Prerequisites are real inner products, eigenvectors, determinants and induction. For the first eigenvector we also use the fundamental theorem of algebra and basic complex conjugation; the resulting eigenbasis will be real.
What needs to be shown
Definition 7.1.1
(Orthogonal matrix and orthonormal eigenbasis)
A real square matrix is orthogonal if , equivalently if its columns are orthonormal. An orthonormal eigenbasis of a matrix is a basis of consisting of pairwise orthogonal unit vectors that are eigenvectors of .
Theorem 7.1.2
(Real spectral theorem)
For a real symmetric matrix , the real spectral theorem states that there is a real orthogonal matrix and real diagonal such that
The columns of are an orthonormal basis of eigenvectors. This is stronger than finding one eigenvalue or proving orthogonality of eigenvectors with different eigenvalues: we must show there are enough eigenvectors to span the entire space, including repeated-eigenvalue cases. Conversely, every real matrix with real diagonal is symmetric.
First obtain a real eigenvector
Lemma 7.1.3
(Eigenvalues of a real symmetric matrix are real)
Let be a real symmetric matrix with . Then every eigenvalue of in is real, and has a real unit eigenvector.
Proof
()
Assume . By the fundamental theorem of algebra, the characteristic polynomial has a complex root . Thus , so a nonzero complex vector satisfies . Write for conjugate transpose. Since is real symmetric, ; consequently
The scalar is real. Also , and the eigenvector equation gives , so is real. Write with . Taking real and imaginary parts of shows and . At least one of is nonzero. Normalize that vector to obtain a real unit eigenvector .
The use of complex numbers is a proof tool, not a claim that a real matrix always has a real eigenvalue. Symmetry was essential in showing real.
Preserve the perpendicular complement and induct
Lemma 7.1.4
(The perpendicular complement of an eigenvector is invariant)
Let be real symmetric, let be a nonzero real vector with , and let . Then for every .
Proof
(Symmetry applied to )
Let . For every , symmetry and give
Hence .
Proof
(Induction on the dimension)
By the invariance lemma, maps into itself. Its restriction to is self-adjoint for the inherited Euclidean product. The space has dimension , since every vector is uniquely with and . Let the columns of the matrix form an orthonormal basis of , so . Invariance gives , where is symmetric. If , then ; furthermore preserves inner products, since .
Induct on dimension. Dimension zero has an empty orthonormal eigenbasis. In positive dimension, the first lemma gives . Apply induction to the real symmetric matrix to obtain an orthonormal eigenbasis in its coordinate space. Multiplying those eigenvectors by lifts them to an orthonormal eigenbasis of . Adding yields one for all of . Put those eigenvectors in and their eigenvalues in : and , proving . Repeated eigenvalues cause no gap: the induction finds as many orthonormal directions as the corresponding eigenspace needs.
Proof X-ray
Where symmetry is used, and why the induction closes
Symmetry is used twice, and complex numbers once. First, is real because , which makes real, so the real and imaginary parts of are real eigenvectors. Second, maps into itself because uses . The non-obvious move is to restrict to : this lowers the dimension by one while keeping the same kind of object, since is again symmetric, so the induction hypothesis applies. Repeated eigenvalues cause no trouble, because each step asks only for one new eigenvector inside the complement of those already found. The argument closes at dimension zero. Without symmetry the second step already fails: for and , the vector is sent to .
Orthogonality and quadratic forms
If , and , then symmetry gives , so . Vectors with the same eigenvalue need not already be perpendicular; choose an orthonormal basis within that eigenspace.
Theorem 7.1.5
(Extreme values on the unit sphere)
Let be real symmetric with eigenvalues . For every unit vector , . The value is attained exactly at the unit vectors of the eigenspace for , and exactly at the unit vectors of the eigenspace for .
Proof
(A weighted average of eigenvalues)
For any , and . On the unit sphere this quadratic form is a weighted average of the eigenvalues, so its minimum and maximum are the smallest and largest eigenvalues. For any eigenvalue , is equivalent to ; hence its eigenspace is exactly the span of those columns of corresponding to diagonal entries of equal to . The weights are nonnegative and sum to one, so equality holds exactly when is a unit vector in the corresponding extreme eigenspace.
Counterexample mode
A real matrix need not have a real eigenvalue
Claim: every real square matrix has a real eigenvalue, so the first step of the proof does not need symmetry. Take the quarter-turn . Then , which has no real root, so has no real eigenvalue and no real eigenvector; its eigenvalues are . This matrix is not symmetric, since . For a real symmetric matrix the same step succeeds because is real. Symmetry cannot be dropped.
Worked examples
Worked example 7.1.6
(A singular symmetric matrix)
Find an orthogonal diagonalization of and compute .
Solution. has eigenvalue and has eigenvalue . They are perpendicular and each has norm , so
The column equations verify and . Therefore . The zero eigenvalue makes singular; it does not obstruct orthogonal diagonalization.
Worked example 7.1.7
(Repeated eigenvalues and all extrema)
Let . Find an orthonormal eigenbasis and the extrema of on , including all attaining vectors.
Solution. , with the all-ones matrix. It acts by on , and by on the plane . An orthonormal eigenbasis is
Their eigenvalues are ; dot products and norms verify orthonormality. Thus the maximum is , attained exactly at , while the minimum is , attained at every unit vector in the sum-zero plane, equivalently with . The repeated eigenvalue gives a whole circle of minimizing directions, not a unique eigenvector.
Worked example 7.1.8
(Extreme values on the unit circle)
Let , so . The eigenvalues are with and with . By the theorem on extreme values, the maximum of over unit vectors is , attained at , and the minimum is , attained at . Check with : , whose range is , equal to at and to at .
Common mistakes
Common mistake
(Equal eigenvalues do not give perpendicular eigenvectors)
For , both and are eigenvectors for the eigenvalue (for instance ), but their dot product is . Symmetry gives orthogonality only across different eigenvalues; inside one eigenspace an orthonormal basis must be chosen.
Exercises
Quick check
(Why is not orthogonally diagonalizable?)
Solution · Guided solution
If with orthogonal and diagonal, then , so would be symmetric. But . In fact is not even diagonalizable: its only eigenvalue is , and has rank , so the eigenspace is a line.
Quick check
(Find the maximum and minimum of over unit vectors for , and the vectors where they occur.)
Solution · Guided solution
The eigenvalues are with and with (trace , determinant ). So the maximum is , attained at , and the minimum is , attained at . Check: for , .
Quick check
(In the induction step, why is the matrix symmetric when is?)
Solution · Answer
Transposing reverses the order of the factors: , using .