Determinants

Why we care (before the formula)

When you treat a matrix A as a linear transformation, it takes shapes in ℝⁿ and maps them somewhere else. A crucial question is:

• •Does A preserve dimension (so the map is invertible), or does it collapse some directions (so information is lost)?
• •How much does A scale area/volume?
• •Does A flip orientation (like a mirror reflection)?

The determinant packages these geometric behaviors into a single scalar.

Definition (intuition first)

For an n×n matrix A, the determinant det(A) is a real number with three core behaviors (these can be taken as axioms that uniquely characterize it):

1) Multilinear in rows (or columns)

• •If you fix all rows except one, det is linear in that row.
• •“Linear” here means scaling and addition distribute.

2) Alternating

• •If you swap two rows, the determinant changes sign.
• •If two rows are equal (or one is a multiple of another), det(A) = 0.

3) Normalization

• •det(I) = 1 for the identity matrix I.

These properties are not random: they are exactly what you’d want from a “signed volume scaling factor.”

Geometric meaning: signed volume scaling

Interpret A as acting on vectors in ℝⁿ. Consider the unit square (in 2D) or unit cube (in 3D), or in general the unit n-dimensional parallelepiped.

• •|det(A)| tells you how the n-dimensional volume scales.
• •If |det(A)| = 3, volumes become 3× larger.
• •If |det(A)| = 1/2, volumes shrink by factor 1/2.
• •sign(det(A)) tells you whether orientation is preserved or reversed.
• •det(A) < 0 means there is a flip (like a reflection composed with some stretching).

The key singularity test

A matrix A is singular (non-invertible) exactly when det(A) = 0.

Intuition: if A collapses at least one dimension, the n-dimensional volume must collapse to 0.

Formally we’ll connect this to row dependence:

• •det(A) = 0 ⇔ rows (or columns) are linearly dependent ⇔ A is not invertible.

Small concrete example (2×2)

For A = [[a, b], [c, d]],

det(A) = ad − bc.

This number is the signed area scaling factor in 2D.

• •If A = [[2, 0], [0, 3]], then det(A) = 6: it stretches x by 2 and y by 3, so areas scale by 2·3.
• •If A = [[1, 0], [0, −1]], then det(A) = −1: it reflects across the x-axis; area magnitude preserved, orientation flipped.

Takeaway: determinants are not merely a computational trick—they are the algebraic encoding of how linear maps scale and orient space.

Why row operations matter

In practice, you often compute determinants by simplifying a matrix via row operations (the same moves used in Gaussian elimination). The determinant is designed to react predictably to these moves.

This gives you a powerful workflow:

1) Reduce A to an upper triangular form U.

2) Adjust for the row operations you used.

3) Use det(U) = product of diagonal entries.

The three fundamental row operations

Below is the rule set you should memorize—because it turns determinant computation into bookkeeping.

A subtle but important note: “Add a multiple of one row to another” keeps determinant unchanged, but it can drastically change entries—this is the move that makes elimination so useful.

Triangular matrices: the determinant shortcut

For an upper triangular matrix U (everything below diagonal is 0),

det(U) = ∏ᵢ uᵢᵢ.

Same for lower triangular.

Why this is true (idea): a triangular matrix maps basis vectors in a way that stacks scaling along diagonal directions without mixing that creates additional volume contributions.

Worked derivation idea (with multilinearity and alternating)

To see why some rules make sense, consider two quick consequences of the axioms:

(1) If two rows are equal, det = 0

Swap the two equal rows:

• •By alternating: det(A) becomes −det(A)
• •But the matrix is unchanged (rows equal)

So:

det(A) = −det(A)

⇒ 2 det(A) = 0

⇒ det(A) = 0.

(2) Adding a multiple of one row to another doesn’t change det

Let rows be r₁, …, rⱼ, …, rᵢ, …, rₙ. Replace rⱼ with rⱼ + k rᵢ.

By multilinearity in row j:

det(r₁,…, rⱼ + k rᵢ, …, rₙ)

= det(r₁,…, rⱼ, …, rₙ) + k det(r₁,…, rᵢ, …, rₙ)

But the second determinant has rᵢ appearing twice (once as row i and once in row j’s slot). Two equal rows ⇒ determinant 0. Hence:

det(new) = det(old) + k·0 = det(old).

This is the algebraic reason elimination works so cleanly.

Computing det(A) via elimination (the safe version)

A common reliable method:

• •Use only row-addition operations (don’t change determinant).
• •If you must swap rows, track a sign flip.
• •If you scale a row, track the scaling factor.

At the end, you reach triangular U, then use product of diagonal.

The determinant and invertibility (connection to elimination)

Gaussian elimination tells you:

• •A is invertible ⇔ you can reduce it to I without ever creating a zero pivot.
• •A is singular ⇔ you hit a zero pivot (a row becomes dependent).

Determinant encodes this in one number:

• •det(A) = 0 ⇔ some elimination step forces a zero pivot (after appropriate swaps) ⇔ rank < n.

So: elimination gives you a procedural test; determinant gives you a scalar certificate.

Why formulas exist at all

The axioms define det(A), but to compute it directly from entries, we need formulas.

For 2×2 and 3×3, explicit closed forms are convenient. For larger n, direct formulas exist (cofactor expansion / Leibniz formula), but computationally they become expensive; elimination is typically preferred.

2×2 formula

For A = [[a, b], [c, d]]:

det(A) = ad − bc.

Geometric sanity checks:

• •If rows are multiples (e.g., [a, b] = t[c, d]), then ad − bc = 0.
• •If A is diagonal diag(α, β), det(A) = αβ.

3×3 formula (cofactor expansion along a row)

Let

A =

[ [a, b, c],

[d, e, f],

[g, h, i] ].

One common expression is:

det(A) = a(ei − fh) − b(di − fg) + c(dh − eg).

Notice the alternating + − + pattern, which comes from the “alternating” nature of det.

Cofactors and minors (the general pattern)

To expand det(A) along row r (or column c), you use minors and cofactors.

• •The minor Mᵣc is the (n−1)×(n−1) matrix you get by deleting row r and column c.
• •The cofactor Cᵣc is:

Cᵣc = (−1)^(r+c) det(Mᵣc).

Then cofactor expansion along row r is:

det(A) = ∑_{c=1 to n} aᵣc Cᵣc.

Similarly, expansion along column c:

det(A) = ∑_{r=1 to n} aᵣc Cᵣc.

Why the sign pattern (−1)^(r+c)?

Swapping rows/columns changes sign. When you “move” an element aᵣc to the top-left conceptually, you perform (r−1) row swaps and (c−1) column swaps, totaling (r+c−2) swaps. Parity of swaps controls the sign, producing the checkerboard pattern.

When cofactor expansion is useful

Cofactor expansion is rarely the fastest general method, but it shines when:

• •A row/column has many zeros (sparse).
• •You’re doing symbolic determinants (with variables) and want structure.

A practical rule:

• •If n ≥ 4 and the matrix isn’t very sparse, prefer elimination.

Determinant as a sum over permutations (conceptual, not computational)

There is also the Leibniz formula:

det(A) = ∑_{σ ∈ Sₙ} sgn(σ) ∏_{i=1 to n} a_{i, σ(i)}.

This makes the “alternating” behavior explicit: each permutation contributes with a sign depending on whether it’s an even or odd permutation.

Why mention it?

• •It explains why det is multilinear and alternating.
• •It connects determinants to combinatorics and to proofs like det(AB) = det(A)det(B).

But computing with it directly costs O(n!) terms—prohibitively large.

Multiplicativity (a central property)

A major theorem:

det(AB) = det(A) det(B).

Interpretation:

• •Apply B: volumes scale by det(B).
• •Then apply A: volumes scale by det(A).
• •Net scaling is the product.

A powerful corollary:

• •A invertible ⇔ det(A) ≠ 0.

Because if A is invertible, AA⁻¹ = I.

Take determinants:

det(A) det(A⁻¹) = det(I) = 1

⇒ det(A) ≠ 0.

And conversely, if det(A) ≠ 0, A must be invertible (can be shown via elimination / rank).

Determinant and transpose

Another helpful identity:

det(Aᵀ) = det(A).

So you can work with rows or columns interchangeably.

Determinants as a geometric “health check”

In applied linear algebra, det(A) answers fast questions:

• •Near singular? If det(A) is close to 0, the transformation nearly collapses volume, and solving Ax = b may be ill-conditioned.
• •Orientation flip? If det(A) < 0, a handedness flip occurred.

Important nuance: “det close to 0” is not always a reliable numerical indicator of ill-conditioning by itself (scaling matters), but conceptually it’s a key signal.

Connection to eigenvalues

If λ₁, …, λₙ are the eigenvalues of A (counted with algebraic multiplicity), then:

det(A) = ∏ᵢ λᵢ.

Why this matters:

• •If any eigenvalue is 0, the product is 0 ⇒ det(A) = 0 ⇒ A is singular.
• •The sign/magnitude of det relates to how eigenvalues scale space in their eigen-directions.

A related identity uses the characteristic polynomial:

p(λ) = det(A − λI).

Roots of p(λ) are eigenvalues. Determinants are literally the engine behind the eigenvalue equation.

Connection to LU / QR decompositions

Many matrix decompositions expose the determinant cheaply.

LU decomposition (with possible permutation P):

PA = LU

where L is lower triangular with 1s on diagonal (often), and U is upper triangular.

Taking det:

det(PA) = det(L) det(U)

det(P) det(A) = det(L) det(U)

So:

det(A) = det(P)⁻¹ det(L) det(U).

Since det(L) is usually 1 (if L has unit diagonal), and det(U) is product of U’s diagonal entries, determinants become almost free once LU is computed.

QR decomposition:

A = QR

where Q is orthonormal (QᵀQ = I) and R is upper triangular.

Then:

det(A) = det(Q) det(R).

Key facts:

• •det(Q) is ±1 (orthonormal transforms preserve volume; they may flip orientation).
• •det(R) is product of diagonal entries.

Determinant and change of variables (multivariable calculus)

If a differentiable map has Jacobian matrix J at a point, then |det(J)| gives the local volume scaling factor. This is why determinants appear in integrals under substitution.

Even if you haven’t studied Jacobians yet, keep the mental model:

• •determinant = local “volume magnifier” (signed).

Determinants in solving systems (Cramer’s Rule)

There is a formula for solving Ax = b using determinants:

xᵢ = det(Aᵢ) / det(A)

where Aᵢ is A with column i replaced by b.

This is elegant theoretically, but computationally inefficient for large systems. Still, it reinforces:

• •det(A) ≠ 0 is the condition for a unique solution.

Big picture summary

Determinants sit at a crossroads:

• •Geometry (volume/orientation)
• •Algebra (invertibility, rank)
• •Computation (elimination, LU/QR)
• •Spectral theory (eigenvalues)

Once you can compute and reason about det(A), the next nodes—eigenvalues/eigenvectors and decompositions—feel much more motivated and connected.