Skip to content

Characteristic Numbers, Eigenfunctions and the Resolvent Kernel

CSIR NET integral equations: iterated kernels, Neumann series, resolvent kernel, Fredholm theorems, symmetric kernels, Hilbert-Schmidt and trace formulas.

  • 21 min read
  • Updated
On this page · 14 sections
  1. Prerequisites
  2. The Hook: A Geometric Series of Operators
  3. Characteristic Numbers and Eigenfunctions
  4. Iterated Kernels
  5. The Resolvent Kernel and the Neumann Series
  6. Fredholm’s Theory in One Table
  7. Symmetric Kernels
  8. Memory Tricks
  9. Don’t Confuse
  10. Common Mistakes
  11. Practice
  12. Quick Revision Table
  13. Key Takeaways
  14. What’s Next

Prerequisites

The Hook: A Geometric Series of Operators

Characteristic Numbers and Eigenfunctions

Basic facts, valid for any continuous (or square-integrable) kernel on a finite interval:

FactReason
\( \lambda = 0 \) is never a characteristic number\( \varphi = 0 \cdot \int K\varphi \) forces \( \varphi = 0 \)
Each characteristic number has finite multiplicityFredholm theory (separable case: \( n - \operatorname{rank}(I - \lambda A) \))
Characteristic numbers are isolated; finitely many in any bounded regionthey are the zeros of an entire function \( D(\lambda) \) (below)
Their number may be 0, finite or countably infinite\( \sin x\cos t \) on \( [0, \pi] \); \( xt \); \( \min(x, t) \)
\( K(x, t) \) and \( K(t, x) \) have the same characteristic numbers, with the same multiplicitiesFredholm’s second theorem
A Volterra kernel (continuous) has noneunique solvability for every \( \lambda \)

Iterated Kernels

For the Volterra case the limits \( t \le s \le x \) come from the triangular shape: \( K(x, s) = 0 \) unless \( s \le x \), and \( K_n(s, t) = 0 \) unless \( t \le s \).

KernelIterated kernel \( K_n \)
\( 1 \) on \( [0, 1] \) (Fredholm)\( 1 \) for all \( n \)
\( xt \) on \( [0, 1] \)\( \frac{xt}{3^{n-1}} \)
\( a(x)b(t) \), \( \alpha = \int ab \)\( \alpha^{n-1}a(x)b(t) \)
\( x + t \) on \( [0, 1] \)\( K_2 = xt + \frac{x + t}{2} + \frac{1}{3} \)
\( 1 \) (Volterra, \( \int_0^x \))\( \frac{(x - t)^{n-1}}{(n-1)!} \)
\( e^{x - t} \) (Volterra)\( e^{x - t}\frac{(x - t)^{n-1}}{(n-1)!} \)
\( k(x - t) \) (Volterra)the \( n \)-fold convolution of \( k \); Laplace transform \( \hat k(s)^n \)

The Resolvent Kernel and the Neumann Series

Successive approximation \( \varphi_0 = f \), \( \varphi_{n+1} = f + \lambda\int K\varphi_n \) gives

$$\varphi_n(x) = f(x) + \sum_{m=1}^{n} \lambda^m\int_a^b K_m(x, t)\,f(t)\,dt.$$

Some books write the resolvent with the factor \( \lambda \) included, \( R = \lambda\Gamma \); the solution is then \( f + \int Rf \). Check which convention the options use before comparing.

When does the Neumann series converge?

SettingSufficient conditionComment
Fredholm, \( \lvert K \rvert \le M \) on \( [a, b]^2 \)\( \lvert\lambda\rvert < \frac{1}{M(b - a)} \)from \( \lvert K_n \rvert \le M^n(b - a)^{n-1} \)
Fredholm, square-integrable\( \lvert\lambda\rvert < \frac{1}{B} \), \( B = \left(\int_a^b\int_a^b \lvert K \rvert^2\,dx\,dt\right)^{1/2} \)usually sharper
Exact radius (Fredholm)\( \lvert\lambda\rvert < \lvert\lambda_1\rvert \), the smallest characteristic number in modulus\( \Gamma \) has a pole there
Volterra, continuous kernelevery \( \lambda \)\( \lvert K_n \rvert \le M^n\frac{(x - t)^{n-1}}{(n-1)!} \)

The \( L^2 \) bound also gives \( \lvert\lambda_1\rvert \ge \frac{1}{B} \): no characteristic number lies inside the disc \( \lvert\lambda\rvert < 1/B \).

Resolvent of a separable kernel

From the method on the separable kernels page, \( c = (I - \lambda A)^{-1}F \), which gives

$$\Gamma(x, t; \lambda) = \sum_{i, j = 1}^{n} a_i(x)\left[(I - \lambda A)^{-1}\right]_{ij} b_j(t).$$

The entries of \( (I - \lambda A)^{-1} \) are cofactors divided by \( \det(I - \lambda A) \). So \( \Gamma \) is a ratio of polynomials in \( \lambda \), and its poles are the characteristic numbers.

Fredholm’s Theory in One Table

For a general continuous kernel, Fredholm built two entire functions of \( \lambda \): the Fredholm determinant \( D(\lambda) \) and the first minor \( D(x, t; \lambda) \), with

$$\Gamma(x, t; \lambda) = \frac{D(x, t; \lambda)}{D(\lambda)}, \qquad D(\lambda) = 1 - \lambda\int_a^b K(x, x)\,dx + \cdots$$

For a separable kernel, \( D(\lambda) = \det(I - \lambda A) \) (its \( \lambda \)-coefficient is \( -\operatorname{tr}A = -\int K(x, x)\,dx \)). The consequences:

Fredholm theorem (numbering varies between texts)Content
Resolvent is meromorphic\( \Gamma \) is analytic in \( \lambda \) except for poles at the zeros of \( D \)
Characteristic numbers = zeros of \( D(\lambda) \)at most countably many, no finite limit point
First theorem\( D(\lambda) \ne 0 \): unique solution for every \( f \), given by the resolvent
Second theoremat a zero of \( D \): the homogeneous equation and its transpose have the same finite number of independent solutions
Third theoremat a zero of \( D \): solvable iff \( f \perp \) every solution of the transposed homogeneous equation

The second and third theorems together are the Fredholm alternative.

Symmetric Kernels

Let \( K \) be real, continuous (or square-integrable), not identically zero, with \( K(x, t) = K(t, x) \). The theory is the infinite-dimensional version of the spectral theorem for real symmetric matrices.

Proof of orthogonality (one line). If \( \varphi_m = \lambda_m\mathcal{K}\varphi_m \) and \( \varphi_n = \lambda_n\mathcal{K}\varphi_n \), symmetry gives \( \langle\mathcal{K}\varphi_m, \varphi_n\rangle = \langle\varphi_m, \mathcal{K}\varphi_n\rangle \), that is \( \frac{1}{\lambda_m}\langle\varphi_m, \varphi_n\rangle = \frac{1}{\lambda_n}\langle\varphi_m, \varphi_n\rangle \). If \( \lambda_m \ne \lambda_n \), the inner product is 0. Taking \( \varphi_n = \overline{\varphi_m} \) the same way proves reality.

Solving with the eigenfunctions

If \( \lambda \) is not a characteristic number, the solution of \( \varphi = f + \lambda\int K\varphi \) is

$$\varphi(x) = f(x) + \lambda\sum_n \frac{f_n}{\lambda_n - \lambda}\,\varphi_n(x), \qquad f_n = \int_a^b f(t)\,\varphi_n(t)\,dt,$$

and correspondingly \( \Gamma(x, t; \lambda) = \sum_n \frac{\varphi_n(x)\varphi_n(t)}{\lambda_n - \lambda} \) (in the mean-square sense). At \( \lambda = \lambda_k \), the term \( \frac{f_k}{\lambda_k - \lambda} \) blows up unless \( f_k = 0 \): the Fredholm alternative again.

Trace formulas

For a symmetric kernel, by property 5:

$$\boxed{\sum_n \frac{1}{\lambda_n^2} = \iint K(x, t)^2\,dx\,dt = \int_a^b K_2(x, x)\,dx.}$$

If in addition \( K \) is continuous and positive (all \( \lambda_n > 0 \)), Mercer’s theorem makes the expansion in property 6 converge uniformly, and then

$$\boxed{\sum_n \frac{1}{\lambda_n} = \int_a^b K(x, x)\,dx.}$$
Kernel on \( [0, 1] \)\( \lambda_n \)\( \int K(x, x)\,dx \)\( \sum \frac{1}{\lambda_n} \)\( \iint K^2 \)\( \sum \frac{1}{\lambda_n^2} \)
\( \min(x, t) - xt \)\( n^2\pi^2 \)\( \frac{1}{6} \)\( \frac{1}{\pi^2}\cdot\frac{\pi^2}{6} = \frac{1}{6} \)\( \frac{1}{90} \)\( \frac{1}{\pi^4}\cdot\frac{\pi^4}{90} = \frac{1}{90} \)
\( \min(x, t) \)\( \frac{(2n-1)^2\pi^2}{4} \)\( \frac{1}{2} \)\( \frac{4}{\pi^2}\cdot\frac{\pi^2}{8} = \frac{1}{2} \)\( \frac{1}{6} \)\( \frac{16}{\pi^4}\cdot\frac{\pi^4}{96} = \frac{1}{6} \)
\( x + t \) (not positive)\( -6 \pm 4\sqrt3 \)\( 1 \)\( 1 \) (finite sum, exact)\( \frac{7}{6} \)\( \frac{7}{6} \)

For a separable kernel the identity \( \sum \frac{1}{\lambda_n} = \operatorname{tr}A = \int K(x, x)\,dx \) holds with no positivity assumption, provided each characteristic number is counted as often as it is a root of \( \det(I - \lambda A) \): the trace of \( A \) is the sum of its eigenvalues, and the zero eigenvalues contribute nothing. That is why the \( x + t \) row works even though the kernel is not positive. These identities are how Part C turns integral equations into series sums: the first row proves \( \sum\frac{1}{n^2} = \frac{\pi^2}{6} \) and \( \sum\frac{1}{n^4} = \frac{\pi^4}{90} \).

graph TD
    A["Kernel K on [a, b]"] --> B{"Volterra (variable upper limit)?"}
    B -->|yes| C["No characteristic numbers; resolvent entire in lambda; Neumann series converges for all lambda"]
    B -->|no| D{"Separable?"}
    D -->|yes| E["det(I - lambda A) = 0: at most n characteristic numbers; Gamma rational in lambda"]
    D -->|no| F["Fredholm theory: characteristic numbers = zeros of D(lambda); possibly infinitely many"]
    E --> G{"Symmetric?"}
    F --> G
    G -->|yes| H["At least one; all real; orthogonal eigenfunctions; sum of 1/lambda_n^2 = double integral of K^2"]
    G -->|no| I["Maybe none, maybe complex; use transposed kernel in the alternative"]

Memory Tricks

Don’t Confuse

ThisvsThat
characteristic number \( \lambda \)operator eigenvalue \( \mu = \frac{1}{\lambda} \) (largest \( \lvert\mu\rvert \) = smallest \( \lvert\lambda\rvert \))
resolvent \( \Gamma \) with \( \varphi = f + \lambda\int\Gamma f \)\( R = \lambda\Gamma \) with \( \varphi = f + \int R f \)
Neumann series diverges at \( \lambda \)equation unsolvable at \( \lambda \) (only if \( \lambda \) is a characteristic number)
\( \sum \frac{1}{\lambda_n^2} = \iint K^2 \) (any symmetric \( L^2 \) kernel)\( \sum \frac{1}{\lambda_n} = \int K(x, x) \) (needs positivity and continuity, or a finite-rank kernel)
non-zero symmetric kernel: at least one characteristic numbernon-symmetric: possibly none (\( \sin x\cos t \) on \( [0, \pi] \))

Common Mistakes

Practice

All questions below are original practice questions written for this page. They are not from any CSIR NET paper. In Level 3, as in Part C of Mathematical Sciences, one or more options may be correct, and each option is justified in the answer.

Level 1: Foundation

  1. Find the iterated kernel \( K_2 \) of \( K = 1 \) on \( [0, 1] \) (Fredholm).
  2. How many characteristic numbers does \( \varphi(x) = \lambda\int_0^x \sin(x - t)\,\varphi(t)\,dt \) have?
  3. Write the resolvent of \( K = xt \) on \( [0, 1] \).
Answers
  1. \( K_2 = \int_0^1 1 \cdot 1\,ds = 1 \). All iterated kernels equal 1, \( \Gamma = \frac{1}{1 - \lambda} \), characteristic number 1.
  2. None. It is a Volterra equation with a continuous kernel.
  3. \( \Gamma = \frac{3xt}{3 - \lambda} \), \( \lambda \ne 3 \).

Level 2: Part B Style

Practice Q1. The resolvent kernel of \( \varphi(x) = f(x) + \lambda\int_0^x \varphi(t)\,dt \) is:

(a) \( e^{\lambda(x - t)} \) (b) \( e^{x - t} \) (c) \( \frac{1}{1 - \lambda} \) (d) \( \frac{1}{1 - \lambda(x - t)} \)

Answer: (a)

Volterra kernel 1: \( K_m = \frac{(x - t)^{m-1}}{(m-1)!} \), so \( \Gamma = \sum_{m \ge 1}\frac{\lambda^{m-1}(x - t)^{m-1}}{(m-1)!} = e^{\lambda(x - t)} \). (c) is the Fredholm resolvent of the same kernel on \( [0, 1] \).

Practice Q2. The Neumann series for \( \varphi(x) = f(x) + \lambda\int_0^1 e^{x + t}\varphi(t)\,dt \) converges (for every continuous \( f \)) exactly when:

(a) \( \lvert\lambda\rvert < 1 \) (b) \( \lvert\lambda\rvert < \frac{1}{e^2} \) (c) \( \lvert\lambda\rvert < \frac{2}{e^2 - 1} \) (d) for all \( \lambda \)

Answer: (c)

Rank 1 with \( \alpha = \int_0^1 e^{2t}\,dt = \frac{e^2 - 1}{2} \). \( K_m = \alpha^{m-1}e^{x + t} \), so the series \( \sum(\lambda\alpha)^{m-1} \) converges iff \( \lvert\lambda\alpha\rvert < 1 \). That is \( \lvert\lambda\rvert < \frac{2}{e^2 - 1} \approx 0.313 \), the characteristic number. (\( B = \left(\iint e^{2x + 2t}\right)^{1/2} = \alpha \) gives the same bound: for a rank-1 symmetric kernel the \( L^2 \) test is exact.)

Practice Q3. The smallest positive characteristic number of \( K(x, t) = \min(x, t) - xt \) on \( [0, 1] \) is:

(a) \( \pi \) (b) \( \pi^2 \) (c) \( \frac{\pi^2}{4} \) (d) \( 6 \)

Answer: (b)

It is the Green’s function of \( -y'' \) with \( y(0) = y(1) = 0 \), so \( \lambda_n = n^2\pi^2 \), and the smallest is \( \pi^2 \approx 9.87 \). (c) belongs to \( \min(x, t) \). (d) is \( \frac{1}{\int K(x,x)dx} \), a lower bound for \( \lambda_1 \) (from \( \frac{1}{\lambda_1} < \sum\frac{1}{\lambda_n} = \frac{1}{6} \)), not its value.

Practice Q4. If \( \varphi(x) = e^x + \int_0^x e^{x - t}\varphi(t)\,dt \), then \( \varphi(1) \) equals:

(a) \( e \) (b) \( 2e \) (c) \( e^2 \) (d) \( e^2 - 1 \)

Answer: (c)

With \( \lambda = 1 \), the worked example gives \( \varphi = e^{(1 + \lambda)x} = e^{2x} \), so \( \varphi(1) = e^2 \approx 7.389 \). Check by differentiating: \( \varphi' = e^x + \varphi(x) + \int_0^x e^{x - t}\varphi(t)\,dt \), and \( e^x + \int_0^x e^{x - t}\varphi \) is \( \varphi \) itself, so \( \varphi' = 2\varphi \) with \( \varphi(0) = 1 \).

Level 3: Part C Style

Practice Q5. Let \( K(x, t) = \min(x, t) \) on \( [0, 1] \), and let \( \lambda_1 < \lambda_2 < \cdots \) be its characteristic numbers. Which statements are true?

(a) \( \lambda_n = \frac{(2n - 1)^2\pi^2}{4} \). (b) The eigenfunctions are \( \sin\frac{(2n - 1)\pi x}{2} \). (c) \( \sum_n \frac{1}{\lambda_n} = \frac{1}{2} \). (d) \( \sum_n \frac{1}{\lambda_n^2} = \frac{1}{3} \).

Answer: (a), (b) and (c)
  • (a), (b) True. \( \varphi = \lambda\int_0^1 \min(x, t)\varphi \) is equivalent to \( \varphi'' + \lambda\varphi = 0 \), \( \varphi(0) = 0 \), \( \varphi'(1) = 0 \) (differentiate twice; the conditions come from \( x = 0 \) and from \( \varphi'(x) = \lambda\int_x^1 \varphi \)). So \( \varphi = \sin\sqrt\lambda x \) with \( \cos\sqrt\lambda = 0 \).
  • (c) True. The kernel is continuous, symmetric and positive, so Mercer applies: \( \sum\frac{1}{\lambda_n} = \int_0^1 \min(x, x)\,dx = \int_0^1 x\,dx = \frac{1}{2} \). Directly: \( \frac{4}{\pi^2}\sum\frac{1}{(2n-1)^2} = \frac{4}{\pi^2}\cdot\frac{\pi^2}{8} \).
  • (d) False. \( \sum\frac{1}{\lambda_n^2} = \iint\min(x, t)^2 = 2\int_0^1\int_0^x t^2\,dt\,dx = \frac{1}{6} \).

Practice Q6. Let \( K \) be a real, continuous, symmetric kernel on \( [0, 1]^2 \), not identically zero, with \( \int_0^1\int_0^1 K(x, t)^2\,dx\,dt = \frac{1}{4} \). Which statements are necessarily true?

(a) \( K \) has at least one characteristic number. (b) Every characteristic number \( \lambda \) satisfies \( \lvert\lambda\rvert \ge 2 \). (c) The Neumann series for \( \varphi = f + \lambda\int_0^1 K\varphi \) converges for \( \lvert\lambda\rvert < 2 \). (d) \( K \) has infinitely many characteristic numbers.

Answer: (a), (b) and (c)
  • (a) True. A non-zero symmetric kernel always has one.
  • (b) True. \( B = \frac{1}{2} \), and \( \frac{1}{\lambda_1^2} \le \sum\frac{1}{\lambda_n^2} = B^2 \) gives \( \lvert\lambda_1\rvert \ge 2 \).
  • (c) True. The \( L^2 \) criterion \( \lvert\lambda\rvert < \frac{1}{B} = 2 \).
  • (d) False. \( K = \frac{3}{2}xt \) satisfies the hypotheses (\( \iint\frac{9}{4}x^2t^2 = \frac{9}{4}\cdot\frac{1}{9} = \frac{1}{4} \)) and has exactly one characteristic number, \( \lambda = 2 \). Note that (b) is sharp for this kernel.

Practice Q7. Consider \( \varphi(x) = x + \lambda\int_0^1 (x + t)\,\varphi(t)\,dt \). Which statements are true?

(a) The Neumann series converges for every \( \lambda \) with \( \lvert\lambda\rvert < 4\sqrt3 - 6 \). (b) The \( L^2 \) criterion alone guarantees convergence of the Neumann series for \( \lvert\lambda\rvert < \sqrt{6/7} \). (c) For \( \lambda = 1 \) the equation has no solution. (d) For \( \lambda = 1 \) the Neumann series converges.

Answer: (a) and (b)
  • (a) True. \( \Gamma \) is rational in \( \lambda \) with poles at \( -6 \pm 4\sqrt3 \); the nearer one has modulus \( 4\sqrt3 - 6 \approx 0.928 \), which is the radius of convergence of the power series in \( \lambda \).
  • (b) True. \( \iint(x + t)^2 = \frac{7}{6} \), so \( \frac{1}{B} = \sqrt{6/7} \approx 0.926 \), just inside the true radius.
  • (c) False. \( 1 \) is not a characteristic number, so there is a unique solution: \( \varphi = -6x - 4 \) (worked example above).
  • (d) False. \( 1 > 0.928 \): outside the radius of convergence. Having a solution and having a convergent Neumann series are different things.

Quick Revision Table

IdeaFact
Characteristic number\( \varphi = \lambda\int K\varphi \) has \( \varphi \ne 0 \); never \( \lambda = 0 \)
Iterated kernel\( K_{n+1} = \int K(x, s)K_n(s, t)\,ds \) (Volterra: \( \int_t^x \))
Resolvent\( \Gamma = \sum\lambda^{m-1}K_m \); \( \varphi = f + \lambda\int\Gamma f \)
Resolvent equation\( \Gamma = K + \lambda\int K(x, s)\Gamma(s, t)\,ds \)
Rank 1\( \Gamma = \frac{a(x)b(t)}{1 - \lambda\int ab} \)
Separable\( \Gamma = \sum a_i[(I - \lambda A)^{-1}]_{ij}b_j \)
Volterra \( K = 1 \), \( e^{x - t} \)\( \Gamma = e^{\lambda(x - t)} \), \( e^{(1 + \lambda)(x - t)} \)
Neumann convergence\( \lvert\lambda\rvert < \frac{1}{M(b - a)} \) or \( < \frac{1}{B} \); exact: \( < \lvert\lambda_1\rvert \); Volterra: all \( \lambda \)
Fredholm\( \Gamma = D(x, t; \lambda)/D(\lambda) \); characteristic numbers = zeros of \( D \)
Symmetricat least one, real, orthogonal eigenfunctions, finite iff separable
Traces\( \sum\lambda_n^{-2} = \iint K^2 \); positive (Mercer): \( \sum\lambda_n^{-1} = \int K(x, x) \)
Hilbert–Schmidt solution\( \varphi = f + \lambda\sum\frac{f_n}{\lambda_n - \lambda}\varphi_n \)

Key Takeaways

What’s Next

Advertisement
Advertisement