1 Differential calculus This introductory chapter is mainly devoted to the differential calculus in Banach spaces. In addition to being a fundamental tool later on, the treatment of the calculus at this level permits better understanding at certain aspects, which might otherwise be neglected. - We discuss in Section 1 the Fréchet and Gateaux derivatives as well as their elementary properties. The differentiability of the Nemitski operator is investigated in Section 2 and higher and partial derivatives are introduced in Sections 3 and 4, respectively. 1 Fréchet and Gateaux derivatives ‘The Fréchet-differential is nothing else than the natural extension to Banach spaces of the usual definition of differential of a map in Buclidean spaces. = Let U be an open subset of X and consider a map F:U 4 Y. Definition 1.1 Let u © U. We say that F is (Fréchet-) differentiable at u if there exists A € L(X,¥) such that, if we set Rh) = F(u +h) — Fu) - A(h), there results Rih) = of|lhll), (1.1) that is IRC)I) ‘Atl 0 as ||hj] +0.10 1 Differential calculus Such an A is uniquely determined and will be called the (Fréchet) dif ferential of F at u and denoted by A=dF(u). If F is differentiable at all u € U we say that F is differentiable in U. Hereafter, when there is no possible misunderstanding, Fréchet differ- entiability will be referred to simply as differentiability. A few comments on the preceding definition are in order. (i) Let us verify that A is unique. Supposing the contrary, let B € L(X,Y) satisfy Definition 1.1 and A # B. It follows that Ah ~ Bh (all If A # B there exists h* € X such that a Taking h = th*, t € IR — {0}, one has |A(th*) — B(th*) I} ||An* — Bre a Hehe ll o [aT erly? a constant, in contradiction with (1.2). (ii) If F is differentiable at u then F(u +h) = F(u) + dF(u)h + of|hll) and F is continuous at the same point. Conversely if F € C(U,Y) then it is not necessary to require in Definition 1.1 the continuity of A. In fact (1.1) yields A(h) = P(u + h) — F(u) - o(||hll) and the continuity of F implies the continuity of A. (iii) ‘The definition of differentiability depends not on the norms but on the topology of X and Y only. That is if, for example, ||.|| and |||-{I| are two equivalent norms on X then F is differentiable at u in (X, |].|]) if and only if F is in (X, ||l.|||) and the differential is the same. a 0 as [lhl] +0. (1.2) Ah — Bh*|| #0. Remark 1.2 The preceding comment (iii) could suggest the idea of extending the notion of Fréchet differentiability to locally convex topo- logical spaces. The most natural way would be the following: let the topology of X (respectively Y) be induced by an infinite family of semi- norms |-|x,¢ (resp. |I.liy.); define the differential of F as the linear continuous map A with the property that for all |.|yj there exists a seminorm |.[x.; such that |F(u +h) — F(u) — Ahly, = o(|hi x2). With such a definition all the main properties of the differential (below) hold true. Unfortunately, in dealing with the higher derivatives, there are1.1 Fréchet and Gateaus derivatives W strong difficulties and people introduced new classes of spaces, such as the “pseudotopological spaces”, where a differential calculus suitable for the purposes of analysis can be carried out. These kind of topics, how- ever, are beyond the purposes of our book. Examples 1.3 (a) The constant map F(u) = c is differentiable at any u and dF(u) = 0 for all ue X (b) Let A€ L(X,¥). Since A(u+h) — A(u) = ACh), it follows that 4 is diffcrentiable in X and dA(u) = A. (c) Let B: X x ¥ s Z be a bilinear continuous map. There results Blu + hw +k) — B(u,v) = Blh, v) + Blu, k) + B(h, k) From the continuity at the origin it follows that NB, &)Il < eA (eH. ‘Then B is differentiable at any (u,v) € X x ¥ and dB(u,v) is the map (h, k) + B(h,») + B(u, k). {d) Let H be a Hilbert space with scalar product (.[.) and consider the map Fi u— lull? = (ula). From [ee All? — el]? = Cee) + ill? it follows that F is differentiable at any w and dF(u)h = 2(ulh) Note that, |].|| is not differentiable at u = 0. For, otherwise, ||hl] = Ah + o(l|hi]) for some A € L(H,IR). Replacing h with —h we would deduce that tl] = —Ah + 0(|jhl]) and hence ||Al| = o(||h!]), a contradiction. (ec) If X =R,U = (a,b) and F : U = Y is differentiable at t € U, the differential dF(t) can be identified with dF(¢)[1] € Y though the canonical isomorphism i: L(R,Y) > Y, i(A) = AQ). For example, if Y = R" and F(t) = (f.(0))ia1,..m, dF (t) “is” the vector with components df;/dt. ‘The main differentiation rules are collected in the following proposi- tion, Proposition 1.4 (i) Let F,G:U 4Y. If F and G are differentiable at u € U then aF +0G is differentiable at u for any a,b € IR and d(aF + bG)(u)h = adF(u)h + bdG(u)h. (ii) (Composite-map formula) Let F : U + Y andG:V — Z with2 1 Differential calculus V > FU), U and V open subsets of X and Y, respectively, and consider the composite map GoFiU + Z, GoFlu) = G(F(u)). Uf F is differentiable at u € U and G is differentiable at y := Flu) €V, then Go F is differentiable at u and A(G o F)(u)h = dG(v) [dF (u)h]. dn other words the differential of Go F at u is the composition of the linear maps dF (u) and dG(v), with v = F(u). ‘The proofs of (i) and (ii) do not differ from those of the differentiation rules in R". Definition 1.5 Let F:U 3 Y bea differentiable in U. The map FU LX,Y), Psu dF(u), is called the (Fréchet) derivative of F. If F" is continuous as a map from U to L(X,Y) we will say that F is C? and write F’ € C'(U,Y). Let us introduce the concept of variational (or potential) operator. If Y = R, maps J : U — R aro usually called functionals and J’ turns out to be a map from U to £(X,IR) = X* (the dual of X). In particular, if X = H is a Hilbert space, J'(u) € H* for all u and the Riesz Representation Theorem allows us to identify J’(u) with an element of H., To be precise, we give the following definition. Definition 1.6 Given a differentiable functional J : U — R the gra- dient of J at u, denoted by VJ(u), is the element of H defined by (VJ{u)|h) = dI{u)h, for all h € H. (1.3) A map F : U + H with the property that there exists a differentiable functional J: U — R such that F = VJ is called a variational (or potential) operator. As for maps in IR", we can also define here a directional derivative, usually called the Gateanx differential (for short, G-differential). Definition 1.7 Let F:U 4 Y be given and let z € U. We say that F is G-differentiable at w if there exists A € L(X,Y) such that for all heX there results Fe Fas snc 0, (14)1.1 Fréchet and Géteaur derivatives "13 ‘The map A is uniquely determined, called the G-differential of F at*u and denoted by de Fu). Clearly, if F is Fréchet-differentiable at u then F is G-differentiable there and the two differentials coincide. Conversely, the G-diferentiabil- ity does not imply the continuity of F; even: recall the elementary example F : IR? — R? defined by sty? =|=—| if F(t) = [seta] ite eo, F(s,0) =0. The following result replaces the elementary “Mean-Value Theorem” and plays a fundamental role in what follows. For u,v € U let [u,v] denote the segment {tu + (1 — tv: € [0, i]}. Theorem 1.8 Let F: U — Y be G-differentiable at any point of U. Given u,v € U such that [u,v] CU, there results IF(u) — Feil < sup{lldcF(w)l] : w € [u,v]} lle - vl. Proof. Without loss of generality we can assume that F(u) # F(v). By a well-known corollary of the Hahn-Banach Theorem there exists wey’, lvl = 1, such that (b, P(u) — Fle) = |P(u) — Fe)ll. (1.5) Let y(t) =tu+(1—t)v, t € [0,1], and consider the map h : [0,1] + R defined by setting A(t) = (B. Fly) = (d, F(tu + (1 — tv). From 7(t +7) = 7(t) + 7(u— ») it follows that h(t+7) = h(t) (@ Fo(t+7))- fatal T T = (y, FR FHO), ag T Since F is G-differentiable in U, passing to the limit in (1.6) as 7 — 0, we find Al(t) = (W,deF (tu + (1-)v)(u—v)). (1.7) Applying the Mean-Value Theorem to ft one has A(1) — h(0) = h'(@) for some @ € (0, 1). (1.8)14 1 Differential calculus Substituting (1.5) and (1.7) into (1.8) we get IF (u) — F()I| = AG) — RO) = 4°) 1, doF (Ou + (1 ~ 8v)(u— )) < livll ldcF (Gu + (1 - 0)v)}) [lu — vl]. Since ||] = 1 and 6u + (1 — 8)u € [u,»] the theorem follows. As a consequence we can find a classic criterion of Fréchet differentia- bility. Theorem 1.9 Suppose F: U —Y is G-differentiable in U and let Fo :U + (X,Y), Fe(u) =doF(u), be continuous at u’. Then F is Fréchet-differentiable at u” and dF(u*) = doF(u’). Proof. We set R(h) == Fu" +h) — Flu’) ~ da F(u")h. Plainly, R is G-ditferentiable in Be, for ¢ > 0 small enough, and dgR(h) sk + dg F(w" + h)k — dg F(u")k. (1.9) Applying Theorem 1.8 with (u, vj = [0, A], we find (note that R(0) = 0) IR@)I| < aap Ide R(EA)|| |All. (1.10) From (1.9) with th instead of h, we deduce {| dg R(th) || = ||dg F(u" + th) £ de Pu"). Substituting into (1.10) we find ACA} < sup [de F(u* + th) - de F(u")|| Wah. osts1 Since F’,is continuous, sup ||dcF(u* + th) —deF(u")|| + 0 as ||hl| +0 ostsi and therefore R(h) = o((|hi]). In view of Theorem 1.8, to find the Fréchet differential of F one can determine dc F and show that F% is continuous. Let F € C([a,5], X) and set (see subsection 0.3) ; ae = [ Fed It is immediately verifiable that © is differentiable and #/(t) = F(t)1.2 Continuity and differentiability of Nemitski operators 15 (we are using the canonical identification between L(IR, X) and X; see Example 1.3 (e)). is called a primitive of F. From Theorem 1.8 it follows that Bt) — O(s) || < sup (IlF(OIl (¢- 5): s SE < t} Hence, if F(€) = 0 for all € € [a,4], one has & =constant. In particular, ® is, up to a constant, the unique primitive of F. As a consequence we can obtain the following useful formula. Suppose that [u,v] € U and let F € C'(U,Y). The map Foy: [0,1] 4 Y, Fo a(t) = F(tu + (1 — tv) is Cl and (Foy) (6) = F'(tu+ (1- tv) [ Integrating from 0 to 1 we get 1 Fe) Fw) = f Fut duu ef 3 i Note that in the last integral F” is meant to take values in L(X,¥). 1 [/ Pitut(1- anal (u-v). Qa) 2 Continuity and differentiability of Nemitski operators In this section we want to study the differentiability of an important class of operators arising in nonlinear analysis: the so called “Nemitski operators” we are going to define. Nemitski operators Let & be an open bounded subset of IR” and let M() denote the class of real-valued functions u : 8% + IR that are measurable on ©. Here, and always hereafter, the measure is the Lebesgue one and will be denoted by y; all the functions we will deal with in this section are taken in M(Q). Let f: x RR be given. Definition 2.1 The Nemitski operator associated to f is the map de- fined on M(Q) by setting ux) > f(z, u(x)). The same symbol f will be used to denote both f and its Nemitski operator.16 1 Differential calculus We shall assume that f is.a Carathéodory function. More precisely, we will say that f satisfies (C) if (i) s+ f(z,s) is continuous for almost every z € 2, (i) 2— f(z,s) is measurable for all s€ R. For the purpose of analysis it is particularly interesting when the Nemitski operators act on Lebesgue spaces L? = L?() (hereafter we will write L? for L?(9)) and we shall discuss this case in some detail. Let us start by noticing that . f(u) € M(Q) for all ue M(Q). (2.1) Indeed, if u € M(Q) there is a sequence Xp of simple functions such that xn + ua.e. in 2. From (C) it follows that f(Xn) is measurable and f(xn) > f(u) ae. in ©, and from this we deduce that f(u) € M(). Continuity of Nemitski operators Let p,g > 1 and suppose If(zs)| 0. Theorem 2.2 Let 2.¢ RR” be bounded and suppose f satisfies (C) and (2.2). Then the Nemitski operator f is a continuous map from L? to L?. For the proof we need the following measure-theoretic result (see, for example, [Br], Theorem LV.9). Theorem 2.3. Let j:(9) < 00 and let un + win LP. Then there exist a sub-sequence tn, and h € LP such that Un, uae. in Q, (2.3) Iuul Sh ae. in 2. (2.4) Proof of Theorem 2.2 From (2.1)-(2.2) it follows immediately that. Sf (u) © £4 whenever u € L?. To show that f is continuous from LP to L*, let un,u € L? be such that lun — elle» 0. Using Theorem 2.3 we can find a sub-sequence {u,,} of {un} and h € L?1.2 Continuity and differentiability of Nemitski operators 17 satisfying (2.3)-(2.4). Since un, converges almost everywhere to u, it readily follows from (C) that Sling) + f(t) ae. in (2.5) Moreover, from assumption (2.2) and (2.4) we infer [f(un,)| < a+ b|tny|? 2 and suppose f has partial derivative fj = 8f/0s satisfying (C) and such that {fa(z,s)| < a+ bfs|P-? (2.7) for some constants a,b > 0. Since f, satisfies (2.7), Theorem 2.2 applies and the Nemitski operator f, is continuous from L? to L", with r = p/(p — 2). As a consequence, for the function f,(u)v defined by fa(u)o: 2 — f.lz,u(z))v(z) one has that f,(u)v € L?’ for u,v € L, where p’ = p/(p — 1) is the conjugate exponent of p. Theorem 2.6 Let 2 CR" be bounded and suppose that p > 2 and f satisfies (C). Moreover, we suppose that f(z,0) is bounded and that f has partial derivative f, satisfying (C) and (2.7). Then f : L” + L” is Fréchet-differentiable on L? with differential df(u) uv — fa(u)v.18 1 Differential catculus Proof. Integrating (2.7) we find constants c,d > 0 such that \f(@,8)| Oast 0. (2.12) Fa(u)e| u2 i fauo =v [alae Gt) = fale. Letting i one has “= os0) = flue Gt) ~ flea, j , | = [ou Y a2 1 < fra J ac ce) = falu)|2d¢. a a When ¢ — 0 then t¢v > 0 ae. in 2 and hence Salut t6v) = fa(u) 4 0 ae. in 2. fut tv) — f(u) PH fluo Since Mala, u(2) + tGo(2)) f.(2,ul2)))? < const the Lebesgue Dominated-Convergence Theorem implies fists + Ctv) — f.(u)|?d¢ > 0, as t — 0, (2.13) a and (2.12) follows.20 1 Differential calculus ‘The preceding theorem is completed by the following proposition. Proposition 2.8 Let {2 CIR" be bounded, suppose f and f, satisfy (C) and (2.11) and let f be Fréchet-differentiable at some u” € L?. Then there erists a(x), b(z) € M() such that f(z, u) = a(z)u + b(z). Proof. Suppose first that u* = 0 and f(z,0) = 0. Let D(y,6) denote the ball centred at y € © with measure 6. Given z* €Qand ER, consider the functions vs(z) € L?() given by vs(z) = A, for z € D(z*,6), v6(z) = 0, for z € Q\D(z*, 6) Obviously vs + 0 in L? as 5 +0. Recall that if f is Fréchet-differentiable at u* = 0 then f“(0) = def (0). Hence f/(0)u = f,(O)v, and therefore Uses) = fa(O)velcz lleeilz2 By a direct catculation one finds 70, wh 0. 72 tote) sO rabes | fue.) - ston D(z 6) and hence 2 1 2 re | f W@-h0aF] 0,060 (214) D(z" 6) Tf we set va(z) = £(2,d) = fal, O)A)? BN (2.14) becomes ; | da(z)dz 0, as 6 +0. (2.15) D(z" ,6) Since for all y € L' and almost every y € Q there results 1 7 W(x)dz > Vy), asd 0, Divs) we deduce from (2.15) that for all € IR there exists a null set Ny such that for all y ¢ Ny one has yay) = 0.1.2 Continuity and differentiability of Nemitski operators 21 ‘Taking A in a countable dense subset A of Rand letting N =Usea Na one infers that for all y ¢ N and all A € A there results y,(y) = 0 that is, LWA) = fly 0)d. (2.16) Using the continuity of f(z,.), one deduces that (2.16) holds true for all \ € R and almost every y. Lastly, let us set o(z,u) = flc,u+u*) = f(u') One has that g is Fréchet-difforentiable at 0 and (2,0) = 0. Applying the preceding arguments to g, we find f(eutut)—f(ut) = fa(evut)u and the proposition follows. Potential operators We end this section by dealing with potential operators (Definition 1.6). The results will not be used in the remainder of this book but are im- portant in connection with variational problems. Let Hj = H2{9) (where Q is @ bounded domain of IR") denote the usual Sobolev space (sce Subsection 0.5) with scalar product (.|.) 402 and norm ||.||1712. Let n > 2 (ifn = 2,2, the arguments we are going to expand apply as well, with some modifications; see Remark 2.10 below). Suppose f satisfies (C) and n+2_,, sagt (2.17) Here 2° = 2n/(n — 2) and, by the Sobolev Embedding Theorem (see Theorem 0.3 (i)), Hg 4 L?” and ella S const. lel rsa By Theorem 2.2 it follows that, IF(a,s)| f(v) in L20/+2), Set F(z,s) = f f(x,t)dt. a From (2.17) it follows that there exist c,d > 0 such that |F(z,8)| fF(.u2))de where a and / stand for the embedding of H@ into L?”, of L?7/(™+?) into L respectively. Since F satisfies (2.20), Theorem 2.6 applies to F as a map from L?" to L7"/"+2), From Proposition 1.4(b) (derivative of the composite map) it follows that ¢ is differentiable with ou) iv— f fleu(a)ve(ayaz 4 Then, recalling the definition of “gradient” (Definition 1.6), we have the following Theorem 2.9 Let CR" be bounded and suppose f satisfies (C) and (2.17). Then @ is aC functional on H} with gradient Vd(u) = N(u), where N is defined in (2.19). Remark 2.10 If n = 2 then the same result holds assuming f satisfies (2.17) with any o < oo. It suffices to repeat the above arguments using the stronger form of the Sobolev Embeddings when 2 C R?.