7
$\begingroup$

Problem: We have matrix $A\in M_{n-1\times n}(\mathbb Z)$ so that the sum of elements in each row is zero. Prove that $\det(AA^T)=nk^2$, where $k\in \mathbb Z$.

What have I considered so far:

  1. First I thought, since sum of all elements in each row is zero, zero is eigenvalue of $A$ but $A\in M_{(n-1)\times n}(\mathbb Z)$ confused me.
  2. I see that $AA^T$ will be $(n-1)$ by $(n-1)$, so I tried calculating $AA^T$ but I failed to see any connection.

I did some research and found this question but we had no mention of it, which makes me believe I am on the wrong track.

Thank you all for your help.

$\endgroup$
11
  • $\begingroup$ $A$ does not have any eigenvalues at all since it is not a square matrix. If you are suspecting that $A$ has a nontrivial kernel, how should a kernel element look like? $\endgroup$ – Roland Jun 8 '16 at 23:34
  • 2
    $\begingroup$ Hint: Cauchy-Binet theorem. $\endgroup$ – darij grinberg Jun 9 '16 at 0:04
  • 2
    $\begingroup$ Hint to alternative proof: Attach a new row to the bottom of $A $; fill this row with $1$s. Call the resulting $n\times n $-matrix $B $. How does $BB^T $ compare with $AA^T $? What can you say about the determinant of $BB^T $? $\endgroup$ – darij grinberg Jun 9 '16 at 0:08
  • $\begingroup$ @Roland There will be one element all other elements are dependent of, lets say $a_{n-1}$. I will try to think more about this. $\endgroup$ – Asleen Jun 9 '16 at 0:13
  • $\begingroup$ @darijgrinberg We did not mention that theorem, sadly. I will try to think about your second advice. $\endgroup$ – Asleen Jun 9 '16 at 0:13
5
$\begingroup$

Let $A = (B,\mathbf{b})$, where $B \in M_{n-1\times n-1}(\mathbb{Z})$ is a "left square part" of matrix $A$, i.e. $B_i^j = A_i^j$, and $\mathbf{b} \in \mathbb{Z}^{n-1}$ is a "right" part of $A$, i.e. $b_i = A_i^n$. As sum of each row is zero, we have $b_i = -\sum_{k=1}^{n-1}B_i^k$. Let $\mathbf{e}\in \mathbb{Z}^{n-1}$ be a vector with each union component, i.e. $e_i = 1$. Then $\mathbf{b} = -B\mathbf{e}$. Now we may multiply $A$ with $A^{\top}$ as block matrices: $$ AA^{\top} = (B\;\;-B\mathbf{e}) \begin{pmatrix} B^\top \\ -\mathbf{e}^\top B^\top \end{pmatrix} = BB^\top + B\mathbf{e}\mathbf{e}^\top B^\top = B(I + \mathbf{e}\mathbf{e}^\top)B^\top $$ where $I$ is identity matrix. As bouth $B$ and $I + \mathbf{e}\mathbf{e}^\top$ are square matrices we now have $$ \det(AA^\top) = \det\left(B(I + \mathbf{e}\mathbf{e}^\top)B^\top\right) = \det (B) \det(I + \mathbf{e}\mathbf{e}^\top)\det(B^\top) = \det(I + \mathbf{e}\mathbf{e}^\top)\left(\det B\right)^2. $$

$B \in M_{n-1\times n-1}(\mathbb{Z})$ thus $\det B \in \mathbb{Z}$.

Now all we need to proof is that $\det(I + \mathbf{e}\mathbf{e}^\top) = n$. This matrix (denote it with $E_{n-1}$ where $n-1$ is dimension of the space or count of raws in $E$) looks like $$ E_{n-1} = \begin{pmatrix} 2 & 1 & \dots & 1 \\ 1 & 2 & \dots & 1\\ \vdots & \vdots & \ddots & \vdots \\ 1 & 1 & \dots & 2 \end{pmatrix}. $$

Let $E^i_n$ be a matrix $E_n$ in which we replaced $2$ in $i$-th row with $1$. It's easy to see that $\det{E^1_n} = 1$ (using Gaussian process). Thus if $i$ is even $\det E_n^i = 1$ and if i is odd $\det E^i_n = -1$, i.e. $\det E^i_n = (-1)^i$. Now let's suppose we know that $\det E_{n-1} = n$. We may decompose $\det E_n$ using Laplace expansion: $$ \det E_n = 2\det E_{n-1} + \sum_{i=1}^{n-1} (-1)^{i}\det E^i_{n-1} = 2n - (n-1) = n+1. $$ QED.

$\endgroup$
3
  • 2
    $\begingroup$ Another way to prove that the last determinant is $n$ is from its eigenvalues: it has $1$ as an eigenvalue with multiplicity $n-2$, and $n$ as an eigenvalue with multiplicity 1. $\endgroup$ – Semiclassical Jun 9 '16 at 1:03
  • $\begingroup$ @Semiclassical I understand why is $n$ an eigenvalue. Could you please explain why is $1$ an eigenvalue with multiplicity $n-2$? Thank you! $\endgroup$ – Asleen Jun 11 '16 at 21:42
  • 1
    $\begingroup$ @Asleen Well, $(1,-1,0,\cdots)^T$ is such an eigenvector, and you can work out $n-1$ other examples which are all linearly independent. (More heuristically, note that subtracting $1I$ from $E_{n-1}$ gives a matrix which has multiple rows identical. That signals that $1$ will be an eigenvalue of multiplicity bigger than one.) $\endgroup$ – Semiclassical Jun 11 '16 at 21:53
1
$\begingroup$

Rather than partitioning $A$ as in the other answer, we may let a row vector and $A$ adjoin instead. Let $e=(1,\ldots,1)^T\in\mathbb R^n$. Then \begin{align} &\pmatrix{e^T\\ A}\left(\begin{array}{c|c}e&\begin{array}{c}0\\ \hline I_{n-1}\end{array}\end{array}\right)=\pmatrix{n&\ast\\ 0&\ast},\tag{1}\\ &\pmatrix{e^T\\ A}\pmatrix{e&A^T}=\pmatrix{n&0\\ 0&AA^T},\tag{2} \end{align} where the RHS of $(1)$ is an integer matrix. It follows from $(1)$ that $\det\pmatrix{e^T\\ A}=nk$ for some integer $k$, and from $(2)$ that $(nk)^2=n\det(AA^T)$. Thus $\det(AA^T)=nk^2$.

$\endgroup$

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.