The equation $Lu = f$ — a linear operator applied to an unknown equals a given datum — is the central problem of linear PDE theory. When $L$ is coercive (or can be shifted to be so), the [Lax–Milgram theorem](/theorems/91) guarantees a unique solution for every $f$. But many natural problems are not coercive: the operator may have a nontrivial kernel, and existence fails for generic right-hand sides. The question is then not *whether* the equation is solvable, but *exactly when* it is solvable, and *how large* the failure of uniqueness and existence can be.

In finite dimensions, the answer is given by the rank–nullity theorem: a square matrix $A$ is either invertible (trivial kernel, full range) or singular (nontrivial kernel, range of codimension equal to the dimension of the kernel), and solvability of $Ax = b$ in the singular case is controlled by orthogonality to the kernel of $A^T$. In infinite dimensions, this tidy picture breaks down for general bounded operators — the range can fail to be closed, the left and right kernels can have different dimensions, and injectivity need not imply surjectivity (the right shift on $\ell^2$ is injective but not surjective). The Fredholm Alternative is the theorem that restores the finite-dimensional picture for operators of the form $I - K$ with $K$ [compact](/page/Linear%20Operators%20on%20Banach%20Spaces). Compactness is the hypothesis that forces the operator to behave like a matrix: the kernel is finite-dimensional, the range is closed, the defect is the same on both sides, and the dichotomy between invertibility and finite-dimensional failure is exact.

## Motivation

[motivation] ### The Vibrating Membrane Consider a membrane stretched over a bounded domain $U \subset \mathbb{R}^2$, fixed at the [boundary](/page/Boundary). The eigenvalue problem for the Laplacian, \begin{align*} -\Delta w &= \lambda w \quad \text{in } U, \\ w &= 0 \quad \text{on } \partial U, \end{align*} has a discrete [sequence](/page/Sequence) of eigenvalues $0 < \lambda_1 < \lambda_2 \leq \lambda_3 \leq \cdots \to \infty$ with corresponding eigenfunctions $w_k$ — the resonant modes of the membrane. Now suppose we apply an external forcing $f$ and ask whether the steady-state problem \begin{align*} -\Delta u - \lambda u &= f \quad \text{in } U, \\ u &= 0 \quad \text{on } \partial U \end{align*} has a solution. If $\lambda$ is not an eigenvalue, one expects a unique solution for every $f$ — the system is not at resonance, and the forcing produces a well-defined response. If $\lambda = \lambda_k$ for some $k$, the homogeneous equation $-\Delta u - \lambda_k u = 0$ has nontrivial solutions (the eigenfunctions $w_k$), so uniqueness fails. But existence does not fail completely: a solution still exists provided the forcing $f$ does not "feed energy into" the resonant mode, which is made precise by the condition $\int_U f \, w_k \, d\mathcal{L}^2 = 0$ for every eigenfunction $w_k$ associated to $\lambda_k$. ### What Could Go Wrong in Infinite Dimensions The vibrating membrane example might suggest that the situation is always this clean. But for a general bounded operator $T$ on an infinite-dimensional Hilbert space, several pathologies can occur that have no finite-dimensional analogue. The kernel of $T$ might be infinite-dimensional. The range of $T$ might fail to be closed, in which case the orthogonal complement of the range does not capture the solvability obstruction. The dimensions of the kernels of $T$ and $T^*$ might differ — the "defect" on the left and right sides could be different. Any of these would make a clean solvability criterion impossible. ### Compactness Restores the Finite-Dimensional Picture The hypothesis that saves the theory is compactness. If $K: H \to H$ is a [compact operator](/page/Linear%20Operators%20on%20Banach%20Spaces) and $T = I - K$, then $T$ behaves like a finite-dimensional perturbation of the identity: its kernel is finite-dimensional, its range is automatically closed, the kernel dimensions of $T$ and $T^*$ are equal, and injectivity implies surjectivity. The analogy with matrices is not a coincidence — compact operators are precisely those that can be approximated by finite-rank operators, and the Fredholm Alternative is the [limit](/page/Limit) of the rank–nullity theorem along this approximation. In the PDE context, compactness enters through the [Rellich–Kondrachov theorem](/theorems/64): the embedding $H^1_0(U) \hookrightarrow L^2(U)$ is compact for bounded $U$, which makes the resolvent operator $K = j \circ L_\gamma^{-1} \circ \iota$ a compact operator on $L^2(U)$. Rewriting the elliptic equation $Lu = f$ in the form $(I - \gamma K)u = Kf$ puts it into the framework of the abstract Fredholm Alternative. The membrane example above is the simplest instance of this reduction. [/motivation]

## Setup

The Fredholm Alternative concerns equations of the form $Tu = f$ where $T = I - K$ for a compact operator $K$ on a Hilbert space. We recall the definitions of the two key ingredients; full developments are on the [Linear Operators on Banach Spaces](/page/Linear%20Operators%20on%20Banach%20Spaces) and [Hilbert Space](/page/Hilbert%20Space) pages.

A bounded linear operator $K: H \to H$ on a Hilbert space is **compact** if every bounded sequence $\{u_n\} \subseteq H$ has a subsequence for which $\{Ku_{n_k}\}$ converges in $H$. Compact operators form a closed two-sided ideal $\mathcal{K}(H) \subseteq \mathcal{L}(H)$ in the operator norm. The **adjoint** $K^*: H \to H$ is the unique bounded linear operator satisfying $(Ku, v)_H = (u, K^*v)_H$ for all $u, v \in H$; if $K$ is compact, so is $K^*$.

## The Theorem

[quotetheorem:72]

The four conclusions of the theorem have distinct roles, and it is worth understanding each in turn.

**Conclusion (1): finite-dimensionality of the kernel.** The kernel $\ker(I - K)$ is the space of solutions to the homogeneous equation $u = Ku$ — the "eigenspace at eigenvalue $1$" of $K$. For a general bounded operator, this kernel could be infinite-dimensional (consider the identity operator, whose kernel under $I - I = 0$ is the entire space). Compactness of $K$ forces the kernel to be finite-dimensional: the closed unit ball of $\ker(I - K)$ coincides with its image under $K$ (since $u = Ku$ on the kernel), so it is precompact, which by the Riesz lemma characterisation means the space is finite-dimensional. In the PDE context, this says that the eigenspace of the elliptic operator at any given eigenvalue is finite-dimensional — there are only finitely many linearly independent resonant modes at each frequency.

**Conclusion (2): closed range and the solvability condition.** The identity $\operatorname{Range}(I - K) = \ker(I - K^*)^\perp$ is the heart of the theorem from the applications point of view. It says that the equation $(I - K)u = f$ is solvable if and only if $f$ is orthogonal to every solution of the adjoint homogeneous equation $(I - K^*)v = 0$. This is a concrete, checkable criterion: to determine whether a given $f$ admits a solution, one solves the adjoint homogeneous problem, finds a basis $v_1, \ldots, v_d$ for $\ker(I - K^*)$, and checks whether $(f, v_j)_H = 0$ for each $j$. (The adjoint kernel is finite-dimensional because $K^*$ is compact whenever $K$ is, so conclusion (1) applies to $T^* = I - K^*$ as well.) In the PDE context, this becomes the condition $\int_U f \, v_j \, d\mathcal{L}^n = 0$ where $v_j$ are the eigenfunctions of the adjoint operator — exactly the orthogonality condition in the membrane example. The closedness of the range is what makes this work: for a general operator, the range need not be closed, and the orthogonal complement of the range would not capture the full solvability obstruction.

**Conclusion (3): dimension equality.** The equality $\dim \ker(I - K) = \dim \ker(I - K^*)$ ties conclusions (1) and (2) together quantitatively. Conclusion (2) says the range of $T = I - K$ has codimension $\dim \ker(I - K^*)$ inside $H$ (since $\operatorname{Range}(T) = \ker(T^*)^\perp$, the codimension is $\dim \ker(T^*)$). Conclusion (3) says this codimension equals $\dim \ker(T)$. The practical consequence is: if the homogeneous equation $(I - K)u = 0$ has a $d$-dimensional solution space, then the solvability of $(I - K)u = f$ is controlled by exactly $d$ orthogonality conditions on $f$ (one for each basis element of $\ker(I - K^*)$), and when those conditions are satisfied, the solution is determined up to an arbitrary element of the $d$-dimensional kernel. The number of "degrees of freedom lost" in uniqueness is exactly the number of "conditions gained" for existence — no more, no less. When $K$ is self-adjoint ($K = K^*$), conclusion (3) is immediate because $\ker(T) = \ker(T^*)$. In the non-self-adjoint case, the equality requires the full power of compactness — the proof constructs a finite-rank perturbation that converts a hypothetical dimension mismatch into a contradiction via Step 4 (injectivity implies surjectivity).

**Conclusion (4): the dichotomy.** The two alternatives are mutually exclusive and exhaustive. Either $\ker(I - K) = \{0\}$ — in which case the three preceding conclusions combine to give that $I - K$ is bijective with bounded inverse (the [Open Mapping Theorem](/theorems/631) supplies the bounded inverse) — or $\ker(I - K) \neq \{0\}$, and solvability is governed by the finitely many orthogonality conditions from conclusion (2). There is no third possibility: the equation cannot be "partially solvable" with an infinite-dimensional obstruction, nor can the kernel be nontrivial but the range still be all of $H$. In the membrane language: at a non-resonant frequency, the problem is uniquely solvable for every forcing; at a resonant frequency, the problem is solvable precisely when the forcing is orthogonal to the resonant modes.

## Examples

### All Four Conclusions in a PDE

The following example shows how each of the four conclusions manifests concretely for an elliptic boundary value problem.

[example: Helmholtz Equation At Resonance In One Dimension] Let $U = (0, \pi) \subset \mathbb{R}$ and consider the Dirichlet eigenvalue problem for $-u''$ on $H^1_0((0, \pi))$. The eigenvalues and normalised eigenfunctions are \begin{align*} \lambda_k &= k^2, \quad w_k(x) = \sqrt{\frac{2}{\pi}} \sin(kx), \quad k = 1, 2, 3, \ldots \end{align*} Now consider the Helmholtz equation at the first eigenvalue $\lambda_1 = 1$: \begin{align*} -u'' - u &= f \quad \text{in } (0, \pi), \\ u(0) = u(\pi) &= 0, \end{align*} for a given $f \in L^2((0, \pi))$. The operator $L = -\partial_{xx}$ is self-adjoint on $H^1_0((0, \pi))$, and the equation $Lu - u = f$ can be rewritten in Fredholm form $(I - \tilde{K})u = h$ by shifting: choose $\gamma = 2$ so that $L_\gamma = -\partial_{xx} + 2I$ is coercive. The resolvent $K = j \circ L_\gamma^{-1} \circ \iota: L^2 \to L^2$ is compact and self-adjoint (since $L$ is self-adjoint), with $\tilde{K} = (\gamma + 1)K = 3K$. The four conclusions of the Fredholm Alternative give the following. **Conclusion (1)** (finite-dimensional kernel): The kernel of $I - \tilde{K}$ corresponds to the eigenspace of $-\partial_{xx}$ at $\lambda_1 = 1$. The homogeneous equation $-u'' - u = 0$ with Dirichlet conditions has general solution $u(x) = A\sin(x) + B\cos(x)$; the boundary condition $u(0) = 0$ forces $B = 0$, and $u(\pi) = 0$ is automatically satisfied. So \begin{align*} \ker(I - \tilde{K}) &= \operatorname{span}\{\sin\}, \end{align*} which is one-dimensional, as guaranteed by the theorem. **Conclusion (2)** (solvability condition): The range of $I - \tilde{K}$ is the orthogonal complement of the adjoint kernel. Since $L$ is self-adjoint, $\tilde{K}$ is self-adjoint, so $\ker(I - \tilde{K}^*) = \ker(I - \tilde{K}) = \operatorname{span}\{\sin\}$. Therefore \begin{align*} \operatorname{Range}(I - \tilde{K}) &= \{f \in L^2((0, \pi)) : (f, \sin)_{L^2} = 0\} = \{\sin\}^\perp. \end{align*} The equation $-u'' - u = f$ is solvable if and only if \begin{align*} \int_0^\pi f(x) \sin(x) \, d\mathcal{L}^1(x) &= 0. \end{align*} For example, $f(x) = \sin(2x)$ satisfies this condition (since $\int_0^\pi \sin(2x) \sin(x) \, d\mathcal{L}^1(x) = 0$ by orthogonality of the eigenfunctions), so $-u'' - u = \sin(2x)$ is solvable. One can verify directly: substituting $u(x) = \frac{1}{3}\sin(2x)$ gives $-u'' - u = \frac{4}{3}\sin(2x) - \frac{1}{3}\sin(2x) = \sin(2x)$, confirming the solution. On the other hand, $f(x) = \sin(x)$ fails the condition ($\int_0^\pi \sin^2(x) \, d\mathcal{L}^1(x) = \pi/2 \neq 0$), so $-u'' - u = \sin(x)$ has no solution — the forcing resonates with the eigenmode. **Conclusion (3)** (dimension equality): Here $\dim \ker(I - \tilde{K}) = 1$ and $\dim \ker(I - \tilde{K}^*) = 1$, so there is exactly one solvability condition on $f$ (orthogonality to $\sin$), and when it is satisfied the solution is unique up to a one-dimensional family. Concretely, if $u_0$ is any particular solution of $-u'' - u = f$, then the general solution is \begin{align*} u &= u_0 + c \sin, \quad c \in \mathbb{R}. \end{align*} One free parameter in the solution, one condition on the data — these are matched by conclusion (3). **Conclusion (4)** (the dichotomy): Since $\ker(I - \tilde{K}) = \operatorname{span}\{\sin\} \neq \{0\}$, we are in the resonance case. If instead we considered $\lambda = 1.5$ (not an eigenvalue), the kernel would be trivial, and the theorem would give unique solvability of $-u'' - 1.5 \, u = f$ for every $f \in L^2((0, \pi))$. [/example]

### Alternative A: Unique Solvability

[example: Volterra Integral Equation] Let $H = L^2([0, 1])$ and define the Volterra [integral](/page/Integral) operator \begin{align*} K: L^2([0, 1]) &\to L^2([0, 1]) \\ u &\mapsto \int_0^x u(y) \, d\mathcal{L}^1(y). \end{align*} The operator $K$ is compact (it maps bounded [sets](/page/Set) in $L^2$ to bounded equicontinuous subsets of $C([0,1])$, which are precompact in $L^2$ by the Arzelà–Ascoli theorem and the compactness of the embedding $C \hookrightarrow L^2$). We check whether $\ker(I - K) = \{0\}$. Suppose $(I - K)u = 0$, i.e. \begin{align*} u(x) &= \int_0^x u(y) \, d\mathcal{L}^1(y) \quad \text{for a.e. } x \in [0,1]. \end{align*} The right-hand side is absolutely [continuous](/page/Continuity), so $u$ is absolutely continuous with $u(0) = 0$. Differentiating both sides gives $u'(x) = u(x)$ with initial condition $u(0) = 0$. The unique solution to this ODE is $u \equiv 0$, so $\ker(I - K) = \{0\}$. By the Fredholm Alternative (conclusion 4, first case), $I - K: L^2([0,1]) \to L^2([0,1])$ is a bijection with bounded inverse. The equation \begin{align*} u(x) - \int_0^x u(y) \, d\mathcal{L}^1(y) &= f(x) \end{align*} has a unique solution $u \in L^2([0,1])$ for every $f \in L^2([0,1])$. [/example]