# 2 TIONS ORMA TRANSF

26 Chapter 2 This is called the inner product, and is preserved by the group of orthogonal transfor-mations, namely, matrices Usatisfying (2.2) U⊤U= I...

2

CANONICAL TRANSFORMATIONS As for a generic system of differential equations, coordinate transformations can be used in order to bring the system to a simpler form. If the system is Hamiltonian it is desirable to keep the Hamiltonian form of the equations when the system is transformed. The search for a class of transformations satisfying the latter property leads to introducing the group of the so called canonical transformations. The condition of canonicity can be expressed in terms of Poisson brackets, Lagrange brackets and differential forms. In this chapter I will discuss five different criteria of canonicity. Precisely, a coordinate transformation is canonical in case: (i) the Jacobian matrix of the transformation is a symplectic matrix; (ii) the transformation preserves the Poisson brackets; (iii) the transformation preserves the Lagrange brackets; P (iv) the transformation preserves the differential 2–form j dpj ∧ dqj ; (v) P the transformation preserves the integral over a closed curve of the 1–form j pj dqj . An useful method for constructing canonicals transformations is furnished by the theory of generating functions. This is the basis for further development of the integration methods, eventually leading to the Hamilton–Jacobi’s equation.

2.1

Elements of symplectic geometry

Before entering the discussion of canonical transformations we recall a few aspects of symplectic geometry that will be useful. A relevant role is played by the bilinear antisymmetric form induced by the matrix J introduced in sect. 1.1.3, formula (1.16). Symplectic geometry is characterized by the skew symmetric matrix J in the same sense as Euclidean geometry is characterized by the identity matrix I. Indeed, Euclidean geometry is characterized by transformations preserving the bilinear symmetric form X xj y j . (2.1) hx, yi = j

26

Chapter 2

This is called the inner product, and is preserved by the group of orthogonal transformations, namely, matrices U satisfying (2.2)

U⊤ U = I .

Similarly, symplectic geometry is characterized by transformations preserving the bilinear form (2.3)

[x, y] = hx, Jyi ,

where J is the matrix defined by (1.16). This is called the symplectic product. It is antisymmetric, i.e., [x, y] = −[y, x], and non degenerate, i.e., [x, y] = 0 for all y ∈ R2n implies x = 0. 2.1.1 The symplectic group Let us consider a linear mapping in R2n (2.4)

x = Uy ,

where U is a 2n × 2n nonsingular matrix. The matrix U is said to be symplectic if (2.5)

U⊤ JU = J .

This condition is actually equivalent to (2.6)

UJU⊤ = J .

For, by (2.5) we have U⊤ = JU−1 J−1 , and so also UJU⊤ = UJ2 U−1 J−1 = −J−1 = J in view of J2 = −I. By the way, this shows that if U is symplectic then U⊤ is symplectic, too. The set of symplectic matrices forms a group with respect to matrix multiplication. For, the identity matrix I is clearly symplectic; if U and V are symplectic then (UV)⊤ J(UV) = V⊤ U⊤ JUV = V⊤ JV = J, so that (UV) is symplectic; if U is symplectic then J = IJI = (U−1 )⊤ U⊤ JUU−1 = (U−1 )⊤ JU−1 , so that U−1 is symplectic. Finally, if U is symplectic so is U⊤ . Going back to the canonically conjugated coordinates (q, p) it is immediately seen that the symplectic 2n vectors z = (q, p) and z′ = (q′ , p′ ) is given Pnproduct′ of two ′ by the expression j=1 (qj pj − qj pj ). That is, the symplectic product is obtained by projecting the parallelogram with sides z, z′ onto each of the planes qj , pj , and then adding up algebraically the oriented areas of all these projections. The set of linear transformations which preserve the symplectic product is characterized as the group of symplectic matrices. For, [Ux, Uy] = hUx, JUyi = hx, U⊤ JUyi, and this coincides with [x, y] if and only if U satisfies (2.5). 2.1.2 Symplectic spaces and symplectic–orthogonality A symplectic space is a real linear vector space V equipped with a bilinear antisymmetric nondegenerate form [·, ·]. We shall consider here only vector spaces of finite

Canonical transformations

27

dimension. Two vectors x, y are said to be symplectic–orthogonal in case [x, y] = 0. We shall write x6 y. To any subspace W of V we associate the set (2.7)

W

6

= {x ∈ V | x6 y for all y ∈ W } ; 6

by the bilinearity of the symplectic form W is a subspace, which is said to 1 be symplectic–orthogonal to W . We emphasize that the concept of symplectic– orthogonality presents sharp differences with respect to the concept of orthogonality in Euclidean geometry. Here are three main differences. (i) Every vector is self symplectic–orthogonal, since the symplectic product is antisymmetric; therefore, any one–dimensional subspace is self symplectic– orthogonal, too. (ii) The restriction of the symplectic product to a subspace is still a bilinear form, but in general it fails to be nondegenerate. For instance, the restriction to a one–dimensional subspace is clearly degenerate. 6 (iii) The subspaces W and W need not be complementary. Two basic properties that are common to both geometries are given by the following Lemma 2.1: Let W be a subspace of a symplectic space V . Then (2.8)

dim W + dim W

6

= dim V .

and (2.9)

W

6

6

=W

Proof. If dim W = 0 or dim W = dim V , then the statement is trivial. So, let us suppose that dim W = m with 0 < m < dim V . Denoting n = dim V , let {u1 , . . . , un } be a P basis of V , with {u1 , . . . , um } a basis of W . Writing a generic P vector v ∈ V as n v = j=1 vj uj , the symplectic product takes the form [v, w] = j,k ajk vj wk , where ajk = [uj , uk ] is an element of an antisymmetric nondegenerate matrix. If w ∈ W , then wm+1 = . . . = wn = 0, because by hypothesis {u1 , . . . , um } is a basis of W . If 6 moreover v ∈ W , then the relation of symplectic orthogonality is m X j=1

βj wj = 0 ,

βj =

n X

ajk vk .

k=1

Since w ∈ W is arbitrary, the first equation implies β1 = . . . = βm = 0 and βm+1 , . . . , βn arbitrary; so, there are n − m independent solutions. Since the matrix {ajk } is nondegenerate, the second relation above guarantees the existence of exactly n − m independent vectors symplectic–orthogonal to the subspace W , so that 6 dim W = n − m , as claimed. 6 Coming to (2.9), if w ∈ W , then v6 w for all v ∈ W , because by definition every 6 element of W is symplectic–orthogonal to every element of W , and so also to w, so 1

Some authors use the name left–orthogonality. See for instance [7].

28

Chapter 2 6

6

6

that W ⊂ (W ) . On the other hand, by (2.8), dim(W ) follows.

6

= dim W , so that (2.9) Q.E.D.

We now go deeper into the concept of symplectic–orthogonality, pointing out some properties that do not appear in Euclidean geometry. To this end, we first need some definitions. A subspace W of a symplectic space V is said to be: 6 (i) isotropic in case W ⊂ W ; 6

(ii) coisotropic in case W ⊃ W , namely, if its symplectic–orthogonal subspace is isotropic; 6 (iii) Lagrangian in case W = W , namely, if it is both isotropic and coisotropic; (iv) symplectic in case the symplectic product restricted to W is still nondegenerate. These definitions are easily illustrated by considering the space R2n . We denote by {e1 , . . . , en , d1 , . . . , dn } the canonical basis of R2n and by (q1 , . . . , qn , p1 , . . . , pn ) the coordinates, so that a vector x ∈ R2n is represented as q1 e1 + . . . + qn en + p1 d1 + . . . + pn dn . The symplectic bilinear form is defined by the relations [ej , ek ] = [dj , dk ] = 0 ,

[ej , dk ] = δj,k ,

j, k = 1, . . . , n .

We shall refer to the basis {e1 , . . . , en , d1 , . . . , dn } satisfying the latter relations as the canonical symplectic basis. Example 2.1: Arithmetic planes. Let us call arithmetic plane the subspace spanned by any subset of the vectors {e1 , . . . , en , d1 , . . . , dn }. Formally, given any subsets J, K of {1, . . . , n}, we consider the plane spanned by the vectors {ej }j∈J ∪ {dk }k∈K . The following examples are easily understood: (i) the arithmetic plane spanned by any subset of the vectors {e1 , . . . , en } is an isotropic subspace; (ii) the direct sum of any of the arithmetic planes of the point (i) with the arithmetic plane span(d1 , . . . , dn ) is a coisotropic subspace; (iii) the arithmetic plane span(e1 , . . . , en ) is a Lagrangian subspace. (iv) for every j ∈ {1, . . . , n}, the arithmetic 2–dimensional plane span(ej , dj ) is symplectic; further symplectic subspaces are generated by direct sum of such arithmetic planes. In the examples (i)–(iii) the role of the vectors ej and dj can be exchanged, of course. A more interesting example is the following: Example 2.2: Lagrangian arithmetic planes. Consider any partition of the indexes {1, . . . , n} into two disjoint subsets J and K (i.e., J ∩ K = ∅ and J ∪ K = {1, . . . , n}). Then the arithmetic plane spanned by the n vectors {ej }j∈J ∪{dk }k∈K is a Lagrangian plane. There are 2n different Lagrangian planes that are generated that way.2 2

There are 2n different partitions of n objects into two disjoint subsets.

Canonical transformations

29

2.1.3 Canonical basis of a symplectic space The examples above are in fact quite general. For, any symplectic space can be equipped with a canonical symplectic basis. This is stated by the following Proposition 2.2: Let V be a symplectic space. Then dim V is even, say 2n, and there exists a canonical basis {e1 , . . . , en , d1 , . . . , dn } satisfying (2.10)

[ej , ek ] = [dj , dk ] = 0 ,

[ej , dk ] = δjk ,

j, k = 1, . . . , n .

The proof depends on some properties that are of independent interest, and are isolated in the following two lemmas. Lemma 2.3: A subspace W of V is symplectic if and only if the subspaces W and 6 6 W are complementary. In that case W is a symplectic subspace. 6

Proof. Let W be a symplectic subspace. Then W ∩W = {0}. For, we have [v, w] = 6 0 for all w ∈ W in view of v ∈ W , and this implies v = 0 in view of v ∈ W and of the nondegeneracy of the symplectic product. On the other hand, by (2.8), we have 6 6 6 dim(W ⊕ W ) = dim W + dim W = dim V , so that W and W are complementary. 6 Conversely, let W and W be complementary, and let v ∈ W be such that [v, w] = 0 6 for all w ∈ W . By definition of W this implies [v, w] = 0 for all w ∈ V , and so also v = 0 by nondegeneracy. We conclude that W is symplectic. 6 6 It remains to prove that W is symplectic. Since W and W are complementary, if 6 [v, w] = 0 for all w ∈ W then v = 0 by the same argument as above. Q.E.D. Lemma 2.4: Let V be a symplectic space. Then dim V ≥ 2, and there exists a decomposition of V in two complementary symplectic subspaces V1 and V2 , with 6 dim V1 = 2 and V2 = V1 . Moreover, in V1 there exists a symplectic canonical basis e1 , d1 .

Proof. Let e1 6= 0 be an arbitrary vector. Then, by nondegeneracy, there exists a vector d1 independent of e1 such that [e1 , d1 ] 6= 0. This proves that the dimension must be at least 2. In view of linearity, by a trivial rescaling we can choose d1 such that [e1 , d1 ] = 1. Let V1 = span(e1 , d1 ) , so that dim V1 = 2 and e1 , d1 is a basis of V1 . We prove that it is a symplectic subspace. To this end, first check that the decomposition of any vector w ∈ V1 over the basis e1 , d1 is w = [w, d1]e1 − [w, e1]d1 ; this is elementary. Suppose now that [w, v] = 0 for all v ∈ V1 . Then we have in particular [w, e1] = [w, d1 ] = 0, and so also w = 0 in view of the decomposition above. This proves that the restriction of the symplectic form to the subspace V1 is nondegenerate. Therefore, V1 is a symplectic subspace, and e1 , d1 is a canonical basis of it. 6 By lemma 2.3, V2 = V1 is a symplectic subspace complementary to V1 . Q.E.D. Proof of proposition 2.2. If dim V = 2 just apply once lemma 2.4. If dim V > 2, then apply recursively lemma 2.4. With an obvious change of notation, first write V = V1 ⊕V ′ , where V1 admits a symplectic canonical basis e1 , d1 , and V ′ is symplectic, with dim V ′ = dim V − 2. Next, apply again the lemma to V ′ , getting V ′ = V2 ⊕ V ′′ with V2 admitting a canonical basis e2 , d2 . Proceeding the same way, end up with a

30

Chapter 2

decomposition V = V1 ⊕ · · · ⊕ Vn (recall that we assume that V is a vector space of finite dimension), where Vj is a 2–dimensional symplectic subspace equipped with a canonical basis ej , dj satisfying [ej , ej ] = [dj , dj ] = 0 and [ej , dj ] = 1 (j = 1, . . . , n) . This implies that {e1 , . . . , en , d1 , . . . , dn } is a basis for V , and so also dim V = 2n . By construction, all subspaces V1 , . . . , Vn are pairwise symplectic–orthogonal, which implies [ej , ek ] = [dj , dk ] = [ej , dk ] = 0 for j 6= k. This proves (2.10). Q.E.D. 2.1.4 Properties of the subspaces of a symplectic space We establish two properties that will be relevant in the discussion concerning integrable systems. Lemma 2.5: If W is isotropic, then dim W ≤ n; if W is coisotropic, then dim W ≥ n; if W is Lagrangian, then dim W = n. An immediate consequence of this lemma is the Corollary 2.6: An isotropic subspace W is Lagrangian if and only if dim W = n. The same holds true for a coisotropic subspace. Proof of lemma 2.5. Just use lemma 2.1, formula (2.8). If W is isotropic, i.e., 6 6 W ⊂ W , then dim W ≤ dim W , which implies dim W ≤ n. If W is coisotropic, just reverse the argument. If W is Lagrangian, it is both isotropic and coisotropic, so that both the inequalities dim W ≤ n and dim W ≥ n apply. Q.E.D. Lemma 2.7: Let V be a symplectic space and {e1 , . . . , en , d1 , . . . , dn } be a canonical basis. Then any Lagrangian subspace W of V is complementary to at least one of the Lagrangian arithmetic planes of example 2.2. Proof. Consider the n–dimensional Lagrangian plane D = span{d1 , . . . , dn }, and let P = D ∩ W and m = dim P . Since P is a subspace of D, we have 0 ≤ m ≤ n, and there exist n − m vectors in {d1 , . . . , dn } spanning a (n − m)–dimensional arithmetic plane complementary to P in D. That is: there is a subset K ⊂ {1, . . . , n}, with #K = n − m, such that the arithmetic plane DK = span{dk }k∈K satisfies DK ∩ P = {0} and DK ⊕ P = D (the set K needs not be unique). Let now J = {1, . . . , n} \ K, so that {J, K} is a partition of {1, . . . , n}, and let EJ = span{ej }j∈J . We prove that the Lagrangian subspace L = EJ ⊕ DK is complementary to W , namely, it is the Lagrangian arithmetic plane we are looking for. Since dim W = dim L = n, it is enough to prove that W ∩ L = {0}. This is seen as follows. On the one hand, by P ⊂ W and W 6 W we have P 6 W ; on the other hand, by DK ⊂ L and L6 L we have DK 6 L. Using these relations we get3 D = P ⊕ DK 6 W ∩ L, and so W ∩ L ⊂ D in view of D being Lagrangian. Therefore, W ∩ L = (W ∩ D) ∩ (L ∩ D) = P ∩ DK = {0}, by construction of DK . Q.E.D.

3

Let v ∈ W ∩ L. Since v ∈ W , we have v6 P ; since v ∈ L, we have v6 DK . We conclude v 6 P ⊕ DK .

Canonical transformations

2.2

31

Transformations preserving the Hamiltonian form of the equations.

Let us now turn to the main argument of this chapter, namely to characterize a class of transformation that allows us to stay in the framework of the Hamiltonian formalism. It may be useful to make again a connection with the Lagrangian formalism. It is well known that Lagrange equations have the nice property of being invariant with respect to point transformation (i.e., changes of the coordinates in configuration space, which by differentiation generate the corresponding transformations on the generalized velocities). The Hamiltonian formalism removes the tie between generalized coordinates and velocities, so that arbitrary transformations involving all the canonical coordinates may be devised. However, an arbitrary transformation will likely produce equations which are not in Hamiltonian form, in the sense that the second members are not expressed as derivatives of a unique function. The problem then is to characterize a restricted class of transformation which keep the form of Hamilton’s equations. 2.2.1 Conditions for canonicity Let us first look for a class of transformations (q, p) = C(Q, P ) satisfying the following

Condition 1: to every Hamiltonian function H(q, p) one can associate another function K(Q, P ) such that the canonical system of equations q˙j =

∂H , ∂pj

p˙j = −

∂H , ∂qj

j = 1, . . . , n ,

is changed into the system ∂K , Q˙ j = ∂Pj

∂K P˙ j = − , ∂Qj

j = 1, . . . , n ,

which is still canonical. We shall say that such a transformation preserves the canonical form of the equations. Condition 2: the transformation preserves the canonical form of the equations with the new Hamiltonian K(Q, P ) = H(q, p) q=q(Q,P ) , p=p(Q,P ) . The difference with respect to the first condition is that the new Hamiltonian is constructed by a direct substitution of the transformation in the old one. Transformations satisfying condition 2 will be called time–independent canonical transformations, or simply canonical.4 4

There is no general agreement about the use of the term canonical transformation. Some authors call canonical any transformation satisfying the requisite of preserving the canonical form of the equations, as stated by condition 1 in the text. For example, this is the attitude of Wintner in [100]. Others follow the attitude of the present notes. Example 2.4 below illustrates the difference.

32

Chapter 2

The case of time–dependent transformations appears to be a little more complex: we need to go back to condition 1, since the new Hamiltonian is not determined by a mere substitution of variables. I will discuss this matter at the end of the chapter, where it is shown how the case of time–dependent transformations can be reduced to the time–independent one, and how the new Hamiltonian K(Q, P ) is constructed. The examples below show that there are transformations satisfying the conditions above. As already anticipated, I will consider here only time–independent transformations. Example 2.3: Translation. The transformation qj = Qj + aj ,

pj = Pj + bj ,

1≤j≤n,

where a = (a1 , . . . , an ) and b = (b1 , . . . , bn ) are constants transforms the equations ∂H ∂H , p˙ j = − ∂q into q˙j = ∂p j j ∂ ∂H d = , (q, p) H(q, p) Q˙ j = (qj − a)= dt ∂pj ∂Pj q=Q+a , p=P +b q=Q+a , p=P +b d ∂H ∂ ˙ Pj = (pj − b)= − (q, p) =− H(q, p) , dt ∂qj ∂Qj q=Q+a , p=P +b q=Q+a , p=P +b

Hence it preserves the canonical form of the equations with the new Hamiltonian K(Q, P ) = H(q, p) . q=Q+a , p=P +b

Thus, the transformation is canonical.

Example 2.4: Scaling transformation. The transformation qj = αQj ,

pj = βPj ,

1≤j≤n,

with real constants α and β produces the transformed equations 1 ∂H 1 ∂pj ∂ q˙j = · (q, p) = · · H(q, p) , Q˙ j = α α ∂pj αβ ∂Pj ∂pj q=αQ , p=βP q=αQ , p=βP p˙j 1 ∂ 1 ∂H ∂qj ˙ Pj = ·= − · =− . (q, p) · H(q, p) · β β ∂qj αβ ∂Qj ∂qj q=αQ , p=βP q=αQ , p=βP

Hence it preserves the canonical form of the equations with the new Hamiltonian 1 H(q, p) . K(Q, P ) = αβ q=αQ , p=βP

This transformation satisfies condition 1 for any choice of the constants α , β. However, condition 2 is fulfilled only in case αβ = 1, and so we shall call it canonical in strict sense only in the latter case. The example shows that restricting the use of the adjective canonical only to transformations satisfying the second condition essentially reduces to excluding the class of scaling transformations for which αβ 6= 1. Including all such transformations makes the definition more general, of course. However, it also introduces some complications in the expositions that are unnecessary, and this is

Canonical transformations

33

what I want to avoid. The more general framework is recovered by keeping in mind that a canonical transformation in strict sense can always be composed with a scaling transformation, which sometimes is an useful device. Example 2.5: Again on the scaling transformation. The transformation of example 2.4 may be given the more general form qj = αj Qj ,

pj = βj Pj ,

1≤j≤n,

with real constants αj and βj provided the condition α1 β1 = . . . = αn βn = κ is satisfied. The new Hamiltonian is 1 . K(Q, P ) = H(q, p) κ q=αQ , p=βP The transformation is canonical in strict sense in case κ = 1. Example 2.6: Exchange of conjugated coordinates. qj = Pj ,

pj = −Qj ,

The transformation

1≤j≤n,

preserves the canonical form of the equations with the new Hamiltonian K(Q, P ) = H(q, p)|q=P , p=−Q . This is a canonical transformation. 2.2.2 Preservation of Poisson brackets As we remarked in sect 1.2.2, the Hamiltonian formalism can be expressed in terms of Poisson brackets, saying that the time evolution of any dynamical variable f is given by equation f˙ = {f, H} . This leads to a characterization of canonical transformations as possessing the property of leaving invariant the form of the Poisson bracket. Let (q, p) = C (Q, P ) be a coordinate transformation, and denote by C f the transformed function  . C f (Q, P ) = f (q, p) (q,p)=C (Q,P )

Also, denote by {·, ·}q,p and by {·, ·}Q,P the Poisson bracket with respect to the conjugate variables q, p and Q, P , respectively. Consider now the class of transformations satisfying the condition that the following diagram is commutative for every pair of functions f and g:

(2.11)

f, g   {·,·}y

{f, g}q,p

C

−→

C f, C g   y{·,·}

 −→ C {f, g}q,p = {C f, C g}Q,P . C

In words, one obtains the same result both (a) by computing the Poisson bracket with respect to the variables q, p and then changing the variables in the result, or (b) by changing the variables and then computing the Poisson bracket with respect to the

34

Chapter 2

new variables Q, P . If this happens to be true, we shall say that the transformations preserves the Poisson brackets. Proposition 2.8: A transformation (q, p) = C (Q, P ) is canonical if and only if it preserves the Poisson brackets, i.e., the diagram (2.11) is commutative. The difficult part of the proof is the “only if”, i.e., that the condition is necessary. In order to see it, we need to investigate to which extent we can transform the criterion expressed by the latter proposition to a practically applicable criterion. Let us consider the coordinates q, p as functions on the phase space; it is an easy matter to check that the relations (2.12)

{qj , qk } = {pj , pk } = 0

{qj , pk } = δjk

1≤j≤n, 1≤k≤n

hold true, where δjk is the Kronecker symbol. These expressions are sometimes called the fundamental Poisson brackets. We prove the following Lemma 2.9: A transformation preserves the Poisson bracket between any two functions if and only if it preserves the fundamental Poisson brackets. As a direct consequence, proposition 2.8 can be reformulated in a more useful manner as Corollary 2.10: A transformation (q, p) = C (Q, P ), is canonical if and only if it preserves the fundamental Poisson brackets, i.e., (2.13)

{qj , qk }Q,P = {pj , pk }Q,P = 0

{qj , pk }Q,P = δjk ,

1≤j≤n,

1≤k≤n.

Proof of lemma 2.9. Assume that the Poisson bracket between any two functions is preserved; then the fundamental Poisson brackets are preserved, too. So, we must prove only the converse. To this end, first check that if we are given a function f (ϕ1 , . . . , ϕr ), where, in turn, ϕ1 . . . , ϕr are functions of the canonical variables q, p, then r X ∂f {f, g} = {ϕl , g} , ∂ϕl l=1

where g(q, p) is any function. This is just matter of straightforward calculations. Put now r = 2n and ϕ1 (Q, P ) = q1 (Q, P ), . . . , ϕ2n (Q, P ) = pn (Q, P ), and consider also g(q, p) as function of the new variables Q, P through ϕ1 , . . . , ϕ2n . Then, using the identity above, compute X  ∂f ∂g ∂f ∂g {qj , qk }Q,P + {qj , pk }Q,P {f, g}Q,P = ∂qj ∂qk ∂qj ∂pk j,k  ∂f ∂g ∂f ∂g + {pj , qk }Q,P + {pj , pk }Q,P . ∂pj ∂qk ∂pj ∂pk

Canonical transformations

35

In view of the preservation of the fundamental Poisson bracket we immediately get {f, g}Q,P = {f, g}q,p, namely that the Poisson bracket between f and g is preserved. Q.E.D. Proof of proposition 2.8. Let the transformation to preserve the Poisson brackets. Denoting Q = Q(q, p) , P = P = (q, p) the inverse transformation, let f (q, p) be any of the new coordinates, e.g., f (q, p) = Qj (q, p) for some j. Then f˙ = {f, H}q,p. On the other hand, by preservation of the Poisson brackets we also have, after changing the ∂K variables, f˙ = {f, H}Q,P , that is, Q˙ j = ∂P where K(Q, P ) = H(q, p) (q,p)=C (Q,P ) . j Therefore, the transformed equations keep the canonical form by just transforming the hamiltonian, as required by condition 2. Conversely, let the transformation be canonical, and let, e.g., f = qj and H = pk for some j and k. Thus, f˙ = {qj , pk } = δj,k . On the other hand, after transforming to new variables we have f˙ = {f, H}Q,P , because in the new variables the equations are still in canonical form. Since the time derivative of f must be the same after the transformation, we conclude {qj , pk }Q,P = δj,k . The argument applies to any pair of canonical coordinates q, p, and this means that the fundamental Poisson bracket are preserved. By lemma 2.9 this implies that the Poisson brackets are preserved. Q.E.D. Example 2.7: The case of one degree of freedom. In the case n = 1 the canonicity condition in order the transformation to be canonical may be written as   ∂q ∂q  ∂Q ∂P  =1, {q, p} = det   ∂p ∂p  ∂Q

∂P

which means that the transformation must be area–preserving. For instance, the scaling transformation (2.14)

q = αQ ,

p=

1 P α

is canonical. An example of a transformation which is common in geometry but is not canonical is the transformation to polar coordinates (in the phase plane), which is not area preserving. A similar transformation which however is canonical is √ √ (2.15) q = 2I cos ϕ , p = 2I sin ϕ . The variables I, ϕ thus defined are called action–angle variables for the harmonic oscillator. Example 2.8: Harmonic oscillators. The Hamiltonian of a system of harmonic oscillators is n X  1 2 pj + ωj2 qj2 , H(q, p) = 2 j=1

with ω = (ω1 , . . . , ωn ) ∈ Rn . A more symmetric Hamiltonian is constructed by apply-

36

Chapter 2

ing the rescaling transformation Qj qj = √ , ωj

√ pj = Pj ωj ,

1≤j≤n,

which is clearly canonical. The transformed Hamiltonian is X ωj  Pj2 + Q2j . K(Q, P ) = 2 j

The form of the Hamiltonian can be further simplified by using the transformation to action–angle variables, namely p p Qj = 2Ij cos ϕj , Pj = 2Ij sin ϕj , 1 ≤ j ≤ n . By this, the Hamiltonian is transformed to H(I, ϕ) =

X

ωj I j ,

j

which is trivially integrated. 2.2.3 Symplecticity of the Jacobian matrix The canonicity condition on the fundamental Poisson brackets is often written as a condition on the Jacobian matrix of the transformation. It is convenient to use the compact notation of sect. 1.1.3. Let us write the transformation as x = u(y), and let K(y) = H ◦ u(y) be the Hamiltonian expressed in the new coordinates y.

Proposition 2.11: The transformation x = u(y) is canonical if and only if the Jacobian matrix uy of the transformation satisfies u⊤ y Juy = J ,

(2.16)

Here, J is the matrix defined by (1.16). In the language of sect. 2.2 we say that the Jacobian matrix of the transformation must be symplectic.5 Proof.

Using coordinates, compute

or, in compact notation,

2n X ∂K ∂xj ∂H , = ∂yk ∂yk ∂xj x=u(y) j=1 ∂ y K = u⊤ y ∂x H ◦ u .

Using J2 = −I, write the canonical equations for H as −Jx˙ = ∂x H, and using also x˙ = uy y˙ compute −Juy y˙ = ∂x H ◦ u. Finally, multiply both sides of the latter relation 5

In the recent literature the name “canonical transformation” is frequently replaced by “symplectic transformation”. In the present notes I prefer to use the old fashioned name ”canonical”. However, I will sometimes use the adjective “symplectic” when dealing with linear transformations, which involve symplectic matrices.

Canonical transformations

37

⊤ ⊤ by u⊤ y , and get −uy Juy y˙ = uy ∂x H ◦ u = ∂y K. Therefore, the equations for y are written −u⊤ y Juy y˙ = ∂y K .

They are in canonical form with the Hamiltonian K(y) provided u⊤ y Juy = J.

Q.E.D.

As a matter of fact, the relation between conditions (2.13) and (2.16) is very simple: they are the same thing, as the reader may verify by writing down in explicit form the product of matrices in (2.16). The reader may also remark that the explicit calculation of the latter product doubles the number of expressions to be handled, unless one realizes that it is enough to calculate the fundamental Poisson brackets. Moreover, the fact that the Poisson bracket between any two function must be preserved is of interest in itself, but remains hidden behind condition (2.16) unless one proves 2.9. 2.2.4 Preservation of Lagrange brackets Suppose that the canonical coordinates q, p are given as functions of two variables u, v. The Lagrange bracket [u, v] is defined as (see [66])  n  X ∂qj ∂pj ∂qj ∂pj . − (2.17) [u, v] = ∂u ∂v ∂v ∂u j=1 There is a strict relation between Lagrange brackets and Poisson brackets. Let the canonical coordinates q, p be expressed as differentiable and invertible functions of 2n independent variables u1 , . . . , u2n as (2.18)

qj = qj (u1 , . . . , u2n ) ,

pj = pj (u1 , . . . , u2n ) ,

j, k = 1, . . . , n ,

so that the variables u can be expressed as differentiable functions of the canonical coordinates q, p , by inversion. Lemma 2.12: Let the 2n × 2n matrices A = Aj,k and B = {Bj,k } , j, k = 1, . . . , 2n be defined as Ajk = {uj , uk } and Bjk = [uj , uk ]. Then we have AB⊤ = I ,

(2.19) the identity matrix. Proof.

For r, s = 1, . . . , 2n compute

2n X {ul , ur }[ul , us ] l=1

2n X n X n  X ∂ul ∂ur

  ∂ul ∂ur ∂qk ∂pk ∂qk ∂pk = − − ∂qj ∂pj ∂pj ∂qj ∂ul ∂us ∂us ∂ul l=1 j=1 k=1 ! ! " 2n n X n 2n X X X ∂ul ∂pk ∂ur ∂qk ∂ul ∂qk ∂ur ∂pk − = ∂q ∂u ∂p ∂u ∂qj ∂ul ∂pj ∂us j l j s j=1 k=1

l=1

l=1

38 − =

n  X j=1

2n X ∂ul ∂pj l=1

∂ur ∂pj ∂pj ∂us

Chapter 2 ! ! # 2n X ∂qk ∂ur ∂pk ∂ul ∂pk ∂ur ∂qk + ∂ul ∂qj ∂us ∂pj ∂ul ∂qj ∂us l=1  ∂ur ∂qj = δrs ; + ∂qj ∂us

here, use has been made of the identities 2n X ∂ul ∂pk l=1

2n X l=1

∂qj ∂ul

=

2n X ∂ul ∂qk =0, ∂pj ∂ul l=1 2n

X ∂ul ∂pk ∂ul ∂qk = = δjk , ∂qj ∂ul ∂pj ∂ul l=1

which hold true because the functions ul (q1 , . . . , qn , p1 , . . . , pn ) are the inverse of the functions (2.18). Q.E.D. It is now convenient to rename (u1 , . . . , u2n ) as (Q1 , . . . , Qn , P1 , . . . , Pn ). If the transformation (2.18) is the identity, then we have [Qj , Qk ] = [Pj , Pk ] = 0 [Qj , Pk ] = δjk , 1 ≤ j ≤ n , 1 ≤ k ≤ n .

(2.20)

These expressions are called the fundamental Lagrange brackets. Proposition 2.13: A transformation (q, p) = C (Q, P ) is canonical if and only if it preserves the fundamental Lagrange brackets. Proof. Recall that a canonical transformation preserves the fundamental Poisson brackets (corollary 2.10), and use lemma 2.12. If the fundamental Poisson brackets are preserved, then A = J, the symplectic matrix defined by (1.16); using JJ⊤ = I, by (2.19) we conclude B = J. Conversely, if the fundamental Lagrange brackets are preserved, then B = J, and by (2.19) we conclude A = J. Q.E.D.

2.3

Invariant differential forms

We consider the differential form ω2 =

(2.21)

n X j=1

dqj ∧ dpj .

Proposition 2.14: A transformation (q, p) = C (Q, P ) is canonical if and only if it P 2 preserves the 2–form ω = j dqj ∧ dpj .

Proof.

Using the formula for changing variables in a differential form, compute    X X X  ∂qj ∂qj ∂pj ∂pj dqj ∧ dpj = dQk + dPk ∧ dQl + dPl ∂Q ∂P ∂Q ∂P k k l l j j k,l

Canonical transformations =

X k
39

([Qk , Ql ]dQk ∧ dQl + [Pk , Pl ]dPk ∧ dPl ) +

X k,l

[Qk , Pl ]dQk ∧ dPl ,

which shows that the coefficients of the transformed differential form are the Lagrange brackets. If the transformation is canonical, then by proposition 2.13 we get X X dPj ∧ dQj . dpj ∧ dqj = j

j

Conversely, if the latter identity is fulfilled, then the Lagrange brackets are preserved. By proposition 2.13 the claim follows. Q.E.D. Using Stokes theorem one proves Proposition 2.15: (2.22)

A transformation (q, p) = C (Q, P ) is canonical if and only if Z X Z X pj dqj = Pj dQj , γ

C (γ)

j

j

where γ is a closed curve. *** Questo paragrafo deve essere ampliato in modo consistente. Aggiungere in appendice un richiamo sulle forme differenziali e relativi teoremi. Vedi Moser-Zehnder. ***

2.4

Generating functions

The characterization of canonical transformations by propositions 2.14 and 2.15 offers us an explicit method for constructing canonical transformations. Let us write the transformation as q = q(Q, P ), p = p(Q, P ), and let us assume that the relation q = q(Q, P ) can be inverted (at least locally) with respect to P , so that ∂(q1 , . . . , qn ) (2.23) det 6= 0 . ∂(P1 , . . . , Pn ) Transformations satisfying this condition are called free canonical transformations. By proposition 2.15 we have Z X Z X Pj dQj , pj dqj = γ

j

γ

j

where γ is an arbitrary closed curve. This means that there exists a function S(q, Q) such that X (2.24) (pj dqj − Pj dQj ) = dS , j

The function S(q, Q) is called the generating function. We have indeed the following

40

Chapter 2

Proposition 2.16: (2.25)

Let the function S(q, Q) satisfy the condition   ∂ 2S 6= 0 . det ∂qj ∂Qk

Then the transformation defined by (2.26)

∂S (q, Q) , ∂qj

pj =

Pj = −

∂S (q, Q) , ∂Qj

1≤j≤n

is a free canonical transformation. Proof. The transformation clearly satisfies the condition of proposition 2.15 for canonicity and condition (2.23) for being a free canonical transformation. Q.E.D. It should be remarked that the transformation is given in implicit form; however, by condition (2.25), we are allowed to do the inversions needed in order to express either the old variables q, p as functions of the new ones Q, P or the new variables as functions of the old ones. For, inverting the second of (2.26) with respect to q we get q = q(Q, P ), and replacing this in the first of (2.26) we get also p = p(q, Q) q=q(Q,P ), as required. Conversely, inverting the first of (2.26) with respect to Q and replacing the result in the second of (2.26) we obtain the inverse transformation. Example 2.9: Exchange of conjugated coordinates. X (2.27) S(q, Q) = qj Qj

The generating function

j

generates the canonical transformation (2.28)

pj = Qj ,

Pj = −qj ,

1≤j≤n,

exchanging the coordinates with the momenta. The class of free canonical transformation does not exhaust all possibilities. For instance, one will immediately realize that the identity is not free, so that it can not be represented by a generating function of the form above. A different form of the canonical transformation can be constructed using the Legendre transformation. *** Aggiungere trasformata di Legendre *** Proposition 2.17: (2.29)

Let the generating function S(P, q) satisfy the condition   2 ∂ S 6= 0 . det ∂Pj ∂qk

Then the transformation implicitly defined by (2.30)

pj =

∂S (P, q) , ∂qj

Qj =

∂S (P, q) , ∂Pj

1≤j≤n.

Canonical transformations

41

Proof. We check that the canonicity condition of propositionP2.15 is satisfied. For, introducing the Legendre transform of S(P, q), namely S˜ = S − j Pj Qj , we compute  X  ∂S ∂S dPj + dqj − Pj dQj − Qj dPj dS˜ = ∂P ∂q j j j X = (pj dqj − Pj dQj ) , j

so that (2.22) follows. Q.E.D. The form S(P, q) of the generating function is actually the most common.6 This because many useful transformations are expressed with a generating function of this form. The examples which follows illustrate some interesting cases. Example 2.10: P The identity and the scaling transformations. The generating function S(P, q) = α j Pj qj generates the scaling transformation pj = αPj ,

Qj = αqj ,

1≤j≤n.

For α = 1 this is the identity. Example 2.11: Extended point transformation. Suppose that we are given a point transformation Q = Q(q) which is a diffeomorphism,7 so that it admits an inverse Q = Q(q) and (2.31)

det

∂(Q1 , . . . , Qn ) 6= 0 , ∂(q1 , . . . , qn )

det

∂(q1 , . . . , qn ) 6= 0 . ∂(Q1 , . . . , Qn )

A corresponding canonical transformation can be constructed using the generating function X S(P, q) = Pk Qk Q=Q(q) . k

For, the complete transformation is qj = qj (Q) ,

pj =

X k

Pk

∂Qk (q) , ∂qj

1≤j≤n.

On the other hand, the invertibility condition (2.29) of proposition 2.17 is satisfied in view of (2.31), since     2 ∂Qk ∂ S = det 6= 0 . det ∂Pk ∂qj ∂qj This extension is not unique. The most general extended point transformation is generated by the function X (2.32) S(P, q) = Pk Qk Q=Q(q) + W (q) , k

6 7

Several books report only this form. This is the class of transformation which are allowed in the framework of the Lagrangian formalism.

42

Chapter 2

where W (q) is an arbitrary function. Example 2.12: Near the identity canonical transformations. Consider the generating function X (2.33) S(P, q) = Pj qj + εf (P, q) , j

where f (P, q) is an arbitrary function and ε a real parameter, which is assumed to be small. The invertibility condition (2.29) of proposition 2.17 is clearly satisfied for ε small enough. The corresponding canonical transformation in implicit form is pj = Pj + ε

∂f (P, q) , ∂qj

Qj = qj + ε

∂f (P, q) , ∂Pj

1≤j≤n.

The explicit form can be found, e.g., by inverting the second relation with respect to q and replacing the result in the first one. This gives ∂f (P, Q) + ε2 . . . ∂Pj ∂f pj = Pj + ε (P, Q) + ε2 . . . . ∂qj qj = Qj − ε

For ε = 0 the transformation is the identity, while for ε 6= 0 the coordinates are changed by a little amount. Such a kind of transformations is the basic tool for the development of perturbation theory. However, it can be remarked that the inversion required in order to put the transformation in explicit form is a quite unpleasant aspect, mainly if one plans to perform an explicit calculation. We shall see that inversions can be avoided by using the algorithm of Lie transforms. The generating functions discussed till now do not actually exhaust the class of canonical transformations. The following example illustrates this point. Example 2.13: 2n canonical transformations. Let J, K be a partition of the set {1, . . . , n} into two disjoint subsets, J ∪ K = {1, . . . , n}, J ∩ K = ∅, and consider the canonical transformation pj = Qj , Pj = −qj for j ∈ J (2.34) pk = Pk , Qk = qk for k ∈ K .

There are 2n different transformation of this type. In particular, the exchange of conjugated coordinates of example 2.9 is found by setting J = {1, . . . , n} , K = ∅ , and the identity is found by setting J = ∅ , K = {1, . . . , n} . We know that the latter two examples are covered by propositions 2.16 and 2.17, respectively, but a generating function of the form above can not be found in all other cases. All the examples above are actually covered by the following Proposition 2.18: Take any partition of the integers {1, . . . , n} into two disjoint sets J, K. Assume that the generating function S = S(QJ , PK , q) satisfy the condition   ∂ 2S 6= 0 . (2.35) det ∂(QJ , PK )∂q

Canonical transformations

43

Then the transformation implicitly defined by Pj = − (2.36)

Qk = pl =

∂S ∂Qj ∂S ∂Pk ∂S ∂ql

for

j∈J

for

k∈K

for

1≤l≤n

is canonical. Conversely, for any canonical transformation one can find a partition J, K and a generating function of the form above. Proof. The proof that the transformation is canonical requires checking that proposition 2.15 applies. This is just a minor modification of the proof of proposition 2.17, and is left to the reader. We prove that all canonical transformations are covered. To this end first remark that the canonical transformation (2.36) is characterized by the condition   ∂(p1 , . . . , pn ) (2.37) det 6= 0 . ∂(QJ , PK ) If a canonical transformation is given, then the n functions p1 (Q, P ), . . . , pn (Q, P ) are independent, which means that the n × 2n Jacobian matrix (2.38)

∂(p1 , . . . , pn ) ∂(Q1 , . . . , Qn , P1 , . . . , Pn )

has  of {pj , pk } = 0 for j, k = 1, . . . , n, the n vectors  rank n. Moreover, in view ∂pj ∂pj ∂pj ∂pj J ∂Q1 , . . . , ∂Qn , ∂P1 , . . . , ∂Qn span a n–dimensional Lagrangian subspace of the tangent space to F at every point (Q, P ). On the other hand, at every point (Q, P ) we can define a canonical basis {e1 , . . . , en , d1 , . . . , dn } by just setting     ∂Qj ∂Qj ∂Pj ∂Pj ej = ,..., , 0, . . . , 0 , dj = 0, . . . , 0, ,..., , j = 1, . . . , n . ∂Q1 ∂Qn ∂P1 ∂Pn By lemma 2.7, the Lagrangian subspace above is complementary to at least one of the Lagrangian arithmentic planes of the canonical basis. This means that there exists a partition {J, K} of {1, . . . , n} such that the 2n × 2n matrix obtained by adding to (2.38) the n rows {ek }k∈K ∪ {dj }j∈J has non zero determinant. On the other hand, the determinant turns out to be exactly (2.37), so that we conclude that J, K is the wanted partition. Q.E.D.

2.5

Time–depending canonical transformations

We show here how the theory of canonical transformations can be generalized so that the cases of non autonomous Hamiltonians and of time depending transformations are taken into account.

44

Chapter 2

We use the extension of the phase space discussed in sect. 1.1.1, i.e., we introduce two further canonical variables q+ , p+ and for a given Hamiltonian H(q, p, t) we consider the Hamiltonian in the extended phase space8 (2.39)

˜ p, q+ , p+ ) = H(q, p, q+ ) + p+ . H(q,

On the extended space phase we can perform canonical transformations of the form q = q (Q, P, Q+, P+ ), p = p (Q, P, Q+, P+ ), q+ = q+ (Q, P, Q+, P+ ), p+ = p+ (Q, P, Q+, P+ ), to which the theory developed till now applies. However, this means that we change also the time variable, in the sense that the new variable Q+ will not evolve uniformly in time. The natural choice is to consider a restricted class of transformations which keeps the coordinate q+ invariant, namely q+ = Q+ ; in turn, p+ = p+ (Q, P, Q+, P+ ) will be determined so as to fulfill the canonicity conditions. A first consequence is that the condition {q+ , p+ } = 1 implies p+ = P+ + f (Q, P, q+), with some function f . A second consequence is that the canonicity conditions {q+ , qj } = {q+ , pj } = 0 for 1 ≤ j ≤ n imply that qj , pj do not depend on P+ . This also means that the Poisson brackets {qj , qk } , {qj , pk } and {pj , pk } are actually computed by differentiating only with respect to the variables Q, P . The general scheme is the following: in order to perform a time dependent transformation we first consider the extended phase space and the Hamiltonian (2.39), and perform a transformation satisfying the conditions above. This means that the transformed Hamiltonian will take the form K(Q, P, Q+, P+ ) = H(q, p, q+ ) q=q(Q,P,Q ), p=p(Q,P,Q ),q =Q + P+ + f (Q, P, Q+) . +

+

+

+

In view of the linear dependence on P+ we remove the extension of the phase space by setting again Q+ = t and removing the term P+ , thus obtaining the transformed Hamiltonian K(Q, P, t) = H(q, p, t) q=q(Q,P,t), p=p(Q,P,t) + f (Q, P, t) .

We emphasize that the new Hamiltonian is not merely the transformed function of the old one: there is an extra term that must be computed. The following propositions show that the canonicity conditions to be checked are the ones discussed till now, that must be fulfilled identically in t, and explain how to determine the extra term in the Hamiltonian.

Proposition 2.19: Let q = q(Q, P, t), p = p(Q, P, t) be a time dependent transformation which preserves the fundamental Poisson brackets identically in t. Then the transformation is canonical, and there exists a function F (q, p, t) such that the transformed Hamiltonian is   (2.40) K(Q, P, t) = H(q, p, t) − F (q, p, t) q=q(Q,P,t),p=p(Q,P,t) 8

We emphasize the particular role played by the variables q+ , p+ by denoting the canonical coordinates in phase space as (q, p, q+ , p+ ), where q = (q1 , . . . , qn ) and p = (p1 , . . . , pn ).

Canonical transformations

45

Proof. We just prove that in the extended phase space there is a function F (q, p, Q+) such that the extended transformation (2.41)

q = q(Q, P, Q+) ,

p = p(Q, P, Q+) ,

q+ = Q+

p+ = P+ − F (q, p, Q+) q=q(Q,P,Q

,

+ ) , p=p(Q,P,Q+ )

is canonical. Differentiating with respect to Q+ the relations {qj , qk } = {pj , pk } = 0 , {qj , pk } = δjk for 1 ≤ j ≤ n (true in view of the assumed preservation of the fundamental Poisson brackets) we get     ∂qj ∂qk , qk + qj , =0, ∂Q+ ∂Q+     ∂pk ∂qj , pk + qj , =0, ∂Q+ ∂Q+     ∂pk ∂pj , pk + pj , =0. ∂Q+ ∂Q+ Recalling that q, p do not depend on P+ , and denoting fj = we write the identities above as ∂fk ∂fj − =0, ∂pk ∂pj

∂qj , ∂Q+

gj =

∂pj , ∂Q+

∂fj ∂gk − =0, ∂qk ∂pj

∂gj ∂gk − =0. ∂qk ∂qj

This implies the (local) existence of a function F (q, p, Q+) such that ∂qj ∂F = , ∂Q+ ∂pj

(2.42)

∂pj ∂F =− . ∂Q+ ∂qj

With this function we complete the transformation as in (2.41). The canonicity of the extended transformation is checked by remarking that {q+ , qj } = {q+ , pj } = 0, {q+ , p+ } = 1 in view of the assumed preservation of the fundamental Poisson brackets for all t, and so for all Q+ , and that {p+ , qj } =

∂qj − {F, qj } , ∂Q+

{p+ , pj } =

∂pj − {F, pj } ∂Q+

are zero in view of (2.42). Replacing the transformation in the Hamiltonian (2.39) and removing the extension of the phase space the claim follows. Q.E.D. Proposition 2.20: Let S(P, q, t) be a function satisfying   2 ∂ S 6= 0 . det ∂Pj ∂qk Then the transformation implicitly defined by Qj =

∂S , ∂Pj

pj =

∂S , ∂qj

1≤j≤n

46

Chapter 2

is canonical, and the transformed Hamiltonian takes the form ∂S (P, q, t) . (2.43) K(Q, P, t) = H(q, p, t) + ∂t q=q(Q,P,t) q=q(Q,P,t),p=p(Q,P,t) Proof.

In the extended phase space consider the generating function ˜ S(P, q, P+ , q+ ) = P+ q+ + S(P, q, q+ ) .

The corresponding transformation is Qj = (2.44)

∂S , ∂Pj

Q+ = q+ ,

pj =

∂S , ∂qj

p+ = P+ +

1≤j≤n, ∂S . ∂q+

Replacing the transformation in the Hamiltonian (2.39) and removing the extension of the phase space the claim follows. Q.E.D.

2.6

The equation of Hamilton–Jacobi

The integration of the canonical equations can be performed by looking for a generating function of a canonical transformation giving the Hamiltonian a particularly simple form. It is customary to use the formalism of time dependent canonical transformation. Having given the Hamiltonian H(q, p, t) we look for a function S which is a solution of the Hamilton–Jacobi equation ([44], [46], [52], [53])   ∂S ∂S ,t + =0. (2.45) H q, ∂q ∂t We are actually looking for the generating function of a transformation such that the transformed Hamiltonian is identically zero. The problem is to find a solution of eq. (2.45) depending on q1 , . . . , qn , t and on n arbitrary parameters α1 , . . . , αn ; this is said to be a complete integral. Proposition 2.21: Consider the Hamiltonian H(q, p, t), and assume that we are given a complete integral S(α, q, t) of Hamilton–Jacobi’s equation (2.45), depending on n arbitrary parameters α1 , . . . , αn and satisfying   2 ∂ S 6= 0 . det ∂αj ∂qk Then the solutions of the canonical equations are written in implicit form as (2.46)

βj =

∂S (α, q, t) , ∂αj

pj =

∂S (α, q, t) , ∂qj

1≤j≤n,

where α1 , . . . , αn , β1 , . . . , βn are constants depending on the initial data.

Canonical transformations

47

Proof. The function S(α, q, t) satisfies the conditions of proposition 2.20; therefore, it is the generating function of a canonical transformation, actually the transformation (2.46). Since the transformed Hamiltonian is identically zero, the corresponding canonical equations are α˙ j = 0 ,

β˙j = 0 ,

1≤j≤n,

i.e., α, β are constants depending on the initial data. By inversion of (2.46) with respect to q, p one gets functions q = q(α, β, t) ,

p = p(α, β, t) ,

i.e., the wanted solutions of the canonical equations. Example 2.14: Free particle

Q.E.D.

Let the Hamiltonian be

1 2 (p + p2y + p2z ) . 2m x The corresponding Hamilton–Jacobi’s equation is " 2  2  2 # 1 ∂S ∂S ∂S ∂S + + + =0. 2m ∂x ∂y ∂z ∂t H=

We use the method of separation of variables. We look for a solution of the form S(x, y, z, t) = X(x) + Y (y) + Z(z) + T (t) , so that the equation is rewritten as " 2  2  2 # dX dY dZ dT 1 + + + =0. (2.47) 2m dx dy dz dt Therefore, we obtain the equations9 dX = αx , dx

dY = αy , dy

α2x + α2y + α2z dT =− , dt 2m

dZ = αz , dz

with αx , αy , αz arbitrary constants. By integration we construct the generating function α2x + α2y + α2z t, S(αx , αy , αz , x, y, z, t) = αx x + αy y + αz z − 2m so that the transformation is px = αx , py = αy , pz = αz αx αy αz βx = x − t , βy = y − t , βz = z − t. m m m This is the solution of the canonical equations. 9

2

dX Differentiating (2.47) with respect to x we get ddxX 2 = 0, so that dx must be constant. Similarly, differentiating with respect to y, z and t we get that dY , dZ and dT are condy dz dt stants, too. In view of (2.47), only three of these constants are arbitrary.

48

Chapter 2