Suppose someone hands you a brand-new matrix A and asks: "what are its eigenvalues?" You can't just guess a vector and check whether A\vec{v} happens to land back on the same line — for anything bigger than a 2×2 matrix, random guessing would take forever. You need a systematic search tool that hunts down every eigenvalue on its own, without you first knowing what it is.

This is exactly the tool that sits behind Google ranking billions of web pages, engineers checking whether a bridge will resonate itself apart, and physicists reading off the allowed energy levels of a quantum system. In every one of those settings, someone needs to answer "what are this matrix's eigenvalues?" with no starting guess to work from — and the answer is always the same equation.

Start from the defining equation and rearrange it so every term is on one side:

Now ask: for which values of \lambda does this equation have a non-zero solution \vec{v}? If A - \lambda I were invertible, we could multiply both sides by its inverse and get \vec{v} = \vec{0} — the boring solution, not an eigenvector. So the only way to get an interesting, non-zero \vec{v} is for A - \lambda I to be non-invertible, which happens exactly when its determinant is zero:

This is the characteristic equation. It turns "find the eigenvalues" into "solve this one polynomial equation" — a completely mechanical search that never requires a lucky guess. For a 2×2 matrix A = \begin{bmatrix} a & b \\ c & d \end{bmatrix} it expands into a neat quadratic in \lambda, written using the trace T = a + d and the determinant D = ad - bc:

Worked example: hunting down both eigenvalues

Let A = \begin{bmatrix} 5 & 4 \\ 1 & 2 \end{bmatrix}. Build A - \lambda I by subtracting \lambda from each diagonal entry, then take its determinant and set the result to zero:

Multiply out the bracket and tidy up:

That factors cleanly — look for two numbers that add to 7 and multiply to 6, namely 1 and 6:

Two roots, two eigenvalues — and notice they land exactly on the trace/determinant shortcut below: 1 + 6 = 7 = T and 1 \times 6 = 6 = D. The characteristic equation found both of them without a single guess.

It's worth pausing on how much that one equation just handed you. Before writing it down, all we knew about A's eigenvalues was that some existed. After writing it down and solving a plain old quadratic — the same kind of equation taught in early algebra — we know both of them exactly, with a built-in double-check (the trace/determinant shortcut) thrown in for free.

Roots are eigenvalues

Plotted against \lambda, the characteristic polynomial is a parabola, and the eigenvalues are exactly where it crosses zero. Slide the trace and determinant and watch the roots slide too. Push the curve until it just kisses the axis (a repeated root — worked below) or lifts clear off it entirely (complex eigenvalues — also worked below, and the matrix turns out to be a pure rotation).

A handy shortcut

The quadratic hands you two tidy facts for free: the eigenvalues add up to the trace (\lambda_1 + \lambda_2 = T) and multiply to the determinant (\lambda_1\lambda_2 = D). So a quick glance at a 2×2 matrix often gives its eigenvalues by inspection — no need to expand the full determinant every time. With the eigenvalues in hand, the eigenvectors come next.

Beyond 2×2: the same idea, more work

Nothing about the derivation cared that A was 2×2 — subtract \lambda down the diagonal, take the determinant, set it to zero. The difference for a 3×3 matrix is just that the determinant of a 3×3 matrix takes more bookkeeping to expand (a sum of three 2×2 determinants along a row), and the result is a cubic in \lambda instead of a quadratic:

A cubic can still be factored by hand when the roots are nice whole numbers — you just have more terms to track. An n \times n matrix always produces a degree- n characteristic polynomial this same way; the idea never changes; only the amount of arithmetic does. (There's a real limit to how far "by hand" can go, though — see the vignette below.)

One family makes the 3×3 case genuinely easy: a triangular matrix (everything above or below the diagonal is zero) has its eigenvalues sitting in plain sight on the diagonal. Take A = \begin{bmatrix} 2 & 1 & 3 \\ 0 & 5 & 4 \\ 0 & 0 & 1 \end{bmatrix}. Subtracting \lambda from the diagonal keeps the matrix triangular, and the determinant of a triangular matrix is just the product of its diagonal entries:

No expansion along a row needed, no cubic to factor by trial and error — the off-diagonal entries (1, 3, 4) don't affect the eigenvalues at all for a triangular matrix. They will matter again once you go looking for the matching eigenvectors, though.

Worked example: a repeated eigenvalue

Not every matrix gives two different roots. Take A = \begin{bmatrix} 4 & 1 \\ 0 & 4 \end{bmatrix}:

The characteristic equation has a single double root, \lambda = 4, counted twice. On the chart above this is exactly the moment the parabola just grazes the axis at one point instead of slicing through it at two.

A repeated eigenvalue is a warning sign, not just a curiosity: it can mean the matrix has only one independent eigen-direction to offer instead of two, even though the equation "used up" both roots on the same number. (This particular matrix does — every eigenvector points along (1,0).) That shortage is exactly what later trips up diagonalization, which needs a full, independent set of eigen-directions to work with.

The trace/determinant shortcut spots a repeated root before you even factor anything: here T = 4 + 4 = 8 and D = 4 \times 4 = 16, and the quadratic formula's discriminant T^2 - 4D = 64 - 64 = 0 is exactly zero — the tell-tale sign of a double root, matching the parabola just grazing the axis on the chart above.

Worked example: no real eigenvalues at all

Some matrices don't cross the axis anywhere. Take the 90° rotation matrix A = \begin{bmatrix} 0 & -1 \\ 1 & 0 \end{bmatrix}:

No real number squares to -1, so this characteristic equation has no real roots — only the complex pair \lambda = \pm i. That makes perfect geometric sense: a 90° rotation swings every vector off its original line, so there genuinely is no real direction that "stays put" the way an eigenvector must. The algebra and the picture agree — when a matrix only rotates (no stretching along any real axis), its characteristic equation floats free of the \lambda-axis entirely.

The discriminant confirms it instantly: here T = 0 and D = 1, so T^2 - 4D = -4 — a negative discriminant, the signature of a parabola that never touches the axis and therefore a matrix with no real eigenvalues at all.