6.1 Functionals and the Euler–Lagrange Equation
Functionals and their first variation, the fundamental lemma, the Euler–Lagrange equation and its first integrals; shortest paths, the brachistochrone and soap films; several functions, higher derivatives and several variables; natural boundary conditions, transversality and isoperimetric problems with Lagrange multipliers.
Calculus finds the points at which a function of finitely many variables is largest or smallest. Many questions in geometry, physics and engineering ask instead for the best function or the best curve: the curve of shortest length joining two points, the profile down which a bead slides fastest, the shape of a soap film stretched between two rings, the path a light ray takes through a medium of varying refractive index, the shape of a hanging chain, the deflection of a loaded beam. In each case a number — a length, a time, an area, an energy — is assigned to every admissible curve, and we want the curve that makes this number extreme. Such a "function of a function" is called a functional, and the calculus of variations is the theory of its extrema.
The central result is Euler's and Lagrange's discovery (1744–1755) that an extremising curve must satisfy a differential equation, the Euler–Lagrange equation. It turns an optimisation problem over an infinite-dimensional set of curves into a boundary value problem for an ODE — exactly the kind of problem studied in Chapters 1–2. This section derives the equation with full proofs, solves the classical problems, extends the method to several unknown functions, higher derivatives and several independent variables, and handles free end points and integral constraints. CSIR NET asks for extremals of given functionals, for first integrals, for natural boundary conditions and for isoperimetric extremals in Part B, and tests the logical status of the Euler–Lagrange equation (necessary, not sufficient; existence can fail) in Part C. The question of when an extremal really is a minimum is the subject of §6.2.
Functionals, Variations and the Fundamental Lemma
Throughout, $[a,b]$ is a compact interval, $C^1[a,b]$ the space of continuously differentiable functions on it, and $F(x,y,p)$ a given function of three variables, the Lagrangian or integrand. We always assume $F$ is of class $C^2$ on an open set containing all the points $(x,y(x),y'(x))$ that occur.
A functional is a map $J:\mathcal{A}\to\mathbb{R}$ defined on a set $\mathcal{A}$ of functions. The basic functional of the calculus of variations is $$J[y]=\int_a^bF\big(x,y(x),y'(x)\big)\,dx,$$ defined on the class of admissible functions $$\mathcal{A}=\{y\in C^1[a,b]:\ y(a)=A,\ y(b)=B\}$$ (the fixed end-point problem). Functions $\eta\in C^1[a,b]$ with $\eta(a)=\eta(b)=0$ are called variations (admissible increments): if $y\in\mathcal{A}$ then $y+\varepsilon\eta\in\mathcal{A}$ for every real $\varepsilon$.
Examples: the length $\int_a^b\sqrt{1+y'^2}\,dx$ of the graph of $y$; the area $2\pi\int_a^by\sqrt{1+y'^2}\,dx$ of the surface obtained by rotating it about the $x$-axis; the "energy" $\int_a^b(y'^2+12xy)\,dx$.
To speak of local extrema we need a notion of nearness of functions. Two natural norms on $C^1[a,b]$ give two different notions: $$\|y\|_0=\max_{[a,b]}|y|,\qquad \|y\|_1=\max_{[a,b]}|y|+\max_{[a,b]}|y'|.$$
A function $y_0\in\mathcal{A}$ is a weak local minimum of $J$ if there is $\delta\gt0$ such that $J[y]\ge J[y_0]$ for all $y\in\mathcal{A}$ with $\|y-y_0\|_1\lt\delta$ (curves close to $y_0$ and with close slopes). It is a strong local minimum if the same holds for all $y\in\mathcal{A}$ with $\|y-y_0\|_0\lt\delta$ (close curves, slopes arbitrary). Maxima are defined by reversing the inequality, and "extremum" means minimum or maximum.
Since $\|y\|_0\le\|y\|_1$, a $\|\cdot\|_1$-ball is contained in the $\|\cdot\|_0$-ball of the same radius, so every strong extremum is a weak extremum; the converse fails (§6.2). Necessary conditions proved for weak extrema therefore apply to both kinds.
Proof. Fix a variation $\eta\ne0$ and put $\varphi(\varepsilon)=J[y+\varepsilon\eta]$. Since $\|(y+\varepsilon\eta)-y\|_1=|\varepsilon|\,\|\eta\|_1$, the functions $y+\varepsilon\eta$ lie in the $\delta$-neighbourhood of Definition 6.2 for $|\varepsilon|\lt\delta/\|\eta\|_1$, and they are admissible. Hence the real function $\varphi$ has a local extremum at $\varepsilon=0$. It is differentiable, so $\varphi'(0)=0$ by Fermat's theorem; and $\varphi'(0)=\delta J[y;\eta]$ by Definition 6.3. ∎
To turn "$\int(F_y\eta+F_{y'}\eta')=0$ for every $\eta$" into a pointwise equation we need two lemmas. They are the engine of the whole subject.
Proof. Suppose $M(x_0)\ne0$ for some $x_0$; replacing $M$ by $-M$ we may assume $M(x_0)\gt0$, and by continuity we may take $x_0\in(a,b)$. Again by continuity there is an interval $[x_1,x_2]\subset(a,b)$ around $x_0$ on which $M\gt0$. Let $$\eta(x)=\begin{cases}(x-x_1)^2(x_2-x)^2,& x_1\le x\le x_2,\\0,&\text{otherwise}.\end{cases}$$ Both one-sided derivatives at $x_1$ and $x_2$ are $0$, so $\eta\in C^1[a,b]$, $\eta(a)=\eta(b)=0$, and $\eta\gt0$ on $(x_1,x_2)$. Then $\int_a^bM\eta\,dx=\int_{x_1}^{x_2}M\eta\,dx\gt0$, because the integrand is continuous, nonnegative and positive inside — a contradiction. For $C^k$ use the exponent $k+1$ instead of $2$; for $C^\infty$ use $\eta=\exp\!\big(-1/((x-x_1)(x_2-x))\big)$ on $(x_1,x_2)$ and $0$ outside. ∎
Proof. Let $c=\frac1{b-a}\int_a^bM\,dx$ be the mean value of $M$ and choose the particular variation $$\eta(x)=\int_a^x\big(M(t)-c\big)\,dt.$$ It is $C^1$ with $\eta'=M-c$, $\eta(a)=0$, and $\eta(b)=\int_a^bM-c(b-a)=0$. Using the hypothesis and $\int_a^b\eta'\,dx=\eta(b)-\eta(a)=0$, $$\int_a^b(M-c)^2\,dx=\int_a^b(M-c)\,\eta'\,dx=\int_a^bM\eta'\,dx-c\int_a^b\eta'\,dx=0.$$ A continuous nonnegative function with zero integral vanishes identically, so $M\equiv c$. ∎
The Euler–Lagrange Equation and Its First Integrals
Let $F\in C^2$ and let $y\in\mathcal{A}$ be a weak local extremum of $J[y]=\int_a^bF(x,y,y')\,dx$ on the fixed end-point class $\mathcal{A}$. Then:
- (Integrated form.) There is a constant $c$ such that for all $x\in[a,b]$ $$F_{y'}\big(x,y(x),y'(x)\big)=\int_a^xF_y\big(t,y(t),y'(t)\big)\,dt+c.$$
- (Euler–Lagrange equation.) Consequently $x\mapsto F_{y'}(x,y,y')$ is $C^1$ and $$F_y-\frac{d}{dx}F_{y'}=0\quad\text{on }[a,b].$$
- (Regularity, Hilbert.) If $F_{y'y'}\big(x,y(x),y'(x)\big)\ne0$ for all $x$, then $y\in C^2[a,b]$, and the equation can be written in the expanded second-order form $$F_{y'y'}\,y''+F_{yy'}\,y'+F_{xy'}-F_y=0.$$
Proof. (1) By Theorem 6.1, $\int_a^b(F_y\eta+F_{y'}\eta')\,dx=0$ for every variation $\eta$. Put $G(x)=\int_a^xF_y\,dt$, a $C^1$ function with $G'=F_y$. Integration by parts gives $\int_a^bF_y\eta\,dx=\big[G\eta\big]_a^b-\int_a^bG\eta'\,dx=-\int_a^bG\eta'\,dx$, because $\eta(a)=\eta(b)=0$. Hence $\int_a^b\big(F_{y'}-G\big)\eta'\,dx=0$ for every variation, and $F_{y'}-G$ is continuous. By du Bois-Reymond's lemma (Theorem 6.3), $F_{y'}-G\equiv c$.
(2) The right-hand side $G+c$ of (1) is $C^1$ with derivative $F_y$; therefore so is the left-hand side, and differentiating gives $\frac{d}{dx}F_{y'}=F_y$. Note that we did not assume $y\in C^2$: the existence of $\frac{d}{dx}F_{y'}$ is part of the conclusion.
(3) Fix $x_0$ and consider the equation $\Phi(x,p)=F_{y'}\big(x,y(x),p\big)-G(x)-c=0$. It is satisfied by $p=y'(x)$, and $\partial\Phi/\partial p=F_{y'y'}\ne0$ at $(x_0,y'(x_0))$. Since $\Phi$ is $C^1$ in $(x,p)$ (here $y\in C^1$, $G\in C^1$, $F\in C^2$), the implicit function theorem gives a $C^1$ function $p(x)$ near $x_0$ solving $\Phi=0$, and by uniqueness in the implicit function theorem (and continuity of $y'$) $y'=p$ near $x_0$. So $y'$ is $C^1$, i.e. $y\in C^2$, and expanding $\frac{d}{dx}F_{y'}(x,y,y')$ by the chain rule gives the second-order form. ∎
Two facts must be kept apart. Theorem 6.4 is a necessary condition: an extremum (if one exists) is an extremal. It says nothing about whether an extremal is an extremum, or whether an extremum exists at all — Example 5 and the ⚠️ box below show that both can fail. Sufficient conditions are developed in §6.2; for many problems a direct argument (as in Example 1) settles the matter.
When the integrand does not involve one of its three arguments the Euler–Lagrange equation can be integrated once. These first integrals are the most-used tools for computing extremals.
Let $y\in C^2$ be an extremal of $\int F\,dx$.
- $F=F(x,y')$ does not contain $y$: $F_{y'}(x,y')=c$ (a first-order equation).
- $F=F(y,y')$ does not contain $x$ (Beltrami identity): $F-y'F_{y'}=c$. Conversely, a $C^2$ solution of $F-y'F_{y'}=c$ satisfies the Euler–Lagrange equation on every interval on which $y'\ne0$.
- $F=F(x,y)$ does not contain $y'$: the Euler–Lagrange equation is $F_y(x,y)=0$, an implicit equation for $y$, not a differential equation; in general its solutions do not meet two prescribed end conditions.
- $F$ is linear in $y'$, $F=M(x,y)+N(x,y)\,y'$: the Euler–Lagrange equation reduces to $M_y-N_x=0$, again not a differential equation. If $M_y\equiv N_x$ on a rectangle (or any simply connected domain), then $J[y]$ depends only on the end points: every admissible curve is an extremal.
- $F=F(y')$ only: $y''\,F''(y')=0$; if $F''\ne0$, the extremals are the straight lines.
Proof. (1) The equation is $\frac{d}{dx}F_{y'}=F_y=0$. (2) For any $C^2$ function $y$, by the chain rule, $$\frac{d}{dx}\big(F-y'F_{y'}\big)=F_x+F_yy'+F_{y'}y''-y''F_{y'}-y'\frac{d}{dx}F_{y'},$$ that is, $$\frac{d}{dx}\big(F-y'F_{y'}\big)=F_x+y'\Big(F_y-\frac{d}{dx}F_{y'}\Big).$$ If $F_x=0$, the left side vanishes along an extremal, which gives the identity. Conversely, if $F-y'F_{y'}$ is constant, then $y'\big(F_y-\frac{d}{dx}F_{y'}\big)=0$, and where $y'\ne0$ the Euler–Lagrange equation follows. (3) is immediate. (4) $F_y=M_y+N_yy'$ and $\frac{d}{dx}F_{y'}=\frac{d}{dx}N(x,y(x))=N_x+N_yy'$, so the equation is $M_y-N_x=0$. If $M_y=N_x$ throughout a simply connected domain, the form $M\,dx+N\,dy$ is exact, $M\,dx+N\,dy=d\Phi$, and $J[y]=\int_a^b\frac{d}{dx}\Phi(x,y(x))\,dx=\Phi(b,B)-\Phi(a,A)$. (5) Here $\frac{d}{dx}F'(y')=F''(y')\,y''=0$; if $F''\ne0$ then $y''=0$. ∎
- Read off $F(x,y,p)$ and look for a missing variable: no $y$ ⇒ $F_{y'}=c$; no $x$ ⇒ Beltrami $F-y'F_{y'}=c$ (watch for spurious constant solutions); no $y'$ or $F$ linear in $y'$ ⇒ not a differential equation.
- Otherwise write $F_y-\frac{d}{dx}F_{y'}=0$; for quadratic integrands it is a linear second-order ODE — use Chapter 2. Terms such as $2yy'$ or $g(x)y'+g'(x)y$ are exact derivatives and drop out.
- Impose the boundary conditions: fixed values; natural conditions $F_{y'}=0$ at free ends; transversality on a curve; and, with a constraint $\int G=\ell$, replace $F$ by $F+\lambda G$ and use the constraint to find $\lambda$.
- Decide whether the candidate is a minimum: direct expansion $J[y_0+\eta]-J[y_0]$, convexity, or the Legendre and Jacobi conditions of §6.2.
Several Functions, Higher Derivatives and Several Variables
Proof. Vary one function at a time: for fixed $i$ and a variation $\eta$, the function $\varepsilon\mapsto J[y_1,\dots,y_i+\varepsilon\eta,\dots,y_n]$ has a local extremum at $0$, and the proof of Theorem 6.4 applies verbatim to the $i$-th component. The first integral follows from $\frac{d}{dx}\big(F-\sum y_i'F_{y_i'}\big)=F_x+\sum_iy_i'\big(F_{y_i}-\frac{d}{dx}F_{y_i'}\big)$, exactly as in Theorem 6.5(2). ∎
With $x$ renamed $t$ and $F=L=T-V$ (kinetic minus potential energy) these are Lagrange's equations of motion, and the first integral is the energy; Hamilton's principle (§8.2) states that the motion of a mechanical system makes $\int L\,dt$ stationary. Curves that are not graphs are handled the same way in parametric form $\big(x(t),y(t)\big)$, with $J=\int G(x,y,\dot x,\dot y)\,dt$ and two Euler–Lagrange equations; for the answer to be independent of the parametrisation, $G$ must be positively homogeneous of degree $1$ in $(\dot x,\dot y)$, as the length integrand $\sqrt{\dot x^2+\dot y^2}$ is.
Proof (for $n=2$; the general case is identical). Admissible variations satisfy $\eta(a)=\eta'(a)=\eta(b)=\eta'(b)=0$, and $\delta J=\int_a^b\big(F_y\eta+F_{y'}\eta'+F_{y''}\eta''\big)dx=0$. Integrate the second term by parts once and the third twice: $$\int_a^bF_{y'}\eta'=\Big[F_{y'}\eta\Big]_a^b-\int_a^b\big(F_{y'}\big)'\eta,$$ $$\int_a^bF_{y''}\eta''=\Big[F_{y''}\eta'-\big(F_{y''}\big)'\eta\Big]_a^b+\int_a^b\big(F_{y''}\big)''\eta.$$ All boundary terms vanish, so $\int_a^b\big(F_y-(F_{y'})'+(F_{y''})''\big)\eta\,dx=0$ for every such $\eta$. The fundamental lemma (Theorem 6.2, in its $C^2$ form with the bump vanishing to second order) gives the equation. ∎
Proof. For $\eta\in C^1(\overline\Omega)$ vanishing on $\partial\Omega$, $\delta J=\iint_\Omega\big(F_u\eta+F_{u_x}\eta_x+F_{u_y}\eta_y\big)=0$. The divergence theorem applied to the vector field $\eta\,(F_{u_x},F_{u_y})$ gives $$\iint_\Omega\big(F_{u_x}\eta_x+F_{u_y}\eta_y\big)=\oint_{\partial\Omega}\eta\,\big(F_{u_x}n_1+F_{u_y}n_2\big)\,ds-\iint_\Omega\eta\Big(\frac{\partial F_{u_x}}{\partial x}+\frac{\partial F_{u_y}}{\partial y}\Big),$$ and the boundary integral is $0$. So $\iint_\Omega\big(F_u-\partial_xF_{u_x}-\partial_yF_{u_y}\big)\eta=0$ for all such $\eta$, and the two-dimensional fundamental lemma (same proof, with a bump $\big(r_0^2-|{\bf x}-{\bf x}_0|^2\big)^2$ on a small disc) finishes. ∎
Natural Boundary Conditions and Variable End Points
If the value of $y$ at an end point is not prescribed, the class of competitors is larger, so the minimiser must satisfy an extra condition there. Remarkably, this condition comes out of the variational principle itself.
Let $J[y]=\displaystyle\int_a^bF(x,y,y')\,dx+\Gamma\big(y(b)\big)$ with $F\in C^2$, $\Gamma\in C^1$, on the class $\{y\in C^1[a,b]: y(a)=A\}$ ($y(b)$ free). If $y\in C^2$ is a weak extremum, then $y$ satisfies the Euler–Lagrange equation and the natural boundary condition $$F_{y'}\big(b,y(b),y'(b)\big)+\Gamma'\big(y(b)\big)=0.$$ In particular, without the boundary term ($\Gamma=0$): $F_{y'}=0$ at a free end. The same holds at $x=a$ (with $+\Gamma'$ replaced by $-\Gamma'$ for a term $\Gamma(y(a))$). For the Euler–Poisson problem with $y(b)$ and $y'(b)$ both free the natural conditions are $F_{y''}=0$ and $F_{y'}-\frac{d}{dx}F_{y''}=0$ at $b$.
Proof. Now every $\eta\in C^1$ with $\eta(a)=0$ is an admissible variation, and $$\delta J=\int_a^b\Big(F_y-\frac{d}{dx}F_{y'}\Big)\eta\,dx+F_{y'}(b)\,\eta(b)+\Gamma'(y(b))\,\eta(b)=0,$$ after integrating $F_{y'}\eta'$ by parts (the term at $a$ vanishes). First take variations with $\eta(b)=0$ as well: the fundamental lemma gives the Euler–Lagrange equation. Then the integral is $0$ for every $\eta$, and choosing one with $\eta(b)=1$ gives the stated condition. The Euler–Poisson case is the same computation with the two boundary terms of the proof of Theorem 6.7. ∎
For the "energy" $\int(py'^2+qy^2)$ the natural condition $F_{y'}=2py'=0$ is the Neumann condition, and a boundary term $\Gamma=\sigma y(b)^2$ produces the Robin condition $p(b)y'(b)+\sigma y(b)=0$: the Neumann and Robin conditions of the Sturm–Liouville problems of §2.3 are exactly the ones a variational principle imposes "for free". Engineers call prescribed conditions essential and the ones that come from minimisation natural; the distinction is crucial in the Rayleigh–Ritz method of §6.2.
Let the left end $(a,A)$ be fixed and let the right end be required to lie on the curve $y=\varphi(x)$, $\varphi\in C^1$, the abscissa $b$ being free. If $y\in C^2$ is an extremum (among curves $y$ on $[a,b]$ with $y(b)=\varphi(b)$, $b$ varying), then $y$ is an extremal and at $x=b$ $$F+\big(\varphi'-y'\big)F_{y'}=0.$$
- End point on a vertical line $x=b$ (only $y(b)$ free): $F_{y'}=0$, the natural condition.
- End point on a horizontal line ($\varphi'=0$): $F-y'F_{y'}=0$.
- For $F=f(x,y)\sqrt{1+y'^2}$ with $f\ne0$: $y'(b)\,\varphi'(b)=-1$, i.e. the extremal meets the curve orthogonally.
Proof (sketch). Extend $y$ to a $C^2$ function slightly beyond $b$ and consider competitors $y+\varepsilon\eta$ on $[a,b(\varepsilon)]$ with $\eta(a)=0$, $b(0)=b$ and $y(b(\varepsilon))+\varepsilon\eta(b(\varepsilon))=\varphi(b(\varepsilon))$. Differentiating $J(\varepsilon)=\int_a^{b(\varepsilon)}F(x,y+\varepsilon\eta,y'+\varepsilon\eta')\,dx$ at $\varepsilon=0$ and integrating by parts, $$J'(0)=F\big|_{b}\,b'(0)+\int_a^b\Big(F_y-\frac{d}{dx}F_{y'}\Big)\eta\,dx+F_{y'}\big|_b\,\eta(b).$$ Variations with $b'(0)=0$, $\eta(b)=0$ give the Euler–Lagrange equation. Differentiating the end condition gives $y'(b)b'(0)+\eta(b)=\varphi'(b)b'(0)$, so $\eta(b)=(\varphi'-y')b'(0)$, and $J'(0)=\big[F+(\varphi'-y')F_{y'}\big]_b\,b'(0)=0$ with $b'(0)$ arbitrary. For $F=fw$, $w=\sqrt{1+p^2}$, $F_{y'}=fp/w$, the condition reads $f\big(w^2+(\varphi'-p)p\big)/w=f(1+\varphi'p)/w=0$. ∎
For instance, the shortest path from a point to a curve is the normal line through the point, and the fastest descent from the origin to a vertical line $x=b$ is the cycloid that meets the line horizontally (Exercise III.8). A free initial point is handled the same way.
Isoperimetric Problems and Lagrange Multipliers
Queen Dido, according to Virgil, was allowed as much land as she could enclose with an ox-hide; cutting it into a thin strip, she had to choose the shape of greatest area for a curve of given length. Problems in which a functional is extremised subject to another functional being held fixed are called isoperimetric ("same perimeter").
Proof. Write $E_F$ and $E_G$ for the Euler–Lagrange expressions $F_y-\frac{d}{dx}F_{y'}$, $G_y-\frac{d}{dx}G_{y'}$ evaluated along $y$; both are continuous, and for every variation $\eta$, $\delta J[y;\eta]=\int E_F\eta$, $\delta K[y;\eta]=\int E_G\eta$ (integration by parts). Since $y$ is not an extremal of $K$, $E_G\not\equiv0$, and by the fundamental lemma there is a variation $\eta_2$ with $\delta K[y;\eta_2]\ne0$. Fix it, take an arbitrary variation $\eta_1$, and set $$\Phi(\varepsilon_1,\varepsilon_2)=J[y+\varepsilon_1\eta_1+\varepsilon_2\eta_2],\qquad\Psi(\varepsilon_1,\varepsilon_2)=K[y+\varepsilon_1\eta_1+\varepsilon_2\eta_2]-\ell.$$ Both are $C^1$ near $(0,0)$, $\Psi(0,0)=0$ and $\partial_{\varepsilon_2}\Psi(0,0)=\delta K[y;\eta_2]\ne0$. By the implicit function theorem the constraint $\Psi=0$ defines a $C^1$ curve $\varepsilon_2=\gamma(\varepsilon_1)$ through $0$, and along it the functions $y+\varepsilon_1\eta_1+\gamma(\varepsilon_1)\eta_2$ are admissible and close to $y$ in $\|\cdot\|_1$. So $\varepsilon_1\mapsto\Phi(\varepsilon_1,\gamma(\varepsilon_1))$ has a local extremum at $0$, and the Lagrange multiplier rule for functions of two variables gives a $\lambda$ with $\nabla\Phi(0,0)+\lambda\nabla\Psi(0,0)=0$. Its second component shows $\lambda=-\delta J[y;\eta_2]/\delta K[y;\eta_2]$ — a number that does not depend on $\eta_1$. The first component is $\delta J[y;\eta_1]+\lambda\,\delta K[y;\eta_1]=\int(E_F+\lambda E_G)\eta_1=0$, valid for every $\eta_1$; the fundamental lemma gives $E_F+\lambda E_G\equiv0$. ∎
With several constraints $\int G_k=\ell_k$ one uses $F+\sum\lambda_kG_k$ (and the regularity hypothesis becomes linear independence of the $E_{G_k}$). For a holonomic constraint $\phi(x,y,z)=0$ the multiplier becomes a function $\lambda(x)$, and one writes the Euler–Lagrange equations of $F+\lambda(x)\phi$ for $y$ and $z$; this is how geodesics on a surface are found (great circles on a sphere).
- An extremal need not be an extremum. $y\equiv0$ is an extremal of $\int_0^{2\pi}(y'^2-y^2)\,dx$ with $y(0)=y(2\pi)=0$, but $y=t\sin(x/2)$ gives $J=t^2\big(\tfrac{\pi}{4}-\pi\big)\to-\infty$: no minimum at all. The deep catenary is an extremal that is not a minimum either.
- An extremum need not exist even when $J$ is bounded below and $F$ is smooth: Weierstrass's $\int_{-1}^1x^2y'^2\,dx$ (Example 5); the soap film when $h/a\lt1.5089$.
- Extremals with given end points may not exist or may not be unique. $\int_0^\pi(y'^2-y^2)$: no extremal with $y(0)=0$, $y(\pi)=1$; infinitely many with $y(\pi)=0$.
- Integrands linear in $y'$ give no differential equation. $\int_0^1(y^2+2xyy')\,dx=y(1)^2$ for every admissible curve: $M_y-N_x=2y-2y=0$ identically, so every curve is an extremal and $J$ is constant.
- Exact derivatives do not change the extremals. Adding $2yy'=(y^2)'$ or $\frac{d}{dx}\Phi(x,y)$ to $F$ changes $J$ by a constant (fixed ends) and leaves the Euler–Lagrange equation unchanged — but it does change natural boundary conditions at a free end.
- Beltrami's identity can introduce spurious solutions. For $F=y\sqrt{1+y'^2}$, $y\equiv0$ satisfies $F-y'F_{y'}=y/\sqrt{1+y'^2}=0$, but $F_y-\frac{d}{dx}F_{y'}=1\ne0$ there. Beltrami is equivalent to Euler–Lagrange only where $y'\ne0$ (and only when $F_x=0$).
- Only free ends get natural conditions. Impose $F_{y'}=0$ at an end where $y$ is prescribed and you over-determine the problem; forget it at a free end and you under-determine it.
- The multiplier rule needs a regular constraint. If $y$ is an extremal of $K$ the rule can fail: with $K[y]=\int_0^1\sqrt{1+y'^2}=1$, $y(0)=y(1)=0$, the only admissible curve is $y\equiv0$; it minimises (and maximises) $J=\int_0^1y\,dx$ trivially, although it is not an extremal of $J+\lambda K$ for any $\lambda$ (the Euler–Lagrange expression of $y+\lambda\sqrt{1+y'^2}$ at $y\equiv0$ is $1$).
- Euler–Poisson needs twice as many conditions. For $\int F(x,y,y',y'')$ prescribe $y$ and $y'$ at both ends (or use the two natural conditions); two conditions in total do not determine a solution of the fourth-order equation.
- The fastest path is not the shortest, and may dip below the target. The brachistochrone to $B$ with $x_B/y_B\gt\pi/2$ goes below $B$.
- A minimiser among smooth curves may be "broken". For $\int_0^1(y'^2-1)^2\,dx$, $y(0)=y(1)=0$, the smooth extremal $y\equiv0$ gives $J=1$, but the zigzag with slopes $\pm1$ gives $J=0$, and $C^1$ curves come arbitrarily close to $0$ (Exercise II.7).
Exercises
- 1 $\int_0^1(y'^2+2xy)\,dx$, $y(0)=0$, $y(1)=1$.
- 2 $\int_0^{\pi/2}(y'^2-y^2)\,dx$, $y(0)=0$, $y(\pi/2)=1$.
- 3 $\int_0^1\big(y'^2+y^2+2ye^x\big)\,dx$, $y(0)=0$, $y(1)=e/2$.
- 4 $\int_1^2(x^2y'^2+6y^2)\,dx$, $y(1)=1$, $y(2)=4$.
- 5 $\int_1^2\dfrac{y'^2}{x^3}\,dx$, $y(1)=0$, $y(2)=15$.
- 6 Show that every $C^1$ function is an extremal of $\int_0^1(y^2+2xyy')\,dx$, and compute $J[y]$ in terms of the end values. Why is there no contradiction with Theorem 6.4?
- 7 $\int_0^1(xy'+y'^2)\,dx$, $y(0)=y(1)=0$.
- 8 Prove that the extremals of $\int f(x,y)\sqrt{1+y'^2}\,dx$ ($f\gt0$) satisfy $f\,y''=(1+y'^2)\big(f_y-y'f_x\big)$.
- 1 Using cylindrical coordinates, show that the geodesics (shortest curves) on the cylinder $x^2+y^2=R^2$, written as $z=z(\theta)$, are helices $z=k\theta+m$.
- 2 Find the extremals of $\int\dfrac{\sqrt{1+y'^2}}{y}\,dx$ ($y\gt0$), and the one through $(-1,\sqrt3)$ and $(3,\sqrt3)$.
- 3 Show that the extremals of $\int\sqrt y\,\sqrt{1+y'^2}\,dx$ are parabolas $y=c^2+\dfrac{(x-d)^2}{4c^2}$ (the paths of projectiles with given energy).
- 4 For the brachistochrone to $B=(\pi a,2a)$, show that the descent time along the cycloid is $\pi\sqrt{a/g}$ and along the straight line $\sqrt{(\pi^2+4)a/g}$. What is their ratio?
- 5 Rings of radius $h=2$ at $x=\pm1$: find the two catenaries numerically, and decide which of the three candidates (two catenoids, two discs) has the least area.
- 6 Light travels in a layered medium with refractive index $n(x)$ (speed $c_0/n$). Show that along a ray $n(x)\sin\psi=$ const, where $\psi$ is the angle between the ray and the $x$-axis (Snell's law).
- 7 For $J[y]=\int_0^1(y'^2-1)^2\,dx$, $y(0)=y(1)=0$: show that $y\equiv0$ is the only extremal in $\mathcal{A}$ with $J=1$, but that $\inf_{\mathcal{A}}J=0$.
- 1 Find the extremals of $\int_0^{\pi/2}(y'^2+z'^2+2yz)\,dx$ with $y(0)=z(0)=0$, $y(\pi/2)=1$, $z(\pi/2)=-1$.
- 2 A clamped beam under uniform load: minimise $\int_0^1(y''^2-48y)\,dx$ with $y=y'=0$ at both ends.
- 3 Find the extremal of $\int_0^{\pi/2}(y''^2-y^2)\,dx$ with $y(0)=1$, $y'(0)=0$, $y(\pi/2)=0$, $y'(\pi/2)=-1$, and show it is unique.
- 4 Minimise $\int_0^1(y'^2+2y)\,dx$ with $y(0)=0$ and $y(1)$ free.
- 5 Find the extremal of $\int_0^1(y'^2+y^2-2xy)\,dx$ when both end values are free.
- 6 Minimise $\int_0^1y'^2\,dx+y(1)^2$ with $y(0)=1$, $y(1)$ free. Which kind of boundary condition appears at $x=1$?
- 7 Use the transversality condition to find the shortest path from the origin to the line $y=2-x$.
- 8 Find the curve of fastest descent from the origin (start at rest) to the vertical line $x=b$, and the descent time.
- 9 Write the Ostrogradsky equations for (a) $\iint(u_x^2+u_y^2+2fu)$; (b) $\iint(u_x^2-u_y^2)$; (c) $\iint\sqrt{1+u_x^2+u_y^2}$.
- 1 Find the extremal of $\int_0^1y'^2\,dx$ subject to $\int_0^1y\,dx=1$, $y(0)=y(1)=0$, and the value of $J$ on it.
- 2 Minimise $\int_0^1y'^2\,dx$ subject to $\int_0^1y^2\,dx=1$, $y(0)=0$, $y(1)$ free.
- 3 A chain of length $2\sinh1$ hangs from $(\pm1,0)$. Minimising its potential energy $\int y\sqrt{1+y'^2}\,dx$ at fixed length, find its shape and its sag.
- 4 Show that the extremal of Dido's problem with $\ell=\pi$ is the semicircle and compute the multiplier and the area.
- 5 Prove: if $M\in C[a,b]$ and $\int_a^bM\eta\,dx=0$ for every $\eta\in C[a,b]$ with $\int_a^b\eta\,dx=0$, then $M$ is constant.
- 6 Let $F=F(p)$ with $F''\gt0$. Prove with Jensen's inequality that the straight line is the unique global minimiser of $\int_a^bF(y')\,dx$ in the fixed end-point class.
- 7 (Bolza) Show that $J[y]=\int_0^1\big(y^2+(y'^2-1)^2\big)\,dx$ with $y(0)=y(1)=0$ has infimum $0$ over $C^1$ functions but no minimiser.
E1. CSIR Part B GATE MA Let $y$ be the extremal of $J[y]=\int_0^1\big(y'^2+2yy'+4y^2\big)\,dx$ with $y(0)=0$, $y(1)=\sinh2$. Then $y(1/2)$ equals (A) $\sinh1$; (B) $\tfrac12\sinh2$; (C) $\sin1$; (D) $e^{-1/2}\sinh1$.
E2. CSIR Part B Let $y$ minimise $J[y]=\int_0^1\big(\tfrac12y'^2+xy\big)\,dx$ over $C^1$ functions with $y(0)=0$ and $y(1)$ free. Then $y(1)$ equals (A) $-\tfrac13$; (B) $\tfrac16$; (C) $0$; (D) $-\tfrac12$.
E3. CSIR Part B The minimum of $\int_0^2y'^2\,dx$ over $C^1$ functions with $y(0)=y(2)=0$ and $\int_0^2y^2\,dx=1$ is (A) $\pi^2$; (B) $\pi^2/4$; (C) $\pi^2/2$; (D) $1$.
E4. CSIR Part C Let $F\in C^2$ and consider extremals of $J[y]=\int_a^bF(x,y,y')\,dx$. Which of the following statements are true? (One or more options may be correct.) (A) If $F=F(y,y')$, then $F-y'F_{y'}$ is constant along every $C^2$ extremal. (B) If $F=F(x,y)$, the Euler–Lagrange equation is a first-order ODE. (C) If $F=M(x,y)+N(x,y)y'$ with $M_y=N_x$ on $\mathbb{R}^2$, then $J$ takes the same value on all $C^1$ curves with the same end points. (D) Every $C^2$ function satisfying $F-y'F_{y'}=c$ for some constant $c$ is an extremal (when $F=F(y,y')$).
E5. CSIR Part C Rings of radius $h$ are placed at $x=\pm1$ and we look for $C^1$ curves $y\gt0$, $y(\pm1)=h$, making $S[y]=2\pi\int_{-1}^1y\sqrt{1+y'^2}\,dx$ stationary. Which are true? (A) For $h=1.4$ there is no such curve. (B) For $h=2$ there are exactly two. (C) For $h=2$ the deeper catenary (smaller $y(0)$) has the smaller area. (D) For $h=2$ the shallower catenoid has smaller area than the two discs of radius $2$.
E6. CSIR Part C Let $y$ minimise $J[y]=\int_0^1y''^2\,dx$ over $C^4$ functions with $y(0)=0$, $y'(0)=0$, $y(1)=1$ and $y'(1)$ free. Which are true? (A) $y''''=0$ on $[0,1]$. (B) $y''(1)=0$. (C) The minimum value of $J$ is $3$. (D) $y'(1)=0$.
Suggested plan — 3 lectures × 60 min.
- Functionals and the Euler–Lagrange equation. Start from the brachistochrone story (project the chapter-intro teaser race). Definitions 6.1–6.3, Theorem 6.1, the fundamental lemma and du Bois-Reymond (Theorems 6.2–6.3, draw the bump), Theorem 6.4 with its integrated form. Example 1 on the board, then Interactive 1 (modes 1 and 2) to show $\varphi'(0)=0$.
- First integrals and the classical problems. Theorem 6.5 with the Beltrami computation, Examples 2–4 (brachistochrone in full; for the catenoid, sketch $\cosh t/t$ and count roots). Interactive 1 in the deep-catenary mode: "stationary is not minimal". Example 5 and the pitfalls 1–6.
- Extensions and constraints. Theorems 6.6–6.8 (one example each: III.1, Example 6(a), the Dirichlet integral), natural boundary conditions and transversality (Theorems 6.9–6.10, Example 6(b)), the multiplier rule (Theorem 6.11 with the two-parameter proof) and Example 7 with Interactive 2. End with E4–E6 done live.
Discussion questions. (1) Why does the proof of the Euler–Lagrange equation need variations that vanish at both ends, and what changes when an end is free? (2) The Euler–Lagrange equation for the isoperimetric problem of Example 7(b) is the eigenvalue equation of §2.3. Which quantity plays the role of the eigenvalue, and why is the minimum the first eigenvalue? (3) A soap film between two rings snaps when the rings are pulled apart. Which mathematical event in Example 4 corresponds to the snap?
Misconceptions to watch for. Treating every extremal as a minimum; forgetting that existence can fail; using Beltrami when $F$ depends on $x$, or accepting its spurious constant solutions; imposing natural conditions at fixed ends; forgetting that the multiplier is unknown and must be found from the constraint; differentiating $F_{y'}$ partially in $x$ instead of totally.
Suggested homework. I(1, 3, 4, 6, 8), II(2, 4, 5, 7), III(1, 2, 4, 6, 8), IV(1, 3, 6, 7); Exam corner E1, E4, E5.
Research question. At what separation does a soap film between two equal rings snap, and does the film that survives always have the least possible area?
Why it matters. Soap films are nature's minimal surfaces. Joseph Plateau's experiments in Belgium in the 1870s started a branch of mathematics; Frei Otto's lightweight tensile roofs for the 1972 Munich Olympic stadium were form-found with soap-film models; and the same equations describe biological membranes and the shape of liquid bridges in microfluidic devices. The catenoid is the simplest case in which the stability of an extremal can be tested on a kitchen table.
What to do.
- Bend two wire rings of the same radius $R$ (5–8 cm), dip them together in soap solution (water, washing-up liquid, a little glycerine) and pull them slowly apart, keeping them coaxial. Measure the separation $d$ at which the film collapses onto the two rings; repeat five times.
- Predict the critical separation from Example 4: with $a=d/2$, $h=R$, a catenoid exists iff $\cosh t/t=h/a$ has a root. Find $t^*$ from $t\tanh t=1$ and compute the critical ratio $d/R=2t^*/\cosh t^*$. Compare with your measurement.
- For several $d/R$ below the critical value, compute the areas of both catenoids and of the two discs. Find numerically the ratio $d/R$ at which the shallow catenoid and the two discs have equal area.
- Explain why films with $d/R$ between this ratio and the critical one are observed even though they do not have the least area. Which kind of minimum (weak, strong, global) are they?
Tools: wire, soap solution, a ruler; Python (scipy's brentq) or a spreadsheet with goal-seek for the root finding.
Hand in: A one-page report: measured and predicted critical ratios with an error estimate, a graph of the three areas against $d/R$, and a paragraph on local versus global minima.
Going further. Read about the Jacobi condition (§6.2): the instability of the deep catenoid is detected by a conjugate point. Explore numerically the second variation of the area along $\eta=\cos(\pi x/2a)$ for both catenaries.
🧪 Self-check quiz
Q1. The Euler–Lagrange equation of $\int_a^bF(x,y,y')\,dx$ is
Q2. For $F=y\sqrt{1+y'^2}$ (no explicit $x$) the extremals are
Q3. If $y(b)$ is not prescribed, a minimiser of $\int_a^bF\,dx$ must satisfy at $x=b$
Q4. The extremal of $\int_0^1y'^2\,dx$ with $y(0)=y(1)=0$ and $\int_0^1y\,dx=1$ is
Q5. Which statement about the Euler–Lagrange equation is correct?