Eigenvalues: The DNA of Matrices

Watch a Matrix Reveal Its Secret

Take the matrix A = [[2, 1], [1, 2]]. Pick any vector you like — say v₀ = (1, 0) — and multiply it by A, over and over, feeding each answer back in as the next input. Most vectors, when you do this, wander around, changing direction unpredictably. But something strange happens here. Track the direction of the vector at each step (compare how much bigger the y-coordinate is relative to the x-coordinate):

A = [[2, 1],
     [1, 2]]

def mat_vec(A, v):
    return [A[0][0]*v[0] + A[0][1]*v[1],
            A[1][0]*v[0] + A[1][1]*v[1]]

v = [1, 0]
for step in range(5):
    print(step, v)
    v = mat_vec(A, v)

Tracing this by hand: v₀ = (1, 0). Multiply: (2·1+1·0, 1·1+2·0) = (2, 1). Multiply again: (2·2+1·1, 1·2+2·1) = (5, 4). Again: (2·5+1·4, 1·5+2·4) = (14, 13). Again: (2·14+1·13, 1·14+2·13) = (41, 40). The code's output is exactly:

0 [1, 0]
1 [2, 1]
2 [5, 4]
3 [14, 13]
4 [41, 40]

Look at the ratio of the two coordinates at each step: 0/1 = 0, then 1/2 = 0.5, then 4/5 = 0.8, then 13/14 ≈ 0.929, then 40/41 ≈ 0.976. That ratio is climbing steadily toward 1 — meaning the vector's direction is being dragged, step by step, toward the 45° line where x = y. No matter which starting vector you try (as long as it isn't a very specific "wrong" choice), repeated multiplication by this particular A always steers the direction toward the same line: the direction of (1, 1).

Why does every vector get pulled toward this one direction, and why that direction specifically? That is the question this chapter answers. The direction (1, 1) is special for A in a way that has nothing to do with which starting vector you picked — it is baked into the matrix itself, the way a strand of DNA carries information independent of which cell reads it. That special direction, and the number that describes exactly how much A stretches it, are called an eigenvector and its eigenvalue.

Refresher: Multiplying a Matrix by a Vector

Before going further, pin down the mechanics. For a 2×2 matrix A = [[a, b], [c, d]] and a vector v = (x, y) written as a column, the product Av is computed row by row: the first entry of Av is a·x + b·y, and the second entry is c·x + d·y. That is all matrix-vector multiplication is — each row of the matrix "reads" the vector and produces one number.

You will also need the determinant of a 2×2 matrix: for M = [[a, b], [c, d]], det(M) = ad − bc. This single number tells you whether M is invertible (det(M) ≠ 0, meaning Mv = 0 only when v = 0) or not (det(M) = 0, meaning some nonzero vector gets crushed to the origin). That fact — that a zero determinant is precisely the signature of a matrix collapsing some nonzero vector to zero — is the hinge on which this entire chapter turns.

The Formal Definition

A nonzero vector v is called an eigenvector of a square matrix A if multiplying by A does not change v's direction — it only scales v by some number λ (lambda):

A v = λ v

The number λ is called the eigenvalue associated with that eigenvector. Geometrically: most vectors, when hit by A, both rotate and stretch. An eigenvector is a survivor — it refuses to rotate. It only gets longer, shorter, or flipped to point the opposite way. The word "eigen" is German for "own" or "characteristic" — these are the directions a matrix keeps for itself, unaffected by the transformation's rotational part.

Deriving the Characteristic Equation

The definition Av = λv is not yet something you can solve, because both v and λ are unknown, and v appears on both sides. The standard trick is to bring everything to one side:

Av = λv
Av − λv = 0
Av − λ(Iv) = 0        (since Iv = v, where I is the identity matrix)
(A − λI) v = 0

This is now a system where a matrix, (A − λI), multiplies a vector v and produces the zero vector. There is always one boring solution: v = 0. But v = 0 is not an eigenvector by definition (eigenvectors must be nonzero) — a stationary point isn't a "direction." For a nontrivial solution (some v ≠ 0) to exist, the matrix (A − λI) must fail to be invertible. Why? Because if (A − λI) were invertible, you could multiply both sides of (A − λI)v = 0 by its inverse and get v = (A − λI)⁻¹·0 = 0 — forcing v to be zero, which is exactly the case we want to rule out. And a matrix fails to be invertible precisely when its determinant is zero. That gives the single most important equation in this chapter, the characteristic equation:

det(A − λI) = 0

Every eigenvalue of A is a root of this equation, and nothing else is.

A Free Shortcut From Vieta's Formulas

For a general 2×2 matrix A = [[a, b], [c, d]], write out det(A − λI) directly:

A − λI = [[a−λ,  b   ],
          [c,    d−λ ]]

det(A − λI) = (a−λ)(d−λ) − bc
            = λ² − (a+d)λ + (ad−bc)

Compare this to the standard quadratic λ² − (sum of roots)λ + (product of roots), which you already know from solving quadratics with Vieta's formulas. Matching term by term:

• The sum of the two eigenvalues equals a + d — the sum of the diagonal entries of A, called the trace, written tr(A).
• The product of the two eigenvalues equals ad − bc — exactly the determinant of A.

This is not a coincidence to memorize separately — it falls straight out of Vieta's formulas applied to the characteristic polynomial. It is also an excellent error-check: once you compute two eigenvalues, add them and confirm you get tr(A); multiply them and confirm you get det(A). Competitive exams reward exactly this kind of quick self-verification.

Worked Example 1: Full Solve

Return to A = [[2, 1], [1, 2]] from the opening. tr(A) = 2 + 2 = 4, det(A) = 2·2 − 1·1 = 3. The characteristic equation is λ² − 4λ + 3 = 0, which factors as (λ − 1)(λ − 3) = 0. So the eigenvalues are λ₁ = 1 and λ₂ = 3. Check: 1 + 3 = 4 = tr(A) ✓, and 1 × 3 = 3 = det(A) ✓.

Now find the eigenvector for each λ by substituting back into (A − λI)v = 0.

For λ = 3: A − 3I = [[2−3, 1], [1, 2−3]] = [[−1, 1], [1, −1]]. Solving [[−1, 1], [1, −1]]·(x, y) = (0, 0) gives −x + y = 0, i.e. y = x. Any vector of the form (t, t) works; the simplest is (1, 1). Verify directly: A(1,1) = (2·1+1·1, 1·1+2·1) = (3, 3) = 3·(1, 1). ✓ This is exactly the direction the iteration in the hook converged toward — because repeatedly multiplying by A stretches the (1,1) direction by a factor of 3 every single step, it eventually dominates every other direction, no matter where you start.

For λ = 1: A − I = [[1, 1], [1, 1]]. Solving gives x + y = 0, i.e. y = −x. The eigenvector is (1, −1). Verify: A(1,−1) = (2·1+1·(−1), 1·1+2·(−1)) = (1, −1) = 1·(1, −1). ✓ This direction is untouched by A at all — it neither stretches nor shrinks, since its eigenvalue is exactly 1.

Misconception Check: "Doesn't an Eigenvector Have to Stay Exactly the Same?"

A common mistake is to think an eigenvalue must be positive, or that the eigenvector must be visually "unchanged." Neither is true — λ can be negative, meaning the vector flips to point exactly opposite, and it can even be zero. Consider the reflection matrix R = [[0, 1], [1, 0]], which reflects every point across the line y = x. Here tr(R) = 0, det(R) = 0·0 − 1·1 = −1. The characteristic equation is λ² − 0λ − 1 = 0, so λ² = 1, giving λ = 1 and λ = −1.

For λ = 1: R − I = [[−1, 1], [1, −1]], giving y = x, eigenvector (1, 1). Check: R(1,1) = (1, 1) = 1·(1,1). This makes sense — points already sitting on the mirror line y = x don't move when reflected.

For λ = −1: R + I = [[1, 1], [1, 1]], giving y = −x, eigenvector (1, −1). Check: R(1,−1) = (−1, 1) = −1·(1, −1). This also makes sense — a point on the line perpendicular to the mirror gets reflected straight through the origin to the opposite side, which is exactly what multiplying by −1 does. So a negative eigenvalue is not an exception to the rule; it is the rule working correctly on a direction that gets flipped rather than merely stretched.

When a Matrix Has No Real Eigenvectors At All

A second misconception is assuming every square matrix must have real eigenvalues. Consider the 90° rotation matrix S = [[0, −1], [1, 0]], which rotates every vector in the plane by 90° counterclockwise. tr(S) = 0, det(S) = 0·0 − (−1)·1 = 1. The characteristic equation is λ² + 1 = 0, so λ² = −1 — there is no real number whose square is −1. The eigenvalues are the complex numbers λ = i and λ = −i.

This should actually feel obvious once you think about what an eigenvector would mean here: an eigenvector of S would be a direction that rotation by 90° leaves unchanged. But rotating anything by exactly 90° always changes its direction — no real, nonzero vector in the plane survives a 90° turn pointing the same way (or exactly opposite). So the absence of a real eigenvector isn't a flaw in the theory; the algebra (complex λ) is faithfully reporting a true geometric fact (no real fixed direction exists). This is why eigenvalues, in general, must be allowed to be complex numbers — the definition Av = λv is incomplete if you restrict λ to real numbers only.

A Free Shortcut: Triangular Matrices

Extend to 3×3 with one especially useful case. Let B = [[2, 1, 0], [0, 3, 4], [0, 0, 5]] — an "upper triangular" matrix, with zeros everywhere below the main diagonal. Then B − λI = [[2−λ, 1, 0], [0, 3−λ, 4], [0, 0, 5−λ]] is also upper triangular. The determinant of any triangular matrix is simply the product of its diagonal entries (expand along the first column repeatedly and every off-diagonal cofactor vanishes because of the zeros). So det(B − λI) = (2−λ)(3−λ)(5−λ) = 0 immediately, giving λ = 2, 3, 5 — you can read the eigenvalues straight off the diagonal, with no expansion needed. This shortcut is worth memorizing: for any triangular (or diagonal) matrix, the eigenvalues are exactly its diagonal entries.

Misconception Check: Is the Eigenvector Unique?

In Worked Example 1, the eigenvector for λ = 3 was written as (1, 1) — but (2, 2), (5, 5), and (−3, −3) all satisfy A(t,t) = 3(t,t) equally well, for any nonzero t. An eigenvector is never a single unique vector; it is really an entire line through the origin (called the eigenspace for that λ), and any nonzero vector on that line is a valid eigenvector. When a textbook or an exam gives "the" eigenvector as (1, 1), it means the simplest representative of that whole family, not the only correct answer. If you compute (2, 2) instead of (1, 1) on an exam, you have not made an error.

The Payoff: Cayley–Hamilton and Why JEE Advanced Cares

The CBSE, JEE, and BITSAT matrices syllabus stops short of naming "eigenvalues" explicitly — but it does not stop short of the characteristic equation itself, which appears in JEE Advanced under the Cayley–Hamilton theorem: every square matrix satisfies its own characteristic equation. For a 2×2 matrix A, this means substituting the matrix A itself in place of λ in λ² − tr(A)λ + det(A) = 0 produces the zero matrix:

A² − tr(A)·A + det(A)·I = 0

Verify this on A = [[2, 1], [1, 2]] from Worked Example 1, where tr(A) = 4 and det(A) = 3. First compute A²:

A² = A·A = [[2·2+1·1, 2·1+1·2], [1·2+2·1, 1·1+2·2]] = [[5, 4], [4, 5]]

Then A² − 4A + 3I:

[[5,4],[4,5]] − [[8,4],[4,8]] + [[3,0],[0,3]]
= [[5−8+3, 4−4+0], [4−4+0, 5−8+3]]
= [[0, 0], [0, 0]]

It works — exactly zero, as the theorem guarantees. JEE Advanced regularly asks you to find A⁻¹ or a matrix expression using this identity instead of computing the inverse directly by cofactors, because once you know A² = 4A − 3I, you can rearrange to A·(A − 4I) = −3I, giving A⁻¹ = −(A − 4I)/3 in one line, without ever computing a single cofactor. This is the direct, practical, exam-relevant payoff of everything derived above.

Why "DNA"? Diagonalization and the Fibonacci Sequence

The title of this chapter calls eigenvalues the "DNA" of a matrix because of what they let you do with powers of a matrix — computing A, A², A³, ... A¹⁰⁰ — without doing a hundred matrix multiplications. If a matrix A can be written as A = PDP⁻¹, where D is a diagonal matrix holding A's eigenvalues, then Aⁿ = PDⁿP⁻¹, and raising a diagonal matrix to the nth power just means raising each diagonal entry to the nth power individually — no repeated matrix multiplication required.

Here is a genuinely striking use of this fact. Define the matrix M = [[1, 1], [1, 0]]. Compute Mⁿ for small n and something remarkable appears:

M¹ = [[1,1],[1,0]]
M² = [[2,1],[1,1]]
M³ = [[3,2],[2,1]]

Every entry is a Fibonacci number. In fact Mⁿ = [[F(n+1), F(n)], [F(n), F(n−1)]] exactly, where F(0)=0, F(1)=1, F(2)=1, F(3)=2, F(4)=3, F(5)=5, ... . So finding the 100th Fibonacci number is really a question about the 100th power of M — and eigenvalues are exactly the tool that makes that power fast to compute.

Find M's eigenvalues: tr(M) = 1, det(M) = 1·0 − 1·1 = −1. Characteristic equation: λ² − λ − 1 = 0. By the quadratic formula, λ = (1 ± √5)/2. These two numbers are famous in their own right: φ = (1+√5)/2 ≈ 1.6180 is the golden ratio, and ψ = (1−√5)/2 ≈ −0.6180 is its conjugate. Once M is diagonalized using φ and ψ, the formula for Mⁿ reduces (after working through P and P⁻¹, which is beyond this chapter's scope but is a standard next step) to Binet's formula:

F(n) = (φⁿ − ψⁿ) / √5

Check it for n = 5: φ⁵ ≈ 11.0902, ψ⁵ ≈ −0.0902 (both computed by repeated squaring/multiplying: φ²≈2.6180, φ³≈4.2361, φ⁴≈6.8541, φ⁵≈11.0902, and similarly for ψ). So (φ⁵ − ψ⁵)/√5 ≈ (11.0902 − (−0.0902))/2.2361 ≈ 11.1804/2.2361 ≈ 5.000 — exactly F(5) = 5. Two irrational numbers, raised to the 5th power and combined, land exactly on an integer. This is not a coincidence; it is a direct consequence of φ and ψ being the eigenvalues that structure how M grows. This is what "eigenvalues are the DNA of a matrix" means precisely: they are the two numbers that completely control how the matrix behaves no matter how many times you apply it, encoded once and read out repeatedly.

Where This Shows Up Beyond the Exam

Structural and aerospace engineers use eigenvalues of a structure's stiffness matrix to find its natural vibration frequencies — the same computation done here, just on a much larger matrix. Every launch vehicle and building in India is checked this way, because if an external force (wind, an earthquake, engine vibration) matches one of these natural frequencies, the structure resonates and the oscillation amplifies dangerously rather than settling down; keeping the eigenvalues of a design away from expected forcing frequencies is standard practice in structural safety analysis. In data science, the eigenvectors of a dataset's covariance matrix (used in Principal Component Analysis) point along the directions where the data varies the most, and the corresponding eigenvalues measure how much variance lies along each direction — this is how large datasets get compressed to their most informative few dimensions. And the dominant eigenvector of the web's link-structure matrix is, essentially, what an early version of Google's PageRank algorithm computed to rank pages.

Test Yourself

1. Find the eigenvalues of C = [[4, 2], [1, 3]] using trace and determinant, then verify by finding both eigenvectors. (Answer: tr = 7, det = 10, so λ² − 7λ + 10 = 0 → λ = 2, 5. For λ=5: eigenvector (2,1). For λ=2: eigenvector (1,−1).)
2. Without expanding a determinant, state the eigenvalues of D = [[−1, 0, 0], [3, 4, 0], [7, 8, 6]]. (Answer: D is lower triangular, so eigenvalues are the diagonal entries: −1, 4, 6.)
3. A 2×2 matrix has trace 6 and determinant 9. What are its eigenvalues, and what's special about them? (Answer: λ² − 6λ + 9 = 0 → (λ−3)² = 0 → λ = 3, 3 — a repeated eigenvalue.)
4. Explain, without computing anything, why a matrix with det(A) = 0 must have λ = 0 as one of its eigenvalues. (Answer: product of eigenvalues = det(A); if det(A) = 0, at least one eigenvalue must be 0.)
5. Using Cayley–Hamilton, if a 2×2 matrix E has tr(E) = 5 and det(E) = 6, write an expression for E⁻¹ in terms of E and I. (Answer: E² − 5E + 6I = 0 → E(E−5I) = −6I → E⁻¹ = −(E−5I)/6 = (5I−E)/6.)

Summary

An eigenvector of a square matrix A is a nonzero direction that A does not rotate — it only scales it by a factor λ, its eigenvalue, following Av = λv. Rearranging gives (A − λI)v = 0, and since v must be nonzero, (A − λI) cannot be invertible, forcing det(A − λI) = 0 — the characteristic equation, whose roots are precisely the eigenvalues. For a 2×2 matrix, the sum of the eigenvalues always equals the trace and the product always equals the determinant, a direct consequence of Vieta's formulas. Eigenvalues can be negative (the eigenvector flips direction), zero (the eigenvector gets crushed to the origin), or complex (no real fixed direction exists, as with pure rotations); and an eigenvector is never unique — it represents an entire line of directions, the eigenspace. Triangular matrices hand you their eigenvalues for free, as their diagonal entries. Every matrix satisfies its own characteristic equation (Cayley–Hamilton), a fact JEE Advanced uses to compute inverses and powers efficiently. And because a matrix can be rewritten in terms of its eigenvalues (diagonalization), those eigenvalues completely determine how the matrix behaves under repeated application — from computing Fibonacci numbers via the golden ratio to predicting whether a bridge will resonate. That is the sense in which eigenvalues are a matrix's DNA: a compact set of numbers that, once known, tell you everything about how the matrix will act, no matter how many times it acts.