Chapter 13

Hamiltonian Formulation of
Local Symmetry
For several purposes, it is sometimes preferable to work on a lattice with different lattice spacings in different directions, as we have done in analyzing the
plaquette. We can write the action as a sum over sites with the appropriate
space-time volume,
Z
X
d4 x ax ay az at
n

iga A

Un,n+ = e

2
F

on each link


1

Tr Up + Up 2 1I
g 2 a 2 a 2

for each placquette

and so1
(



1 X 1

+
U
Tr
U

1
I
(1)0 +0
p
p
2
2 a 2
4g
a
n

)
X 1
.
+i
Un,n+ n m
n+
a

S = ax ay az at

We now turn to a discussion of the Hamiltonian formulation, as well as the

Lagrangian. This will give us a quantum mechanics in the usual language,
1

I am leaving out a discussion of fermion doubling. In general we shall concentrate on

the gauge field U .

147

148. Last Latexed: April 4, 2016 at 14:56

Joel A. Shapiro

but of a rather
P 2 unusual form, because the kinetic energy term will not be
the usual
pi /2m, because the degrees of freedom, the group elements on
each link, do not live in a Euclidean space, but instead live on the group
manifold. The way the derivative operators enter the Hamiltonian will lead
us to a metric and a measure on the group manifold.
We will derive the Hamiltonian formulation of the lattice gauge theory by
considering the at 0 limit. First we must use gauge invariance to remove
some of the degrees of freedom.
Quantum mechanics is an integral of the action over all possible configurations of the physical degrees of freedom. Gauge degrees of freedom have
no effect on the action, and are therefore not physical degrees of freedom.
Given any configuration U, on all the links and sites, it is possible to find
a new configuration, gauge transformed, satisfying some imposed gauge requirements but representing the same physical configuration. In our case, we
pick a particular time, say t = 0. We now perform gauge transformations
at each site with t 6= 0 in order to make all Un,n+t = 1I. This can be done
iteratively. For example, at t = at , at site n = ~r, at we examine U(~r,0),t = ei~r ,
and then make a gauge transformation at the site (~r, at ) with ~r , with no
change at (~r, 0), so the new U on the timelike link is

i~r i~r 0
U(~
e e = 1I.
r ,0),t = e

This set of transformations changes all the spacelike Us at t = at to new

values, but it does not affect the integral over all such Us, provided the
weight of integration is group invariant. We shall say more about this later.
Reiterating this step forwards in time to t = , and also going backwards
from t = 0, we wind up with a configuration of Us and s equivalent to our
original one, but subject to the gauge constraint Un,t = 1I. In the continuum
limit, this means A0 = 0 everywhere, which is called the temporal gauge.
x
y
z
Now let us consider the limit at
R 0 with a =
R a = a = a held fixed.
t
The sum over times, times a , is just dt, so S = dtL as usual, with

X


1 X
3
L =
Tr
U

1
I
+
h.c.
a
P
0j
2g 2a2 at 2 j
nx ,ny ,nz

1 X 
+ 2 4
Tr Upij + Upij 2 1I + fermion terms.
4g a ij
The first term is the sum over placquettes in a time-space plane, while the

618:

Last Latexed: April 4, 2016 at 14:56

149

second is in a space-space plane (at a fixed time). The first of these involves
Up0j = Un ,0 Un +at ,j Un +aj ,0 Un,j .
| {z }
| {z }
1I

Expand

so

1I

U(~n,t),j 1 t 2 2 U
U(~n,t+at ),j = U(~n,t),j + at
+ a
,
2
t
2
t
!
1
Un1
,j
at 2 2 Un ,j

t
Un ,j +
Un ,j .
U(~n,t+at ),j U(~n,t),j 1I = a
t
2 t2

Adding the hermitean conjugate gives

!


1
2
U
,j
,j
U
at 2 2 U 1
n
n
t
1
1 U
a
+
Un ,j + Un ,j
U +U
.
t
t
2
t2
t2

(U 1 U) = at t
1I = 0, while from the second we may
The first term is at t
subtract




2
at 2 2 U 1
at 2 U 1
U 1 U
at 2
1 U
1 U
=
0=
U +U
+2
+U
2 t
2 t
t
t
2
t2
t t
t2

which leaves
Tr UP0j

 1



U U

t2
1 + h.c. = Tr Ut+at Ut 1I + h.c. = a Tr
.
t t

,
We can improve this by using our expression for zero again, Ut U = U 1 U
t
U 1
1 U 1
so t = U t U and





t2
1 1
Tr Ut+at Ut 1I + h.c. = a Tr U U U U ,
and


X a 
1 1
L=
T r U UU U V {U},
2g 2
~
n

where V contains the spacelike placquette terms, without any time derivatives.
To get the Hamiltonian we need first to find the canonical momenta =
L
. In our case the coordinates q are the matrix elements of U on each
q

150. Last Latexed: April 4, 2016 at 14:56

Joel A. Shapiro

spacelike link U ab , where specifies the link (both spacelike direction and
spatial position) and ab are the matrix indices. For convenience we will define
the matrix P as the transpose, so

a 
L
(P )cb :=   = 2 U1 U U1 ,
cb
g
U
bc

H=

U bc P cb L =

,c,b

X g 2

2a

Tr (U P U P ) + V ({U}).

Quantum mechanically, we write the momentum conjugate to q as i q

,

so Pcb = i Ubc . But P only enters H together with U. So let us define,

for each link,

.
(E )ab = i (U P )ab = U ac
(U )bc
P
In terms of of the Es, the kinetic energy term in L becomes 2ga2 Tr E2 .
The Es are a scaled electric field (ag times the usual field). As a differential operator it is very interesting.
But before we pursue the group theoretic arguments over the interpretation of E, let us meet an objection to the above procedure, that the Uab are
i
not independent real coordinates2 . Instead, we should write U = ei Li and
treat the real, independent i as the coordinates. Then we get

i =

U ab L
Uab L
L
=
=
i
i

Uab
i U ab

ab
i.
where I have made use of U ab = U
i


a
L
= 2 U 1 U U 1 , so
But

g
ba
Uab

a
i = 2 Tr
g

"
!
#
U 1 X j U
U

U 1
j
i





a
a
i U
1 j U
1
1 U U 1
and i = 2 Tr
=

U
Tr
UU
g
i
j
g2
i

Of course we have a U and an for each link, but we will drop the index , or ~n,
that specifies which link, in what follows.

618:

Last Latexed: April 4, 2016 at 14:56

151

in agreement with what we found by careless manipulation.

One sees, however, that there are only N 2 1 coordinates and momenta
for SU(N), not N 2 as on might expect from P cb . The requirement that U be
unitary and of determinant 1 means there are constraints on the P cb s we
can vary U only preserving the unitarity and determinant. Interpreting the
P cb quantum mechanically as i U bc would be varying into a dimension for
which the action is undefined, where det U 6= 1. Our E cb do not have that
problem. Consider any function f on n n unitary matrices u. Each such
matrix is u = ei U with U SU(N), so we may write f (u) = f(, U) and


f ucd 
f
f 
= ei
,
 =

Uab 
ucd Uab
uab

so if Eab

f
= iuac
, Eab f = iei Uac
= iUac
 and has no depenubc
ubc
Ubc 

dence on how f varies off the det u = 1 subset. So we see that the Eab does
not involve a derivative in the direction away from det U = 1, but does form
a complete set of N 2 1 first order differential operators on functions of
SU(N) matrices.
The N 2 1 momenta conjugate to i are quantum mechanically
Uab

= i i
.
i = i
i
Uab
Collectively these clearly span the same space as the Eab . To understand the
connection, consider different ways of asking how a function on G varies as
we move the element g G.
We see that i is not the same as E. To understand the connections,
consider a function f : G C defined on the group. Then derivatives of
f mean finding the change in f (U) under small changes in U. If we write
i
U = ei Li , then

f (U) = ii f.
i
On the other hand, we could consider the change in f as U U () =
i
ei Li U, for infinitesimal i . If we define the differential operators



 ij Lj 
f (U ) Uab
 = f (U) i(Lj )ac Ucb
Ej f (U) = i
f
e
U
=
i


j
U
j 
Uab
=0

= (Lj )ac Ucb

f = Tr(Lj E)f.
Uab

ab

j =0

152. Last Latexed: April 4, 2016 at 14:56

Joel A. Shapiro

So we see that the electric fields Eab = U ac P cb which we defined earlier

are just
P linear combinations of the left-handed differential operators Ej , with
E = j Ej Lj .
For completeness, let us also define a right derivative on the space of
functions from G C by








ij Lj 
ij Lj

i i Ue
=
f Ue
Rj f (U) = i

cb 
i
Ucb

=0
=0
f
=
Uca (Lj )ab .
Ucb
This is not exactly the same as Ej If we make a matrix from Rj in the
same way as we did for Ej ,
Rab =

Rj j ab = Ucb

and recall Eab = Uac

Uca

,
Ubc

so

f
f
U = U 1 U T U = U 1 (Ef )U.
T
U
U
Note that the Ej , Rj and j are each sets of N 2 1 first order differential
operators on the space of functions from G C, an N 2 1 dimensional
space. Thus they are linear combinations of each other. The connection
between them is associated with the adjoint representation of the group.
We will write [S(L )]jk for the adjoint representation of the generators,
so [S(L )]jk = icjk . Let us call the adjoint representation of the group S:



i L
i
S(L
)

.
Sjk (e
)= e
Rf =

jk

[Note: S and S were both previously called adj . Also it is hard to distinguish
S from S. ] These will come up in the connections we are seeking.

13.1

Relation of Rj to Ej

Now U 1 L U is a linear combination of the Ls, so U 1 L U = Mm (U)Lm .

We can show Mm = Sm .

618:

Last Latexed: April 4, 2016 at 14:56

j

Proof: Let U() = ei Lj , M() := M(U()). Then ei

Mm ()Lm . Differentiate with respect to :
ei

jL
j

[L , i k L ] ei
| {z k}

jL
j

jL
j

153
kL
k

k S(L )
k

L ei

dMm ()
Lm = k ckn Mnm ()Lm .
d

k ckn Ln

The Lm are linearly independent, so



dM()
= k ckn Mnm () = i k [S(Lk )M]m .
d
m
This is a differential equation with solution
k
M() = ei S(Lk ) K

with initial condition Mm (0) = m = 1I = K, so M() = ei

S(U()). Thus
k
k
b
ei Lk L ei Lk = Sm (ei Lb )Lm .
k

So U 1 L U = Sm (U)Lm and by exponentiation, U 1 ei Lk U = ei

Now reconsider

 ik Lk 
f Ue
R f (U) = i


=0

 ik Skm (U 1 )Lm 
= i
f e
U 

=0





n


i m Lm

= i
f
e
U



| {z }

S (U 1 )
n

where

= Skm (U

kS

km (U )Lm

=0

). So

R f (U) = Sm (U 1 )Em f (U),

R = Sm (U 1 )Em ,

and Em = Sm (U)R ,

as S 1 (U 1 ) = S(U).
P 2
is E better than R ? In the Hamiltonian we have
E , why not
P Why
2
R ?
X
2
Em
= Sm (U)R Smk (U)Rk


 

i j Lj 
Rk .
Smk Ue
= Sm (U)Smk (U)R Rk iSm


=0

154. Last Latexed: April 4, 2016 at 14:56

Joel A. Shapiro
j

Note that S is hermitean and imaginary so S = ei S(Lj ) is unitary and real,

j
hence orthogonal, and Sm (U)Smk (U) =k . Also Smk (Uei Lj ) =
j
Smn (U)Snk (ei Lj ), so i Smk (UeiL )=0 = Smj (U)Sjk (L ), so the second
term is iSm (Smj ) Sjk (L )Rk = j (icjk )Rk . Thus
X
X
X
2
Em
=
R Rk k + im cmj Rj =
Rj2 ,

by the antisymmetry
P 2
Pof 2c. So in fact the Hamiltonian does not distinguish
between
E and
R , as they are the same thing.
We can also ask what the connection is between E and . To find the
A(x)
e
for an operator A which
connection we need to develop a formula for x
need not commute with its derivative. The formula is
Z 1
A(x)
A(x) (1)A(x)
e
=
e
(13.1)
d eA(x)
x
x
0

which you will prove for homework. Then around a fixed element U0 = ei0 L ,
we may write

ei L = ei L ei0 L
with 0 as 0 . Thus

m
 i L  m
im Lm
iEm f.
f
e
=
f
e
U
=
0

m

The partial derivative matrix can be evaluated with this formula as applied
to f (U) = U,
Z 1
m
ik Lk
k
k
i k Lk
e
= i Lm e
=
ei Lk iL ei(1) Lk

or

m
Lm =

or

i k Lk

d e

i k Lk

L e

m
=

so

1
0

i
h
k
d ei S(Lk )

1
0



k
d Sm ei Lk Lm ,

"

1 ei S(Lk )
=
i k S(Lk )

#
"
k
1 ei S(Lk )

=

k S(Lk )

Em

618:

Last Latexed: April 4, 2016 at 14:56

155

gives the connection of the left derivative with the ordinary partial derivative.
[Note: the fraction in square brackets is well defined even though k S(Lk )
is likely not an invertible matrix, so dividing by it is dubious. But the
expression in the bracket has a well defined power series expansion and thus
a well defined meaning.]
We have given a great deal of discussion to our Hamiltonian operator. We
now turn to the states on which this operator acts. A state can be specified
by giving
(a) at each site, the occupation number for the fermions at that site
(b) on each link, a group element. We focus on this.
We are not accustomed to having dynamics with canonical coordinates
defined on a compact space like U G, except in advanced courses in classical
mechanics (e.g. . the orientation of a rigid body). In quantum mechanics we
know that the
Pmomentum is related to translations, so the ordinary kinetic
energy term i p2i /2m is connected with a space of qs which is translation
invariant in each pi . Our kinetic energy is not of that form, as the E s
dont commute and the / s do not enter H in this simple fashion. This
is consistent with the fact that our space of Us is just not a Euclidean
N 2 1 dimensional space. Now quantum mechanics can be viewed as a
functional integral over all possible configurations, which requires a weight
for the configurations. In a deep sense this weight is connected with the
kinetic energy term in the Lagrangian density. We can sidestep this issue
(temporarily) by observing that the only sensible way to integrate over the
Us is in a group invariant way. In terms of eigenstates |Uab i of U (on a given
link), we need to be able to write
Z
1I = d(U) |Uab i hUab | ,
or in termsRof the parameters R , d(U) = d( ), the measure must be group
invariant, d()f (U()) = d()f (GU()) for any function f : G C
and any fixed element G in the group. This is also what we needed long ago
to define the orthogonality of group representations.
So now we leave physical applications and turn to the problem of finding
the Haar or Hurwitz measure d( ) satisfying the invariance.
Let us write d() = h()d 1 d n for an n dimensional Lie group. For
j
a fixed G and , the group element U = GU() = U( ) = ei Lj , where
the j are functions of the s (and also of G, but we are holding G fixed).

156. Last Latexed: April 4, 2016 at 14:56

Joel A. Shapiro

If we change variables on the right hand side of our invariance equation, we

get
 j
Z
Z


n
  n
h()f (U()) d = h()f (U( ))  k  d

where |/| is the Jacobian determinant. Invariance for arbitrary functions

clearly requires
 j
 

h( ) = h()  k  .

Let h(0) = c. Then h at any other point is determined by taking G = ei

 k 1
 
,
h() = c  j 
=0

where ei Lk = ei Lk ei Lk .
This is a necessary condition for the
invariance. To show it is also sufficient,
we need to show that the relation holds
for all . Consider the measure at two
points, and . We need to show that
if U( ) = GU( ) for a fixed G,
 



 
.
h( ) = h( )  
=
k

jL
j

Let G = ei Lk and G = ei Lk and U( ) = G U(v) and U( ) = G U().

Then U( ) = G (G )1 U( ) = GU( ), so G = G (G )1
 
    

 

 
 
 
h( )
c
c


 =



=
=
 
 
 
h( )
h( )
h( )
0
0
=
which verifies the invariance in general
k
k
k
Now we need to evaluate the Jacobian at = 0. As ei Lk = ei Lk ei Lk ,


j ik Lk i k Lk 
j ik Lk
i k Lk 
=
=
e
e
e
e
iLj


k
k
j
k
=0
=0
Z 1

i(1) k Lk
i k Lk 
=
d e
iLj e

0

618:

so

Last Latexed: April 4, 2016 at 14:56

j
Lj =
k

Thus

i j Lj

de

i j Lj

Lk e

h() = c det = c det

1 ei S(Lm )
=
i m S(Lm )

157

Lj .

kj


m
ei S(Lm ) 1
.
i m S(Lm )

For any fixed (real) , m S(Lm ) is a hermitean matrix and can be diagonalized. If j are the real eigenvalues, we have
h() = c

Y eij 1 Y sin(j /2) Y

=
eij /2 .
i

/2
j
j
j
j
j

But m S(Lm ) is also imaginary, while still hermitean, P

so it is antisymmetric
and has zero trace. This is preserved by similarity, so j j = 0, and
h() = c

Y sin(j /2)
j

j /2

Let us consider an explicit example, SU(2). Any U = ei~ ~ /2 is equivalent

(by similarity) to one along the z axis, with the same ~ 2 . Then j S(Lj ) =
||S(L3 ). S is the adjoint representation, so L3 has the eigenvalues 1, 0.
The zero eigenvalue gives 1 in the product (recall the note on the definition
of the [] expression) and the others each give 2 sin(||/2)/||, so all together
sin2 (||/2)
h() = 4c
.
2
Choosing c = 1, we may take the invariant volume to be
Z
2
1
2
3 4 sin (||/2)
.
d d d
||2
||2
If we used spherical coordinates for
~ , this gives
Z 2
d d2 4 sin2 (||/2),
0

where d = sin d d as usual. We note that the volume with || 2

3 , the same as the volume with || , and not times the ordinary
is 4
3

158. Last Latexed: April 4, 2016 at 14:56

Joel A. Shapiro

area of the sphere, that is, 4(2)2 . The metric says points with 2,
even in very different directions, are close
 together,
 because for = 2, all
1 0
these s correspond to the same U =
.
0 1
Some comments:
(a) For SU(3), not all directions are equivalent. So there is no equivalent of
||.
(b) For many applications, we needed the existence of the invariant measure
but not the actual expression for it. For example, in our orthogonality
proofs.
(c) Most discussions of SU(2) parameterize the group in terms of Euler angles rather than in the way we did, the angle of the rotation and the
direction of the axis.
(d) Note that the factor in h is just the determinant of the matrix which
connects the left derivatives to the ordinary derivatives. This is not
an accident, but rather has to do with this relation defining a metric
(not just a measure) on the group space, and the measure given by the
determinant of the metric.