This action might not be possible to undo. Are you sure you want to continue?
00
TENSOR ANAL YSIS
BY EDWARD NELSON
Preliminary Informal Notes
of University Coarses and Seminars in Mathematics
MATHEMATICAL NOTES PRINCE TON UNIVERSITY PRESS
ANNALS OF MATHEMATICS STUDIES
Edited by Robert C. Gllnnlng, 101m C. btoore. and Marston Morse
1. Algelmuc Theory of Numbers By HER~L.J\NN WEYL
3. Consistency of the Continuum Hypothesis By KURT GbDEL
11. Introduction to Nonlinear Mechanics
By N. KRYLOFF and N. BOGOLIUBOFF
20. Contributions to the Theory of Nonlinear Oscillations, Vol. I Edited by S. LEFSCHETZ
21. Functional Operators, Vol. I By JOHN VON NEUMANN
24. Contrilmticns to the Theory of Games, Vol. I Edited by H. W. KU.HN and A. W. TUCKER
25. Contribntions ttl Fourier Analysis
Edited by A. ZYCMUND, \Y. TRANSUE, M. !.,[ORSE, A. P. CALDERON, and S. BOCHNER
27. Isopprimetric Inequalities in Mathematical Physics By G. P6LYA and G. SZEGo
28. Contriblitions to the Theory of Games, Vol. II Edited by H. W. KUHN and A. W. TUCKER
30. Contrtbutions to the Theory of Riemann Surfaces Edited by L. AHLFORS et aI.
33. Contributions to the Theory of Partial Differential Equations Edited by L. J3EaS, S. BOCHNER, and F. JOHN
34. Automata Studies
Edited by c. E. SIIANNON and 1. McCARTHY
38. Linear Inequalities and Related Systems
Edited by H. W. KUHN and A. W. TUCKER
39. Contributions to the Theory of Games, Vol. III
Edited hy xr. DBESHER, A. W. TUCKER, and P. WOLFE
40. Contributlons to the Theory of Games, Vol. IV Edited by R. DUNCAN LUCE and A. W. TUCKER
41. Contributions to the Theory of Nonlinear Oscillations, Vol. IV Edited bv S. LEFSCHETIZ
42.. Leotnres 'on Fourier Integrals
By S. BOCHNER
43. Ramification Theoretic Methods In. Algebraic Geometry By S. ABHYANKAR
44. Stationary Processes and Prediction Theory By H. FURSTENBE1~G
45. Contributions to the Theory of Nonlinear Oscillations, Vol. V Edited by L. CESARI, 1 LASALLE, and S. .LEFSCHETl
46. Seminar on Transformation Groups By A. BOREL et al.
47. Theory of Formal Systems Bv R. ShfULLYAN
Continued on iT!side back cover
PRINCETON UNIVERSITI PRESS Princeton, New Jersey
ANNALS OF MATHEMATICS STUDIES
Continued [rom back cover
48. Lectures on Modular Forms By R. C. GUNNING
49. Composition Methods in Homotopy Groups of Spheres By H. TODA
50. Cohomology Operations
Lectures by J\'. E. STEENROD, written and revised by D. B. A. EPSTEIN
51. Morse Theorv
By J. W. MILNOR
52. Advances in Game Theory
Edited by .M. DRESHER, L. SHAPLEY, and A, W. TUCKER
53. Flows on Homogeneous Spaces
By L. AUSLANDER, L. GREEN, F. HAHN, et nl.
54. Elementary Differential Topology By J. R. MUNKRES
55. Degrees of Unsolvabihty By G. E. SACKS
56. Knot Groups
By L. P. NEUWIRTH
57. Seminar on the AtiyahSinger Index Theorem By R. S. PALAIS
58. Continuous Model Theory
By C. C. CHANG and H. J. KEISLER
59. Lectures on Curves on an Algebraic Surface By DAVID MUMFORD
60, Topology Seminar, Wisconsin, 1965
Edited by R. H. BING and R. J. BEAN
TENSOR ANALYSIS BY EDWARD NELSON
PRINCETON UNIVERSITY PRESS AND THE UNIVERSITY OF TOKYO PRESS
PRINCETON, NEW JERSEY 1967
Copyright (§) 1967, by Princeton University Press All Rights Reserved
Published in Japan exclusively by the University or Tokyo Press;
in other parts or the world by Princeton University Press
Printed in the United States or America
Preface
These are the lecture notes for the first part of a oneterm course on differential geometry given at Princeton in the spring of 1967. They are an expository account of the formal algebraic aspects of tensor analysis using both modern and classical notations.
I gave the course primarily to teach myself. One difficulty in learning differential geometry (as well as the source of its great beauty) is the interplay of algebra, geometry, and analysis. In the first part of the course I presented the algebraic aspects of the study of the most familiar kinds of structure on a differentiable manifold and in the second part of the course (not covered by these notes) discussed some of the geometric and analytic techniques.
These notes m2lf be useful to other beginners in conjunction with a book on differential geometry, such as that of Helgason [2,§1], Nomizu [5,§5], De Rham [7,§7J, Sternberg [9,§8], or Lichnerowicz [ll,§9]. These books, together with the beautiful survey by S. S. Chern of the topics of current interest in differential geometry (Bull. Am. Math. Soc., vol. 72, pp. 167219, 1966) were the main sources for the course.
The principal object of interest in tensor analysis is the module of C~ contravariant vector fields on a C~ manifold over the algebra of C~ real functions on the manifold, the module being equipped with the additional structure of the Lie product. The fact that this module is "totally reflexive" (i.e. that multilinear functionals on it and its dual can be identified with elements of tensor product modules) followsfor a finitedimensional secondcountable
11.
DO
C Hausdorff manifold  by the theorem that such a. manifold has a.
covering by finitely many. coordinate neighborhoods. See J. R. Mlmkres, Elementary Differential Topology, p.18, Annals of Mathematics Studies No. 54, Princeton UniV1:l1'sity Press, 1963.
I wish to tha.nk the members of the class, pa.rticularly Barry
Simon, for many ~roV1:lments, and Elizabeth Epstein for typing the
manuscript so be~tifully.
CONTENTS
§l.
Multilinear algebra • • • • • • . • . . • • • . . 1. The algebra of scalars 2. Modules
3. Tensor products 4. Multilinear functionals 5. Two notions of tensor field 6. Flinear mappings of tensors 7. Contractions 8. The symmetric tensor algebra 9. The Grassmann algebra. 10. Interior multiplication
11. Free modules of finite type 12. Classical tensor notation 13. Tensor fields on manifolds 14. Tensors and mappings
§2.
Derivations on scalars
1. Lie products 2. Lie modules 3. Coordinate Lie modules 4. Vector fields and flows
Derivations on tensors • • • • • • • • • • • • • 1. Algebra derivations 2. Module derivations 3. Lie derivatives 4. Flinear derivations 5. Derivations on modules which are free of finite type
§4.
The exterior derivative . • • • • • • • • • • . .
1. The exterior derivative in local coordinates
2. The exterior derivative considered globally
3. The exterior derivative and interior multiplication 4. The cohomology ring
§5.
Covariant differentiation • • • • • • • • • • •
1. Affine connections in the sense of Koszul
2. The covariant derivative 3. Components of affine connections 4. Classical tensor notation for the covariant derivative
5. Affine connections and tensors 6. Torsion 7. Torsionfree affine connections and the exterior derivative 8. Curvature
9. Affine connections on Lie algebras 10. The Bianchi identities 11. Ricci's identity 12. Twisting and turning
§6.
Holonomy • • • • • • • • . • • • . • • • • • •
1. Principal fiber bundles 2. Lie bundles 3. The relation between the two notions of connection
iii.
Page 1
25
37
47
57
79
iv.
Page
Riemannian metrics . • • • • • • • . • • . • . • •
1. PseudoRiemannian metrics 2. The Riemannian connect ten 3. Raising and lowering indices 4. The RiemannChristoffel tensor
5. The codifferential 6. Divergences
7. The Laplace operator 8. The Weitzenb~ck formula 9. Operators commuting with the Lap1acean 10. Hodge theory
89
§8. Symplectic structures . • . . . . • • • • • • • . . . 111
1. Almost symplectic structures
2. Hamiltonian vector fields and Poisson brackets 3. Symplectic structures in local coordinates 4. Hamiltonian dynamics
§9. Complex structures • • • • • • • • • • • • • • . . . • 117 1. Comp1exification 2. Almost complex
structures 3. Torsion of an almost complex
structure 4. Complex structures in local
coordinates 5. AJJnost complex cormections
6. K'ahler structures
1.
§l. Multilinear algebra
1. The algebra of sca.lBrs
We make the permanent conventions that FO is a field of characteristic 0 and that F is a commutative algebra with identity over FO Elements of F will be called scalars and elements of FO will be called constants.
The madn example we have in mind is
the field
m of real COO manifold
numbers and F the algebra of all real M. In this example the set of all COO
COO functions on a
contravariant vector fields is a
module over F, with the additional structure that the contravariant vector fields act on the scalars via differentiation and on each other via the Lie product. Tensor analysis is the study of this structure. In this section we will consider only the module structure.
2. Modules
The term "module" will always mean a unitary module (ix = X) • Thus an F module E is an Abelian group (written additively) with a ma.pping of FXE into E (indicated by juxtaposition) such that
f(X+Y) fX+fY (f+g)X fX+gX (fg)X = f(gX) ,
lX = X , for all X,Y in E and f,g in F
If E is an F module, the dual module E' is the module 'of all
2.
§l. MULTILINEAR ALGEBRA
Flinear mappings of E into F. If WEE' we denote the value of W on X in E by any of the symbols
W(X) ,
X(W) ,
< w,X > ,
.< x,w >
If A is an Flinear mapping of E into E its dual A' defined by
< A'w,X > = < w,AX >
is an Flinear mapping of E' into E' . There is a natural mapping K: E ~ En defined by
(KX)(w) = < w,X > ,
W € E' ,
and E is called reflexive in case K is bijective. (The mapping K is not in general injective. For example, if F is the algebra of all COO fUnctions on a manifold M and E is the F module of all continuous contravariant vector fields then E' = 0 .)
The notions of submodule, F module homomorphism, and quotient module are defined in the obvious way. If H and K are F modules and ~: H ~ K is an F module hbmomorphism then the quotient module Hiker ~ is canonically isomorphic to the image of ir : See Bourbaki [1].
We will frequently refer to the elements X of an F module E as contravariant vector fields or vector fields and to elements w of the dual module E' as covariant vector fields or lforms.
3. Tensor products
If H and K are two F modules, their tensor product H8K
(over F) is the F module whose Abelian group is the free Abelian group generated by all pairs X®Y with X in H and Y in K modulo the sub" group generated by all elements of the form
§l. MULTILINEAR ALGEBRA
3·
(Xl+~)®y  X1®Y"  X2®Y" , X®(Yl+Y2)  X®Yl  X®Y2 ' (fi)®y  X®(fY)
where f is in F, and the action of F on H8K is given by
f(X®Y) = (fi)®y = X®(fY)
let E be an F module. We define
o
ES = ~ •.• ~'® .•. @!' r
(E r times, E' s times).
If r or s
is
00 00 0
o we sometimes omit it, and we set EO = E = EO = F
Notice that ~l = E, ~l = E'
We also define
0000 E* = z E r=O r
~* = 'i ~ ,
r=O
where the sums are weak direct sums (only finitely many components of any
element axe nonzero).
o 0*
Notice that E* and E are associative graded F algebras with
the tensor product ® as multiplication. We make the identification
~ ~ = ~s~
r r
0*
With this identification, E* is an associative bigraded Falgebra.
4 . Multilinear functionals
Let E be an Fmodule.
We define ES to be the set of all r
Fmultilinear mappings of
E'X ••• XE'XEX ... XE
(E' r times, E s times)
6.
§l. MULTILINEAR ALGEBRA
and extending to all of ~; by Flinearity. In particular, the dual module
Os r
of Er is canonically isomorphic to Es ' so that if E is total!l reflex
ive then each E6 is reflexive.
r
Proof.
The mapping L is welldefined by the definition of tensor
product, and is an F module homomorphism. It is obviously injective and
surjective. QED.
Suppose that E is totally reflexive. A number of special cases
of Theorem I come up sufficiently often to warrant discussion. We identify (Es), and Er , and denote the pairing by any of the expressions < u,v > ,
r s
< v,u >, u(v) , v(u) , as convenient. If A is in Ei we use the same
symbol A for the Flinear transformation L(l,O,I,O)A of E = El into itself, so that
< w,AX > = A(W,X) ,
Notice that the Flinear transformation L(O,l,O,l)A of El into El is the dual A' of A. If A and B are in Ei we write AB for their product as Flinear mappings of E into itself and similarly for An.
The identity mapping of E into itself is denoted I.
If E is totally reflexive then L(2,O,I,O) identifies Ei with the set of all structures of F algebra (not necessarily associative) on E.
If B
. . E2 ·t
1S 1n 1 we wr1 e
B(X,Y)
or BxY for the product in this sense
of the two vector fields X and Y, so that
< w,B(X,Y» = < w,BxY > = B(W,X,y) ,
Also, L(I,O,O,I) identifies E2 with the set of all Flinear mappings of EI into EI, so that
< u(x), Y > = u(x, Y) ,
§l. MULTILINEAR ALGEBRA
7·
Similarly, L(O,l,l,O) identifies E2 with the set of all Flinear mappings of El into El, so that
< ~,v(w» = v(w,~) ,
7. Contractions
let E be an F module, and let J2~ IJ..::?', l<V<s
We define the
contraction
by
where the circumflex denotes omission, and by extending C~
to all of ~s r
by Flinearity. By the definition of tensor product, this is welldefined,
and it is a module homomorphism. The Encyclopaedia Brita=ica calls it an
operation of almost magical efficiency. (See the interesting article on
tensor analysis in the l4th edition.)
If A E Ei then ciA is denoted tr A , and called the trace of A.
8. The symmetric tensor algebra
let E be an F module and let G (r ) be the symmetric group on r letters. For u in E and 0 in G (r ) define up (0) by
r
Then p
is a right representation of G (r)
on E r
that is,
Define Sym on Er by
8.
§l. MULTILINEA,_"\ ALGEBRA
1
Sym u = .. r: up(a) . r. C5
(Since FO is a field of characteristic zero, l/r! makes sense.) Extend
Sym to the contravariant tensor algebra E* by additivity. A contravariant
tensor u is called symmetric in case Sym u = u .
Thus u in E is r
symmetric if and only if u(wl, ... ,wr) is invariant under the transposition
of any pair of wls .
The set of all symmetric teL"lOrS in E is denoted r
Sr and the set of all symmetric tensors in E* is deno+ed S*, so that
where of course So = F .
Theorem 2.
Sym is Flinear and is a projection
2
(Sym = Sym) with
range S*. Consequently S* may be identified with the quotient of E*
by the kernel of Sym. The kernel of Sym is a twosided ideal in E*.
Consequently the multiplication
(4)
uv = Sym 1.fi9v
makes S* into an associative commutative graded algebra over F.
Proof.
Sym is clearly Flinear. That it is a projection follows
from (3)  it is easily checked that the average over a group representation
is a projection. The range of Sym is S* by definition, so that we may
identify S* with the quotient of E* by the kernel of Sym.
By the definitions of Q9 and Sym, if u € Er and v € Es then
_ 1 e (0(1) wo(r)) ( (r+L) wo(r+s))
r.::::::\T .... uw , ... , va , ... ,
\:r+Sj. a
where 0 ranges over :G; (r+s). If Sym u or Sym v is 0 this is clearly
§l. MULTILINEAR ALGEBRA
9·
o , so that the kernel of Sym is a twosided ideal, and the quotient alge
bra is an associative commutative graded F algebra. QED.
The algebra S* is called the (contravariant) symmetric tensor
algebra.
One ~ also construct the covariant symmetric tensor algebra
* S .
9. The Grassmann algebra
The discussion of the (covariant) Gr~ssmann algebra, given an F module E, proceeds along similar lines. For 0: in Er and a in e (r ), define ap(a) by
where sgn a is 1 for a an even permutation and 1 for a an odd per
mutation. Then P is a right representation of G (r) on Er Alt on Er by
Define
Alt 0: = ~ ~ o:p(a)
r. a
* and extend Alt by additivity to E
*
An element 0: of E such that
Alt 0: = 0: is called alternate or antisymmetric and is also called an exterior form. The set of alternate tensors in Er is denoted Ar , and
elements of Ar are called rforms.
* The set of all alternate tensors in E
*
is denoted A ,so that
A* ~ Ar
r=O
Notice that A 0 = F and A 1 = El. A covariant tensor 0: of rank r is
alternate if and only if O:(Xl, ... ,Xr) changes sign under the transposition of ar~ two XIS.
Theorem 3.
* Alt is Flinear and is a projection with range A
* *
Consecl.'J.ently A ~ be identified with the quotient of E by the kernel
10.
§l. MULTILINEAR A:r.GEBRA
of Alt.
*
The kernel of Alt is a twosided ideal in E and the multi
plication
a....t3 = Alt CX@t3
makes A* into an associative graded algebra over F satisfying
(6)
Proof.
The proof is quite analogous to the proof of Theorem 2.
Instead of (5) we have, for 0: in Er and t3 in ES ,
(Al t o:@t3)(Xl, .•. ,X )
r+s
= ~ 1: (Sgll cr)o:(X (1)'" .,X ( ))t3(x ( 1)'" .,X ( ))
,r+sJ' cr cr cr r cr r+ cr r+s
'lED.
*
The algebra A is called the (covariant) Grassmann algebra. One
can also construct the contravariant Grassmann algebra A*
Warning: As we have defined the notion, an rform is simply a co
variant tensor of rank r which is alternate. However it is customary in
the literature, and we will follow the custom because it is convenient, to
make from time to time conventions about rforms which differ from conven
tions already made about tensors. These special conventions have the pur
pose of ridding the notation of factors r! , etc.
If ex
k
is an exterior form we denote by ex the exterior product of
ex with itself k times, ~ = exA ••• Aex. If k> 1 this is 0 for ex in Al or for ex an exterior product of Iforms, but not for general elements
* in A
A graded algebra whose multiplication satisfies (6) is sometimes
called "commutative," but this miserable terminology will not be used here.
§L MULTILINEAR ALCEBRA
11.
10. Interior multiplication
Let X be in the F module E and let 0: be an rform.· We define
X J 0: by
(8)
XJo:=O
o
if 0: € A ,and we define X J 0: by additivity if 0: is a
* general element of A
The mapping 0: ~ X J 0: is F1inear from Ar to
Ar1 , and it follows from (7) that it is an antiderivation of A*; that is,
X J (0:,,13)
* 13 € A
11. Free modules of finite type
An F module E is free of finite type if there exist XF· .. ,Xn
in E , called a basis, such that every element Y in E has a unique
expression of the form
y = 1: Y~i ' yi € F • (Unless indicated otherwise, 1: always denotes summation over all repeated
indices. )
Theorem 4.
Let E be free of finite type, with a basis Xl'·· .,Xn·
Then E is totally reflexive. The dual module has a unique basis 1 n
w, ... ,w (called the dual basis) such that
i i
< w ,X. > = 5. ,
J J
where 5~ is 1 if i j and 0 otherwise. The
J
j~ js
X. ® •.. ®lC. ®.&l ••• ®.&l
~l ~
r
are a basis of ES , so that every u in ES has a unique expression of
r r 12.
§l. MULTILINEAR ALGE:BRA
the form
The coefficients in this expression (called the components of u with res
pect to the given basis of E
are given by
il i
u(w , ... ,w r,Xj , ... ,X. ) 1 Js
The
(10)
are a basis of Ar , so that every r~ a has a unique expression of the
form
a=
The coefficients in this expression (called the components of a as an r
form or simply the components of a) are given by
r~a(x. , ... ,X. )
11 1r
so that the components of 0: as an rform· are r! times the "components of
a regarded as an element of Er . If r>n then Ar = o .
If u € ES, s' then Ul9v has components
r v € Er' , If l~~r, l<V<s and u € E: then c~u has components
§1. MULTILnlEAR ALGEBRA
13·
r+s ' If' u e: Es+r'
and
then
(tu)v
s' in Er'
has components
il, .. i kl· .. k , j1'" j
L s r r
ujl' . ,j £1"" .e ,vi .
r s l"'~s
If u € Er then Sym u has components
where the e is 1 if the j' s are a permutation of the i' s and is 0 otherwise. If' a € Er then A1t Q has components as an element of' Er
given by
where the 0 is 1 if' the j , s are an even permutation of the i' s , is 1
if the j's are an odd permutation of' the i's, and is 0 otherwise, If'
a is an rform and t3 is an sform then the (r+s) form a....f) has components
If X € E1 and a ~ rform, the (r1)~ X J a has components
Let
X~"., "Xn'
be another basis of E, and define
J~'
~
and ~,
Then
14.
§l. MULTILINEAR ALGEBRA
If u E ES the components of n with respect to the new basis are r
Proof. The proof is trivial. QED.
Notice that the primed indices do not take values in the set
(l, ... ,n} but in a disjoint set (l', ... ,n'} of the same cardinality.
This notation is very convenient, as it makes it impossible to make a mis
take in writing the transformation laws.
12. Classical tensor notation
Despite the profusion of indices, the classical tensor notation is
frequently quite useful, especially in computations involving contractions.
The vector fields over a coordinate neighborhood in a finite dimensional
manifold are a free module of finite type, but the module of all vector
fields does not in general have a basis. (If it does, the manifold is called
parallelizable.) However, it is possible to use the classical tensor nota
tion globally, without any choice of local coordinates, if we make the fol
lowing conventions.
Let E be an F module, and let
Consider an expression
of the form
(11)
Instead of "il", •.. ,"js" we may use any other r+s indices, provided they are distinct indices. The upper indices are called contravariant indices,
the lower indices are covariant indices. Next we suppose that the contra
variant indices are covariant vector fields and the covariant indices are
contravariant vector fields. Then we define (11) to be the scalar
§l. MULTILINEAR ALGEBRA
15·
il·· .ir
u. . = U(il, .•. ,ir,jl, •.• ,js) .
Jl·· ·Js
(It would perhaps be better to write jl ... jS, but we don't.) Notice that
although the indices are required to be distinct indices, the mathematical
objects they denote need not be distinct. (Thus we may have il = i2 as covariant vector fields although obviously "il" f. "i2" .) However, for an rform a we make the special convention that
a. . = r!a(il,·· .,ir) ,
~l" '~r
Now suppose that E is totally reflexive, so that contractions of
tensor fields are meaningful.
If u E ES we define r
Instead of "a" we may use any other index, provided it is distinct from
the other indices occurring. An index which occurs precisely twice, once
as an upper index and once as a lower index, is called a dummy index.
Notice that there is no summation sign in (13). This is because nothing is
being summed. (When dealing with components with respect to a basis of a
free module of finite type, we will continue to write summation signs when
summations occur , ) We may have more than one dummy index, provided they are
all distinct from each other and the remaining indices, to indicate repeated
contractions. The notation is unambiguous because, from the definition of
contraction, the order in which the contractions are performed is immaterial.
Here are some examples of the use of this notation. In all but the
first example we assume that E is totally reflexive.
If u E ES and r
s' vEEr' then
16.
§l. MULTILINEAR ALGE:BRA
l<V<s and u € ES then r
(C~u)il···~~···ir = uil···i~_lai~+l···ir
V jl···jV···js jl •.. jV_lajV+l .•. js
I'+s' s
If u € ~s+r" v € Er and L = L(r,s,r',s') then
(14)
The notation here is abusive. The right hand side of (14) is not the product
of two scalars but is written instead of
il···i kl···k ,jl···j
('L'®v) s r r
jl,,·j tl,,·t ,il• "i
r s s
We will indulge :freely in this abuse of notation. Now let Symr be the
restriction of Sym to Er .
Since Sym is Flinear,
r! Sym = L(r,O,r,O)E r
for a unique tensor E in E~, and if u € Er then il" .ir 1 il· .. ir jl" ·jr
(Sym u) = ~ Ej j U
r. 1'" r
The tensor E me;y be computed explicitly, and one finds
where perm denotes the permanent. (The permanent of a square arre;y of
scalars is defined in the same we;y as the determinant except that there are no minus signs.) Similarly, if 0: € ~ then
for a unique tensor 5
in Er , and r
§l. MULTILINEAR ALGE:BRA
17·
where det denotes the determinant. If a € Ar and ~ € AS then (recall
(J2) )
and if X € E then
13. Tensor fields on manifolds
Let p be a point in the COO manifold M
A tangent vector at p
is an equivalence class of differentiable mappings x: lR ~ M with
x(o) = p ,where x and y are equivalent in case the coordinates of x( t)
and yet) differ by oCt)
One verifies that this condition is independ
ent of the choice of local coordinates, and that addition and multiplication
by constants are welldefined on tangent vectors. Thus the set of all tan
gent vectors at p
forms a real vector space M , called the tangent space p
at p. A cotangent vector at p is the dual notion: an equivalence class
of differentiable mappings f: M ~ JR with f(p) = 0 ,where f and g are equivalent if g(q) and g(q) differ by little 0 of the difference
in coordinates of q and p
Again, the condition is independent of the
choice of local coordinates, and the cotangent vectors form a vector space
M' which is in a natural way the dual vector space to M
p P
The set T(M) of all tangent vectors at all points of M has a
natural structure of COO manifold as qoes the set
*
T (M)
of all cotangent
Vectors. They are called the tangent bundle and cotangent bundle. They are
equipped with natural projections onto M, the projections which assign to each vector the point p at which it lives. A COO section of the tangent bundle is called a contravariant vector field or vector field and a COO
18.
§l. MULTILINEAR ALGEBRA
,/
I
Eigure 1.
Pictures of a tangent vector X and a cotangent p
A tangent vector gives a direction and speed of
vector w p
motion, a cotangent vector is a linear approximation to a
scalar. The tangent vector 2X would be indicated by an p
arrow twice as long,
2w would be indicated a relabeling p
of the hyperplanes (twice as dense).
In the figure X p
and w look as if they are in some sense the same, but p
this has no meaning unless the tangent space is equipped
with additional structure, such as a pseudoRiemannian
metric or symplectic structure.
§l. MULTILINEAR AI.GE1!RA
19·
section of the cotangent bundle is called a covariant vector field or Iform. They form modules E and E' over the algebra F of all scalar!;! (C" real i'unctinns on M). Therefore we have the notions of tensor fields on M and
tensors at a point p.
Tensors are of great importance in differential geometry because they are invariantly defined geometrical objects (independent of any coordinate system) which live at points. Both characteristic's are necessary in
order for an object to be a tensor. Suppose for example we attempt to de
fine a tensor u, contravariant of rank 2, by requiring, in local coordi
nates,
u(w,~) = E 5ijwi~j ,
where 5 ij is 1 if i = j and 0 otherwise. This lives at points but
is not inva.:riantly defined, since in new coordinates
l' n'
x , ... ,x
it would
have components
kl axi' o.x:j' E 5 r
<:Jxk o.x:.fi
(On the other hand, 5 ~ are in each coordinate system, the components of a certain tensor.) As another example, let X be a fixed contraVariant
vector field other than 0 and define e on E by e(y) = [X,Y] , where [X, y] is the Lie product of X and Y (§2) • This is invariantly defined but it does not live at points, because in order'to know e(y) at a point
p we need to know something about Y in a neighborhood of p in order to
differentiate it. In fact, e is lRlinear but not Flinear, since
e(fY) = fe(Y)+(X·f)Y , so that e is not a tensor field. The condition of
Flinearity is in fact the condition that an lRmultilinear object live at
points.
If for example w is a Iform and X = Y then (w(X))(p) p p
(w(Y))(p) , since we may write XY = fZ with f(p) = 0, and so
20.
§1. MULTILINEAR ALGEl3RA
(W(X»(p)(W(y»(p) = (w(£Z»(p) = f(p)(w(Z»(p) = 0 •
The example of the Lie product shows that not all ~nteresting geometrical objects are tensors. Affine connections are another example of secondorder geometrical objects. Tensor fields are firstorder geometrical objects since the notion of tangent vector involves one derivative.
14. Tensors and mappings
Suppose we have two FO algebras F and 1', an F module E wtth dual E' , and an l' module Ef with dual 'M' • We shall use the word homomorphism for any ot the following:
an FO algebra homomorphism;
(16) p: E ..;:. Ef ,
a group homomorphism (and similarly for p: E' ~ l')
(17) p: (F,E) ~ (1',l)
where p: F ~ l' and p: E ~ Ef are homomorphisms satisfying the compatability condition
(18)
pCf'}() = p(f)p(X) ,
f € F, X € E ,
(and similarly for p: (F,E') ~ (1',Ef'» ; and finally for
(19)
where p: F ~~, p: E ~ If, p: E' ~ Ef' are homomorphislliS satisfying the compatabi1ity conditions (18) and
(20)
p(fW) = p(f)p(w) ,
f € F, W € E' ,
§l. MULTILINEAR· ALGEBRA
21.
(21)
III E E', X E E •
NoV let
be a C<Xl mapping of the manifold M into the manifold l:t. Then
defined by (~~)(p) = f(~(p» is a homomorphism. If we recall what a
tangent vector at a point p in M iSI we see that ~ induces a vector space homomorphism (linear transformation)
It is called the differential of ~ at p
By duality,
If we define
«~~) ,X> = <~~( ),d~ (X »
p p ... p p p
then
is a homomorphism. In the same way we obtain a homomorphism
of the covariant tensor algebras, which preserves the grading add products ® , and sends the Grassmann algebra r into A * However, it is 1mme
diately clear to almost anyone that we do not in general obtain a homomorphism (F,E) _.,. (]I,Ef) since ~ is not necessarily onto, we may not have d~ (X ) = d~ (X) whenever ~(p) = ~(q) ) and even if these difficulties
p p q q
do not arise we may not get a C<Xl section of T(l:t) (see Exercise A.4 on
p.83 of Helgason [2]). The mapping ~ induces C<Xl maps
22.
§l. MULTILINEAR ALGEBRA
~*: T(M) ~ T~) } q;*: T*(1.t) ~ T*(M)
but ~* does not in general induce a mapping on C~ sections of T(M) • Suppose now that ~ is a diffeomorphism of M onto ~. Then we obtain a homomorphism (in fact, an isomorphism)
(22)
~ * (F,E,E') ~ (F,E,E')
as follows. On ~ and E",
*
~ is as def'Lned above.
For ~ in E we
define
This homomorphism extends in a natural way to the mixed tensor algebras.
In the same way we obtain a homomorphism (22) if ~ is an imbedding of M
in ~ .
It is unfortunate that covariant tensor fields transform contra
variantly under point mappings of manifolds, but it is too late to change
the terminology. Early geometers were more concerned with coordinate
changes than point mappings, and coordinates are scalars, which transform
the same way as covariant tensor fields.
*
Notice that we have used the notation E for covariant tensor
fields in keeping with the fact that they transform the opposite way to
point mappings. For example, the cohomology ring is formed from the Grass
* *
mann algebra A and it is universally denoted H
In our study of tensor analysis we shall make no use of points except at one point in the discussion of harmonic forms (§7), where we will4
need the following notion.
§l. MULTILINEAR ALGEBRA
23·
Definition. The F module E is punctual if there exists a sepa
rating family 01' homomorphisms of the 1'orm
p : (F,E,E') ~ (F ,E ,E')
p p p p
where Fp = FO and Ep is a finite dimensional FO vector space.
The module of contravariant vector fields on a ~ifold is punctual:
take Pp to be evaluation at the point p space at p.
and E to be the tangent p
References
[1] N. Bourbaki, El~ments de math~ma.tique, Hermann, Paris. See especially
Book 2, Alge bre, Chaps. 2 and 3.
[2] Sigurc!ur Helgason, Differential Geometry and Symmetric S:paces,
Academic Press, New York, 1962.
§2. Derivations on sca.lars
25·
1. Lie products
A derivation of F is an FO_linear mapping X: F ~ F such that
X(fg) = (Xf)g + f(Xg) ,
f,g e F .
(See §l.l for the assumptions on FO and F.)
If X and Y are derivations then so is X+Y defined by
(X+Y)f = Xf + Yf ,
and if h is in F then hX defined by
(hX)f = h(Xf)
is also a derivation. Thus the set of all derivations of F is an F module. It is denoted D .
If XeD then Xl = Xl +Xl , so that Xl = O. By FOlinearity, Xa = 0 for all a in FO
If X and Y are in D, we define their Lie product [X,Y] by
[X, Y]f = XYf  YXf
This is again a derivation:
[X, Y](fg) = XY(fg)  YX~fg)
= ~{(Yf)g+f(Yg)}Y{(Xf)g+ f(Xg)} (XYf)g+ (Yf)(Xg) + (xr) (Yg) + f(XYg)
 (YXf)g  (Xf)(Yg) .. (Yf)(Xg)  f(YXg) ([X,Y]f)g + f([X,Y]g) .
The set of all FOlinear mappings of F into itself forms an associative ring, and D is a subset of it. In any asso~iative ring we define the Lie product of any two elements X and Y to be [X, y] =XYYX •
26.
§2. DERIVATIONS ON SCAIARS
A simple computation shows that the Jacobi identity
(1)
[[X,Y],Z] + [[Y,Z],X] + [[Z,X],y] = 0
holds in any associative ring. Since D is closed under the formation of
Lie products, the Jacobi identity holds in it. (However, with respect to
the Lie product as multiplication D is not in general associative.) The
Jacobi identity may be rewritten as
[X,[Y,Z]]  [Y,[X,Z]]
[[X;~],Z] .
Define eX on D by
Then the Jacobi identity (2) is
(4)
More generally, if Y is an iterated Lie product of n elements
Xl"",Xn' associated in any way, then by induction ~ is the sum of n terms, in the ~th of which X is replaced by a~ . For example,
~ W~
[w, [[X,Y],Z]
[[ [w ,X], Y],Z] + [[X, [W, Y] ],Z] + [[X, Y], [W ,Z]]
If X and Y are in D then clearly
(6)
[Y,X] = [X,Y] .
The Lie product is FObilinear, so that with the Lie product as multiplication D is an FO algebra (not in general associative). An FO algebra
satisfying (6) and (1) is called a Lie algebra, so that D is a Lie algebra
over
However, the Lie product is not Fbillnear. In fact,
[f'X, gY] = fg[X, y] + f(Xg)Y  g(Yf)X
§2. DERIVATIONS ON SCAIARS
27·
2. Lie modules
Definition. A Lie module is an F module E together with an Flinear mapping X ~ X· of E into derivations of F and a mapping (X,Y) ~ [X,Y] of EX& into E such that with respect to it E is a Lie algebra over Fa and
(8) [X,fY] = f[X,Y] + (X·f)Y ,
(9) [X,YJ·f = X·Y·f Y·X·f
for all f in F and X, Y in E •
Notice that from (8) and the fact that [Y,X] is a Lie algebra over pO) the more general relation
[X,Y] (since E
(10) [fX,gY] = fg[X,Y] +f(X·g)yg(Y·f)X
holds. We have the following.
Theorem 1.
Let D be the module of all derivations of F, with
X·f = Xf and [X,Y] = XYYX. Then D is a Lie module.
The main example of interest is the Lie module E of all derivations of the algebra F of COO functions on a manifold, which may be identified with the set of all contravariant vector fields. In the definition of Lie module we did not assume that the mapping X ~ X· is injective. Other examples of a Lie module are a Lie algebra over FO when F = FO , and the set of all vector fields on a manifold invariant under some COO action of
c Lie group over the algebra F of all invariant scalars.
Roughly speaking, a Lie module is like a Lie algebra except that the elements of the module act on the coefficients by derivations in a natural ves .
28.
§2 . DERIVATIONS ON SCAIARS
Definition. let E be a Lie module, f a scalar. The differential
of f, df , is defined by
(df)(X) = X·f ,
X € E
Since X ~ X· is Flinear, the differential of f is a Iform
(§1.2).
3. Coordinate Lie modules
Definition. A Lie module E is called a coordinate Lie module in case there exist scalars xl, ... ,xn (called coordinates) whose differentials
are a basis for the module of Iforms.
Let E be a coordinate Lie module with coordinates
I n
x , ... ,x
I
Then E , and consequently E = EI ' is free of finite type (§1.11) and
therefore totally reflexive (§1.5). The
(11)
I n dx, ... ,dx
are a basis of EI •
The dual basis, which is a basis of E, is denoted
(12)
, ... ,
oxn
so that
Observe that the elements of the basis (12) commute (Le. their Lie
products are 0) since
k 0 0 0 0 k
< ax '[1' .J> =. . x
oX (3xJ ox~ oxJ
o 0 k
..X
oxJ (3x~
§2.
DERIVATIONS ON SCAIARS
29·
due to the fact that the
k k
OJ' 0i are constants (lor 0).
Notice that the
symbol O/Ox1, for example, has meaning only if the entire coordinate sys
1 n 1
tem x , .•. ,x is given (in contrast to ax ,which is simply the differ
1 1 2 l' 1
ential of the scalar x) . Thus if x ,x are coordinates so are x = x ,
l' 1 even though x = x .
Let f be a scalar, let X be a vector field with components xi, and let Y be a vector field with components yi. Then df has com
ponents
X·f is the scalar
and [X, y] has components
l' n' If x , ... ,x
are also coordinates then the Jacobian matrix
J~'
].
of §l.ll is
and similarly for its inverse, so that we have the familiar formulas
i' axi' = .E Ox. axj oxJ
o oxj 0 oxil = .E axil axj
30. §2. DERIVATIONS ON SCAIARS
4. Vector fields and flows
Let us discuss heuristically but in some detail why contravariant
vector fields on a manifold are the same as derivations of the algebra of
scalars, and what the geometrical mean:ing of scalar multiples, sums, and
Lie products is.
Let X be a (contravariant) vector field on a manifold M, everything being assumed to be of class Coo. Thus X is an assignment of
velocities at each point p of the manifold. The fundamental existence
theorem for differential equations shows how to integrate to obta:in the
motnon of a particle starting at any point. In this way we obta:in (at
least if M is compact) a oneparameter family of diffeo=rphisms ~(t) of M onto itself such that ~(t+s) = ~(t)~(s) • We call this a flow,
and in this discussion we shall ignore the fact that in general ~(t) is only defined locally if M is not compact. Thus ~(t)p is the
position of a particle at time t if it starts at ~(O)p = P at time O. As discussed in §1.14, if f is a scalar (COO function) we have ~(t)*f *
given by (~(t) f)(p) = f(~(t)p. But ~(t)p is a representative of
the tangent vector X at p p
*
(~(t) f)(p) at t = 0 is known if X
p
(see §1.13), so that the derivative of
is known. We let
*
(Xf)(p) = lim (~(t) f)(p)  f(p)
t~ t
This gives the action of vector fields on scalars. It is perhaps worth
recalling why this is a derivation:
(X(fg»(p) = lim (fg)(~(t)p)  (fg)(p)
t~O t
= l' (f(~(t)P)f(P) g(~(t)p)+g(p) + f(~(t)p) + f(p) g(~(t)p)g(p»)
1m t 2 2 t
t~o
= (Xf)g + f(Xg»)(p) ..
§2. DERIVATIONS ON SCAIARB
3.l·
Conversely, if X is a derivation on the scalars and p is a point then
the f(p) + t(Xf)(p) , where
f is allowed to be
1 n
x , •.. ,x for a set of
local coordinates at p, determines a curve whose equivalence class (the
tangent vector at p) is independent of the choice of local coordinates.
Con~equently we identify the vector fields on a manifold with the deri
vat ions of the algebra of scalars. Notice that the assumption that everything is of class COO is necessary for this identification.
Now we shall discuss the meaning of scalar multiplication, addf,,
tion, and Lie products in terms of flows.
Let X generate the flow ~,let h be a scalar, and let bX
generate the flow 'f'. Then 'f' is the same as ~ except for a change
of time scale. The new velocity is bX, so we have (.letting s be the
new time parameter)
dt )
ds = h(~(t p) .
Thus
'f(s)p ~(t)p
where
ft 1
s = 0 h(~(r)p) dr •
If h(p) = 0 then 'f(s)p = p
Now let X generate ~,Y generate 'f', and X+Y generate !2.
If X and Y commute (i.e., if [X,Y] = 0) then ~ and i commute (locally) and !2 is given by
!2(t)p = ~(t)'f'(t)p
The general case is more complicated, and
32·
§2 • DERIVATIONS ON SCAIARS
Roughly speaking, the flow n is the simultaneous action of the flows <fl
and ~. To see (13) formally, notice that formally
* tx <fl(t) = e ,
* * *
since <fl(t+s) = <fl(t) <fl(s) and
d *
dt <fl(t) 't=O = X
Then the expansion of
(.!X !Y)
n n n
e e
t(X+Y) e
in powers of
in powers of t
is the same as the expansion of
t , except for a fraction of terms in each order such that the fraction
tends to 0 as n ~ 00. The product formula (13) in the very general
setting of semigroups on Banach spaces is due to Trotter [3].
Finally, let X generate the flow <fl, Y generate the flow~,
and lX, y] generate the flow e. Then
(14)
To see (14) forma.J.ly, we make a formal computation to second order in t;
tx tY tx tY e e e e
~~ ~_9 ~~ ~~ 2
= (l+tx + 2 X)(l+tY + 2 r)(ltx + 2 X)(ltY + 2 r) + oCt )
= 1 +t2:J? +t2y2 +t2xY_t2:J? _ t2xy _ t2yx_ t2y2 +t2xy + 0 ( t2 ) =1+t2[X,Y]+o(t2) .
If we replace t2 by t this gives, formally, et[X,y] = e"txe"tYe~tXe~tY + oCt)
. ~Ax AY Ax Ay)n
= l~ e e e e
n~oo
§2. DERIVATIONS ON SCALARS
33·
This computation concerns the action of the flows on scalars, and the
result is e~uivalent to (14) for the point flows. Helgason essentially gives a proof of (14) for the case when X and Y are in the Lie
algebra of a Lie group (see pages 96, 97 and 105 of [2, §1J). It should
not be difficult to establish (14) in the general case of vector fields
on a manifold.
We conclude with an example. Consider a car. The configuration
space of a car is the four dimensional manifold
parameterized by (x,y,~,e), where (x,y) are the Cartesian coordinates
of the center of the front axle, the angle ~ measures the direction in
which the car is headed, and e is the angle made by the front wheels
with the car. (MOre realistically, the configuration space is the open
submanifold
e < e < e .)
max max
See Figure 2.
There are two distinguished vector fields, called Steer and
Drive, on M corresponding to the two ways in which we can change the
configuration of a car.
Clearly
d Steer = de
since in the corresponding flow e changes at a uniform rate while
x,y and ~ remain the same. To compute Drive, suppose that the car, starting in the configuration (x,y,~,e), moves an infinitesimal dis
tance h in the direction in which the front wheels are pointing. In
the notation of Figure 3,
D = (x+h cos(~e)+o(h), y+h sin(cp+e)+o(h)).
Let t = AB be the length of the tie rod (if that is the name of the
34.
§2. DERIVATIONS ON SCAIARS
I
I~e .: , \ /
I / /
Figure 2. A car
A
Figure 3. A car in motion
§2 • DERIVATIONS ON SCAIARS
35·
thing connecti~g the front and rear axles). Then CD = t too since the
tie rod does not change length (in nonrelativistic mechanics). It is
readily seen that CE = t+o(h) , and since DE = h sin e+o(h) the angle BCD (which is the increment in ~) is h sine/l ,while e remains the
s arne. let us choose units so that t = 1. Then
(16)
Drive = cos(q>+e) fx + sin(q>+e) ~ + sin e ~ ..
By (15) and (16),
000 [Steer, Drive] = sin(q>+e) OX + coslq>+e) oy + cos e ~
Slide = sin ~ Jk + cos
o Rotate = ~
Then the Lie product of steer and Drive is equal to Slide + Rotate on
e = 0 , and generates a flow which is the simultaneous action of sliding
and rotating. This motion is just what is needed to get out of a tight parking spot. By formula (14) this motion may be approximated arbitrar
ily closely, even with a restriction emax < e < emax with emax arbitrarily small, in the following way: steer, drive, reverse steer,
reverse drive, steer, drive, reverse steer, •••
What makes the process
so laborious is the square roots in (14).
let us denote the Lie product (17) of Steer and Drive by Wriggle.
Then further s~le computations show that we have the commutation
relations
[Steer, Drive] = Wriggle, [Steer, Wriggle] Drive,
[Wriggle, Drive] Slide,
36.
§2. DERIVATIONS ON SCALARS
and the commutator of Slide with Steer, Drive, and Wriggle is zero. Thus the four vector fields span a four dimensional solvable Lie algebra over
lR.
To get out of an extremely tight parking spot, Wriggle is Lnsurficient because it may produce too much rotation. The last commutation relation shows, however, that one may get out of an arbitrarily tight parking spot in the following way: wriggle, drive, reverse wriggle (this requires a cool head), reverse drive, wriggle, drive, ...
The example illustrates a phenomenon of frequent occurrence in differential geometry, namely holonomy, or rather the lack of holonomy. The vector fields Steer and Drive, which at first sight give the only possible motions of a car, span a module over the scalars which is not closed under the formation of Lie products. That is, the field of two dimensional planes in the tangent bundle is not integrable (not involutive, not holonomic) and so is not the fiela of tangent planes to a family of two dimensional surfaces. Motions which at first sight are impossible can in fact be approximated arbitrarily closely (in the CO topology but not the Cl topology) by possible motions.
Reference
[3] H. F. Trotter, On the product of semigroups of operators, Proceedings of the American Mathematical Society 10(1959), 545551.
§3. Derivations on tensors
37.
1. Algebra derivations
let K be an If algebra, not necessarily commutative or associati ve , A derivation X of K is an FO linear mapping of K into itself such that X(uv) = (Xu)v +u(Xv) for all u and v in K. The
computation in §2.1 used neither commutativity nor associativity, so if
X and Y are derivations of K so is [X, yJ
The derivations lie in
the associative algebra of endomorphisms of K as an If vector space,
so the Jacobi identity holds. Thus the derivations of K form a Lie algebra over FO•
Now suppose that K is a graded algebra. That is, K is the
weak direct sum
00
where each KrKs C Kr+s' The Kr with r < 0 are usUally but not necessarily O. An FO linear mapping X of K into itself is homo
geneous of degree a if each :xx C K I and homogeneous if it is
r r+a
homogeneous of degree a for some a. The notions of a bigraded algebra, and bihomogeneous mappings of bidegree (a,b) , are defined similarly. An antiderivation of a graded algebra K is an FO linear mapping
of K into itself such that
X( uv) = (Xu)v + (_l)ru(Xv) I
U e Kr' v € K •
The anticommutator of X and Y is XY + YX. A simple calculation
establishes the following theorem.
Theorem 1. let X and Y be antider1vations on the graded
algebra K, homogeneous of add degrees a and b respectively. Then
the anticommutator XY + YX is a derivation of K, homogeneous of de
~ a+b.
§ 3. DERIVATIONS ON TENSORS
2. Module derivations
Let E be an F module. In §1.14 we defined the notion of a homomorphism of (F,E) and consequently we have the notion of an auto
morphism of (F,E). Formally, let pet) be a oneparameter group of automorphisms of (F,E) and let cp be the derivative of p at t =0 • *
(For example, pet) may be ~(t) where ~(t) is a flow on a mani
fold.) Ey the product rule for differentiation we obtain, formally,
cp(iX) = fcp(X) + cp(f)X ,
f E F, X E E •
This motivates the following definition.
Definition. Let E be an F module. A derivation cp of (F,E) is a derivation cp of F and an FO_linear mapping cp: E ~ E such that (1) holds. A derivation cp of (F,E,E') is a derivation cp of (F,E) and an FO linear mapping q>: E' ~ E' such that in addition
cp(fW) = fcp(w) + cp(f)w ,
f E F, WEE',
cp<w,X> = <cP,w X> + <w,qOC> ,
WEE', X E E •
The motivation for (2) and (3) is again the product rule for
differentiation, as it is of course the motivation for the definition of derivation of an FO algebra. We shall find it convenient to indicate
derivations of a number of different F modules by the same symbol cp.
This is legitimate, provided of course they all give the same derivation
of F, since we may regard cp as defined on the disjoint union of F
and the various F modules.
Theorem 2. Let E be an F module and let cp be a derivation
of (F,E). Then:
(a) cp has a unique extension to a derivation cp of (F,E,E').
§ 3. DERIVATIONS ON TENSORS
39·
(b) If' we extend cp to E" ~ cp(KX:) = Kcp(X) for a.ll X
in E, ~ K is the natural mapping of E into E" .
(c) cp ~ (F,E,E') has a unique extension which is a deriva
~ 0 ~
tion cp of E* as an F algebra, and cp is a derivation on (F,E) r
for all '[' and s .
Os r
If' we extend cp to be a derivation on (F,E ,E) and r s
define
* *
cp on all of E* by additivity, then cp is a derivation of E*
as an ~ algebra.
For all u
r 1 ~ r
+ Zu(w , ..• ,cpw , ... ,w ,~, •.. ,Xs)
~=l
a 1
+ Z u(w , ... ,(,,r,xl, ... ,q:DC , ... ,x)
~=l ~ s
Proof. We define cp on E'
Then each cpw is a
lform since if we replace X by fX the a.d.ditiona.l terms cp(f) <>w,X > on the two sides of (3) canc~l. Similarly, (2) holds, so that cp is a
derivation of (F,E,E')
The uniqueness is clear.
By (a) we do have a undque extension of cp to E" (such that it is a derivation of (F,E' ,Eli». For all X in E,
cp<w,KX:> = <cpw,KX:> + <w,cp(KX:) >
By definition of K, K may be dropped from the first two terms of this, and by comparison with (3) we see that <w,cp(KX:) > = < w,q:DC> =
< w ,K(q:DC) > for all winE' , so that cp(KX:) = K(q:DC) •
The uniqueness assertion in (c) is clear since F, E, and E'
0* 0
generate E* as an l algebra. To prove existence, we need only show
40.
§3. DERIVATIONS ON TENSORS
that if H and K are F modules, cp a derivation on (F,H) and cp a derivation on (F,K) (agreeing on F), then
cp(X®Y) = cpX®Y + X®cpY
is welldefined on H®K and extends by additivity to a derivation of (F,H8K). To see this, notice that (5) is obviously welldefined on the free Abelian group used in the definition of tensor product (§1.3) and that cp sends (fX)®y  X®(fY) into the subgroup generated by the re
lations imposed in §1.3, and so is welldefined on H8K. It is then
clear that cp is a derivation on (F,H8K)
By (a) and (c) and the fact (§1.6) that Er is the dual of Os
s Er'
0
cp has a unique extension as a derivation of (F ES Er) We extend cp
, r' s
* u € ES, s' or or'
by additivity to E* . Let v € Er" y € Es' z € Es' Then,
r cp<~,;y0z >
<cp(~) ,;y0z > + < ~,cp,y®z +;y®pz > ,
so that
«q:ro.)(y»v(z) + u(cpy)v(z) +u(y)(_(cpv){z» +u{y)v(cpz)
= < cp{~),;y0z > + u(cpy)v(z) + u(y)v(cpz) .
* That is, cpt.&" + U8tpv = cp(~) , so that cp is a derivation of E* as
an FO algebra.
The final formula '4) is simply (3) for (}® ... ®J ®Xl ® ... ®Xs
u in ES and r
2r in j!; s
QED.
Notice that by (b) if E is totally reflexive, so that we may identi:t'y ~s and ES , the two definitions of cp agree. The various
r r
derivations given by Theorem 2 will be called the derivations induced
§3. DERIVATIONS ON TENSORS
41.
by the derivation ~ of (F,E)
Theorem 3. L:t E be an F module and let ~1'~2 be derivations of (F,E). Then the commutator [~1'~2] of ~l and ~2 ~
derivation of (F,E)
The commutators of the derivations induced by <lJ..
and ~2 are the derivations induced by [~1'~2]'
Proof. We have
~1~2(fX) = ~l (f~2(X) + ~2(f)X}
= f~l ~2 (X) + ~l (f)~2 (X) + ~2 (f)~l (X) + ~l (CP2 (f»X ,
and similarly for CP2~1 (rx) , so that
and we know that [~1'~2] is a derivation of F. The last statement of the theorem is an immediate consequence of the uniqueness assertions
in Theorem 2. QED.
By Theorem 2 and §1.6, if ~ is a derivation of (F,E) ,
rst
u E s+r ' , and
L = ~(r,s,rt,st) we have the following diagram:
~s LU > Est
r rl
In general it does not commute, and [~, LU] = cpo (LU)  (LU)o~ is not O.
Theorem 4. L:t E be an F module, ~ a derivation of (F,E) ,
rSI
U E s+rl' and t
L(r,f;.,r' ,s,)
Then
(6)
42.
§ 3. DERIVATIONS ON TENSORS
Proof. By the definition (§1.6) of t, if y € ~; , Z € ~I , and v = (tu)(y) then v(z) =u(;Y®Z). Therefore, by (4),
( cpv) ( z.) = cp( v( a)  v( cpz )
= cp(u(y®z» u(y®cpz) = (cpu)(y®z) +u(cpy®z)
so that (6) holds. QED.
3. Lie derivatives
Suppose tha.t E is a Lie module (§2.2). By the definition of
Lie module, if X € E and we let
exf = X·f , eXY = [X,Y] ,
l' € F
Y € E ,
then eX is a derivation of (F,E). The induced derivation eX on the mixed tensor algebra. is ca.lled the Lie derivative. Thus ex is defined on lforms w by (3), which gives
x·<w,Y> = <e:xw,Y> + <w,[X,Y]>,
Y € E
and for tensors u contravariant of rank r and covariant of rank s
by
s 1 r
+ L u(w , ... ,w ,Xl' •.. , [X,X 1, ... ,X )
~=l ~ s
If X is a vector field on a manifold, generating the flov ~ ( t),
then for any tensor field u, exu is the deri vati ve at t = 0 of *
~(t) u (see [4]).
§3. DERIVATIONS ON TENSORS
4. F1inear derivations
:ret E be an F module and let A be an Fl1near transfo:rms.
tion of E into itself. (If E is reflexive the set of Flinear transformations of E into itself can be identified with Ei, by §1.6). Define (JlA on (F,E) by
(PAf = 0 , f E F ,
(PAX = AX , X E E .
Then (PA is clearly a derivation of (F,E) , and every derivation of (F,E which is Flinear and. 0 on F is of this type. The induced deriva
tions are also denoted (PI..' By (3), CPA on E' is _At, where A'
is the dual of A •
By (4) we have for
We shall have occasion later (§7) to use a related notion.
Theorem ,. :ret E be a totaJ..ly reflexive F module. There is a unique Flinear mapping ~ of ~ into the F module of all
 *
Flinear mappings of the mixed tensor algebra E* into itself such that
1
for all B and C in El,
(10) ~B8'C = (Plf(Jlc
For each A in ~, ~A maps each ES into itself and each Ar into
r
itse1fz and ~A is 0 on F • It a is in Ar then 44.
§ 3. DERIVATIONS ON TENSORS
(11)
+ 2.L: (_l)ll+v Aba Q ,,"
~" i iv abi_ ... i ... i ... i
v~ Il ~ V Il r
Proof.
Since E is totalzy reflexive, ~ = El®E1. The map""211
ping
which sends (B/O) to ~B.~C is Fbilinear and so (by the definition
)11 of tensllr product has a unique F1inear extension ~ to E1 ®E1 .
Since
1 s r
~B for B in Ei sends each Er and A into itself and is
o on F, ~A has the same properties.
The notation in (11) is that of §1.12. To prove (11) it suf
fices to consider the case A = B®C
By (9),
(ia)
r
(~~) = .L: (_1)1l ca a " ,
c il.··i 1 i ail··.i .•• i
r Il= Il Il r
where we have made use of the fact that a is Blternating. If we use
(ia) to compute ~B~cf' we find (11) for the case A = B® C QED.
5. Derivations on modules which are free of finite type
Theorem 6. ~ E ~ F module, free of finite type with basis X1, •.. ,Xn. Let ~ be a derivation of (F,E) and define ri
~(X ) = 1::f'JX
i i J
If u € E~ the components of cpu are
§ 3. DERIVATIONS ON TENSORS
j' _k
If Xl'" ",Xnl is another basis and Ji and ~.i' are defined as in
§l.ll, and if ri: is defined by
then
Let E be a coordinate Lie module with coordinates
1 n
x , .. '/X
and let the vector field X have components Xi. Then the ~i ~sponding to the Lie derivative eX are
i' If X has components X
with respect to new coordinates
l' n
x , ... ,x
then
where we use the notation
to mean
Proof. The proof is trivial. QED.
Formula (13) shows the basic fact that the partial derivatives
of the components of a tensor do not in general form the components of
46.
§ 3. DERIVATIONS ON TENSORS
a tensor. This was what led Christoffel to the notion of covariant differentiation (§5).
Reference
[4] R. S. Palais, A definition of the exterior derivative in terms of Lie derivatives, Proceedings of the American Mathematical Society ~(l954), 902908. In the definition on p.908, ~ should be identified with (k+l)de rather than with de itself.
§4. The exterior derivative
1. The exterior derivative in local coordinates *
let E be an F module, A the Grasema.nn algebra (§1.9).
By an exterior derivative we mean an ~ linear mapping d of A * into
itself such that
[Dl] [D2] [D3]
d(~~) = d~~ + (_l)r~d~ , 2
d = 0 •
* ~ € A ,
Theorem 1. let E be a coordinate Lie module. Then there is
a unique exterior derivative d such that for all scalars f, driB
the differential of f.
Proof.
let
1 n
x , ... ,x be coordinates.
Then each 0 in Ar
is uniquely of the form
a=
If [tel and [:D3] hold tnen we must have
a i il ir
L: L: . ai i dx ~dx A. ••• A.dx
il<·· .<ir ax~ 1'" r
da =
That is, the components of dO are
(1)
This' proves uniqueness.
To prove existence, choose coordinates and define d by (1), extending to all of A * by additivity. Then d is FO linear, and [Dl] holds. The relation [D3] holds since the olCbt" and olaV conmru.te.
To prove [D2], let a € Ar, ~ € AS. By the explicit formula in §l.ll
48.
§4. THE EXTERIOR DERIVATIVE
for the components of ~~ and by (1), [re] holds. QED.
The proof shows, of course, that (1) holds for any choice of
coordinates. This is certainly the quickest approach to the exterior
derivative on a manifold, for once d is known locally it is trivial
to define it globally. However, a coordinatefree treatment of the
exterior derivatiye is worthwhile for several reasons. For one, it
applies to Lie modules which do not have coordinates (even locally, such as a Lie aJ.ge bra over F c FO) • The invariant expres s ions for d
are useful. Finally, it deepen@ one's understanding of the exterior
derivative and shows it to be the natural dual object to the Lie pro
duct.
2. The exterior deri vati ve considered globally
Theorem 2. If E is a totally reflexive Lie module there is a
unique exterior derivative d such that for all scalars f, df is
the differential of f and for all Iforms W and vector fields X
and Y,
(2) 2dw(X,y) = X.w(y)  y.w(X) w«(X,y])
*
If E is any Lie module and we define d on A by
r+l +1 "
E (_l)1l X .a:(x_, ... ,X , ... ,X 1)
1l=1 Il JL Il r+
*
and by extending d to all of A by additivity, then d is an exterior
derivative.
§4. THE EXTERIOR DERIVATIVE
49.
Proof.
o 1 *
For E totally reflexive, A and A generate A
as an FO algebra, so the uniqueness assertion is clear. (Notice that
the requirements that df be the differential of f and (2) are the
special cases r = 0 and r = 1 of (3).) Therefore we need only prove that d defined by (3) is an exterior derivative.
Let ~ denote the set of all FOmultilinear (not necessarily F multilinear) mappings of E X ... X E (r times) into F. Thus
If Z = (~, ••• 'Zr) is in E x •.. x E (r times), X is in
E , and 0 is in Er , we use the notation a([x,z]) as an abbrevi
ation for
For a in ~ we define
(oo)(X,Z) = x·a(z)  ~ a([x,z])
and (aa)(X) = x·a if r = 0
Then 00 is in ~+ 1 and is not in
general in ~ 1 even if 0 is in Er. Let P! be the set of all alternating elements in ?" , so that Ar C ? c? For 0 in K we define dO = Alt Oa ; that is,
For a in Ar this definition of da agrees with (3). It is trivial that d: Ar ~ Ar+l , and a simple computation shows that
We claim that d: Ar ~ Ar+l that is, if a is Fmultllinear so is dO. To see this, let 0 E Ar, f E F , and let
50.
§4. THE EXTERIOR DERIVATIVE
~ = (r+l)dO(fX1,X2,""Xr+1) , r = f(r+l)dO(X1'X2"",Xr+l) ,
r+l ~+1 h
5 = z (1) (X ·f)<X(X1,···,X ?."Xr+l)
~=l ~ ~
We must show that ~=7. But
1 r+l
f3 = fX1·O(X2, ••. ,Xr+l)  2' ~~Q(X2"",[~,X~]"."Xr+l)
r+l
( )~+1 ( h ) 1 ( ] h )
+ i: ~1 (X·o ~, •.. ,X , ... ,X 1 ~ 2' 0: [X , fXl , •.• ,X , ..• ,X ·1
~=2 ~ ~ r+ ~ ~ r+
 ~ z 0:(fX1,.··, [X ,XV]"· .,X +1)}
Vfl,~ ~ r
1 1
=72'5+52'5=7,
which proves the eJ.a.im.
It remains only to show that d2 = O. A simple computation
hr shows that for 0: in A ,
Let us therefore compute
for X and Y in E and Z in E X ••• X E (r times). Let f3(Y,Z):= da(Y,Z) , so that cfo:(x,y,z) = df3(X,Y,Z) • We have
~(y,Z) = Y·O(Z)  ~ o([y,Z]) ,
so that
§4o THE EXTERIOR DERIVATIVE
51.
02ex(X,y,Z) = xrotz) _ ~[X,Y]'ex(Z)  ~ Y'ex([X,Z])  ~ Xoa:([y,Z]) + ~ ex( [[X, y],Z]) + ~ ex( [Y, [x,Z])
Since Xoyoex(Z)  yoXoa:(Z) = [X, y] ·a:(Z) we find that
~ ex([[X,Y],Z]) + ~ a([y, [X,Z)])  ~ a([[Y,x],Z)  ~ e(Ix, [y,zll) = ~ a([[x,y],Z]) ,
where we have used the Jacobi identity to cancel the first, second, and
fourth terms of the second line and the antisymmetry of the Lie product
to rewrite the third termo Consequently (4) is equal to
and glJ.VA = (_l)IJ.+V+A+l if IJ.,V,A are in their natural order, glJ.VA changes sign under transpositions, and glJ.VA = 0 if IJ.,V,A are not
distinct. By the Jacobi identity again, the last sum is 0, so that
d2 '" O. QED 0
52.
§4. THE EXTERIOR DERIVATIVE
If E is any Lie module we shall call the operator d defined by (3) the exterior derivative.
Theorem 3. :ret E be a totaJ.:Qr reflexive F module, d ~
*
exterior derivative on A • Define X·f for X a vector field and f
a scalar by
(5) df(X) = X·f
and define [X,Y] ~ X and Y vector fields by
(6)
2dW(X,y) = X·w(y) y,w(X) w([X,Y]) ,
W e E' •
Then with respeot to these operations E is a Lie module. Thus for Ii
a totaJ.:Qr reflexive F module there is a onetoone correspondence be
tween structures of Lie module on E and exterior deri vati ves on the
Grassmann algebra.
Proof. Recall the definition of Lie module in §2.2.
Since df is a lform, X ~ X· is Flinear. Since
d(fg) = df,..g + f"dg , each X· is a derivation of F. Next we need to show that [X,Y] is in E; that is, that w([X,Y]) is Flinea.r
in w. But
(fW)([X,Y]) = X.«fW)(Y» Y'«fW)(X» 2(d(fw»)(X,Y)
= fX·w(y) + (X·f)w(Y)  ty·w(X)  (Y·f)w(X) 2(df"w)(X,Y)  f2dW(X,Y) ,
and by definition of the exterior product, 2(df"w)(X,Y) = df(X)w(Y) + w(X)df(Y) = (X·f)W(Y)+ (Y·f)w(y). Thus (tw)([X,Y]) = f(w([X,Y]),
and [X,Y] is in E. Also,
0= 2d2f(X,Y) = X.df(Y) Y·df(X) df([X,Y]) = X·Y·f  Y·X·f  [X,Y]·f .
§4. THE EXTERIOR DERIVATIVE
53·
Since 2dW(Y,X) = 2dw(X,Y) , we have [~,X] = [X>Y]. Since dw is
a 2form,
f2dw(X, Y) = 2dW(X, rr)
.. X·w(f'Y)  :ry.w(X)  w.([X,f'Y])
so that [X,f'Y] = f[x,Y] + (X·f)Y. It is clear that [X,Y] is FO_bi_
linear in X and Y, so it remains only to prove the Jacobi identity.
Define d by formula (3), so that d equals d on scalars
and lforms. The proof given in Theorem 2 tha.t d maps Ar into A r+ 1 and that d is an antiderivation did not use toe Jacobi ~dentity and
so remains valid under our present assumptions. Since E is totaJ.ly reflexive, A * is generated as an pP algebra by A 0 and Al. Consequently, d = d. Therefore we may use (3) to compute 0 = d2w(X,~Z) for w a lform. ret us use e to denote cyclic sums,
G? K(X, Y,Z) = K(X,Y,Z) + K(Y,Z,X) + K(Z,X, Y) ,
where K is aJJY :f'u.nction from E XE XE to an additive group. If' a is a 2form, (3) mB\Y be written
(8)
3da(X, r.z) = G {x·a(y,z)  o( lx, y],z)} •
If' we let a = 2dW ,where w is a l .. form, we obtain
o = 6d2w(X, Y,Z)
= 6 {X'Y'w(z)  X'Z,w(Y) • X'w([Y,z]) [x,y],w(Z) + z·w([x,y]) + w([[X,Y],z])} 6w([[x,Y],Z]) •
Since this is true for aJ.l lforms w, the Jacobi identity holds. QED.
54.
§4. THE EXTERIOR DERIVATIVE
More generally, the proof shows that if E is an arbitr~ F
module, if
*
d is an exterior derivative on A ,and we define
X·f
for X in El = ~' and f in F by (5) (i.e., (3) for r=O) and
[X,Y] for X and Y in El by (6) (i.e., (3) for r=l), and if (3) holds, then E1 is a Lie module. Thus Lie products and exterior derivatives are dual notions, and the Jacobi identity corresponds to the fac1
that an exterior derivative has square o.
3. The exterior derivative and interior multip1ication
Theorem 4. Let E be a Lie module, X in E. Then the anti
commutator of the exterior derivative and interior multiplication by X
is the Lie deri vati ve eX on exterior forms. That is,
* Q€A
If a is a closed exterior form, eXQ is exact.
Proof. We give the proof first under the assumption that E
is totally reflexive, since this is the CBBe of interest in differential
geometry and the proof is less computational. Since d is an anti*
derivation of A which is homogeneous of degree 1 and interior multi
*
plication by X is an antiderivation of A which is homogeneous of
* degree 1 (§l.lO), their anticommutator is a derivation of A which
is homogeneous of degree 0 (Theorem 1, §3.1). If E is totally reflexive, A * is generated as an FO algebra lIy A 0 and A 1 , so
we need only verif'y (9) for a a scalar or Iform. If Q = f is a scalar, (9) says that X·f = df(X) , which is true. If a = W is a Iform, (9) s~s that for a1l vector fields Y,
< exw,y> = <d(W(X»,Y> + 2clW(X,Y) ,
§4. THE EXTERIOR DERIVATIVE 55.
which is the same as (6). Thus (9) is true if E is totally reflexive.
The proof for the general case is similar: one verifies (9) for Q an
rform by the formula (3) for d, the definition (formula (8), §l.lO)
of interior multiplication, and the formula {§3.3) for Lie derivatives.
The computation is omitted.
The last assertion in the theorem is an immediate consequence
of (9). QED.
4. The cohomology ring
* Let d be an exterior derivative on the Grassmann algebra A
of an F module E. An exterior form Q is called closed if dQ = 0 ,
exact if a = d~ for some exterior form ~
Theorem 5. Let E be an F module, d an exterior derivative
* on the Grassmann algebra A
The set of closed exterior forms is a
graded ~ algebra in which the set of exact exterior forms is a homogeneous ideal, so that the quotient is a graded FO algebra.
Proof. The proof is trivial. QED.
*
The quotient algebra is denoted H ,with homogeneous subspaces
W. It is called the cohomology ring. The dimension of W as an ~ vector space is denoted br and called the rth Betti number. De Rham's theorem asserts that the cohomology ring formed from the COO exterior
forms on a COO manifold is a topological invariant, the cohomology ring
of the manifold with real coefficients.
§5. Covariant differentiation
57.
1. Affine connections in the sense of Koszul
On a differentiable manifold there is no intrinsic way of dif
ferentiating tensor fields to obtain tensor fields, covariant of one
rank higher, and to obtain such a "covariant derivative" we must impose
additional structure. In Riemannian geometry there is a natural notion
of covariant differentiation (§7) discovered by Christoffel. Many years
later LeviCivita discovered the geometrical meaning of covariant dif
ferentiation by integrating to obtain parallel translation along curves.
A number of people, especially Elie Cartan, studied nonRiemannian
"affine connections," and the notion was axiomatized in a convenient
way by Koszul, as follows.
Definition. Let E be a Lie module. An affine connection ~
on E is a function X ~~X from E to the set of mappings of E into itself satisfying
[V'l]
~ = :f'i7.X + rFy ,
fX+gY
for all vector fields X,Y,Z and scalars f and g.
Thus an a.ffine connection is an Flinear mapping X ~ ~X of
E
into derivations on (F,E)
such that for all scalars
f ,
'\1 f = X·f X
(see §3.2).
*
The derivation of the mixed tensor algebra E* induced by
~X (§3.2) is also denoted ~X ' and is called the covariant derivative in the direction of X (with respect to the given affine oonnection).
In particular we have
(1)
W € E', Y € E •
58.
§5. COVARIANT DIFFERENTIATION
2 . The covariant derivative
Covariant derivatives have an important property which is not
enjoyed by Lie derivatives:
Definition. Let 'V be an affine connection on the Lie moduJ.e
E , and let u be in ES r
We define the covariant derivative 'Vu by
Theorem 1. Let "V be an affine connection on the Lie module E , and let u be in ES• Then 'Vu is in ES+ 1 •
r r
Proof. Since "Vx u is a tensor the only point at issue is the 1
Flinearity of (2) in ~, and this is an immediate consequence of [\11]
(which remains true for the induced derivation of the mixed tensor
algebra). QED.
3. Components of affine connections
Theorem 2. Let E be a free Lie module of finite type, with
basis Xl' ... ,Xn ' and let "V be an affine connection on E. Define
'VX X. = .r.r1:.x_ .
i J ~JK
If u is in E~ the components of 'Vu are
(4)
r i1···i lai 1 ... i i
+ Z ~ ~+ rr ~
lUjl ... j ia
~= s
s L ... r
z ra u . .L .r .
1 ij J1···J laJ l···j
~= ~ ~ ~+ s
§5. COVARIANT DIFFERENrIATION
59·
If the ~: j' are defined in the same Way with respect to an
other basis Xl" ... ,Xn, then

and }:,
 J
are defined as in §l.ll
Proof. This follows eas ily from § 3.5. QED.
Notice that if the ~j are an arbitrary set of n3 scalars there is a unique affine connection on E satisfying (3) for the given basis. The ~j are called the components of the affine connection (with respect to the given basis). The components may be 0 with re
spect to one basis but not with respect to another.
If E is a coordinate Lie module (§2.3) with coordinates xl, .•. ,xn then (5) takes the more familiar form
(6)
For some applications in differential geometry, a coordinate system
does not give the most convenient local basis for the vector fields.
For example, on a Lie group it is usually convenient to choose a basis
of leftinvariant vector fields. These do not in general come from a
coordinate system since they do not in general commute.
It is customary to denote the left hand side of (4) by
The notations
and
are also in use to mean the same thing. However, it is convenient in
60.
§5. COVARIANT DIFFERENTIATION
connection with the exterior derivative (see paragraph 7) to have the
new covariant index be the first covariant index.
4. Classical tensor notation for the covariant derivative.
Ie t "il be an affine connection on the Lie module E and recall
the global meaning of the classical tensor notation (.§l.12). Following
convention, we write
Notice therefore that
On the other hand, if X and y are vector fields and u is in ES
r '
"ilX"ilyU is again in ES and is in general quite different from the
r
tensor in ES obtained by substituting X and Y in the first two
r contravariant arguments of Wu. See paragraph 11.
5. Affine connections and tensors
wt E be a totally reflexive Lie module. We have seen (§l.6) that tensors B in ~ may be regarded as Fbilinear mapp:iIgsof E XE into E, and we write B(X,Y) or ExY for the image vector field.
Let E2 be the set of all FObilinear (not necessarily Fbilinear) map 1
pings of E X E into E. Thus Ei is an FO vector subspace of ~ .
is an FO vector space, and ii other examp.Iea of elements of Ei
are affine connections and the Lie product. The Lie module E is
called a Lie algebra over
F
in case the Lie product is in
E2 ; that 1
element of
[X, y] is Fbilinear in X and Y. When we say that an "2
E: is a tensor we mean that it lies in the vector subspace 1
is, in case
~.
§5. COVARIANT DIFFERENTIATION
61.
Theorem 3. let E be a totally reflexive Lie mod .:J.e which 15
not a Lie algebra over F. Then no affine connection is a tensor. ~e
difference of any two affine connections is a tensor, and if ~ is an affine connection and B is a tensor in :S ~ ~+ B is an affine connection.
Let ~(a) be a finite set of affine connections and let ~(a)
be scalars. Then
np(a)~(a)
a
is an affine connection if np(a) = 1 . a
Proof. By ['\72 ] an affine connection ~ is a tensor if and only if (X'f)Y = 0 for all vector fields X and Y and scalars f
L, e., if and only if E is a Lie algebra over F.
let ~(l) and ~(2) be affine connections. and let
B = ~(l) _ ~(2). By Ivi l, BxY is Flinear in X and by ['\72] it is Flinear in Y, so it is a tensor. If V is an affine connection and 2
B is in El then ~+B satisfies [~O], [V'l), and ['\72], and so is an
affine connection. The second paragraph of the theorem is obvious. QED.
By the theorem, the set of all affine connections on a totally
reflexive Lie module E is either empty or is an affine subspace parallel to but disjoint from :S' Thus the choice of any affine connection on E establishes a onetoone correspondence between all affine
connections and all tensor fields contravariant of rank 1 and covariant
of rank 2.
Affine connections always exist on (paracompact) COO manifolds.
The proof of the theorem shows that if E is totally reflexive
and is a Lie algebra over F then the set of all affine connection on E is ~.
62.
§5. COVARIANT DIFFERENTIATION
6. Torsion
Definition. Let ~ be an affine connection on the Lie module
E. The torsion of ~ is the mapping T of E XE into E defined
by
(8)
T(X,Y) = ~XY  ~..f  [X, y] .
The torsion tensor of ~ is the mapping T of E' XE XE into F
defined by
<to),T(X,Y» •
Theorem 4.
The torsion tensor of an affine connection is a
tensor.
Proof. Since (9) is clearly Flinear in to) , we need only show that T(X, Y) is Fbilinear in X and Y, and since
(10)
T(Y,X) = T(X,Y)
by (8), we need only show that T(X,Y) is Flinear in X. But
T(fX, Y) = ~ fXY  ~yCfX)  Irx, xl
= fT(X,Y)  (Y·f)X + (Y·f)X = fT(X,Y) QED.
If E is totally reflexive we identify the torsion tensor and the torsion, as in the preceding paragraph (since T € ~). The torsion T maps E2 into El so its dual T' maps El into ~ (if
E is totally reflexive), but El = Al and by (10) it is clear that T' Al into A2 •
maps
An affine connection is called torsionfree if its torsion is
O. (It is sometimes called symmetric, but this term is generally
reserved for a torsionfree affine connection ~ such that in addition
§5. COVARIANT DIFFERENTIATION
Y'R = 0 where R is the curvature tensor.) The most important class
of affine connections, the Riemannian connections (§7), are torsionfree.
Theorem 5. Let E be a coordinate Lie module with coordinates
1 n J
x , ... ,x ,and let ~ be an affine connection with cOmponents Iij .
Then the components of its torsion tensor T are
(ll)
Proof.
k (0 0 0
<dx ,T i' j >
Ox ox
If E is merely a free Lie module of finite type with basis
Xl' •.• ,Xn we define the structure scalars
in F by
Then (11) must. be modified to read
Theorem 6. Let ~ be an affine connection on the Lie module
E. Then so is ~r defined by
and 'ii'" = ~
The affine connection
is torsionfree. If T is the torsion of ~ then T is the torsion
of ~I , and ~, = ~T •
64.
§5· COVARIANT DIFFERENTIATION
Every a.1'fine connection "V on a. Lie module E can be written
uniquely in the form
(14)
"
where "V is a torsionfree affine connection and T is a skewsymmetric
Fbilinear mapping of E XE into E. In this decomposition T is the
"
torsion "V and "V is given by (13).
Proof. ret "V be an affine connection with torsion T. By
the definitions of T and "V' , 'i7' = "VT , and since T is Fbilinear
'i7' is an a.1'fine connection. The torsion of 'i7' is given by
"VxY  'i7yx  Ix, yJ = "VxY  T(X, y)  'i7f + T(Y,X)  lx, y] = T(Y,X) T(X, r).
By the second paragraph of Theorem 3, ~ defined by (13) is an a.1'f'ine
1 1
connection, and its torsion is clearly 2 the torsion of 'i7 plus 2 the
torsion of "V' , which is O. Thus every a.1'fine connection 'i7 can be
"
written in the form (14).mere 'i7 is the torsionfree affine connection
(13) and T is the torsion of "V. It remains to prove the uniqueness of the decomposition (14), so suppose "V is represented in the form (14) with ~ a torsionfree affine connection and T a skewsymmetric
Fbilinear mapping of E XE into E. Then the torsion of "V is the torsion of ~ (which is 0) plus ~(X, Y)  ~(Y,x) = T(X, y) ,so T
is indeed the torsion of "V . QED.
7. Torsionfree affine connections and the exterior derivative
Theorem 7. Let "V be a torsionfree affine connection on the
Lie module E and let d be the exterior derivative. Then d =Alt "V
* on A .
§5. COVARIANT DIFFERENTIATION
~ Define d on A* by d = Alt 'V
Suppose first that
E is totally reflexive, so that A* is generated as an rP algebra
by AO and AI. It is trivial to verity that d is an antiderivation
* A
of A ,so we need only verify that d equals d on scalars and
lforms. On scalars, 'V is just the differential. let II) be a lform.
Then
so that, since 'V is torsionfree,
2(Alt ~)(X, y)
If E is not totally reflexive, recall that the exterior de
rivative d is defined by formuJ.a (3) of §4.2.
A
One verifies that d = d
by direct computation. QED.
This is a useful theorem. The exterior d.erivative was defined to be Alt d ,but d is not a mapping of tensor fields into tensor
fields whereas 'V is •
8. Curvature
Definition. let 'V be an affine connection on the Lie module E. The curtature of 'V is the function R from E XE into the set
of mappings of E into itself given by
66.
§5. COVARIANT DIFFERENTIATION
..
The curvature tensor of 'V is the mapping of E' XE XE XE into F
given by
R(W,Z,X,y)
<w,R(X,Y)Z> .
Theorem 8,
The curvature tensor of an affine connection is a
tensor.
Proof.
Since R(X,y)Z is in E, (16) is clearly Flinear in
w. If we replace Z by fZ the coefficient is differentiated via
X·Y·fY·X·f [X,Y].f = ° , so (16) is Flinear in Z. Similarly, if
we replace X by fX the coefficient is differentiated via
0 Y·f + Y·f = ° , so (16) is Flinear in X. Since
R(Y,X) = R(X,Y)
it is Flinear in Y as well, and so is a tensor in
QED.
Notice the order of the contravariant vector fields in the defi
nit ion (16) of the curvature tensor. If E is totally reflexive the
above identification of the curvature R with the curvature tensor is not the same as the identification L(2,0,1,i) of Ei with the set of
Flinear mappings of E2
1
(where El in turn is identified
with the set of Flinear mappings of E into itself) but is the composition of L(2,0,1,1) with a permutation. The definition (16) is the
universally adopted convention.
Theorem 9.
Let E be a coordinate Lie module with coordinates
1 n j
x , •.. ,x and let 'V be an affine connection with components F1j.
The components of the curvature tensor R are given by
(18)
§5. COVARIANT DIFFERENTIATION
Proof.
and the computation showing that this is the same as (18) is not only
trivial but easy. QED.
If E is merely a free Lie module of finite type wlth basis Xl' ... ,Xn we must modify (18) by replacing d/~k by ~. and d/~l by Xl' I and by adding the term
_L:ca ri kl aj
where the c~ are the structure scalars defined by (12).
9. Affine connections on Lie algebras
Let E be a Lie algebra over F = FO
As we have seen (in
paragraph 5), an affine connection 'V on E is the same as an FO bi
linear map B of EX E into itself.
If G is a Lie group its Lie algebra ~ (over JR) may be identified with the set of leftinvariant vector fields on G. Affine
connections B on 'J such B(X,X) = 0 for all X in 7J (which is the same as B(Y,X) = B(X, Y) for all X, Y in ~ ) are of particular interest, for this means that each left, invariant vector field X is
autoparallel and the geodesics (with respect to the affine connection)
issuing from the identity of G are precisely the oneparameter sub
groups. The cases
B(X,Y) = 0 (the () connection) ,
1 (the (0) connection) ,
B(X,Y) = '2[X,y]
B(X, Y) = [X, xl (the « ) connection) ,
.. 68.
§5. COVARIANT DIFFERENTIATION
have been studied by E. Cartan and Bchoirten, They make sense as affine
connections whenever E is a Lie algebra over F.
From the definition (8) of torsion, the torsions of the above three affine connections are respectively [X,Y), 0, and [X,Y). The curvature (15) of the 1) connection is 0 since the connection itself is 0 and the curvature of the (+) connection is 0 by the Jacobi identity. For the (0) connection the curvature is given by
again by the Jacobi identity. This is 0 for Abelian Lie algebras and certain nilpotent Lie algebras. Let us, as an exercise, compute ~ Since R(w,X, Y,z) is a scalar and the differe",tial of any scalar is 0 (in the Lie algebra case) we have, for any WinE,
o = (~wR)(w,Z,X,y) + R(~WW,Z,X,y) + R(W,~WK'Y'Z) + R(W,X,~wY'Z) + R(W,X,y,~wZ) = (~wR)(w,z,X,y)
+ ~W, {[W, ux, y] ,Z])[[X,X), [W ,Z])[[ [W ,X), Y],Z)  [[X, [W, y)] ,Z]} > = (~wR)(w,Z,X,y)
Since, by the Jacobi identity, Sw is a derivation (cf.§2.1). Thus the (0) connection is symmetric; that is, the torsion is 0 and VR = 0 • We shall not return to this subject. See Helgason [2,§1].
10. The Bianchi identities
Following Nomizu [5] we prove some identities relating curvature and tors ion. As in §4. 2, G denotes cyclic sums.
§5· COVARIANT DIFFERENTIATION
Theorem 10. ~ \1 be an afi"ine connection on a Lie moduil:e E wi th torsion T and curvature R. Then, for all vector fields X, Y,Z I the following identities hold:
(19) (20) (21) (22)
T(X,Y) = T(Y,X) , R(X,Y) = R(Y,X) ,
e R(X, Y)Z = G T(T(X, Y) ,Z) + G (\1XT)(Y,z) , 6 (\1zR)(x,y) + GR(T(X,Y),Z) = 0
In Earticular, if \1 is torsionfree
~ R(X,Y)Z= 0 , (0 (\1zR)(X, Y) = 0
Proof. We have already noted the trivial identities (19) and (20). To prove (21) and (22) we need
(\1ZT)(X, Y) = \1Z(T(X, Y»  T(\1ZX,Y)  T(X, \1zY) , (\1za)(X,y) = [\1Z,R(X,y)]  R(\1ZX,Y)  R(X,\1ZY)
The relation (25) is an immediate consequence of the fact that \1Z is a derivation of the mixed tensor algebra. For (26) we need Theorem 3, §3.2 as well (notice that [\1Z,R(X,y)] is not a Lie product of vector field but a commutator of operators on vector fields).
To prove (21), we use the definition (8) of torsion and (19) to
find
T(T(X,Y),Z) = T(\1XY'Z) + T(Z,\1yX)  T([X,Y],Z) •
We use (25) to reexpress the first two terms on the right of (27) and
make the discovery that
70.
§5· COVARIANT DIFFERENTIATION
GT(T(X,y),Z) =  C9 ('VZT)(X,y) + G)'VZ(T(X,y))  GT([X,Y],z)
8 (('VZT)(X,y) +'Vz'VxY'Vz'V..f 'Vz[X,y] _'V[X,y]z+'Vz[X,y]  [[x,y],zJ} ~ (('VZT)(X,Y) + R(X,Y)Z}
by the Jacobi identity and the definition (15) of curvature. Thus (21)
holds.
To prove (22), use the definition (8) of torsion and (20) to
find
R(T(X, y),z) = R('VXY  'V0  [X, x l.z)
= R('VXY'Z) + R(z''V0)  R([X,Y],z)
Sum cyclically and use (26) to obtain
By the definition (15) of curvature and the Jacobi identity,
~(['VZ,R(X,y)]  R«(X,Y],z)}
8 (['Vz'['Vx,'Vy] _'V[X,y]]  ['V[X,y]''Vz] + 'V[[X,y],Z]} 8 (['Vz'['Vx,'VyJ] + 'V[[X,y],Z]} = 0 •
QED.
The relation (24) (and sometimes (23)) for a torsionfree affine
connection is called Bianchi's identity. If 'V is a torsionfree
affine connection with curvature tensor R then
(28) (29 ) (30)
§5. COVARIANT DIFFERENTIATION
71.
since these are merely (23), (24), and (20) in a different notation.
11. Ricci's identity
The identities we prove next will be used frequently. Recall
the definition of ~A given in §3.4, where A is an Flinear mapping of an F module E into itself. If \l is an affine connection on a
Lie modul.e E
and u is in ES we will use the notations r
(V'u) (X)
and (VVU)(X,Y), it being understood that X and X,Y are the first contravariant arguments in V'u and vvu respectively. Thus (V'u)(X)
and (Wu) (X, Y)
are again tensors in ES • r
Theorem ll.
Let ~ be an affine connection on the Lie module
E with torsion T and curvature R. Then for all vector fields X,Y
and all mixed tensors u,
(Wu)(X,y)  (Wu)(Y,X) = ~R(X,Y) ~T(X,y)U
In particular, if ~ is torsionfree then
(Wu) (X, Y)  (Wu)(Y,X) ~R(X,y)u
is Flinear in u.
Proof.
For given X,Y in E, let
By §3.2 this is the derivation on the mixed tensor algebra induced by
the derivation [~X'~Y]  ~[X,y] on E, so to prove (31) we need verify it only on scalars (where both sides are 0) and vector fields (where
it is the definition of curvature).
72·
§5. COVARIANT DIFFERENTIATION
To prove ( 32 ), let u be in ES g; in ~r. Then
r' s
(Wu)(x, Y, z )  (Wu)(Y,X, z )
= (~x~)(Y,z)  (~Y~)(X,z)
= X·«~)(y,z»  (~)(~xY,z)  (~)(Y,~xZ) Y'«~)(X,z» + (~)(~0'Z) + ('V1J.)(x'~yZ)
= (~X~yu)(z)  (~~xyu)(z)  (~y~xu)(Z) + (~~0u)(z)
so that, by the definition (8) of torsion,
(Wu)(x, Y)  (Wu) (Y,X)
= (~X~y  ~y~X  V'[X,y])u  ~T(X,y)U .
By (31), this proves (32). QED.
If ~ is torsionfree we may write (33) as
s i1 ... i
ERa u r
1 j klj1···j 1aj l···j
~= ~ ~ ~+ s
This is Ricci' s identity. We emphas ize again that we follow the custom of writing
for
§5. COVARIANr DIFFERENTIATION
73·
J2 • Twisting and turning
Suppose we have an affine connection on the manifold M. Let
C be a parameterized curve t ~ Ct in M with tangent vector X, and let Y be a vector field on M. In local coordinates X has components dxk(t)/dt at Ct, where xk(t) is the ktih coordinate of Ct' Let Y nave componenbs yk(t) at Ct' Then 'V'XY at Ct has components
Tha.t is, we don't need to know Y off C to compute 'V'XY since X points in the Q1rection of C.
If we prescribe yk( 0) as the components of a tangent vector
Yp at the initial point p = Co of C and set (35) equal to 0, we have a wellposed initial value problem. When we integrate we obtain
a family of vectors Y along C satisfying 'V'XY = 0 and Y = Yp at p. let q " 01 be the final point of the curve segment C (<>.:::t51) • Then we obtain a. mapping
by setting ~ Y = Y • This mapping is linear and invertible, and we
c p q
obtain a groupoid in this way. The mapping ~ c is called parallel
translation along C
A particUlar case of (35) is 'V'0 along C, which has compc
nents
This is called the acceleration vector. Notice that although the velocity
vector of a particle moving on a differentiable manifold makes sense,
74.
§5· COVARIANT DIFFERENTIATION
there is no meaning to the notion of acceleration vector unless we have
additional structure such as an affine connection. Notice that (36) is
the same for aIfine connections having the same torsionfree part (see Theorem 6), since (dxi/dt) (dxj/dt) is symmetric in i and j. If
we start a particle at p at time 0 with prescribed initial velocity
X and required the acceleration (36) to be 0, we have a wellposed p
initial value problem for a secondorder differential equation. This
secondorder differential equation has the special property that if we
multiply X by a constant k the particle travels in the same trap
jectory with velocity k times the previous velocity, and is called a
spray. In fact, this is the geodesic spray of the affine connection
and the trajectories are called the geodesics of the affine connection.
Affine connections with the same torsionfree part give rise to the
same geodesic spray and the same geodesics. Any spray is the geodesic
spray of a unique torsionfree affine connection. On a manifold with
an affine connection (or with a spray) Newton's dynamical law F = rna is meaningful. (The proper setting for Hamiltonian dynamics is a sym
plectic manifold (§8).)
Suppose we have a torsionfree affine connection ~ on M.
let P
be a point in M
and let X, Y p p
be two tangent vectors at p
Choose curves at p with tangent vectors X, Y and let q and r p p
be the points on these curves corresponding to the parameter value t
(see Figure 4).
Parallel translate
q to obtain Yq and
parallel translate Xp to r to obtain Yr' and choose curves at r and q with these tangent vectors. Since ~ is torsionfree and ~XY
and ~0 are 0 at p, [X, xl is 0 at p
Therefore (§2.4) the
end points (corresponding to the parameter value t) of the curves
§5. COVARIANT DJl'FERENTIATION
75·
figure 4. A parallelogram
representing Xr and Yq are equal to second order in t. This is false if the affine connection has torsion.
Gi ven a spray or affine connection on M, let Yp be a tangent
vector at p
and let a particle start at p with velocity Y and p
travel for unit time with zero acceleration. Its position at unit time
is denoted exp Y , so that exp (the exponential mapping) is a mapping p
of the tangentspace M into the manifold M. It is always wellp
defined locally (and locally is a diffeomorphism) and is sometimes well
defined globally (in which case the spray or affine connection is called
complete) .
let C be a curve segment in M starting at p with tangent
vectors X , and given an affine connection '1 on M let Y (t) be
p
the parallel translate of Y in M along C to Ct For Y
p p p
sufficiently small, the curves exp Y (t) sweep out a tubular neighbor
p
hood of C . If '1 is torsionfree then the tangent vectors to these 76.
§5. COVARIANT DIFFERENTIATION
curves are, to first order, obtained by pare.l.lel translation of X as
in Figure 4. Now consider the affine connection with the same torsion
free part but with torsion T. Then in (35) we add the term
Since ~i = 0 ,
I k dxi
2' m'ij F is a linear transformation acting trans
versally to the direction of the curves
exp Y (t) , and it twists them p
around C. This is torsion.
Now let US consider curvature. A frame at a point p in the
manifold M is simply a basis of the tangent space M p
all frames at p is the principal homogeneous space of the general
The set of
linear group GL(n, It). If any frame is singled out, there is a unique element of GL(n, It) taking it into any given frame. The set of e.l.l
frames at all points is a principal fiber bundle over M with structural
group GL(n, lR). The projection maps each frame to the point Ii in M
at which it lives.
Let \l be an affine connection on M, C a curve segment on
M. Then parallel translation along C takes frames at the initial
point p into frames at the final point q. The frame travels along
C to q, picks up a tensor and brings it back to p for differentiation
or other purposes. It is similar to the repair truck which leaves the
Mobil station to tow back a car and in fact the notion is called the
rep ere mobile. If C is a closed loop at p parallel translation
around C gives an automorphism of M , and the set of such automorp
phisms forms a group called the holonomy group.
Rece.l.l from §2.4 the geometrical meaning of the Lie product of
t,wo vector fields X and Y, which may be expressed by saying that
§5. COVARIANT DID'ERENrIATION
77·
,
Figure 5. A loop
Figure 5 is a closed loop to second order in t
If we parallel trans
late a frame around this loop we will find in general that it has turned.
In fact, parallel translation of a tangent vector Z around the loop
gives to second order in t
2· 2
Z + t ('VX'VyZ  'Vy'VxZ  'V[X, y]Z) = Z + t R(X, Y)Z.
This is curvature.
Reference
[5] K. Nomizu, Lie Groups and Differential Geometry, ~l1cation of the Mathematical Society of Japan 2, (1956).
79.
§6. Holonomy
1. Principal fiber bundles
The frame bundle discussed in §5.J2 is an example of a principal fiber bundle. In a principal fiber bundle we have a c<X> mapping
71": P ~ B of the bundle P onto the base B, where P and B are c<X> manifolds. Each fiber 71"l(p), P EO B , is a principaJ. homogeneous space of a Lie group G, the structure group of the bundle. That is, each fiber is the same as G except that it has forgotten which element is the identity (cf'. frames vs , the general linear group). The structure group G acts by translations on each fiber and therefore on the whole bundle P, and this action is C<X>.
One can define the notion of connection in this setting and discuss curvature and holonomy and the relation of connections to reductions of the structure group (see Nomizu [5,§5]). The discussion is simpler if we confine ourselves entirely to Ginvariant objects on the bundle P. The invariant scalars on P are isomorphic to the algebra
of all scalars on B, and the invariant vector fields on P form a Lie module over them.
Each invariant vector field on P projects onto a vector field on B, and some of them (those which lie along the fibers) project
onto 0
We shall discuss this situation algebraically. Because of
our restriction to Ginvariant objects, the terminology is a bit different from the standard terminology [5, §5] • The usual treatment is complicated by the notion of the connection form, a notion which we avoid.
At each point in a principal fiber bundle we have a distinguished linear subspace of the tangent space, called the vertical space, consisting of those tangent vectors tangent to the fiber. A connection is
80.
§6. HOLONOMY
p
B
Figure 6. A horizontaJ. curve
a Ginvariant choice of' a. complementary linear subspace at each point, called the horizontaJ. space. The f'ield of' horizontal subspaces is not in generaJ. integrable, and this is where the notion of' holonomy enters (see §2.4). A curve in P is called horizontal if' its tangent vector aJ.ways lies in 'bhe horizontaJ. space. A horizontaJ. curve may cut a given f'iber in many points (see Fig.5), and the curvature of the connection describes the verticaJ. directions of' motion which can be approximated
by horizontaJ. curves.
2. Lie bundles
The collection of' all Lie modules over a f'ixed F f'orms a category if' we define a morphism of two Lie modules Q and P to be
an Fmodule homomorphism p: Q __.;:. P (i.e., an Flinear mapping of' Q into P such that
§6. HOLONOMY
(pX)·f = X·f , p[X,Y] = [pX,pY] ,
X e Q, f e F , X, Y € Q .
A morphism in this category will be called a Lie homomorphism.
An ideal V in a Lie module P is a submodule such that
Ix.z) € V ,
X € P, Z € V ,
and such that
Z·f = 0 ,
Z € V, f € F •
If P is a Lie module then (X € P: X·f = 0 for all f E F} is clearly an ideaJ., and every ideal is contained in it.
If V is an ideal in the Lie module P then the quotient module p/v is in the obvious way a Lie module, and the projection ~: P > p/v is a Lie homomorphism. If' p: P .;:.. Q is any Lie homomorphism then the kernel V of p is an ideal, the image pep) is a Lie module, and the induced mapping of p/v onto pcp) is an isomorphism.
A submodule K of a Lie module P is called a Lie submodule in case [X, rl € K whenever X and Y are in K. Then K is a Lie module and the inclusion L: K .;:.. P is a Lie homomorphism.
Let P be a Lie module and H a submodule (not necessarily a Lie submodule), and let p be the projection of P onto the module p/H For X and Y in H define
R(X,Y) = p[X,Y]
We call R the curvature of the submodule.
Theorem 1.
Let H be a submodule of the Lie Dl9dule P. Then
the curvature R is Fbilinear and antisymmetric, and is 0 ~ only if H is a Lie submodule.
82.
§6. HOIDNOMY
Proof. The curvature is Flinear in X since
R(fi,Y) = pffi"Y] = f'p[X,Y] ~ (Y'f')pX = f'p[X,Y] = f'R(X,Y) •
Since R is clearly antisymmetric it is also Flinear in Y. The last
statement in the theorem is obvious: QED.
Definition. A Lie bundle 'IT": P ~ B is an epimorphism of Lie modules. The kernel V of 'IT" is called the vertical module and ele
ments of' it are called vertical. A reduction of' a Lie bundle 'IT": P ~ B
is a commutative diagrsm
Q __ t.;.._.,;. P
B
where t: Q ~ P is a monomorphism of Lie modules and a: Q ~ B is
a Lie bundle. A Lie bundle 7f: P ~ B is trivial if there is a redu~
tion with Q = B and a the identity. A connection in a Lie bundle
'IT": P __,.. B is a submodule H of' P such that P is the module direct
sum of' H and the vertical module V. The module H in a connection
is also called the horizontal module and elements of' it are called hori
zontal. The projections of P onto V and H are denoted v and h
respectively. The curvature of' a connection is the mapping R of H XH
into V given by
R(X,Y) = v[X,Y] ,
X,Y E H •
A connection is flat in case its curvature is _0. The holononzy module
L of' a connection is the smallest submodul.e of P such that
§6. HOIDNOMY
R(X, Y) € L ,
x, Y € H ,
and
(2 )
[X,Z] € L ,
X € H, Z € L •
Notice that we may identify the module p/R with V, so that
the definition of curvature agrees with the previous one. We may if we
wish extend R to a mapping of P XP into V by setting
R(X,Y) c v[hX,hY] , X,Y € P .
This has the ef.feot of setting R(X,Y) 0 if X or Y is vertical.
The horizontal module H in a connection of the Lie bundle
1T': P ~ B is isomorphic as a module to B, since it is complemente:t'y
to the kernel V of 1T'. However, if the connection is not flat then
H is not a Lie lll.Odule whereas B is. If X is in B the element Jc' of H such that 'lf1. = X is called the lift of X. We also define the curvature as a mapping of B X B into V by setting R(X, Y) = R(X, Y) •
Theorem 2.
Let 1T': P ~ B be a Lie lnmdle with vertical
module V. Suppose there is a connection with horizontal module H, and let L be the holonomy module.
Then L is contained in V, the lll.Odules L and H+ L are Lie
modules, L is an ideal in H+ L , and
H+ L _ _.;;_....;;> P
where t is the inclusion map and a is the restriction of 1T', ~
reduction of the Lie bundle 1T': P"""';' B. Also, H is a connection in
84.
§6. HOI£>NOMY
(1: H+L""":'> B end its holonomy module is L. A Lie bundle is trivial if and only if it has a flat connection.
~ BY definition of the curvature, V satisfies (l) end since V is an ideal (being the kernel of n ) it satisfies (2). But L is the smallest module satisfying (l) and (2), so LeV.
The holonomy module L is spanned as a module by the set of
aJ..l vector fieldS of the form
since each
is an Flinear combination of elements of the form (3).
Let Z be given by (3) and let
(4)
be of the same form. BY the Jacobi identity,
e12 R(Yl,yO)'Z] e12R(Yl'YO),eYmZ]
[w,Z] = ey [ey
m ml
Assume as an induction hypothesis on m that foraJ..l W of the form
(4) and all Z of the form (3) for any value of n, the Lie product [W,Z] is in L. The relation (5) shows that if m>2 the induction
hypothesis is true for m if it is true for ml
Consequently, to
show that [W,Z] is alw~s in L we need only show that
§6. HOLONOMY
is in L.. By the definition of R,
But h[Yl,YO] is horizontal, so that [h[Yl,YO]'Z] is in L. Therefore we need only show that
is in L. But by the Jacobi identity
and the right hand side is in L. Therefore L Ls a Lie module.
Since L is conta.ined in V, the moduJ.e H+L· is a direct sum
of modules. Let :JS.'~ € H and Zl'~ € L. Then
Since L is a Lie module, [Zl'~] is in L, and [Xl'~] + [~,~] is in L by the definition of L. The formu.la. (6) applied to [Xl'~] shows that it is in H+L. Therefore H+L is a Lie module, and the argument shows that L is an ideal in H+L (if we notice in addition
that Z·f = 0 for all Z in L, and scalars f since L is contained in the ideal V). Oonsequently, if we let a be the restriction of 11"
to H+L then a: H+L~ B is a Lie bundle and, together with the in
elusion ~: H+L ~ P , gives a reduction of 11": P ~ B. It is clear
that H is a connection in a: H+L ~ B and that its holonomy module
is again L.
In particular, if the connection is :f'la.t then L = 0 since 0 satisfies (1) and (2). Then H is a Lie module, a: H ~ B is an isomorphism of Lie modules, and ~ 0 al: B ~ P is a monomorphism of
86.
§6. HOLONOMY
Lie modules, So that 71": P ~ B is trivial. Conversely, if 71": P ~ B is trivial we may take as our connection H the image of B in P in
the trivialization, and H is a flat connection. QED.
3. The relation between the two notions of connection
On a manifold we have the notion of an affine connection in the sense of Koszul (§5.1) and the notion of a connection in the frame bundle, and the notions have a close affinity. There is a onetoone correspondence between GL(n,JR)invariant vector fields on the frame bundle and derivations (§3.2) of the module of vector fields, and this motivates the following construction.
Theorem 3. Let B be the Lie module of all derivations of F and let P be the module of all derivations of (F,B). For Cjl in P define mp in B to be the restriction of Cjl to F. With respect
to the operations Cjl·f = (mp)·f and [Cjll,Cjl2] = CjllCjl2  Cjl2Cjll ' the module P is a Lie module and 71": P ~ B is a Lie bundle. The vertical module V is the set of all derivations of (F,B) which are 0 on F and Flinear on B.
Let H be a connection in 71": P ~ B. Then the mapping of
B into H which takes X into its lift ~ is an affine connection (in the sense of Koszul). Conversely, if ~ is an affine connection the set H ~ ~X for X in B is a connection in 71": P ~ B , and ~X is the lift of X. The curvature R as a mapping B XB into V is identical with the curvature of the affine connection ~.
Proof. By Theorem 1, §2.2, B is indeed a Lie module. ~ §3·2 the set P of all derivations of (F,B) is an FO Lie algebra and it
§6. HOLONOMY
is readily seen to be a Lie module, and 7r: P ?: B is clearly a Lie bundle, with vertical module as described.
let H be a connection, and define "V'X to be the lift Jt' of X. Then X ~ "V'X is Flinear and each "V'X is a derivation of (F,B) so "V' is an affine connection. Conversely, if "V' is an affine con
nection the set H of all "V'X is a submodule of P, since X ~ "V'X is Flinear. We need to show that any derivation cp of (F,B) is the sum of some "V'X and an element of V, but if we let X = mp
then cp  "V'X is Clearly in V. Equally clearly, V n H = 0 , so H is a connection. By definition "V'X is horizontal and since 71Y'x = X "V'x is the lift of X. Thus we may identify the two notions of connection.
The curvature as a mapping of B XB into V is by definition
R(X,Y) = R(X,Y) = v[X,y] = [X,y]  h[X,y] •
Since 7r[X,y] = [nx,iY] = [X,Y] , the horizontal part of [X,Y] is the lift of [X,yJ. That is,
QED.
§7. Riemannian metrics
l. PseudoRiemannian metrics
ret E be an F module and let u be a covariant tensor of
rank 2. Ii' X is e.ny vector field we let u(X) be the Iform defined
by
<u(X),Y> = u(X,Y) •
The tensor u is called nondegenerate in case the mapping u: E ~ E'
is bijective. Ii' u is nondegenerate u: E :.> E' is an isomorphism
of F modules, so that E is reflexive. Therefore the dual u' is
also a mapping of E to E', and u is syllD'lletric if and only if u = U' •
The inverse me;pping of u: E ~ E' for u nondegenerate is denoted
1 1 1
u ,so that u : E' ~ E. The inverse me.pp:l.ng u determines the
contravariant tensor ul of rank 2 given by ul(w, n) = <ulw, Tj > •
Definition. A pseudoRiemannian metric on an F module E is
a symmetric nondegenerate covariant tensor of rank 2.
On a manifold M, a Riemannian metric g is a pseudoRiemannian metric such thB:t g(X,X)?: 0 for all vector fields X (and consequently g(X,X)(p) > 0 at all points p such that Xp I: 0) Riemannian metrics enjoy important geometrical and analytic properties not shared by pseudo
Riemannian metrics, but the formal algebraic properties, which we will
consider now, are the same.
If the F module E is free of finite type with basis
g is a pseudoRiemannian metric we denote its compo
nents by gij
1 i·
and those of g by g J , so that
90.
§ 7. RIEMANNIAN METRICS
and the gij are given explicitly in terms of the ~ by the familiar but extra.ordina.:t'ily complicated formula. for the inverse of a matrix.
2. The Riema.tmie:n connection
Theorem 1. Let g be a pseudoRiemannian metric on a Lie
module E
Then there is a unique torsionfree affine connection ~
such that ~g = 0 •
Proof. First we prove uniqueness. Let X, Y , and Z be vector
fields. Since ~g = 0 we have ~ zg = 0 , so that
Since ~ is torsionfree,
(1) Z·g(X,Y) = g(~~,y) + g([Z,X],Y) + g(xSzY) ,
(2) X·g(Y,Z) = g(~0'Z) + g([X,Y],Z) + g(y,~~) ,
(3) yog(Z,x) = g(~zy,x) + g([Y,z],x) + g(z'~0) ,
where (2) and (3) are obtained from (1) by cyclic pernru.tation. Now subtra.ct (2) from (1) plus (3). We obtain
(4)
2g(X,~ZY) = Z.g(X,Y) + Y·g(Z,X)  X·g(Y,Z) g([Z,X],Y)  g([Y,Z],X) + g([X,Y],Z) •
The right hand side of (4) does not involve ~, so we have a formula for g(~zy) on X. Since g is nondegenerate and X is arbitrary, ~ZY is uniquely determined, so ~ is unique.
To prove existence, consider the right hand side of (4). This
is Flinear in X, for if we replace X by fX then f is differen
tiated via
§7. RIEMANNIAN METRICS
91.
(Z'f)g(X,Y) + (yof)g(Z,X)  ~Zof)g(X,y)  (Y.f)g(X,Z) = 0 0
Therefore if we fix Y and Z the right hand side of (4) is a lform w and we may define 'V ZY to be ~ g l( w). Then 'V ZY is Flinear in Z , for if we replace Z by fZ then f is differentiated via
(Y.f)g(Z,X)  (Xof)g(y,Z) + (Xof)g(Z,y)  (Y·f)g(Z,X) = 0 0
Finally, if we replace Y by fY then f is differentiated via
(Z·f)g(X,Y)  (Xof)g(y,Z) + (Z.f)g(Y,X) + (X'f)g(y,Z) = 2(Z·f)g(X,y)
so that
Thus 'V is an affine connection.
let e(X, Y,Z) be the left hand side of (1) minus the right hand side of (1). We have seen that (4) is equivalent to
e(X,Y,Z)  e(Y,Z,X) + e(Z,X,y) = 0 •
Therefore (5) is true, and by cyclic permutation
(6)
e(Y,Z,X)  e(Z,X,Y) + e(X,Y,Z) = 0
By (5) and (6), e(X,y,Z) = o. That is, (1) holds. let T be the torsion of 'V 0 By (1),
We also have
(8) Zog(X,Y) = ('Vzg)(X,y) + g('VZX'Y) + g(x,'Vzy)
By (7) and (8)
('Vzg)(X,y) = g(T(X,Z),y) 0
92.
§7. RIEMANNIAN METRICS
We need to show that both sides of (9) are o. Let u(X,Y,Z) denote the common value. The u is symmetric in the first two variables (by the left hand side of (9» and antisymmetric :In the first and third variables (by the right hand side of (9». Therefore
u(X,Y,Z) = u(Z,Y,X) = u(y,Z,X) = u(X~Z,y) = u(Z,X,Y) = u(Y,X,Z) = u(x,Y,Z)
and u = o. Thus \7g = 0 and s:Ince g is nondegenerate, T = 0 • QED.
The connection \7 of Theorem 1 is ca.lled the Riemannian connec
tion or the Le~Oivita connection. It is due to Christoffel.
Theorem 2. Let g be a pseudoRiemannian metric on a coordinate
Lie module E with coordinates xl, ••• ,xn
Then the components of the
Riemannian connection \7 are given by
(10)
Proof. Apply (4) to X = o/axa, Z = o/axi, Y = o/axj Then
ka
J:f we multiply by g and sum over a we obtain (10). QED.
If E is merely free of finite type we obtain three additional
terms :Involving the structure scalars due to the fact that the vector
fields of the basis may not commute.
The discovery of the Christoffel symbols (10) and of covariant
differentiation was a marvellous discovery. Despite the fact that tensor
analysis has been studied so intensively that it has become easy it re
mains an amazing subject.
This action might not be possible to undo. Are you sure you want to continue?
We've moved you to where you read on your other device.
Get the full title to continue reading from where you left off, or restart the preview.