Jl(u) = f
u(x), Du(x)) dx,
92 c 1R".
a Let u be an F-extremal of class CZ(S2, R'), and set Q
(3)
'E(u) :_ {v e C1(D, IR"): JI v - u ll o, n < s, V I DO = u 1 aft } .
In order that M becomes a calibrator for IF, u, WP}, 0 < a << 1, we have to construct a geodesic slope field )6(x, z) = (x, z, 9(x, z)) in a neighbourhood G of graph u in 1R" x IR" such that u fits 4. Because of the highly nonlinear character of M with respect to p the corresponding field theory, due to Caratheodory, is much more involved than the De Donder-Weyl field theory corresponding to the divergence-type calibrators used in 4.1. For instance Legendre's transformation is a very appropriate tool for the De Donder-Weyl theory while it seems to be much less useful for Caratheodory's field theory. Here Caratheodory replaced it by another involutory transformation which he called a generalized Legendre transformation. This terminology is somewhat misleading since for n = 1 Caratheodory's transformation does not reduce to the ordinary Legendre transformation but to Haar's transformation (cf. 10,3.2), a composition of a
Legendre transformation and a Holder transformation. While under suitable (and reasonable) conditions on F Haar's transformation generated by F is a global diffeomorphism, similar results seem still to be lacking for Caratheodory's
transformation in case that n > 1; here we only can formulate natural conditions guaranteeing that it is a local diffeomorphism. We begin our discussion by developing some parts of Caratheodory's transformation formalism needed for the purpose of the calculus of variations. We do not use Carathbodory's ingenious notation which is very suggestive but requires a certain interpretation skill since the difference between dependent and independent variables is not always clear. Our admittedly much less elegant notation might be more instructive in this regard.
4.2. Caratheodory's Field Theory
107
As in 4.1 we fix a domain G in IR" x IRN as configuration space and a C2-Lagrangian F(x, z, p) on G := G x IR"N. Then we define the following expressions as functions of (x, z, p) E G, i.e. as fields on d: 7rf := F t
(4)
7r = (70)
,
Qa:=pang-FSa, a=(aa); A:=deta;
(5) (6)
b #":= cofactor of of in det a,
(7) (8)
ck
b = (bi);
Park - 5,F, c = (ck); C:= det c ;
(9)
of :=
(10)
A
b'7rf,
-707rk], R=(Rk)
Rk '- Fik
(11)
7 = (rlf);
(Rows are indicated by lower indices, columns by upper indices.) Our basic assumptions on F are the following:
(i) The functions F and A satisfy
FOO and AO0.
(12)
(ii) The matrix function (Rk) is positive definite, R > 0, that is, (13)
0.
(iii) The mapping -4F : G - IR" X 1RN x ]R"N defined by
-qF(x, z, P) :_ (x, z, q),
(14)
q:= 1(x, z, P),
is a (local) diffeomorphism of G onto d* :_ R (d). As we shall see later, assumptions (i) and (ii) imply that RF is a local diffeo-
morphism. For n = 1 these two assumptions even imply that QF is a global diffeomorphism, but we do not know whether this conclusion holds also true for
n> 1. The transformation RF is denoted as Caratheodory transformation, and the function K(x, z, q) defined by (15)
K = (-F)"-l
o'
F
A
is said to be the Caratheodory transform of F. As Caratheodory transformation in the general sense we denote the two-step procedure passing first from the variables (x, z, p) to the variables (x, z, q) via (14), i.e. by (16)
of = Abbd g
Chapter 7. Legendre Transformation, Hamiltonian Systems, Convexity, Field Theories
108
and then from the Lagrangian F(x, z, p) to the Caratheodory function K(x, z, q). (This resembles the Legendre transformation IF where one first passes from (x, z, p) to (x, z, it) via 70 = F i(x, z, p) and then from F(x, z, p) to the Hamiltonian O(x, z, 1r) defined by fi = (p,,FF, - F) o F t ) Now we want to derive several formulas describing Caratheodory's trans_ First we note that formation -RF and its inverse RF1
Aa-1
(17)
= b,
whence
as bf = 6.1A,
(18)
albs = 45,1,'A.
Let e be the n x n-unit matrix. Then we have
Ae = ab, whence
A" = det(Ae) = det ab = (det a)(det b) = A det b and therefore
detb=A' '
(19)
We infer from (10) and (17) that
= na-1,
(20)
whence (21)
la,
7r
that is 7ri =
(21')
ase1,a
Note that
(-1r'A=
1
0
Sk
0
p
-a
Pa
- am
Since b" id = nfl, -aQ + paick = 6.0F, an obvious transformation of this determinant yields
(-1)"A (22)
gk
,re
pa
5F
and similarly we prove (23)
(-1)NC =
8kF
ruff
Ps
as
From these two equations we derive
" NA (-1)F
F6;`
Fire
Pa
FSQ
= (-1)NF"C,
4.2. Caratheodory's Field Theory
whence
(-F)NA = (-F)"C.
(24)
On account of (12) it follows that 0.
C
(25)
Moreover equations (5) and (10) yield
and now (18) implies (26)
?7apa=S;+Abfi,
and therefore (27)
bp' =
A
F (f1a pi - S0') .
By introducing
9# := 5f - i p8, 9 = (9B),
(28)
9:= det g,
(29)
Fn t
Y':=(
,
A
K=Tog?F,t,
i.e.
F
9o=-Abp By virtue of (19) it follows that
=(-F/A)"detb=(-F)"A, that is, (32)
9_ -FT' 0 0,
and, because of (33)
(-F)n-2
A
F
Y'
equation (31) is equivalent to F)"-2
(34)
Then we write (34) as
b"p
2
P
go
.
109
110
Chapter 7. Legendre Transformation, Hamiltonian Systems, Convexity, Field Theories
go = F(-
(35)
F)n ' be
Set
det g,
in
hB := cofactor of g,
(36)
h := (h").
Since
h=g-' T
and gh =hg=19e,
we have ga hP = hPgo = bQ I
(37)
Then (35), (37) and (18) imply n+ib°aa by
lay = qb; as = g°hpaa = P(_F)-n-'Abahp =
(-F)-n+1'cAhQ
=
and therefore
hQ = Pa!.
(38)
In conjunction with (21') it follows that (39)
Yrca = hQ n
and by virtue of (37) we infer that VIga it
(40)
-
On account of (32) we then obtain
-F>7a = gpir,.
(41)
By (28) we have
1IkPk-Sa-ga P P P and from (8) we derive Ciylk = pk7rpqka
On account of (40) we then obtain a ka that is.
1t = Cj tlk
(42)
The above formulas suffice to construct the Caratheodory calibrator M. However the reader might like to see why Caratheodory's transformation is an involution. To this end we introduce the field C(x, z, p) _ z, p)) on d by F"-z
(43)
:_
(-1)"-1
A
ad Po = F. as Pa
4.2. Caratheodory's Field Theory
By virtue of(38) we can write Ca as
Fps = ha p'
(44)
From (10) and (43) we infer that =(-1)n-1Fn zA 2Rnpobva,
and (5) yields a17 + FSQ .
It follows that '1a (p
(-1)n-'F" ZA 2(a° + FSQ )bpa .
=
Since
aQb, a = Ab,aB =
6,b, a' = bQao = ASB,
we obtain
a _ (F) "
l (6
+
-a;) = (Fbp +as).
On account of (5) we arrive at
////
FrlaC@ = snap'.
(45)
Let us introduce the fields 2(x, z, p) and uc(x, z, p) on G by (46)
(47)
C - `PBS, A_ (AP, 4:= cofactor of Aa in det,, 14=(4), 2
and set
A:=detA.
(48)
We have on account of (45)
a = (F 11%' - Fbo I = 1V
and in conjunction with (5) we arrive at (49)
.1° Y1
=
(n°pl - FSf ),
F
ae°
F
AMY-" = AF-"
F-n+16p .
111
112
Chapter 7. Legendre Transformation, Hamiltonian Systems, Convexity, Field Theories
Moreover, (44) implies IN6. p p
F9a ce = 9a hP p'
and thus we infer from (31) that
Pa = (F/5)(-F/A)ba
.
By (32) we obtain
Pi= Fb!('
(52)
=
Finally we infer from (50) and (30) that
il^' = AA-'F' = (-1)n-'AF(-1)n-to-1Fn-t (-1)n-1AFY',
_ whence
F-
(_ _)n-1
(53)
A
Thus we infer from (52) by means of (51) and (53) that
Pai-1 -
(54)
i
µa Ce
and (21'), (49), (53) yield n 2
a=
(55)
1)n-1
A
a_
A511
Now we take the total differential of the determinant I of (gd). It follows that
dT = h, dgs , and (32) implies
-dg =Fd'l'+WdF. By (28) we have
-dg" = n° dpi + p' dn5 and thus we see that
F dtP+'I'dF = hPrle dp, + hap' drl" On account of (39) and (44) we arrive at (56)
F(dtW - dad,) + V(dF-nadpp)=0.
This is the key identity from which we shall derive the involutory character of Caratheodory's transformation. Because of (4) we can write equation (56) as (57)
F(dW - Ca drl;) +
dx" + Y'F=; dzi = 0.
4.2. Caratheodory's Field Theory
113
Recall that MF(x, z, p) = (x, z, q), (58)
q=i(x,z,p),
K(x, z, q) = `F(-qF1(x, z, p)),
and set (59)
va
aa
9F1,
v(x, z, q) = (va(x, z, q)).
Taking the pull-back of (57) under MF', we then obtain (60)
(FoMF1)[dK-v'dqF]+K(FX,oAF')dx"+K(F2,oMF')dz`=0,
whence
vi = Kq;
(61) and
(FoRF1)KX,= -K(Fx,o. F'), (62)
(Fo9PF')Kz,= -K(FZ,oRF').
From (4)-(7) and (10) and the corresponding equations (61), (46)-(48) and (54) we read off that BF' is obtained in the same way from K as R. is generated by F, that is,
AF1 = Rx, where K is the Caratheodory transform of F. If we write (63)
ay = (64)
F. A = det(a'),
bb = cofactor of a# in A , and
eo =gfKga - K, E =det(e!), (65)
f! = cofactor of e' in E,
then the full symmetry of Caratheodory's formalism is expressed by the following relations:
KA=(-F)"-', FE=(-K)"-1, EP=AK", q;= Ibp"Fi, pa=Ef/Kq,,
(66)
FKX,+KF,,=0, FKZ,+KFZ,=0, Fq,Kgo = Kp,6Fp;.
Here we use the following sloppy but rather instructive notation: The quantities in (64) mean the values
F=F(x,z,p), Fz,=Fx,(x,z,p), ...,
bp" =bb(x,z,p),
Chapter 7. Legendre Transformation, Hamiltonian Systems, Convexity, Field Theories
114
and similarly we use in (65) the abbreviations
K = K(x, z, q), K,()c, z, q),
,
f"I = fa(x, z, q),
and the variables (x, z, p) and (x, z, q) are linked by (x, z, q) = -RF(x, z, p)
(Note that E = A o
or (X, z, P) = g?K(x, z, q).
RF1.)
Now we want to study the invertibility of Caratheodory's mapping AF : (x, z, p) H ()c, z, q), q = n (x, z, p), where
)1 =Abbnk (see (4)-(14)). In our basic assumptions on F we had required that (i) FA # 0, (ii) (R k) > 0, and (iii) AF is a diffeomorphism or at least a local diffeomorphism. Now we want to show that (iii) is superfluous since it follows from (i) and (ii), more precisely, we shall prove that (i) and (ii) imply that 9F is a local diffeomorphism. Since QF is given by the system of equations
x=x,
z=z,
q=rl(x,x,p),
it sufficies to show that the Jacobian det nP does not vanish. Let us introduce the functions WB(x, z, p, q) defined by
Wa := ns -(pank - FSk)q¢.
(67)
Since the system of equations IF(x,z,p)=qa
is equivalent to the system
Wa(x,z,p,q)=0, 1<x
det
(68)
apa9
\
W° I A 0 on {q = r1 (x, z, p)} .
We have
and a
a
appkk
[po1ri ] = gf'nk + q°po a k r p6
thus (69)
Wa = ap6
app
70 + qi nk - qk nk - q9 po
ni . ape
4.2. Caratheodory's Field Theory
115
Now we introduce the matrix L k(x, z, p, q) by
a
L 'fl
(70)
c; 8Q pk
W,° = c, ap
W"'
Since the nN x nN-matrix c' b,' has the determinant CN where C := det(c;) 0 0, we conclude that (68) holds true if and only if
det(L k) # 0
(71)
on {q = ri(x, z, p)}.
We are now going to verify that assumptions (i) and (ii) imply inequality (71). From (69) and (70) it follows that a
(72)
a
L k = c ask + c, q;rc - c;
Cr q: Po
PB
ani
pp
Suppose now that p and q are related by qa = r1"(x, z, p). Then we have proved earlier (cf. (42)) that
a; =cigka k
and thus it follows from (72) that a
Lit = -(
G971'
+ 7rFpa) apk +(r<°nk V B
By virtue of (8) we have
-.+p;n,°=6jF, and so we obtain (73)
L"k =
ana
1
F lepkkF
nenk)
.
B
On account of (11) we thus have proved that (74)
L aft = -FR,k on { (x, z, p, q) : q = ri(x, z, p)}.
Then the basic assumptions (i) and (ii) imply that (75)
(Lk)>0 ifF<0, (Lk) <0 ifF> 0,
provided that q = ri(x, z, p),
and in particular we have (75')
det(Lak) 0 0
provided that q = ri(x, z, p).
Thus we have verified that AF : G -+ G* is a local djeomorphism, i.e. for any (xo, zo, Po) e G there is an open neighbourhood do of (xo, zo, Po) in G such that AFIG0 furnishes a d(eomorphisrn of do onto some open neighborhood d0* of (xo, zo, q0), q0 = f1(xo, zo, po)
Now we want to define a transversality relation between N-dimensional surfaces 9 and n-dimensional surfaces .K in the configuration space W" =
116
Chapter 7. Legendre Transformation, Hamiltonian Systems, Convexity, Field Theories
1R" x R'. Suppose that X is the graph of a smooth mapping u : 0 -+ IR" £2cIR",i.e. ," = {(x, Z): Z = u(x), X E S2},
and that 9 is the graph of a smooth map : 92' -+ R", £2' c IR", i.e. ,92 = {(X, Z): x = (Z), Z E SZ' } .
Let P = (x, z) be a point in .9' n .N'. The tangent space T Y of .N' at P is spanned by the rows of the n x (n + N)-matrix (8Q,
while the tangent
space TF.92 of .9' at P is spanned by the rows of the N x (n + N) matrix (C: (z), b;`). Hence we can characterize T.K and TpY by the "elements" e = (x, z, P), Pa = u{-(x}, and e = (x, z, q), q, = i (z) Assume now that e = (x, z, p) is an arbitrary element in G, and let e = (x, z, q) be an arbitrary element in G*.
Definition 1. Two elements e and a are said to be transversal (in the sense of Caratheodory) if e = AF(e).
Note that transversality is a one-to-one relation between the elements of d
and G* if and only if AF is a global diffeomorphism. If PlF is only a local diffeomorphism, then only the elements e of sufficiently small neighborhoods do are in 1-1 relation to the elements a of yPF(Go) = Go.
Definition 2. Let .K and 9 be n- and N-dimensional surfaces as described above. We say that .N' and So intersect transversally (in the sense of Caratheodory) if for all P = (x, z) e X r ),V the tangential elements e and s of Tp.K and Tp.9' respectively are transversal. (Here we have tacitly assumed that the tangential elements of .N' and .9' lie in G and G* respectively.)
The reader may check that, for n = 1, Carathbodory's transversality reduces to Kneser's transversality, i.e. to the notion of free transversality between curves and hypersurfaces introduced in 2,4 (see 10,3.4). It is interesting to check whether two surfaces .4" and 6P intersect transversally in the sense of algebraic geometry (i.e. TEA" + T .9' = IR"+" for P e X n.9', or equivalently Tp.N n T,.9' = {0} since n = dim N = dim Tp') if they intersect transversally in the sense of Carath6odory. This is in fact true as we can
see by the following reasoning. Let P = (x, z) E .N n .P, and assume that the elements e = (x, z, p) and e = (x, z, q) describing TpK and T. are transversal in the sense of Carath6odory, i.e. q = ri(x, z, p). Consider the determinant (76)
8
q#
Pal ajr
we have to show that d 0 0. This follows from
4 2. Caratheodory's Field Theory
sa-Paga
q; -
6i
,
Ri
P
Pa
ga
117
i ,
0
Pa
bji
S;
since the two basic assumptions (i) and (ii) imply
$ 0, cf. (32).
Now we turn to the construction of a Caratheodory calibrator M(x, z, p) of the form (1),
M(x, z, p) := det(S. (x, z) + S";(x, z)p'),
for a suitably chosen mapping S(x, z) = (S' (x, z), ... , S"(x, z)). We try again the approach of 4.1; to this end we assume without loss of generality that Go is of the form Go = Go x B,(Po), where Go is a ball in IR" x IR" centered at Po = (xo, zo), zo = u(xo), and Br(Po) _ {p e IR"". Ip - Pol < r}, r > 0, po = Du(xo), xo e Q. Note that eo = (xo, zo, Po)
is a tangential element of the n-dimensional surface 9 := graph u. Consider now mappings S : Go -+ 1R" and slope fields A : Go -+ Go,
A(x, z) = (x, z, 9(x, z)),
9(x, z) = (9a(x, z)).
We try to find a pair IS, 91 such that u fits y, that is, (78)
{(x, u(x), Du(x)): x e A
} c do,
£. a sufficiently small neighbourhood of xo in 1R", and
Du(x) = 1(x, u(x)) for all x e S2o
(79)
and that the null Lagrangian M defined on Go x 1R"" satisfies
M(x, z, 9(x, z)) = F(x, z, P(x, z)) for all (x, z) e Go
(I)
and M(x, z, p) < F(x, z, p) for all (x, z, p) e Go x R"'.
(II)
Then M is a calibrator for IF, uo, leE(uo)}, uo := ujno, 0 < e << 1, since (78), (79)
and (I) imply M(x, u(x), Du(x)) = F(x, u(x), Du(x)) for all x e Slo , while (II) yields M(x, v(x), Dv(x)) < F(x, v(x), Dv(x)) for all v e'tf (uo) and all xo a fo
Note that we require (II) for all (x, z, p) e Go x 1R"" since we want to prove that uo is a strong minimizer; this matches with the "Legendre condition" R(x, z, p) > 0 of the basic assumption (ii) which is supposed to hold for all (x, z, p) E G x IR" = G. However, Caratheodory's transformation 3PF operates only on Go = Go x Br(po) and not necessarily on Go x lR". This transformation is used for constructing the field A(x, z) = (x, z, 9(x, z)), (x, z) a Go, whose range will lie in Go.
Chapter 7. Legendre Transformation, Hamiltonian Systems, Convexity, Field Theories
118
As in 4.1 we begin our construction of M by deriving some necessary consequences of (I) and (II). Set
F*(x, z, p) := F(x, z, p) - M(x, z, p)
(80)
for (x, z, p) e Go X R"N.
Then (1) and (II) are equivalent to
F*(x, z, 2(x, z)) = 0 for all (x, z) e Go
(1*)
and
F*(x, z, p) > 0
(11*)
for all (x, z, p) e Go x I[ZnN
To simplify the notation we introduce (81)
.Z (x, z, P) := S (x, z) +
(x, z)P
E = (E« )
Then the null Lagrangian (1) can be written as M(x, z, p) = det(Ea (x, z, p)).
(82)
Let Ts be the cofactor of E." in det(I ). Then we have
EQ Tf = 6, M and EPTY = Sa M.
(83)
Furthermore the differentiation rule of determinants yields MP,
and thus we have Mp, = S=,T?.
(84)
We also introduce 17(x, z) = (17,(x, z)) for (x, z) e Go by (85)
17,'(x, z):= Fi(x, z, P(x, z)),
that is,
I1a:=Fp.oA=fi*Fi,
(85')
and q(x, z) = (x, z, 2(x, z)), (x, z) e Go, by
:_ 1F o fz -
(86)
*91F,
i.e.
(86')
The composition of quantities depending on (x, z, p) e do with the mapping 1z will be denoted by the superscript , e.g. (87)
F := F o fe,
F., o fz,
IIi = Fpa := F1, o /1, etc.,
while for quantities depending on (x, z, q) a Go the superscript position with ", e.g. (88)
K:=Ko j,
K=,:=KX,o9,
etc.
means "com-
4.2. Caratheodory's Field Theory
119
Now we are going to exploit (1*) and (11*). If these two relations are satisfied we necessarily have
F* = 0 and Fp = 0,
(89)
F=M Fpi = MPa
Equations (90) and (91) are called Caratheodory's equations for {S, g}. Definition 3. A slope field A (x, z) = (x, z, 9(x, z)) on Go is said to be a geodesic
slope field (in the sense of Caratheodory), or briefly: a Caratheodory field, if there is a map S e CZ(G, IR") such that {S, 9} solves the Caratheodory equations (90), (91). We call S an eikonal map associated with the geodesic field A.
Let us now derive some further relations to be satisfied by geodesic fields. Lemma 1. Suppose that fi : Go --> Go is a geodesic slope field with an eikonal map S and 95 = -4, o fi. Then the null Lagrangian M defined by (81), (82) satisfies
M=Mo/i
(92)
0,
whence
det(Ts) =
(93)
Let
M"_'
0.
(x, z) = (x, z, 9(x, z)), 9(x, z) _ (x, z, 2(x, z)) and
(94)
17i'F-,=a.2f.
Then Caratheodory's equations are equivalent to (95)
F = M,
17ia = Mp;,
and we obtain (96)
as = -S,TP,
(97)
S .92 + Sz' = 0,
(98)
A = (-1)" det(S )Fn-1,
(99)
det(Sxe)
0.
Proof. Relations (92) and (93) follow from F 0 0 and M = F (cf. the proof of (19)), and from M = det(Lf) we infer that (Ea) is invertible. Furthermore (94) is an immediate consequence of (10), while (95) is obviously equivalent to (90) and (91) on account of (94). By (5) and (11) we have
as = Fp - F6. = aMpe - F8R
.
Chapter 7. Legendre Transformation, Hamiltonian Systems, Convexity, Field Theories
120
Taking (83) and (84) into account we infer that Qa = (S7'.1.9
- -'a )T
,
and (81) yields
-S,=SlAi-Z'Q. Combining the last two equations we obtain (96); then (98) is a consequence of (93) and (96), and inequality (99) follows from (12), (92) and (98). Finally, (94) and (96) imply that - IIZY = T'S.""Y,
and Caratheodory's equation (952) in conjunction with (84) yields IT? = Ty Sz;
.
Adding these two equations we arrive at
0 = E's7 [SS{ + SX,2°]
= R6. 11S.", + S2°] = M[S., + Z2°]. By M 96 0 we obtain (97).
Consider now the system of equations (100)
S"(x, z) = 6°,
a = 1, ..., n,
and set 00 := S"(x0, zo) and 00 = (00, ..., 00). Then, for any 0 = (0', ..., on) with 10 - 001 << 1 and any z with Iz - zoo << 1, there is a uniquely determined solution x = (0, z) of (100) satisfying e C2 and (60i zo) = xo, by virtue of (99) and the implicit function theorem, and we can assume that there is an open neighbourhood r0 of (00, zo) in 1R" x 1RN such that Go = {(x, z): x = (0, z), (0, z) E r0}, if
we replace Go by an appropriate neighbourhood of (00, z0) which is again denoted by G0. If we do not insist in G0 being a ball we can even assume that TO = CO x I0 where 70 is an open neighbourhood of zo in lR", and CO c 1R" denotes an open cube in 1R" centered at 00 which is of the form
BobO.
(101)
We can also assume that the mapping r : (0, z) H (x, z), x = g(0, z), is a C2diffeomorphism of r0 onto G0; then (-, z) is a C2-diffeomorphism of Co onto a domain _q(z) in 1R", hence (102)
G0 = {(x, z): z e I, x e 2(z)}.
Therefore each of the surface (103)
,9 := {(x, z) a GO: S(x, z) = 6}, 0 e CO,
4.2. Caratheodory's Field Theory
121
is an n-dinemsional manifold representable by x = l;(O, z), z a lo, and the family {.Po}occ,, yields a foliation of Go.
From
z)) = B" we obtain by differentiating with respect to z` that SXB(x, z) '(8, z) + 5",(x, z) = 0,
x = 4(B, Z).
On account of (97) and (98) it follows that Z (0, z) = 2;'(x(0, z), z)
(104)
for all (0, z) e To.
This means, the surfaces 500 represented by x = (O, z), z a 10, fit the slope field 9 : G0 -+ G0 x 1R"N which is transversal to the geodesic slope field /i : G0
G0 in
the sense that /(x, z) and 9(x, z) are transversal for every (x, z) E G0, see Definition 1. We denote each manifold Yo = {(x, z) E G0: S(x, z) = 0} as a transversal surface with regard to the geodesic slope field A, and the family {9 }9Eco is said to be a transversal foliation with respect to fi.
Thus we have found quite a satisfactory geometric interpretation of the eikonal map S and of relation (97); this relation expresses the fact that the surfaces 500 = IS = 0} are transversal to the geodesic slope field associated with S.
Proposition 1. Any C2-mapping v : field /1 is an F-extremal.
R, * ' c IR", fitting a geodesic slope
Proof. Since M is a null Lagrangian, we have (105)
DaMpi(x, v(x), Dv(x)) - MZ,(x, v(x), Dv(x)) = 0.
Furthermore IS, 9) is a solution of the Caratheodory equations
M=F,
Mpi=Fp;,
where indicates the composition with /k. Differentiating F and M with respect to z` we obtain
MZr+Mpaa
'Ia
stF=FZ;+Fp.,,p
apM=
'q
and therefore also MZ, = FZ'
.
Since fi(x, v(x)) = (x, v(x), Dv(x)), equation (105) implies DDFF;(x, v(x), Dv(x)) - FZi(x, v(x), Dv(x)) = 0.
11
is a slope field, 9, = 9F o,4, and A(x, z) = Proposition 2. Suppose that / : (x, z, Y(x, z)), 9(x, z) = (x, z, 2(x, z)). Then fi is a geodesic slope field if and only if there is a mapping s e CZ(G0, IR"), s = (S1, ..., S"), such that the following holds true: (106)
S0.2f + Sz, = 0,
Chapter 7. Legendre Transformation, Hamiltonian Systems, Convexity, Field Theories
122
K det(S) + 1 = 0.
(107)
Here we have as usually set: (108)
K(x, z) := K(x, z, 2(x, z)) = W(x, z, 9(x, z)) =:
z).
Furthermore, if {S, 2} is a.solution of (106), (107), then S is an eikonal map associated with fE, and det(SSp) 94- 0.
Proof. (i) Suppose that A is a geodesic slope field with an eikonal map S. Then, by virtue of Lemma 1, IS, 9} satisfy (97) and (98). However, these equations are equivalent to (106) and (107), since we have
k = U = (-
F)"-'/A
on account of (30).
(ii) Conversely, suppose that {S, 2} are solutions of (106), (107). Then we infer from (81) and (106) that
EB = SS1 - SS121?4 = SS,[SJ - 2°gpl and by (28) and (32) we have (109)
gp = 8'OP - 2°g;' ,
= -FK.
Thus we obtain (110)
411 = SS,g;
and
M = det(EB) = det(SS,)-W = (-1/K)(-FK), i.e.
M=F. Furthermore (110) implies
ElT?h° = SX 9°h.TB whence
RE =Ss,'WSaTB=S"T19, and (109..) now leads to
Mi = -FKS "Tfi. it
Since we have already verified that M = F, it follows that
hQ = -S"T?. By (39) we have
Klla = 2 ,°ha,
4.2. Caratheodory's Field Theory
123
whence
17,° = - °SXTz and (106) now implies 17,a=SETA.
We have FPa = 17,a, and (84) yields
Sk T = MPa .
Thus we arrive at the second desired formula
a
FPi = M, An immediate consequence of Proposition 2 is the following result.
Proposition 3. Suppose that A : Go --> Go is a geodesic slope field with an eikonal map S G C2(Go, IR"). Then S satisfies the partial differential equation (111)
K(x,z,
S,,+ 1 = 0.
Conversely, let S e C2(Go, IR") be a solution of (111), and define 9 : Go --* Go by 9(x, z) = (x, z, .2(x, z)) and (112)
2 := -SZSX i, i.e.
2fSX, _ -SS,.
Then 4 := RFt o q : Go -+ Go is a geodesic field in the sense of Caratheodory, an S is an associated eikonal map of A.
We denote the first-order partial differential equations (111) for the eikonal map S as Vessiot-Caratheodory equation. For n = 1 it does not reduce to the Hamilton-Jacobi equation but to Vessiot's equation, which under appropriate assumptions on F is "equivalent" to Hamilton-Jacobi's equation (see 10,2.5 and 10,3).
Let us now summarize what we so far have achieved for the solution of our main problem. We try to find a Caratheodory calibrator M, given by M(x, z, p) = det[SX,(x, z) + S.,,(x, z)p'],
for {F, uo, W,(uo)} where uo = u1no, and £2o is a neighbourhood of xo a Q, f2o c S2, such that graph uo c Go. Since we want to carry out such a construction for each xo e S2, we have to assume that u is an F-extremal, according to Proposition 1. Let xo be an arbitrary point in Q. Then for sufficiently small neighbourhoods 00 of xo and Go of (xo, zo), zo = u(xo), with graph uo c Go, uo := ulna, we try to find a solution S e C2(Go, IR") of Vessiot-Caratheodory's equation (111) such that u fits the geodesic field k : Go -r Go generated by S as we have described in Proposition 3. Note that this fitting problem for uo is a highly underdetermined problem since (I11) is a simple scalar equation for n unknown function S', ..., S". The fitting problem can be interpreted in the
124
Chapter 7. Legendre Transformation, Hamiltonian Systems, Convexity, Field Theories
following way: Given a sufficiently small part uo of an F-extremal u, we have to find a foliation of Go by level surfaces .. = {(x, z) e Go: S(x, z) = 0} of a solution S of (111) such that each leaf 5° intersects the given extremal surface ,9o := graph uo transversally. Let us presently assume that the fitting problem for uo as described above is solved by means of a suitable solution S of Vessiot-Caratheodory's equation (111). We then want to show that the null Lagrangian M constructed in (1) in terms of S is a calibrator. This will be achieved by establishing (I) and (II) for the geodesic slope field /i(x, z) = (x, z, 9(x, z)) generated by S, see Proposition 3. Note that {S, 9} satisfies the Caratheodory equations (90) and (91); hence (I)
holds true as it means M = F. Thus we only have to make sure that (II) is satisfied, or equivalently that
F*(x, z, p) > 0 for all (x, z, p) E Go x IRnN
(113)
where F* denotes the modified Lagrangian F* = F - M which we had already introduced in (80). The function F* plays in Caratheodory's field theory the same role as Weierstrass's excess function IiF(x, z, po, p), po = .9(x, z), in De Donder-Weyl's theory. In general we have to add condition (113) to our basic assumptions (i) and (ii) on F (cf. (12) and (13)) to make certain that M is a calibrator for {F, uo, We(uo)}. This corresponds to the assumption OF >_ 0 in the
field theory for one-dimensional variational problems (cf. Chapters 6 and 8). Assumption (113) looks rather unpleasant because it not only involves F but also S which is still to be constructed; however, for the local fitting problem the
situation is not as bad as it may first appear. In addition we shall see that assumption (13) "almost" follows form (113); in fact we shall prove that the assumption F* >_ 0 implies 0. In other words, the two conditions R >- 0 and R > 0 play a similar role in Caratheodory's theory as the "necessary" Legendre condition Fpp > 0 and the "sufficient" Legendre condition Fpp > 0 for one-dimensional variational integrals fQ F(x, u(x), u'(x)) dx. In the sequel we shall always use our standard notation
F=Fo/, as =aa o 1z, Ili"=F,,=Fio j,etc. We begin by deriving a second expression for M = det(Ef) assuming that S is an eikonal map for a geodesic slope field it. Interestingly enough only terms in F and h enter in this expression while S has completely disappeared. Proposition 4. Let A : Go -+ Go be a geodesic slope field, /(x, z) = (x, z, 91(x, z)), and let M(x, z, p) be a null Lagrangian of form (1) where S is an eikonal map for S. Then M can be written as (114)
M(x, z, p) = F'-"(x, z) det {F(x, z)ba + [p! - 9a(x, z)] II;p(x, z)} .
Proof. We have (115)
jo-r = Tp/SX, + .Psz pi P
cc
4.2. Caratheodory's Field Theory
125
From (952) and (84) we obtain PI S,.
(116)
= II,I ,
and therefore
=tea TIS°,-Mba
Zip
Ir Ia PI S?i - EQ PI = PI (9aS=; - -Ta )
whence
aQ= - PISSa.
(117)
Combining (115)-(117) we find that
PIE. = -as + TI pa = -Mi Ass - F6.) + n" Pa, whence (118)
PIEQ = Fa + (pai:
- Ya)17,I
It follows that (119)
(det T)(det E) = det[FSQ + (pa -
Furthermore we have
M = det E, M = F,
M"-1
= det T,
and thus (119) yields (120)
MF"-1
= det[F81 + (pa - 9,',)17;I].
An immediate consequence of Proposition 4 is Proposition 5. Condition (113) is equivalent to (121)
F-F1-"det[FS,
on GoxR"N
if the assumptions of Proposition 4 are satisfied.
Lemma 2. We have
dTI = M-1(TITz - TIT")
(122)
Proof. From M = det E we infer that
dM=TdEf. Furthermore we have
&M=E"T"". Therefore
6.TdE; =6 dM=d(S. M)=d(L,;T) =(dE)T"+E,;dT".
0
126
Chapter 7. Legendre Transformation, Hamiltonian Systems, Convexity, Field Theories
Multiplying both sides by TO and summing over p we find
TA TPdEo = T!TPdE; +MdTO, and therefore M dT1 = T.1 T111 dEµ - Tx T" dE,,;,
whence we obtain (122).
Lemma 3. We have (123)
MPiPA = M (MPaMP8 - MPOMPa ).
Proof. By (84) we have (124)
MPa = SZ; T"'
whence (125)
mk'P' = SZ Uk TY . app
From (122) we infer that (126)
apeTy = M(Tr TA - TL
)ap",Eu
and E,, = S,, + S. -',p' implies that (127)
apk E,; = S." 616,.' = S.- bµ
.
Combining (125)-(127) we obtain (128)
1
MP:Pg = MSZ+S= (Tz,"TB - T"T°),
and in conjunction with (124) we arrive at (123).
11
From (11), (90), (91), and Lemma 3 we obtain for F* = F - M the following relations.
Proposition 6. If the assumptions of Proposition 4 hold true we have F* = 0,
Fp;=0,and (129)
Fv'P; = R k .
Forming the Taylor expansion of F(x, z, p) at p = 9(x, z) for fixed x and z we therefore obtain
4.2. Caratheodory's Field Theory
127
F*(x, z, p) = zFp*apa(x, z, 9p + (1 - 9)9)(9Q - pi)(°.p - p') for some 9 e (0, 1), Y = e(x, z). Hence (113) implies that (130)
Fnapg (x, z) >_ 0
for all (x, z) E Go,
which by (129) is equivalent to (131)
R k (x, z) >_ 0
for all (x, z) E Go
.
Thus we have proved: Proposition 7. The condition F* > 0 on Go implies that (R k) > 0 on Go.
Now we can formulate the following result summarizing the preceding propositions: Theorem 1. Suppose that F 0 0, A j4 0, and (R k) > 0. Moreover let u : 0 -- 1R" 92 c 1R", be an F-extremal, xo e Q, zo = u(xo), po = Du(xo). Then there exist open neighbourhoods Go = {(x, z): z e lo, x e.9 (z)} of (xo, zo) in lR" x RN and Go = Go x B,.(po) in 1R" x IRN x IR" such that Caratheodory's transformation "F yields a diffeomorphism of do onto some domain GO*. Choose a sufficiently small open neighborhood Qo of xo in Q such that graph uo c Go where uo = uIno, and suppose also that uo fits a geodesic slope field fe : Go --> Go with an associated eikonal map S : Go Finally assume that (132)
F - Ft-" dett
+ (pi - °Ja)Fpo] >_ 0 for (x, z, p) a Go x IRnN,
where ye(x, z) = (x, z, 3(x, z)), F = F o /i, Fpa = Fpa o fe. Then the null Lagrangian M(x, z, p) = det[S,xp(x, z) + p'S=;(x, z)] is a calibrator for {F, uo, W,(uo)}, 0 < a << 1, and therefore uo is a strong minimizer of f no F(x, v(x), Dv(x)) dx among all v e WE(u). In fact we can prove more. For convenience we assume that F > 0 (instead of F # 0). Consider the mapping x f-+ 0 = 9(x) where 9(x) := S(x, uo(x)), x e ffo. We have 9sx(x) = EB(x, uo(x), Duo(x)),
whence (133)
det D9(x) = M(x, uo(x), Duo(x)) = M(x, uo(x)) > 0.
Since we have chosen 0o as a sufficiently small neighborhood of zo we can assume that 9 is a diffeomorphism of Q0 onto do*- where 520 := 9(d2o). Consider the tube Z defined by
:r.= U Ys(x) = r(8920* x lo), xePao
cf. (101)-(103). Suppose also that 0520 is a smooth manifold, and let
f is a smooth manifold of dimension n + N - 1 containing the boundary ad'o of the extremal surface Bo = cp(Sfo) = graph uo.
Chapter 7. Legendre Transformation, Hamiltonian Systems, Convexity, Field Theories
128
We infer from (133) that
det D9(x) dx
.11(u0) = L. M(x, uo(x), Duo(x)) dx = n"
=measd2a=J rdB'nde2n ndB" ao
=f,200*d[01d62A...AdO1]=J zao
whence
J(uo) = meas Sdp = J
(134)
A d8".
dB' A dO2 A Bo
Furthermore consider an arbitrary map v e C2(U, RI), U c 92, Ii(x) := (x, v(x)), and suppose that
9':= graph v = O(U) c Go. From dS°(x, v(x)) = EB (x, v(x), Dv(x)) dx,,
we infer that
v, Dv) dx = rG*[dS' n dS2 n .
A dS"],
whence
..k(v) = J M(x, v(x), Dv(x)) dx = J ,*[dS' A u
A dS"]
u
=Lu)
J f d[S'dS2A...AdS"] 3.
a'
S1
Thus by introducing the (n - 1)-form a = S' dS2 A .,K(v) =
(135)
A dS" we find
a.
or.
Suppose now that the boundary 59- of 9' lies on the tube .1 and that the mapping t : a.°T -+ 3Qo defined by (x, v(x)) i-. 0 = S(x, v(x)) is one-to-one. Then we infer from (135) that ../1(v) = J
B' dO2 n
n dO".
ano
Then it follows in conjunction with (134) that (136)
.11(uo) = ..1((v) if a9' c T and t : a9"
3Q,* is 1-1.
More generally we have .1(uo) _ .11(v) if a9 c 9 and if 0-T and ado are homologous in 26y, since in this case there is an n-chain aB in . such that &'I = a9' - ado, and thus we obtain by Stokes's theorem .11(v)-.11(uo)=La
a
-
Jad"a=J
a =Jyda,
and W c 9' implies that
da= J dS' n f,
AdS"=0.
4.2 Caratheodory's Field Theory
129
Thus, we have
.41(uo) = W(v) if 8°l c T and
(136)
&o in f,
which leads to the following Supplement of Theorem 1. Let v be a comparison map of class C' (U, 1R"), U c 12, whose graph, J 'T, satisfies PT c Go, 89- c Y, and 037' - 88o in ! where 61o = graph uo. Then we have
.F(uo) :=
F(x, uo(x), Duo(x)) dx
Ja"
(137)
<
r
F(x, v(x), Dv(x)) A := 97(v).
u
Furthermore, if v fits the geodesic field /L, then JF(uo) = ! (v) We can view this result as a generalization of A. Kneser's transversality theorem (see Chapter 6).
There is no comparable result in De Donder-Weyl's field theory which is taylored to variational problems with fixed boundaries, and H. Boerner [3] has proved that Caratheodory's theory plays a distinguished role among all possible field theories (cf. 4.3) introduced by Lepage as it is the only one
allowing a treatment of free boundary problems analogously to the case n = 1. Let us finally sketch how the local fitting problem can be solved for Carathdodory's theory. The first solution of this problem was given by H. Boerner [5]; his approach is similar to the one we have presented in 4.1 for solving the fitting problem in the framework of De Donder-Weyl's theory, only
that the underlying formalism is now much more involved. Here we want to indicate another method based on ideas of E. Holder [2] which lead to a considerable formal simplification and a better geometric understanding of the problem. We begin by looking at a special situation. For solving the fitting problem we have to find a solution S(x, z) _ (S' (x, z), .. , S"(x, z)) of the Vessiot-Caratheodory equation (111) in Go such that uo = ula" fits = 31F' g, where g(x, z) = (x, z, 2(x, z)), 2 = -Se' S,-', i.e. uo has to satisfy uo.. _ 9(x, uo) where jl(x, z) _ (x, z, 9(x, z)), or equivalently uo must fulfil the equations (138)
2°(x, uo(x))Sxa(x, uo(x)) = -SS,(x, uo(x)),
cf (112). We try the Ansatz
S''(x,z)=x', 2