Chapter 14 · Introduction

Chapter 14 — Quadratic Forms — Reduction & Classification

A quadratic form is a symmetric matrix in disguise: complete squares or rotate to principal axes, and it becomes a sum of signed squares whose numbers of plus, minus and zero signs classify it completely.

Is the point where a function of two variables has zero gradient a minimum, a maximum or a saddle? In 1759 Joseph-Louis Lagrange, studying maxima and minima of functions of several variables, reduced the question to the second-order terms of the Taylor expansion — an expression such as $ax^2+2hxy+by^2$ — and settled it by completing squares. Seventy years later Augustin-Louis Cauchy, looking for the principal axes of quadric surfaces, proved that a real symmetric matrix has only real eigenvalues, so that a rotation of the coordinates removes every cross term. And in 1852 James Joseph Sylvester proved, and named, the law of inertia (Carl Jacobi had found it too): however you reduce a real quadratic form to a sum of squares $d_1y_1^2+\dots+d_ny_n^2$, the numbers of positive, negative and zero coefficients are always the same.

Those three numbers — the inertia $(n_+,n_-,n_0)$, or equivalently the rank and the signature — are all there is to know about a real quadratic form up to a change of variables. They decide whether a form is positive definite, which is what the Hessian test of §7.2 really asks; they say whether $5x^2-4xy+8y^2=1$ is an ellipse, a hyperbola or a pair of lines, and which central quadric — ellipsoid, hyperboloid of one or two sheets, cylinder — a surface $x^TAx=1$ is; they tell you whether two symmetric matrices are congruent and how many congruence classes there are. Quadratic forms are also the energy of a vibrating system and the variance $a^T\Sigma a$ of a combination of random variables, and in every case positive definiteness is the property that matters. This one-section chapter puts the spectral theorem of Chapter 13 to work, alongside the quick, exact arithmetic of Lagrange's method and the determinant test of Sylvester's criterion.

Bilinear forms→Symmetric matrix of $q$→Congruence $P^TAP$→Lagrange's reduction→Principal axes→Sylvester's law of inertia→Rank, index, signature→Sylvester's criterion→Conics & quadrics
🎯 By the end of this chapter you can…
  • pass between a bilinear or quadratic form and its symmetric matrix, and see how a change of variables $x=Py$ turns $A$ into $P^TAP$ (congruence, not similarity);
  • reduce a form to a sum of signed squares by Lagrange's method — including forms with no square terms, such as $2xy$ — and by an orthogonal change of variables to principal axes;
  • compute the rank, index and signature of a form, use Sylvester's law of inertia to decide whether two real symmetric matrices are congruent, and count congruence classes over $\mathbb{R}$ and over $\mathbb{C}$;
  • test positive and negative (semi)definiteness by eigenvalues, by leading principal minors (Sylvester's criterion) and by Jacobi's sign-change rule, and know why leading minors alone cannot detect semidefiniteness;
  • classify plane conics and central quadrics from their equations, including equations with linear terms and one-parameter families.
Teaser — A Quadratic Form on the Unit Circle

A unit vector $x$ runs round the circle (left) while the right panel plots $q(x)=x^TAx$ against its angle. The circle is coloured red where $q\gt0$ and blue where $q\lt0$; the purple curves are the level sets $q=1$ (solid) and $q=-1$ (dashed), the green dashed lines are the principal axes, and the amber lines are the directions where $q=0$. Watch where the graph reaches its highest and lowest values, and how the signs alone decide the shape of $q=1$.

x—
q(x)—
eigenvalues of A—
inertia (n₊, n₋, n₀)—
rank, signature—
type—

How this is examined

Quadratic forms appear in almost every CSIR NET paper, usually as quick but trap-laden questions. Part B asks for the matrix of a given form, its rank, index or signature, the number of positive eigenvalues of a symmetric matrix, the range of a parameter for which a form is positive definite, the type of conic or quadric defined by a second-degree equation, or the number of congruence classes of real symmetric $n\times n$ matrices (of all ranks, or of a given rank). Part C, where several options can be correct, asks you to judge statements: "two real symmetric matrices with the same rank and signature are congruent", "congruent matrices are similar", "a positive definite matrix has positive diagonal entries", "if all leading principal minors are $\ge0$ the form is positive semidefinite", "over $\mathbb{C}$ congruence is decided by the rank alone", "the form is positive definite if and only if all its eigenvalues are positive". The false options are exactly the counterexamples in the ⚠️ box of §14.1, and the fastest tools are the ones the teaser shows: the signs of the eigenvalues, or the signs left after completing squares. GATE MA asks the same material, often combined with the Hessian test for extrema.

Sections in this chapter

✅ Before you start

This chapter rests on Chapter 13. You need the spectral theorem for real symmetric matrices and the definition and characterisations of positive (semi)definite matrices (§13.2), and orthonormal bases and orthogonal matrices (§13.1). From earlier chapters you need eigenvalues and the characteristic polynomial (§11.1), determinants, minors and rank (§10.1), and change of basis for linear maps (§9.2) — so that you can see why a form changes by $P^TAP$ and an operator by $P^{-1}AP$. The Hessian test for local extrema (§7.2) is the main application in analysis. Completing the square in one variable and the standard equations of the ellipse, hyperbola and parabola are assumed.