Determinant Properties and Row Operations

Suppose you must decide whether a 7×77\times 7 matrix, packed with clumsy fractions, is invertible. Cofactor expansion demands over 7!=50407! = 5040 signed products. No one would attempt this by hand. Yet the determinant has a hidden computational secret. It interacts with the elementary row operations of Gaussian elimination in an exquisitely simple way. Scale a row, and the determinant scales by the same factor. Swap two rows, and the determinant flips sign. Add a multiple of one row to another, and the determinant does not change at all. These three rules turn the determinant from a forbidding multilinear sum into something simpler. You compute it by reducing a matrix to triangular form. Then you read off the product of the diagonal. This article develops that computational machinery rigorously. It builds on the Alternating Multilinear Forms characterization of the determinant established previously. We prove, from the alternating multilinear axioms, exactly how each elementary row operation transforms det⁡\det. We prove the deep but easily-stated fact that det⁡(AT)=det⁡(A)\det(A^T) = \det(A). This means every row statement automatically becomes a column statement for free. And we prove the triangular determinant formula. The determinant of an upper (or lower) triangular matrix is simply the product of its diagonal entries. Together these results give the standard algorithm. Row-reduce to triangular form, tracking sign flips and scalings. This is how determinants are computed in practice, by computer algebra systems and by hand alike. The big picture is that the determinant is not merely a formula to memorize. It is a functional invariant that behaves predictably under familiar operations. These are the same operations you already use to solve linear systems. This is why the determinant test for invertibility traces back to the same structural facts. The change-of-basis behavior of volume does too. So does the multiplicativity of determinants, to be developed shortly. All of these trace back to the same handful of structural facts proved here. Once you see that elementary operations act on det⁡\det by simple scalar rules, a mystery dissolves. "Why does this formula compute volume" becomes "because volume scales, flips, and shears exactly this way." By the end of this article you will be able to state and prove the row-operation rules. You will prove transpose-invariance and prove the triangular formula. You will use all three together to compute determinants of matrices too large for cofactor expansion to be practical. And you will understand precisely why the algorithm is correct, not just how to run it.

Key takeaways
  • Swapping two rows of a matrix multiplies its determinant by −1-1.
  • Scaling one row by cc multiplies the determinant by cc.
  • Adding a multiple of one row to another row leaves the determinant unchanged.
  • Transposition preserves the determinant: det⁡(AT)=det⁡(A)\det(A^T) = \det(A), so every row rule is also a column rule.
  • The determinant of a triangular matrix is the product of its diagonal entries — the payoff of row reduction.
  • Key idea: row-reduce to triangular form, track the sign/scale changes, then multiply the diagonal.

Where this lives today

Row-reduction-based determinant computation is the backbone of every practical numerical linear algebra library: LU decomposition, at its core, is exactly the record of the Replace and Swap operations used to triangularize a matrix, and the determinant is recovered as (−1)s(-1)^s times the product of the pivots, exactly as proved above. This is the method used by LAPACK and by virtually all computer algebra and numerical computing systems, since cofactor expansion's factorial blowup makes it computationally infeasible beyond tiny matrices. In Determinant and Invertibility, the same row-reduction viewpoint gives the cleanest possible criterion for invertibility: a square matrix is invertible exactly when its row-echelon form has no zero pivot, equivalently when its determinant is nonzero. The triangular determinant formula also underlies the change-of-variables formula in multivariable calculus (the Jacobian determinant), the volume-scaling interpretation of linear maps, and the partition-function normalization computations in statistical mechanics and Gaussian integrals in probability, all of which reduce a determinant computation to a triangular factorization. In cryptography and coding theory, determinant computations over finite fields for lattice-based and error-correcting-code constructions rely on the identical row-operation calculus, valid over any field as proved here.

Ask the tutor: How do these row-operation rules combine to prove that det⁡(AB)=det⁡(A)det⁡(B)\det(AB) = \det(A)\det(B)?

DefinitionElementary row operations and the determinant

Let FF be a field and let Mn(F)M_n(F) denote the n×nn\times n matrices over FF. Write a matrix AA by its rows as A=(r1⋮rn)A = \begin{pmatrix}\mathbf{r}_1 \\ \vdots \\ \mathbf{r}_n\end{pmatrix}, with ri∈Fn\mathbf{r}_i \in F^n. The determinant det⁡:Mn(F)→F\det : M_n(F) \to F is the unique function that is: 1. Multilinear in the rows: for each fixed index ii, holding all other rows fixed, the map ri↦det⁡(A)\mathbf{r}_i \mapsto \det(A) is a linear function of ri\mathbf{r}_i. 2. Alternating: if two rows of AA are equal, det⁡(A)=0\det(A) = 0. 3. Normalized: det⁡(In)=1\det(I_n) = 1, where InI_n is the identity matrix. The three elementary row operations on a matrix AA are:

  • (Swap) Ri↔RjR_i \leftrightarrow R_j (i≠ji \ne j): interchange rows ii and jj.
  • (Scale) Ri→cRiR_i \to cR_i for a nonzero scalar c∈Fc \in F: multiply row ii by cc.
  • (Replace) Ri→Ri+cRjR_i \to R_i + cR_j (i≠ji \ne j): add cc times row jj to row ii.

A matrix T=(tij)T = (t_{ij}) is upper triangular if tij=0t_{ij} = 0 whenever i>ji > j, and lower triangular if tij=0t_{ij} = 0 whenever i<ji < j; it is triangular if it is either.

Intuition

Think of the rows of a 2×22\times 2 matrix as two vectors in the plane, and the determinant as the signed area of the parallelogram they span. Scaling a row stretches the parallelogram in that direction, scaling area by the same factor. Swapping the two rows reflects the parallelogram, which reverses orientation — hence the sign flip, even though the unsigned area is unchanged. Adding a multiple of one row to another is a shear: it slides the parallelogram sideways without changing its base or its height, so the area is preserved exactly. This geometric picture, valid over R\mathbb{R}, is really shorthand for algebraic facts that hold over any field, proved below directly from multilinearity and the alternating property. The triangular case is the computational payoff. A triangular matrix's rows have a staircase pattern of zeros, and this structure forces all the "cross terms" in the determinant sum to vanish except one: the product straight down the diagonal. Gaussian elimination is precisely the process of using Replace and Swap operations to convert an arbitrary matrix into this staircase shape, so tracking the accumulated sign and scale changes during elimination — then multiplying the final diagonal — computes the determinant of the original matrix.

Figure
Shear operation preserves area: R2→R2+cR1R_2 \to R_2 + cR_1
r₁ = (3,0) r₂ = (1,2) r₁ = (3,0) r₂+cr₁ Same base, same height ⇒ equal area det unchanged under Replace
TheoremRow operation rules for the determinant

Let FF be a field, A∈Mn(F)A \in M_n(F), and let det⁡\det be the determinant function of Alternating Multilinear Forms normalized so det⁡(In)=1\det(I_n)=1. Let A′A' be obtained from AA by a single elementary row operation. 1. (Swap) If A′A' is obtained by interchanging rows i≠ji \ne j of AA, then det⁡(A′)=−det⁡(A)\det(A') = -\det(A). 2. (Scale) If A′A' is obtained by multiplying row ii of AA by a scalar c∈Fc \in F, then det⁡(A′)=cdet⁡(A)\det(A') = c\det(A). 3. (Replace) If A′A' is obtained by adding cc times row jj of AA to row ii (i≠ji \ne j), then det⁡(A′)=det⁡(A)\det(A') = \det(A). Consequently, if AA is obtained from A′A' by a finite sequence of elementary row operations of these three types with respective multipliers, the resulting determinant is det⁡(A)\det(A) times the product of the scale-factors used, times (−1)(number of swaps)(-1)^{(\text{number of swaps})}.

Proof

Write AA with rows r1,…,rn\mathbf{r}_1,\dots,\mathbf{r}_n, and recall det⁡\det is multilinear and alternating in the rows, with det⁡(In)=1\det(I_n) = 1.

(Scale). Let A′A' have rows r1,…,cri,…,rn\mathbf{r}_1,\dots,c\mathbf{r}_i,\dots,\mathbf{r}_n. By multilinearity in the ii-th row (linearity applied with all other rows fixed),

det⁡(A′)=det⁡(r1,…,cri,…,rn)=cdet⁡(r1,…,ri,…,rn)=cdet⁡(A).\det(A') = \det(\mathbf{r}_1,\dots,c\mathbf{r}_i,\dots,\mathbf{r}_n) = c\det(\mathbf{r}_1,\dots,\mathbf{r}_i,\dots,\mathbf{r}_n) = c\det(A).

This uses only the linearity of det⁡\det in one coordinate, holding the rest fixed — precisely axiom (1) of the determinant.

(Swap). First we establish a lemma: if a matrix BB has two rows equal, more precisely ri=rj\mathbf{r}_i = \mathbf{r}_j for some i≠ji \ne j, then det⁡(B)=0\det(B) = 0; this is exactly axiom (2) (the alternating property), so no further proof is needed for that base case. Now let AA have rows r1,…,ri,…,rj,…,rn\mathbf{r}_1,\dots,\mathbf{r}_i,\dots,\mathbf{r}_j,\dots,\mathbf{r}_n and let A′A' be AA with rows ii and jj swapped. Consider the matrix CC obtained from AA by replacing both row ii and row jj with ri+rj\mathbf{r}_i + \mathbf{r}_j in those two positions (all other rows as in AA). Then CC has two equal rows (positions ii and jj both equal ri+rj\mathbf{r}_i+\mathbf{r}_j), so by the alternating property det⁡(C)=0\det(C) = 0. Expand det⁡(C)\det(C) using multilinearity separately in row ii and row jj. Writing det⁡(…,ri+rj,…,ri+rj,… )\det(\dots,\mathbf{r}_i+\mathbf{r}_j,\dots,\mathbf{r}_i+\mathbf{r}_j,\dots) and expanding linearity in the row-ii slot first:

det⁡(C)=det⁡(…,ri,…,ri+rj,… )+det⁡(…,rj,…,ri+rj,… ).\det(C) = \det(\dots,\mathbf{r}_i,\dots,\mathbf{r}_i+\mathbf{r}_j,\dots) + \det(\dots,\mathbf{r}_j,\dots,\mathbf{r}_i+\mathbf{r}_j,\dots).

Now expand linearity in the row-jj slot in each term:

det⁡(C)=det⁡(…,ri,…,ri,… )+det⁡(…,ri,…,rj,… )+det⁡(…,rj,…,ri,… )+det⁡(…,rj,…,rj,… ).\det(C) = \det(\dots,\mathbf{r}_i,\dots,\mathbf{r}_i,\dots) + \det(\dots,\mathbf{r}_i,\dots,\mathbf{r}_j,\dots) + \det(\dots,\mathbf{r}_j,\dots,\mathbf{r}_i,\dots) + \det(\dots,\mathbf{r}_j,\dots,\mathbf{r}_j,\dots).

The first and last terms each have a repeated row (row ii appears in both slot ii and slot jj in the first; row jj appears in both slots in the last), so both vanish by the alternating property. The middle two terms are det⁡(A)\det(A) and det⁡(A′)\det(A') respectively (the second term is AA itself with rows i,ji,j as ri,rj\mathbf{r}_i,\mathbf{r}_j, and the third term is AA with rows i,ji,j swapped to rj,ri\mathbf{r}_j,\mathbf{r}_i, which is exactly A′A'). Hence

0=det⁡(C)=0+det⁡(A)+det⁡(A′)+0,0 = \det(C) = 0 + \det(A) + \det(A') + 0,

so det⁡(A′)=−det⁡(A)\det(A') = -\det(A).

(Replace). Let A′A' have rows r1,…,ri+crj,…,rn\mathbf{r}_1,\dots,\mathbf{r}_i + c\mathbf{r}_j,\dots,\mathbf{r}_n (row jj, j≠ij\ne i, unchanged). By multilinearity in row ii:

det⁡(A′)=det⁡(…,ri,…,rj,… )+cdet⁡(…,rj,…,rj,… ),\det(A') = \det(\dots,\mathbf{r}_i,\dots,\mathbf{r}_j,\dots) + c\det(\dots,\mathbf{r}_j,\dots,\mathbf{r}_j,\dots),

where in the second term row ii has been replaced by rj\mathbf{r}_j, so rows ii and jj are now both rj\mathbf{r}_j. The first term is det⁡(A)\det(A). The second term vanishes by the alternating property, since it is cc times the determinant of a matrix with two equal rows (positions ii and jj). Hence det⁡(A′)=det⁡(A)+0=det⁡(A)\det(A') = \det(A) + 0 = \det(A).

Conclusion. Each individual rule is proved. For a finite sequence of elementary operations transforming A0=AA_0 = A into Ak=A′A_k = A' via A0→A1→⋯→AkA_0 \to A_1 \to \cdots \to A_k, apply the appropriate rule at each step: each Swap contributes a factor of −1-1, each Scale by cmc_m contributes a factor of cmc_m, and each Replace contributes a factor of 11. Multiplying these factors across all kk steps gives

det⁡(Ak)=(−1)s(∏mcm)det⁡(A0),\det(A_k) = (-1)^{s}\left(\prod_m c_m\right)\det(A_0),

where ss is the number of swaps used and the product runs over the scale factors used, establishing the final claim. ■\blacksquare

∎
TheoremTranspose invariance

Let FF be a field and A∈Mn(F)A \in M_n(F). Then

det⁡(AT)=det⁡(A).\det(A^T) = \det(A).
Proof

We use the Computing Determinants by Cofactor Expansion formula, which expresses the determinant via the Leibniz sum over permutations:

det⁡(A)=∑σ∈Snsgn⁡(σ)∏i=1nai,σ(i),\det(A) = \sum_{\sigma \in S_n} \operatorname{sgn}(\sigma) \prod_{i=1}^n a_{i,\sigma(i)},

where SnS_n is the symmetric group on {1,…,n}\{1,\dots,n\} and sgn⁡(σ)∈{1,−1}\operatorname{sgn}(\sigma) \in \{1,-1\} is the sign of the permutation σ\sigma. This formula is equivalent to the multilinear alternating characterization; it is the concrete expansion of det⁡\det derived from those axioms and is not itself circular here since we invoke it as an already-established computational formula for det⁡\det, not as something depending on the present theorem. Let B=ATB = A^T, so bij=ajib_{ij} = a_{ji} for all i,ji,j. Then

det⁡(B)=∑σ∈Snsgn⁡(σ)∏i=1nbi,σ(i)=∑σ∈Snsgn⁡(σ)∏i=1naσ(i),i.\det(B) = \sum_{\sigma \in S_n} \operatorname{sgn}(\sigma) \prod_{i=1}^n b_{i,\sigma(i)} = \sum_{\sigma \in S_n} \operatorname{sgn}(\sigma) \prod_{i=1}^n a_{\sigma(i),i}.

Fix σ∈Sn\sigma \in S_n. In the product ∏i=1naσ(i),i\prod_{i=1}^n a_{\sigma(i),i}, reindex the product by setting j=σ(i)j = \sigma(i), i.e. i=σ−1(j)i = \sigma^{-1}(j); as ii ranges over {1,…,n}\{1,\dots,n\}, so does jj (since σ\sigma is a bijection of {1,…,n}\{1,\dots,n\}). Hence

∏i=1naσ(i),i=∏j=1naj,σ−1(j).\prod_{i=1}^n a_{\sigma(i),i} = \prod_{j=1}^n a_{j,\sigma^{-1}(j)}.

Also, sgn⁡(σ)=sgn⁡(σ−1)\operatorname{sgn}(\sigma) = \operatorname{sgn}(\sigma^{-1}), since σσ−1=e\sigma\sigma^{-1} = e has sign 1=sgn⁡(σ)sgn⁡(σ−1)1 = \operatorname{sgn}(\sigma)\operatorname{sgn}(\sigma^{-1}) (sign is a group homomorphism Sn→{±1}S_n \to \{\pm1\}), forcing sgn⁡(σ)\operatorname{sgn}(\sigma) and sgn⁡(σ−1)\operatorname{sgn}(\sigma^{-1}) to be equal (both are ±1\pm 1 and their product is 11). Therefore

det⁡(B)=∑σ∈Snsgn⁡(σ−1)∏j=1naj,σ−1(j).\det(B) = \sum_{\sigma \in S_n} \operatorname{sgn}(\sigma^{-1}) \prod_{j=1}^n a_{j,\sigma^{-1}(j)}.

As σ\sigma ranges over all of SnS_n, so does τ:=σ−1\tau := \sigma^{-1}, bijectively (inversion is a bijection Sn→SnS_n \to S_n). Substituting τ\tau for σ−1\sigma^{-1},

det⁡(B)=∑τ∈Snsgn⁡(τ)∏j=1naj,τ(j)=det⁡(A).\det(B) = \sum_{\tau \in S_n} \operatorname{sgn}(\tau)\prod_{j=1}^n a_{j,\tau(j)} = \det(A).

Hence det⁡(AT)=det⁡(B)=det⁡(A)\det(A^T) = \det(B) = \det(A). ■\blacksquare

∎
CorollaryColumn operations

Let FF be a field and A∈Mn(F)A \in M_n(F). The rules of the previous theorem hold verbatim with "row" replaced by "column": swapping two columns negates det⁡\det, scaling a column by cc scales det⁡\det by cc, and adding a multiple of one column to another leaves det⁡\det unchanged.

Proof

Let A′A' be obtained from AA by an elementary column operation of one of the three types. Then (A′)T(A')^T is obtained from ATA^T by the corresponding row operation of the same type, since transposition converts columns of AA into rows of ATA^T (this is immediate from the definition (AT)ij=Aji(A^T)_{ij} = A_{ji}: column kk of AA becomes row kk of ATA^T, and any linear operation performed on that column becomes the identical linear operation on that row). By the Row Operation Rules Theorem applied to ATA^T,

det⁡((A′)T)={−det⁡(AT)swapcdet⁡(AT)scale by cdet⁡(AT)replace.\det((A')^T) = \begin{cases} -\det(A^T) & \text{swap} \\ c\det(A^T) & \text{scale by } c \\ \det(A^T) & \text{replace.}\end{cases}

By transpose invariance, det⁡((A′)T)=det⁡(A′)\det((A')^T) = \det(A') and det⁡(AT)=det⁡(A)\det(A^T) = \det(A). Substituting gives exactly the claimed column rules. ■\blacksquare

∎
TheoremTriangular determinant formula

Let FF be a field and let T=(tij)∈Mn(F)T = (t_{ij}) \in M_n(F) be upper triangular (so tij=0t_{ij} = 0 for i>ji > j). Then

det⁡(T)=t11t22⋯tnn,\det(T) = t_{11}t_{22}\cdots t_{nn},

the product of the diagonal entries. The same formula holds if TT is lower triangular.

Proof

We use the Leibniz formula again:

det⁡(T)=∑σ∈Snsgn⁡(σ)∏i=1nti,σ(i).\det(T) = \sum_{\sigma \in S_n} \operatorname{sgn}(\sigma)\prod_{i=1}^n t_{i,\sigma(i)}.

Fix a permutation σ∈Sn\sigma \in S_n with σ≠id\sigma \ne \mathrm{id}. We claim the corresponding term ∏i=1nti,σ(i)\prod_{i=1}^n t_{i,\sigma(i)} is zero. Since σ≠id\sigma \ne \mathrm{id}, there exists some index with σ(i)≠i\sigma(i) \ne i. We show, in fact, that there must exist an index ii with σ(i)<i\sigma(i) < i. Suppose, for contradiction, that σ(i)≥i\sigma(i) \ge i for all i=1,…,ni = 1,\dots,n. Since σ\sigma is a bijection on the finite set {1,…,n}\{1,\dots,n\}, and σ(n)≥n\sigma(n) \ge n forces σ(n)=n\sigma(n) = n; then among {1,…,n−1}\{1,\dots,n-1\}, σ\sigma restricts to a bijection of {1,…,n−1}\{1,\dots,n-1\} (as σ(n)=n\sigma(n)=n is already used), and σ(n−1)≥n−1\sigma(n-1)\ge n-1 together with σ(n−1)≠n\sigma(n-1) \ne n forces σ(n−1)=n−1\sigma(n-1) = n-1. Iterating this argument downward (formally, by induction on n−in - i), we conclude σ(i)=i\sigma(i) = i for every ii, contradicting σ≠id\sigma \ne \mathrm{id}. Hence there is some index ii with σ(i)<i\sigma(i) < i, i.e. i>σ(i)i > \sigma(i). Since TT is upper triangular, ti,σ(i)=0t_{i,\sigma(i)} = 0 whenever the row index exceeds the column index, so ti,σ(i)=0t_{i,\sigma(i)} = 0. Therefore the product ∏k=1ntk,σ(k)\prod_{k=1}^n t_{k,\sigma(k)} contains this zero factor and vanishes. Consequently every term in the Leibniz sum except the term for σ=id\sigma = \mathrm{id} vanishes, and

det⁡(T)=sgn⁡(id)∏i=1nti,id(i)=1⋅∏i=1ntii=t11t22⋯tnn.\det(T) = \operatorname{sgn}(\mathrm{id})\prod_{i=1}^n t_{i,\mathrm{id}(i)} = 1 \cdot \prod_{i=1}^n t_{ii} = t_{11}t_{22}\cdots t_{nn}.

For the lower triangular case, let LL be lower triangular. Then LTL^T is upper triangular with the same diagonal entries (since (LT)ii=Lii(L^T)_{ii} = L_{ii}), so by the upper triangular case just proved, det⁡(LT)=ℓ11⋯ℓnn\det(L^T) = \ell_{11}\cdots\ell_{nn}. By transpose invariance, det⁡(L)=det⁡(LT)=ℓ11⋯ℓnn\det(L) = \det(L^T) = \ell_{11}\cdots \ell_{nn}. ■\blacksquare

∎

Worked examples

Example 1: A 3×33\times 3 matrix with a fraction-inducing pivot, computed by row reduction. Let

A=(24−21−13321).A = \begin{pmatrix} 2 & 4 & -2 \\ 1 & -1 & 3 \\ 3 & 2 & 1 \end{pmatrix}.

We compute det⁡(A)\det(A) by reducing to upper triangular form, tracking multipliers by the theorem above. Step 1 (Swap R1↔R2R_1 \leftrightarrow R_2, sign factor −1-1) — chosen to get a pivot of 11 in the top-left and avoid fractions:

A1=(1−1324−2321),det⁡(A)=−det⁡(A1).A_1 = \begin{pmatrix} 1 & -1 & 3 \\ 2 & 4 & -2 \\ 3 & 2 & 1 \end{pmatrix}, \qquad \det(A) = -\det(A_1).

Step 2 (Replace R2→R2−2R1R_2 \to R_2 - 2R_1, factor 11):

A2=(1−1306−8321),det⁡(A1)=det⁡(A2).A_2 = \begin{pmatrix} 1 & -1 & 3 \\ 0 & 6 & -8 \\ 3 & 2 & 1 \end{pmatrix}, \qquad \det(A_1) = \det(A_2).

Step 3 (Replace R3→R3−3R1R_3 \to R_3 - 3R_1, factor 11):

A3=(1−1306−805−8),det⁡(A2)=det⁡(A3).A_3 = \begin{pmatrix} 1 & -1 & 3 \\ 0 & 6 & -8 \\ 0 & 5 & -8 \end{pmatrix}, \qquad \det(A_2) = \det(A_3).

Step 4 (Replace R3→R3−56R2R_3 \to R_3 - \tfrac{5}{6}R_2, factor 11):

R3new=(0, 5−56⋅6, −8−56⋅(−8))=(0, 0, −8+406)=(0,0,−83).R_3^{\text{new}} = (0,\, 5 - \tfrac56\cdot 6,\, -8 - \tfrac56\cdot(-8)) = \left(0,\ 0,\ -8 + \tfrac{40}{6}\right) = \left(0, 0, -\tfrac{8}{3}\right).
A4=(1−1306−800−83),det⁡(A3)=det⁡(A4).A_4 = \begin{pmatrix} 1 & -1 & 3 \\ 0 & 6 & -8 \\ 0 & 0 & -\tfrac83 \end{pmatrix}, \qquad \det(A_3) = \det(A_4).

Now A4A_4 is upper triangular, so by the Triangular Determinant Formula,

det⁡(A4)=1⋅6⋅(−83)=−16.\det(A_4) = 1 \cdot 6 \cdot \left(-\tfrac83\right) = -16.

Unwinding the accumulated factor of −1-1 from the single swap:

det⁡(A)=−det⁡(A1)=−det⁡(A2)=−det⁡(A3)=−det⁡(A4)=−(−16)=16.\det(A) = -\det(A_1) = -\det(A_2) = -\det(A_3) = -\det(A_4) = -(-16) = 16.

Check by cofactor expansion along the first row of the original AA:

det⁡(A)=2det⁡(−1321)−4det⁡(1331)+(−2)det⁡(1−132).\det(A) = 2\det\begin{pmatrix}-1 & 3\\ 2& 1\end{pmatrix} - 4\det\begin{pmatrix}1&3\\3&1\end{pmatrix} + (-2)\det\begin{pmatrix}1&-1\\3&2\end{pmatrix}.

Compute each minor: det⁡(−1321)=(−1)(1)−(3)(2)=−1−6=−7\det\begin{pmatrix}-1&3\\2&1\end{pmatrix} = (-1)(1)-(3)(2) = -1-6=-7; det⁡(1331)=1−9=−8\det\begin{pmatrix}1&3\\3&1\end{pmatrix} = 1-9=-8; det⁡(1−132)=2−(−3)=5\det\begin{pmatrix}1&-1\\3&2\end{pmatrix}=2-(-3)=5. So

det⁡(A)=2(−7)−4(−8)+(−2)(5)=−14+32−10=8.\det(A) = 2(-7) - 4(-8) + (-2)(5) = -14+32-10 = 8.

This disagrees with 1616 — a deliberate check that reveals an arithmetic slip must be hunted down; redo Step 4 carefully: −8−56(−8)=−8+406=−8+203=−24+203=−43-8 - \tfrac56(-8) = -8 + \tfrac{40}{6} = -8+\tfrac{20}{3} = \tfrac{-24+20}{3} = -\tfrac43, not −83-\tfrac83. Correcting: det⁡(A4)=1⋅6⋅(−43)=−8\det(A_4) = 1\cdot 6\cdot(-\tfrac43) = -8, and det⁡(A)=−(−8)=8\det(A) = -(-8) = 8, matching the cofactor computation exactly. This illustrates precisely why one always cross-checks a row-reduction computation: arithmetic slips in the multipliers are the single most common source of error, not a flaw in the method itself.

Example 2: Using column operations and transpose invariance together, 4×44\times 4. Let

B=(120325−16−10213618).B = \begin{pmatrix} 1 & 2 & 0 & 3\\ 2 & 5 & -1 & 6\\ -1 & 0 & 2 & 1\\ 3 & 6 & 1 & 8 \end{pmatrix}.

Notice column 4 is close to 33 times column 1 plus something; rather than search for such patterns, apply row reduction directly, using Replace operations only (so no sign or scale tracking is needed). Step 1: R2→R2−2R1R_2 \to R_2 - 2R_1, R3→R3+R1R_3 \to R_3 + R_1, R4→R4−3R1R_4 \to R_4 - 3R_1:

B1=(120301−100224001−1),det⁡(B)=det⁡(B1).B_1 = \begin{pmatrix} 1 & 2 & 0 & 3\\ 0 & 1 & -1 & 0\\ 0 & 2 & 2 & 4\\ 0 & 0 & 1 & -1 \end{pmatrix}, \qquad \det(B) = \det(B_1).

(Check row 2: (2,5,−1,6)−2(1,2,0,3)=(0,1,−1,0)(2,5,-1,6) - 2(1,2,0,3) = (0,1,-1,0). Row 3: (−1,0,2,1)+(1,2,0,3)=(0,2,2,4)(-1,0,2,1)+(1,2,0,3) = (0,2,2,4). Row 4: (3,6,1,8)−3(1,2,0,3)=(0,0,1,−1)(3,6,1,8) - 3(1,2,0,3) = (0,0,1,-1).) Step 2: R3→R3−2R2R_3 \to R_3 - 2R_2:

new R3=(0,2,2,4)−2(0,1,−1,0)=(0,0,4,4).\text{new } R_3 = (0,2,2,4) - 2(0,1,-1,0) = (0,0,4,4).
B2=(120301−100044001−1),det⁡(B1)=det⁡(B2).B_2 = \begin{pmatrix} 1 & 2 & 0 & 3\\ 0 & 1 & -1 & 0\\ 0 & 0 & 4 & 4\\ 0 & 0 & 1 & -1 \end{pmatrix}, \qquad \det(B_1) = \det(B_2).

Step 3: R4→R4−14R3R_4 \to R_4 - \tfrac14 R_3:

new R4=(0,0,1,−1)−14(0,0,4,4)=(0,0,0,−2).\text{new }R_4 = (0,0,1,-1) - \tfrac14(0,0,4,4) = (0,0,0,-2).
B3=(120301−100044000−2),det⁡(B2)=det⁡(B3).B_3 = \begin{pmatrix} 1 & 2 & 0 & 3\\ 0 & 1 & -1 & 0\\ 0 & 0 & 4 & 4\\ 0 & 0 & 0 & -2 \end{pmatrix}, \qquad \det(B_2) = \det(B_3).

B3B_3 is upper triangular, so by the Triangular Determinant Formula,

det⁡(B3)=1⋅1⋅4⋅(−2)=−8.\det(B_3) = 1\cdot 1\cdot 4\cdot(-2) = -8.

Since only Replace operations were used, no correction factors are needed:

det⁡(B)=det⁡(B1)=det⁡(B2)=det⁡(B3)=−8.\det(B) = \det(B_1) = \det(B_2) = \det(B_3) = -8.

To illustrate the Column Operations Corollary and transpose invariance simultaneously, compute det⁡(BT)\det(B^T) independently. By the Transpose Invariance Theorem, det⁡(BT)=det⁡(B)=−8\det(B^T) = \det(B) = -8 automatically, with no further computation required — this is exactly the labor-saving content of that theorem: one never needs to separately row-reduce the transpose.

CounterexampleWhy the field/scalar hypothesis on Scale matters

The Scale rule requires the multiplier cc to be an arbitrary field scalar, and the theorem's derivation used only that FF is a field (so every nonzero element has an inverse and there are no zero divisors). If one is tempted to relax the elementary-operation framework by allowing "scaling by c=0c=0" as an elementary operation (rather than merely evaluating the linearity rule at c=0c=0, which is fine), disaster follows for the row-reduction algorithm, though not for the determinant identity itself. Consider

A=(1234),det⁡(A)=1⋅4−2⋅3=−2.A = \begin{pmatrix} 1 & 2 \\ 3 & 4\end{pmatrix}, \qquad \det(A) = 1\cdot4 - 2\cdot 3 = -2.

The identity det⁡(A′)=cdet⁡(A)\det(A') = c\det(A) still holds formally at c=0c = 0: scaling row 11 by 00 gives

A′=(0034),det⁡(A′)=0⋅4−0⋅3=0=0⋅(−2),A' = \begin{pmatrix} 0 & 0 \\ 3 & 4\end{pmatrix}, \qquad \det(A') = 0\cdot4 - 0\cdot 3 = 0 = 0\cdot(-2),

consistent with the formula. The real danger is definitional: elementary row operations used in the row-reduction algorithm for solving linear systems and for invertibility are required to have c≠0c \ne 0 precisely because scaling by 00 is not invertible as an operation — it destroys information (two different matrices can have the same image after a 00-scale), so it cannot be undone, and it collapses an invertible matrix's row space. If c=0c=0 were permitted as an elementary operation, the reduction A→A′A \to A' above would assert that AA (invertible, det⁡=−2≠0\det = -2 \ne 0) is "row-equivalent" to A′A' (singular, det⁡=0\det = 0), destroying the entire correspondence between row-equivalence and having the same determinant up to a nonzero scalar. This is why the definition of elementary Scale operations always insists c≠0c \ne 0 as a standing hypothesis — dropping it breaks the invertibility of each elementary step, not the determinant formula itself.

Scope, edge cases & common traps

SituationWhat happensCommon trap
Swap rows twiceTwo sign flips cancel: (−1)2=1(-1)^2=1Forgetting to track parity of swaps across a long reduction
Scale row by c=1c=1No change, factor is 11Confusing "scale" with "replace" bookkeeping
Row reduce with only Replace opsNo correction factor needed at allAssuming a factor is needed even when no swaps/scales occurred
Triangular matrix with a zero on diagonaldet⁡=0\det = 0 automaticallyTrying to continue reducing past a zero pivot without swapping first
Applying rules to a non-square arrayUndefined — determinant only for square matricesAttempting "row reduction determinant" on a rectangular matrix
Column vs row operationsBoth valid, by the Column Operations CorollaryAssuming column swaps behave differently in sign from row swaps (they do not)

The mathematician behind it

Origins · 1 minAugustin-Louis Cauchy and how this was discovered
AC
Augustin-Louis Cauchy
1789–1857 · French mathematician, École Polytechnique and the Paris Academy of Sciences

Cauchy gave the first systematic treatment of determinants as objects satisfying algebraic laws in their own right, in his 1812 memoir on the subject, unifying earlier scattered results of Leibniz, Cramer, Vandermonde, and Laplace. He introduced much of the modern permutation-based notation and proved the multiplication theorem for determinants, laying groundwork that made the row-operation calculus — the practical engine for computing determinants used throughout linear algebra — a rigorous and teachable subject rather than a collection of ad hoc tricks.

History
- 1750: Gabriel Cramer publishes Cramer's Rule, using determinant-like expressions to solve linear systems.
- 1772: Alexandre-Théophile Vandermonde treats determinants as an independent subject, not merely a tool for solving equations.
- 1812: Augustin-Louis Cauchy's memoir systematizes determinant theory, proves the multiplication rule, and introduces permutation notation.
- 1841: Carl Gustav Jacobi popularizes determinants of derivatives (the Jacobian), cementing their role beyond linear systems.
- 1867: Charles Dodgson (Lewis Carroll) publishes an elementary condensation method exploiting row-operation-like identities.
Why does swapping rows flip the sign of the determinant, geometrically?

Geometrically, the rows of a matrix are vectors, and the determinant is the signed volume of the parallelepiped they span, where the sign records orientation. Swapping two of the spanning vectors is a reflection, and reflections reverse orientation while preserving unsigned volume. This is why ∣det⁡(A′)∣=∣det⁡(A)∣|\det(A')| = |\det(A)| after a swap, but the sign flips. Algebraically, the proof does not depend on geometry at all: it follows purely from the alternating property det⁡(…,v,…,v,… )=0\det(\dots,\mathbf v,\dots,\mathbf v,\dots)=0 applied to the sum ri+rj\mathbf r_i+\mathbf r_j placed in both slots, expanded by multilinearity, as shown in the theorem's proof.

Do the row operation rules also work for complex matrices, or only real ones?

Yes — the entire development in this article is carried out over an arbitrary field FF, which includes R\mathbb{R}, C\mathbb{C}, and finite fields such as Fp\mathbb{F}_p. The proofs use only the field axioms: additive and multiplicative structure, and the existence of inverses for nonzero scalars in the Scale rule. Nothing in the argument invokes properties special to the real numbers, such as order or completeness, so all three row-operation rules, transpose invariance, and the triangular formula hold verbatim over C\mathbb{C} and over finite fields.

Why is the identity matrix's determinant defined to be exactly 1, rather than derived?

The normalization det⁡(In)=1\det(I_n)=1 is one of the three defining axioms (alongside multilinearity and the alternating property), not a derived fact, because multilinear alternating forms on nn-tuples of vectors in FnF^n form a one-dimensional vector space. Any nonzero alternating multilinear form is a scalar multiple of any other. Fixing the value at InI_n to be 11 pins down a single specific form out of that one-dimensional family, which is what makes "the" determinant unique rather than "a" determinant unique only up to scalar.

If I make an arithmetic mistake tracking the sign from swaps, how can I catch it?

The most reliable check is to recompute the same determinant by a second, structurally different method — cofactor expansion along a different row or column, or a different sequence of row operations — and confirm the two answers agree, exactly as demonstrated in the first worked example above, where an initial slip was caught this way. A useful sanity check specific to row reduction is to count swaps and scalings separately as you go, writing down the running correction factor at each step rather than trying to recall it at the end, since forgetting a single swap is the most common source of a sign error.

Does the column operations corollary need a separate proof from scratch, or does it really just follow from transpose invariance?

It follows entirely from transpose invariance together with the already-proved row rules; no new argument about columns is needed. The key observation is that performing a column operation on AA is the same, after transposing, as performing the corresponding row operation on ATA^T, since transposition exchanges the roles of rows and columns entry-by-entry. Combined with det⁡(AT)=det⁡(A)\det(A^T)=\det(A), this transfers every row-operation fact automatically into the corresponding column-operation fact, which is exactly the labor-saving content of proving transpose invariance in the first place.

Can I mix row and column operations in the same computation?

Yes, and doing so is often more efficient, though it is important to track the correction factors from row operations and column operations separately (or in a single combined running product), since both types affect the same determinant. For example, one might use a column swap to bring a convenient pivot to the top-left, then proceed with row Replace operations to clear the rest of that column, still arriving at a triangular matrix whose diagonal product, corrected by all sign flips and scalings from both row and column operations used, equals the original determinant.

Why does the triangular determinant proof need to rule out every non-identity permutation, rather than just checking a few?

Because the Leibniz formula sums over all n!n! permutations, and correctness of the triangular formula requires that every single non-identity term vanishes, not just some of them. The proof handles this uniformly: for any permutation σ≠id\sigma \ne \mathrm{id}, it shows there must exist some index ii with σ(i)<i\sigma(i) < i, forcing a zero entry ti,σ(i)t_{i,\sigma(i)} into that term's product by the triangularity condition. This single argument, using only the pigeonhole-style descending induction shown in the proof, disposes of all n!−1n!-1 non-identity terms simultaneously, rather than requiring a case-by-case check.

Before this

Where this leads