For a finite matrix, an equation can fail to have a unique solution only because a kernel or an adjoint-kernel compatibility condition appears. Infinite-dimensional operators can behave worse: their ranges need not even be closed. Compact perturbations of the identity recover much of the finite-dimensional picture, without requiring the compact operator to be self-adjoint.
The alternative
Let be a real or complex Hilbert space, with inner product linear in the first variable, and let be compact. Put . We use the Hilbert projection theorem, the bounded inverse theorem, and the facts proved in the compact-operator notes that is compact and compactness is preserved by scalar multiples.
Theorem 2.6.1
(Fredholm alternative on a Hilbert space)
The spaces and are finite dimensional and have equal dimension. The range of is closed and
Consequently is solvable exactly when . Whenever it is solvable, all solutions form and precisely one lies in . In particular, either both kernels are zero and has a bounded inverse, or both are nonzero and solvability requires the stated compatibility conditions.
This formulation allows non-self-adjoint . The two kernels have the same dimension but need not be the same subspace. We first establish closed range and finite nullities, then prove the equality of dimensions rather than assuming that a bounded-below operator is automatically onto.
Finite kernel and closed range
Lemma 2.6.2
(Kernel, lower bound and range of )
Let . Then is finite dimensional, there is a constant such that
and is closed. The same holds for .
Proof
(Orthonormal kernels and almost-kernel vectors)
If were infinite dimensional, it would contain an orthonormal sequence . Since , compactness would give a convergent subsequence of , impossible because distinct terms are distance apart. Hence is finite dimensional.
Suppose no constant works. Then we can choose with and . Compactness gives a convergent subsequence of , and then converges along it to some unit vector , since is closed. Continuity gives , so , contradicting . If the estimate is vacuous.
Every image equals for the component . If converges and each is chosen in , the lower bound makes Cauchy. Its limit maps to the limit of the images, so is closed. The operator is compact, so the same argument applies to . For the theorem, orthogonality gives , and closedness yields the displayed range identity .
Proof X-ray
The lower bound belongs on the kernel complement
Compactness rules out almost-kernel unit vectors orthogonal to the actual kernel. It does not produce on all of when the kernel is nonzero. Projecting away the kernel gives both the closed-range proof and the unique minimum-norm solution once compatibility holds.
Why the two nullities agree
The lemma gives closed range and finite nullities, but it does not compare with . That comparison uses an integer that cannot jump under small perturbations.
Definition 2.6.3
(Index of an operator with finite kernel and cokernel)
Let be bounded, with finite-dimensional kernel and closed range of finite codimension. Its index is
For , closed range gives , so the index is , and the theorem's claim is exactly that this index is zero.
Lemma 2.6.4
(Local constancy of the index)
Let have finite-dimensional kernel and closed range of finite codimension. Then every bounded with sufficiently small has the same properties, and .
Proof
(Block factorization near )
Put , , . The restriction is a bounded bijection: injectivity follows from the definition of , and every image has a preimage in after its kernel component is removed. Since is closed, hence Hilbert, the bounded inverse theorem gives a bounded inverse.
Use the decompositions in the domain and in the codomain, and write
The block satisfies because orthogonal projections have norm at most one. Hence stays invertible for sufficiently small perturbations by the Neumann-series criterion. With , direct multiplication gives
The outer factors are bounded isomorphisms. Since are finite dimensional, has closed range, and this factorization shows that has closed range, , and its range codimension is . Finite-dimensional rank–nullity therefore gives
This also covers zero-dimensional summands.
With both lemmas available, the equality of nullities follows by moving from the identity to along a continuous path of operators of the same form.
Proof
(From the identity to along a path)
Put for . For every , and its adjoint are compact, so the kernel-and-range lemma applies to and ; in particular is defined. The path is norm-continuous, and by the local-constancy lemma its integer-valued index is locally constant in . A locally constant function on the connected interval is constant. Since has index zero, has index zero, proving .
If , the range identity makes onto. Equal nullities then make it injective, and the bounded inverse theorem applies. Conversely, an invertible has both kernels zero. For a general solution , its orthogonal projection onto has the same image and is the unique minimum-norm solution.
Proof X-ray
What the path argument contributes
The compactness argument supplies closed range and finite-dimensional kernels at every point of the path. The block factorization then reduces small perturbations to a finite-dimensional map, whose kernel-minus-cokernel dimension is fixed by rank–nullity. Starting at the identity fixes the index at zero. No self-adjointness is used.
Examples and a failure without compactness
Worked example 2.6.5
(A non-self-adjoint rank-one operator)
On define . Its range is one dimensional, so it is compact. Then
For the adjoint, gives , whose kernel is . Thus is solvable exactly when . In that case every solution has , , and for . Minimizing gives the minimum-norm choice . The kernels have equal dimension but are not equal. For example fails the compatibility condition.
Worked example 2.6.6
(Diagonal resonance)
Let on . Finite-rank truncations converge to in operator norm, so is compact. For , if is a positive integer, its kernel and adjoint kernel are both ; the equation is solvable exactly when . For , set and choose freely. The nonzero denominators are bounded away from zero, since , so the resulting sequence is in . If is not a positive integer, all denominators are nonzero and bounded away from zero, and is invertible. This example includes complex ; the adjoint of is .
Worked example 2.6.7
(A self-adjoint rank-one case where the kernels coincide)
Let be a unit vector, let be a real scalar, and put . Then .
If , try . Since ,
and makes the bracket zero. So is invertible with ; both kernels are zero.
The edge case gives , where is the orthogonal projection onto . Now , so . The equation is solvable exactly when ; then the solutions are , and the minimum-norm solution is itself. Here the kernels coincide, unlike the first example.
Two errors recur when the theorem is applied to these examples.
Common mistake
(Testing compatibility against the wrong kernel)
The condition is , not . In the first example, is orthogonal to , yet , so has no solution. Conversely is not orthogonal to , but . Only the dimensions of the two kernels are guaranteed to agree.
Common mistake
(Expecting the compact operator itself to have closed range)
The closed-range conclusion is about . The compact operator has dense range, because its range contains every finitely supported sequence. Its range is not closed: lies in , but would force for all , which is not in .
The identity part of is essential, but so is compactness of the perturbation: a closed range by itself does not force the two nullities to agree.
Counterexample mode
A closed range need not have index zero
The right shift has zero kernel and closed range of codimension one. Its adjoint has kernel . Thus , and cannot be written as identity minus a compact operator. Closedness alone does not imply the equal-nullity conclusion.
Practice with solutions
Quick check
(For the rank-one operator and , decide whether and are solvable.)
Use the compatibility condition rather than solving coordinate by coordinate.
Solution · Answer
Here , so is solvable exactly when . For the second coordinate is , so there is no solution; indeed the second coordinate of every is zero. For the condition holds, and is one solution since .
Quick check
(For the same operator and , find the minimum-norm solution of .)
Parametrize all solutions by the kernel direction.
Solution · Answer
All solutions are . The squared norm is minimized at , giving . As a check, , so lies in , as the theorem predicts.
Quick check
(For on , describe all solutions of .)
Work coordinatewise: .
Solution · Answer
The coefficient vanishes only at , so solvability requires . Then for (for instance ), while is arbitrary. The coefficients for have absolute value at least , so . The minimum-norm solution takes .
Quick check
(Why is a nonzero kernel of a compatibility condition rather than merely a statement about uniqueness?)
Compare with .
Solution · Answer
If , then for every , so every image is orthogonal to . A right-hand side with a nonzero component in cannot be reached. The kernel of instead describes nonuniqueness among solutions once one exists. The two conditions are linked only through the equality of their dimensions.