S. D. V. Stephens

notes / Studies in Algebra and Group Theory / chapter 3: Linear Maps and Operators / From my book

B7.4 Eigenvalues and Eigenvectors

The simplest invariant subspaces are one-dimensional. These give rise to the fundamental notion of eigenvalues.

7.4.1 Definitions and Basic Properties

Definition 7.4.1 (Eigenvector and eigenvalue). Let . A nonzero vector is an eigenvector of if for some scalar . The scalar is called the eigenvalue corresponding to , and is an eigenvalue of if it corresponds to some eigenvector.

Note. The condition says that acts on the line spanned by as scaling by . Equivalently, is a one-dimensional -invariant subspace. Conversely, a one-dimensional invariant subspace has for some , so it consists of eigenvectors with the same eigenvalue, together with zero.

Proposition 7.4.2 (Characterization of eigenvalues). For and , the following are equivalent:

  1. is an eigenvalue of .
  2. is not injective.
  3. .

If is finite-dimensional, they are also equivalent to:

  1. is not surjective (equivalently, not invertible).
  2. .

Proof. (a) (b) (c): is an eigenvalue iff there exists nonzero with , iff for some nonzero , iff , iff is not injective (Theorem 6.1.4).

(b) (d): For an operator on a finite-dimensional space, injective, surjective and invertible are equivalent (Corollary 6.1.16).

(d) (e): An operator on a finite-dimensional space is invertible if and only if its determinant (Definition 5.4.20) is nonzero (Theorem 5.4.8). ∎

Definition 7.4.3 (Eigenspace). For , the eigenspace of for is It is a subspace of , and it is nonzero exactly when is an eigenvalue; its nonzero elements are the eigenvectors with eigenvalue (the zero vector belongs to it but is not an eigenvector).

Theorem 7.4.4 (Eigenvectors with distinct eigenvalues are linearly independent). Let be eigenvectors of with distinct eigenvalues . Then are linearly independent.

Proof. Induction on . Base case : a single eigenvector is nonzero, hence linearly independent.

Suppose the result holds for eigenvectors. Assume . Apply : Multiply the original equation by and subtract: By the induction hypothesis (since are eigenvectors with distinct eigenvalues ), we have for . Since , we get . Substituting back: , so (since ). ∎

Corollary 7.4.5 (Bound on the number of eigenvalues). A linear operator on an -dimensional space has at most distinct eigenvalues.

Proof. If are distinct eigenvalues with eigenvectors , then are linearly independent. In an -dimensional space, any linearly independent set has at most elements. ∎

Corollary 7.4.6 (The sum of eigenspaces is direct). If are distinct eigenvalues of , then the sum is direct: If is finite-dimensional, then .

Proof. Suppose with . If some were nonzero, the nonzero ones would be eigenvectors with distinct eigenvalues satisfying a linear relation with all coefficients equal to , contradicting Theorem 7.4.4. So every , which is what it means for the sum to be direct. The dimension of a direct sum is the sum of the dimensions of the summands (Proposition 4.7.14), and it is at most . ∎

7.4.2 The Characteristic Polynomial

Eigenvalues are the roots of a polynomial built from the determinant (Section 5.4). The matrix has entries in the polynomial ring , and determinants of matrices over a commutative ring are defined by the same Leibniz formula and have the same basic properties as over a field (Example 5.4.25 and the discussion before it).

Definition 7.4.7 (Characteristic polynomial). Let . The characteristic polynomial of is the determinant of the matrix . For with finite-dimensional, the characteristic polynomial of is , where for a basis of . (For we set .)

Proposition 7.4.8 (Basic properties of the characteristic polynomial). Let with , and let .

  1. is a monic polynomial of degree in , and
  2. For , the value of at is .
  3. does not depend on the choice of basis .
  4. The roots of in are exactly the eigenvalues of .

Proof. (a) and (b): By the Leibniz formula, is a sum, over permutations, of a product of entries of , one from each row and each column. Each entry is a polynomial in , so the sum is a polynomial in ; and substituting term by term gives the Leibniz formula for , which proves (b) (see also Example 5.4.25). The identity permutation contributes . Every other permutation moves at least two indices, so its product contains at most diagonal entries and has degree at most . Hence is monic of degree with -coefficient . Its constant term is by (b).

(c) If is another basis, then for an invertible matrix (Corollary 5.3.4). Then in , so by multiplicativity of the determinant (Theorem 5.4.7, which holds over the commutative ring ),

(d) By (b), , which vanishes if and only if is an eigenvalue (Proposition 7.4.2). ∎

Example 7.4.9 (Characteristic polynomial of a matrix). For :

Definition 7.4.10 (Algebraic and geometric multiplicity). Let be an eigenvalue of , finite-dimensional. The algebraic multiplicity of is the largest integer such that divides . The geometric multiplicity of is .

When is a product of linear factors, the algebraic multiplicity simply counts how often occurs:

Lemma 7.4.11 (Multiplicity of a root in a product of linear factors). Let . The largest such that divides is the number of indices with .

Proof. Let be the number of with , and write with . Then . Clearly divides the product. If divided it, say , then cancelling (possible because has no zero divisors, Lemma 4.3.3) would give , hence , a contradiction. ∎

7.4.3 Diagonalizable Operators

The best-case scenario: has a basis of eigenvectors.

Definition 7.4.12 (Diagonalizable operator). An operator on a finite-dimensional space is diagonalizable if there exists a basis of consisting entirely of eigenvectors of .

Note. If is a basis of eigenvectors with , then the matrix of is The eigenvalues appear on the diagonal (possibly with repetition).

Theorem 7.4.13 (Characterizations of diagonalizability). Let with , and let be the distinct eigenvalues of . The following are equivalent:

  1. is diagonalizable.
  2. The matrix of with respect to some basis is diagonal.
  3. .
  4. .

Proof. (a) (b): The matrix in the basis is diagonal exactly when each is a multiple of , i.e., when every is an eigenvector.

(a) (c): Every vector of a basis of eigenvectors lies in some , so these eigenspaces span ; their sum is direct by Corollary 7.4.6.

(c) (d): Take dimensions: .

(d) (a): The sum is direct (Corollary 7.4.6) and has dimension , so it equals . Combining bases of the gives a basis of consisting of eigenvectors. ∎

Corollary 7.4.14 (Enough distinct eigenvalues implies diagonalizability). If has distinct eigenvalues, then is diagonalizable.

Proof. Each eigenspace has dimension at least 1 (it contains an eigenvector). With distinct eigenvalues, . But this sum is also (Corollary 7.4.6), so equality holds, and criterion (d) above is satisfied. ∎

7.4.4 Examples: Diagonalizable and Non-Diagonalizable

Example 7.4.15 (Diagonal matrices). The identity matrix has eigenvalue with eigenspace all of . More generally, a diagonal matrix has the standard basis vectors as eigenvectors with eigenvalues . Such a matrix is (trivially) diagonalizable—it is already diagonal.

Example 7.4.16 (A non-diagonalizable matrix). Consider with matrix The characteristic polynomial is , so is the only eigenvalue.

The eigenspace: .

This has dimension 1, but . So is not diagonalizable.

Geometrically, is a “shear”: it fixes the -axis pointwise and slides other points parallel to the -axis. The only eigenvector direction is along the -axis.

Example 7.4.17 (No eigenvectors over ). Consider rotation by in : The characteristic polynomial is , which has no real roots. So has no eigenvalues over , hence no eigenvectors. Over , the eigenvalues are with eigenvectors , and is diagonalizable as a complex matrix.

This illustrates that diagonalizability depends on the scalar field.

Note. The rotation example shows that over an operator may have no eigenvalues at all; Section 7.10 explains what can be said instead. Even over , where every polynomial has roots, an operator may fail to be diagonalizable (like the shear above). Understanding non-diagonalizable operators requires the Jordan normal form, which we develop in Section 7.8.

Question 7.4.18 (Exercises on quotient spaces, operators, and eigenvalues).

  1. Let and . Find , , and describe explicitly.
  2. Prove that if commute (), then every eigenspace of is -invariant.
  3. Let be diagonalizable with distinct eigenvalues . Prove that where .
  4. Show that the quotient map induces a bijection between and .
  5. Let . Find the eigenvalues and eigenspaces of . Is diagonalizable? If so, find a matrix such that is diagonal.
Show solution

(a) The map defined by is linear and has . It is surjective, since . By Rank–Nullity, ; a basis is .

By the dimension formula: .

By the First Isomorphism Theorem, is an isomorphism. Explicitly, the coset is determined by the pair .

(b) Let , i.e., . We need , i.e., .

Compute: . ✓

(We used commutativity in the first equality. This is also a special case of Proposition 7.3.3, since commutes with .)

(c) Since is diagonalizable, . Let . The factors of commute (Proposition 7.2.2), so we may apply first: Since this holds on each eigenspace and the eigenspaces span , we have .

(d) Define by . This lands in the right set, because .

Injective: If , then for all , so .

Surjective: Given with , Theorem 7.1.7 gives with , so .

(Both sides are vector spaces and is linear, so is even an isomorphism of vector spaces.)

(e) Characteristic polynomial: .

Eigenvalues: , .

Eigenspaces:

  • : gives , so . Eigenspace: .
  • : gives , so . Eigenspace: .

Since has distinct eigenvalues, it is diagonalizable (Corollary 7.4.14).

Let (columns are eigenvectors). Then .