Introductory Lectures To Loop Quantum Gravity: Pietro Doná and Simone Speziale
Introductory Lectures To Loop Quantum Gravity: Pietro Doná and Simone Speziale
Introductory Lectures To Loop Quantum Gravity: Pietro Doná and Simone Speziale
a
Centre de Physique Théorique †, CNRS-Luminy Case 907, 13288 Marseille Cedex 09, France
b
Scuola Normale Superiore, Piazza dei Cavalieri 7, 56126 Pisa, Italy
Résumé
Nous présentons une introduction de base à la gravité quantique à boucles, des va-
riables ADM aux etats du réseau de spin. Nous incluons une discussion sur la géométrie
quantique sur un graphique fixe et sa relation avec une approximation discrète de la rela-
tivité générale.
Abstract
∗ Lectures given at the 3eme Ecole de Physique Theorique de Jijel, Algeria, 26 Sep – 3 Oct, 2009
† Unité
Mixte de Recherche (UMR 6207) du CNRS et des Universites Aix-Marseille I, Aix-Marseille II et du Sud
Toulon-Var. Laboratoire affilié à la FRUMAM (FR 2291).
15
16 G RAVITATION QUANTIQUE
TABLE DES MATIÈRES 17
2 Tetrad formulation 31
2.1 The action in terms of tetrads . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.2 Hamiltonian analysis of tetrad formulation . . . . . . . . . . . . . . . . . . . 35
2.3 Smearing of the algebra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
2.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
5 On quantum geometry 59
18 G RAVITATION QUANTIQUE
where Z
1 √
SEH (gµν ) = d4 x −g g µν Rµν (Γ (g)) (1.2)
16πG
is the Einstein-Hilbert action. Here g is the determinant of the metric, and Γ(g) is the Levi-
Civita connection entering the covariant derivative of vectors and tensors,
1 νλ
∇µ v ν = ∂µ v ν + Γνρµ (g)v ρ , Γνρµ (g) = g [∂ρ gλµ + ∂µ gλρ − ∂λ gρµ ]. (1.3)
2
From courses in QFT, we know how to properly define only the Gaussian integral
Z Z
1
Dϕ exp − ϕϕ . (1.4)
2
This corresponds to a free theory. An interacting theory can be treated in perturbation theory
from a generating functional based on (1.4). However, there is no quadratic term in the
Einstein-Hilbert action (in fact, it is not even polynomial). This obstruction can be bypas-
sed if one performs a perturbative expansion of the metric. Consider the field redefinition
where ηµν is fixed to be the Minkowski metric, and hµν the new dynamical field. The intro-
duction by hand of the background field ηµν is crucial : if we treat hµν as a small fluctuation,
a perturbative expansion of the action gives (in the De Donder gauge)
Z
1
SEH (gµν ) = d4 x hµν hµν + o(h3 ). (1.6)
32πG
The background metric can be also used to define the Wick rotation. This action is amenable
to the machinery of QFT. Among the main results of this background-dependent perturbative
approach, let us highlight the following ones :
1. The quanta of the gravitational field are interpreted as massless spin-2 particles (the
“gravitons”).
1 A PATH TO LOOP QUANTUM GRAVITY 19
2. In the low energy static limit, one can compute quantum corrections to the classical
Newton potential [1], recovering the classical relativistic correction o(G2 ) and a new,
purely quantum one o(~G2 ) :
m1 m2 m1 + m2 127 ~G −3
VN = −G 1+G − + o(r ) (1.7)
r r 30π 2 r2
3. As for other quantum field theories, one can define the S-matrix to describe the funda-
mental observables, and compute scattering amplitudes of gravitons with matter fields
and with themselves [2]
This brief list does not exhaust the interesting results obtained with this approach. On
the other hand, there is a fundamental difficulty with it. As for any quantum field theory,
infinities appear when considering the effects of arbitrarily small (“ultraviolet”) field fluc-
tuations. These divergences are usually dealt with via the procedure of renormalization. This
technique, successful for λφ4 or gauge theories, fails in the case of gravity. This can be ex-
pected from dimensional arguments, and the rigorous proof that the perturbative quantization
of general relativity fails because of non-renormalizable ultraviolet divergences was obtained
in the late eighties, by Goroff and Sagnotti [3].
This means that the theory constructed as sketched above can be used for low-energy
calculations, but it would be inconsistent if taken seriously at all energy scales. Heuristically,
the inconsistence can be explained as follows. The key field redefinition (1.5) assumes that we
can quantize the field hµν , and that its dynamics takes place on a fixed classical background
ηµν . However, as we increase the energy of the fluctuations its backreaction increases, and
it becomes inconsistent to assume that the background stays fixed, unperturbed. Although
gravitons capture correctly the low-energy physics of the gravitational field, they might not
be the right quantities to describe the quanta of gravity at Planckian energies.
For the reader less familiar with this approach, see [4]. Given the overwhelming success
of the functional integral approach to the quantization of (non-gravitational) field theories,
the canonical formalism is often underappreciated in many courses, and a certain taste of
“old fashion” is associated with it. It was applied to general relativity by Arnowitt, Deser
and Misner (ADM), Dirac, Wheeler and De Witt, among many others. The loop approach is
related to this original idea, which we will now review in some details. For more on it, see
e.g. [5].
20 G RAVITATION QUANTIQUE
X µ + δX µ
Σt+δt
N nµ δt
Xµ Σt
N µ δt
where a = 1, 2, 3 are spatial indices and are contracted with the 3-dimensional metric gab .
Notice that the spatial part gab is not in general the intrinsic metric on Σt . The latter is
given by
qµν = gµν − nµ nν . (1.12)
Tensors on the spatial slide Σ, since their scalar product with n vanishes, can be equivalently
contracted with g or q. The quantity qνµ = g µρ qρν acts as a projector on Σt , allowing us to
define the tensorial calculus on Σt from the one on M. An important quantity is the extrinsic
curvature of Σt , ′ ′
Kµν = qµµ qνν ∇µ′ nν ′ . (1.13)
This tensor is symmetric and it is connected to the Lie derivative of the intrinsic metric
Ln qµν = 2Kµν . It enters the relation between the Riemann tensor of Σt (R) and that of
M (R),
′
Rµνρσ = qµµ′ qνν qρρ qσσ Rνµ′ ρ′ σ′ − Kνσ Kρµ − Kνρ Kσµ
′ ′ ′
(1.14)
This formula, proved in the Appendix and known as the Gauss-Codazzi equation, lets us
rewrite the action (1.2) in the following form, 1
Z Z
√
S = dt d3 x qN R − K 2 + Tr(KK) . (1.15)
Σ
A key consequence of this analysis is the fact that time derivatives of N and N a do not
enter in the Lagrangian L. This implies that N and N a are Lagrange multipliers, with null
conjugate momenta δL/δ Ṅ = δL/δ Ṅ a = 0. The true dynamical variables are the spatial
components qab only, with conjugate momenta
δL √
π ab ≡ = q K ab − Kq ab . (1.16)
δ q̇ab
One can finally evaluate the Legendre transform, which after some calculations gives
Z Z
1
SEH qab , π ab , N, N a = dt d3 x π ab q̇ab − N a Ha − N H , (1.17)
16πG
where (the covariant derivative ∇ contains only spatial indexes)
b
√ π
Ha = −2 q ∇b √a (1.18)
q
and
1 √
H = √ Gabcd π ab π cd − q R, Gabcd = qac qbd + qad qbc − qab qcd . (1.19)
q
1. We are assuming for simplicity of exposition that Σ has no boundary, the case with boundary is treated
extensively in [8] and [9].
22 G RAVITATION QUANTIQUE
The quantity Gabcd is often called the supermetric, or DeWitt metric. The variation of the
action (1.17) with respect to the Lagrange multipliers gives the equations
which are called respectively vector, or space-diffeomorphism constraint, and scalar or Ha-
miltonian constraint. In the following we will also use the compact notation Hµ = (H, Ha ).
Physical configurations, also called on-shell configurations with particle physics jargon, must
satisfy these constraints.
From (1.17) we see that the Hamiltonian of general relativity is
Z
1
H= d3 x N a Ha + N H. (1.21)
16πG
This Hamiltonian is peculiar, since it is proportional to the Lagrange multipliers and thus
vanishes on-shell. Hence, there is no dynamics and no physical evolution in the time t. This
puzzling absence of a physical Hamiltonian is in fact a consequence of what we discussed
earlier : the diffeomorphism-invariance of the theory tells us that t is a mere parameter devoid
of an absolute physical meaning, thus there is no physical dynamics in t. This is the root of
the problem of time in general relativity. For discussions of this problem, see [10] and [11].
From there we can evaluate the following brackets among the constraints (for the explicit
calculation see in the Appendix)
Notice that the right-hand sides vanish on the constraint surface (1.20). This means that the
Poisson flows generated by the constraints preserve the constraint hypersurface. Constraints
with this characteristic are said to be first class, as opposed to second class constraints whose
Poisson brackets do not vanish on-shell. First class constraints generate gauge transforma-
tions on the constraint surface (See e.g. [12] for details).
To see what the gauge transformations look like in our case, consider the smearing of the
constraints
Z Z
H(N~) = H a (x)Na (x)d3 x, H(N ) = H(x)N (x)d3 x. (1.24)
Σ Σ
1 A PATH TO LOOP QUANTUM GRAVITY 23
which means that the vector constraint is the generator of space-diffeomorphism on Σ. The
situation is somewhat subtler for the Hamiltonian constraint. We now have (see Appendix)
We see that outside the constraint surface, the "structure constants" on the right-hand side
contain the field g ab itself. Hence they are not constants at all, and the algebra (1.23) is
not a Lie algebra, unlike Diff(M). Instead, (1.23) shows that when we introduce a foliation
to define the canonical formalism, we still have the symmetry of diffeomorphisms, which
acts changing the foliation, and this new one, which acts deforming it. The two symmetries
coincide on physical configurations.
(12 − 4)·∞3 = 8 ·∞3 . The fact that the algebra of constraints is first class guarantees that the
gauge transformations generated by the constraints preserve the constraint surface. We shall
refer to the trajectories drawn by the gauge transformations as orbits. Points along one orbit
correspond to the same physical configuration, only described in different coordinate systems.
Hence, to select the physical degrees of freedom we have to divide by the gauge orbits, in
a manner identical to what happens in gauge theories. Since the orbits span a dimension 4
manifold at each space point, dividing by the orbits gives (8 − 4) · ∞3 = 4 · ∞3 . This is
the physical phase space. It has four dimensions per space point, namely the theory has two
physical degrees of freedom per space point (or simply two degrees of freedom, for brevity,
with the space dependence tacitly implied), a result consistent with the linearized analysis.
See e.g. [12] or [14] for more details.
qab , π ab = 12 · ∞3
qab , π ab = 8 · ∞3
H = Ha = 0
qab , π ab phys
= 4 · ∞3
This has far as the counting goes. However, in the case of the linearized analysis we are
also able to identify the 2 degrees of freedom as the two helicities, and associate a physical
trajectory to each point in phase space, thanks to the fact that we are able to solve the dyna-
mics. Compare this to the mechanics of a point particle : the dynamics is given by φ[φ0 ,p0 ] (t),
and the degree of freedom corresponds to varying the initial conditions (φ0 , p0 ), spanning in
this way all possible physical evolutions.
Therefore, if we want to know what the two physical degrees of freedom of general rela-
tivity are, we need to control the general solution of the theory. This is a formidable task due
to the high non-linearity of the equations, and in spite of the effort in this direction, still little
is known. See [15] for a review of some attempts.
There is a particular reformulation of the dynamics that plays a role in quantization. Let
us go back to the Gauss-Codazzi equation (1.14). A simple manipulation gives the so-called
first Gauss-Codazzi identity,
where G is the Einstein tensor. One can show (but we will not do it here) that the left hand
side of (1.29) is equivalent to a linear combination of the constraints. This means that if gµν
1 A PATH TO LOOP QUANTUM GRAVITY 25
Ĥ µ ψ = 0 ∀ψ ∈ Hphys . (1.31)
These steps should then be completed with an explicit knowledge of the scalar product in
Hphys and a physical interpretation of the quantum observables.
Dirac’s procedure is more general than gravity : it applies to any fully constrained system.
Let us now try to apply it to the ADM formulation of general relativity. We then look for a
space of functionals carrying a representation of the quantum Poisson algebra
q̂ab (x), π̂ cd (y) = i~δ(ab)
cd
δ 3 (x, y), (1.32)
[q̂ab (x), ĝcd (y)] = 0,
π̂ ab (x), π̂ cd (y) = 0.
Formally, we can proceed by analogy with better known cases, such as a scalar field theory,
and consider a Schrödinger representation q̂ab (x) = qab (x), π̂ ab (x) = −i~ δqabδ (x) , acting on
wave functionals
ψ[qab (x)] (1.33)
of the 3-metric. This procedure, which works well for the scalar field [4], encounters a number
of difficulties when applied to the gravitational context. For instance, to define the (kinemati-
cal) Hilbert space we need a scalar product, formally
Z
dg ψ[g] ψ ′ [g] ≡ hψ | ψ ′ i . (1.34)
Let us ignore these and other similar issues, and try to proceed formally assuming that a
well-defined Hkin exist. The next step is to promote the constraints (1.20) to operators, and
characterize their space of solutions. Let us proceed in two halves,
Ĥ a = 0 Ĥ = 0
Hkin −−−−−−→ HDif f −−−−−−→ Hphys . (1.35)
Consider first the vector constraint. In the Schrödinger representation defined above, the
smeared version gives
Z
δψ
Ĥ(Na )ψ[qab ] = 2i~ d3 x∇b Na = 0, (1.36)
Σ δqab
namely the solution of the vector constraint are those functionals of the metric invariants
under diffeomorphism. This is very nice, as it realizes at the quantum level the correct action
of the classical constraints. However, the space of solutions HDif f is again ill-defined, since
it inherits from the kinematical one the lack of measure theory or other means of control over
it.
For the Hamiltonian constraint we can write
2
~ abcd 1 δ2 p
Ĥψ[qab ] = − G : √ : − detĝR(ĝ) ψ[qab ], (1.37)
2 detĝ δqab (x)δqcd (x)
where the colon : means that an ordering of the operators needs to be prescribed. The situa-
tion is more complicated for the Hamiltonian constraint, since it requires the definition of
products of operators at the same point, notoriously very singular objects. The formal ex-
pression (1.37) is known as the Wheeler-DeWitt equation. Even if manage to give a suitable
ordering prescription and regularize the differential operator, the problem with the equation is
that we do not have any characterization of the solutions, not even formally as for the diffeo-
constraints above. And of course, again no clue on the knowledge of the physical Hilbert
space and scalar product (see however minisuperspace models [16]).
Loop quantum gravity is an approach to the problem which improves significantly the
situation, and gives a number of answers to these open questions. The key to LQG and to
such improvement is surprisingly simple : instead of changing the gravitational theory or the
quantization paradigm, we just use different variables to describe gravity. After all, we are
familiar with the fact that not all choices of fundamentals variables work out as well when
quantizing a classical theory. Consider for instance the harmonic oscillator. Classically, the
most elegant description of the system is in terms of action-angle variables, which parame-
trize the phase space as {φ, J} = 1 instead of {q, p} = 1. However, it is more convenient to
quantize the system using the (q, p) variables than the action-angle ones, which require extra
care in constructing the operator algebra and dealing with unitarity.
We now introduce the variables that allows us to reformulate general relativity in a way
more amenable to Dirac’s quantization procedure.
1 A PATH TO LOOP QUANTUM GRAVITY 27
1.3 Appendix
In this Appendix we report some calculations which are used in the main text.
Proof of the Gauss-Codazzi equation. We start the computation from the characterization of R in terms
˜ defined as the projection of ∇ on Σ via qνµ
of covariant derivatives ∇
′
˜ [µ ∇
2∇ ˜ ν] wρ = −wρ′ Rρρµν (1.38)
′ ′ ′ ′′ ′′
′ ′
˜ µ∇
∇ ˜ ν wρ = ∇
˜ µ qνν qρρ ∇ν ′ wρ′ = qµµ qνν qρρ ∇µ′ qνν′′ q ρ′′ ∇ν ′ wρ′ (1.39)
ρ
Antisymmetrizing in a,b
′ ′ ′ ′ ′ ′ ′
˜ [µ ∇
∇ ˜ ν] wρ = q µ qν]
ν ρ ν
qρ ∇µ′ ∇ν ′ wρ′ + qρρ nν K[µν] ∇ν ′ wρ′ − q[ν Kµ]ρ wρ′ ∇ν ′ nρ = (1.45)
[µ
′ ′ ′ ′
ρ
= qµµ qνν qρρ ∇[µ′ ∇ν ′ ] wρ′ − K[ν Kµ]ρ wρ′ = (1.46)
1 ′ ′ ′ ′′ ρ′
= − qµµ qνν qρρ Rρρ′ µ′ ν ′ wρ′′ − K[ν Kµ]ρ wρ′ (1.47)
2
or in terms of the Riemann tensor using (1.38)
′
′ ′ ′′ ′
ρ′
Rρρµν wρ′ = qµµ qνν qρρ Rρρ′′ µ′ ν ′ + 2K[ν Kµ]ρ wρ′ (1.48)
′
this equation state for w ∈ Σ for make it true for all v ∈ M it is sufficient to take wµ = qµµ vµ′
′ ′ ′ ′
Rσρµν = qµµ qνν qρρ qσσ′ Rρσ′ µ′ ν ′ + 2K[ν
σ
Kµ]ρ (1.49)
Or equivalently
′ ′ ′ ′
Rσρµν = qµµ qνν qρρ qσσ Rσ ′ ρ′ µ′ ν ′ + 2Kσ[ν Kµ]ρ (1.50)
′ ′ ′ ′
R = q σµ q νρ qµµ qνν qρρ qσσ Rσ ′ ρ′ µ′ ν ′ + q σµ q νρ Kσν Kµρ − q σµ q νρ Kσµ Kνρ = (1.51)
σ ′ µ′ ν ′ ρ′ 2
=q q Rσ ′ ρ′ µ′ ν ′ + Tr(KK) − K = (1.52)
ν′ ρ′ σ′ µ′ ν′ ρ′ 2
= R + 2n n Rρ′ ν ′ + n n n n Rσ ′ ρ′ µ′ ν ′ + Tr(KK) − K (1.53)
28 G RAVITATION QUANTIQUE
′ ′ ′ ′
the term nσ nµ nν nρ Rσ ′ ρ′ µ′ ν ′ vanishes because of the symmetries of the Riemann tensor, using
′ ′
the fact that nν nρ gρ′ ν ′ = −1 we can write
′ ′
g ρ′ ν ′
R = 2nν nρ Rρ′ ν ′ − R + Tr(KK) − K 2 (1.54)
2
That is the first Gauss-Codazzi equation.
we want to show that it generates infinitesimal spatial diffeomorphism (it’s sufficient to do it on the
phase space variables)
n o Z
~ ) = d3 xN l δHl (x) = 2∇(a Nb) = L ~ qab
qab (y), H(N (1.57)
N
δπ ab (y)
Note that the functional derivative is easy to compute with a by part integration. The last term vanishes
for metric compatibility and in the Lie derivative you can substitute ∂ with ∇.
n o Z k
~ ) = −N l δHl (x) = −2π ca ∇c N b + √q∇k N
π ab , H(N √ π
ab
=
δqab (y) q
√
q √ π ab √ π ab
= −2π ca ∇c N b √ + q(∇k N k ) √ + qN k ∇k √ =
q q q
√ π ab π ab √
= qLN~ √ + √ (∇k N k ) q =
q q
√ π ab π ab √
= qLN~ √ + √ LN~ q ≡ LN~ π ab
q q
Note that π ab is a tensorial density so its lie derivative has to be defined as above. Given A an arbitrary
function of q and π using Leibniz we can compute the Poisson bracket easily
n o Z δA δH(N ~) δH(N ~ ) δA
~) =
A, H(N − =
δqab δπ ab δqab δπ ab
Z
δA n ~
o n o
~ ) δA =
= qab , H(N) − π ab , H(N
δqab δπ ab
Z
δA δA
= L ~ qab − LN~ π ab ab = LN~ A
δqab N δπ
1 A PATH TO LOOP QUANTUM GRAVITY 29
√ Hb Hb √
LN~ Hb = qLN~ √ + √ LN~ q =
q q
√ Hb
= qN a ∂a √ + Ha ∂b N a + Hb ∇a N a =
q
√ 1
= N a ∂a Hb + N a Hb q∂a √ + Ha ∂b N a + Hb ∇a N a =
q
√ 1
= N a ∂a Hb − N a Hb ∂a ( q) √ + Ha ∂b N a + Hb ∇a N a =
q
Na √
= N a ∂a Hb + Ha ∂b N a + Hb (∇a N a − √ ∂a q) =
q
= N a ∂a Hb + Ha ∂b N a + Hb ∂a N a
In fact
∇a N a = ∂a N a + Γaba N b
Na √ 1 a ab a b
√ ∂a q = N q ∂a qab = Γba N
q 2
Where the last equality is due to metric compatibility. We can conclude that
Z
N a Ha , Hb = −N a ∂a Hb − Ha ∂b N a − Hb ∂a N a (1.58)
~ ) is the generator
Moreover also the second Poisson bracket is easy to compute using the fact that H(N
~.
of infinitesimal diffeomorphism along N
√ H H √
LN~ H = qLN~ √ + √ LN~ q = (1.60)
q q
√ H H 1 √ cd
= qN a ∂a √ + √ ( qq LN~ qcd ) =
q q 2
√
√ 1 H 2 q cd
= N a ∂a H + qN a H∂a √ + √ q qca ∇d N a =
q q 2
√ −1 1 cd
= N a ∂a H + H∂a N a + qN a H( ) √ q ∂a qcd + HΓaba N b =
2 q
√ −1 1 cd
= N a ∂a H + H∂a N a + qN a H( ) √ q ∇a qcd = N a ∂a H + H∂a N a
2 q | {z }
=0
that is equivalent to
{Ha (x), H(y)} = H∂a δ(x − y) (1.62)
30 G RAVITATION QUANTIQUE
To ultimate the computation of the second Poisson bracket we need some preliminary results. Given
an arbitrary function f (π, h)
nn o o
f, H(N ~ 1 ) , H(N
~ 2) = L ~ L ~ f (1.63)
N1 N2
Z Z Z
N1a Ha , N2a Ha = [N1 , N2 ]a Ha (1.65)
Finally we compute the last Poisson bracket, if N 6= N ′ (if they are equal it vanish)
Z
δH(N ) δH(N ′ ) δH(N ′ ) δH(N )
H(N ), H(N ′ ) = d3 x ab
− (1.66)
δqab (x) δπ (x) δqab (x) δπ ab (x)
We notice that all the algebraical terms in q and π simplify each other. In fact
δH(x)
= f h (x)ab δ(x − y) (1.67)
δqab (y) algebraic
part
δH(x)
= fπ (x)ab δ(x − y) (1.68)
δπab (y) algebraic
part
Z
H(N )|algebraic , H(N ′ )algebraic = d3 xN fab
h
N ′ fπab − N ′ fπab N fab
h
=0 (1.69)
part part
The non algebraic part in q are in R conversely the Hamiltonian constraint is algebraic in π.
non non
Z
3
δH(N )algebraic δH(N ′ ) δH(N ′ )algebraic δH(N )
H(N ), H(N ′ ) = d x term
− term
(1.70)
δqab (x) δπ ab (x) δqab (x) δπ ab (x)
Substituting this variation in the Poisson bracket the first term becomes
Z R R Z
δ N H δ N ′H √ a b 2 ab
′ πab πqab
= − q ∇ ∇ N − ∇ (N q ) 2N √ − √ =
δqab δπ ab q q(3 − 1)
Z
π q ab qab ′
= −2 (∇a ∇b N )πab N ′ − (∇2 N )πN ′ − (∇2 N )N ′ + (π∇2 N ) N =
3−1 3−1
Z
3 1
= 2 (∇a ∇b N )πab N ′ + − − 1 πN ′ ∇2 N =
3−1 3−1
Z Z
√ a b πab ′ √ b a πab N ′
= −2 q(∇ ∇ N ) √ N = 2 q∇ N ∇ √ (1.71)
q q
2 Tetrad formulation
A tetrad is a quadruple of 1-forms, eIµ (x), I = 0, 1, 2, 3 such that
By its definition, it provides a local isomorphism between a general reference frame and an
inertial one, characterized by the flat metric ηIJ . A local inertial frame is defined up to a
Lorentz transformation, and in fact notice that the definition is invariant under
This means that the “internal” index I carries a representation of the Lorentz group. Contrac-
ting vectors and tensors in spacetime with the tetrad, we get objects that transform under the
Lorentz group, e.g. eIµ nµS= nI . The tetrad thus provides an isomorphism between the tangent
bundle of M, T (M) = ρ (ρ, Tρ (M)), and a Lorentz principal bundle F = (M, SO(3, 1)).
On this bundle we have a connection ωµIJ , that is a 1-form with values in the Lorentz algebra,
which we can use to define covariant differentiation of the fibres,
This is the analogue of the covariant derivative ∇µ = ∂µ + Γµ for vectors in T (M). We can
also define the derivative for objects with both indices, such as the tetrad,
where we separed the spacetime indices into their symmetric and antisymmetric combina-
tions, and used the fact that the Levi-Civita connection Γ(g) has no antisymmetric part.
From these equations we immediately obtain the following relation between the spin and
Levi-Civita connections,
I
ωµJ = eIν ∇µ eνJ , (2.6)
as well as the fact that the spin connection satisfies
dω eI = deI + ω I J ∧ eJ = ∂µ eIν + ωµJ
I
eJν dxµ ∧ dxν = 0. (2.7)
This equation is known as Cartan’s first structure equation. Here we introduced the exterior
calculus of forms, with d the exterior derivative, dω the covariant exterior derivative, and ∧
the wedge product. See [17] for an introduction to this formalism.
Given the connection, we define its curvature
F IJ = dω IJ + ω I K ∧ ω KJ , (2.8)
where Rµνρσ (e) is the Riemann tensor constructed out of (the metric defined by) the tetrad
eIµ . This relation is known as Cartan second structure equation. It shows that general relativity
is a gauge theory whose local gauge group is the Lorentz group, and the Riemann tensor is
nothing but the field-strength of the spin connection.
Cartan second structure equation. Starting from the definition of F in (2.9) and inserting (2.6), we have
IJ
Fµν =∂µ eIρ ∂ν eρJ + ∂µ eIρ Γρσν eσJ + eIρ ∂µ (Γρσν ) eσJ + eIρ Γρσν ∂µ eσJ + eIρ ∂µ eρK eK
σ ∂ν e
σJ
+
+ eIρ Γρδµ eδK eK
σ ∂ν e
σJ
+ eIρ ∂µ eρK eK σ δJ
σ Γδν e + eIρ Γρδµ eδK eK σ ηJ
σ Γην e − (µ ↔ ν).
Relation between the determinants. The determinant g of gµν is related to the determinant of the tetrad
e by the simple relation
g = −e2 . (2.11)
2 T ETRAD FORMULATION 33
This expression can be easily derived recalling Caley’s formula for the determinant of a matrix,
1 µνρσ αβγδ
g = detgµν = ε ε gµα gνβ gργ gσδ . (2.12)
4!
If we substitute the expression of gµν in terms of tetrads we get
1
g = det eIµ eJν ηIJ = εµνρσ εαβγδ eIµ eJα ηIJ eK L M N O P
ν eβ ηKL eρ eγ ηM N eσ eδ ηOP =
4!
1
= εµνρσ eIµ eK M O αβγδ J L N P
ν eρ eσ ε eα eβ eγ eδ ηIJ ηKL ηM N ηOP =
4!
1
= e2 εIKM O εJ LNP ηIJ ηKL ηM N ηOP = −e2
4!
Rewriting the actions in terms of tetrads. Explicitly using (2.10) and (2.11)
Z Z
√
SEH (gµν (e)) = d4 x −g g µν Rµν = d4 x e eµI eνI Rµρνσ eρJ eσJ =
Z Z
1
= d4 xeeµI eρJ Fµρ IJ
(ω(e)) = d4 x εIJ KL εµραβ eK L IJ
α eβ Fµρ (ω(e))
4
Z
1
= εIJ KL eI ∧ eJ ∧ F KL (ω(e))
2
A fact which plays an important role in the following is that we can lift the connection to be
an independent variable, and consider the new action
Z
1
S(eIµ , ωµIJ ) = εIJKL eI ∧ eJ ∧ F KL (ω) . (2.14)
2
Although it depends on extra fields, this action remarkably gives the same equations of motion
as the Einstein-Hilbert one (2.13). This happens because the extra field equations coming
from varying the action with respect to ω do not add anything new : they simply impose the
form (2.6) of the spin connection, and general relativity is thus recovered.
Deriving the fields equations. One first verifies the Palatini identity
δω F KL (ω) = dω δω KL . (2.15)
34 G RAVITATION QUANTIQUE
As it gives the same field equations, (2.15) can be used as the action of general relativity.
Notice that only first derivatives appears, thus it provides a first order formulation of gene-
ral relativity. Furthermore, the action is polynomial in the fields, a desiderable property for
quantization. On the other hand, there are two non-trivial aspects to take into account :
– The equivalence with general relativity holds only if the tetrad is non-degenerate, i.e.
invertible. On the other hand, (2.15) is also defined for degenerate tetrads, since inverse
tetrads never appear. Compare the situation with the Einstein-Hilbert action, where the
inverse metric appears explicitly. Hence, the use of (2.15) leads naturally to an exten-
sion of general relativity where a sector with degenerate tetrads, and thus degenerate
metrics, exists.
– If we insist on the connection being an independent variable, there exists a second term
that we can add to the Lagrangian that is compatible with all the symmetries and has
mass dimension 4 :
δIJKL eI ∧ eJ ∧ F KL (ω), (2.18)
where δIJKL ≡ δI[K δL]J . This term is not present in the ordinary second order metric,
since when (2.6) holds,
δIJKL eI ∧ eJ ∧ F KL (ω(e)) = εµνρσ Rµνρσ (e) ≡ 0. (2.19)
Adding this second term to (2.15) with a coupling constant 1/γ leads to the so-called Holst
action [18] Z
1 1
S (e, ω) = εIJKL + δIJKL eI ∧ eJ ∧ F KL (ω) . (2.20)
2 γ
Assuming non-degenerate tetrads, this action leads to the same field equations of general
relativity,
ωµIJ = eIν ∇µ eJν , Gµν (e) = 0. (2.21)
This result is completely independent of the value of γ, which is thus a parameter irrelevant in
classical vacuum general relativity. It will however turn out to play a key role in the quantum
theory, where it is known as the Immirzi parameter. 2
Deriving the fields equations for the Holst action. The vanishing of the variations give
1 1
εIJ KL + δIJ KL eI ∧ dω eJ = 0, (2.22)
2 γ
1 1
εIJ KL + δIJ KL eJ ∧ F KL (ω) = 0. (2.23)
2 γ
2. The Immirzi parameter becomes relevant also at the classical level if source of torsion are present [19, 20].
2 T ETRAD FORMULATION 35
For invertible tetrads, the first one is again uniquely solved by (2.6). Substituting this solution into the
second equation we get
1 1
eIν Gα
ν + δIJ KL εµσδα eJµ Fσδ
KL
= eIν Gαν + δIJ KL εµσδα eJµ eKν eLρ Rσδνρ =
γ γ
1 µσδα Iν ρ 1 σδµα Iν
= eIν Gα
ν + ε e δµ − eIρ δµν Rσδνρ = eIν Gα ν − ε e Rσδµν = eIν Gα
ν = 0,
2γ γ
which is equivalent to Gµν = 0 thanks to the non-degeneracy of the tetrad. In the last step we used the
first Bianchi identity εσδµα Rσδµν = 0.
The “triad” eia is the spatial part of the tetrad. As before, we want to identify canonically
conjugated variables and perform the Legendre transform, but we now have two new features
which complicate the analysis. The first one is the tetrad formulation, which in particular has
introduced a new symmetry in the action : the invariance under local Lorentz transformations.
As a consequence, we expect more constraints to appear, corresponding to the generators
of the new local symmetry. The second one is the use of the tetrad and the connection as
independent fields. Therefore, the conjugate variables are now functions of both eIa and ωaIJ
(and their time derivatives), as opposed to be functions of the metric gab only.
The consequence of these novelties is a much more complicated structure than in the
metric case. In particular, the constraint algebra is second class. However, there is a particular
choice of variables which simplifies the analysis, making it possible to implement a part of
the constraint and reducing the remaining ones to first class again. These are the famous
Ashtekar variables, which we now introduce. 3
To simplify the discussion, it is customary to work in the “time gauge” eIµ nµ = δ0I , where
e0µ = (N, 0) −→ eI0 = N, N a eia . (2.25)
The crucial change of variables is the following : we define the densitized triad
1
Eia = eeai = εijk εabc ejb ekc , (2.26)
2
and the Ashtekar-Barbero connection
1
Aia = γωa0i + εijk ωajk . (2.27)
2
3. On the general analysis with the second class constraints see [21].
36 G RAVITATION QUANTIQUE
These variables turns out to be conjugated. In fact, we can rewrite the action (2.20) in terms
of the new variables as [21, 13]
Z Z h i
1
S(A, E, N, N a ) = dt d3 x Ȧia Eia − Ai0 Da Eia − N H − N a Ha , (2.28)
γ Σ
where
Gj ≡ Da Eia = ∂a Eja + εjkℓ Aja E aℓ , (2.29)
2
1 j b 1+γ
Ha = F E − Kai Gi , (2.30)
γ ab j γ
h i ε EaEb 1 + γ2 i Ea
j jkℓ k ℓ
H = Fab − γ 2 + 1 εjmn Kam Kbn + G ∂a i . (2.31)
detE γ detE
The resulting action is similar to (1.17), with (A, E) as canonically conjugated variables, as
opposed to (q, π). Lapse and shift are still Lagrange multipliers, and consistently we still
refer to H(A, E) and Ha (A, E) as the Hamiltonian and space-diffeomorphism constraints.
The algebra is still first class. The new formulation in terms of tetrads has introduced the
extra constraint (2.29). The reader familiar with gauge theories will recognize it as the Gauss
constraint. Just as the H µ constraints generate diffeomorphisms, the Gauss constraint gene-
rates gauge transformations. It is in fact easy to check that Ejb and Aia transform respectively
as an SU(2) vector and as an SU(2) connection under this transformation.
The algebra generated by the Gauss constraint. We define the smearing of the Gauss constraint
Z
G (Λ) = d3 xGi (x)Λi (x). (2.32)
Z Z n o
3
d xΛ j
(x)Gj (x), Aia (y) = d3 xΛj (x) ∂b Ejb + εmjn Am bn i
b E , Aa (y) =
Hence,
n o
G(Λ), ∂a Eia (y) + εjik Aja E ak (y) =
= γεijn ∂a Λj (y)E an (y) − γεjik Aja (y)εkℓn Λℓ (y)E an (y)+
+ γεjik E ak ∂a Λj (y) + εmℓj Λℓ (y)Am a (y) =
= γεijn Λj (y)∂a E an (y) + γ (εijk εℓnk − εink εℓjk ) Λℓ (y)E an (y)Aja (y) =
= γεiℓk Λℓ (y)∂a E ak (y) + γεiℓk Λℓ (y)εjkn E an (y)Aja (y) = γεiℓk Λℓ (y)Gk (y).
2 T ETRAD FORMULATION 37
Finally, smearing also the second term in the above bracket we get
γ
{G(Λ1 ), G(Λ2 )} = G ([Λ1 , Λ2 ]) , (2.33)
2
which we recognize as the su(2) algebra structure equations.
We should not be surprised of the appearance of this extra constraint. When we use the
tetrad formalism, we introduce a new symmetry in the theory, the invariance under local
gauge transformation. The Gauss constraint is there to enforce this invariance at the canonical
level. On the other hand, it might be more puzzling that although the local gauge invariance
of the covariant action was the full Lorentz group, the Legendre transform (2.28) is only
invariant under SU(2).
The origin of this puzzle lies precisely in the change of variables (2.26-2.27). The Ashtekar-
Barbero connection is an SU (2) connection, not a Lorentz connection. In particular, it is not
the pull-back of the space-time Lorentz connection [22]. It should then be seen as an auxiliary
variable, useful to recast the algebra in a first class form.
Summarizing, in this formulation of General Relativity the theory is described by an
extended phase space of dimension 18 · ∞3 with the fundamental Poisson bracket
i
Aa (x), Ejb (y) = γδab δji δ 3 (x, y) (2.34)
This new internal index i corresponds to the adjoint representation of SU(2), and we can
recover the old 12 · ∞3 dimensional phase space on the constraint surface Gi = 0 dividing by
gauge orbits generated by G. This group SU(2) should be seen as an auxiliary local symmetry
group, since the connection with the original Lorentz group of the tetrad formulation is hidden
in the change of variables (2.26-2.27). It is only for the special case γ = i (which is the
original one introduced by Ashtekar) that the relation is manifest : in that case, the SU(2)
corresponds to the self-dual subgroup of the Lorentz group. Notice that this case also leads to
a simplification of the Hamiltonian constraint (2.31). On the other hand, the variables are now
complex, and to recover general relativity one needs to impose reality conditions. These are
particularly difficult to deal with at the quantum level, and for this reason most developments
in LQG have focused on γ real. This is what we do also in this review.
On the definition of the holonomy. More precisely, we call holonomy the solution of the differential
equation
d
hγ (t) − hγ (t)A(γ(t)) = 0, hγ (0) = 1. (2.39)
dt
If we integrate the equation by iteration we have
Z t
hγ (t) = 1 + A(γ(s))hγ (s)ds = (2.40)
0
Z t Z tZ 1
=1+ A(γ(s1 ))hγ (s1 )ds1 + A(γ(s1 ))A(γ(s2 ))hγ (s2 )ds1 ds2 =
0 0 s1
= ...
therefore formally
∞
X ZZZ
hγ (t) = A(γ(s1 )) · · · A(γ(sn ))ds1 · · · dsn . (2.41)
n=0
t>s1 >···>sn >0
To complete the proof, one needs to show that the series is well defined. Indeed, it converges with
respect to sup norm
∞ ZZZ ∞
X X tn
||hγ || ≤ ||A(γ(s1 )) · · · A(γ(sn ))|| ds1 · · · dsn ≤ ||A||n . (2.42)
n=0 n=0
n!
t>s1 >···>sn >0
For further reference, let us also notice that the terms of the series can be written as integrals over square
domains (s1 , . . . , sn ) ∈ [0, t]n , instead triangle domains t > s1 > · · · > sn > 0. This gives
I X ∞ ZZZ
1
hγ (t) = P exp A = P (A(γ(s1 )) · · · A(γ(sn )) . (2.43)
γ n=0
n!
2 T ETRAD FORMULATION 39
where s(γ) and t(γ) are respectively the source and target points of the line γ.
– Under the action of diffeomorphism, the holonomy transforms as
Basic proofs. If not otherwise specified, we assume for simplicity that the source is t0 = 0. For two
composable paths α and β, we define the composition
β(t), if t ∈ [0, S]
βα = (2.48)
α(t − S), if t ∈ [S, T + S]
R S+T RS R S+T
If we split the integrals as 0
= 0
+ S
, (2.43) reads
ZZZ n
!Z Z Z Z
S S T T
X n
= dt1 · · · dti dti+1 · · · dtn . (2.49)
i=0
i 0 0 0 0
P (A(βα(t1 )) · · · A(βα(tn ))) = P (A(β(ti+1 )) · · · A(β(tn ))) P (A(α(t1 )) · · · A(α(ti ))) . (2.50)
= hβ hα ,
2.4 Summary
In this section we have taken the two key steps needed to prepare general relativity for the
loop quantization. The first step was to reformulate the theory in terms of the tetrad field and
an SU(2) independent connection, the Ashtekar-Barbero connection (2.27). The second step
was to regularize the resulting Poisson algebra using paths and surfaces, instead of the all of
space as in traditional smearings (cf. what was done in the ADM formulation). The resulting
smeared algebra of hγ [A] and Ei (S) is called holonomy-flux algebra. It provides the most
natural regular (i.e. no delta functions appear) version of the Poisson algebra (2.34).
The difference with a gauge theory is of course in the dynamics : In gauge theory, after
imposing the Gauss law, we have a physical Hamiltonian. Here instead we still have a fully
constrained system.
Let us first briefly review some basic steps in the quantization of gauge theories. In a
nutshell, the usual procedure goes along these lines :
– Use the Minkowski metric to define a Gaussian measure δA on the space of connections
modulo gauge-transformations.
– Consider the Hilbert space L2 (A, δA) ∋ ψ[A] and define the Schrödinger representa-
tion
– Impose the Gauss law constraint, Ĝi ψ[A] = 0, which selects the gauge-invariant states.
– Study the dynamics with the physical Hamiltonian,
LA in plane waves
In practice, the easiest way to deal with the dynamics is to decompose
and the Hilbert space into a Fock space, L2 (A, δA) = HF ock = n Hn , the direct
sum over n-particles spaces.
The key difference in the case of general relativity is that we do not have a background
metric at disposal to define the integration measure, since now the metric is a fully dynamical
quantity. Hence, we need to define a measure on the space of connections without relying on
any fixed background metric. The key to do this is the notion of cylindrical functions, which
we introduce next.
Commutator.
h i δ δ
Âia (x), Êjb (y) ψ[A] = −i~γAia (x) ψ[A] + i~γ j Aia (x)ψ[A]
δAjb (y) δAb (y)
δAia (x)
= i~γ ψ[A] = i~γδab δji δ 3 (x, y)ψ[A]
δAjb (y)
defined as a collection of oriented paths e ⊂ Σ (we will call these paths the links of the
graph) meeting at most at their endpoints. Given a graph Γ ⊂ Σ we denote by L the total
number of links that it contains. A cylindrical function is a couple (Γ, f ) of a graph and a a
smooth function f : SU (2)L −→ C, and it is given by a functional of the connection defined
as
hA | Γ, f i = ψ(Γ,f ) [A] = f (he1 [A], . . . , heL [A]) ∈ CylΓ (3.4)
This space of functionals can be turned into an Hilbert space if we equip it with a scalar
product. The switch from the connection to the holonomy is crucial in this respect, because
the holonomy is an element of SU(2), and the integration over SU(2) is well-defined. In parti-
cular, there is a unique gauge-invariant and normalized measure dh, called the Haar measure.
Using L copies of the Haar measure, we define on CylΓ the following scalar product,
Z Y
ψ(Γ,f ) | ψ(Γ,f ′ ) ≡ dhe f (he1 [A], . . . , heL [A])f ′ (he1 [A], . . . , heL [A]). (3.5)
e
Hkin = ⊕ HΓ . (3.6)
Γ⊂Σ
The scalar product on Hkin is easily induced from (3.5) in the following manner : if ψ and
ψ ′ share the same graph, then (3.5) immediately applies. If they have different graphs, say Γ1
and Γ2 , we consider a further graph Γ3 ≡ Γ1 ∪ Γ2 , we extend f1 and f2 trivially on Γ3 , and
define the scalar product as (3.5) on Γ3 :
ψ(Γ1 ,f1 ) | ψ(Γ2 ,f2 ) ≡ ψ(Γ1 ∪Γ2 ,f1 ) | ψ(Γ1 ∪Γ2 ,f2 ) . (3.7)
The key result, due to Ashtekar and Lewandowski, is that (3.6) defines an Hilbert space over
(suitably generalized, see [25] for details) gauge connections A on Σ, i.e.
The integration measure dµAL over the space of connections is called the Ashtekar-Lewandowski
measure. What (3.8) means is that (3.7) can be seen as a scalar product between cylindrical
functionals of the connection with respect to the Ashtekar-Lewandowski measure :
Z
dµAL ψ(Γ1 ,f1 ) (A)ψ(Γ2 ,f2 ) (A) ≡ ψ(Γ1 ,f1 ) | ψ(Γ2 ,f2 ) . (3.9)
Now that we have a candidate kinematical Hilbert space which does not require a back-
ground metric, let us look for a representation of the holonomy-flux algebra on it. To that
end, it is convenient to introduce an orthogonal basis in the space. This can be done easily
thanks to the Peter-Weyl theorem. It states that a basis on the Hilbert space L2 (G, dµHaar )
of functions on a compact group G is given by the matrix elements of the unitary irreducible
representation of the group. For the case of SU (2),
X j = 0, 12 , 1, . . .
f (g) = fˆmn
j (j)
Dmn (g) (3.10)
m = −j, . . . , j
j
(j)
where the Wigner matrices Dmn (g) give the spin-j irreducible matrix representation of the
group element g. This immediately applies to HΓ , since the latter is just a tensor product of
L2 (SU (2), dµHaar ). That is, the basis elements are
(j1 ) (jn )
hA | Γ; je , me , ne i ≡ Dm 1 n1
(he1 ) . . . Dm n nn
(hen ), (3.11)
On this basis, we can give a Schrödinger representation like (3.1) for the regularized
holonomy-flux version of the algebra. Consider for simplicity the fundamental representation,
1
he ≡ D( 2 ) (he ). The holonomy acts by multiplication,
Here e1 and e2 are the two new edges defined by the point at which the triad acts and the sign
depends on the relative orientation of e and S. The action vanishes, Ê[S]he1 [A] = 0, when e
is tangential to S or e ∩ S = 0.
Consider now the action of the scalar product of two fluxes acting inside the link,
Êi (S)Ê i (S)he [A] = −~2 γ 2 he1 [A]τ i τi he2 [A]. (3.17)
On the right hand side, we see the appearance of the scalar contraction of algebra generators,
τ i τi ≡ C 2 . This scalar product is known as the Casimir operator of the algebra. In the
fundamental representation considered here, C 2 = − 43 12 . The Casimir clearly commutes
with all group elements, thus (3.17) can be written as
Êi (S)Ê i (S)he [A] = −~2 C 2 γ 2 he1 [A]he2 [A] = −~2 C 2 γhe [A]. (3.18)
This expression will be useful below.
On the other hand, if two consecutive fluxes act on one endpoint, say the target, we get
Êi (S)Êj (S)he [A] = −~2 γ 2 he [A]τi τj . (3.19)
From this result we immediately find that two flux operators do not commute,
[Êi (S), Êj (S)]he [A] = −~2 γ 2 he [A][τi , τj ] = −~2 γ 2 ǫij k he [A]τk . (3.20)
The actions (3.13) of the holonomy-flux algebra trivially extends to a generic basis ele-
ment D(j) (h). The action (3.13a) is unchanged, and in the right hand side of (3.13b) one
simply has to replace τi by the generator Ji in the arbitrary irreducible j. Consequently, in
(3.18) we have the Casimir Cj2 = −j(j + 1)12j+1 on a generic irreducible representation,
Ĝ = 0 Ĥ a = 0 Ĥ = 0
Hkin −−i−−−→ Hkin
0
−−−−−−→ HDif f −−−−−−→ Hphys . (3.22)
3 L OOP QUANTUM GRAVITY : K INEMATICS 45
Example : The theta graph. To give a constructive example, let us consider the following graph,
2
Γ
j1
j2
j3
1
A generic cylindrical function can be expressed in terms of the orthonormal basis using (3.12). Since
the gauge transformations act only on the group elements, the gauge-invariant part is obtained looking
at the gauge-invariant part of the product of Wigner matrices,
X h i
finv (h1 , . . . , h3 ) = fˆm
j1 ,j2 ,j3
1 ,m 2 ,m ,n ,n
3 1 2 3,n Dm
(j1 )
n
1 1
(h1 )Dm
(j2 )
n
2 2
(h2 )Dm
(j3 )
n
3 3
(h3 ) .
inv
jl ,ml ,nl
This projector can be written in terms of normalized Clebsch-Gordan coefficients, or Wigner’s 3j-m
symbols, as
Z
(j1 ) (j2 ) (j3 ) j1 j2 j3 j1 j2 j3
dg1 Dm 1 α1
(g1 )Dm 2 α2
(g1 )Dm 3 α3
(g1 ) = . (3.27)
m1 m2 m3 α1 α2 α3
where in is a short-hand notation for the 3j-m symbols. Notice that these are the invariant tensors in the
space of ⊗ je of all the spins that enters in the node n. Finally, we have
e∈n
XY Y X j1 j2 j3 j1 j2 j3
finv = D (je )
(he ) in fˆm
j1 ,j2 ,j3
1 ,m2 ,m3 ,n1 ,n2 ,n3
m1 m2 m3 n1 n2 n3
je e n me ne
X Y Y
= fˆj1 ,j2 ,j3 D (je )
(he ) in , (3.28)
je e n
with the new coefficients fˆj1 ,j2 ,j3 including the sums over the magnetic numbers me , ne .
The group averaging amounts to inserting on each node n the following projector,
Z Y
P = dg D(je ) (g). (3.29)
e∈n
Here the integrand is an element in the tensor product of SU (2) irreducible representations,
Y O
(je )
Dm e ne
(he ) ∈ V (je ) . (3.30)
e e
N
Then, the integration in (3.29) selects the gauge invariant part of e V (je ) , namely the singlet
space V (0) , if the latter exists. Since P is a projector, we can decompose it in terms of a basis
3 L OOP QUANTUM GRAVITY : K INEMATICS 47
of V (0) . Denoting iα a vector (“ket”) in this basis, α = 1, . . . , dimV (0) , and i∗α the dual
(“bra”),
(0)
dimV
X
P= iα i∗α . (3.32)
α=1
These invariants are called intertwiners. For the case of a 3-valent node as in the above
example, dimV (0) = 1 and the unique intertwiner i is given by Wigner’s 3j-m symbols
(cf. (3.27)). More precisely in the case of a three-valent node the space
h i
V (j1 ) ⊗ V (j2 ) ⊗ V (j3 )
inv
|j2 − j3 | ≤ j1 ≤ j2 + j3 . (3.33)
For an n-valent node, the space V (0) can have a larger dimension. To visualize the intertwi-
ners, it is convenient to add first two irreps only, then the third, and so on. This gives rise to a
decomposition over virtual links, which for n = 4 and n = 5 looks as follows :
j3
j1 j3 j1 j4
k k1 k2
j2 j4 j2 j5
Y Z Y Z Y
P D(je ) (g1 ) = dg2 D(je ) (g2 )D(je ) (g1 ) = dg2 D(je ) (g2 g1 ) = P (3.37)
e∈n e∈n e∈n
The facts that P acts only on the nodes of the graph that label the basis of Hkin and equa-
tion (3.32) implies that the result of the action of P on elements of Hkin can be written as
a linear combination of products of representation matrices D(j) (he ) contracted with intert-
winers, generalizing the result (3.28). The states labeled with a graph Γ, with an irreducible
representation D(j) (h) of spin-j of the holonomy h along each link, and with an element i of
48 G RAVITATION QUANTIQUE
the intertwiner space Hn ≡ Inv[ ⊗ V (je ) ] in each node, are called spin network states, and
e∈n
are given by
ψ(Γ,je ,in ) [he ] = ⊗D(je ) (he ) ⊗ in . (3.38)
e n
Here the indices of the matrices and of the interwiners are hidden for simplicity of notation.
Their contraction pattern can be easily reconstructed from the connectivity of the graph.
j2
j3
i1
j4
j1
j5
i2
To complete the discussion, let us show that imposing the gauge-invariance amounts to
solving the Gauss law constraint, Ĝi ψ = 0. Consider a gauge invariant node n, and a surface
S centered in n of radius ǫ. The action of the total flux operator through S on n vanishes
identically :
lim Ê(S) |ni = 0. (3.39)
ǫ→0
In fact, using (3.13b) at each link one notices that (3.39) produces the infinitesimal gauge
transformation gα = 1 − αi τ i ∈ SU (2) at the node, and because the node is gauge invariant
such action vanishes.
Summarizing, spin network states (3.38) form a complete basis of the Hilbert space of
0
solutions of the quantum Gauss law, Hkin . The structure of this space is nicely organized by
the spin networks basis. As before, different graphs Γ select different orthogonal subspaces,
0
thus Hkin decomposes as a direct sum over spaces on a fixed graph,
0
Hkin = ⊕ HΓ0 . (3.40)
Γ⊂Σ
Furthermore, the Hilbert space on a fixed graph decomposes as a sum over intertwiner spaces,
These two equations are the analogue in loop gravity of the Fock decomposition of the Hilbert
space of a free field in Minkowski spacetime into a direct sum of n-particle states, and play
an equally important fundamental role.
3 L OOP QUANTUM GRAVITY : K INEMATICS 49
The area at classical level. We start from the standard definition of area in terms of the metric,
Z s
∂xa ∂xb
A(S) = dσ1 dσ2 det gab α α, β = 1, 2 (3.43)
S ∂σ ∂σ β
∂xa
using the notation ∂1 xa = ∂σ 1
∂xa ∂xb h
a b c d a b c d
i
det gab α = g ab g cd ∂ 1 x ∂ 1 x ∂ 2 x ∂ 2 x − ∂ 1 x ∂ 2 x ∂ 1 x ∂ 2 x = (3.44)
∂σ ∂σ β
= gab gcd 2∂1 xa ∂1 x[b ∂2 xc] ∂2 xd = 2ga[b gc]d ∂1 xa ∂1 xb ∂2 xc ∂2 xd = (3.45)
ef
= gg ne nf (3.46)
1 ∂xa ∂xb
ga[b gc]d = εace εbdf gg ef ne = εeab (3.47)
2 ∂σ 1 ∂σ 2
Using the definition of tetrad and densitized tetrad we have
Z q Z q
A(S) = dσ1 dσ2 e2 eei ef i ne nf = dσ1 dσ2 Eie E f i ne nf (3.48)
S S
At the quantum level, we know from (3.13b) that the triad operators act as functional
derivatives. The action of the scalar product of two triads operator was also studied, see
equation (3.21), for the case of a surface intersected only once by the holonomy path. The
case of a generic graph can be easily dealt with if we regularize the expression for the area in
the following way. We introduce a decomposition of S in N two-dimensional cells, and write
the integral as the limit of a Riemann sum,
Here N is the number of cells, and Ei (SI ) is the flux of Ei through the I-th cell.
50 G RAVITATION QUANTIQUE
where in AN (S) we simply replace the classical flux Ei (SI ) by the operator Êi (SI ). This
operator now acts on a generic spin network state ψΓ , where the graph Γ is generic and
can intersect S many times. We already know that Êi (SI )Ê i (SI ) gives zero if SI is not
intersected by any link of the graph. Therefore once the decomposition is sufficiently fine
so that each surface SI is punctured once and only once, taking a further refinement has
no consequences. Therefore, the limit amounts to simply sum the contributions of the finite
number of punctures p of S caused by the links of Γ. That is,
N q
X X q
Â(S) ψΓ = lim Êi (SI )Ê i (SI ) ψΓ = ~ γ 2 jp (jp + 1) ψΓ . (3.55)
N →∞
I=1 p∈S∪Γ
There are two key remarks to make to this formula : first of all, the spectrum of the area
operator is completely known 4 and quantized : the area can only take up discrete values,
with minimal excitation being proportional to the squared Planck length ℓ2P = ~G, restoring
Newton’s constant. This result can be compared with other celebrated quantizations, such as
the radii of electron’s orbitals in atoms.
Second, the operator has a diagonal action on spin networks. Therefore, spin network
states are eigenstates of the area operator.
4. In this expression, we assumed that each puncture is caused by a link crossing the surface. However, it could
also happen that S is puncture by a node of the graph. A closed expression for the area spectrum is known also in
this general case. See the literature for details [11, 13, 27].
3 L OOP QUANTUM GRAVITY : K INEMATICS 51
where the quantity in absolute value can be recognized as the determinant of the densitised
triad, det Eia .
Two distinct mathematically well-defined volume operators have been proposed in the
literature. One is due to Rovelli and Smolin, and the other to Ashtekar and Lewandowski. We
will refer to them as RS and AL respectively. Both of them act non-trivially only at the nodes
of a spin network state. Let us begin reviewing the construction by RS.
As we did for the area, we replace the integral over R by the limit of a Riemann sum.
R we consider a partition of the region in cubic cellsPCI so that R ⊂ ∪I CI , and the
Specifically,
integral R d3 x can be approximated from above by the sum I volume(CI ). This partition
allows us to rewrite (3.56) in terms of fluxes. In fact, consider the following integral,
Z Z Z
1
WI ≡ 2
d σ1 2
d σ2 d2 σ3 εijk Eia (σ1 )na (σ1 )Ejb (σ2 )nb (σ2 )Ekc (σ3 )nc (σ3 ) .
48 ∂CI ∂CI ∂CI
In the continuum limit where we send the size of the cell ε 7→ 0 and we shrink the cell to a
point x, we obtain
1 abc
WI = ε na nb nc det Eia (x) ε6 ≃ det Eia (x) ε6 ≃ volume2 (CI ).
48
Hence, we have Xp
V (R) = lim WI . (3.57)
ε7→0
I
For the sake of notation, let us subdivide each ∂CI into surfaces S α such that ∂CI =
∪α SIα . Then, we can write WI as a sum of fluxes over three surfaces, and
s
X 1 X β γ
V (R) = lim εijk Ei (SIα )Ej (SI )Ek (SI ). (3.58)
ε7→0 48
I α,β,γ
means that the volume does not act on links, since if no node is present, two of the SIα have to
be the same. We thus obtain the important result that the volume operator acts only on nodes
of the graph.
Let us now focus on a single node, i.e. the I-th contribution to (3.59). We consider the
cubic operator
1 X
Û = εijk Êi (S α )Êj (S β )Êk (S γ ) . (3.60)
48
α,β,γ
Let us restrict to gauge-invariant spin networks, where the Gauss law holds and each node is
labeled by an intertwiner. First of all, it is immediate to see that the action of (3.60) on a 3-
valent node is zero. In fact, the Gauss law tells us that the sum of the fluxes through a surface
around a gauge invariant node is zero. In the 3-valent case, only three S α give non-zero fluxes,
thus
Êi (S α ) + Êi (S β ) + Êi (S γ ) |ii = 0, (3.61)
which implies
Êi (S α ) |ii = − Êi (S β ) + Êi (S γ ) |ii . (3.62)
Using this result in (3.60) we get zero because two identical fluxes always appear,
εijk Êi (S α )Êj (S β )Êk (S γ ) |ii = −εijk Êi (S β ) + Êi (S γ ) Êj (S β )Êk (S γ ) |ii = 0.
Non-trivial contributions to the volume comes from nodes of valency 4 or higher. Notice
that these are the cases for which the intertwiner is not unique, but a genuine independent
quantum number. Thus the operator (3.60) probes precisely the degrees of freedom hidden
in the intertwiners. Consider the 4-valent case. Of all the surface cells S α , only four are
punctured by the links. Hence, we have only four contributions to the Gauss law, which we
can then use to eliminate one flux in favor of the remaining three,
The sum over the S α in (3.60) then reduces to the sole contributions from the four punctured
surfaces. Using (3.63) these contributions are all equal. A simple combinatorial shows that
there are 48 terms, all equal, thus
Û = εijk Êi (S 1 )Êj (S 2 )Êk (S 3 ) = ~3 γ 3 εijk Ji1 Jj2 Jk3 . (3.64)
In the last step we have used the result (3.13b), which gives the action of the fluxes in terms of
the generators J~a in the spin ja representation. Notice that the orientation factors ± (between
the edge and the surface) are irrelevant because of the modulus.
This is a well-defined cubic operator, whose action spectrum is again discrete, with mini-
mal excitation proportional to (~G)3/2 . The explicit formula is much more complicated than
that of the area. We refer the interested reader to the literature [29].
The alternative proposal by Ashtekar and Lewandowski has the extra property of being
sensitive to the differential structure of the graph at the node. The subdivision of the region
R into cubic cells is the same as above, the key difference is in how to choose the surfaces
3 L OOP QUANTUM GRAVITY : K INEMATICS 53
to smear the triad filed. Instead of subdividing the boundary ∂CI into cells S α , we now
consider for each CI only three surfaces. We assign a local coordinate system xa , and take
the three surfaces to be any three surfaces SIa inside the cube and orthogonal to each other.
The Ashtekar and Lewandowski volume operator is defined as
s
X 1
V̂ (R) = lim εijk εabc Ê i (S a )Ê j (S b )Ê k (S c ). (3.65)
ε7→0 3! I I I
CI
Unlike (3.56), this expression is already written in terms of fluxes, so we do not need the
manipulation in terms of WI . We can directly use (3.13b) to write
1 ~3
εijk εabc Ê i (SIa )Ê j (SIb )Ê k (SIc ) = εijk εabc κa κb κc Jai Jbj Jck , (3.66)
3! 48
where again Jeia is the SU (2) generator in the spin ja representation, and κS is the relative
orientation between the cell SIa and the edge puncturing it, i.e.
1 if e lies above S
κS (e) = −1 if e lies below S (3.67)
0 otherwise
The operator (3.66) strongly depends on the choice of the triple of surfaces SIa . In order
to define a volume operator independent from the regulator, we average (3.66) on all the
possible choices of orthogonal triples SIa (see the [27] for the details). The result is a sum
over the nodes n of the graph,
v
u
Xu u κ 0
X
V̂ (R) = ~ 3/2
t εijk ǫ(e, e , e ) J (e)J (e )J (e ),
′ ′′ i j ′ k ′′ (3.68)
n∈Γ 48 e,e′ ,e′′
where e,e′ ,e′′ runs over the set of edges passing through the node n and ǫ is the orientation
function, which equals 0 if the tangent directions of the three edges are linearly dependent
and ±1 if they are linearly independent and oriented positively or negative respect to the
orientation of Σ. κ0 is an undetermined constant introduced by the averaging procedure.
This alternative volume differs from (3.59) in two ways : firstly, it depends on an arbitrary
constant, κ0 , unlike (3.59) which is unique. Secondly, the absolute value is outside the internal
summation, unlike (3.59) where it is inside. This second differences is what makes the AL
operator sensible to the differential structure of the graph : in fact, the action of (3.68) vanishes
on nodes whose edges lie on a plane, because in that case ǫ ≡ 0.
On the other hand, both operators annihilate a 3-valent node, and coincide (up to a pro-
portionality constant) in the 4-valent case, where they are both proportional to εijk Ji1 Jj2 Jk3 .
Summarizing, both volume operators act only on nodes of the graph. Their matrix ele-
ments vanish between different intertwiner spaces, and since every intertwiner space is finite
dimensional, their spectra are discrete with minimal excitations proportional to the Planck
length cube ℓ3P .
54 G RAVITATION QUANTIQUE
Together with the discreteness of the area operator, these results show that in loop quan-
tum gravity the space geometry is discrete at the Plack scale. Each spin network describes a
quantum geometry, where each face dual to a link has an area proportional to the spin je , and
each region around a node has a volume determined by the intertwiner in as well as the spins
of the link sharing the node.
It is important to stress that this is not a built-in discretization, as in lattice approaches to
quantum gravity. It is a result of the quantum theory, similar to the quantization of the energy
levels of an harmonic oscillator, or the radii of the atomic orbitals.
Thanks to this fundamental discreteness, where the minimal geometric excitations are
proportional to the Planck length, the theory is expected not to have ultraviolet divergences,
and to resolve the problem of the classical singularities of general relativity at the big bang
(e.g. [31]), or at the center of black holes (e.g. [32]). Addressing these issues of course re-
quires taking into account the dynamics of the theory, to which we turn next.
0
the diffeomorphism constraint can be described in terms of linear functionals on Hkin . Let
0∗ 0 0∗
us denote Hkin the space of linear functionals on Hkin . Then, η ∈ Hkin is a diffeomorphic-
invariant functional if
0
η(φ̂ψ) = η(ψ) ∀ψ ∈ Hkin . (4.1)
∗
The space of such functionals is denoted HDif f , and the Hilbert space HDif f solution of
the constraint is constructed by duality. The condition (4.1) should remind the reader of the
analogue condition of gauge-invariance studied in the previous section. There, we were able
to implement it via a simple group averaging procedure. In the same manner, we would like
to define a projector PDif f on HDif f such that
X
hψ|ψ ′ iDif f ≡ hψ|PDif f |ψ ′ i = hφ̂ψ|ψ ′ i, (4.2)
φ∈Diff/TDiff Γ
where the sum is over all the diffeomorphism mapping Γ into Γ′ except those corresponding
to the trivial ones TDiffΓ . See [30] for more details.
The result of this procedure are spin network states defined on equivalence classes of
graphs under diffeomorphisms. These equivalence classes are called knots, see Figure 1. The
study of knots forms an elegant branch of mathematics. The diff-invariant Hilbert space of
loop quantum gravity is spanned by knotted spin networks.
Diffeo
F IGURE 1 – Left panel. A diffeomorphism can change the way a graph is embedded in Σ, but
not the presence of knots within the graph. Right panel. The first few knots (without nodes),
taken from Wikipedia.
where we introduced the shorthand notation H E (N ) and T (N ). As with the ADM Hamil-
tonian constraint (1.19), this expression is non-linear, which anticipates difficulties to turn it
into an operator. However, a trick due to RThiemann
p [33] allows us to rewrite (4.3) in a way
amenable to quantization. Denoting V = det(E) the volume of Σ, and
Z
K̄ = Kai Eia , (4.4)
we can use the classical brackets (2.34) to establish the following identities,
1 1 i
Kai = Aia − Γia (E) = Aa , K̄ , (4.5)
γ γ
1 E
K̄ = 3/2
H (N ≡ 1), V , (4.6)
γ
EaEb 4 k
p i j ǫijk ǫabc = Aa , V . (4.7)
det(E) γ
The advantage of this reformulation is that the non-linearity is mapped into Poisson bra-
ckets. The next step is to rewrite these expressions in terms of holonomies and fluxes, so that
we can turn them into operators. Notice that we already know the volume operator, and its
spectrum can be explicitly computed. This is very promising towards the prospect of kno-
wing the action of the Hamiltonian constraint. Next, the connection and curvature have to be
written in terms of holonomies. This requires a regularization procedure. We describe it only
for the first term, H E , and refer the reader to the literature [13] for T (N ).
The connection can be easily expressed in terms of holonomies. Writing explicitly (2.38),
we have that for a path ea of length ǫ along the xa coordinate, hea [A] ≃ 1 + εAia τi + O(ǫ2 ),
therefore we also have i 2
h−1
ea {hea , V } = ǫ Aa , V + O(ǫ ). (4.9)
For the curvature, consider an infinitesimal triangular loop αab , lying on the plane ab, and
with coordinate area ǫ2 . At lowest order in ǫ,
1
hαab = 1 + ǫ2 Fab
i i
τ + O(ǫ4 ), (4.10)
2
thus
2 i i 4
hαab − h−1
αab = ǫ Fab τ + O(ǫ ). (4.11)
At this point, we proceed as we did for the geometric operators in the previous section. We
introduce a cellular decomposition of Σ, and regularize the integral as a Riemann sum over
4 L OOP QUANTUM GRAVITY : DYNAMICS 57
the cells CI ,
X
H E = lim ǫN
I 3ǫ
abc
Tr (Fab {Ac , V }) = (4.12)
ǫ→0
I
X −1
= lim NI ǫabc Tr hαab − h−1
αab hec {hec , V } . (4.13)
ǫ→0
I
This time is more convenient to specify the cellular decomposition in terms of a triangulation,
namely a collection of tetrahedral cells. The loop αab can then be adapted to the triangular
faces of this decomposition, as in the following figure.
ec
αab
This is a well-defined operator, whose action is explicitly known. It inherits the property of
the volume operator of acting only on the nodes of the spin network. From the holonomies,
it modifies the spin network by creating new links carrying spin 1/2 around the node, see the
following figure.
m
o 1
2
ǫ2
j n
Finally, its amplitude depends on the details of the action of the volume operator. See the
literature [11, 13, 30] for details.
Since the only dependence on ǫ is on the position of the new link, in the Hilbert space
HDif f where the position of the link is irrelevant, the limit can be safely taken without
58 G RAVITATION QUANTIQUE
affecting the result. Not only this operator is well-defined, also the Dirac algebra (1.23) can be
realized at the quantum level. To see this, we need first to note that the new “exceptional” links
added by (4.14) carry zero volume and are thus invisible to a further action by (4.14). This
crucial property requires the Ashtekar-Lewandowski version of the volume quantum operator,
because the new nodes at the junction of the new link with the old ones are planar, and the
AL volume operator vanishes on planar nodes. Thanks to this, H(N b 1 ) and H(N
b 2 ) commute
on the space HDif f of diffeomorphic-invariant spin networks. Hence for diffeomorphism
invariant states hφ| ∈ HDif f ,
h i
b 1 ), H(N
hφ| H(N ~ 2] ) |ψi ,
b 2 ) |ψi = 0 = hφ| H(N[1 ∇N ∀ψ ∈ Hkin . (4.15)
This is the correct commutator algebra on-shell. Thanks to this non-trivial property, the quan-
tization is said to be anomaly-free.
Another key result is that an infinite number of states solutions of H b are known : any
b E b
graph without nodes is in the kernel of H and T . This result [35], together with the discre-
teness of the spectra [34], started the whole interest in the loop quantum gravity approach.
More in general, a formal solution of the Hamiltonian constraint will be a particular linear
combination (generally infinite) of spin networks with an arbitrary number of exceptional
links, whose coefficients depend on the details of the quantum operator and on the spins
carried by the spin network,
= α +...+ ω +...
This construction defines a new spin network with a “dressed node” as the solution. However,
this procedure is only formal, and no explicit solutions are known in general.
In conclusion, we have a perfectly well-defined Hamiltonian constraint, whose action
is explicitly known and finite. An infinite number of states solutions of H b are known, and
the Dirac algebra is anomaly-free on physical states. Compare this with the old-fashioned
Wheeler-De Witt equation (1.37), which was badly ill-defined, and we see the full force of
the use of the Ashtekar variables to quantize general relativity.
Nevertheless, the program of quantization is far from being complete : the complete cha-
racterization of Hphys nor the full spectrum of H b are known. In spite of the successes of
this approach, some limitations need to be stressed. In particular, although well-defined, the
Hamiltonian constraint is plagued by a number of ambiguities :
– We can change the spin of the exceptional edges to an arbitrary value j. This change
have significant physical effects in loop quantum cosmology for example.
– We can regularize the connection and the curvature with different paths and still make a
consistent theory [30]. In particular, one can envisage a different construction in which
the constraint acts on more than one node simultaneously, as advocated in [36].
– Alternative orderings can be explored.
5 ON QUANTUM GEOMETRY 59
5 On quantum geometry
Let us review what we have learned in Section 3. The golden tools of loop quantum gra-
vity are the spin network states. These form a basis in the kinematical Hilbert space, and
diagonalize geometric operators. In particular, the quantum numbers carried by a spin net-
work, (Γ, je , in ), define a notion of quantum geometry, 5 in the same way as the quantum
number of an harmonic oscillator defines its quantum state. These quantum numbers can be
compared with the kinematical metric gab , defining the classical geometry of space. We use
the word kinematical to mean that gab is an arbitrary metric, not necessarily a solution of
Einstein’s equations, just like an arbitrary spin network spans the kinematical Hilbert space,
not necessarily solving the diffeomorphisms and Hamiltonian constraints. The quantum geo-
metry described by (Γ, je , in ) is very different than a classical geometry gab . Specifically, it
is largely insufficient to reconstruct a metric gab . We can highlight three key differences :
(i) Quantized spectra : the spectra of geometric operators are discrete, as opposed to the
continuum values of their classical counterparts. This is a standard situation in quantum
mechanics, think for instance of the discretization of the energy levels of an harmonic
oscillator.
(ii) Non-commutativity : not all geometric operators commute among themselves. This is
a consequence of the non-commutativity of the fluxes (3.20). This is also standard, like
the incompatibility of position and momentum observables.
(iii) Distributional nature : the states capture only a finite number of components of the
original fields, that is their values along paths (for the connection) and surfaces (for the
triad). This is reminiscent of what happens in lattice theories, where the continuum field
theory is discretized on a fixed lattice and only a finite number of degrees of freedom
are captured.
5. Each face dual to a link has an area proportional to the spin je , and each region around a node has a volume
determined by the intertwiner in as well as the spins of the link sharing the node.
60 G RAVITATION QUANTIQUE
In spite of these differences, if the theory is correct it must admit a semiclassical regime
where a smooth geometry emerges. And furthermore, the dynamics of this smooth geometry
should be given by general relativity, at least in some approximation. This is what we expect
from any quantum theory : that the ~ 7→ 0 limit is well-defined, and reproduces classical
physics. The points (i-iii) show that this limit might not be trivial to obtain in loop quantum
gravity. Although we started from genuine general relativity and quantum mechanics, without
any exotic ingredient, we ended up with a description in terms of quantities, (Γ, je , in ), which
are very far from the original ones, gab . This is the problem of the semiclassical limit.
In this final Section, we would like to comment on the first part of this problem, namely
the emergence of a smooth geometry. This requires understanding the points (i-iii). The first
point is easier to deal with : also the orbitals of the hydrogen atoms are quantized, placed at
distances labeled by an integer n. The classical Keplerian behavior is recovered if we look at
the large n limit. Similarly, continuum spectra are recovered in the large spin limit jl 7→ ∞.
Points (ii) and (iii) are more subtle. The key to deal with them is the use of coherent states,
namely linear superpositions of spin network states peaked on a smooth geometry.
Recall that coherent states are peaked on a point in the classical phase space. For instance,
the coherent state |zi for an harmonic oscillator is peaked on the initial position q = Re(z)
and momentum p = Im(z). The phase space is the familiar cotangent bundle to the real line,
(q, p) ∈ T ∗ R. Analogously, coherent states for loop quantum gravity are peaked on a point
(Aia (x), Eia (x)) in the classical phase space, which defines an intrinsic (through the triad)
and extrinsic (through the connection) 3-geometry. Such coherent states for loop quantum
gravity were introduced in [39]. These states minimize the uncertainty of the flux operators,
thus addressing (ii). In order to address (iii) and recover a smooth geometry everywhere on
Σ, the coherent states have support over an infinite number of graphs.
On the other hand, dealing with an infinite number of graphs is a formidable task, and
for practical purposes, one often needs to rely on approximations. A convenient one is to
allow only states living on a fixed finite graph Γ. The Hilbert space HΓ provides a truncation
of the theory, which may be sufficient to capture the physics of appropriate regimes. Within
this truncation, one considers coherent states in HΓ . However, in which sense one can assign
a classical geometrical interpretation to these states ? They will be peaked on points in the
phase space corresponding to HΓ0 , which consists of classical holonomies and fluxes on Γ.
These quantities capture only a finite number of components of a continuum geometry. To be
able to interpret these truncated coherent states in a physical sense, we need to use these data
to approximate a continuum geometry. The problem is similar to approximating a continuous
function f (x) if we are given a finite number of its values, fn = f (xn ). Various interpolating
procedures are common. We now describe a notion of interpolating geometry which emerges
naturally.
The associated gauge-invariant Hilbert space on a fixed graph is a sum over intertwiner
spaces Hn ≡ Inv[ ⊗ V (je ) ] associated to the nodes (see (3.41)),
e∈n
relevant for us is that this phase space can be parametrized as follows [40],
SΓ = × T ∗ S 1 × SF (n) . (5.2)
l n
Here T ∗ S 1 is the cotangent bundle to a circle, F the valency of the node n, and SF is a 2(F −
3)-dimensional phase space, introduced by Kapovich and Millson [41]. This parametrization
mimics the decomposition (5.1). In particular, the quantization of T ∗ S 1 gives the quantum
number jl associated to each link, and the quantization of SF (n) gives the F − 3 intertwiners
associated to an F -valent node.
The parametrization defines a notion of interpolating geometry associated to a cellular
decomposition of Σ dual to Γ, called twisted geometry [40]. The key to the geometric inter-
pretation of (5.2) is the node space SF (n) . Fix the spins jl on the links connected to the node,
and consider F unit vectors nl in R3 constrained to satisfy the following closure condition,
F
X
C= jl nl = 0. (5.3)
l=1
Because of the closure, the vectors jl nl span a polygon in R3 , see Figure 2. The space of all
j4 n 4
j3 n 3
j2 n 2
j1 n 1
O
F IGURE 2 – A polygon in R3 , with F sides of fixed lengths. Because of the closure condition,
not all the components of the normals are independent, but only 2F −3. If we further consider
the polygon up to rotations, that is the space SF , this is only specified by 2F −3−3 = 2(F −3)
numbers. These can be taken to be the lengths of F −3 diagonals and their conjugated dihedral
angles (which measure the way the polygon hinges on each diagonal).
polygons with fixed jl up to SO(3) rotations forms the phase space SF , called by Kapovich
and Millson the space of shapes of the polygon.
This means that we can interpret the intertwiners as “quantized polygons”. Such inter-
pretation is appealing, but not particularly useful for our purposes. More promising would be
a description in terms of polyhedra, rather than polygons : Polyhedra can be glued together
to approximate a smooth manifold, and their geometry will induce a discrete metric of some
sort.
The good news [43] is that this is precisely what happens ! In fact, each (non-coplanar)
configuration of vectors nl describes also a unique polyhedron in R3 , with jl as the areas of
62 G RAVITATION QUANTIQUE
F IGURE 3 – A generic polyhedron with 100 faces. Notice that most faces are hexagons.
the faces, and nl as their unit normals. More precisely, a convex bounded polyhedron. The
explicit reconstruction of the polyhedron geometry from holonomies and fluxes is studied in
[43]. Therefore, up to the degenerate configurations corresponding to coplanar normals, we
can visualize SF as a polyhedron with F faces.
Going back to (5.2), the T ∗ S 1 space on the links carries information needed to define the
parallel transport between adjacent polyhedra. This unique correspondence between SF in
(5.2) and polyhedra means that each (non-degenerate) classical holonomy-flux configuration
on a fixed graph can be visualized as a collection of adjacent polyhedra, each one with its
own frame, and with a notion of parallel transport between any two polyhedra. Since twisted
geometries parametrize the classical phase space on a fixed graph, coherent states on a fixed
graph are peaked on the approximated smooth geometry described by a twisted geometry, i.e.
a collection of polyhedra.
Twisted geometries have a peculiar characteristic which justifies their name : they define
a metric which is locally flat, but discontinuous. To understand this point, consider the link
shared by two nodes. Its dual face has area proportional to jl . However, its shape is determined
independently by the data around each node (the normals and the other areas), thus generic
configurations will give different shapes. In other words, the reconstruction of two polyhedra
from holonomies and fluxes does not guarantee that the shapes of shared faces match. Hence,
the metric of twisted geometries is discontinuous in the sense that the shape of a face “jumps”
when going from one polyhedra to the next. See left panel of Figure 4.
Notice that one can also consider a special set of configurations for which the shapes
match, see right panel of Figure 4. This is a subset of SΓ which corresponds to piecewise
flat and continuous metrics. For the special case in which all the polyhedra are tetrahedra,
this is the set-up of Regge calculus, and those holonomies and fluxes describe a 3d Regge
geometry. The matching conditions in this case were studied in [44]. For an arbitrary graph,
the matching conditions are studied in [43], and the result would be Regge calculus on generic
cellular decompositions.
The relation between twisted geometry and Regge calculus implies that holonomies and
fluxes carry more information than the phase space of Regge calculus. We stress that this
is not in contradiction with the fact that the Regge variables and the LQG variables on a
fixed graph both provide a truncation of general relativity : simply, they define two distinct
R ÉFÉRENCES 63
F IGURE 4 – Two high-valency nodes and the link connecting them. Left panel : In general,
the two adjacent polyhedra, defined by holonomies and fluxes on the nodes, don’t glue well.
Even if the area of a polygonal face is the same because there is a unique spin associated
with the link, the shapes will be different in general. Right panel : in order for the shapes to
match, one needs to impose appropriate conditions on the polyhedra such as the matching of
the normals to the edge in the plane of the face. These conditions, studied in [43], affect the
global shape of the polyhedron.
truncations of the full theory. See [45] for a discussion of these aspects.
Summarizing, when we truncate the theory to a single graph, we capture only a finite num-
ber of degrees of freedom of the geometry. This finite amount of information can be suitably
interpolated to give an approximate description of a smooth geometry. The reparametriza-
tion in terms of twisted geometries achieves a prescription for this interpolation procedure :
Starting from holonomies and fluxes on a graph, we can assign to them a specific twisted
geometry, that is, a bunch of polyhedra stuck together.
These results offers on the one hand a new way to visualize the quantum geometry of
loop quantum gravity in terms of fuzzy polyhedra. On the other hand, they open the door to
new beautiful mathematical ingredients, from the geometry of polyhedra [42, 43], to twistors
[46], to new emerging symmetries [47].
These intriguing aspects, together with the striking recent developments in the study of
the dynamics and in the applications to cosmology and the physics of black holes, make of
loop quantum gravity an excellent arena for a student interested in the problems of physics at
the Planck scale.
Acknowledgements SS wishes to thank the organizers of the School for their kind in-
vitation, and for their splendid hospitality and great help in making the time in Algeria so
enjoyable. J’espere que pour la prochaine ecole, nous aurons aussi le tournoi de baby-foot !
Références
[1] J. Donoghue, Leading quantum correction to the Newtonian potential, Phys. Rev. D 50,
3874 (1994).
64 G RAVITATION QUANTIQUE
[34] C. Rovelli and L. Smolin, Discreteness of area and volume in quantum gravity, Nucl.
Phys. B 442 (1995) 593 [Erratum-ibid. B 456 (1995) 753] [arXiv :gr-qc/9411005].
[35] C. Rovelli and L. Smolin, Loop Space Representation of Quantum General Relativity,
Nucl. Phys. B 331 (1990) 80.
[36] L. Smolin, The classical limit and the form of the Hamiltonian constraint in non-
perturbative quantum general relativity, arXiv :gr-qc/9609034.
[37] T. Thiemann, “Quantum spin dynamics. VIII : The master constraint,” Class. Quant.
Grav. 23 (2006) 2249 [arXiv :gr-qc/0510011].
[38] C. Rovelli, “A new look at loop quantum gravity,” arXiv :1004.1780 [gr-qc].
[39] A. Ashtekar, J. Lewandowski, D. Marolf, J. Mourao and T. Thiemann, “Coherent state
transforms for spaces of connections,” J. Funct. Anal. 135 (1996) 519 [arXiv :gr-
qc/9412014].
T. Thiemann, “Gauge field theory coherent states (GCS). I : General properties,” Class.
Quant. Grav. 18 (2001) 2025 [arXiv :hep-th/0005233].
[40] L. Freidel and S. Speziale, “Twisted geometries : A geometric parametrisation of SU(2)
phase space,” arXiv :1001.2748 [gr-qc].
[41] M. Kapovich and J. J. Millson, “The symplectic geometry of polygons in Euclidean
space,” J. Differential Geom. 44, 3 (1996), 479-513.
[42] A. D. Alexandrov, “Convex Polyhedra,” Springer (2005).
[43] E. Bianchi, P. Doná and S. Speziale, “Polyhedra in loop quantum gravity,”
arXiv :1009.3402 [gr-qc].
[44] B. Dittrich and S. Speziale, “Area-angle variables for general relativity,” New J. Phys.
10 (2008) 083006 [arXiv :0802.0864 [gr-qc]].
[45] C. Rovelli and S. Speziale, “On the geometry of loop quantum gravity on a graph,”
arXiv :1005.2927 [gr-qc].
[46] L. Freidel and S. Speziale, “From twistors to twisted geometries,” arXiv :1006.0199
[gr-qc].
[47] L. Freidel and E. R. Livine, “U(N) Coherent States for Loop Quantum Gravity,”
arXiv :1005.2090 [gr-qc].