This problem showed up on UCLA's basic exam for Fall 2018:

Let $X$ be an $n \times n$ symmetric (real) matrix and $z \in \mathbb{C}$ with $\text{Im } z > 0$. Define $G = (X - zI)^{-1}.$ Show that $$\sum_{1 \leq j \leq n} |G_{ij}|^2 = \frac{\text{Im } G_{ii}}{\text{Im }z}.$$

I worked on this for a little bit where I applied the real spectral theorem to $X$ which in turn gives you that $G = Q^T D Q$ where $Q$ is real orthogonal ($Q^T Q = Q Q^T = I$) and $D$ is diagonal satisfying $D_{ii} = (\lambda_i - z)^{-1}$. Where $\lambda_1, \ldots, \lambda_n$ are the eigenvalues for $X$. Couldn't really see any immediate way out from there. Would be interested in seeing peoples' solutions to this problem and any connections to the study of matrix resolvents: https://en.wikipedia.org/wiki/Resolvent_formalism


By considering $\tilde{X}=X-\text{Re}(z)I$ in place of $X$, we may assume that $z=\mu i$ for some $\mu>0$. Now observe that $$ (X-\mu iI)(X+\mu iI) = X^2+\mu^2 I>0. $$ This gives $$ G=(X-\mu iI)^{-1}=(X^2+\mu^2 I)^{-1}(X+\mu iI). $$ Let $e_i$ be the vector whose $i$-th coordinate is $1$ and other coordinates are all $0$'s. We can observe that $$\begin{eqnarray} G_{ii}=e_i'Ge_i&=&e_i'(X^2+\mu^2 I)^{-1}(X+\mu iI)e_i\\&=&e_i'(X^2+\mu^2 I)^{-1}Xe_i+i\mu \cdot e_i'(X^2+\mu^2 I)^{-1}e_i \end{eqnarray}$$ and hence $$ \text{Im}(G_{ii})=\mu\cdot e_i'(X^2+\mu^2 I)^{-1}e_i. $$ This gives $$ \frac{\text{Im}(G_{ii})}{\text{Im}( z)}=\frac{\mu\cdot e_i'(X^2+\mu^2 I)^{-1}e_i}{\mu}= e_i'(X^2+\mu^2 I)^{-1}e_i. $$ We can also see that $$ GG^*=(X-\mu iI)^{-1}(X+\mu iI)^{-1}=(X^2+\mu^2 I)^{-1} $$ and $\sum_{j=1}^n |G_{ij}|^2$ can be represented as $$ \sum_{j=1}^n |G_{ij}|^2 =|G^*e_i|^2=e_i'GG^*e_i. $$ Thus it follows $$ \sum_{j=1}^n |G_{ij}|^2 =e_i'GG^*e_i=e_i'(X^2+\mu^2 I)^{-1}e_i=\frac{\text{Im}(G_{ii})}{\text{Im}( z)}. $$ This proves the desired result.

Note: $e_i'$ means the transpose of $e_i$ and $G^*$ means conjugate transpose of $G$.

  • $\begingroup$ Awesome. So it seems to be the key insight is that the diagonal entries of $G G^{*}$ give us the values for the sums we are concerned with, and by considering what $G G^{*}$ looks like in terms of $X$ with some clever manipulation we get what we want. $\endgroup$ – BenB Jan 19 at 21:01
  • $\begingroup$ @BenB I think so. Perhaps such manipulations were the key. I hope this will help. $\endgroup$ – Song Jan 19 at 21:03

Here's an alternative solution:

We first note that $(GG^*)_{ii} = \sum_{j = 1}^n |G_{ij}|^2$. Thus our statement reduces to showing that $$(GG^*)_{ii} = \frac{\text{Im } G_{ii}}{\text{Im }z} \iff \text{Im }[(zG^*G)_{ii}] = \text{Im }G_{ii}.$$ Well $$G = (X - zI)^{-1} \implies G(X - zI) = I \implies GXG^* - zGG^* = G^* \implies GXG^* - G^* = zGG^*$$ Thus $$\text{Im }[(zGG^*)_{ii}] = \text{Im }[(GXG^* - G^*)_{ii}] = \text{Im }[(GXG^*)_{ii}] + \text{Im }[(-G^*)_{ii}] = \text{Im }[(GXG^*)_{ii}] + \text{Im }[G_{ii}].$$ Well we then notice that since $GXG^*$ is self-adjoint that $$(GXG^*)_{ii} = e_i^*GXG^*e_i = \langle GXG^*e_i, e_i\rangle \in \mathbb{R}$$ where $e_i$ be the $i$th standard basis vector, i.e. the vector containing all zeros except for a one in the $i$th component. Thus $$\text{Im }[(zGG^*)_{ii}] = \text{Im }[(GXG^*)_{ii}] + \text{Im }[G_{ii}] = \text{Im }[G_{ii}]$$ as desired.


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.