The determinant of the previous unit detects when a square matrix is singular. Applied to , it detects the special scalars for which acts on some nonzero vector as a pure scaling. This unit sets up the vocabulary for eigenvalues precisely: the characteristic polynomial, eigenspaces, the two notions of multiplicity, and the exact condition under which a matrix is similar to a diagonal matrix. It prepares the applications in the next unit, which use .
Definitions
Definition 2.2.1
(Eigenvalues, eigenspaces, multiplicities and diagonalizability)
Let be an matrix with real or complex entries. A scalar is an eigenvalue of if for some nonzero vector , called an eigenvector for . Equivalently, has a nonzero solution, that is, is singular, that is, . The polynomial
is the characteristic polynomial of . It is monic of degree , and its roots are the eigenvalues. The eigenspace of is , a subspace that contains and all eigenvectors for .
The algebraic multiplicity is the multiplicity of as a root of , and the geometric multiplicity . A matrix is diagonalizable if there is an invertible and a diagonal with .
Proposition 2.2.2
(Diagonalization is an eigenbasis)
If with , then the columns of are eigenvectors, . Conversely, if are linearly independent eigenvectors with eigenvalues , then for and .
Proof
(Comparing columns of and )
says that the -th column of , which is , equals the -th column of , which is . Invertibility of is equivalent to independence of its columns.
So a matrix is diagonalizable exactly when has a basis of eigenvectors.
Basic facts
Theorem 2.2.3
(Basic facts about eigenvalues)
(a) Eigenvectors belonging to distinct eigenvalues are linearly independent. (b) and , counting algebraic multiplicities over . (c) and have the same characteristic polynomial. (d) If , then for , and if is invertible, then and .
Proof
(Four short arguments)
(a) Suppose are eigenvectors for distinct , and suppose they are dependent. Choose a nontrivial relation with the fewest nonzero coefficients, say with all , (a single term is impossible because ). Apply for some : . The term vanishes and the others keep nonzero coefficients, giving a nontrivial relation with fewer terms, a contradiction. (b) Compare coefficients in : the coefficient of is on the left and on the right, and gives . (c) . (d) Induction on ; and gives , which forces .
Proof X-ray
The minimal-relation argument for distinct eigenvalues
Part (a) is a minimal-counterexample argument: among all nontrivial dependences, take one with the fewest nonzero coefficients. The key move is applying . It annihilates the -th term and multiplies each other coefficient by , which is nonzero precisely because the eigenvalues are distinct; this is the only place distinctness enters. The result is a nontrivial relation with fewer terms, contradicting minimality. The case of a single term is excluded because with forces . With a repeated eigenvalue the multiplier can vanish, and indeed the independent vectors are eigenvectors of for the same eigenvalue.
Theorem 2.2.4
(Geometric multiplicity is at most algebraic multiplicity)
For every eigenvalue, .
Proof
(A block triangular form)
Let and choose a basis of . Extend it to a basis of and let . Since for ,
for some blocks and . Similar matrices have the same characteristic polynomial, and the block triangular form gives . Hence divides , so .
The diagonalizability criterion
Theorem 2.2.5
(The diagonalizability criterion)
Let be an matrix with distinct eigenvalues over . Then is diagonalizable (over ) if and only if for every . In particular, distinct eigenvalues suffice. A real matrix is diagonalizable over if and only if, in addition, all its eigenvalues are real.
Proof
(Direct sum of eigenspaces and a dimension count)
The sum of the eigenspaces is direct: if with , the nonzero are eigenvectors for distinct eigenvalues, hence independent by the independence statement in the basic facts, so every . Therefore , and a basis of eigenvectors exists if and only if this sum is . By the multiplicity inequality and (over , the polynomial has roots with multiplicity), , with equality exactly when for all . With distinct eigenvalues each , so forces .
In practice: find the roots of ; for each repeated root compute and compare it with the multiplicity.
Counterexample mode
Two eigenvalues counted with multiplicity do not give an eigenbasis
Claim: a matrix with two eigenvalues counted with multiplicity is diagonalizable. This is false. Let . Then , so , but has rank one, so . If then (the only eigenvalue is ), and so , which it is not. The repaired condition is for every eigenvalue.
Worked examples
Worked example 2.2.6
(Distinct eigenvalues)
Let . Then . For : gives . For : gives . With we have , and
Also and .
Worked example 2.2.7
(A defective matrix)
Let . Then , so and . For : has rank , so . By the diagonalizability criterion, is not diagonalizable: the eigenvectors span only a -dimensional space.
Worked example 2.2.8
(Repeated but diagonalizable)
Let where is the all-ones matrix. The eigenvalues of are , so has eigenvalues . For : has rank , so . Hence is diagonalizable. An eigenbasis is for and , for . Check: and ; the vector is checked in the same way.
Common mistakes and subtle points
Common mistake
(A repeated eigenvalue does not forbid diagonalization)
The repeated-but-diagonalizable example has a repeated eigenvalue and is diagonalizable; what matters is .
- Distinct eigenvalues are sufficient, not necessary. The identity matrix has one eigenvalue and is diagonalizable.
- Eigenvalues of a sum or product. In general the eigenvalues of or are not obtained from those of and separately.
Common mistake
(Real versus complex eigenvalues)
The rotation has , no real eigenvalue and so no real eigenvector, but it is diagonalizable over with eigenvalues .
- The zero vector is not an eigenvector. The eigenspace contains , but eigenvectors are the nonzero elements.
Common mistake
(Row operations do not preserve eigenvalues)
Elimination does not preserve eigenvalues; only similarity does.
Exercises
Quick check
(Find the eigenvalues and eigenvectors of and decide whether is diagonalizable.)
Solution · Answer
The trace is and the determinant is , so and the eigenvalues are and . For : gives . For : gives . Since the eigenvalues are distinct, is diagonalizable with and . Check: and .
Quick check
(Show that is not diagonalizable.)
Solution · Answer
The characteristic polynomial is , so . The matrix has rank , so . By the diagonalizability criterion the matrix is not diagonalizable.
Quick check
(Prove that and have the same eigenvalues, but give an example in which they have different eigenvectors.)
Solution · Answer
gives the same characteristic polynomial and hence the same eigenvalues. For , the eigenvalue has eigenvector for , but for it has eigenvector (indeed ), and is not an eigenvector of since .
Quick check
(For which real is diagonalizable?)
Solution · Answer
The only eigenvalue is with . The matrix has rank if and rank if . So exactly when , and the matrix is diagonalizable if and only if (when it is the identity).