Analisi funzionale. Ð On weak Hessian determinants.Nota di LUIGI D'ONOFRIO, FLAVIAGIANNETTIe LUIGIGRECO, presentata (*) dal Socio C. Sbordone.
ABSTRACT. Ð We consider and study several weak formulations of the Hessian determinant, arising by formal integration by parts. Our main concern are their continuity properties. We also compare them with the Hessian measure.
KEY WORDS: Hessian determinant; Schwartz distributions; Hessian measure.
RIASSUNTO. ÐSui determinanti hessiani deboli.Consideriamo ed esaminiamo varie formulazioni deboli del determinante hessiano, definite come distribuzioni di Schwartz mediante integrazione per parti, principalmente riguardo alle loro proprietaÁ di continuitaÁ. Confrontiamo inoltre tali formulazioni deboli con la misura hessiana.
1. INTRODUCTION
In this paper we discuss several weak formulations of the Hessian determinant which still attract a great interest, seee.g.[3-5, 17, 10, 12]. We begin with the case of functions of two variables. LetVR2be an open set andube a function defined onVwith second order derivatives. We denote byHuthe Hessian determinant
HudetD2u uxx uxy
uyx uyy
:
Ifuis sufficiently smooth, the following identities can be easily checked: Hu(uxuyy)x (uxuxy)y(uyuxx)y (uyuxy)x
1
2(u uxx)yy 1
2(u uyy)xx (u uxy)xy
(uxuy)xy 12(u2x)yy 12(u2y)xx:
On the other hand, each of the above expressions can be used to define a Schwartz distribution onV, providedubelongs to an appropriate Sobolev space. The first one we consider is the distribution of order zero defined by the rule
H0u[W]
Z
V
Hu(x;y)W(x;y)dx dy
for all test functionsW2D(V). Of course, the natural assumption for definingH0uis u2Wloc2;2(V), as it implies that the Hessian determinantHuis locally integrable onV.
Other distributions will result integrating formally by parts. Precisely, we define H1u[W] Z V (uxuxyWy uxuyyWx)dx dy Z V (uyuxyWx uyuxxWy)dx dy H2u[W]12 Z V (u uxxWyyu uyyWxx 2u uxyWxy)dx dy H 2u[W]12 Z V (2uxuyWxy ux2Wyy u2yWxx)dx dy for allW2D(V).
In general,H1uis a distribution of order one. It is well defined foru2W2; 4 3
loc. Indeed, by Sobolev imbeddingux;uy2L4loc, thusuxuxy anduxuyy are locally integrable and the first integral in the definition of H1uconverges. The same argument shows that the
second integral also converges for u2W2;43
loc. We remark that the two integrals give the same distribution. Therefore, we can also write
H1u[W] 12
Z
V
Wx(uxuyy uyuxy)Wy(uxxuy uxyux)dx dy:
REMARK1.1. The definition ofH1uis related to the well-known concept of weak (or distributional) Jacobian introduced by Ball [2]. Actually, H1u is precisely the weak
Jacobian of the gradient mapDu.
To justify the definitions ofH2uandH2uwe again use Sobolev imbeddings: Wloc2;2W2;43
locWloc2;1Weloc2;1 C
0 Wloc1;2
1:1
where we denoted byWeloc2;1the space ofWloc1;1-functions whose second order distributional derivatives are Radon measures onV. Therefore, clearlyH2umakes sense foru2Wloc2;1, or even foru2We2;1
loc (with obvious interpretation ofuxx,uyy anduxyRadon measures). Surprisingly, H2u makes sense for u2Wloc1;2 and does not require second order derivatives.
Thus, we are imposing weaker and weaker conditions onu. It will be clear that, if we are in a position to define two of the above expressions, they yield the same distribution. For example, ifu2Weloc2;1, we can define bothH2uandH2u, and we have
H2u[W]H2u[W] 1:2
1.1. Weak Hessian inRn.
To deal with functions of several variables, we will use the formalism of differential forms which is quite effective. LetVRnbe an open subset and letu2C1(V). The
Hessian determinantHudetD2uinduces an-form:
Hu dxdux1^ ^duxn:
Forj1;. . .;n, setvjdux1^ ^du^ ^duxn, whereduis at thej-th factor in the
wedge product, while fori6jthei-th factor isduxi. Then
@
@xj vj dux1xj ^ ^du^ ^duxn
dux1^ ^duxj^ ^duxn dux1^ ^du^ ^duxnxj : 1:3
The first term in the right hand side (assumingj61) will appear also in@@x
1v1, but with
opposite sign becausedux1xj duxjx1 andduare exchanged to each other. Similarly for
the other terms in the right hand side of (1.3), except for
dux1^ ^duxj ^ ^duxnHu dx: Hence we have Xn j1 @ @xj vj n dux1^ ^duxn; that is dux1^ ^duxn 1 n Xn j1 @ @xj dux1^ ^du^ ^duxn: 1:4
Identity (1.4) yields the definition of the distributionH1u. We multiply both sides by a
test functionWand integrate by parts in the right hand side: H1u[W] 1n Xn j1 Z V Wxjdux1^ ^du^ ^duxn: 1:5
Using Laplace theorem for determinants, this is easily seen to equal H1u[W] 1n Xn j1 Z V uxjdux1^ ^dW^ ^duxn:
As above,dWis thej-th factor.
We proceed further from (1.4). One way is to notice that for eachj1;. . .;nwe have
vj( 1)j 1d(u dux1^ ^duxj 1^duxj1^ ^duxn):
Inserting this into (1.4), multiplying by the test functionWand integrating by parts twice will lead to the definition ofH2u:
H2u[W]1n Xn j1 Z V u dux1^ ^duxj 1^dWxj ^duxj1^ ^duxn: 1:6
To defineH
2u, fori6j we write
vj( 1)i 1d(uxidux1^ ^du^ ^duxi 1^duxi1^ ^duxn):
Now we sum with respect toi and insert into (1.4). In a routine manner, this gives (1.7) n(n 1)H 2u[W] X i6j Z V uxidux1^ ^du^ ^duxi 1^dWxj ^duxi1^ ^duxn:
We recall that in each termduis thej-th factor anddWxj is thei-th factor of the wedge product.
Now we consider functions uin Sobolev spaces. If u2W2;nn21
loc (V), then clearly the minors of ordern 1 ofD2uare locally integrable with exponentn2=(n2 1), while by
Sobolev imbedding theorem the first order derivatives of uare locally integrable with exponentn2. Noticing that these exponents are HoÈlder conjugate, we easily see thatH1u
makes sense:
H1u:W2;
n2 n1
loc (V)!D0(V):
For H2u we recall that functions in Wloc2;n 1(V) are continuous (actually, HoÈlder continuous with exponent (n 2)=(n 1) if n>2). For H2u we remark that if
u2W2;nn22
loc (V), then the minors of ordern 2 ofD2uand the first order derivatives of uare locally integrable respectively with exponentsn2=(n2 4) andn2=2. Therefore,
H2u:Wloc2;n 1(V)!D0(V); H2u:W2;
n2 n2
loc (V)!D0(V):
We notice that in dimensionn>2 the definition ofH2urequires thatuhas second order derivatives.
1.2. The Hessian measure.
Another concept of weak Hessian is classically introduced for defining generalized solutions of the Monge-AmpeÁre equation, see [9] and references therein; see also [18] for further developments. Now, we assume thatVRnis a convex set anduis a real-valued convex function defined onV. We denote by@uthe subdifferential (also called normal mapping) ofuand by@u(E) the image by this multifunction of the subsetEV:
@u(E)[
x2E
@u(x):
The Hessian measure ofu, which we denote byMu, is a Borel measure onVdefined by
Mu(E)Ln(@u(E))
1:8
for each Borel subsetE.Lnis the Lebesgue measure inRn. It is well known that for
sense that the following equality holds Z V WdMu Z V WHu dx; 1:9
for all continuous Wwith compact support.
We recall that a convex functionuis locally Lipschitz and its gradientDuhas locally bounded variation, see [6]. Henceu2Weloc2;1and in dimension 2 we can considerH2u. In
the following theorem we compare these notions.
THEOREM 1.2. If u is in the domain of two of the above weak formulations of the
Hessian, these formulation coincide.
As an illustration, in dimension n2 if u is convex, then MuH2uH2u.
Theorem 1.2 will be an easy consequence of the continuity properties we discuss in the next section.
2. CONTINUITY PROPERTIES
One of our concern is to elucidate on the continuity properties of those weak formulations of the Hessian. First, we note trivially thatH1,H2andH2are continuous
in their corresponding domains with respect to the strong convergences. More precisely:
Ifuh !uinW2; n2 n1
loc , thenH1uh!H1uin the sense of distributions. This means
thatH1uh[W]!H1u[W], for allW2D(V);
Ifuh!uinWloc2;n 1, thenH2uh!H2uin the sense of distributions.
Forn2, ifuh!uinWloc1;2, thenH2uh !H2uin the sense of distributions. Forn>2, ifuh!uinW2;
n2
n2
loc , thenH2uh!H2uin the sense of distributions.
It is also easy to prove convergence in D0 of the weak Hessians of u
h coupling a suitable weak convergence of the minors of the Hessian matrix D2u
h with a strong convergence of the functionuhor its first order derivativesDuh. As an example, we have
If Duh!Du strongly inL n2
2
locand (forn>2) the minors of order n 2 ofD2uh converge to corresponding minors ofD2uweakly inLn2 4n2
loc , thenH2uh !H2uin the
sense of distributions.
Next, we examine continuity properties with respect to weak convergences. It is enough to consider the weakest formulation;i.e.H
2. We first consider the casen2. In
[10] it is shown by an example that H
2 is not continuous with respect to weak
convergence inWloc1;2. Here we present another simple example of this type. EXAMPLE2.1. The functions
clearly converge uniformly to zero andDuhweaklyinL1. On the other hand,H2uhare not converging to zero in the sense of distributions. To see this we consider a test function
Wsupported in the unit diskDR2, and compute
H2uh[W] Z D
(4xyWxy 2x2Wyy 2y2Wxx)cos2h(x2y2)dx dy: Integrating in polar co-ordinates and then by parts, we find that
lim h H 2uh[W] Z D (2xyWxy x2W yy y2Wxx)dx dy2 Z D Wdx dy:
The above example and the following arguments indicate clearly that in order to ensure convergence ofH
2uh toH2uone really needs strong convergence inWloc1;2. THEOREM 2.2. Let uh 2Wloc2;1, h1;2;. . ., and u2Wloc2;1 be given. Then the dis-tributionH2uhH2uhconverge toH2uH2u under each of the assumptions below:
(a)uh*u weakly in Wloc2;1;
(b)the sequence fuhg is bounded in W2;1
loc and converges to u uniformly on compact subsets.
The proof follows easily when we realize that both assumptions (a) and (b) yield
uh!ustrongly inWloc1;2. Actually, we have
LEMMA2.3. (a))(b)) uh!u strongly in Wloc1;2. We remark that both imbeddings ofWloc2;1C0andW2;1
loc Wloc1;2are not compact, in the sense that bounded sequences inWloc2;1may not have subsequences converging locally uniformly, or strongly inW1;2
loc.
EXAMPLE 2.4. Given a function inu2W2;1(R2) with u60 and lim
z!1u(z)0, we
define uh(z)u(hz), 8h2N. Then uh(z)!0, 8z2R2 f0g, Duh!0 in L1,
kD2u
hk1 kD2uk1for allh2NandD2uh* Cdweakly in the sense of measures, with
Ca constant matrix. Moreover fuhg has no subsequence converging locally uniformly and, askDuhk2 kDuk2 >0 for allh, no subsequence strongly converging inW1;2
loc. On the other hand, the assumption of weak convergence inWloc2;1is essentially stronger than mere boundedness. By Ascoli-ArzelaÁ theorem, under assumption (a) we can prove uniform convergenceuh!uon compact subsets, hence condition (b) holds.
PROOF OFLEMMA2.3. There is no loss of generality in assumingu0, otherwise we consider the functionsuh u.
Let us prove (a))(b). For simplicity, we assumeV]0;1[2. ForW2C1 0 (V), we
definevh Wuh,h1;2;. . ., and we have
DvhuhDWWDuh;
D2v
huhD2WDWDuhDuhDWWD2uh; hencevh*0 weakly inW2;1(V).
The point is that weak convergence offuhggives equi-integrability offDuhg. It is well known that W2;1(V) is continuously imbedded into C(V), see [1, p. 100]. If v is a
continuous function of classW2;1(V) vanishing on@V, then for all (x
1;y1) and (x2;y2) in V, we have jv(x1;y1) v(x2;y2)j4 Zx1 0 ds Zy1 y2 vxy(s;t)dt Zx1 x2 ds Zy2 0 vxy(s;t)dt : In particular, kvk14kvxyk1:
These inequalities yield equiboundedness and equicontinuity of any bounded family of functions inW02;1(V) with equi-integrable second order derivatives. We are in a position to apply Ascoli-ArzelaÁ theorem and get vh!0 uniformly in V. This implies uniform convergenceuh!0 on compact subsets ofV. IfKVis compact, it suffices to choose
W1 onK.
Now we assume that condition (b) holds and prove strong convergence inWloc1;2(V). If
v2W02;1(V), integrating by parts we find Z V jDvj2 Z V vDv4kvk1kD2vk 1:
Therefore, defining vhWuh as above, we see that vh!0 strongly in W1;2(V). To deduce from this strong convergence ofuhon an arbitrary compactKV, as before we only need to takeW1 onK.
Notice that (b) implies strong convergence of fuhg inWloc1;2 can be seen also as a
consequence of inequality (1.16) of [13]. p
Theorem 2.2 can be slightly generalized.
THEOREM2.5. Letfuhgbe a sequence bounded inWeloc2;1and converging to u uniformly
on compact subsets. Then u2Weloc2;1andH2uh H2uhconverge toH2uH2u in the sense of distributions.
Now we assumen>2. We have the following continuity property.
THEOREM2.6. If uh *u weakly in Wloc2;pfor some p>nn22 , thenH2uh!H2u in the
PROOF. As p> n
2
n2 the imbedding Wloc2;pW1; n2
2
loc is compact. Then Duh!Du strongly in Ln22
loc. On the other hand, the minors of order n 2 of D2uh converge to corresponding minors of D2u in the sense of distributions, see [11, Theorem 8.2.1,
p. 173], and are bounded inLnp2
loc, where we notice thatn p2> n 2
n2 4>1, hence they
are weakly converging inLnp2
loc and therefore we can pass to the limit in the expression
definingH
2u[W]. p
REMARK2.7. In [5] it is shown that the conditionp> n
2
n2in Theorem 2.6 is sharp, in the sense that the conclusion cannot be drawn in general (in dimensionn>2) assuming
uh*u weakly in W2; n2 n2
loc :
2:1
With this condition, we still have weak convergence inLn2 4n2
loc of the minors of ordern 2
ofD2u
h, but we need to reinforce (2.1) to guarantee the strong convergence inL n2
2 locof first
derivatives. For 1<n
2<r<n, we have the Sobolev imbeddings
Wloc2;rWloc1;r C0;2 nr
loc ;
neither of which is compact. Clearly, ifuh!ustrongly inW1;r
loc, then it converges also strongly in the HoÈlder space. On the other hand, by the interpolation inequality (5) of [16], for a bounded sequencefuhginW2;r
locthe convergence in the HoÈlder space implies strong convergence of first derivatives, that is, the two convergences are equivalent. For
rnn22, we have the following generalization of Theorem 2.6, in the spirit of Theorem 2.5:
Let condition(2.1)hold together with uh!u strongly in C0;1
2
n
loc . ThenH2uhconverge
toH
2u in the sense of distributions.
We conclude this section by recalling some continuity properties of the Hessian measure, see [9].
THEOREM2.8. Let uh, h1;2;. . ., and u be convex functions and let the sequence uh
converge to u uniformly on compact subsets. Then Muh converges to Mu weaklyin the
sense of measure, that is,
Z V WdMuh ! Z V WdMu for all continuous functionsWwith compact support.
It is clear that in the two-dimensional context Theorem 2.8 follows from Theorem 2.5.
3. THEL1-PART OF THEHESSIAN
One concern of the study of the weak formulations of Jacobian and Hessian determinant is to relate them with the point-wise definitions. In [14] it is proved that the point-wise Jacobian coincides with the distributional one, provided the latter is a locally integrable function. More generally, assuming that the distributional Jacobian is a Radon measure, then the point-wise Jacobian is the density of its absolutely continuous part with respect to the Lebesgue measure. In a different direction, in [8] the identity between distributional and point-wise Jacobian is established without assuming a priori that the former is a distribution of order 0, but under suitable integrability conditions for the differential matrix.
We mention also that in [15] examples are given showing that the support of the singular part of the distributional Jacobian inRncan be a set of any prescribed Hausdorff dimension less thann.
The result of [14] immediately gives that the point-wise Hessian is the density of the absolutely continuous part ofH1uforu2W2;
n2 n1
loc , assuming thatH1uis a measure. This is generalized in [10] showing that the point-wise Hessian is the regular part of the distributionH2u. A further extension is given in [12] forH2u. Here we remark that an
analogous result holds for the Hessian measureMu. Given a Radon measuremonV, we denote by ma the density of its absolutely continuous part with respect to Lebesgue measure. We recall that
ma(x)limh mrh(x); for a:e: x2V;
where frhg is a sequence of mollifiers. If u is a convex function, its second order derivativeD2uis a (matrix-valued) measure.
THEOREM3.1. For every convex function u the density of the absolutely continuous part
of the Hessian measure Mu isdet(D2u)
a, that is (Mu)adet(D2u)a:
3:1
In particular,det(D2u)ais locally integrable.
DenotingMubydetD2u, we formally rewrite (3.1) as
(detD2u)
adet(D2u)a:
Therefore, we can state the result of Theorem 3.1 by saying that the two operations of computing the Hessian determinant and of passing to the density of the absolutely continuous part commute.
To prove Theorem 3.1, we begin with a lemma.
LEMMA 3.2. Let fh2L1loc(V), h1;2;. . .; be functions verifying fh50, 8h,
fh(x)!f(x)a.e. and fh* mweakly in the sense of measures. Then we have f(x)4ma(x)
PROOF. For any continuous functionW50 with compact support, by Fatou lemma we find Z V Wf dx4lim h Z V Wfhdx Z V Wdm:
This inequality with arbitrary W implies that f 2L1
loc(V). Moreover, considering a
sequencefrhgof mollifiers, for a.e.x2Vwe have
f(x)lim
h f rh(x)4limh mrh(x)ma(x):
p
More generally, ifmh,h1;2;. . ., are nonnegative Radon measures andmh* m, then lim inf
h (mh)a4ma:
PROOF OFTHEOREM3.1. By convolution, we can find a sequence of smooth convex functionsuhsuch thatuh!uuniformly on compact subsets ofV,D2uh!(D2u)aa.e. in
V and hence also detD2u
h !det(D2u)a a.e. in V. Notice that for all h we have
MuhdetD2u
h. Recalling the continuity property of Theorem 2.8, by Lemma 3.2 we get
det(D2u)
a4(Mu)a; a:e:inV:
To prove the opposite inequality, we use the change of variable formula, see [7, Theorem 1, p. 72]. Accordingly, there exists a Borel set RuV such that uis twice differentiable inRu,Ln(VZnRu)0, and A det(D2u)adx Z Rn Nu(y;A)dy; 3:2
for every measurable subsetAofV. HereNu(y;A) denotes the Banach's indicatrix
Nu(y;A)#f x2A\Ru : Du(x)yg: From (3.2) we deduce Z A det(D2u) adx5Ln Du(A\Ru)Mu(A\Ru); which clearly implies
det(D2u)a5(Mu)a:
p
REMARK3.3. Taking into account the continuity property of Hessian measures stated in Theorem 2.8, it is interesting to notice that weak convergence of measures does not imply point-wise convergence of the densities of absolutely continuous parts. As an example, we recall that the Lebesgue integral of a continuous function is the limit of integral sums, which can be considered as integrals with respect to purely atomic measures.
REFERENCES
[1] R.A. ADAMS,Sobolev spaces. Pure and Applied Mathematics, vol. 65, Academic Press, 1975.
[2] J.M. BALL, Convexity conditions and existence theorems in nonlinear elasticity. Arch. Rational Mech. Anal., 63, n. 4, 1976-77, 337-403.
[3] J.M. BALL- J.C. CURIE- P.J. OLVER,Null Lagrangians, weak continuity, and variational problems of arbitrary
order. J. Funct. Anal., 41, 1981, 135-174.
[4] R. COIFMAN- P.-L. LIONS- Y. MEYER- S . SEMMES,Compensated compactness and Hardy spaces. J. Math. Pures Appl., 72, 1993, 247-286.
[5] B. DACOROGNA- F. MURAT,On the optimality of certain Sobolev exponents for the weak continuity of
determinants. J. Funct. Anal., 105, n. 1, 1992, 42-62.
[6] L.C. EVANS- R.F. GARIEPY,Measure Theory and Fine Properties of Functions. CRC Press, Boca Raton, FL 1992. [7] M. GIAQUINTA- G. MODICA- J. SOUCÏEK,Area and the area formula. Rend. Sem. Mat. Fis. Milano, 62, 1992,
53-87.
[8] L. GRECO,A remark on the equalitydetDfDetDf. Differential Integral Equations, 6, n. 5, 1993, 1089-1100.
[9] C.E. GUTIEÂRREZ,The Monge-AmpeÁre Equation. BirkhaÈuser, Boston 2001.
[10] T. IWANIEC,On the concept of the weak Jacobian and Hessian. Report. Univ. JyvaÈskylaÈ, 83, 2001, 181-205. [11] T. IWANIEC- G. MARTIN,Geometric Function Theory and Nonlinear Analysis. Oxford University Press,
Oxford 2001.
[12] J. MALYÂ,From Jacobians to Hessian: distributional form and relaxation. In:Proceedings of the Trends in the
Calculus of Variations (Parma, September 15-18, 2004). Riv. Mat. Univ. Parma, in press; http:// www.math.cmu.edu/nw0z/publications/05-CNA-003/003abs/05-CNA-003.pdf.
[13] C. MIRANDA,Su alcuni teoremi di inclusione. Annales Polonici Math., 16, 1965, 305-315.
[14] S. MUÈLLER, Detdet, a remark on the distributional determinant. C.R. Acad. Sci. Paris, SeÂr. I Math., 311, 1990, 13-17.
[15] S. MUÈLLER,On the singular support of the distributional determinant. Ann. Inst. H. Poincare Anal. Non LineÂaire, 10, 1993, 657-698.
[16] L. NIRENBERG,An extended interpolation inequality. Ann. Sc. Norm. Sup. Pisa Sci. fis. mat., 20, 1966, 733-737.
[17] P.J. OLVER,Hyper-Jacobians, determinantal ideals and weak solutions to variational problems. Proc. Roy. Soc. Edinburgh Sect. A, 95, 1983, n. 3-4, 317-340.
[18] N.G. TRUDINGER,Weak solutions of Hessian equations. Comm. Partial Differential Equations, 22, 1997, n. 7-8, 1251-1261.
______________
Pervenuta il 15 febbraio 2005, in forma definitiva il 31 maggio 2005.
L. D'Onofrio: Dipartimento di Statistica e Matematica per la Ricerca Economica UniversitaÁ degli Studi di Napoli «Parthenope» Via Medina, 40 - 80100 NAPOLI luigi.donofrio@uniparthenope.it F. Giannetti, L. Greco: Dipartimento di Matematica e Applicazioni «R. Caccioppoli» UniversitaÁ degli Studi di Napoli «Federico II» Via Cintia - 80126 NAPOLI giannett@unina.it luigreco@unina.it