A is a $4*4$ matrix with eigenvalues $\lambda_A$. Consider a block matrix $B = \left( \begin{array}{ccc} A & I \\ I & A \end{array} \right) $. Then how can we find eigenvalue $\lambda_B$ of matrix B in terms of $\lambda_A$ ?
I know that to find $\lambda_B$, we have to do $|B-\lambda_BI| = 0$, and by properties of block matrices, we have equivalently $|(A-\lambda_BI)^2 - I^2| = 0$.
But I am stuck at this point, I don't know how to proceed further.


It looks like you are trying to find a determinant of block matrix using the rule for 2x2 matrices, but it is incorrect.

According to http://en.wikipedia.org/wiki/Determinant#Block_matrices

$$ |B-\lambda_bI| = \left| \left(\begin{matrix}A-\lambda_bI & I \\ I & A - \lambda_bI\end{matrix}\right) \right|= |A-\lambda_bI - I||A-\lambda_bI + I|= \\ =|A-(\lambda_b+1)I||A-(\lambda_b - 1)I| = 0 $$

Then you can determine $\lambda_b$ using the fact that $\lambda_a$ is a solution of $|A - \lambda_aI|=0$.

  • $\begingroup$ ...learing to $\sigma(B)=\{\lambda \in \mathbb{C}: \lambda + 1 \in \sigma(A) \mbox{ or } \lambda - 1 \in \sigma(A)$. $\endgroup$
    – Roland
    Feb 6 '14 at 8:12

Or to look at it from the point of view of normal forms, if $A=TJT^{-1}$, then using


the block matrix can be decomposed into smaller 2x2 blocks if $J$ is diagonal, and correspondingly larger structured blocks if not. In any event, the eigenvalues are determined as the eigenvalues of the 2x2 blocks $\begin{pmatrix}λ_a&1\\1&λ_a\end{pmatrix}$, leading of course to the same eigenvalues as in Steelravens answer.


I think I can see two ways to proceed with this. But before going further, I wish to point out that what I am about to say applies to any $N \times N$ matrix $A$, so I will remove the restriction that $\text{size} (A) = 4$ until further notice. Having clarified that point, the first option is to work directly from the definitions of eigenvalue and eigenvector, as follows: suppose $\lambda_A$ is an eigenvalue of $A$, with eigenvector $v_A \ne 0$, so that we have

$Av_A = \lambda_A v_A, \tag{1}$

and note that the $2N$-vectors

$V_{A+} = \begin{pmatrix} v_A \\ v_A \end{pmatrix} \tag{2}$


$V_{A-} = \begin{pmatrix} v_A \\ -v_A \end{pmatrix} \tag {3}$

are both eigenvectors of $B$, thus:

$BV_{A+} = \begin{bmatrix} A & I \\ I & A \end{bmatrix} \begin{pmatrix} v_A \\ v_A \end{pmatrix} = \begin{pmatrix} (A + I) v_A \\ (I + A) v_A \end{pmatrix} = \begin{pmatrix} (\lambda_A + 1) v_A \\ (\lambda_A + 1) v_A \end{pmatrix} = (\lambda_A + 1) V_{A+}, \tag{4}$

and thus:

$BV_{A-} = \begin{bmatrix} A & I \\ I & A \end{bmatrix} \begin{pmatrix} v_A \\ -v_A \end{pmatrix} = \begin{pmatrix} (A - I) v_A \\ (I - A) v_A \end{pmatrix} = \begin{pmatrix} (\lambda_A - 1) v_A \\ -(\lambda_A - 1) v_A \end{pmatrix} = (\lambda_A - 1) V_{A-}. \tag{5}$

(4) and (5) show that the eigenvalues of $B$ are at least those of the form $\lambda_A \pm 1$, but could there be more? We may investigate this possibility via a sort of unravelling of (4)-(5), in the sense that if

$V =\begin{pmatrix} v_1 \\ v_2 \end{pmatrix} \ne 0, \tag{6}$

where $v_1$, $v_2$ are $N$-vectors, is an eigenvector of $B$ with eigenvalue $\mu$, that is,

$BV = \mu V, \tag{7}$

then if we write out (7) in terms of $v_1$, $v_2$ using the definition of $B$ in terms of $A$ we see that

$BV = \begin{bmatrix} A & I \\ I & A \end{bmatrix} \begin{pmatrix} v_1 \\ v_2 \end{pmatrix} = \begin{pmatrix} Av_1 + v_2 \\ v_1 + A v_2 \end{pmatrix} = \mu \begin{pmatrix} v_1 \\ v_2 \end{pmatrix}, \tag{8}$

so that (7)-(8) breaks out into the two equations

$Av_1 + v_2 = \mu v_1 \tag{9}$


$Av_2 + v_1 = \mu v_2. \tag{10}$

We analyze (9)-(10) by first addressing the cases $v_2 = \pm v_1$; in the event that $v_2 = v_1$, (9)-(10) become the same equation

$Av_i + v_i = \mu v_i \tag{11}$


$Av_i = (\mu - 1)v_i \tag{12}$

for $i = 1, 2$. If, on the other hand, $v_2 = -v_1$, (9)-(10) coalesce into

$Av_i - v_1 = \mu v_i \tag{13}$


$Av_i = (\mu + 1)v_i. \tag{14}$

Having dispensed with the possibility that $v_2 = \pm v_1$, we next proceed to the case $v_2 \ne \pm v_1$; then adding (9) and (10) we obtain

$A(v_1 + v_2) + (v_1 + v_2) = \mu (v_1 + v_2) \tag{15}$

whereas subtracting yields

$A(v_1 - v_2) - (v_1 - v_2) = \mu (v_1 - v_2). \tag{16}$

(15) and (16) are equivalent to

$A(v_1 + v_2) = (\mu - 1) (v_1 + v_2) \tag{17}$


$A(v_1 - v_2) = (\mu + 1)(v_1 - v_2), \tag{18}$

respectively. It should be observed that, by (6), at least one of $v_1, v_2 \ne 0$, so that in the case $v_2 = \pm v_1$, (12) and (14) are bona fide eigen-equations for the $v_i$ and $\mu$. Likewise, when $v_2 \ne \pm v_1$, neither $v_1 + v_2 = 0$ nor $v_1 - v_2 = 0$ may hold, so that (17)-(18) are legitimate eigen-equations as well.

A careful scrutiny of (1)-(5) reveals that $\lambda_A \pm 1$ are eigenvalues of $B$ for every eigenvalue $\lambda_A$ of $A$. Furthermore, if $\mu$ is an eigenvalue of $B$, so that (8) holds for some $v_1, v_2$, we see from (9)-(18) that at least one of $\lambda_A = \mu \pm 1$ is in fact an eigenvalue of $A$; then by setting up an equation of the from (4) or (5) we see that $\mu = \lambda_A \mp 1$ in fact arises from an eigenvalue of $A$, so every eigenvalue $\mu$ of $B$ is of this form. The eigenvalues of $B$ are precisely the $\lambda_A \pm 1$, where $\lambda_A$ ranges over the eigenvalues of $A$.

Finally, a few comments concerning the eigenvectors of $A$ and $B$.

Almost Done!!! Your patience is appreciated! Meanwhile,

Hope this helps; Cheerio,

and as always,

***Fiat Lux!!!


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.