Why eigenvectors are not always enough
For , an eigenvector with gives the solution . If we can find a basis of eigenvectors, linear combinations of these solutions give every initial state. Repeated eigenvalues do not guarantee such a basis, however: algebraic multiplicity counts roots of the characteristic polynomial, while eigenspace dimension counts independent eigenvectors. The missing directions require a different construction.
Consider . Its only eigenvectors are multiples of . Yet an initial value problem in the plane has two independent initial coordinates. Direct integration gives and . The factor supplies a second independent solution, although it does not produce another eigenvector.
Chains and the exponential
Definition 4.2.1
(Jordan chain)
For an eigenvalue , a Jordan chain of length is a sequence with , , and for . The vectors after are generalized eigenvectors.
The order matters: a chain is built backwards from the eigenvector. Adding a multiple of to produces another valid second vector, so a chain is generally not unique. Chain vectors are independent: if , apply to obtain , and proceed downwards.
Theorem 4.2.2
(Solutions from a chain)
For a chain of a constant complex matrix , define
Then each solves , and these solutions are independent. If all chosen chains together form a basis of , their solutions form a fundamental set over .
Proof
(Differentiation moves one link down the chain)
Put . On the chain, and . Differentiating the polynomial factor yields precisely times that factor: the derivative of the th term becomes the st term with . Thus . At , ; their independence follows from independence of the chain vectors. If a collection of chains is a basis at time zero, the corresponding solution columns form a fundamental matrix: uniqueness for the linear initial value problem prevents a nontrivial combination from vanishing at any later time.
Proof X-ray
Where the polynomial comes from
The proof uses only the two chain relations and . The key move is the choice of weights : differentiation lowers the power of by one, exactly as moves one link down the chain, so and are the same sum. The sum closes without a leftover term only because the last vector of the shift is the genuine eigenvector , which annihilates; a sequence that does not end in an eigenvector would leave the term .
Independence is checked only at , and uniqueness for the linear initial-value problem carries it to every . Equivalently, : on the span of the chain, with nilpotent, so the series for terminates. The powers of count links of the chain; they are not extra eigenvalues.
The Jordan basis theorem guarantees a basis of chains over for any finite square complex matrix. In calculations it is often enough to solve successive linear systems for the particular chains needed. The theorem does not say that arbitrary independent eigenvectors can always be extended by arbitrary generalized eigenvectors into chains.
Worked example 4.2.3
(A chain of length three)
Let
With standard basis vectors , the nilpotent part sends to , to , and to zero. The solution with is therefore
The initial value is immediate; differentiating each coordinate verifies the system. Despite the polynomial factors, as , because every fixed polynomial is dominated by .
Worked example 4.2.4
(Multiplicity three, but chains of lengths two and one)
Let
The eigenvalue has algebraic multiplicity three, and the eigenspace has dimension two. Since , no chain is longer than two. A second chain vector requires , so must lie in the range of , which is . Edge case: the eigenvector cannot be extended, because is never . Taking the chains and , the solution with is
Only appears, although the multiplicity is three: the highest power of is one less than the longest chain, not one less than the multiplicity.
Common mistake
(Using as the second solution)
The guess fails: , while , so the residual never vanishes. The generalized eigenvector is needed to absorb it: works because .
Complex eigenvalues and real data
Suppose is real and a complex solution is , with real vectors . Since is real, comparing real and imaginary parts of shows that both and are real solutions.
Theorem 4.2.5
(Real solutions from a complex eigenpair)
Let be a real matrix with eigenvalue , , and eigenvector with . Then
are linearly independent real solutions of .
Proof
(Real and imaginary parts of one complex solution)
Expanding shows that the two displayed functions are its real and imaginary parts, hence real solutions. Their values at are and . If were real-dependent, both would be real multiples of one real vector , so with , . Then is an eigenvector, and with and real forces real, contradicting . Hence the initial values are independent, and so are the solutions.
Complex generalized chains can be handled in the same way. Pair conjugate chains and take real and imaginary parts; using both the original complex columns and their real/imaginary parts would count the same solution directions twice.
Worked example 4.2.6
(Rotation with growth)
For , has eigenvalue . Its real and imaginary solution parts are
They have independent initial values. Thus the real solution through is . The angular motion and the factor describe different features: rotation does not cancel the growth.
Common mistake
(Taking parts of both conjugate solutions)
In the rotation example, the conjugate eigenvector for gives the same real part and the negative imaginary part. Its parts add nothing new, so a real fundamental set uses one eigenvector from each conjugate pair.
Counterexample mode
Negative eigenvalues do not imply monotone decay
For and , the solution is . It eventually tends to zero, but its norm is not monotone: , while . Every solution of a constant-coefficient system tends to zero as exactly when every eigenvalue has strictly negative real part; this condition does not imply monotone decay of each solution's Euclidean norm.
Practice
Quick check
(For , solve with .)
Solution · Answer
The standard basis forms a chain. Hence . Substitution gives and , and the initial value is correct.
Quick check
(If and , verify directly that is a solution.)
Solution · Answer
Differentiation gives . The chain equations give , proving equality.
Quick check
(For , find the solution through and decide whether its norm changes.)
Solution · Answer
The solution is . Its squared norm is , either by expanding or by observing that for this skew-symmetric matrix.