Linear AlgebraDeterminants6 of 6
Determinants · 6

The Invertible Matrix Theorem: One Fact, Many Faces

We have met determinants, inverses, span, linear independence, null space, and rank as separate tests, and now we want to see how they all connect for a square matrix.

1

One Fact, Many Faces

The Invertible Matrix Theorem says that for an n×nn \times n square matrix, all of those tests ask the exact same geometric question: does the matrix keep nn-dimensional space whole, or does it squish space into a lower dimension?

Here is why matching the input and output dimensions leaves only those two possibilities, and why undoing a square matrix from just one side is enough.

No Squish Means Invertible
The Invertible Matrix Theorem
1.\htmlClass{lc-mute}{1.}
A is invertibleA \text{ is invertible}
2.\htmlClass{lc-mute}{2.}
det⁡(A)≠0\htmlClass{lc-blue}{\det(A)} \neq 0
3.\htmlClass{lc-mute}{3.}
independent columns\text{independent columns}
4.\htmlClass{lc-mute}{4.}
one-to-one\text{one-to-one}
5.\htmlClass{lc-mute}{5.}
onto\text{onto}
6.\htmlClass{lc-mute}{6.}
RREF(A)=I\text{RREF}(A) = I
7.\htmlClass{lc-mute}{7.}
Null(A)={0}\htmlClass{lc-orange}{\text{Null}(A)} = \{0\}
⋮\htmlClass{lc-mute}{\vdots}

Because a square matrix cannot squish space into a lower dimension and then be un-squished back to full dimension, finding a matrix that undoes AA from just one side proves that AA does not squish space, so the same matrix automatically works on the other side too:

CA\displaystyle \undo{C}A =IorAB\displaystyle = I \quad \text{or} \quad A\undo{B} =I  ⟹  A−1 exists and C\displaystyle = I \implies \undo{A^{-1}} \text{ exists and } \undo{C} =B\displaystyle = \undo{B} =A−1\displaystyle = \undo{A^{-1}}
2

Determinants and the Columns of AA

Now we want to spot whether a square matrix squishes space by checking a single number, its determinant, or by looking at where its columns land.

Determinants and the Columns of A
11
Text solution
Independent columns keep the grid openIn matrix AA, the two columns point in different directions, so the parallelogram between them has nonzero area and the output still fills 22 dimensions.
A=[2112],det⁡(A)≠0,Rank(A)=2\displaystyle A = \begin{bmatrix} \colone{2} & \coltwo{1} \\ \colone{1} & \coltwo{2} \end{bmatrix}, \quad \area{\det(A)} \neq 0, \quad \text{Rank}(A) = 2
Dependent columns squish the grid onto a lineIn matrix BB, the second column is twice the first column, so both grid lines collapse onto the same line and the area drops to zero.
B=[2412],[42]=2[21],det⁡(B)=0,Rank(B)=1\displaystyle B = \begin{bmatrix} \colone{2} & \coltwo{4} \\ \colone{1} & \coltwo{2} \end{bmatrix}, \quad \coltwo{\begin{bmatrix} 4 \\ 2 \end{bmatrix}} = 2\colone{\begin{bmatrix} 2 \\ 1 \end{bmatrix}}, \quad \area{\det(B)} = 0, \quad \text{Rank}(B) = 1

Now try dragging the second column around the plane to see the exact moment the grid collapses.

The 2D grid covers the whole plane unless the second column lands on the first column's line and squishes everything flat.

When the nn columns of an n×nn \times n matrix point in nn independent directions, no dimension is lost and the columns reach every vector in Rn\mathbb{R}^n:

det⁡(A)≠0  ⟺  columns of Aare linearlyindependent, span Rn, and forma basis for Rn  ⟺  Rank(A)=n\displaystyle \begin{aligned} & \area{\det(A)} \neq 0 \\ & \iff{} \text{columns of } A \text{are linearly} \\ & \text{independent, span } \mathbb{R}^n\text{, and form} \\ & \text{a basis for } \mathbb{R}^n \\ & \iff{} \text{Rank}(A) = n \end{aligned}
3

Solving Ax=bAx = \target{b} and the Null Space

Next, we want to know how many solutions the equation Ax=bAx = \target{b} has for a target vector b\target{b}, and which inputs get crushed into the origin when b=0\target{b} = 0.

Solving Ax = b and the Null Space
Ax=bAx = \htmlClass{lc-blue}{b}
A\htmlClass{lc-violet}{A}
x=?x = \enspace ?
b\htmlClass{lc-blue}{b}

Try moving the target vector b\target{b} on both grids to see when Ax=bAx = \target{b} has one solution, no solution, or a whole line of solutions.

An open grid reaches every output point with exactly one input, while a squished grid misses points off the line and crowds a whole line of inputs onto each point on the line.

Because the origin x=0x = 0 always maps to 00, checking Ax=0Ax = 0 tells us whether any other vectors also land on the origin, which decides whether Ax=bAx = \target{b} has a unique solution everywhere:

Ax=0  ⟹  x=0(Null(A)={0}, Nullity(A)=0)  ⟺  A is one-to-oneand onto  ⟺  Ax=bhas a uniquesolution for every b∈Rn\displaystyle \begin{aligned} & Ax = 0 \implies x = 0 \\ & (\lost{\text{Null}(A)} = \{0\}, \ \lost{\text{Nullity}(A)} = 0) \\ & \iff{} A \text{ is one-to-one} \text{and onto} \\ & \iff{} Ax = \target{b} \text{has a unique} \\ & \text{solution for every } \target{b} \in \mathbb{R}^n \end{aligned}
4

Pivots, Row Reduction, and the Rows of AA

For a 3×33 \times 3 or larger matrix, we usually cannot tell just by looking whether one column is a combination of the others, so we row reduce AA to see whether every dimension survives as a pivot.

Remember that transposing a square matrix flips its rows into columns without changing the determinant:

det⁡(AT)\displaystyle \area{\det(A^T)} =det⁡(A)\displaystyle = \area{\det(A)}

A squished dimension shows up as a row of all zeros during row reduction, and det⁡(AT)=det⁡(A)\area{\det(A^T)} = \area{\det(A)} guarantees that the rows of AA behave the exact same way as its columns.

Pivots, Row Reduction, and the Rows of A
[121251130]\begin{bmatrix} \htmlClass{lc-green}{1} & \htmlClass{lc-pink}{2} & \htmlClass{lc-violet}{1} \\ \htmlClass{lc-green}{2} & \htmlClass{lc-pink}{5} & \htmlClass{lc-violet}{1} \\ \htmlClass{lc-green}{1} & \htmlClass{lc-pink}{3} & \htmlClass{lc-violet}{0} \end{bmatrix}
3D whole,\text{3D whole,}
or squished?\text{or squished?}
x\htmlClass{lc-mute}{x}
y\htmlClass{lc-mute}{y}
z\htmlClass{lc-mute}{z}
Text solution
Clear below the first pivotUse the top-left pivot 1\pivot{1} to eliminate the entries below it in the first column.
[121251130]→R2←R2−2R1R3←R3−R1[12101−101−1]\displaystyle \begin{bmatrix} \pivot{1} & 2 & 1 \\ 2 & 5 & 1 \\ 1 & 3 & 0 \end{bmatrix} \xrightarrow{\substack{R_2 \leftarrow R_2 - 2R_1 \\ R_3 \leftarrow R_3 - R_1}} \begin{bmatrix} \pivot{1} & 2 & 1 \\ 0 & 1 & -1 \\ 0 & 1 & -1 \end{bmatrix}
Eliminate below the second pivot to expose a zero rowSubtracting R2R_2 from R3R_3 turns the bottom row into all zeros, leaving only 22 pivots and showing that 3D space got squished onto a 2D plane.
[12101−101−1]→R3←R3−R2[12101−1000]\displaystyle \begin{bmatrix} 1 & 2 & 1 \\ 0 & 1 & -1 \\ 0 & 1 & -1 \end{bmatrix} \xrightarrow{R_3 \leftarrow R_3 - R_2} \begin{bmatrix} \pivot{\boxed{1}} & 2 & 1 \\ 0 & \pivot{\boxed{1}} & -1 \\ \lost{0} & \lost{0} & \lost{0} \end{bmatrix}
Compare with three surviving pivotsWhen the bottom-right entry of the starting matrix is 22 instead of 00, row reduction produces 33 pivots, one in every row and column, and simplifies all the way to the identity matrix II.
[121251132]→R2←R2−2R1R3←R3−R1[12101−1011]→R3←R3−R2[12101−1002]→RREF[100010001]=I\displaystyle \begin{aligned} & \begin{bmatrix} 1 & 2 & 1 \\ 2 & 5 & 1 \\ 1 & 3 & 2 \end{bmatrix} \\ &\xrightarrow{\substack{R_2 \leftarrow R_2 - 2R_1 \\ R_3 \leftarrow R_3 - R_1}} \begin{bmatrix} 1 & 2 & 1 \\ 0 & 1 & -1 \\ 0 & 1 & 1 \end{bmatrix} \\ &\xrightarrow{R_3 \leftarrow R_3 - R_2} \begin{bmatrix} \pivot{\boxed{1}} & 2 & 1 \\ 0 & \pivot{\boxed{1}} & -1 \\ 0 & 0 & \pivot{\boxed{2}} \end{bmatrix} \\ &\xrightarrow{\text{RREF}} \begin{bmatrix} 1 & 0 & 0 \\ 0 & 1 & 0 \\ 0 & 0 & 1 \end{bmatrix} = I \end{aligned}
5

Three Ways to Use the Theorem

On a test, some questions look like a lot of algebra while others take a single check. We use the Invertible Matrix Theorem to answer several slow questions with one pivot count, or to figure out everything a square matrix does from a single clue, even when we aren't given any entries of the matrix.

Three Ways to Use the Theorem
v1=[121]\htmlClass{lc-green}{v_1 = \begin{bmatrix} 1 \\ 2 \\ 1 \end{bmatrix}}
v2=[253]\htmlClass{lc-pink}{v_2 = \begin{bmatrix} 2 \\ 5 \\ 3 \end{bmatrix}}
v3=[112]\htmlClass{lc-violet}{v_3 = \begin{bmatrix} 1 \\ 1 \\ 2 \end{bmatrix}}
(a) span R3?\text{(a) span } \mathbb{R}^3?
(b) independent?\text{(b) independent?}
(c) Ax=[7−311]: unique?\text{(c) } Ax = \htmlClass{lc-blue}{\begin{bmatrix} 7 \\ -3 \\ 11 \end{bmatrix}}\text{: unique?}
(d) Null(A)=?\text{(d) } \htmlClass{lc-orange}{\text{Null}(A)} = \enspace ?
Text solution
Answer four vector questions with one pivot countPlacing v1=⟨1,2,1⟩\colone{v_1} = \colone{\langle 1, 2, 1 \rangle}, v2=⟨2,5,3⟩\coltwo{v_2} = \coltwo{\langle 2, 5, 3 \rangle}, and v3=⟨1,1,2⟩\colthree{v_3} = \colthree{\langle 1, 1, 2 \rangle} into the columns of a 3×33 \times 3 matrix AA gives 33 pivots (no squish), which immediately answers every part without using ⟨7,−3,11⟩\target{\langle 7, -3, 11 \rangle}.
A=[121251132] has 3 pivots  ⟹  span R3, independent,unique xfor Ax=[7−311], and Null(A)={0}\displaystyle \begin{aligned} & A = \begin{bmatrix} \colone{1} & \coltwo{2} & \colthree{1} \\ \colone{2} & \coltwo{5} & \colthree{1} \\ \colone{1} & \coltwo{3} & \colthree{2} \end{bmatrix} \text{ has } 3 \text{ pivots} \\ & \implies{} \text{span } \mathbb{R}^3\text{, independent,} \\ &\quad \text{unique } x \text{for } Ax \\ &\quad = \target{\begin{bmatrix} 7 \\ -3 \\ 11 \end{bmatrix}}\text{, and } \lost{\text{Null}(A)} \\ &\quad = \{0\} \end{aligned}
Deduce every property from a no-squish clue (4×44 \times 4 matrix AA)Knowing that Ax=0Ax = 0 has only the solution x=0x = 0 tells us that no dimension was crushed, so every statement in the theorem holds for AA.
Null(A)={0}  ⟹  Ax=bsolvable for all b∈R4,Rank(A)=4,rows independent (AT invertible),RREF(A)=I4\displaystyle \begin{aligned} & \lost{\text{Null}(A)} = \{0\} \\ \implies{} & Ax = \target{b} \\ & \text{solvable for all } \target{b} \in \mathbb{R}^4, \\ & \text{Rank}(A) = 4, \\ & \text{rows independent } \\ & (A^T \text{ invertible}), \\ & \text{RREF}(A) = I_4 \end{aligned}
Deduce every failure from a squish clue (5×55 \times 5 matrix MM)Knowing that the columns of MM are linearly dependent tells us that 5D space collapsed into at most 4 dimensions, so every statement in the theorem fails together.
Dependent columns  ⟹  det⁡(M)=0,Rank(M)≤4(zero row in RREF),Mx=0 has ∞ solutions,Mx=b has 0 or ∞ solutions\displaystyle \begin{aligned} & \text{Dependent columns} \\ \implies{} & \area{\det(M)} = 0, \\ & \text{Rank}(M) \le 4 \\ & \text{(zero row in RREF)}, \\ & Mx = 0 \text{ has } \infty \text{ solutions}, \\ & Mx = \target{b} \text{ has } 0 \text{ or } \infty \text{ solutions} \end{aligned}

Watch out for a common exam trap when a square matrix MM squishes space: Mx=0Mx = 0 never has "no solution." Because x=0x = 0 always works, squishing a dimension means an entire line or plane of nonzero vectors lands on 00, so Mx=0Mx = 0 jumps straight from one solution to infinitely many solutions.

6

Why the Matrix Must Be Square

Before using the Invertible Matrix Theorem, we must check that the matrix is square (n×nn \times n). For a rectangular m×nm \times n matrix (m≠nm \neq n), the input and output spaces have different dimensions, so "not squishing" (one-to-one) and "filling the output space" (onto) split apart.

Let's see how a tall 3×23 \times 2 matrix and a wide 2×32 \times 3 matrix break the link between one-to-one and onto, and then put the two cases for a square matrix side by side.

Why the Matrix Must Be Square
T=[100100]T = \begin{bmatrix} \htmlClass{lc-green}{1} & \htmlClass{lc-pink}{0} \\ \htmlClass{lc-green}{0} & \htmlClass{lc-pink}{1} \\ \htmlClass{lc-green}{0} & \htmlClass{lc-pink}{0} \end{bmatrix}
W=[102013]W = \begin{bmatrix} \htmlClass{lc-green}{1} & \htmlClass{lc-pink}{0} & \htmlClass{lc-violet}{2} \\ \htmlClass{lc-green}{0} & \htmlClass{lc-pink}{1} & \htmlClass{lc-violet}{3} \end{bmatrix}
3×2\htmlClass{lc-mute}{3 \times 2}
2×3\htmlClass{lc-mute}{2 \times 3}
one-to-one?\text{one-to-one?}
onto?\text{onto?}
one-to-one?\text{one-to-one?}
onto?\text{onto?}
Text solution
Tall 3×23 \times 2 matrix TT (one-to-one, but not onto)The two columns of TT point along the xx- and yy-axes without collapsing, so Tx=0Tx = 0 has only the trivial solution x=0x = 0, but the bottom row of zeros forces every output to have z=0z = 0, so the columns cannot span R3\mathbb{R}^3.
T=[100100]:independent columns (Tx=0  ⟹  x=0),Col(T)≠R3 (not onto)\displaystyle \begin{aligned} & T = \begin{bmatrix} \colone{1} & \coltwo{0} \\ \colone{0} & \coltwo{1} \\ \colone{0} & \coltwo{0} \end{bmatrix}: \\ & \text{independent columns } \\ & (Tx = 0 \implies x = 0), \\ & \covered{\text{Col}(T)} \neq \mathbb{R}^3 \text{ (not onto)} \end{aligned}
Wide 2×32 \times 3 matrix WW (onto, but not one-to-one)The first two columns of WW already span all of R2\mathbb{R}^2, which puts the third column ⟨2,3⟩\colthree{\langle 2, 3 \rangle} in the same 2D plane and forces a dimension of 3D input space to be squished.
W=[102013]:columns span R2 (onto),dependent columns(not one-to-one)\displaystyle \begin{aligned} & W = \begin{bmatrix} \colone{1} & \coltwo{0} & \colthree{2} \\ \colone{0} & \coltwo{1} & \colthree{3} \end{bmatrix}: \\ &\quad \text{columns span } \mathbb{R}^2 \text{ (onto)}, \\ &\quad \text{dependent columns} \\ &\quad \text{(not one-to-one)} \end{aligned}

Key concepts

Before using the theorem on a problem, look at the size of the matrix. If the matrix is rectangular (m×nm \times n with m≠nm \neq n), do not use the Invertible Matrix Theorem. Check one-to-one and onto separately, because a tall matrix like T=[100100]T = \begin{bmatrix} \colone{1} & \coltwo{0} \\ \colone{0} & \coltwo{1} \\ \colone{0} & \coltwo{0} \end{bmatrix} can have independent columns (Tx=0  ⟹  x=0Tx = 0 \implies x = 0) without spanning R3\mathbb{R}^3, and a wide matrix like W=[102013]W = \begin{bmatrix} \colone{1} & \coltwo{0} & \colthree{2} \\ \colone{0} & \coltwo{1} & \colthree{3} \end{bmatrix} can span R2\mathbb{R}^2 while forcing its three columns in 2D to be dependent.

When the matrix AA is square (n×nn \times n), pick whichever test is fastest to decide which of the two cases you are in:

For a 2×22 \times 2 matrix

check whether one column is a multiple of the other, like ⟨4,2⟩=2⟨2,1⟩\coltwo{\langle 4, 2 \rangle} = 2\colone{\langle 2, 1 \rangle} in B=[2412]B = \begin{bmatrix} \colone{2} & \coltwo{4} \\ \colone{1} & \coltwo{2} \end{bmatrix}, or compute det⁡(A)\area{\det(A)}.

For a 3×33 \times 3 or larger matrix of numbers

row reduce just far enough to count pivots: nn pivots (such as [12101−1002]\begin{bmatrix} \pivot{1} & 2 & 1 \\ 0 & \pivot{1} & -1 \\ 0 & 0 & \pivot{2} \end{bmatrix}) means no squish, whereas a row of all zeros means space got squished.

For a conceptual clue without entries

check which of the two cases the clue puts you in:

No squish (all hold):A−1 exists (CA=I or AB=I)  ⟺  det⁡(A)=det⁡(AT)≠0  ⟺  n pivots (RREF(A)=I, Rank(A)=n)  ⟺  Null(A)={0}(Nullity(A)=0)\displaystyle \begin{aligned} & \text{No squish (all hold):} \\ & \undo{A^{-1}} \text{ exists } (\undo{C}A = I \text{ or } A\undo{B} = I) \\ \iff{} & \area{\det(A)} = \area{\det(A^T)} \neq 0 \\ \iff{} & n \text{ pivots } \\ & (\text{RREF}(A) = I, \ \text{Rank}(A) = n) \\ \iff{} & \lost{\text{Null}(A)} = \{0\} \\ & \text{(\lost{Nullity}}\lost{(A)} = 0\text{)} \end{aligned}

When AA does not squish space, both the columns and the rows of AA are linearly independent, span Rn\mathbb{R}^n, and form a basis for Rn\mathbb{R}^n, ATA^T is invertible, and AA is one-to-one and onto (Ax=bAx = \target{b} has a unique solution for every b∈Rn\target{b} \in \mathbb{R}^n).

Squish (all fail):A not invertible  ⟺  det⁡(A)=det⁡(AT)=0  ⟺  Rank(A)<n(zero row in RREF)  ⟺  Nullity(A)≥1(Ax=0 has ∞ solutions)\displaystyle \begin{aligned} & \text{Squish (all fail):} \\ & A \text{ not invertible} \\ \iff{} & \area{\det(A)} = \area{\det(A^T)} = 0 \\ \iff{} & \text{Rank}(A) < n \\ & \text{(zero row in RREF)} \\ \iff{} & \lost{\text{Nullity}(A)} \ge 1 \\ & \text{(} Ax = 0 \text{ has } \infty \text{ solutions)} \end{aligned}

When AA squishes space, the columns and rows are linearly dependent and fail to span Rn\mathbb{R}^n, ATA^T is not invertible, and Ax=bAx = \target{b} has either 00 solutions (when b\target{b} is off the squished output space) or ∞\infty solutions (when b\target{b} is on it), never a unique solution.

7

Practice

1

Consider the matrix A=[3−6−12]A = \begin{bmatrix} \colone{3} & \coltwo{-6} \\ \colone{-1} & \coltwo{2} \end{bmatrix}. By looking at its columns without row reducing, decide whether AA squishes 2D space, and find det⁡(A)\area{\det(A)}, whether the columns of AA form a basis for R2\mathbb{R}^2, and Rank(A)\text{Rank}(A).

Compare the second column ⟨−6,2⟩\coltwo{\langle -6, 2 \rangle} with the first column ⟨3,−1⟩\colone{\langle 3, -1 \rangle}: is one a scalar multiple of the other?

The second column ⟨−6,2⟩\coltwo{\langle -6, 2 \rangle} is −2-2 times the first column ⟨3,−1⟩\colone{\langle 3, -1 \rangle}, so both columns lie on the same line and AA squishes 2D space onto a 1D line. Therefore det⁡(A)=0\area{\det(A)} = 0, and the columns do not form a basis for R2\mathbb{R}^2 (they are linearly dependent and do not span R2\mathbb{R}^2). Be careful not to say Rank(A)=0\text{Rank}(A) = 0 just because det⁡(A)=0\area{\det(A)} = 0: the area of the parallelogram collapsed to 00, but the grid squished onto a 1D line rather than a single point, so Rank(A)=1\text{Rank}(A) = 1.
2

Let v1=⟨1,2,−1⟩\colone{v_1} = \colone{\langle 1, 2, -1 \rangle}, v2=⟨2,5,1⟩\coltwo{v_2} = \coltwo{\langle 2, 5, 1 \rangle}, and v3=⟨1,3,3⟩\colthree{v_3} = \colthree{\langle 1, 3, 3 \rangle}, and let AA be the 3×33 \times 3 matrix with columns v1\colone{v_1}, v2\coltwo{v_2}, and v3\colthree{v_3}. Row reduce AA to count its pivots, and use the result to answer all four questions: (a) Do v1,v2,v3\colone{v_1}, \coltwo{v_2}, \colthree{v_3} span R3\mathbb{R}^3? (b) Are v1,v2,v3\colone{v_1}, \coltwo{v_2}, \colthree{v_3} linearly independent? (c) Does Ax=⟨4,−9,7⟩Ax = \target{\langle 4, -9, 7 \rangle} have a unique solution? (d) What is Null(A)\lost{\text{Null}(A)}?

Place v1,v2,v3\colone{v_1}, \coltwo{v_2}, \colthree{v_3} as the columns of AA and eliminate below the first two pivots. Do you need to include ⟨4,−9,7⟩\target{\langle 4, -9, 7 \rangle} in the row reduction to know whether the solution is unique?

Placing the vectors into the columns gives A=[121253−113]A = \begin{bmatrix} \colone{1} & \coltwo{2} & \colthree{1} \\ \colone{2} & \coltwo{5} & \colthree{3} \\ \colone{-1} & \coltwo{1} & \colthree{3} \end{bmatrix}. Applying R2←R2−2R1R_2 \leftarrow R_2 - 2R_1 and R3←R3+R1R_3 \leftarrow R_3 + R_1 (watch the sign on R3+R1R_3 + R_1: subtracting R1R_1 instead of adding would leave −2-2 in the bottom-left entry instead of 00) gives [121011034]\begin{bmatrix} 1 & 2 & 1 \\ 0 & 1 & 1 \\ 0 & 3 & 4 \end{bmatrix}, and then R3←R3−3R2R_3 \leftarrow R_3 - 3R_2 gives [121011001]\begin{bmatrix} \pivot{1} & 2 & 1 \\ 0 & \pivot{1} & 1 \\ 0 & 0 & \pivot{1} \end{bmatrix}, which has 33 pivots (no squish). By the Invertible Matrix Theorem: (a) Yes, they span R3\mathbb{R}^3. (b) Yes, they are linearly independent. (c) Yes, Ax=⟨4,−9,7⟩Ax = \target{\langle 4, -9, 7 \rangle} has a unique solution. (d) Null(A)={0}\lost{\text{Null}(A)} = \{0\} (write the set containing the zero vector {0}\{0\}, not the number 00, which is Nullity(A)\lost{\text{Nullity}(A)}).
3

Suppose AA is a 4×44 \times 4 matrix whose reduced row echelon form is I4I_4. Without knowing any entries of AA, find: (a) whether Ax=bAx = \target{b} is solvable for every b∈R4\target{b} \in \mathbb{R}^4, (b) Nullity(A)\lost{\text{Nullity}(A)}, (c) whether ATA^T is invertible, and (d) whether the rows of AA form a basis for R4\mathbb{R}^4.

Count the pivots of I4I_4 to see whether AA squishes R4\mathbb{R}^4, and use det⁡(AT)=det⁡(A)\area{\det(A^T)} = \area{\det(A)} for the questions about ATA^T and the rows of AA.

Since RREF(A)=I4\text{RREF}(A) = I_4, AA has 44 pivots and does not squish R4\mathbb{R}^4. Therefore: (a) Yes, Ax=bAx = \target{b} has a unique solution for every b∈R4\target{b} \in \mathbb{R}^4. (b) Nullity(A)=0\lost{\text{Nullity}(A)} = 0 (don't confuse nullity with rank and answer 44: all 44 dimensions survived as Rank(A)=4\text{Rank}(A) = 4, so 00 dimensions were crushed into the null space). (c) Yes, ATA^T is invertible because det⁡(AT)=det⁡(A)≠0\area{\det(A^T)} = \area{\det(A)} \neq 0. (d) Yes, the rows of AA are linearly independent, span R4\mathbb{R}^4, and form a basis for R4\mathbb{R}^4.
4

Suppose MM is a 5×55 \times 5 matrix whose columns are linearly dependent. Determine: (a) det⁡(M)\area{\det(M)}, (b) what can be said about Rank(M)\text{Rank}(M) and the bottom row of RREF(M)\text{RREF}(M), (c) how many solutions Mx=0Mx = 0 has, and (d) how many solutions Mx=bMx = \target{b} can have for a vector b∈R5\target{b} \in \mathbb{R}^5.

Dependent columns mean MM squishes 5D space into a lower dimension. What happens to the subspace of inputs that gets crushed into 00, and what happens when a target b\target{b} lies off versus on the squished output space?

Dependent columns mean MM squishes 5D space into a lower dimension, so every statement of the Invertible Matrix Theorem fails together: (a) det⁡(M)=0\area{\det(M)} = 0. (b) Rank(M)≤4\text{Rank}(M) \le 4 (strictly less than 55), and RREF(M)\text{RREF}(M) has at least one row of all zeros at the bottom. (c) Mx=0Mx = 0 has infinitely many solutions (watch out for the mistake of answering "zero solutions": x=0x = 0 always works, and squishing space crushes an entire line or higher-dimensional subspace of nonzero vectors into the origin). (d) Mx=bMx = \target{b} has either zero solutions (if b\target{b} lies off the squished output space) or infinitely many solutions (if b\target{b} lies on it); it can never have a unique solution.
5

Suppose AA and C\undo{C} are 4×44 \times 4 matrices such that CA=I4\undo{C}A = I_4. Do you need to check whether AC=I4A\undo{C} = I_4 to know if AA is invertible? Explain using the grid picture, state what ACA\undo{C} equals, and state how many solutions Ax=bAx = \target{b} has for each b∈R4\target{b} \in \mathbb{R}^4.

Follow any vector x∈R4x \in \mathbb{R}^4 as AA acts on it first and C\undo{C} acts on it second to produce CAx=x\undo{C}Ax = x: could C\undo{C} bring every vector back to xx if AA had squished R4\mathbb{R}^4?

No, you do not need to check AC=I4A\undo{C} = I_4 (thinking you still need to multiply ACA\undo{C} is only true for non-square matrices; for a square matrix, a one-sided inverse automatically works on both sides). Because AA is square, CA=I4\undo{C}A = I_4 means applying AA followed by C\undo{C} returns every vector xx in R4\mathbb{R}^4 back to xx, which would be impossible if AA squished R4\mathbb{R}^4 into a lower dimension. Thus AA does not squish space, so AA is invertible, C=A−1\undo{C} = \undo{A^{-1}} works on both sides so AC=I4A\undo{C} = I_4, and Ax=bAx = \target{b} has a unique solution for every b∈R4\target{b} \in \mathbb{R}^4.
6

Consider the tall 3×23 \times 2 matrix T=[100100]T = \begin{bmatrix} \colone{1} & \coltwo{0} \\ \colone{0} & \coltwo{1} \\ \colone{0} & \coltwo{0} \end{bmatrix} and the wide 2×32 \times 3 matrix W=[10−2014]W = \begin{bmatrix} \colone{1} & \coltwo{0} & \colthree{-2} \\ \colone{0} & \coltwo{1} & \colthree{4} \end{bmatrix}. (a) A student checks that Tx=0Tx = 0 has only the trivial solution x=0x = 0 and concludes by the Invertible Matrix Theorem that the columns of TT must span R3\mathbb{R}^3. Is the student right? (b) The student sees that the columns of WW span R2\mathbb{R}^2 and concludes that WW must be one-to-one. Is the student right?

Check the dimensions of TT and WW before linking "one-to-one" with "onto": can a 2D plane fill 3D space, or can 3D space fit into a 2D plane without squishing a dimension?

Neither conclusion is right; the student's mistake is applying the Invertible Matrix Theorem to non-square matrices where input and output dimensions do not match. (a) TT is 3×23 \times 2: its two columns are linearly independent and TT is one-to-one (it does not squish the 2D input plane), but a 2D plane inside 3D space (z=0z = 0) cannot fill R3\mathbb{R}^3, so its columns do not span R3\mathbb{R}^3 and TT is not onto. (b) WW is 2×32 \times 3: its first two columns ⟨1,0⟩\colone{\langle 1, 0 \rangle} and ⟨0,1⟩\coltwo{\langle 0, 1 \rangle} already span R2\mathbb{R}^2 so WW is onto, but mapping 3D space into 2D space forces at least one dimension to be squished (the third column ⟨−2,4⟩\colthree{\langle -2, 4 \rangle} is a combination of the first two in the same 2D plane), so WW has dependent columns and is not one-to-one.