Chapter 2 · Linear Ordinary Differential Equations

2.4 Green’s Functions

Green's functions for second-order boundary value problems: construction from two one-sided solutions, the jump condition, symmetry, the solution formula $u=\int G f$ and when it exists, the Fredholm alternative and the modified Green's function, eigenfunction expansions of $G$, and Green's functions for initial value problems and higher-order equations.

Hang a weight on a taut string at the point $t$: the string bends into a tent, straight on either side of the load, with a kink under it. Load the string everywhere with force density $f(x)$ and, because the deflection depends linearly on the load, the result is the superposition of tents — one for each element $f(t)\,dt$ of the load: $$u(x)=\int_0^1G(x,t)\,f(t)\,dt,$$ where $G(x,t)$ is the deflection at $x$ caused by a unit load at $t$. This kernel is the Green's function of the boundary value problem, named after George Green, a self-taught miller from Nottingham who introduced it in 1828 for electrostatics. It is the continuous analogue of an inverse matrix: if the boundary value problem is "$Au=f$", then $G$ is the matrix of $A^{-1}$, with $x$ the row index and $t$ the column index.

Green's functions turn a differential equation with boundary conditions into an explicit integral formula, valid for every right-hand side at once. They exist exactly when the homogeneous problem has only the trivial solution (§2.3, Theorem 2.21); they are symmetric when the problem is self-adjoint; and their expansions in eigenfunctions connect them with Sturm–Liouville theory and with the integral equations of Chapter 7. In CSIR NET you are asked to construct $G$ for a given problem, to evaluate it at a point, to recognise which of several kernels is the Green's function (continuity, jump, boundary conditions, symmetry), to solve a problem through $G$, and to decide when no Green's function exists and what replaces it.

A First Example: the Loaded String

Consider $-u''=f$ on $[0,1]$ with $u(0)=u(1)=0$, $f$ continuous. Integrating twice from $0$, $$u'(x)=u'(0)-\int_0^xf(t)\,dt,\qquad u(x)=u'(0)\,x-\int_0^x(x-t)f(t)\,dt$$ (the double integral $\int_0^x\int_0^sf$ equals $\int_0^x(x-t)f(t)\,dt$ by interchanging the order of integration). The condition $u(1)=0$ gives $u'(0)=\int_0^1(1-t)f(t)\,dt$, so $$u(x)=\int_0^1x(1-t)f(t)\,dt-\int_0^x(x-t)f(t)\,dt=\int_0^xt\,(1-x)\,f(t)\,dt+\int_x^1x\,(1-t)\,f(t)\,dt,$$ since $x(1-t)-(x-t)=t(1-x)$. That is, $u(x)=\int_0^1G(x,t)f(t)\,dt$ with $$G(x,t)=\begin{cases}x\,(1-t),&0\le x\le t\le1,\\ t\,(1-x),&0\le t\le x\le1,\end{cases}\qquad\text{i.e.}\qquad G(x,t)=\min(x,t)\,\big(1-\max(x,t)\big).$$ Read off its properties, all of which will become the definition below:

  1. $G$ is continuous on the square $[0,1]^2$;
  2. for fixed $t$, $x\mapsto G(x,t)$ is linear on $[0,t]$ and on $[t,1]$, so it solves $-u''=0$ for $x\ne t$, and it satisfies both boundary conditions;
  3. $\partial_xG$ jumps at $x=t$: it equals $1-t$ just left of $t$ and $-t$ just right of it, a jump of $-1$ — the unit point force (formally, $-\partial_x^2G=\delta(x-t)$);
  4. $G(x,t)=G(t,x)$ (the deflection at $x$ due to a load at $t$ equals the deflection at $t$ due to a load at $x$ — Maxwell's reciprocity);
  5. $G\ge0$, with maximum $G(t,t)=t(1-t)\le\tfrac14$; hence $f\ge0$ implies $u\ge0$, and since $\int_0^1G(x,t)\,dt=\tfrac{x(1-x)}{2}\le\tfrac18$, every solution obeys $\max|u|\le\tfrac18\max|f|$.

Green's Functions for Second-Order Boundary Value Problems

We work with the self-adjoint operator $L[u]=(p\,u')'+q\,u$ of §2.3 on $[a,b]$, where $p\in C^1[a,b]$, $p\gt0$, $q\in C[a,b]$, and with separated homogeneous boundary conditions $$B_a[u]:=\alpha_1u(a)+\alpha_2u'(a)=0,\qquad B_b[u]:=\beta_1u(b)+\beta_2u'(b)=0.$$ (Any equation $a_2u''+a_1u'+a_0u=g$ with $a_2\gt0$ is brought to this form by Theorem 2.22; the right-hand side becomes $f=\frac{p}{a_2}\,g$.)

📘 Definition 2.14 — Green's function of a two-point boundary value problem
A function $G:[a,b]\times[a,b]\to\mathbb{R}$ is a Green's function of the problem $L[u]=f$, $B_a[u]=B_b[u]=0$ if, for every $t\in(a,b)$:
  1. $x\mapsto G(x,t)$ is continuous on $[a,b]$, twice continuously differentiable on $[a,t)$ and on $(t,b]$, and satisfies $L_x[G(\cdot,t)]=0$ there;
  2. $B_a[G(\cdot,t)]=0$ and $B_b[G(\cdot,t)]=0$;
  3. (jump condition) $\ \partial_xG(t^+,t)-\partial_xG(t^-,t)=\dfrac{1}{p(t)}$.
Formally $L_xG(x,t)=\delta(x-t)$: $G(\cdot,t)$ is the response to a unit point source at $t$.
📐 Theorem 2.30 — Construction, uniqueness and symmetry of $G$; the solution formula
Suppose the homogeneous problem $L[u]=0$, $B_a[u]=B_b[u]=0$ has only the trivial solution. Let $u_1\not\equiv0$ solve $L[u_1]=0$ with $B_a[u_1]=0$, and $u_2\not\equiv0$ solve $L[u_2]=0$ with $B_b[u_2]=0$. Then:
  1. $u_1,u_2$ are linearly independent, and $C:=p(x)\,W(u_1,u_2)(x)=p\,(u_1u_2'-u_1'u_2)$ is a nonzero constant;
  2. the unique Green's function is $$G(x,t)=\begin{cases}\dfrac{u_1(x)\,u_2(t)}{C},&a\le x\le t\le b,\\[2mm]\dfrac{u_1(t)\,u_2(x)}{C},&a\le t\le x\le b;\end{cases}$$
  3. $G(x,t)=G(t,x)$ for all $x,t$;
  4. for every $f\in C[a,b]$, $$u(x)=\int_a^bG(x,t)\,f(t)\,dt=\frac{1}{C}\left[u_2(x)\int_a^xu_1(t)f(t)\,dt+u_1(x)\int_x^bu_2(t)f(t)\,dt\right]$$ is the unique solution of $L[u]=f$, $B_a[u]=B_b[u]=0$.

Proof. Existence of $u_1,u_2$. The solution of $L[u]=0$ with $u(a)=\alpha_2$, $u'(a)=-\alpha_1$ is nontrivial (its data are not both $0$) and $B_a[u]=\alpha_1\alpha_2-\alpha_2\alpha_1=0$; similarly at $b$.

(1) By Lagrange's identity (§2.3, Theorem 2.23), $\big(p\,W(u_1,u_2)\big)'=u_1L[u_2]-u_2L[u_1]=0$, so $C$ is constant. If $u_1,u_2$ were dependent, $u_2=cu_1$, then $u_1$ would satisfy both boundary conditions — a nontrivial solution of the homogeneous problem, which is excluded. Hence $W\ne0$ (Theorem 2.4) and $C\ne0$.

(2) For fixed $t$, $G(\cdot,t)$ is a constant multiple of $u_1$ on $[a,t]$ and of $u_2$ on $[t,b]$: it solves $L=0$ there, satisfies $B_a$ (because $u_1$ does) and $B_b$ (because $u_2$ does), and the two formulas agree at $x=t$. The jump is $$\partial_xG(t^+,t)-\partial_xG(t^-,t)=\frac{u_1(t)u_2'(t)-u_1'(t)u_2(t)}{C}=\frac{W(t)}{p(t)W(t)}=\frac1{p(t)}.$$ Uniqueness. If $\tilde G$ is another Green's function, $D=G(\cdot,t)-\tilde G(\cdot,t)$ is continuous with a continuous derivative at $x=t$ (the jumps cancel), and solves $L[D]=0$ on $[a,t)$ and on $(t,b]$. The solution of $L=0$ on $[t,b]$ with the data $D(t),D'(t)$ coincides, by uniqueness for the initial value problem at $t$, with the continuation of the left piece, so $D$ solves $L[D]=0$ on all of $[a,b]$ and satisfies both boundary conditions: $D\equiv0$.

(3) Swapping $x$ and $t$ exchanges the two lines of the formula.

(4) Write $u=\frac1C\big[u_2(x)A(x)+u_1(x)B(x)\big]$ with $A=\int_a^xu_1f$, $B=\int_x^bu_2f$, so $A'=u_1f$, $B'=-u_2f$. Then $$u'=\tfrac1C\big[u_2'A+u_1'B\big]+\tfrac1C\big[u_2u_1f-u_1u_2f\big]=\tfrac1C\big[u_2'A+u_1'B\big],$$ $$(pu')'=\tfrac1C\big[(pu_2')'A+(pu_1')'B\big]+\tfrac{p}{C}\big[u_2'u_1-u_1'u_2\big]f=-q\,u+\frac{pW}{C}\,f=-q\,u+f,$$ using $(pu_i')'=-qu_i$. So $L[u]=f$. At $x=a$: $A(a)=0$, so $u=\frac{B(a)}{C}u_1$ and $u'=\frac{B(a)}{C}u_1'$ there, giving $B_a[u]=\frac{B(a)}{C}B_a[u_1]=0$; similarly $B_b[u]=\frac{A(b)}{C}B_b[u_2]=0$. Uniqueness: the difference of two solutions solves the homogeneous problem. ∎

Notice how the construction mirrors variation of parameters (§2.2, Theorem 2.16): the same two integrals appear, but the lower limits are now chosen so that the boundary conditions hold automatically — $\int_a^x$ goes with the solution adapted to the right end, $\int_x^b$ with the one adapted to the left end.

🧭 Problem-solving strategy — constructing and using a Green's function
  1. Write the operator in self-adjoint form $(pu')'+qu=f$ (Theorem 2.22), and remember that $f$ is the new right-hand side. For $-u''=f$ use $L[u]=u''$ with right side $-f$ (or keep $-u''$ and take jump $-1$).
  2. Check that $G$ exists: the homogeneous problem must have only $u\equiv0$ (otherwise see Theorem 2.31).
  3. Find $u_1$ satisfying the left condition and $u_2$ satisfying the right condition; compute $C=p\,(u_1u_2'-u_1'u_2)$ — it must come out constant (a free check).
  4. Write $G$ = $u_1(\text{smaller})\,u_2(\text{larger})/C$. Verify continuity, the conditions, the jump $1/p(t)$, and symmetry.
  5. Solve by splitting $\int_a^b=\int_a^x+\int_x^b$. For non-homogeneous boundary values, add the solution of $L[w]=0$ with those boundary values.
✏️ Example 1 — A string with a free end
Find the Green's function of $-u''=f$, $u(0)=0$, $u'(1)=0$, and use it to solve the problem with $f\equiv1$. Interpret physically.
Existence $-u''=0$ gives $u=A+Bx$; $u(0)=0$ and $u'(1)=0$ force $A=B=0$. So $G$ exists.
Build Take $L[u]=u''$ ($p=1$) with right side $-f$. $u_1=x$ satisfies $u_1(0)=0$; $u_2=1$ satisfies $u_2'(1)=0$; $C=x\cdot0-1\cdot1=-1$. So $G_L(x,t)=-\min(x,t)$, and the solution of $-u''=f$ is $u=\int G_L\cdot(-f)$: the Green's function of $-u''$ is $$G(x,t)=\min(x,t).$$ Check: continuous; $G(0,t)=0$; for $x\gt t$, $G=t$ is constant, so $\partial_xG(1,t)=0$; slope $1$ left of $t$ and $0$ right of it — jump $-1$, as for $-u''$; symmetric.
Solve $$u(x)=\int_0^x t\,dt+\int_x^1x\,dt=\frac{x^2}{2}+x(1-x)=x-\frac{x^2}{2}.$$ Check: $-u''=1$, $u(0)=0$, $u'(1)=1-1=0$. ✓
Meaning A unit load at $t$ lifts every point beyond $t$ by the same amount $t$: the part of the string past the load is free and stays horizontal. The same problem is the steady temperature in a rod held at $0$ at $x=0$ and insulated at $x=1$, with heat source $f$ (Exercise IV.7).
✏️ Example 2 — Green's function of $u''+u=f$ on $[0,\tfrac{\pi}{2}]$
Construct the Green's function of $u''+u=f$, $u(0)=0$, $u(\tfrac{\pi}{2})=0$, and solve the problem for $f(x)=x$. Evaluate $u(\tfrac{\pi}{4})$.
Existence $A\cos x+B\sin x$ with $A=0$ and $B\sin\tfrac{\pi}{2}=B=0$: only the trivial solution (equivalently, $1$ is not a Dirichlet eigenvalue on $[0,\tfrac\pi2]$, whose eigenvalues are $4n^2$).
Build $u_1=\sin x$, $u_2=\cos x$, $p=1$, $C=\sin x(-\sin x)-\cos x\cos x=-1$: $$G(x,t)=\begin{cases}-\sin x\cos t,&x\le t,\\-\sin t\cos x,&x\ge t.\end{cases}$$
Solve $u(x)=-\cos x\int_0^xt\sin t\,dt-\sin x\int_x^{\pi/2}t\cos t\,dt$. Since $\int_0^xt\sin t\,dt=\sin x-x\cos x$ and $\int_x^{\pi/2}t\cos t\,dt=\big[t\sin t+\cos t\big]_x^{\pi/2}=\tfrac{\pi}{2}-x\sin x-\cos x$, $$u=-\sin x\cos x+x\cos^2x-\tfrac{\pi}{2}\sin x+x\sin^2x+\sin x\cos x=x-\tfrac{\pi}{2}\sin x.$$
Check $u''+u=\tfrac{\pi}{2}\sin x+x-\tfrac{\pi}{2}\sin x=x$, $u(0)=0$, $u(\tfrac{\pi}{2})=\tfrac{\pi}{2}-\tfrac{\pi}{2}=0$. And $u(\tfrac{\pi}{4})=\tfrac{\pi}{4}-\tfrac{\pi}{2}\cdot\tfrac{\sqrt2}{2}=\dfrac{\pi(1-\sqrt2)}{4}\approx-0.3253$. Note $G\le0$ here, so a positive $f$ gives a negative $u$.
✏️ Example 3 — An Euler equation: divide before you integrate
Solve $x^2u''+xu'-u=x^3$ on $[1,2]$, $u(1)=u(2)=0$, by a Green's function, and show what goes wrong if the right-hand side is not adjusted.
Self-adjoint form $a_2=x^2$, $a_1=x$: $p=e^{\int dx/x}=x$, multiplier $p/a_2=1/x$. The problem becomes $$(xu')'-\frac{u}{x}=x^2,\qquad u(1)=u(2)=0,$$ with right-hand side $x^3/x=x^2$.
Build Homogeneous solutions $x$, $x^{-1}$ (indicial roots $\pm1$). $u_1=x-\tfrac1x$ ($u_1(1)=0$), $u_2=x-\tfrac4x$ ($u_2(2)=0$). Then $W=\left(x-\tfrac1x\right)\left(1+\tfrac4{x^2}\right)-\left(1+\tfrac1{x^2}\right)\left(x-\tfrac4x\right)=\tfrac6x$, so $C=pW=6$ — constant, as it must be: $$G(x,t)=\frac{(x^2-1)(t^2-4)}{6\,x\,t}\quad(x\le t),\qquad G(x,t)=G(t,x).$$
Solve $u=\int_1^2G(x,t)\,t^2\,dt$ gives $$u(x)=\frac{x^3}{8}-\frac{5x}{8}+\frac{1}{2x}.$$ Check by the direct method: $\tfrac{x^3}{8}$ is a particular solution ($x^2\cdot\tfrac{6x}{8}+x\cdot\tfrac{3x^2}{8}-\tfrac{x^3}{8}=x^3$), and $c_1x+c_2/x$ with $u(1)=u(2)=0$ gives $c_1=-\tfrac58$, $c_2=\tfrac12$. ✓
The trap Integrating $G$ against the original right side $x^3$ gives $\dfrac{3x^5-31x^2+28}{45x}$, which satisfies the boundary conditions but not the equation. The Green's function of the self-adjoint form acts on the right-hand side of the self-adjoint form.
Interactive 1 — The Green's Function Is the Response to a Point Load

The solid curve is $x\mapsto G(x,t)$, the response to a unit point source at $x=t$ (arrow). The short segments show the one-sided slopes at $t$: their difference is the jump $1/p(t)$ ($-1$ for $-u''$, $+1$ for $u''+\dots$). The dashed curve is $x\mapsto G(t,x)$ — the arguments swapped; it lies on top of the solid curve exactly when $G$ is symmetric. Press Play to move the load (in the $u''+k^2u$ mode, Play sweeps $k$ instead: watch $G$ blow up as $k$ passes the eigenvalue $\pi$ and change sign).

0.351.50
G(t, t)—
slopes ∂ₓG(t−), ∂ₓG(t+)—
jump—
max |G(x, t) − G(t, x)|—

When the Homogeneous Problem Has Nontrivial Solutions

If the homogeneous problem has a solution $\varphi\not\equiv0$, there is no Green's function: the construction of Theorem 2.30 breaks down because $u_1$ and $u_2$ become proportional. Theorem 2.21 says the problem $L[u]=f$ is then solvable for some $f$ and not for others. Self-adjointness tells us exactly which.

📐 Theorem 2.31 — Fredholm alternative for second-order boundary value problems
Let $L$, $B_a$, $B_b$ be as above (separated conditions), and suppose the homogeneous problem has a nontrivial solution $\varphi$. Then:
  1. every solution of the homogeneous problem is a constant multiple of $\varphi$;
  2. for $f\in C[a,b]$, the problem $L[u]=f$, $B_a[u]=B_b[u]=0$ has a solution if and only if $$\int_a^bf(x)\,\varphi(x)\,dx=0;$$
  3. in that case the solutions are exactly $u^*+c\,\varphi$ ($c$ arbitrary), where $u^*$ is any one of them.
So exactly one of the following holds: either the problem is uniquely solvable for every $f$ (and has a Green's function), or the homogeneous problem has a nontrivial solution and the problem is solvable precisely for the $f$ orthogonal to it.

Proof. (1) Two solutions of the homogeneous problem both satisfy the condition at $a$, so their Wronskian vanishes at $a$ and they are dependent (as in Theorem 2.24(3)). (2) Necessity. If $u$ solves the problem, Green's formula (Theorem 2.23(3)) gives $\int_a^b\varphi f=\int_a^b\varphi L[u]=\int_a^buL[\varphi]=0$. Sufficiency. Let $u_0$ be the solution of $L[u_0]=f$ with $u_0(a)=u_0'(a)=0$ (Theorem 2.1). The solutions of $L[u]=f$ satisfying $B_a[u]=0$ are $u_0+c\varphi$: indeed $u-u_0$ solves $L=0$ with $B_a=0$, and these solutions form the one-dimensional space spanned by $\varphi$ (the condition $\alpha_1u(a)+\alpha_2u'(a)=0$ cuts out a line in the space of initial data). Since $B_b[\varphi]=0$, the problem is solvable iff $B_b[u_0]=0$. Now apply Green's formula to $u_0$ and $\varphi$: the terms at $a$ vanish because $u_0(a)=u_0'(a)=0$, so $$\int_a^b\varphi f\,dx=\int_a^b\big(\varphi L[u_0]-u_0L[\varphi]\big)dx=p(b)\big(\varphi(b)u_0'(b)-\varphi'(b)u_0(b)\big).$$ If $\beta_2\ne0$, then $\varphi'(b)=-\frac{\beta_1}{\beta_2}\varphi(b)$ and $\varphi(b)\ne0$ (otherwise $\varphi(b)=\varphi'(b)=0$ and $\varphi\equiv0$), so the right side is $\frac{p(b)\varphi(b)}{\beta_2}B_b[u_0]$. If $\beta_2=0$, then $\varphi(b)=0\ne\varphi'(b)$ and the right side is $-\frac{p(b)\varphi'(b)}{\beta_1}B_b[u_0]$. In both cases it is a nonzero multiple of $B_b[u_0]$, so $\int\varphi f=0$ iff $B_b[u_0]=0$ iff the problem is solvable. (3) Theorem 2.21(2) together with (1). ∎

When the problem is solvable, one particular solution can still be written with a kernel, provided the null solution is "projected out" of the source:

📘 Definition 2.15 — Modified (generalised) Green's function
Suppose the homogeneous problem has the nontrivial solution $\varphi$, normalised by $\int_a^b\varphi^2\,dx=1$. A modified Green's function is a function $G_M(x,t)$ that satisfies conditions (1)–(3) of Definition 2.14 except that, for $x\ne t$, $$L_x\big[G_M(\cdot,t)\big](x)=-\varphi(x)\varphi(t)\qquad\text{(formally }L_xG_M=\delta(x-t)-\varphi(x)\varphi(t)\text{)},$$ together with the normalisation $\int_a^bG_M(x,t)\,\varphi(x)\,dx=0$.
📐 Theorem 2.32 — The modified Green's function (stated, with proof sketch)
Under the hypotheses of Definition 2.15 there is exactly one modified Green's function; it is symmetric, $G_M(x,t)=G_M(t,x)$. If $\int_a^bf\varphi=0$, then $u=\int_a^bG_M(x,t)f(t)\,dt$ is the unique solution of $L[u]=f$, $B_a[u]=B_b[u]=0$ with $\int_a^bu\varphi=0$, and all solutions are $u+c\varphi$.

Proof (sketch). The source $\delta(\cdot-t)-\varphi(\cdot)\varphi(t)$ is orthogonal to $\varphi$, so by Theorem 2.31 (applied piecewise, with the jump condition handled as in Theorem 2.30) the defining problem is solvable; the solution is determined up to adding $c(t)\varphi(x)$, and the normalisation fixes $c(t)$. Symmetry follows from Green's formula applied to $G_M(\cdot,s)$ and $G_M(\cdot,t)$ exactly as for symmetric kernels. Finally $L_x\int G_M(x,t)f(t)\,dt=f(x)-\varphi(x)\int\varphi f=f(x)$ when $\int\varphi f=0$, and $\int u\varphi=\int f(t)\int G_M(x,t)\varphi(x)\,dx\,dt=0$. ∎

✏️ Example 4 — The insulated rod: Neumann conditions
Consider $-u''=f$ on $[0,1]$ with $u'(0)=u'(1)=0$. (a) Show that there is no Green's function and find the solvability condition. (b) Solve the problem for $f(x)=x-\tfrac12$. (c) Verify that $G_M(x,t)=\tfrac{x^2+t^2}{2}-\max(x,t)+\tfrac13$ is the modified Green's function, and find the solution it produces for this $f$.
(a) $-u''=0$, $u'(0)=u'(1)=0$ is solved by every constant: $\varphi=1$ (normalised: $\int_0^11^2=1$). So no Green's function exists, and by Theorem 2.31 the problem is solvable iff $\int_0^1f(x)\,dx=0$. Physically: the steady temperature in an insulated rod exists only if the heat sources and sinks balance.
(b) $\int_0^1\left(x-\tfrac12\right)dx=0$ ✓. Integrating $u''=\tfrac12-x$: $u'=\tfrac x2-\tfrac{x^2}2+c_1$, and $u'(0)=0$ gives $c_1=0$; then $u'(1)=0$ holds automatically (that is the compatibility condition at work). So $$u=\frac{x^2}{4}-\frac{x^3}{6}+c.$$
(c) Check $G_M$ For $x\lt t$: $G_M=\tfrac{x^2+t^2}{2}-t+\tfrac13$, so $\partial_xG_M=x$ and $-\partial_x^2G_M=-1$; for $x\gt t$: $\partial_xG_M=x-1$, $-\partial_x^2G_M=-1$. So $-\partial_x^2G_M=-\varphi(x)\varphi(t)=-1$ off the diagonal (the operator here is $-u''$). The slope jumps from $t$ to $t-1$, by $-1$ ✓; $\partial_xG_M(0,t)=0$, $\partial_xG_M(1,t)=0$ ✓; $G_M$ is continuous and symmetric ✓; and $\int_0^1G_M\,dx=\tfrac16+\tfrac{t^2}{2}-\left(\tfrac{t^2}2+\tfrac12\right)+\tfrac13=0$ ✓, using $\int_0^1\max(x,t)\,dx=t^2+\tfrac{1-t^2}{2}$.
(c) Solution $\int_0^1G_M(x,t)\left(t-\tfrac12\right)dt=\dfrac{x^2}4-\dfrac{x^3}6-\dfrac1{24}$ — the member of the family in (b) with mean value $0$ (indeed $\int_0^1\left(\tfrac{x^2}4-\tfrac{x^3}6\right)dx=\tfrac1{12}-\tfrac1{24}=\tfrac1{24}$).

Green's Functions and Eigenfunctions

Green's function and the Sturm–Liouville eigenfunctions of §2.3 are two descriptions of the same operator. In matrix language: if $A=\sum\lambda_nv_nv_n^T$ is symmetric with orthonormal eigenvectors, then $A^{-1}=\sum\lambda_n^{-1}v_nv_n^T$. The analogue for $G$ is the bilinear (Mercer) expansion.

📐 Theorem 2.33 — Eigenfunction expansion of $G$; poles at the eigenvalues
Let $\lambda_1\lt\lambda_2\lt\cdots$ and $\varphi_1,\varphi_2,\dots$ be the eigenvalues and eigenfunctions of the regular SL problem $L[\varphi]+\lambda r\varphi=0$, $B_a[\varphi]=B_b[\varphi]=0$, normalised by $\int_a^br\varphi_n^2=1$. Let $\mu\in\mathbb{R}$ be different from every $\lambda_n$.
  1. The problem $L[u]+\mu\,r\,u=f$, $B_a[u]=B_b[u]=0$ has the Green's function $$G(x,t;\mu)=\sum_{n=1}^{\infty}\frac{\varphi_n(x)\,\varphi_n(t)}{\mu-\lambda_n},$$ the series converging absolutely and uniformly on $[a,b]^2$. In particular ($\mu=0$, when $0$ is not an eigenvalue) the Green's function of $L[u]=f$ is $G=-\sum_n\varphi_n(x)\varphi_n(t)/\lambda_n$.
  2. As a function of $\mu$, $G(x,t;\mu)$ has a simple pole at each eigenvalue $\lambda_n$, with residue $\varphi_n(x)\varphi_n(t)$; at $\mu=\lambda_n$ no Green's function exists (Theorem 2.31).
  3. (Trace identities.) $\displaystyle\int_a^bG(x,x;\mu)\,r(x)\,dx=\sum_n\frac{1}{\mu-\lambda_n}$ and $\displaystyle\int_a^b\!\!\int_a^bG(x,t;\mu)^2\,r(x)r(t)\,dx\,dt=\sum_n\frac{1}{(\mu-\lambda_n)^2}$.

Proof (sketch). Let $u$ solve $L[u]+\mu ru=f$ with the boundary conditions. By Green's formula, $$\int_a^b\varphi_nf=\int_a^b\varphi_n\,(L[u]+\mu ru)=\int_a^bu\,(L[\varphi_n]+\mu r\varphi_n)=(\mu-\lambda_n)\int_a^bru\varphi_n,$$ so the coefficients of $u$ are $c_n=\frac{\int\varphi_nf}{\mu-\lambda_n}$, and by the expansion theorem (§2.3, Theorem 2.28) $$u(x)=\sum_n\frac{\varphi_n(x)}{\mu-\lambda_n}\int_a^b\varphi_n(t)f(t)\,dt=\int_a^b\Big(\sum_n\frac{\varphi_n(x)\varphi_n(t)}{\mu-\lambda_n}\Big)f(t)\,dt.$$ The interchange is justified by uniform convergence: for a regular problem the normalised $\varphi_n$ are uniformly bounded and $\lambda_n\sim cn^2$ (§2.3, Theorem 2.26(4)), so the series is dominated by a constant times $\sum n^{-2}$. Comparing with $u=\int Gf$ for all $f$ identifies the kernel. (2) is read off from (1): only the $n$-th term is singular at $\mu=\lambda_n$. (3) Integrate the uniformly convergent series termwise, using $\int r\varphi_n\varphi_m=\delta_{nm}$. ∎

✏️ Example 5 — Two famous sums from one Green's function
For $-u''=f$, $u(0)=u(1)=0$, write $G(x,t)=\min(x,t)(1-\max(x,t))$ as a series in the eigenfunctions $\sqrt2\sin n\pi x$, and deduce $\sum_{n\ge1}\frac1{n^2}=\frac{\pi^2}6$ and $\sum_{n\ge1}\frac1{n^4}=\frac{\pi^4}{90}$.
Expansion In the notation of Theorem 2.33, $L=d^2/dx^2$, $r=1$, $\lambda_n=n^2\pi^2$, $\varphi_n=\sqrt2\sin n\pi x$. The Green's function of $L[u]=g$ is $-\sum\varphi_n(x)\varphi_n(t)/\lambda_n$, and $-u''=f$ means $L[u]=-f$, so $$G(x,t)=\min(x,t)\big(1-\max(x,t)\big)=\sum_{n=1}^\infty\frac{2\sin n\pi x\,\sin n\pi t}{n^2\pi^2}.$$ (Numerically, at $(x,t)=(0.3,0.7)$ both sides equal $0.09$.)
Trace Put $t=x$ and integrate over $[0,1]$: the left side gives $\int_0^1x(1-x)\,dx=\tfrac16$; the right side gives $\sum\frac{2\cdot\frac12}{n^2\pi^2}=\sum\frac{1}{n^2\pi^2}$. Hence $\sum\frac1{n^2}=\frac{\pi^2}6$.
Square $\int_0^1\!\int_0^1G^2=2\int_0^1\!\int_0^tx^2(1-t)^2\,dx\,dt=\tfrac23\int_0^1t^3(1-t)^2\,dt=\tfrac23\cdot\tfrac1{60}=\tfrac1{90}$, while the series gives $\sum\frac1{n^4\pi^4}$ (orthonormality). Hence $\sum\frac1{n^4}=\frac{\pi^4}{90}$.
Bonus At $x=t=\tfrac12$: $\tfrac14=\sum\frac{2\sin^2(n\pi/2)}{n^2\pi^2}=\frac2{\pi^2}\sum_{n\text{ odd}}\frac1{n^2}$, so $\sum_{n\text{ odd}}\frac1{n^2}=\frac{\pi^2}8$ — consistent with $\left(1-\tfrac14\right)\frac{\pi^2}{6}$.
✏️ Example 6 — Prove or disprove: positivity and the parameter $k$
For $u''+k^2u=f$, $u(0)=u(1)=0$ ($k\gt0$) decide: (a) a Green's function exists iff $k\notin\pi\mathbb{N}$; (b) for $0\lt k\lt\pi$, $f\ge0$ implies $u\le0$; (c) as $k\to\pi$, $G(x,t;k)\approx\dfrac{2\sin\pi x\sin\pi t}{k^2-\pi^2}$; (d) for every $k\notin\pi\mathbb{N}$, $f\ge0$ implies $u\le0$.
Formula $u_1=\sin kx$, $u_2=\sin k(1-x)$, $C=W=-k\sin kx\cos k(1-x)-k\cos kx\sin k(1-x)=-k\sin k$. So for $x\le t$, $$G(x,t;k)=-\frac{\sin kx\;\sin k(1-t)}{k\sin k},$$ and $G(x,t)=G(t,x)$.
(a) True The homogeneous problem has the solution $\sin kx$ iff $\sin k=0$, i.e. $k=n\pi$; otherwise Theorem 2.30 applies.
(b) True For $0\lt k\lt\pi$ and $0\lt x,t\lt1$: $kx$, $k(1-t)$ and $k$ lie in $(0,\pi)$, so all three sines are positive and $G\lt0$. Hence $u=\int Gf\le0$ when $f\ge0$.
(c) True Theorem 2.33(2) with $\mu=k^2$, $\lambda_1=\pi^2$, $\varphi_1=\sqrt2\sin\pi x$: residue $2\sin\pi x\sin\pi t$. Directly from the closed form, $\lim_{k\to\pi}(k^2-\pi^2)G=2\sin\pi x\sin\pi t$ (Exercise III.4).
(d) False For $k=\tfrac{3\pi}{2}$ and $f\equiv1$: $u=\dfrac{1-\cos kx+\sin kx}{k^2}$ (check: $u''+k^2u=1$, $u(0)=0$, $u(1)=\frac{1-0-1}{k^2}=0$), and $u(\tfrac12)=\dfrac{1+\sqrt2}{k^2}\gt0$. Past the first eigenvalue, positivity is lost: $G$ changes sign, as Interactive 1 shows.

Initial Value Problems and Higher-Order Equations

The same idea — a kernel that solves the homogeneous equation away from the diagonal and has a prescribed jump on it — works for initial value problems and for equations of any order. For an IVP the kernel is the impulse response $K(x,t)$ of §2.2 (Duhamel's principle), switched on at $x=t$.

📐 Theorem 2.34 — Green's functions for initial value problems and for $n$-th order boundary value problems
Let $L[u]=u^{(n)}+p_{n-1}(x)u^{(n-1)}+\dots+p_0(x)u$ with $p_k$ continuous on $[a,b]$.
  1. (Causal Green's function.) The solution of $L[u]=f$, $u(a)=u'(a)=\dots=u^{(n-1)}(a)=0$ is $u(x)=\int_a^bG(x,t)f(t)\,dt$ with $$G(x,t)=\begin{cases}K(x,t),&t\le x,\\0,&t\gt x,\end{cases}$$ where $K(\cdot,t)$ solves $L=0$ with $K=\partial_xK=\dots=\partial_x^{n-2}K=0$ and $\partial_x^{n-1}K=1$ at $x=t$. For constant coefficients $K(x,t)=k(x-t)$.
  2. ($n$-th order BVP.) Suppose $n$ homogeneous boundary conditions (each a linear combination of $u,\dots,u^{(n-1)}$ at $a$ or at $b$) are given and the homogeneous problem has only the trivial solution. Then there is a unique $G$ such that for each $t\in(a,b)$: $G(\cdot,t)$ solves $L=0$ on $[a,t)$ and $(t,b]$ and satisfies the boundary conditions; $G,\partial_xG,\dots,\partial_x^{n-2}G$ are continuous at $x=t$; and $\partial_x^{n-1}G$ jumps by $1$ there. Then $u=\int_a^bG(x,t)f(t)\,dt$ is the unique solution of $L[u]=f$ with these conditions. (If the leading coefficient is $a_n(x)$ instead of $1$, the jump is $1/a_n(t)$.)

Proof (sketch). (1) is the kernel form of variation of parameters (§2.2, Theorem 2.16(2)–(3)): $u(x)=\int_a^xK(x,t)f(t)\,dt$, and cutting the kernel off for $t\gt x$ writes the variable upper limit as a fixed one. (2) Write $G(x,t)=G_c(x,t)+\sum_{j=1}^nc_j(t)y_j(x)$, where $G_c$ is the causal kernel and $\{y_j\}$ a fundamental set. The sum is smooth in $x$, so continuity and the jump are inherited from $G_c$; the $n$ boundary conditions give an $n\times n$ linear system for $c_1(t),\dots,c_n(t)$ whose matrix is the same as for the homogeneous problem — nonsingular by the analogue of Theorem 2.21. Uniqueness and the solution formula follow as in Theorem 2.30. ∎

A causal Green's function is not symmetric — the future does not influence the past — and that is consistent with the theory: an initial value problem is not self-adjoint. Symmetry of $G$ is the fingerprint of a self-adjoint boundary value problem, of any order.

✏️ Example 7 — A causal kernel and a cantilever beam
(a) Find the Green's function of $u''+u=f$, $u(0)=u'(0)=0$ on $[0,b]$, and show it is not symmetric. (b) A cantilever beam of unit length (clamped at $x=0$, free at $x=1$, unit flexural rigidity) has deflection $u$ with $u''''=f$, $u(0)=u'(0)=0$, $u''(1)=u'''(1)=0$. Verify that $$G(x,t)=\begin{cases}\tfrac16x^2(3t-x),&x\le t,\\\tfrac16t^2(3x-t),&x\ge t,\end{cases}$$ is its Green's function, and find the deflection of the free end under a uniform load $f\equiv1$.
(a) $K(\cdot,t)$ solves $u''+u=0$ with $K(t,t)=0$, $\partial_xK(t,t)=1$: $K=\sin(x-t)$. So $G(x,t)=\sin(x-t)$ for $t\le x$ and $0$ for $t\gt x$, and $u(x)=\int_0^x\sin(x-t)f(t)\,dt$. Not symmetric: $G(1,0.5)=\sin0.5\approx0.479$ but $G(0.5,1)=0$.
(b) Conditions For fixed $t$, each piece is a cubic in $x$, so $\partial_x^4G=0$ off the diagonal. Left piece: $G(0,t)=0$, $\partial_xG=\tfrac16(6tx-3x^2)$ vanishes at $0$. Right piece: $\partial_x^2G=0$ and $\partial_x^3G=0$ identically (it is linear in $x$), so the free-end conditions hold.
(b) Jumps At $x=t$: values $\tfrac13t^3$ from both sides; $\partial_xG$: left $tx-\tfrac12x^2\to\tfrac12t^2$, right $\tfrac12t^2$; $\partial_x^2G$: left $t-x\to0$, right $0$; $\partial_x^3G$: left $-1$, right $0$ — a jump of $+1$, as Theorem 2.34(2) requires for $u''''$. By uniqueness this is the Green's function; it is symmetric (Maxwell–Betti reciprocity), as it must be for this self-adjoint problem.
(b) Uniform load $u(x)=\int_0^x\tfrac16t^2(3x-t)\,dt+\int_x^1\tfrac16x^2(3t-x)\,dt=\dfrac{x^2(x^2-4x+6)}{24}$, so the free end deflects by $u(1)=\tfrac{3}{24}=\tfrac18$ — the classical $\frac{wL^4}{8EI}$. A single unit load at the tip gives $G(1,1)=\tfrac13$, the classical $\frac{PL^3}{3EI}$.
Interactive 2 — Building $u=\int G f$ from Point Loads or from Eigenmodes

The thick grey curve is the exact solution $u=\int_0^1G(x,t)f(t)\,dt$. In point loads mode the load is replaced by $N$ point loads $f(t_i)/N$ at the midpoints $t_i$ (bars), each producing the tent $G(x,t_i)f(t_i)/N$ (thin curves for small $N$); their sum $u_N$ is a Riemann sum of the solution formula. In eigenmodes mode $u_N$ is the partial sum of the bilinear series of Theorem 2.33 — each mode is damped by $1/\lambda_n$, so very few are needed. Press Play to increase $N$.

3
N—
u(½) exact—
u_N(½)—
max |u_N − u|—
⚠️ Counterexamples and pitfalls
  1. A Green's function need not exist. $-u''=f$ with $u'(0)=u'(1)=0$ (constants solve the homogeneous problem), $u''+\pi^2u=f$ with $u(0)=u(1)=0$ ($\sin\pi x$), and $u''=f$ with periodic conditions have none. Then the problem is solvable only for $f$ orthogonal to the null solution (Theorem 2.31).
  2. Sign conventions differ by a minus sign. For $u''=f$, $u(0)=u(1)=0$ the Green's function is $-\min(x,t)(1-\max(x,t))\le0$ (jump $+1$); for $-u''=f$ it is $+\min(x,t)(1-\max(x,t))\ge0$ (jump $-1$). Read the problem.
  3. The jump is $1/p(t)$ only in self-adjoint form — and then $G$ acts on the transformed right-hand side. For $a_2u''+a_1u'+a_0u=g$ in normal form the jump is $1$ and the kernel acts on $g/a_2$. Mixing the two (Example 3) gives a function that satisfies the boundary conditions but not the equation.
  4. $G$ is symmetric only for self-adjoint problems. For $u''+u'=f$, $u(0)=u(1)=0$: $G(0.3,0.6)\approx-0.1352$ but $G(0.6,0.3)\approx-0.1001$. Its self-adjoint form $(e^xu')'=e^xf$ has a symmetric kernel $G_s$, and $G(x,t)=G_s(x,t)\,e^t$.
  5. The causal Green's function of an IVP is never symmetric: it vanishes for $t\gt x$ but not for $t\lt x$.
  6. $G$ is continuous but not $C^1$. $\partial_xG$ jumps on the diagonal; a candidate kernel with continuous $\partial_xG$ (or a discontinuous $G$) is not a Green's function of a second-order problem. Yet $u=\int Gf$ is $C^2$.
  7. Positivity of $G$ is not automatic. $G\ge0$ for $-u''+qu$ with $q\ge0$ and Dirichlet conditions (a maximum principle), but for $u''+k^2u$ the kernel changes sign once $k\gt\pi$ (Example 6(d)).
  8. $u=\int Gf$ satisfies homogeneous boundary conditions only. For $u(0)=A$, $u(1)=B$ add the solution of the homogeneous equation with those boundary values (for $-u''=f$: $A(1-x)+Bx$).
  9. The order of the arguments matters when $G$ is not symmetric: $u(x)=\int G(x,t)f(t)\,dt$, integrating over the second variable.
  10. The modified Green's function gives one solution, not the solution. It produces the solution orthogonal to the null space; all others differ by multiples of $\varphi$.

Exercises

I
Constructing Green's functions.
  • 1 $-u''=f$, $u'(0)=0$, $u(1)=0$.
  • 2 $-u''=f$, $u(0)=0$, $u(1)+u'(1)=0$.
  • 3 $u''+4u=f$, $u(0)=0$, $u(\tfrac{\pi}{4})=0$.
  • 4 $u''-u=f$, $u(0)=u(1)=0$. Is $G$ of one sign?
  • 5 $(xu')'=f$ on $[1,e]$, $u(1)=u(e)=0$.
  • 6 $x^2u''+2xu'=f$ on $[1,2]$, $u(1)=u(2)=0$.
  • 7 Prove directly from Definition 2.14 that a Green's function, if it exists, is unique, and that it can exist only if the homogeneous problem has only the trivial solution.
  • 8 Use the Green's function of $-u''=f$, $u(0)=u(1)=0$ to solve $-u''=x$, and compute $u(\tfrac12)$.
1 $u_1=1$ ($u_1'(0)=0$), $u_2=1-x$; for $L=d^2/dx^2$, $C=1\cdot(-1)-0=-1$, so $G_L=-(1-\max(x,t))$ and for $-u''$: $G(x,t)=1-\max(x,t)$.
2 $u_1=x$, $u_2=2-x$ (then $u_2(1)+u_2'(1)=1-1=0$); $C=x(-1)-1(2-x)=-2$. For $-u''$: $G(x,t)=\tfrac12\min(x,t)\,\big(2-\max(x,t)\big)$.
3 $u_1=\sin2x$, $u_2=\cos2x$ ($u_2(\tfrac\pi4)=0$), $C=-2\sin^22x-2\cos^22x=-2$: $G=-\tfrac12\sin2x\cos2t$ ($x\le t$), $-\tfrac12\sin2t\cos2x$ ($x\ge t$).
4 $u_1=\sinh x$, $u_2=\sinh(1-x)$, $C=-\sinh x\cosh(1-x)-\cosh x\sinh(1-x)=-\sinh1$: $G=-\dfrac{\sinh x\,\sinh(1-t)}{\sinh1}$ ($x\le t$). Yes: $G\lt0$ on the open square ($u''-u=f$ is $-(-u''+u)=f$, and $-u''+u$ has a positive kernel).
5 $u_1=\log x$, $u_2=1-\log x$, $p=x$: $C=x\left(\log x\cdot\left(-\tfrac1x\right)-\tfrac1x(1-\log x)\right)=-1$. $G=-\log x\,(1-\log t)$ for $x\le t$, symmetric.
6 Already self-adjoint: $(x^2u')'=f$, $p=x^2$. Homogeneous solutions $1$, $1/x$: $u_1=1-\tfrac1x$, $u_2=\tfrac1x-\tfrac12$; $W=-\tfrac1{2x^2}$, $C=-\tfrac12$. $G=-2\left(1-\tfrac1x\right)\left(\tfrac1t-\tfrac12\right)=\dfrac{(x-1)(t-2)}{xt}$ for $x\le t$.
7 Uniqueness: as in the proof of Theorem 2.30(2). Necessity: if $\varphi\not\equiv0$ solves the homogeneous problem and $G$ existed, then $u=\int Gf$ would solve $L[u]=f$ for every $f$ (the computation of Theorem 2.30(4) uses only properties (1)–(3)); taking $f=\varphi$, Green's formula would give $\int\varphi^2=\int\varphi L[u]=\int uL[\varphi]=0$, a contradiction.
8 $u=\int_0^xt(1-x)t\,dt+\int_x^1x(1-t)t\,dt=\dfrac{x(1-x^2)}{6}$; check $-u''=x$. $u(\tfrac12)=\tfrac{1}{16}$.
II
Solving with Green's functions; the alternative.
  • 1 Solve $u''+u=1$, $u(0)=u(\tfrac{\pi}{2})=0$ with the Green's function of Example 2.
  • 2 Solve $-u''=1$, $u(0)=1$, $u(1)=2$ using the Green's function plus a correction for the boundary values.
  • 3 For which of $f=1$, $f=\cos\pi x$, $f=x-\tfrac12$ is $u''+\pi^2u=f$, $u(0)=u(1)=0$ solvable?
  • 4 Show that $u''=f$ with periodic conditions $u(0)=u(1)$, $u'(0)=u'(1)$ is solvable iff $\int_0^1f=0$, and that the solution is then unique up to an additive constant.
  • 5 With the modified Green's function of Example 4, solve $-u''=\cos\pi x$, $u'(0)=u'(1)=0$.
  • 6 Let $p\gt0$, $q\ge0$. Prove that the Green's function of $-(pu')'+qu=f$, $u(a)=u(b)=0$ is positive on $(a,b)^2$; deduce that $f\ge0$ implies $u\ge0$.
  • 7 Find the Green's function of $u''+u'=f$, $u(0)=u(1)=0$ and show that it is not symmetric. What is its relation to the Green's function of $(e^xu')'=g$?
  • 8 Show that $y''+\lambda y=0$, $y(0)=y(1)=0$ is equivalent to the integral equation $y(x)=\lambda\int_0^1G(x,t)y(t)\,dt$ with $G=\min(x,t)(1-\max(x,t))$.
1 $u=-\cos x\int_0^x\sin t\,dt-\sin x\int_x^{\pi/2}\cos t\,dt=-\cos x(1-\cos x)-\sin x(1-\sin x)=1-\cos x-\sin x$. Check: $u''+u=1$, $u(0)=0$, $u(\tfrac\pi2)=0$.
2 $\int_0^1G\cdot1=\tfrac{x(1-x)}2$, and $w=1\cdot(1-x)+2x=1+x$ solves $-w''=0$ with the boundary values. $u=1+\tfrac32x-\tfrac12x^2$.
3 Null solution $\sin\pi x$. $\int_0^1\sin\pi x\,dx=\tfrac2\pi\ne0$: not solvable for $f=1$. $\int_0^1\cos\pi x\sin\pi x=0$: solvable. $\int_0^1\left(x-\tfrac12\right)\sin\pi x=\tfrac1\pi-\tfrac12\cdot\tfrac2\pi=0$: solvable.
4 Every solution is $u=\int_0^x(x-t)f(t)\,dt+A+Bx$. $u(1)=u(0)$: $\int_0^1(1-t)f+B=0$. $u'(1)=u'(0)$: $\int_0^1f+B=B$, i.e. $\int_0^1f=0$. Then $B$ is determined and $A$ is free. (The null solutions are the constants; $L=d^2/dx^2$ is symmetric under periodic conditions, so this is the Fredholm alternative again.)
5 $\int_0^1\cos\pi x=0$ ✓. The zero-mean solution is $u=\dfrac{\cos\pi x}{\pi^2}$ (direct check: $-u''=\cos\pi x$, $u'=-\tfrac{\sin\pi x}{\pi}$ vanishes at $0,1$, $\int u=0$); the general solution adds a constant.
6 Let $u_1$ solve $(pu')'-qu=0$, $u_1(a)=0$, $u_1'(a)=1$. If $u_1(c)=0$ for some $c\in(a,b]$, then multiplying by $u_1$ and integrating over $[a,c]$ gives $\int_a^c(pu_1'^2+qu_1^2)=0$, so $u_1\equiv0$ — impossible. Thus $u_1\gt0$ on $(a,b]$; likewise $u_2$ with $u_2(b)=0$, $u_2'(b)=-1$ is positive on $[a,b)$. For $L=(pu')'-qu$, $C=pW$ evaluated at $a$ equals $-p(a)u_2(a)\lt0$, so $G_L=u_1u_2/C\lt0$ and the Green's function of $-(pu')'+qu=-L$ is $-G_L\gt0$. Hence $u=\int Gf\ge0$ if $f\ge0$.
7 $u_1=1-e^{-x}$, $u_2=1-e^{1-x}$, $W=(e-1)e^{-x}$ (normal form, jump $1$): $G=\dfrac{u_1(x)u_2(t)}{W(t)}$ for $x\le t$ and $\dfrac{u_1(t)u_2(x)}{W(t)}$ for $x\ge t$. The factor $1/W(t)=e^t/(e-1)$ depends on $t$ only, so $G(0.3,0.6)\approx-0.1352\ne G(0.6,0.3)\approx-0.1001$. The self-adjoint form has $p=e^x$, $C=e^xW=e-1$ constant, kernel $G_s=u_1(x)u_2(t)/(e-1)$ (symmetric), and since $g=e^xf$: $G(x,t)=G_s(x,t)\,e^t$.
8 If $y$ solves the BVP, then $-y''=\lambda y$, so $y=\int_0^1G(x,t)\,\lambda y(t)\,dt$ by Theorem 2.30. Conversely, a continuous solution of the integral equation is $C^2$ (the right side is) and, by the same theorem, satisfies $-y''=\lambda y$ and the boundary conditions. This is the bridge to Fredholm integral equations with symmetric kernels (§7.1–7.2).
III
Eigenfunction expansions and the parameter.
  • 1 Use the bilinear series of Example 5 at $x=t=\tfrac14$ to evaluate $\sum_{n\ge1}\frac{\sin^2(n\pi/4)}{n^2}$.
  • 2 For $-u''=f$, $u(0)=0$, $u'(1)=0$ (Green's function $\min(x,t)$), use the trace identity to evaluate $\sum_{n\ge1}\frac1{(2n-1)^2}$.
  • 3 Use $\int\!\!\int\min(x,t)^2$ to evaluate $\sum_{n\ge1}\frac1{(2n-1)^4}$.
  • 4 Verify directly from the closed form of Example 6 that $(k^2-\pi^2)G(x,t;k)\to2\sin\pi x\sin\pi t$ as $k\to\pi$.
  • 5 Solve $u''+\frac{\pi^2}{4}u=1$, $u(0)=u(1)=0$, and confirm that $u\le0$.
  • 6 Show that for $k=\tfrac{3\pi}2$ the solution of $u''+k^2u=1$, $u(0)=u(1)=0$ is nonnegative on $[0,1]$.
  • 7 Let $G$ be continuous and symmetric on $[a,b]^2$ and $(Tf)(x)=\int_a^bG(x,t)f(t)\,dt$. Prove $\int_a^b(Tf)\,g=\int_a^bf\,(Tg)$ for continuous $f,g$, and deduce that eigenfunctions of $T$ for distinct eigenvalues are orthogonal.
1 $G(\tfrac14,\tfrac14)=\tfrac14\cdot\tfrac34=\tfrac3{16}=\frac2{\pi^2}\sum\frac{\sin^2(n\pi/4)}{n^2}$, so the sum is $\frac{3\pi^2}{32}$.
2 Eigenpairs: $\lambda_n=\left(n-\tfrac12\right)^2\pi^2$, $\varphi_n=\sqrt2\sin\left(n-\tfrac12\right)\pi x$, and $G=\sum\varphi_n(x)\varphi_n(t)/\lambda_n$. Trace: $\int_0^1x\,dx=\tfrac12=\sum\frac{4}{(2n-1)^2\pi^2}$, so $\sum\frac1{(2n-1)^2}=\frac{\pi^2}8$.
3 $\int_0^1\!\int_0^1\min(x,t)^2=2\int_0^1\!\int_0^tx^2\,dx\,dt=\tfrac16=\sum\frac{16}{(2n-1)^4\pi^4}$, so $\sum\frac1{(2n-1)^4}=\frac{\pi^4}{96}$.
4 For $x\le t$, $(k^2-\pi^2)G=-\dfrac{(k+\pi)\sin kx\sin k(1-t)}{k}\cdot\dfrac{k-\pi}{\sin k}$. As $k\to\pi$: $\dfrac{k-\pi}{\sin k}\to\dfrac{1}{\cos\pi}=-1$, $\sin k(1-t)\to\sin\pi(1-t)=\sin\pi t$, so the limit is $\dfrac{2\pi}{\pi}\sin\pi x\sin\pi t=2\sin\pi x\sin\pi t$.
5 $u=\tfrac4{\pi^2}+A\cos\tfrac{\pi x}2+B\sin\tfrac{\pi x}2$ with $A=B=-\tfrac4{\pi^2}$: $u=\tfrac{4}{\pi^2}\left(1-\cos\tfrac{\pi x}{2}-\sin\tfrac{\pi x}{2}\right)$. Since $\cos\theta+\sin\theta=\sqrt2\sin(\theta+\tfrac\pi4)\ge1$ for $\theta\in[0,\tfrac\pi2]$, $u\le0$ — as predicted by $G\lt0$ for $k=\tfrac\pi2\lt\pi$.
6 As in Example 6(d), $u=\frac{1-\cos kx+\sin kx}{k^2}$. With $\theta=kx\in[0,\tfrac{3\pi}2]$: $1-\cos\theta+\sin\theta=1+\sqrt2\sin\left(\theta-\tfrac\pi4\right)$, which is $\ge0$ iff $\sin(\theta-\tfrac\pi4)\ge-\tfrac1{\sqrt2}$, true for $\theta-\tfrac\pi4\in[-\tfrac\pi4,\tfrac{5\pi}4]$ — exactly our range. Equality only at $\theta=0$ and $\theta=\tfrac{3\pi}{2}$, i.e. at the ends.
7 By Fubini and symmetry, $\int(Tf)g=\int\!\!\int G(x,t)f(t)g(x)\,dt\,dx=\int f(t)\int G(t,x)g(x)\,dx\,dt=\int f\,(Tg)$. If $T\varphi=\alpha\varphi$, $T\psi=\beta\psi$, $\alpha\ne\beta$: $\alpha\int\varphi\psi=\int(T\varphi)\psi=\int\varphi(T\psi)=\beta\int\varphi\psi$, so $\int\varphi\psi=0$.
IV
Initial value problems, higher order and modelling.
  • 1 Find the causal Green's function of $u''+3u'+2u=f$, $u(0)=u'(0)=0$, and solve for $f\equiv1$.
  • 2 Find the causal Green's function of $x^2u''-2xu'+2u=g$ on $[1,\infty)$, $u(1)=u'(1)=0$ (normal form first!), and solve for $g=x^3$.
  • 3 Show that the Green's function of $u'''=f$, $u(0)=u'(0)=u''(0)=0$ is $\tfrac12(x-t)^2$ for $t\le x$ and $0$ otherwise, and recover Cauchy's formula for repeated integration.
  • 4 A simply supported beam: $u''''=f$, $u(0)=u''(0)=u(1)=u''(1)=0$. Verify that $G(x,t)=\tfrac16x(1-t)(2t-t^2-x^2)$ for $x\le t$ (extended symmetrically) is its Green's function, and find the midspan deflection under $f\equiv1$.
  • 5 Prove that the Green's function of the cantilever in Example 7 is symmetric, and explain why symmetry was to be expected.
  • 6 Find the Green's function of $-u''+u=f$ on $\mathbb{R}$ for bounded solutions, and use it to solve the problem with $f=\cos x$.
  • 7 (Modelling.) A steel rod of length $L=0.4\ \text{m}$ and thermal conductivity $k=50\ \text{W m}^{-1}\text{K}^{-1}$ is held at $0\,^\circ\text{C}$ (relative temperature) at $x=0$ and insulated at $x=L$; it generates heat uniformly at $q_0=5000\ \text{W m}^{-3}$. The steady temperature satisfies $-ku''=q_0$, $u(0)=0$, $u'(L)=0$. Use the Green's function to find the temperature at the insulated end.
1 Roots $-1,-2$; $k(s)$ with $k(0)=0$, $k'(0)=1$ is $e^{-s}-e^{-2s}$. $G=e^{-(x-t)}-e^{-2(x-t)}$ for $t\le x$, $0$ otherwise. $u=\int_0^x\left(e^{-(x-t)}-e^{-2(x-t)}\right)dt=\tfrac12-e^{-x}+\tfrac12e^{-2x}$; check $u''+3u'+2u=1$, $u(0)=u'(0)=0$.
2 Normal form $u''-\tfrac2xu'+\tfrac2{x^2}u=\tfrac{g}{x^2}$; fundamental set $x$, $x^2$, $W=x^2$. $K(x,t)=\dfrac{y_1(t)y_2(x)-y_1(x)y_2(t)}{W(t)}=\dfrac{x^2}{t}-x$. For $g=x^3$ the normal-form forcing is $x$: $u=\int_1^x\left(\tfrac{x^2}{t}-x\right)t\,dt=x^2(x-1)-\tfrac{x(x^2-1)}2=\tfrac{x^3}{2}-x^2+\tfrac{x}{2}$. Check: $x^2(3x-2)-2x\left(\tfrac32x^2-2x+\tfrac12\right)+2\left(\tfrac{x^3}2-x^2+\tfrac x2\right)=x^3$ ✓, $u(1)=u'(1)=0$ ✓.
3 $K(\cdot,t)$ solves $u'''=0$ with $K=K_x=0$, $K_{xx}=1$ at $x=t$: $K=\tfrac12(x-t)^2$. So the solution is $\int_0^x\tfrac12(x-t)^2f(t)\,dt$, which therefore equals the triple integral $\int_0^x\!\int_0^{s_1}\!\int_0^{s_2}f$ — Cauchy's formula $\int_0^x\frac{(x-t)^{n-1}}{(n-1)!}f(t)\,dt$ for $n=3$.
4 Each piece is a cubic in $x$. Left: $G(0,t)=0$, $\partial_x^2G=-x(1-t)$ vanishes at $0$; right piece (swap $x,t$): $G=\tfrac16t(1-x)(2x-x^2-t^2)$ vanishes at $x=1$, and $\partial_x^2G=-t(1-x)$ vanishes at $1$. At $x=t$, $G$, $\partial_xG$, $\partial_x^2G$ match and $\partial_x^3G$ jumps from $-(1-t)$ to $t$, by $+1$. Uniform load: $u=\tfrac{x(1-2x^2+x^3)}{24}$, so $u(\tfrac12)=\tfrac{5}{384}$ (the classical $\frac{5wL^4}{384EI}$).
5 For $x\le t$, $G(t,x)$ is given by the second line with the roles exchanged: $\tfrac16x^2(3t-x)=G(x,t)$. Expected because $u''''$ with these four conditions is self-adjoint: integrating $\int(vu''''-uv'''')$ by parts four times leaves $[vu'''-v'u''+v''u'-v'''u]_0^1$, which vanishes when both $u,v$ are clamped at $0$ and free at $1$.
6 Bounded solutions of $-u''+u=0$ on each side of $t$: $e^{x}$ for $x\lt t$ and $e^{-x}$ for $x\gt t$. Continuity and jump $-1$ in $\partial_xG$ give $G(x,t)=\tfrac12e^{-|x-t|}$. For $f=\cos x$: $u=\tfrac12\int_{\mathbb{R}}e^{-|s|}\cos(x-s)\,ds=\cos x\cdot\tfrac12\int_{\mathbb{R}}e^{-|s|}\cos s\,ds=\tfrac12\cos x$ (the sine part vanishes by oddness, and $\int_0^\infty e^{-s}\cos s\,ds=\tfrac12$). Check: $-u''+u=\tfrac12\cos x+\tfrac12\cos x=\cos x$.
7 By Example 1 (scaled to $[0,L]$) the Green's function of $-u''$ with these conditions is $\min(x,t)$, so $u(x)=\frac{q_0}{k}\int_0^L\min(x,t)\,dt=\frac{q_0}{k}\left(Lx-\frac{x^2}2\right)$. At $x=L$: $u(L)=\frac{q_0L^2}{2k}=\frac{5000\times0.16}{2\times50}=8\ \text{K}$, i.e. the insulated end is $8\,^\circ\text{C}$ above the held end.
🎯 Exam corner CSIR Part BCSIR Part CGATE MA

E1. CSIR Part B GATE MA Let $G(x,t)$ be the Green's function of $u''=f$, $u(0)=u(1)=0$ (so that $u=\int_0^1G(x,t)f(t)\,dt$). Then $G(\tfrac14,\tfrac12)$ equals (A) $\tfrac18$; (B) $-\tfrac18$; (C) $-\tfrac{3}{16}$; (D) $\tfrac{3}{16}$.

Idea For $u''$ (not $-u''$): $u_1=x$, $u_2=x-1$, $C=x\cdot1-1\cdot(x-1)=1$, so $G=x(t-1)$ for $x\le t$ and $t(x-1)$ for $x\ge t$.
Compute $x=\tfrac14\le t=\tfrac12$: $G=\tfrac14\left(\tfrac12-1\right)=-\tfrac18$. (A) is the value for $-u''$ (sign convention); (C) $=-\tfrac14\cdot\tfrac34$ is $G(\tfrac14,\tfrac14)$ — evaluating on the diagonal instead of at $(\tfrac14,\tfrac12)$; (D) combines both slips.
Answer (B).

E2. CSIR Part B The Green's function of $-u''=f$, $u(0)=0$, $u'(1)=0$ is (A) $\min(x,t)$; (B) $\max(x,t)$; (C) $\min(x,t)\big(1-\max(x,t)\big)$; (D) $1-\max(x,t)$.

Check the conditions (B) fails $G(0,t)=0$ ($\max(0,t)=t$). (C) belongs to $u(0)=u(1)=0$: its slope for $x\gt t$ is $-t\ne0$. (D) satisfies $u'(0)=0$, $u(1)=0$ — the mirror problem (Exercise I.1). (A): $G(0,t)=0$, constant $=t$ for $x\gt t$ so $\partial_xG(1,t)=0$, slope jumps from $1$ to $0$ (by $-1$).
Answer (A) (Example 1).

E3. CSIR Part B If $u$ solves $u''+u=x$, $u(0)=0$, $u(\tfrac{\pi}{2})=0$, then $u(\tfrac{\pi}{4})$ equals (A) $\dfrac{\pi(1-\sqrt2)}{4}$; (B) $\dfrac{\pi(\sqrt2-1)}{4}$; (C) $\dfrac{\pi}{4}$; (D) $0$.

Idea Either through the Green's function $-\sin(\min)\cos(\max)$ (Example 2), or directly: $u=x+A\cos x+B\sin x$, $A=0$, $B=-\tfrac\pi2$.
Compute $u(\tfrac\pi4)=\tfrac\pi4-\tfrac\pi2\cdot\tfrac{\sqrt2}2=\tfrac{\pi(1-\sqrt2)}4\approx-0.325$. The sign is negative because $G\le0$ and $f=x\ge0$ — option (B) has the wrong sign.
Answer (A).

E4. CSIR Part C Let $G$ be the Green's function of $-u''=f$, $u(0)=u(1)=0$. Which of the following are true? (One or more options may be correct.) (A) $G(x,t)=G(t,x)$ for all $x,t\in[0,1]$. (B) $G(x,t)\ge0$ on $[0,1]^2$. (C) $\partial G/\partial x$ is continuous on $[0,1]^2$. (D) $\int_0^1\!\int_0^1G(x,t)\,dx\,dt=\tfrac16$.

(A) True $G=\min(x,t)(1-\max(x,t))$ is symmetric (self-adjoint problem).
(B) True Both factors are $\ge0$.
(C) False $\partial_xG$ jumps by $-1$ across the diagonal.
(D) False $\int_0^1G(x,t)\,dt$ is the solution for $f\equiv1$, namely $\tfrac{x(1-x)}2$, whose integral is $\tfrac1{12}$. ($\tfrac16$ is the trace $\int_0^1G(x,x)\,dx$.)
Answer (A) and (B).

E5. CSIR Part C Consider $u''+k^2u=f$, $u(0)=u(\pi)=0$, with $k\gt0$ and $f$ continuous. Which of the following are true? (One or more options may be correct.) (A) The problem has a unique solution for every $f$ if and only if $k$ is not an integer. (B) For $k=1$ it is solvable if and only if $\int_0^\pi f(x)\sin x\,dx=0$. (C) For $k=\tfrac12$ the Green's function is symmetric. (D) For $k=\tfrac32$, $f\ge0$ implies $u\le0$ on $[0,\pi]$.

(A) True The Dirichlet eigenvalues on $[0,\pi]$ are $n^2$, so the homogeneous problem is nontrivial iff $k^2=n^2$, i.e. $k\in\mathbb{N}$ (Theorems 2.21, 2.30).
(B) True Fredholm alternative (Theorem 2.31) with $\varphi=\sin x$.
(C) True The problem is self-adjoint; $G=-\dfrac{\sin(\frac x2)\sin(\frac{\pi-t}2)}{\frac12\sin\frac\pi2}$ for $x\le t$, symmetric by Theorem 2.30(3).
(D) False For $k=\tfrac32$, $G(x,t)=-\dfrac{\sin\frac{3x}2\sin\frac{3(\pi-t)}2}{\frac32\sin\frac{3\pi}2}=\tfrac23\sin\tfrac{3x}{2}\sin\tfrac{3(\pi-t)}{2}$ for $x\le t$, which changes sign: e.g. $G(0.5,2.5)\approx0.373\gt0$ while $G(2.3,2.5)\approx-0.166\lt0$. A nonnegative $f$ concentrated near $t=2.5$ then gives $u(0.5)\gt0$. (We are past the first eigenvalue $k^2=1$.)
Answer (A), (B) and (C).

E6. CSIR Part C Consider $-u''=f$ on $[0,1]$ with $u'(0)=u'(1)=0$. Which of the following are true? (One or more options may be correct.) (A) The problem has a Green's function. (B) It is solvable if and only if $\int_0^1f(x)\,dx=0$. (C) When it is solvable, any two solutions differ by a constant. (D) For $f(x)=x$, $u(x)=-\tfrac{x^3}{6}$ is a solution.

(A) False Constants solve the homogeneous problem (Theorem 2.30 needs only the trivial solution; ⚠️ 1).
(B) True Theorem 2.31 with $\varphi=1$.
(C) True Theorem 2.31(3): the null space is spanned by $1$.
(D) False $-u''=x$ holds, but $u'(1)=-\tfrac12\ne0$; indeed $\int_0^1x\,dx=\tfrac12\ne0$, so no solution exists at all.
Answer (B) and (C).
🧑‍🏫 For the teacher

Suggested plan — 2 lectures × 60 min.

  1. Construction. Derive the loaded-string Green's function by integrating twice (ten minutes) and list its five properties on the board; then Definition 2.14 and Theorem 2.30 with the full proof — emphasise that $pW$ is constant, which students can use as a check. Project Interactive 1 (first mode, then the mixed and Robin modes: ask the class to predict the shape before switching). Examples 1 and 2 on the board; Example 3 as the cautionary tale about the right-hand side.
  2. Beyond the basic case. Theorem 2.31 (Fredholm alternative) with the insulated rod (Example 4) as motivation; mention the modified Green's function briefly. Theorem 2.33 and Example 5 — students enjoy getting $\zeta(2)$ and $\zeta(4)$ from a string; then Interactive 1 in the $u''+k^2u$ mode to show the pole at $k=\pi$ and the loss of positivity (Example 6). Finish with Theorem 2.34, the causal kernel and the cantilever (Example 7), and Interactive 2 in both modes.

Discussion questions. (1) In what sense is $G$ "the inverse matrix" of the boundary value problem, and what plays the role of the identity matrix? (2) Why must a Green's function of a self-adjoint problem be symmetric, and what physical principle does this express for beams and strings? (3) The causal Green's function of $u''+u=f$ and the Green's function of the boundary value problem on $[0,\frac\pi2]$ are both built from $\sin$ and $\cos$. What is the essential structural difference?

Misconceptions to watch for. Using the jump $+1$ for $-u''$ (or forgetting the factor $1/p$); applying $G$ of the self-adjoint form to the original right-hand side; believing every BVP has a Green's function; expecting symmetry for non-self-adjoint problems or IVPs; forgetting to correct for non-homogeneous boundary values; confusing $\int\!\!\int G$ with the trace $\int G(x,x)$.

Suggested homework. I(1, 2, 5, 8), II(1, 3, 6, 7), III(2, 4, 6), IV(1, 2, 4, 7); Exam corner E1, E5.

🔬 Research–Teaching NexusResearch-tutored · optional, about 60 minutes

Research question. Where should a heavy vehicle stand on a bridge to make a given point of the deck sag the most? Bridge engineers answer this with influence lines: the deflection at a fixed point $x_0$ caused by a unit load at a variable point $t$, i.e. $t\mapsto G(x_0,t)$. How does Maxwell–Betti reciprocity, $G(x_0,t)=G(t,x_0)$, turn this into a question about a single deflection curve?

Why it matters. Codes of practice for highway and railway bridges (for example the load models of the Indian Roads Congress or of the Eurocodes) prescribe moving vehicle loads, and the worst position of each vehicle is found from influence lines. Maxwell stated the reciprocal theorem in 1864 and Betti generalised it in 1872; in our language it is the symmetry of the Green's function of a self-adjoint problem.

What to do.

  1. For a simply supported beam of unit length and unit rigidity, $u''''=f$, $u(0)=u''(0)=u(1)=u''(1)=0$, take the Green's function of Exercise IV.4. Check its symmetry numerically on a grid.
  2. Plot the influence line $t\mapsto G(\tfrac14,t)$ for the deflection at the quarter point, and find the load position $t^*$ that maximises it — analytically (it is a cubic in $t$) and numerically. Explain, using reciprocity, why $t^*$ is also the point where the deflection curve $x\mapsto G(x,\tfrac14)$ under a load at $\tfrac14$ is largest.
  3. Model a two-axle vehicle as two unit loads a distance $0.2$ apart. Find the position that maximises the midspan deflection and compare it with a single load of size $2$ at midspan.
  4. Repeat step 2 for the taut string ($-u''=f$, Dirichlet). Why is the worst position for the string exactly at the point itself, but not for the beam?

Tools: Python or a spreadsheet; Interactive 1 for the second-order analogue.

Hand in: The influence-line plot, the values of $t^*$ and the maximal deflections, and a short explanation of reciprocity in your own words.

Going further. Influence lines for bending moment and shear are Green's functions of the same beam for different "observables". Read about the Müller-Breslau principle, which obtains them by imposing a unit kink or a unit jump at the observation point — a mechanical picture of the jump conditions of Definition 2.14.

Instructor note: for $x\le t$, $G(\tfrac14,t)=\tfrac1{24}(1-t)\left(2t-t^2-\tfrac1{16}\right)$; setting the derivative to $0$ gives $3t^2-6t+\tfrac{33}{16}=0$, so $t^*=1-\tfrac{\sqrt5}{4}\approx0.4410$ (not $0.25$!) with $G\approx0.014558$; the same point maximises the deflection curve under a load at $\tfrac14$, by reciprocity. Two unit loads $0.2$ apart: the maximum midspan deflection is at the symmetric position $t=0.4,\,0.6$, value $2G(\tfrac12,0.4)\approx0.039333$, versus $2G(\tfrac12,\tfrac12)=\tfrac1{24}\approx0.041667$ for a double load at midspan. For the string, $G(x_0,t)$ is a tent with its peak at $t=x_0$, because the second-order Green's function is piecewise linear in $t$.

🧪 Self-check quiz

0/5 correct

Q1. For $L[u]=(p\,u')'+q\,u$ with separated boundary conditions, $\partial G/\partial x$ at $x=t$

Q2. The Green's function of $-u''=f$, $u(0)=u(1)=0$, for $0\le x\le t\le1$, is

Q3. Which boundary value problem has no Green's function?

Q4. The Green's function of the initial value problem $u''+u=f$, $u(0)=u'(0)=0$ (so that $u=\int_0^\infty G(x,t)f(t)\,dt$) is

Q5. Using $\min(x,t)(1-\max(x,t))=\sum_{n\ge1}\frac{2\sin n\pi x\sin n\pi t}{n^2\pi^2}$, the value of $\sum_{n\ge1}\frac1{n^2}$ follows from