2
$\begingroup$

This question may seem a little dumb, but I really googled around with no success.

Let $A$ and $B$ be two square matrix of size $n$. Denote with $\lambda_1(A) \leq \lambda_2(A) \cdots \leq \lambda_n(A)$ and $\lambda_1(B) \leq \lambda_2(B) \cdots \leq \lambda_n(B)$ the generic eigenvalues of A and B, respectively. What can I say about $\lambda_i(AB)$?

Thank you in advance!

$\endgroup$
3
$\begingroup$

Here are two results (this is exercise III.6.14 in Matrix Analysis by R. Bhatia):

Let $\lambda^\uparrow(A)$ be the vector of eigenvalues of $A$ in increasing order and $\lambda^\downarrow(A)$ be the vector of eigenvalues of $A$ in decreasing order.

Then, if $A$ and $B$ are positive definite: $$\lambda^\downarrow(A)\lambda^\uparrow(B) \prec \lambda(AB) \prec \lambda^\downarrow(A)\lambda^\downarrow(B).$$ The $\prec$ symbol means majorized; For $n$-dimensional vectors $x,y$, $x \prec y$ means that for $1 \leq k < n$: $$\sum_{i = 1}^k x_i^\downarrow \leq \sum_{i = 1}^k y_i^\downarrow$$ and $$\sum_{i = 1}^n x_i = \sum_{i =1}^n y_i$$ i.e., if we sum the $k$ largest elements in each vector, the sum of elements from $x$ will be smaller than or equal to the sum of elements from $y$, but if we sum all the elements in each vector, the sums are equal.

More generally, for any Hermitian (symmetric, if $A$ and $B$ are real) matrices $A,B$: $$\langle \lambda^\downarrow(A), \lambda^\uparrow(B) \rangle \leq \operatorname{tr} AB \leq \langle \lambda^\downarrow(A), \lambda^\downarrow(B) \rangle$$ and we have in general, for any square matrix $X$, that $$\operatorname{tr} X = \sum_{i=1}^n \lambda_i(X)$$.

Special case if $AB$ is symmetric

An interesting case occurs if the product $AB$ is also symmetric. This happens if and only if $AB = BA$. Since $A$ and $B$ are diagonalizable (since they are symmetric), this means that $A$ and $B$ are simultaneously diagonalizable, i.e. there exists $T$ such that $$\begin{align*} A &= TD_AT^{-1} \\ B &= TD_BT^{-1} \end{align*}$$ and hence $$AB = TD_AT^{-1}TD_BT^{-1} = TD_AD_BT^{-1}$$ and the product $D_AD_B$ will be a diagonal matrix, containing the eigenvalues of $AB$. Hence an eigenvalue of $AB$ will be a product of one eigenvalue from $A$ and one eigenvalue from $B$.

$\endgroup$
  • $\begingroup$ can you develop the meaning of "majorized"? $\endgroup$ – the_candyman Nov 19 '13 at 21:08
  • $\begingroup$ I have added a short explanation and some important notation I had missed. $\endgroup$ – Calle Nov 19 '13 at 21:21

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.