Determinant

The formula

The determinant is written det(A) or |A|. For a square matrix it collapses the whole table of numbers down to a single scalar.

2x2:   | a  b |
       | c  d |  = a·d - b·c

Reads as: cross-multiply the diagonals and subtract. For [[3, 8], [4, 6]] that is 3·6 - 8·4 = -14.

3x3 (Sarrus):
| a b c |
| d e f |  = aei + bfg + cdh - ceg - bdi - afh
| g h i |

Reads as: sum the three "down-right" diagonal products, subtract the three "down-left" ones. Sarrus is a memory trick that works for 3x3 only — do not try to extend it to 4x4.

Laplace (cofactor) expansion along row i:
det(A) = sum over j of  (-1)^(i+j) · a_ij · M_ij

Reads as: walk across one row, and for each entry multiply it by the determinant of the smaller matrix you get by deleting that entry's row and column (M_ij, the minor), with a checkerboard of plus/minus signs. This is the general definition — it recurses down to 2x2 blocks.

Multiplicative property:   det(A·B) = det(A) · det(B)
Inverse:                   det(A^-1) = 1 / det(A)
Transpose:                 det(A^T)  = det(A)
Triangular matrix:         det = product of the diagonal entries

Reads as: determinants of a product multiply; a matrix is invertible exactly when its determinant is nonzero; transposing changes nothing; and for a triangular matrix you just multiply the diagonal.

Where you meet it

A control review board, checking stability. The characteristic equation of a linear system is det(sI - A) = 0. The roots are the poles. When you factor an aircraft or missile pitch model into a state matrix A, the determinant of sI - A produces the polynomial whose roots decide whether the response settles or diverges. The Routh-Hurwitz table you build to check stability is a determinant test in disguise — its leading minors are determinants.
A test stand, solving a multi-channel calibration. Fit a strain-gauge bridge or a six-axis load cell and you end up solving A·x = b for the calibration coefficients. If det(A) comes back at or near zero, your excitation vectors were not independent — you loaded the fixture along directions that the math cannot separate. The determinant is the early warning that the calibration matrix is singular or ill-conditioned before the coefficients come out as garbage.
A coordinate change in an integral. Moving a surface or volume integral from Cartesian to polar, cylindrical, or spherical coordinates drops a Jacobian determinant into the integrand. In polar coordinates that factor is r — the area element dx·dy becomes r·dr·dθ. Miss the Jacobian and your integrated antenna pattern, thermal load, or mass property is off by a position-dependent factor.
Mesh and CAD sanity checks. Finite-element pre-processors flag elements whose Jacobian determinant goes negative. A negative determinant means the element got turned inside-out during meshing — the coordinate map folded over on itself. The solver will either refuse to run or return nonsense stresses at that element.

How it works

The cleanest way to read a determinant is geometric. Treat the columns of the matrix as vectors. In 2D, det is the signed area of the parallelogram those two column vectors span. In 3D, it is the signed volume of the parallelepiped the three columns span. That single idea explains almost every property:

Zero determinant means collapse. If the columns are linearly dependent — one is a combination of the others — the parallelepiped is flat, volume zero, and det = 0. That is the same condition as "the matrix is singular" and "the system A·x = b has no unique solution." [[1, 2], [2, 4]] has determinant 1·4 - 2·2 = 0 because the second row is just twice the first.
The sign is orientation. A negative determinant means the transformation flips handedness — a right-handed frame becomes left-handed. Swap two rows and the sign flips; that mirrors the parallelepiped. This is why a swapped pair of columns in a rotation matrix quietly turns it into a reflection.
Magnitude is the volume scale factor. det(A) = 6 means A multiplies every volume by 6. That is exactly why the Jacobian determinant is the right correction factor when you change variables in an integral: it is the local volume-stretch of the coordinate map.

The gotchas are mostly about not confusing the determinant with things it resembles:

Determinant is not conditioning. A tiny determinant does not by itself mean the matrix is numerically bad, and a large one does not mean it is good. Scale a well-behaved 1000x1000 identity-like matrix by 0.1 and its determinant is 0.1^1000, an underflow-to-zero number, while the matrix is perfectly invertible. For "is this system safe to solve numerically," use the condition number, not the determinant.
Never solve linear systems with Cramer's rule. Cramer's rule expresses each unknown as a ratio of determinants, and it is a fine theoretical statement. As an algorithm it is catastrophic: cofactor expansion of an n x n determinant costs on the order of n! operations. A 20x20 system that way is roughly 10^18 operations; the same system by LU decomposition is a few thousand. In practice determinants themselves are computed by LU factorization — reduce to triangular form, multiply the diagonal, track the sign from row swaps. That is O(n^3), the same cost as just solving the system.
Watch the overflow/underflow. For large matrices the determinant can be astronomically large or vanishingly small even when nothing is wrong. Numerical libraries return the log-determinant and a separate sign for exactly this reason. If you need det of anything past maybe 10x10, reach for slogdet rather than the raw value.
Determinant of a sum is not the sum of determinants. det(A + B) has no simple relationship to det(A) and det(B). Only the product rule is clean.

History

Determinants showed up before matrices did, which surprises people — the number got invented to solve systems of equations, and the rectangular array it summarizes came later. The earliest known work is Japanese: Seki Takakazu, working out elimination methods for polynomial systems, wrote out determinant calculations for cases from 2x2 up through 5x5 in a 1683 treatise, entirely independent of anything in Europe ^[1]^[2]. Ten years later, in 1693, Leibniz described the same idea in a letter to l'Hopital, using it to state when a linear system has a solution — but the work sat unpublished and had little immediate reach ^[1]^[2].

The idea entered the European mainstream through Gabriel Cramer, who in 1750 published the rule that still carries his name, stated without proof in an appendix on algebraic curves ^[1]^[2]. Over the next generation Vandermonde (1771) treated the determinant as an object worth studying in its own right rather than a bookkeeping trick, and Laplace (1772) gave the cofactor expansion now named after him ^[1]^[2]. Gauss used the word "determinant" in his 1801 Disquisitiones Arithmeticae, though he applied it to the discriminant of a quadratic form, not quite to the matrix quantity we mean today ^[1]^[2]. The modern meaning, and the first proof of the multiplication theorem det(AB) = det(A)det(B), came from Augustin-Louis Cauchy in 1812 ^[1]^[2]. The word stuck; the theory Cauchy laid down is essentially the one engineers use now.

Related tools

/tools/strain-gauge-bridge
/tools/wheatstone-bridge
/tools/euler-buckling
/tools/vibration-natural-freq
/tools/series-rlc-impedance

Sources

Written by HE in our own words from the cited sources — engineering judgment included, your stamp still required. All entries →