Link Search Menu Expand Document

Volumes and determinants

Determinants also relate to volumes of geometric objects. It may seem silly but let’s look at the one-dimensional setting first. (In general, if you don’t understand the 1-d setting, then you should start your learning there.)

In $\mathbb{R}^1 = \mathbb{R}$, a vector is just a scalar or a element of $\mathbb{R}$. The length of the vector $a$ is $|a|$. In 1-d, length is volume. So we have the following relation between the volume of the line segment $P(a)$ spanned by $a$ is \(\operatorname{Vol}(P(a)) = |\det(a)|\) The determinant is a signed volume in 1-d.

With some confidence with one dimension, we turn to 2-d. Let’s start with a $2 \times 2$ matrix \(A = \begin{pmatrix} a & b \\ c & d \end{pmatrix}\) The columns give up two vectors \(\mathbf{v}_1 = \begin{pmatrix} a \\ c \end{pmatrix}, \ \mathbf{v}_2 = \begin{pmatrix} b \\ d \end{pmatrix}\) The vectors $\mathbf{v}_1, \mathbf{v}_2$ span a parallelogram $P = P(A)$.

Let’s check that \(\operatorname{Vol}(P(A)) = |\det(A)|\)

Neither the volume nor the determinant depend on where we place the parallelogram in the plane so assume that the unlabelled vertex is at the origin.

Let’s rotate $P$ so that $\mathbf{v}_1$ lies along the $x$-axis.

Notice that \(\mathbf{v}_1^\prime = \begin{pmatrix} a^\prime \\ 0 \end{pmatrix}, \ \mathbf{v}_2^\prime = \begin{pmatrix} b^\prime \\ d^\prime \end{pmatrix}\) so for the rotated matrix \(\det A^\prime = a^\prime d^\prime\) which is exactly the volume of the rotated parallelogram $P^\prime$ (base $\times$ height).

To finish, we want to see that the determinant also does not change up when rotating in $\mathbb{R}^2$. However, a rotation is a linear transformation! So it can be represented by a matrix when using the standard basis. For rotating counter-clockwise through an angle $\theta$ from $x$-axis, we have \(R_\theta := \begin{pmatrix} \cos(\theta) & -\sin(\theta) \\ \sin(\theta) & \cos(\theta) \end{pmatrix}\) as the matrix. Since \(A^\prime = R_\theta A\) we get \(\det A^\prime = \det(R_\theta) \det(A)\) and \(\det R_\theta = \cos^2 \theta + \sin^2 \theta = 1.\) So the determinant is unchanged also.

We can conclude that \(\operatorname{Vol} P = |\det(A)|\) in 2-d, like 1-d.

We feel pretty good about making the following statement.

For any $n \times n$ matrix $A$ we have \(\operatorname{Vol} P(A) = |\det(A)|\) where $P(A)$ is the parallelopiped spanned by the column vectors of $A$.

In three dimensions, the parallelopiped looks like

For higher dimensions, we rely on our intuition from dimensions 1 through 3.

The proof of \(\operatorname{Vol} P(A) = |\det(A)|\) can go one of two ways:

  • We can brute force things along lines of two dimensions above. Use explicit formulas and manipulations to match up the results.
  • Or we can reach for the power of abstraction and use the properties that completely capture the behavior $\operatorname{Vol}$ and $|\det(A)|$.

Proposition. If \(f,g : \operatorname{Mat}_{n \times n}(\mathbb{R}) \to \mathbb{R}\) are two functions which both satisfy

  • \[\phi(L(i,j,c)A) = \phi(A)\]
  • \[\phi(P(i,j)A) = \phi(A)\]
  • \[\phi(S(i,c)A) = |c|\phi(A)\]
  • \[\phi(I_n) = 1.\]
  • and \(\phi(A) = 0\) for singular $A$.

Then, $f = g$.

Proof. (Expand to view)

From Gaussian Elimination, we know that any matrix can be written as \(A = C U\) where $C$ is a product of $L(i,j,c), S(i,c)$, and $P(i,j)$ and $U$ is in reduced row echelon form. Thus, \(\phi(A) = \phi(U)\) and $\phi$ only depends on the reduced row echelon form. We have seen that $A$ is invertible if and only $U = I_n$. Thus, we have specified $\phi(A)$ for invertible $A$ completely from the properties.

For non-invertible $A$, we know that $\phi(A) = 0$.

Corollary. For any $n \times n$ matrix $A$ we have \(\operatorname{Vol} P(A) = |\det(A)|\) where $P(A)$ is the parallelopiped spanned by the column vectors of $A$.

Proof. (Expand to view)

Both $\operatorname{Vol}(P(-))$ and $|\det(-)|$ satisfy the conditions of the previous proposition.

We have already seen that $\det(A)$ satisfies all the conditions with the exception that \(\det(P(i,j)A) = -\det(A)\) But the absolute value doesn’t change.

For the volume, we know it doesn’t change with rigid motions. Let’s write $\mathbf{v}1,\ldots,\mathbf{v}_n$ be the _row vectors of $A$. Then, we have an intuitively clear formula \(\operatorname{Vol} (P(A)) = \operatorname{ht}(\mathbf{v}_i}) \operatorname{Vol} (P_i)\) where $P_i$ is the parallelopiped spanned by all the vectors $\mathbf{v}1,\ldots,\mathbf{v}_n$ _except $\mathbf{v}_i$ and $\operatorname{ht}(\mathbf{v}_i})$ is height of $\mathbf{v}_i$ above that parallelopiped. (We should really prove this too but we will rely a bit on “geometric intuition” here and below.)

From this formula, we see that \(\operatorname{Vol}(P(P(i,j)A)) = \operatorname{Vol}(P(A))\) and \(\operatorname{Vol}(P(S(i,c)A)) = |c|\operatorname{Vol}(P(A))\) using induction on $n$.

For \(\operatorname{Vol}(P(L(i,j,c)A)) = \operatorname{Vol}(P(A))\) we note that the height of $\mathbf{v}_j + c\mathbf{v}_i$ has the same height above $P_j$ as $\mathbf{v}_j$.

Finally, the volume of $P(I_n)$ is the volume of the unit cube in $n$-dimensions which is $1$. And if $A$ is singular, then it spans at most a $(n-1)$-dimensional subspace and must have $0$ volume.

This corollary tells us that the determinant can also be viewed as a way to provide signed volumes in all dimensions. This is useful for multi-variable integration, for example, which is why you find determinants showing up in change of variable formulae for multi-variable integrals.