Integral equations ask for an unknown function that appears under an integral sign. Their structure is simultaneously analytic and operator-theoretic: kernels define operators, compactness shapes spectra, and solvability becomes a question about invertibility in function space.
This guide owns R27.07. It develops Fredholm and Volterra equations, kernel operators, compactness, Neumann series, resolvent kernels, finite-rank approximations and the Fredholm alternative. The aim is to show how integral equations become a bridge between analysis, operator theory, numerical approximation and boundary-value problems.
Prerequisites: Linear Operators and Spectra, Hilbert Spaces, and basic measure/integration.
Reading route: equation classes → kernel operators → Fredholm second kind → Neumann series → resolvent kernel → first kind ill-posedness → Volterra structure → successive approximation → compact operators → Fredholm alternative → self-adjoint kernels → Hilbert–Schmidt ideas → finite-rank kernels → degenerate kernels → numerical quadrature → conditioning and regularisation → applications → practice and solutions.
1. Integral equations are operator equations in disguise
A linear integral equation often has the form
u(x)−λ∫_a^b K(x,y)u(y)dy=f(x).
Define the integral operator
(Tu)(x)=∫_a^bK(x,y)u(y)dy.
Then the equation is simply
(I−λT)u=f.
Solving the integral equation means deciding whether I−λT is invertible on the chosen function space.
2. First-kind and second-kind equations have different stability
A first-kind equation is
Tu=f.
A second-kind equation is
u−λTu=f.
Second-kind equations often inherit stable invertibility away from the spectrum of T. First-kind equations with compact T are frequently ill-posed because inversion magnifies small singular values.
3. Fredholm equations integrate over a fixed interval
A Fredholm equation typically uses a fixed integration domain:
u(x)−λ∫_a^bK(x,y)u(y)dy=f(x).
The same global range [a,b] contributes for every x. This differs from Volterra equations, where the upper integration limit usually depends on x.
4. Volterra equations encode causal or triangular structure
A standard second-kind Volterra equation is
u(x)−λ∫_a^xK(x,y)u(y)dy=f(x).
Only values y≤x contribute. This triangular structure often gives stronger existence and uniqueness results than the corresponding Fredholm problem.
5. Continuous kernels define bounded operators on common spaces
If K is continuous on [a,b]², then T maps C([a,b]) to itself and
||Tu||∞≤(b−a) sup|K| · ||u||∞.
Thus T is bounded in the sup norm.
6. Square-integrable kernels define Hilbert–Schmidt operators
If K∈L²([a,b]²), then
(Tu)(x)=∫K(x,y)u(y)dy
defines a bounded operator on L², with a norm estimate controlled by ||K||_{L²}.
Such operators are Hilbert–Schmidt and therefore compact.
7. Compactness is the structural reason Fredholm theory resembles matrix algebra
Compact operators behave spectrally like infinite matrices with eigenvalues tending to zero.
For compact T, nonzero spectral points are eigenvalues of finite multiplicity, and zero is the only possible accumulation point.
This is the operator-theoretic foundation of the Fredholm alternative.
8. Small operator norm gives a direct Neumann-series solution
If
|λ| ||T||<1,
then
(I−λT)^{-1}=Σ_{n=0}^∞ λ^nT^n.
Hence
u=f+λTf+λ²T²f+⋯.
This is the operator form of successive substitution.
9. Iterated kernels make the Neumann series explicit
Define K₁=K and recursively
K_{n+1}(x,y)=∫_a^bK(x,z)K_n(z,y)dz.
Then T^n has kernel K_n, and the Neumann series can be rewritten as an integral with a resolvent kernel when the series converges.
10. The resolvent kernel packages the inverse
Formally define
R_λ(x,y)=K(x,y)+λK₂(x,y)+λ²K₃(x,y)+⋯.
Then the solution becomes
u(x)=f(x)+λ∫_a^bR_λ(x,y)f(y)dy.
The resolvent kernel is the integral-equation counterpart of the operator resolvent.
11. Worked rank-one kernel example
Let K(x,y)=p(x)q(y). Then
(Tu)(x)=p(x)∫q(y)u(y)dy.
The range of T lies in span{p}, so T has rank at most one.
For u−λTu=f, write c=∫q(y)u(y)dy. Then u=f+λcp. Integrating against q gives
c=∫qf + λc∫qp.
If 1−λ∫qp≠0, solve c explicitly and recover u.
12. Finite-rank kernels reduce to finite linear systems
If
K(x,y)=Σ_{j=1}^m p_j(x)q_j(y),
then Tu lies in the m-dimensional span of p_j.
Introduce coefficients c_j=∫q_j(y)u(y)dy. The integral equation reduces to an m×m linear system for the c_j.
This is the simplest exact bridge from integral equations to matrix algebra.
13. The Fredholm alternative separates unique solvability from homogeneous obstruction
For compact T on suitable Banach or Hilbert spaces, the second-kind equation
(I−λT)u=f
has a sharp alternative: either the homogeneous equation has only the zero solution and the inhomogeneous problem is uniquely solvable for every f, or nontrivial homogeneous solutions exist and solvability of the inhomogeneous equation requires compatibility conditions.
The exact compatibility statement is naturally expressed using the adjoint.
14. The adjoint determines compatibility conditions
On Hilbert space, if I−λT fails to be invertible because of a nontrivial kernel, then f must be orthogonal to the kernel of the adjoint operator associated with the obstructing spectral value.
For real λ and suitable conventions, a typical condition is
f ⟂ Ker(I−λ̄T*).
This is the infinite-dimensional analogue of consistency conditions for a singular linear system.
15. Self-adjoint kernels simplify the spectral picture
If K(x,y)=overline{K(y,x)} in an L² setting, then the associated operator is self-adjoint.
Its nonzero compact spectrum is real, eigenspaces for distinct eigenvalues are orthogonal, and an orthonormal eigenbasis is available on the relevant closed range structure.
Integral equations can then be solved by eigenfunction expansion away from reciprocal eigenvalues.
16. Eigenfunction expansion gives a diagonal solution formula
Suppose Tφ_n=μ_nφ_n with orthonormal eigenfunctions and f=Σf_nφ_n.
Then
(I−λT)u=f
gives coefficient equations
(1−λμ_n)u_n=f_n.
Hence u_n=f_n/(1−λμ_n) whenever denominators are nonzero.
17. Resonance occurs when λμ_n=1
If 1−λμ_n=0 for some n, the inverse coefficient formula blows up.
At that spectral value, the homogeneous equation has the nonzero solution φ_n.
The inhomogeneous problem is solvable only when f has no incompatible component in the corresponding adjoint eigendirection.
18. First-kind compact equations are often ill-posed
Suppose Tu=f with compact T and singular/eigenvalues μ_n→0.
Formally, u_n=f_n/μ_n.
Small μ_n amplify noise dramatically. Even when an inverse exists on the range, it may be unbounded.
This is a classic mechanism of ill-posed inverse problems.
19. Regularisation trades exact inversion for stability
Tikhonov regularisation replaces Tu=f by minimising
||Tu−f||²+α||u||²
with α>0.
The penalty suppresses unstable components associated with tiny singular values.
The parameter α controls the trade-off between fidelity and stability.
20. Volterra equations often admit global successive approximation
Because the integration domain is triangular, iterated Volterra kernels acquire factorial-type volume factors.
This often makes the Neumann series converge for every finite λ on bounded intervals, even when |λ|||T|| is not less than one.
The triangular geometry creates stronger convergence than a generic Fredholm operator.
21. Differentiating a Volterra equation can recover an ODE
Consider
u(x)=f(x)+∫_a^x u(y)dy.
Differentiating under suitable regularity gives
u'(x)=f'(x)+u(x).
The original equation also determines u(a)=f(a).
Thus Volterra integral equations can encode initial-value differential equations.
22. Green functions convert differential equations into integral equations
If L is a differential operator and G(x,y) a Green kernel adapted to boundary conditions, then solutions can often be written as
u(x)=∫G(x,y)f(y)dy.
More complicated equations may become second-kind integral equations after isolating a principal inverse.
R28 owns the differential-equation theory; R27.07 owns the operator/integral-equation formulation.
23. Boundary integral methods reduce dimension
For some elliptic boundary-value problems, representation formulas rewrite a PDE in a volume as an integral equation on its boundary.
This reduces geometric dimension but introduces singular kernels and operator-theoretic questions.
Such methods are powerful precisely because differential structure is transferred into a boundary operator.
24. Nyström methods discretise the integral directly
Choose quadrature nodes y_j and weights w_j. Approximate
∫K(x_i,y)u(y)dy≈Σ_j w_jK(x_i,y_j)u(y_j).
The integral equation becomes a linear system for sampled values u(y_j).
Accuracy depends on kernel regularity, quadrature order and conditioning of the resulting matrix.
25. Galerkin methods project the operator equation
Choose a finite-dimensional trial space V_N with basis φ_j. Seek u_N∈V_N such that the residual is orthogonal to a chosen test space.
This converts the operator equation into a matrix problem built from inner products such as
⟨(I−λT)φ_j,φ_i⟩.
The method is a direct continuation of Hilbert-space projection.
26. Degenerate-kernel approximation explains why finite matrices work
Approximate K(x,y) by a finite sum
K_N(x,y)=Σ_{j=1}^N p_j(x)q_j(y).
The resulting finite-rank operator is exactly reducible to linear algebra.
If K_N→K in an appropriate operator norm, the finite-dimensional solutions may converge to the true operator solution away from spectral singularities.
27. Conditioning deteriorates near spectral obstruction
Even when I−λT is invertible, the inverse norm can become large when λ approaches a reciprocal eigenvalue.
Small perturbations in f or numerical quadrature can then produce large changes in u.
Near-singularity is a stability issue before it becomes an exact solvability failure.
28. Resolvent identities help compare parameter values
For A(λ)=I−λT, resolvent identities relate inverses at different λ values whenever both exist.
This supports perturbation analysis and reveals how poles of the resolvent correspond to spectral obstructions for compact operators.
29. Verification should combine residual and spectral checks
- substitute the computed u back into the integral equation;
- measure the residual in the same function norm used by the theory;
- refine quadrature or basis size;
- compare against a known finite-rank or separable-kernel case;
- estimate distance to a problematic eigenvalue;
- test regularisation sensitivity for first-kind equations;
- compare direct and operator-series solutions when both are available.
30. Common theorem-boundary errors
- First-kind and second-kind integral equations have different stability properties.
- Neumann-series inversion requires a convergence argument; formal iteration alone is not proof.
- Compact does not mean finite rank.
- The Fredholm alternative applies under operator hypotheses, not to every arbitrary integral equation.
- A singular homogeneous problem can make the inhomogeneous problem incompatible or nonunique.
- Small residual does not guarantee small solution error in an ill-conditioned first-kind problem.
- Volterra convergence benefits from triangular structure and should not be copied blindly to Fredholm equations.
31. A dependable integral-equation workflow
Choose the function space and define the kernel operator. Prove boundedness and, when relevant, compactness. Decide whether the equation is first kind, second kind, Fredholm or Volterra. Check the parameter against the operator spectrum.
Use Neumann series when norm control is available, spectral expansion for self-adjoint compact structure, or finite-rank/Galerkin/Nyström approximation numerically. Always measure the residual and conditioning separately.
32. Independent practice: twenty questions
- Write a Fredholm second-kind equation as (I−λT)u=f.
- Distinguish first-kind and second-kind equations.
- Distinguish Fredholm and Volterra equations.
- Give a sup-norm bound for a continuous-kernel operator.
- What property do Hilbert–Schmidt operators have?
- When does the Neumann series invert I−λT?
- Define the second iterated kernel K₂.
- What is a resolvent kernel?
- Why does a rank-one kernel reduce to one scalar equation?
- How does a finite-rank kernel reduce to matrix algebra?
- State the Fredholm alternative in words.
- What role does the adjoint play in compatibility?
- What spectral advantage does a self-adjoint compact kernel have?
- What happens when 1−λμ_n=0?
- Why are first-kind compact equations often ill-posed?
- What does Tikhonov regularisation minimise?
- Why can Volterra Neumann series converge more strongly?
- How can a Green function produce an integral equation?
- What is the Nyström idea?
- Name two independent numerical checks.
33. Worked solutions and checks
1. Define Tu(x)=∫K(x,y)u(y)dy; then u−λTu=f.
2. First kind is Tu=f; second kind is u−λTu=f.
3. Fredholm uses a fixed integration range; Volterra usually integrates only up to x.
4. ||T||≤(b−a)sup|K| on C([a,b]) in the sup norm.
5. They are compact.
6. When |λ|||T||<1, or under another proven convergence criterion.
7. K₂(x,y)=∫K(x,z)K(z,y)dz.
8. A kernel summing the iterated-kernel series so that the inverse solution can be written as f plus one integral term.
9. The operator range is one-dimensional, so all unknown dependence collapses to a single coefficient.
10. Introduce finitely many integral coefficients; they satisfy a finite linear system.
11. Either the homogeneous equation is trivial and the inhomogeneous problem is uniquely solvable, or a nontrivial homogeneous kernel creates compatibility/nonuniqueness conditions.
12. Solvability requires f to be orthogonal to the relevant adjoint nullspace.
13. Real eigenvalues, orthogonal eigenspaces and an orthonormal spectral decomposition on the appropriate subspace.
14. The inverse coefficient blows up and the homogeneous equation has a nonzero solution.
15. Compact singular values/eigenvalues tend to zero, so inversion amplifies small data errors.
16. ||Tu−f||²+α||u||².
17. Iterated triangular integration produces factorial-type decay in operator iterates.
18. It represents the inverse differential operator and turns forcing into an integral against a kernel.
19. Replace the integral by quadrature at nodes, producing a linear algebra system.
20. Examples: residual norm, quadrature refinement, spectral-distance check, comparison with separable-kernel solution, regularisation sensitivity.
34. Where this guide hands off
R27.08 leaves linearity behind. Fixed-point theorems, monotone operators, variational principles and topological degree provide existence methods when superposition fails and the operator equation becomes nonlinear.
Sources and further study
For the operator-theoretic background, continue through MIT OpenCourseWare 18.102 Introduction to Functional Analysis. Standard integral-equation references develop Fredholm, Volterra and compact-kernel theory in greater depth.