The Transpose

Imagine a spreadsheet with one row per student and one column per subject — each cell holds a test score. Now imagine you need the opposite layout: one row per subject, one column per student, so a teacher marking Maths can scan straight down a single row. You don't have to retype a single number. You just flip the table over its diagonal, and rows become columns automatically. That flip is called the transpose.

We mark the transpose of a matrix A with a small \mathsf{T}:

A = \begin{bmatrix} 2 & -1 & 0 \\ 3 & 5 & 4 \end{bmatrix} \;\longrightarrow\; A^{\mathsf{T}} = \begin{bmatrix} 2 & 3 \\ -1 & 5 \\ 0 & 4 \end{bmatrix}.

Entry by entry, the rule is simply (A^{\mathsf{T}})_{ij} = A_{ji} — swap the row and column index. Whatever sat in row i, column j of A now sits in row j, column i of A^{\mathsf{T}}. A 2\times 3 matrix transposes into a 3\times 2 one; the shape turns on its side. It looks like a purely cosmetic rearrangement — but it's also the secret ingredient that lets us write the dot product of two vectors as ordinary matrix multiplication, as you'll see below.

Watch the flip

Step through the entries. Each highlighted number in A lands in the mirror-image position of A^{\mathsf{T}} — row and column swapped. Notice that the two diagonal entries, 2 and 5, never leave their spot; only the off-diagonal entries trade places across the fold line.

Worked example: transposing by hand

Let's build a transpose from scratch, one entry at a time, with a fresh matrix:

C = \begin{bmatrix} 7 & 1 & -2 \\ 0 & 4 & 9 \end{bmatrix}.

C is 2\times 3 — two rows, three columns. To build C^{\mathsf{T}}, take row 1 of C, (7,\,1,\,-2), and lay it down as column 1 of C^{\mathsf{T}}. Then take row 2, (0,\,4,\,9), and lay it down as column 2:

C^{\mathsf{T}} = \begin{bmatrix} 7 & 0 \\ 1 & 4 \\ -2 & 9 \end{bmatrix}.

Check the shape: C was 2\times 3 (2 rows, 3 columns) and C^{\mathsf{T}} is 3\times 2 (3 rows, 2 columns) — the two dimensions have swapped places. That's not a coincidence: an m\times n matrix always transposes into an n\times m one.

Worked example: pivoting a real table

Suppose a weather app records the temperature (in °C) at three cities over five days — rows are days, columns are cities:

T = \begin{array}{c|ccc} & \text{London} & \text{Paris} & \text{Berlin} \\ \hline \text{Mon} & 14 & 17 & 12 \\ \text{Tue} & 15 & 18 & 13 \\ \text{Wed} & 13 & 16 & 11 \\ \text{Thu} & 16 & 19 & 14 \\ \text{Fri} & 15 & 17 & 12 \end{array}

Reading straight across a row of T answers "what was the weather like on Tuesday, everywhere?" Now transpose it. Cities become the rows, days become the columns:

T^{\mathsf{T}} = \begin{array}{c|ccccc} & \text{Mon} & \text{Tue} & \text{Wed} & \text{Thu} & \text{Fri} \\ \hline \text{London} & 14 & 15 & 13 & 16 & 15 \\ \text{Paris} & 17 & 18 & 16 & 19 & 17 \\ \text{Berlin} & 12 & 13 & 11 & 14 & 12 \end{array}

Now reading straight across a row answers a completely different question: "how did Paris change over the week?" No number moved to the wrong city or the wrong day — the transpose just changed which direction you read in. This exact operation has a name in every spreadsheet program: it's called a pivot.

The double flip returns you home

What happens if you transpose a matrix, then transpose the result again? Try it on a small example:

D = \begin{bmatrix} 3 & -5 \\ 8 & 2 \end{bmatrix} \;\longrightarrow\; D^{\mathsf{T}} = \begin{bmatrix} 3 & 8 \\ -5 & 2 \end{bmatrix} \;\longrightarrow\; (D^{\mathsf{T}})^{\mathsf{T}} = \begin{bmatrix} 3 & -5 \\ 8 & 2 \end{bmatrix} = D.

Flipping a page over swaps its front and back — flip it a second time and you're back to reading the front, exactly as it started. The transpose behaves the same way: it's its own perfect undo.

For any matrix A:

(A^{\mathsf{T}})^{\mathsf{T}} = A — transposing twice recovers the original matrix exactly.
If A is m\times n, then A^{\mathsf{T}} is n\times m, and (A^{\mathsf{T}})^{\mathsf{T}} is back to m\times n.

Worked example: the dot product in disguise

Here's the payoff promised earlier. Take two column vectors:

\vec{u} = \begin{bmatrix} 1 \\ 2 \\ 3 \end{bmatrix}, \qquad \vec{v} = \begin{bmatrix} 4 \\ -1 \\ 2 \end{bmatrix}.

Their ordinary dot product multiplies matching entries and adds: \vec{u}\cdot\vec{v} = 1(4) + 2(-1) + 3(2) = 4 - 2 + 6 = 8. Now transpose \vec{u} — a 3\times 1 column becomes a 1\times 3 row, \vec{u}^{\mathsf{T}} = \begin{bmatrix} 1 & 2 & 3 \end{bmatrix} — and multiply it by \vec{v} as a matrix product:

\vec{u}^{\mathsf{T}}\vec{v} = \begin{bmatrix} 1 & 2 & 3 \end{bmatrix}\begin{bmatrix} 4 \\ -1 \\ 2 \end{bmatrix} = 1(4) + 2(-1) + 3(2) = 8.

Exactly the same number. A 1\times 3 row times a 3\times 1 column is the only shape combination that multiplies down to a bare 1\times 1 number — which is exactly what a dot product is supposed to produce. The transpose is what makes that shape line up.

When the flip changes nothing: symmetric matrices

Most of the time A^{\mathsf{T}} looks nothing like A — the shape itself might even change. But every so often, flipping a matrix across its diagonal lands every entry back exactly where it started. Such a matrix is called symmetric: A = A^{\mathsf{T}}.

Here's a symmetric matrix with a very natural meaning — road distances (in km) between three cities:

D = \begin{array}{c|ccc} & A & B & C \\ \hline A & 0 & 120 & 340 \\ B & 120 & 0 & 260 \\ C & 340 & 260 & 0 \end{array}

Entry D_{AB} = 120 is the distance from city A to city B. Entry D_{BA} = 120 too — the distance from B back to A is exactly the same road, walked the other way. Because D_{ij} = D_{ji} for every pair, D = D^{\mathsf{T}}: transposing this matrix does absolutely nothing to it. The zeros running down the diagonal make sense too — the distance from any city to itself is zero, and diagonal entries never move under a transpose anyway.

Symmetric matrices aren't a curiosity — they're everywhere a quantity is measured between pairs of things rather than belonging to one thing alone, which is also exactly why the dot product hides a transpose inside it: writing two column vectors \vec{u} and \vec{v}, their dot product is \vec{u}\cdot\vec{v} = \vec{u}^{\mathsf{T}}\vec{v} — a 1\times n row multiplied by an n\times 1 column, which is precisely the shape matrix multiplication needs to produce a single number.

Two mistakes catch almost everyone the first time they meet the transpose:

The dimensions don't survive unless the matrix is square. An m\times n matrix becomes n\times m — not another m\times n matrix. Only when m = n (a square matrix) does the shape stay the same after transposing — the entries can still move even then, just within the same square outline.
The diagonal entries never move. The rule (A^{\mathsf{T}})_{ij}=A_{ji} only relocates an entry when i \ne j. When i = j (a diagonal entry) it just swaps with itself — that is, it stays exactly where it is. Only the off-diagonal pairs actually trade places.

Once you know to look, symmetric matrices show up constantly. A correlation matrix in statistics records how strongly every pair of variables moves together — the correlation between height and weight is the same number as between weight and height, so the matrix is automatically symmetric. An adjacency matrix of an ordinary (non-directional) friendship network works the same way: if the matrix entry says "Amir is friends with Beatriz," the mirrored entry says "Beatriz is friends with Amir" — and since friendship goes both ways, the two entries always agree.

And that "pivot" move from the temperature table above isn't just a maths-class trick — it's the literal name of the button in Excel and Google Sheets that turns a "rows of records" table into a "summary by category" table. Every time you click Pivot Table, you're asking the spreadsheet to transpose (and often also summarise) your data for you.