Rank of a matrix — definition, row rank = column rank theorem: Visual walkthrough

Step 0 — The words we will use (built before we use them)

Before any symbol appears in a formula, let us anchor it to a picture.

A vector is just an arrow from the origin, or equally a list of numbers like $(1, 2, 3)$ . The numbers tell you how far to walk along each axis.
A linear combination of some vectors means: scale each one by a number, then add them up. Picture stretching several arrows and laying them tip-to-tail.
The span of a set of vectors is every point you can reach by linear combinations of them. One nonzero arrow spans a line; two arrows pointing different ways span a plane.
Vectors are independent if none of them is reachable from the others — each adds a genuinely new direction. If one arrow lies inside the span of the rest, they are dependent.
A basis of a space is a smallest set of independent arrows whose span is the whole space. The number of arrows in a basis is the dimension.

These live in different worlds ( $R^{n}$ vs $R^{m}$ ), possibly of different sizes. Our goal: show these two dimensions are the same integer. That is genuinely surprising, which is why it deserves a picture-proof.

Figure — Rank of a matrix — definition, row rank = column rank theorem

More on these spaces: Row Space and Column Space. On independence and bases: Linear Independence and Basis.

Step 1 — Look at a concrete matrix and count both ways by hand

WHAT. Before any theorem, let us physically count independent rows and independent columns for one small matrix, so the claim isn't abstract.

$A = 123246134$

WHY. A believed example is worth ten proofs of faith. We want to witness the equality once, then explain it.

PICTURE. In the figure below, the three columns are drawn as arrows. Columns $1$ and $3$ point in genuinely different directions; column $2 = 2 \times$ column $1$ (it lies on top of column 1's line, just longer). So only 2 independent column-directions.

$(2 1) 3$ — call this $a_{1}$ , the first direction.
$(4 2) 6 = 2 a_{1}$ — no new direction (dependent).
$(3 1) 4$ — call this $a_{3}$ , points off the $a_{1}$ line.

Column rank $= 2$ . We will confirm row rank $= 2$ in Step 2.

Step 2 — Elimination reveals the count without guessing

WHAT. Reduce $A$ using row operations — replacing a row by (itself minus a multiple of another row) — until each nonzero row starts further right than the one above (echelon form). Then count the nonzero rows.

123246134 R_{2} - 2 R_{1} R_{3} - 3 R_{1} 100200111 R_{3} - R_{2} \underline{1} 00 200 1 \underline{1} 0

Term by term in the arrows:

$R_{2} - 2 R_{1}$ means new row 2 $=$ old row 2 $- 2 \times$ row 1. The $2 \times$ is chosen exactly to zero out the leading $2$ under the leading $1$ .
$R_{3} - 3 R_{1}$ zeros the $3$ under the same leading $1$ .
$R_{3} - R_{2}$ removes the duplicate $(0, 0, 1)$ so a hidden dependence becomes an honest zero row.

The underlined entries $\underline{1}$ are the pivots — the first nonzero entry of each surviving row.

WHY row operations are legal here. Each new row is a linear combination of old rows, and old rows are recoverable from new ones. So the span of the rows never changes — the row space, and therefore its dimension (row rank), is untouched. (Full story: Gaussian Elimination & Echelon Form.)

PICTURE. The figure shows the row space as a shaded plane. Every operation slides a row within that plane; the plane itself does not tilt. When the third row collapses to $0$ , we learn the plane was only 2-dimensional all along: row rank $= 2$ . It matches the column count from Step 1 — but why it must match is Steps 3–7.

Recall Check your grip

Why is counting nonzero rows in the original matrix wrong? ::: They can be nonzero yet dependent (like $(1, 2)$ and $(2, 4)$ ); only echelon form exposes true independence. What quantity did the row operations preserve? ::: The row space, hence the row rank.

Step 3 — Give the column rank a name and grab a column basis

WHAT. From now on, general matrix. Let

$r = column rank of A .$

Here $r$ is just "how many independent column-directions exist". Pick $r$ actual columns that are independent and span all the others — call them $c_{1}, \dots, c_{r}$ . This is a basis of the column space.

WHY. A basis is the minimal instruction set: the parent note's LEGO idea. Every column of $A$ can be built from these $r$ arrows and nothing fewer.

PICTURE. The column space is drawn as an $r$ -dimensional flat (for $r = 2$ , a plane inside $R^{m}$ ). The basis arrows $c_{1}, \dots, c_{r}$ are highlighted; every other column is a dot lying inside that flat, reachable as a mix of the highlighted arrows.

Stack these basis columns side by side into one skinny matrix:

$C = [c_{1} c_{2} \dots c_{r}], C is m \times r .$

Each $c_{k}$ has $m$ entries (it lives in $R^{m}$ ) → $m$ rows.
There are $r$ of them → $r$ columns. That is why $C$ is $m \times r$ .

Step 4 — Every column of $A$ is $C$ times a short recipe

WHAT. Because $c_{1}, \dots, c_{r}$ span the column space, column $j$ of $A$ (call it $A_{j}$ ) is some combination of them:

$A_{j} = x_{1 j} c_{1} + x_{2 j} c_{2} + \dots + x_{r j} c_{r} .$

Reading each symbol:

$A_{j}$ — the $j$ -th column of the original matrix (an arrow in $R^{m}$ ).
$x_{k j}$ — the amount of basis arrow $c_{k}$ needed to build $A_{j}$ . A single number.
The whole right side — "mix the $r$ ingredients in these proportions". This is exactly matrix multiplication read as combining columns.

WHY. We are turning " $A_{j}$ lives in the span" (a picture) into " $A_{j} = C \cdot (recipe)$ " (an equation), so algebra can carry it forward.

PICTURE. For one chosen column, the figure shows the target arrow $A_{j}$ reconstructed tip-to-tail from scaled copies of $c_{1}$ and $c_{2}$ . The little numbers $x_{1 j}, x_{2 j}$ label the scaling.

Collect the recipe of every column into one matrix $R$ , whose $j$ -th column is $(x_{1 j}, \dots, x_{r j})$ :

$A = C R, C : m \times r, R : r \times n .$

$R$ has $r$ rows (one per basis ingredient) and $n$ columns (one recipe per original column) → $r \times n$ .
The shared inner dimension $r$ is the column rank. Hold that thought — it is the whole trick.

Step 5 — The same factorization secretly controls the rows

WHAT. Multiplication $A = C R$ can be read by rows instead of by columns. Row $i$ of $A$ equals row $i$ of $C$ multiplied into $R$ :

$(row i of A) = (row i of C) R .$

Symbol by symbol:

$(row i of C)$ — a short list of just $r$ numbers.
Multiplying it by $R$ mixes the $r$ rows of $R$ using those $r$ numbers as weights.
So the result — row $i$ of $A$ — is a linear combination of the $r$ rows of $R$ .

WHY this finishes the hard direction. Every row of $A$ is built from the same $r$ rows of $R$ . So all the rows of $A$ live inside the span of only $r$ vectors. A span of $r$ vectors has dimension at most $r$ :

$row rank (A) \leq r = column rank (A) .$

PICTURE. Left panel: the columns of $A$ pour out of the $r$ columns of $C$ . Right panel: the rows of $A$ pour out of the $r$ rows of $R$ . Same tiny bottleneck $r$ feeds both worlds — that shared waist is the reason the two ranks are chained together.

Step 6 — Flip the matrix to get the other inequality

WHAT. We proved row rank $\leq$ column rank for every matrix. Now apply that same result to the transpose $A^{T}$ (rows become columns, columns become rows).

$row rank (A^{T}) \leq column rank (A^{T}) .$

Translate each side back:

Rows of $A^{T}$ are the columns of $A$ ⇒ $row rank (A^{T}) = column rank (A)$ .
Columns of $A^{T}$ are the rows of $A$ ⇒ $column rank (A^{T}) = row rank (A)$ .

Substituting:

$column rank (A) \leq row rank (A) .$

WHY. Step 5 gave " $\leq$ " one way; transposing gives " $\leq$ " the other way. Two opposite inequalities can only both hold if the two numbers are equal.

$row rank (A) = column rank (A) = rank (A)$

PICTURE. The figure shows $A$ and $A^{T}$ as mirror images across the diagonal, with the two inequality arrows meeting in the middle to pin the ranks together.

Step 7 — Degenerate and edge cases (the reader must never be surprised)

WHAT. Check every corner where the argument might wobble.

Case $r = 0$ (the zero matrix). No independent columns. Then the column space is just the single point ${0}$ , dimension $0$ . Row space is also ${0}$ . Both ranks $= 0$ ; the factorization $A = C R$ degenerates but the equality $0 = 0$ holds. ✔

Case full row rank ( $r = m$ ). All rows independent; $C$ is $m \times m$ and invertible, $R$ has $m$ rows. Column rank also $= m \leq n$ . Consistent with the bound below.

Case full column rank ( $r = n$ ). All columns independent; nothing collapses. Here $rank = n \leq m$ .

Case square and $r = n = m$ (full rank square). Then the map loses no directions — it is invertible.

The universal bound. In all cases,

$0 \leq rank (A) \leq min (m, n) .$

Why? The columns are $n$ arrows living in $R^{m}$ : you cannot have more independent arrows than you have arrows ( $\leq n$ ), nor more than the dimension of the room they sit in ( $\leq m$ ). So the rank is squeezed under $min (m, n)$ .

PICTURE. A number line from $0$ to $min (m, n)$ with a slider marking where a matrix's rank can sit, and the three named landmarks (zero matrix, full column rank, full row rank) pinned on it.

The one-picture summary

Everything above compresses to a single diagram: the matrix $A$ squeezed through a waist of width $r$ . Columns flow out the left face of the waist; rows flow out the right face; the waist width is the rank, shared by both.

Recall Feynman retelling of the whole walkthrough

Picture a matrix as a doorway. On one side you count how many truly different columns squeeze through — call that number $r$ . Now the trick: you can rebuild the entire matrix as " $r$ core columns, then a recipe sheet telling how to mix them" — that's $A = C R$ . But look at that recipe sheet from the side: it also has exactly $r$ rows, and every actual row of the matrix is just a blend of those $r$ recipe-rows. So the rows can't need more than $r$ independent directions either. Rows $\leq$ columns. Then flip the matrix on its diagonal and run the identical argument to get columns $\leq$ rows. The only way both "less-than-or-equal" statements can be true at once is if the two counts are the same number — the rank. Rows and columns shook hands at the waist of width $r$ , and that handshake is the theorem.

Connections

Row Space and Column Space — the two spaces we measured.
Gaussian Elimination & Echelon Form — why row ops preserve row rank (Step 2).
Linear Independence and Basis — meaning of the $r$ basis columns (Step 3).
Matrix Multiplication as Linear Combination — the reading behind $A = C R$ (Steps 4–5).
Rank–Nullity Theorem — where the bound $rank \leq n$ leads next.
Invertible Matrix Theorem — full-rank square case (Step 7).

Visual walkthrough — Rank of a matrix — definition, row rank = column rank theorem