International Journal o f Theoretical Physics, Iiol. 30, No.

9, 1991

Holonomy and Path Structures in General Relativity

and Yang-Mills Theory
J. W . Barrett ~

Received November 12, 1990

This article is about a different representation of the geometry of the gravitational

field, one in which the paths of test bodies play a crucial role. The primary
concept is the geometry of the motion of a test body, and the relation between
different such possible motions. Space-time as a Lorentzian manifold is regarded
as a secondary construct, and it is shown how to construct it from the primary
data. Some technical problems remain. Yang-Mills fields are defined by their
holonomy in an analogous construction. I detail the development of this idea in
the literature, and give a new version o f the construction of a bundle and connec-
tion from holonorny data. The field equations of general relativity are discussed
briefly in this context.

PROLOGUE
The following text was written in 1985, forming the author's Ph.D.
thesis. In preparing it for publication I have edited and amended it in a
number of places, particularly updating the citations of work which has
appeared since 1985, or has been brought to my attention since then. But I
have tried to keep to my original intentions regarding the particular point
of view about fundamental physics which I tried to express in 1985.
The work in Section 2, on Yang-Mills fields (connections), has been
developed in innumerable ways over the years from Dirac onward, with
many of the developers working separately, it seems, particularly those in
the physics community. The introduction provides a guide to the literature
which I know about on this subject--but is surely not complete, and I expect
that more of this fragmented literature will come to light as the years pass.
The material in Section 2 provides the necessary synthesis for the two sec-
tions which follow.

~Department of Physics, The University, Newcastle upon Tyne, NE1 7RU England.
1171
0020-7748/91/0900-1171$06.50/0 ct) 1991 PlenumPublishingCorporation
1172 Barrett

In Sections 3 and 4 1 apply the same philosophy and method to gravi-

tational fields (Riemannian metrics), following the physical reasoning I out-
line in the Introduction. This departure appeared to be novel at the time,
and still does. There are some loose parallels with the physical motivation
of the twistor program in general relativity, as regards recasting the gravi-
tational field in terms of the incidence of lines. Still looser is any similarity
to string theories, where lines and loops wander on manifolds. If there is a
connection at any technical level, it remains to be made.
I paid attention to the intuitive physical ideas (and the Yang-Mills
analogy), sacrificing the ability to calculate and, probably, the mathematical
precision and comprehensiveness which would allow the theory to develop
further. However, from the point of view of theoretical physics, the next
input must be quantum theory, and in quantum theory points and lines on
manifolds as physical objects cannot be defended. Thus, I do not believe
that a tighter mathematical formulation of the ideas I present here will
become a part of fundamental physics in a direct way. Rather, my hope
is that the ideas will suggest a physical reformulation, involving perhaps
discreteness or quantum theory, or both.

Terms Assumed Throughout. Suppose B and C are two topological

spaces, and M a differentiable manifold with a singled-out basepoint .. Then
I use the following notation.

Map(B, C) denotes the set of continuous maps B ~ C

I the unit interval [0, 1]
TM the tangent bundle of M
TxM the tangent space at x
MI the unpointed path space: the set of piecewise smooth maps
I~M
PM the pointed path space: the subset of M *with p(O) = . , p ~PM;
note that, in places indicated, when E is a bundle over M, PE
means the space of paths with p(0) anywhere in the fiber
over *
~M the loop space: the subset of P M with p ( l ) = *
-1 for p ~ M x, p-1 is the reverse path: p - l ( i ) =p(1 - i)
P
o denotes the composition of paths: for P2, P~ ~ M x with Pl (1) =
p 2 ( 0 ) , P2 o p ! is the path

9 ~p,(2i) i_<1/2
l ~ p 2 ( 2 i - 1) i>1/2
Holonomy and Path Structures in GR and YM Theory 1173

the tilde convention: for any m a p ~: A ~ M a p ( I , B), the same

symbol ~, without the tilde represents the associated m a p
~; A x I~B

(u, i) --* ~t(u)[i]

Ill the e n d - o f - p r o o f symbol

1. I N T R O D U C T I O N

The idea o f "gravitational field" is more or less accepted as the basic

n o t i o n o f Einstein's general relativity. One imagines, in the absence o f
matter, a space-time manifold obeying the v a c u u m Einstein equations.
Observers, or test particles, m a y venture into parts o f this manifold to make
such measurements as they please, and, so long as their mass is negligibly
small, they m a y do so without disturbing the gravitational field. The field
has an existence as a physical quantity independent o f the test particles,
whose proper lengths the field purports to indicate.
The mathematics o f gravitation has changed very little since Einstein's
original f o u n d a t i o n o f general relativity--one might mention the later intro-
duction o f spin d e n s i t y - - b u t the conception o f the gravitational field has
altered significantly, if we can judge f r o m Einstein's b o o k on relativity theory
(Einstein, 1922) :
For the concept of space the following seems essential. We can form new bodies
by bringing bodies B, C,... up to body A; we say that we continue body A. We
can continue body A in such a way that it comes into contact with any other
body, X. The ensemble of all continuations of body A we can designate as the
"space of the body A". Then it is true that all bodies are in the "space of the
(arbitrarily chosen) body A". In this sense we cannot speak of space in the
abstract, but only of the "space belonging to a body A". The Earth's crust plays
such a dominant role in our daily life in judging the relative positions of bodies
that it has led to an abstract conception of space which cannot be defended. In
order to free ourselves from this fatal error we shall speak only of "bodies of
reference" or "space of reference". It was only through the theory of general
relativity that the refinement of these concepts became necessary, as we shall see
later... [Einstein (1922, p. 2)]

In this passage, Einstein is against the notion o f an abstract space or space-

time having an existence independent o f the measuring bodies. In other
places, it is true, he did support the analogy between the electromagnetic
field and the gravitational field, an analogy which p r o b a b l y did most to
establish the now-accepted notion o f the gravitational field. However, a bit
o f selective quotation serves to make this particular point.
1174 Barrett

Quantum gravity has very much taken its cue from the conventional
field point of view. One considers the space of all metrics on a given manifold
and tries to form a statistical, probabilistic, theory on this space, which
satisfies the conventional ideals of a quantum theory. In other words, the
attempted theory purports to deal only with the pure field, and does not
consider as part of its scope the measuring bodies which Einstein considered
to be basic to his conception of "space." We know now that this conception
of quantum gravity has all sorts of objectional features (which, of course,
may be due to other aspects of the theory) (see, e.g., Isham, 1981). One of
the unpleasant aspects is the necessity of considering the diffeomorphism
symmetry of the configuration space. This arises because quantum field
theory is based on the local vector space structure of the configuration
space (Isham, 1984), and so is based on the space of all metrics. However,
diffeomorphism-related metrics are considered to describe the same physical
object, and so one is forced to "factor through" the diffeomorphism sym-
metry. Now this symmetry is an essentially unphysical one, and arises
because the space of all metrics is really the wrong concept, it does not have
a one-to-one correspondence with real physical quantities. What one should
really work with are the set of distinct geometries. The diffeomorphism
symmetry is a symmetry of a representation, via a particular differentiable
manifold.
This article is about a different representation of the geometry of the
gravitational field, one in which the test bodies play a crucial role. It is an
analysis of what the relationships between the measured quantities of the
motion of the test bodies are. It describes the effect of the geometry of space-
time on the geometries of the test-particle motions. The key idea is suggested
by the quote of Einstein above: a point on a manifold is defined as the set
of all particles which arrive "there." In other words, we have to say what a
particle path is, and what it means for particle paths to be "in coincidence."
Particle paths shall be defined so that they all start at the same place.
A particle path is defined by its geometry--the specification of the angles
through which the particle bends, in which directions, and at what proper
distances along the path. In other words, the intrinsic geometry (proper
distances along the path) and extrinsic geometry (parallel transport of vec-
tors along the path) are specified.
The information in the gravitational field is involved in grouping to-
gether, in a particular way which will be described more fully below, all the
possible particle paths into sets which represent the particles whose paths
are "in coincidence." These sets form the points of the space-time manifold.
This is actually all that is required! What one finds is that the geometry of
the particle paths is sufficient to specify the geometry of the resulting space-
time. The result is a theory of gravity in which the test particles are firmly
Holonomy and Path Structures in GR and YM Theory 1175

mixed up with the phenomenon of the field itself. In fact, the field is "made
out of" the motions of the test bodies. There is no conception of an inde-
pendent gravitational field divorced from the theory of the propagation of
matter.
So the presence of gravity is felt by the fact that it alters the sets of
coincident particle motions. The coincidence is specified by an equivalence
relation on the set of all possible particle geometries. I have termed the
information that goes into this equivalence relation the "holonomy" of the
gravitational field, the word holonomy being using in rather a loose sense.
The point is that two particle paths end at the same point if they form a
closed loop on the space-time manifold, which starts and ends at their mutual
starting points. So to specify the equivalence relation one needs to know the
geometry of the particles which move in closed loops only, and in addition
the Lorentz group holonomy element (in the strict sense of the word holono
omy) of the closed loop. This is the way in which the gravitational "field"
is specified. The mathematical details of this are best left to the following
sections.
So, with the motivation the analysis of gravity, why is a large section
of the article about classical Yang-Mills fields? The reason is that Yang-
Mills theory is very similar to general relativity, but, as is usually the case,
it is a simpler theory. The Yang-Mills result (the representation theorem of
Section 2) both provides a comparison with general relativity, and also a
technical result which is of use in building the linear connection of the
gravitational field. Moreover, the technical tools used on the way in Section
2 bear a strong resemblance to, and in fact motivate, the techniques of
Section 3.
Table I contains a comparison of the main features of Yang-Mills
theory in Section 2 and gravity in Section 3; the notation has been deliber-
ately chosen to enhance the similarity.
The strong resemblances displayed in the comparison throw fresh light
on the debate about the status of gravity as a gauge theory. From the point
of view of holonomy, gravity has no diffeomorphism symmetry: one just
specifies the set P and the map h. The "gauge freedom" of gravity is still
present, however, in a rather subtle form. The set P (see Section 3) contains
the geometries of the paths which form closed loops. Suppose, for the sake
of argument, that the paths are composed of n linear sections (straight lines).
Then, to specify an element of P (and hence give information about the
gravitational field) one can arbitrarily specify the first n - 1 pieces. Then the
geometry of the last section of the path is entirely fixed by the requirement
that it be in P. We can say that the gauge freedom of gravity is the freedom
to specify the first n - 1 sections at will, and the real information (the "holon-
omy") is the fixed value of the geometry of the last section. This notion of
1176 Barrett

Table I
Construct YM Gr
G Structure group Lorentz group
M Base space Tangent space
Geometry of loops f~M P
Holonomy mapping H: ~ M ~ G h: P - ~ G
Equivalence p(1) =p'(1), There exists 7r~P:
relation on P M x G, g' = H ( p - ~ o p')g zr= ( H ( ~r)p- ~) o p',
(p, g) ~(p', g') g' = h(zc)g
Constructed set P M x G / ~ Bundle Manifold and frame bundle
Lifting Lifting function l. Inverse development
map A, lifting in
frame bundle
whose smoothness gives Chartson the bundle Charts on the manifold
and which d e f i n e s Connection Metric and connection

gauge freedom is of a different character from the rather empty notion

of "diffeomorphism symmetry." It clearly relates to choices made by the
experimenter about the information which is to be gathered. Thus, it is a
symmetry with an explicit physical significance. This interpretation of the
"gauge freedom" of gravity is only possible if one has in mind the physical
identification of the paths of the holonomy description (which might be
regarded as mere mathematical artifacts) with particle path geometries.
The last section is concerned with the field equations of gravity, which
have so far not been introduced. The equations are presented in a form which
relates the (infinitesimal) holonomy of the field to the matter momentum and
angular momentum. The details are best left to that section. It suffices to
remark here that by taking a very specific point of view on the construction
of the gravitational field, as is described in this paper, one gets a very specific
point of view about the field equations: the metric-compatible aspect of the
connection is a matter of definition, but the torsion-flee aspect is a field
equation, with the same status as the Einstein field equation. In fact, the
torsion equation is naturally paired with the Einstein equation: it is part of
the "angular m o m e n t u m " field equation, the Einstein equation being the
"linear m o m e n t u m " field equation.

2. T H E H O L O N O M Y R E P R E S E N T A T I O N OF GAUGE FIELDS

2.1. Introduction

Classical Yang-Mills fields are usually described by potentials, or con-

nections on a principal fiber bundle. A physical field configuration is really
a set of potentials related by gauge transformations, so that in other words
Holonomy and Path Structures in GR and YM Theory 1177

the configuration space is the space of orbits of the action of the gauge group
f9 in the space d of all connections on some given bundle. This orbit space
has been much studied" because one can view Yang-Mills quantum field
theory as a theory of integration on the orbit space, albeit in an imprecise
fashion (Atiyah, 1980; Singer, 1981; Babelon and Viallet, 1981). There is
also a canonical version, where geodesics on orbit space correspond to the
Hamiltonian flow (see also Narasimhan and Ramadas, 1979).
For good physical reasons, then, one would like to have more infor-
mation about the configuration space d/fg. To achieve a different charac-
terization of the configuration space, it is possible to consider the holonomy
mapping of the connection as the fundamental object. This mapping takes
the loop space f~M of the base manifold M into the structure group G of
the bundle by mapping a loop into its holonomy element. The loop space is
the set of all piecewise smooth paths in M which start and end at an arbi-
trarily singled-out point, d e n o t e d . , in M, with a topology which will be
discussed later. The holonomy element of a loop is defined in terms of the
horizontal lifting of the loop into the total space of the bundle. The two
endpoints of the lifting define a G-element which translates one of the end-
points to the other, using the G-action on the bundle. This is the holonomy
element. Section 2.2 describes how a connection gives rise to the concepts
of horizontal lifting and holonomy, and why the machinery of bundles is
the most appropriate.
Discussion of the holonomy in electromagnetism goes back to Dirac
(1931) in the context of magnetic monopoles. Aharonov and Bohm (1959)
discussed parallel transport as a phase shift of a Schr6dinger wavefunction,
and recognized the physical significance of the holonomy operator for a
closed loop, being more directly involved in the behavior of the Schr6dinger
wavefunction than magnetic forces. Some of these ideas were used in an
attempt to find gauge-invariant quantizations of electromagnetic and other
gauge fields (Mandelstam 1962a,b, 1968a,b; Bialynicki-Birula, 1963).
There are two important facts about the holonomy mapping. First,
different physical configurations give rise to different holonomy mappings,
and second, the set of restricted gauge transformations fg. c fg, that is, ones
which act trivially at the basepoint, leave the holonomy mapping invariant.
This means that, if we consider gauge equivalence to be restricted in this
way, the physical configurations are faithfully and uniquely represented by
their holonomy mappings. The residual gauge freedom of field rotations at
the basepoint is parametrized by G ~ fg/fg., a finite-dimensional Lie group,
in contrast to the original infinite-dimensional gauge group f#. Thus, the
restricted, or pointed, configuration space d/fg. is equivalent to a set of
certain types of mappings OM~G, namely those which arise as holonomy
mappings.
1178 Barrett

The results demonstrated in this section show that there is a simple

set of conditions H1-H3, given below, defining the relevant subset
j g c Map(f~M, G) and a simple reconstruction of both the principal bundle
over M and the connection on it.
This works equally well for any principal G-bundle over M, so that it
is more natural to consider as the configuration space the disjoint union over
inequivalent bundles of the orbit spaces for each bundle. This can be much
more elegantly stated as the set ~ , of triples (B, F, b), where B is any princi-
pal G-bundle, F a connection on it, and b a point in the fiber over the
b a s e p o i n t . , with equivalence

(g, r , b) ~ (B', r', b')

if there is a bundle isomorphism B~B' taking F to F' and b to b', and such
that it is the identity on M. This definition is more useful here, since the
holonomy mapping involves only the base manifold M in its definition, the
reconstruction process manufactures "new" bundles as well as connections.
Note that since a preferred point in the fiber over 9 is preserved, these
triples are equivalent to orbits of the restricted gauge group in the space of
potentials. To be precise, if C is a particular bundle with basepoint c, and
~ - c ~ ~- is defined as the subset of triples (B, F, b) where B is isomorphic
to C, and d and f#, are the potentials and restricted gauge group of (C, c),
then ~ .c is in bijective correspondence with d / f # . , in the obvious way.
The two main results are as follows.

Reconstruction Theorem. Suppose M is a connected manifold with base-

point ,, and H: f~M~G satisfies conditions H I - H 3 below; then there exists
a differentiable principal fiber bundle B=(E, zr, M, G), a point beTr-l(*)
and a connection F on B such that H is the holonomy mapping of (B, F, b).

Representation Theorem. If M is a connected, Hausdorff manifold, the

correspondence of the reconstruction theorem between the Yang-Mills con-
figuration space ~ , defined as the set of triples (B, F, b) as above, and the
set of holonomy maps ~r G), defined by the conditions H1-
H3, is a bijection.

Results of this nature can be proved in a variety of settings, for example,

for differentiable or topological bundles. The infinitesimal connection (as
described above) might be replaced by some more general gadget, e.g., a
lifting function. In the literature, generalizations of infinitesimal connections
are rather vaguely called "connections."
The first mention of a result such as the reconstruction theorem occurs
in Kobayashi (1954) in a short note without proofs. The setting is rather
Holonomy and Path Structures in GR and YM Theory 1179

similar to the one used here, differentiable bundles and infinitesimal connec-
tions, but Kobayashi's axioms do not seem to include any mention of differ-
entiability (see H3 below). The reconstruction theorem appears in a more
general guise in papers on bundle topology which appeared shortly after-
ward. Milnor (1956) considered locally trivial bundles ("fiber" bundles) over
a simplicial complex, with a bundle slicing function playing the role of a
connection. Milnor's analogue of H2 is particularly elegant. The construc-
tions are all topological, rather than differentiable.
The most general setting is that of Lashof (1956), who considered topo-
logical principal bundles (not even necessarily locally trivial) with a general
lifting function playing the role of the connection. The lifting function is
much more general than an infinitesimal connection because the lifting may
not commute with the structure group action. This is also connected with
the fact that in this formulation, paths which differ only by reparametriza-
tions and other similar operations (cf. H2) are not considered to be equiva-
lent. Lashof has a notion of equivalence of maps ~ M ~ G which renders the
equivalence classes in a bijective correspondence with the set of inequivalent
G-bundles over M. This is thus a coarse version of the representation
theorem, which does not distinguish between different connections on the
same bundle. Interestingly, Lashof's constructions are rather close to the
gravitational constructions presented here.
The differential aspect of the subject was explored by Teleman
(1960, 1963) shortly afterward. He later wrote two papers which are
probably the most comprehensive overview of all of these sorts of results.
In the first (Teleman, 1969a), the reconstruction theorem appears (Theo-
rem 3) in a general guise, a connection being a lifting function which
satisfies analogues of H1 and H2. Various special cases (simplicial, differ-
ential, complex analytic) are discussed in the second paper (Teleman,
1969b).
The constructions were later rediscovered by physicists (Giles, 1981;
Anandan, 1983; Barrett, 1985, 1989; Fischer, 1986). For the most part, they
confined themselves to the algebraic part of the constructions rather than
the topological aspect. Chan and Tsou (1986) and Chan et al. (1986) found
an interesting extension of the reconstruction to presenting the data in the
form of a connection on loop space f~M, with applications to monopoles._
The last paper also contains references to many other associated papers in
the physics literature. It is interesting to note that Dirac's (1931) paper on
monopoles discussed a homomorphism f~M~ U1, with M three-dimensional
Euclidean space minus a point, which by the reconstruction theorem immedi-
ately translates into a connection on a bundle, the modern understanding of
monopoles. The inequivalent bundles correspond to the different monopole
charges which Dirac found.
1180 Barrett

Many physicists were interested in the character of the homomorphism

~ M ~ G in the case that G acts on a vector space, which goes under
the name "Wilson loop" in the physics literature (Wilson, 1974). See,
e.g., Durhuus (1980) for a lattice version of a result which goes back to
Teleman (1969a, Theorem 5) at least. Polyakov (1979) discusses some of
the uses of Wilson loops in quantum field theory; it is a subject that was
taken up by many other authors.

2.1.1. Conditions H1-H3

The first condition on a holonomy map H is straightforward:

H1. H is a homomorphism of the composition law of loops

n(co2 o o)1) = n ( c o l ) H ( ~ )

H1 is consistent with the holonomy element being defined as the G-

element which translates the point b to its image point under parallel trans-
port backward around the loop. This will be taken as the definition of the
holonomy element. To be more explicit, if co' is a horizontal lift of co with
co'(1) = b, then bH(co) = (0'(0) (Kobayashi and Nomizu, 1963).
The second condition axiomatizes several related features of H. H takes
the same value on loops which differ by a reparametrization, or by the
addition or removal of path sections which "double back" on themselves.
This is formalized by the definition of thin loops: 0 is a thin loop if there
exists a homotopy of 0 to the trivial loop, with the image of the homotopy
lying entirely within the image of 0. This makes precise the notion that a
thin loop in M does not enclose any area of M.

H2. H takes the same value on thinly equivalent loops: col ~ a h if

o91 o coy1 is thin.
The path co-1 is the reverse path of co: co-l(i)= c0(1- i).
This property, together with HI, implies that H(co-1) = H-l(co). In fact
we can consider the whole H-group structure of f~M (Spanier, 1966), and
since the homotopies which make loop composition homotopy-associative
and homotopy-invertible are actually thin homotopies describing thin
equivalences, ~qM factored by thin equivalence is a group, and H defines a
homomorphism of groups. This is explained further in Section 2.3. We can
think of thin equivalence as a restricted notion of homotopy equivalence,
intuitively similar to homotopy theory on a very fine mesh or sieve. The
holes of the mesh stop the homotopies sweeping out across areas of M.
Holonomy and Path Structures in GR and YM Theory 1181

H1 and H2 are all the algebraic properties required of H. It remains to

formulate a notion of smoothness for the holonomy mapping for axiom H3.
One should not arbitrarily impose conditions of continuity and differ-
entiability for purely technical reasons, but in an ideal world, these things
should relate to questions of physical importance, about the way in which
the measurements that the field represents are made. This point was made
to me by Chris Isham. In Yang-Mills theory this is a little difficult to inter-
pret, since except for the case of electromagnetism, one does not measure
the classical field directly, it just appears as a construct in the quantum field
theory. Nevertheless, the question is certainly important for classical gravity
and electromagnetism.
Any classical field configuration represents an infinite amount of infor-
mation, and since one can only ever measure a finite set of values, it is an
idealization of the true situation. Any continuum of quantities represents a
sequence (in the mathematical sense of a countable set in an order labeled
by the integers) of measurements performed to greater accuracy as the
number of measurements increases. Perhaps it would be more accurate to
say that the field configuration represents a measurement algorithm, rather
than any set of measurements themselves. Now in any theory envisaging an
infinite set of possible outcomes this set should at least be a topological
space, so that one can meaningfully discuss the convergence of a set of
approximations representing the process of measurement. Clearly, the topol-
ogy involved should relate closely to the way in which the theory envisages
the measurements being made. Conventionally, when discussing theories
involving an infinite amount of information, such as a classical field theory,
there seem to be two distinct possible "sources" of the infinity. On one hand,
there may be an infinite amount of information contained in a given field
configuration, and on the other, there may be an infinite range of possibilities
for that field configuration. So for a field theory, say involving real-valued
functions on space-time, one needs both a topology on space-time to define
a field as a continuous function, and a topology on the configuration space
to determine which field configurations neighbor each other. Clearly, when
one relates these topologies to the measurement process, one can see that
they should be intimately related to each other. Similar sorts of remarks
apply to the smooth structures of these spaces.
The most immediate problem is to find the structure on f~M which
makes holonomy mappings continuous, at least, and hopefully differentiable.
Since much of the theory has a strong algebraic topological flavor to it, the
first thing to try, and in fact discard, is the compact-open topology. The
problem with this is simple: holonomy mappings are not in general continu-
ous with this topology. A simple example will serve to demonstrate this. Let
space-time be R 2 and consider an electromagnetic field on it with constant
1182 Barre~

field strength F. For any closed loop the holonomy element is

H(co) = exp f v iF dx I A dx 2

where V is the volume bounded by co. Consider the family of loops

~t: I ~ f ~ R 2 given by

~f(s)t(1 - t) exp(it/s), sr 0
Vt(s, t)
(0, s=0

where R 2 has been identified with the complex numbers. The parameter t is
the time along the path, and s is the family parameter. If f is a continuous
real function with f ( 0 ) = 0, then ~t is continuous, and so fit is continuous in
the compact-open topology. The family describes a continuous set of loops
which, as s ~ 0 , simultaneously wind around faster and faster, and shrink to
zero radius. The area integral is easily computed, putting r(t)=f(s)t(1- t)
and O(t)= t/s, then the area enclosed by loop s is

f If2(s)/6Os, s~ 0
It2 dO = ~0, s=0

Clearly, the choice f ( s ) = s ~/2 renders the holonomy Hqt: I ~ U1 a discontinu-

ous function. The consequence of this is that although many of the construc-
tions to be used further on are very strongly linked with algebraic topology,
the standard topology used in algebraic topology is not relevant. On a
technical level, the reason is that essentially the parallel transport operator
is the solution of an ordinary differential equation, and so it is important
that the topology controls the derivatives of the paths. In the above example,
the derivative of ~t is not continuous at s = 0. On a physical level, this failure
stems from the fact that there were no good physical reasons for supposing
that the compact-open topology is relevant.
To return to the physical motivation, what is needed is a topology on
f~M which relates more closely to the physics of the measurements. The
physics is based on entirely classical (nonquantum) ideas. Suppose we just
limit attention to the subspace of loops which are piecewise linear, i.e.,
composed of a finite number of straight-line segments. Then if space-time is
again R 4, and we consider the loops with n + 1 sections, these can be parame-
trized b the positions of the "corners," i.e., (~4)n. Thus, measuring the shape
of the loop is reduced to measuring a finite number of particle positions,
Holonomy and Path Structures in GR and YM Theory 1183

and this is exactly where the topology and differentiable structure of space-
time come into use. 2 Thus, we want to postulate that the holonomy is a
smooth map (~4)n~G. So, to consider the space o f piecewise linear loops,
the smoothness requirement is:
For any integer n > 0, if V: (~4) n--*~2M is the family of piecewise linear
loops with n + 1 linear pieces, then the composite map HVt: ~ 4 ~ M ~ G is
a smooth map.
It would be quite possible, and physically well motivated, to work with
the space of piecewise linear loops. However, the linear structure of the base
space (here R 4) is not really relevant, and we want the results to hold equally
well on any differentiable manifold. It is also technically rather inconvenient.
So the remedy is to allow the above situation to relax under arbitrary
diffeomorphisms of the base space.
The final form of the axiom H3 encapsulates the notion that a smooth
finite-dimensional family of loops has a smoothly-varying holonomy image
in G. A smooth finite-dimensional family of loops is a map ~: U~f~M with
U an open subset of ~ for any n, which is smooth in the sense that the
associated map
~t: Ux I ~ M
(u, i) ~ ~(u)[i]
is continuous, and smooth (C ~) on the subintervals
U• for io=O<il<...ik<l=i~+l, n=0,1 ..... k
H3. For any smooth finite-dimensional family of loops ~t: U~f~M, the
composite map H ~ : U ~ f ~ M ~ G is smooth.
The collection of all such maps ~t from any such U into f~M defines the
induced topology on f~M. This is the finest topology which makes all these
maps continuous. Any map which obeys the axiom H3 is continuous
(Spanier, 1966; Dugundji, 1966).

2.1.2. The Representation Map

If H is the holonomy mapping of (B, F, b ) ~ , , then it has to be shown
that properties H 1 - H 3 hold. H1 is straightforward. To prove H2, we have
to show first that H(O)= identity when 0 is a thin loop.

21 amindebted to Dr. R. W. Tucker for bringing to my attention the fact that if parts of the
loop consideredare spacelike,then it does not have the direct interpretation of being composed
of particle paths. One just has to consider it as a figure in space-time, perhaps as marked out
by, and measured with, the aid of auxiliary particles.
1184 Barrett

fibers of bundle

(o, 1) ~o
I !
I
Fig. 1

Let/~: I--,~M be the smooth homotopy of 0 which makes it a thin loop,

so that 0= if(l), ~(0) is the trivial loop, and the map h: I x I ~ M has Im(h) =
Im(0) (using the tilde convention).
We can study the lifting in the pullback bundle Bh = (Eh, Jrh, I x / , G)
of B by the homotopy h. Since the image of h is one-dimensional, the curvat-
ure form in Bh is zero, and Bh has the canonical flat connection. It follows
that the path p: i~(1, i) in I x I can be deformed to the path

~(1-3i, 0), 0<i_<1/3

q: i ~ ](0, 3 i - 1), 1/3<i~2/3
( ( 3 i - 2 , 1), 2/3<i_< 1

without altering the endpoints of the lift in Bh. The map h carries p to the
path 0 in M and q to the trivial path in M (Figure 1). Since the canonical
homomorphism B h ~ B carries horizontal lifts into horizontal lifts, it follows
that there is a horizontal lift of 0 in E which has the same endpoints as the
trivial lift of the trivial loop in M, and so the endpoints of the 0 lift'are the
same point. Hence the holonomy of 0 is the identity element of G. It should
be noted that the homotopy h may be smooth only on subintervals
I x [in, in+ 1], but that the conclusion is unaffected.
If r then H(c01 o c0~-l)=id, and since H(c0zl)=H(co2) -1, it
follows, using H1, that H(o,~)= H(c0~). Property H3 follows from a slight
modification of the proof in Kobayashi and Nomizu (1963, p. 74).
Since these conditions are satisfied, the representation map ~ : ~,~,~/f,
which takes (B, F, b) to its holonomy map, is defined.

2.1.3. Bundle Construction

In this section it is supposed that H~oUf, i.e., it is a map ~ M ~ G
satisfying H1-H3. Using H, we can construct the total space E of a principal
Holonomy and Path Structures in GR and YM Theory 1185

bundle B as the set

E = P M • G/R
where P M is the path space of M, piecewise smooth paths with p(0) = ,, and
R is an equivalence relation quotienting P M x G:
(p,g)~(p',g') if p(1)=p'(1) and g'=H(p-lop')g
The loop p - l o p , is formed by composing p' with p parametrized in the
inverse direction. By virtue of properties H1 and H2, R is an equiv-
alence relation. In the following the equivalence class of (p, g) will be
denoted {p, g}.
It is interesting to note that this construction is analogous to the associ-
ated bundle construction, where ~M--+PM~ M is analogous to the principal
bundle with fiber ~ M and H defines the action of ~ M on G. This analogy
is made precise by turning ~ M quotiented by thin equivalence into a topo-
logical group (Teleman, 1960; Milnor, 1956).
The projection map and G-action are
Jr: E--+M
{p, g} p(1)
p: Ex G~G
{p, g}h- {p, gh}
The preferred point in the fiber over *~M is
b = {t, id}, where t is the trivial path
The lifting function is also implicit in the construction
l,: PMxG~PE

(P, g)-~q
with PE the space of paths in E starting over ,, and q is the path
q: I~E
i--+{K(p, i), g}
the contraction K(p, i) representing the section of path p from 0 to i:
K: P M x I-+PM
K(p, i)[j] =p(ij)
Now having a preferred point b in the fiber z - l ( , ) provides an iso-
morphism between this fiber and the group G. So the lifting function l, gives
1186 Barrett

for each path p in the base starting at * and a point g in the fiber over 9 a
path in E starting at g which is carried back onto p by the projection zr. For a
differentiable bundle with a connection a lifting function is given by assigning
l , ( p , g) the unique horizontal curve starting at g which covers p. The lifting
described here may, a priori, not be of this type. But with the aid of axioms
H I - H 3 we will show that it is.
As yet E has no differentiable structure. Suppose U is a contractible
open set o f M. Then there exists a smooth family of paths gt: U--*PM such
that gt(u) ends at u. Using the endpoint map e: P M ~ M , e ( p ) = p ( 1 ) , this
condition can be restated as egt= id.
The chart Cv for Jr- 1(U) c E is

Cv: Ux G (ui,,,id)' PMxG

a
...... > PM•

where a is the canonical projection. C~, is a bijection as a map

U x G ~ r c - ~ ( U ) . It is interesting to note, and will be useful later, that a =
e'l,, where e': P E - - , E is the endpoint map of PE, so that the chart Cv is the
map
(~,id) l* e'
C~: U xG ' PMxG ' PE ) E
Hence we can say that the lifting function is the relevant structure of E
which is used to define the charts.
L e m m a 1. The charts have smooth transition functions.
P r o o f Suppose there is a second chart C~, with gt': V ~ P M ; then,
putting W = Uc~ V and defining )?: W ~ M
w ~ g,- ~( W ) o g,(w)

C'(w, g) = { O'(w), g} = { O(w), H ( 2 ( w ) )g} = C(w, H ( 2 ( w ) )g)

So C-~C' is the map
(diag,id) (id,H;?,id) (id,compose)
WxG , WxWxG , WxGxG ' WxG
which is smooth by property H3. 9
Since C v commutes with the G-action on E, we have the following
result.
rc
Proposition. E ~ M is a principal fiber bundle with group G. []
In a similar way, it is possible to show that the lifting function l, is
smooth, in the sense that a smooth family of paths in M lifts to a smooth
family of paths in E.
Hoionomy and Path Structures in GR and YM Theory 1187

For the present, we shall assume the proof of Lemma 3, which states
that l, is the horizontal lifting of a connection on B. It is clear that the
connection is uniquely specified by the lifting function. Thus, the construc-
tion of a triple (B, F, b) is complete. This defines the construction map
~: ~ - o ~ , .
Proof o f the Reconstruction Theorem. To complete this proof, it remains
to show that ~cg= id, or, in other words, that if (B, F, b) is constructed
from H ~ and has holonomy mapping H', then H = H ' . This is
straightforward. []
Proof o f the Representation Theorem. The reconstruction theorem
shows that ~cg=id, so it remains to show that <g~=id. Suppose that
~(B, F, b) = H and C~(H) = (B, F, b), b = (e, Jr, m, g), a n d / ~ = (E, ~, M, G),
and that l, is the lifting function of (B, F, b). Consider the map
[* e'
PM x G ~ PE ' E
This factors through the relation Rn on P M • G to give a map ~b:/7~E. The
map ~b is a bijection and commutes with the G-actions on E and E.
Lemma 2. c~ is a diffeomorphism.
Proof To show that ~bis a smooth bijection, it is enough to show that
the charts Cv defined for/~ map smoothly to E. The map ~bCv is
(cP,id) l, e'
UxG ~ PMxG" ' PE ' E
which is smooth due to the fact that a smooth family of paths lifts to a
smooth family. Since ~b is the identity on M and an isomorphism of the
fibers, it is clear that ~bCV,is a chart for E, and so q~is a diffeomorphism. []
It is also easy to check that the induced mapping of paths takes the
lifting function of (B, F, b) into the lifting function of (B, F, b). Hence 4) is
a bundle isomorphism preserving M, F, and b.

2.1.4. The Reconstructed Lifting Function Defines a Connection

Now the missing lemma in the proof of the reconstruction theorem will
be dealt with.
Lemma 3. Let B = (E, It, M, G) with lifting function l, be constructed
as in Section 2.1.3. Suppose q is the lift of any path p ~ P M , and q passes
through point c~E, q(i)= c. Then (dq/di)(i) depends only on (dp/di)(i). In
fact there is a linear injection Fc: T,(c)M~ TeE such that Fc(dp/di)= dq/di.
The images of F,. for e ~ E define a smooth distribution (Kobayashi and
Nomizu, 1963) on E.
1188 Barrett

Since the lifting property of l, guarantees that ~,Fc = id, it follows from
Lemma 3 that F defines a connection.
Proof of Lemma 3. The first point is that the lifting defined by l, is
local in the sense that the lift of a section J c I of a path p c P M depends
only on the section pllJ o f p . More precisely, i f p ' c P M agrees with p on J,
then on J the lifts o f p and p' are related by right translation with a fixed G-
element. So the tangent vector of a lift at a point c c E can depend only on
the classes of paths in M which agree in some open neighborhood of ~(c).
For example, it might depend on any of the derivatives of the path p at the
point 7c(c). It shall be shown, however, through Lemmas 4-8, that the three
conditions H1-H3 in combination are sufficiently restrictive for the tangent
to the lift to be determined only by a linear mapping of the first derivative
o f p at r
Lemma 4 eliminates a possible pathology of the holonomy mapping
(see also Section 2.5).
Lemma 4. Suppose ~t: I ~ f ~ M is a smooth one-parameter family of
loops with r t, the trivial loop; then (d/di)(H~t)(O)= O.
Proof Since only the behavior of ~t near the p o i n t , c M is important,
we can assume without loss of generality that M is the m-dimensional vector
space Nm with * as the origin. The strategy is to embed t~ in an m-dimensional
family of loops &: Im"--~'~ m. This is defined by its associated map co
co: I m x I--* ~m

(SI, $2 . . . . . Sm, t)~(Vh(&, t), Vt2(s2, t) . . . . . IVrn(Sm, t))

In this formula ~tn is the nth coordinate of the function ~/, t denotes the time
along the paths, and the s's are the shrinking parameters. The original family
~t is the diagonal of the parameter space Im: ~t=~A with A : I ~ I m,
s ~ ( s , s , . . . , s). Starting from a point d on the diagonal of I m and moving
toward a face of I m by keeping all coordinates of I " except the nth fixed,
we see that the function cb smoothly collapses the nth coordinate of the loop
so that on the face & - 0 of I m the loop has no variation in its nth coordinate.
It is the projection onto the plane x, = 0 of Nm of the original loop cb(d)
(Figure 2).
Now 6~ is a smooth family of loops and so Hch is a smooth mapping.
Consequently,

(Hv,) (H A)
n GSn

Evaluating these derivatives at i = 0, each term on the right is just a derivative

along a coordinate axis ( 0 , 0 , . . . , s . . . . . . 0) of Im. But the loop
Holonomyand Path Structures in GR and YM Theory 1189

~'"I ~ . (1,1)

M
M IxI ) ~(O,s)
Y

Fig.2
t ~ ( 0 , 0 , . . . , N(s,, t) . . . . ,0) is a thin loop, and the holonomy element is
the identity for all values ors,. Hence ff/Ss,)(H(o)(O)= 0. Then the conclu-
sion of Lemma 4 follows. 9
Returning to the proof of Lemma 3, the local property of the lifting
means that it is possible to define the lifting of all paths in M, not only the
ones which start at .. This lifting has the product property: If q lifts p, q'
lifts p', and q(1) =q'(0), then q'o q lifts p' op. It also enjoys the obvious G-
invariance and reparametrization invariances. There is a smoothness prop-
erty which is discussed more below, and a thin equivalence property: If q
lifts p, q' lifts p', p is thinly equivalent to p' [p- 1 op, is a thin loop based at
p(0)], and q(0)=q'(0), then q(1)=q'(1).
Instead of considering directly the lifting of tangent vectors, it is useful
to consider, as a generalization, the lifting of one-parameter families of paths
~: I ~ M I [MIc Map(/, M ) denoting the unpointed piecewise smooth path
space] which shrink at parameter zero to the constant path at some point
in M (Figure 3). Such a family is specified by the smooth map u/: I x I ~ M ,
with N0, t)= x. The first factor will be denoted s, the shrinking parameter,
the second, t, is the time along the path. The quantity
v'-v ~ 1) 8~,(0, O)
Os Os
generalizes the tangent vector of a path in M. For ifp: I--*M is a path, then
putting ~s, t) =p(st), it follows that v ~= (dp/di)(O) and v~= 0. The vectors
v~ v ~ will be called the boundary vectors of gt.
Now suppose that ~: I x I ~ E lifts ~t, i.e., for fixed s, the path ~. (s, t) is
a lift of ~t(s, t). The function s~;~ (s, 0) determining the starting point of the
1190 Barrett

t
a/~ 1)

lxI (0 0) M -

generalized
tangent
vector
Fig. 3

lifts is arbitrary. If this is chosen to be smooth, then ,~ is also smooth.

Consider the boundary vectors of ;L,

wl _ ~,~ (0, 1) w0.__ ~ (0, 0)

Os ' 0s

Since t ~ Vt(0, t) is the trivial path, putting c = A (0, t), both w ~ and w~ lie
in T~E.
The first part of Lemma 3 is a consequence of the following result.

L e m m a 5. There exists a linear injection Fc: T , ~ c ) M ~ T~E such that for

any functions ~t and A defined as above

W 1 -- W0 ~--.rC(~)l -- /.)0)

Proof The proof spans Lemmas 6-8.

L e m m a 6. Suppose we have a number of smooth one-parameter famil-

ies of paths @k in M, k = 1, 2 . . . . , n, shrinking to the point *, with lifts '~k
shrinking to the point b e ~r-l(,), defined as above, and such that

Vtk(s, 1)=Vk+l(s, 0) and Vt,(s, 1)=Vtl(S, 0) = *

so that the composition 7: i ~ n ( i ) ~ ~-1(i) o " " " @~(i) is a map I ~ M

with O--*t; then ~ = 1 w ~ - w ~
Holonomy and Path Structures in GR and YM Theory 1191

Proof Define g~: I ~ G to be the maps such that

2.~+1(s, O)=A.k(S, 1)g~(s), k = 0 , 1. . . . . n

putting ~(s, t ) = Z . + l ( s , t ) = b (Figure 4). Then gk(0)=id, and H(7(s)) =
go(s)gl(s)g2(s) 9 9 9 g.(s). Lemma 4 then gives

k=o
ag~(~
ds

Using the Leibnitz rule on the defining equation of the g's gives

w~ = w~ + b dgk(O___))
ds

so that ~"k = O w ~ and the result follows since Wo~= w,+l

0 =0. 9

L e m m a 7. Suppose we have two smooth one-parameter~families of

paths ~ and ~' in M shrinking to the point *, with lifts 2 and 2' shrinking
to the point b s r c - l ( , ) , and the boundary vectors of ~t, p,' satisfy v'1 = v 1,
v'~ v~ then it follows that the boundary vectors of Z,/l.' satisfy w' 1 - w'~
W 1 - - W o.

P r o o f Families of paths f/, t~, J~: I ~ M 1 will be defined in such a way

that the composition
r(i) = f / - l ( i ) o ~(i) o ~-1(i)o ~(i)o ~t'(i) o fT(i)

is a map I--*f~M. The fl and a will be defined so that w ~ - w ~ and

w u1 - w a0 = 0. Then Lemma 6 will give the desired result since the contribu-
tions to ~ w I - w~ will cancel from ~/and f/- 1.
Since only the behavior of the functions in a neighborhood of 9 is
required, it can be assumed without loss of generality that M = ~m, with *

g.

bo

@~r ~176

Fig. 4. Pictureat a fixedvalue of s.

1192 Barrett

as the origin, The functions are

rl(s, t) = t~t'(s, O)
a(s, t ) = ( 1 - t ) ~ t ' ( s , 1) + t~t(s, 1)
fl(s, t ) = ( 1 - t)gt(s, O) + tgt'(s, O)
Since the values and the derivatives of the two curves s ~ V t ' ( s , 1) and
s-.gt(s, 1) are equal at s--O, the function
A: Ix I~R m

I , t ,
V(s, 1 ) + ~ [V(s, 1 ) - ~ (s, 1)1, s#O

/
~2 \ ds 2 ds 2 ]' s=O

is smooth, a can be factored through A: it is the map a = A to,

a: (s,t) ~ (s, s2t ) A M

Consequently, if )-A is a lift of A, then )~atr is a lift of a, using the reparametriz-

ation invariance of the lifting. Since d~c(O, 1 ) / d s is the same tangent vector
as dtr O)/Os, it follows that

w~_wo_aa(o,1) aa(O,O)=o
8s 8s

Similarly, w ~ - w~ = 0. 9
The next lemma shows that only the difference v 1 - v~ is important.
L e m m a 8. Suppose we have two smooth one-parameter.families of
paths ~ and ~t' in M shrinking to the point *, with lifts 2 and Z' shrinking
to the point b ~ n - l ( * ) just as in Lemma 7, but with v ' 1 - v'~ v I - v~ then
it follows that w 'l - w '~ = w ~ - w ~
P r o o f Again we can suppose that M = Rm, with * = 0. Suppose four
vectors v 1, v~ v'1, v'~ ~m are given, with v'~ - v '~ v I - v~ The two functions
v a n d ~': I x I ~ M
g: (s, t)--*s[(1 - t ) v ~ tv 1]

V': (s, t)--*s[(1-t)v'~ '1]

have boundary vectors vl=8g(O, 1)~as, etc. The two-parameter family

: I x I ~ M I is defined to interpolate between the one-parameter families ~'
Holonomy and Path Structures in GR and YM Theory 1193

and ~'. The 9 is defined by mapping I x I x I into R'~ by the linear map which
extends

(1, 1, 0)--+v~ (0, 1, 0 ) ~ v '~

(1, 1, 1 ) ~ v 1, (0, 1, 1 ) ~ v '~

~t and ~,' factor through q~: gt=q~cr and ~'=qbcr' with cr and
or': I x I ~ I x I x I defined by

cr(s, t) = (s, s, st)

a'(s, t) = (0, s, st)

The point of the construction is that (Figure 5)

~ 8o-(0,1) 8o-(0,0) 8or'(0,1) 0o-'(0,0)

8s 8s Os 8s

If 2 . lifts q~, then ~ , a and 2oct' lift ~t and gt', and hence ~*(0, 0, 0)(4) =
w1 - w~ w' 1 - w'~ The lemma is proved for the particular maps N, N' con-
structed here by linear interpolation in t. But Lemma 7 shows that any other
families of paths will give the same result. I

To complete the proof of Lemma 5, we note that Lemmas 6-8 show

that the map Fb: T , M ~ ThE is well-defined. The fact that it is linear hinges
on the product property of the lifting. Suppose that Z: I ~ M ~is the composi-
tion of two families ~t and ~' shrinking to the point ,, i.e.,

2 ( 0 = ~(i)~ [?'(i) and t~(O) = ~'(0) = 9

lxl lxlxl M

Fig. 5
1194 Barrett

If v ~, v ~ Vtl, and v'~ are the boundary vectors of ~ and g', then v~ v'1, and
v ~ and v '~ are the boundary vectors of Z. The product property of the lifting
implies that

F ( v I - v '~ = F ( v 1 - v ~ + F ( v " - v '~

and so F is linear.
The conditions of the lemma are thus satisfied at the point c = b. The
same proof can be applied at other points because the holonomy mapping
based at other points in E obeys the same axioms H1-H3. 9

The smoothness of the distribution on E given by the image of F is a

straightforward consequence of the smoothness property of l.. This com-
pletes Lemma 3. 9

2.1.5. Conclusion

The representation theorem provides an alternative model for the

Yang-Mills configuration space, and gives some useful insights into its struc-
ture. It is the development of much earlier work on topological bundles
(Milnor, 1956; Lashof, 1.956), which used homotopy classes of lifting
functions to classify topologically distinct bundles.
There is a similar type of representation for the gravitational field, which
is the subject of the next section. The results here form the basis of the
gravitational construction. Apart from this technical role, they also provide
an interesting comparison of the construction and structure of the classical
Yang-Mills and gravitational fields. This gives a fresh view of the deep
similarities and the differences of principle between these two types of phys-
ical fields.
Section 2.4 gives some applications of the Yang-Mills results. Section
2.5 relates the curvature two-form to the holonomy of families of loops, and
shows how this fits in with the result of Lemma 4.

2.2. Bundles and Liftings

Many physicists are quite happy with the idea that gauge fields are Lie-
algebra-valued one-forms on the base manifold, and regard bundles as an
extra complication, which one can happily do without. For them, all bundles
are trivial, and the connection is a G-invariant one-form on M x G. The
extra G's worth looks rather redundant. So why is the language of bundles
natural and appropriate here? To answer this question, first the idea of
horizontal lifting will be described.
Hoionomy and Path Structures in GR and YM Theory 1195

In general terms, if z: E ~ M is a function, then a lifting assigns a path

;t in E to each path p in M in such a way that zr projects the path 2, onto p:

I/ p M

Usually one can freely specify the starting point 5t (0) of 2 in z-~(p(O)).
The lifting is usually required to vary continuously with respect to both the
path in M and the starting point in E, but the details of this will be omitted
here.
For a gauge theory (defined by potentials on the base) one can define
the parallel transport map r: MI~G. The horizontal lift of a path p into the
trivial bundle M x G starting at (p(0), g) is the path i~(p(i), gr(K(p, i))),
where K is the contraction in path space: K(p, i) is the path which traverses
the section [0, i] of path p, i.e., K(p, i)[j] =p(ij). The holonomy of a loop,
which was defined in Section 2.1, can now be seen to be just the parallel
transport operator for the loop.
Any lifting function can be specified by giving, for each path p in the
base space, a vector field w in the section of the bundle above the path, so
that the lifts A, are the integral curves of the vector field. The vector field
must project down onto the tangent vector field of p in the base, and so it
is determined up to an arbitrary vertical component. The horizontal lifting
is a very special type of lifting function. For each point ceE, there is a linear
mapping Fc: T~(c)M~T~E such that for any path p, w=d~/di=F(dp/di).
The image of F is called the horizontal distribution of TE, and at a point e
it defines the horizontal subspace of TeE. To put this another way, the
important point is that for any two paths passing through the same point,
p(i) = z(c), and having the same tangent vectors at that point, the tangent
vectors of the lifts agree. The vector w depends only on the tangent vector
of the path at that point, and on no other feature, local or nonlocal, of the
path. That is what distinguishes the horizontal lifting of a connection from
any other lifting.
After this slight digression on lifting functions, we come back to the
original question: why are bundles essential? The short answer is that in the
holonomy representation nontrivial bundles automatically appear alongside
the trivial ones (we are supposing that the base space actually has some
nontrivial bundles, so that it is not a contractible space, as, for example, ~n
is). The reason for this is that the concept of parallel transport "unifies" the
infinitesimal aspect of the connection one-form (on the base space) with the
topologically nontrivial global aspect of the finite transformations of the
transition functions on overlapping charts. In bundle language it "unifies"
l 196 Barrett

the connection and the topology of the bundle. To put it crudely, the only
difference between an electromagnetic field on a circle and a Moebius band,
as far as parallel transport and holonomy are concerned, is that in one case
the structure group (U1 or ~) is connected and in the other case (Z2) it is
not. Conventionally, one would ascribe the holonomy in the first case to the
presence of a field, and in the second case to the topology of the bundle.
When one comes to consider the set of all holonomy mappings, it is rather
artificial to separate out those that belong to different bundles. For different
connections on the same bundle, the holonomy mappings are in the same
homotopy class, so perhaps one could restrict attention to only one homo-
topy class of maps. This seems a cumbersome idea, and since there is nothing
to be lost by considering different bundles, it is not pursued.
There is a second reason for using bundles in Yang-Mills theory. The
bundle constructions have a very deep analogy with the reconstruction of
the manifold of general relativity, which is described in later sections. The
analogy is that the base space of Yang-Mills theory corresponds to the
tangent space in gravity, and the total space of the bundle corresponds to
the differentiable manifold of general relativity. Thus, to explore the relation
between the two theories, it is essential to use the theory of bundles for
Yang-Mills fields.

2.3. The H-Group Structure of the Loop Space

In homotopy theory the H-group structure of the loop space f~M is
important. The H-group property of the composition law is that the follow-
ing three homotopy equivalences hold:

0),-~ a ) o t

(.0 o ( 0 - 1 ~ t

( a o 13) o ? , ~ a o (13o ~')

where t is the trivial loop, so that if the loop space is factored by homotopy
equivalence, the result is an algebraic group.
The important thing to notice is that these homotopies are all of the
form
h
IxI 'I ' M

where f~ is the loop on the left-hand side of the three homotopy equivalences.
The map h is not continuous, but the overall map f~h is. The homotopies
are thus actually all thin equivalences, and so f ~ M / 0 (0 is the thin equivalence
Holonomyand Path Structures in GR and YM Theory 1197

relation) is a group. For the analogous constructions in the space of piece-

wise linear loops on a simplicial complex, see Milnor's work on universal
bundles (Milnor, 1956).

2.4. Flat Bundles in Physics

There are many situations in physics where the information about a
gauge field is given by specifying its holonomy. One example is a gauge
field which has zero curvature, giving homotopic loops the same holonomy
element, but with nontrivial holonomy elements for noncontractible loops.
The holonomy mapping descends to a group homomorphism
H: 7q(K)--,G

Now any such holonomy mapping obeys conditions H1-H3. H1 is

immediate, H2 follows because thin equivalence is a restricted notion of
homotopy equivalence, and H3 follows because if ~: U--,DM is a family of
loops, then Hq~ is constant on connected components of U, and hence
smooth. The reconstruction theorem can be applied, and so it is clear that
the gauge field is properly specified by just the information in H.
This type of situation occurs in several different physical models. Some
examples are the Aharonov-Bohm effect, vacuum configurations for gauge
theories in a Kaluza-Klein context, where the holonomy elements are a
symmetry-breaking mechanism, and again in the context of topological field
theories.

2.5. The Curvature Two-Form

It is a well-known "fact" that the holonomy of a "small" loop expands
as

where the integral is over the region V which the loop bounds. One can
make this expansion precise: If gt: I ~ M is a smooth family of loops with
gt(0) = t, then the second derivative of the holonomy is related to the curva-
ture tensor F on the base manifold by
d2(H~)
as 2 (0)=F(A)

A is a bivector, which contracts with F to give a Lie algebra element. A

indicates the asymptotic shape of the area of the loops as they shrink to
1198 Barrett

zero. Its definition is

= -~2
- ~o xtUdx vl = -~2
- f v dx EuA dx vl
A'v Os2 v Os2

where x ~ is the coordinate function, and the derivatives are taken at s = 0.

These relations can be proved by differentiating the differential equation for
the parallel transport operator. The linear term of the expansion of the
holonomy of a family of loops shrinking to the trivial loop is zero:

d(H(g)
- - (0) = 0
ds
This is the content of Lemma 4 of Section 2.1.

Note: The quantity A vv is a tensor at the point ,. One can take the
~2/0s2 inside the integral to define it as
fo dxt~ d2xVl
dt ds ds dt

with the derivatives evaluated at s = 0, which is an integral over vectors at

*, since xV(0, t) = ,. The other two terms of the differentiation vanish.

3. H O L O N O M Y AND GRAVITY

3.1. Introduction
The last section was concerned with Yang-Mills theory in its own right.
This section presents the analogous constructions for general relativity. The
Yang-Mills work provides both the general framework for a similar treat-
ment of gravity and some specific results which are of use in this section.
The work on gravity is not complete in the same way as the Yang-Mills
results, and so a less formalized presentation is used. The difficulties are
of a fairly technical nature; mainly questions about differentiability and
differentiable structure, and are still open problems. However, the goals are
the same: we are looking for the holonomy representation of a gravitational
field configuration, an axiomatization of holonomy, a reconstruction theo-
rem for the field configuration, up to a diffeomorphism, and finally a repre-
sentation theorem for the gravitational configuration space. Perhaps the
reader can keep in mind the more formal development of the Yang-Mills
theory.
Holonomy and Path Structures in GR and YM Theory 1199

In order to present properly the analogy which motivates the gravi-

tational constructions, the basic idea of the Yang-Mills theory is presented
afresh.

3.2. Y a n g - M i i l s

A Yang-Mills field is a connection on a principal bundle B=

(E, lr, M, G). The base space M can be any differentiable manifold, but we
have in mind mainly Minkowski space. A curved space-time M involves
gravity, and since the point of this section is to represent a gravitational field
in a different way, it is better to think of the Yang-Mills theory being i-a the
absence of gravity.
The important aspect of the connection here is that it provides a lifting:
given a path in the base space and a point in the fiber over the initial point
of the path, it gives a path in the bundle starting at the given point. This
point is the one whose tangent vectors are horizontal at every point, and
project down onto the tangent vectors of the original path in the base.
Using the idea of lifting, we can describe the points of the total space
of the bundle in an unusual manner. We fix a basepoint 9 in the base once
and for all, and also one in the fiber over., so that the fiber becomes a copy
of G. A point c E E is to be described by an equivalence class of P M x G. It
will be convenient to call this equivalence class c also. A pair (p, g) is in c
ifp starts at 9 and ends at ~c), and the lift ofp starting at g ends precisely
at c.
This description of the bundle B describes the connection neatly, too.
The best way to see this is to exhibit the lifting, just using the information
given in the equivalence classes. This was described in Section 2.
The final part of the story is that there is an equivalence relation which
describes the classes in P M x G of the points in E. R is the relation
(p, g) ~ (p', g') if the endpoints o f p and p' coincide, and g' = H ( p - 1 o p,)g.
H is the holonomy mapping of closed loops in M to their holonomy elements
in G. Thus, both the bundle and the Yang-Mills field configuration, the
connection, are described by the holonomy mapping. The holonomy map-
ping involves only the base space M and the group G. To calculate the
field configuration it describes, the bundle has to be constructed using the
equivalence relation R:

E = PM • G/R

and then the lifting function follows.

Section 2 presents this in much more detail, together with axioms for
the subspace Jt~cMap(f~M, G) which represents all possible holonomy
mappings.
1200 Barrett

So, to recap, Yang-Mills theory can be said to be about the holonomy

of loops in M. The lifts of these loops do not dose, the endpoints being in
the same fiber, and mapped onto each other by the holonomy element in G.
In gravity we shall see that the "lifted" loops do not close by a translation
in a space-time direction, and by a rotation of frames.

3.3. Gravity
The gravitational field is described by a differentiable manifold X with
a metric, and a connection on the bundle O(X) of orthonormal frames. As
before, we pick an arbitrary basepoint * in X and a frame f a t ,, so that the
fiber over X is identified with the Lorentz group G. To achieve the holonomy
description of gravity, we need the idea of development. The development of
a curve in the base space p s P X is defined by horizontally lifting the curve
into O(X). Then, using the canonical ~n-valued l-form on O(X) (the
t
"vierbein"), e, the integral C(t)= ~o e, integrating along the lift of p, gives a
curve in ~ , which is identified v i a f w i t h the tangent space at ,, viewed as
an affine space with a metric, i.e., Minkowski space M. For the rest of the
section, M is now definitely Minkowski space. Intuitively, we can think of
the development as the curve in PM with the same geometry as the path p,
that is, it bends through the same angles at the same proper distances as p.
It has the same intrinsic and extrinsic geometry as p. One can also define
the curve C as the unique path in M obeying C(0)--0 and

dC_.dtrK(P'i~(~t)

where r/~w.,) is the parallel transport map: Tp(,)X~ T,X. Note that a closed
loop in PX will not in general develop a closed loop in M.
Now we can see how to describe the manifold X:
point of X = subset of PM
where p is in x e X if p is the development of a path which ends at x. The
gravitational field is described:
point of O(X) = subset of PM x G
where the G element is defined in exactly the same manner as for Yang-
Mills theories: the frame is parallel transported back to the fiber above ,,
using the connection. This second set of subsets is a very powerful object.
It describes the set X, its differentiable structure, its metric, the frame bundle
O(X), and the connection on it, as will be shown in detail below. In the
same way as for Yang-Mills theory, there is an underlying equivalence
relation yielding these subsets of PM x G as equivalence classes. This relation
Hoionomy and Path Structures in GR and YM Theory 1201

is specified by the holonomy set ~ c P M • G, which is just the subset of

P M x G which corresponds to the f r a m e r One can decompose ~ into a set
P c P M and a mapping h: P ~ G , so that (p, g) ~ i f f p ~ P and g = h(p). The
set P is the subset of P M corresponding to the point 9 6X, or in other words,
the paths in M which are the development of closed loops in X. The function
h is the Lorentz holonomy mapping of P, and is simply the holonomy map-
ping, in the sense described for Yang-Mills theory, of the Lorentz connection
in the bundle O ( X ) for the loops in X corresponding to the elements of P.
The (affine) holonomy mapping H of P takes a point (p, g) to the iso-
metry of M given by a rotation of g about the origin followed by a translation
by an amount p(1), the value of the endpoint of the path:
H: P~A
p~p(1)h(p)
where A is the Poincar6 group. H is in fact the holonomy mapping of the
affine connection in the bundle of affine frames. Note that H can be derived
from the information in ~, or, equivalently, the information in P and h.
Now we come to the description of the two equivalence relations, on
P M and on P M x G, which reconstruct the manifold and the bundle of
orthonormal flames, purely in terms of the information in P and h. The
relation R on P M x G is
R: (p, g) ~ (p', g') if there exists zr~P such that
re= (H0r)p '-1) op
and g = h(~r)g'
The first is rather a curious equation, since ~r "appears on both sides."
If we want to consider 0r, h0r)) as a transformation acting on the space
P M x G, the condition can be rewritten: There exists lr~ P such that
p ~ (HOOp')~ Jr
g = h( rc)g'
with the ~ denoting thin equivalence (see Section 2).
With the relation R we can form the set
E = PM x G/R
which is identical to the total space of the bundle. Using the relation R' on
P M defined in a similar way,
R': p~p' if there exists zr~P such that
zr= ( HOr)p'- ') o p
1202 Ba~e~

it is clear that P M / R ' is the set X, and that the projection Jr: E---,X defined
by {p, g} R~ {p} R' is intrinsic to the construction. It is also extremely plausible
that the differentiable structure, metric, and connection are implicit in the
constructions. As far as the metric goes, we can see that X is "made of"
sections of paths in M, and so distances can be measured by measuring the
corresponding distances in M. This can be formalized by defining the inverse
development map A,

A: PM--* PX
A(p)[i] = {K(p, i)}

where K is the contraction in path space introduced in Section 2:

K(p, i)[j] =p(/j). The map A gives paths A(p) in the set X with a known
metric geometry, i.e., that of p e P M , and hence can be expected to give a
metric to X. The differential structure should follow because A allows the
construction of smooth families of paths in X from smooth families of paths
in M. In other words, the differentiable structure is that of M = ~4, patched
together in a way determined by the equivalence relation. From these consid-
erations, it is seen that h plays the role for the manifold X that l. plays for
the bundle in Yang-Mills theory.
Finally, the connection in the bundle E ~ X should be reconstructable
in much the same way as for Yang-Mills theory. In fact the strategy is to
construct the space-time manifold, and then the holonomy mapping of the
(Lorentz) connection, satisfying the axioms H1-H3 of the Yang-Mills
theory. Then the Yang-Mills results can be used to construct the connection
on the frame bundle.

3.4. Axiomatization and Reconstruction

The previous section showed how to develop the holonomy representa-
tion of gravity by analogy with the Yang-MiUs construction. It exhibited
the holonomy information (P, h), and showed that it is plausible that a
complete reconstruction of the gravitational field configuration should fol-
low from it. This section seeks to axiomatize the set P c P M and the mapping
h : P ~ G , and provide a reconstruction from the axioms. Actually, the basic
ingredients of the reconstruction were spelt out in the previous section. The
axioms are not yet complete, and so parts of the reconstruction cannot be
done rigorously. The difficulties are pointed out.
Since the Lorentz group is used throughout, it is to be expected that
the reconstruction will yield a bundle with a metric compatible connection.
However, the notion of torsion has no basic construction in the holonomy
scheme of things, and so the connection will be independent of the metric,
and the torsion may take any value. Later, when the field equations are
Holonomy and Path Structures in GR and YM Theory 1203

discussed, the equation torsion = 0 will be seen to have a similar status to

the Einstein field equation.
While discussing exactly what the gravitational configuration space will
be taken to be, there are two further properties of a gravitational field to be
mentioned, namely completeness and the Hausdorff property. The question
of what happens to the holonomy of a non-Hausdorff manifold is rather
interesting. It seems likely, however, that the reconstruction process from
the holonomy information will only yield Hausdorff manifolds. This is
because in a non-Hausdorff manifold there exist paths [0, 1)~X, open at
one end, with two possible endpoints to complete the curve to a path
[0, 1]-*X. These paths have the same development, which is of finite
(Euclidean) length. [It is to be supposed that an arbitrary positive-definite
metric is used in the tangent space M, as is usual in discussions of complete-
ness (Hawking and Ellis, 1973).] Such paths can be constructed by the
following method. Because the manifold is non-Hausdorff, there exist two
points x and x' which cannot be separated by open sets. Now in two charts
on open sets U and U' containing the points x and x', respectively, we can
consider two sequences of open balls Bn and B ' , centered on x and x'. The
intersection of Bn with B" must be nonempty, and we choose a sequence of
points pn, with p~ lying in the intersection of Bn and B ' . If the radii of the
balls are chosen to converge to zero as n ~ , the sequence of points pn
converges to both x and x'. The path [0, 1)~X is constructed by joining
together the points in a suitable fashion, for example, by linear interpolation
in a chart. If the radii of the balls are chosen to converge sufficiently fast, then
the development of the path will have finite length. Now in the holonomy
description the points of space-time are labeled by the developments of the
paths of particles which arrive at that point; the points have no a priori
existence themselves. The construction of the manifold consists in just
grouping together the paths of all the particles which are in coincidence,
defined by the equivalence relation R'. So it would be impossible if in the
resulting manifold there were two points which were the endpoint of just
one particle path. So it would seem that non-Hausdorff manifolds are
unphysical, if we believe that the holonomy construction actually represents
the physics of the definition of the manifold, as an abstraction from the
behavior of particles.
We can still contemplate a non-Hausdorff manifold, however, and ask
what the essential difference is in the behavior of the holonomy, in distinction
to a Hausdorff manifold. It seems to me that, because two distinct paths in
the manifold may have the same development, probably one of two things
can happen. Either the holonomy does not obey the axioms to be presented
below, or it does, and the reconstruction process identifies regions which
have an identical geometry. The standard example of a non-Hausdorff mani-
fold is made by taking two copies of the real line ~, and identifying them
1204 Barrett

on the open interval (0, ~ ) . The two points 0 in each copy are the two
points which cannot be separated by open sets. The tangent space is R. If
the metric on the manifold is taken to be just the standard metric on ~, the
development of a path on the manifold is the path in E one gets by identify-
ing the two "bottom legs" ( - ~ , 0] of the manifold. Now the holonomy set
will be identical to that of R, as the base manifold, and so the reconstruction
process will give back as the manifold only ~. In other words, the two
geometrically indistinguishable "legs" ( - ~ , 0] get identified. If, on the other
hand, the two legs are given different geometries [this is not possible in one
dimension!--so imagine a two-dimensional example E x R identified
on R • (0, ~)], then the holonomy description may be inconsistent. For
example, two loops with identical development but which are partly in
different "legs" may have different Lorentz group holonomy elements. Hence
the mapping h: P ~ G would not be definable.
Completeness of a manifold can be defined in terms of the development
mapping 6: P X ~ P M which takes paths to their developments. A manifold
is said to be complete if I m ( 8 ) = PM. In physical terms, every conceivable
particle motion, as defined by its geometry, can take place, and ends at some
point in the manifold. This notion of completeness coincides with the notion
of b-completeness (Hawking and Ellis, 1973). Completeness is defined in
this way because for a pseudo-Riemannian metric the manifold does not
have the metric space structure that a positive metric would give. In the
case of a positive metric, the metric space completeness coincides with
development completeness (Kobayashi and Nomizu, 1963).
For the reasons given above, then, a gravitational field configuration
will be defined as a connected, Hausdorff manifold which is complete in the
sense that the development map is complete, with a metric of Lorentz signa-
ture, and a connection on the bundle of orthonormal frames, so that the
connection is metric compatible, but may have nonzero torsion.
Now we turn to the details of the axiomatization. We start with P and
h. From it H: P ~ A is defined in the same way as before: H(p)=p(1)h(p).
Then a product operation 9 on P is defined:
p~ * p2 = (H(p2)pl) op2
and an inverse operation p~/~:
p = H ( p ) - I p -1
The first and second axioms are as follows.
G1 (Homomorphism). If p~,pz6P, then pl *pz~P, and h(pl *p2) =
h(p2)h(pO.
G2 (Inverse). I f p ~ P , then p~P.
Hoionomy and Path Structures in GR and YM Theory 1205

Now at some stage it has to be shown that if P and h are derived from
a gravitational field, i.e., manifold X, metric, etc., then the axioms hold. In
order to motivate the axioms, this will be done as the axioms are stated.
Proof of G1 and G2 for a gravitational field configuration. P is the
development of the loop space of the manifold, and the product and inverse
operations on P are just the product and inverse operations of the loop space.
The homomorphism condition on h is just the corresponding condition for
the holonomy mapping of the Lorentz connection. 9
Since a path which is thinly equivalent to a loop is also a loop, and the
holonomy mapping of a connection agrees on thinly equivalent loops, it is
natural to propose the following.
G3 (Thin equivalence). P contains complete thin equivalence classes of
paths. The map h agrees on these equivalence classes.
Attempted Proof of G3for a Gravitational Field. The proof of this is not
complete, and is one of the technical problems referred to at the beginning of
the section. The proof rests on the following conjecture: For p, p'~PX, p is
thinly equivalent to p' iff S(p) is thinly equivalent to 6(p'). The problem in
proving this is that in general the homotopy which establishes one of the thin
equivalences does not develop (or inverse-develop) to a homotopy which
establishes the other one. This is easily seen to be the case if the image of
one of the thin loops is not a simply connected set. However, one can
establish the proof for a special class of thin loops, ones which can be
transformed to the trivial loop by a finite number of operations of either (1)
reparametrization of the loop, or (2) replacing subsections of the path of
the form p-1 op with the constant path at p(0). It is not actually clear
whether this is a more restricted class of thin loops than the original defini-
tion. There may be a pathological example which shows that it is a restricted
class.
There are three possible ways out of the problem: establish the conjec-
ture, modify the definition of thin loops to a more restricted set (all the
proofs would work with the restricted notion of thin loop given above), or
prove that this restricted set is actually all the thin loops. In any event, the
lack of a proof does not seem too serious. The proof will, however, be
assumed in the following. 9
Returning to the axiomatic development, a few facts can be established,
from the axioms.
Proposition. H agrees on thin equivalence classes.
Proposition. If P is not empty, t~P and h(t)=id, t being the trivial
loop.
1206 Barrett

Proof There exists peP, hence p , p = p - ~ ~ which is thinly

equivalent to t, so teP. Then h(t) = h(p 9 p) = h(p)h(p). But h(t) = h(t 9 t) =
h(t) 2, so h(t)=id. 9

The proof also shows the following result.

Proposition. h(p) = h(p) - ~.

Now we consider the relation R' introduced in the last section. The
three axioms introduced are sufficient to prove that R' is an equivalence
relation. The proof is rather tedious, and not particularly illuminating, and
so is omitted. Then the set X is defined by X = PM/R', with basepoint 9 =
{t}, and the inverse development map A: P M ~ P X is defined as in the last
section: A(p)[i] = {K(p, i)}. Actually it is a bit premature using the space
PX, as the set X does not even yet have a topology, let alone a differentiable
structure. The map A still exists, however, as a map into the space of
functions I ~ X , and so for the moment PX will be defined as just the image
of the map A, as a subspace of the space of functions I ~ X . Likewise, D..X
is defined as the subset {p: p ( 1 ) = .}.

Proposition. If pl, p2~PM and pl ~P2 by thin equivalence, then p~ ~pz

by the relation R'.

Proof Define O=p ;-1 op2. Then OsP and H(O)=id, and so
(H(O)p;-') op2=O
that is, p~ "~pz by R'. 9

Proposition. A maps P onto D.X, and pl * P2 is mapped to the composi-

tion of loops A(pj) o A(p2).

Proof For a general peP, A(p)(1)= {p}, so to show that A(p) is a

loop, it is necessary to show that {p} = {t}. If q is the path t o p , which is
just a reparametrization of p, then q is thinly equivalent to p, and q~P.
Moreover, q = ( H ( q ) t - ' ) o p , which means that, by definition, (t}={p}.
Clearly, the reverse argument holds; if {t} = {p}, then peP, and so A(p) is
a loop.
For the product property, consider two general paths p~ ,p2~P. Then
A(pl,p2)=A(H(p2)plop2)=qloA(p2), where ql is the path ql(i) =
{(H(p2)K(pl, i)) op2}. But it is easy to show that
R'
(H(p2)q)op2 ~ q

for any path q, and so q~(i)= {K(p, i)}, i.e., qi = A(p~). 9

Holonomy and Path Structures in GR and YM Theory 1207

Before going any further, this is the right point to introduce the smooth-
ness axioms. Clearly, so far there is no reason to suppose that X is a mani-
fold; in fact, with the axioms presented so far it may be very far removed
from a smooth finite-dimensional manifold. For example, take P to be a
proper subgroup of f~R n, for example, say by removing the loops passing
through some of the points of Rn: P = f~U, with U a subset o f R n, and h to
be trivial: h(m)= id for all og~P. Then P and h will satisfy the axioms G 1 -
G3, but the reconstruction will yield something rather bizarre, certainly not
a manifold. The paths which end at points outside U will not be related by
R' to any other paths. So if p is such a path, it will form an independent
point of X. The set X will contain "too many points" in the sense that if
one considers deforming p locally, then X is locally the same "dimension"
as the path space PM. Similarly, one can also imagine that too many points
may be identified, leading too small a dimension for X. It may also happen
that the paths are identified by R' in a chaotic, discontinuous manner, not
admitting any smooth structure for the set X.
With these points in mind, the axiom G4 should be tentatively (and
imprecisely) stated as:

P is a smooth submanifold of PM, of codimension four.

Before discussing what this might mean, a simple example serves to

motivate the axiom. Consider p~4, with * = 0. This is a vector space with
pointwise addition of paths. The loop space ~'-~4 is a linear subspace of PR 4,
and so by any decent definition of the manifold structure o f P ~ 4, this would
be a submanifold. The quotient space Q = P ~ 4 / ~ 4 is a vector space of
dimension four, and coincides with set X = P~a/R'. In fact, Q is isomorphic
to ~4 by the endpoint map of the path space.
To arrive at a more precise notion of what a smooth submanifold is,
we have to examine the smooth structure, which is relevant here, of the space
PM. There is a well-defined notion of a smooth map into P M from a finite-
dimensional manifold: the notion of a smooth family of paths. There is also
a well-defined notion of a smooth map from P M into a manifold Z: the
smooth families V: U ~ P M give a smooth map U ~ P M ~ Z . So it is natural
to suppose that the canonical projection map a: P M ~ P M / R ' = X should
be smooth in this sense. Now a = eA, where e is the endpoint map P X ~ X .
This means that the differentiable structure of X should be defined in such
a way that for any smooth family of paths @: U ~ P M , the map

U , PM ~ PX , X

is smooth. At this point, to check that we are on the right track, we can
1208 Barrett

compare this expression with the construction used in the Yang-Mills theory
to construct charts on the total space E of the bundle
(~,id) 1. e'
Cv: U x G , PMx G ' PE' ' E
which is very similar. It has already been remarked that the inverse-develop-
ment function A plays the same role for the manifold construction as l, did
for the bundle construction of the Yang-Mills theory.
Without going too deeply into the technical problems involved [which
start with the fact that ~ U) is not open in PM], we shall assume, as axiom
G4, what should be the principal conclusions of the above tentative proposal.
G4. There is a unique four-dimensional smooth structure on X such
that for any smooth family fi: U ~ P M the map eAfi: U ~ X is smooth. Paths
are mapped nondegenerately into X: if p ~ P M and ( d p / d i ) ( i ) ~ 0 , then
[dA(p)/di](i) # 0 .
The smoothness for the map h is straightforward:
G5. For any smooth family fit: U ~ P , the map hfi: U---,G is smooth.
The proof strategy only for the remainder of the reconstruction will be
sketched, as the work is not yet complete, and a more thorough exposition
would be premature. The main points to show are:
1. A can be used to construct "Riemann normal coordinates" around
any point x ~ X . For example, around the point ,, the family p: M ~ P M ,
p(m)[i] = im (i.e,, radial straight lines) should, in some neighborhood of the
origin, map invertibly to X. The derivative at the origin is nonzero on
account of the nondegeneracy condition in G4.
2. A is a 1-1 mapping of P M to PX.
3. For paths p, q ~ P M , p is thinly equivalent to q if and only if A(p) is
thinly equivalent to A(q).
4. The mapping h can, by virtue of the 1-1 correspondence of P and
D_.X, be regarded as a holonomy mapping I'bY~G. The axioms H1-H3 are
established by the points above.
5. The reconstruction theorem is applied to h. The construction is the
same as that given by the relation R directly on P M x G. The resulting G-
principal bundle needs to be interpreted as the frame bundle of X. At a point
e~ E, the frame is defined as the mapping
Oc: M ~ T.(c)X

dp dA(p) (1)
g~;i (1)-~ di
for any point (p, g ) E P M x G which is in the equivalence class of c. One has
Holonomy and Path Structures in GR and YM Theory 1209

to show that this mapping agrees on all the different possible paths p, and
is linear. This is the same as the problem of defining a connection in the
Yang-Mills theory, and one should be able to use the Yang-Mills result by
defining the affine holonomy mapping D.X~A, and constructing its bundle
(which should be the isometric affine frame bundle) and connection. Its
connection splits into two parts, the Lorentz connection, and the inverse of
the map 0.
The nondegenerateness of 0, which is essential to the notion of a frame,
follows directly from the nondegenerate property in G4.
6. Finally, the metric on X follows once the frames are established,
essentially from the fact that the reconstructed bundle contains only a subset
of frames, the orthonormal ones. The map 0c is used to map the metric on
the Minkowski space M to the tangent space at x on X. These metrics will
agree for all the points c in the fiber above x, because the frames are all
related by Lorentz transformations in M.

4. THE FIELD EQUATIONS OF GRAVITY

4.1. Introduction

In previous sections, the holonomy description gave rise to a gravi-

tational field with a connection that was naturally metric-compatible, but
the fields were otherwise arbitrary. In other words, the Einstein field equation
was not imposed, and the torsion was also arbitrary. The aim of this section
is to demonstrate a form of the field equations which is naturally suited to
the holonomy scheme. There are really two field equations, the Einstein
equation and the torsion equation, and, as we shall see, they are naturally
paired as the linear momentum field equation and the angular momentum
field equation.
So far, the holonomy description has proceeded with the minimal use
of tensors; there are displacement vectors and Lorentz group elements, but
nothing more complicated than this. The plethora of different tensor types
which usually accompanies general relativity is rather a foreign element in
this approach. So, to continue in this spirit, maximal use will be made of
differential forms in expressing the equations of motion. To give an example,
the energy-momentum tensor pab is best expressed as a vector-valued three-
form pa=pabl~bcdeeCA edA e e, where e is the unit vector-value one-form. This
has a more geometric meaning than the former; when integrated with a
"small" three-surface element (over which the curvature can be ignored) it
gives the energy-momentum passing through that surface. It also has the
technical advantage that the covariant exterior derivative can be applied,
1210 Barrett

giving different treatment to vectors and differential forms, when there is

torsion in general.
Going further than this, the spirit is to introduce the relevant geometric
objects over which the differential forms are to be integrated. This has been
pursued to its logical conclusion already in previous sections as far as the
connection one-form is concerned: one-forms are integrated along paths,
and the result is a theory formulated in terms of functions on path space.
The strategy here is then to introduce "small" three-surface elements and
their bounding two-spheres over which the momenta are integrated. Precise
results can be expressed in terms of the limit of a family of spheres which
shrink to a point in a smooth way.

4.2. Field Equations

The two equations of motion are (Kibble, 1961; Sciama, 1962)
1 l.~abA eCg,abcd=Pd
~,. (energy-momentum density)

1 .ca A e b ~abcd= Scd (spin density)

where R is the curvature two-form and r the torsion two-form. The two
quantities on the left-hand sides are the Einstein tensor and the modified
torsion tensor. When the right-hand sides are set to zero the equations
become the vacuum Einstein equation and the equation r = 0, expressing the
connection form in terms of the metric.
Let us examine the Einstein equation first. What is needed is a precise
expression of the idea that over a sufficiently "small" three-surface V, so
that the curvature over its extent can be neglected, the integral of the Einstein
tensor over V is equal to the matter energy-momentum passing through the
surface V,

89~ R ab/~ ece~bcd.~energy-momentumthrough V

,) V

To perform this integral properly, what is needed is a notion of parallel

transport for the vector index of the integrand. The idea is that V is
"sufficiently small" for any reasonable parallel transport of the vectors to
one point in V to produce a result differing only by corrections of higher
order in the size of V from the result of the integration itself. Suppose that
V is topologically a three-disk, so that the boundary ~3V is a topological two-
sphere. Then V can be filled with a family of curves (a "spray") which each
start at the same point v in the interior of V and end at the different points
of t3V. then the vector in the integrand at a point u ~ V is parallel transported
Holonomy and Path Structures in GR and YM Theory 1211

along the unique path which links it with v. An alternative way of looking
at this is that a special type of gauge has been picked, that in which the
connection form o~ is zero in the directions along the curves, and the inte-
gration is performed in this gauge. At the point v, curves radiate in all
directions, and so a~= 0.
The quantity I(V) = 89Sv R~bA ecgabcd' defined by the aid of a particular
spray of curves from v, is the same up to third order in a small parameter s,
for any spray of curves. What this means precisely is that if one has a smooth
one-parameter family of three-disks V(s) shrinking to the point v at s = 0
(i.e., a smooth map to M of a regular cone whose base is a standard three-
disk), and one attaches two different smooth families of sprays to these, then
the two different integrals I(V(s)) agree up to third order in s at s = 0. Clearly,
since the integration is over a three-dimensional region, this just follows
from the equality of the integrands at the point v. Note that for this result
it is important that V(s) is parametrized smoothly, as defined above.
This formula is not particularly exciting as it stands, but can be rewritten
in an interesting way. The same spray of curves can be developed into
Minkowski space M. Suppose that the manifold has a b a s e p o i n t . , and that
c is an arbitrary curve which connects 9 to v. The paths that were chosen in
V are connected to c and then developed into M, the tangent space of the
point .. Then the region Vcan be mapped into M by mapping a point u s V
to the endpoint of the development of the path leading to u.
Let us formalize this briefly. We started with a family of paths
~: V ~ P o X (PoX is the path space of the manifold X, based at v). These
ended at the point in question: ~(u, 1)=u. The map x: V ~ M was defined
by x(u)= 6(~t(u) o c)(1), where 6 is the development map P , X ~ P M .
Again, one can regard this as a special type of coordinate gauge fixing,
x providing a particular type of coordinate chart, such that dx = e for vectors
along the family of curves. The point is that co and x obey

co(v) = O, dx(v) = e

so that the "coordinates" are very particular ones: they are adapted to the
geometry of the manifold at v.
The local conservation of energy-momentum dp"(v) = 0 (which actually
only holds if the torsion or curvature vanishes at v) suggests that the integral
over V can be rewritten as a boundary term. In fact, the relevant expression
is

1E'abcdIt?V Rabxc ~--~" VPd

1212 Barrett

which again holds up to third order. This formula is true because, using
Stokes' formula, the integrand at v is equal to 89 times

d(RabxC)(v) = (dRabx C+ R ab A eC)(v)

At the point v, exterior differentiation d is equivalent to exterior covariant

differentiation (D), because co(v)= 0. Hence the Bianchi identity gives d R =
0. The formula is known as Cartan's moment of rotation (Cartan, 1922;
Misner et al., 1972). It relates the energy-momentum threading through the
two-sphere a V to the integral of the "moment of rotation" R[abx c]. So this
is our first field equation, that the integral formula holds to third order for
any smooth family of two-spheres which shrinks smoothly to a point v e X .
It is perhaps a highly inefficient way of stating the Einstein equation, but
physically it gives a very appealing picture. Note that, in contrast to the
conservation law D p a = O , which only holds when r=0, the integral
expression gives the correct formula with torsion.
The formula for Cartan's moment of rotation found an application to
Regge's theory of discrete general relativity (Regge, 1961), which for my
part, was a result of considering the formulas here (Barrett, 1985, 1986,
1987, 1988; Miller, 1986). Regge's equations of motion can be understood
as a discrete version of the two-sphere expression which rather surprisingly
turns out to be an exact formula, rather than approximate to third order as
it is here.
In Cartan's time the relativists did not think of the torsion equation as
a second field equation, although Cartan and other mathematicians worked
out the mathematical theory of torsion. The relativists just set ~-= 0, and
probably did not ask for the same sort of physical picture that Cartan's
moment of rotation gives to the Einstein equation. As far as I am aware,
the presentation of the torsion equation as a "small two-sphere" integral
expression is a new idea.
The modified torsion tensor, which forms the left-hand side of the tor-
sion equation, is not covariantly conserved, since it is equal only to the
matter spin density and not the total angular momentum. However, by using
the idea that the total angular momentum should be covariantly conserved,
one can arrive at the following formula: If we set

Mab = Sab -]- x[aPb]

where the quantities S, p, and x ( v ) are evaluated at v, then

1 cd 2 1 c d
M.b=d(E~,bcd(~R X .-k~C x ))(~)
where C" is the translational curvature C a ( u ) = R a b x b - r a. The C a is the
Holonomy and Path Structures in GR and YM Theory 1213

translational part of the affine holonomy for a "small" loop at the end of
the curve e.
The corresponding integral expression is

M a b ~ ~abcd ; V
gl~ X "1"-~l.~ X

with the equality again holding to third order. The two terms in the integrand
might be called the first moment of translational curvature and a second
moment of (rotational) curvature. The equation states that the integral of
these two moments of curvature over a small two-sphere t?V is equal to the
matter angular momentum passing through the two-sphere.
This equation can be regarded as the second field equation because if the
first field equation holds, one can subtract the "orbital" angular momentum
X(V[a)Pb] from the total angular momentum Mab, and the result is the equa-
tion relating the modified torsion tensor to the spin density.
The origin dependence of the total angular momentum Mab is what one
would expect. If the angular momentum is measured from a different origin,
or indeed the same origin but connected to it by a different path c', then the
coordinate of v changes by x ' - x , and the angular momentum changes by
M ' - M = (x' - X[a)Pb]. In addition, if there is a change of frame, all the vector
indices are rotated by the Lorentz transformation. The linear momentum
was independent of change of origin, but behaved as a vector under a change
of frame.
Due to the origin dependence of M~b, it is possible to sort out the
intrinsic spin of the matter from the orbital angular momentum, thus resolv-
ing the ambiguity noted by Kibble (1961). Roughly speaking, the spin part
of the angular momentum is the part that cannot be transformed away by
the change of origin mentioned above. The details of how this works in
ordinary flat space is contained in Penrose and MacCullum (1973).
The two small-sphere expressions for linear and angular momenta pre-
sented here have appeared previously in different contexts, due to the fact
that the small-sphere limit is the same as the weak-field limit (Penrose, 1982).
The expressions, for zero torsion, therefore agree with the twistor expressions
for quasilocal momenta (Penrose, 1982; Kelly et al., 1986).
A suitable representation of the equations of motion has been achieved
with the aid of the notion of gravitational holonomy, particularly the use of
development. The linear and angular momenta of matter passing through
small two-spheres are equated with integrals of moments of the two curva-
tures, rotational and translational. All these quantities have fairly immediate
significance in the holonomy description. The curvatures are essentially the
holonomy elements for small loops, as explained in Section 2.5. The other
1214 Barrett

quantity is the displacement vector x defined using the notion of develop-

ment. Thus, the equations are constraints on the quantities of holonomy
and development of paths in a fairly direct way.

ACKNOWLEDGMENTS
Thanks are due to the following for discussing various areas: Jeeva
Anandan, Rob Baston, Bernard Kay, Michael Singer (Yang-Mills axioms),
Chris Clarke (Hausdorff property), David Elworthy (path space topology),
and Luke Hodgkin and Graeme Segal (literature). I am indebted to Chris
Isham, Tom Kibble, Robin Tucker, and Nick Webber for general discussion
and criticism.

REFERENCES
Aharonov, Y., and Bohm, D. (1959). Physical Review, 115, 485-491.
Anandan, J. (1983). Holonomy groups in gravity and gauge fields, in Proceedings Conference
Differential Geometric Methods in Physics, Trieste 1981, G. Denardo and H. D. Doebner,
eds., World Scientific, Singapore.
Atiyah, M. F. (1980). Geometrical aspects of gauge theories, in Proceedings International
Congress Mathematics, 0. Lehto, ed., Helsinki, pp. 881-885.
Babelon, O., and Viallet, C. M. (1981). Communications in Mathematical Physics, 81, 515-525.
Barrett, J. W. (1985). The holonomy description of classical Yang-Mills theory and general
relativity, Ph.D. thesis, University of London.
Barrett; J. W. (1986). Classical and Quantum Gravity, 3, 203-206.
Barrett, J. W. (1987). Classical and Quantum Gravity, 4, 1565-1576.
Barrett, J. W. (1988). Classical and Quantum Gravity, 5, 1187-1192.
Barrett, J. W. (1989). General Relativity and Gravitation, 21, 457466.
Bialynicki-Birula, I. (1963). Bulletin de FA cadbmie Polonaise des Sciences, 11, 135.
Cartan, E. (1922). Comptes Rendus, 174, 437-439.
Chan, H.-M., and Tsou, S. T. (1986). Acta Physica Polonica B, 17, 259-276.
Chan, H.-M., Scharbach, P., and Tsou, S. T. (1986). Annals of Physics, 166, 396~21.
Dirac, P. A. M. (1931). Proceedings of the Royal Society of London A, 133, x-xxi.
Dugundji, J. (1966). Topology, Allyn and Bacon, Boston.
Durhuus, B. (1980). Letters in Mathematical Physics, 4, 515-522.
Einstein, A. (1922). The Meaning of Relativity, 6th ed., Chapman and Hall, London.
Fischer, A. E. (1986). General Relativity and Gravitation, 18, 597-608.
Giles, R. (1981 ). Physical Review D, 24, 2160-2168.
Hawking, S. W., and Ellis, G. F. R. (1973). The Large Scale Structure of Space-Time,
Cambridge University Press, Cambridge.
Isham, C. J. (1981). Quantum gravity--An overview, in Quantum Gravity 2, A Second Oxford
Symposium, C. J. Isham, R. Penrose, and D. W. Sciama, eds., Clarendon Press, Oxford.
Isham, C. J. (1984). Topological and global aspects of quantum theory, in 1983 Les Houches
Summer School Lectures "'Relativity Groups and Topology", North-Holland, Amsterdam.
Kelly, R. M., Tod, K. P., and Woodhouse, N. M. J. (1986). Classicaland Quantum Gravitation,
3, 1151-1167.
Kibble, T. W. B. ( 1961). Journal of Mathematical Physics, 2, 212-221.
Holonomy and Path Structures in GR and YM Theory 1215

Kobayashi, S. (1954). Comptes Rendus, 238, 443-444.

Kobayashi, S., and Nomizu, K. (1963). Foundations of Differential Geometry, Volume 1,
Interscience, New York.
Lashof, R. (1956). Annals of Mathematics, 64, 436-446.
Mandelstam, S. (1962a). Annals of Physics, 19, 1-24.
Mandelstam, S. (1962b). Annals of Physics, 19, 25-66.
Mandelstam, S. (1968a). Physical Review, 175, 1580-1603.
Mandelstam, S. (1968b). Physical Review, 175, 1604-1623.
Miller, W. A. (1986). Foundations of Physics, 16, 143-169.
Milnor, J. (1956). Annals of Mathematics, 63, 272-284.
Misner, C. W., Thorne, K. S., and Wheeler, J. A. (1972). Gravitation, Freeman, San Francisco.
Narasimhan, M. S., and Ramadas, T. R. (1979). Communications in Mathematical Physics, 67,
121-136.
Penrose, R. (1982). Proceedings of the Royal Society of London A, 381, 53-63.
Penrose, R., and MacCullum, M. A. H. (1973). Physics Reports, 6C, 241-316.
Polyakov, A. M. (1979). Nuclear Physics B, 164, 171-188.
Regge, T. (1961). Nuovo Cimento, 19, 558-571.
Sciama, D. W. (1962). On the analogy between charge and spin in general relativity, in Recent
Developments in General Relativity, Pergamon, Oxford.
Singer, I. M. (1981). Physica Scripta, 24, 817-820.
Spanier, E. H. (1966). Algebraic Topology, McGraw-Hill, New York.
Teleman, C. (1960). Annates Scientifiques de l'Ecole Normale Sup~rieure 3, 77, 195-234.
Teleman, C. (1963). Annali di Matematica, Pura ed Applicata, LXII, 379-412.
Teleman, C. (1969a). Indagationes Mathematicae, 31, 89-103.
Teleman, C. (1969b). Indagationes Mathematicae, 31, 104-112.
Wilson, K. G. (1974). Physical Review D, 10, 2445-2459.