You are on page 1of 19

Legendre Transformation

sci.physics, sci.math
1997 August 28 01:23:30 GMT

This is a follow-up on a reply given by Chris Hillman [1,2] to the question “what is the Legendre Transformation”. It
may also be regarded as a pedestrian summary of LNP 107 [3].

Hillman’s reply was not very clear, did not touch on the main point of the original poster’s question, and only applied to
a limited case of the general issue being asked about in the original. Therefore, I expanded on and generalized the reply,
making it more directly relevant to the question, while also clarifying what Hillman said. A revision of Hillman’s reply
is thus also included, with the notation altered to match the intervening discussion (and to match what is more commonly
in use).

One of the most important additions made, here, is the expansion of the discussion to field theory, á la LNP 107. So, I’m
also going to describe the two Legendre transforms in common use there – those associated with the Hamiltonian and the
de Donder-Weyl Hamiltonian. This is illustrated with the 1-mode scalar field, whose treatment is generic and applicable
(as an enveloping generalization) to all 1-mode scalar fields whose dynamics are describable by a Lagrangian. This leads
to a brief discussion of classical renormalization.

Finally, the Lagrangian and Hamiltonian formalisms are special cases of a much more general and robust approach that
is firmly tied to symplectic geometry and Poisson manifolds. Here, the Legendre transform is seen as a special case of a
more general relation that pertains to “canonical” transformations. This is discussed briefly at the end, following the
revision of Hillman’s article; where we also show how the complementary set of p ‘s and q ‘s arise as a consequence
within this formalism, and not as axioms. (Hint: the Darboux Theorem). This is applied to classifying all the “elementary
systems” for non-relativistic theories – the classification being into tardions, synchrons and vacuons – and in deriving
(rather than postulating) their coordinate representations in terms of complementary variables.

Richard writes:
I’m trying to study Hamilton’s Canonical equations but I need to first master Legendre Transformations. I’ve tried
several sources but find them all mysterious (e.g., Arnold, Lanczos, Goldstein, Chow). Can someone offer a clear
explanation of what a Legendre Transformation is along with a simple proof. Please be as intuitive as possible without
loss of mathematical rigor. Any help is very much appreciated. Thank you.

1. The General Form of a Lagrangian System


In a (first-order) Lagrangian formulation of dynamics for a given system, there are 4 types of quantities that make their
appearance:
 the configuration coordinates  q a : a  0, , N  1
 the configuration velocities  v a : a  0, , N  1
 the conjugate momenta  pa : a  0, , N  1 , and
 the conjugate forces  fa : a  0, , N  1 .
The  q , v
a a
 , together, comprise the kinematics of the system and are governed by a kinematic law,
dq a
va .
dt
The  pa , f a  comprise the dynamics and are governed by a dynamic law,
dpa
fa  .
dt

The kinematic and dynamic variables, and the kinematic and dynamic laws, together, give us the basic elements that
apply universally to any N degree of freedom system describable by a Lagrangian. The features specific to a particular
system, that distinguish it from other N degree-of-freedom systems, are encapsulated entirely within the constitutive
relations. These relate the dynamic variables to the kinematic variables. For a Lagrangian system, the constitutive
relations arise from a single function, the Lagrangian L  L  t , q, v  , via the equations
L L
pa 
, fa  a .
v a
q
The function L is therefore referred to as a generating function for the system’s dynamics.

If one were to formalize this mathematically, one would start with a 1-dimensional continuum1 M :  t  , attaching at each
time t a copy Qt :  q a  of configuration space; resulting in a configuration bundle Q :  t , q a  . The system is represented
as a section q :  t   M t, q t   Q on the bundle (i.e., its “trajectory”), while the kinematic law is effected by first
a

broadening the configuration space to a jet bundle J Q :  t , q , v  and then extending the system’s trajectory to its jet
1 a a

prolongation j q :  t   M
1
t, q t  , q t  .
a a

The dynamic variables would be obtained by broadening the configuration bundle to a phase bundle P :  t , q a , pa  and
this, in turn, to a Legendre bundle J 1 P :  t , q a , pa , v a , f a  . Within this jet bundle, the “dynamic law” can be similarly
encapsulated.

2. Dynamics with Everyone on Equal Footing


Other generating functions can be produced that yield the same dynamics, by posing other constitutive laws that take
different subsets of these variables as the independent ones. The question then arises: how to correctly transform between
different sets of independent variables so as to preserve the dynamics. The key feature is captured by the differential of
the Lagrangian, L  f a q a  pa va .2 The expression on the right is the canonical 1-form   f a q a  pa va . The
condition above stipulates that it be an exact differential. Formally, when subject to the kinematic and dynamic laws, this
reduces to a total time derivative
 dq a  d a a d
   p q  
dp
  a q a  pa      pa q a .
dt  dt  dt dt
In fact, as we’ll see: this relation is not merely “formal” but can be expressed as a derivative – a Lie derivative.

The relation   L can be thought of as residing in the space of differential forms spanned by the differentials of the
fiber coordinates V *t J 1 P :  q a , pa , v a , f a  (i.e. within the “vertical bundle” V * J 1 P ). The relation says that a certain
first-order differential form is exact. If we take the exterior derivative of it, we obtain 0 as a result,
0  2 L  f a  q a  pa  v a .
This is the master relation that governs all the possible first-order relations that can be written in terms of the differentials.
It’s the ground out of which all generating functions are forged.

The second-order differential


  f a  q a  pa  va

is the canonical 2-form. The dynamics are defined by the condition   0 . Stated in this form, the condition is
independent of which variables are treated as dependent and which as independent – all 4 sets of quantities are on equal
footing.

LNP 107 treats the dot only as a formal derivative, but as pointed out above, it’s not just formal. Both  and  can be
represented as a Lie derivatives:

1We will present manifolds, here, by listing their natural coordinatizations alongside.
2The summation convention is used both here and below. Differentials on the fiber coordinates are presented with  , for the sake of intuitive appeal;
where we’re always thinking of these in the context of the variational problem.
  Ld dt ,   Ld dt ,
namely, the Lie derivative with respect to the vector field
d   
  va a  fa .
dt t q pa
The 1-form and 2-forms they are respectively derived from are
  pa qa ,   pa  q a .
The condition Ld dt   0 then identifies d dt as a “canonical field” or “symmetry” of  . The closed 2-form  is the
symplectic form associated with the dynamics; the underlying geometry of the phase space being that of a symplectic
manifold.

The question of how one actually arrives at the  q a , pa  coordinates for a phase bundle, in the first place, along with the
construction of  , is addressed by way of a couple examples at the end of this article, where we deal with the more
general question of symmetries.

3. Changes of Variables – Legendre Transform


The Legendre transform passes over from one set of independent variables to another and converts the corresponding
generating functions appropriately.

If we take the variables  q a , pa  as independent, then we can integrate out the relation
0  f a  q a  pa  va  f a  q a  v a  pa    f a q a  v a pa  .
This produces an exact differential for another generating function
H  f a q a  v a pa ;
which is the Hamiltonian. Under this change of independent variables, the constitutive laws now take the form
H H
fa   a , va  .
q pa
The appropriate conversion is determined by comparing the respective first-order differential relations,

L  f a q a  pa v a 
  L  H  pa v  v pa    pa v  .
a a a

H  f a q  v pa 
a a

Thus, up to choice of constant, we have
L  H  pa v a .

Making the arguments of the functions explicit, this actually expresses two relations, rather than one:
L  t , q, v  t , q, p    H (t , q, p)  pa v a t , q, p  , L t , q, v   H  q, p t , q, v  , t   pa t , q, v  v a .
The first relation passes over from L  H and requires that we can solve the equations
L
pa  t , q, v   a  t , q, v   v a  v a t , q, v  .
v
The necessary condition for this, of course, is that the matrix of second derivatives
2 L p p
mab  a b  ba  ab
v v v v
be non-singular. These are the coefficients of inertia. Hence, a Legendre transform from velocities to momenta requires
that there be no zero-mass modes in the system. A system having such modes possesses first class constraints.

The second relation passes over from H  L and requires that we can solve the equations
H
v a  t , q, p   t , q, p   pa  pa t , q, v  .
pa
Here, the necessary condition is that the matrix of second derivatives
2 H v a vb
wab   
pa pb pb pa
be non-singular. These are the coefficients of dispersion and are the inverse of the coefficients of inertia. The requirement
comes down to the condition that there be no infinite-mass modes. In a quantum system, this matrix also gives you the
classical limit of the coordinate-velocity commutators
 qˆ a , vˆb 
 q a , vb   wab .
i
Thus, each infinite-mass mode corresponds to an essentially classical degree of freedom, defining what’s referred to as a
second class constraint.

Any other combination of variables can be taken as independent, provided they enable one to construct a first-order
differential identity from the second canonical form  . For instance, one can take
0  f a  q a  pa  va  q a  f a  v a  pa    q a f a  v a pa 
which leads to a generating function (the “dual Lagrangian”)

J  q a f a  v a pa  J J
  qa ,  va .
J  L  q f a  v pa   f a pa
a a

Or, one can take


0  f a  q a  pa  va  q a  f a  pa  v a    q a f a  v a pa 
leading to the generating function
 K
K  q a f a  v a pa  K
  q a ,  pa .
L  K  q fa  f a
 v a
a

In addition, one can take other combinations that mix and match some coordinate with some conjugate forces, and some
velocities with some momenta.

Example 1 (The Kepler Problem)


The archetypical Lagrangian dynamics problem is the Kepler problem, which is governed by the dynamics
r
mv  p,  3  f
r
where
r   q1 , q 2 , q3  , v   v1 , v 2 , v3  , p   p1 , p2 , p3  , f   f1 , f 2 , f 3  ,

r  r r  q   q   q 
1 2 2 2 3 2
, m  0,   0.

First, we note that the Legendre transform between the Lagrangian L and the Hamiltonian H modes is non-singular
with the coefficients of inertia and dispersion given by
p v a 1 ab
mab  ba  mab , wab   
v pb m
are expected (where ab  ab is the Kroenecker delta).

As we know, these constitutive relations indeed define a Lagrangian system. The generating function can be produced
by integrating out the corresponding first order differential relations. In Lagrangian mode, we have
r  mv 2   mv 2 
L  p  v  f  r  mv  v  3  r     L  .
r  2 r 2 r
In Hamiltonian mode, we have
p r  p2   p2 
H  v  p  f  r   p  3  r     H   .
m r  2m r  2m r
The transformation between the two is verified explicitly by
p 2 mv 2 p  v p  v
LH      p v .
2m 2 2 2

Interestingly, the other modes also produce cohesive results. We can solve configuration coordinates in terms of the
conjugate forces
  v  f
 f r  ,  3 f r  .
r2 f r f f
Then, we have
 f p  p2  p2
J  r  f  v  p    f   p  d  2 f    J  2 f  ,
r f m  2m  2m
 f  mv 2  mv 2
K  r  f  p  v   f  mv  v    2 f    K  2 f  .
r f  2  2
Again, the transformation may be verified explicitly
mv 2 p 2  
J L    2 f  p  v  f  r, L  K   2 f  f  r .
2 2m r r

4. Generalization to Field Theory


All of the above generalizes to field theory. The only real change is that in place of one “evolution parameter” t , there
are as many  x :   0, , n  1 as there are dimensions n in the underlying “evolution space” M :  x  . Hence,
mechanics (with time-dependent Lagrangians) is “field theory” over a 1-dimensional evolution space. Static field theory
does not involve time, but still entails the description of the “unfolding” of the field over space. So, here the evolution
space is 3-dimensional, comprising the coordinates r   x1 , x 2 , x3  for space. Field dynamics does involve evolution over


time as well, and so has an evolution space that is 4-dimensional, x0  t ,  x1 , x 2 , x3   r . 
Now, the roster of players is upgraded to the following 4 types of quantities
 the configuration coordinates  q a : q  0, , N  1 , as before,
 the configuration velocities  va : a  0, , N  1;   0, , n  1 ,
 the conjugate momenta Pa : a  0, , N  1;   0, , n  1 , and
 the conjugate forces  Fa : a  0, , N  1 .
The “velocities” are now gradients; correspondingly, the momenta now acquire an extra index. Both the momenta and
forces are now densities. All the quantities naturally arrange themselves in terms of differential forms over the evolution
space
q a , va  va dx , pa  Pa dx , f a  Fa dx .
The densities are the coefficients for the volume n -form dx  dx0   dxn 1 , and the  n  1 -forms defined as its
contractions
dx   _| dx   1 dx0    0, , n  1 .

 dx1  dx1   dx n 1

The kinematic and dynamic laws become, respectively


 q a  va ,  Pa  Fa
or, in the language of differential forms, v a  dq a and f a  dpa . One now also has a generalized form of the Bianchi
identity,
dva  0   va    va  0 .
The Lagrangian becomes a density L  L  x, q, v  associated with the n -form L  Ldx , and its variation now assumes
the form
L  Pa va  Fa q a , L  va  pa  q a  f a .
Note that we now have a mixture of coordinate differentials dx , and differentials of the fiber coordinates
q a , va  va dx , pa  Pa dx , f a  Fa dx .
The wedge product is taken to apply to both sets of differentials. The 2-form  now becomes
  Pa  va  Fa  q a .
This is now also a density and is the kernel of the n -form
dx  pa  va  f a  q a .
This works out consistently, if we take the coordinate differentials and fibre coordinate differentials to also anti-commute
with one another, as well as stipulating so for each set with one another amongst themselves.

As for the univariate case, the overall condition is that   L , where the canonical 1-form now becomes
  Pa va  Fa q a .
Again, under application of the kinematic law and dynamic law, this reduces to a total differential – this time taking on
the form of a divergence:
  Pa    q a     Pa  q a   Pa q a     Pa q a dx  q a  pa .

Example 2 (The General 1-Mode Scalar Field)


The archetypical Lagrangian field dynamics is scalar field theory. The variables, here, are
  q 0 ,   v00 , A   v10 , v20 , v30  ,
  F0 ,   P00 , J   P01 , P02 , P03  .
The kinematic and dynamic laws become, respectively,
   
   , A    , J   ,
 t  t
and the Bianchi identity,
 A 
 × A  0,    0 .
 t 

The variational of the Lagrangian can then be written as


L    J  A   .
The canonical 2-form becomes
      J  A     ,
where, here, the 3-vectors J and A are combined with both the dot product and wedge product.

If we take the underlying space M to be a 4-dimensional Minkowski space, then we would assume that the dynamics
are Lorentz-invariant and are given by a Lagrangian whose only functional dependence on the kinematic variables is
through their Lorentz invariant combinations:
2 2  c 2 A2
I1  , I2  .
2 2
In that case, the variational reduces to the following
L  I 2  I 2 ,
where each of the two constitutive coefficients     I1 , I 2  and     I1 , I 2  are also functions of the invariants.
Substituting the invariants into this expression, we then obtain
L      c2 A A
thus yielding the following constitutive laws
  ,   , J  c2 A .
The field equation, itself, may be obtained by substituting the kinematic law into these expressions and then substituting
these into the dynamic law; the result being:
    2
    c         0 ,
t  t 
or a Sturm-Liouville equation.

5. Legendre Transform for Field Theory


5.1. The 3+1-Dimensional Hamiltonian
The fact that we can take different combinations of variables as independent comes fully into play here. The form of
Hamiltonian field theory used in Physics swaps only one of the gradients for its conjugate momentum:    . The others
are kept intact.3 This leads to the following first-order differential equation
      J  A         JA     H    JA   .
Thus, we arrive at the equations for the Hamiltonian,
H H H 
   ,  J  c 2 A,    .
 A  
Upon integration, this yields,
   2  1  2   A2c 2 
H    c 2 A  A              .
  2   2  2 

If the constitutive coefficients are constants, then this reduces to the quadratic expression
 2 2 A2c 2
,  constant  H     .
2 2 2
If the coefficients are not constant, we can still reduce  to a constant by renormalizing the field variable  ; i.e., making
an adjustment   Z  by a suitable variable factor Z . This works as long as  is bounded away from 0. The exceptional
case occurs where   0 , which may be identified as a generalization of the Landau Pole phenomenon.4

5.2. The deDonder-Weyl Hamiltonian


On the other hand, one can swap all the components of momentum and gradients. The result is the deDonder-Weyl
Hamiltonian. Unlike the other, more familiar, field Hamiltonian, this one places all the coordinates  x  on equal footing,
and does not even require that any of the dimensions be “time”. Hence, for instance, if could also be used to describe
statics, or “dynamics” in a 4-dimensional space-“time”, where the time dimension has become spatial5.

The variational, in general form, is given by the deDonder-Weyl Hamiltonian density H  H  x, q, P  and its
corresponding n -form H  Hdx by
H  Pava  Faqa  H  q a  f a  va  pa .
Note that the sign on the last term switches back to – because of the transposition of the 1-form Pa over the 1-form v a
. The corresponding equations are
H H
 Fa   Pa ,  va    q a .
q a
Pa
This Legendre transform generalizes the relation between the Lagrangian and Hamiltonian to:
L  H  Pava  L  H    v a  pa  .

3 Technically, this means that we’re actually dealing with what’s called a Routhian, not a Hamiltonian. A Routhian is whatever results from doing a
partial, but not complete, Legendre transform on the velocity and momentum coordinates.
4 This exercise also helps to show that though such concepts as the Landau Pole and renormalization are normally presented in t he context of Quantum

field theory, they actually have nothing to do specifically with quantum anything. Rather, they are classically grounded and properly regarded as a part
of classical field theory.
5 Which, if you believe Hawking, Hartle, et al., is what characterizes the transition that takes place at the initial cosmological singularity, or “Big Bang”.
Example 3 (One-Mode Scalar Fields)
Returning to the scalar field dynamics example, these developments lead to the following
      A  J         A  J     H    A  J  
resulting in the equations
H    H H 
        J  ,  A  ,   
  t  J  t
leading to the following integral
  2  1  2c 2  J 2 
H        .
 2    2c
2

Assuming the coefficients are constant, this reduces to
2 1  2 J 2 
,  constant  H      2  .
2 2  c 

6. Form-Valued Fields and the “Exterior Jet Bundle”


As already suggested by the notational conventions we adopted, the treatment for field dynamics can be brought close in
line with the univariate case, provided we write the field quantities in the language of differential forms. This leads
naturally to the generalization in which the configuration coordinates  q a  are, themselves, taken to be form-valued.

As far as I’m aware, there is no literature reference available that fully embodies this idea. What it requires is that in place
of the Jet Bundle, one takes the “exterior jet bundle”. Thus, if we start with a bundle whose natural coordinates are
Q :  x , q a  , where the  q a  are now form-valued, then the corresponding jet bundle would have coordinates
J d Q :  x , q a , v a  . A section q :  x   M  x , q  x   Q
 a
then extends to its “exterior jet prolongation”
j d q :  x   M   x , q a  x  , dq a  x    J d Q . This then serves to formalize the “kinematric law” v a  dq a . Again, we
obtain a version of the Bianchi identity, dv a  0 .

The Legendre transform then takes place over the phase bundle, P :  x , q a , pa  and its exterior jet bundle
J d P :  x , q a , pa , v a , f a  . This serves to additionally formalize the “dynamic law” f a  dpa . Finally, a feature that is novel
to the extension of the configuration coordinates to form-valued quantities, is that the forms  fa  are no longer of
maximal degree. Consequently, alongside the Bianchi identities dv  0 are a second set of identities that express a
a

“conservation law”: the continuity equation df a  0 .

Assume that  q a  are homogeneous forms of degree q . The canonical 2-form  then generalizes to
  q a  f a   1 v a  pa .
q

Its integrals yield the Lagrangian and deDonder-Weyl Hamiltonian


L  q a  f a   1 va  pa , H  q a  f a  v a  pa .
q

Upon subtraction, we obtain


  L  H    1 v a  pa  v a  pa   1   v a  pa   L  H   1 v a  pa ,
q q q

which generalizes the Legendre transform.

Example 4 (1-Form Fields and Maxwell’s Equations)


An important example occurs where q  1 : the 1-form field. For the base space M , we adopt the following coordinates
t  x0 , r   x, y, z    x1 , x 2 , x3  ,
and use the following conventions for the coordinate differentials
dr   dx, dy, dz  , dS   dy  dz, dz  dx, dx  dy  , dV  dx  dy  dz .
The variables, here, can be represented as
q a  Aa  A a  dr  a dt , v a  F a  Ba  dS  Ea  dr  dt ,
f a  J a  a dV  J a  dS  dt , pa  Ga  Da  dS  H a  dr  dt.
The kinematic and dynamic laws become, respectively:
 Ba    Aa     Da  a 
   
dq  v   a
a a
 A a
 ; dp  f    D  ,
 × H a  t  J a 
a a
E   
a a

 t 
while the Bianchi identity and continuity equation, respectively, become
   Ba  0 
   a 
dv  0  
a
B a
 , df a  0      J a  0 .
  E 
a
 0  t 
 t 
Together, these comprise one set of “Maxwell equations”
dAa  F a , dF a  0, dGa  J a , dJ a  0
for each value of the index a  0, , N  1 ; which gives us the general form of a dynamics for what is known as an
Abelian gauge field.6

The variational of a prospective Lagrangian L  Ldt  dV and Hamiltonian H  Hdt  dV reduce, respectively, to the
forms
L  Aa  J a  F a  Ga , H  Aa  J a  F a  Ga
resulting in
L  Aa  J a  aa  Ba  Ha  Ea  Da , H  Aa  J a  aa  Ba  Ha  Ea  Da .
For the Lagrangian form, this leads to the following constitutive laws
L L L L
a   a , J a  , Da  a , H a   a ,
 A a
E B
while for the Hamiltonian form, we obtain the following as the “Hamilton’s equations”:
H H Da H Aa H
     D ,   J    H  ,  E a
  a
 ,  B a    A a .
 A t Da t H a
a a a a a a

The Legendre transform is, here, expressed as


L  H  F a  Ga  L  H  Ba  Ha  Ea  Da .

An example in which a partial Legendre transform occurs is the one which historically arose in connection with
Maxwell’s development of electromagnetic theory. In this form, the independent variables are chosen to include the
“potentials”  a , A a  and the “force” fields  Ea , H a  , while the “induction” fields  Da , Ba  are considered as derived,
and the “sources”  a , J a  as either derived or external to the dynamics.

Here, instead of a Lagrangian or Hamiltonian, one has a Routhian R , which is related to the Lagrangian by
R  Ba  H a  L ,
thus yielding the following variational
R  Aa  J a  aa  Ba  Ha  Ea  Da .

Example 5 (Constitutive Relations: Isotropic Fields)


In a similar vein to example 2, we may enquire what the most general constitutive law is for 1-form fields that respects
certain symmetries. For isotropic fields, the symmetry in question is rotational symmetry for the spatial coordinates r .
For the sake of simplicity, we will restrict our attention to 1-mode fields, suppressing the extra index a . In addition, we

6As we will shortly see, these variables also suffice to handle non-Abelian gauge fields, but it is common to adjust the kinematic and dynamic laws by
directly incorporating gauge invariance into them.
will also assume that the field is gauge invariant; i.e., invariant under the transformation A  A  d  . Then, it follows
that the Lagrangian can only depend on scalar combinations of the force and induction fields:
 E2  1  B2 
L        E  B     .
 2    2 
In the Routhian formulation, this corresponding variational takes on the form
 E2   H2 
R        E  H     .
 2   2 
These yield the following sets of constitutive relations
D  E  B 
  D  E  H 
L : B ; R :  ;
 H   E   B  E  H 
  
the two sets of constitutive coefficients being related by
    2 ,    .

Historically in electromagnetic theory, the coefficients  and  were introduced by Maxwell to draw an analogy with
the spring coefficient k and mass m for a simple harmonic oscillator. They are, respectively, the dielectric coefficient
and permeability. In the Lagrangian formulation, the coefficient  appears in place of  . This is the permittivity. Finally,
the remaining two coefficients  ,   are present only if the field dynamics fail to be invariant under parity r  r . For
gauge fields, the coefficient  assumes a fundamental importance, underlying what are known as the theta-vacuua. In
contrast, the electromagnetic field is assumed to respect parity.

Finally: the Legendre transform from the Lagrangian to the Routhian modes requires that 0     . To make a full
Legendre transform to the deDonder-Weyl Hamiltonian, we need to be able to express  B, E  in terms of  D, H  .
Working from the Routhian form of the constitutive laws, we have
D  H   2  D  H
E , B  D    H  .
     
This shows that we need for the dielectric coefficient to be regular: 0     .

Example 6 (Constitutive Relations: Gauge-Invariant Fields)


Gauge fields are invoked to localize the symmetries that may apply to other fields. Suppose we subject a form-valued
field q to a gauge symmetry q  q and assume that a similar symmetry v  v should apply to its field velocities
v . In order for this to be consistent with a variable symmetry generator  , it is necessary to adjust the definition of v
from v  dq to v  dq  Aq , with the addition of a gauge field A whose transformation will provide the appropriate
compensation:
v  v    dq  Aq     dq  Aq   d  q    A q  A q     dq   Aq .
Then, expanding d  q     d   q    dq  , this yields the following relation
 A q   d   A  A q .
In order for this to apply universally for all possible fields, q , we should then require the following transformation
property
A  d   A  A .
If the symmetry arises from a (finite dimensional) Lie algebra in which A  A   A,  , it may be decomposed with
respect to a given basis Ya  by    aYa and A  AaYa , and we may express the transformation property in component
form by
Ac  d c  f abc Aab
where the structure coefficients  f abc  are defined by the Lie bracket relations Ya ,Yb   f abc Yc .
From this, we may infer the following transformation for the field velocities F  dA ,
F  d  A 
 d  d   A  A 
  dA   A  d   d   A    dA
 F   F  A  d   d   A.
In component form, this becomes
F c  f abc  F ab  Aa  d b  .
The foregoing may then be applied to seek out the most general gauge-invariant Lagrangian for a gauge field. Under a
gauge transformation, we have
L  Ac  J c  F c  Gc
  d  c  f abc Aa b   J c  f abc  F a  b  Aa  d  b   Gc
 d  c  J c  f abc Aa b  J c  f abc F a  b  Gc  f abc Aa  d  b  Gc .
With a readjustment of indices and reordering of terms, this becomes
L  d b   J b  f abc Aa  Gc   b f abc  Aa  J c  F a  Gc 

   
 d b   J b  f abc Aa  Gc   b f abc  Aa  J c  F a  Gc   d  J b  f abc Aa  Gc  .
Separately equating the order 0 and 1 terms to 0, this leads to the following two conditions
J b  0, f abc F a  Gc  0 ,
where
1
F c  F c  f abc Aa  Ab , J b  J b  f abc Aa  Gc
2
comprise the above-mentioned adjustment of the kinematic and dynamic variables. Note that under this adjustment, we
have the following as the Lagrangian variational
L  Ab  J b  F c  Gc
 
 Ab   J b  f abc Aa  Gc     F c  f abc Aa  Ab   Gc
1
 2 
 Ab  J b  f abc Ab  Aa  Gc  F c  Gc  f abc  Aa  Ab  Aa  Ab   Gc
1
2
 Ab  J b  F c  Gc .
The cubic terms cancel on account of the anti-symmetry of the structure coefficients, f abc   fbac .

In vector form, the modified kinematic and dynamic laws read:


 c 1 c a b 
 B    A  2 f ab A  A 
c

1
F c  dAc  f abc Aa  Ab   ,
Ec  c  A  f c a A b 
c
2
 t
ab

   Db  f abc A a  Dc  b 
 
dGb  f ab A  Gc  J b  
c a
Db ;
 × H b   f ab  A  H c   Dc   J b 
c a a

 t 
and the “Bianchi identity” and “continuity equation” become:
   B c  f abc A a  Bb  0 
 
dF  f ab A  F  0  
c c a b
B c
,
  E 
c
 f ab  A  E   B   0 
c a b a b

 t 
  
dJ b  f abc Aa  J c  f abc F a  Gc   b    J b  f abc  A a  J c  a c   f abc  B a  H c  Ea  Dc  .
  t 
With this modification in mind, the constitutive laws from Example 5 will apply intact, with suitable generalization for
the indices
Da  abEb  abBb , Ha   1  Bb  abEb .
ab

As with Example 5, we’re assuming the regularity of the permeability matrix     ab  and the Legendre transform
requires the regularity of the dielectric coefficient matrix    ab  , whose components are given by
ab  ab  cd acbd .

It bears to mention here that these are not the most general constitutive relations. A feature novel to non-Abelian gauge
fields is that invariants also exist involving cubic combinations of  B, E  . This leads to the addition of extra terms
quadratic in  B, E  into the constitutive relations for  D, H  .

A point worth making here is that the LNP 107 takes the above-mentioned formalism beyond the application of dynamics,
citing examples from statics and one from thermodynamics. This serves to illustrate, that the phenomenon underlying the
Legendre transform generalizes beyond dynamics (as does the rest of what’s described above). That’s part of the point
of bringing up the case, n  3 , of statics where systems “unfold” in 3-space, rather than “evolve” in 1-dimensional time
or 4-dimensional space-time.

7. Hillman’s Reply, Revised


On second thought, I think I’ll try to explain two geometric ways of understanding Legendre duality (c.f. [4]).

Definition: Let L : R  R be differentiable and concave up. Define


H  L* : p  R sup  vp  L  v   .
vR

The concavity of L is the following condition on the second derivative


m  L   v   0 .
More generally, the requirement is that the coefficients of inertia m be bounded away from 0. So the case m  L   v   0
is also understood to be included in this, modulo a switch of sign.

Lemma:
 L  c  H  c
*
(a)
(b) H  p   vp  L  v  , where p  L   v  .
Proof:
(a) First, note that convexity is preserved under addition by constants. Thus, we may write
 L  c   p   sup  vp  L  v   c   sup  vp  L  v    c  H  p   c .
*

xR xR

(b) Note that


v  L  c   v    L  c  v   vL   v   L  v   c  H  L   v    c   L  c   L   v   .
*

Since  L  c   v   L   v  , the proposition is translation-invariant. So, normalize L by subtracting out L  0  , so that we


have L  0   0 .

So now we have
vL   v   L  v    vL   v   L  v     0L   v   L  0     VL   v   L V   dV    L   v   L  V   dV .
v d v

0 dV 0

This is the maximum value for VL   v   L V   over all V  R . For V  v :

 vL  v   L  v   VL  v   L V     L  v   L  v  dv  0
v

since L  V   L   v  by convexity. For V  v :

 vL  v   L  v   VL  v   L V   L  v  v  V    L v   L V     L v   L v  dv  0
V

using the expansion above for vL   v   L  v  .

Graphically,
 L V  is the area between  p  L   v  , v  V , p  0  in the vp plane.
 H  P  is the area between  p  L   v  , p  P, v  0  in the vp plane.

A result of the lemma is:


Corollary: For all v, p ,
 vp  L  v   H  p  , with equality if and only if p  L   v  .
 p  L   v  if and only if v  H   p  .
Proof:
Defining m  v   L   v  , we have

  vL   v   L  v    vm  v  .
dp d
H   p
dv dv
Thus
dp dp
p  L v   m v  
 H   p   v, H   p   v  m  v    p  L v  .
dv dv
The constant of integration in the second relation is fixed by our stipulation that L  0   0 . Indeed, we could have written

H   L v   H   p  c    H  p  c  H   p .
d
dp
Thus, we have the result
p  L v   v  H   p  .

Lemma: L is convex, if and only if H is convex.


Proof:
This is a direct consequence of the result, just established, that L  and H  are inverses. The condition that L and H
be convex, is that their respective derivatives be increasing. The inverse of an increasing function is increasing, thus their
respective derivatives
m  L   v   0, w  H   p   0
are what had been previously referred to as the coefficient of inertia and coefficient of dispersion, respectively.

Therefore, we can define a second transformation H *  L** , which leads us to the following result:

Theorem: (Duality property)


H*  L .
Proof:
Suppose v  H   p  and p  L   v  . Then
H *  v   pH   p   H  p   pv   vL   v   L  v    L  v  .
This recovers a first understanding of the Legendre duality.
There is another way to understand Legendre duality. Consider the set of points (the line) given by
T  pv  E .
This can also be written as
E  pv  T
which expresses E as a function of p . Whereas, l  p, E    v, pv  E  : v  R defines a line l  p, E  as a set of points,
l  p, pv  E  : p  R defines a set of lines, given by the function E  h  p   pv  T . These are the lines that pass
through  v, T  .

More generally, given a curve T  L  v  , define a similar function E  h  p  that yield the lines tangent to T  L  v  .
This set yields the envelope of the curve and provides an equivalent way to represent the curve.

The tangent through  v0 , L  v0   is just T   v  v0  L   v0   L  v0  , or


T0  pv  E , where p  L   v0  and E  v0 L   v0   L  v0  .
Thus
E  h  p   v0 L   v0   L  v0  , where p  L   v0  ,
or
E  H  p .

The Legendre transform provides a duality between these two representations of the curve; one, as the set of points
T  L  v  , and the other as the envelope, E  H  p  .

The first definition of Legrende dual H  L* provided above can be applied to any convex function. Then it’s called the
Legendre-Fenchel dual. But the inequality vp  L  v   H  p  has to be proven some other way, though it is still true.

Example 7
A. Consider the following 5 cases,
1 2 3 4 5
2 3
v v
L v , L  v   , L  v   log v, L  v   v log v, L  v   v log v  1  v  log 1  v  .
2 3
B. From the relation p  L  v  , we obtain the following expressions for p ,

1 2 3 4 5
1 1 
p  v, p  v2 , p , p   log v  1  log 1  v   1  log   1 .
p  1  log v,
v v 
C. Inverting, we obtain the following as equations for H   p   v
1 2 3 4 5
1 1
H   p   v  p, H   p   v  p, H   p  v  , H   p   v  e p 1 , H   p   v  p .
p e 1
D. Solving for H  p  , we obtain (up to constants of integration)
1 2 3 4 5
2
, H  p   p3 2 , H  p   1  log p, H  p   e p 1 , H  p    log 1  e  p  .
p 2
H  p 
2 3
E. Finally, in each case we may verify that L  v   H  p   vp , given the relations listed in B and C:
v 2 p 2 vp vp
1: L v  H  p      vp,
2 2 2 2
v 3 2 p p vp 2vp
2: L v  H  p      vp,
3 3 3 3
3: L  v   H  p   log v  1  log p  1  log vp  1  vp,
4: L  v   H  p   v log v  e p 1  v  p  1  v  vp,
1   v 
5 : L  v   H  p   v log v  1  v  log 1  v   log 1  e  p   v log   1  log 1  v   log 1    vp.
v   1 v 

In conjunction with the discussion following the cite to the LNP 107 reference, the fifth example has application in statics
– particularly in thermodynamics. It describes the free information I   log Z of the Fermi-Dirac partition function,
with the following correspondences
 Z  1  e p = the partition function,
 I  H = the free information,
 L = the entropy
 v = the energy per particle, with
 p    1  kT  = the inverse temperature.
This example is a case in point where we consider a concave-down L , rather than one that is concave up.

The relation of free information to inverse temperature is given, in this example, by the H versus p curve. The Legendre
transform yields the L versus v curve, which gives us the relation of entropy to energy. That is, the tangents to the curve
 I ,     log Z ,1  kT     H , p  yields intercepts for the v axis that give us the entropy L  H * versus the energy
v  dH dp .

8. Construction of the Canonical 2-Form


8.1. Canonical Symmetries
The canonical 2-form  , like  , can be expressed as a Lie derivative   Ld dt  , where
  pa  q a .
The pairs  q a , pa  are always in the dynamics of any system governed by the geometry of a Poisson manifold or
symplectic manifold. In the latter case, the Darboux Theorem guarantees you the ability to locally coordinatize the
manifold by conjugate pairs  q a , pa  in such a way that its symplectic form  has the form indicated above. The
dynamics then come out of this by selecting a distinguished differential operator d dt and taking the Lie derivative to
obtain
  Ld dt  .
In turn, this generates the kinematic law v a   d dt  q a and the dynamic law f a   d dt  pa .

The condition that governs the dynamics, itself, is just that   0 . That is, it’s the requirement that d dt be a canonical
“symmetry”:
Ld dt   0 .
This, in turn, is what makes   Ld dt  exact7.

This can be generalized in one direction as follows: in place of d dt , one can take the vector fields associated with a set
of symmetries. For example, one could consider the symmetries of space-time. Then, instead of just one symmetry d dt

7 Or: at least closed, if the underlying space is not simply connected.


associated with time translation, one would have an entire set of vector fields – 3 for spatial translation, 3 for rotation, 3
for boosts and 1 for time translation. In that generalized approach, various objects of importance appear, including the
contractions
X _| , X _| 
and the Lie derivatives
LX , LX  .
The requirement that the symmetry generated by X be canonical is that LX   0 . The other quantities are closely related
to (and generalize) the Hamiltonian, Lagrangian and “Cartan form”   pa v a . You can determine which goes with which
by substituting in X  d dt and seeing what happens. Thus,
   
 _|  pa  q   f a q  v pa  H ,
d
X _|   _|     v a a  f a a a a

dt  t q pa 
   
 _|  pa q   pa v  ,
d
X _|   _|     v a a  f a a a

dt  t q pa 
LX   Ld dt     L.
Hence, X _|  generalizes the Hamiltonian; X _|  , the Cartan form; and LX  , the Lagrangian. Then, noting that
   , the Legendre transform relation, itself,
L    H
generalizes to the relation
LX     X _|   X _| 
which is an identity.

8.2. Poisson Manifolds


A different approach on the construction of  is as follows: If the dynamics are governed by a Poisson manifold, and
not just by a symplectic manifold, then the manifold partitions into a union or “layering” of symplectic manifolds, called
its “symplectic leaves”. Each one has its own coordinatization in terms of complementary pairs  q a , pa  . The leaves,
themselves, are indexed by a set of quantities which are invariant on each leaf, but may vary from layer to layer. On each
layer is a symplectic form.

This will be illustrated by two examples.

8.3. Examples
Example 8 (Classical Spin)
The Lie group associated with rotations is SO  3 . Its Lie algebra is given by

 
so  3  S   S x , S y , Sz  : Sx , S y , Sz  R ,  S y , Sz   Sx ,  Sz , Sx   S y ,  Sx , S y   Sz .
This extends to the following Poisson bracket
f g
 f  S  , g S   S  S  S .
The invariant is S  S . For S  0 , the symplectic leaves are each 2-dimensional and may be written in coordinate form
as
f g f g
S   S cos  sin , S sin  sin , S cos   ,  f  S  , g  S    ,  q, p    , S cos   .
q p p q
The corresponding symplectic form is then just
  p  q   S sin        S sin     .

For S  0 , all the components S x , S y , S z are invariants (namely, each is 0), and the symplectic leaf is 0-dimensional
with 0 pairs of coordinates. The symplectic form is just   0 .
Example 9 (Non-Relativistic “Particles”)
A second example in the same mould is the Poisson manifold given by the Lie algebra of the (extended) Galilei group.
Here, in place of 3 generators, are 11: 10 governing space-time symmetries, plus a central charge.

Starting from first principles, the Galilei group may be defined as the subgroup of GA  4 t ,r  GL  4 h,p which has the
following 3 quadratic invariants
dt 2 , p 2 , p  dr  hdt .
The most general transformation is given, in infinitesimal form, by
  dr   ω  dr  υdt ,   dt   0, p  ω  p, h  υ  p .
The vectors ω and υ correspond, respectively, to infinitesimal rotations and infinitesimal boosts. Upon integration, the
transformations on the two differentials become
r  ω  r  υt  ε, t   ,
where the constants of integration ε and  respectively yield infinitesimal forms of spatial translations and time
translations. Together, this gives us a 10-parameter group whose general vector may be written as
  ω  J  υ  K  ε  P  H ,
the corresponding “Hamiltonian field” being labeled    . Under Noether’s Theorem, the 10 basis elements
correspond respectively to
J = rotation generator  angular momentum,
K = boost generator  center of mass moment,
P = spatial translation generator  linear momentum,
H = time translation generator  kinetic energy.

The Lie bracket on the generators is obtained by the condition that ,     ,   , where    . Upon application to
 
the defining representation, we find that
  r, t , p, h     ω  r  υt  ε ,  , ω  p,  υ  p 
  ω   ω  r  υt  ε   υ, 0, ω   ω  p  ,  υ   ω  p   ,
  r, t , p, h    ω   ω  r  υt  ε   υ  , 0, ω   ω  p  , υ   ω  p   ,
,    r, t , p, h     ω  ω   r   ω  υ  υ  ω  t   ω  ε  ε  ω    υ  υ  , 0,  ω  ω   p,   ω  υ  υ  ω   p  .

Thus,
,     ω  ω   J   ω  υ  υ  ω   K   ω  ε  ε  ω   P   υ  υ   P .

The Lie algebra has a non-trivial central extension involving the K , P  brackets
υ  K, ε  P  υ  εM .
The central charge M corresponds to the mass. To make a Noether’s Theorem type correspondence work requires also
extending the defining representation by adding in the mass m and a coordinate u conjugate to it. This leads to the 5-
dimensional representation of the Galilei group – where the defining relations are now given by modified forms of the 3
quadratic invariants plus the linear invariant for the central charge:
dr 2  2dtdu, p 2  2mh, p  dr  hdt  mdu, m .
The transformations that yield the extended Galilei group can then be written as
r  ω  r  υt  ε, t  , u  υ  r  ,
p  ω  p  υm, h  υ  p, m  0.
The extended Lie vector is   ω  J  υ  Κ  ε  P  H  M , and the Lie bracket relation becomes
,   ω  ω  J  ω  υ  υ  ω  K  ω  ε  ε  ω   P   υ  υ  P   υ  ε  ε  υ M ,
where   ω  J  υ  Κ  ε  P  H  M .
From this, the Poisson bracket relation may be read off:
f f  f f f f   f f f f 
 
f  J , K , P, H , M  , f  J , K , P, H , M   J    K  
J J
     P     
 K J J K   P J J P 
 f f f f   f f f f 
P    M     .
 K H H K   K P P K 

The symplectic leaves are determined by the functional invariants, which are found by the following two steps. First, the
translation invariants are
M , H , P, W  MJ  P  K, W0  P  J .
That is, the general solution to  f , P  0 ,  f , H   0 and  f , M   0 is f  f  M , H , P, W, W0  . Second, the
combinations of these which are also invariant under boosts and rotations are
M , P  2MH , W 2 .
The classification of the symplectic leaves is then given by the following conditions:
P  0, M  0  Vacuon (Translation-invariant media)
P  0, M  0  Synchron (Media that support action at a distance transfers of momentum)
M  0  Tardion (Media with well-defined centers of mass – or “particles”)
In the cases of synchrons and vacuons, further invariants exist besides the three just listed.

The tardions have a decomposition into complementary coordinates that readily emerges by carrying out the following
decomposition
 p2 
 J, K, P, H , M    r  p  S, mr, p,  U , m  .
 2m 
The 3 invariants become  S 2 , U , m  ; and under this reduction, the Poisson brackets reduce to
 
 f , f   S  Sf  Sf   fr  pf  pf  fr  .
 
Thus, r, p yield the 3 sets of “Heisenberg” complementary coordinates, while from S may be derived a fourth, as in
the previous example. The symplectic form of the previous example then expands out to
    S cos      p  r  S  0  ,
  p  r  S  0 .
For synchrons, there is a different decomposition into complementary coordinates. One of the coordinate pairs (as
indicated by the family’s name) is the time coordinate and arises by way of the decomposition
WP PK
K 2
 Pt , t   2 .
P P
The Poisson bracket decomposes into a part involving J, W, P and a part involving H , t :
     
 f , f   J  Jf  Jf  W   Jf  Wf  Wf  Jf   P   Jf  Pf  Pf  Jf    Hf ft  ft Hf  .
     
The two subclasses that occur are where W  0 , in which case there are 3 complementary pairs of variables that can be
extracted from J, W, P ; and where W  0 , in which case W0 becomes an invariant and 2 complementary pairs of
coordinates are to be found from within J, P .

Finally, for the vacuons, H becomes an invariant, as do J  K and K 2 . There are 3 subclasses of symplectic leaves,
determined by the following conditions:
Vacuum J  0, Here, there are 0 coordinate pairs and the symplectic form is just   0 .
Boost-invariant isotropic K0
media
Quasi-Vacuum J  0, This is the same case as SO  3 , J 2  0 is an invariant and there is one
Boost-invariant K0 coordinate pair in  .
anisotropic media
Generic Vacuon J  0, Here, J  K and K 2  0 are the invariants and there are 2 coordinate pairs
Boost non-invariant K  0 involved in the definition of  .
anisotopic media

References
[1] Christopher Hillman; Legendre Transformation; sci.physics, sci.math; 1997 August 28 01:23:30
http://groups.google.com/group/sci.physics/browse_thread/thread/87548d550f654c21/a2a88f90aac5155b
[2] Christopher Hillman; Lagrangians I; sci.physics.research; 1998 April 24 (A thread that follows up on [1].)
http://groups.google.com/group/sci.physics.research/browse_thread/thread/ca59127e93fc5de4/c2487fc484
[3] Jerzy Kijowski; Wlodzimierz M. Tulczyjew, “A Symplectic Framework for Field Theories”, Lecture Notes in
Physics, 107, Springer-Verlag, Berlin, Heidelberg, New York.
[4] “Legendre-Fenchel duality” in R. Tyrrell Rockafellar, Conjugate duality and optimization, Philadelphia: Society
for Industrial and Applied Mathematics [1974].

You might also like