in which we do not need to know $\mathbf{b}$? In other words: my intuition tells me that there could be a closed-form expression for $\mathbf{D}$ in which $\mathbf{b}$ does not appear.

Thanks in advance,

1

There are 1 best solutions below

0
On BEST ANSWER

I will self-answer my question, but I would thank anyone who can provide a proof, even if the proof relies on basic linear algebra... So:

We write

$\mathbf{D}^T \, \mathrm{diag}(\mathbf{b}) \, \mathbf{D} = \sum_{m=1}^{M} (\mathbf{D}_{m*})^T\mathbf{D}_{m*}b_m$,

so therefore, since

$\sum_{m=1}^{M} (\mathbf{D}_{m*})^T\mathbf{D}_{m*}b_m = \sum_{m=1}^{M}b_mc_m\mathbf{A}_m$,

then

$(\mathbf{D}_{m*})^T\mathbf{D}_{m*} = c_m\mathbf{A}_m$.

This means that we can construct matrix $\mathbf{D}$, element-wise, as

$\mathbf{D}_{mn} = (c_m \mathbf{A}_{m,nn})^{1/2}$,

where $\mathbf{A}_{m,nn}$ represents the $n$-th diagonal entry of matrix $\mathbf{A}_m$.