4
$\begingroup$

I have a matrix of the following form:

$$A=\begin{bmatrix} 2 & 0 & 0 & 0 \\-1 & 2 & 0 & 0\\ 0 & -1 & 2 & 0\\ 0 & 0 & -1 & 2 \end{bmatrix}$$

which, in general, can be any size with the pattern continued parallel the diagonal. (I didn't know how to show ellipsis in a matrix with LaTeX, feel free to edit if you know how).

I want to find an inverse for this matrix in general for any given size of the matrix. What should I do? I had thought of maybe splitting the matrix up into two matrices, one with all the diagonal terms and one with all the off-diagonal terms and finding the inverse of their sum but I didn't know if that would be helpful/possible.

$\endgroup$
  • $\begingroup$ i would suggets to use induction to generaliase from a few initial cases (if of course a pattern is there, you should check it if you find the pattern) $\endgroup$ – Nikos M. Mar 8 '15 at 2:07
4
$\begingroup$

HINT: The inverse of your matrix is $$\dfrac{1}{16}\begin{bmatrix}8&0&0&0\\4&8&0&0\\2&4&8&0\\1&2&4&8\end{bmatrix}$$ i.e. for a $4\times 4$ matrix, $$\dfrac{1}{2^4}\begin{bmatrix}2^3&0&0&0\\2^2&2^3&0&0\\2^1&2^2&2^3&0\\2^0&2^1&2^2&2^3\end{bmatrix}$$

You can generalize this result for the inverse of an $n\times n$ matrix that follows your pattern (where $2^4\to 2^n$, $2^3\to 2^{n-1}$ and so on).

QUESTION: Can you prove it?

$\endgroup$
  • 1
    $\begingroup$ That's the question I have too. $\endgroup$ – NeutronStar Mar 8 '15 at 2:45
2
$\begingroup$

Instead of using the matrix write it as linear transformation: so the given one is the matrix of $T$ satisfying $Te_i = 2e_i -e_{i+1}$ for $i<n$ and $Te_n = 2e_n$.

Now the inverse $S$ of this $T$ has to send the vectors in the rhs to those in the corresponding lhs. $S(2e_n) = e_n$ and so $ S(e_n) =\frac12e_n$. Then $S(2e_{n-1})= Se_n+e_{n-1}$ And so $Se_{n-1} = \frac12e_{n-1} +\frac12e_n$. Ths way you can work it out for any size without going through inverse calculation (though both are esentally the same).

$\endgroup$
1
$\begingroup$

Here's sort of an abstract algebra- flavored approach:

Notice that the given matrix may be written as

$A = 2I - N, \tag{1}$

where $I$ is the $4 \times 4$ identity matrix and $N$ is the $4 \times 4$ matrix with ones along the subdiagonal and zeroes elsewhere. The present solution extends to similarly structured matrices of any size $n$, so we will use the same notation to represent the corresponding matrices of size $n$. Indeed, much of what is said here will hold independently of $n$; there is a very general, abstract solution to this problem. In fact, there is no essential reason to restrict the multiplier of the matrix $I$ in (1) to the value $2$; we will in fact consider the case

$A = \alpha I - N \tag{2}$

since the essential operations are the same for any non-zero value of $\alpha$. We observe that $N$ is in fact always a nilpotent matrix; that is, no matter what the value of the size $n$ may be, there is always a positive integer $m$ such that

$N^m = 0. \tag{3}$.

For the case $n = 4$ we may in fact take $m = 4$ as well, as direct calculation reveals; indeed, we have in this case

$N = \begin{bmatrix} 0 & 0 & 0 & 0 \\ 1 & 0 & 0 & 0 \\ 0 & 1 & 0 & 0 \\ 0 & 0 & 1 & 0 \end{bmatrix}, \tag{4}$

$N^2 = NN =\begin{bmatrix} 0 & 0 & 0 & 0 \\ 1 & 0 & 0 & 0 \\ 0 & 1 & 0 & 0 \\ 0 & 0 & 1 & 0 \end{bmatrix} \begin{bmatrix} 0 & 0 & 0 & 0 \\ 1 & 0 & 0 & 0 \\ 0 & 1 & 0 & 0 \\ 0 & 0 & 1 & 0 \end{bmatrix} = \begin{bmatrix} 0 & 0 & 0 & 0 \\ 0 & 0 & 0 & 0 \\ 1 & 0 & 0 & 0 \\ 0 & 1 & 0 & 0 \end{bmatrix}, \tag{5}$

$N^3 = N^2 N = \begin{bmatrix} 0 & 0 & 0 & 0 \\ 0 & 0 & 0 & 0 \\ 1 & 0 & 0 & 0 \\ 0 & 1 & 0 & 0 \end{bmatrix}\begin{bmatrix} 0 & 0 & 0 & 0 \\ 1 & 0 & 0 & 0 \\ 0 & 1 & 0 & 0 \\ 0 & 0 & 1 & 0 \end{bmatrix} = \begin{bmatrix} 0 & 0 & 0 & 0 \\ 0 & 0 & 0 & 0 \\ 0 & 0 & 0 & 0 \\ 1 & 0 & 0 & 0 \end{bmatrix}; \tag{6}$

from this point showing that

$N^4 = N^3 N = 0 \tag{7}$

is so simple that I'll omit the details. The essential fact we shall need here is that, for $N$ of size $n$, we have

$N^n = 0, \tag{8}$

but

$N^m \ne 0 \tag{9}$

for $0 \le m < n$. Furthermore, the pattern exhibited in the powers of $N$ (4) - (8), in which successive multiplications by $N$ each move the diagonal of ones a step further toward the lower left-hand corner, until $N^{n - 1}$ finally has but a lonely one in $n, 1$ postion, and zeroes everywhere else:

$(N^{n - 1})_{n1} = 1, \;\; (N^{n -1})_{ij} = 0, (i, j) \ne (n, 1). \tag{10}$

Knowledge of these particular properties of the $N^k$ will play a useful role in the explicit calculation of $(\alpha I - N)^{-1}$, as will be made evident in what follows. The preceding assertions are indeed quite well-known and so I have forgone the effort of offering detailed proofs. Finally, we observe that $N$ is not invertible; to see this, note that since $N^{n - 1} \ne 0$, there is some vector $v$ with $N^{n - 1} v \ne 0$; then $N(N^{n - 1} v) = N^n v = 0$; thus $0 \ne N^{n - 1} v \in \ker N$; $\ker N \ne \{0 \}$; $N^{-1}$ does not exist.

Bearing the above in mind, we compute the inverse of $\alpha I - N$, $\alpha \ne 0$, as follows (note that we have just eliminated the case $\alpha = 0$ by showing that $N$ has no inverse): we write

$A = \alpha I - N = \alpha (I - \alpha^{-1}N), \tag{11}$

and note that we have

$(I - \alpha^{-1}N)(\sum_0^{n - 1} (\alpha^{-1}N)^i) = I - (\alpha^{-n}N^n) = I; \tag{12}$

(12) shows that $I - \alpha^{-1}N$ is invertible with inverse $\sum_0^{n - 1} (\alpha^{-1}N)^i$, which sum terminates with $(\alpha^{-1} N)^{n - 1}$ since $N^n = 0$. Furthermore, $(I - \alpha^{-1}N)^{-1} = \sum_0^{n - 1} (\alpha^{-1}N)^i$ is relatively easy to calculate by virtue of the peculiar structure of the matrices $N^i$ pointed out in the above; indeed, having understood the structure of the powers of $N$ in general terms, it is not even necessary to formally add or multiply any matrices; we merely need to insert the powers of $\alpha^{-1}$ in the appropriate places, since the sum $\sum_0^{n - 1} (\alpha^{-1} N)^i$ almost directly describes the entries of a lower triangular matrix. Thus, the inverse of $I - \alpha^{-1}N$ is most easily had.

Based on the above, and observing that

$\alpha I - N = \alpha(I - \alpha^{-1}N), \tag{13}$

we see that

$(\alpha I - N)^{-1} = \alpha^{-1}(I - \alpha^{-1}N)^{-1} = \alpha^{-1}\sum_0^{n - 1} (\alpha^{-1}N)^i = \sum_0^{n - 1} \alpha^{-(i + 1)} N^i. \tag{14}$

Formula (14) presents a relatively simple and efficient algorithm for calculating $(\alpha I - N)^{-1}$; we apply it to the matrix $A$ of (1) by taking $\alpha = 2$, $n = 4$; then, using (4)-(6) in the role of what is essentially a template, we simply write down the result as

$(2 I - N)^{-1} = \begin{bmatrix} \dfrac{1}{2} & 0 & 0 & 0 \\ \dfrac{1}{4} & \dfrac{1}{2} & 0 & 0 \\ \dfrac{1}{8} & \dfrac{1}{4} & \dfrac{1}{2} & 0 \\ \dfrac{1}{16} & \dfrac{1}{8} & \dfrac{1}{4} & \dfrac{1}{2} \end{bmatrix}, \tag{15}$

in accord with the answer given by Demosthene, who asked

Can you prove it?

The preceeding arguments present both a description of the general pattern and a proof that this pattern applies. QED.

$\endgroup$
  • $\begingroup$ Very well done, sir. Here, have my upvote :) $\endgroup$ – Demosthene Mar 9 '15 at 7:48
  • $\begingroup$ @Demosthene: Thank you, my friend! Cheers! $\endgroup$ – Robert Lewis Mar 9 '15 at 9:36

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.