Fredholm alternative

If V is an n-dimensional vector space and ${\displaystyle T:V\to V}$  is a linear transformation, then exactly one of the following holds:

1. For each vector v in V there is a vector u in V so that ${\displaystyle T(u)=v}$ . In other words: T is surjective (and so also bijective, since V is finite-dimensional).
2. ${\displaystyle \dim(\ker(T))>0.}$ 

A more elementary formulation, in terms of matrices, is as follows. Given an m×n matrix A and a m×1 column vector b, exactly one of the following must hold:

1. Either: A x = b has a solution x
2. Or: A^T y = 0 has a solution y with y^Tb ≠ 0.

In other words, A x = b has a solution ${\displaystyle (\mathbf {b} \in \operatorname {Im} (A))}$  if and only if for any y such that A^T y = 0, it follows that y^Tb = 0 ${\displaystyle (i.e.,\mathbf {b} \in \ker(A^{T})^{\bot })}$ .

Let ${\displaystyle K(x,y)}$  be an integral kernel, and consider the homogeneous equation, the Fredholm integral equation,

${\displaystyle \lambda \varphi (x)-\int _{a}^{b}K(x,y)\varphi (y)\,dy=0}$ 

and the inhomogeneous equation

${\displaystyle \lambda \varphi (x)-\int _{a}^{b}K(x,y)\varphi (y)\,dy=f(x).}$ 

The Fredholm alternative is the statement that, for every non-zero fixed complex number ${\displaystyle \lambda \in \mathbb {C} ,}$  either the first equation has a non-trivial solution, or the second equation has a solution for all ${\displaystyle f(x)}$ .

A sufficient condition for this statement to be true is for ${\displaystyle K(x,y)}$  to be square integrable on the rectangle ${\displaystyle [a,b]\times [a,b]}$  (where a and/or b may be minus or plus infinity). The integral operator defined by such a K is called a Hilbert–Schmidt integral operator.

Results about Fredholm operators generalize these results to complete normed vector spaces of infinite dimensions; that is, Banach spaces.

The integral equation can be reformulated in terms of operator notation as follows. Write (somewhat informally) $${\displaystyle T=\lambda -K}$$  to mean $${\displaystyle T(x,y)=\lambda \;\delta (x-y)-K(x,y)}$$  with ${\displaystyle \delta (x-y)}$  the Dirac delta function, considered as a distribution, or generalized function, in two variables. Then by convolution, ${\displaystyle T}$  induces a linear operator acting on a Banach space ${\displaystyle V}$  of functions ${\displaystyle \varphi (x)}$  $${\displaystyle V\to V}$$  given by $${\displaystyle \varphi \mapsto \psi }$$  with ${\displaystyle \psi }$  given by $${\displaystyle \psi (x)=\int _{a}^{b}T(x,y)\varphi (y)\,dy=\lambda \;\varphi (x)-\int _{a}^{b}K(x,y)\varphi (y)\,dy.}$$ 

In this language, the Fredholm alternative for integral equations is seen to be analogous to the Fredholm alternative for finite-dimensional linear algebra.

The operator ${\displaystyle K}$  given by convolution with an ${\displaystyle L^{2}}$  kernel, as above, is known as a Hilbert–Schmidt integral operator. Such operators are always compact. More generally, the Fredholm alternative is valid when ${\displaystyle K}$  is any compact operator. The Fredholm alternative may be restated in the following form: a nonzero ${\displaystyle \lambda }$  either is an eigenvalue of ${\displaystyle K,}$  or lies in the domain of the resolvent $${\displaystyle R(\lambda ;K)=(K-\lambda \operatorname {Id} )^{-1}.}$$ 

The Fredholm alternative can be applied to solving linear elliptic boundary value problems. The basic result is: if the equation and the appropriate Banach spaces have been set up correctly, then either

(1) The homogeneous equation has a nontrivial solution, or

(2) The inhomogeneous equation can be solved uniquely for each choice of data.

The argument goes as follows. A typical simple-to-understand elliptic operator L would be the Laplacian plus some lower order terms. Combined with suitable boundary conditions and expressed on a suitable Banach space X (which encodes both the boundary conditions and the desired regularity of the solution), L becomes an unbounded operator from X to itself, and one attempts to solve

${\displaystyle Lu=f,\qquad u\in \operatorname {dom} (L)\subseteq X,}$ 

where f ∈ X is some function serving as data for which we want a solution. The Fredholm alternative, together with the theory of elliptic equations, will enable us to organize the solutions of this equation.

A concrete example would be an elliptic boundary-value problem like

${\displaystyle (*)\qquad Lu:=-\Delta u+h(x)u=f\qquad {\text{in }}\Omega ,}$ 

supplemented with the boundary condition

${\displaystyle (**)\qquad u=0\qquad {\text{on }}\partial \Omega ,}$ 

where Ω ⊆ Rⁿ is a bounded open set with smooth boundary and h(x) is a fixed coefficient function (a potential, in the case of a Schrödinger operator). The function f ∈ X is the variable data for which we wish to solve the equation. Here one would take X to be the space L²(Ω) of all square-integrable functions on Ω, and dom(L) is then the Sobolev space W ^2,2(Ω) ∩ W^1,2
₀(Ω), which amounts to the set of all square-integrable functions on Ω whose weak first and second derivatives exist and are square-integrable, and which satisfy a zero boundary condition on ∂Ω.

If X has been selected correctly (as it has in this example), then for μ₀ >> 0 the operator L + μ₀ is positive, and then employing elliptic estimates, one can prove that L + μ₀ : dom(L) → X is a bijection, and its inverse is a compact, everywhere-defined operator K from X to X, with image equal to dom(L). We fix one such μ₀, but its value is not important as it is only a tool.

We may then transform the Fredholm alternative, stated above for compact operators, into a statement about the solvability of the boundary-value problem (*)–(**). The Fredholm alternative, as stated above, asserts:

• For each λ ∈ R, either λ is an eigenvalue of K, or the operator K − λ is bijective from X to itself.

Let us explore the two alternatives as they play out for the boundary-value problem. Suppose λ ≠ 0. Then either

(A) λ is an eigenvalue of K ⇔ there is a solution h ∈ dom(L) of (L + μ₀) h = λ⁻¹h ⇔ –μ₀+λ⁻¹ is an eigenvalue of L.

(B) The operator K − λ : X → X is a bijection ⇔ (K − λ) (L + μ₀) = Id − λ (L + μ₀) : dom(L) → X is a bijection ⇔ L + μ₀ − λ⁻¹ : dom(L) → X is a bijection.

Replacing -μ₀+λ⁻¹ by λ, and treating the case λ = −μ₀ separately, this yields the following Fredholm alternative for an elliptic boundary-value problem:

• For each λ ∈ R, either the homogeneous equation (L − λ) u = 0 has a nontrivial solution, or the inhomogeneous equation (L − λ) u = f possesses a unique solution u ∈ dom(L) for each given datum f ∈ X.

The latter function u solves the boundary-value problem (*)–(**) introduced above. This is the dichotomy that was claimed in (1)–(2) above. By the spectral theorem for compact operators, one also obtains that the set of λ for which the solvability fails is a discrete subset of R (the eigenvalues of L). The eigenvalues’ associated eigenfunctions can be thought of as "resonances" that block the solvability of the equation.

• Spectral theory of compact operators
• Farkas' lemma

• Fredholm, E. I. (1903). "Sur une classe d'equations fonctionnelles". Acta Math. 27: 365–390. doi:10.1007/bf02421317.
• A. G. Ramm, "A Simple Proof of the Fredholm Alternative and a Characterization of the Fredholm Operators", American Mathematical Monthly, 108 (2001) p. 855.
• Khvedelidze, B.V. (2001) [1994], "Fredholm theorems", Encyclopedia of Mathematics, EMS Press
• "Fredholm alternative", Encyclopedia of Mathematics, EMS Press, 2001 [1994]