Invertible matrix theorem — 12+ equivalent conditions: Visual walkthrough

Step 1 — What a matrix does to an arrow

WHAT. A square matrix $A$ is a machine. You feed it an arrow $x$ (a point in the plane, written $x = (x_{1}, x_{2})$ ), and it hands back a new arrow, which we write $A x$ . "Square" means it takes plane-arrows to plane-arrows (same number of inputs as outputs) — that word will matter later.

WHY. Everything in the theorem is secretly a question about this machine: can you always undo it? does it ever crush two arrows into one? can it reach every arrow? So before any algebra, we must picture the machine acting.

PICTURE. Below, the faint grid is the input plane. The bright grid is where $A$ sends that plane. The amber arrow is one input $x$ ; the cyan arrow is its output $A x$ .

Figure — Invertible matrix theorem — 12+ equivalent conditions

Step 2 — $A x$ is just a recipe mixing the columns

WHAT. Split $A$ into its columns — the arrows you get by feeding it the pure directions $(1, 0)$ and $(0, 1)$ . Call them $a_{1}$ and $a_{2}$ . Then the machine obeys one rule:

$A x = how much of a_{1} x_{1} first column a_{1} + how much of a_{2} x_{2} second column a_{2} .$

WHY. This is the single most important re-reading in all of linear algebra. It says the output is nothing but a weighted blend of the columns. So questions about $A$ become questions about its columns $a_{1}, a_{2}$ : do they point independent ways? do their blends reach everywhere?

PICTURE. The output arrow (cyan) is built by walking $x_{1}$ steps along $a_{1}$ , then $x_{2}$ steps along $a_{2}$ — a little parallelogram of instructions.

Link to the span picture: the set of all possible $A x$ is exactly the span of the columns.

Step 3 — The good machine vs. the collapsing machine

WHAT. Two cases, and only two.

Good (invertible): the two columns $a_{1}, a_{2}$ point in genuinely different directions. The bright grid is a full, un-squashed plane.
Collapsing (singular): $a_{2}$ lies along the same line as $a_{1}$ (say $a_{2} = 3 a_{1}$ ). Every blend $x_{1} a_{1} + x_{2} a_{2}$ still lands on that one line. The whole plane gets flattened onto a line.

WHY. This flattening is the root cause of every failing condition in the theorem. Once you see the plane crush to a line, all 17 conditions fail at once — that's what the walkthrough will unpack.

PICTURE. Left: independent columns, area preserved. Right: parallel columns, the shaded square collapses to a segment (zero area).

Step 4 — The null space: which arrows get crushed to zero?

WHAT. Ask the machine: which inputs $x$ come out as the zero arrow $0$ ? That collection is the null space, $Nul A = {x : A x = 0}$ .

Good machine: only $x = 0$ maps to $0$ . The null space is a single point.
Collapsing machine: a whole line of inputs maps to $0$ (all the arrows the flattening squashes onto the origin).

WHY. "Can we undo the machine?" is the same as "does the machine ever crush two different inputs together?" If two inputs $u \neq = v$ give the same output, then $A u = A v$ , so $A (u - v) = 0$ with $u - v \neq = 0$ — a nonzero crushed arrow. So a fat null space is exactly why you cannot undo the machine. This is condition (4) meeting condition (6), one-to-one.

PICTURE. For the collapsing machine, the amber line is the null space: every arrow on it lands on the origin dot.

See injective: one-to-one $⟺$ null space is just ${0}$ .

Step 5 — Independence and "crushing" are the very same fact

WHAT. Chain Steps 2–4 together. $A x = 0$ means $x_{1} a_{1} + x_{2} a_{2} = 0$ — a blend of columns equalling zero. Compare with the definition in Step 3: independent columns means the only such blend is the all-off one.

$null space is a point (Step 4) A x = 0 only at x = 0 ⟺ columns point distinct ways (Step 3) a_{1}, \dots, a_{n} independent$

WHY. We didn't prove two facts — we noticed they are one sentence written two ways. This is condition (4) $⟺$ (5). No formula to memorise; the equality of the two boxes is forced by Step 2's column-blend reading.

PICTURE. Two views of the same scene: on the left, the null space as a point/line; on the right, the columns as independent/parallel arrows. An amber bracket ties them.

Step 6 — Row reduction: turning "independent" into " $A \sim I$ "

WHAT. Row reduction hunts for pivots — the leading nonzero entry in each row after we tidy the matrix. Independent columns of a square matrix mean no column is free, so every column earns a pivot. For an $n \times n$ matrix, $n$ pivots fill every row and every column, and the tidied form (RREF) can only be the identity $I_{n}$ .

$each E_{i} = one reversible row move E_{k} \dots E_{2} E_{1} A = I .$

WHY. Each row move is reversible (you can always swap back, unscale, or un-add), so each $E_{i}$ is itself invertible. Setting $B = E_{k} \dots E_{1}$ gives $B A = I$ — an honest inverse built by hand. That's the arrow $2 \Rightarrow 1$ : row-equivalent-to- $I$ manufactures the inverse.

PICTURE. A staircase of pivots descending the diagonal; the bright cells are pivots, one per row and column.

Step 7 — The determinant: the area the machine keeps

WHAT. The determinant $det A$ is the signed area (in 2-D) or volume (in higher-D) of the shape you get by feeding the machine the unit square. The good machine stretches/rotates that square into a parallelogram of some nonzero area. The collapsing machine flattens it to a line — area $= 0$ .

$det A \neq = 0 ⟺ unit square keeps positive area ⟺ no collapse .$

WHY. This is why $det$ is the cheapest test for small matrices: it measures the flattening directly with one number. Row moves only scale area by nonzero factors and $det I = 1$ , so $A \sim I ⟺ det A \neq = 0$ . Note the theorem cares only whether the area is zero, never how small (a $det$ of $1 0^{- 9}$ is still fully invertible).

PICTURE. Left parallelogram: shaded, nonzero area, $det \neq = 0$ . Right: the parallelogram collapsed to an amber segment, area $0$ .

Step 8 — The eigenvalue $0$ : an arrow the machine leaves on its own line but shrinks to nothing

WHAT. An eigenvalue $λ$ (see Eigenvalues and Eigenvectors) is a number for which some nonzero arrow $v$ obeys $A v = λ v$ — the machine keeps $v$ on its own line and just scales it by $λ$ . If $λ = 0$ , then $A v = 0$ for a nonzero $v$ : a crushed arrow, exactly the fat null space of Step 4.

$det A = λ_{1} λ_{2} \dots λ_{n} .$

WHY. A single zero factor kills the whole product, so $0$ being an eigenvalue $⟺ det A = 0 ⟺$ collapse. This is condition (14) sliding neatly into the same picture: "0 is not an eigenvalue" is one more mask for "no crush."

PICTURE. The collapsing machine with its zero-eigenvector $v$ (amber) shrinking straight to the origin.

Step 9 — The degenerate warning: non-square machines break the deal

WHAT. Everything above quietly used square. A non-square machine — say $3$ inputs, $2$ outputs, or $2$ inputs, $4$ outputs — can be one-to-one or onto but never both, so the theorem simply does not apply.

Tall ( $4 \times 2$ ): only $2$ columns living in $4$ -D space; their blends fill at most a $2$ -D plane, so they cannot reach every output. Not onto, hence not invertible even if columns are independent.
Wide ( $2 \times 4$ ): $4$ columns in $2$ -D space must be dependent (too many arrows in too little room), so there is always a nonzero crushed input. Not one-to-one.

WHY. The IMT's magic — injective and surjective coinciding — comes from " $n$ pivots fill both all rows and all columns," which needs equal counts. Break the square, break the equivalence.

PICTURE. Left: tall map, columns fill only a plane inside 3-D space. Right: wide map, four arrows in the plane forced to overlap.

The one-picture summary

Every mask is the same coin. The good machine keeps area, keeps information, reaches everywhere; the collapsing machine flattens, crushes a line to zero, and misses most of space. One flattening event trips all the failing conditions simultaneously.

Recall Feynman: the whole walkthrough in plain words

Think of $A$ as a machine that reshuffles arrows. Its columns are where it sends the pure directions, and every output is just a blend of those columns with your dials $x_{1}, x_{2}, \dots$ . A good machine has columns pointing genuinely different ways: the only blend that gives the zero arrow is turning every dial off. That means nothing gets crushed, you can always rebuild the input, the unit square keeps real area (so the determinant isn't zero), no arrow gets scaled to nothing (so $0$ is not an eigenvalue), and the blends can reach every arrow in space. A bad machine has a repeated direction: the plane flattens onto a line, a whole line of inputs gets crushed to the origin, the area drops to zero, $0$ becomes an eigenvalue, and huge regions of space become unreachable — every failing test fires together. And all of this only works when the machine has as many inputs as outputs: a non-square machine can be reversible on one side but never both, so the theorem quietly steps aside.

Recall Quick self-check

Why does a nonzero null-space vector immediately kill invertibility? ::: It means two different inputs share an output ( $A (u - v) = 0$ ), so information is lost and the machine can't be undone. Why must RREF equal $I$ for an invertible square matrix? ::: Independent columns give a pivot in every column; $n$ pivots in an $n \times n$ grid fill every row and column, leaving only the identity. What does $det A = 0$ mean geometrically? ::: The machine flattens the unit square/cube to zero area/volume — space collapses to a lower dimension. Why doesn't the IMT apply to a $4 \times 2$ matrix? ::: It isn't square; its columns can be independent but can never span $R^{4}$ , so injective and surjective no longer coincide.