Rank, inverse, and elementary operations

A matrix hides a single honest number that tells you how much real information it carries: its rank. Stack three "independent" equations on top of each other and you might discover, after a little arithmetic, that one of them was secretly a copy of the other two — the rank drops, and so does the number of genuinely independent constraints. This article builds the toolkit that finds that number and puts it to work: elementary row operations — swap, scale, add-a-multiple — which never change a matrix's rank and which, run systematically on the augmented pair $[A\,|\,I]$ , also hand you the inverse $A^{-1}$ for free. The same three moves you use to compute rank are exactly the moves that turn $A$ into $I$ and $I$ into $A^{-1}$ , and both computations are the direct setup for deciding, in the next article, whether a linear system has no solution, one solution, or infinitely many.

The core insight

Elementary row operations reduce a matrix to row echelon form, whose count of nonzero rows is the rank, and applying the same operations to the augmented matrix $[A\,|\,I]$ turns it into $[I\,|\,A^{-1}]$ whenever $A$ is invertible.

The core idea

Elementary row operations on a matrix are exactly three moves, each reversible: (i) swap two rows, (ii) multiply a row by a nonzero scalar, (iii) add a scalar multiple of one row to another. None of the three changes the rank of the matrix, and none changes the solution set of a linear system the matrix represents — this is the single fact that makes row reduction trustworthy as a computational tool rather than just a bookkeeping trick. The rank of a matrix $A$ is the number of nonzero rows once $A$ is reduced to row echelon form (REF) — a form in which each row's leading (leftmost nonzero) entry sits strictly to the right of the leading entry in the row above, with any all-zero rows pushed to the bottom. Rank equals the maximum number of linearly independent rows, and — a fact worth trusting without reproving it every time — this always equals the maximum number of linearly independent columns too:

\operatorname{rank}(A) = \operatorname{rank}(A^T).

where: $A^T$ is the transpose of $A$ . An equivalent definition treats rank as the order of the largest nonzero minor (the determinant of a square submatrix) of $A$ ; for an $m\times n$ matrix this forces $\operatorname{rank}(A)\le\min(m,n)$ , and a square $n\times n$ matrix is full rank — and therefore invertible — exactly when $\det A\neq0$ .

Inverse by Gauss–Jordan reduction. Augment $A$ with the identity to build $[A\,|\,I]$ , then apply row operations to the whole augmented row (both sides at once) until the left block becomes $I$ :

[A\,|\,I]\ \xrightarrow{\text{row ops}}\ [I\,|\,A^{-1}].

where: the operations applied to the left block to turn $A$ into $I$ are, when applied to the right block starting from $I$ , precisely $A^{-1}$ . If at any stage a row of the left block collapses to all zeros, no amount of further reduction can produce $I$ there — $A$ is singular and has no inverse. This method scales far better than the cofactor/adjoint formula once $A$ is bigger than $2\times2$ or $3\times3$ . One more product rule is worth keeping in view because it is a fast sanity check on arithmetic: for two matrices whose product is defined, $\operatorname{rank}(AB)\le\min(\operatorname{rank}(A),\operatorname{rank}(B))$ — multiplying can only hold rank steady or shrink it, never grow it.

Figure

Row reduction of an augmented matrix from A beside the identity to the identity beside the inverse

Derivation AWhy row-reducing the augmented matrix produces the inverse

The goal is to show, without invoking any inverse formula as a starting assumption, that if a sequence of elementary row operations turns $A$ into $I$ , then applying that same sequence to $I$ must produce exactly $A^{-1}$ — which is why building $[A\,|\,I]$ and reducing the left block to $I$ is a legitimate, general method for finding a matrix inverse rather than a coincidence that happens to work on textbook examples.

Step 1. Represent each elementary row operation as left-multiplication by an elementary matrix. Swapping two rows of $A$ , scaling a row, or adding a multiple of one row to another can each be produced by multiplying $A$ on the left by a suitable matrix $E$ , so that the new matrix is $EA$ . A sequence of $k$ operations is then a chain of left-multiplications.

Step 2. Suppose the full sequence $E_1, E_2, \dots, E_k$ (applied in that order) reduces $A$ all the way to $I$ . Written as one product, this says $E_k\cdots E_2E_1 A = I$ . Call the combined product $P = E_k\cdots E_2E_1$ , so the statement is simply $PA = I$ .

Step 3. By the definition of a matrix inverse, an equation of the form $PA = I$ with $A$ square means $P$ is $A^{-1}$ — there is only one matrix that does this job, because if $Q$ also satisfied $QA = I$ then $Q = QI = Q(AP) = (QA)P = P$ using associativity. So $P = A^{-1}$ exactly.

Step 4. Now apply that same product $P$ to $I$ instead of $A$ : $PI = P = A^{-1}$ . But $P$ was built as "the same row operations, in the same order," and applying a chain of row operations to $I$ one at a time is exactly what the right-hand block of the augmented matrix experiences as the left block goes from $A$ to $I$ . So the right block, which starts as $I$ and receives every operation the left block receives, ends up equal to $P I = A^{-1}$ . This is the whole justification for reading $A^{-1}$ straight off the right-hand block once the left-hand block reaches $I$ . This argument also explains the failure mode: if no sequence of row operations can reduce $A$ to $I$ (a zero row appears first), then no $P$ with $PA=I$ exists, which is exactly the statement that $A$ is singular.

Worked example

Find $A^{-1}$ for $A=\begin{pmatrix}1&1&1\\0&2&3\\5&5&1\end{pmatrix}$ using Gauss–Jordan reduction on $[A\,|\,I]$ , and confirm $A$ is full rank on the way.

Step 1. Since $A$ is $3\times3$ , full rank means $\operatorname{rank}(A)=3$ , which will show up automatically as the reduction succeeding all the way to $I$ with no zero row along the way — no separate rank computation is needed once the inverse is found. Start from

\left[\begin{array}{ccc|ccc}1&1&1&1&0&0\\0&2&3&0&1&0\\5&5&1&0&0&1\end{array}\right].

Step 2. Clear column 1 below the pivot: $R_3\to R_3-5R_1$ turns row 3 into $(0,0,-4\,|\,-5,0,1)$ :

\left[\begin{array}{ccc|ccc}1&1&1&1&0&0\\0&2&3&0&1&0\\0&0&-4&-5&0&1\end{array}\right].

Step 3. Normalize the remaining pivots: $R_2\to\tfrac12R_2$ and $R_3\to-\tfrac14R_3$ give

\left[\begin{array}{ccc|ccc}1&1&1&1&0&0\\0&1&1.5&0&0.5&0\\0&0&1&1.25&0&-0.25\end{array}\right].

Every pivot is nonzero, so the reduction is not stuck — this already certifies $\operatorname{rank}(A)=3$ .

Step 4. Clear upward from the last pivot: $R_2\to R_2-1.5R_3$ gives row 2 $=(0,1,0\,|\,-1.875,0.5,0.375)$ ; then $R_1\to R_1-R_3$ gives row 1 $=(1,1,0\,|\,-0.25,0,0.25)$ ; then $R_1\to R_1-R_2$ gives row 1 $=(1,0,0\,|\,1.625,-0.5,-0.125)$ . The left block is now $I$ , so the right block is the inverse:

A^{-1}=\begin{pmatrix}1.625&-0.5&-0.125\\-1.875&0.5&0.375\\1.25&0&-0.25\end{pmatrix}.

Step 5. Spot check: row 1 of $A$ , $(1,1,1)$ , dotted with column 1 of $A^{-1}$ , $(1.625,-1.875,1.25)^T$ , gives $1.625-1.875+1.25=1$ — matching the $(1,1)$ entry of $I$ , since $AA^{-1}=I$ . Answer: $A^{-1}=\begin{pmatrix}1.625&-0.5&-0.125\\-1.875&0.5&0.375\\1.25&0&-0.25\end{pmatrix}$ , and $\operatorname{rank}(A)=3$ .

Competitive stretch

If instead $A=\begin{pmatrix}1&2&3\\2&4&7\\3&6&10\end{pmatrix}$ , the same elimination pattern ( $R_2\to R_2-2R_1$ , $R_3\to R_3-3R_1$ , then $R_3\to R_3-R_2$ ) reduces the left block to $\begin{pmatrix}1&2&3\\0&0&1\\0&0&0\end{pmatrix}$ — a zero row appears, so the reduction can never reach $I$ . Reading rank off REF gives $\operatorname{rank}(A)=2$ , and the augmented-matrix method correctly reports "no inverse exists" rather than producing a wrong answer, consistent with $\det A=0$ here ( $\text{col}_2=2\,\text{col}_1$ ).

Common traps

Confusing row echelon form (enough to read off rank) with the fully reduced form needed to read off an inverse — rank only needs the pivot pattern, but Gauss–Jordan for $A^{-1}$ needs the left block reduced all the way to $I$ .
Forgetting to apply a row operation to the augmented ( $I$ ) side as well as the $A$ side — the two blocks must move together at every step.
Assuming a square matrix is automatically full rank; a nonzero determinant (or a clean reduction to $I$ ) must actually be checked, since proportional rows or columns silently lower the rank.
Mixing up $\operatorname{rank}(A)$ with the number of rows $m$ or columns $n$ ; rank is always $\le\min(m,n)$ , never larger, and a zero row appearing mid-reduction is the signal, not a mistake to erase.

Ask the tutor: want to work through reducing a specific matrix to echelon form or a Gauss–Jordan inverse step by step? Ask the AI tutor to walk it with you row by row.

Why does rank stop increasing once I have found the number of nonzero rows, even if I keep row-reducing further

Because elementary row operations never change rank — they only reveal it. Once the matrix reaches row echelon form, the count of nonzero rows already equals the rank; further reduction toward reduced row echelon form (needed for reading off an inverse or a solution set directly) rearranges entries within those rows but cannot create or destroy a nonzero row, so the count stays fixed.

Is Gauss-Jordan on the augmented matrix always faster than the adjoint or cofactor formula for A inverse

For $2\times2$ matrices the adjoint formula is usually quicker by hand. From $3\times3$ upward, Gauss–Jordan wins clearly: the cofactor method needs a full set of minors and determinants whose count grows fast with size, while row reduction is a fixed, mechanical sequence of arithmetic that also tells you immediately, via a stuck zero row, when no inverse exists.

If det(A) is nonzero, do I still need to check rank separately before computing the inverse

No — for a square matrix the two checks are the same fact stated two ways. $\det A\neq0$ is exactly equivalent to $\operatorname{rank}(A)=n$ for an $n\times n$ matrix, and both are exactly equivalent to the Gauss–Jordan reduction succeeding without a zero row. In practice, just start the row reduction: if it completes, $A$ was full rank and invertible all along.

Rank, inverse, and elementary operations

The core idea

Worked example

Competitive stretch

Common traps

Before this

Where this leads