The principal axis theorem
Completing the square diagonalizes by any invertible P, which distorts lengths and angles. The principal axis theorem says you can do better for a real symmetric form: there is an orthogonal P (a rotation/reflection, P^T = P^-1) with P^T A P diagonal. Because P is orthogonal, congruence and similarity coincide, so the diagonal entries are exactly the eigenvalues of A.
Classifying conics and quadrics
Now the geometric payoff. A general conic a x^2 + b xy + c y^2 + ... = 1 is q(x) plus linear terms; the principal axis theorem rotates away the cross-term, completing-the-square shifts away the linear terms, and the curve appears in standard position. The signature of the quadratic part then dictates the shape — no plotting required.
[[conic-classification]] in R^2 by the signature of the quadratic part: signature (2, 0) both eigenvalues same sign -> ELLIPSE signature (1, 1) eigenvalues opposite signs -> HYPERBOLA signature (1, 0) one zero eigenvalue -> PARABOLA (degenerate axis) Worked example: 5 x^2 + 4 xy + 5 y^2 = 9 A = [5, 2; 2, 5] eigenvalues 3 and 7 (both > 0) rotate to principal axes: 3 u^2 + 7 v^2 = 9 signature (2, 0) -> an ELLIPSE, semi-axes sqrt(9/3), sqrt(9/7).
The same machine runs in any dimension. In R^3 the classification of quadrics — ellipsoids, hyperboloids of one or two sheets, paraboloids, cones — is read off the signature of a 3x3 symmetric matrix. One invariant triple (p, q, z), computed once, names a surface you would otherwise struggle to picture.
Beyond the reals: Hermitian and symplectic forms
Over C, x^T A x can be complex and useless for measuring length, so we conjugate one slot: a Hermitian form is linear in one argument and conjugate-linear in the other, with A satisfying A* = A (conjugate transpose). Its values q(x) = x* A x are always real, it has a real signature, and its own spectral theorem — Hermitian matrices are unitarily diagonalizable with real eigenvalues. This is the form underlying quantum mechanics.
And the skew thread from Guide 2 reaches its destination: a nondegenerate alternating form is a symplectic form, and inertia has a startling analogue — all symplectic forms of a given even dimension are congruent. There is exactly one, with standard matrix [0, I; -I, 0]; there is no signature, no shape to classify, because there is only one shape. This is the rigid geometry beneath Hamiltonian dynamics.