Note del corso of Calcolo delle Variazioni, a.a. 2012-13 — A. Visintin Redazione del febbraio 2013 Queste note costituiscono solo una traccia delle lezioni. Per una presentazione piu` com- pleta il lettore `e rimandato alle opere in bibliografia. L’asterisco indica i complementi. Table of Contents I. Classical one-dimensional calculus of variations I.1. Introduction I.2. Integral functionals I.3. Classical conditions for minimization I.4. First integral and Lagrange multipliers I.5. Examples I.6. Boundary conditions I.7. Noether Theorem I.8. Legendre transformation and canonical equations II. Analytical mechanics and optics II.1. Analytical mechanics * II.2. Analytical optics * II.3. From the Maxwell equations to geometrical optics II.4. Hamilton-Jacobi equation III. Minimization, variational inequalities and Γ-Convergence III.1. Direct method of minimization (Tonelli theorem) III.2. Variational inequalities (Lions-Stampacchia theorem) III.3. The obstacle problem * III.4. De Giorgi’s Γ-convergence in metric spaces * III.5. Ekeland’s minimization principle IV. Optimal control IV.1. Control problems * IV.2. Linear optimal control problems (Lions’s theory) IV.3. Dynamic programming and time-discrete Bellman’s theory IV.4. Time-continuous Bellman’s theory IV.5. Pontryagin’s maximum principle IV.6. Pontryagin’s equations * IV.7. Differential games (Isaacs’s theory) V. Elements of convex calculus V.1. Convex lower semicontinuous functions * V.2. Fenchel’s transformation * V.3. Subdifferential 1 Acronyms B-function: Bellman function B-equation: Bellman equation E-L equation: Euler-Lagrange equation H-J equation: Hamilton-Jacobi equation H-J-B equation: Hamilton-Jacobi -Bellman equation PDP: principle of dynamic programming ODE: ordinary differential equation PDE: partial differential equation We shall denote by D ,...,D the partial derivatives with respect to x ,..,x . If the 1 N 1 N latter are M-dimensional vectors, we shall denote by D the partial derivatives with respect ij to the jth component of x , for i = 1,..,N and j = 1,..,M. i I. ONE-DIMENSIONAL CALCULUS OF VARIATIONS I.1. Introduction This course will touch several issues of the classical and the modern theory of calculus of variations, which also includes aspects that are at the interface with optimization, convex analysis, variational inequalities, and other subjects. This field if relevant in itself as well as for applications. We start with the classical one-dimensional theory, including fundamental results of Euler, Lagrange, and others. We then illustrate applications to analytical mechanics and analytical optics, including the Hamilton and Jacobi theories. Afterwards we introduce Tonelli’s direct method of the calculus of variations and Stam- pacchia’s variational inequalities. We also outline De Giorgi’s theory of Γ-convergence and Ekeland’s variational principle in metric spaces. We then deal with optimal control, including fundamental results of Lions, Bellman and Pontryagin; we also outline differential games. Notice that differential games generalize optimalcontrol,whichinturngeneralizesthebasicBolzaproblemofcalculusofvariations. A commonthreadmaybefoundinthecalculusofvariations,optimalcontrol,differentialgames and their applications to physics and economics, and so on: the possibility of formulating a Hamilton-Jacobi-type PDE and the equivalent approach in terms of a canonical-type system of ODEs. In the final part we introduce some basic elements of convex calculus: the Legendre- Fenchel transformation and the notion of subdifferential. I.2. Integral functionals Let us define the Lagrange functional (cid:90) b F(u) = f(x,u(x),u(cid:48)(x))dx (2.1) a ∀u ∈ V := {v ∈ C1([a,b]) : v(a) = α,v(b) = β}, with a < b, α,β ∈ R prescribed, and f ∈ C0([a,b]×R×R) a given function. More generally, one may deal with: 2 (i)vector-valuedfunctions[a,b] → RN,forsomeintegerN ≥ 1. Inthiscaseoneassumes that f ∈ C0([a,b]×RN×RN), and α,β ∈ RN; (ii) either scalar- or vector-valued functions u that depend on several variables. In this case the ordinary derivative u(cid:48) is replaced by the gradient ∇u; (iii) functions u that belong to other function spaces, in particular Sobolev spaces; (iv) more general boundary conditions; (v) more general functionals than F; and so on. On the history of the Calculus of Variations. Classical one-dimensional problems include Dido’s isoperimetric problem, the problem of the curve of minimal length joining two prescribed points, the problem of the curve brachystochrone, and so on (see references). AmongthemultidimensionalproblemswementiontheclassicalDirichlet problem, which in its original formulation consisted in minimizing the integral (cid:90) I(u) = |∇u(x)|2dx u ∈ C1(Ω)∩C0(Ω¯) (2.2) Ω (with Ω domain of R3), and prescribing the value of u on (a part of) the boundary. One may show that any minimizing function of (2.2) is harmonic, namely fulfills the equation ∆u = 0 in Ω. Problems like this need not have a solution in Ck spaces. Traditionally one distinguishes between: (i) classical or indirect methods (essentially before 1900): study of either necessary or sufficientconditionsforminimizers, typicallyformulatedintermsofequationsorinequalities; this assumed the existence of a minimizer; (ii) direct methods (essentially after 1900): these are aimed to establish the existence of minimizers, typically via the use of minimizing sequences and the topological notions of compactness and lower semicontinuity. Here larger spaces than the Cks are usually involved, e.g., Lp and Sobolev spaces. The Calculus of Variations was developed by analysts and mathematical-physicists, at a lucky time when there was no sharp distinction between these disciplines. This left traces in terminology and notation. Examples. Let us define the functionals (cid:90) 1 F (u) = x2u(cid:48)(x)2dx 1 (2.3) −1 ∀u ∈ V := {v ∈ C1([−1,1]) : v(−1) = 0,v(1) = 1}, 1 (cid:90) 1 1 F (u) = dx 2 1+u(cid:48)(x)2 (2.4) 0 ∀u ∈ V := {v ∈ C1([0,1]) : v(0) = 0,v(1) = 1}. 2 It is easy to see that both functionals have no minimum and no maximum. Two lemmata. Lemma 2.1. (Fundamental lemma of the Calculus of Variations) If g ∈ C0([a,b]) is such that (cid:90) b g(x)v(x)dx = 0 ∀v ∈ C0([a,b]),v(a) = v(b) = 0, (2.5) a 3 then g ≡ 0 in [a,b]. Proof. By contradiction, let x¯ ∈ [a,b] exist such that g(x¯) (cid:54)= 0. Then g has constant sign in a suitable neighborhood U of x¯. Selecting a function v which is positive in U and vanishes (cid:82)b elsewhere, we get g(x)v(x)dx (cid:54)= 0, at variance with the assumption. (cid:117)(cid:116) a Lemma 2.2. (Du Bois-Reymond lemma) If g ∈ C0([a,b]) is such that (cid:90) b g(x)v(cid:48)(x)dx = 0 ∀v ∈ C1([a,b]),v(a) = v(b) = 0, (2.6) a then g is constant in [a,b]. If g ∈ C1([a,b]) this assertion is a simple consequence of the previous lemma. If g is just continuous, this result may be proved via a small trick, or by a regularization procedure. Variations. Let a,b,α,β ∈ R with a < b, assume that f ∈ C1([a,b]×R×R), and search for a minimizer of the functional (cid:90) b F(u) = f(x,u(x),u(cid:48)(x))dx (2.7) a ∀u ∈ V := {v ∈ C1([a,b]) : v(a) = α,v(b) = β}. α,β The linear space V is associated to the affine space V . We shall deal with the min- 0,0 α,β imization of this functional in this (affine) function space; this is known as the Lagrange problem.. Let us introduce the auxiliary function ϕ (t) := F(u+tv) ∀u ∈ V ,∀v ∈ V ,∀t ∈ R, (2.8) u,v α,β 0,0 and set δF(u,v) := ϕ(cid:48) (0) ∀u ∈ V ,∀v ∈ V . (2.9) u,v α,β 0,0 When existing, δF(u,v) is called the first variation of F in u, with respect to the (infinite- dimensional)vectorv. Thisextendsthenotionofderivativewithrespecttofinite-dimensional vectors, that is already known to the reader for functions RN → R. Similarly, we set δ2F(u,v) := ϕ(cid:48)(cid:48) (0) ∀u ∈ V ,∀v ∈ V ; (2.10) u,v α,β 0,0 ifexisting, thisiscalledthesecondvariationofF atuwithrespecttothe(infinite-dimension- al) vector v. It is easily checked that (cid:90) b δF(u,v) = (vD f +v(cid:48)D f)dx ∀u ∈ V ,∀v ∈ V , (2.11) 2 3 α,β 0,0 a (cid:90) b δ2F(u,v) = [v2D2f +2vv(cid:48)D D f +(v(cid:48))2D2f]dx 2 2 3 3 (2.12) a ∀u ∈ V ,∀v ∈ V . α,β 0,0 4 Exercise. Let f : R2 → R be of class C2. If (0,0) is a point of relative minimum for the restrictions of f to all straight lines through the origin, does it follow that (0,0) is also a local minimizer of f in the plane? (i.e., is f minimized in some neighborhood of (0,0)?) Hint: consider Peano’s function f(x,y) = (y−x2)(y−2x2) for any (x,y) ∈ R2. I.3. Classical conditions for minimization Proposition 3.1. (GeneralizedFermatprinciple)LetthefunctionalF bedefinedasin(2.7), with f ∈ C1([a,b]×R×R). For any extremum u of F (i.e., a point of either maximum or minimum), then δF(u,v) = 0 ∀v ∈ V . (3.1) 0,0 Proof. By the definition (2.9) of the first variation, it suffices to apply the classical Fermat principle to the function ϕ (cf. (2.8)). (cid:117)(cid:116) u,v Because of the theorem of passage to the limit under integral, it is easy to see that (3.1) entails the Euler-Lagrange equation (more shortly, “E-L equation”) in weak form: (cid:90) b (δF(u;v) =) [D f(x,u(x),u(cid:48)(x))v(x)+D f(x,u(x),u(cid:48)(x))v(cid:48)(x)]dx = 0 2 3 (3.2) a ∀v ∈ V . 0,0 This equation is also equivalent to the following integral equation (named after Du Bois- Reymond, and easily deducible from the Du Bois-Reymond Lemma 2.2), for a suitable real constant C: (cid:90) x D f(x,u(x),u(cid:48)(x)) = C + D f(s,u(s),u(cid:48)(s))ds ∀x ∈ [a,b]. (3.3) 3 2 a If f ∈ C2([a,b]×R×R) and u ∈ V ∩C2([a,b]), differentiating (3.3) with respect to x α,β one gets the E-L equation in strong form: (1) d D f(x,u(x),u(cid:48)(x)) = D f(x,u(x),u(cid:48)(x)) ∀x ∈ [a,b]. (3.4) 3 2 dx Conversely, if f ∈ C2([a,b]×R×R), u ∈ V ∩C2([a,b]) and the strong form of the E-L α,β equation is fulfilled, then obviously the weak form of the E-L equation is also satisfied. By what we saw, any extremum fulfills the E-L equation in weak form if u ∈ C1([a,b]), and that in strong form if f ∈ C2([a,b]×R×R) and u ∈ C2([a,b]). Any solution of the E-L equation in weak (strong, resp.) form is called a weak (strong, resp.) extremal.(2) Therefore, under suitable regularity conditions, any extremum is an extremal, but not conversely. (3.5) However we have the next statement. Proposition 3.2. If F is convex, then any extremal is a minimizer (so it is an extremum). (3.6) (1) In alternative, (3.4) may be derived directly from (3.1) via the fundamental lemma of the calculus of variations. (2) So there are extrema, extremals, and also ... the extremes of the interval. 5 If f(x,·,·) is convex for any x, then F is convex. If f(x,·,·) is strictly convex for any x, then the minimizer (if existing) is unique. Proof. (3.6) is obvious for functions R → R. In our case it suffices to notice that for any u ∈ V and any v ∈ V the auxiliary function R → R : t (cid:55)→ ϕ (t) := F(u+tv) is convex. α,β 0,0 u,v The proof of the second part is straightforward. (cid:117)(cid:116) Classically the E-L equation has been used in the search for extrema: if an extremum exists then it fulfills the E-L equation, and may thus be searched for among the solutions of that equation. * On the integral formulation of the Euler-Lagrange equation. Theintegralequation (3.3) and the equivalent weak equation (3.4) characterize extremals of F also if C1([a,b]) is replaced either by C1 ([ab]) (namely the space of continuous functions [a,b] → R that are pw piecewise of class C1), or by the wider space of Lipschitz functions. In this case (3.3) just holds a.e. in ]a,b[. In order to write the strong equation (3.4) it is not needed that f,u be of class C2. (3) The hypothesis that D f(x,u(x),u(cid:48)(x)) be of class C1 is actually needed for the application 3 of the fundamental lemma; anyway, instead of that lemma one may use Du Bois-Reymond’s lemma. Moreover existence and continuity of the left member of (3.4) follow from (3.4) itself, by comparing the terms of this equation. In any case, without further regularity of f and u, the derivative at the left side cannot be developed. One may then write (3.4) in either weak or integral form. E-L equation for vector functions. Let f ∈ C1([a,b]×RN×RN) and u ∈ C1([a,b])N, for some integer N ≥ 1. With reference to the function [a,b]×RN×RN → R : (x,u,ξ) (cid:55)→ f(x,u,ξ), let us denote by D (D , resp.) the partial derivative with respect to the j-th component 2j 3j of the vector u (ξ, resp.), for j = 1,...,N. By an obvious extension of the previous procedure, if u is an extremum of F (defined as in (2.1)) we infer the E-L equation in weak form (cid:90) b [D f(x,u(x),u(cid:48)(x))v (x)+D f(x,u(x),u(cid:48)(x))v(cid:48)(x)]dx = 0 2j j 3j j (3.7) a ∀v ∈ C1([a,b])N such that v(0) = v(b) = 0, implying the sum (from 1 to N) over repeated indices (Einstein’s convention). This weak equation is also equivalent to an integral equation which is analogous to (3.3). If f ∈ C2([a,b]×RN×RN) and u ∈ V ∩ C2([a,b])N, then here also the weak E-L α,β equation is equivalent to the strong E-L equation d D f(x,u(x),u(cid:48)(x)) = D f(x,u(x),u(cid:48)(x)) ∀x ∈ [a,b], for j = 1,...,N. (3.8) 3j 2j dx Alsointhiscaseonespeaksofweakandstrongextremals, andtheaboveresultsareextended to vector-valued functions in an obvious way. (3) Anyway so is written in several textbooks, where by writing (3.4) it is implied that the derivative at the left side is developed via the theorem of the composed function. 6 The E-L equation can also be extended to functionals that depend on higher-order derivatives of u, to different boundary conditions, to functions u of several variables, and so on. For instance, let us consider the problem (cid:90) b F(cid:98)(u) = f(x,u(x),u(cid:48)(x),u(cid:48)(cid:48)(x))dx (3.9) a ∀u ∈ V := {v ∈ C2([a,b]) : v(a) = α,v(b) = β} for a given function f ∈ C4([a,b]×R×R×R), and prescribed a < b, α,β ∈ R. (Different boundary conditions may also be considered.) It is easy to see that any extremal u ∈ V ∩C4([a,b]) of F(cid:98) fulfills the following E-L equation in strong form (written implying the argument (x,u(x),u(cid:48)(x),u(cid:48)(cid:48)(x)) of f and of its derivatives): d dx2 D f − D f = D f ∀x ∈ [a,b], for j = 1,...,N. (3.10) dx 3 dx2 4 2 More generally, if the integrand depends on derivatives up to an order k and f ∈ C2k, then the E-L equation in strong form is an ODE of order 2k. A necessary condition for minimization. Theorem 3.3. (Legendre condition) If f ∈ C2([a,b]×RN×RN), then for any minimizer u of F the Hessian matrix D2f(x,u,u(cid:48)) is positive semidefinite. (3.11) 3 Proof. If u is a minimizer of F then δ2F(u,v) ≥ 0 ∀v ∈ V . (3.12) 0,0 Iff ∈ C3([a,b]×RN×RN)andu ∈ V ∩C2([a,b]), noticingthat2vv(cid:48) = (v2)(cid:48) andintegrating α,β by parts the second addendum of (2.12), we get (implying the argument (x,u(x),u(cid:48)(x)) of f and its derivatives) (cid:90) b(cid:104) (cid:16) d (cid:17) (cid:105) δ2F(u,v) = v2 D2f − D D f +(v(cid:48))2D2f dx ∀u ∈ V ,∀v ∈ V . (3.13) 2 dx 2 3 3 α,β 0,0 a These derivatives of f are obviously bounded. As (v(cid:48))2 may be arbitrarily large also where v2 is small, (3.11) must then hold for any minimizer (i.e., point of minimum) of F. (cid:117)(cid:116) The Legendre condition (3.11) is just a necessary condition for the minimization of F; actually, freely speaking, the v2-term may prevail over the (v(cid:48))2-term in (3.13): (v(cid:48))2 may be arbitrarily small also where v2 is large.. Even if the Hessian D2f(x,u,u(cid:48)) were positive 3 definite for any x ∈ [a,b], one could not infer that u is a minimizer. A counterexample is provided by the quadratic functional F(u) = (cid:82)2π[(u(cid:48))2 − u2]dx; 0 here the minimization sees the first-order term competing with the zero-order one. In this case u ≡ 0 fulfills the E-L equation, and is thus an extremal of F. On the other hand (as 0 the integrand is a homogeneous function of degree two) (cid:90) 2π δ2F(u ,v) = [(v(cid:48))2 −v2]dx ∀v ∈ V ; (3.14) 0 0,0 0 7 setting v(x) = sin(x/2), it is easily checked that (cid:90) 2π(cid:110)1 (cid:111) δ2F(u ,v) = cos2(x/2)−sin2(x/2) dx < 0. (3.15) 0 4 0 Exercise. Study the minimization of the functionals (cid:90) 1 G (u) = u(x)2[1−u(cid:48)(x)]2dx 1 (3.16) −1 ∀u ∈ W := {v ∈ C1 ([−1,1]) : v(−1) = 0,v(1) = 1}, 1 pw G (u) = G (u) 2 1 (3.17) ∀u ∈ W := {v ∈ C1([−1,1]) : v(−1) = 0,v(1) = 1}. 2 (cid:90) 1 G (u) = u(x)2xdx 3 (3.18) 0 ∀u ∈ W := {v ∈ C1([0,1]) : v(0) = 0,v(1) = 1}. 3 (G has a minimum in u (x) = (x − 1)+. As u (cid:54)∈ C1([−1,1]), one may see that 1 0 0 infG = 0 but G has no minimizer. G has no minimizer too, also if infG = 0.) 2 2 3 3 Relative minimizers. So far, we just dealt with absolute (or global) minimizers. Next we consider relative (or local) minimizers, with respect to a prescribed topology; here we confine ourselves to the scalar setting (i.e., N = 1). Any u ∈ C1([a,b]) is called a strong (weak, 0 resp.) relative minimizer if and only if there exists a neighborhood U of u in the topology 0 of C0([a,b])(4) (C1([a,b]), resp.) such that u is an absolute minimizer for the restriction of 0 F to U. As the topology of C1([a,b]) is strictly finer than that of C0([a,b]), every strong relative minimizer is also a weak relative minimizer. The converse fails; see e.g. the counterexample in [De 100]. Using this terminology (which is standard in the Calculus of Variations), thus absolute minimizer ⇒ strong relative minimizer ⇒ weak relative minimizer, and all of these implications are strict. Other second order conditions. (5) Let us consider the functional F defined in (2.1). In analogy with what is known from the basic course of analysis for functions of C2(RN), any strong relative minimizer u of F fulfills: (i) the first-order extremality condition δF(u,v) = 0 ∀v ∈ V (equivalent to the E-L equation), (3.19) 0,0 (ii) the second-order condition(6) δ2F(u,v) ≥ 0 ∀v ∈ V . (3.20) 0,0 (4) This is the topology that C0([a,b]) induces onto the subset C1([a,b]). (5) This outline is especially synthetical. A more detailed presentation may be found e.g. in [De 98-113]. (6) here we just deal with scalar functions u scalari. 8 The converse fails, even if (at variance with what is known from the basic course of analysis for functions of C2(RN)) the latter condition is replaced by the stronger δ2F(u,v) > 0 ∀v ∈ V ; (3.21) 0,0 see e.g. the counterexample in [De 104]. At the end of this section we shall see that however a suitable strengthening of this inequality provides a sufficient condition for an extremal to be a strong relative minimizer. In any case, if the functional F is convex then any extremal is a minimizer. Theorem 3.4. (Legendre) If f ∈ C2 and u is a weak relative minimizer, then (compare 0 with Theorem 3.3 for N = 1) D2f(x,u (x),u(cid:48)(x)) ≥ 0 ∀x ∈ ]a,b[. (3.22) 3 0 0 The converse fails (we already saw a counterexample). In order to display the next result, first for any f ∈ C1 let us define the Weierstrass excess function E(x,u,ξ,η) := f(x,u,η)−f(x,u,ξ)−(η−ξ)D f(x,u,ξ) 3 (3.23) ∀(x,u,ξ,η) ∈ [a,b]×R3. Theorem 3.5. (Weierstrass) If f ∈ C1 and u is a strong relative minimizer, then 0 E(x,u (x),u(cid:48)(x),η) ≥ 0 ∀x ∈ ]a,b[,∀η ∈ R. (3.24) 0 0 This fails for weak relative minimizers. The converse implication also fails. Notice that (3.24) means that, for any x, in the plane (u,f) the straight line that is tangent to the graph of f(x,u (x),·) at the point (u(cid:48)(x),f(x,u (x),u(cid:48)(x)) stays be- 0 0 0 0 low that graph everywhere. (3.24) then entails that D2f(x,u (x),u(cid:48)(x)) ≥ 0, whenever 3 0 0 D2f(x,u (x),u(cid:48)(x)) exists. If f ∈ C2, then the Weierstrass condition (3.24) entails the 3 0 0 Legendre condition (3.22); in this case the Weierstrass Theorem 3.5 thus follows from the Legendre Theorem 3.4. A sufficient condition: If f ∈ C2, u fulfills the weak E-L equation, and 0 (cid:90) b ∃λ > 0 : ∀v ∈ V , δ2F(u,v) ≥ λ [v(x)2 +v(cid:48)(x)2]dx, (3.25) 0,0 a then u is a weak relative minimizer. The converse obviously fails. (7) On the other hand, 0 the weaker condition (3.21) does not even entail that u is a weak relative minimizer; for a 0 counterexample see e.g. [De 104]. Synthesis. (See above for the regularity assumptions) — For absolute minimizers (for any N): if F is convex, then any extremal is a minimizer, (7) This result is at the basis of a sophisticated theory of Legendre and Jacobi (see e.g. [De 115-124]). 9 for any minimizer u, the Hessian matrix D2f(x,u,u(cid:48)) is positive semidefinite. 3 — For relative minimizers (for N = 1): for any weak relative minimizer u, D2f(x,u,u(cid:48)) ≥ 0, 3 for any strong relative minimizer u, E(x,u (x),u(cid:48)(x),η) ≥ 0, 0 0 any extremal u that fulfills (3.25) is a weak relative minimizer. The converse of these statements fails. I.4. First integral and Lagrange multipliers First integrals. For any f ∈ C2 and any solution u ∈ V ∩C2([a,b]) of the strong E-L α,β equation of the functional (2.1), a simple computation [De 70] shows that d [f(x,u(x),u(cid:48)(x))−u(cid:48)(x)D f(x,u(x),u(cid:48)(x))] = D f(x,u(x),u(cid:48)(x)) 3 1 dx (4.1) ∀x ∈ [a,b]. In analogy with what we saw for the integral formulation of the E-L equation, by integrating this equation we get (cid:90) x f(x,u(x),u(cid:48)(x))−u(cid:48)(x)D f(x,u(x),u(cid:48)(x)) = C + D f(s,u(s),u(cid:48)(s))ds 3 1 (4.2) a ∀x ∈ [a,b], and this holds also if f ∈ C1 and u ∈ C1([a,b]). More generally, if u ∈ C1 ([a,b]) then (4.2) pw holds a.e. in [a,b]. If f does not explicitly depend on x then, defining the Hamiltonian function, H(u,ξ) := ξD f(u,ξ)−f(u,ξ) ∀(u,ξ) ∈ R2, (4.3) 3 by (4.1) we get d H(u(x),u(cid:48)(x)) = 0 ∀x ∈ [a,b]. dx That is, for any extremal u = u(x) of F, H(u(x),u(cid:48)(x)) = C : constant ∀x ∈ [a,b]. (4.4) The constant C may be determined by the boundary conditions. (4.4) is also named a conservation law, and H(u,u(cid:48)) is also called a first integral (of the motion). Establishing a first integral may be useful for the integration of the E-L equation. For instance the equation H(u,u(cid:48)) = C is of first order, whereas the E-L equation is of second order. If the equation H(u,u(cid:48)) = C can be made explicit with respect to u(cid:48), to wit if it may be rewritten in the normal form u(cid:48) = g(u,C), then a standard theory may be applied to the study of this equation. Analogous remarks apply in the case of D f(x,u,u(cid:48)) ≡ 0, that is the function f(...) does 2 not depend on u. The E-L equation is then reduced to D f˜(x,u(cid:48)) = C. If this equation 2 can be made explicit with respect to u(cid:48), then also in this case one is reduced to a first-order equation. Functionals of a different form from (2.1) may have several first integrals, or no one at all. 10