For a matrix, eigenvalues often describe the obstruction to solving . In infinite dimensions, an operator can be injective at every scalar and still fail to have a bounded inverse. The spectrum records all failures of bounded invertibility, not just eigenvectors. Its radius, surprisingly, can be recovered from the long-term growth of powers of the operator.
Definition 5.2.1
(Resolvent set, spectrum and resolvent)
Let be a nonzero complex Banach space and . The resolvent set consists of for which has a bounded inverse on all of ; the spectrum is . If , the resolvent at is .
By the bounded inverse theorem for Banach spaces, a bijective automatically has bounded inverse; surjectivity and injectivity both matter. A scalar is an eigenvalue when fails to be injective; the spectrum may contain points of other kinds.
Neumann series and the resolvent
Theorem 5.2.2
(Neumann series)
If and , then is invertible and
Proof
(Telescoping partial sums)
Since , completeness of gives norm convergence of . Multiplying its finite partial sums by on either side gives , which tends to ; passing to the limit shows the sum is a two-sided inverse, and the geometric series bounds its norm.
Proposition 5.2.3
(Analytic structure of the resolvent)
For ,
At any , with , every with lies in and
For ,
Proof
(Factor out an invertible operator)
For , write and apply (1) with ; the bound is . Near , factor
and use (1) with ; inverting the product gives the norm-convergent expansion (3). For (4), insert and between the two inverses: .
Thus is open, is closed and lies in the disk , and is locally represented by an operator-norm power series, so it is norm-continuous on . Resolvents at different points commute, because the operators and themselves commute.
Nonempty spectrum
Theorem 5.2.4
(Nonempty compact spectrum)
is nonempty and compact.
Proof
(Liouville applied to scalar matrix coefficients)
Closedness and boundedness were proved above. Suppose instead that the spectrum is empty. For and , equation (3) makes an entire scalar function. The estimate (2) makes as . It is therefore bounded: outside a large disk use (2), and inside use continuity on a compact disk. By the scalar Liouville theorem, is identically zero. Hahn–Banach separation then implies for all , impossible because is the inverse of and .
This proof uses the scalar Liouville theorem and complex Hahn–Banach as prior results. Both are statements about complex scalars, and the hypothesis cannot be dropped.
Counterexample mode
Over the real field the spectrum can be empty
Claim to test: the theorem holds for real Banach spaces with real . On , let , the rotation by . For real ,
so is invertible for every real and the real spectrum is empty. The repair is to work over : on the same matrix has exactly at , so .
Growth of powers
Define , the spectral radius, which exists because the spectrum is nonempty and compact.
Theorem 5.2.5
(Spectral-radius formula)
Proof
(Submultiplicativity, spectral mapping, then a Cauchy estimate)
First, the limit and infimum agree by submultiplicativity. If some , all later powers vanish, and both the limit and the infimum are zero. Otherwise put . For fixed , write , ; then (where ). Taking th roots and a limsup gives . Since is arbitrary and every term is at least the infimum, the limit is the infimum.
For the lower bound, if then . Indeed, with . If had inverse , then commutes with , and would be both a right and a left inverse of . Applying the disk bound to yields and hence .
For the reverse bound, fix . The circle lies in the resolvent set, and continuity makes finite. For and , the scalar function
is analytic on , with the radius interpreted as infinite when . Near zero this follows from (2), whose series is . On we have , so the scalar Cauchy coefficient estimate gives
Hahn–Banach yields . Taking th roots and then proves the reverse inequality in (5).
Proof X-ray
How analyticity converts a radius into a growth rate
The three parts use different inputs. Equality of limit and infimum is pure norm algebra: only is used. The lower bound needs only the factorization of and the disk bound applied to . Neither part uses complex analysis.
The key move is in the upper bound. The power series is the Neumann series (2) in the variable , and a priori it is known to converge only for . But is also given by the resolvent, which is analytic on the larger disk . A scalar power series converges on every disk where its sum is analytic, and Cauchy's estimate on the circle turns this into the coefficient bound . Uniformity in and , then Hahn–Banach, upgrade the scalar bound to an operator-norm bound. The argument closes by letting decrease to : the factor tends to one for each fixed , even though may blow up as approaches the spectrum.
Common mistake
(Confusing the norm with the spectral radius)
measures the largest single-step amplification, whereas measures the asymptotic exponential rate of powers. Always , but equality can fail badly. For on with Euclidean norm, but and . The radius is zero even though the first step can be large; a bound says nothing about whether powers decay.
Worked example 5.2.6
(Transient growth before decay)
On with Euclidean norm, let with . The matrix is upper triangular with diagonal entries , so and . Since and commutes with , the binomial formula gives
The off-diagonal entries are for , so the powers do not shrink at first. To check (5), bound the norm by the entry and by the triangle inequality:
Both bounds have th roots tending to , because and . Hence , although .
Worked example 5.2.7
(A diagonal operator whose spectrum contains a non-eigenvalue)
On , let . Each is an eigenvalue, with eigenvector . If , then because is closed, and the diagonal operator with entries is bounded by and inverts . So . Since the spectrum is closed and contains every , it also contains , so .
The point is not an eigenvalue, because forces every ; instead is injective but not surjective: would need the preimage . Finally for every (attained at ), so (5) gives , matching .
Spectrum without eigenvectors
Worked example 5.2.8
(The unilateral right shift)
On , let , the unilateral right shift. It is an isometry, so and its spectrum lies in the closed unit disk. If , the equation reads and for , which forces , not square-summable. At , is outside the range of . Thus the open unit disk lies in , and closedness gives the entire closed disk. Nevertheless has no eigenvalues: the coordinate equations read and , which force both for (inductively from ) and for . Failure of surjectivity supplies the spectrum here.
The diagonal example had a single non-eigenvalue spectral point; the shift has a whole disk of them.
Practice
Quick check
(On , take . Find and .)
Compute when is invertible.
Solution · Answer
is invertible exactly when , so and .
Quick check
(If , what can its spectrum be? Justify nonemptiness.)
Use the polynomial factorization from the lower-bound step.
Solution · Answer
The factorization argument gives for every spectral , hence . The spectrum is nonempty by the theorem, so .
Quick check
(For the right shift , compute and check (5).)
Each power shifts by places.
Solution · Answer
Every power is an isometry, so and (5) yields , consistent with being the closed unit disk.
Quick check
(If , bound the error after replacing in (2) by the sum through .)
The error is a geometric tail.
Solution · Answer
With , the error is , whose norm is at most .