The real spectral theorem proved that every real symmetric matrix has an orthonormal basis of eigenvectors. This unit is about using that theorem: how to produce and with in practice, why eigenvectors for different eigenvalues need no extra work to be orthogonal, what to do when an eigenvalue is repeated, and how the same data rewrite as a sum of rank-one projectors. We use eigenvectors and characteristic polynomials (5.1 and 5.3), Gram-Schmidt (6.3) and the real spectral theorem (7.1).
Statement and a converse
Definition 7.3.1
(Orthogonally diagonalizable matrix)
A real matrix is orthogonally diagonalizable if there are a real orthogonal matrix (so ) and a real diagonal matrix with .
Theorem 7.3.2
(Real spectral theorem (recalled))
If is a real symmetric matrix, there is a real orthogonal matrix (so ) and a real diagonal matrix with . The columns of are orthonormal eigenvectors and the diagonal entries of are the corresponding eigenvalues.
The theorem is proved in the unit on the real spectral theorem.
Proposition 7.3.3
(Converse of the spectral theorem)
If with orthogonal and diagonal real, then is symmetric.
Proof
(Transposing )
because .
Hence a real matrix is orthogonally diagonalizable if and only if it is symmetric. In particular, a non-symmetric matrix such as cannot be orthogonally diagonalized.
Eigenvectors of different eigenvalues are automatically orthogonal
Proposition 7.3.4
(Eigenvectors of distinct eigenvalues are orthogonal)
Let be real symmetric, , with . Then .
Proof
(Moving A across the dot product)
Using ,
So , and forces .
Proof X-ray
The single step that uses symmetry
The proof computes in two ways. One way uses . The other moves across the dot product, , which holds for every real matrix; symmetry is used only to replace by , after which gives . Both computations give the same number, so , and the hypothesis is what allows the factor to be canceled. The proof says nothing when , which is why orthogonality inside a repeated eigenspace needs Gram-Schmidt. For a non-symmetric the quantity is not , and the argument breaks.
Consequently, eigenspaces belonging to different eigenvalues are mutually orthogonal. The only work needed is inside each eigenspace.
The procedure
For a real symmetric :
- Compute the characteristic polynomial and find all eigenvalues (they are real by the real spectral theorem).
- For each eigenvalue find a basis of . By the real spectral theorem, the dimension of this eigenspace equals the multiplicity of .
- Apply Gram-Schmidt inside each eigenspace to obtain an orthonormal basis of that eigenspace. If is simple, normalize the single vector.
- Collect all these vectors as the columns of and put the eigenvalues, in the same order, on the diagonal of .
The vectors collected in step 4 are orthonormal because vectors inside an eigenspace are orthonormal by construction and vectors from different eigenspaces are orthogonal by the orthogonality proposition above.
Spectral decomposition
Theorem 7.3.5
(Spectral decomposition)
Let with orthogonal and with columns , and . Then , where is the matrix of the orthogonal projection onto the line spanned by . Grouping equal eigenvalues, , where is the orthogonal projection onto the eigenspace for , with for and .
Proof
(Expanding column by column)
Let be the orthonormal columns of with eigenvalues . Expanding column by column gives
Each is the matrix of orthogonal projection onto the line spanned by ; indeed . Grouping the terms with equal eigenvalues, where is the orthogonal projection onto the eigenspace for . These projections satisfy for and .
Concept lensGeometric
A as stretching along perpendicular axes
Geometrically, splits into components along the orthonormal axes , multiplies the component along by (reversing it when , collapsing it to zero when ), and reassembles the result. In the first worked example the axes and are stretched by and , so the unit circle is mapped to an ellipse with semi-axes and along those lines. The matrix factorization and this picture describe the same map.
This expression makes functions of easy: , and if all , then is a symmetric matrix whose square is .
Worked examples
Worked example 7.3.6
(Distinct eigenvalues)
Let . The characteristic polynomial is . For : gives . For : gives , and as predicted by the orthogonality proposition. So
The spectral decomposition is , and the entries add up to .
Worked example 7.3.7
(A repeated eigenvalue and Gram-Schmidt)
Let , where is the all-ones matrix. The matrix has eigenvalue on and eigenvalue on the plane . So has eigenvalue (simple) and (multiplicity ). For : . For the eigenspace is the plane , with the basis , . These are not orthogonal (), so apply Gram-Schmidt: , then
which has norm , so . Check: : . Also and . Therefore and .
Worked example 7.3.8
(A singular matrix and a scalar matrix)
Let . Its trace is and its determinant is , so the eigenvalues are and , with unit eigenvectors and . The spectral decomposition has a single term,
so . The zero eigenvalue only makes the second term vanish; it does not obstruct orthogonal diagonalization. At the other extreme, has a single eigenvalue whose eigenspace is all of , so every orthonormal basis works: both and give . The matrix is therefore not unique when an eigenvalue repeats.
Common mistakes and subtle points
Counterexample mode
Distinct eigenvalues, eigenvectors not perpendicular
Claim: eigenvectors belonging to distinct eigenvalues of a real matrix are orthogonal. This needs symmetry. Take , which has the distinct eigenvalues and . The vector satisfies , and satisfies . Their dot product is . The matrix is diagonalizable, but it is not orthogonally diagonalizable: by the converse proposition, would force to be symmetric, and . The repair is the hypothesis .
Common mistake
(Orthogonality inside an eigenspace is not automatic)
In the second worked example, and are both eigenvectors for , but . Normalizing them and using them as columns of would not give . Orthogonality for distinct eigenvalues does not extend to a basis of one eigenspace; Gram-Schmidt is needed there.
Common mistake
(Unnormalized columns and mismatched order)
Row reduction for gives the eigenvector for , of length ; with it as a column, has a on the diagonal rather than , so . Also the -th diagonal entry of must belong to the -th column of : if the column comes first, then , not . The check is and .
Exercises
Quick check
(Orthogonally diagonalize .)
Solution · Guided solution
The trace is and the determinant is , so . For : gives . For : gives . These are orthogonal (). Normalizing, and .
Quick check
(Find the symmetric matrix with eigenvalue for the eigenvector and eigenvalue for .)
Solution · Guided solution
The unit eigenvectors are and (orthogonal). Then . Check: and .
Quick check
(Using (the matrix of the first worked example), find a symmetric with .)
Solution · Guided solution
Take . Then is symmetric and
Quick check
(Let be symmetric with . Show that is the matrix of an orthogonal projection.)
Solution · Guided solution
Since , , and gives , so every eigenvalue satisfies , i.e. . By the spectral decomposition, , the sum of the projectors onto the orthonormal vectors with eigenvalue . This is the orthogonal projection onto their span, which is the column space of .
Quick check
(Orthogonally diagonalize .)
Solution · Guided solution
The matrix is block diagonal with the block and the block , whose eigenvalues are with and with . So the eigenvalues are (simple) and (multiplicity ). The eigenvalue has . The eigenspace for is spanned by and , which are already orthogonal, so the orthonormal basis is and . Hence and .