A change of basis can diagonalize a real symmetric form by congruence. If there are two forms, we need the same basis to handle both. A positive-definite first form supplies exactly the extra structure we need: it defines an inner product, and the second form becomes a self-adjoint operator for that inner product. This yields simultaneous congruence and the generalized eigenvalue equation.
The bilinear-form notes establish congruence and real inertia, and the inner-product notes establish self-adjoint spectral diagonalization and positive square roots. We combine those tools here to find the common basis for a definite pair of forms.
Conventions
Throughout, are real symmetric matrices, , and is positive definite unless an example explicitly tests failure of that hypothesis. Column coordinates are used. The first form is , and its norm is . A basis matrix is -orthonormal when . This is different from Euclidean orthogonality .
Definition 11.1.1
(Generalized eigenvalue and eigenvector)
Let be positive definite and symmetric. A scalar is a generalized eigenvalue of the pair if for some nonzero ; such a is a generalized eigenvector. The eigenvalues are the roots of .
Normalize the positive form, then use the spectral theorem
Theorem 11.1.2
(Simultaneous reduction of a definite pair)
Let be real symmetric matrices with positive definite. Then there is an invertible real matrix with and , where is real diagonal.
Proof
(Normalize, then diagonalize)
The positive square root exists and is invertible. Its inverse is symmetric. Define
This matrix is symmetric, because transposing reverses the product and each factor is symmetric. Therefore an orthogonal matrix gives , with real diagonal. Put
Direct multiplication proves the simultaneous congruences
Proof X-ray
Where each hypothesis is spent
The reduction has two steps, and each hypothesis is used once. Positivity of is spent in the first step: it provides , so . Symmetry of is spent in the second step: it makes symmetric, so the spectral theorem supplies an orthogonal with , and . The key move is to normalize to the identity first, because the identity is unchanged by every later orthogonal substitution; this is why the second, Euclidean-orthogonal diagonalization does not disturb the first. Nothing says that and commute: diagonalizes , not . Without positivity there is no , which is exactly the failure displayed in the counterexample below.
The first congruence normalizes the metric; the second diagonalizes the other form in the same coordinates. This does not claim that and commute or that a Euclidean orthogonal matrix diagonalizes both. The change of basis need not be orthogonal in the original coordinates.
Conversely, any invertible with these two congruences yields the same conclusions about the pair. The diagonal entries are intrinsic up to order: they are the eigenvalues of the operator considered next, counted with multiplicity.
The operator is self-adjoint in the correct metric
Let . It need not be symmetric in ordinary coordinates. Nevertheless,
Thus , so is self-adjoint for the inner product defined by . The simultaneous congruences also imply
Indeed gives , and multiplying by and gives the identity. If is column of , we obtain
These are generalized eigenvectors, orthonormal in the reference metric. They form a basis, and all generalized eigenvalues are real. Equivalently they are the roots, with multiplicity, of , since
For distinct eigenvalues, orthogonality can be checked without diagonalizing. If and , symmetry gives and also . Hence . Within a repeated eigenspace, one must choose an -orthonormal basis; arbitrary eigenvectors there need not already be orthogonal.
Congruence and similarity carry different invariants
An individual form matrix changes by . Numerical ordinary eigenvalues are not invariants of arbitrary congruence. But if both matrices change together,
then
Thus the generalized eigenvalues of a pair are unchanged. This is the reason they classify definite pairs up to simultaneous congruence. After normalizing the first form to the identity, only the multiset of diagonal entries in the second remains.
The matrix can be nonsymmetric without having nonreal eigenvalues: its self-adjointness belongs to the metric. Treating it as a Euclidean self-adjoint matrix would use the wrong hypothesis, even when the final eigenvalues happen to be correct.
Quotient bounds and every equality case
Theorem 11.1.3
(Generalized Rayleigh quotient)
Let and be real symmetric with generalized eigenvalues . For every ,
with equality on the left exactly when lies in the generalized eigenspace of , and on the right exactly when it lies in that of .
Proof
(A weighted average of eigenvalues)
Write . The two forms become
For , their quotient is therefore a weighted average:
Equality at the lower bound occurs exactly when every nonzero coordinate belongs to a minimal-eigenvalue direction; equality at the upper bound has the analogous maximal-eigenvalue condition. In the original coordinates, the extremizers are the nonzero vectors in the corresponding generalized eigenspaces. Repeated extreme eigenvalues give a whole subspace of extremizers.
Concept lensGeometric
Axes of one ellipsoid read through another form
The set is an ellipsoid, and the quotient bounds say that restricted to ranges over . A Lagrange condition for a critical point of on is exactly the generalized eigenvalue equation, and the critical value is . So the -orthonormal eigenvectors are the principal directions of as seen by . In coordinates the ellipsoid becomes the unit sphere and the diagonal form . For the pair in the examples, the values of on fill exactly .
The same change of basis handles a pencil of forms:
Its inertia is read from the signs of , with multiplicity. In particular it is positive definite exactly when , positive semidefinite exactly when , and singular exactly at the generalized eigenvalues. Inertia concerns signs under congruence; it does not identify the ordinary numerical eigenvalues of the unnormalized pencil.
Why mere invertibility of the reference form is insufficient
Counterexample mode
An invertible reference form is not enough
A tempting claim is that invertibility of the real symmetric reference form suffices for the reduction. Take
Both matrices are real symmetric and is invertible, but
has characteristic polynomial , hence no real eigenvalue, so no real basis of generalized eigenvectors exists. Equally, no real gives : , which is inertia at work. The repair is the hypothesis , which makes an inner product.
If is singular, even writing is invalid. Simultaneous analysis of general singular or indefinite pairs requires other tools; the theorem here is explicitly the positive-definite case.
Worked examples
Worked example 11.1.4
(Reducing a pair with singular second form)
Take , which is positive definite because and , and . Then has determinant , so the generalized eigenvalues are and . For , gives , and , so . For , gives and , so . Check: and . With we get and . The eigenvalue is allowed: only must be definite.
Worked example 11.1.5
(Sharp bounds for a quotient)
For the same pair, , so the theorem gives values in . The upper bound is attained at , where the quotient is , and the lower bound at , where it is . A vector such as gives , strictly between.
Worked example 11.1.6
(Inertia of a pencil)
For the same pair, . Thus is positive definite for , positive semidefinite and singular at , indefinite for , negative semidefinite and singular at , and negative definite for . Check at : has determinant , so it is indefinite.
Theorem checklist and common errors
- For real symmetric and , some invertible satisfies and real diagonal.
- The columns satisfy and are orthonormal for the inner product. These eigenvalues are those of , not generally those of .
- Simultaneous congruence of the pair changes by similarity, so its eigenvalues with multiplicity are invariants.
- The generalized Rayleigh quotient has the stated sharp extreme values and equality spaces. Pencil inertia counts the signs of .
Common mistake
(Generalized eigenvalues are not the eigenvalues of B)
The numbers are the eigenvalues of , not of . For and the eigenvalues of are , but has roots and . Indeed gives and .
Do not diagonalize the two forms using unrelated bases and call that simultaneous reduction. Keep positivity of the reference form in every theorem statement, and distinguish a repeated generalized eigenvalue from an instruction to choose just one eigenvector.
Exercises
Quick check
(For and , find the generalized eigenvalues and the range of .)
Solve .
Solution · Answer
, so and . The quotient takes exactly the values in : it is at and at .
Quick check
(Why does have real eigenvalues and a basis of eigenvectors even when it is not symmetric?)
Which inner product makes it self-adjoint?
Solution · Answer
It is self-adjoint for the inner product , because for . The spectral theorem in that inner-product space gives real eigenvalues and an -orthonormal eigenbasis.
Quick check
(If the generalized eigenvalues of are and , for which is positive definite?)
Read the signs of .
Solution · Answer
Both and must be positive, so exactly when .