7.2 Eigenvalues, Eigenfunctions and the Resolvent Kernel
Characteristic numbers and eigenfunctions of a kernel, bounds on them, iterated kernels, the Neumann series and the resolvent kernel $R(x,t;\lambda)$, entire resolvents of Volterra kernels, and the spectral theory of symmetric kernels: real characteristic numbers, orthogonal eigenfunctions, the Hilbert–Schmidt and Mercer theorems.
In §7.1 we solved Fredholm equations with separable kernels by linear algebra and found that everything hinges on the numbers $\lambda$ at which the homogeneous equation $y=\lambda\int_a^bK(x,t)y(t)\,dt$ has nonzero solutions. For a general continuous kernel there is no finite matrix to diagonalise, but the same objects survive: characteristic numbers and eigenfunctions, the integral-equation analogues of the eigenvalues and eigenvectors of a matrix. Away from them we would like a formula for the solution valid for every right-hand side at once, $y=f+\lambda\int_a^bR(x,t;\lambda)f(t)\,dt$, just as $(I-\lambda A)^{-1}$ solves every linear system. The kernel $R$ is the resolvent kernel, and the most direct way to build it is to iterate: $y=f+\lambda\mathcal{K}f+\lambda^2\mathcal{K}^2f+\cdots$, the Neumann series, which is the geometric series $(I-\lambda\mathcal{K})^{-1}$ $=\sum\lambda^n\mathcal{K}^n$ for operators.
The section answers the natural questions. How small must $|\lambda|$ be for the Neumann series to converge, and what is its true radius of convergence? Why does the resolvent of a Volterra kernel exist for every $\lambda$? And what extra structure appears when the kernel is symmetric, $K(x,t)=K(t,x)$ — the case of every Green's function of a self-adjoint boundary value problem? The answer to the last question, due to Hilbert and Schmidt, is a complete analogue of the spectral theorem for symmetric matrices: real characteristic numbers, orthogonal eigenfunctions, and expansions of the kernel and of the solution in eigenfunctions. CSIR NET asks for resolvents of simple kernels, for the range of $\lambda$ in which the Neumann series converges, for characteristic numbers and eigenfunctions, and for true/false statements about symmetric kernels.
Characteristic Numbers and Eigenfunctions
Throughout, $K$ is continuous on $Q=[a,b]^2$ (possibly complex-valued), and we write $$(\mathcal{K}y)(x)=\int_a^bK(x,t)\,y(t)\,dt,\qquad (f,g)=\int_a^bf(x)\,\overline{g(x)}\,dx,\qquad\|f\|_2=(f,f)^{1/2}.$$ Two sizes of a kernel will appear constantly: the maximum $M=\max_Q|K|$ and the Hilbert–Schmidt norm $$B=\|K\|_2=\left(\int_a^b\!\!\int_a^b|K(x,t)|^2\,dx\,dt\right)^{1/2}\le M(b-a).$$
- Every characteristic number satisfies $$|\lambda|\ge\frac{1}{B}\ge\frac{1}{M(b-a)}.$$ In particular, for $|\lambda|B\lt1$ the homogeneous equation has only the trivial solution.
- The multiplicity $m$ of a characteristic number $\lambda$ is finite, and $m\le|\lambda|^2B^2$.
Proof. (1) Let $y=\lambda\mathcal{K}y$, $y\not\equiv0$. By the Cauchy–Schwarz inequality, for each $x$ $$|y(x)|^2=|\lambda|^2\left|\int_a^bK(x,t)y(t)\,dt\right|^2\le|\lambda|^2\int_a^b|K(x,t)|^2dt\;\|y\|_2^2.$$ Integrating over $x$ gives $\|y\|_2^2$ $\le|\lambda|^2B^2\|y\|_2^2$. Since $y$ is continuous and not identically zero, $\|y\|_2\gt0$, so $|\lambda|B\ge1$. Finally $B^2\le M^2(b-a)^2$.
(2) Let $\varphi_1,\dots,\varphi_m$ be linearly independent eigenfunctions for $\lambda$; by Gram–Schmidt we may take them orthonormal. Fix $x$ and apply Bessel's inequality to the function $g_x(t)=\overline{K(x,t)}$ and the orthonormal set $\varphi_1,\dots,\varphi_m$: since $(\varphi_i,g_x)$ $=\int K(x,t)\varphi_i(t)\,dt$ $=\varphi_i(x)/\lambda$, $$\sum_{i=1}^m\frac{|\varphi_i(x)|^2}{|\lambda|^2}\le\int_a^b|K(x,t)|^2\,dt.$$ Integrating over $x$ and using $\|\varphi_i\|_2=1$ gives $m/|\lambda|^2\le B^2$. ∎
Both bounds can be sharp. For $K=xt$ on $[0,1]$, $B=\tfrac13$ and the only characteristic number is $3=1/B$. For $\cos(x-t)$ on $[0,2\pi]$, $B^2=2\pi^2$ and the characteristic number $\frac1\pi$ has multiplicity $2=\frac{1}{\pi^2}\cdot2\pi^2$.
For a separable kernel the characteristic numbers are the zeros of the polynomial $D(\lambda)=\det(I-\lambda A)$ (the Fredholm alternative for separable kernels in §7.1). Fredholm showed that the whole theory survives for every continuous kernel, with $D$ replaced by an entire function.
- the characteristic numbers of $K$ are exactly the zeros of $D$; they form a finite or countably infinite set with no finite accumulation point, and each has finite multiplicity;
- $\lambda$ is a characteristic number of $K$ if and only if it is one of $K^{T}$, with the same multiplicity;
- if $D(\lambda)\ne0$, then $y=f+\lambda\mathcal{K}y$ has exactly one solution for every $f\in C[a,b]$, namely $y=f+\lambda\int_a^bR(x,t;\lambda)f(t)\,dt$ with $R=D(x,t;\lambda)/D(\lambda)$ for a kernel $D(x,t;\lambda)$ entire in $\lambda$;
- if $D(\lambda)=0$, the equation is solvable if and only if $(f,\bar z)=\int_a^bf\,z\,dx=0$ for every solution $z$ of the transposed homogeneous equation $z=\lambda\int_a^bK(t,x)z(t)\,dt$.
Proof (sketch). Fix $r\gt0$; we prove (1)–(4) for $|\lambda|\lt r$. By the Weierstrass approximation theorem in two variables there is a polynomial $P(x,t)$ with $\max|K-P|\lt\frac{1}{r(b-a)}$. Write $K=P+H$. For $|\lambda|\lt r$, $|\lambda|\max|H|(b-a)\lt1$, so by Theorem 7.12 below the operator $I-\lambda\mathcal{H}$ is invertible, with inverse $I+\lambda\mathcal{R}_H$ given by a Neumann series. The equation $y-\lambda\mathcal{P}y-\lambda\mathcal{H}y=f$ is then equivalent to $$y-\lambda(I+\lambda\mathcal{R}_H)\mathcal{P}y=(I+\lambda\mathcal{R}_H)f,$$ and $(I+\lambda\mathcal{R}_H)\mathcal{P}$ is an integral operator with a separable kernel (because $P$ is). The Fredholm alternative for separable kernels (§7.1) now gives the alternative for $|\lambda|\lt r$, with a determinant that is analytic in $\lambda$; the transposed statement follows from the same reduction applied to $K^T$, and the determinants can be matched to Fredholm's series. Since $r$ was arbitrary, the zeros of an analytic function that is not identically zero are isolated, and finite-dimensionality is Theorem 7.10(2). ∎
Iterated Kernels, the Neumann Series and the Resolvent Kernel
Successive approximations $y_0=f$, $y_{n+1}=f+\lambda\mathcal{K}y_n$ give $y_n=f+\lambda\mathcal{K}f+\dots+\lambda^n\mathcal{K}^nf$. The powers $\mathcal{K}^n$ are again integral operators, whose kernels are obtained by "matrix multiplication" of kernels.
The identity $K_{m+n}=\int K_mK_n$ is associativity of the operator product $\mathcal{K}^{m+n}$ $=\mathcal{K}^m\mathcal{K}^n$ (Fubini's theorem). For a Volterra kernel ($K=0$ for $t\gt x$) all iterated kernels are again Volterra kernels: $K_{n+1}(x,t)$ $=\int_t^xK(x,s)K_n(s,t)\,ds$.
- $|K_n(x,t)|\le M^n(b-a)^{n-1}$, $\ \|K_n\|_2\le B^n$, and $|K_{n+2}(x,t)|$ $\le k(x)\,B^n\,\tilde k(t)$.
- If $|\lambda|B\lt1$ (in particular if $|\lambda|M(b-a)\lt1$), the series $R(x,t;\lambda)$ $=\sum\lambda^{n-1}K_n$ converges absolutely and uniformly on $[a,b]^2$ to a continuous kernel satisfying the resolvent equations $$R(x,t;\lambda)=K(x,t)+\lambda\int_a^bK(x,s)R(s,t;\lambda)\,ds,$$ $$R(x,t;\lambda)=K(x,t)+\lambda\int_a^bR(x,s;\lambda)K(s,t)\,ds.$$
- For such $\lambda$ and every $f\in C[a,b]$, the equation $y=f+\lambda\mathcal{K}y$ has exactly one continuous solution, $$y(x)=f(x)+\lambda\int_a^bR(x,t;\lambda)\,f(t)\,dt,$$ and the successive approximations converge to it uniformly.
Proof. (1) By induction, $|K_{n+1}(x,t)|$ $\le\int_a^bM\cdot M^n(b-a)^{n-1}ds$ $=M^{n+1}(b-a)^n$. Next, Cauchy–Schwarz in $s$ gives $|K_{n+1}(x,t)|^2$ $\le\int|K(x,s)|^2ds\int|K_n(s,t)|^2ds$; integrating over $x$ and $t$, $\|K_{n+1}\|_2^2\le B^2\|K_n\|_2^2$. Finally $K_{n+2}(x,t)$ $=\iint K(x,s)K_n(s,r)K(r,t)\,ds\,dr$, and Cauchy–Schwarz in $(s,r)$ gives $|K_{n+2}(x,t)|$ $\le k(x)\tilde k(t)\|K_n\|_2$ $\le k(x)\tilde k(t)B^n$.
(2) $k$ and $\tilde k$ are bounded by $M\sqrt{b-a}$, so the terms of $\sum_{n\ge3}\lambda^{n-1}K_n$ are bounded by $M^2(b-a)\,|\lambda|\,(|\lambda|B)^{n-2}$: a convergent geometric series when $|\lambda|B\lt1$. The Weierstrass M-test gives absolute and uniform convergence, so $R$ is continuous. Multiplying the series by $K(x,s)$ and integrating term by term, $$\lambda\int_a^bK(x,s)R(s,t;\lambda)\,ds=\sum_{n\ge1}\lambda^nK_{n+1}(x,t)=R-K;$$ the second resolvent equation is proved the same way using $K_{n+1}=\int K_nK$.
(3) Let $y=f+\lambda\int Rf$. Then, by Fubini and the first resolvent equation, $$\lambda\mathcal{K}y=\lambda\int_a^bK f\,dt+\lambda^2\!\int_a^b\!\left[\int_a^bK(x,s)R(s,t)\,ds\right]f(t)\,dt=\lambda\int_a^bR(x,t)f(t)\,dt=y-f.$$ So $y$ is a solution. It is the only one: the difference of two solutions solves the homogeneous equation, which has only the trivial solution because $|\lambda|B\lt1$ (Theorem 7.10(1)). The successive approximations are the partial sums $f+\sum_{n\le N}\lambda^n\mathcal{K}^nf$, which converge uniformly by (2). ∎
The condition $|\lambda|B\lt1$ is sufficient, not necessary. The truth is described by Fredholm's theory.
For a separable kernel all of this is explicit: by the Fredholm alternative for separable kernels (§7.1), $R=\sum_{i,j}a_i(x)[(I-\lambda A)^{-1}]_{ij}b_j(t)$, a rational function of $\lambda$ whose denominator is $D(\lambda)=\det(I-\lambda A)$. When the Neumann series diverges ($|\lambda|\gt|\lambda_1|$) the resolvent still exists at every non-characteristic $\lambda$ — the series is just the wrong way to compute it.
For Volterra kernels the factorials of §7.1 make the resolvent exist for every $\lambda$.
Proof. If $K_n(s,t)=0$ for $s\lt t$, then $K_{n+1}(x,t)$ $=\int_a^bK(x,s)K_n(s,t)\,ds$ has a nonzero integrand only when $t\le s\le x$; so $K_{n+1}(x,t)=0$ for $t\gt x$ and $K_{n+1}(x,t)$ $=\int_t^xK(x,s)K_n(s,t)\,ds$ otherwise. The estimate follows by induction: $|K_{n+1}(x,t)|$ $\le\int_t^xM\frac{M^n(s-t)^{n-1}}{(n-1)!}\,ds$ $=\frac{M^{n+1}(x-t)^n}{n!}$. Thus $\sum|\lambda|^{n-1}|K_n|$ $\le M\sum\frac{(|\lambda|M(x-t))^{n-1}}{(n-1)!}$ $=Me^{|\lambda|M(x-t)}$, and the M-test gives uniform convergence for $|\lambda|\le r$, any $r$; a uniformly convergent series of polynomials in $\lambda$ is entire. The resolvent equations and the solution formula follow exactly as in Theorem 7.12, and uniqueness is the existence theorem for Volterra equations (§7.1). ∎
- Separable kernel: compute $A$, invert $I-\lambda A$, and form $R=\sum a_i[(I-\lambda A)^{-1}]_{ij}b_j$. For rank one, $K=a(x)b(t)$, $K_n=\alpha^{n-1}K$ with $\alpha=\int ab$, and $R=\frac{K}{1-\lambda\alpha}$.
- Look for a pattern in $K_2$, $K_3$: often $K_n=c^{n-1}K$ (e.g. $e^{x-t}$ on $[0,1]$, Fredholm) or $K_n=\frac{(x-t)^{n-1}}{(n-1)!}\times(\dots)$ (Volterra); then sum the series.
- Convolution Volterra kernel: $\hat R=\hat k/(1-\lambda\hat k)$, then invert the Laplace transform.
- Symmetric kernel with known eigenfunctions: $\int R f$ $=\sum\frac{(f,\varphi_n)}{\lambda_n-\lambda}\varphi_n$ (Theorem 7.17).
- Check: $R(x,t;0)=K(x,t)$, the poles of $R$ are the characteristic numbers, and $R$ satisfies $R=K+\lambda\int KR$.
Symmetric Kernels
Green's functions of self-adjoint boundary value problems, covariance functions in probability, and kernels like $\min(x,t)$, $\cos(x-t)$, $e^{-|x-t|}$ and $x+t$ are symmetric. For them the theory becomes as clean as that of real symmetric matrices.
- Every characteristic number of $K$ is real, and the eigenfunctions can be chosen real.
- Eigenfunctions belonging to different characteristic numbers are orthogonal: $(\varphi,\psi)=0$.
- $K$ has at least one characteristic number. The one of smallest modulus satisfies $$\frac{1}{|\lambda_1|}=\max_{\|\varphi\|_2=1}\big|(\mathcal{K}\varphi,\varphi)\big|,$$ and $\|\mathcal{K}\varphi\|_2$ $\le\|\varphi\|_2/|\lambda_1|$ for every $\varphi$.
- The characteristic numbers, repeated according to multiplicity, can be listed as $|\lambda_1|$ $\le|\lambda_2|$ $\le\cdots$ with real orthonormal eigenfunctions $\varphi_1,\varphi_2,\dots$; the list is finite or $|\lambda_n|\to\infty$, and $\displaystyle\sum_n\frac{1}{\lambda_n^2}$ $\displaystyle \le\iint K^2$.
- If $K$ is positive, all its characteristic numbers are positive.
Proof. (1) Let $\varphi=\lambda\mathcal{K}\varphi$, $\varphi\not\equiv0$. Then $(\mathcal{K}\varphi,\varphi)$ $=\frac1\lambda\|\varphi\|_2^2$. By symmetry $(\mathcal{K}\varphi,\varphi)$ $=(\varphi,\mathcal{K}\varphi)$ $=\overline{(\mathcal{K}\varphi,\varphi)}$, so $(\mathcal{K}\varphi,\varphi)$ is real, and so is $\lambda$ (as $\|\varphi\|_2\gt0$; it is nonzero because $\lambda\ne0$). Since $K$ and $\lambda$ are real, the real and imaginary parts of $\varphi$ are again solutions, and at least one is nonzero.
(2) If $\varphi=\lambda\mathcal{K}\varphi$ and $\psi=\mu\mathcal{K}\psi$ with real $\lambda\ne\mu$, then $$\frac1\lambda(\varphi,\psi)=(\mathcal{K}\varphi,\psi)=(\varphi,\mathcal{K}\psi)=\frac1\mu(\varphi,\psi),$$ so $(\varphi,\psi)=0$.
(3) (Sketch.) Let $\mu_1=\sup_{\|\varphi\|=1}|(\mathcal{K}\varphi,\varphi)|$; one shows $\mu_1\gt0$ (otherwise polarisation gives $(\mathcal{K}f,g)=0$ for all $f,g$, hence $K\equiv0$). Take $\varphi_k$ with $\|\varphi_k\|=1$ and $(\mathcal{K}\varphi_k,\varphi_k)\to\pm\mu_1$. The functions $\mathcal{K}\varphi_k$ are uniformly bounded and equicontinuous (by uniform continuity of $K$ and Cauchy–Schwarz), so by the Arzelà–Ascoli theorem a subsequence converges uniformly; the identity $\|\mathcal{K}\varphi_k\mp\mu_1\varphi_k\|^2$ $=\|\mathcal{K}\varphi_k\|^2\mp2\mu_1(\mathcal{K}\varphi_k,\varphi_k)+\mu_1^2$ $\le2\mu_1^2\mp2\mu_1(\mathcal{K}\varphi_k,\varphi_k)\to0$ then shows that the limit $\psi$ satisfies $\mathcal{K}\psi=\pm\mu_1\psi$, $\psi\ne0$. So $\pm1/\mu_1$ is a characteristic number; and no characteristic number has smaller modulus, because $\varphi=\lambda\mathcal{K}\varphi$ with $\|\varphi\|=1$ gives $\frac1{|\lambda|}$ $=|(\mathcal{K}\varphi,\varphi)|$ $\le\mu_1$. (Here we used the standard fact that for symmetric kernels $\|\mathcal{K}\varphi\|$ $\le\mu_1\|\varphi\|$, which follows from the polarisation identity.)
(4) Choose an orthonormal basis of each (finite-dimensional, Theorem 7.10) eigenspace; by (2) the union is orthonormal. For any finitely many of them the Bessel argument of Theorem 7.10(2) gives $\sum_n\frac{\varphi_n(x)^2}{\lambda_n^2}$ $\le\int_a^bK(x,t)^2\,dt$, and integrating, $\sum_n\frac1{\lambda_n^2}$ $\le\iint K^2$. Hence only finitely many $\lambda_n$ lie in any disc, so they can be ordered by modulus and $|\lambda_n|\to\infty$ if there are infinitely many. (5) If $\varphi=\lambda\mathcal{K}\varphi$ is real and nonzero, $\frac1\lambda\|\varphi\|^2$ $=(\mathcal{K}\varphi,\varphi)\ge0$, so $\lambda\gt0$. ∎
- (Hilbert–Schmidt.) Every function of the form $g=\mathcal{K}h$, $h\in C[a,b]$ ("sourcewise representable"), has the expansion $$g(x)=\sum_n(g,\varphi_n)\,\varphi_n(x)=\sum_n\frac{(h,\varphi_n)}{\lambda_n}\,\varphi_n(x),$$ converging absolutely and uniformly on $[a,b]$.
- (Iterated kernels and traces.) For $m\ge2$, $\displaystyle K_m(x,t)$ $\displaystyle =\sum_n\frac{\varphi_n(x)\varphi_n(t)}{\lambda_n^m}$ uniformly, and $\displaystyle\int_a^bK_m(x,x)\,dx$ $\displaystyle =\sum_n\frac{1}{\lambda_n^m}$. In particular $$\sum_n\frac1{\lambda_n^2}=\int_a^b\!\!\int_a^bK(x,t)^2\,dx\,dt,$$ and the bilinear series $\sum_n\frac{\varphi_n(x)\varphi_n(t)}{\lambda_n}$ converges to $K$ in the mean square on $[a,b]^2$.
- $K$ has only finitely many characteristic numbers if and only if it is separable, and then $K(x,t)$ $=\sum_n\frac{\varphi_n(x)\varphi_n(t)}{\lambda_n}$ exactly.
- (Mercer.) If $K$ is positive (all $\lambda_n\gt0$), the bilinear series converges to $K$ absolutely and uniformly on $[a,b]^2$, and $$\sum_n\frac1{\lambda_n}=\int_a^bK(x,x)\,dx.$$
Proof (of (1)–(3), with (1) sketched; (4) stated). (1) Uniform convergence. By Cauchy–Schwarz and the inequality $\sum_n\varphi_n(x)^2/\lambda_n^2$ $\le\int K(x,t)^2dt$ $\le M^2(b-a)$ from the proof of Theorem 7.15(4), $$\sum_{n=p}^q\left|\frac{(h,\varphi_n)\varphi_n(x)}{\lambda_n}\right|\le\Big(\sum_{n=p}^q(h,\varphi_n)^2\Big)^{1/2}M\sqrt{b-a},$$ and the right side tends to $0$ as $p,q\to\infty$ by Bessel's inequality for $h$, uniformly in $x$. The limit is $g$. The coefficients are right: $(g,\varphi_n)$ $=(\mathcal{K}h,\varphi_n)$ $=(h,\mathcal{K}\varphi_n)$ $=(h,\varphi_n)/\lambda_n$. The kernel $K^{(N)}$ $=K-\sum_{n\le N}\varphi_n(x)\varphi_n(t)/\lambda_n$ is symmetric, kills $\varphi_1,\dots,\varphi_N$ and has characteristic numbers $\lambda_{N+1},\lambda_{N+2},\dots$ only; by the extremal property in Theorem 7.15(3), $\|K^{(N)}h\|_2$ $\le\|h\|_2/|\lambda_{N+1}|\to0$ (or $K^{(N)}\equiv0$ if the list stops at $N$). So $g-\sum_{n\le N}(g,\varphi_n)\varphi_n$ $=\mathcal{K}^{(N)}h\to0$ in mean square, and the uniform limit must be $g$.
(2) $K_2(\cdot,t)$ $=\mathcal{K}[K(\cdot,t)]$ is sourcewise representable with $(K(\cdot,t),\varphi_n)$ $=\varphi_n(t)/\lambda_n$, so (1) gives $K_2(x,t)$ $=\sum\varphi_n(x)\varphi_n(t)/\lambda_n^2$, uniformly in $x$ (and in $t$, by the estimate above); similarly for $m\gt2$. Put $x=t$ and integrate term by term: $\int K_m(x,x)\,dx$ $=\sum1/\lambda_n^m$. For $m=2$, $\int K_2(x,x)\,dx$ $=\iint K(x,s)K(s,x)\,ds\,dx$ $=\iint K^2$ by symmetry. Finally, expanding the square and using orthonormality, $$\iint\Big(K-\sum_{n\le N}\frac{\varphi_n(x)\varphi_n(t)}{\lambda_n}\Big)^2=\iint K^2-\sum_{n\le N}\frac1{\lambda_n^2}\to0.$$ (3) If there are finitely many $\lambda_n$, the remainder kernel $K^{(N)}$ is symmetric with no characteristic numbers, hence zero by Theorem 7.15(3); the converse follows from the Fredholm alternative for separable kernels in §7.1. (4) is Mercer's theorem (1909); its proof uses positivity to show that the partial sums of $\sum\varphi_n(x)^2/\lambda_n$ increase to $K(x,x)$ and then Dini's theorem. ∎
- If $\lambda$ is not a characteristic number, the unique solution of $y=f+\lambda\mathcal{K}y$ is $$y(x)=f(x)+\lambda\sum_n\frac{(f,\varphi_n)}{\lambda_n-\lambda}\,\varphi_n(x),$$ the series converging absolutely and uniformly. Formally, $R(x,t;\lambda)$ $=\sum_n\frac{\varphi_n(x)\varphi_n(t)}{\lambda_n-\lambda}$.
- If $\lambda=\lambda_k$, the equation is solvable iff $(f,\varphi_j)=0$ for every eigenfunction $\varphi_j$ belonging to $\lambda_k$; then the solutions are the same series, summed over $n$ with $\lambda_n\ne\lambda_k$, plus an arbitrary combination of those $\varphi_j$.
Proof. (1) Necessity of the formula. If $y$ solves the equation, $y-f=\mathcal{K}(\lambda y)$ is sourcewise representable, so by Theorem 7.16(1) $y-f=\sum c_n\varphi_n$ with $c_n=(y-f,\varphi_n)$ $=\lambda(\mathcal{K}y,\varphi_n)$ $=\frac{\lambda}{\lambda_n}(y,\varphi_n)$ $=\frac{\lambda}{\lambda_n}\big((f,\varphi_n)+c_n\big)$. Solving, $c_n(\lambda_n-\lambda)$ $=\lambda(f,\varphi_n)$, which gives the formula when no $\lambda_n$ equals $\lambda$, and also uniqueness. Sufficiency. Write $\frac{(f,\varphi_n)}{\lambda_n-\lambda}$ $=\frac{\lambda_n}{\lambda_n-\lambda}\cdot\frac{(f,\varphi_n)}{\lambda_n}$; the factors $\frac{\lambda_n}{\lambda_n-\lambda}$ are bounded (they tend to $1$), so the series converges absolutely and uniformly by the estimate in the proof of Theorem 7.16(1). Apply $\mathcal{K}$ term by term ($\mathcal{K}\varphi_n$ $=\varphi_n/\lambda_n$) and use the expansion $\mathcal{K}f$ $=\sum_n\frac{(f,\varphi_n)}{\lambda_n}\varphi_n$ of Theorem 7.16(1): $$\lambda\mathcal{K}y=\lambda\sum_n(f,\varphi_n)\varphi_n\left[\frac1{\lambda_n}+\frac{\lambda}{\lambda_n(\lambda_n-\lambda)}\right],$$ and the bracket equals $\frac{1}{\lambda_n-\lambda}$, so $\lambda\mathcal{K}y=y-f$. (2) If $\lambda=\lambda_k$, the relation $c_n(\lambda_n-\lambda)$ $=\lambda(f,\varphi_n)$ for the $\varphi_n$ belonging to $\lambda_k$ forces $(f,\varphi_n)=0$ and leaves those $c_n$ free; this is the Fredholm alternative with $K^{T}=K$. ∎
- "Eigenvalue" means two things. In $y=\lambda\int Ky$, $\lambda$ is a characteristic number; the eigenvalue of the operator is $1/\lambda$. Exam options often list both — for $xt$ on $[0,1]$ the characteristic number is $3$ and the operator eigenvalue $\frac13$.
- No characteristic numbers at all. $\sin x\cos t$ on $[0,\pi]$ ($K_2\equiv0$, so $R=K$ for every $\lambda$), and every continuous Volterra kernel. An entire resolvent does not mean the kernel is Volterra.
- Only non-real ones. $x-t$ on $[0,1]$: $\pm2\sqrt3\,i$. Reality is guaranteed only for symmetric (Hermitian) kernels.
- The Neumann condition is sufficient, not necessary. $|\lambda|M(b-a)\lt1$ and $|\lambda|B\lt1$ are guarantees; the true radius is $|\lambda_1|$ (Theorem 7.13): $3$ for $xt$ (bound $1$), $\tfrac32$ for $xt+x^2t^2$ on $[-1,1]$ (bounds $\tfrac14$ and $1.286$).
- Divergent Neumann series, existing solution. Beyond $|\lambda_1|$ the series diverges, but for non-characteristic $\lambda$ the resolvent and the solution still exist (Example 6(d)).
- Non-symmetric kernels: no orthogonality. $K=1-2x+3t$ on $[0,1]$ has characteristic numbers $1$ and $2$ with eigenfunctions $1-x$ and $4x-3$, and $\int_0^1(1-x)(4x-3)\,dx$ $=-\tfrac56\ne0$.
- Hilbert–Schmidt expands only the range of $\mathcal{K}$. For $K=xt$ on $[0,1]$ ($\varphi_1=\sqrt3x$), the "expansion" of $f=1$ is $(1,\varphi_1)\varphi_1$ $=\tfrac32x\ne1$: $1$ is not of the form $\int_0^1xt\,h(t)\,dt$.
- Uniform convergence of the bilinear series needs Mercer. For general symmetric kernels it converges only in mean square, and the trace formula $\sum1/\lambda_n=\int K(x,x)\,dx$ is guaranteed only for positive (or separable) kernels; for $m\ge2$ the trace formulas always hold.
- Count with multiplicity. For $\cos(x-t)$ on $[0,2\pi]$, $\sum1/\lambda_n^2$ must count $\frac1\pi$ twice: $2\pi^2=\iint\cos^2(x-t)$. Counting it once gives $\pi^2$, which is wrong.
- $K_2$ is not $K$ squared pointwise. $K_2(x,t)=\int K(x,s)K(s,t)\,ds$; its characteristic numbers are $\lambda_n^2$, all positive for symmetric $K$.
- The null space is not an eigenspace. $\int_0^1xt\,y(t)\,dt=0$ for every $y\perp t$ — an infinite-dimensional null space — but $0$ is never a characteristic number.
Exercises
- 1 Find the characteristic numbers and eigenfunctions of $e^{x+t}$ on $[0,1]$.
- 2 Find the characteristic numbers and eigenfunctions of $x+t$ on $[-1,1]$, and check that the eigenfunctions are orthogonal.
- 3 Find the characteristic numbers, with multiplicities, of $\cos^2(x-t)$ on $[0,\pi]$.
- 4 Find the characteristic numbers and eigenfunctions of $1-\max(x,t)$ on $[0,1]$.
- 5 For $K=xt$ on $[0,1]$ compute $B$, and show that the bound $|\lambda|\ge1/B$ is attained.
- 6 Show that the set of solutions of $\int_0^1xt\,y(t)\,dt=0$ is infinite-dimensional, and explain why this does not make $0$ a characteristic number.
- 7 Find the characteristic numbers and eigenfunctions of $K=1-2x+3t$ on $[0,1]$ and of its transpose, and show that $\int_0^1fz\,dx=0$ fails for $f=1$ at both characteristic numbers.
- 8 Show directly (without the general Volterra theorem of §7.1) that the Volterra kernel $e^{x-t}$ on $0\le t\le x\le1$ has no characteristic numbers.
- 1 Find the resolvent of $xt$ on $[0,1]$ and solve $y(x)=x^2+\lambda\int_0^1xt\,y(t)\,dt$.
- 2 For $K=x+t$ on $[0,1]$ compute $K_2$ and $K_3$.
- 3 For the Fredholm kernel $e^{x-t}$ on $[0,1]$ show $K_n=K$, find $R$, and solve $y(x)=1+\lambda\int_0^1e^{x-t}y(t)\,dt$.
- 4 For $K=\sin x\cos t$ on $[0,\pi]$ show $K_2\equiv0$. What are $R$ and the solution of $y=f+\lambda\mathcal{K}y$?
- 5 For $K=x+t$ on $[0,1]$ compare $\frac1{M(b-a)}$, $\frac1B$ and the radius of convergence of the Neumann series.
- 6 Find the resolvent of the Volterra kernel $x-t$ and solve $y(x)=1+\lambda\int_0^x(x-t)\,y(t)\,dt$ for $\lambda\gt0$.
- 7 Find the resolvent of the Volterra kernel $K\equiv1$ ($0\le t\le x$).
- 8 Use the Laplace transform to find the resolvent of the Volterra kernel $\sin(x-t)$ for $\lambda=1$, and solve $y(x)=x+\int_0^x\sin(x-t)\,y(t)\,dt$ with it.
- 9 Derive the resolvent equation $R(x,t;\lambda)-K(x,t)$ $=\lambda\int_a^bK(x,s)R(s,t;\lambda)\,ds$ from the series, and show that, conversely, a continuous $R$ satisfying both resolvent equations gives the unique solution $y=f+\lambda\int Rf$.
- 1 For $K=x+t$ on $[-1,1]$ (Exercise I.2) verify $\sum1/\lambda_n^2=\iint K^2$ and $\sum1/\lambda_n=\int K(x,x)\,dx$.
- 2 Solve $y(x)=x+\lambda\int_0^1G_0(x,t)\,y(t)\,dt$, $G_0=\min(x,t)(1-\max(x,t))$, by Schmidt's formula, and identify the sum in closed form. For which $\lambda$ is there no solution?
- 3 Solve $y(x)=x+\lambda\int_{-1}^1(1+xt)\,y(t)\,dt$ by Schmidt's formula; what happens at $\lambda=\tfrac12$ and at $\lambda=\tfrac32$?
- 4 Prove that a real continuous symmetric kernel with $K_2\equiv0$ is identically zero. Is this true without symmetry?
- 5 Prove that a real symmetric separable kernel of rank $n$ has exactly $n$ characteristic numbers, counted with multiplicity.
- 6 Prove that $\min(x,t)$ is a positive kernel on $[0,1]$ by showing $\iint\min(x,t)h(x)h(t)\,dx\,dt$ $=\int_0^1\left(\int_s^1h\right)^2ds$.
- 7 Check the trace formula for $e^{x+t}$ on $[0,1]$ (Exercise I.1).
- 8 For $\cos(x-t)$ on $[0,2\pi]$ compute $K_2$, the resolvent, and the bilinear expansion, and check $\sum1/\lambda_n^2=\iint K^2$.
- 1 Every continuous kernel on $[a,b]^2$ has at least one characteristic number.
- 2 If $\lambda$ is a characteristic number of a separable kernel $K$, it is a characteristic number of $K^{T}$.
- 3 For a real continuous symmetric kernel, the Neumann series converges for $|\lambda|\lt|\lambda_1|$ and diverges (for suitable $f$) for $|\lambda|\gt|\lambda_1|$.
- 4 If $K$ is real, continuous and symmetric with characteristic numbers $\lambda_n$, then the characteristic numbers of $K_2$ are exactly the $\lambda_n^2$.
- 5 There is a continuous kernel on $[0,1]^2$ whose only characteristic numbers are $\pm2\sqrt3\,i$.
E1. CSIR Part B GATE MA The resolvent kernel of $K(x,t)=xt$ on $[0,1]$ is: (A) $\frac{3xt}{3-\lambda}$; (B) $\frac{xt}{1-\lambda}$; (C) $xt\,e^{\lambda/3}$; (D) $\frac{3xt}{3+\lambda}$.
E2. CSIR Part B The resolvent kernel of the Volterra equation $y(x)=f(x)+\lambda\int_0^xy(t)\,dt$ is: (A) $e^{\lambda(x-t)}$; (B) $e^{\lambda(x+t)}$; (C) $\frac1{1-\lambda}$; (D) $\cosh\lambda(x-t)$.
E3. CSIR Part B Let $\lambda_1,\lambda_2,\dots$ be the characteristic numbers of $K(x,t)=e^{-|x-t|}$ on $[0,1]$, repeated according to multiplicity. Then $\sum_n\frac1{\lambda_n^2}$ equals: (A) $\frac{1+e^{-2}}{2}$; (B) $1$; (C) $\frac{1-e^{-2}}{2}$; (D) $1-e^{-1}$.
E4. CSIR Part C Let $K(x,t)=x-t$ on $[0,1]$. Which of the following are true? (One or more options may be correct.) (A) $K$ has no real characteristic number. (B) The Neumann series for $y=f+\lambda\int_0^1(x-t)y(t)\,dt$ converges for every $|\lambda|\lt2\sqrt3$. (C) For every real $\lambda$ and every continuous $f$ the equation has exactly one solution. (D) $K_2\equiv0$.
E5. CSIR Part C Let $K\not\equiv0$ be a real continuous symmetric kernel on $[a,b]^2$, with orthonormal eigenfunctions $\varphi_n$. Which of the following are necessarily true? (A) $K$ has at least one characteristic number. (B) All characteristic numbers of $K$ are real. (C) Every $f\in C[a,b]$ satisfies $f=\sum(f,\varphi_n)\varphi_n$ uniformly. (D) If $\lambda$ is a characteristic number of $K$, then $\lambda^2$ is a characteristic number of $K_2$.
E6. CSIR Part C Consider $y(x)=f(x)+\lambda\int_0^xe^{x-t}y(t)\,dt$ on $[0,1]$. Which of the following are true? (A) The resolvent kernel is $e^{(1+\lambda)(x-t)}$. (B) For $\lambda=-1$ the solution is $y=f$. (C) For $\lambda=2$ and $f\equiv1$ the solution is $y=\frac{1+2e^{3x}}{3}$. (D) For some $\lambda$ the homogeneous equation has a nonzero continuous solution.
Suggested plan — 3 lectures × 60 min.
- Characteristic numbers. Definition 7.4 and the two meanings of "eigenvalue"; Theorem 7.10 with full proof (Cauchy–Schwarz, then Bessel); Examples 1 and 2 on the board. State Fredholm's theorems (7.11) and explain the proof idea "polynomial plus small" in five minutes.
- Neumann series and resolvents. Definition 7.5, Theorem 7.12 with proof; Example 3 (three radii); project Interactive 1 and move $\lambda$ across $1/B$ and then $|\lambda_1|$ — let the class watch the series diverge while the dashed solution stays perfectly finite. Theorem 7.14 and Example 4; the strategy box.
- Symmetric kernels. Theorem 7.15(1), (2), (5) with proofs, (3) as a sketch; Hilbert–Schmidt and Mercer (7.16) with the proof of the trace identity; Interactive 2 (start with $t_0$ near the end point, then at $t_0=0.5$); Examples 5–7; finish with E4 and E5 live.
Discussion questions. (1) For matrices, a symmetric matrix always has real eigenvalues and an orthonormal eigenbasis. Which part of this survives for symmetric kernels, and which part fails (think of $xt$)? (2) Why is the radius of convergence of the Neumann series the distance to the nearest characteristic number, and not $1/B$? (3) The kernel $\min(x,t)$ is the covariance of Brownian motion. What does Mercer's theorem say about random functions?
Misconceptions to watch for. Confusing $\lambda$ with $1/\lambda$; believing every kernel has characteristic numbers, or that they are always real; treating the sufficient condition $|\lambda|M(b-a)\lt1$ as the exact range; thinking that divergence of the Neumann series means there is no solution; expanding an arbitrary $f$ in eigenfunctions of a kernel of finite rank; counting a multiple characteristic number once in $\sum1/\lambda_n^2$; computing $K_2$ as $K^2$.
Suggested homework. I(1, 2, 3, 7), II(1, 3, 5, 6, 8), III(1, 3, 6, 8), IV(1, 3, 5); Exam corner E1, E3, E4, E6.
Research question. Brownian motion $W(x)$, $0\le x\le1$ — the model behind stock prices in option pricing, diffusing molecules and noise in electronic circuits — has covariance $\mathbb{E}[W(x)W(t)]=\min(x,t)$. How many eigenfunctions of this kernel are needed to reproduce Brownian paths, and how well does a computer recover its characteristic numbers?
Why it matters. Mercer's theorem for a covariance kernel becomes the Karhunen–Loève expansion $W(x)=\sum_nZ_n\frac{\varphi_n(x)}{\sqrt{\lambda_n}}$ with independent standard normal $Z_n$ — the continuous version of principal component analysis, used in signal compression, uncertainty quantification in engineering, and in generating random fields for climate and groundwater models.
What to do.
- Using Example 2, write the Karhunen–Loève expansion of $W$, and check that $\mathbb{E}[W(x)W(t)]$ $=\sum_n\frac{\varphi_n(x)\varphi_n(t)}{\lambda_n}$ $=\min(x,t)$.
- The fraction of the total variance $\int_0^1\mathbb{E}[W(x)^2]\,dx$ $=\tfrac12$ captured by the first $N$ terms is $\frac{2}{1}\sum_{n\le N}\frac1{\lambda_n}$. Compute it for $N=1,2,3,5,10$.
- Simulate paths with $N=5$ and $N=100$ terms (seeded random numbers), and compare them with a scaled random walk with $1000$ steps.
- Nyström's method. Replace $\int_0^1$ by the trapezoidal rule on $m$ equally spaced points, so that the integral equation becomes a matrix eigenvalue problem, and compute the smallest characteristic number for $m=11,21,41,81$. How fast does it approach $\pi^2/4$?
Tools: Python with NumPy (random numbers and numpy.linalg.eigvalsh), or any language with an eigenvalue routine.
Hand in: The variance table, two plots of simulated paths, and a table of Nyström approximations with their errors and your estimate of the order of convergence.
Going further. Find the eigenfunctions of the Brownian bridge covariance $\min(x,t)-xt$ (it is $G_0$!), and of the exponential covariance $e^{-|x-t|}$ used for random fields, whose characteristic numbers solve a transcendental equation.
🧪 Self-check quiz
Q1. The characteristic number of $K(x,t)=xt$ on $[0,1]$ is
Q2. Which condition guarantees that the Neumann series of a continuous kernel converges?
Q3. The resolvent kernel of a continuous Volterra kernel, as a function of $\lambda$, is
Q4. For a real continuous symmetric kernel, the characteristic numbers are always
Q5. For a real continuous symmetric kernel with characteristic numbers $\lambda_n$ (repeated by multiplicity), $\sum_n1/\lambda_n^2$ equals