7.1 Fredholm and Volterra Equations; Separable Kernels
Linear integral equations of Fredholm and Volterra type: classification, the passage from initial and boundary value problems to integral equations and back, separable kernels and the Fredholm alternative, successive approximations for Volterra equations, convolution kernels, the Laplace transform and Abel's equation.
An initial value problem such as $y''+q(x)\,y=0$, $y(0)=1$, $y'(0)=0$ asks for a function whose second derivative is prescribed in terms of the function itself. Integrating twice turns it into a single equation, $$y(x)=1-\int_0^x(x-t)\,q(t)\,y(t)\,dt,$$ in which the unknown function appears under an integral sign and the initial data are built into the right-hand side. A boundary value problem, solved with its Green's function from §2.4, turns into an equation of the same shape in which the integral runs over the whole interval. Equations of this kind — linear integral equations — are often easier to analyse than the differential equations they come from: they need no differentiability of the unknown, they contain the side conditions automatically, and they can be attacked by iteration, by linear algebra and by transforms. They also arise directly, without any differential equation behind them: in Abel's problem of recovering the shape of a slide from the descent times, in population models in which births depend on the whole age structure, in potential theory, in tomography and in the theory of materials with memory.
This section sets up the basic theory. We classify integral equations (first and second kind, Fredholm and Volterra), convert initial and boundary value problems into integral equations and back, and solve two large classes completely. For separable kernels $K(x,t)=\sum a_i(x)b_i(t)$ the equation reduces to a linear system $(I-\lambda A)c=F$, and the familiar dichotomy of linear algebra becomes the Fredholm alternative. For Volterra equations of the second kind, successive approximations converge for every value of the parameter, so the solution always exists and is unique; for convolution kernels the Laplace transform gives it in closed form. CSIR NET questions on this topic are short and frequent: an equivalent integral equation for an IVP, the characteristic numbers of a kernel like $x+t$ or $\cos(x+t)$, the solution of a convolution equation, and Part C statements about when solutions exist.
Classification of Linear Integral Equations
Throughout, $[a,b]$ is a bounded interval, $f$ is a given continuous function, $\lambda$ is a real or complex parameter, and the kernel $K(x,t)$ is a given function of two variables. We look for a continuous unknown function $y$.
Let $K$ be continuous on the square $Q=[a,b]\times[a,b]$ (Fredholm case) or on the triangle $T=\{(x,t):a\le t\le x\le b\}$ (Volterra case), and let $f\in C[a,b]$.
- In a Fredholm equation the limits of integration are fixed; in a Volterra equation the upper limit is the variable $x$.
- In an equation of the first kind the unknown appears only under the integral; in one of the second kind it also appears outside. (An equation $h(x)y(x)=f(x)+\lambda\int K y$ with $h$ vanishing somewhere is sometimes said to be of the third kind; we shall not need it.)
- A second-kind equation is homogeneous if $f\equiv0$, and non-homogeneous otherwise.
- The equation is linear because the unknown enters linearly; $y(x)=1+\int_0^xy(t)^2\,dt$ is a nonlinear Volterra equation. It is singular if the interval is infinite or the kernel is unbounded, as in Abel's equation $\int_0^x\frac{y(t)}{\sqrt{x-t}}\,dt=f(x)$.
A Volterra equation is a Fredholm equation in disguise: extend the kernel by $K(x,t)=0$ for $t\gt x$, and $\int_a^x K(x,t)y(t)\,dt$ $=\int_a^bK(x,t)y(t)\,dt$. The extended kernel is in general discontinuous on the diagonal, but it is bounded, and this "triangular" structure is exactly what makes Volterra equations so much better behaved (Theorem 7.7). The two kinds behave very differently. A second-kind equation is a perturbation of the identity, $y-\lambda\mathcal{K}y=f$, and behaves like a well-conditioned linear system. A first-kind equation $\mathcal{K}y=f$ is typically ill-posed: integration smooths, so $f$ must already be smooth to be in the range, and tiny wiggles in $f$ correspond to huge wiggles in $y$. For Volterra equations there is a way out — differentiate.
Proof. We use the Leibniz rule (Unit 1): if $K$ and $K_x$ are continuous on $T$ and $y$ is continuous, then $\Phi(x)=\int_a^xK(x,t)y(t)\,dt$ is continuously differentiable with $$\Phi'(x)=K(x,x)\,y(x)+\int_a^xK_x(x,t)\,y(t)\,dt.$$ If $\Phi=f$ on $[a,b]$, then $f(a)=\Phi(a)=0$, and differentiating gives $K(x,x)y(x)+\int_a^xK_xy\,dt=f'(x)$; dividing by $K(x,x)\ne0$ gives the second-kind equation. Conversely, if $f(a)=0$ and $y$ solves the second-kind equation, multiplying by $K(x,x)$ shows $\Phi'(x)=f'(x)$ on $[a,b]$; since also $\Phi(a)=0=f(a)$, the mean value theorem gives $\Phi\equiv f$. The last sentence follows because the second-kind equation has exactly one continuous solution. ∎
If $K(x,x)\equiv0$ but $K_x(x,x)\ne0$ (for instance $K=x-t$), differentiate twice; each differentiation adds a compatibility condition ($f(a)=0$, then $f'(a)=0$, …).
From Differential Equations to Integral Equations and Back
The key tool for initial value problems is the formula that collapses repeated integrals into a single one.
Proof. Let $u$ be the right-hand side. For $n=1$ it is $\int_a^xg$. For $n\ge2$ the Leibniz rule gives $u'(x)=\frac{1}{(n-2)!}\int_a^x(x-t)^{n-2}g(t)\,dt$ (the boundary term vanishes because $(x-t)^{n-1}=0$ at $t=x$). Repeating, $u^{(k)}(x)$ $=\frac{1}{(n-1-k)!}\int_a^x(x-t)^{n-1-k}g(t)\,dt$ for $k\le n-1$, so $u^{(n-1)}=\int_a^xg$, $u^{(n)}=g$, and $u^{(k)}(a)=0$ for $k\le n-1$. The $n$-fold integral $I_ng$ has the same properties, and a function with $n$-th derivative $g$ and vanishing derivatives of orders $0,\dots,n-1$ at $a$ is unique (the difference is a polynomial of degree $\lt n$ with all derivatives zero at $a$). ∎
- (Equation for $y$.) If $p\in C^1[a,b]$, a function $y\in C^2[a,b]$ solves (IVP) if and only if $y\in C[a,b]$ solves $$y(x)=F(x)+\int_a^xK(x,t)\,y(t)\,dt,$$ where $$K(x,t)=(x-t)\big[p'(t)-q(t)\big]-p(t),$$ $$F(x)=\alpha+\big[\beta+p(a)\alpha\big](x-a)+\int_a^x(x-t)\,g(t)\,dt.$$ In particular, for $y''+q(x)y=g$ the kernel is simply $-(x-t)\,q(t)$.
- (Equation for $u=y''$.) With no extra smoothness of $p$: $y$ solves (IVP) if and only if $u=y''$ is continuous and solves $$u(x)=g(x)-\beta p(x)-q(x)\big[\alpha+\beta(x-a)\big]-\int_a^x\big[p(x)+q(x)(x-t)\big]u(t)\,dt,$$ and then $y(x)=\alpha+\beta(x-a)+\int_a^x(x-t)\,u(t)\,dt$.
Proof. (1) Suppose $y$ solves (IVP). Integrate $y''=g-py'-qy$ from $a$ to $x$, integrating the term $py'$ by parts: $$y'(x)-\beta=\int_a^xg\,dt-\Big[p\,y\Big]_a^x+\int_a^xp'(t)y(t)\,dt-\int_a^xq(t)y(t)\,dt,$$ so $$y'(x)=\beta+p(a)\alpha-p(x)y(x)+\int_a^x\big[p'(t)-q(t)\big]y(t)\,dt+\int_a^xg\,dt.$$ Integrate once more from $a$ to $x$ and use Theorem 7.2 ($n=2$) on the two integrals: $$y(x)=\alpha+\big[\beta+p(a)\alpha\big](x-a)-\int_a^xp(t)y(t)\,dt+\int_a^x(x-t)\big[p'(t)-q(t)\big]y(t)\,dt+\int_a^x(x-t)g(t)\,dt,$$ which is the stated equation.
Conversely let $y\in C[a,b]$ solve the integral equation. The right-hand side is $C^1$ (Leibniz rule), so $y\in C^1$, and since $K(x,x)=-p(x)$ and $K_x(x,t)=p'(t)-q(t)$, $$y'(x)=\beta+p(a)\alpha+\int_a^xg\,dt-p(x)y(x)+\int_a^x\big[p'(t)-q(t)\big]y(t)\,dt.$$ The right-hand side is again $C^1$ (now that $y\in C^1$ and $p\in C^1$), so $y\in C^2$ and $y''=g-p'y-py'+(p'-q)y=g-py'-qy$. Finally $y(a)=F(a)=\alpha$ and $y'(a)=\beta+p(a)\alpha-p(a)\alpha$ $=\beta$.
(2) If $y$ solves (IVP) and $u=y''$, then by Theorem 7.2, $y'=\beta+\int_a^xu$ and $y=\alpha+\beta(x-a)+\int_a^x(x-t)u\,dt$; substituting into $u=g-py'-qy$ gives the equation for $u$. Conversely, if $u$ is a continuous solution, define $y$ by the last formula; then $y''=u$, $y(a)=\alpha$, $y'(a)=\beta$, and the equation for $u$ says exactly that $y''=g-py'-qy$. ∎
The integral equation is a complete substitute for the IVP: the initial data sit in $F$, and by Theorem 7.7 it has exactly one continuous solution on the whole of $[a,b]$ — a second proof of the global existence theorem for linear equations (§2.1).
For boundary value problems the right tool is the Green's function of §2.4. There we worked with $L[u]=(p\,u')'+q\,u$ on $[a,b]$ ($p\in C^1$, $p\gt0$, $q$ continuous) and separated homogeneous conditions $B_a[u]$ $=\alpha_1u(a)+\alpha_2u'(a)=0$, $B_b[u]=\beta_1u(b)+\beta_2u'(b)=0$, and proved: if the homogeneous problem has only the trivial solution, there is a unique, symmetric Green's function $G(x,t)$, and for every $h\in C[a,b]$ the function $u=\int_a^bG(x,t)h(t)\,dt$ is the unique solution of $L[u]=h$ with the boundary conditions.
Proof. If $y\in C^2$ solves the boundary value problem, then $L[y]=h-\lambda ry$ with $h-\lambda ry$ continuous, and the solution formula of §2.4 gives $y=\int_a^bG\,(h-\lambda ry)\,dt$. Conversely, if $y$ is continuous and satisfies the integral equation, put $\tilde h=h-\lambda ry\in C[a,b]$. The equation says $y=\int_a^bG\tilde h\,dt$, and by §2.4 this function is $C^2$, satisfies both boundary conditions, and $L[y]=\tilde h=h-\lambda ry$. ∎
Two remarks. First, the kernel $G(x,t)r(t)$ is not symmetric unless $r$ is constant; if $r\gt0$, the substitution $z=\sqrt{r}\,y$ produces the symmetric kernel $\sqrt{r(x)}\,G(x,t)\sqrt{r(t)}$, which is how Sturm–Liouville problems with a weight become symmetric integral equations (§7.2). Second, the hypothesis on the homogeneous problem matters: for Neumann conditions $u'(0)=u'(1)=0$ the operator $u''$ has no Green's function (constants solve the homogeneous problem), and one must first move a multiple of $y$ to the other side (Exercise II.5).
- IVP → Volterra. Solve for the highest derivative, integrate $n$ times from the initial point, integrate terms like $p\,y'$ by parts, and collapse repeated integrals with $\int_a^x\!\int_a^s g$ $=\int_a^x(x-t)g(t)\,dt$. The initial values go into the free term; the kernel never contains them.
- BVP → Fredholm. Find the Green's function $G$ of the part of the operator you keep on the left (it must have no nontrivial homogeneous solution — shift a multiple of $y$ across if necessary), then write $y=\int G\cdot(\text{everything else})$. Inhomogeneous boundary values: subtract a function satisfying them first (Exercise II.3).
- Integral → differential. Split the kernel at $t=x$, differentiate with the Leibniz rule until the integrals disappear, and read off the initial or boundary conditions by putting $x=a$ (and $x=b$) in the equation and in its derivatives.
- Check by putting $\lambda=0$: the free term must solve the problem with the $\lambda$-term removed.
Separable Kernels and the Fredholm Alternative
When the kernel is a finite sum of products, the integral $\int_a^bK(x,t)y(t)\,dt$ is a combination of finitely many fixed functions of $x$, with coefficients that are numbers depending on $y$. The equation then reduces to finitely many unknowns.
Examples: $xt$ (rank $1$), $x+t=x\cdot1+1\cdot t$ (rank $2$), $1+3xt$, $\cos(x-t)$ $=\cos x\cos t+\sin x\sin t$, $\sin(x+t)$, $e^{x+t}=e^xe^t$, and every polynomial in $x$ and $t$. The kernels $e^{xt}$, $\min(x,t)$ and $\frac1{1+(x-t)^2}$ are not separable.
- If $y\in C[a,b]$ solves $y(x)=f(x)+\lambda\int_a^bK(x,t)y(t)\,dt$, then the numbers $c_i=\int_a^bb_i(t)y(t)\,dt$ satisfy $$(I-\lambda A)\,c=F,\qquad\text{and}\qquad y(x)=f(x)+\lambda\sum_{i=1}^nc_i\,a_i(x).$$
- Conversely, if $c\in\mathbb{C}^n$ solves $(I-\lambda A)c=F$, then $y=f+\lambda\sum c_ia_i$ solves the integral equation, and $\int_a^b b_iy\,dt=c_i$.
Proof. (1) $\int_a^bK(x,t)y(t)\,dt$ $=\sum_ia_i(x)\int_a^bb_i(t)y(t)\,dt$ $=\sum_ic_ia_i(x)$, so $y=f+\lambda\sum_jc_ja_j$. Multiply this identity by $b_i$ and integrate: $c_i=F_i+\lambda\sum_jA_{ij}c_j$, i.e. $(I-\lambda A)c=F$. (2) If $(I-\lambda A)c=F$ and $y=f+\lambda\sum_jc_ja_j$, then $\int_a^bb_iy\,dt$ $=F_i+\lambda\sum_jA_{ij}c_j=c_i$; hence $\int_a^bKy\,dt=\sum_ia_ic_i$ and $f+\lambda\int Ky$ $=f+\lambda\sum c_ia_i=y$. Starting from $y$, (1) produces $c$ and the formula $y=f+\lambda\sum c_ia_i$ returns the same $y$; starting from $c$, (2) produces $y$ with $\int b_iy=c_i$. So the maps are mutually inverse. ∎
To state the alternative we need two definitions that §7.2 develops much further.
- $D$ is a polynomial of degree at most $n$ with $D(0)=1$. The characteristic numbers of $K$ are exactly the roots of $D$; there are at most $n$ of them, and there may be none, or only non-real ones.
- If $D(\lambda)\ne0$, then for every $f\in C[a,b]$ the equation $y=f+\lambda\int_a^bKy\,dt$ has exactly one solution, namely $$y(x)=f(x)+\lambda\int_a^bR(x,t;\lambda)\,f(t)\,dt,\qquad R(x,t;\lambda)=\sum_{i,j=1}^na_i(x)\big[(I-\lambda A)^{-1}\big]_{ij}\,b_j(t).$$
- If $D(\lambda)=0$, then the homogeneous equation has exactly $r=n-\operatorname{rank}(I-\lambda A)\ge1$ linearly independent solutions; the transposed homogeneous equation also has exactly $r$ linearly independent solutions $z_1,\dots,z_r$; and the non-homogeneous equation is solvable if and only if $$\int_a^bf(x)\,z_k(x)\,dx=0\qquad(k=1,\dots,r).$$ In that case its general solution is $y_p+\sum_{k}C_ky_k$, where $y_p$ is any one solution and $y_1,\dots,y_r$ are independent eigenfunctions.
Proof. (1) Every entry of $I-\lambda A$ is a polynomial of degree $\le1$ in $\lambda$, so $D$ has degree $\le n$, and $D(0)=\det I=1$. By Theorem 7.5 with $f=0$, the solutions of the homogeneous equation are the functions $y=\lambda\sum c_ia_i$ with $(I-\lambda A)c=0$. For $\lambda\ne0$ the map $c\mapsto\lambda\sum c_ia_i$ is injective because the $a_i$ are linearly independent; so a nonzero solution exists iff $(I-\lambda A)c=0$ has a nonzero solution, i.e. iff $D(\lambda)=0$.
(2) If $D(\lambda)\ne0$, the system has the unique solution $c=(I-\lambda A)^{-1}F$, so by Theorem 7.5 the integral equation has exactly one solution, $y=f+\lambda\sum_ia_i\sum_j[(I-\lambda A)^{-1}]_{ij}F_j$. Substituting $F_j=\int b_j(t)f(t)\,dt$ gives the formula with $R$.
(3) The injective map above identifies the solution space of the homogeneous equation with the null space of $I-\lambda A$, of dimension $n-\operatorname{rank}(I-\lambda A)=r\ge1$. The transposed kernel $K(t,x)=\sum_ib_i(x)a_i(t)$ is separable with the roles of $a_i$ and $b_i$ exchanged, so its matrix is $A'_{ij}=\int a_ib_j=A_{ji}$, i.e. $A'=A^{T}$. Its homogeneous solutions are $z=\lambda\sum d_ib_i$ with $(I-\lambda A^{T})d=0$, and since $\operatorname{rank}(I-\lambda A^{T})$ $=\operatorname{rank}(I-\lambda A)$, there are again exactly $r$ independent ones. By Theorem 7.5 the non-homogeneous equation is solvable iff $(I-\lambda A)c=F$ is solvable, and by linear algebra this holds iff $F\cdot d=0$ for every $d$ with $(I-\lambda A)^{T}d=0$ (the column space of a matrix is the orthogonal complement of the null space of its transpose). For such $d$ and the corresponding $z=\lambda\sum d_ib_i$, $$F\cdot d=\sum_id_i\int_a^bb_i f\,dx=\int_a^bf\sum_id_ib_i\,dx=\frac1\lambda\int_a^bf\,z\,dx,$$ so the condition is $\int fz=0$ for every solution $z$ of the transposed homogeneous equation. The description of the general solution is linearity. ∎
This is the integral-equation version of the Fredholm alternative for boundary value problems (§2.4), and of the fact that $Ax=b$ is solvable iff $b\perp\ker A^{T}$. Fredholm's great discovery (§7.2) is that the same three statements hold for every continuous kernel, with $D(\lambda)$ replaced by an entire function. Note that the solvability condition involves the eigenfunctions of the transposed kernel; they coincide with those of $K$ only when $K$ is symmetric.
- Write $K=\sum a_i(x)b_i(t)$ with as few terms as possible (expand $\cos(x\pm t)$, $\sin(x\pm t)$, $(x+t)^2$, …).
- Name the unknown constants $c_i=\int b_i y$, so that $y=f+\lambda\sum c_ia_i$. Multiply by each $b_i$ and integrate: this gives $(I-\lambda A)c=F$ with $A_{ij}=\int b_ia_j$ — mind the order of the indices.
- Characteristic numbers: roots of $D(\lambda)=\det(I-\lambda A)$. Eigenfunctions: $\sum c_ia_i$ for $c$ in the null space of $I-\lambda A$.
- At a characteristic number, test solvability with the transposed equation (or simply look for inconsistency in the linear system), and add the eigenfunctions to a particular solution.
- Check by substituting the answer back: two or three integrals, and they catch nearly every sign error.
Volterra Equations: Successive Approximations and the Laplace Transform
For Fredholm equations, iteration converges only for small $|\lambda|$ (§7.2). Volterra equations are different: each integration over $[a,x]$ produces a factor $(x-a)^n/n!$, and factorials beat every geometric growth.
- $|\varphi_n(x)|$ $\le\|f\|\,\dfrac{M^n(x-a)^n}{n!}$ for all $n$ and $x$;
- the series $y=\sum_{n=0}^\infty\lambda^n\varphi_n$ converges uniformly on $[a,b]$, and $y$ is the unique continuous solution of $$y(x)=f(x)+\lambda\int_a^xK(x,t)\,y(t)\,dt;$$
- the successive approximations $y_0=f$, $y_{n+1}=f+\lambda\int_a^xKy_n\,dt$ satisfy $y_n=\sum_{k=0}^n\lambda^k\varphi_k$ and $$\|y-y_n\|\le\|f\|\sum_{k=n+1}^{\infty}\frac{\big(|\lambda|M(b-a)\big)^k}{k!}\le\|f\|\,\frac{\big(|\lambda|M(b-a)\big)^{n+1}}{(n+1)!}\,e^{|\lambda|M(b-a)}.$$
Proof. Each $\varphi_n$ is continuous (an integral of a continuous function over $[a,x]$ with a continuous kernel). (1) by induction: true for $n=0$; if it holds for $n-1$, then $$|\varphi_n(x)|\le\int_a^xM\,\|f\|\frac{M^{n-1}(t-a)^{n-1}}{(n-1)!}\,dt=\|f\|\frac{M^n(x-a)^n}{n!}.$$ (2) By (1), $\|\lambda^n\varphi_n\|$ $\le\|f\|\frac{(|\lambda|M(b-a))^n}{n!}$, the terms of a convergent exponential series, so the Weierstrass M-test gives uniform convergence of $\sum\lambda^n\varphi_n$ to a continuous $y$. Uniform convergence lets us integrate term by term: $$\lambda\int_a^xK(x,t)y(t)\,dt=\sum_{n\ge0}\lambda^{n+1}\int_a^xK\varphi_n\,dt,$$ $$\sum_{n\ge0}\lambda^{n+1}\int_a^xK\varphi_n\,dt=\sum_{n\ge0}\lambda^{n+1}\varphi_{n+1}=y-f.$$ So $y$ is a solution. For uniqueness let $y,\tilde y$ be two continuous solutions and $w=y-\tilde y$, $W=\max|w|$. Then $w=\lambda\int_a^xKw\,dt$, so $|w(x)|\le|\lambda|MW(x-a)$, and inductively $|w(x)|$ $\le W\frac{(|\lambda|M(x-a))^n}{n!}$ for every $n$. The right side tends to $0$, so $w\equiv0$. (3) $y_n=\sum_{k\le n}\lambda^k\varphi_k$ by induction ($y_{n+1}$ $=f+\lambda\int K\sum_{k\le n}\lambda^k\varphi_k$ $=f+\sum_{k\le n}\lambda^{k+1}\varphi_{k+1}$), and the tail estimate follows from (1); the last inequality uses $\sum_{k\ge n+1}\frac{u^k}{k!}$ $\le\frac{u^{n+1}}{(n+1)!}e^u$. ∎
Compare Picard's theorem (§1.1): there the existence interval had length $h=\min(a,b/M)$ because the nonlinear right-hand side was controlled only near the initial point. In the linear case the Lipschitz condition holds on the whole strip, and the solution exists on all of $[a,b]$.
When the kernel depends only on $x-t$ and $a=0$, the integral is a convolution, $(k*y)(x)=\int_0^xk(x-t)y(t)\,dt$, and the Laplace transform $\mathcal{L}[g](s)$ $=\int_0^\infty e^{-sx}g(x)\,dx$ turns it into a product.
Proof (sketch). Theorem 7.7, applied on every $[0,b]$, gives a unique continuous solution on $[0,\infty)$. Put $\tilde y=e^{-\sigma x}y$; it satisfies a Volterra equation with right-hand side $e^{-\sigma x}f$ (bounded by $C$) and kernel $e^{-\sigma(x-t)}k(x-t)$ (bounded by $C$), so the estimate in Theorem 7.7(2) gives $|\tilde y(x)|\le Ce^{|\lambda|Cx}$, and $y$ is of exponential order. The convolution theorem $\mathcal{L}[k*y]=\hat k\,Y$ then turns the equation into $Y=F+\lambda\hat kY$; for large $s$, $\hat k(s)\to0$, so $1-\lambda\hat k(s)\ne0$. ∎
The most famous convolution equation has a singular kernel. In Abel's problem a bead slides from rest at height $h$ along a smooth wire to the bottom; if $\phi(y)=\frac{ds}{dy}$ describes the wire, the descent time is $T(h)=\frac{1}{\sqrt{2g}}\int_0^h\frac{\phi(y)}{\sqrt{h-y}}\,dy$. Prescribing $T$ and finding $\phi$ is the following equation.
Proof. The key integral is, for $t\lt x$, $$\int_t^x\frac{ds}{\sqrt{(x-s)(s-t)}}=\pi$$ (substitute $s=t+(x-t)\sin^2\theta$, $0\le\theta\le\tfrac\pi2$: then $ds=2(x-t)\sin\theta\cos\theta\,d\theta$ and $\sqrt{(x-s)(s-t)}$ $=(x-t)\sin\theta\cos\theta$). Uniqueness and the formula. If $y$ is a continuous solution, divide the equation (with $x$ replaced by $s$) by $\sqrt{x-s}$ and integrate over $s\in[0,x]$; interchanging the order of integration (the integrand is absolutely integrable) gives $$\int_0^x\frac{f(s)}{\sqrt{x-s}}\,ds=\int_0^xy(t)\int_t^x\frac{ds}{\sqrt{(x-s)(s-t)}}\,dt,$$ which equals $\pi\int_0^xy(t)\,dt$, and differentiating gives the first formula, so $y$ is determined by $f$. Writing $\int_0^xf(t)(x-t)^{-1/2}dt$ $=\int_0^xf(x-u)u^{-1/2}du$ and differentiating under the integral (the boundary term $f(0)x^{-1/2}$ vanishes) gives the second form, which is continuous. Existence. For this $y$, the same interchange gives $\int_0^x\frac{y(t)}{\sqrt{x-t}}dt$ $=\frac1\pi\int_0^xf'(s)\int_s^x\frac{dt}{\sqrt{(x-t)(t-s)}}\,ds$ $=\int_0^xf'(s)\,ds$ $=f(x)-f(0)=f(x)$. ∎
For example $f(x)=x$ gives $y=\frac1\pi\int_0^x\frac{dt}{\sqrt{x-t}}$ $=\frac{2\sqrt x}{\pi}$; check: $\int_0^x\frac{2\sqrt t}{\pi\sqrt{x-t}}\,dt$ $=\frac{2x}{\pi}B\!\left(\tfrac32,\tfrac12\right)$ $=\frac{2x}{\pi}\cdot\frac\pi2=x$. For the tautochrone, $T$ is constant, so $f$ is a nonzero constant $C$ and $f(0)\ne0$; the first formula still applies and gives $\phi(y)=\frac{C}{\pi\sqrt y}$, an integrable but unbounded solution — and the curve with $\frac{ds}{dy}\propto y^{-1/2}$ is a cycloid.
- First-kind Volterra equations need compatible data. $\int_0^xy(t)\,dt=1+x$ has no continuous solution ($x=0$ gives $0=1$); $\int_0^x(x-t)y(t)\,dt=\sin x$ has none either, because differentiating once gives $\int_0^xy=\cos x$, false at $x=0$. Theorem 7.1 needs $f(a)=0$ and $K(x,x)\ne0$.
- First-kind Fredholm equations are worse. $\int_0^1y(t)\,dt=x$ has no solution (the left side is constant); $\int_0^1y(t)\,dt=0$ has infinitely many; and $\int_0^1(x+t)y(t)\,dt=f(x)$ is solvable only when $f$ is linear (Exercise III.8). Second-kind equations have no such range restriction.
- Volterra second kind is always uniquely solvable (continuous kernel) — but this fails for unbounded kernels: with $K(x,t)=\frac1x$ on $0\lt t\le x\le1$, every $\lambda\ge1$ has the eigenfunction $x^{\lambda-1}$ (Exercise IV.7).
- A nonzero kernel may have no characteristic numbers. $\sin x\cos t$ on $[0,\pi]$: $A=\int_0^\pi\cos t\sin t\,dt=0$, so $D(\lambda)\equiv1$. Every continuous Volterra kernel is another example.
- Characteristic numbers may all be non-real. $\sin(x-t)$ on $[0,2\pi]$ gives $D=1+\pi^2\lambda^2$; $x-t$ on $[0,1]$ gives $D=1+\frac{\lambda^2}{12}$, roots $\pm2\sqrt3\,i$. Only symmetric kernels are guaranteed real ones (§7.2).
- Rank $n$ does not mean $n$ characteristic numbers. $D$ has degree $\le n$; the degree drops when $\det A=0$ (e.g. $1+\sin\pi x\cos\pi t$ on $[0,1]$ has rank $2$ but $D=1-\lambda$), and a root may be repeated ($\cos(x-t)$ on $[0,2\pi]$ has the double characteristic number $\frac1\pi$ with the two eigenfunctions $\cos x$, $\sin x$).
- The solvability condition uses the transposed kernel. For $K=t$ on $[0,1]$ (characteristic number $2$, eigenfunction $1$), the transposed eigenfunction is $x$, and $y=f+2\int_0^1t\,y(t)\,dt$ is solvable iff $\int_0^1xf\,dx=0$. With $f=3x-2$ we have $\int_0^1f\cdot1\,dx=-\tfrac12\ne0$ but $\int_0^1xf\,dx=0$: solvable.
- Index order in $A$. $A_{ij}=\int b_i\,a_j$, not $\int a_ib_j$; using the wrong one gives $A^{T}$, which has the same determinant (so the same characteristic numbers) but wrong eigenfunctions and wrong solutions.
- Green's function must exist. $y''+\lambda y=0$, $y'(0)=y'(1)=0$ cannot be converted with the Green's function of $u''$ — none exists, since constants solve the homogeneous Neumann problem. Write $y''-y=-(\lambda+1)y$ instead (Exercise II.5).
- Do not drop the initial data. For $y''=g-qy$ with $y(a)=\alpha$, $y'(a)=\beta$ the free term is $\alpha+\beta(x-a)+\int(x-t)g$; forgetting $\beta(x-a)$, or writing $(t-x)$ for $(x-t)$, are the commonest errors in Part B options.
- Laplace transforms need convolution form. They apply to $\int_0^xk(x-t)y(t)\,dt$, not to Fredholm integrals $\int_0^1$, nor to kernels like $k(x+t)$ or $xt$.
- Nonlinear is different. $y=1+\int_0^xy^2$ blows up at $x=1$; Theorem 7.7 is a theorem about linear equations.
Exercises
- 1 Classify: (a) $y(x)=e^x+\int_0^x\sin(x-t)\,y(t)\,dt$; (b) $\int_0^1e^{xt}y(t)\,dt=g(x)$; (c) $y(x)=\lambda\int_{-1}^1(1+xt)\,y(t)\,dt$; (d) $y(x)=x+\lambda\int_0^1\sqrt{y(t)}\,dt$.
- 2 Show that $y=\cos x$ solves $y(x)=1-\int_0^x(x-t)\,y(t)\,dt$.
- 3 Convert $\int_0^x(2+x-t)\,y(t)\,dt=x$ into a second-kind equation and solve it.
- 4 Convert $y''+3y'+2y=0$, $y(0)=1$, $y'(0)=0$, into a Volterra equation for $y$, and verify it with the exact solution.
- 5 Show that $y'''-2y=0$, $y(0)=1$, $y'(0)=y''(0)=0$, is equivalent to $y(x)=1+\int_0^x(x-t)^2y(t)\,dt$.
- 6 Find the initial value problem equivalent to $y(x)=1+\int_0^x(x-2t)\,y(t)\,dt$, and solve it.
- 7 Solve $\int_0^x\cos(x-t)\,y(t)\,dt$ $=\sin x$ (a) by Theorem 7.1, (b) by the Laplace transform.
- 8 Show that neither $\int_0^xy(t)\,dt=1+x$ nor $\int_0^x(x-t)\,y(t)\,dt=\sin x$ has a continuous solution.
- 1 Show that $y''+\lambda y=0$, $y(0)=y(1)=0$, is equivalent to $y=\lambda\int_0^1G_0(x,t)\,y(t)\,dt$.
- 2 Convert $y''+\lambda y=1$, $y(0)=y(1)=0$, into a Fredholm equation.
- 3 Convert $y''+\lambda y=0$, $y(0)=0$, $y(1)=1$ (inhomogeneous boundary condition) into a Fredholm equation.
- 4 Show that $y''+\lambda y=0$, $y'(0)=0$, $y(1)=0$, is equivalent to $y=\lambda\int_0^1\big(1-\max(x,t)\big)\,y(t)\,dt$.
- 5 Convert $y''+\lambda y=0$, $y'(0)=y'(1)=0$, into a Fredholm equation by writing it as $y''-y=-(\lambda+1)y$.
- 6 Find the boundary value problem equivalent to $y(x)=x+\lambda\int_0^1\big(1-\max(x,t)\big)\,y(t)\,dt$.
- 7 The IVP $y''+\lambda y=0$, $y(0)=0$, $y'(0)=1$ has a solution for every $\lambda$, but the BVP $y''+\lambda y=1$, $y(0)=y(1)=0$ does not. Explain this in terms of the two integral equations.
- 8 Whirling string. A string of length $L$, tension $T$ and linear density $\rho(x)\gt0$ rotates with angular speed $\omega$ about its equilibrium axis; a small steady deflection $y$ satisfies $Ty''+\omega^2\rho(x)y=0$, $y(0)=y(L)=0$. Write this as a Fredholm equation, show how to make its kernel symmetric, and find the critical speeds when $\rho\equiv\rho_0$.
- 1 Solve $y(x)=x+\lambda\int_0^1xt\,y(t)\,dt$. For which $\lambda$ is there no solution?
- 2 Solve $y(x)=x^2+\int_0^1(x+t)\,y(t)\,dt$.
- 3 Find the characteristic numbers and eigenfunctions of $\cos(x-t)$ on $[0,2\pi]$.
- 4 Find the characteristic numbers and eigenfunctions of $1+xt$ on $[-1,1]$.
- 5 For which $a$ is $y(x)=x+a+2\int_0^1t\,y(t)\,dt$ solvable? Find all solutions for that $a$.
- 6 Find the characteristic numbers and eigenfunctions of $\sin(x+t)$ on $[0,\pi]$, and solve $y(x)=1+\lambda\int_0^\pi\sin(x+t)\,y(t)\,dt$.
- 7 Prove that a separable kernel of rank $n$ has at most $n$ characteristic numbers. Give a kernel of rank $2$ with no real characteristic number.
- 8 Show that $\int_0^1(x+t)\,y(t)\,dt=f(x)$ has a continuous solution if and only if $f$ is a polynomial of degree $\le1$, and find all solutions for $f(x)=x$.
- 1 Solve $y(x)=1+\int_0^x(x-t)\,y(t)\,dt$ by successive approximations.
- 2 Solve $y(x)=1+2\int_0^xt\,y(t)\,dt$ by successive approximations, and check by converting to an IVP.
- 3 Solve $y(x)=\sin x+2\int_0^x\cos(x-t)\,y(t)\,dt$ by the Laplace transform.
- 4 Solve $y(x)=x+\int_0^x\sin(x-t)\,y(t)\,dt$.
- 5 Solve Abel's equation $\int_0^x\frac{y(t)}{\sqrt{x-t}}\,dt=x^2$.
- 6 Prove, using Gronwall's inequality (§1.1), that the homogeneous equation $y(x)=\lambda\int_a^xK(x,t)y(t)\,dt$ with continuous $K$ has only the trivial solution.
- 7 Show that for the unbounded kernel $K(x,t)=\frac1x$ ($0\lt t\le x\le1$), every $\lambda\ge1$ admits a nonzero continuous solution of $y(x)=\lambda\int_0^x\frac{y(t)}{x}\,dt$. Which hypothesis of Theorem 7.7 fails?
- 8 Renewal equation. In a simple population model the birth rate $B(t)$ satisfies $B(t)=e^{-t}+\int_0^t\beta e^{-(t-s)}B(s)\,ds$ ($t$ in generations, $\beta\gt0$ the reproduction rate). Solve it and decide when the population grows.
E1. CSIR Part B GATE MA The initial value problem $y''-xy=0$, $y(0)=1$, $y'(0)=2$, is equivalent to: (A) $y(x)=1+2x+\int_0^x(x-t)\,t\,y(t)\,dt$; (B) $y(x)=1+2x-\int_0^x(x-t)\,t\,y(t)\,dt$; (C) $y(x)=2+x+\int_0^x(x-t)\,t\,y(t)\,dt$; (D) $y(x)=1+2x+\int_0^x(x-t)\,x\,y(t)\,dt$.
E2. CSIR Part B The characteristic numbers of the kernel $K(x,t)=1+3xt$ on $[0,1]$ are: (A) $4\pm2\sqrt3$; (B) $2\pm\sqrt3$; (C) $\tfrac{2\pm\sqrt3}{2}$; (D) $-4\pm2\sqrt3$.
E3. CSIR Part B The continuous solution of $y(x)=e^x-\int_0^xe^{x-t}\,y(t)\,dt$ is: (A) $1$; (B) $e^x$; (C) $e^{2x}$; (D) $\cosh x$.
E4. CSIR Part C Consider $y(x)=f(x)+\lambda\int_0^1x\,y(t)\,dt$ on $[0,1]$. Which of the following statements are true? (One or more options may be correct.) (A) $\lambda=2$ is the only characteristic number. (B) For $\lambda=2$ and $f(x)=2x-1$ the equation has infinitely many solutions. (C) For $\lambda=2$ and $f(x)=3x-2$ the equation has a solution. (D) For $\lambda=1$ and $f\equiv1$ the solution is $y=1+2x$.
E5. CSIR Part C Let $K$ be continuous on $[0,1]^2$, not identically zero. Which of the following are necessarily true? (A) For every $\lambda$ and every $f\in C[0,1]$, $y(x)=f(x)+\lambda\int_0^xK(x,t)y(t)\,dt$ has exactly one continuous solution. (B) If $|\lambda|\max|K|\lt1$, then $y(x)=f(x)+\lambda\int_0^1K(x,t)y(t)\,dt$ has exactly one continuous solution for every $f$. (C) $\int_0^xK(x,t)y(t)\,dt=f(x)$ has a continuous solution for every $f\in C^1[0,1]$ with $f(0)=0$. (D) $K$ has at least one characteristic number.
E6. CSIR Part C GATE MA Consider the boundary value problem $y''+\lambda y=0$, $y(0)=0$, $y'(1)=0$. Which of the following are true? (A) It is equivalent to $y(x)=\lambda\int_0^1\min(x,t)\,y(t)\,dt$. (B) Its characteristic numbers are $n^2\pi^2$, $n\in\mathbb{N}$. (C) $y''+\lambda y=x$, $y(0)=0$, $y'(1)=0$ is equivalent to $y(x)=\frac{x^3}{6}-\frac{x}{2}+\lambda\int_0^1\min(x,t)\,y(t)\,dt$. (D) For $\lambda=\frac{\pi^2}{4}$ the equation $y(x)=x+\lambda\int_0^1\min(x,t)\,y(t)\,dt$ has a solution.
Suggested plan — 3 lectures × 60 min.
- Classification and conversion. Open with Abel's bead problem and the population-with-memory story; Definition 7.1 and Example 1 (ask the class to classify before you do; part (e) catches those who forget linearity). Theorem 7.2 on the board, then Theorem 7.3 with Example 2 — let students derive both forms and compare. Finish with Theorem 7.1 and Exercise I.8.
- Boundary value problems and separable kernels. Recall the Green's function from §2.4; Theorem 7.4, Examples 3–4 and the strategy box. Then Definition 7.2, Theorem 7.5 (the proof is five lines — do it fully) and Example 5. Project Interactive 1 with the kernel $x+t$ and let the class predict where $y$ blows up.
- The alternative and Volterra theory. Theorem 7.6 with its linear-algebra proof; Example 6 and Interactive 1 with $\cos(x+t)$ (solvable at $\frac1\pi$, not at $-\frac1\pi$) and with $1-2x+3t$ (non-symmetric). Theorem 7.7 with full proof, Example 7, Interactive 2; Laplace transforms and Abel's equation as a short coda. Do E4 and E5 live.
Discussion questions. (1) Why does an IVP lead to a Volterra equation and a BVP to a Fredholm equation — what is it about "information at one end" versus "information at both ends"? (2) The Fredholm alternative says "uniqueness implies existence". Where else have you met this principle (square matrices, Green's functions)? (3) Why are first-kind equations ill-posed? What happens to $\int_0^1\sin(n\pi t)\,y(t)\,dt$-type data as $n\to\infty$?
Misconceptions to watch for. Forgetting the $\beta(x-a)$ term or writing $(t-x)$ for $(x-t)$; testing solvability against the eigenfunctions of $K$ rather than of $K^{T}$; believing every kernel has characteristic numbers, or that they must be real; using $A_{ij}=\int a_ib_j$; assuming a first-kind equation is solvable whenever the second-kind one is; applying the Laplace transform to a Fredholm integral.
Suggested homework. I(3, 4, 6, 8), II(2, 3, 4, 8), III(1, 2, 5, 6, 8), IV(1, 3, 5, 7, 8); Exam corner E2, E4, E6.
Research question. A flame, a plasma column or a gas cloud around a star is often radially symmetric, but a camera only sees its brightness integrated along lines of sight. Can the radial emission profile be recovered from the side-on image, and how badly do measurement errors spoil the answer?
Why it matters. This is Abel inversion, used daily in combustion research, laser-plasma diagnostics in fusion laboratories, and astrophysics. Its mathematics is Abel's equation — an integral equation of the first kind — so it inherits the ill-posedness of first-kind problems: the inversion formula contains a derivative of measured data.
What to do.
- For a radial profile $\varepsilon(r)$ on $0\le r\le1$ the side-on projection is $P(y)=2\int_y^1\frac{\varepsilon(r)\,r}{\sqrt{r^2-y^2}}\,dr$. Show that the substitutions $x=1-y^2$, $t=1-r^2$ turn this into an equation of Abel type, and verify that $\varepsilon(r)=1-r^2$ gives $P(y)=\tfrac43(1-y^2)^{3/2}$.
- Implement the inversion formula $\varepsilon(r)$ $=-\frac1\pi\int_r^1\frac{P'(y)}{\sqrt{y^2-r^2}}\,dy$ numerically (finite differences for $P'$, a quadrature that copes with the square-root singularity), and recover $\varepsilon$ from exact data on a grid of $100$ points.
- Add random noise of relative size $1\%$ to $P$ and invert again. Measure the error of $\varepsilon$ near $r=0$ and near $r=1$.
- Smooth the noisy data first (for instance by a least-squares fit of a low-degree polynomial in $y^2$) and repeat. Explain the trade-off you see between fitting the data and stabilising the inversion.
Tools: Python with NumPy (or any language with arrays); a plotting library.
Hand in: Plots of the true and recovered profiles with and without noise and smoothing, and a paragraph explaining why the noise is amplified.
Going further. Read about Tikhonov regularisation for first-kind equations, and about the Herglotz–Wiechert inversion that recovers the seismic velocity inside the Earth from travel times — another Abel equation.
🧪 Self-check quiz
Q1. The equation $\int_0^1K(x,t)\,y(t)\,dt=f(x)$ is
Q2. How many characteristic numbers does a continuous Volterra kernel $K(x,t)$ have?
Q3. The IVP $y''+y=0$, $y(0)=0$, $y'(0)=1$ is equivalent to
Q4. For a separable kernel $K=\sum_{i=1}^na_i(x)b_i(t)$, solving $y=f+\lambda\int Ky$ reduces to
Q5. If $\lambda$ is a characteristic number of a separable kernel $K$, the equation $y=f+\lambda\int_a^bKy$ is solvable if and only if