The Fredholm Alternative: Either Solve Always or Solve Almost Never

What the theorem actually claims

Let X be a Banach space and K : X → X a compact linear operator (K maps bounded sets to relatively compact sets). Set T = I − K. The Fredholm alternative asserts a bundle of facts, all equivalent to the single slogan ind(T) = 0:

• Fredholm property: ker(T) is finite-dimensional and ran(T) is closed with finite codimension.
• Index zero: dim ker(T) = codim ran(T) = dim ker(T*), where T* is the adjoint (Banach transpose).
• Dichotomy: either T is bijective — so (I − K)x = y has a unique solution for every y, and (I − K)⁻¹ is bounded — or ker(T) ≠ {0}, in which case (I − K)x = y is solvable precisely for those y satisfying ⟨y, φ⟩ = 0 for every φ ∈ ker(T*).

The name captures the flavor: solvability is all-or-conditional. You never get "generic solvability with occasional failure." Uniqueness for one y forces existence for all y.

The intuition: compact ≈ finite-dimensional

In ℝⁿ every square matrix A satisfies the rank–nullity dichotomy: A is injective iff surjective, and Ax = b solves iff b ⊥ ker(Aᵀ). The Fredholm alternative says a compact perturbation of the identity behaves exactly like a finite matrix, even in infinite dimensions.

Why should it? A compact operator K is, in a precise sense, "almost finite-rank": it is the norm-limit of finite-rank operators (on Hilbert space, always; on Banach spaces with the approximation property). Its image concentrates in a nearly finite-dimensional set — bounded sequences get squeezed so that Kxₙ has convergent subsequences. The identity contributes the infinite-dimensional "bulk" that is invertible, while K contributes a finite-dimensional "defect." The eigenvalue λ = 1 of K can only produce a finite-dimensional eigenspace, and the failure of I − K to be onto is confined to a matching finite-dimensional obstruction. Infinite dimensions add no new pathology at this particular point of the spectrum.

The mechanism of the proof (Riesz–Schauder)

The engine is a sequence of lemmas due to F. Riesz. (1) Finite-dimensional kernel: on ker(T) = ker(I − K) we have Kx = x, so K restricts to the identity there; but K is compact and the identity is compact only on finite-dimensional spaces, forcing dim ker(T) < ∞.

(2) Closed range. One shows ‖Tx‖ ≥ c·dist(x, ker T) for some c > 0. If not, a normalized sequence xₙ (modulo the kernel) has Txₙ → 0; compactness of K extracts Kxₙ → z, hence xₙ = Txₙ + Kxₙ → z with Tz = 0 and ‖z‖ = 1 in the quotient — contradiction. This coercivity gives closedness.

(3) Index zero. Riesz's lemma shows the ascending chains ker(Tⁿ) and descending ran(Tⁿ) stabilize (finite ascent and descent), so X = ker(Tᵏ) ⊕ ran(Tᵏ). Comparing dimensions across this splitting, plus the duality identity ran(T)⊥ = ker(T*), yields dim ker(T) = dim ker(T*) and codim ran(T) = dim ker(T). Solvability of Tx = y ⇔ y annihilates ker(T*) is then just the closed-range theorem: ran(T) = ⊥ker(T*).

Canonical example: the Fredholm integral equation

Fredholm's original setting. On X = C[a,b] or L²[a,b], take (Kx)(s) = ∫ₐᵇ k(s,t) x(t) dt with a continuous (or L²) kernel k. Such K is compact (Arzelà–Ascoli / Hilbert–Schmidt). The equation of the second kind

x(s) − λ ∫ₐᵇ k(s,t) x(t) dt = y(s)

is exactly (I − λK)x = y. The alternative says: for a fixed λ, either the homogeneous equation x = λKx has only x ≡ 0 — then the inhomogeneous equation has a unique continuous solution for every data y — or λ⁻¹ is an eigenvalue of K, the homogeneous equation has a finite basis φ₁,…,φₙ of solutions, and the inhomogeneous equation is solvable iff ∫ₐᵇ y(t) ψⱼ(t) dt = 0 for each solution ψⱼ of the transposed equation ψ = λK*ψ. This is precisely how one proves solvability of boundary-value problems after converting them to integral equations via a Green's function.

Why compactness is essential — and what breaks

Drop compactness and the alternative collapses instantly. Let X = ℓ² and let S be the right shift, S(x₁, x₂, …) = (0, x₁, x₂, …). Then S is an isometry: injective but not surjective — ind(S) = −1. Its adjoint, the left shift S*, is surjective but not injective — ind(S*) = +1. Neither is I − K for compact K; both flatly violate "injective ⇔ surjective." So the alternative genuinely fails for general bounded operators.

The right generalization keeps only the essence: T is a Fredholm operator if dim ker T < ∞ and codim ran T < ∞, with index ind(T) = dim ker T − codim ran T. The index is a homotopy invariant, stable under compact perturbation (Atkinson's theorem): ind(T + K) = ind(T). Compact perturbations of the identity have ind = ind(I) = 0 — that zero index is the Fredholm alternative. Nonzero-index operators (like the shift) live outside it.

Why it matters: existence theory for PDEs and beyond

The Fredholm alternative is the workhorse of linear existence theory. For an elliptic boundary-value problem Lu = f (e.g. −Δu + cu = f on a domain Ω with u|∂Ω = 0), the solution operator of the principal part is compact (Rellich–Kondrachov gives compact Sobolev embeddings), converting the problem to (I − K)u = g. The alternative then reads: either Lu = f is uniquely solvable for all f, or 0 is an eigenvalue of L and f must be orthogonal to the finite-dimensional eigenspace — the classic "solvability condition" seen in the Poisson–Neumann problem (∫ f = 0) and in resonance for forced oscillators.

It underlies the Riesz–Schauder spectral theory of compact operators (nonzero spectrum is discrete eigenvalues of finite multiplicity), the Fredholm theory of the resolvent, and — via the index — the Atiyah–Singer index theorem, which turns the analytic index of an elliptic operator into a topological invariant. From bifurcation theory to boundary integral methods, it is the first tool one reaches for.

The dichotomy for T = I − K with K compact: exactly one column holds for each value of λ = 1.

Property	Case A: T injective	Case B: T not injective
ker(I − K)	{0}	finite-dim, dimension n ≥ 1
ker(I − K*)	{0}	same dimension n
Range of I − K	all of X (surjective)	closed, codimension n
(I − K)x = y	unique solution ∀ y	solvable iff y ⊥ ker(I − K*)
Bounded inverse	(I − K)⁻¹ exists, bounded	no inverse