You find the determinant of a 3x3 matrix by expanding along any row or column: multiply each entry by the 2x2 determinant left when you cross out its row and column, apply the sign pattern + − + / − + − / + − +, and add the three results. For matrix A with rows (a, b, c), (d, e, f), (g, h, i), that gives a(ei − fh) − b(di − fg) + c(dh − eg). The rule of Sarrus reaches the same number faster, but it works for 3x3 only — never for 4x4 or larger.
What the Determinant Actually Measures Before You Compute It
The determinant is a single number attached to a square matrix. For a 3x3 matrix it equals the signed volume of the parallelepiped whose edges are the matrix's row vectors, which is why geometry courses lean on it so heavily. A determinant of 6 means that box holds six cubic units.
Two quick facts follow from that picture, and they're worth checking before you grind through arithmetic. If any row is a multiple of another row, the box is flat and the determinant is zero. The same happens when any row is all zeros or when one row is a sum of the other two.
Zero also tells you the matrix has no inverse. A determinant-like value called the minor exists for non-square matrices, but the determinant itself is defined only for square ones, so a 2x3 or 3x2 array has no determinant at all. Keep that distinction straight and half the confusion around this topic evaporates.
The number can be negative, and the sign carries meaning. Swapping two rows of A flips the sign of det(A) while leaving the magnitude untouched, per standard linear algebra treatment such as David C. Lay's Linear Algebra and Its Applications. That's a useful self-check: if two versions of your work differ only in sign, suspect a row swap or a sign slip.
Name the Entries First: The Row-Column Map That Prevents Most Errors
Label the nine entries before touching arithmetic. Reading left to right, top to bottom gives a, b, c in row 1; d, e, f in row 2; g, h, i in row 3. Write them down on paper with brackets around them and the labels above each column.
Mislabeling is the most common source of wrong answers, especially with negative entries. If one entry in your matrix is −4, decide now whether that sign belongs to the entry or to the cofactor sign in front of it. Treat every entry as a signed number in its own right; the cofactor signs are a separate layer.
The generic 3x3 template looks like this: row 1 is (a, b, c), row 2 is (d, e, f), row 3 is (g, h, i). Every formula in this article uses those letters, so a matrix like (2, 1, 3; 0, −4, 5; 1, 2, −1) maps to a = 2, b = 1, c = 3, d = 0, e = −4, f = 5, g = 1, h = 2, i = −1.
Write the sign grid beside the matrix: the top-left cell is positive, the one to its right negative, and the pattern alternates across and down. That grid is sometimes called the checkerboard, and it applies no matter which row or column you expand along.
The Cofactor Expansion, Step by Step in the Order You Actually Do It
Here is the full procedure, in the sequence that minimizes backtracking.
1: Write out the labeled matrix and the nine-entry template above it. 2: Choose the row or column with the most zeros — that choice is legal because every row and column gives the identical determinant. 3: For the first entry in your chosen line, cross out its row and its column, then write the four remaining entries as a 2x2 determinant in the form (top-left × bottom-right) − (top-right × bottom-left). 4: Multiply that 2x2 result by the entry, then apply the cofactor sign (+ or −) from the checkerboard for that position. 5: Repeat for the second and third entries of the chosen line. 6: Add the three signed terms and check the result against a mental estimate of the volume.
Applied to (2, 1, 3; 0, −4, 5; 1, 2, −1) with expansion along row 1: the first term is 2 × ((−4)(−1) − (5)(2)) = 2 × (4 − 10) = −12. The second is −1 × ((0)(−1) − (5)(1)) = −1 × (−5) = 5. The third is 3 × ((0)(2) − (−4)(1)) = 3 × 4 = 12. The determinant is −12 + 5 + 12 = 5.
If any intermediate 2x2 comes out as a clean positive-looking number and you're not sure about its sign, recompute it in isolation before folding it into the sum. A single flipped sign inside a 2x2 costs you the entire answer, and the error is invisible in the final line.
Why the Sign Pattern Trips People Up and How to Keep It Straight
The checkerboard is not optional decoration; it's baked into the expansion formula. For row 1 the signs are +, −, +. For row 2 they shift to −, +, −. For row 3 they return to +, −, +. Columns follow the same grid: column 1 is +, −, +, column 2 is −, +, −, column 3 is +, −, +.
Two habits protect you. First, write the signs above the entries in the row you're expanding before doing any multiplication, so each term gets its sign at the moment you compute it. Second, when an entry itself is negative, keep the two signs separate: a negative entry in a minus cofactor position produces a plus term, and that double negative is precisely where answers go astray.
Try expanding the example matrix along the first column instead and confirm you get 5 again. Column 1 signs are +, −, +, so the terms are 2 × (4 − 10) = −12, then −0 × (…), then +1 × (5 − (−12)) = 17, and −12 + 0 + 17 = 5. The zero entry kills its whole term, which is the efficiency argument for scanning for zeros first.
This is also why an entry of zero is worth hunting: a zero wipes out a 2x2 computation entirely, not just the multiplication by the entry. Clever row or column selection can shrink the arithmetic by a third without changing the answer.
Sarrus' Rule in 75 Seconds: When It Helps and Where It Fails
The rule of Sarrus is a shortcut specific to 3x3 matrices. Rewrite the first two columns to the right of the matrix, then add three diagonal products running top-left to bottom-right and subtract three running the other way. For the template, the six products are aei, bfg, cdh (added) and ceg, afh, bdi (subtracted), so the determinant is aei + bfg + cdh − ceg − afh − bdi.
On the same matrix, the diagonals give 2(−4)(−1) + 1(5)(1) + 3(0)(2) − 3(−4)(1) − 2(5)(2) − 1(0)(−1) = 8 + 5 + 0 − (−12) − 20 − 0 = 5. Same answer, six multiplications, no cofactor signs to track.
That speed comes with a hard boundary. Sarrus does not generalize: a 4x4 determinant has 24 signed permutation terms, and writing out "extended columns" for it produces garbage. If a course or exam expects cofactor expansion or an LU-based method on anything larger, Sarrus is a dead end beyond 3x3.
Anecdotally, the diagonals are physically easy to draw — three lines that slant one way, three that slant the other. The subtraction set is the one running bottom-left to top-right, and forgetting to subtract it is the classic Sarrus error. If your answer is exactly double the diagonal-sum of one direction, you've almost certainly added where you should have subtracted.
Expand Along the Zero Row: A Worked Example With Column Arithmetic
Take the matrix (1, 2, 3; 2, 4, 6; 7, 8, 9) and compute the determinant directly. Expanding along row 1: 1 × (4·9 − 6·8) − 2 × (2·9 − 6·7) + 3 × (2·8 − 4·7) = (36 − 48) − 2(18 − 42) + 3(16 − 28) = −12 + 48 − 36 = 0. The zero is expected, because row 2 is exactly twice row 1 — a linearly dependent row, flat volume.
Now a second technique, useful when fractions or awkward numbers make a direct expansion painful. Add a multiple of one row to another to manufacture zeros, then expand. Start with (3, 1, 2; 6, 4, 5; 9, 7, 8). Subtract twice row 1 from row 2 to get (0, 2, 1), and three times row 1 from row 3 to get (0, 4, 2). Expanding along column 1 now leaves a single term: 3 × (2·2 − 1·4) = 0.
That result again encodes dependency: row 3 of the new matrix is twice row 2, so a zero determinant is correct. Both examples exist to show that row operations aim zeros at one line, converting three 2x2 computations into one, and that a zero answer is a real answer, not a mistake.
Practice sets on this pattern — especially matrices where one row is a visible combination of the others — build the instinct that the answer will land on zero. Gilbert Strang's Introduction to Linear Algebra treats it as a first-line sanity check: an obviously singular-looking matrix should return zero, and if it doesn't, retrace your steps.
Checking Your Work: Four Independent Verifications
A determinant answer is worth checking four ways, and each takes under a minute.
1: Expand along a different row or column and compare. Any consistent answer across two expansions is very likely correct, because systematic sign errors rarely survive both paths. 2: Apply Sarrus to the same matrix if you expanded by cofactors; the two methods share no arithmetic, so agreement is strong evidence. 3: Swap two rows of the matrix, recompute, and confirm the new determinant is the negative of the original. This tests your sign discipline from the outside. 4: Estimate the volume magnitude from the row lengths (roughly the product of the row norms), and reject any answer wildly outside that ballpark.
Try a matrix with a variable entry to see the payoff. Let the matrix be (k, 1, 0; 1, k, 1; 0, 1, k). Expansion along row 1 gives k(k² − 1) − 1(k) + 0 = k³ − 2k. Setting that to zero shows the matrix is singular at k = 0 and k = ±√2 — the kind of question that appears on discrete math and eigenvalue quizzes far more often than in pure arithmetic practice.
If your answer fails two checks, don't patch numbers. Rewrite the expansion from a freshly copied matrix, ideally expanding along a different line, and compare the term structure before the sum. Repeating the same path tends to reproduce the same slip.
Worked Numerical Check: One Matrix, Three Methods, One Answer
Take A = (5, −2, 1; 0, 3, −1; 2, 0, 4). This one has two zeros, one advantage to each method depending on how you deploy them.
By cofactor expansion along row 2 (signs −, +, −): term 1 is −0 × (…), term 2 is +3 × ((5)(4) − (1)(2)) = 3 × 18 = 54, term 3 is −(−1) × ((5)(0) − (−2)(2)) = +1 × 4 = 4. Total = 58.
By Sarrus: 5·3·4 + (−2)(−1)(2) + 1·0·0 − 1·3·2 − 5(−1)·0 − (−2)·0·4 = 60 + 4 + 0 − 6 − 0 − 0 = 58.
By full expansion along row 1: 5(12 − 0) − (−2)(0 − (−2)) + 1(0 − 6) = 60 − 4 − 6 = 50. That's different — and the reason is instructive. The second 2x2 minor should be (0·4 − (−1)·2) = 2, so −(−2)(2) = +4, and the third is (0·0 − 3·2) = −6, so +1(−6) = −6. Correct total: 60 + 4 − 6 = 58.
The arithmetic is the point: two methods agreeing at 58, with the third delivered only after a sign-driven error, is exactly the pattern that should make you distrust a lone computation. When numbers look "nearly right," the small difference is almost always an entry sign applied with the wrong cofactor sign.
Platform Presets Worth Knowing: TI-84, NumPy, and Hand Arithmetic
On a Texas Instruments TI-84 Plus or TI-84 Plus CE, the built-in det( function lives under MATRX → MATH, and it accepts a matrix you've entered through the MATRX → EDIT screen. The function returns the determinant directly, but it inherits the display precision limits of the calculator, so entries that produce fractions may round; for exact symbolic work the handheld isn't the right tool.
In Python, NumPy's numpy.linalg.det computes a floating-point determinant using LU decomposition, and floats follow IEEE 754. A matrix whose exact determinant is 0 can return something like 1.1102e-16 instead of 0, a rounding residue, not a mathematical answer. Check whether the result is within tolerance of zero rather than equal to zero.
Hand computation still wins when you need the intermediate minors, when the entries are symbolic, or when an exam forbids calculators. Mixed entries like k and √2 make software awkward and hand expansion natural. Practically, the matrix (2.5, 0, 1; 0.25, 3, 0; 1, 0.5, 4) is fine on a TI-84 but slows a hand expansion into minute-counting, while (1, 0, 2; 0, 1, 0; 0, 0, 1) collapses instantly on paper.
Whichever route you take, record the intermediate terms. A calculator gives you the final number but no trace of the path, and a wrong determinant with no shown work is harder to debug than a correct one you can't explain — and coursework exams usually require the shown work anyway.
Adjacent Ideas After You Can Compute It: Inverse, Cramer's Rule, Cross Product
Once the determinant is reliable, three neighbouring topics open up. The inverse of A is the adjugate divided by det(A), and that division is why a zero determinant blocks inversion — no amount of rearrangement recovers it. The nine cofactors you already computed are the entries of that adjugate, transposed, so most of the work is reused.
Cramer's rule then solves a linear system Ax = b by replacing one column at a time and dividing determinants. For the first unknown x, you take det of the matrix with column 1 swapped for b, and divide by det(A). The denominators are all the same, so a single zero determinant kills the whole solution attempt.
The cross product in vector calculus runs through the same 2x2 minors: u × v has components (u₂v₃ − u₃v₂, u₃v₁ − u₁v₃, u₁v₂ − u₂v₁), which is the cofactor expansion of a 3x3 whose top row is the standard basis vectors. If you've done cross products, you've already used this machinery.
Matrix determinants connect to eigenvalues through the characteristic equation det(A − λI) = 0. Warning before you compute: determinants and eigenvalues do not distribute — det(A + B) is not det(A) + det(B), and det(kA) for a 3x3 matrix is k³ times det(A), not k times it. A scalar multiple multiplies each of the three rows, and each row multiplies the determinant by that scalar, so a factor of k in every entry contributes k³.
FAQ
What is the fastest way to do a 3x3 determinant by hand?
Pick the line with the most zeros and expand along it, or use Sarrus if the matrix has no zeros and the course allows the shortcut. Both reach the answer in under a minute for typical integer entries. The maximum speedup comes from spotting zeros before computing anything.
Can Sarrus' rule be used on a 4x4 matrix?
No. Sarrus relies on six signed permutation terms that exist only for 3x3 matrices; 4x4 has 24, and the extended-column diagram doesn't generate them. Use cofactor expansion, row reduction to upper triangular form, or software for larger sizes.
Why is the determinant of a singular matrix exactly zero?
Because a singular matrix collapses space into a lower dimension, so its row vectors are linearly dependent and the parallelepiped they span has zero volume. Row 2 being twice row 1 in the example (1, 2, 3; 2, 4, 6; 7, 8, 9) is exactly this situation, and the determinant computes to 0 as expected.
Is the determinant of A + B equal to det(A) + det(B)?
No, and the failure can be dramatic: 2x2 identity matrices give det(I) = 1, but det(I + I) = det(2I) = 4, not 2. For 3x3 matrices the same scaling rule gives det(kA) = k³ det(A), so doubling every entry multiplies the determinant by 8, not 2.
What does a negative determinant mean geometrically?
The sign records orientation: a negative determinant means the transformation reverses handedness, like a mirror reflection of the coordinate system. The magnitude is still the volume scale factor. Swapping two rows in the matrix flips that sign, which is a convenient check on your arithmetic.
Why does NumPy return 1.1102e-16 instead of 0 for a singular matrix?
NumPy's LU decomposition works in floating point, so an exact zero can leave a rounding residue around 1e-16. Treat magnitudes below roughly 1e-12 as zero when testing singularity, and follow NumPy's own documentation on numpy.linalg.det rather than treating the raw float as exact.
Can I compute a determinant without a calculator for an exam?
Yes, and most linear algebra courses expect it. Carry the nine labeled entries, write the checkerboard signs before multiplying, and expand along the zero-rich line. Check by expanding a second line; if the entries are small integers the two answers should match exactly.
Do I need the cofactor signs if I expand along a column rather than a row?
Yes, the same checkerboard applies to columns — column 1 is +, −, +, column 2 is −, +, −, column 3 is +, −, +. Choosing a column changes which entries multiply, not the sign pattern that accompanies each position.
Which method should I use when the matrix contains variables like k?
Cofactor expansion is usually cleanest, because each minor stays symbolic and you can track k through the terms. Sarrus works and gives a single polynomial, but it's easier to drop a sign when all six products contain letters.