A matrix like A = \begin{bmatrix} 2 & 1 \\ 1 & 2 \end{bmatrix} looks genuinely tangled: multiply it into a vector and both coordinates get mixed together. But put on the right glasses — view the plane using A's own eigenvectors as the coordinate axes instead of the usual x- and y-axes — and the tangle disappears. In that special basis, A does nothing but stretch each axis by its own eigenvalue. No mixing, no rotation, just plain multiplication along two independent directions.

That change of glasses is called diagonalization, and it is written as a single compact equation:

where the columns of P are the eigenvectors and D is the diagonal matrix of the matching eigenvalues. Read the right-hand side right to left: P^{-1} rewrites a vector in eigen-coordinates, D stretches along each eigen-axis, and P translates the result back into ordinary coordinates. This isn't just a tidy formula — it's the key that turns "apply this matrix a hundred times" into "raise two numbers to the hundredth power," as the payoff below shows.

Built from one idea, repeated for every eigenvector at once

The whole recipe falls out of a single fact you already know: apply A to one of its own eigenvectors, and it doesn't rotate or twist — it just scales, A\vec{v} = \lambda\vec{v}. Now do that for every eigenvector at once, packing them side by side as the columns of a matrix P. Applying A to each column separately is exactly the same as multiplying A by the whole matrix P:

Each column just picked up its own scale factor — and multiplying a matrix by a diagonal matrix on the right is exactly "scale column i by D's i-th entry." That's precisely what PD means. So AP = PD. As long as P is invertible — which it is exactly when the eigenvectors are independent — multiply both sides on the right by P^{-1} and out pops A = PDP^{-1}. No new ideas were needed: just "eigenvectors scale," applied to every eigen-direction simultaneously.

A stretch along its own axes

The faint grid below is aligned to the eigenvectors of A = \begin{bmatrix} 2 & 1 \\ 1 & 2 \end{bmatrix} — the very matrix we'll diagonalize by hand next. Drag the slider to apply A: the grid doesn't rotate or shear, it simply stretches — ×3 along one eigen-axis, ×1 (unchanged) along the other. Seen through the eigenvectors' eyes, this matrix was never complicated at all.

Worked example: diagonalize a 2×2 matrix, start to finish

Take A = \begin{bmatrix} 2 & 1 \\ 1 & 2 \end{bmatrix} and walk the whole recipe through by hand.

Step 1 — eigenvalues from the characteristic equation. Solve \det(A - \lambda I) = 0:

So \lambda_1 = 3 and \lambda_2 = 1.

Step 2 — eigenvectors via the null-space method. For each \lambda, solve (A - \lambda I)\vec{v} = \vec{0}.

For \lambda_1 = 3: \begin{bmatrix} -1 & 1 \\ 1 & -1 \end{bmatrix}\vec{v} = \vec{0} gives v_2 = v_1, so \vec{v}_1 = \begin{bmatrix}1\\1\end{bmatrix}.

For \lambda_2 = 1: \begin{bmatrix} 1 & 1 \\ 1 & 1 \end{bmatrix}\vec{v} = \vec{0} gives v_2 = -v_1, so \vec{v}_2 = \begin{bmatrix}1\\-1\end{bmatrix}.

Step 3 — assemble P and D, eigenvectors as columns of P, eigenvalues in the matching order along D:

Inverting the 2×2 matrix P (determinant -2) gives P^{-1} = \begin{bmatrix} 1/2 & 1/2 \\ 1/2 & -1/2 \end{bmatrix}.

Step 4 — verify. Multiply it back out:

It really does come back to A. Every step was ordinary matrix multiplication — the cleverness was all in choosing P and D correctly in the first place.

The payoff: A^100 in three easy pieces

Here is why anyone bothers with all of this. Suppose you need A^5 for the same matrix. Multiplying A by itself five times by hand is tedious and the numbers grow messy fast. Diagonalization sidesteps all of it:

because every inner P^{-1}P pair cancels to the identity, leaving only the outer P and P^{-1} and a chain of D's. And a diagonal matrix to a power is trivial — just raise each entry:

Multiplying back out gives the exact answer with no repeated matrix-by-matrix grinding:

(Quick sanity check: the trace 122+122=244 matches 3^5+1^5, and the determinant 122^2-121^2=243 matches 3^5\cdot1^5. It checks out.)

Now imagine wanting A^{100} instead of A^5. Multiplying the matrix by itself, entry by entry, ninety-nine more times (or even using repeated squaring) is a real computation with numbers exploding in size at every step. Diagonalized, nothing changes in difficulty at all: D^{100} just needs 3^{100} and 1^{100} — two numbers raised to a power, done once — sandwiched between the same fixed P and P^{-1} as before. That single trick evaluates matrix powers instantly, solves systems of differential equations, computes the long-run behaviour of Markov processes, and sets up the decompositions that data science leans on. Eigenvectors turn "apply this matrix a hundred times" into "raise two numbers to the hundredth power."