30
$\begingroup$

I have been wondering about this question since I was in school. How can one number tell so much about the whole matrix being invertible or not?

I know the proof of this statement now. But I would like to know the intuition behind this result and why this result is actually true.


The proof I have in mind is:

If $A$ is invertible, then

$$ 1 = \det(I) = \det(AA^{-1}) = \det(A)\cdot\det(A^{-1})$$

whence $\det(A) \neq 0$.

Conversely, if $\det(A) \neq 0$, we have

$$ A adj(A) = adj(A)A = \det(A)I$$

whence $A$ is invertible.

$adj(A)$ is the adjugate matrix of $A$.

$$ adj(A)_{ji} = (-1)^{i+j}\det(A_{ij})$$

where $A_{ij}$ is the matrix obtained from $A$ by deleting $ith$ row and $jth$ column.

Any other insightful proofs are also welcome.

$\endgroup$
1
  • 12
    $\begingroup$ The way I tend to remember it is that the determinant gives you the scale factor associated with the transformation represented by the matrix. And any figure scaled to "zero" looks the same… so there's not really enough information left to invert the transformation. $\endgroup$ Sep 28, 2013 at 7:18

7 Answers 7

50
$\begingroup$

Here's an explanation for three dimensional space ($3 \times 3$ matrices). That's the space I live in, so it's the one in which my intuition works best :-).

Suppose we have a $3 \times 3$ matrix $\mathbf{M}$. Let's think about the mapping $\mathbf{y} = f(\mathbf{x}) = \mathbf{M}\mathbf{x}$. The matrix $\mathbf{M}$ is invertible iff this mapping is invertible. In that case, given $\mathbf{y}$, we can compute the corresponding $\mathbf{x}$ as $\mathbf{x} = \mathbf{M}^{-1}\mathbf{y}$.

Let $\mathbf{u}$, $\mathbf{v}$, $\mathbf{w}$ be 3D vectors that form the columns of $\mathbf{M}$. We know that $\det{\mathbf{M}} = \mathbf{u} \cdot (\mathbf{v} \times \mathbf{w})$, which is the volume of the parallelipiped having $\mathbf{u}$, $\mathbf{v}$, $\mathbf{w}$ as its edges.

Now let's consider the effect of the mapping $f$ on the "basic cube" whose edges are the three axis vectors $\mathbf{i}$, $\mathbf{j}$, $\mathbf{k}$. You can check that $f(\mathbf{i}) = \mathbf{u}$, $f(\mathbf{j}) = \mathbf{v}$, and $f(\mathbf{k}) = \mathbf{w}$. So the mapping $f$ deforms (shears, scales) the basic cube, turning it into the parallelipiped with sides $\mathbf{u}$, $\mathbf{v}$, $\mathbf{w}$.

Since the determinant of $\mathbf{M}$ gives the volume of this parallelipiped, it measures the "volume scaling" effect of the mapping $f$. In particular, if $\det{\mathbf{M}} = 0$, this means that the mapping $f$ squashes the basic cube into something flat, with zero volume, like a planar shape, or maybe even a line. A "squash-to-flat" deformation like this can't possibly be invertible because it's not one-to-one --- several points of the cube will get "squashed" onto the same point of the deformed shape. So, the mapping $f$ (or the matrix $\mathbf{M}$) is invertible if and only if it has no squash-to-flat effect, which is the case if and only if the determinant is non-zero.

$\endgroup$
3
  • 5
    $\begingroup$ I signed up just so I could upvote this. Thank you. Very informative and intuitive. $\endgroup$
    – Matt Klein
    Oct 9, 2013 at 2:37
  • $\begingroup$ Thanks, Matt. I'm glad it helped. $\endgroup$
    – bubba
    Oct 9, 2013 at 6:14
  • $\begingroup$ nice explanation ! +1 only for "3D is where we live, so it's where our intuition works best" I was wondering what would it be if we live in higher dimensional space ? :) $\endgroup$
    – Houda
    Jun 8, 2015 at 20:27
15
$\begingroup$

I know this is pretty old, but for the people who might find this in a google search (I know I did), I thought I'd add this.

Remember that the space of all $n \times n$ matrices is isomorphic to the space of all operators on an $n$-dimensional vector space. In other words, the matrix is just some linear operator. Recall that the determinant is the product of the eigenvalues. Both $ \mathbb{C} $ and $ \mathbb{R} $ are integral domains so if the determinant is $0$ then that means you have a zero eigenvalue. That means, if your matrix is $A$, there exists a vector $0 \neq \overline{v} $ such that $ (A-0I)\overline{v}=\overline{0} $ which means $ A\overline{v}=\overline{0} $. Clearly $ \overline{0} $ gets sent to $ \overline{0} $ but so does some other non-zero vector. So the transformation isn't injective and, thus, non-invertible. This is just the intuition though.

$\endgroup$
2
  • 5
    $\begingroup$ This is not just intuition. This is a rigorous proof and a very good one at that! $\endgroup$ Jul 8, 2015 at 23:57
  • $\begingroup$ @VishalGupta With the caveat that linear maps $\mathbb{R}^n\to \mathbb{R}^n$ might not have any eigenvalues at all. For instance the linear map on $\mathbb{R}^2$ or rotation by $\pi/2$. $\endgroup$ Jun 13, 2021 at 13:00
4
$\begingroup$

The absolute value of the determinant of a matrix is the volume of the parallelepiped spanned by the column vectors of that matrix.

Michael

$\endgroup$
3
  • $\begingroup$ Then I would beg to ask for an intuition as to why the determinant is the volume and also how the invertibility of the matrix is related to the volume of the parellelopiped being non-zero. $\endgroup$ Sep 28, 2013 at 8:53
  • $\begingroup$ For the volume thing remember that the determinant of a diagonal matrix equals the product of its diagonal elements. For the invertibility the column vectors need to be lin. indep. What is the volume of a parallelepiped spanned by lin. dep. vectors? $\endgroup$ Sep 28, 2013 at 10:34
  • $\begingroup$ @Vishal Relevant answer to determinant/volume. For the invertibility/volume think about the nullspace of a non-invertible matrix. $\endgroup$
    – jkn
    Sep 28, 2013 at 10:39
2
$\begingroup$

Another classical way is more understandable: note that a determinant is not changed if we add one row to other and one column to another. Thus we obtain a diagonal matrix $B$. This matrix differs from $A$ by matrix-multipliers which correspond to elementary transformations and are invertible. So $A$ is invertible iff $B$ is invertible iff $\det(B) \neq 0$ iff $\det(A) \neq 0$.

$\endgroup$
4
  • $\begingroup$ Sorry, I do not understand the import of your statement. Do you mean to use the row-reduced echelon form? $\endgroup$ Sep 28, 2013 at 7:29
  • $\begingroup$ No, of course, I mean adding also columns (one to another). $\endgroup$ Sep 28, 2013 at 7:36
  • $\begingroup$ OK. But that does the case when $A$ is invertible. If $\det(A) \neq 0$, are we sure that we will land up with identity by performing these operations? $\endgroup$ Sep 28, 2013 at 7:41
  • $\begingroup$ I will write the answer more detaily. $\endgroup$ Sep 28, 2013 at 7:48
1
$\begingroup$

I personally think of the determinant as a function $f(A)$ that has the following three properties:

  1. $f(AB) = f(A) f(B)$

  2. $f(T)$ = product of diagonals for triangular matrix $T$

  3. $f(E) \neq 0$ for an elementary matrix $E$

An elementary matrix $E$ is a matrix such that $EA$ either multiples a row, swaps a row, or swaps and adds a row of $A$.

Now Gaussian elimination is the process of applying elementary matrixes to $A$ and obtaining a upper triangular matrix. That is, every matrix can be written as $$ A = (E_1 \cdots E_n)^{-1} T$$

Which means $$f(A) = f(E_1)^{-1} \cdots f(E_n)^{-1} f(T)$$ and is nonzero if and only if $f(T)$ is nonzero. But a matrix is invertible if and only if the gaussian eliminated form $T$ has nonzero elements in the diagonal. That is, exactly, that $f(T) \neq 0$.


In my mind, any other definitions of the determinant, such as the practical way of computing it with adjunct matrices and such, are not the "real" or "primary" definition of the determinant. We want the determinant function to satisfy the above three properties. We also probably want the determinant to be a polynomial in the entries of our matrix. This then should give us the method of computing the determinant.

$\endgroup$
1
$\begingroup$

With all of the elegant ways of thinking about the determinant, I fear it is sometimes not emphasized enough that the easiest way to discover the determinant is to just solve the linear system \begin{align} a_{11} x_1 + a_{12} x_2 &= b_1 \\ a_{21} x_1 + a_{22} x_2 &= b_2 \end{align} by hand. When you do this, out pops the determinant! And we see immediately that if the determinant is $0$ then our solution doesn't work (because then we would be dividing by $0$). A similar computation can be done for $3 \times 3$ systems, etc.

This is probably how the determinant was first discovered. It is so easy that a middle school student can do it, and at the moment of discovery we see that a nonzero determinant guarantees that $Ax = b$ has a unique solution.

$\endgroup$
0
$\begingroup$

Consider $A$, an $n$ x $n$ matrix to be a mapping from $R^n$ to $R^n$. If the determinant is zero, the columns of $A$ are linearly dependent. This means that the nullspace of $A$ is nonempty. Hence the linear mapping $A$ is non-invertible, since several $x$ get mapped to the same value b, i.e there are multiple solutions (infinite in fact) to $Ax=b$ when a single solution exists. Example - both $x$ and $x+v$ get mapped to b, where $v$ is a vector in the nullspace of $A$

$\endgroup$

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.