6.2 Sufficient Conditions and Variational Methods
When is an extremal really a minimum? The second variation, Legendre's and Jacobi's conditions and conjugate points, weak versus strong minima and the Weierstrass excess function; then the variational view of boundary value problems — energy minimisation, the Rayleigh–Ritz and Galerkin methods, and the Dirichlet principle for Poisson's equation.
The Euler–Lagrange equation of §6.1 is only a necessary condition: it finds the curves at which the first variation vanishes, just as $f'(x)=0$ finds the critical points of a function. In one variable we then look at $f''$. The same idea works for functionals, but with a twist that has no finite-dimensional analogue: the second variation can be positive in every "short" direction and still negative in a long one. Legendre (1786) found the pointwise condition $F_{y'y'}\ge0$; Jacobi (1837) discovered that one must also watch for conjugate points, where neighbouring extremals from the starting point cross the extremal again — the reason a great circle on a sphere stops being the shortest path once it passes the antipodal point. Weierstrass (1879) added a condition that distinguishes minima among nearby curves from minima among nearby curves with nearby slopes.
The second half of the section turns the logic around. Instead of solving a differential equation to minimise a functional, we minimise a functional to solve a differential equation. The boundary value problem $-(pu')'+qu=f$ is equivalent to minimising an energy; approximating the minimiser in a finite-dimensional space gives the Rayleigh–Ritz method (Ritz, 1908), its generalisation by Galerkin (1915), and ultimately the finite element method used in every engineering design office. The same idea for Laplace's and Poisson's equations is the Dirichlet principle, whose rescue by Hilbert around 1900 shaped modern analysis. CSIR NET asks for conjugate points, Legendre's condition and Ritz coefficients in Part B, and for the logical relations between the conditions (necessary, sufficient, weak, strong) in Part C. We build on the Rayleigh quotient of §2.3 and use the notation of §6.1.
The Second Variation and Legendre's Condition
Throughout, $J[y]=\int_a^bF(x,y,y')\,dx$ on the fixed end-point class $\mathcal{A}$ of §6.1, $F\in C^3$, and $y\in C^2[a,b]$ is an extremal in $\mathcal{A}$. Variations $\eta\in C^1[a,b]$ vanish at $a$ and $b$; the norms $\|\cdot\|_0$, $\|\cdot\|_1$ and weak/strong extrema are as in Definition 6.2.
- $$J[y+\eta]=J[y]+\delta J[y;\eta]+\tfrac12\delta^2J[y;\eta]+R[\eta]$$, where $|R[\eta]|\le\rho\big(\|\eta\|_1\big)\int_a^b(\eta^2+\eta'^2)\,dx$ for a function $\rho(r)\to0$ as $r\to0$.
- If $y$ is a weak local minimum, then $\delta^2J[y;\eta]\ge0$ for every variation $\eta$.
- For every variation, $\delta^2J[y;\eta]=\displaystyle\int_a^b\big(P\eta'^2+Q\eta^2\big)\,dx$.
Proof. (1) Taylor's formula with integral remainder for $F$ in the variables $(y,p)$ gives $$F(x,y+\eta,y'+\eta')=F+F_y\eta+F_{y'}\eta'+\tfrac12\Phi+r(x),$$ $$\Phi=F_{yy}\eta^2+2F_{yy'}\eta\eta'+F_{y'y'}\eta'^2,$$ where $r$ is the same quadratic expression with the second derivatives replaced by their differences between an intermediate point and $(x,y,y')$. The second derivatives of $F$ are uniformly continuous on a compact neighbourhood of the graph, so these differences are at most $\rho(\|\eta\|_1)\to0$, and $2|\eta\eta'|\le\eta^2+\eta'^2$. Integrate. (2) $\varphi(\varepsilon)=J[y+\varepsilon\eta]$ is $C^2$ with a local minimum at $0$, so $\varphi''(0)\ge0$. (3) Since $\eta^2$ vanishes at both ends, $\int2F_{yy'}\eta\eta'=\int F_{yy'}(\eta^2)'=-\int\big(\tfrac{d}{dx}F_{yy'}\big)\eta^2$. ∎
The reduced form shows the competition between $P\eta'^2$ and $Q\eta^2$. A variation can have small values and large slopes — a short, steep bump — and then $P\eta'^2$ dominates. This observation gives Legendre's condition.
Proof. Suppose $P(x_0)\lt0$. By continuity there are $m\gt0$ and an interval $I_0$ of positive length with end point $x_0$, contained in $[a,b]$, on which $P\le-m$; say $I_0=[x_0,x_0+h_0]$ (otherwise reflect). For $0\lt h\le h_0$ let $$\eta_h(x)=\sin^2\frac{\pi(x-x_0)}{h}\ \ \text{on }[x_0,x_0+h],\qquad\eta_h=0\ \text{elsewhere}.$$ It is $C^1$, $0\le\eta_h\le1$, $|\eta_h'|\le\pi/h$ and $\int\eta_h'^2\,dx=\frac{\pi^2}{2h}$. With $M_1=\max|F_{yy}|$ and $M_2=\max|F_{yy'}|$ along $y$, using the form of Definition 6.6, $$\delta^2J[y;\eta_h]\le M_1h+2M_2\cdot\frac{\pi}{h}\cdot h-\frac{m\pi^2}{2h}=M_1h+2\pi M_2-\frac{m\pi^2}{2h},$$ which is negative for small $h$. This contradicts Theorem 6.12(2). ∎
Legendre's condition alone is far from sufficient (Example 2 below). One case where everything is easy is convexity.
Proof. A $C^1$ convex function lies above its tangent plane: $$F(x,y_0+\eta,y_0'+\eta')\ge F(x,y_0,y_0')+F_y\eta+F_{y'}\eta'$$. Integrating, $J[y_0+\eta]\ge J[y_0]+\delta J[y_0;\eta]=J[y_0]$, since the first variation vanishes at an extremal (integrate by parts, using Theorem 6.4). If the convexity is strict, equality forces $(\eta,\eta')=(0,0)$ at every point. ∎
Jacobi's Condition and Conjugate Points
By Theorem 6.12(3), the extremal $y$ can be a weak minimum only if the quadratic functional $\eta\mapsto\int(P\eta'^2+Q\eta^2)$ is nonnegative on variations. Its own Euler–Lagrange equation decides this.
Proof. Differentiate the identity $$F_y\big(x,y(x,c),y_x(x,c)\big)-\frac{d}{dx}F_{y'}\big(x,y(x,c),y_x(x,c)\big)=0$$ with respect to $c$ at $c_0$: $$F_{yy}u+F_{yy'}u'-\frac{d}{dx}\big(F_{yy'}u+F_{y'y'}u'\big)=0.$$ Since $F_{yy'}u'-\frac{d}{dx}(F_{yy'}u)=-\big(\frac{d}{dx}F_{yy'}\big)u$, this is $Qu-(Pu')'=0$. For the pencil, $y(a,c)=y(a)$ for all $c$ and $y_x(a,c)=c$, so $u(a)=0$ and $u'(a)=1$. ∎
Proof (sketch). Suppose $a^*\in(a,b)$ is conjugate to $a$, and let $\eta=u_a$ on $[a,a^*]$, $\eta=0$ on $[a^*,b]$. Integrating by parts and using the Jacobi equation, $\delta^2J[y;\eta]=\big[Pu_au_a'\big]_a^{a^*}=0$. Since $\delta^2J\ge0$ for all variations (Theorem 6.12), $\eta$ minimises the quadratic functional $\delta^2J$; it is piecewise $C^1$ and $\delta^2J$ is approximated by $C^1$ variations, so the Weierstrass–Erdmann corner condition (Theorem 6.18 below, applied to $\delta^2J$) requires $P\eta'$ to be continuous at $a^*$. But $P(a^*)u_a'(a^*)\ne0$ (a nontrivial solution cannot have a double zero), while $\eta'=0$ to the right of $a^*$ — a contradiction. ∎
The decisive point is that the strengthened conditions are sufficient.
- (L⁺) $P(x)=F_{y'y'}\big(x,y,y'\big)\gt0$ for all $x\in[a,b]$ (strengthened Legendre condition);
- (J⁺) there is no point conjugate to $a$ in the half-open interval $(a,b]$ (strengthened Jacobi condition).
Proof. Step 1: a positive solution. The solution $u_a$ is positive on $(a,b]$. By continuous dependence of solutions of linear ODEs on initial data and parameters (§1.1), for small $\alpha\gt0$ and $\delta\gt0$ the solution $v$ of $-\big((P-\alpha)v'\big)'+Qv=0$ with $v(a)=\delta$, $v'(a)=1$ is still positive on $[a,b]$: near $a$ because $v(a)\gt0$ and $v'\approx1$, and on the rest of $[a,b]$ because $v$ is uniformly close to $u_a$, which has a positive minimum there. Choose $\alpha$ also so small that $\tilde P=P-\alpha\gt0$.
Step 2: Picone's identity. For a variation $\eta$, a direct differentiation using $(\tilde Pv')'=Qv$ gives $$\frac{d}{dx}\Big(\frac{\tilde Pv'}{v}\eta^2\Big)=Q\eta^2+\frac{2\tilde Pv'\eta\eta'}{v}-\frac{\tilde Pv'^2\eta^2}{v^2},$$ hence $$\tilde P\eta'^2+Q\eta^2=\tilde P\Big(\eta'-\frac{v'}{v}\eta\Big)^2+\frac{d}{dx}\Big(\frac{\tilde Pv'}{v}\eta^2\Big).$$ Integrating, the boundary term vanishes and $\int(\tilde P\eta'^2+Q\eta^2)\ge0$, that is $$\delta^2J[y;\eta]\ \ge\ \alpha\int_a^b\eta'^2\,dx.$$
Step 3: from the second variation to $J$. Since $\eta(x)=\int_a^x\eta'$, the Cauchy–Schwarz inequality gives $\int_a^b\eta^2\le(b-a)^2\int_a^b\eta'^2$. By Theorem 6.12(1), with $\delta J=0$, $$J[y+\eta]-J[y]\ge\Big(\frac\alpha2-\rho(\|\eta\|_1)\big(1+(b-a)^2\big)\Big)\int_a^b\eta'^2\,dx,$$ which is positive whenever $\|\eta\|_1$ is so small that the bracket is positive and $\eta\not\equiv0$. ∎
Weak and Strong Extrema: the Weierstrass Excess Function
Theorem 6.17 compares $y$ only with curves whose slopes are also close to $y'$. A strong neighbourhood also contains curves with small, steep zigzags, and these can lower $J$ even when every smooth nearby curve raises it. Weierstrass measured the effect of replacing the slope $p=y'$ by an arbitrary slope $q$ on a short stretch.
- (Necessary.) If the extremal $y$ is a strong local minimum, then $E\big(x,y(x),y'(x),q\big)\ge0$ for all $x\in[a,b]$ and all $q\in\mathbb{R}$.
- (Sufficient, statement.) If $y$ satisfies (L⁺) and (J⁺) — so that it can be embedded in a field of extremals with slope function $\psi(x,Y)$ covering a neighbourhood of its graph — and $E\big(x,Y,\psi(x,Y),q\big)\ge0$ for all $(x,Y)$ in that neighbourhood and all $q$, then $y$ is a strong local minimum. In particular this holds if $F_{y'y'}(x,Y,q)\ge0$ for all $(x,Y)$ near the graph and all $q$.
- (Weierstrass–Erdmann corner conditions.) If a piecewise $C^1$ curve minimises $J$ among piecewise $C^1$ curves, then at every corner both $F_{y'}$ and $F-y'F_{y'}$ are continuous.
Proof (ideas). (1) Fix $x_0$ and $q$. For small $h\gt0$ replace $y$ on $[x_0,x_0+h]$ by a segment of slope $q$ starting at $(x_0,y(x_0))$ followed by a segment rejoining $y$ at $x_0+\sqrt h$ (a "needle"). The new curve is within $O(h)$ of $y$ in $\|\cdot\|_0$, though its slope differs by $q-y'$. A direct expansion gives $$J[\text{new}]-J[y]=h\,E\big(x_0,y(x_0),y'(x_0),q\big)+o(h)$$, which must be $\ge0$. (2) In a field, Hilbert's invariant integral yields the Weierstrass formula $J[Y]-J[y]=\int_a^bE\big(x,Y,\psi(x,Y),Y'\big)\,dx$ for every curve $Y$ in the neighbourhood with the same end points. (3) The integrated Euler–Lagrange equation of Theorem 6.4(1) holds for piecewise $C^1$ minimisers, and its right-hand side is continuous; the second condition comes from moving the corner horizontally. ∎
Variational Formulation of Boundary Value Problems
We now reverse the direction of §6.1. Consider the self-adjoint boundary value problem $$-\big(p(x)u'\big)'+q(x)u=f(x),\qquad u(a)=u(b)=0,$$ with $p\in C^1[a,b]$, $p\gt0$, $q,f\in C[a,b]$, $q\ge0$. (In the notation $(py')'+qy+\lambda ry=0$ of §2.3 the coefficient $q$ there is $-q$ here; $q\ge0$ here corresponds to $q\le0$ there.) Physically $u$ is the deflection of a string or the temperature in a rod; $p$ is the tension or conductivity, $q$ an elastic foundation or heat loss, $f$ the load or heat source. Let $\mathcal{A}_0$ be the set of continuous, piecewise $C^1$ functions on $[a,b]$ vanishing at $a$ and $b$, and define the energy functional and the energy inner product $$I[v]=\int_a^b\big(pv'^2+qv^2-2fv\big)\,dx,$$ $$a(u,v)=\int_a^b\big(pu'v'+quv\big)\,dx,\qquad\|v\|_E^2=a(v,v).$$
- If $u\in C^2[a,b]\cap\mathcal{A}_0$ solves the boundary value problem, then for every $\eta\in\mathcal{A}_0$ $$I[u+\eta]=I[u]+\int_a^b\big(p\eta'^2+q\eta^2\big)\,dx,$$ so $u$ is the unique minimiser of $I$ on $\mathcal{A}_0$, and $I[u]=-\int_a^bfu\,dx$.
- Conversely, if $u\in C^2[a,b]\cap\mathcal{A}_0$ minimises $I$ on $\mathcal{A}_0$, then $u$ solves the boundary value problem.
- The boundary value problem has exactly one solution, and it satisfies the weak form $a(u,v)=\int_a^bfv\,dx$ for all $v\in\mathcal{A}_0$.
Proof. (1) Expanding the squares, $$I[u+\eta]-I[u]=2\int_a^b\big(pu'\eta'+qu\eta-f\eta\big)\,dx+\int_a^b\big(p\eta'^2+q\eta^2\big)\,dx.$$ Integrating $pu'\eta'$ by parts on each interval where $\eta$ is $C^1$ (the boundary contributions telescope because $\eta$ and $pu'$ are continuous, and vanish at $a,b$), the first integral becomes $2\int_a^b\big(-(pu')'+qu-f\big)\eta\,dx=0$. The second is $\ge0$, and it is $0$ only if $\eta'\equiv0$ (since $p\gt0$), i.e. $\eta\equiv0$. Finally $I[u]=a(u,u)-2\int fu$ and $a(u,u)=\int fu$ (multiply the equation by $u$ and integrate by parts). (2) The Euler–Lagrange equation of $pv'^2+qv^2-2fv$ is $2qu-2f-(2pu')'=0$ (Theorem 6.4). (3) If $u_0$ solves the homogeneous problem, multiplying by $u_0$ and integrating by parts gives $\int(pu_0'^2+qu_0^2)=0$, so $u_0'\equiv0$ and $u_0\equiv0$. By the solvability theorem for two-point problems (§2.3; equivalently, the Green's function of §2.4 exists), the problem has exactly one solution. The weak form is the computation in (1) with $\eta=v$. ∎
Natural boundary conditions appear exactly as in Theorem 6.9: minimising $I$ over functions with only $v(a)=0$ prescribed produces $p(b)u'(b)=0$; adding $\sigma v(b)^2$ ($\sigma\ge0$) to $I$ produces the Robin condition $p(b)u'(b)+\sigma u(b)=0$. Similarly, by §2.3 and Theorem 6.11, the eigenvalues of $-(pu')'+qu=\lambda ru$ with Dirichlet conditions are the stationary values of the Rayleigh quotient $R[v]=a(v,v)\big/\int rv^2$, the smallest being $\lambda_1=\min R$.
Let $\varphi_1,\dots,\varphi_n\in\mathcal{A}_0$ be linearly independent and $V_n=\operatorname{span}\{\varphi_1,\dots,\varphi_n\}$ (the trial space; its functions must satisfy the essential boundary conditions). The Ritz approximation $u_n$ is the minimiser of $I$ on $V_n$. Writing $u_n=\sum_jc_j\varphi_j$, it is given by the linear system $$Kc=F,\qquad K_{ij}=a(\varphi_i,\varphi_j),\qquad F_i=\int_a^bf\varphi_i\,dx.$$ For a general (not necessarily symmetric) linear problem $Lu=f$, the Galerkin approximation is the $u_n\in V_n$ whose residual is orthogonal to the trial space: $\int_a^b\big(Lu_n-f\big)\varphi_i\,dx=0$, $i=1,\dots,n$ (after an integration by parts, $a(u_n,\varphi_i)=\int f\varphi_i$). For the symmetric problem above Galerkin and Ritz coincide.
- $K$ is symmetric and positive definite, so $u_n$ exists and is unique.
- Galerkin orthogonality: $a(u-u_n,v)=0$ for all $v\in V_n$.
- Energy identity: $I[v]=I[u]+\|v-u\|_E^2$ for every $v\in\mathcal{A}_0$. Hence $u_n$ is the best approximation of $u$ from $V_n$ in the energy norm, $I[u_n]\ge I[u]$, and for nested spaces $V_1\subset V_2\subset\cdots$ the energies $I[u_n]$ decrease.
- Convergence: if every $w\in\mathcal{A}_0$ can be approximated in the energy norm by elements of $\bigcup_nV_n$ (for example $V_n=\operatorname{span}\{\sin\frac{k\pi(x-a)}{b-a}:k\le n\}$), then $\|u-u_n\|_E\to0$ and $\max|u-u_n|\to0$.
Proof. (1) $K_{ij}=K_{ji}$, and for $c\ne0$, $c^{\mathsf T}Kc=a(v,v)=\int(pv'^2+qv^2)\gt0$ with $v=\sum c_i\varphi_i\ne0$ (as in Theorem 6.19, $a(v,v)=0$ forces $v\equiv0$). On $V_n$, $I\big[\sum c_j\varphi_j\big]=c^{\mathsf T}Kc-2c^{\mathsf T}F$, a strictly convex quadratic whose unique critical point solves $Kc=F$. (2) $Kc=F$ says $a(u_n,\varphi_i)=\int f\varphi_i$, and the weak form says $a(u,\varphi_i)=\int f\varphi_i$. (3) Using $\int fv=a(u,v)$: $$I[v]=a(v,v)-2a(u,v)=a(v-u,v-u)-a(u,u)$$, and $I[u]=-a(u,u)$. Minimising $I$ over $V_n$ is therefore minimising $\|v-u\|_E$. Larger spaces give smaller minima. (4) By (3), $\|u-u_n\|_E\le\|u-v\|_E$ for every $v\in V_n$, which tends to $0$ by hypothesis. For $w\in\mathcal{A}_0$, $$|w(x)|=\Big|\int_a^xw'\Big|\le\sqrt{b-a}\,\Big(\int_a^b w'^2\Big)^{1/2},$$ which is at most $\sqrt{(b-a)/\min p}\ \|w\|_E$, so energy convergence implies uniform convergence. Density of the sine system in the energy norm follows from Fourier series of $w'$ (§4.2). ∎
Proof. For $k=1$: $$\Lambda_1^{(n)}=\min_{c\ne0}\frac{c^{\mathsf T}Kc}{c^{\mathsf T}Mc}=\min_{v\in V_n\setminus\{0\}}R[v]\ge\min_{v\in\mathcal{A}_0\setminus\{0\}}R[v]=\lambda_1$$ by the minimum principle of §2.3. For general $k$ we use (statement) the Courant–Fischer min–max principle $\lambda_k=\min_{S}\max_{v\in S\setminus\{0\}}R[v]$, the minimum over $k$-dimensional subspaces $S\subset\mathcal{A}_0$; the same formula with $S\subset V_n$ gives $\Lambda_k^{(n)}$ (the min–max characterisation of the generalised eigenvalues of the symmetric pencil $(K,M)$), and minimising over fewer subspaces gives a larger value. Nested spaces: $V_n\subset V_{n+1}$ means more subspaces to minimise over. ∎
- Energy. Put the equation in self-adjoint form $-(pu')'+qu=f$ (§2.3) and write $I[v]=\int(pv'^2+qv^2-2fv)$ — or, for an eigenvalue problem, $R[v]=\int(pv'^2+qv^2)/\int rv^2$.
- Trial functions. Choose $\varphi_i$ satisfying the essential (Dirichlet) conditions: $x(1-x),\,x^2(1-x),\dots$ or $\sin k\pi x$ on $[0,1]$. Natural conditions need not be imposed. Exploit symmetry.
- Assemble and solve. $K_{ij}=\int(p\varphi_i'\varphi_j'+q\varphi_i\varphi_j)$, $F_i=\int f\varphi_i$ (and $M_{ij}=\int r\varphi_i\varphi_j$); solve $Kc=F$ or $\det(K-\Lambda M)=0$.
- Interpret. $I[u_n]=-c^{\mathsf T}F\ge I[u]$; $\Lambda_k\ge\lambda_k$. Pointwise values carry no one-sided guarantee.
Galerkin for a non-symmetric problem. For $-u''+u'=1$, $u(0)=u(1)=0$ there is no energy to minimise (the operator is not self-adjoint), but Galerkin's condition still makes sense. With $\varphi_1=x(1-x)$ and $u_1=c\varphi_1$ the residual is $R=2c+c(1-2x)-1$, and $\int_0^1R\varphi_1\,dx=\tfrac c3-\tfrac16=0$ gives $c=\tfrac12$, $u_1(\tfrac12)=0.125$; the exact solution $u=x-\frac{e^x-1}{e-1}$ has $u(\tfrac12)\approx0.1225$. Galerkin's method is the basis of the finite element method for convection–diffusion, elasticity and fluid flow, where the operator is often non-symmetric; the price is that the monotonicity of Theorem 6.20(3) is lost.
Poisson's Equation and the Dirichlet Principle
Everything above extends to several variables. Let $\Omega\subset\mathbb{R}^2$ be a bounded domain with piecewise smooth boundary, $f\in C(\overline\Omega)$ and $g$ continuous on $\partial\Omega$, and let $\mathcal{A}_g=\{v\in C^1(\overline\Omega):v=g\text{ on }\partial\Omega\}$, assumed nonempty. The Dirichlet energy with a source is $$D[v]=\iint_\Omega\Big(\tfrac12|\nabla v|^2-fv\Big)\,dx\,dy.$$
- If $u\in C^2(\overline\Omega)$ solves $-\Delta u=f$ in $\Omega$, $u=g$ on $\partial\Omega$, then $D[u+\eta]=D[u]+\tfrac12\iint_\Omega|\nabla\eta|^2$ for every $\eta\in C^1(\overline\Omega)$ vanishing on $\partial\Omega$; hence $u$ is the unique minimiser of $D$ on $\mathcal{A}_g$.
- Conversely, a minimiser $u\in C^2(\overline\Omega)\cap\mathcal{A}_g$ of $D$ solves $-\Delta u=f$, $u=g$ on $\partial\Omega$.
- (Neumann problem.) For $h\in C(\partial\Omega)$, a minimiser $u\in C^2(\overline\Omega)$ of $D_h[v]=D[v]-\oint_{\partial\Omega}hv\,ds$ over all $v\in C^1(\overline\Omega)$ solves $-\Delta u=f$, $\partial u/\partial n=h$ (a natural condition). $D_h$ is bounded below only if $\iint_\Omega f+\oint_{\partial\Omega}h=0$.
Proof. (1) $$D[u+\eta]-D[u]=\iint(\nabla u\cdot\nabla\eta-f\eta)+\tfrac12\iint|\nabla\eta|^2$$. By Green's first identity, $\iint\nabla u\cdot\nabla\eta=\oint\eta\,\partial_nu\,ds-\iint\eta\,\Delta u$, and the boundary term vanishes; so the first integral is $\iint(-\Delta u-f)\eta=0$. The remaining term is $\ge0$ and vanishes only for $\nabla\eta\equiv0$, i.e. $\eta\equiv0$ (as $\eta=0$ on $\partial\Omega$ and $\Omega$ is connected). (2) Ostrogradsky's equation (Theorem 6.8) for $F=\tfrac12(u_x^2+u_y^2)-fu$ is $-f-u_{xx}-u_{yy}=0$. (3) Now $\eta$ is arbitrary on $\partial\Omega$: the first variation is $\iint(-\Delta u-f)\eta+\oint(\partial_nu-h)\eta$; interior variations give the equation, and then arbitrary boundary values give $\partial_nu=h$. Finally $D_h[v+c]=D_h[v]-c\big(\iint f+\oint h\big)$ for constants $c$, which is unbounded below unless the bracket is $0$. ∎
A cautionary history. Riemann (1851, 1857) built his theory of complex functions on the "Dirichlet principle": since $D[v]\ge0$ when $f=0$, there must be a minimiser, and it is harmonic. In 1870 Weierstrass pointed out that a functional bounded below need not attain its infimum — his counterexample is Example 5 of §6.1. Worse, Hadamard observed (1906) that $u=\sum_{n\ge1}\frac{r^{n!}\cos(n!\,\theta)}{n^2}$ is harmonic in the unit disc and continuous on the closed disc (the series converges uniformly), yet its Dirichlet integral $\pi\sum_n\frac{n!}{n^4}$ is infinite: for its boundary values no admissible function has finite energy, and the principle is empty. Hilbert (1900–1904) showed that the principle is correct under reasonable hypotheses on $\Omega$ and $g$, and the modern formulation in the Sobolev space $H^1_0(\Omega)$ — completed with respect to the energy norm, where the Riesz representation (Lax–Milgram) theorem gives a unique weak solution for every square-integrable $f$ — is the foundation of the finite element method. Hilbert's 19th and 20th problems (1900) are about exactly these questions.
- Legendre's condition is not sufficient. For $\int_0^{2\pi}(y'^2-y^2)$, $P=2\gt0$ everywhere, yet $y\equiv0$ is not a minimum: the conjugate point $\pi$ lies inside.
- Legendre's necessary condition is $P\ge0$, not $P\gt0$. $\int_0^1y'^4\,dx$ with $y(0)=y(1)=0$: along $y\equiv0$, $P=12y'^2=0$, yet $y\equiv0$ is the global minimum.
- The sufficient conditions are not necessary. On $[0,\pi]$, $y\equiv0$ minimises $\int(y'^2-y^2)$ (Wirtinger), although $\pi$ is conjugate to $0$ and (J⁺) fails.
- Weak minimum does not imply strong minimum. $\int_0^1y'^3$: $y=x$ satisfies (L⁺), (J⁺), but $E\lt0$ for slopes $q\lt-2$ and $\inf J=-\infty$ in every strong neighbourhood.
- Convexity in $y'$ alone is not enough for a global minimum. $y'^2-y^2$ is convex in $y'$, not jointly in $(y,y')$; joint convexity is what Theorem 6.14 needs.
- A conjugate point is a property of the extremal and the starting point, not of the integrand alone. For $\int(y'^2+y^2)$ there are none; for $\int(y'^2-\omega^2y^2)$ they are at multiples of $\pi/\omega$ from $a$.
- Rayleigh–Ritz bounds are one-sided for energies and eigenvalues only. $\Lambda_k\ge\lambda_k$ and $I[u_n]\ge I[u]$; pointwise values can be too high (torsion centre value) or oscillate around the truth (Example 4).
- Trial functions must satisfy the essential conditions. For the Dirichlet problem, $u\equiv1$ gives $R=0\lt\pi^2$; it is not admissible. Natural conditions need not be satisfied.
- Without positivity there is no minimum principle. $-u''-10u=f$, $u(0)=u(1)=0$ is uniquely solvable ($10$ is not an eigenvalue $n^2\pi^2$), but $I[v]=\int(v'^2-10v^2-2fv)$ is unbounded below ($v=t\sin\pi x$, since $\pi^2\lt10$): the solution is a saddle point, and Ritz approximations lose their meaning as minimisers.
- Galerkin is not Ritz for non-symmetric operators. For $-u''+u'=f$ there is no energy, approximations need not improve monotonically, and the stiffness matrix is not symmetric.
- The Dirichlet principle can fail. The infimum may not be attained (Weierstrass), and for some continuous boundary data no admissible function has finite energy (Hadamard's series).
Exercises
- 1 For the length functional at the line $y=mx+k$, show that $\delta^2J[\eta]=\displaystyle\int\frac{\eta'^2}{(1+m^2)^{3/2}}\,dx$.
- 2 Compute $F_{y'y'}$ for $F=\sqrt{1+y'^2}/x$ ($x\gt0$) and $F=e^x\sqrt{1+y'^2}$, and state what Legendre's test says about their extremals.
- 3 Show that $y=x$ is a strict weak maximum of $\int_0^1(y'^2-y'^4)\,dx$, $y(0)=0$, $y(1)=1$.
- 4 Which integrands are jointly convex in $(y,p)$? (a) $p^2+y^2$; (b) $p^2-y^2$; (c) $\sqrt{1+p^2}$; (d) $y^2p^2$; (e) $e^p+y^2$; (f) $y\sqrt{1+p^2}$ on $y\gt0$.
- 5 Use convexity to prove that $y=x^4-1$ is the unique global minimiser of $\int_1^2y'^2x^{-3}\,dx$ with $y(1)=0$, $y(2)=15$.
- 6 Show that $y\equiv0$ is the global minimum of $\int_0^1y'^4\,dx$ ($y(0)=y(1)=0$) although $P=0$ along it.
- 7 Prove the reduced form $\delta^2J=\int(P\eta'^2+Q\eta^2)$ in detail, stating the regularity you need.
- 1 Find the point conjugate to $0$ for $\int(y'^2-9y^2)\,dx$. For which $b$ is $y\equiv0$ a strict weak minimum with $y(0)=y(b)=0$?
- 2 Show that $\int_0^b(y'^2+9y^2)\,dx$ has no conjugate points and that $y\equiv0$ is the global minimum for every $b$.
- 3 Find the point conjugate to $1$ for $\int_1^b\big(x^2y'^2-\tfrac{17}{4}y^2\big)\,dx$.
- 4 For the pencil $y=c\sin x$ of extremals of $\int(y'^2-y^2)$ through the origin, verify that $\partial y/\partial c$ solves the Jacobi equation, and find where the pencil refocuses.
- 5 On $[0,\tfrac{3\pi}{2}]$, $y=\sin x$ is the extremal of $\int(y'^2-y^2)$ with $y(0)=0$, $y(\tfrac{3\pi}2)=-1$. Show directly that it is not a minimum using $\eta=\sin\tfrac{2x}{3}$.
- 6 Use Picone's identity with $v=\cos\big(\sqrt2(x-\tfrac12)\big)$ to prove $\int_0^1\eta'^2\,dx\ge2\int_0^1\eta^2\,dx$ for $\eta(0)=\eta(1)=0$. What is the best constant?
- 7 For $0\lt b\lt\pi$, find the minimiser and the minimum value of $\int_0^b(y'^2-y^2)\,dx$ with $y(0)=0$, $y(b)=1$. What happens as $b\to\pi^-$?
- 1 Compute $E$ for $F=\sqrt{1+p^2}$, show $E\ge0$, and conclude that straight lines are strong minima of the length.
- 2 Compute $E$ for $F=p^3$ and for $F=p^4$. Which has $E\ge0$ for all $p,q$?
- 3 For $\int_0^1y'^3\,dx$, compute $J$ for the polygon through $(0,0)$, $(0.01,-0.1)$, $(1,1)$ and its maximal distance from $y=x$.
- 4 Prove that $E(x,y,p,q)=\tfrac12(q-p)^2F_{y'y'}(x,y,\xi)$ for some $\xi$ between $p$ and $q$.
- 5 Show that broken extremals of $\int(y'^2-1)^2\,dx$ with slopes $\pm1$ satisfy both Weierstrass–Erdmann corner conditions.
- 1 Write the energy functional whose minimiser solves $-\big((1+x^2)u'\big)'+2u=e^x$, $u(0)=u(1)=0$.
- 2 Find the boundary value problem solved by the minimiser of $\int_1^2(x^2u'^2+2u^2-2xu)\,dx$, $u(1)=u(2)=0$.
- 3 One-term Ritz for $-u''=x$, $u(0)=u(1)=0$, with $\varphi=x(1-x)$; compare with the exact solution at $x=\tfrac12$ and in energy.
- 4 Show that the two-term Ritz approximation with $x(1-x)$, $x^2(1-x)$ for the problem of 3 is exact. Why?
- 5 For $-y''=\lambda y$, $y(0)=0$, $y'(1)=0$, compute $R[v]$ for $v=x$ and $v=2x-x^2$ and compare with $\lambda_1$. Is $v=x$ admissible?
- 6 Estimate $\lambda_1$ for $-\big((1+x)y'\big)'=\lambda y$, $y(0)=y(1)=0$, with the trial function $\sin\pi x$.
- 7 One-term Galerkin for $-u''+u'=1$, $u(0)=u(1)=0$, with $\varphi=x(1-x)$ (verify the value in the text).
- 8 Show that $u=\tfrac14(1-x^2-y^2)$ minimises $\iint_{x^2+y^2\lt1}\big(\tfrac12|\nabla v|^2-v\big)$ among $v$ vanishing on the unit circle, and find the minimum value.
- 9 Show that $\int_0^1(u'^2-2fu)\,dx$ (no boundary conditions) is bounded below only if $\int_0^1f=0$, and that a minimiser then solves $-u''=f$, $u'(0)=u'(1)=0$. Is it unique?
- 10 Prove that the Ritz matrix $K$ is positive definite when $p\gt0$, $q\ge0$ and the $\varphi_i\in\mathcal{A}_0$ are independent. Does the proof survive if $q$ takes negative values?
E1. CSIR Part B GATE MA For $J[y]=\int_0^b\big(y'^2-4y^2+2xy\big)\,dx$, the first point conjugate to $x=0$ along any extremal is (A) $\pi/4$; (B) $\pi/2$; (C) $\pi$; (D) there is none.
E2. CSIR Part B The one-term Ritz approximation $u_1=c\,x(1-x)$ for $-u''+u=1$, $u(0)=u(1)=0$ has $c$ equal to (A) $\tfrac5{11}$; (B) $\tfrac5{22}$; (C) $\tfrac12$; (D) $\tfrac6{11}$.
E3. CSIR Part B For $J[y]=\int_0^1y'^3\,dx$, $y(0)=0$, $y(1)=1$, the extremal $y=x$ is (A) a strong minimum; (B) a weak minimum but not a strong minimum; (C) neither a weak nor a strong minimum; (D) a weak maximum.
E4. CSIR Part C Consider $J[y]=\int_0^b(y'^2-y^2)\,dx$ with $y(0)=y(b)=0$. Which are true? (One or more options may be correct.) (A) If $b=3$, then $y\equiv0$ is a strict minimum. (B) If $b=4$, there is an admissible $y$ with $J[y]\lt0$. (C) If $b=\pi$, then $J\ge0$ on admissible functions, and $J=0$ for some $y\not\equiv0$. (D) Legendre's condition fails when $b\gt\pi$.
E5. CSIR Part C Let $q,f\in C[0,1]$ and $I[v]=\int_0^1\big(v'^2+qv^2-2fv\big)\,dx$. Which are true? (A) If $q\ge0$, the minimiser of $I$ over $v\in C^2$ with $v(0)=v(1)=0$ solves $-u''+qu=f$. (B) If $q\ge0$, the minimiser over $v$ with only $v(0)=0$ prescribed satisfies $u'(1)=0$. (C) If $q\equiv-10$ and $f\equiv0$, then $I$ is bounded below on $\{v:v(0)=v(1)=0\}$. (D) Ritz approximations from nested trial spaces have non-increasing energies $I[u_n]$.
E6. CSIR Part C For $-y''=\lambda y$, $y(0)=y(1)=0$ (eigenvalues $n^2\pi^2$), which are true? (A) Every $C^1$ trial function $v\ne0$ with $v(0)=v(1)=0$ has $\int v'^2/\int v^2\ge\pi^2$. (B) The trial function $x(1-x)$ gives the value $10$. (C) Rayleigh–Ritz with the basis $\{x(1-x),\,x^2(1-x)^2\}$ gives an approximation of $\lambda_1$ below $\pi^2$. (D) With the basis $\{x(1-x),\,x^2(1-x)\}$ the two Ritz values are $10$ and $42$, and $42\ge\lambda_2$.
Suggested plan — 3 lectures × 60 min.
- Second variation, Legendre, convexity. Recall $f''$ in calculus; Definition 6.6 and Theorem 6.12. Draw the thin bump of the Legendre proof and compute $\int\eta_h'^2$ with the class. Theorem 6.14 and Example 1. Close with the question: $\int_0^{2\pi}(y'^2-y^2)$ satisfies Legendre — is $y\equiv0$ a minimum?
- Jacobi and Weierstrass. Answer the question with Interactive 1 (sweep $b$ through $\pi$). Theorems 6.15–6.17 (prove Picone's identity on the board), Example 2. Then the excess function, Theorem 6.18 and Example 3 (draw the steep "needle"); the summary rule-cards.
- Variational methods. Theorem 6.19 (the energy identity is a two-line computation), Definition 6.9 and Theorem 6.20 (stress Galerkin orthogonality as a projection), Example 4 by hand and then Interactive 2 (smooth load vs step load). Theorem 6.21 with Example 5; Theorem 6.22, the history of the Dirichlet principle and Example 6. Do E4–E6 live.
Discussion questions. (1) Why can a second variation be positive for every short bump and still negative for a long one? Relate this to conjugate points and to the size of the interval. (2) Galerkin orthogonality says $u_n$ is the energy-orthogonal projection of $u$ onto $V_n$. What does this mean geometrically, and why does it make the energy (not the pointwise error) the natural measure of accuracy? (3) An engineer uses Rayleigh–Ritz to estimate the lowest vibration frequency of a bridge and the stiffness of a beam. Which estimates are on the safe side?
Misconceptions to watch for. Treating Legendre's condition as sufficient; forgetting that (L⁺), (J⁺) give only a weak minimum; computing $Q$ as $F_{yy}$ when $F_{yy'}$ is not constant; using trial functions that violate essential conditions or imposing natural ones unnecessarily; believing Ritz values bound pointwise values; applying the energy principle when $q$ is very negative.
Suggested homework. I(2, 3, 4, 5), II(1, 3, 5, 6, 7), III(1, 2, 3), IV(1, 3, 5, 6, 7, 8, 9); Exam corner E1, E2, E5, E6.
Research question. How much load can a tapered column carry before it buckles, and how quickly do Rayleigh–Ritz estimates of that load converge?
Why it matters. Euler's buckling load $P=\pi^2EI/L^2$ (1744) — derived in the same treatise that founded the calculus of variations — governs the design of columns, aircraft struts, wind-turbine towers and bicycle frames. Real members are often tapered to save material, and then no closed formula exists. Design codes in Europe, India and the USA rely on energy (Rayleigh–Ritz) estimates, and finite element software is a refined version of the same idea.
What to do.
- A pin-ended column of length $1$ with bending stiffness $EI(x)=1+x$ buckles under the load $P$ when $y''+\dfrac{P}{1+x}y=0$, $y(0)=y(1)=0$ has a nonzero solution. Recognise this as a Sturm–Liouville problem $-y''=P\,r(x)\,y$ with weight $r=1/(1+x)$, and write its Rayleigh quotient.
- Compute the one-term Rayleigh–Ritz estimate with $\sin\pi x$ (numerical integration is fine), then the two-term estimate with $\sin\pi x,\sin2\pi x$ by solving $\det(K-PM)=0$.
- Find the critical load accurately by shooting: solve the ODE with $y(0)=0$, $y'(0)=1$ and adjust $P$ until $y(1)=0$. Tabulate the errors of the two estimates.
- Compare with the uniform columns $EI\equiv1$ and $EI\equiv2$. Is the tapered column closer to the weaker or the stronger one? Explain with the Rayleigh quotient where the stiffness "matters most".
Tools: Python with numpy and scipy (quad, solve_ivp, brentq, scipy.linalg.eigh), or any spreadsheet with a Simpson-rule column.
Hand in: A table of the estimates and errors, a plot of the buckled shape (the Ritz eigenvector next to the shooting solution), and a short paragraph on why every estimate is on the unsafe side of the true load and what an engineer must do about it.
Going further. Repeat with the energy form $P\le\int EI\,y''^2/\int y'^2$ (valid for other end conditions too) and compare; read about how the finite element method uses piecewise linear "hat" trial functions and why its energy error is $O(h^2)$.
🧪 Self-check quiz
Q1. Legendre's necessary condition for an extremal $y$ to be a weak minimum of $\int F(x,y,y')\,dx$ is
Q2. For $J=\int_0^b(y'^2-25y^2)\,dx$, the first point conjugate to $0$ is
Q3. The Weierstrass excess function is
Q4. The solution of $-u''=f$, $u(0)=u(1)=0$ is the minimiser over $v(0)=v(1)=0$ of
Q5. Rayleigh–Ritz approximations of the first eigenvalue of a regular Sturm–Liouville problem are