Definition

Let H0,H1H_0,H_1 be . A T:H0H1T:H_0\to H_1 is Fredholm if its range is closed and both kerT\ker T and the cokernel H1/ranTH_1/\operatorname{ran}T are finite-dimensional. Equivalently, kerT\ker T and kerT\ker T^* are finite-dimensional and ranT\operatorname{ran}T is closed. Its Fredholm index is

ind(T)=dimkerTdimkerT.\operatorname{ind}(T)=\dim\ker T-\dim\ker T^*.

The closed-range hypothesis is essential: finite-dimensional kernel and dense, nonclosed range do not make an operator Fredholm.

Atkinson characterization

Atkinson's theorem says that TT is Fredholm exactly when it is invertible modulo . Concretely, this means there is a bounded S:H1H0S:H_1\to H_0 such that

STIH0andTSIH1ST-I_{H_0} \quad\text{and}\quad TS-I_{H_1}

are compact. Such an SS is called a parametrix modulo compact operators. When H0=H1=HH_0=H_1=H, the image of TT is therefore invertible in the Calkin algebra B(H)/K(H)B(H)/K(H).

Stability and index

Fredholmness is open in the , and the index is locally constant on the set of Fredholm operators. If K:H0H1K:H_0\to H_1 is compact, then

T+K is Fredholmandind(T+K)=ind(T).T+K\ \text{is Fredholm} \qquad\text{and}\qquad \operatorname{ind}(T+K)=\operatorname{ind}(T).

If T:H0H1T:H_0\to H_1 and S:H1H2S:H_1\to H_2 are Fredholm, then

ind(ST)=ind(S)+ind(T).\operatorname{ind}(ST)=\operatorname{ind}(S)+\operatorname{ind}(T).

These facts make the index insensitive to many analytic perturbations while retaining global information.

Example

The unilateral shift U:2(N)2(N)U:\ell^2(\mathbb N)\to\ell^2(\mathbb N),

U(x0,x1,)=(0,x0,x1,),U(x_0,x_1,\ldots)=(0,x_0,x_1,\ldots),

has zero-dimensional kernel and a one-dimensional cokernel. Hence UU is Fredholm and ind(U)=1\operatorname{ind}(U)=-1.

Unbounded convention

A closed is also sometimes called Fredholm when it has closed range and finite-dimensional kernel and cokernel. That is a distinct unbounded notion: boundedness, domains, and adjoints must be handled explicitly. Under suitable regularity, an unbounded Fredholm operator can be represented by a bounded transform, but the two definitions should not be silently identified.

References