Introduction
In this paper we study the role of Dirac structures in the geometric formulation and symmetry reduction of a class of optimal control problems. More specifically, we show how the reduction theory for Dirac structures developed by Yoshimura and Marsden [41 ] can be extended to optimal control systems admitting a Lie group of symmetries. We prove that, as in the case of implicit Hamiltonian and Lagrangian systems, the reduction of the variational principle and the reduction of the Dirac structure underlying the Pontryagin Maximum Principle (PMP) lead to the same reduced equations, and we then use the Dirac formulation to treat the singular case, in which the control cannot be eliminated and a constraint algorithm is required.
The reduced equations themselves are not new. They are the Hamiltonian equations on the dual of the Atiyah algebroid T P / G TP/G T P / G , obtained in that setting by Martínez [29 ] and, in the wider context of almost Lie algebroids, by Grabowski and Jóźwikowski [23 ] ; for a Lie group they go back to Jurdjevic [27 ] . Section 3.7 records the dictionary explicitly. What the Dirac formulation adds is a framework in which the constraint algorithm and the symmetry reduction interact, and it is there that our main new result lies.
Optimal control problems often possess a rich geometric structure. Throughout this paper we restrict our attention to smooth geometric optimal control problems, where both the state space and the control space are smooth manifolds (assumed, for simplicity, to be without boundary), the controlled dynamics is described by a control-dependent vector field, and the cost functional is defined by a Lagrangian density. This setting naturally admits a differential-geometric description and provides the appropriate framework for studying additional geometric features, such as symmetries and reduction.
Suppose that a Lie group acts on the state and control spaces in such a way that both the dynamics and the cost functional are invariant under the action. We then say that the optimal control problem is symmetric with respect to the given Lie group. The presence of symmetry is a powerful tool for simplifying the analysis of dynamical systems. In Hamiltonian and Lagrangian mechanics, conserved quantities associated with symmetries often allow one to reduce the dimension of the problem. More generally, reduction techniques replace the original system with a family of lower-dimensional systems whose solutions can later be reconstructed to recover the solutions of the original dynamics.
The theory of reduction for systems with symmetry has a long history, and we do not attempt to review it here. Instead, we focus on the work of Yoshimura and Marsden, who developed a unified framework for implicit Lagrangian and Hamiltonian systems based on Dirac structures. Their contributions include the formulation of implicit Lagrangian mechanics using Dirac structures [37 , 38 ] , the reduction of the canonical Dirac structure on the cotangent bundle of a Lie group [40 ] , the treatment of implicit Hamiltonian and Lagrangian systems [39 ] , and, ultimately, the reduction of Dirac structures on cotangent bundles [41 ] . These developments provide the main geometric inspiration for the present work.
The Pontryagin Maximum Principle admits a natural presymplectic formulation, which has been extensively developed in the literature (see, for example, [6 , 5 ] and the references therein). Since every presymplectic manifold carries a canonical Dirac structure, the PMP can also be formulated entirely in terms of Dirac geometry. This observation makes it natural to ask whether the reduction theory for Dirac structures can be used to reduce optimal control problems while preserving not only the dynamics but also their variational character.
A straightforward approach would be to perform presymplectic reduction of the Hamiltonian system associated with the PMP. Although this reduces the number of degrees of freedom, it generally loses the variational information encoded in the original optimal control problem. Consequently, the reduced dynamics alone does not guarantee that its solutions correspond to extremals of the original cost functional. What is needed is a reduction procedure that simultaneously preserves the geometric dynamics and the underlying variational principle.
Dirac structures provide precisely such a framework. As shown by Yoshimura and Marsden, the reduction of Dirac structures is compatible with the reduction of the associated variational principles. This makes Dirac geometry particularly well suited to the study of symmetric optimal control problems, where one seeks a reduced formulation that retains both the geometric and variational content of the original problem.
Following this perspective, we extend the Dirac reduction framework of [41 ] to geometric optimal control. The geometric formulation of the PMP naturally leads to the introduction of the Pontryagin bundle , a bundle over the state manifold that plays a role analogous to the Pontryagin bundle T Q ⊕ T ∗ Q TQ \oplus T^{*}Q T Q ⊕ T ∗ Q arising in implicit Lagrangian mechanics. We formulate the first-order necessary conditions of the PMP as an implicit Hamiltonian system determined by the Pontryagin Hamiltonian and the canonical Dirac structure on this bundle. Assuming the optimal control problem admits a symmetry group G G G , we study the reduction of both the variational principle and the associated Dirac structure.
The reduced PMP equations are obtained in two independent ways. First, we derive them by reducing the variational principle obtained through the method of Lagrange multipliers. Second, we obtain the same equations by directly reducing the Dirac structure defining the PMP, the reduced canonical Dirac structure being derived in Section 4.3 rather than postulated. The agreement between these two constructions shows that the reduced Dirac system provides the correct geometric realization of the reduced optimal control problem.
It may be helpful to state at the outset what we claim and what we do not. Our main new result is Theorem 5: symmetry reduction commutes with the constraint algorithm, so that for a singular problem the algorithm may be run on the reduced space and the unreduced constraint submanifolds recovered by pulling back. To formulate it we give a Dirac-geometric form of the algorithm, in which each stage is the Dirac structure induced by the constraint distribution and the constraint forces appear in its annihilator. Alongside this we supply the reduction picture in full: reconstruction (Theorem 1), independence of the choice of principal connection (Theorem 2), and the relation between criticality and optimality (Proposition 4). A second new result is Theorem 8, which decides in reduced terms whether an abnormal extremal is strictly abnormal : it is not, precisely when its projection to the base satisfies Wong’s equation for some parallel coadjoint multiplier. The criterion separates cleanly what the curvature alone determines from what the cost contributes, and it is decided by a second constraint algorithm run in the coadjoint fiber (Proposition 17). § 8.9 then supplies the second-order theory in the same variables: for an abelian structure group the index form of an abnormal extremal is ∫ ⟨ μ , F A ( ∇ z ˙ v , v ) + ( ∇ v F A ) ( z ˙ , v ) ⟩ d t \int\langle\mu,F_A(\nabla_{\dot{z}}v,v) + (\nabla_vF_A)(\dot{z},v)\rangle\,dt ∫ ⟨ μ , F A ( ∇ z ˙ v , v ) + ( ∇ v F A ) ( z ˙ , v )⟩ d t (Theorem 9), which involves the curvature and its covariant derivative but not the metric. The abelian hypothesis is not removable: for nonabelian holonomy the second variation acquires time-ordered commutator terms that we do not compute, and Remark 24 exhibits an S O ( 3 ) SO(3) S O ( 3 ) example in which they destroy positivity on long arcs. Location, rigidity and strictness are then three reduced tests (Theorem 10), the first and third criteria and the second a sufficient certificate, and the deformed Martinet family of [1 ] together with the rank-two example of § 8.8 shows that the last two are genuinely independent.
By contrast, several ingredients are recastings of known results in the present language, and we have tried to attribute them carefully. The constraint algorithm is that of Gotay, Nester and Hinds [21 ] , applied to the PMP in the presymplectic setting by Barbero-Li~nan and Mu~noz-Lecanda [5 ] , who also identified its secondary constraints with the classical Goh and generalized Legendre–Clebsch conditions. The device of adjoining the running cost to the state, which brings abnormal extremals within the same framework, is due to Ohsawa [34 ] and Jóźwikowski and Respondek [25 ] ; we only observe that it is itself an instance of the reduction theory developed here, the cost multiplier being a momentum map. The characterization of abnormal extremals in § 8.5 is the specialization to a principal bundle of criteria that are classical [2 , 25 ] ; what is ours is the observation that the obstruction is exactly the curvature term by which the reduced PMP differs from the unreduced one.
The remainder of the paper is organized as follows. In § 2 we introduce the geometric framework for the class of optimal control problems under consideration. We review the necessary notions of Pontryagin bundles, Dirac structures, and symmetries, and establish the notation used throughout the paper.
In § 3 we study the reduction of the variational formulation of the Pontryagin Maximum Principle. We derive the reduced Pontryagin bundle, formulate the reduced variational principle, and obtain the corresponding first-order necessary conditions, which we refer to as the reduced Pontryagin Maximum Principle. The section closes with the four results that complete the reduction picture: reconstruction, independence of the connection, the relation between criticality and optimality, and the comparison with Lie algebroid reduction.
§ 4 approaches the reduction problem from the perspective of Dirac geometry. We construct the reduced Dirac structure associated with the Pontryagin bundle, derive its connection-dependent description, and relate it to the reduction of the canonical Dirac structure on the cotangent bundle. We show that the reduced Dirac formulation yields exactly the same reduced equations obtained from the variational approach.
We then illustrate the theory with two complementary examples. In § 5 we treat the optimal control of a rigid spacecraft carrying three internal reaction wheels, in which the symmetry group S O ( 3 ) SO(3) S O ( 3 ) is nonabelian but the connection is flat, so that the reduced equations are governed by the coadjoint term. In § 7 we treat the sub-Riemannian geodesic problem on the Heisenberg group, in which the symmetry group is the abelian center and the coadjoint term vanishes, while the curvature of the natural connection is nowhere zero. Between them the two examples exercise both of the terms by which the reduced Pontryagin Maximum Principle differs from its unreduced counterpart.
§ 8 removes the regularity assumption made in § 2 –§ 7 . It is here that the Dirac formulation does work that the presymplectic one does not: we give the constraint algorithm in Dirac form, prove that reduction commutes with it, and show that for control-linear systems the abnormal extremals occupy the leaf on which the base costate vanishes, the first secondary constraint being the degeneracy of the curvature pairing. The Martinet problem, in which the curvature vanishes on a hypersurface and the abnormal extremals are exactly the curves lying in it, is worked out as a third example, in both its abnormal and its normal leaf. The section closes with the strictness criterion, with an example, of structure group R 2 \mathbb{R}^{2} R 2 , in which strictness is decided only at the second stage of the algorithm, and with the reduced index form and the resulting trichotomy.
Geometrical and Variational Aspects of Optimal Control Problems with Symmetry
As discussed in the introduction, we restrict our attention to a class of geometric optimal control problems for which the underlying differential-geometric structures can be described in a transparent way. Throughout the paper, all manifolds are assumed to be smooth and without boundary, and all maps, vector fields, differential forms, and curves are assumed to be smooth.
For simplicity, we consider in § 2 –§ 7 only regular optimal control problems, excluding abnormal extremals and singular control systems (see, for example, [2 , 26 ] ); the regularity assumption is removed in § 8 . Moreover, admissible controls are assumed to be smooth. These assumptions allow us to focus on the geometric aspects of the theory without addressing the additional analytical difficulties arising from nonsmooth controls or more general admissible classes. Most of the constructions presented here extend, with appropriate modifications, to more general settings, including systems defined on manifolds with boundary or singular optimal control problems.
We begin by recalling the geometric framework and fixing the notation used throughout the paper.
A geometric formulation of optimal control problems
We follow the standard geometric formulation of optimal control (see, for instance, [2 , 26 , 13 ] ).
Definition 1 (Control system). A control system is a triple ( P , C , Γ ) (P,C,\Gamma) ( P , C , Γ ) consisting of:
(i) A smooth manifold P P P , called the state space. Local coordinates on P P P are denoted by ( x i ) (x^i) ( x i ) , i = 1 , … , n i=1,\ldots,n i = 1 , … , n , and a generic point by x x x .
(ii) A locally trivial fiber bundle π : C ⟶ P \pi:C\longrightarrow P π : C ⟶ P , called the control bundle. For each state x ∈ P x\in P x ∈ P , the fiber C x = π − 1 ( x ) C_x=\pi^{-1}(x) C x = π − 1 ( x ) is the manifold of admissible controls at x x x . Adapted local coordinates on C C C are denoted by ( x i , u a ) (x^i,u^a) ( x i , u a ) , where a = 1 , … , r a=1,\ldots,r a = 1 , … , r , and a generic point is written ( x , u ) (x,u) ( x , u ) .
(iii) A vector field along the projection π : C ⟶ P \pi:C\longrightarrow P π : C ⟶ P , that is, a smooth map Γ : C ⟶ T P \Gamma:C\longrightarrow TP Γ : C ⟶ T P satisfying τ P ∘ Γ = π \tau_P\circ\Gamma=\pi τ P ∘ Γ = π , where τ P : T P ⟶ P \tau_P:TP\longrightarrow P τ P : T P ⟶ P denotes the canonical projection. The map Γ \Gamma Γ describes the controlled dynamics; in local coordinates, Γ = f i ( x , u ) ∂ / ∂ x i \Gamma=f^i(x,u)\partial/\partial x^i Γ = f i ( x , u ) ∂ / ∂ x i , where natural coordinates on T P TP T P are denoted by ( x i , v i ) (x^i,v^i) ( x i , v i ) , or equivalently by ( x i , x ˙ i ) (x^i,\dot{x}^i) ( x i , x ˙ i ) . An admissible control is a smooth curve c : I ⟶ C c:I\longrightarrow C c : I ⟶ C , defined on an interval I ⊂ R I\subset\mathbb{R} I ⊂ R , with state trajectory x = π ∘ c : I ⟶ P x=\pi\circ c:I\longrightarrow P x = π ∘ c : I ⟶ P . Writing locally c ( t ) = ( x ( t ) , u ( t ) ) c(t)=(x(t),u(t)) c ( t ) = ( x ( t ) , u ( t )) , the controlled dynamics is described by
x ˙ i ( t ) = f i ( x ( t ) , u ( t ) ) , i = 1 , … , n . (1) \dot{x}^i(t)=f^i(x(t),u(t)),\qquad i=1,\ldots,n. \tag*{(1)} x ˙ i ( t ) = f i ( x ( t ) , u ( t )) , i = 1 , … , n . (1)
Definition 2 (Optimal control problem). An optimal control problem ( O C P ) (OCP) ( O C P ) is a quadruple ( P , C , Γ , L ) (P,C,\Gamma,L) ( P , C , Γ , L ) consisting of
(i) A control system ( P , C , Γ ) (P,C,\Gamma) ( P , C , Γ ) .
(ii) A smooth cost function L : C ⟶ R L:C\longrightarrow\mathbb{R} L : C ⟶ R , called the running cost (or Lagrangian), which defines the objective functional
S ( c ) = ∫ 0 T L ( c ( t ) ) d t = ∫ 0 T L ( x ( t ) , u ( t ) ) d t , (2) S(c)=\int_0^T L(c(t))\,dt=\int_0^T L(x(t),u(t))\,dt, \tag*{(2)} S ( c ) = ∫ 0 T L ( c ( t )) d t = ∫ 0 T L ( x ( t ) , u ( t )) d t , (2) where c : [ 0 , T ] ⟶ C c:[0,T]\longrightarrow C c : [ 0 , T ] ⟶ C is an admissible control satisfying π ( c ( t ) ) = x ( t ) \pi(c(t))=x(t) π ( c ( t )) = x ( t ) , and the associated state trajectory x ( t ) x(t) x ( t ) satisfies the controlled dynamics (1) together with the prescribed boundary conditions x ( 0 ) = x 0 x(0)=x_0 x ( 0 ) = x 0 , x ( T ) = x T x(T)=x_T x ( T ) = x T . The optimal control problem consists of finding admissible controls minimizing the functional (2).
The presymplectic formulation of optimal control problems
The fundamental tool in optimal control theory is the Pontryagin Maximum Principle (PMP), which provides first-order necessary conditions for optimality. Throughout this paper we assume that optimal trajectories exist and, until Section 7, restrict our attention to regular optimal control problems.
The geometric formulation of the PMP naturally leads to the introduction of the following bundle.
Definition 3 (Pontryagin bundle). The Pontryagin bundle associated with the control system ( P , C , Γ ) (P,C,\Gamma) ( P , C , Γ ) is the fiber product M = T ∗ P × P C M=T^*P\times_P C M = T ∗ P × P C , whose fiber over a point x ∈ P x\in P x ∈ P is M x = T x ∗ P × C x M_x=T_x^*P\times C_x M x = T x ∗ P × C x . Equivalently,
M = { ( α x , c x ) ∈ T ∗ P × C ∣ π P ( α x ) = π ( c x ) = x } . M=\{(\alpha_x,c_x)\in T^*P\times C\mid\pi_P(\alpha_x)=\pi(c_x)=x\}. M = {( α x , c x ) ∈ T ∗ P × C ∣ π P ( α x ) = π ( c x ) = x } .
The Pontryagin bundle inherits natural local coordinates ( x i , p i , u a ) (x^i,p_i,u^a) ( x i , p i , u a ) , induced from the canonical coordinates ( x i , p i ) (x^i,p_i) ( x i , p i ) on T ∗ P T^*P T ∗ P and the adapted coordinates ( x i , u a ) (x^i,u^a) ( x i , u a ) on the control bundle C C C .
We denote by pr P : M ⟶ T ∗ P \operatorname{pr}_P:M\longrightarrow T^*P pr P : M ⟶ T ∗ P and pr C : M ⟶ C \operatorname{pr}_C:M\longrightarrow C pr C : M ⟶ C the canonical projections onto the two factors,
pr P ( x , p , u ) = ( x , p ) , pr C ( x , p , u ) = ( x , u ) . (3) \operatorname{pr}_P(x,p,u)=(x,p),\qquad\operatorname{pr}_C(x,p,u)=(x,u). \tag*{(3)} pr P ( x , p , u ) = ( x , p ) , pr C ( x , p , u ) = ( x , u ) . (3)
These maps fit into the commutative diagram
M = T ∗ P × P C → pr C C pr P ↓ ↓ π T ∗ P → π P P \begin{CD}
M=T^*P\times_P C @>{\operatorname{pr}_C}>> C \\
@V{\operatorname{pr}_P}VV @VV{\pi}V \\
T^*P @>{\pi_P}>> P
\end{CD} M = T ∗ P × P C pr P ↓ ⏐ T ∗ P pr C π P C ↓ ⏐ π P
The Pontryagin bundle carries a natural presymplectic form obtained by pulling back the canonical symplectic form ω P \omega_P ω P on T ∗ P T^*P T ∗ P , namely
Ω = pr P ∗ ω P = d x i ∧ d p i . (4) \Omega=\operatorname{pr}_P^*\omega_P=dx^i\wedge dp_i. \tag*{(4)} Ω = pr P ∗ ω P = d x i ∧ d p i . (4)
The control vector field Γ \Gamma Γ and the running cost L L L determine the Pontryagin Hamiltonian
H P ( x , p , u ) = ⟨ p , Γ ( x , u ) ⟩ − L ( x , u ) = p i f i ( x , u ) − L ( x , u ) . (5) H_P(x,p,u)=\langle p,\Gamma(x,u)\rangle-L(x,u)=p_i f^i(x,u)-L(x,u). \tag*{(5)} H P ( x , p , u ) = ⟨ p , Γ ( x , u )⟩ − L ( x , u ) = p i f i ( x , u ) − L ( x , u ) . (5)
The presymplectic system ( M , Ω , H P ) (M,\Omega,H_P) ( M , Ω , H P ) provides a geometric formulation of the Pontryagin Maximum Principle. Indeed, its integral curves are precisely the extremals of the optimal control problem ( P , C , Γ , L ) (P,C,\Gamma,L) ( P , C , Γ , L ) . The characteristic distribution of Ω \Omega Ω is
K = ker Ω = ker ( pr P ) ∗ = span { ∂ ∂ u a ; a = 1 , … , r } , K=\ker\Omega=\ker(\operatorname{pr}_P)_*=\operatorname{span}\left\{\frac{\partial}{\partial u^a};\ a=1,\ldots,r\right\}, K = ker Ω = ker ( pr P ) ∗ = span { ∂ u a ∂ ; a = 1 , … , r } ,
that is, the vertical bundle of the projection pr P : M ⟶ T ∗ P \operatorname{pr}_P:M\longrightarrow T^*P pr P : M ⟶ T ∗ P .
The dynamics is determined by the presymplectic equation
i X P Ω = d H P . (6) \mathord{\mathrm{i}}_{X_P}\Omega=dH_P. \tag*{(6)} i X P Ω = d H P . (6)
Since Ω \Omega Ω is degenerate, solutions of (6) do not exist everywhere. A necessary condition for solvability [21 ] is ⟨ Z , d H P ⟩ = 0 \langle Z,dH_P\rangle=0 ⟨ Z , d H P ⟩ = 0 , ∀ Z ∈ K \forall Z\in K ∀ Z ∈ K , which, in local coordinates, becomes
φ a = ∂ H P ∂ u a = 0 , a = 1 , … , r . (7) \varphi_a=\frac{\partial H_P}{\partial u^a}=0,\qquad a=1,\ldots,r. \tag*{(7)} φ a = ∂ u a ∂ H P = 0 , a = 1 , … , r . (7)
The functions φ a \varphi_a φ a are called the primary constraints , and their common zero set
M 1 = { ( x , p , u ) ∈ M ∣ φ a ( x , p , u ) = 0 , a = 1 , … , r } (8) M_1=\{(x,p,u)\in M\mid\varphi_a(x,p,u)=0,\ a=1,\ldots,r\} \tag*{(8)} M 1 = {( x , p , u ) ∈ M ∣ φ a ( x , p , u ) = 0 , a = 1 , … , r } (8)
is the primary constraint submanifold . Under the regularity assumptions considered in this paper, M 1 M_1 M 1 is a smooth submanifold of codimension r r r .
Let i 1 : M 1 ↪ M i_1:M_1\hookrightarrow M i 1 : M 1 ↪ M denote the canonical inclusion. We write
Ω 1 = i 1 ∗ Ω , H 1 = i 1 ∗ H P , (9) \Omega_1=i_1^*\Omega,\qquad H_1=i_1^*H_P, \tag*{(9)} Ω 1 = i 1 ∗ Ω , H 1 = i 1 ∗ H P , (9)
for the induced presymplectic form and Hamiltonian on M 1 M_1 M 1 , and by p 1 = pr P ∘ i 1 : M 1 ⟶ T ∗ P p_1=\operatorname{pr}_P\circ i_1:M_1\longrightarrow T^*P p 1 = pr P ∘ i 1 : M 1 ⟶ T ∗ P the restriction of pr P \operatorname{pr}_P pr P to M 1 M_1 M 1 . Then
p 1 ∗ ω P = Ω 1 , K 1 : = ker ( p 1 ) ∗ = T M 1 ∩ K ⊆ ker Ω 1 , p_1^*\omega_P=\Omega_1,\qquad K_1:=\ker(p_1)_*=TM_1\cap K\subseteq\ker\Omega_1, p 1 ∗ ω P = Ω 1 , K 1 := ker ( p 1 ) ∗ = T M 1 ∩ K ⊆ ker Ω 1 ,
with equality precisely at those points where ( p 1 ) ∗ ( T M 1 ) (p_1)_*(TM_1) ( p 1 ) ∗ ( T M 1 ) is a symplectic subspace of T ( T ∗ P ) T(T^*P) T ( T ∗ P ) , hence in particular at every regular point (see Proposition 1 below). The inclusion may be strict at singular points: for n = r = 1 n=r=1 n = r = 1 and H P = p u − 1 3 u 3 H_P=pu-\frac{1}{3}u^3 H P = p u − 3 1 u 3 one has M 1 = { p = u 2 } M_1=\{p=u^2\} M 1 = { p = u 2 } and Ω 1 = 2 u d x ∧ d u \Omega_1=2u\,dx\wedge du Ω 1 = 2 u d x ∧ d u , so that at u = 0 u=0 u = 0 the radical ker Ω 1 = T M 1 \ker\Omega_1=TM_1 ker Ω 1 = T M 1 is two-dimensional whereas K 1 = span { ∂ u } K_1=\operatorname{span}\{\partial_u\} K 1 = span { ∂ u } .
Where the differentials d φ 1 , … , d φ r d\varphi_1,\ldots,d\varphi_r d φ 1 , … , d φ r are linearly independent, so that T M 1 = ker d φ TM_1=\ker d\varphi T M 1 = ker d φ , a vector C a ∂ / ∂ u a C^a\partial/\partial u^a C a ∂ / ∂ u a of K K K is tangent to M 1 M_1 M 1 if and only if C a W a b = 0 C^aW_{ab}=0 C a W ab = 0 , where
W = W a b d u a ⊗ d u b , W a b = ∂ 2 H P ∂ u a ∂ u b , W=W_{ab}\,du^a\otimes du^b,\qquad W_{ab}=\frac{\partial^2H_P}{\partial u^a\partial u^b}, W = W ab d u a ⊗ d u b , W ab = ∂ u a ∂ u b ∂ 2 H P ,
is the symmetric bilinear form along the fibers of pr P \operatorname{pr}_P pr P ; hence K 1 ( ξ ) ≠ { 0 } K_1(\xi)\ne\{0\} K 1 ( ξ ) = { 0 } if and only if det W a b ( ξ ) = 0 \det W_{ab}(\xi)=0 det W ab ( ξ ) = 0 . We say that a point ξ ∈ M 1 \xi\in M_1 ξ ∈ M 1 is regular if W W W is nondegenerate at ξ \xi ξ , and singular otherwise. The following characterization is proved in [20 ] .
Proposition 1. Let ξ ∈ M 1 \xi\in M_1 ξ ∈ M 1 and assume that the differentials d φ 1 , … , d φ r d\varphi_1,\ldots,d\varphi_r d φ 1 , … , d φ r are linearly independent at ξ \xi ξ , so that near ξ \xi ξ the set M 1 M_1 M 1 is an embedded submanifold of M M M of codimension r r r with T M 1 = ker d φ TM_1=\ker d\varphi T M 1 = ker d φ (this is automatic at a regular point, since the d u du d u -block of the matrix of the d φ b d\varphi_b d φ b is W W W ). Then the following statements are equivalent: (i) ξ \xi ξ is a regular point; (ii) K 1 ( ξ ) = { 0 } K_1(\xi)=\{0\} K 1 ( ξ ) = { 0 } ; (iii) M 1 M_1 M 1 is transverse at ξ \xi ξ to the fibers of pr P \operatorname{pr}_P pr P ; (iii) ker Ω 1 ( ξ ) = { 0 } \ker\Omega_1(\xi)=\{0\} ker Ω 1 ( ξ ) = { 0 } ; (iv) p 1 p_1 p 1 is a local symplectomorphism between ( M 1 , Ω 1 ) (M_1,\Omega_1) ( M 1 , Ω 1 ) and ( T ∗ P , ω P ) (T^*P,\omega_P) ( T ∗ P , ω P ) in a neighbourhood of ξ \xi ξ . In that case the equation i X Ω 1 = d H 1 \mathord{\mathrm{i}}_X\Omega_1=dH_1 i X Ω 1 = d H 1 has a unique local solution X X X near ξ \xi ξ , whose integral curve through ξ \xi ξ is the unique extremal of the optimal control problem through ξ \xi ξ .
If ξ \xi ξ is a singular point, then K 1 ( ξ ) ≠ { 0 } K_1(\xi)\ne\{0\} K 1 ( ξ ) = { 0 } and the presymplectic equation i X P Ω 1 = d H 1 \mathord{\mathrm{i}}_{X_P}\Omega_1=dH_1 i X P Ω 1 = d H 1 may fail to admit a local solution, so that there may exist no optimal trajectory passing through ξ \xi ξ . Extremals passing through singular points are called singular arcs .
Throughout § 2 –§ 7 we restrict our attention to regular optimal control problems , namely those for which every point of M 1 M_1 M 1 is regular. The Dirac formulation developed below remains valid in the singular case, which is taken up in § 8 .
Since Ω = pr P ∗ ω P \Omega=\operatorname{pr}_P^*\omega_P Ω = pr P ∗ ω P , a solution X P X_P X P of (6) enters the equation only through its projection T pr P ∘ X P T\operatorname{pr}_P\circ X_P T pr P ∘ X P . The dynamics of the Pontryagin Maximum Principle is therefore naturally regarded as a vector field along the projection pr P : M ⟶ T ∗ P \operatorname{pr}_P:M\longrightarrow T^*P pr P : M ⟶ T ∗ P , defined on the locus M 1 M_1 M 1 on which (6) is solvable, that is, as a smooth map X : M 1 ⟶ T ( T ∗ P ) X:M_1\longrightarrow T(T^*P) X : M 1 ⟶ T ( T ∗ P ) satisfying τ T ∗ P ∘ X = pr P ∘ i 1 = p 1 \tau_{T^*P}\circ X=\operatorname{pr}_P\circ i_1=p_1 τ T ∗ P ∘ X = pr P ∘ i 1 = p 1 ; this is the form in which the dynamics enters § 2.5 and § 4.4 .
The Lagrange–Pontryagin principle
We now describe the variational formulation underlying the presymplectic picture developed above. Consider the optimal control problem ( P , C , Γ , L ) (P,C,\Gamma,L) ( P , C , Γ , L ) and let M = T ∗ P × P C M=T^*P\times_P C M = T ∗ P × P C be its Pontryagin bundle. The optimal control problem may be regarded as a constrained variational problem, where the functional (2) is minimized subject to the controlled dynamics (1). Under suitable regularity assumptions, the method of Lagrange multipliers shows that the extremals are precisely the critical points of the functional S : C x 0 , x T ( M ) ⟶ R S:\mathcal{C}_{x_0,x_T}(M)\longrightarrow\mathbb{R} S : C x 0 , x T ( M ) ⟶ R defined by
S ( γ ) = ∫ 0 T ( L ( x ( t ) , u ( t ) ) + ⟨ p ( t ) , x ˙ ( t ) − Γ ( x ( t ) , u ( t ) ) ⟩ ) d t , (10) S(\gamma)=\int_0^T\left(L(x(t),u(t))+\langle p(t),\dot{x}(t)-\Gamma(x(t),u(t))\rangle\right)\,\mathrm{d}t, \tag*{(10)} S ( γ ) = ∫ 0 T ( L ( x ( t ) , u ( t )) + ⟨ p ( t ) , x ˙ ( t ) − Γ ( x ( t ) , u ( t ))⟩ ) d t , (10)
where C x 0 , x T ( M ) \mathcal{C}_{x_0,x_T}(M) C x 0 , x T ( M ) denotes the space of smooth curves γ : [ 0 , T ] ⟶ M \gamma:[0,T]\longrightarrow M γ : [ 0 , T ] ⟶ M , γ ( t ) = ( x ( t ) , p ( t ) , u ( t ) ) \gamma(t)=(x(t),p(t),u(t)) γ ( t ) = ( x ( t ) , p ( t ) , u ( t )) , such that x ( 0 ) = x 0 x(0)=x_0 x ( 0 ) = x 0 and x ( T ) = x T x(T)=x_T x ( T ) = x T .
Although the objective of an optimal control problem is to minimize the functional (2), in the present geometric framework we are interested in the extremals of S S S , assuming that a minimum exists in the class of admissible curves.
Let ξ ∈ T γ C x 0 , x T ( M ) \xi\in T_\gamma\mathcal{C}_{x_0,x_T}(M) ξ ∈ T γ C x 0 , x T ( M ) be a variation of the curve γ \gamma γ . A standard computation (see, for example, [19 ] ) shows that
d ξ S ( γ ) = ∫ 0 T ⟨ ι γ ˙ ( t ) Ω γ ( t ) − d H P ( γ ( t ) ) , ξ ( t ) ⟩ d t . \mathrm{d}_{\xi}S(\gamma)=\int_0^T\left\langle\mathord{\iota}_{\dot{\gamma}(t)}\Omega_{\gamma(t)}-\mathrm{d}H_P(\gamma(t)),\xi(t)\right\rangle\mathrm{d}t. d ξ S ( γ ) = ∫ 0 T ⟨ ι γ ˙ ( t ) Ω γ ( t ) − d H P ( γ ( t )) , ξ ( t ) ⟩ d t .
Thus, the differential of S S S is locally represented by the one-form along γ \gamma γ
ι γ ˙ Ω − d H P , \mathord{\iota}_{\dot{\gamma}}\Omega-\mathrm{d}H_P, ι γ ˙ Ω − d H P ,
which is precisely the geometric object defining the presymplectic dynamics. Consequently, a curve γ \gamma γ is an extremal of S S S if and only if it satisfies
ι γ ˙ ( t ) Ω γ ( t ) = d H P ( γ ( t ) ) , t ∈ [ 0 , T ] . (11) \mathord{\iota}_{\dot{\gamma}(t)}\Omega_{\gamma(t)}=\mathrm{d}H_P(\gamma(t)),\qquad t\in[0,T]. \tag*{(11)} ι γ ˙ ( t ) Ω γ ( t ) = d H P ( γ ( t )) , t ∈ [ 0 , T ] . (11)
Following Yoshimura and Marsden [41 ] , a variational principle of this type is usually called the Hamilton–Pontryagin principle . Here, however, we shall refer to (10) as the Lagrange–Pontryagin principle , since it is naturally derived from the Lagrangian cost function and is defined on the Pontryagin bundle; rearranging the integrand as ⟨ p , x ˙ ⟩ − H P ( x , p , u ) \langle p,\dot{x}\rangle-H_P(x,p,u) ⟨ p , x ˙ ⟩ − H P ( x , p , u ) gives the equivalent Hamilton–Pontryagin form .
Dirac structures and optimal control
The presymplectic and variational formulations developed above admit a natural interpretation in terms of Dirac geometry. We briefly recall the basic definitions needed in the sequel; for further details we refer to [17 , 18 ] .
Definition 4 (Dirac structure on a vector bundle). Let E ⟶ N E\longrightarrow N E ⟶ N be a vector bundle. A Dirac structure on E E E is a maximally isotropic subbundle
D ⊂ E ⊕ E ∗ D\subset E\oplus E^* D ⊂ E ⊕ E ∗ with respect to the symmetric bilinear form
⟨ ⟨ ( u , α ) , ( v , β ) ⟩ ⟩ = ⟨ α , v ⟩ + ⟨ β , u ⟩ , \left\langle\left\langle(u,\alpha),(v,\beta)\right\rangle\right\rangle=\langle\alpha,v\rangle+\langle\beta,u\rangle, ⟨ ⟨ ( u , α ) , ( v , β ) ⟩ ⟩ = ⟨ α , v ⟩ + ⟨ β , u ⟩ , where u , v ∈ E u,v\in E u , v ∈ E and α , β ∈ E ∗ \alpha,\beta\in E^* α , β ∈ E ∗ .
Definition 5 (Integrability). A Dirac structure D ⊂ T N ⊕ T ∗ N D\subset TN\oplus T^*N D ⊂ T N ⊕ T ∗ N on the tangent bundle of a smooth manifold N N N is integrable if its space of sections is closed under the Courant bracket [18 ]
[ ( X , α ) , ( Y , β ) ] = ( [ X , Y ] , L X β − L Y α + 1 2 d ( α ( Y ) − β ( X ) ) ) ; [(X,\alpha),(Y,\beta)]=\left([X,Y],\mathcal{L}_X\beta-\mathcal{L}_Y\alpha+\frac{1}{2}\mathrm{d}(\alpha(Y)-\beta(X))\right); [( X , α ) , ( Y , β )] = ( [ X , Y ] , L X β − L Y α + 2 1 d ( α ( Y ) − β ( X )) ) ; an integrable Dirac structure on T N TN T N is simply called a Dirac structure on the manifold $N.
Two important classes of Dirac structures are provided by presymplectic and Poisson manifolds. In particular, if Ω \Omega Ω is a closed 2-form on a smooth manifold N N N , the graph of Ω \Omega Ω defines the Dirac structure
D Ω = { ( X , α ) ∈ T N ⊕ T ∗ N ∣ α = ι X Ω } , D_{\Omega}=\{(X,\alpha)\in TN\oplus T^*N\mid\alpha=\mathord{\iota}_X\Omega\}, D Ω = {( X , α ) ∈ T N ⊕ T ∗ N ∣ α = ι X Ω } ,
and conversely the integrability of D Ω D_{\Omega} D Ω is equivalent to the closedness of Ω \Omega Ω . If D D D is a Dirac structure on N N N and H ∈ C ∞ ( N ) H\in C^\infty(N) H ∈ C ∞ ( N ) , the corresponding dynamics is defined by the vector fields X X X satisfying ( X , d H ) ∈ D (X,\mathrm{d}H)\in D ( X , d H ) ∈ D .
Definition 6 (Canonical Dirac structure associated with an optimal control problem). Let ( P , C , Γ , L ) (P,C,\Gamma,L) ( P , C , Γ , L ) be an optimal control problem and let M = T ∗ P × P C M=T^*P\times_P C M = T ∗ P × P C be its Pontryagin bundle endowed with the canonical presymplectic form Ω = pr P ∗ ω P \Omega=\operatorname{pr}_P^*\omega_P Ω = pr P ∗ ω P . The associated Dirac structure D Ω ⊂ T M ⊕ T ∗ M D_{\Omega}\subset TM\oplus T^*M D Ω ⊂ T M ⊕ T ∗ M will be called the canonical Dirac structure associated with the optimal control problem.
With this definition, the presymplectic equation (6) reads ( X P , d H P ) ∈ D Ω (X_P,dH_P)\in D_{\Omega} ( X P , d H P ) ∈ D Ω : the dynamics of the Pontryagin Maximum Principle is the Dirac system on the Pontryagin bundle determined by the canonical Dirac structure and the Pontryagin Hamiltonian, and its integral curves are precisely the extremals of the optimal control problem.
Dirac structures for regular optimal control problems
For regular optimal control problems the previous Dirac formulation admits a more intrinsic description, which will play a fundamental role in the reduction process developed later. Let D P ⊂ T ( T ∗ P ) ⊕ T ∗ ( T ∗ P ) D_P\subset T(T^*P)\oplus T^*(T^*P) D P ⊂ T ( T ∗ P ) ⊕ T ∗ ( T ∗ P ) denote the canonical Dirac structure determined by the canonical symplectic form ω P \omega_P ω P on T ∗ P T^*P T ∗ P . Since the problem is regular, p 1 : M 1 ⟶ T ∗ P p_1:M_1\longrightarrow T^*P p 1 : M 1 ⟶ T ∗ P is a local symplectomorphism (Proposition 1) and the presymplectic equation admits a unique solution at every point of M 1 M_1 M 1 , so that the dynamics of Section 2.2 is a vector field X : M 1 ⟶ T ( T ∗ P ) X:M_1\longrightarrow T(T^*P) X : M 1 ⟶ T ( T ∗ P ) along p 1 p_1 p 1 . Likewise, since p 1 ∗ : T ∗ ( T ∗ P ) ⟶ T ∗ M 1 p_1^*:T^*(T^*P)\longrightarrow T^*M_1 p 1 ∗ : T ∗ ( T ∗ P ) ⟶ T ∗ M 1 is an isomorphism of vector bundles, the differential of H 1 = i 1 ∗ H P H_1=i^*_{1}H_P H 1 = i 1 ∗ H P may be regarded as a covector on T ∗ P T^*P T ∗ P along p 1 p_1 p 1 ,
d H 1 = ( p 1 ∗ ) − 1 ( i 1 ∗ d H P ) : M 1 ⟶ T ∗ ( T ∗ P ) , dH_1=(p_1^*)^{-1}(i_1^*dH_P):M_1\longrightarrow T^*(T^*P), d H 1 = ( p 1 ∗ ) − 1 ( i 1 ∗ d H P ) : M 1 ⟶ T ∗ ( T ∗ P ) ,
and we use the symbol d H 1 dH_1 d H 1 for both; the constraint manifold M 1 M_1 M 1 thus plays the role of the graph of the Legendre transformation in the implicit Lagrangian formalism of Yoshimura and Marsden [41 ] . The implicit Dirac formulation of the regular optimal control problem is
( X , d H 1 ) ∈ D P on M 1 , (X,dH_1)\in D_P\qquad\text{on }M_1, ( X , d H 1 ) ∈ D P on M 1 ,
both entries being taken along p 1 p_1 p 1 . This formulation is completely equivalent to the presymplectic description obtained after imposing the primary constraints (7).
Symmetries of optimal control problems
We now introduce the notion of symmetry for an optimal control problem. Let ( P , C , Γ , L ) (P,C,\Gamma,L) ( P , C , Γ , L ) be an optimal control problem, and let π : C ⟶ P \pi:C\longrightarrow P π : C ⟶ P be the corresponding control bundle.
Suppose that a Lie group G G G acts on the right on the bundle C C C by fiber-preserving transformations. We denote this action by Ψ : C × G ⟶ C \Psi:C\times G\longrightarrow C Ψ : C × G ⟶ C , ( c , g ) ↦ c ⋅ g (c,g)\mapsto c\cdot g ( c , g ) ↦ c ⋅ g , and the induced action on the state manifold by Φ : P × G ⟶ P \Phi:P\times G\longrightarrow P Φ : P × G ⟶ P , ( x , g ) ↦ x ⋅ g (x,g)\mapsto x\cdot g ( x , g ) ↦ x ⋅ g ; writing Ψ g ( c ) = Ψ ( c , g ) \Psi_g(c)=\Psi(c,g) Ψ g ( c ) = Ψ ( c , g ) and Φ g ( x ) = Φ ( x , g ) \Phi_g(x)=\Phi(x,g) Φ g ( x ) = Φ ( x , g ) , the fact that the action on C C C is fibered means that π ∘ Ψ g = Φ g ∘ π \pi\circ\Psi_g=\Phi_g\circ\pi π ∘ Ψ g = Φ g ∘ π for all g ∈ G g\in G g ∈ G .
Definition 7. An optimal control problem ( P , C , Γ , L ) (P,C,\Gamma,L) ( P , C , Γ , L ) is said to be G G G -invariant, or equivalently to admit G G G as a symmetry group, if the fibered action ( Φ , Ψ ) (\Phi,\Psi) ( Φ , Ψ ) satisfies
(i) the running cost is invariant,
L ∘ Ψ g = L , ∀ g ∈ G ; L\circ\Psi_g=L,\qquad\forall\ g\in G; L ∘ Ψ g = L , ∀ g ∈ G ; (ii) the controlled dynamics is equivariant,
T Φ g ∘ Γ = Γ ∘ Ψ g , ∀ g ∈ G . T\Phi_g\circ\Gamma=\Gamma\circ\Psi_g,\qquad\forall\ g\in G. T Φ g ∘ Γ = Γ ∘ Ψ g , ∀ g ∈ G .
The action of G G G on P P P induces, by tangent and cotangent lift, right actions on T P TP T P and T ∗ P T^*P T ∗ P , namely v ⋅ g = T Φ g ( v ) v\cdot g=T\Phi_g(v) v ⋅ g = T Φ g ( v ) for v ∈ T x P v\in T_xP v ∈ T x P and p ⋅ g = ( T Φ g − 1 ) ∗ ( p ) p\cdot g=(T\Phi_{g^{-1}})^*(p) p ⋅ g = ( T Φ g − 1 ) ∗ ( p ) for p ∈ T x ∗ P p\in T_x^*P p ∈ T x ∗ P . Consequently, G G G acts naturally on the Pontryagin bundle M = T ∗ P × P C M=T^*P\times_P C M = T ∗ P × P C according to
( x , p , u ) ⋅ g = ( x ⋅ g , p ⋅ g , u ⋅ g ) = ( Φ g ( x ) , ( T Φ g − 1 ) ∗ p , Ψ g ( u ) ) . (x,p,u)\cdot g=(x\cdot g,p\cdot g,u\cdot g)=(\Phi_g(x),(T\Phi_{g^{-1}})^*p,\Psi_g(u)). ( x , p , u ) ⋅ g = ( x ⋅ g , p ⋅ g , u ⋅ g ) = ( Φ g ( x ) , ( T Φ g − 1 ) ∗ p , Ψ g ( u )) .
The presymplectic formulation of the optimal control problem is naturally compatible with the action of the symmetry group.
Proposition 2. Let ( P , C , Γ , L ) (P,C,\Gamma,L) ( P , C , Γ , L ) be a G G G -invariant optimal control problem. Then the induced action of G G G on the Pontryagin bundle M = T ∗ P × P C M=T^*P\times_P C M = T ∗ P × P C preserves both the canonical presymplectic form Ω = pr P ∗ ω P \Omega=\operatorname{pr}_P^*\omega_P Ω = pr P ∗ ω P and the Pontryagin Hamiltonian H P H_P H P ; in particular,
H P ( ( x , p , u ) ⋅ g ) = H P ( x , p , u ) , ∀ g ∈ G . (12) H_P((x,p,u)\cdot g)=H_P(x,p,u),\qquad\forall\ g\in G. \tag*{(12)} H P (( x , p , u ) ⋅ g ) = H P ( x , p , u ) , ∀ g ∈ G . (12) Moreover, the action admits the equivariant momentum map J M = pr P ∗ J P : M ⟶ g ∗ J_M=\operatorname{pr}_P^*J_P:M\longrightarrow\mathfrak{g}^* J M = pr P ∗ J P : M ⟶ g ∗ , ⟨ J M ( x , p , u ) , ξ ⟩ = ⟨ p , ξ P ( x ) ⟩ \langle J_M(x,p,u),\xi\rangle=\langle p,\xi_P(x)\rangle ⟨ J M ( x , p , u ) , ξ ⟩ = ⟨ p , ξ P ( x )⟩ , where J P : T ∗ P ⟶ g ∗ J_P:T^*P\longrightarrow\mathfrak{g}^* J P : T ∗ P ⟶ g ∗ is the canonical momentum map of the cotangent-lifted action and ξ P \xi_P ξ P the infinitesimal generator of the action on M M M . It satisfies ι ξ M Ω = d ⟨ J M , ξ ⟩ \iota_{\xi_M}\Omega=d\langle J_M,\xi\rangle ι ξ M Ω = d ⟨ J M , ξ ⟩ for all ξ ∈ g \xi\in\mathfrak{g} ξ ∈ g , where ξ M \xi_M ξ M denotes the infinitesimal generator of the action on M M M .
Proof. Cotangent lifts are symplectic, so Ω = pr P ∗ ω P \Omega=\operatorname{pr}_P^*\omega_P Ω = pr P ∗ ω P is G G G -invariant, and J M J_M J M is the pull-back of the canonical momentum map, with which it shares its equivariance. The invariance of H P H_P H P follows from the equivariance of Γ \Gamma Γ and the invariance of L L L : H P ( ( x , p , u ) ⋅ g ) = ⟨ ( T Φ g − 1 ) ∗ p , T Φ g Γ ( x , u ) ⟩ − L ( x , u ) = H P ( x , p , u ) . H_P((x,p,u)\cdot g)=\langle(T\Phi_{g^{-1}})^*p,T\Phi_g\Gamma(x,u)\rangle-L(x,u)=H_P(x,p,u). H P (( x , p , u ) ⋅ g ) = ⟨( T Φ g − 1 ) ∗ p , T Φ g Γ ( x , u )⟩ − L ( x , u ) = H P ( x , p , u ) .
Throughout the paper we shall assume that the quotient space P / G P/G P / G is a smooth manifold and that the canonical projection ρ : P ⟶ P / G \rho:P\longrightarrow P/G ρ : P ⟶ P / G is a surjective submersion. Whenever explicit local expressions are required, we shall further assume that the action of G G G on P P P is free and proper, so that ρ \rho ρ is a principal G G G -bundle.
For the same reason, we assume that the control bundle is obtained from an associated bundle. More precisely, let U U U be a smooth manifold endowed with a left action Σ : G × U ⟶ U \Sigma:G\times U\longrightarrow U Σ : G × U ⟶ U . The associated bundle over the quotient is U ~ = P × G U = ( P × U ) / G \widetilde{U}=P\times_G U=(P\times U)/G U = P × G U = ( P × U ) / G , and the control bundle is assumed to be its pull-back to P P P , C ≃ ρ ∗ U ~ C\simeq\rho^*\widetilde{U} C ≃ ρ ∗ U . Whenever no confusion is possible, we shall use the same notation U ~ \widetilde{U} U for its pull-back to P P P or to T ∗ P T^*P T ∗ P . Notice that, as bundles over P P P , ρ ∗ U ~ ≃ P × U \rho^*\widetilde{U}\simeq P\times U ρ ∗ U ≃ P × U , although the notation U ~ \widetilde{U} U is convenient for emphasizing its origin as the pull-back of the associated bundle over P / G P/G P / G .
Reduction of the Lagrange–Pontryagin principle
Let ( P , C , Γ , L ) (P,C,\Gamma,L) ( P , C , Γ , L ) be a G G G -invariant optimal control problem. Since the running cost L L L is G G G -invariant and the controlled dynamics Γ \Gamma Γ is G G G -equivariant, the Lagrange–Pontryagin functional S : C ( M ) ⟶ R S:C(M)\longrightarrow\mathbb{R} S : C ( M ) ⟶ R defined in (10) is invariant under the natural action of G G G on the space of smooth curves in the Pontryagin bundle, ( γ ⋅ g ) ( t ) = ( x ( t ) ⋅ g , p ( t ) ⋅ g , u ( t ) ⋅ g ) (\gamma\cdot g)(t)=(x(t)\cdot g,p(t)\cdot g,u(t)\cdot g) ( γ ⋅ g ) ( t ) = ( x ( t ) ⋅ g , p ( t ) ⋅ g , u ( t ) ⋅ g ) , that is, S ( γ ⋅ g ) = S ( γ ) S(\gamma\cdot g)=S(\gamma) S ( γ ⋅ g ) = S ( γ ) for all γ ∈ C ( M ) \gamma\in C(M) γ ∈ C ( M ) and g ∈ G g\in G g ∈ G .
Two endpoint conditions must be distinguished here, because the space of curves with genuinely fixed endpoints x ( 0 ) = x 0 x(0)=x_0 x ( 0 ) = x 0 , x ( T ) = x T x(T)=x_T x ( T ) = x T is not preserved by the G G G -action: for a free action a nontrivial group element fixes no endpoint. The only formulation on which G G G acts is the orbit-endpoint problem, in which x ( 0 ) x(0) x ( 0 ) and x ( T ) x(T) x ( T ) are required only to lie in the orbits G ⋅ x 0 G\cdot x_0 G ⋅ x 0 and G ⋅ x T G\cdot x_T G ⋅ x T ; there S S S descends to the quotient of the corresponding curve space. Its endpoint variations are vertical, so that δ z ( 0 ) = δ z ( T ) = 0 \delta z(0)=\delta z(T)=0 δ z ( 0 ) = δ z ( T ) = 0 automatically, while the reduced vertical variations η ( 0 ) \eta(0) η ( 0 ) and η ( T ) \eta(T) η ( T ) of Section 3.2 remain free and produce the transversality conditions μ ( 0 ) = μ ( T ) = 0 \mu(0)=\mu(T)=0 μ ( 0 ) = μ ( T ) = 0 .
The fixed-endpoint problem is treated by reduction as follows. Its admissible curves are exactly the curves obtained by reconstructing a reduced curve from the fixed initial lift x ( 0 ) = x 0 x(0)=x_0 x ( 0 ) = x 0 and imposing on it the terminal holonomy condition x ( T ) = x T x(T)=x_T x ( T ) = x T , and its variations are those with δ z ( 0 ) = δ z ( T ) = 0 \delta z(0)=\delta z(T)=0 δ z ( 0 ) = δ z ( T ) = 0 and η ( 0 ) = η ( T ) = 0 \eta(0)=\eta(T)=0 η ( 0 ) = η ( T ) = 0 . The two problems therefore have the same interior equations, namely the reduced Lagrange–Pontryagin equations (26) derived below, and differ only in their boundary conditions: no condition on μ \mu μ at the endpoints in the fixed case, and μ ( 0 ) = μ ( T ) = 0 \mu(0)=\mu(T)=0 μ ( 0 ) = μ ( T ) = 0 in the orbit case. Except where stated otherwise we work with the fixed-endpoint problem, and all the endpoint conditions written below, in Section 3.2 and in Proposition 4, are its. Strictly speaking, then, it is the interior variational problem, and not the fixed-endpoint problem itself, that is G G G -invariant; this is what is used in what follows.
On the unconstrained curve space C ( M ) C(M) C ( M ) , and likewise on the orbit-endpoint curve space, the functional S S S therefore descends to the quotient C ( M ) / G C(M)/G C ( M ) / G . We denote the induced functional by [ S ] G [S]_G [ S ] G and refer to it as the reduced Lagrange–Pontryagin functional . Since the quotient map C ( M ) ⟶ C ( M ) / G C(M)\longrightarrow C(M)/G C ( M ) ⟶ C ( M ) / G identifies curves belonging to the same G G G -orbit, the critical points of [ S ] G [S]_G [ S ] G are in one-to-one correspondence with G G G -equivalence classes of extremals of S S S . For the fixed-endpoint problem the corresponding statement is the one supplied by Theorem 1: reduced solutions correspond to extremals through a prescribed initial point x 0 x_0 x 0 , and the terminal condition x ( T ) = x T x(T)=x_T x ( T ) = x T is imposed afterwards.
Our goal is to characterize these critical points in terms of the geometry of the reduced Pontryagin bundle. The corresponding Euler–Lagrange equations on the quotient space will be called the reduced Lagrange–Pontryagin equations . Since they represent the reduction of the first-order necessary conditions of the Pontryagin Maximum Principle, we shall also refer to them simply as the reduced Pontryagin Maximum Principle .
Canonical reduced bundles
We begin by introducing several canonical bundles naturally associated with the symmetry reduction, which will be used throughout the remainder of this section.
Unless otherwise stated, we assume that the right action of the Lie group G G G on P P P is free and proper. Hence the quotient B = P / G B=P/G B = P / G is a smooth manifold and the canonical projection ρ : P ⟶ B \rho:P\longrightarrow B ρ : P ⟶ B defines a principal G G G -bundle. The orbit through a point x ∈ P x\in P x ∈ P will be denoted either by x ⋅ G x\cdot G x ⋅ G or by [ x ] = { x g ∣ g ∈ G } [x]=\{xg\mid g\in G\} [ x ] = { xg ∣ g ∈ G } , and we shall write z = [ x ] ∈ B z=[x]\in B z = [ x ] ∈ B . Local coordinates on B = P / G B=P/G B = P / G will be denoted by ( z α ) (z^\alpha) ( z α ) , α = 1 , … , dim B \alpha=1,\ldots,\dim B α = 1 , … , dim B .
The adjoint and coadjoint bundles
Associated with the principal G G G -bundle ρ : P ⟶ B \rho:P\longrightarrow B ρ : P ⟶ B are the adjoint and coadjoint vector bundles, corresponding to the adjoint and coadjoint representations of G G G on its Lie algebra g \mathfrak{g} g and its dual g ∗ \mathfrak{g}^* g ∗ , respectively.
The adjoint bundle is the associated vector bundle Ad P = P × G g = ( P × g ) / G \operatorname{Ad}P=P\times_G\mathfrak{g}=(P\times\mathfrak{g})/G Ad P = P × G g = ( P × g ) / G , where the quotient is taken with respect to the left action
g ⋅ ( x , ξ ) = ( x g − 1 , Ad g ξ ) . (13) g\cdot(x,\xi)=(xg^{-1},\operatorname{Ad}_g\xi). \tag*{(13)} g ⋅ ( x , ξ ) = ( x g − 1 , Ad g ξ ) . (13)
Following the notation introduced in § 2.6 , we shall denote this bundle by g ~ \widetilde{\mathfrak{g}} g . An element of g ~ \widetilde{\mathfrak{g}} g will be written simply as ( z , ξ ) (z,\xi) ( z , ξ ) , where z = [ x ] ∈ B z=[x]\in B z = [ x ] ∈ B .
Similarly, the coadjoint bundle is the associated vector bundle Ad ∗ P = P × G g ∗ \operatorname{Ad}^*P=P\times_G\mathfrak{g}^* Ad ∗ P = P × G g ∗ , where the quotient is defined by the left action
g ⋅ ( x , μ ) = ( x g − 1 , Ad g − 1 ∗ μ ) . (14) g\cdot(x,\mu)=(xg^{-1},\operatorname{Ad}^*_{g^{-1}}\mu). \tag*{(14)} g ⋅ ( x , μ ) = ( x g − 1 , Ad g − 1 ∗ μ ) . (14)
Since the coadjoint representation is the dual of the adjoint representation, Ad ∗ P \operatorname{Ad}^*P Ad ∗ P is naturally the dual vector bundle of Ad P \operatorname{Ad}P Ad P . Accordingly, we shall denote it by g ~ ∗ \widetilde{\mathfrak{g}}^* g ∗ ; equivalently,
g ~ ∗ ≅ g ∗ ~ . \widetilde{\mathfrak{g}}^*\cong\widetilde{\mathfrak{g}^*}. g ∗ ≅ g ∗ .
The principal bundles T P TP T P and T ∗ P T^*P T ∗ P
Since the action of G G G on P P P is free and proper, its tangent lift ( x , v x ) ⋅ g = ( x g , T Φ g ( v x ) ) (x,v_x)\cdot g=(xg,T\Phi_g(v_x)) ( x , v x ) ⋅ g = ( xg , T Φ g ( v x )) and its cotangent lift ( x , p x ) ⋅ g = ( x g , ( T Φ g − 1 ) ∗ p x ) (x,p_x)\cdot g=(xg,(T\Phi_{g^{-1}})^*p_x) ( x , p x ) ⋅ g = ( xg , ( T Φ g − 1 ) ∗ p x ) are free and proper right actions, so that T P → T P / G TP\to TP/G T P → T P / G and T ∗ P → T ∗ P / G T^*P\to T^*P/G T ∗ P → T ∗ P / G are again principal G G G -bundles, and the tangent map T ρ : T P → T ( P / G ) T\rho:TP\to T(P/G) T ρ : T P → T ( P / G ) covers ρ \rho ρ . The tangent group T G TG T G is identified, by left trivialization ( g , v g ) ↦ ( g , T L g − 1 v g ) (g,v_g)\mapsto(g,T L_{g^{-1}}v_g) ( g , v g ) ↦ ( g , T L g − 1 v g ) , with the semidirect product G ⋉ g G\ltimes\mathfrak{g} G ⋉ g , whose multiplication is ( g , ξ ) ( h , ζ ) = ( g h , Ad h − 1 ξ + ζ ) (g,\xi)(h,\zeta)=(gh,\operatorname{Ad}_{h^{-1}}\xi+\zeta) ( g , ξ ) ( h , ζ ) = ( g h , Ad h − 1 ξ + ζ ) , and the induced right action of T G TG T G on T P TP T P reads
( x , v x ) ⋅ ( g , ξ ) = ( x g , T Φ g ( v x ) + ξ P ( x g ) ) , (15) (x,v_x)\cdot(g,\xi)=(xg,T\Phi_g(v_x)+\xi_P(xg)), \tag*{(15)} ( x , v x ) ⋅ ( g , ξ ) = ( xg , T Φ g ( v x ) + ξ P ( xg )) , (15)
where ξ P \xi_P ξ P denotes the fundamental vector field on P P P generated by ξ ∈ g \xi\in\mathfrak{g} ξ ∈ g : for any curve with g ( 0 ) = g g(0)=g g ( 0 ) = g and T L g − 1 g ˙ ( 0 ) = ξ T L_{g^{-1}}\dot{g}(0)=\xi T L g − 1 g ˙ ( 0 ) = ξ one has d d t ∣ t = 0 x g ( t ) = d d t ∣ t = 0 ( x g ) exp ( t ξ ) = ξ P ( x g ) \left.\frac{d}{dt}\right|_{t=0}xg(t)=\left.\frac{d}{dt}\right|_{t=0}(xg)\exp(t\xi)=\xi_P(xg) d t d t = 0 xg ( t ) = d t d t = 0 ( xg ) exp ( t ξ ) = ξ P ( xg ) . With this convention (15) is a right action of the group above, since T Φ h ( ξ P ( x g ) ) = ( Ad h − 1 ξ ) P ( x g h ) T\Phi_h(\xi_P(xg))=(\operatorname{Ad}_{h^{-1}}\xi)_P(xgh) T Φ h ( ξ P ( xg )) = ( Ad h − 1 ξ ) P ( xg h ) .
The reduced tangent and cotangent bundles
The quotient T P / G TP/G T P / G is the middle term of the Atiyah exact sequence of vector bundles over B = P / G B=P/G B = P / G ,
0 ⟶ g ~ ⟶ T P / G → T ρ T B ⟶ 0 , (16) 0\longrightarrow\widetilde{\mathfrak{g}}\longrightarrow TP/G\xrightarrow{T\rho}TB\longrightarrow0, \tag*{(16)} 0 ⟶ g ⟶ T P / G T ρ T B ⟶ 0 , (16)
and is therefore canonically an affine bundle over T ( P / G ) T(P/G) T ( P / G ) modeled on the pullback τ P / G ∗ ( g ~ ) \tau^*_{P/G}(\widetilde{\mathfrak{g}}) τ P / G ∗ ( g ) of the adjoint bundle, where τ P / G : T ( P / G ) → P / G \tau_{P/G}:T(P/G)\to P/G τ P / G : T ( P / G ) → P / G is the tangent bundle projection: two classes [ x , v x ] [x,v_x] [ x , v x ] and [ x , w x ] [x,w_x] [ x , w x ] lying over the same v z ∈ T z B v_z\in T_zB v z ∈ T z B differ by the vertical class [ x , v x − w x ] ∈ g ~ z [x,v_x-w_x]\in\widetilde{\mathfrak{g}}_z [ x , v x − w x ] ∈ g z , but no class over v z v_z v z is distinguished. Its identification with T ( P / G ) ⊕ g ~ T(P/G)\oplus\widetilde{\mathfrak{g}} T ( P / G ) ⊕ g requires a splitting of (16), that is, a principal connection on ρ : P → P / G \rho:P\to P/G ρ : P → P / G , and is therefore not canonical. As before, we shall use the same notation g ~ \widetilde{\mathfrak{g}} g for the adjoint bundle and its pullback whenever no confusion is likely to arise.
Let A A A be a principal connection on P P P , and consider the map
α A : T P ⟶ T ( P / G ) × ( P × g ) , α A ( x , v x ) = ( T ρ ( v x ) , x , A x ( v x ) ) . \alpha_A:TP\longrightarrow T(P/G)\times(P\times\mathfrak{g}),\qquad\alpha_A(x,v_x)=(T\rho(v_x),x,A_x(v_x)). α A : T P ⟶ T ( P / G ) × ( P × g ) , α A ( x , v x ) = ( T ρ ( v x ) , x , A x ( v x )) .
The group G G G acts trivially on the factor T ( P / G ) T(P/G) T ( P / G ) and on P × g P\times\mathfrak{g} P × g by the adjoint action (13). Since α A \alpha_A α A is G G G -equivariant, it descends to the quotient, giving rise to the vector bundle isomorphism
[ α A ] : T P / G ⟶ T ( P / G ) ⊕ g ~ , [ α A ] ( [ x , v x ] ) = ( T ρ ( v x ) , [ x , A x ( v x ) ] ) , (17) [\alpha_A]:TP/G\longrightarrow T(P/G)\oplus\widetilde{\mathfrak{g}},\qquad[\alpha_A]([x,v_x])=(T\rho(v_x),[x,A_x(v_x)]), \tag*{(17)} [ α A ] : T P / G ⟶ T ( P / G ) ⊕ g , [ α A ] ([ x , v x ]) = ( T ρ ( v x ) , [ x , A x ( v x )]) , (17)
under which the class [ x , v x ] [x,v_x] [ x , v x ] is represented by the triple ( z , v z , ζ ) (z,v_z,\zeta) ( z , v z , ζ ) , where z = [ x ] ∈ P / G z=[x]\in P/G z = [ x ] ∈ P / G , v z = T ρ ( v x ) ∈ T z ( P / G ) v_z=T\rho(v_x)\in T_z(P/G) v z = T ρ ( v x ) ∈ T z ( P / G ) and ζ = [ x , A x ( v x ) ] ∈ g ~ z \zeta=[x,A_x(v_x)]\in\widetilde{\mathfrak{g}}_z ζ = [ x , A x ( v x )] ∈ g z .
By duality, the connection identifies the quotient T ∗ P / G T^*P/G T ∗ P / G , the Weinstein space [36 ] , with T ∗ ( P / G ) ⊕ g ~ ∗ T^*(P/G)\oplus\widetilde{\mathfrak{g}}^* T ∗ ( P / G ) ⊕ g ∗ , which is a vector bundle over T ∗ ( P / G ) T^*(P/G) T ∗ ( P / G ) , namely the pullback of the coadjoint bundle g ~ ∗ \widetilde{\mathfrak{g}}^* g ∗ to the cotangent bundle projection π P / G : T ∗ ( P / G ) → P / G \pi_{P/G}:T^*(P/G)\to P/G π P / G : T ∗ ( P / G ) → P / G ; as before, we shall use the notation T ∗ ( P / G ) ⊕ g ~ ∗ T^*(P/G)\oplus\widetilde{\mathfrak{g}}^* T ∗ ( P / G ) ⊕ g ∗ for this bundle whenever no confusion may arise. Explicitly, let H x = ker A x ⊂ T x P H_x=\ker A_x\subset T_xP H x = ker A x ⊂ T x P be the horizontal subspace determined by the connection. Since the restriction of T ρ x : H x → T [ x ] ( P / G ) T\rho_x:H_x\to T_{[x]}(P/G) T ρ x : H x → T [ x ] ( P / G ) is an isomorphism, it induces an identification H x ∗ ≅ T [ x ] ∗ ( P / G ) H_x^*\cong T^*_{[x]}(P/G) H x ∗ ≅ T [ x ] ∗ ( P / G ) , and the decomposition T x P = H x ⊕ V x T_xP=H_x\oplus V_x T x P = H x ⊕ V x , where V x V_x V x denotes the vertical subspace, defines a projection π ~ G : T ∗ P → T ∗ ( P / G ) \widetilde{\pi}_G:T^*P\to T^*(P/G) π G : T ∗ P → T ∗ ( P / G ) , π ~ G ( x , p x ) = ( z , p z ) \widetilde{\pi}_G(x,p_x)=(z,p_z) π G ( x , p x ) = ( z , p z ) , obtained by restricting covectors to H x H_x H x and identifying them with covectors on T [ x ] ( P / G ) T_{[x]}(P/G) T [ x ] ( P / G ) . The momentum map J : T ∗ P → g ∗ J:T^*P\to\mathfrak{g}^* J : T ∗ P → g ∗ associated with the cotangent lift of the G G G -action on P P P then provides the second component of the identification. More precisely, consider the map
β J : T ∗ P ⟶ T ∗ ( P / G ) × ( P × g ∗ ) , β J ( x , p x ) = ( z , p z ; x , J ( x , p x ) ) . \beta_J:T^*P\longrightarrow T^*(P/G)\times(P\times\mathfrak{g}^*),\qquad\beta_J(x,p_x)=(z,p_z;x,J(x,p_x)). β J : T ∗ P ⟶ T ∗ ( P / G ) × ( P × g ∗ ) , β J ( x , p x ) = ( z , p z ; x , J ( x , p x )) .
Since β J \beta_J β J is G G G -equivariant, where G G G acts on P × g ∗ P\times\mathfrak{g}^* P × g ∗ by the coadjoint action (14), it descends to the quotient, yielding the vector bundle isomorphism
[ β J ] : T ∗ P / G ⟶ T ∗ ( P / G ) ⊕ g ~ ∗ , [ β J ] ( [ x , p x ] ) = ( z , p z ; [ x , J ( x , p x ) ] ) . (18) [\beta_J]:T^*P/G\longrightarrow T^*(P/G)\oplus\widetilde{\mathfrak{g}}^*,\qquad[\beta_J]([x,p_x])=(z,p_z;[x,J(x,p_x)]). \tag*{(18)} [ β J ] : T ∗ P / G ⟶ T ∗ ( P / G ) ⊕ g ∗ , [ β J ] ([ x , p x ]) = ( z , p z ; [ x , J ( x , p x )]) . (18)
Accordingly, an element of T ∗ P / G T^*P/G T ∗ P / G will be represented by a triple ( z , p z , μ ) (z,p_z,\mu) ( z , p z , μ ) , where ( z , p z ) ∈ T ∗ ( P / G ) (z,p_z) \in T^*(P/G) ( z , p z ) ∈ T ∗ ( P / G ) and μ ∈ g ~ z ∗ \mu\in\widetilde{\mathfrak{g}}_z^* μ ∈ g z ∗ .
The reduced control bundle
If G G G is a symmetry group of the optimal control problem ( P , C , Γ , L ) (P,C,\Gamma,L) ( P , C , Γ , L ) , the quotient C ~ = C / G \widetilde{C}=C/G C = C / G is a bundle over the orbit space P / G P/G P / G with projection π ~ ( [ x , u ] ) = [ x ] \widetilde{\pi}([x,u])=[x] π ([ x , u ]) = [ x ] . As explained in § 2.6 , we assume that the control bundle π : C → P \pi:C\to P π : C → P is the pullback of an associated bundle U ~ = P × G U → P / G \widetilde{U}=P\times_G U\to P/G U = P × G U → P / G , where G G G acts on the typical fiber U U U , so that C = ρ ∗ ( U ~ ) ≅ P × U C=\rho^*(\widetilde{U})\cong P\times U C = ρ ∗ ( U ) ≅ P × U (a trivialization which is not G G G -equivariant), and passing to the quotient yields the associated bundle
C / G ≅ U ~ , [ x , u ] ⟼ [ x , u ] , (19) C/G\cong\widetilde{U},\qquad[x,u]\longmapsto[x,u], \tag*{(19)} C / G ≅ U , [ x , u ] ⟼ [ x , u ] , (19)
which will be regarded as the reduced control bundle over P / G P/G P / G . The right-hand side is the class of ( x , u ) (x,u) ( x , u ) in P × G U P\times_G U P × G U and not the pair ( [ x ] , u ) ([x],u) ([ x ] , u ) , which for a nontrivial action on U U U is defined only once a representative x x x of z = [ x ] z=[x] z = [ x ] has been chosen. We write u ~ ∈ U ~ z \widetilde{u}\in\widetilde{U}_z u ∈ U z for the reduced control, and revert to the notation u u u only in a local trivialization of U ~ \widetilde{U} U or when the action on U U U is trivial.
Unless the contrary is stated, U U U is only a smooth G G G -manifold, so that U ~ \widetilde{U} U is a fiber bundle over B = P / G B=P/G B = P / G and not a vector bundle. From § 4 onwards, and in particular in § 8.8 and § 8.9 , we assume in addition that U U U is a vector space carrying a linear G G G -action, in which case U ~ \widetilde{U} U is an associated vector bundle.
The reduced Pontryagin bundle
Since the action of G G G on the Pontryagin bundle M = T ∗ P × P C M=T^*P\times_P C M = T ∗ P × P C covers the action on P P P , the quotient M / G M/G M / G is naturally a bundle over P / G P/G P / G , and combining the identifications (18) and (19) we obtain the isomorphism of fiber bundles over B = P / G B=P/G B = P / G
M / G ≅ ( T ∗ ( P / G ) ⊕ g ~ ∗ ) × B U ~ . (20) M/G\cong\left(T^*(P/G)\oplus\widetilde{\mathfrak{g}}^*\right)\times_B\widetilde{U}. \tag*{(20)} M / G ≅ ( T ∗ ( P / G ) ⊕ g ∗ ) × B U . (20)
The fiber product × B \times_B × B reduces to the Whitney sum T ∗ ( P / G ) ⊕ g ~ ∗ ⊕ U ~ T^*(P/G)\oplus\widetilde{\mathfrak{g}}^*\oplus\widetilde{U} T ∗ ( P / G ) ⊕ g ∗ ⊕ U when U U U is a vector space with a linear G G G -action, and we write the right-hand side of (20) additively throughout, since that hypothesis is in force from § 4 onwards; in § 3.4 , where it is not yet needed, the additive notation is to be read as the fiber product, no operation being performed on the third factor. An element of the reduced Pontryagin bundle will be represented by a quadruple ( z , p z , μ , u ~ ) (z,p_z,\mu,\widetilde{u}) ( z , p z , μ , u ) , where ( z , p z ) ∈ T ∗ ( P / G ) (z,p_z)\in T^*(P/G) ( z , p z ) ∈ T ∗ ( P / G ) , μ ∈ g ~ z ∗ \mu\in\widetilde{\mathfrak{g}}_z^* μ ∈ g z ∗ , and u ~ ∈ U ~ z \widetilde{u}\in\widetilde{U}_z u ∈ U z .
The reduced Lagrange–Pontryagin principle
We are now in a position to derive the reduced Lagrange–Pontryagin variational principle.
Since the running cost L L L is G G G -invariant, it descends to a smooth function l : U ~ = C / G → R l:\widetilde{U}=C/G\to\mathbb{R} l : U = C / G → R , given by l ( z , u ) = L ( x , u ) l(z,u)=L(x,u) l ( z , u ) = L ( x , u ) , where z = [ x ] ∈ P / G z=[x]\in P/G z = [ x ] ∈ P / G . This definition is independent of the chosen representative because of the G G G -invariance of L L L .
Similarly, the equivariance of the control vector field Γ : C → T P \Gamma:C\to TP Γ : C → T P induces a bundle map [ Γ ] : U ~ → T P / G [\Gamma]:\widetilde{U}\to TP/G [ Γ ] : U → T P / G . After choosing a principal connection and using the identification T P / G ≅ T ( P / G ) ⊕ g ~ TP/G\cong T(P/G)\oplus\widetilde{\mathfrak{g}} T P / G ≅ T ( P / G ) ⊕ g , the reduced dynamics decomposes as
[ Γ ] ( z , u ) = ( Γ ~ ( z , u ) , γ ~ ( z , u ) ) , [\Gamma](z,u)=\bigl(\widetilde{\Gamma}(z,u),\widetilde{\gamma}(z,u)\bigr), [ Γ ] ( z , u ) = ( Γ ( z , u ) , γ ( z , u ) ) ,
where Γ ~ : U ~ → T ( P / G ) \widetilde{\Gamma}:\widetilde{U}\to T(P/G) Γ : U → T ( P / G ) is a vector field along the bundle projection and γ ~ : U ~ → g ~ \widetilde{\gamma}:\widetilde{U}\to\widetilde{\mathfrak{g}} γ : U → g is a section of the adjoint bundle. In terms of the connection,
γ ~ ( z , u ) = [ x , A ( Γ ( x , u ) ) ] , \widetilde{\gamma}(z,u)=[x,A(\Gamma(x,u))], γ ( z , u ) = [ x , A ( Γ ( x , u ))] ,
which is well defined by the equivariance of the connection form.
Using the decompositions T P / G ≅ T ( P / G ) ⊕ g ~ TP/G\cong T(P/G)\oplus\widetilde{\mathfrak{g}} T P / G ≅ T ( P / G ) ⊕ g and T ∗ P / G ≅ T ∗ ( P / G ) ⊕ g ~ ∗ T^*P/G\cong T^*(P/G)\oplus\widetilde{\mathfrak{g}}^* T ∗ P / G ≅ T ∗ ( P / G ) ⊕ g ∗ , the density of the Lagrange–Pontryagin functional becomes
L = l ( z , u ) + ⟨ p z , z ˙ − Γ ~ ( z , u ) ⟩ + ⟨ μ , ξ − γ ~ ( z , u ) ⟩ , (21) \mathcal{L}=l(z,u)+\langle p_z,\dot{z}-\widetilde{\Gamma}(z,u)\rangle+\langle\mu,\xi-\widetilde{\gamma}(z,u)\rangle, \tag*{(21)} L = l ( z , u ) + ⟨ p z , z ˙ − Γ ( z , u )⟩ + ⟨ μ , ξ − γ ( z , u )⟩ , (21)
where z ˙ = T ρ ( x ˙ ) \dot{z}=T\rho(\dot{x}) z ˙ = T ρ ( x ˙ ) , ξ = A ( x ˙ ) \xi=A(\dot{x}) ξ = A ( x ˙ ) and p x ↔ ( p z , μ ) p_x\leftrightarrow(p_z,\mu) p x ↔ ( p z , μ ) under the above identification.
The reduced Lagrange–Pontryagin principle is formulated in terms of a curve ( z ( t ) , p z ( t ) , μ ( t ) , u ~ ( t ) ) (z(t),p_z(t),\mu(t),\widetilde{u}(t)) ( z ( t ) , p z ( t ) , μ ( t ) , u ( t )) in the reduced Pontryagin bundle M / G ≅ T ∗ ( P / G ) ⊕ g ~ ∗ ⊕ U ~ M/G\cong T^*(P/G)\oplus\widetilde{\mathfrak{g}}^*\oplus\widetilde{U} M / G ≅ T ∗ ( P / G ) ⊕ g ∗ ⊕ U , together with the reduced velocity ( z ˙ ( t ) , ξ ( t ) ) ∈ T ( P / G ) ⊕ g ~ (\dot{z}(t),\xi(t))\in T(P/G)\oplus\widetilde{\mathfrak{g}} ( z ˙ ( t ) , ξ ( t )) ∈ T ( P / G ) ⊕ g , where ξ \xi ξ is the vertical component of the velocity determined by the chosen principal connection. The reduced Lagrange–Pontryagin functional is therefore
S G = ∫ 0 T [ l ( z , u ) + ⟨ p z , z ˙ − Γ ~ ( z , u ) ⟩ + ⟨ μ , ξ − γ ~ ( z , u ) ⟩ ] d t . (22) S_G=\int_0^T\left[l(z,u)+\left\langle p_z,\dot{z}-\widetilde{\Gamma}(z,u)\right\rangle+\left\langle\mu,\xi-\widetilde{\gamma}(z,u)\right\rangle\right]\,dt. \tag*{(22)} S G = ∫ 0 T [ l ( z , u ) + ⟨ p z , z ˙ − Γ ( z , u ) ⟩ + ⟨ μ , ξ − γ ( z , u ) ⟩ ] d t . (22)
Equivalently,
S G = ∫ 0 T [ ⟨ p z , z ˙ ⟩ + ⟨ μ , ξ ⟩ − h ( z , p z , μ , u ) ] d t (23) S_G=\int_0^T \left[\langle p_z,\dot z\rangle+\langle\mu,\xi\rangle-h(z,p_z,\mu,u)\right]\,\mathrm{d}t \tag*{(23)} S G = ∫ 0 T [ ⟨ p z , z ˙ ⟩ + ⟨ μ , ξ ⟩ − h ( z , p z , μ , u ) ] d t (23)
where the reduced Pontryagin Hamiltonian is
h ( z , p z , μ , u ) = ⟨ p z , Γ ~ ( z , u ) ⟩ + ⟨ μ , γ ~ ( z , u ) ⟩ − l ( z , u ) . (24) h(z,p_z,\mu,u)=\langle p_z,\widetilde{\Gamma}(z,u)\rangle+\langle\mu,\widetilde{\gamma}(z,u)\rangle-l(z,u). \tag*{(24)} h ( z , p z , μ , u ) = ⟨ p z , Γ ( z , u )⟩ + ⟨ μ , γ ( z , u )⟩ − l ( z , u ) . (24)
The variations of the base variables satisfy δ z ( 0 ) = δ z ( T ) = 0 \delta z(0)=\delta z(T)=0 δ z ( 0 ) = δ z ( T ) = 0 , while the variations δ p z \delta p_z δ p z , δ μ \delta\mu δ μ and δ u \delta u δ u are arbitrary. The variation of the reduced vertical velocity ξ \xi ξ is constrained by the Lagrange–Poincaré reduction [15 ] : if η ( t ) ∈ g ~ z ( t ) \eta(t)\in\widetilde{\mathfrak{g}}_{z(t)} η ( t ) ∈ g z ( t ) denotes the reduced vertical variation, with η ( 0 ) = η ( T ) = 0 \eta(0)=\eta(T)=0 η ( 0 ) = η ( T ) = 0 , then
δ ξ = D η D t + [ η , ξ ] + F A ( δ z , z ˙ ) , (25) \delta\xi=\frac{D\eta}{Dt}+[\eta,\xi]+F_A(\delta z,\dot z), \tag*{(25)} δ ξ = D t D η + [ η , ξ ] + F A ( δ z , z ˙ ) , (25)
where F A F_A F A denotes the curvature of the chosen principal connection A A A .
Two covariant derivatives appear in the computation that follows, and we fix them now. The principal connection A A A induces a linear connection ∇ A \nabla^A ∇ A on the adjoint bundle g ~ \widetilde{\mathfrak{g}} g and, by duality, on g ~ ∗ \widetilde{\mathfrak{g}}^* g ∗ ; the symbols D η / D t D\eta/Dt D η / D t and D μ / D t D\mu/Dt D μ / D t denote the corresponding covariant derivatives along z ( ⋅ ) z(\cdot) z ( ⋅ ) . In a local trivialization over a chart ( z α ) (z^\alpha) ( z α ) on B B B , with { e i } \{e_i\} { e i } a local frame of g ~ \widetilde{\mathfrak{g}} g and ∇ ∂ α A e i = Γ α i k e k \nabla^A_{\partial_\alpha}e_i=\Gamma^k_{\alpha i}e_k ∇ ∂ α A e i = Γ α i k e k ,
( D η D t ) k = η ˙ k + Γ α i k z ˙ α η i , ( D μ D t ) i = μ ˙ i − Γ α i k μ k z ˙ α . \left(\frac{D\eta}{Dt}\right)^k=\dot\eta^k+\Gamma^k_{\alpha i}\dot z^\alpha\eta^i,\qquad
\left(\frac{D\mu}{Dt}\right)_i=\dot\mu_i-\Gamma^k_{\alpha i}\mu_k\dot z^\alpha. ( D t D η ) k = η ˙ k + Γ α i k z ˙ α η i , ( D t D μ ) i = μ ˙ i − Γ α i k μ k z ˙ α .
The connection A A A does not, however, induce a connection on the base B = P / G B=P/G B = P / G , and the symbols D p z / D t Dp_z/Dt D p z / D t , δ z ˙ \delta\dot z δ z ˙ and ∂ h / ∂ z \partial h/\partial z ∂ h / ∂ z below require one. We therefore fix, once and for all, an auxiliary torsion-free linear connection ∇ \nabla ∇ on B B B (any one will do, for instance the Levi-Civita connection of a Riemannian metric; in Sections 7.8 and 7.9 it will be the Levi-Civita connection of the reduced metric), and we understand: D p z / D t Dp_z/Dt D p z / D t as the covariant derivative of p z ( ⋅ ) p_z(\cdot) p z ( ⋅ ) along z ( ⋅ ) z(\cdot) z ( ⋅ ) with respect to the connection induced by ∇ \nabla ∇ on T ∗ B T^*B T ∗ B ; δ z ˙ \delta\dot z δ z ˙ as the covariant variation D ( δ z ) / D t D(\delta z)/Dt D ( δ z ) / D t , which agrees with the covariant ϵ \epsilon ϵ -derivative of z ˙ \dot z z ˙ because ∇ \nabla ∇ is torsion-free; and ∂ h / ∂ z \partial h/\partial z ∂ h / ∂ z as the covariant partial derivative of h h h , namely the derivative of h h h along ∇ ⊕ ∇ A \nabla\oplus\nabla^A ∇ ⊕ ∇ A -parallel translation of the fiber variables ( p z , μ ) (p_z,\mu) ( p z , μ ) , the control u u u being held fixed in a local trivialization of U ~ \widetilde{U} U . The fiber derivatives ∂ h / ∂ p z \partial h/\partial p_z ∂ h / ∂ p z , ∂ h / ∂ μ \partial h/\partial\mu ∂ h / ∂ μ and ∂ h / ∂ u \partial h/\partial u ∂ h / ∂ u require no connection. This is the same device as in the Hamilton–Poincaré equations of Cendra, Marsden, Pekarsky and Ratiu [14 ] . The choice of ∇ \nabla ∇ affects the two symbols separately but not the equations: under a change ∇ ↦ ∇ + S \nabla\mapsto\nabla+S ∇ ↦ ∇ + S , with S S S a symmetric ( 1 , 2 ) (1,2) ( 1 , 2 ) -tensor field, D p z / D t Dp_z/Dt D p z / D t changes by − ⟨ p z , S ( z ˙ , ⋅ ) ⟩ -\langle p_z,S(\dot z,\cdot)\rangle − ⟨ p z , S ( z ˙ , ⋅ )⟩ and ∂ h / ∂ z \partial h/\partial z ∂ h / ∂ z by + ⟨ p z , S ( ⋅ , ∂ h / ∂ p z ) ⟩ +\langle p_z,S(\cdot,\partial h/\partial p_z)\rangle + ⟨ p z , S ( ⋅ , ∂ h / ∂ p z )⟩ , and the two changes cancel in the reduced equations (26) below by the symmetry of S S S , along the first equation z ˙ = ∂ h / ∂ p z \dot z=\partial h/\partial p_z z ˙ = ∂ h / ∂ p z .
The reduced Lagrange–Pontryagin equations. We now derive the reduced Pontryagin Maximum Principle directly from the reduced variational principle. Taking the variation of (23), we obtain
δ S G = ∫ 0 T [ ⟨ δ p z , z ˙ ⟩ + ⟨ p z , δ z ˙ ⟩ + ⟨ δ μ , ξ ⟩ + ⟨ μ , δ ξ ⟩ − ⟨ ∂ h ∂ z , δ z ⟩ − ⟨ ∂ h ∂ p z , δ p z ⟩ − ⟨ ∂ h ∂ μ , δ μ ⟩ − ⟨ ∂ h ∂ u , δ u ⟩ ] d t . \delta S_G=\int_0^T\left[
\langle\delta p_z,\dot z\rangle+\langle p_z,\delta\dot z\rangle+\langle\delta\mu,\xi\rangle+\langle\mu,\delta\xi\rangle
-\left\langle\frac{\partial h}{\partial z},\delta z\right\rangle
-\left\langle\frac{\partial h}{\partial p_z},\delta p_z\right\rangle
-\left\langle\frac{\partial h}{\partial\mu},\delta\mu\right\rangle
-\left\langle\frac{\partial h}{\partial u},\delta u\right\rangle
\right]\mathrm{d}t. δ S G = ∫ 0 T [ ⟨ δ p z , z ˙ ⟩ + ⟨ p z , δ z ˙ ⟩ + ⟨ δ μ , ξ ⟩ + ⟨ μ , δ ξ ⟩ − ⟨ ∂ z ∂ h , δ z ⟩ − ⟨ ∂ p z ∂ h , δ p z ⟩ − ⟨ ∂ μ ∂ h , δ μ ⟩ − ⟨ ∂ u ∂ h , δ u ⟩ ] d t .
We substitute (25) for δ ξ \delta\xi δ ξ , integrate by parts the terms containing δ z ˙ \delta\dot z δ z ˙ and D η / D t D\eta/Dt D η / D t , which is legitimate since δ z \delta z δ z and η \eta η vanish at the endpoints, and use ⟨ μ , [ η , ξ ] ⟩ = − ⟨ ad ξ ∗ μ , η ⟩ \langle\mu,[\eta,\xi]\rangle=-\langle\operatorname{ad}^*_\xi\mu,\eta\rangle ⟨ μ , [ η , ξ ]⟩ = − ⟨ ad ξ ∗ μ , η ⟩ , which is our convention for the coadjoint action. Collecting all terms, we obtain
δ S G = ∫ 0 T [ ⟨ z ˙ − ∂ h ∂ p z , δ p z ⟩ + ⟨ ξ − ∂ h ∂ μ , δ μ ⟩ − ⟨ D p z D t + ∂ h ∂ z , δ z ⟩ + ⟨ μ , F A ( δ z , z ˙ ) ⟩ − ⟨ D μ D t + ad ξ ∗ μ , η ⟩ − ⟨ ∂ h ∂ u , δ u ⟩ ] d t . \delta S_G=\int_0^T\left[
\left\langle\dot z-\frac{\partial h}{\partial p_z},\delta p_z\right\rangle
+\left\langle\xi-\frac{\partial h}{\partial\mu},\delta\mu\right\rangle
-\left\langle\frac{Dp_z}{Dt}+\frac{\partial h}{\partial z},\delta z\right\rangle
+\langle\mu,F_A(\delta z,\dot z)\rangle
-\left\langle\frac{D\mu}{Dt}+\operatorname{ad}^*_\xi\mu,\eta\right\rangle
-\left\langle\frac{\partial h}{\partial u},\delta u\right\rangle
\right]\mathrm{d}t. δ S G = ∫ 0 T [ ⟨ z ˙ − ∂ p z ∂ h , δ p z ⟩ + ⟨ ξ − ∂ μ ∂ h , δ μ ⟩ − ⟨ D t D p z + ∂ z ∂ h , δ z ⟩ + ⟨ μ , F A ( δ z , z ˙ )⟩ − ⟨ D t D μ + ad ξ ∗ μ , η ⟩ − ⟨ ∂ u ∂ h , δ u ⟩ ] d t .
Since δ z \delta z δ z , δ p z \delta p_z δ p z , δ μ \delta\mu δ μ , δ u \delta u δ u and η \eta η are arbitrary, the stationarity condition δ S G = 0 \delta S_G=0 δ S G = 0 yields
z ˙ = ∂ h ∂ p z , D p z D t = − ∂ h ∂ z + ⟨ μ , F A ( ⋅ , z ˙ ) ⟩ , D μ D t = − ad ξ ∗ μ , ξ = ∂ h ∂ μ , 0 = ∂ h ∂ u . (26) \begin{aligned}
\dot{z} &= \frac{\partial h}{\partial p_z},\\
\frac{D p_z}{D t} &= -\frac{\partial h}{\partial z}+\langle\mu,F_A(\cdot,\dot{z})\rangle,\\
\frac{D\mu}{D t} &= -\operatorname{ad}^{*}_{\xi}\mu,\\
\xi&= \frac{\partial h}{\partial\mu},\\
0 &= \frac{\partial h}{\partial u}. \tag*{(26)}
\end{aligned} z ˙ D t D p z D t D μ ξ 0 = ∂ p z ∂ h , = − ∂ z ∂ h + ⟨ μ , F A ( ⋅ , z ˙ )⟩ , = − ad ξ ∗ μ , = ∂ μ ∂ h , = ∂ u ∂ h . (26)
These equations constitute the reduced first-order necessary conditions of the Pontryagin Maximum Principle; by the antisymmetry of the curvature two-form, the curvature term in the second equation may equally be written as − ⟨ μ , F A ( z ˙ , ⋅ ) ⟩ -\langle\mu,F_A(\dot{z},\cdot)\rangle − ⟨ μ , F A ( z ˙ , ⋅ )⟩ .
Remark 1 (the left-action convention). Suppose that G G G acts on P P P on the left. The connection is then normalized by A ( ζ P ) = ζ A(\zeta_P)=\zeta A ( ζ P ) = ζ with ζ P ( x ) = d d t ∣ t = 0 exp ( t ζ ) ⋅ x \zeta_P(x)=\left.\frac{d}{dt}\right|_{t=0}\exp(t\zeta)\cdot x ζ P ( x ) = d t d t = 0 exp ( tζ ) ⋅ x , the adjoint bundle carries the equivalence ( x , ζ ) ∼ ( g ⋅ x , Ad g ζ ) (x,\zeta)\sim(g\cdot x,\operatorname{Ad}_g\zeta) ( x , ζ ) ∼ ( g ⋅ x , Ad g ζ ) , the induced covariant derivative on g ~ \widetilde{\mathfrak{g}} g reads D / D t = ∂ t − ad ξ D/Dt=\partial_t-\operatorname{ad}_{\xi} D / D t = ∂ t − ad ξ , and the structure equation becomes F A = d A − 1 2 [ A , A ] F_A=dA-\frac{1}{2}[A,A] F A = d A − 2 1 [ A , A ] , since the orbit map now pulls A A A back to the right, rather than the left, Maurer–Cartan form. With these conventions the constrained variations (25) read δ ξ = D η / D t + [ ξ , η ] + F A ( δ z , z ˙ ) \delta\xi=D\eta/Dt+[\xi,\eta]+F_A(\delta z,\dot{z}) δ ξ = D η / D t + [ ξ , η ] + F A ( δ z , z ˙ ) , and the reduced Lagrange–Pontryagin equations are (26) with the sign of the coadjoint equation reversed,
D μ D t = + ad ξ ∗ μ (27) \frac{D\mu}{D t}=+\operatorname{ad}^{*}_{\xi}\mu \tag*{(27)} D t D μ = + ad ξ ∗ μ (27)
the other four equations, and in particular the curvature term, are unaffected, since the latter arises from ∫ 0 T ⟨ μ , F A ( δ z , z ˙ ) ⟩ d t \int_0^T\langle\mu,F_A(\delta z,\dot{z})\rangle\,dt ∫ 0 T ⟨ μ , F A ( δ z , z ˙ )⟩ d t in either convention. This is the form taken by the reduced Pontryagin Maximum Principle in body variables, and is consistent with the implicit Lie–Poisson equations of Yoshimura and Marsden [40 Prop. 6.5] .
Reconstruction
Reduction is only half of the correspondence. We now show that a solution of the reduced equations (26), together with an initial condition in P P P , determines a unique extremal of the original optimal control problem, so that no information is lost in passing to the quotient.
Let ( z ( t ) , p z ( t ) , μ ( t ) , u ( t ) ) (z(t),p_z(t),\mu(t),u(t)) ( z ( t ) , p z ( t ) , μ ( t ) , u ( t )) be a solution of (26) on [ 0 , T ] [0,T] [ 0 , T ] , and let ξ ( t ) = ∂ h / ∂ μ ∈ g ~ z ( t ) \xi(t)=\partial h/\partial\mu\in\widetilde{\mathfrak{g}}_{z(t)} ξ ( t ) = ∂ h / ∂ μ ∈ g z ( t ) be the associated reduced vertical velocity. Fix x 0 ∈ P x_0\in P x 0 ∈ P with ρ ( x 0 ) = z ( 0 ) \rho(x_0)=z(0) ρ ( x 0 ) = z ( 0 ) .
Definition 8. The horizontal lift of z ( ⋅ ) z(\cdot) z ( ⋅ ) through x 0 x_0 x 0 is the unique curve x h ( ⋅ ) x^h(\cdot) x h ( ⋅ ) in P P P with
ρ ( x h ( t ) ) = z ( t ) , A ( x ˙ h ( t ) ) = 0 , x h ( 0 ) = x 0 . \rho(x^h(t))=z(t),\qquad A(\dot{x}^h(t))=0,\qquad x^h(0)=x_0. ρ ( x h ( t )) = z ( t ) , A ( x ˙ h ( t )) = 0 , x h ( 0 ) = x 0 . Along it, write ξ h ( t ) ∈ g \xi^h(t)\in\mathfrak{g} ξ h ( t ) ∈ g for the representative of ξ ( t ) \xi(t) ξ ( t ) determined by x h ( t ) x^h(t) x h ( t ) , that is, ξ ( t ) = [ x h ( t ) , ξ h ( t ) ] \xi(t)=[x^h(t),\xi^h(t)] ξ ( t ) = [ x h ( t ) , ξ h ( t )] .
Theorem 1 (Reconstruction). Let g ( ⋅ ) g(\cdot) g ( ⋅ ) be the solution of the initial value problem in G G G
g ˙ ( t ) = T R g ( t ) ξ h ( t ) , g ( 0 ) = e (28) \dot{g}(t)=T R_{g(t)}\xi^h(t),\qquad g(0)=e \tag*{(28)} g ˙ ( t ) = T R g ( t ) ξ h ( t ) , g ( 0 ) = e (28) and set x ( t ) = x h ( t ) ⋅ g ( t ) x(t)=x^h(t)\cdot g(t) x ( t ) = x h ( t ) ⋅ g ( t ) . Then x ( ⋅ ) x(\cdot) x ( ⋅ ) is the unique curve in P P P with x ( 0 ) = x 0 x(0)=x_0 x ( 0 ) = x 0 whose class in T P / G ≃ T ( P / G ) ⊕ g ~ TP/G\simeq T(P/G)\oplus\widetilde{\mathfrak{g}} T P / G ≃ T ( P / G ) ⊕ g is ( z ˙ ( t ) , ξ ( t ) ) (\dot{z}(t),\xi(t)) ( z ˙ ( t ) , ξ ( t )) . Let μ ^ ( t ) ∈ g ∗ \widehat{\mu}(t)\in\mathfrak{g}^{*} μ ( t ) ∈ g ∗ be the representative of μ ( t ) \mu(t) μ ( t ) determined by x ( t ) x(t) x ( t ) , that is, μ ( t ) = [ x ( t ) , μ ^ ( t ) ] \mu(t)=[x(t),\widehat{\mu}(t)] μ ( t ) = [ x ( t ) , μ ( t )] , and let c ( t ) ∈ C x ( t ) c(t)\in C_{x(t)} c ( t ) ∈ C x ( t ) be the representative of the reduced control u ~ ( t ) ∈ U ~ z ( t ) \widetilde{u}(t)\in\widetilde{U}_{z(t)} u ( t ) ∈ U z ( t ) determined by x ( t ) x(t) x ( t ) under (19). Define the costate p ( t ) ∈ T x ( t ) ∗ P p(t)\in T^{*}_{x(t)}P p ( t ) ∈ T x ( t ) ∗ P by
⟨ p ( t ) , v ⟩ = ⟨ p z ( t ) , T ρ ( v ) ⟩ + ⟨ μ ^ ( t ) , A ( v ) ⟩ , v ∈ T x ( t ) P (29) \langle p(t),v\rangle=\langle p_z(t),T\rho(v)\rangle+\langle\widehat{\mu}(t),A(v)\rangle,\qquad v\in T_{x(t)}P \tag*{(29)} ⟨ p ( t ) , v ⟩ = ⟨ p z ( t ) , T ρ ( v )⟩ + ⟨ μ ( t ) , A ( v )⟩ , v ∈ T x ( t ) P (29) which is the unique covector over x ( t ) x(t) x ( t ) whose class is [ β J ] − 1 ( z ( t ) , p z ( t ) , μ ( t ) ) ∈ T ∗ P / G [\beta_J]^{-1}(z(t),p_z(t),\mu(t))\in T^{*}P/G [ β J ] − 1 ( z ( t ) , p z ( t ) , μ ( t )) ∈ T ∗ P / G . Then the curve ( x ( t ) , p ( t ) , c ( t ) ) (x(t),p(t),c(t)) ( x ( t ) , p ( t ) , c ( t )) is an extremal of the unreduced Pontryagin Maximum Principle issuing from x 0 x_0 x 0 ; it solves the fixed-endpoint problem with terminal state x T x_T x T precisely when the terminal holonomy condition x ( T ) = x T x(T)=x_T x ( T ) = x T holds, which is a condition on z ( ⋅ ) z(\cdot) z ( ⋅ ) , ξ ( ⋅ ) \xi(\cdot) ξ ( ⋅ ) and x 0 x_0 x 0 and not merely on z ( T ) z(T) z ( T ) . Conversely, every extremal projects to a solution of (26), and two extremals project to the same solution if and only if they lie on the same G G G -orbit.
The three representatives μ ^ ( t ) \hat{\mu}(t) μ ^ ( t ) , c ( t ) c(t) c ( t ) and p ( t ) p(t) p ( t ) are the point at which the quotient must be undone: μ ( t ) \mu(t) μ ( t ) and u ~ ( t ) \tilde{u}(t) u ~ ( t ) are classes, and it is the choice of the reconstructed point x ( t ) x(t) x ( t ) in the fiber, not the reduced data alone, that selects representatives of them.
Proof. Write x = x h g x=x^h g x = x h g . By (15) ,
x ˙ = T Φ g ( x ˙ h ) + ( T L g − 1 g ˙ ) ( x h g ) , \dot{x}=T\Phi_g(\dot{x}^h)+(TL_{g^{-1}}\dot{g})(x^h g), x ˙ = T Φ g ( x ˙ h ) + ( T L g − 1 g ˙ ) ( x h g ) ,
so that, using the equivariance A x g ∘ T Φ g = Ad g − 1 ∘ A x A_{xg}\circ T\Phi_g=\operatorname{Ad}_{g^{-1}}\circ A_x A xg ∘ T Φ g = Ad g − 1 ∘ A x and A ( x ˙ h ) = 0 A(\dot{x}^h)=0 A ( x ˙ h ) = 0 ,
A x ( t ) ( x ˙ ( t ) ) = Ad g − 1 A ( x ˙ h ) + T L g − 1 g ˙ = T L g − 1 g ˙ . A_{x(t)}(\dot{x}(t))=\operatorname{Ad}_{g^{-1}}A(\dot{x}^h)+TL_{g^{-1}}\dot{g}=TL_{g^{-1}}\dot{g}. A x ( t ) ( x ˙ ( t )) = Ad g − 1 A ( x ˙ h ) + T L g − 1 g ˙ = T L g − 1 g ˙ .
On the other hand the representative of ξ ( t ) \xi(t) ξ ( t ) determined by x ( t ) = x h g x(t)=x^h g x ( t ) = x h g is Ad g − 1 ξ ~ ( t ) \operatorname{Ad}_{g^{-1}}\tilde{\xi}(t) Ad g − 1 ξ ~ ( t ) , by the equivalence ( x , ζ ) ∼ ( x g , Ad g − 1 ζ ) (x,\zeta)\sim(xg,\operatorname{Ad}_{g^{-1}}\zeta) ( x , ζ ) ∼ ( xg , Ad g − 1 ζ ) of (13) . The condition A ( x ˙ ) = Ad g − 1 ξ ˙ h A(\dot{x})=\operatorname{Ad}_{g^{-1}}\dot{\xi}^h A ( x ˙ ) = Ad g − 1 ξ ˙ h is therefore T L g − 1 g ˙ = Ad g − 1 ξ ˙ h TL_{g^{-1}}\dot{g}=\operatorname{Ad}_{g^{-1}}\dot{\xi}^h T L g − 1 g ˙ = Ad g − 1 ξ ˙ h , which, since Ad g − 1 = T L g − 1 ∘ T R g \operatorname{Ad}_{g^{-1}}=TL_{g^{-1}}\circ TR_g Ad g − 1 = T L g − 1 ∘ T R g , is exactly (28) . Moreover T ρ ( x ˙ ) = T ρ ( x ˙ h ) = z ˙ T\rho(\dot{x})=T\rho(\dot{x}^h)=\dot{z} T ρ ( x ˙ ) = T ρ ( x ˙ h ) = z ˙ , so the class of x ˙ \dot{x} x ˙ in T P / G TP/G T P / G is ( z ˙ , ξ ) (\dot{z},\xi) ( z ˙ , ξ ) under (17) . Equation (28) is a time-dependent right-invariant equation on G G G and therefore has a unique solution on all of [ 0 , T ] [0,T] [ 0 , T ] ; uniqueness of x ( ⋅ ) x(\cdot) x ( ⋅ ) follows since the pair ( z ˙ , ξ ) (\dot{z},\xi) ( z ˙ , ξ ) determines x ˙ \dot{x} x ˙ pointwise.
The costate (29) is well-defined because [ β J ] [\beta_J] [ β J ] of (18) is a vector bundle isomorphism, so that [ β J ] − 1 ( z , p z , μ ) [\beta_J]^{-1}(z,p_z,\mu) [ β J ] − 1 ( z , p z , μ ) is a well-defined class in T ∗ P / G T^*P/G T ∗ P / G ; the pairing (29) is exactly the representative of that class lying over the chosen point x ( t ) x(t) x ( t ) , and it is the only one. The same remark applies to c ( t ) c(t) c ( t ) : the class u ~ ( t ) = [ x ( t ) , c ( t ) ] \tilde{u}(t)=[x(t),c(t)] u ~ ( t ) = [ x ( t ) , c ( t )] has exactly one representative in the fiber C x ( t ) C_{x(t)} C x ( t ) .
That the resulting curve satisfies the unreduced first-order conditions is seen in two steps, neither of which uses the Dirac-geometric results of Section 4. First, the state equation. By (24) , the first and fourth equations of (26) read z ˙ = Γ ~ ( z ~ , u ) \dot{z}=\tilde{\Gamma}(\tilde{z},u) z ˙ = Γ ~ ( z ~ , u ) and ξ = γ ~ ( z , u ) \xi=\tilde{\gamma}(z,u) ξ = γ ~ ( z , u ) , so the class of x ˙ \dot{x} x ˙ in T P / G ≃ T ( P / G ) ⊕ g ~ TP/G\simeq T(P/G)\oplus\tilde{\mathfrak{g}} T P / G ≃ T ( P / G ) ⊕ g ~ , shown above to be ( z ˙ , ξ ) (\dot{z},\xi) ( z ˙ , ξ ) , coincides with the class of Γ ( x , u ) \Gamma(x,u) Γ ( x , u ) ; two vectors of T x P T_xP T x P with the same class are equal, the action being free, whence x ˙ = Γ ( x , u ) \dot{x}=\Gamma(x,u) x ˙ = Γ ( x , u ) . Second, the extremal property. By construction S ( γ ) = S G ( [ γ ] ) S(\gamma)=S_G([\gamma]) S ( γ ) = S G ([ γ ]) for γ = ( x , p , u ) \gamma=(x,p,u) γ = ( x , p , u ) , and the constrained variations (25) , together with free variations of ( p z , μ , u ) (p_z,\mu,u) ( p z , μ , u ) , are exactly the projections of the free variations of curves in M M M with fixed endpoints in P P P , every such reduced variation admitting a lift [15 ] . Stationarity of S G S_G S G at [ γ ] [\gamma] [ γ ] under the constrained variations is the system (26) , by Section 3.3, while stationarity of S S S at γ \gamma γ under free variations is the unreduced first-order system (11) ; the correspondence of variations therefore converts the one into the other. The converse statements are immediate from the G G G -invariance of the data. □
Remark 2. Once Theorem 4 is available, the same correspondence can be read off the Dirac side, since the reduced equations (26) are equivalent to the inclusion ( [ X ] G , d H G ) ∈ [ D P ] G ([X]_G,dH_G)\in[D_P]_G ([ X ] G , d H G ) ∈ [ D P ] G , which holds if and only if its representatives ( X , d H P ) (X,dH_P) ( X , d H P ) lie in D P D_P D P ; the proof given above is independent of that result. In the left-action convention of Remark 1 the same computation gives g ˙ = T L g ξ h \dot{g}=TL_g\xi^h g ˙ = T L g ξ h in place of (28) ; the examples of Sections 5 and 6 specialize accordingly. Finally, Theorem 1 makes precise the assertion of Section 3 that the critical points of [ S ] G [S]_G [ S ] G are in bijection with G G G -equivalence classes of extremals of S S S : a reduced solution determines a whole G G G -orbit of extremals, one for each choice of initial point in the fiber over z ( 0 ) z(0) z ( 0 ) , and for the fixed-endpoint problem, whose curve space is not itself G G G -invariant, the terminal condition x ( T ) = x T x(T)=x_T x ( T ) = x T selects, within the reduced solutions, those whose reconstruction through x 0 x_0 x 0 closes up.
Independence of the connection
The reduced equations (26) were written using a principal connection A A A , which is not part of the data of the optimal control problem. We now verify that the choice affects only the presentation.
Let A A A and A ′ A' A ′ be two principal connections on ρ : P → B \rho:P\to B ρ : P → B . Their difference τ = A ′ − A \tau=A'-A τ = A ′ − A is a G G G -equivariant horizontal g \mathfrak{g} g -valued one-form on P P P , hence corresponds to a one-form τ ~ \tilde{\tau} τ ~ on B B B with values in the adjoint bundle, τ ~ ∈ Ω 1 ( B ; g ~ ) \tilde{\tau}\in\Omega^1(B;\tilde{\mathfrak{g}}) τ ~ ∈ Ω 1 ( B ; g ~ ) .
Proposition 3. Under the change of connection A ↦ A ′ = A + τ A\mapsto A'=A+\tau A ↦ A ′ = A + τ the identifications (17) and (18) change by
ξ ′ = ξ + τ ~ ( z ˙ ) , p z ′ = p z − ⟨ μ , τ ~ ⟩ , μ ′ = μ , (30) \xi'=\xi+\tilde{\tau}(\dot{z}),\qquad p'_z=p_z-\langle\mu,\tilde{\tau}\rangle,\qquad\mu'=\mu, \tag*{(30)} ξ ′ = ξ + τ ~ ( z ˙ ) , p z ′ = p z − ⟨ μ , τ ~ ⟩ , μ ′ = μ , (30) while the reduced Hamiltonian is unchanged as a function on M / G M/G M / G ,
h A ′ ( z , p z ′ , μ , u ) = h A ( z , p z , μ , u ) . (31) h^{A'}(z,p'_z,\mu,u)=h^A(z,p_z,\mu,u). \tag*{(31)} h A ′ ( z , p z ′ , μ , u ) = h A ( z , p z , μ , u ) . (31)
Proof. The momentum map J J J is defined by evaluation on fundamental vector fields and so does not involve the connection; hence μ ′ = μ \mu'=\mu μ ′ = μ . For α ∈ T x ∗ P \alpha\in T_x^*P α ∈ T x ∗ P and v ~ ∈ T x P \tilde{v}\in T_xP v ~ ∈ T x P one has ⟨ p z , T ρ ( v ) ⟩ = ⟨ μ , A ( v ) ⟩ \langle p_z,T\rho(v)\rangle=\langle\mu,A(v)\rangle ⟨ p z , T ρ ( v )⟩ = ⟨ μ , A ( v )⟩ , and subtracting the same identity for A ′ A' A ′ gives p z ′ = p z − ⟨ μ , τ ~ ⟩ p'_z=p_z-\langle\mu,\tilde{\tau}\rangle p z ′ = p z − ⟨ μ , τ ~ ⟩ . Likewise ξ ′ = [ x , A ′ ( x ˙ ) ] = ξ + τ ~ ( Γ ~ ) \xi'=[x,A'(\dot{x})]=\xi+\tilde{\tau}(\tilde{\Gamma}) ξ ′ = [ x , A ′ ( x ˙ )] = ξ + τ ~ ( Γ ~ ) .
For (31) , note that Γ ~ = T ρ ∘ Γ \tilde{\Gamma}=T\rho\circ\Gamma Γ ~ = T ρ ∘ Γ does not depend on the connection, whereas γ ~ A ′ = γ ~ A + τ ~ ( Γ ~ ) \tilde{\gamma}^{A'}=\tilde{\gamma}^{A}+\tilde{\tau}(\tilde{\Gamma}) γ ~ A ′ = γ ~ A + τ ~ ( Γ ~ ) . Hence, by (24) ,
h A ′ = ⟨ p z − ⟨ μ , τ ~ ⟩ , Γ ~ ⟩ + ⟨ μ , γ ~ A + τ ~ ( Γ ~ ) ⟩ − l = h A , h^{A'}=\langle p_z-\langle\mu,\tilde{\tau}\rangle,\tilde{\Gamma}\rangle+\langle\mu,\tilde{\gamma}^{A}+\tilde{\tau}(\tilde{\Gamma})\rangle-l=h^A, h A ′ = ⟨ p z − ⟨ μ , τ ~ ⟩ , Γ ~ ⟩ + ⟨ μ , γ ~ A + τ ~ ( Γ ~ )⟩ − l = h A ,
the two τ ~ \tilde{\tau} τ ~ terms canceling.
Theorem 2. The reduced dynamics is independent of the choice of principal connection. More precisely, the map (30) is an isomorphism of vector bundles over B B B carrying the reduced Dirac structure and the reduced equations (26) written with respect to A A A to those written with respect to A ′ A' A ′ , and it commutes with reconstruction: the unreduced extremal produced by Theorem 1 is the same for both connections.
Proof. The reduced Dirac structure [ D Ω ] G [D_{\Omega}]_G [ D Ω ] G and the reduced Hamiltonian H G H_G H G are defined without reference to a connection, as is the intrinsic reduced dynamics ( [ X ] G , [ d H ] ) ∈ [ D Ω ] G ([X]_G,[dH])\in[D_{\Omega}]_G ([ X ] G , [ d H ]) ∈ [ D Ω ] G of (34). The connection enters only through the isomorphism [ β J ] [\beta_J] [ β J ] of (18) used to write that condition in the explicit form (42), and (30) is exactly the composite [ β J A ′ ] ∘ [ β J A ] − 1 [\beta_{J}^{A'}]\circ[\beta_{J}^{A}]^{-1} [ β J A ′ ] ∘ [ β J A ] − 1 . The two systems of equations are therefore the same condition expressed in two coordinatizations, and (31) shows that the Hamiltonian is literally the same function. Reconstruction is likewise connection-free: the class of x ˙ \dot{x} x ˙ in T P / G TP/G T P / G is determined by x ˙ \dot{x} x ˙ , and A A A enters Theorem 1 only through the splitting of that class into a horizontal lift and a group factor.
Criticality and optimality
The Pontryagin Maximum Principle provides necessary conditions, and everything above concerns critical points of the functionals rather than minimizers. We make the relationship explicit, and record what is and is not claimed.
Three standing conventions should be stated plainly. First, we work with the weak form of the maximum principle: the control set is the whole fiber C x C_x C x , controls are smooth, and the maximum condition is replaced by the stationarity condition ∂ H P / ∂ u a = 0 \partial H_P/\partial u^a=0 ∂ H P / ∂ u a = 0 . Second, the normalization p 0 = − 1 p_0=-1 p 0 = − 1 implicit in H P = ⟨ p , Γ ⟩ − L H_P=\langle p,\Gamma\rangle-L H P = ⟨ p , Γ ⟩ − L selects the normal extremals; the abnormal case p 0 = 0 p_0=0 p 0 = 0 is treated in Section 7 through the cost extension of Proposition 12. Third, the endpoints x 0 , x T ∈ P x_0,x_T\in P x 0 , x T ∈ P are prescribed, so that in the reduced principle the vertical variation satisfies η ( 0 ) = η ( T ) = 0 \eta(0)=\eta(T)=0 η ( 0 ) = η ( T ) = 0 .
Proposition 4. Let γ \gamma γ be an admissible curve of ( P , C , Γ , L ) (P,\mathcal{C},\Gamma,L) ( P , C , Γ , L ) from x 0 x_0 x 0 to x T x_T x T and let [ γ ] [\gamma] [ γ ] be its reduction.
(i) S ( γ ) = S G ( [ γ ] ) S(\gamma)=S_G([\gamma]) S ( γ ) = S G ([ γ ]) , and this holds for every lift of [ γ ] [\gamma] [ γ ] ;
(ii) γ \gamma γ is a critical point of S S S if and only if [ γ ] [\gamma] [ γ ] is a critical point of S G S_G S G with respect to the constrained variations (25);
(iii) γ \gamma γ is a local minimizer of S S S among admissible curves from x 0 x_0 x 0 to x T x_T x T if and only if [ γ ] [\gamma] [ γ ] is a local minimizer of S G S_G S G among reduced admissible curves from z 0 z_0 z 0 to z T z_T z T whose reconstruction through x 0 x_0 x 0 , in the sense of Theorem 1, terminates at x T x_T x T ;
(iv) if γ \gamma γ is a critical point of S S S , then the second variation of S S S at γ \gamma γ descends to the second variation of S G S_G S G at [ γ ] [\gamma] [ γ ] ; and for the orbit-endpoint problem introduced at the beginning of Section 3 the tangent space to the G G G -orbit of γ \gamma γ lies in the kernel of the former.
Proof. Part (i) is the definition ℓ ( z , u ) = L ( x , u ) \ell(z,u)=L(x,u) ℓ ( z , u ) = L ( x , u ) together with the G G G -invariance of L L L . Part (ii) is the content of Section 3.3 combined with Theorem 1: the constrained variations (25) are exactly the projections of the free variations of curves in P P P with fixed endpoints, and η ( 0 ) = η ( T ) = 0 \eta(0)=\eta(T)=0 η ( 0 ) = η ( T ) = 0 encodes the fixed endpoints. Part (iii) follows from (i) and the bijection of Theorem 1, which identifies admissible curves from x 0 x_0 x 0 to x T x_T x T with reduced admissible curves whose reconstruction through x 0 x_0 x 0 terminates at x T x_T x T ; the terminal condition is a genuine restriction, since a reduced curve from z 0 z_0 z 0 to z T z_T z T reconstructs to a curve ending anywhere in the fiber over z T z_T z T , the discrepancy being the group element g ( T ) g(T) g ( T ) of (28) (for the Heisenberg problem, the signed area swept by the plane curve, which is the enclosed area when the curve is closed), and this is why the vertical variable ξ \xi ξ is retained in the reduced principle (22) rather than eliminated. Part (iv) follows by differentiating (i) twice. Two qualifications are needed for the second assertion of (iv), and both are essential. First, criticality cannot be dropped: at a noncritical point the second derivative along an orbit is Hess ( γ ˙ , γ ˙ ) + d S ( γ ) \operatorname{Hess}(\dot{\gamma},\dot{\gamma})+dS(\gamma) Hess ( γ ˙ , γ ˙ ) + d S ( γ ) , and invariance annihilates the sum rather than the first term. The rotation-invariant function x 2 + y 2 x^2+y^2 x 2 + y 2 on R 2 \mathbb{R}^2 R 2 at ( 1 , 0 ) (1,0) ( 1 , 0 ) , whose Hessian is 2 I 2I 2 I , is the elementary instance. Second, the orbit direction is tangent to the space of competitors only for the orbit-endpoint problem, since γ ⋅ g \gamma\cdot g γ ⋅ g does not have the endpoints of γ \gamma γ ; for the fixed-endpoint problem the assertion holds with G G G replaced by the common stabilizer G x 0 ∩ G x T G_{x_0}\cap G_{x_T} G x 0 ∩ G x T , which is trivial when the action is free.
Proposition 4 asserts only that reduction neither creates nor destroys minimizers; no sufficient conditions are claimed here, and the question of minimality for the abnormal extremals of Section 7 is taken up in Section 7.9; in general the regularity of abnormal minimizers remains open [2 , 33 ] .
Relation to Lie algebroid reduction
Martínez [29 ] formulates optimal control on Lie algebroids and obtains reduced equations by regarding a G G G -invariant problem on P P P as a problem on the Atiyah algebroid T P / G TP/G T P / G . The correspondence between that formulation and the present one is known: Grabowski and Jóźwikowski [23, §7], in the wider setting of almost Lie algebroids, reduce an invariant optimal control problem on a principal bundle to the Atiyah algebroid and compute the resulting Poisson tensor on E ∗ ≅ T ∗ M × g ∗ E^{*} \cong T^{*}M \times\mathfrak{g}^{*} E ∗ ≅ T ∗ M × g ∗ in a connection-adapted splitting, obtaining a canonical part, a curvature term and a Lie–Poisson part; for P = G P = G P = G this recovers results of Jurdjevic [27 ] . The purpose of this subsection is not to add to those results but to record the dictionary explicitly for (26) , since the correspondence is asserted in Section 1 and has not so far been exhibited.
Recall that the Atiyah algebroid of ρ : P → B \rho: P \to B ρ : P → B is the vector bundle E = T P / G → B E = TP/G \to B E = T P / G → B of § 3.3 , whose sections are the G G G -invariant vector fields on P P P , with the Lie bracket of vector fields and with anchor ϱ = T ρ \varrho= T\rho ϱ = T ρ . Its dual is
E ∗ = T ∗ P / G ≅ T ∗ ( P / G ) ⊕ g ~ ∗ , E^{*} = T^{*}P/G \cong T^{*}(P/G) \oplus\widetilde{\mathfrak{g}}^{*}, E ∗ = T ∗ P / G ≅ T ∗ ( P / G ) ⊕ g ∗ ,
by (18) . Consequently the reduced Pontryagin bundle (20) is
M / G ≅ E ∗ × B U ~ , (32) M/G \cong E^{*} \times_{B} \widetilde{U}, \tag*{(32)} M / G ≅ E ∗ × B U , (32)
the Pontryagin bundle of the Atiyah algebroid, and the reduced Pontryagin Hamiltonian h h h is a function on it.
For a Lie algebroid E → B E \to B E → B with local coordinates ( z α ) (z^{\alpha}) ( z α ) on B B B and a local basis of sections { e a } \{e_{a}\} { e a } with
ϱ ( e a ) = ϱ a α ∂ ∂ z α , [ e a , e b ] E = C a b c e c , \varrho(e_{a}) = \varrho_{a}^{\alpha} \frac{\partial}{\partial z^{\alpha}}, \qquad[e_{a},e_{b}]_{E} = C_{ab}^{c} e_{c}, ϱ ( e a ) = ϱ a α ∂ z α ∂ , [ e a , e b ] E = C ab c e c ,
the Hamiltonian equations on E ∗ E^{*} E ∗ , with fiber coordinates y a y_{a} y a , are
z ˙ α = ϱ a α ∂ h ∂ y a , y ˙ a = − ϱ a α ∂ h ∂ z α − C a b c y c ∂ h ∂ y b , (33) \dot{z}^{\alpha} = \varrho_{a}^{\alpha} \frac{\partial h}{\partial y_{a}}, \qquad
\dot{y}_{a} = -\varrho_{a}^{\alpha} \frac{\partial h}{\partial z^{\alpha}} - C_{ab}^{c}y_{c}\frac{\partial h}{\partial y_{b}}, \tag*{(33)} z ˙ α = ϱ a α ∂ y a ∂ h , y ˙ a = − ϱ a α ∂ z α ∂ h − C ab c y c ∂ y b ∂ h , (33)
together with ∂ h / ∂ u a = 0 \partial h/\partial u^{a} = 0 ∂ h / ∂ u a = 0 in the controlled case [29 ] .
Proposition 5. Let A A A be a principal connection and let { X α h } \{X_{\alpha}^{h}\} { X α h } be the horizontal lifts of a coordinate frame on B B B , { e i } \{e_{i}\} { e i } a local frame of g ~ \widetilde{\mathfrak{g}} g . In the adapted basis { X α h , e i } \{X_{\alpha}^{h},e_{i}\} { X α h , e i } of E E E the anchor and structure functions are
ϱ ( X α h ) = ∂ ∂ z α , ϱ ( e i ) = 0 , \varrho(X_{\alpha}^{h}) = \frac{\partial}{\partial z^{\alpha}}, \qquad\varrho(e_{i}) = 0, ϱ ( X α h ) = ∂ z α ∂ , ϱ ( e i ) = 0 , C α β i = − F α β i , C α i k = − C i α k = Γ α i k , C i j k = − c i j k , C_{\alpha\beta}^{i} = -F_{\alpha\beta}^{i}, \qquad C_{\alpha i}^{k} = -C_{i\alpha}^{k} = \Gamma_{\alpha i}^{k}, \qquad C_{ij}^{k} = -c_{ij}^{k}, C α β i = − F α β i , C α i k = − C i α k = Γ α i k , C ij k = − c ij k , where F α β i F_{\alpha\beta}^{i} F α β i are the components of the curvature F A F_{A} F A , Γ α i k \Gamma_{\alpha i}^{k} Γ α i k those of the induced connection on g ~ \widetilde{\mathfrak{g}} g introduced in § 3.4 , and c i j k c_{ij}^{k} c ij k the structure constants of the fiberwise bracket of the adjoint bundle, [ e i , e j ] = c i j k e k [e_{i},e_{j}] = c_{ij}^{k}e_{k} [ e i , e j ] = c ij k e k , together with the values forced by antisymmetry of the bracket, C β α i = F α β i C_{\beta\alpha}^{i} = F_{\alpha\beta}^{i} C β α i = F α β i and C j i k = c i j k C_{ji}^{k} = c_{ij}^{k} C j i k = c ij k . All structure functions not determined by these vanish.
Proof. The anchor statements are immediate. Since [ X α h , X β h ] [X_{\alpha}^{h},X_{\beta}^{h}] [ X α h , X β h ] projects to [ ∂ α , ∂ β ] = 0 [\partial_{\alpha},\partial_{\beta}] = 0 [ ∂ α , ∂ β ] = 0 it is vertical, and its vertical part is A ( [ X α h , X β h ] ) = − F A ( ∂ α , ∂ β ) A([X_{\alpha}^{h},X_{\beta}^{h}]) = -F_{A}(\partial_{\alpha},\partial_{\beta}) A ([ X α h , X β h ]) = − F A ( ∂ α , ∂ β ) by Lemma 4. (Readers comparing with [23 ] should note the opposite curvature convention there: F ∇ ( X , Y ) = [ ∇ X , ∇ Y ] E − ∇ ( [ X , Y ] ) F_{\nabla}(X,Y) = [\nabla X,\nabla Y]_{E} - \nabla([X,Y]) F ∇ ( X , Y ) = [ ∇ X , ∇ Y ] E − ∇ ([ X , Y ]) equals A ( [ X h , Y h ] ) A([X^{h},Y^{h}]) A ([ X h , Y h ]) , so that F ∇ = − F A F_{\nabla} = -F_{A} F ∇ = − F A with the Cartan structure equation F A = d A + 1 2 [ A , A ] F_{A} = dA + \frac{1}{2}[A,A] F A = d A + 2 1 [ A , A ] used here.) The bracket of a horizontal lift with a vertical section is the covariant derivative of that section, [ X α h , X σ ] = X ∇ ∂ α ∇ α σ [X_{\alpha}^{h},X_{\sigma}] = X_{\nabla_{\partial_{\alpha}}^{\nabla_{\alpha}\sigma}} [ X α h , X σ ] = X ∇ ∂ α ∇ α σ , which one checks in a local trivialization. The bracket of two invariant vertical fields is the opposite of the fiberwise bracket of g ~ \widetilde{\mathfrak{g}} g : for a right action the invariant vertical fields restrict on each fiber to right-invariant vector fields, and right-invariant fields bracket anti-homomorphically, so [ X σ , X τ ] = − X [ σ , τ ] [X_{\sigma},X_{\tau}] = -X_{[\sigma,\tau]} [ X σ , X τ ] = − X [ σ , τ ] and C i j k = − c i j k C_{ij}^{k} = -c_{ij}^{k} C ij k = − c ij k . (For a left action the invariant vertical fields are the left-invariant ones and C i j k = + c i j k C_{ij}^{k} = +c_{ij}^{k} C ij k = + c ij k , consistently with the sign change of the coadjoint equation in Remark 1.) □ \square □
Theorem 3. Under the identification (32) , and writing y = ( p α , μ ) y = (p_{\alpha},\mu) y = ( p α , μ ) , the algebroid Hamiltonian equations (33) for the Atiyah algebroid coincide with the reduced Lagrange–Pontryagin equations (26) . In particular the curvature term of (26) is the structure function C α β i = − F α β i C_{\alpha\beta}^{i} = -F_{\alpha\beta}^{i} C α β i = − F α β i , the coadjoint term is c i j k = − c i j k c_{ij}^{k} = -c_{ij}^{k} c ij k = − c ij k , and the structure functions C α i k = Γ α i k C_{\alpha i}^{k} = \Gamma_{\alpha i}^{k} C α i k = Γ α i k are absorbed by the covariant derivative D μ / D t D\mu/Dt D μ / D t and by the covariant partial derivative ∂ h / ∂ z α \partial h/\partial z^{\alpha} ∂ h / ∂ z α of § 3.4 .
Proof. Substituting Proposition 5 into (33) and writing y α = p α y_{\alpha} = p_{\alpha} y α = p α , y i = μ i y_{i} = \mu_{i} y i = μ i , ∂ h / ∂ μ i = ξ i \partial h/\partial\mu_{i} = \xi^{i} ∂ h / ∂ μ i = ξ i , the first equation gives z ˙ α = ∂ h / ∂ p α \dot{z}^{\alpha} = \partial h/\partial p_{\alpha} z ˙ α = ∂ h / ∂ p α . The equation for p α p_{\alpha} p α reads
p ˙ α = − ∂ h ∂ z α − C α β i μ i ∂ h ∂ p β − C α i k μ k ∂ h ∂ μ i = − ∂ h ∂ z α + F α β i μ i z ˙ β − Γ α i k μ k ξ i , \dot{p}_{\alpha} = -\frac{\partial h}{\partial z^{\alpha}} - C_{\alpha\beta}^{i}\mu_{i}\frac{\partial h}{\partial p_{\beta}} - C_{\alpha i}^{k}\mu_{k}\frac{\partial h}{\partial\mu_{i}} = -\frac{\partial h}{\partial z^{\alpha}} + F_{\alpha\beta}^{i}\mu_{i}\dot{z}^{\beta} - \Gamma_{\alpha i}^{k}\mu_{k}\xi^{i}, p ˙ α = − ∂ z α ∂ h − C α β i μ i ∂ p β ∂ h − C α i k μ k ∂ μ i ∂ h = − ∂ z α ∂ h + F α β i μ i z ˙ β − Γ α i k μ k ξ i ,
in which the middle term is − ⟨ μ , F A ( z ˙ , ∂ α ) ⟩ -\langle\mu,F_{A}(\dot{z},\partial_{\alpha})\rangle − ⟨ μ , F A ( z ˙ , ∂ α )⟩ and the last is the connection term by which the naive partial derivative ∂ h / ∂ z α \partial h/\partial z^{\alpha} ∂ h / ∂ z α , taken at fixed ( p α , μ i ) (p_{\alpha},\mu_{i}) ( p α , μ i ) , differs from the covariant partial derivative of § 3.4 : since the reduced equations are independent of the auxiliary connection, we may take ∇ \nabla ∇ flat in the chart, and then ( D p z / D t ) α = p ˙ α (D p_z/Dt)_\alpha=\dot{p}_\alpha ( D p z / D t ) α = p ˙ α while ( ∂ h / ∂ z ) α = ∂ h / ∂ z α + Γ α i κ μ k ξ i (\partial h/\partial z)_\alpha=\partial h/\partial z^\alpha+\Gamma^\kappa_{\alpha i}\mu_k\xi^i ( ∂ h / ∂ z ) α = ∂ h / ∂ z α + Γ α i κ μ k ξ i ; the displayed equation is therefore the second equation of (26) . The equation for μ i \mu_i μ i reads, using C i α k = − Γ α i k C^k_{i\alpha}=-\Gamma^k_{\alpha i} C i α k = − Γ α i k ,
μ ˙ i = − C i j k μ k ∂ h ∂ μ j − C i α k μ k ∂ h ∂ p α = C i j k μ k ξ j + Γ α i k μ k z ˙ α , \dot{\mu}_i=-C^k_{ij}\mu_k\frac{\partial h}{\partial\mu_j}-C^k_{i\alpha}\mu_k\frac{\partial h}{\partial p_\alpha}=C^k_{ij}\mu_k\xi^j+\Gamma^k_{\alpha i}\mu_k\dot{z}^\alpha, μ ˙ i = − C ij k μ k ∂ μ j ∂ h − C i α k μ k ∂ p α ∂ h = C ij k μ k ξ j + Γ α i k μ k z ˙ α ,
which is D μ / D t = − ad ξ ∗ μ D\mu/Dt=-\operatorname{ad}^*_\xi\mu D μ / D t = − ad ξ ∗ μ , since ( ad ξ ∗ μ ) i = ⟨ μ , [ ξ , e i ] ⟩ = − c i j k μ k ξ j (\operatorname{ad}^*_\xi\mu)_i=\langle\mu,[\xi,e_i]\rangle=-c^k_{ij}\mu_k\xi^j ( ad ξ ∗ μ ) i = ⟨ μ , [ ξ , e i ]⟩ = − c ij k μ k ξ j . The stationarity condition ∂ h / ∂ u a = 0 \partial h/\partial u^a=0 ∂ h / ∂ u a = 0 is common to both. □
Remark 3. The reduction of the variations, rather than only of the equations, has also been studied in the algebroid setting; see Grabowska and Grabowski [22 ] for constrained variational calculus on general algebroids and the attendant distinction between the vakonomic and nonholonomic principles.
Remark 4. The comparison also explains the status of the two constructions. Martínez’s formulation is intrinsic, the algebroid T P / G TP/G T P / G requiring no choice of connection; the price is that the equations are written in terms of structure functions relative to a chosen local basis. The formulation of Section 3.3 fixes a connection once and for all, which converts the structure functions into the geometric objects F A F_A F A and ad ∗ \operatorname{ad}^* ad ∗ and makes the vertical variable ξ \xi ξ available for reconstruction. Proposition 3 is the statement that changing the connection is a change of adapted basis, which is why the resulting dynamics is the same.
Remark 5. The Dirac formulation of § 4 does not have an evident algebroid counterpart. The reduced object [ D Ω ] G [D_\Omega]_G [ D Ω ] G is a Dirac subbundle of the Courant algebroid ( T M ⊕ T ∗ M ) / G (TM\oplus T^*M)/G ( T M ⊕ T ∗ M ) / G , and the natural question of whether it can be described as a Dirac structure on a suitable prolongation of T P / G TP/G T P / G appears to be open; we do not pursue it here.
Reduction of the Dirac structure of optimal control problems with symmetry
Lie–Dirac reduction of the Dirac structure
Let ( P , C , Γ , L ) (P,C,\Gamma,L) ( P , C , Γ , L ) be a G G G -invariant optimal control problem, and let D Ω ⊂ T M ⊕ T ∗ M D_\Omega\subset TM\oplus T^*M D Ω ⊂ T M ⊕ T ∗ M be the Dirac structure on the Pontryagin bundle M = T ∗ P × P C M=T^*P\times_P C M = T ∗ P × P C associated with the canonical presymplectic form Ω \Omega Ω .
Since the action of G G G on M M M lifts naturally to the tangent and cotangent bundles, it also induces an action on the Whitney sum T M ⊕ T ∗ M TM\oplus T^*M T M ⊕ T ∗ M . It is therefore natural to ask whether the Dirac structure descends to the quotient. The following result shows that this is indeed the case.
Proposition 6. Let ( P , C , Γ , L ) (P,C,\Gamma,L) ( P , C , Γ , L ) be a G G G -invariant optimal control problem. Then the canonical Dirac structure D Ω ⊂ T M ⊕ T ∗ M D_\Omega\subset TM\oplus T^*M D Ω ⊂ T M ⊕ T ∗ M is G G G -invariant.
Proof. If α = i X Ω \alpha=i_X\Omega α = i X Ω then g ∗ − 1 α = i g ∗ X Ω g^{*-1}\alpha=i_{g_*X}\Omega g ∗− 1 α = i g ∗ X Ω , since g ∗ Ω = Ω g^*\Omega=\Omega g ∗ Ω = Ω by Proposition 2; hence ( g ∗ X , g ∗ − 1 α ) ∈ D Ω (g_*X,g^{*-1}\alpha)\in D_\Omega ( g ∗ X , g ∗− 1 α ) ∈ D Ω . □
Since the action of G G G on T M ⊕ T ∗ M TM\oplus T^*M T M ⊕ T ∗ M is fiber preserving, the quotient ( T M ⊕ T ∗ M ) / G (TM\oplus T^*M)/G ( T M ⊕ T ∗ M ) / G , which is canonically the Whitney sum of the quotient bundles T M / G TM/G T M / G and T ∗ M / G T^*M/G T ∗ M / G , is a vector bundle over the reduced Pontryagin bundle M / G M/G M / G , and the G G G -invariance of D Ω D_\Omega D Ω implies that it descends to a vector subbundle of the quotient bundle.
Definition 9. The vector subbundle [ D Ω ] G ⊂ T M / G ⊕ T ∗ M / G [D_\Omega]_G\subset TM/G\oplus T^*M/G [ D Ω ] G ⊂ T M / G ⊕ T ∗ M / G is called the Lie–Dirac reduction of the Dirac structure D Ω D_\Omega D Ω .
From now on we assume, as holds in all the examples of this paper, that the standard fiber U U U of the control bundle is a vector space on which G G G acts linearly, so that U ~ \widetilde{U} U is a vector bundle and A A A induces linear connections on g ~ \widetilde{\mathfrak g} g , g ~ ∗ \widetilde{\mathfrak g}^* g ∗ and U ~ \widetilde{U} U ; in particular, A A A identifies T M / G TM/G T M / G with T ( M / G ) ⊕ g ~ T(M/G)\oplus\widetilde{\mathfrak g} T ( M / G ) ⊕ g and T ∗ M / G T^*M/G T ∗ M / G with T ∗ ( M / G ) ⊕ g ~ ∗ T^*(M/G)\oplus\widetilde{\mathfrak g}^* T ∗ ( M / G ) ⊕ g ∗ , as in § 3.3 .
Notice that, in general, [ D Ω ] G [D_\Omega]_G [ D Ω ] G is not a Dirac structure on the reduced Pontryagin bundle M / G M/G M / G , since it is not a subbundle of
T ( M / G ) ⊕ T ∗ ( M / G ) . T(M/G)\oplus T^*(M/G). T ( M / G ) ⊕ T ∗ ( M / G ) .
Indeed, using the canonical identifications T M / G ≅ T ( M / G ) ⊕ g ~ TM/G\cong T(M/G)\oplus\widetilde{\mathfrak g} T M / G ≅ T ( M / G ) ⊕ g and T ∗ M / G ≅ T ∗ ( M / G ) ⊕ g ~ ∗ T^*M/G\cong T^*(M/G)\oplus\widetilde{\mathfrak g}^* T ∗ M / G ≅ T ∗ ( M / G ) ⊕ g ∗ , we obtain
[ D Ω ] G ⊂ T ( M / G ) ⊕ T ∗ ( M / G ) ⊕ g ~ ⊕ g ~ ∗ . [D_\Omega]_G\subset T(M/G)\oplus T^*(M/G)\oplus\widetilde{\mathfrak g}\oplus\widetilde{\mathfrak g}^*. [ D Ω ] G ⊂ T ( M / G ) ⊕ T ∗ ( M / G ) ⊕ g ⊕ g ∗ .
It is, however, a Dirac structure in a precise sense. The bundle
E = ( T M ⊕ T ∗ M ) / G ⟶ M / G E=(TM\oplus T^*M)/G\longrightarrow M/G E = ( T M ⊕ T ∗ M ) / G ⟶ M / G
carries the quotients of the symmetric pairing of Definition 4 and of the Courant bracket of Definition 5, both of which are preserved by the tangent and cotangent lifts of the G G G -action; hence E E E is a Courant algebroid over M / G M/G M / G in the general sense of Liu, Weinstein and Xu [28 ] , although it is not of the form T ( M / G ) ⊕ T ∗ ( M / G ) T(M/G) \oplus T^{*}(M/G) T ( M / G ) ⊕ T ∗ ( M / G ) . Accordingly, if D ⊂ T M ⊕ T ∗ M D \subset TM \oplus T^{*}M D ⊂ T M ⊕ T ∗ M is a G G G -invariant Dirac structure, then D / G D/G D / G is a Dirac subbundle of E E E , and D / G D/G D / G is integrable whenever D D D is, since G G G preserves the Courant bracket (cf. [40 p. 400] ). In particular Ω = pr P ∗ ω P \Omega= \operatorname{pr}_{P}^{*}\omega_{P} Ω = pr P ∗ ω P is closed, so that D Ω D_{\Omega} D Ω is integrable and hence so is [ D Ω ] G [D_{\Omega}]_{G} [ D Ω ] G .
We shall keep the notation [ D Ω ] G [D_{\Omega}]_{G} [ D Ω ] G and refer to it as the reduced Dirac structure , understanding it as a Dirac subbundle of the Courant algebroid E E E ; the pullback and pushforward introduced below are then the corresponding morphisms of Courant algebroids.
Reduced Dirac equations of an optimal control problem with symmetry
Once the reduced Dirac structure [ D Ω ] G [D_{\Omega}]_{G} [ D Ω ] G has been constructed, the corresponding reduced dynamics is simply determined by the inclusion
( [ X ] , [ d H ] ) ∈ [ D Ω ] G . (34) ([X], [dH]) \in[D_{\Omega}]_{G}. \tag*{(34)} ([ X ] , [ d H ]) ∈ [ D Ω ] G . (34)
Here [ X ] [X] [ X ] denotes a section of the reduced tangent bundle T M / G TM/G T M / G , while [ d H ] [dH] [ d H ] is the corresponding section of T ∗ M / G T^{*}M/G T ∗ M / G determined by the differential of the reduced Pontryagin Hamiltonian H G : M / G → R H_{G}: M/G \to\mathbb{R} H G : M / G → R , induced by the G G G -invariant Hamiltonian H P H_{P} H P (cf. (12) ). Under the identifications of § 3.3 , the reduced vector field takes values in T ( M / G ) ⊕ g ~ T(M/G) \oplus\tilde{\mathfrak{g}} T ( M / G ) ⊕ g ~ and the reduced covector field in T ∗ ( M / G ) ⊕ g ~ ∗ T^{*}(M/G) \oplus\tilde{\mathfrak{g}}^{*} T ∗ ( M / G ) ⊕ g ~ ∗ .
Although equation (34) provides the intrinsic reduced dynamics, its explicit computation is rather involved because of the structure of the reduced bundle. A considerable simplification follows from observing that the presymplectic form on the Pontryagin bundle is the pullback of the canonical symplectic form on T ∗ P T^{*}P T ∗ P , namely Ω = pr P ∗ ω P \Omega= \operatorname{pr}_{P}^{*}\omega_{P} Ω = pr P ∗ ω P , and does not involve the control directions. Consequently, after projecting these out, the Lie–Dirac reduction of D Ω D_{\Omega} D Ω is directly related to the Lie–Dirac reduction of the canonical Dirac structure on T ∗ P T^{*}P T ∗ P associated with ω P \omega_{P} ω P .
To make this relation precise we recall the notions of pullback and pushforward of Dirac structures [41 ] , which for the quotients considered here are the corresponding morphisms of Courant algebroids.
Definition 10 (Pushforward and pullback of Dirac structures). Let ψ : N → N ′ \psi: N \to N^{\prime} ψ : N → N ′ be a smooth map, and let Dir ( N ) \operatorname{Dir}(N) Dir ( N ) and Dir ( N ′ ) \operatorname{Dir}(N^{\prime}) Dir ( N ′ ) denote the sets of Dirac structures on N N N and N ′ N^{\prime} N ′ , respectively.
The forward map F ψ : Dir ( N ) → Dir ( N ′ ) F_{\psi} : \operatorname{Dir}(N) \to\operatorname{Dir}(N^{\prime}) F ψ : Dir ( N ) → Dir ( N ′ ) is defined by
F ψ ( D ) = { ( ψ ∗ u , α ) ∣ ( u , ψ ∗ α ) ∈ D } , F_{\psi}(D) = \{(\psi_{*}u,\alpha) \mid(u,\psi^{*}\alpha) \in D\}, F ψ ( D ) = {( ψ ∗ u , α ) ∣ ( u , ψ ∗ α ) ∈ D } , where u ∈ T N u \in TN u ∈ T N and α ∈ T ∗ N ′ \alpha\in T^{*}N^{\prime} α ∈ T ∗ N ′ .
Similarly, the backward map B ψ : Dir ( N ′ ) → Dir ( N ) B_{\psi} : \operatorname{Dir}(N^{\prime}) \to\operatorname{Dir}(N) B ψ : Dir ( N ′ ) → Dir ( N ) is defined by
B ψ ( D ′ ) = { ( v , ψ ∗ β ) ∣ ( ψ ∗ v , β ) ∈ D ′ } , B_{\psi}(D^{\prime}) = \{(v,\psi^{*}\beta) \mid(\psi_{*}v,\beta) \in D^{\prime}\}, B ψ ( D ′ ) = {( v , ψ ∗ β ) ∣ ( ψ ∗ v , β ) ∈ D ′ } , where v ∈ T N v \in TN v ∈ T N and β ∈ T ∗ N ′ \beta\in T^{*}N^{\prime} β ∈ T ∗ N ′ .
The pushforward and pullback of Dirac structures are then defined by
ψ ∗ D = F ψ ( D ) , ψ ∗ D ′ = B ψ ( D ′ ) . \psi_{*}D = F_{\psi}(D), \qquad\psi^{*}D^{\prime} = B_{\psi}(D^{\prime}). ψ ∗ D = F ψ ( D ) , ψ ∗ D ′ = B ψ ( D ′ ) .
Both formulas are, in the first place, pointwise prescriptions, and for an arbitrary smooth ψ \psi ψ the subspaces they produce are maximally isotropic at each point but need not have constant rank, hence need not be smooth subbundles; a forward image must in addition be constant along the fibers of ψ \psi ψ in order to be well defined on N ′ N^{\prime} N ′ . We therefore read Definition 10 as producing pointwise Dirac relations , and call ψ ∗ D \psi_{*}D ψ ∗ D or ψ ∗ D ′ \psi^{*}D^{\prime} ψ ∗ D ′ a Dirac structure only under the usual clean-intersection and constant-rank hypotheses (see [8 ] and Remark 11 below). All the maps to which the definition is applied in this paper are surjective submersions and all the structures involved are graphs of pullbacks of two-forms, for which those hypotheses are automatic: the backward image of the graph of Ω ~ \widetilde{\Omega} Ω along ψ \psi ψ is the graph of ψ ∗ Ω ~ \psi^{*}\widetilde{\Omega} ψ ∗ Ω , of constant rank dim N \dim N dim N , and the forward image of the graph of ψ ∗ Ω ~ \psi^{*}\widetilde{\Omega} ψ ∗ Ω is again the graph of Ω ~ \widetilde{\Omega} Ω .
The Dirac structures D Ω D_{\Omega} D Ω and D P D_{P} D P are related by pullback and pushforward through the canonical projection pr P : M → T ∗ P \operatorname{pr}_{P}: M \to T^{*}P pr P : M → T ∗ P .
Lemma 1. Let D P D_{P} D P denote the canonical Dirac structure on T ∗ P T^{*}P T ∗ P associated with the symplectic form ω P \omega_{P} ω P . Then
pr P ∗ D P = D Ω , pr P ∗ D Ω = D P . \operatorname{pr}_{P}^{*}D_{P} = D_{\Omega}, \qquad\operatorname{pr}_{P*}D_{\Omega} = D_{P}. pr P ∗ D P = D Ω , pr P ∗ D Ω = D P .
Proof. Since Ω = pr P ∗ ω P \Omega= \operatorname{pr}_{P}^{*}\omega_{P} Ω = pr P ∗ ω P and D P D_{P} D P is the graph of ω P \omega_{P} ω P , the condition ι X Ω = pr P ∗ α \iota_{X}\Omega= \operatorname{pr}_{P}^{*}\alpha ι X Ω = pr P ∗ α is equivalent to ι pr P ∗ X ω P = α \iota_{\operatorname{pr}_{P*}X}\omega_{P} = \alpha ι pr P ∗ X ω P = α , which by Definition 10 is both equalities at once. □ \square □
The previous result is a particular instance of a more general situation. Let ( M ~ , Ω ~ ) (\widetilde{M},\widetilde{\Omega}) ( M , Ω ) be a symplectic manifold and let p : M → M ~ p : M \to\widetilde{M} p : M → M be a surjective submersion. If Ω = p ∗ Ω ~ \Omega= p^{*}\widetilde{\Omega} Ω = p ∗ Ω , then exactly the same argument shows that D Ω = p ∗ D Ω ~ D_{\Omega}=p^{*}D_{\widetilde{\Omega}} D Ω = p ∗ D Ω : the presymplectic Dirac structure D Ω D_{\Omega} D Ω projects onto the symplectic Dirac structure D Ω ~ D_{\widetilde{\Omega}} D Ω .
Assume now that a Lie group G G G acts on both M M M and M ~ \widetilde{M} M in such a way that the projection p p p is equivariant,
p ( x ⋅ g ) = p ( x ) ⋅ g , x ∈ M , g ∈ G , (35) p(x\cdot g)=p(x)\cdot g,\qquad x\in M,\quad g\in G, \tag*{(35)} p ( x ⋅ g ) = p ( x ) ⋅ g , x ∈ M , g ∈ G , (35)
and that the action preserves the symplectic form, g ∗ Ω ~ = Ω ~ g^{*}\widetilde{\Omega}=\widetilde{\Omega} g ∗ Ω = Ω . Then both D Ω ~ D_{\widetilde{\Omega}} D Ω and D Ω D_{\Omega} D Ω are G G G -invariant. Consequently, the Lie–Dirac reduction of D Ω ~ D_{\widetilde{\Omega}} D Ω defines a Dirac subbundle of the corresponding Courant algebroid over the reduced space,
[ D Ω ~ ] G ⊂ T ( M ~ / G ) ⊕ T ∗ ( M ~ / G ) ⊕ g ~ ⊕ g ~ ∗ . [D_{\widetilde{\Omega}}]_{G}\subset T(\widetilde{M}/G)\oplus T^{*}(\widetilde{M}/G)\oplus\widetilde{\mathfrak{g}}\oplus\widetilde{\mathfrak{g}}^{*}. [ D Ω ] G ⊂ T ( M / G ) ⊕ T ∗ ( M / G ) ⊕ g ⊕ g ∗ .
Moreover, the equivariance condition (35) induces a smooth map p G : M / G → M ~ / G p_{G}:M/G\to\widetilde{M}/G p G : M / G → M / G satisfying
p G ( [ x ] ) = [ p ( x ) ] . (36) p_{G}([x])=[p(x)]. \tag*{(36)} p G ([ x ]) = [ p ( x )] . (36)
Before stating the corresponding statement after reduction, we must say what “pullback” means there, since [ D Ω ] G [D_{\Omega}]_{G} [ D Ω ] G and [ D Ω ~ ] G [D_{\widetilde{\Omega}}]_{G} [ D Ω ] G are not Dirac structures on manifolds but subbundles of the quotient Courant algebroids
E M = ( T M / G ) ⊕ ( T ∗ M / G ) , E M ~ = ( T M ~ / G ) ⊕ ( T ∗ M ~ / G ) , \mathcal{E}_{M}=(TM/G)\oplus(T^{*}M/G),\qquad\mathcal{E}_{\widetilde{M}}=(T\widetilde{M}/G)\oplus(T^{*}\widetilde{M}/G), E M = ( T M / G ) ⊕ ( T ∗ M / G ) , E M = ( T M / G ) ⊕ ( T ∗ M / G ) ,
over M / G M/G M / G and M ~ / G \widetilde{M}/G M / G respectively. A class [ α ] ∈ T ∗ M / G [\alpha]\in T^{*}M/G [ α ] ∈ T ∗ M / G is not an ordinary covector on M / G M/G M / G — under a principal connection it is a pair consisting of a covector on M / G M/G M / G and a section of the coadjoint bundle — so [ p ∗ β ] [p^{*}\beta] [ p ∗ β ] is not an ordinary cotangent pullback and no map M / G → M ~ / G M/G\to\widetilde{M}/G M / G → M / G can by itself induce it. What does induce it is the following pair of bundle maps, both consequences of the equivariance of p p p .
Since T p : T M → T M ~ Tp:TM\to T\widetilde{M} T p : T M → T M is G G G -equivariant, it descends to a vector bundle morphism
T p / G : T M / G ⟶ T M ~ / G , [ v ] ⟼ [ T p ( v ) ] , (37) T p/G:TM/G\longrightarrow T\widetilde{M}/G,\qquad[v]\longmapsto[Tp(v)], \tag*{(37)} T p / G : T M / G ⟶ T M / G , [ v ] ⟼ [ T p ( v )] , (37)
covering p G p_{G} p G . Dually, if β \beta β is a covector at p ( x ) p(x) p ( x ) then p ∗ β p^{*}\beta p ∗ β is a covector at x x x , and p ∗ ( β ⋅ g ) = ( p ∗ β ) ⋅ g p^{*}(\beta\cdot g)=(p^{*}\beta)\cdot g p ∗ ( β ⋅ g ) = ( p ∗ β ) ⋅ g , so the assignment
( T p ) ∗ / G : p G ∗ ( T ∗ M ~ / G ) ⟶ T ∗ M / G , [ β ] ⟼ [ p ∗ β ] , (38) (Tp)^{*}/G:p_{G}^{*}(T^{*}\widetilde{M}/G)\longrightarrow T^{*}M/G,\qquad[\beta]\longmapsto[p^{*}\beta], \tag*{(38)} ( T p ) ∗ / G : p G ∗ ( T ∗ M / G ) ⟶ T ∗ M / G , [ β ] ⟼ [ p ∗ β ] , (38)
is a well-defined vector bundle morphism over $M/G: the representative β \beta β is taken at p ( x ) p(x) p ( x ) for the same representative x x x used on the left, and changing x x x to x g xg xg changes both sides by the action of g g g . In the splitting determined by a principal connection, (37) is ( T p G , id g ) (Tp_{G},\operatorname{id}_{\mathfrak{g}}) ( T p G , id g ) and (38) is ( p G ∗ , id g ∗ ) (p_{G}^{*},\operatorname{id}_{\mathfrak{g}^{*}}) ( p G ∗ , id g ∗ ) , the adjoint and coadjoint components being the identity because p p p is equivariant and therefore matches infinitesimal generators: T p ( ζ M ) = ζ M ~ Tp(\zeta_{M})=\zeta_{\widetilde{M}} T p ( ζ M ) = ζ M for every ζ ∈ g \zeta\in\mathfrak{g} ζ ∈ g . This is the precise content of the informal statement that p G p_{G} p G “acts as the identity on the adjoint factor”.
The pair (37)–(38) defines a Courant relation
R p = { ( [ v ] , [ α ] ) , ( [ w ] , [ β ] ) : [ w ] = ( T p / G ) [ v ] , [ α ] = ( ( T p ) ∗ / G ) [ β ] } ⊂ E M × M / G p G ∗ E M ~ , (39) R_{p}=\left\{([v],[\alpha]),([w],[\beta]):[w]=(Tp/G)[v],\ [\alpha]=((Tp)^{*}/G)[\beta]\right\}\subset\mathcal{E}_{M}\times_{M/G}p_{G}^{*}\mathcal{E}_{\widetilde{M}}, \tag*{(39)} R p = { ([ v ] , [ α ]) , ([ w ] , [ β ]) : [ w ] = ( T p / G ) [ v ] , [ α ] = (( T p ) ∗ / G ) [ β ] } ⊂ E M × M / G p G ∗ E M , (39)
and we write B R p BR_{p} B R p for the associated backward image,
B R p ( D ′ ) = { ( [ v ] , [ α ] ) ∈ E M : ∃ [ β ] , ( ( T p / G ) [ v ] , [ β ] ) ∈ D ′ and [ α ] = ( ( T p ) ∗ / G ) [ β ] } . BR_{p}(D')=\left\{([v],[\alpha])\in\mathcal{E}_{M}:\exists[\beta],\ ((Tp/G)[v],[\beta])\in D'\ \text{and}\ [\alpha]=((Tp)^{*}/G)[\beta]\right\}. B R p ( D ′ ) = { ([ v ] , [ α ]) ∈ E M : ∃ [ β ] , (( T p / G ) [ v ] , [ β ]) ∈ D ′ and [ α ] = (( T p ) ∗ / G ) [ β ] } .
This is the reduced counterpart of the backward map B ψ B\psi B ψ of Definition 10, and it reduces to it when G G G is trivial.
Proposition 7. With the notation introduced above, the Lie–Dirac reduction of the presymplectic Dirac structure D Ω D_{\Omega} D Ω is the backward image, along the Courant relation (39), of the Lie–Dirac reduction of the symplectic Dirac structure D Ω ~ D_{\widetilde{\Omega}} D Ω .
[ D Ω ] G = B R p ( [ D Ω ~ ] G ) = : p G ∗ [ D Ω ~ ] G , [D_{\Omega}]_{G}=BR_{p}([D_{\widetilde{\Omega}}]_{G})=:p_{G}^{*}[D_{\widetilde{\Omega}}]_{G}, [ D Ω ] G = B R p ([ D Ω ] G ) =: p G ∗ [ D Ω ] G , where p G ∗ p_{G}^{*} p G ∗ denotes B R p BR_{p} B R p and not a cotangent pullback. In particular, for the Pontryagin bundle M = T ∗ P × P C M=T^{*}P\times_{P}C M = T ∗ P × P C of a G G G -invariant optimal control problem and p = pr P p=\operatorname{pr}_{P} p = pr P ,
[ D Ω ] G = B R pr P ( [ D P ] G ) = : pr G ∗ [ D P ] G , [D_{\Omega}]_{G}=BR_{\operatorname{pr}_{P}}([D_{P}]_{G})=:\operatorname{pr}_{G}^{*}[D_{P}]_{G}, [ D Ω ] G = B R pr P ([ D P ] G ) =: pr G ∗ [ D P ] G , where pr G : M / G ⟶ T ∗ P / G \operatorname{pr}_{G}:M/G\longrightarrow T^{*}P/G pr G : M / G ⟶ T ∗ P / G is the map induced by the canonical projection pr P \operatorname{pr}_{P} pr P .
Proof. By Lemma 1, extended to p p p as above, ( X , α ) ∈ D Ω (X,\alpha)\in D_{\Omega} ( X , α ) ∈ D Ω if and only if α = p ∗ β \alpha=p^{*}\beta α = p ∗ β for some β \beta β with ( p ∗ X , β ) ∈ D Ω ~ (p_{*}X,\beta)\in D_{\widetilde{\Omega}} ( p ∗ X , β ) ∈ D Ω . Passing to G G G -orbits, which is legitimate since D Ω D_{\Omega} D Ω and D Ω ~ D_{\widetilde{\Omega}} D Ω are G G G -invariant and the action is free, and using [ p ∗ X ] = ( T p / G ) [ X ] [p_{*}X]=(Tp/G)[X] [ p ∗ X ] = ( T p / G ) [ X ] and [ p ∗ β ] = ( ( T p ) ∗ / G ) [ β ] [p^{*}\beta]=((Tp)^{*}/G)[\beta] [ p ∗ β ] = (( T p ) ∗ / G ) [ β ] , which are the definitions (37) and (38), this becomes: ( [ X ] , [ α ] ) ∈ [ D Ω ] G ([X],[\alpha])\in[D_{\Omega}]_{G} ([ X ] , [ α ]) ∈ [ D Ω ] G if and only if [ α ] = ( ( T p ) ∗ / G ) [ β ] [\alpha]=((Tp)^{*}/G)[\beta] [ α ] = (( T p ) ∗ / G ) [ β ] for some [ β ] [\beta] [ β ] with ( ( T p / G ) [ X ] , [ β ] ) ∈ [ D Ω ~ ] G ((Tp/G)[X],[\beta])\in[D_{\widetilde{\Omega}}]_{G} (( T p / G ) [ X ] , [ β ]) ∈ [ D Ω ] G , which is the membership condition for B R ρ ( [ D Ω ~ ] G ) \mathcal{BR}_{\rho}([D_{\widetilde{\Omega}}]_G) B R ρ ([ D Ω ] G ) . The two inclusions are this equivalence read in its two directions. □
The reduced canonical Dirac structure
We now derive the connection-dependent description of [ D P ] G [D_P]_G [ D P ] G that will be used in the sequel. Throughout, A A A is a fixed principal connection on ρ : P → B = P / G \rho: P \to B = P/G ρ : P → B = P / G , with curvature F A = d A + 1 2 [ A , A ] F_A = dA + \frac{1}{2}[A,A] F A = d A + 2 1 [ A , A ] , and we use the identification T ∗ P / G ≅ T ∗ ( P / G ) ⊕ g ~ ∗ T^*P/G \cong T^*(P/G) \oplus\widetilde{\mathfrak{g}}^* T ∗ P / G ≅ T ∗ ( P / G ) ⊕ g ∗ of Section 3.1.3.
At a point ( z , p z , μ ) (z,p_z,\mu) ( z , p z , μ ) we write an element of the reduced tangent bundle as v = ( v z , v p z ; ξ , v μ ) v=(v_z,v_{p_z};\xi,v_\mu) v = ( v z , v p z ; ξ , v μ ) , where
( v z , v p z ) ∈ T ( z , p z ) T ∗ ( P / G ) , ξ ∈ g ~ z , v μ ∈ g ~ z ∗ . (v_z,v_{p_z}) \in T_{(z,p_z)}T^*(P/G), \qquad\xi\in\widetilde{\mathfrak{g}}_z, \qquad v_\mu\in\widetilde{\mathfrak{g}}_z^*. ( v z , v p z ) ∈ T ( z , p z ) T ∗ ( P / G ) , ξ ∈ g z , v μ ∈ g z ∗ .
and an element of the corresponding dual bundle as α = ( π z , π p z ; α ξ , α v μ ) \alpha=(\pi_z,\pi_{p_z};\alpha_\xi,\alpha_{v_\mu}) α = ( π z , π p z ; α ξ , α v μ ) . The pair notation ( v z , v p z ) (v_z,v_{p_z}) ( v z , v p z ) for a tangent vector to T ∗ ( P / G ) T^*(P/G) T ∗ ( P / G ) refers to the splitting T T ∗ ( P / G ) ≅ T B ⊕ T ∗ B TT^*(P/G)\cong TB\oplus T^*B T T ∗ ( P / G ) ≅ T B ⊕ T ∗ B determined by the auxiliary torsion-free connection ∇ \nabla ∇ fixed in Section 3.2; equivalently, all formulas below may be read in cotangent-natural local coordinates. Because ∇ \nabla ∇ is torsion-free, the canonical symplectic form of T ∗ ( P / G ) T^*(P/G) T ∗ ( P / G ) takes in this splitting the familiar pair form appearing as the first two terms of (41) . The duality is
⟨ α , v ⟩ = ⟨ π z , v z ⟩ + ⟨ π p z , v p z ⟩ + ⟨ α ξ , ξ ⟩ + ⟨ α v μ , v μ ⟩ . (40) \langle\alpha,v\rangle=\langle\pi_z,v_z\rangle+\langle\pi_{p_z},v_{p_z}\rangle+\langle\alpha_\xi,\xi\rangle+\langle\alpha_{v_\mu},v_\mu\rangle. \tag*{(40)} ⟨ α , v ⟩ = ⟨ π z , v z ⟩ + ⟨ π p z , v p z ⟩ + ⟨ α ξ , ξ ⟩ + ⟨ α v μ , v μ ⟩ . (40)
Lemma 2. Let Θ \Theta Θ denote the canonical one-form on T ∗ P T^*P T ∗ P , so that ω P = − d Θ \omega_P=-d\Theta ω P = − d Θ . Under the identification (18) , the form ω P \omega_P ω P descends to the fiberwise skew-symmetric bilinear form
Ω r e d ( v , w ) = ⟨ w p z , v z ⟩ − ⟨ v p z , w z ⟩ − ⟨ v μ , ζ ⟩ + ⟨ w μ , ξ ⟩ − ⟨ μ , [ ξ , ζ ] ⟩ − ⟨ μ , F A ( v z , w z ) ⟩ , (41) \Omega_{\mathrm{red}}(v,w)=\langle w_{p_z},v_z\rangle-\langle v_{p_z},w_z\rangle
-\langle v_\mu,\zeta\rangle+\langle w_\mu,\xi\rangle-\langle\mu,[\xi,\zeta]\rangle-\langle\mu,F_A(v_z,w_z)\rangle, \tag*{(41)} Ω red ( v , w ) = ⟨ w p z , v z ⟩ − ⟨ v p z , w z ⟩ − ⟨ v μ , ζ ⟩ + ⟨ w μ , ξ ⟩ − ⟨ μ , [ ξ , ζ ]⟩ − ⟨ μ , F A ( v z , w z )⟩ , (41) where w = ( w z , w p z ; ζ , w μ ) w=(w_z,w_{p_z};\zeta,w_\mu) w = ( w z , w p z ; ζ , w μ ) .
Proof. For α x ∈ T x ∗ P \alpha_x\in T_x^*P α x ∈ T x ∗ P and v ∈ T x P v\in T_xP v ∈ T x P the connection splits v = hor v + ver v v=\operatorname{hor}v+\operatorname{ver}v v = hor v + ver v , with ver v = ( A ( v ) ) P ( x ) \operatorname{ver}v=(A(v))_P(x) ver v = ( A ( v ) ) P ( x ) , so that
⟨ α x , v ⟩ = ⟨ p z , T ρ ( v ) ⟩ + ⟨ J ( x , α x ) , A ( v ) ⟩ , \langle\alpha_x,v\rangle=\langle p_z,T\rho(v)\rangle+\langle J(x,\alpha_x),A(v)\rangle, ⟨ α x , v ⟩ = ⟨ p z , T ρ ( v )⟩ + ⟨ J ( x , α x ) , A ( v )⟩ ,
where p z ∈ T z ∗ B p_z\in T_z^*B p z ∈ T z ∗ B is the horizontal part of α x \alpha_x α x and J J J is the momentum map appearing in (18) . Consequently
Θ = Θ B + ⟨ J , A ⟩ \Theta=\Theta_B+\langle J,A\rangle Θ = Θ B + ⟨ J , A ⟩
in the obvious sense, the two terms being pulled back to T ∗ P T^*P T ∗ P along the maps of (18) . Taking − d -d − d and using the structure equation,
− d ⟨ J , A ⟩ = − ⟨ d J ∧ A ⟩ − ⟨ J , F A ⟩ + 1 2 ⟨ J , [ A , A ] ⟩ , -d\langle J,A\rangle=-\langle dJ\wedge A\rangle-\langle J,F_A\rangle+\frac{1}{2}\langle J,[A,A]\rangle, − d ⟨ J , A ⟩ = − ⟨ dJ ∧ A ⟩ − ⟨ J , F A ⟩ + 2 1 ⟨ J , [ A , A ]⟩ ,
Passing to the quotient, the second term produces the curvature term − ⟨ μ , F A ( v z , w z ) ⟩ -\langle\mu,F_A(v_z,w_z)\rangle − ⟨ μ , F A ( v z , w z )⟩ , and the third produces + ⟨ μ , [ ξ , ζ ] ⟩ +\langle\mu,[\xi,\zeta]\rangle + ⟨ μ , [ ξ , ζ ]⟩ , since [ A , A ] ( v , w ) = 2 [ ξ , ζ ] [A,A](v,w)=2[\xi,\zeta] [ A , A ] ( v , w ) = 2 [ ξ , ζ ] . The first term requires care: the fiber components v μ , w μ v_\mu,w_\mu v μ , w μ of the splitting are the covariant ones, and along a curve through the point the naive derivative of the momentum in the trivialization differs from them by the coadjoint correction of the transport, d J ( v ) = v μ + ad ξ ∗ μ dJ(v)=v_\mu+\operatorname{ad}^*_\xi\mu dJ ( v ) = v μ + ad ξ ∗ μ . Hence
− ⟨ d J ∧ A ⟩ ( v , w ) = − ⟨ v μ , ζ ⟩ + ⟨ w μ , ξ ⟩ − 2 ⟨ μ , [ ξ , ζ ] ⟩ , -\langle dJ\wedge A\rangle(v,w)=-\langle v_\mu,\zeta\rangle+\langle w_\mu,\xi\rangle-2\langle\mu,[\xi,\zeta]\rangle, − ⟨ dJ ∧ A ⟩ ( v , w ) = − ⟨ v μ , ζ ⟩ + ⟨ w μ , ξ ⟩ − 2 ⟨ μ , [ ξ , ζ ]⟩ ,
and the bracket contributions of the first and third terms combine into the single term − ⟨ μ , [ ξ , ζ ] ⟩ -\langle\mu,[\xi,\zeta]\rangle − ⟨ μ , [ ξ , ζ ]⟩ of (41) , while − d Θ B -d\Theta_B − d Θ B gives the canonical part ⟨ w p z , v z ⟩ − ⟨ v p z , w z ⟩ \langle w_{p_z},v_z\rangle-\langle v_{p_z},w_z\rangle ⟨ w p z , v z ⟩ − ⟨ v p z , w z ⟩ . Skew-symmetry of (41) is immediate, the last two terms being skew by the antisymmetry of [ ⋅ , ⋅ ] [\cdot,\cdot] [ ⋅ , ⋅ ] and of F A F_A F A . For the detailed verification we refer to [41 ] . □
Proposition 8. The Lie–Dirac reduction [ D P ] G [D_P]_G [ D P ] G of the canonical Dirac structure on T ∗ P T^*P T ∗ P is the graph of (41) , that is,
[ D P ] G = { ( v , α ) ∣ α = i v Ω r e d } , [D_P]_G=\{(v,\alpha)\mid\alpha=i_v\Omega_{\mathrm{red}}\}, [ D P ] G = {( v , α ) ∣ α = i v Ω red } , and ( v , α ) ∈ [ D P ] G (v,\alpha)\in[D_P]_G ( v , α ) ∈ [ D P ] G if and only if
{ v z = π p z , v p z + π z = − ⟨ μ , F A ( v z , ⋅ ) ⟩ , ξ = α v μ , v μ + α ξ = − ad ξ ∗ μ . (42) \begin{cases}
v_z=\pi_{p_z},\\
v_{p_z}+\pi_z=-\langle\mu,F_A(v_z,\cdot)\rangle,\\
\xi=\alpha_{v_\mu},\\
v_\mu+\alpha_\xi=-\operatorname{ad}^*_\xi\mu. \tag*{(42)}
\end{cases} ⎩ ⎨ ⎧ v z = π p z , v p z + π z = − ⟨ μ , F A ( v z , ⋅ )⟩ , ξ = α v μ , v μ + α ξ = − ad ξ ∗ μ . (42) Moreover [ D P ] G [D_P]_G [ D P ] G is maximally isotropic with respect to the pairing induced by (40) .
Proof. Write ⟨ α , w ⟩ = Ω r e d ( v , w ) \langle\alpha,w\rangle=\Omega_{\mathrm{red}}(v,w) ⟨ α , w ⟩ = Ω red ( v , w ) and evaluate on the four families w = ( w z , 0 ; 0 , 0 ) w=(w_z,0;0,0) w = ( w z , 0 ; 0 , 0 ) , ( 0 , w p z ; 0 , 0 ) (0,w_{p_z};0,0) ( 0 , w p z ; 0 , 0 ) , ( 0 , 0 ; ζ , 0 ) (0,0;\zeta,0) ( 0 , 0 ; ζ , 0 ) and ( 0 , 0 ; 0 , w μ ) (0,0;0,w_\mu) ( 0 , 0 ; 0 , w μ ) in turn. The second gives π p z = v z \pi_{p_z}=v_z π p z = v z . The first gives ⟨ π z , w z ⟩ = − ⟨ v p z , w z ⟩ − ⟨ μ , F A ( v z , w z ) ⟩ \langle\pi_z,w_z\rangle=-\langle v_{p_z},w_z\rangle-\langle\mu,F_A(v_z,w_z)\rangle ⟨ π z , w z ⟩ = − ⟨ v p z , w z ⟩ − ⟨ μ , F A ( v z , w z )⟩ , that is, v p z + π z = − ⟨ μ , F A ( v z , ⋅ ) ⟩ v_{p_z}+\pi_z=-\langle\mu,F_A(v_z,\cdot)\rangle v p z + π z = − ⟨ μ , F A ( v z , ⋅ )⟩ . The fourth gives α v μ = ξ \alpha v_\mu=\xi α v μ = ξ . Finally the third gives
⟨ α ξ , ζ ⟩ = − ⟨ v μ , ζ ⟩ − ⟨ μ , ad ξ ζ ⟩ = − ⟨ v μ + ad ξ ∗ μ , ζ ⟩ \langle\alpha\xi,\zeta\rangle=-\langle v_\mu,\zeta\rangle-\langle\mu,\operatorname{ad}_\xi\zeta\rangle=-\langle v_\mu+\operatorname{ad}^{*}_\xi\mu,\zeta\rangle ⟨ α ξ , ζ ⟩ = − ⟨ v μ , ζ ⟩ − ⟨ μ , ad ξ ζ ⟩ = − ⟨ v μ + ad ξ ∗ μ , ζ ⟩
for all ζ ∈ g ~ z \zeta\in\widetilde{\mathfrak g}_z ζ ∈ g z , that is, v μ + α ξ = − ad ξ ∗ μ v_\mu+\alpha\xi=-\operatorname{ad}^{*}_\xi\mu v μ + α ξ = − ad ξ ∗ μ . These are precisely the four relations (42).
As for the last assertion, let ( v , α ) (v,\alpha) ( v , α ) and ( w , β ) (w,\beta) ( w , β ) belong to [ D P ] G [D_P]_G [ D P ] G . Then
⟨ ⟨ ( v , α ) , ( w , β ) ⟩ ⟩ = ⟨ α , w ⟩ + ⟨ β , v ⟩ = Ω r e d ( v , w ) + Ω r e d ( w , v ) = 0 \langle\langle(v,\alpha),(w,\beta)\rangle\rangle=\langle\alpha,w\rangle+\langle\beta,v\rangle=\Omega_{\mathrm{red}}(v,w)+\Omega_{\mathrm{red}}(w,v)=0 ⟨⟨( v , α ) , ( w , β )⟩⟩ = ⟨ α , w ⟩ + ⟨ β , v ⟩ = Ω red ( v , w ) + Ω red ( w , v ) = 0
by skew-symmetry, so that [ D P ] G [D_P]_G [ D P ] G is isotropic; and it is maximally so, being the graph of a bundle map and therefore of fiber rank equal to that of the reduced tangent bundle. □
Remark 6. Both of the signs that carry the content of the reduction, namely that of ad ξ ∗ μ \operatorname{ad}^{*}_\xi\mu ad ξ ∗ μ and that of the curvature coupling, originate in the single term − d ⟨ J , A ⟩ -\mathrm{d}\langle J,A\rangle − d ⟨ J , A ⟩ of Lemma 2, through the structure equation F A = d A + 1 2 [ A , A ] F_A=\mathrm{d}A+\frac{1}{2}[A,A] F A = d A + 2 1 [ A , A ] . They are not determined by isotropy alone: replacing − ad ξ ∗ μ -\operatorname{ad}^{*}_\xi\mu − ad ξ ∗ μ by + ad ξ ∗ μ +\operatorname{ad}^{*}_\xi\mu + ad ξ ∗ μ , or reversing the sign of the curvature term, again yields a maximally isotropic subbundle of the same rank. This is why (41) is derived here, rather than the relations (42) being posited.
Remark 7 (the left-action convention). With the conventions of Remark 1, that is, for a left action of G G G on P P P , the same computation gives the relations (42) with the last one replaced by
v μ + α ξ = + ad ξ ∗ μ (43) v_\mu+\alpha\xi=+\operatorname{ad}^{*}_\xi\mu \tag*{(43)} v μ + α ξ = + ad ξ ∗ μ (43)
which for P = G P=G P = G is the relation of Yoshimura and Marsden [40 Prop. 4.1] ; substituting the reduced vector field and the reduced differential then returns (26) with the coadjoint equation (27), so that Theorem 4 below holds verbatim in this convention.
The Lie–Dirac reduction and the reduced Pontryagin Maximum Principle
In Section 3.3 we derived the reduced Pontryagin Maximum Principle by reducing the Lagrange–Pontryagin variational principle. We now show that exactly the same equations are obtained from the Lie–Dirac reduction of the implicit Dirac formulation of the optimal control problem. By Proposition 7, the reduced presymplectic Dirac structure [ D Ω ] G [D_\Omega]_G [ D Ω ] G is the pullback of the reduced canonical Dirac structure [ D P ] G [D_P]_G [ D P ] G , so that it is sufficient to describe the dynamics using [ D P ] G [D_P]_G [ D P ] G .
Let H G : M / G ⟶ R H_G:M/G\longrightarrow\mathbb{R} H G : M / G ⟶ R be the reduced Pontryagin Hamiltonian. Since the optimal control problem is regular, the primary constraint manifold M 1 = { ∂ H P / ∂ u a = 0 } M_1=\{\partial H_P/\partial u^a=0\} M 1 = { ∂ H P / ∂ u a = 0 } is G G G -invariant and therefore descends to the quotient [ M 1 ] G = M 1 / G [M_1]_G=M_1/G [ M 1 ] G = M 1 / G , given by ∂ H G / ∂ u a = 0 \partial H_G/\partial u^a=0 ∂ H G / ∂ u a = 0 . Accordingly, the reduced implicit Dirac system is defined by
( [ X ] G , d h ) ∈ [ D P ] G , (44) ([X]_G,\mathrm{d}h)\in[D_P]_G, \tag*{(44)} ([ X ] G , d h ) ∈ [ D P ] G , (44)
where both entries are taken along the map
p 1 , G : [ M 1 ] G ⟶ T ∗ P / G p_{1,G}:[M_1]_G\longrightarrow T^{*}P/G p 1 , G : [ M 1 ] G ⟶ T ∗ P / G
induced by pr P \operatorname{pr}_P pr P , and h h h is the reduced Hamiltonian on T ∗ P / G T^{*}P/G T ∗ P / G obtained by eliminating the stationary control. Some care is needed with the types here. The differential d H G ∣ [ M 1 ] G \mathrm{d}H_G|_{[M_1]_G} d H G ∣ [ M 1 ] G is a covector on [ M 1 ] G [M_1]_G [ M 1 ] G , whereas the covector component of [ D P ] G [D_P]_G [ D P ] G lives over T ∗ P / G T^{*}P/G T ∗ P / G ; the two are matched by p 1 , G p_{1,G} p 1 , G , which by Proposition 1 is a local diffeomorphism at every regular point of M 1 M_1 M 1 ; since the action is free and proper and pr P \operatorname{pr}_P pr P is equivariant, that property descends to p 1 , G p_{1,G} p 1 , G . Hence
h = ( p 1 , G ∗ ) − 1 ( H G ∣ [ M 1 ] G ) h=(p_{1,G}^{*})^{-1}\left(H_G|_{[M_1]_G}\right) h = ( p 1 , G ∗ ) − 1 ( H G ∣ [ M 1 ] G )
is defined, and the vector component [ X ] G [X]_G [ X ] G is likewise a vector field along p 1 , G p_{1,G} p 1 , G , rather than on [ M 1 ] G [M_1]_G [ M 1 ] G . Since regularity gives only a local diffeomorphism, h h h is defined locally on each branch of the stationary control, that is, of the solution set of ∂ H G / ∂ u a = 0 \partial H_G/\partial u^a=0 ∂ H G / ∂ u a = 0 ; it is defined globally on T ∗ P / G T^{*}P/G T ∗ P / G only under the additional hypothesis that the stationary control is globally unique. All the statements below are to be read on a fixed such branch.
We now show that this implicit formulation is completely equivalent to the reduced Pontryagin Maximum Principle obtained from the reduced variational principle.
Theorem 4. Let ( P , C , Γ , L ) (P,C,\Gamma,L) ( P , C , Γ , L ) be a regular G G G -invariant optimal control problem. The equations of the reduced Pontryagin Maximum Principle obtained from the reduced Lagrange–Pontryagin variational principle coincide with the implicit equations determined by the Lie–Dirac reduction (44).
Proof. Let ( z ( t ) , p z ( t ) , μ ( t ) , u ( t ) ) (z(t),p_z(t),\mu(t),u(t)) ( z ( t ) , p z ( t ) , μ ( t ) , u ( t )) be a curve in [ M 1 ] G [M_1]_G [ M 1 ] G , on a fixed branch of the stationary control. In the splitting of § 4.3 , determined by ∇ \nabla ∇ and by the connection induced by A A A , the velocity of the reduced curve and the differential of h h h are represented by
[ X ] G = ( z ˙ , D p z D t ; ξ , D μ D t ) , ( π z , π p z ; α ξ , α v μ ) = ( ∂ h ∂ z , ∂ h ∂ p z ; 0 , ∂ h ∂ μ ) (45) [X]_G=\left(\dot z,\frac{Dp_z}{Dt};\xi,\frac{D\mu}{Dt}\right),\qquad
(\pi_z,\pi_{p_z};\alpha_\xi,\alpha_{v_\mu})=\left(\frac{\partial h}{\partial z},\frac{\partial h}{\partial p_z};0,\frac{\partial h}{\partial\mu}\right) \tag*{(45)} [ X ] G = ( z ˙ , D t D p z ; ξ , D t D μ ) , ( π z , π p z ; α ξ , α v μ ) = ( ∂ z ∂ h , ∂ p z ∂ h ; 0 , ∂ μ ∂ h ) (45)
where D p z / D t Dp_z/Dt D p z / D t , D μ / D t D\mu/Dt D μ / D t and the covariant partial derivative ∂ h / ∂ z \partial h/\partial z ∂ h / ∂ z are those fixed in § 3.4 , and the derivative with respect to the coadjoint variable μ \mu μ is paired with the variation v μ v_\mu v μ and therefore corresponds to α v μ \alpha_{v_\mu} α v μ , so that α ξ = 0 \alpha_\xi=0 α ξ = 0 . Substituting (45) into the four relations (42) gives, in turn, z ˙ = ∂ h / ∂ p z \dot z=\partial h/\partial p_z z ˙ = ∂ h / ∂ p z , D p z / D t = − ∂ h / ∂ z − ⟨ μ , F A ( z ˙ , ⋅ ) ⟩ Dp_z/Dt=-\partial h/\partial z-\langle\mu,F_A(\dot z,\cdot)\rangle D p z / D t = − ∂ h / ∂ z − ⟨ μ , F A ( z ˙ , ⋅ )⟩ , ξ = ∂ h / ∂ μ \xi=\partial h/\partial\mu ξ = ∂ h / ∂ μ and, since α ξ = 0 \alpha_\xi=0 α ξ = 0 , D μ / D t = − ad ξ ∗ μ D\mu/Dt=-\operatorname{ad}^*_\xi\mu D μ / D t = − ad ξ ∗ μ . Together with the primary constraint ∂ h / ∂ u = 0 \partial h/\partial u=0 ∂ h / ∂ u = 0 these are exactly the reduced Lagrange–Pontryagin equations (26) , the curvature term being written in the equivalent form noted after (26). □
Application: Optimal control of a rigid spacecraft with internal rotors
In this section, we illustrate the reduction procedure developed above by means of the optimal control of a rigid spacecraft equipped with three internal reaction wheels. The mechanical system is adapted from the classical model considered in [10 ] . It provides a particularly convenient example for our purposes, since the configuration manifold carries a natural S O ( 3 ) SO(3) S O ( 3 ) -symmetry, while the torques applied to the internal rotors provide natural control variables.
We first describe the mechanical system and its Hamiltonian formulation. We then regard the resulting Hamiltonian phase space as the state manifold of an optimal control problem. Thus, the configuration manifold of the mechanical system will be denoted by Q Q Q , whereas the state manifold entering the optimal control problem will be P = T ∗ Q P=T^*Q P = T ∗ Q . The reduction procedure developed in the previous sections is therefore applied to the induced S O ( 3 ) SO(3) S O ( 3 ) -action on P P P .
Mechanical model
The configuration manifold of the spacecraft–rotor system is
Q = S O ( 3 ) × ( S 1 ) 3 . Q=SO(3)\times(S^1)^3. Q = S O ( 3 ) × ( S 1 ) 3 .
where R ∈ S O ( 3 ) R\in SO(3) R ∈ S O ( 3 ) describes the attitude of the spacecraft and θ = ( θ 1 , θ 2 , θ 3 ) ∈ ( S 1 ) 3 \theta=(\theta_1,\theta_2,\theta_3)\in(S^1)^3 θ = ( θ 1 , θ 2 , θ 3 ) ∈ ( S 1 ) 3 denotes the angular positions of the three reaction wheels. The body angular velocity is Ω ^ = R − 1 R ˙ ∈ s o ( 3 ) \widehat{\Omega}=R^{-1}\dot R\in\mathfrak{so}(3) Ω = R − 1 R ˙ ∈ so ( 3 ) , written simply as Ω ∈ R 3 \Omega\in\mathbb{R}^3 Ω ∈ R 3 under the standard identification s o ( 3 ) ≃ R 3 \mathfrak{so}(3)\simeq\mathbb{R}^3 so ( 3 ) ≃ R 3 , and the angular velocity of the rotors relative to the spacecraft is Ω r = θ ˙ ∈ R 3 \Omega_r=\dot\theta\in\mathbb{R}^3 Ω r = θ ˙ ∈ R 3 .
Let I ∈ R 3 × 3 I\in\mathbb{R}^{3\times3} I ∈ R 3 × 3 be the inertia tensor of the spacecraft excluding the axial inertia of the reaction wheels, and let J = diag ( J 1 , J 2 , J 3 ) J=\operatorname{diag}(J_1,J_2,J_3) J = diag ( J 1 , J 2 , J 3 ) be the inertia tensor of the three rotors about their rotation axes. The mechanical Lagrangian is taken to be
L ( R , θ , Ω , Ω r ) = 1 2 Ω T I Ω + 1 2 ( Ω + Ω r ) T J ( Ω + Ω r ) . (46) \mathcal{L}(R,\theta,\Omega,\Omega_r)=\frac{1}{2}\Omega^T I\Omega+\frac{1}{2}(\Omega+\Omega_r)^T J(\Omega+\Omega_r). \tag*{(46)} L ( R , θ , Ω , Ω r ) = 2 1 Ω T I Ω + 2 1 ( Ω + Ω r ) T J ( Ω + Ω r ) . (46)
In particular, L \mathcal{L} L is independent of R R R and θ \theta θ when expressed in body variables.
The natural symmetry of the rigid-body system is the left multiplication of S O ( 3 ) SO(3) S O ( 3 ) on the attitude variable, that is, the left action
A ⋅ ( R , θ ) = ( A R , θ ) , A ∈ S O ( 3 ) . (47) A\cdot(R,\theta)=(AR,\theta),\qquad A\in SO(3). \tag*{(47)} A ⋅ ( R , θ ) = ( A R , θ ) , A ∈ S O ( 3 ) . (47)
This is a smooth, free and proper left action. The general reduction theory of Sections 3.3 and 4 was formulated for right actions; the corresponding statements for a left action are those of Remarks 1 and 7, and it is these that we use throughout the present section. With this choice the product connection on P ≃ S O ( 3 ) × B P\simeq SO(3)\times B P ≃ S O ( 3 ) × B is normalized by A ( ζ P ) = ζ A(\zeta_P)=\zeta A ( ζ P ) = ζ with ζ P ( R ) = ζ R \zeta_P(R)=\zeta R ζ P ( R ) = ζ R , so that A ( R ˙ ) = R ˙ R − 1 A(\dot R)=\dot R R^{-1} A ( R ˙ ) = R ˙ R − 1 , trivializing the adjoint bundle by the section R = 1 R=1 R = 1 then gives ξ = Ad R − 1 ( R ˙ R − 1 ) = R − 1 R ˙ \xi=\operatorname{Ad}_{R^{-1}}(\dot R R^{-1})=R^{-1}\dot R ξ = Ad R − 1 ( R ˙ R − 1 ) = R − 1 R ˙ and, dually, μ = Π \mu=\Pi μ = Π .
Indeed, under (47) one has R ↦ A R R\mapsto AR R ↦ A R and R ˙ ↦ A R ˙ \dot R\mapsto A\dot R R ˙ ↦ A R ˙ , so that ( A R ) − 1 ( A R ˙ ) = R − 1 R ˙ (AR)^{-1}(A\dot R)=R^{-1}\dot R ( A R ) − 1 ( A R ˙ ) = R − 1 R ˙ . The body angular velocity Ω \Omega Ω is therefore invariant under the action, as is Ω r = θ ˙ \Omega_r=\dot\theta Ω r = θ ˙ , and consequently L ( A R , θ , Ω , Ω r ) = L ( R , θ , Ω , Ω r ) \mathcal{L}(AR,\theta,\Omega,\Omega_r)=\mathcal{L}(R,\theta,\Omega,\Omega_r) L ( A R , θ , Ω , Ω r ) = L ( R , θ , Ω , Ω r ) : the mechanical Lagrangian is S O ( 3 ) SO(3) S O ( 3 ) -invariant. The quotient of the configuration manifold is Q / S O ( 3 ) ≃ ( S 1 ) 3 Q/SO(3)\simeq(S^1)^3 Q / S O ( 3 ) ≃ ( S 1 ) 3 .
Hamiltonian formulation
We now pass to the Hamiltonian description of the spacecraft. The momenta conjugate to Ω \Omega Ω and Ω r \Omega_r Ω r are
m = ∂ L ∂ Ω = I Ω + J ( Ω + Ω r ) , (48) m=\frac{\partial\mathcal{L}}{\partial\Omega}=I\Omega+J(\Omega+\Omega_r), \tag*{(48)} m = ∂ Ω ∂ L = I Ω + J ( Ω + Ω r ) , (48)
and
l = ∂ L ∂ Ω r = J ( Ω + Ω r ) . (49) l=\frac{\partial\mathcal{L}}{\partial\Omega_r}=J(\Omega+\Omega_r). \tag*{(49)} l = ∂ Ω r ∂ L = J ( Ω + Ω r ) . (49)
It follows that
Ω = I − 1 ( m − l ) , Ω r = J − 1 l − Ω . (50) \Omega= I^{-1}(m-l), \qquad\Omega_r = J^{-1}l-\Omega. \tag*{(50)} Ω = I − 1 ( m − l ) , Ω r = J − 1 l − Ω. (50)
Assuming that I I I and J J J are positive definite, the Legendre transformation is hyperregular. We may therefore work on the Hamiltonian phase space P = T ∗ Q P=T^*Q P = T ∗ Q .
Using the body trivialization of T ∗ S O ( 3 ) T^*SO(3) T ∗ S O ( 3 ) , we identify
P ≃ S O ( 3 ) × ( S 1 ) 3 × s o ( 3 ) ∗ × R 3 , P \simeq SO(3)\times(S^1)^3\times\mathfrak{so}(3)^*\times\mathbb{R}^3, P ≃ S O ( 3 ) × ( S 1 ) 3 × so ( 3 ) ∗ × R 3 ,
and write a point of P P P as x = ( R , θ , m , l ) x=(R,\theta,m,l) x = ( R , θ , m , l ) .
The Hamiltonian is obtained by the Legendre transformation,
H = ⟨ m , Ω ⟩ + ⟨ l , Ω r ⟩ − L . H=\langle m,\Omega\rangle+\langle l,\Omega_r\rangle-\mathcal{L}. H = ⟨ m , Ω ⟩ + ⟨ l , Ω r ⟩ − L .
Using (50), it becomes
H ( R , θ , m , l ) = 1 2 ( m − l ) T I − 1 ( m − l ) + 1 2 l T J − 1 l . (51) H(R,\theta,m,l)=\frac{1}{2}(m-l)^T I^{-1}(m-l)+\frac{1}{2}l^T J^{-1}l. \tag*{(51)} H ( R , θ , m , l ) = 2 1 ( m − l ) T I − 1 ( m − l ) + 2 1 l T J − 1 l . (51)
The corresponding Hamiltonian equations are
R ˙ = R Ω ^ , θ ˙ = Ω r , m ˙ = m × Ω , l ˙ = 0 , (52) \begin{aligned}
\dot{R}&=R\widehat{\Omega},\\
\dot{\theta}&=\Omega_r,\\
\dot{m}&=m\times\Omega,\\
\dot{l}&=0, \tag*{(52)}
\end{aligned} R ˙ θ ˙ m ˙ l ˙ = R Ω , = Ω r , = m × Ω , = 0 , (52)
where
Ω = I − 1 ( m − l ) , Ω r = J − 1 l − I − 1 ( m − l ) . (53) \Omega=I^{-1}(m-l), \qquad\Omega_r=J^{-1}l-I^{-1}(m-l). \tag*{(53)} Ω = I − 1 ( m − l ) , Ω r = J − 1 l − I − 1 ( m − l ) . (53)
The induced symmetry on the state manifold
The S O ( 3 ) SO(3) S O ( 3 ) -action (47) on Q Q Q induces, by cotangent lift, a free and proper action on P = T ∗ Q P=T^*Q P = T ∗ Q .
In the body trivialization introduced above, the action is
A ⋅ ( R , θ , m , l ) = ( A R , θ , m , l ) . (54) A\cdot(R,\theta,m,l)=(AR,\theta,m,l). \tag*{(54)} A ⋅ ( R , θ , m , l ) = ( A R , θ , m , l ) . (54)
Thus the reduced state manifold is
B = P / S O ( 3 ) ≃ ( S 1 ) 3 × s o ( 3 ) ∗ × R 3 . (55) B=P/SO(3)\simeq(S^1)^3\times\mathfrak{so}(3)^*\times\mathbb{R}^3. \tag*{(55)} B = P / S O ( 3 ) ≃ ( S 1 ) 3 × so ( 3 ) ∗ × R 3 . (55)
A point of B B B is written z = ( θ , m , l ) z=(\theta,m,l) z = ( θ , m , l ) . Notice that it is B = P / G B=P/G B = P / G , and not Q / G Q/G Q / G , that is the base manifold entering the reduction of the optimal control problem developed in the previous sections.
The principal bundle ρ P : P → B \rho_P:P\to B ρ P : P → B is trivial, and we choose the principal connection associated with the product decomposition P ≃ S O ( 3 ) × B P\simeq SO(3)\times B P ≃ S O ( 3 ) × B . The adjoint and coadjoint bundles are then trivial,
s o ~ ( 3 ) ≃ B × s o ( 3 ) , s o ~ ( 3 ) ∗ ≃ B × s o ( 3 ) ∗ , \widetilde{\mathfrak{so}}(3)\simeq B\times\mathfrak{so}(3), \qquad\widetilde{\mathfrak{so}}(3)^*\simeq B\times\mathfrak{so}(3)^*, so ( 3 ) ≃ B × so ( 3 ) , so ( 3 ) ∗ ≃ B × so ( 3 ) ∗ ,
and the connection is flat, F A = 0 F_A=0 F A = 0 .
Accordingly, the general identifications obtained in the previous sections become
T P / S O ( 3 ) ≃ T B ⊕ s o ~ ( 3 ) (56) TP/SO(3)\simeq TB\oplus\widetilde{\mathfrak{so}}(3) \tag*{(56)} T P / S O ( 3 ) ≃ T B ⊕ so ( 3 ) (56)
and
T ∗ P / S O ( 3 ) ≃ T ∗ B ⊕ s o ~ ( 3 ) ∗ . (57) T^*P/SO(3)\simeq T^*B\oplus\widetilde{\mathfrak{so}}(3)^*. \tag*{(57)} T ∗ P / S O ( 3 ) ≃ T ∗ B ⊕ so ( 3 ) ∗ . (57)
These are precisely the reduced tangent and cotangent bundles entering the reduction procedure developed above.
Optimal control extension
We now introduce controls through torques applied to the three internal reaction wheels. The control space is U = R 3 U=\mathbb{R}^3 U = R 3 with the trivial S O ( 3 ) SO(3) S O ( 3 ) -action, so the control bundle is C = P × U → P C=P\times U\to P C = P × U → P and a control is written u = ( u 1 , u 2 , u 3 ) ∈ R 3 u=(u_1,u_2,u_3)\in\mathbb{R}^3 u = ( u 1 , u 2 , u 3 ) ∈ R 3 . The applied torques modify the rotor momentum equation to l ˙ = u \dot{l}=u l ˙ = u . Thus the controlled dynamics is described by the bundle map Γ : C = P × U → T P \Gamma:C=P\times U\to TP Γ : C = P × U → T P given by
Γ ( R , θ , m , l , u ) = ( R Ω ^ , Ω r , m × Ω , u ) . (58) \Gamma(R,\theta,m,l,u)=\left(R\widehat{\Omega},\Omega_r,m\times\Omega,u\right). \tag*{(58)} Γ ( R , θ , m , l , u ) = ( R Ω , Ω r , m × Ω , u ) . (58)
where Ω = I − 1 ( m − l ) \Omega= I^{-1}(m-l) Ω = I − 1 ( m − l ) and Ω r = J − 1 l − Ω \Omega_r = J^{-1}l-\Omega Ω r = J − 1 l − Ω . Equivalently, the controlled equations are
R ˙ = R Ω ^ , θ ˙ = J − 1 l − Ω , m ˙ = m × Ω , l ˙ = u . (59) \begin{aligned}
\dot{R} &= R\widehat{\Omega},\\
\dot{\theta} &= J^{-1}l-\Omega,\\
\dot{m} &= m\times\Omega,\\
\dot{l} &= u. \tag*{(59)}
\end{aligned} R ˙ θ ˙ m ˙ l ˙ = R Ω , = J − 1 l − Ω , = m × Ω , = u . (59)
Introducing the input vector fields Y i = ∂ / ∂ l i Y_i=\partial/\partial l_i Y i = ∂ / ∂ l i for i = 1 , 2 , 3 i=1,2,3 i = 1 , 2 , 3 , the controlled dynamics is equivalently Γ ( x , u ) = X H ( x ) + ∑ i = 1 3 u i Y i ( x ) \Gamma(x,u)=X_H(x)+\sum_{i=1}^{3}u_iY_i(x) Γ ( x , u ) = X H ( x ) + ∑ i = 1 3 u i Y i ( x ) , which is control-affine with drift X H X_H X H .
Proposition 9. The bundle map Γ : P × U → T P \Gamma:P\times U\to TP Γ : P × U → T P is S O ( 3 ) \mathrm{SO}(3) SO ( 3 ) -equivariant.
Proof. The Hamiltonian (see (51) ) is invariant under the S O ( 3 ) \mathrm{SO}(3) SO ( 3 ) -action (see (54) ), so its Hamiltonian vector field X H X_H X H is equivariant. The input vector fields Y i = ∂ / ∂ l i Y_i=\partial/\partial l_i Y i = ∂ / ∂ l i are invariant, since the lifted action leaves the rotor momentum variables l i l_i l i unchanged. The action on U U U is trivial, so T Φ A P ∘ Γ = Γ ∘ Ψ A T\Phi_A^P\circ\Gamma=\Gamma\circ\Psi_A T Φ A P ∘ Γ = Γ ∘ Ψ A for every A ∈ S O ( 3 ) A\in\mathrm{SO}(3) A ∈ SO ( 3 ) , with Ψ A ( x , u ) = ( A ⋅ x , u ) \Psi_A(x,u)=(A\cdot x,u) Ψ A ( x , u ) = ( A ⋅ x , u ) . Hence Γ \Gamma Γ is S O ( 3 ) \mathrm{SO}(3) SO ( 3 ) -equivariant. □
We consider the quadratic running cost
L : P × U ⟶ R , L ( x , u ) = 1 2 ∣ u ∣ 2 . (60) L:P\times U\longrightarrow\mathbb{R},\qquad L(x,u)=\frac{1}{2}\lvert u\rvert^2. \tag*{(60)} L : P × U ⟶ R , L ( x , u ) = 2 1 ∣ u ∣ 2 . (60)
Since the S O ( 3 ) \mathrm{SO}(3) SO ( 3 ) -action on U U U is trivial, the running cost is S O ( 3 ) \mathrm{SO}(3) SO ( 3 ) -invariant. We have therefore obtained an S O ( 3 ) \mathrm{SO}(3) SO ( 3 ) -invariant optimal control problem
( P , C , Γ , L ) , P = T ∗ Q , C = P × R 3 . (P,C,\Gamma,L),\qquad P=T^*Q,\qquad C=P\times\mathbb{R}^3. ( P , C , Γ , L ) , P = T ∗ Q , C = P × R 3 .
Pontryagin formulation
We now construct the Pontryagin bundle associated with this optimal control problem. Since the state manifold is P = T ∗ Q P=T^*Q P = T ∗ Q , the Pontryagin bundle is
M = T ∗ P × P C = T ∗ ( T ∗ Q ) × T ∗ Q ( T ∗ Q × R 3 ) . (61) M=T^*P\times_P C=T^*(T^*Q)\times_{T^*Q}(T^*Q\times\mathbb{R}^3). \tag*{(61)} M = T ∗ P × P C = T ∗ ( T ∗ Q ) × T ∗ Q ( T ∗ Q × R 3 ) . (61)
A costate λ ∈ T x ∗ P \lambda\in T_x^*P λ ∈ T x ∗ P is written, in the trivialization introduced above, as λ = ( Π , p θ , p m , p l ) \lambda=(\Pi,p_\theta,p_m,p_l) λ = ( Π , p θ , p m , p l ) , where
Π ∈ s o ( 3 ) ∗ , p θ , p m , p l ∈ R 3 . \Pi\in\mathfrak{so}(3)^*,\qquad p_\theta,p_m,p_l\in\mathbb{R}^3. Π ∈ so ( 3 ) ∗ , p θ , p m , p l ∈ R 3 .
The Pontryagin Hamiltonian is
H P ( x , λ , u ) = ⟨ λ , Γ ( x , u ) ⟩ − 1 2 ∣ u ∣ 2 . (62) H_P(x,\lambda,u)=\langle\lambda,\Gamma(x,u)\rangle-\frac{1}{2}\lvert u\rvert^2. \tag*{(62)} H P ( x , λ , u ) = ⟨ λ , Γ ( x , u )⟩ − 2 1 ∣ u ∣ 2 . (62)
Using (58) , this becomes
H P = ⟨ Π , Ω ⟩ + ⟨ p θ , J − 1 l − Ω ⟩ + ⟨ p m , m × Ω ⟩ + ⟨ p l , u ⟩ − 1 2 ∣ u ∣ 2 , (63) \begin{aligned}
H_P={}&\langle\Pi,\Omega\rangle+\langle p_\theta,J^{-1}l-\Omega\rangle+\langle p_m,m\times\Omega\rangle\\
&+\langle p_l,u\rangle-\frac{1}{2}\lvert u\rvert^2, \tag*{(63)}
\end{aligned} H P = ⟨ Π , Ω ⟩ + ⟨ p θ , J − 1 l − Ω ⟩ + ⟨ p m , m × Ω ⟩ + ⟨ p l , u ⟩ − 2 1 ∣ u ∣ 2 , (63)
where Ω = I − 1 ( m − l ) \Omega=I^{-1}(m-l) Ω = I − 1 ( m − l ) . The primary constraint ∂ H P / ∂ u = 0 \partial H_P/\partial u=0 ∂ H P / ∂ u = 0 gives
u = p l , (64) u=p_l, \tag*{(64)} u = p l , (64)
and ∂ 2 H P / ∂ u 2 = − I d R 3 \partial^2H_P/\partial u^2=-\mathrm{Id}_{\mathbb{R}^3} ∂ 2 H P / ∂ u 2 = − Id R 3 is nondegenerate, so the optimal control problem is regular.
Reduction of the optimal control problem
Since the action on U = R 3 U=\mathbb{R}^3 U = R 3 is trivial, the reduced control bundle is U ~ = C / S O ( 3 ) ≅ B × R 3 \widetilde{U}=C/\mathrm{SO}(3)\cong B\times\mathbb{R}^3 U = C / SO ( 3 ) ≅ B × R 3 , a vector bundle, and an element of it is written ( z , u ) = ( θ , m , l , u ) (z,u)=(\theta,m,l,u) ( z , u ) = ( θ , m , l , u ) . The quotient of the Pontryagin bundle is, by the general construction,
M / S O ( 3 ) ≅ T ∗ B ⊕ s o ( 3 ) ∗ ⊕ U ~ . (65) M/\mathrm{SO}(3)\cong T^*B\oplus\mathfrak{so}(3)^*\oplus\widetilde{U}. \tag*{(65)} M / SO ( 3 ) ≅ T ∗ B ⊕ so ( 3 ) ∗ ⊕ U . (65)
All the bundles appearing here are trivial in the present example, so that M / S O ( 3 ) ≅ T ∗ B ⊕ ( B × s o ( 3 ) ∗ ) ⊕ ( B × R 3 ) M/\mathrm{SO}(3)\cong T^*B\oplus(B\times\mathfrak{so}(3)^*)\oplus(B\times\mathbb{R}^3) M / SO ( 3 ) ≅ T ∗ B ⊕ ( B × so ( 3 ) ∗ ) ⊕ ( B × R 3 ) .
The equivariant controlled vector field Γ \Gamma Γ induces a reduced bundle map [ Γ ] : U ~ → T P / S O ( 3 ) [\Gamma]:\widetilde{U}\to TP/\mathrm{SO}(3) [ Γ ] : U → T P / SO ( 3 ) , which the identification (56) decomposes as [ Γ ] ( z , u ) = ( Γ ~ ( z , u ) , γ ~ ( z , u ) ) [\Gamma](z,u)=(\widetilde{\Gamma}(z,u),\widetilde{\gamma}(z,u)) [ Γ ] ( z , u ) = ( Γ ( z , u ) , γ ( z , u )) . For z = ( θ , m , l ) z=(\theta,m,l) z = ( θ , m , l ) , the reduced base dynamics is
Γ ~ ( θ , m , l , u ) = ( J − 1 l − Ω , m × Ω , u ) , (66) \widetilde{\Gamma}(\theta,m,l,u)=(J^{-1}l-\Omega,m\times\Omega,u), \tag*{(66)} Γ ( θ , m , l , u ) = ( J − 1 l − Ω , m × Ω , u ) , (66)
whereas the vertical component associated with the eliminated attitude variable is
γ ~ ( θ , m , l , u ) = Ω = I − 1 ( m − l ) . (67) \widetilde{\gamma}(\theta,m,l,u)=\Omega=I^{-1}(m-l). \tag*{(67)} γ ( θ , m , l , u ) = Ω = I − 1 ( m − l ) . (67)
Reduced variational formulation
Reduced Lagrange–Pontryagin principle
We now perform explicitly the reduction of the Lagrange–Pontryagin variational principle for the spacecraft optimal control problem. Recall that the unreduced Lagrange–Pontryagin functional is
S = ∫ 0 T [ L ( x , u ) + ⟨ λ , x ˙ − Γ ( x , u ) ⟩ ] d t , (68) S = \int_{0}^{T} \left[L(x,u) + \langle\lambda,\dot{x} - \Gamma(x,u)\rangle\right] \, dt, \tag*{(68)} S = ∫ 0 T [ L ( x , u ) + ⟨ λ , x ˙ − Γ ( x , u )⟩ ] d t , (68)
with x = ( R , θ , m , l ) ∈ P x=(R,\theta,m,l)\in P x = ( R , θ , m , l ) ∈ P , λ ∈ T x ∗ P \lambda\in T_x^*P λ ∈ T x ∗ P and u ∈ R 3 u\in\mathbb{R}^3 u ∈ R 3 . Equivalently, in Hamilton–Pontryagin form,
S = ∫ 0 T [ ⟨ λ , x ˙ ⟩ − H P ( x , λ , u ) ] d t . (69) S = \int_{0}^{T} \left[\langle\lambda,\dot{x}\rangle- H_P(x,\lambda,u)\right] \, dt. \tag*{(69)} S = ∫ 0 T [ ⟨ λ , x ˙ ⟩ − H P ( x , λ , u ) ] d t . (69)
The functional is reduced using the decomposition T P / S O ( 3 ) ≅ T B ⊕ s o ~ ( 3 ) TP/SO(3)\cong TB\oplus\widetilde{\mathfrak{so}}(3) T P / S O ( 3 ) ≅ T B ⊕ so ( 3 ) determined by the principal connection introduced above. A curve x ( t ) = ( R ( t ) , θ ( t ) , m ( t ) , l ( t ) ) x(t)=(R(t),\theta(t),m(t),l(t)) x ( t ) = ( R ( t ) , θ ( t ) , m ( t ) , l ( t )) in P P P has reduced base curve z ( t ) = ( θ ( t ) , m ( t ) , l ( t ) ) ∈ B z(t)=(\theta(t),m(t),l(t))\in B z ( t ) = ( θ ( t ) , m ( t ) , l ( t )) ∈ B , and its velocity x ˙ \dot{x} x ˙ determines the reduced pair ( z ˙ , ξ ) ∈ T B ⊕ s o ~ ( 3 ) (\dot{z},\xi)\in TB\oplus\widetilde{\mathfrak{so}}(3) ( z ˙ , ξ ) ∈ T B ⊕ so ( 3 ) , where ξ \xi ξ is the vertical component carried by the attitude variable. Dually, the costate is represented by ( p z , μ ) ∈ T ∗ B ⊕ s o ~ ( 3 ) ∗ (p_z,\mu)\in T^*B\oplus\widetilde{\mathfrak{so}}(3)^* ( p z , μ ) ∈ T ∗ B ⊕ so ( 3 ) ∗ with p z = ( p θ , p m , p l ) p_z=(p_\theta,p_m,p_l) p z = ( p θ , p m , p l ) .
For the spacecraft the two components of the reduced controlled vector field are
γ ~ ( z , u ) = Ω = I − 1 ( m − l ) , Γ ~ ( z , u ) = ( J − 1 l − Ω , m × Ω , u ) . \widetilde{\gamma}(z,u)=\Omega=I^{-1}(m-l),\qquad
\widetilde{\Gamma}(z,u)=\left(J^{-1}l-\Omega,\ m\times\Omega,\ u\right). γ ( z , u ) = Ω = I − 1 ( m − l ) , Γ ( z , u ) = ( J − 1 l − Ω , m × Ω , u ) .
Under these identifications the pairing splits,
⟨ λ , x ˙ − Γ ( x , u ) ⟩ = ⟨ p z , z ˙ − Γ ~ ( z , u ) ⟩ + ⟨ μ , ξ − γ ~ ( z , u ) ⟩ , \langle\lambda,\dot{x}-\Gamma(x,u)\rangle
=
\left\langle p_z,\dot{z}-\widetilde{\Gamma}(z,u)\right\rangle
+
\left\langle\mu,\xi-\widetilde{\gamma}(z,u)\right\rangle, ⟨ λ , x ˙ − Γ ( x , u )⟩ = ⟨ p z , z ˙ − Γ ( z , u ) ⟩ + ⟨ μ , ξ − γ ( z , u ) ⟩ ,
so that (68) reduces to the density (21) of the general theory, whose running-cost term is here 1 2 ∣ u ∣ 2 \frac{1}{2}|u|^2 2 1 ∣ u ∣ 2 . (In (21) and (24) the symbol l l l denotes the reduced running cost; in the present section l l l is the rotor momentum, a component of z = ( θ , m , l ) z=(\theta,m,l) z = ( θ , m , l ) .) Written out,
S G = ∫ 0 T [ 1 2 ∣ u ∣ 2 + ⟨ p θ , θ ˙ − J − 1 l + Ω ⟩ + ⟨ p m , m ˙ − m × Ω ⟩ + ⟨ p l , l ˙ − u ⟩ + ⟨ μ , ξ − Ω ⟩ ] d t , (70) S_G=\int_{0}^{T}\left[
\frac{1}{2}|u|^2
+\left\langle p_\theta,\dot{\theta}-J^{-1}l+\Omega\right\rangle
+\left\langle p_m,\dot{m}-m\times\Omega\right\rangle
+\left\langle p_l,\dot{l}-u\right\rangle
+\left\langle\mu,\xi-\Omega\right\rangle
\right]\,dt, \tag*{(70)} S G = ∫ 0 T [ 2 1 ∣ u ∣ 2 + ⟨ p θ , θ ˙ − J − 1 l + Ω ⟩ + ⟨ p m , m ˙ − m × Ω ⟩ + ⟨ p l , l ˙ − u ⟩ + ⟨ μ , ξ − Ω ⟩ ] d t , (70)
where Ω = I − 1 ( m − l ) \Omega=I^{-1}(m-l) Ω = I − 1 ( m − l ) throughout. Equivalently, the reduced Pontryagin Hamiltonian (24) is here
h = ⟨ p θ , J − 1 l − Ω ⟩ + ⟨ p m , m × Ω ⟩ + ⟨ p l , u ⟩ + ⟨ μ , Ω ⟩ − 1 2 ∣ u ∣ 2 , (71) h=\left\langle p_\theta,J^{-1}l-\Omega\right\rangle
+\left\langle p_m,m\times\Omega\right\rangle
+\left\langle p_l,u\right\rangle
+\langle\mu,\Omega\rangle-\frac{1}{2}|u|^2, \tag*{(71)} h = ⟨ p θ , J − 1 l − Ω ⟩ + ⟨ p m , m × Ω ⟩ + ⟨ p l , u ⟩ + ⟨ μ , Ω ⟩ − 2 1 ∣ u ∣ 2 , (71)
and hence
S G = ∫ 0 T [ ⟨ p z , z ˙ ⟩ + ⟨ μ , ξ ⟩ − h ( z , p z , μ , u ) ] d t . (72) S_G=\int_{0}^{T}\left[\langle p_z,\dot{z}\rangle+\langle\mu,\xi\rangle-h(z,p_z,\mu,u)\right]\,dt. \tag*{(72)} S G = ∫ 0 T [ ⟨ p z , z ˙ ⟩ + ⟨ μ , ξ ⟩ − h ( z , p z , μ , u ) ] d t . (72)
Reduced variations
We now derive the reduced optimality equations directly from (72), rather than by quoting the general result of Section 3.3.
The variations δ z = ( δ θ , δ m , δ l ) \delta z=(\delta\theta,\delta m,\delta l) δ z = ( δ θ , δ m , δ l ) of the base variables are arbitrary and vanish at the endpoints, δ z ( 0 ) = δ z ( T ) = 0 \delta z(0)=\delta z(T)=0 δ z ( 0 ) = δ z ( T ) = 0 ; the variations δ p θ \delta p_\theta δ p θ , δ p m \delta p_m δ p m , δ p l \delta p_l δ p l , δ μ \delta\mu δ μ and δ u \delta u δ u are arbitrary and unconstrained.
The variation of the vertical variable ξ \xi ξ , however, is constrained by the reduction. With the convention adopted in the general reduction theory, it is of the form
δ ξ = D η D t + [ ξ , η ] + F A ( δ z , z ˙ ) , (73) \delta\xi=\frac{D\eta}{Dt}+[\xi,\eta]+F_A(\delta z,\dot{z}), \tag*{(73)} δ ξ = D t D η + [ ξ , η ] + F A ( δ z , z ˙ ) , (73)
where η ( t ) ∈ s o ( 3 ) \eta(t)\in\mathfrak{so}(3) η ( t ) ∈ so ( 3 ) is arbitrary and vanishes at the endpoints. The product connection chosen in the present example is flat, F A = 0 F_A=0 F A = 0 , and therefore
δ ξ = η ˙ + [ ξ , η ] . (74) \delta\xi=\dot{\eta}+[\xi,\eta]. \tag*{(74)} δ ξ = η ˙ + [ ξ , η ] . (74)
Taking the variation of (72) gives
δ S G = ∫ 0 T [ ⟨ δ p z , z ˙ ⟩ + ⟨ p z , δ z ˙ ⟩ + ⟨ δ μ , ξ ⟩ + ⟨ μ , δ ξ ⟩ − ⟨ ∂ h ∂ z , δ z ⟩ − ⟨ ∂ h ∂ p z , δ p z ⟩ − ⟨ ∂ h ∂ μ , δ μ ⟩ − ⟨ ∂ h ∂ u , δ u ⟩ ] d t . (75) \delta S_G=\int_{0}^{T}\left[
\langle\delta p_z,\dot{z}\rangle
+\langle p_z,\delta\dot{z}\rangle
+\langle\delta\mu,\xi\rangle
+\langle\mu,\delta\xi\rangle
-\left\langle\frac{\partial h}{\partial z},\delta z\right\rangle
-\left\langle\frac{\partial h}{\partial p_z},\delta p_z\right\rangle
-\left\langle\frac{\partial h}{\partial\mu},\delta\mu\right\rangle
-\left\langle\frac{\partial h}{\partial u},\delta u\right\rangle
\right]\,dt. \tag*{(75)} δ S G = ∫ 0 T [ ⟨ δ p z , z ˙ ⟩ + ⟨ p z , δ z ˙ ⟩ + ⟨ δ μ , ξ ⟩ + ⟨ μ , δ ξ ⟩ − ⟨ ∂ z ∂ h , δ z ⟩ − ⟨ ∂ p z ∂ h , δ p z ⟩ − ⟨ ∂ μ ∂ h , δ μ ⟩ − ⟨ ∂ u ∂ h , δ u ⟩ ] d t . (75)
Two integrations by parts remove the derivatives from the variations. The δ z \delta z δ z term gives ∫ 0 T ⟨ p z , δ z ˙ ⟩ d t = − ∫ 0 T ⟨ p ˙ z , δ z ⟩ d t \int_0^T \langle p_z,\delta\dot z\rangle\,dt=-\int_0^T\langle\dot p_z,\delta z\rangle\,dt ∫ 0 T ⟨ p z , δ z ˙ ⟩ d t = − ∫ 0 T ⟨ p ˙ z , δ z ⟩ d t , the boundary term vanishing with δ z \delta z δ z . For the vertical term, substitute (74) , integrate the η \eta η term by parts using η ( 0 ) = η ( T ) = 0 \eta(0)=\eta(T)=0 η ( 0 ) = η ( T ) = 0 , and use the coadjoint convention of this paper, ⟨ μ , [ ξ , η ] ⟩ = ⟨ ad ξ ∗ μ , η ⟩ \langle\mu,[\xi,\eta]\rangle=\langle\operatorname{ad}^{*}_{\xi}\mu,\eta\rangle ⟨ μ , [ ξ , η ]⟩ = ⟨ ad ξ ∗ μ , η ⟩ :
∫ 0 T ⟨ μ , δ ξ ⟩ d t = − ∫ 0 T ⟨ D μ D t − ad ξ ∗ μ , η ⟩ d t . \int_0^T\langle\mu,\delta\xi\rangle\,dt=-\int_0^T\left\langle\frac{D\mu}{Dt}-\operatorname{ad}^{*}_{\xi}\mu,\eta\right\rangle\,dt. ∫ 0 T ⟨ μ , δ ξ ⟩ d t = − ∫ 0 T ⟨ D t D μ − ad ξ ∗ μ , η ⟩ d t .
Since all the remaining variations are arbitrary, the condition δ S G = 0 \delta S_G=0 δ S G = 0 yields
z ˙ = ∂ h ∂ p z , p ˙ z = − ∂ h ∂ z , D μ D t = ad ξ ∗ μ , ξ = ∂ h ∂ μ , 0 = ∂ h ∂ u . (76) \begin{aligned}
\dot z&=\frac{\partial h}{\partial p_z},\\
\dot p_z&=-\frac{\partial h}{\partial z},\\
\frac{D\mu}{Dt}&=\operatorname{ad}^{*}_{\xi}\mu,\\
\xi&=\frac{\partial h}{\partial\mu},\\
0&=\frac{\partial h}{\partial u}. \tag*{(76)}
\end{aligned} z ˙ p ˙ z D t D μ ξ 0 = ∂ p z ∂ h , = − ∂ z ∂ h , = ad ξ ∗ μ , = ∂ μ ∂ h , = ∂ u ∂ h . (76)
This is (26) with F A = 0 F_A=0 F A = 0 and with the sign of the coadjoint equation reversed as in (27) , obtained here directly from the reduced variational principle rather than by quoting the general reduced Pontryagin Maximum Principle. The agreement is the first of the two independent checks this section performs on the general theory; the second is in Section 5.8.
Reduced optimality equations
We now compute the equations in (76) explicitly. Variations with respect to the costates p θ p_\theta p θ , p m p_m p m and p l p_l p l give z ˙ = ∂ h / ∂ p z \dot z=\partial h/\partial p_z z ˙ = ∂ h / ∂ p z , that is,
θ ˙ = J − 1 l − Ω , m ˙ = m × Ω , Ω = I − 1 ( m − l ) , l ˙ = u . (77) \begin{aligned}
\dot\theta&=J^{-1}l-\Omega,\\
\dot m&=m\times\Omega,\qquad\Omega=I^{-1}(m-l),\\
\dot l&=u. \tag*{(77)}
\end{aligned} θ ˙ m ˙ l ˙ = J − 1 l − Ω , = m × Ω , Ω = I − 1 ( m − l ) , = u . (77)
Variation with respect to μ \mu μ gives ξ = ∂ h / ∂ μ \xi=\partial h/\partial\mu ξ = ∂ h / ∂ μ , that is,
ξ = Ω = I − 1 ( m − l ) , (78) \xi=\Omega=I^{-1}(m-l), \tag*{(78)} ξ = Ω = I − 1 ( m − l ) , (78)
and variation with respect to the control gives 0 = ∂ h / ∂ u = p l − u 0=\partial h/\partial u=p_l-u 0 = ∂ h / ∂ u = p l − u , so that the optimal control is
u = p l . (79) u=p_l. \tag*{(79)} u = p l . (79)
Here ∂ 2 h / ∂ u 2 = − Id R 3 \partial^2h/\partial u^2=-\operatorname{Id}_{\mathbb{R}^3} ∂ 2 h / ∂ u 2 = − Id R 3 , so the problem is regular and the stationarity condition determines the control uniquely.
The vertical variation gives the reduced coadjoint equation
D μ D t = ad Ω ∗ μ , equivalently μ ˙ = μ × Ω . (80) \frac{D\mu}{Dt}=\operatorname{ad}^{*}_{\Omega}\mu,\qquad\text{equivalently}\qquad\dot\mu=\mu\times\Omega. \tag*{(80)} D t D μ = ad Ω ∗ μ , equivalently μ ˙ = μ × Ω. (80)
It remains to determine the equations for the base costates. Since
h = ⟨ p θ , J − 1 l − Ω ⟩ + ⟨ p m , m × Ω ⟩ + ⟨ p l , u ⟩ + ⟨ μ , Ω ⟩ − 1 2 ∣ u ∣ 2 , h=\langle p_\theta,J^{-1}l-\Omega\rangle+\langle p_m,m\times\Omega\rangle+\langle p_l,u\rangle+\langle\mu,\Omega\rangle-\frac{1}{2}|u|^2, h = ⟨ p θ , J − 1 l − Ω ⟩ + ⟨ p m , m × Ω ⟩ + ⟨ p l , u ⟩ + ⟨ μ , Ω ⟩ − 2 1 ∣ u ∣ 2 ,
and h h h is independent of θ \theta θ , we immediately obtain
p ˙ θ = 0. (81) \dot p_\theta=0. \tag*{(81)} p ˙ θ = 0. (81)
To compute the remaining derivatives, observe that
δ Ω = I − 1 ( δ m − δ l ) . (82) \delta\Omega=I^{-1}(\delta m-\delta l). \tag*{(82)} δ Ω = I − 1 ( δ m − δ l ) . (82)
Throughout, s o ( 3 ) ∗ ≃ R 3 \mathfrak{so}(3)^*\simeq\mathbb{R}^3 so ( 3 ) ∗ ≃ R 3 by the Euclidean scalar product, and the inertia matrices I I I and J J J are symmetric.
For variations with respect to m m m one has δ m Ω = I − 1 δ m \delta_m\Omega=I^{-1}\delta m δ m Ω = I − 1 δ m , and hence
δ m h = − ⟨ p θ , I − 1 δ m ⟩ + ⟨ p m , δ m × Ω ⟩ + ⟨ p m , m × I − 1 δ m ⟩ + ⟨ μ , I − 1 δ m ⟩ . \delta_m h=-\langle p_\theta,I^{-1}\delta m\rangle+\langle p_m,\delta m\times\Omega\rangle
+\langle p_m,m\times I^{-1}\delta m\rangle+\langle\mu,I^{-1}\delta m\rangle. δ m h = − ⟨ p θ , I − 1 δ m ⟩ + ⟨ p m , δ m × Ω ⟩ + ⟨ p m , m × I − 1 δ m ⟩ + ⟨ μ , I − 1 δ m ⟩ .
The two cross-product terms are rearranged by the identities ⟨ p m , δ m × Ω ⟩ = ⟨ Ω × p m , δ m ⟩ \langle p_m,\delta m\times\Omega\rangle=\langle\Omega\times p_m,\delta m\rangle ⟨ p m , δ m × Ω ⟩ = ⟨ Ω × p m , δ m ⟩ and ⟨ p m , m × I − 1 δ m ⟩ = ⟨ I − 1 ( p m × m ) , δ m ⟩ \langle p_m,m\times I^{-1}\delta m\rangle=\langle I^{-1}(p_m\times m),\delta m\rangle ⟨ p m , m × I − 1 δ m ⟩ = ⟨ I − 1 ( p m × m ) , δ m ⟩ , the second using the symmetry of I I I . This gives
∂ h ∂ m = − I − 1 p θ + Ω × p m + I − 1 ( p m × m ) + I − 1 μ . \frac{\partial h}{\partial m}=-I^{-1}p_\theta+\Omega\times p_m+I^{-1}(p_m\times m)+I^{-1}\mu. ∂ m ∂ h = − I − 1 p θ + Ω × p m + I − 1 ( p m × m ) + I − 1 μ .
Therefore,
p ˙ m = I − 1 p θ − Ω × p m − I − 1 ( p m × m ) − I − 1 μ . (83) \dot p_m=I^{-1}p_\theta-\Omega\times p_m-I^{-1}(p_m\times m)-I^{-1}\mu. \tag*{(83)} p ˙ m = I − 1 p θ − Ω × p m − I − 1 ( p m × m ) − I − 1 μ . (83)
For variations with respect to l l l one has δ l Ω = − I − 1 δ l \delta_l\Omega=-I^{-1}\delta l δ l Ω = − I − 1 δ l , and therefore
δ l h = ⟨ p θ , J − 1 δ l + I − 1 δ l ⟩ − ⟨ p m , m × I − 1 δ l ⟩ − ⟨ μ , I − 1 δ l ⟩ . \delta_l h=\langle p_\theta,J^{-1}\delta l+I^{-1}\delta l\rangle
-\langle p_m,m\times I^{-1}\delta l\rangle-\langle\mu,I^{-1}\delta l\rangle. δ l h = ⟨ p θ , J − 1 δ l + I − 1 δ l ⟩ − ⟨ p m , m × I − 1 δ l ⟩ − ⟨ μ , I − 1 δ l ⟩ .
It follows that
∂ h ∂ l = J − 1 p θ + I − 1 p θ − I − 1 ( p m × m ) − I − 1 μ , \frac{\partial h}{\partial l}=J^{-1}p_\theta+I^{-1}p_\theta-I^{-1}(p_m\times m)-I^{-1}\mu, ∂ l ∂ h = J − 1 p θ + I − 1 p θ − I − 1 ( p m × m ) − I − 1 μ ,
and hence
p ˙ l = − J − 1 p θ − I − 1 p θ + I − 1 ( p m × m ) + I − 1 μ . (84) \dot p_l=-J^{-1}p_\theta-I^{-1}p_\theta+I^{-1}(p_m\times m)+I^{-1}\mu. \tag*{(84)} p ˙ l = − J − 1 p θ − I − 1 p θ + I − 1 ( p m × m ) + I − 1 μ . (84)
Substituting the optimal control u = p l u=p_l u = p l and the reconstruction variable ξ = Ω \xi=\Omega ξ = Ω into the preceding equations gives the complete reduced optimality system
\begin{aligned}
\dot\theta&=J^{-1}l-\Omega,\\
\dot m&=m\times\Omega,\\
\dot l&=p_l,\\
\dot p_\theta&=0,\\
\dot p_m&=I^{-1}p_\theta-\Omega\times p_m-I^{-1}(p_m\times m)-I^{-1}\mu,\\
\dot p_l&=-J^{-1}p_\theta-I^{-1}p_\theta+I^{-1}(p_m\times m)+I^{-1}\mu,\\
\frac{D\mu}{Dt}&=\operatorname{ad}_{\Omega}^{*}\mu,\\
u&=p_l,
\end{aligned}
\qquad
\Omega=I^{-1}(m-l). \tag*{(85)}\end{aligned}
Under the standard identification s o ( 3 ) ∗ ≃ R 3 \mathfrak{so}(3)^{*}\simeq\mathbb{R}^{3} so ( 3 ) ∗ ≃ R 3 the coadjoint equation reads μ ˙ = μ × Ω \dot\mu=\mu\times\Omega μ ˙ = μ × Ω . Two consistency checks are worth recording. First, μ \mu μ and the state variable m m m are both body quantities in s o ( 3 ) ∗ \mathfrak{so}(3)^{*} so ( 3 ) ∗ attached to the same S O ( 3 ) SO(3) S O ( 3 ) symmetry, and (85) gives them the same Lie–Poisson form, m ˙ = ad Ω ∗ m \dot m=\operatorname{ad}_{\Omega}^{*}m m ˙ = ad Ω ∗ m and μ ˙ = ad Ω ∗ μ \dot\mu=\operatorname{ad}_{\Omega}^{*}\mu μ ˙ = ad Ω ∗ μ . Second, since H P H_P H P is S O ( 3 ) SO(3) S O ( 3 ) -invariant, the spatial costate v = Ad R − 1 ∗ μ = R μ v=\operatorname{Ad}_{R^{-1}}^{*}\mu=R\mu v = Ad R − 1 ∗ μ = R μ is conserved, and v ˙ = R ( μ ˙ + Ω × μ ) \dot v=R(\dot\mu+\Omega\times\mu) v ˙ = R ( μ ˙ + Ω × μ ) vanishes precisely when μ ˙ = μ × Ω \dot\mu=\mu\times\Omega μ ˙ = μ × Ω .
The reduced optimality system has therefore been obtained directly by reducing the Lagrange–Pontryagin variational principle. The attitude variable R R R no longer appears among the reduced state variables; its contribution is carried by the vertical variable ξ = Ω \xi=\Omega ξ = Ω and the coadjoint variable μ ∈ s o ( 3 ) ∗ \mu\in\mathfrak{so}(3)^{*} μ ∈ so ( 3 ) ∗ . The next subsection recovers the same system independently, from the Lie–Dirac reduction of the implicit formulation.
Reduced Dirac formulation
The reduced Dirac structure
The canonical Dirac structure D P D_P D P induces a reduced Dirac structure [ D P ] S O ( 3 ) [D_P]_{SO(3)} [ D P ] S O ( 3 ) on T ∗ P / S O ( 3 ) ≃ T ∗ B ⊕ s o ( 3 ) ∗ T^{*}P/SO(3)\simeq T^{*}B\oplus\mathfrak{so}(3)^{*} T ∗ P / S O ( 3 ) ≃ T ∗ B ⊕ so ( 3 ) ∗ . At a point ( z , p z , μ ) (z,p_z,\mu) ( z , p z , μ ) , an element of the reduced tangent bundle is written ( v z , v p z ; ξ , v μ ) (v_z,v_{p_z};\xi,v_\mu) ( v z , v p z ; ξ , v μ ) and an element of the dual bundle is written ( π z , π p z ; α ξ , α v μ ) (\pi_z,\pi_{p_z};\alpha_\xi,\alpha_{v_\mu}) ( π z , π p z ; α ξ , α v μ ) .
According to the reduced canonical Dirac structure (42) in the left-action convention of Remark 7, that is, with the coadjoint relation (43) , these variables satisfy
v z = π p z , v p z + π z = − ⟨ μ , F A ( v z , ⋅ ) ⟩ , ξ = α v μ , v μ + α ξ = ad ξ ∗ μ . (86) \begin{aligned}
v_z&=\pi_{p_z},\\
v_{p_z}+\pi_z&=-\langle\mu,F_A(v_z,\cdot)\rangle,\\
\xi&=\alpha_{v_\mu},\\
v_\mu+\alpha_\xi&=\operatorname{ad}_{\xi}^{*}\mu. \tag*{(86)}
\end{aligned} v z v p z + π z ξ v μ + α ξ = π p z , = − ⟨ μ , F A ( v z , ⋅ )⟩ , = α v μ , = ad ξ ∗ μ . (86)
The product connection chosen in the spacecraft example is flat, F A = 0 F_A = 0 F A = 0 , so the reduced Dirac structure simplifies to
{ v z = π p z , v p z = − π z , ξ = α v μ , v μ + α ξ = ad ξ ∗ μ . (87) \begin{cases}
v_z = \pi_{p_z},\\
v_{p_z} = -\pi_z,\\
\xi= \alpha_{v_\mu},\\
v_\mu+ \alpha_\xi= \operatorname{ad}^{*}_{\xi}\mu. \tag*{(87)}
\end{cases} ⎩ ⎨ ⎧ v z = π p z , v p z = − π z , ξ = α v μ , v μ + α ξ = ad ξ ∗ μ . (87)
Reduced implicit Dirac system
The reduced Pontryagin Hamiltonian is the one already obtained in (71), with Ω = I − 1 ( m − l ) \Omega= I^{-1}(m-l) Ω = I − 1 ( m − l ) . The reduced implicit Dirac system is
( [ X ] S O ( 3 ) , d h ) ∈ [ D P ] S O ( 3 ) , (88) ([X]_{\mathrm{SO}(3)},dh) \in[D_P]_{\mathrm{SO}(3)}, \tag*{(88)} ([ X ] SO ( 3 ) , d h ) ∈ [ D P ] SO ( 3 ) , (88)
together with the reduced primary constraint ∂ h / ∂ u = 0 \partial h/\partial u = 0 ∂ h / ∂ u = 0 . The problem is regular, so that constraint determines the control locally as a function of the remaining reduced variables, and on the reduced primary constraint manifold [ M 1 ] S O ( 3 ) = { ( z , p z , μ , u ) : ∂ h / ∂ u = 0 } [M_1]_{\mathrm{SO}(3)} = \{(z,p_z,\mu,u):\partial h/\partial u=0\} [ M 1 ] SO ( 3 ) = {( z , p z , μ , u ) : ∂ h / ∂ u = 0 } the reduced Pontryagin Hamiltonian determines a Hamiltonian on T ∗ B ⊕ s o ( 3 ) ∗ T^*B \oplus\mathfrak{so}(3)^* T ∗ B ⊕ so ( 3 ) ∗ , as in Section 4.4. Along a curve in [ M 1 ] S O ( 3 ) [M_1]_{\mathrm{SO}(3)} [ M 1 ] SO ( 3 ) the reduced tangent component of the dynamics is represented by
[ X ] S O ( 3 ) = ( z ˙ , p ˙ z ; ξ , D μ D t ) . (89) [X]_{\mathrm{SO}(3)} = \left(\dot{z},\dot{p}_z;\xi,\frac{D\mu}{Dt}\right). \tag*{(89)} [ X ] SO ( 3 ) = ( z ˙ , p ˙ z ; ξ , D t D μ ) . (89)
The differential of the reduced Hamiltonian is represented, in the dual variables introduced above, by
( π z , π p z ; α ξ , α v μ ) = ( ∂ h ∂ z , ∂ h ∂ p z ; 0 , ∂ h ∂ μ ) . (90) (\pi_z,\pi_{p_z};\alpha_\xi,\alpha_{v_\mu}) = \left(\frac{\partial h}{\partial z},\frac{\partial h}{\partial p_z};0,\frac{\partial h}{\partial\mu}\right). \tag*{(90)} ( π z , π p z ; α ξ , α v μ ) = ( ∂ z ∂ h , ∂ p z ∂ h ; 0 , ∂ μ ∂ h ) . (90)
In particular α ξ = 0 \alpha_\xi= 0 α ξ = 0 and α v μ = ∂ h / ∂ μ \alpha_{v_\mu} = \partial h/\partial\mu α v μ = ∂ h / ∂ μ : the derivative with respect to the coadjoint variable μ \mu μ is paired with the variation v μ v_\mu v μ , and so corresponds to α v μ \alpha_{v_\mu} α v μ and not to α ξ \alpha_\xi α ξ .
Substituting (89) and (90) into the defining relations (87), we obtain
{ z ˙ = ∂ h ∂ p z , p ˙ z = − ∂ h ∂ z , ξ = ∂ h ∂ μ , D μ D t = ad ξ ∗ μ , 0 = ∂ h ∂ u . (91) \begin{cases}
\dot{z} = \dfrac{\partial h}{\partial p_z},\\
\dot{p}_z = -\dfrac{\partial h}{\partial z},\\
\xi= \dfrac{\partial h}{\partial\mu},\\
\dfrac{D\mu}{Dt} = \operatorname{ad}^{*}_{\xi}\mu,\\
0 = \dfrac{\partial h}{\partial u}. \tag*{(91)}
\end{cases} ⎩ ⎨ ⎧ z ˙ = ∂ p z ∂ h , p ˙ z = − ∂ z ∂ h , ξ = ∂ μ ∂ h , D t D μ = ad ξ ∗ μ , 0 = ∂ u ∂ h . (91)
The four relations of (87) give the first four equations in order, the fourth using α ξ = 0 \alpha_\xi= 0 α ξ = 0 , and the last equation is the reduced primary constraint.
Thus, before any particular derivative is computed, the reduced Dirac system reproduces the equations obtained from the reduced Lagrange–Pontryagin principle, namely (76). The flat product connection is what removes the curvature term from the reduced costate equation, and the vertical part of the reduced Dirac structure is what produces the coadjoint equation. It remains to write these equations out explicitly for the spacecraft.
Explicit Dirac equations
The five equations of (91) are now evaluated. The reduced Hamiltonian h h h entering them is (71), which is the same function of ( z , p z , μ , u ) (z,p_z,\mu,u) ( z , p z , μ , u ) that appeared in the variational route, so its partial derivatives have already been computed in Section 5.7.3 and need not be recomputed here. The state equations z ˙ = ∂ h / ∂ p z \dot{z} = \partial h/\partial p_z z ˙ = ∂ h / ∂ p z return (77). The vertical equation ξ = ∂ h / ∂ μ \xi= \partial h/\partial\mu ξ = ∂ h / ∂ μ returns ξ = Ω \xi= \Omega ξ = Ω , and the primary constraint 0 = ∂ h / ∂ u = p l − u 0 = \partial h/\partial u = p_l-u 0 = ∂ h / ∂ u = p l − u returns u = p l u=p_l u = p l . The vertical component of the reduced Dirac relation, the fourth line of (87) with α ξ = 0 \alpha_\xi=0 α ξ = 0 , gives the coadjoint equation D μ / D t = ad Ω ∗ μ D\mu/Dt = \operatorname{ad}^{*}_{\Omega}\mu D μ / D t = ad Ω ∗ μ , that is, μ ˙ = μ × Ω \dot{\mu}=\mu\times\Omega μ ˙ = μ × Ω . Finally p ˙ z = − ∂ h / ∂ z \dot{p}_z=-\partial h/\partial z p ˙ z = − ∂ h / ∂ z gives p ˙ θ = 0 \dot{p}_\theta=0 p ˙ θ = 0 , since h h h is independent of θ \theta θ , together with (83) and (84) from ∂ h / ∂ m \partial h/\partial m ∂ h / ∂ m and ∂ h / ∂ l \partial h/\partial l ∂ h / ∂ l .
Substituting u = p l u=p_l u = p l and ξ = Ω \xi=\Omega ξ = Ω , the complete reduced Dirac system is therefore
\begin{aligned}
\dot{\theta} &= J^{-1}l-\Omega,\\
\dot{m} &= m\times\Omega,\\
\dot{l} &= p_l,\\
\dot{p}_{\theta} &= 0,\\
\dot{p}_m &= I^{-1}p_{\theta}-\Omega\times p_m-I^{-1}(p_m\times m)-I^{-1}\mu,\\
\dot{p}_l &= -J^{-1}p_{\theta}-I^{-1}p_{\theta}+I^{-1}(p_m\times m)+I^{-1}\mu,\\
\frac{D\mu}{Dt} &= \operatorname{ad}^{*}_{\Omega}\mu,\\
u &= p_l,
\end{aligned}
\qquad
\Omega=I^{-1}(m-l). \tag*{(92)}\end{aligned}
Comparing (92) with (85), the two systems are term by term the same. Thus, the reduction of the implicit Dirac formulation and the reduction of the Lagrange–Pontryagin variational principle produce exactly the same reduced optimality equations for the spacecraft with internal rotors.
This provides an explicit realization, for the spacecraft optimal control problem, of the equivalence between the reduced variational and reduced Dirac descriptions established in the general theory.
Application: the sub-Riemannian geodesic problem on the Heisenberg group
The spacecraft of Section 5 was reduced by a nonabelian group, but with respect to a flat product connection, so that the curvature term in the reduced equations (26) and in the reduced Dirac structure (42) played no role. We now treat an example of the opposite kind. The symmetry group is the center of the Heisenberg group, hence abelian, so that the coadjoint term drops out; but the natural connection is not flat, its curvature being nowhere zero. (The bundle itself is trivial: σ ( x , y ) = ( x , y , 0 ) \sigma(x,y)=(x,y,0) σ ( x , y ) = ( x , y , 0 ) is a global section, and σ ( x , y ) ⋅ s = ( x , y , s ) \sigma(x,y)\cdot s=(x,y,s) σ ( x , y ) ⋅ s = ( x , y , s ) is a global equivariant trivialization. It is the connection, not the topology of the bundle, that distinguishes this example from the previous one; nonzero curvature is perfectly compatible with a trivial bundle.) Taken together, the two examples exercise both of the terms that distinguish the reduced Pontryagin Maximum Principle from its unreduced counterpart.
A further feature of the present example is that the reduction is not merely a convenience: the vertical variable is precisely the one eliminated by passing to the sub-Riemannian picture, and the reconstruction equation turns out to be the horizontality condition itself.
The Heisenberg group and its horizontal distribution
Let H H H denote the three-dimensional Heisenberg group, that is, R 3 \mathbb{R}^{3} R 3 with coordinates ( x , y , w ) (x,y,w) ( x , y , w ) and the multiplication
( x 1 , y 1 , w 1 ) ⋅ ( x 2 , y 2 , w 2 ) = ( x 1 + x 2 , y 1 + y 2 , w 1 + w 2 + 1 2 ( x 1 y 2 − y 1 x 2 ) ) . (93) (x_1,y_1,w_1)\cdot(x_2,y_2,w_2)=\left(x_1+x_2,y_1+y_2,w_1+w_2+\frac{1}{2}(x_1y_2-y_1x_2)\right). \tag*{(93)} ( x 1 , y 1 , w 1 ) ⋅ ( x 2 , y 2 , w 2 ) = ( x 1 + x 2 , y 1 + y 2 , w 1 + w 2 + 2 1 ( x 1 y 2 − y 1 x 2 ) ) . (93)
A basis of left-invariant vector fields is
X = ∂ ∂ x − y 2 ∂ ∂ w , Y = ∂ ∂ y + x 2 ∂ ∂ w , W = ∂ ∂ w , (94) X=\frac{\partial}{\partial x}-\frac{y}{2}\frac{\partial}{\partial w},
\qquad
Y=\frac{\partial}{\partial y}+\frac{x}{2}\frac{\partial}{\partial w},
\qquad
W=\frac{\partial}{\partial w}, \tag*{(94)} X = ∂ x ∂ − 2 y ∂ w ∂ , Y = ∂ y ∂ + 2 x ∂ w ∂ , W = ∂ w ∂ , (94)
and the only nonvanishing bracket is
[ X , Y ] = W . (95) [X,Y]=W. \tag*{(95)} [ X , Y ] = W . (95)
The distribution
H = span { X , Y } ⊂ T H \mathcal{H}=\operatorname{span}\{X,Y\}\subset TH H = span { X , Y } ⊂ T H
is bracket generating of step two, and the sub-Riemannian structure is obtained by declaring X , Y X,Y X , Y to be orthonormal.
The optimal control problem
We take as state manifold
P = H ≅ R 3 , P=H\cong\mathbb{R}^{3}, P = H ≅ R 3 ,
as control bundle the trivial bundle
C = P × U , U ≅ R 2 , u = ( u 1 , u 2 ) , C=P\times U,\qquad U\cong\mathbb{R}^{2},\qquad u=(u_1,u_2), C = P × U , U ≅ R 2 , u = ( u 1 , u 2 ) ,
and as controlled dynamics the bundle map
Γ ( x , y , w , u ) = u 1 X + u 2 Y , (96) \Gamma(x,y,w,u)=u_1X+u_2Y, \tag*{(96)} Γ ( x , y , w , u ) = u 1 X + u 2 Y , (96)
that is,
x ˙ = u 1 , y ˙ = u 2 , w ˙ = 1 2 ( x u 2 − y u 1 ) . (97) \dot{x}=u_1,\qquad\dot{y}=u_2,\qquad\dot{w}=\frac{1}{2}(xu_2-yu_1). \tag*{(97)} x ˙ = u 1 , y ˙ = u 2 , w ˙ = 2 1 ( x u 2 − y u 1 ) . (97)
The running cost is the sub-Riemannian energy
L ( x , y , w , u ) = 1 2 ∣ u ∣ 2 = 1 2 ( u 1 2 + u 2 2 ) . (98) L(x,y,w,u)=\frac{1}{2}|u|^2=\frac{1}{2}(u_1^2+u_2^2). \tag*{(98)} L ( x , y , w , u ) = 2 1 ∣ u ∣ 2 = 2 1 ( u 1 2 + u 2 2 ) . (98)
Minimizing (98) subject to (97) is the sub-Riemannian geodesic problem on H H H ; see [33 ] . Since the admissible curves are exactly the horizontal ones, the quadruple ( P , C , Γ , L ) (P,C,\Gamma,L) ( P , C , Γ , L ) is an optimal control problem in the sense of Definition 2.
The symmetry group and the reduced bundles
Let
G = Z ( H ) = { ( 0 , 0 , s ) : s ∈ R } ≅ R G=Z(H)=\{(0,0,s):s\in\mathbb{R}\}\cong\mathbb{R} G = Z ( H ) = {( 0 , 0 , s ) : s ∈ R } ≅ R
be the center of H H H , acting on P P P by right translation,
( x , y , w ) ⋅ s = ( x , y , w + s ) . (99) (x,y,w)\cdot s=(x,y,w+s). \tag*{(99)} ( x , y , w ) ⋅ s = ( x , y , w + s ) . (99)
Because G G G is central, this coincides with the corresponding left action, and the left and right conventions of Remarks 1 and 7 agree; we may therefore use the equations of Sections 3.3 and § 4 exactly as they stand.
The action is free and proper, G G G acts trivially on U U U , and neither Γ \Gamma Γ nor L L L depends on w w w , so that ( P , C , Γ , L ) (P,C,\Gamma,L) ( P , C , Γ , L ) is G G G -invariant in the sense of Definition 7. The base of the reduction is
B = P / G ≅ R 2 , z = ( x , y ) , (100) B=P/G\cong\mathbb{R}^2,\qquad z=(x,y), \tag*{(100)} B = P / G ≅ R 2 , z = ( x , y ) , (100)
and ρ : H → R 2 \rho:H\to\mathbb{R}^2 ρ : H → R 2 , ρ ( x , y , w ) = ( x , y ) \rho(x,y,w)=(x,y) ρ ( x , y , w ) = ( x , y ) , is a principal R \mathbb{R} R -bundle.
Since g = R \mathfrak{g}=\mathbb{R} g = R is abelian, the adjoint and coadjoint bundles are trivial,
g ~ ≅ B × R , g ~ ∗ ≅ B × R , ad ξ ∗ μ = 0 for all ξ , μ , (101) \widetilde{\mathfrak{g}}\cong B\times\mathbb{R},\qquad
\widetilde{\mathfrak{g}}^*\cong B\times\mathbb{R},\qquad
\operatorname{ad}^*_\xi\mu=0\quad\text{for all }\xi,\mu, \tag*{(101)} g ≅ B × R , g ∗ ≅ B × R , ad ξ ∗ μ = 0 for all ξ , μ , (101)
and the induced covariant derivative on g ~ \widetilde{\mathfrak{g}} g and g ~ ∗ \widetilde{\mathfrak{g}}^* g ∗ reduces to the ordinary time derivative. The reduced control bundle is U ~ ≅ B × R 2 \widetilde{U}\cong B\times\mathbb{R}^2 U ≅ B × R 2 , so that the reduced Pontryagin bundle is
M / G ≅ T ∗ R 2 ⊕ g ~ ∗ ⊕ U ~ , (102) M/G\cong T^*\mathbb{R}^2\oplus\widetilde{\mathfrak{g}}^*\oplus\widetilde{U}, \tag*{(102)} M / G ≅ T ∗ R 2 ⊕ g ∗ ⊕ U , (102)
with coordinates ( x , y , p x , p y , μ , u ) (x,y,p_x,p_y,\mu,u) ( x , y , p x , p y , μ , u ) .
The connection and its curvature. We take as principal connection the R \mathbb{R} R -valued one-form
A = d w − 1 2 ( x d y − y d x ) . (103) A=dw-\frac{1}{2}(x\,dy-y\,dx). \tag*{(103)} A = d w − 2 1 ( x d y − y d x ) . (103)
It is indeed a principal connection: the fundamental vector field of ζ ∈ g = R \zeta\in\mathfrak{g}=\mathbb{R} ζ ∈ g = R is ζ P = ζ ∂ w \zeta_P=\zeta\,\partial_w ζ P = ζ ∂ w , whence A ( ζ P ) = ζ A(\zeta_P)=\zeta A ( ζ P ) = ζ , and A A A is invariant under (99) because its coefficients do not involve w w w . Moreover
A ( X ) = y 2 − y 2 = 0 , A ( Y ) = − x 2 + x 2 = 0 , A(X)=\frac{y}{2}-\frac{y}{2}=0,\qquad A(Y)=-\frac{x}{2}+\frac{x}{2}=0, A ( X ) = 2 y − 2 y = 0 , A ( Y ) = − 2 x + 2 x = 0 ,
so that the horizontal subspace of A A A is exactly the sub-Riemannian distribution,
ker A = H = span { X , Y } . (104) \ker A=\mathcal{H}=\operatorname{span}\{X,Y\}. \tag*{(104)} ker A = H = span { X , Y } . (104)
Since g \mathfrak{g} g is abelian, [ A , A ] = 0 [A,A]=0 [ A , A ] = 0 and the curvature is
F A = d A = − d x ∧ d y , (105) F_A=dA=-dx\wedge dy, \tag*{(105)} F A = d A = − d x ∧ d y , (105)
which is a nowhere vanishing basic two-form on B = R 2 B=\mathbb{R}^2 B = R 2 . This is the essential difference with the example of Section 5.
The reduced dynamics. The decomposition [ Γ ] ( z , u ) = ( Γ ~ ( z , u ) , γ ~ ( z , u ) ) [\Gamma](z,u)=(\widetilde{\Gamma}(z,u),\widetilde{\gamma}(z,u)) [ Γ ] ( z , u ) = ( Γ ( z , u ) , γ ( z , u )) determined by A A A is
Γ ~ ( z , u ) = u 1 ∂ ∂ x + u 2 ∂ ∂ y , γ ~ ( z , u ) = A ( Γ ( x , y , w , u ) ) = 0 , (106) \widetilde{\Gamma}(z,u)=u_1\frac{\partial}{\partial x}+u_2\frac{\partial}{\partial y},\qquad
\widetilde{\gamma}(z,u)=A\bigl(\Gamma(x,y,w,u)\bigr)=0, \tag*{(106)} Γ ( z , u ) = u 1 ∂ x ∂ + u 2 ∂ y ∂ , γ ( z , u ) = A ( Γ ( x , y , w , u ) ) = 0 , (106)
the vertical part vanishing identically by (104): the controlled vector field is horizontal, which is precisely the statement that admissible curves of the sub-Riemannian problem are horizontal curves.
Pontryagin formulation
Writing a costate as λ = ( λ x , λ y , λ w ) ∈ T ( x , y , w ) ∗ P \lambda= (\lambda_x,\lambda_y,\lambda_w) \in T^*_{(x,y,w)}P λ = ( λ x , λ y , λ w ) ∈ T ( x , y , w ) ∗ P , the Pontryagin Hamiltonian H P = ⟨ λ , Γ ⟩ − L H_P = \langle\lambda,\Gamma\rangle- L H P = ⟨ λ , Γ ⟩ − L is
H P = λ x u 1 + λ y u 2 + λ w 2 ( x u 2 − y u 1 ) − 1 2 ∣ u ∣ 2 . (107) H_P = \lambda_x u_1 + \lambda_y u_2 + \frac{\lambda_w}{2}(x u_2-y u_1)-\frac{1}{2}|u|^2. \tag*{(107)} H P = λ x u 1 + λ y u 2 + 2 λ w ( x u 2 − y u 1 ) − 2 1 ∣ u ∣ 2 . (107)
The primary constraints ∂ H P / ∂ u a = 0 \partial H_P/\partial u^a=0 ∂ H P / ∂ u a = 0 read
u 1 = λ x − y 2 λ w , u 2 = λ y + x 2 λ w , (108) u_1=\lambda_x-\frac{y}{2}\lambda_w,\qquad u_2=\lambda_y+\frac{x}{2}\lambda_w, \tag*{(108)} u 1 = λ x − 2 y λ w , u 2 = λ y + 2 x λ w , (108)
and
∂ 2 H P ∂ u a ∂ u b = − Id R 2 , \frac{\partial^2 H_P}{\partial u^a\partial u^b}=-\operatorname{Id}_{\mathbb{R}^2}, ∂ u a ∂ u b ∂ 2 H P = − Id R 2 ,
so that the problem is regular at every point, and the standing assumptions of § 2 are met. (Regularity of the normal problem does not by itself exclude abnormal extremals, which live on the leaf p 0 = 0 p_0=0 p 0 = 0 where the Hessian in the controls vanishes; that the Heisenberg problem has no nontrivial abnormal extremals is proved in Corollary 1.) The unreduced first-order conditions are (97) together with
λ ˙ x = − λ w 2 u 2 , λ ˙ y = λ w 2 u 1 , λ ˙ w = 0. (109) \dot{\lambda}_x=-\frac{\lambda_w}{2}u_2,\qquad\dot{\lambda}_y=\frac{\lambda_w}{2}u_1,\qquad\dot{\lambda}_w=0. \tag*{(109)} λ ˙ x = − 2 λ w u 2 , λ ˙ y = 2 λ w u 1 , λ ˙ w = 0. (109)
Under the identification (18) determined by A A A , the momentum map is J ( x , λ ) = ⟨ λ , ∂ w ⟩ = λ w J(x,\lambda)=\langle\lambda,\partial_w\rangle=\lambda_w J ( x , λ ) = ⟨ λ , ∂ w ⟩ = λ w and the horizontal part of λ \lambda λ is obtained by evaluating on the frame (94) . Hence the reduced variables are
p x = λ x − y 2 λ w , p y = λ y + x 2 λ w , μ = λ w . (110) p_x=\lambda_x-\frac{y}{2}\lambda_w,\qquad p_y=\lambda_y+\frac{x}{2}\lambda_w,\qquad\mu=\lambda_w. \tag*{(110)} p x = λ x − 2 y λ w , p y = λ y + 2 x λ w , μ = λ w . (110)
Reduced variational formulation
By (106) the reduced Pontryagin Hamiltonian (24) is
h ( z , p z , μ , u ) = ⟨ p z , Γ ~ ( z ~ , u ) ⟩ + ⟨ μ , γ ~ ( z ~ , u ) ⟩ − l ( z , u ) = p x u 1 + p y u 2 − 1 2 ∣ u ∣ 2 . (111) h(z,p_z,\mu,u)=\langle p_z,\widetilde{\Gamma}(\widetilde{z},u)\rangle+\langle\mu,\widetilde{\gamma}(\widetilde{z},u)\rangle-l(z,u)=p_xu_1+p_yu_2-\frac{1}{2}|u|^2. \tag*{(111)} h ( z , p z , μ , u ) = ⟨ p z , Γ ( z , u )⟩ + ⟨ μ , γ ( z , u )⟩ − l ( z , u ) = p x u 1 + p y u 2 − 2 1 ∣ u ∣ 2 . (111)
Substituting (110) shows that h h h coincides with (107) , as it must. Note that h h h is independent of both z z z and μ \mu μ : the entire dependence on the eliminated variable is carried by the curvature. The reduced Lagrange–Pontryagin functional (23) is
S G = ∫ 0 T [ ⟨ p z , z ˙ ⟩ + ⟨ μ , ξ ⟩ − h ( z , p z , μ , u ) ] d t , (112) S_G=\int_0^T\left[\langle p_z,\dot{z}\rangle+\langle\mu,\xi\rangle-h(z,p_z,\mu,u)\right]dt, \tag*{(112)} S G = ∫ 0 T [ ⟨ p z , z ˙ ⟩ + ⟨ μ , ξ ⟩ − h ( z , p z , μ , u ) ] d t , (112)
and, since g \mathfrak{g} g is abelian, the constrained variations (25) reduce to
δ ξ = η ˙ + F A ( δ z , z ˙ ) = η ˙ − ( δ x y ˙ − δ y x ˙ ) . (113) \delta\xi=\dot{\eta}+F_A(\delta z,\dot{z})=\dot{\eta}-(\delta x\dot{y}-\delta y\dot{x}). \tag*{(113)} δ ξ = η ˙ + F A ( δ z , z ˙ ) = η ˙ − ( δ x y ˙ − δ y x ˙ ) . (113)
with η ( 0 ) = η ( T ) = 0 \eta(0)=\eta(T)=0 η ( 0 ) = η ( T ) = 0 . The bracket term [ η , ξ ] [\eta,\xi] [ η , ξ ] , which produced the coadjoint equation in § 5 , is absent here; what survives is exactly the curvature term.
Taking variations and integrating by parts as in Section 3.3,
∫ 0 T ⟨ μ , δ ξ ⟩ d t = − ∫ 0 T ⟨ μ ˙ , η ⟩ d t + ∫ 0 T ⟨ μ , F A ( δ z , z ˙ ) ⟩ d t , \int_0^T\langle\mu,\delta\xi\rangle\,dt=-\int_0^T\langle\dot{\mu},\eta\rangle\,dt+\int_0^T\langle\mu,F_A(\delta z,\dot{z})\rangle\,dt, ∫ 0 T ⟨ μ , δ ξ ⟩ d t = − ∫ 0 T ⟨ μ ˙ , η ⟩ d t + ∫ 0 T ⟨ μ , F A ( δ z , z ˙ )⟩ d t ,
so that stationarity of (112) yields
z ˙ = ∂ h ∂ p z , p ˙ z = − ∂ h ∂ z − ⟨ μ , F A ( z ˙ , ⋅ ) ⟩ , μ ˙ = 0 , ξ = ∂ h ∂ μ , 0 = ∂ h ∂ u . (114) \begin{aligned}
\dot{z}&=\frac{\partial h}{\partial p_z},\\
\dot{p}_z&=-\frac{\partial h}{\partial z}-\langle\mu,F_A(\dot{z},\mathord{\cdot})\rangle,\\
\dot{\mu}&=0,\\
\xi&=\frac{\partial h}{\partial\mu},\\
0&=\frac{\partial h}{\partial u}. \tag*{(114)}
\end{aligned} z ˙ p ˙ z μ ˙ ξ 0 = ∂ p z ∂ h , = − ∂ z ∂ h − ⟨ μ , F A ( z ˙ , ⋅ )⟩ , = 0 , = ∂ μ ∂ h , = ∂ u ∂ h . (114)
which is (26) with ad ξ ∗ μ = 0 \operatorname{ad}^*_{\xi}\mu=0 ad ξ ∗ μ = 0 .
The reduced optimality equations. We compute (114) explicitly. The stationarity condition in the control gives
0 = ∂ h ∂ u = p z − u , that is, u = ( p x , p y ) . (115) 0=\frac{\partial h}{\partial u}=p_z-u,\qquad\text{that is,}\qquad u=(p_x,p_y). \tag*{(115)} 0 = ∂ u ∂ h = p z − u , that is, u = ( p x , p y ) . (115)
Since h h h is independent of μ \mu μ ,
ξ ˙ = ∂ h ∂ μ = 0. (116) \dot{\xi}=\frac{\partial h}{\partial\mu}=0. \tag*{(116)} ξ ˙ = ∂ μ ∂ h = 0. (116)
For the costate equation, ∂ h / ∂ z = 0 \partial h/\partial z=0 ∂ h / ∂ z = 0 and, by (105),
F A ( z ˙ , ⋅ ) = − ( x ˙ d y − y ˙ d x ) = u 2 d x − u 1 d y , F_A(\dot z,\cdot)=-(\dot x\,dy-\dot y\,dx)=u_2\,dx-u_1\,dy, F A ( z ˙ , ⋅ ) = − ( x ˙ d y − y ˙ d x ) = u 2 d x − u 1 d y ,
whence
p ˙ x = − μ u 2 , p ˙ y = μ u 1 . (117) \dot p_x=-\mu u_2,\qquad\dot p_y=\mu u_1. \tag*{(117)} p ˙ x = − μ u 2 , p ˙ y = μ u 1 . (117)
Collecting (115)–(117), the complete reduced optimality system is
x ˙ = p x , y ˙ = p y , p ˙ x = − μ p y , p ˙ y = μ p x , μ ˙ = 0 , ξ = 0 , u = ( p x , p y ) . (118) \begin{aligned}
\dot x&=p_x,\\
\dot y&=p_y,\\
\dot p_x&=-\mu p_y,\\
\dot p_y&=\mu p_x,\\
\dot\mu&=0,\\
\xi&=0,\\
u&=(p_x,p_y). \tag*{(118)}
\end{aligned} x ˙ y ˙ p ˙ x p ˙ y μ ˙ ξ u = p x , = p y , = − μ p y , = μ p x , = 0 , = 0 , = ( p x , p y ) . (118)
Remark 8. Equations (118) may be checked directly against the unreduced conditions (109). Differentiating (110) along an unreduced extremal and using λ w = 0 \lambda_w=0 λ w = 0 gives
p ˙ x = λ ˙ x − y ˙ 2 λ w = − λ w 2 u 2 − u 2 2 λ w = − μ u 2 , p ˙ y = λ ˙ y + x ˙ 2 λ w = μ u 1 , \dot p_x=\dot\lambda_x-\frac{\dot y}{2}\lambda_w=-\frac{\lambda_w}{2}u_2-\frac{u_2}{2}\lambda_w=-\mu u_2,\qquad
\dot p_y=\dot\lambda_y+\frac{\dot x}{2}\lambda_w=\mu u_1, p ˙ x = λ ˙ x − 2 y ˙ λ w = − 2 λ w u 2 − 2 u 2 λ w = − μ u 2 , p ˙ y = λ ˙ y + 2 x ˙ λ w = μ u 1 ,
in agreement with (117). This provides an independent verification of the sign of the curvature term in (26) and in (42), which the example of Section 5 could not supply.
Reduced Dirac formulation
We now recover (118) from the Lie–Dirac reduction. By Proposition 8, and using ad ξ ∗ μ = 0 \operatorname{ad}^{*}_{\xi}\mu=0 ad ξ ∗ μ = 0 together with (105), the reduced canonical Dirac structure on T ∗ R 2 ⊕ g ∗ T^{*}\mathbb{R}^{2}\oplus\mathfrak{g}^{*} T ∗ R 2 ⊕ g ∗ is
v z = π p z , v p z + π z = − ⟨ μ , F A ( v z , ⋅ ) ⟩ , ξ = α v μ , v μ + α ξ = 0. (119) \begin{aligned}
v_z&=\pi_{p_z},\\
v_{p_z}+\pi_z&=-\langle\mu,F_A(v_z,\cdot)\rangle,\\
\xi&=\alpha_{v_\mu},\\
v_\mu+\alpha_\xi&=0. \tag*{(119)}
\end{aligned} v z v p z + π z ξ v μ + α ξ = π p z , = − ⟨ μ , F A ( v z , ⋅ )⟩ , = α v μ , = 0. (119)
Note that, in contrast with (87), the second relation is not the canonical one: the vertical variable μ \mu μ enters the horizontal block through the curvature.
Along a curve in the reduced primary constraint manifold
[ M 1 ] G = { ( z , p z , μ , u ) ∈ M / G | ∂ h ∂ u = 0 } , [M_1]_G=\left\{(z,p_z,\mu,u)\in M/G\ \middle|\ \frac{\partial h}{\partial u}=0\right\}, [ M 1 ] G = { ( z , p z , μ , u ) ∈ M / G ∂ u ∂ h = 0 } ,
the reduced tangent component of the dynamics is
[ X ] G = ( z ˙ , p ˙ z ; ξ , μ ˙ ) , (120) [X]_G=(\dot z,\dot p_z;\xi,\dot\mu), \tag*{(120)} [ X ] G = ( z ˙ , p ˙ z ; ξ , μ ˙ ) , (120)
while the differential of the reduced Hamiltonian is represented by
( π z , π p z ; α ξ , α v μ ) = ( ∂ h ∂ z , ∂ h ∂ p z ; 0 , ∂ h ∂ μ ) . (121) (\pi_z,\pi_{p_z};\alpha_\xi,\alpha_{v_\mu})=
\left(\frac{\partial h}{\partial z},\frac{\partial h}{\partial p_z};0,\frac{\partial h}{\partial\mu}\right). \tag*{(121)} ( π z , π p z ; α ξ , α v μ ) = ( ∂ z ∂ h , ∂ p z ∂ h ; 0 , ∂ μ ∂ h ) . (121)
Substituting (120) and (121) into (119), the first relation gives z ˙ = ∂ h / ∂ p z \dot z=\partial h/\partial p_z z ˙ = ∂ h / ∂ p z ; the second gives
p ˙ z = − ∂ h ∂ z − ⟨ μ , F A ( z ˙ , ⋅ ) ⟩ ; \dot p_z=-\frac{\partial h}{\partial z}-\langle\mu,F_A(\dot z,\cdot)\rangle; p ˙ z = − ∂ z ∂ h − ⟨ μ , F A ( z ˙ , ⋅ )⟩ ;
the third gives ξ = ∂ h / ∂ μ \xi=\partial h/\partial\mu ξ = ∂ h / ∂ μ ; and the fourth, since α ξ = 0 \alpha_\xi=0 α ξ = 0 , gives μ ˙ = 0 \dot\mu=0 μ ˙ = 0 . Together with the reduced primary constraint ∂ h / ∂ u = 0 \partial h/\partial u=0 ∂ h / ∂ u = 0 these are precisely (114), hence (118).
Thus, as in § 5 , the reduction of the implicit Dirac formulation and the reduction of the Lagrange–Pontryagin variational principle produce the same reduced optimality equations; here, however, the agreement is tested on the curvature term rather than on the coadjoint term.
Reconstruction and explicit solution
The reduced system (118) is readily integrated. Since μ \mu μ is constant and
d d t ( p x + i p y ) = i μ ( p x + i p y ) , \frac{d}{dt}(p_x+i p_y)=i\mu(p_x+i p_y), d t d ( p x + i p y ) = i μ ( p x + i p y ) ,
we obtain, for μ ≠ 0 \mu\ne0 μ = 0 ,
p x ( t ) + i p y ( t ) = r e i ( μ t + φ ) , r = ∣ p z ( 0 ) ∣ , (122) p_x(t)+i p_y(t)=r e^{i(\mu t+\varphi)},\qquad r=\lvert p_z(0)\rvert, \tag*{(122)} p x ( t ) + i p y ( t ) = r e i ( μ t + φ ) , r = ∣ p z ( 0 )∣ , (122)
so that ∣ u ∣ = r \lvert u\rvert=r ∣ u ∣ = r is constant and the extremal is parametrized proportionally to arclength. Integrating z ˙ = p z \dot z=p_z z ˙ = p z ,
x ( t ) + i y ( t ) = c + r i μ e i ( μ t + φ ) , c = ( x ( 0 ) + i y ( 0 ) ) + i r μ e i φ , (123) x(t)+i y(t)=c+\frac{r}{i\mu}e^{i(\mu t+\varphi)},\qquad c=(x(0)+i y(0))+\frac{i r}{\mu}e^{i\varphi}, \tag*{(123)} x ( t ) + i y ( t ) = c + i μ r e i ( μ t + φ ) , c = ( x ( 0 ) + i y ( 0 )) + μ i r e i φ , (123)
that is, the reduced base curve is a circle of center c c c and radius r / ∣ μ ∣ r/\lvert\mu\rvert r / ∣ μ ∣ , traversed with angular velocity μ \mu μ . For μ = 0 \mu=0 μ = 0 the momentum p z p_z p z is constant and the base curve is a straight line.
The reconstruction equation is ξ = 0 \xi=0 ξ = 0 , that is, A ( x ˙ ) = 0 A(\dot x)=0 A ( x ˙ ) = 0 : the extremal in P P P is the horizontal lift of the reduced base curve. Explicitly,
w ˙ = 1 2 ( x y ˙ − y x ˙ ) , so that w ( T ) − w ( 0 ) = 1 2 ∫ 0 T ( x d y − y d x ) , (124) \dot w=\frac{1}{2}(x\dot y-y\dot x),\qquad\text{so that}\qquad w(T)-w(0)=\frac{1}{2}\int_0^T(x\,dy-y\,dx), \tag*{(124)} w ˙ = 2 1 ( x y ˙ − y x ˙ ) , so that w ( T ) − w ( 0 ) = 2 1 ∫ 0 T ( x d y − y d x ) , (124)
the signed area swept by the segment joining the origin to the moving base point. One recovers in this way the classical description of Heisenberg geodesics: they project to circles and straight lines in the plane, the vertical increment being the swept area [33 ] . For an open projected path this swept area is not an intrinsic enclosed area: it depends on the choice of origin, changing by 1 2 ( a Δ y − b Δ x ) \frac{1}{2}(a\,\Delta y-b\,\Delta x) 2 1 ( a Δ y − b Δ x ) under the translation of the origin by ( a , b ) (a,b) ( a , b ) , in step with the gauge dependence of A = d w − 1 2 ( x d y − y d x ) A=dw-\frac{1}{2}(x\,dy-y\,dx) A = d w − 2 1 ( x d y − y d x ) . It is the enclosed signed area exactly when the projection is closed, as for a full circle.
Singular and abnormal extremals
Throughout Sections 2–4 the optimal control problem was assumed regular, so that the Hessian W a b = ∂ 2 H P / ∂ u a ∂ u b W_{ab} = \partial^2 H_P/\partial u^a\partial u^b W ab = ∂ 2 H P / ∂ u a ∂ u b was invertible, the control could be eliminated, and p 1 : M 1 → T ∗ P p_1 : M_1 \to T^*P p 1 : M 1 → T ∗ P was a local diffeomorphism. Under that hypothesis the primary constraint submanifold is locally symplectomorphic to T ∗ P T^*P T ∗ P , on each branch of the stationary control, and the Pontryagin Maximum Principle is there an ordinary Hamiltonian system; the Dirac formulation is then a convenient language, but not more than that. Regularity by itself gives no global statement: several nondegenerate solutions of ∂ H P / ∂ u = 0 \partial H_P/\partial u = 0 ∂ H P / ∂ u = 0 may lie over the same ( x , p ) (x,p) ( x , p ) , as for P = R 2 P = \mathbb{R}^2 P = R 2 , Γ ( x , u ) = u \Gamma(x,u) = u Γ ( x , u ) = u and L ( u ) = e u 1 cos u 2 L(u) = e^{u^1}\cos u_2 L ( u ) = e u 1 cos u 2 , where det Hess L = − e 2 u 1 \det\operatorname{Hess} L = -e^{2u^1} det Hess L = − e 2 u 1 never vanishes while p 1 ( x , u ) = ( x , ∇ L ( u ) ) p_1(x,u) = (x,\nabla L(u)) p 1 ( x , u ) = ( x , ∇ L ( u )) is 2 π 2\pi 2 π -periodic in u 2 u_2 u 2 and misses p = 0 p = 0 p = 0 .
Dropping that hypothesis is not generality for its own sake. Already in [24 ] one finds a bracket-generating rank-two system, the Hilbert–Cartan system on R 5 \mathbb{R}^5 R 5 , carrying a five-parameter family of integral curves that admit no compactly supported variations at all; such curves are trivially stationary for every functional, and are in general not solutions of the associated Euler–Lagrange system. They are counterexamples to the assertion, still occasionally made, that a minimizing curve in a sub-Riemannian manifold satisfies the geodesic equations [24 §2] . The regular theory of Sections 2–4 cannot see them.
In this section we drop the regularity assumption. Two things change. First, the control can no longer be eliminated, and the constraint algorithm of Gotay, Nester and Hinds [21 ] must be run; for optimal control this was carried out in the presymplectic setting by Barbero-Liñán and Muñoz-Lecanda [5 ] . Second, and more importantly for the present paper, the reduction of Section 4 cannot be simplified as in Section 4.4: the passage from [ D Ω ] G [D_{\Omega}]_G [ D Ω ] G to [ D P ] G [D_P]_G [ D P ] G used the regularity of the problem, so in the singular case one must work with the reduced Dirac structure [ D Ω ] G [D_{\Omega}]_G [ D Ω ] G on M / G M/G M / G itself. The object constructed in Section 4 and then set aside is precisely the one that is needed here.
The main result of this section is that reduction commutes with the constraint algorithm (Theorem 5): the algorithm may be run on the reduced space, which is smaller, and the answer is the quotient of the answer obtained upstairs. We then show that abnormal extremals occupy a distinguished leaf of the reduced Pontryagin bundle, and that for control-linear systems the first secondary constraint is exactly the vanishing of the curvature pairing ⟨ μ , F A ⟩ \langle\mu,F_A\rangle ⟨ μ , F A ⟩ appearing in the reduced costate equation (26) (Theorem 6). The term that distinguishes the reduced Pontryagin Maximum Principle from its unreduced counterpart is therefore also the term that obstructs abnormality.
Dirac structures induced by a distribution
We first recall the construction, due in this form to Yoshimura and Marsden [40 §7] , by which a two-form and a distribution together determine a Dirac structure in the vector-bundle sense of Definition 4. It is this construction, rather than the graph of a nondegenerate form, that will carry the constraints.
Definition 11. Let N N N be a manifold, Ω \Omega Ω a two-form on N N N and Δ ⊂ T N \Delta\subset TN Δ ⊂ T N a regular distribution. The Dirac structure induced by ( Ω , Δ ) (\Omega,\Delta) ( Ω , Δ ) is
D Ω , Δ ( n ) = { ( v , α ) ∈ T n N ⊕ T n ∗ N ∣ v ∈ Δ ( n ) , α ∣ Δ ( n ) = Ω ( n ) ( v , ⋅ ) ∣ Δ ( n ) } . (125) D_{\Omega,\Delta}(n) = \left\{(v,\alpha) \in T_nN \oplus T_n^*N \mathrel{\big|} v \in\Delta(n),\ \alpha|_{\Delta(n)} = \Omega(n)(v,\cdot)|_{\Delta(n)}\right\}. \tag*{(125)} D Ω , Δ ( n ) = { ( v , α ) ∈ T n N ⊕ T n ∗ N v ∈ Δ ( n ) , α ∣ Δ ( n ) = Ω ( n ) ( v , ⋅ ) ∣ Δ ( n ) } . (125)
Equivalently, ( v , α ) ∈ D Ω , Δ (v,\alpha) \in D_{\Omega,\Delta} ( v , α ) ∈ D Ω , Δ if and only if v ∈ Δ v \in\Delta v ∈ Δ and
α − i v Ω ∈ Δ ∘ , (126) \alpha-i_v\Omega\in\Delta^\circ, \tag*{(126)} α − i v Ω ∈ Δ ∘ , (126)
which is the form in which the constraints will appear; the failure of the unconstrained equation α = i v Ω \alpha=i_v\Omega α = i v Ω is required to lie in the annihilator of Δ \Delta Δ , exactly as a reaction force does in nonholonomic mechanics.
Proposition 10. D Ω , Δ D_{\Omega,\Delta} D Ω , Δ is a Dirac structure on the vector bundle T N TN T N , in the sense of Definition 4, for every two-form Ω \Omega Ω , possibly degenerate, and every regular distribution Δ \Delta Δ . Moreover
G D Ω , Δ ( n ) : = { v ∈ T n N : ( v , 0 ) ∈ D Ω , Δ ( n ) } = Δ ( n ) ∩ Δ ( n ) ⊥ Ω , (127) G_{D_{\Omega,\Delta}}(n) := \{v \in T_nN : (v,0) \in D_{\Omega,\Delta}(n)\} = \Delta(n) \cap\Delta(n)^{\perp_\Omega}, \tag*{(127)} G D Ω , Δ ( n ) := { v ∈ T n N : ( v , 0 ) ∈ D Ω , Δ ( n )} = Δ ( n ) ∩ Δ ( n ) ⊥ Ω , (127) where Δ ⊥ Ω \Delta^{\perp_\Omega} Δ ⊥ Ω denotes the Ω \Omega Ω -orthogonal complement. Taking Δ = T N \Delta= TN Δ = T N recovers the graph D Ω D_\Omega D Ω , with G D Ω = ker Ω G_{D_\Omega} = \ker\Omega G D Ω = ker Ω .
Proof. Isotropy: if ( v , α ) (v,\alpha) ( v , α ) and ( w , β ) (w,\beta) ( w , β ) lie in D Ω , Δ D_{\Omega,\Delta} D Ω , Δ then w ∈ Δ w \in\Delta w ∈ Δ and v ∈ Δ v \in\Delta v ∈ Δ , so
⟨ ⟨ ( v , α ) , ( w , β ) ⟩ ⟩ = ⟨ α , w ⟩ + ⟨ β , v ⟩ = Ω ( v , w ) + Ω ( w , v ) = 0. \langle\!\langle(v,\alpha),(w,\beta)\rangle\!\rangle= \langle\alpha,w\rangle+ \langle\beta,v\rangle= \Omega(v,w) + \Omega(w,v) = 0. ⟨ ⟨( v , α ) , ( w , β )⟩ ⟩ = ⟨ α , w ⟩ + ⟨ β , v ⟩ = Ω ( v , w ) + Ω ( w , v ) = 0.
Maximality: for fixed v ∈ Δ v \in\Delta v ∈ Δ the admissible α \alpha α form an affine space over Δ ∘ \Delta^\circ Δ ∘ , so dim D Ω , Δ = dim Δ + dim Δ ∘ = dim N \dim D_{\Omega,\Delta} = \dim\Delta+ \dim\Delta^\circ= \dim N dim D Ω , Δ = dim Δ + dim Δ ∘ = dim N , which is the rank of a maximally isotropic subbundle of T N ⊕ T ∗ N TN \oplus T^*N T N ⊕ T ∗ N . The kernel: ( v , 0 ) ∈ D Ω , Δ (v,0) \in D_{\Omega,\Delta} ( v , 0 ) ∈ D Ω , Δ means v ∈ Δ v \in\Delta v ∈ Δ and Ω ( v , ⋅ ) ∣ Δ = 0 \Omega(v,\cdot)|_\Delta= 0 Ω ( v , ⋅ ) ∣ Δ = 0 , that is, v ∈ Δ ∩ Δ ⊥ Ω v \in\Delta\cap\Delta^{\perp_\Omega} v ∈ Δ ∩ Δ ⊥ Ω . □
Only the pointwise linear algebra of D Ω , Δ D_{\Omega,\Delta} D Ω , Δ enters the constraint algorithm of the next subsection, so no integrability hypothesis will be needed there; a routine computation with the Courant bracket shows that D Ω , Δ D_{\Omega,\Delta} D Ω , Δ is integrable precisely when Δ \Delta Δ is involutive and d Ω \mathrm{d}\Omega d Ω vanishes on Δ \Delta Δ , but this will not be used. We shall use Definition 11 with Δ = T N \Delta= TN Δ = T N for a submanifold N ⊂ M N \subset M N ⊂ M , in which case (126) reads α − i v Ω ∈ ( T N ) ∘ \alpha-\mathrm{i}_v\Omega\in(TN)^\circ α − i v Ω ∈ ( T N ) ∘ along N N N .
The constraint algorithm in Dirac form
Let D D D be a Dirac structure on a manifold N N N and H ∈ C ∞ ( N ) H\in C^\infty(N) H ∈ C ∞ ( N ) . The implicit Hamiltonian system determined by the pair is ( X , d H ) ∈ D (X,\mathrm{d}H)\in D ( X , d H ) ∈ D . Writing, as in (127) ,
G D ( n ) = { v ∈ T n N : ( v , 0 ) ∈ D ( n ) } , ρ ∗ ( D ) ( n ) = { α ∈ T n ∗ N : ∃ v , ( v , α ) ∈ D ( n ) } , G_D(n)=\{v\in T_nN:(v,0)\in D(n)\},\qquad\rho^*(D)(n)=\{\alpha\in T_n^*N:\exists v,\ (v,\alpha)\in D(n)\}, G D ( n ) = { v ∈ T n N : ( v , 0 ) ∈ D ( n )} , ρ ∗ ( D ) ( n ) = { α ∈ T n ∗ N : ∃ v , ( v , α ) ∈ D ( n )} ,
both understood pointwise , maximal isotropy of D ( n ) D(n) D ( n ) in T n N ⊕ T n ∗ N T_nN\oplus T_n^*N T n N ⊕ T n ∗ N gives ρ ∗ ( D ) ( n ) = G D ( n ) ∘ \rho^*(D)(n)=G_D(n)^\circ ρ ∗ ( D ) ( n ) = G D ( n ) ∘ , so that the system admits a solution at n n n if and only if d H ( n ) ∈ G D ( n ) ∘ \mathrm{d}H(n)\in G_D(n)^\circ d H ( n ) ∈ G D ( n ) ∘ . This is the intrinsic form of the primary constraint.
Definition 12. Let Ω \Omega Ω be a two-form on N N N , let D = D Ω D=D_\Omega D = D Ω be its graph, and let H ∈ C ∞ ( N ) H\in C^\infty(N) H ∈ C ∞ ( N ) . The constraint algorithm for ( N , D Ω , H ) (N,D_\Omega,H) ( N , D Ω , H ) is the sequence
N 0 = N , N k + 1 = { n ∈ N k : d H ( n ) ∣ G D k ( n ) = 0 } , D k = D Ω , T N k , (128) N_0=N,\qquad N_{k+1}=\{n\in N_k:\mathrm{d}H(n)|_{G_{D_k}(n)}=0\},\qquad D_k=D_{\Omega,TN_k}, \tag*{(128)} N 0 = N , N k + 1 = { n ∈ N k : d H ( n ) ∣ G D k ( n ) = 0 } , D k = D Ω , T N k , (128) where D k D_k D k is the structure induced on N N N along N k N_k N k in the sense of Definition 11; by Proposition 10 it is a Dirac structure in the vector-bundle sense, which is all that the pointwise conditions of (128) require. We assume throughout that every N k N_k N k is an embedded submanifold and that the sequence stabilizes, N k + 1 = N k = : N ∞ N_{k+1}=N_k=:N_\infty N k + 1 = N k =: N ∞ for k k k large.
Two features of this definition should be noted, because they are what make it usable and what limit its scope. First, it is stated for a presymplectic graph D = D Ω D=D_\Omega D = D Ω and not for an arbitrary Dirac structure D D D on N N N : the recipe D k = D Ω , T N k D_k=D_{\Omega,TN_k} D k = D Ω , T N k of Definition 11 takes a two-form and a distribution as its data, and for a general D D D there is no such Ω \Omega Ω . For a general Dirac structure the corresponding object is the clean backward image i k ! D \mathrm{i}_k^!D i k ! D along the inclusion i k : N k ↪ N \mathrm{i}_k:N_k\hookrightarrow N i k : N k ↪ N , which requires the constant-rank hypotheses of Remark 11; we do not need that generality, since every structure to which the algorithm is applied in this paper is of the form D Ω D_\Omega D Ω . Second, D k D_k D k is an ambient object: it is a maximally isotropic subbundle of ( T N ⊕ T ∗ N ) ∣ N k (TN\oplus T^*N)|_{N_k} ( T N ⊕ T ∗ N ) ∣ N k , whose covectors are still covectors of N N N , and not an intrinsic Dirac structure on N k N_k N k . This is deliberate, and is what keeps the constraint forces visible in the annihilator, as explained in Remark 11; the intrinsic structure on N k N_k N k is its image σ ( D k ) \sigma(D_k) σ ( D k ) , the graph of i k ∗ Ω \mathrm{i}_k^*\Omega i k ∗ Ω .
Applied to the Pontryagin bundle this reproduces the constraints of Section 2 with no regularity hypothesis.
Proposition 11. Let M = T ∗ P × P C M=T^*P\times_P C M = T ∗ P × P C carry Ω = pr P ∗ ω P \Omega=\operatorname{pr}_{P}^*\omega_P Ω = pr P ∗ ω P and the Pontryagin Hamiltonian H P H_P H P . Then G D Ω = ker Ω = K G_{D_\Omega}=\ker\Omega=K G D Ω = ker Ω = K is the vertical bundle of pr P \operatorname{pr}_P pr P , and the first step of (128) is
M 1 = { ( x , p , u ) ∈ M : ∂ H P ∂ u a = 0 , a = 1 , … , r } , M_1=\left\{(x,p,u)\in M:\frac{\partial H_P}{\partial u^a}=0,\ a=1,\ldots,r\right\}, M 1 = { ( x , p , u ) ∈ M : ∂ u a ∂ H P = 0 , a = 1 , … , r } , which is (7) . At the next step, G D 1 = T M 1 ∩ ( T M 1 ) ⊥ Ω = ker Ω 1 G_{D_1}=TM_1\cap(TM_1)^{\perp_{\Omega}}=\ker\Omega_1 G D 1 = T M 1 ∩ ( T M 1 ) ⊥ Ω = ker Ω 1 and
M 2 = { m ∈ M 1 : d H P ( m ) ∣ ker Ω 1 ( m ) = 0 } , M_2=\{m\in M_1:\mathrm{d}H_P(m)|_{\ker\Omega_1(m)}=0\}, M 2 = { m ∈ M 1 : d H P ( m ) ∣ k e r Ω 1 ( m ) = 0 } , which is the Gotay–Nester secondary constraint. If the problem is regular then ker Ω 1 = 0 \ker\Omega_1=0 ker Ω 1 = 0 and the algorithm stops at M 1 M_1 M 1 .
Proof. The first assertion is the computation of ker Ω \ker\Omega ker Ω in § 2 , and d H P ∣ K = 0 dH_P|_K = 0 d H P ∣ K = 0 is ∂ H P / ∂ u a = 0 \partial H_P/\partial u^a = 0 ∂ H P / ∂ u a = 0 . The second follows from (127) with Δ = T M 1 \Delta= TM_1 Δ = T M 1 , since Ω 1 = i 1 ∗ Ω \Omega_1 = i_1^*\Omega Ω 1 = i 1 ∗ Ω and hence ker Ω 1 = T M 1 ∩ ( T M 1 ) ⊥ Ω \ker\Omega_1 = TM_1 \cap(TM_1)^{\perp_\Omega} ker Ω 1 = T M 1 ∩ ( T M 1 ) ⊥ Ω . The last assertion is Proposition 1(iii). □
Along M k M_k M k the equation ( X , d H P ) ∈ D k (X,dH_P) \in D_k ( X , d H P ) ∈ D k reads, by (126),
X ∈ T M k , i X Ω − d H P ∈ ( T M k ) ∘ , (129) X \in TM_k,\qquad i_X\Omega-dH_P \in(TM_k)^\circ, \tag*{(129)} X ∈ T M k , i X Ω − d H P ∈ ( T M k ) ∘ , (129)
so the constraints enter twice, as a restriction on the admissible velocities and as a relaxation of the covector equation into the annihilator. In the regular case ( T M 1 ) ∘ (TM_1)^\circ ( T M 1 ) ∘ contributes nothing beyond the elimination of the control, and one recovers the formulation of § 2 .
Remark 13. The presymplectic algorithm for the Pontryagin Maximum Principle, including the identification of the secondary constraints with the classical Goh and generalized Legendre–Clebsch conditions, is due to Barbero-Liñán and Muñoz-Lecanda [5 ] ; what is new here is the Dirac formulation (129) of each stage and, in the next subsection, its behavior under symmetry reduction.
Reduction commutes with the constraint algorithm
We now come to the main point. The reduced object [ D Ω ] G [D_\Omega]_G [ D Ω ] G is a Dirac subbundle of the Courant algebroid E = ( T M ⊕ T ∗ M ) / G E=(TM\oplus T^*M)/G E = ( T M ⊕ T ∗ M ) / G over M / G M/G M / G , not a Dirac structure on T ( M / G ) ⊕ T ∗ ( M / G ) T(M/G)\oplus T^*(M/G) T ( M / G ) ⊕ T ∗ ( M / G ) , so Definition 12 does not apply verbatim to the reduced data ( M / G , [ D Ω ] G , H G ) (M/G,[D_\Omega]_G,H_G) ( M / G , [ D Ω ] G , H G ) : the kernel
G [ D ] G ( [ m ] ) = { [ v ] ∈ ( T M / G ) [ m ] : ( [ v ] , 0 ) ∈ [ D ] G ( [ m ] ) } \mathcal{G}_{[D]_G}([m])=\{[v]\in(TM/G)_{[m]}:([v],0)\in[D]_G([m])\} G [ D ] G ([ m ]) = {[ v ] ∈ ( T M / G ) [ m ] : ([ v ] , 0 ) ∈ [ D ] G ([ m ])}
is a subspace of the fiber ( T M / G ) [ m ] (TM/G)_{[m]} ( T M / G ) [ m ] , whereas d H G ( [ m ] ) dH_G([m]) d H G ([ m ]) is a covector on M / G M/G M / G , and the two cannot be paired directly. The pairing is supplied by the anchor map of E E E , which is the anchor of the Atiyah algebroid T M / G TM/G T M / G (cf. § 3.8 , where the case T P / G TP/G T P / G is spelled out),
a : T M / G ⟶ T ( M / G ) , a ( [ v ] ) = T π M ( v ) , a:TM/G\longrightarrow T(M/G),\qquad a([v])=T\pi_M(v), a : T M / G ⟶ T ( M / G ) , a ([ v ]) = T π M ( v ) ,
well defined by the equivariance of T π M T\pi_M T π M . The constraint algorithm for the reduced data is then the sequence
( M / G ) 0 = M / G , ( M / G ) k + 1 = { [ m ] ∈ ( M / G ) k : d H G ( [ m ] ) ∣ a ( G k ( [ m ] ) ) = 0 } , (130) (M/G)_0=M/G,\qquad(M/G)_{k+1}=\{[m]\in(M/G)_k:dH_G([m])|_{a(\mathcal{G}_k([m]))}=0\}, \tag*{(130)} ( M / G ) 0 = M / G , ( M / G ) k + 1 = {[ m ] ∈ ( M / G ) k : d H G ([ m ]) ∣ a ( G k ([ m ])) = 0 } , (130)
where G ‾ k = G D ‾ k \overline{\mathcal{G}}_k=\mathcal{G}_{\overline{D}_k} G k = G D k and D ‾ k \overline{D}_k D k is the structure induced along ( M / G ) k (M/G)_k ( M / G ) k , read fiberwise in E E E as in Definition 11: D ‾ k = D [ Ω ] G , Δ k \overline{D}_k=D_{[\Omega]_G,\Delta_k} D k = D [ Ω ] G , Δ k with two-form [ Ω ] G ( [ v ] , [ w ] ) = Ω ( v , w ) [\Omega]_G([v],[w])=\Omega(v,w) [ Ω ] G ([ v ] , [ w ]) = Ω ( v , w ) on T M / G TM/G T M / G , well defined by the invariance of Ω \Omega Ω , and distribution Δ k = a − 1 ( T ( M / G ) k ) ⊂ T M / G \Delta_k=a^{-1}(T(M/G)_k)\subset TM/G Δ k = a − 1 ( T ( M / G ) k ) ⊂ T M / G . Write ( M / G ) k (M/G)_k ( M / G ) k for the resulting sequence.
Lemma 3. Let G G G act freely and properly on M M M and let D ⊂ T M ⊕ T ∗ M D\subset TM\oplus T^*M D ⊂ T M ⊕ T ∗ M be a G G G -invariant Dirac structure. Then
G [ D ] G = [ G D ] G . \mathcal{G}_{[D]_G}=[\mathcal{G}_D]_G. G [ D ] G = [ G D ] G .
Proof. Both sides are understood pointwise, as in Remark 12. The action of G G G on T M TM T M and on T ∗ M T^*M T ∗ M is by vector bundle automorphisms, so [ v ] = 0 [v]=0 [ v ] = 0 if and only if v = 0 v=0 v = 0 , and likewise for covectors. Since the action is free, a class ( [ v ] , [ α ] ) ([v],[\alpha]) ([ v ] , [ α ]) at [ m ] [m] [ m ] has, for each choice of representative m m m , a unique representative pair ( v , α ) ∈ T m M ⊕ T m ∗ M (v,\alpha)\in T_mM\oplus T_m^*M ( v , α ) ∈ T m M ⊕ T m ∗ M , and ( [ v ] , [ α ] ) ∈ [ D ] G ([v],[\alpha])\in[D]_G ([ v ] , [ α ]) ∈ [ D ] G if and only if ( v , α ) ∈ D ( m ) (v,\alpha)\in D(m) ( v , α ) ∈ D ( m ) . Hence ( [ v ] , 0 ) ∈ [ D ] G ([v],0)\in[D]_G ([ v ] , 0 ) ∈ [ D ] G if and only if ( v , 0 ) ∈ D (v,0)\in D ( v , 0 ) ∈ D , that is, v ∈ G D v\in\mathcal{G}_D v ∈ G D . □
Theorem 5. Let ( P , C , Γ , L ) (P,C,\Gamma,L) ( P , C , Γ , L ) be a G G G -invariant optimal control problem, not assumed regular, with G G G acting freely and properly. Assume that the constraint algorithm for ( M , D Ω , H P ) (M,D_{\Omega},H_P) ( M , D Ω , H P ) produces embedded submanifolds M k M_k M k and stabilizes. Then
(i) every M k M_k M k is G G G -invariant, and the induced Dirac structures D k D_k D k are G G G -invariant;
(ii) ( M / G ) k = M k / G (M/G)_k=M_k/G ( M / G ) k = M k / G for every k k k , and D ‾ k = [ D k ] G \overline{D}_k=[D_k]_G D k = [ D k ] G : the Dirac structure carried by the k k k -th stage of the reduced algorithm is the reduction of D k D_k D k ;
(iii) consequently ( M / G ) ∞ = M ∞ / G (M/G)_\infty=M_\infty/G ( M / G ) ∞ = M ∞ / G , and the reduced dynamics on ( M / G ) ∞ (M/G)_\infty ( M / G ) ∞ is the reduction of the dynamics on M ∞ M_\infty M ∞ .
Proof. We argue by induction on the statement: M k M_k M k is G G G -invariant, ( M / G ) k = M k / G (M/G)_k=M_k/G ( M / G ) k = M k / G , and D ‾ k = [ D k ] G \overline{D}_k=[D_k]_G D k = [ D k ] G . For k = 0 k=0 k = 0 all three assertions hold trivially.
Suppose they hold at stage k k k . Then T M k TM_k T M k is a G G G -invariant distribution and, Ω \Omega Ω being G G G -invariant by Proposition 2, the induced structure D k = D Ω , T M k D_k=D_{\Omega,TM_k} D k = D Ω , T M k is G G G -invariant; this proves the second half of (i) at stage k k k . By Lemma 3, G [ D k ] G = [ G D k ] G \mathcal{G}_{[D_k]_G}=[\mathcal{G}_{D_k}]_G G [ D k ] G = [ G D k ] G , so that
a ( G ‾ k ( [ m ] ) ) = a ( G [ D k ] G ( [ m ] ) ) = T m π M ( G D k ( m ) ) . a(\overline{\mathcal{G}}_k([m]))=a(\mathcal{G}_{[D_k]_G}([m]))=T_{m}\pi_M(\mathcal{G}_{D_k}(m)). a ( G k ([ m ])) = a ( G [ D k ] G ([ m ])) = T m π M ( G D k ( m )) .
Since H P = π M ∗ H G H_P=\pi_M^*H_G H P = π M ∗ H G , one has ⟨ d H P ( m ) , v ⟩ = ⟨ d H G ( [ m ] ) , T m π M ( v ) ⟩ \langle dH_P(m),v\rangle=\langle dH_G([m]),T_m\pi_M(v)\rangle ⟨ d H P ( m ) , v ⟩ = ⟨ d H G ([ m ]) , T m π M ( v )⟩ for every v ∈ T m M v\in T_mM v ∈ T m M . Hence, for m ∈ M k m\in M_k m ∈ M k ,
d H G ( [ m ] ) ∣ a ( G k ( [ m ] ) ‾ ) = 0 ⟺ d H P ( m ) ∣ D G k ( m ) = 0 , \left.dH_G([m])\right|_{a(\overline{G_k([m])})}=0\Longleftrightarrow\left.dH_P(m)\right|_{D_{G_k(m)}}=0, d H G ([ m ]) ∣ a ( G k ([ m ]) ) = 0 ⟺ d H P ( m ) ∣ D G k ( m ) = 0 ,
so that M k + 1 M_{k+1} M k + 1 is G G G -invariant and ( M / G ) k + 1 = M k + 1 / G (M/G)_{k+1}=M_{k+1}/G ( M / G ) k + 1 = M k + 1 / G . Finally, M k + 1 M_{k+1} M k + 1 is G G G -invariant and contains the G G G -orbits of its points, so the vertical spaces lie in T M k + 1 TM_{k+1} T M k + 1 and T M k + 1 = ( T π M ) − 1 ( T ( M k + 1 / G ) ) TM_{k+1}=(T\pi_M)^{-1}(T(M_{k+1}/G)) T M k + 1 = ( T π M ) − 1 ( T ( M k + 1 / G )) along M k + 1 M_{k+1} M k + 1 ; hence Δ k + 1 = a − 1 ( T ( M / G ) k + 1 ) = T M k + 1 / G \Delta_{k+1}=a^{-1}(T(M/G)_{k+1})=TM_{k+1}/G Δ k + 1 = a − 1 ( T ( M / G ) k + 1 ) = T M k + 1 / G , and the two pointwise conditions of (125) pass to the quotient classwise, exactly as in the proof of Lemma 3, giving D ‾ k + 1 = [ D k + 1 ] G \overline{D}_{k+1}=[D_{k+1}]_G D k + 1 = [ D k + 1 ] G . This is (i) and (ii) at stage k + 1 k+1 k + 1 .
For (iii), stabilization upstairs gives stabilization downstairs by (ii), and the reduced dynamics is characterized by ( [ X ] G , [ d H P ] G ) ∈ [ D ‾ ∞ ] G ([X]_G,[dH_P]_G)\in[\overline{D}_\infty]_G ([ X ] G , [ d H P ] G ) ∈ [ D ∞ ] G , with [ d H P ] G = a ∗ d H G [dH_P]_G=a^*dH_G [ d H P ] G = a ∗ d H G the section of T ∗ M / G T^*M/G T ∗ M / G of (34), which by (ii) and the G G G -invariance of D ∞ D_\infty D ∞ is the reduction of ( X , d H P ) ∈ D ∞ (X,dH_P)\in D_\infty ( X , d H P ) ∈ D ∞ . □
Remark 14. Theorem 5 is of practical as well as structural interest: the algorithm may be run on M / G M/G M / G , whose dimension is smaller by dim G \dim G dim G , and the constraints obtained upstairs are then recovered by pulling back. In the regular case the theorem is vacuous, both algorithms stopping at the first stage; it has content precisely when the problem is singular.
Remark 15. The proof uses only that [ D Ω ] G [D_\Omega]_G [ D Ω ] G is a Dirac subbundle of the Courant algebroid E E E , the anchor map a a a of E E E , and that H G H_G H G is a function on M / G M/G M / G . It therefore does not require the identification T ∗ P / G ≃ T ∗ ( P / G ) ⊕ g ∗ T^*P/G\simeq T^*(P/G)\oplus\mathfrak{g}^* T ∗ P / G ≃ T ∗ ( P / G ) ⊕ g ∗ , and in particular no principal connection is chosen. A connection is needed only to write the resulting equations in the explicit form (26).
Remark 16. It is worth contrasting the reduced object used here with the one of Blankenstein and van der Schaft [9 ] and Blankenstein and Ratiu [8 ] . Given a symmetry group acting regularly, they project D D D to an honest Dirac structure D ^ \widehat{D} D on M / G M/G M / G , whose local sections are the pairs ( X ^ , α ^ ) (\widehat{X},\widehat{\alpha}) ( X , α ) admitting a lift ( X , α ) ∈ D (X,\alpha)\in D ( X , α ) ∈ D with α = π ∗ α ^ \alpha=\pi^*\widehat{\alpha} α = π ∗ α ; that is, they retain only those covectors which are pullbacks from the quotient [8 Prop. 5] . This is the appropriate object for a Lie–Poisson type description of the reduced dynamics, and it is a Dirac structure in the ordinary sense.
It is not, however, the object required here. Discarding the covectors that are not basic discards precisely the g ∗ \mathfrak{g}^* g ∗ component, and with it the variables ξ \xi ξ and μ \mu μ through which the reduced equations (26) are expressed; no reconstruction equation survives the passage. The quotient [ D Ω ] G [D_\Omega]_G [ D Ω ] G used in § 4 , a Dirac subbundle of the Courant algebroid E = ( T M ⊕ T ∗ M ) / G E=(TM\oplus T^*M)/G E = ( T M ⊕ T ∗ M ) / G rather than a Dirac structure on M / G M/G M / G , retains that information, which is why it is the one carried through the constraint algorithm above. The two constructions are complementary rather than competing: D ^ \widehat{D} D answers what the reduced dynamics is on M / G M/G M / G , and [ D Ω ] G [D_\Omega]_G [ D Ω ] G answers what it is on M / G M/G M / G together with the data needed to reconstruct the unreduced extremal.
A further caveat from the same source deserves mention. Restriction of an implicit Hamiltonian system to an invariant submanifold carries solutions in one direction only, there being in general no bijection between the solutions of ( M , D , H ) (M,D,H) ( M , D , H ) contained in N N N and those of ( N , D N , H N ) (N,D_N,H_N) ( N , D N , H N ) [8 Prop. 4] ; and not every solution of a reduced implicit Hamiltonian system need be the projection of a solution upstairs, although every solution is when the constraints have index one [8 Prop. 6] . Theorem 5 is a statement about the constraint manifolds and the Dirac structures carried on them, and is unaffected by these subtleties; the correspondence of individual solutions requires them.
The abnormal multiplier as a momentum map
The Pontryagin Hamiltonian H P = ⟨ p , Γ ⟩ − L H_P=\langle p,\Gamma\rangle-L H P = ⟨ p , Γ ⟩ − L used throughout corresponds to the normalization p 0 = − 1 p_0=-1 p 0 = − 1 of the cost multiplier, and therefore describes normal extremals only. Abnormal extremals are recovered by adjoining the running cost to the state, a device introduced in this context by Ohkuwa [34 ] and developed by Jóźwikowski and Respondek [25 ] , who work on the extended space Q × R Q\times\mathbb{R} Q × R of configurations and costs and pass to the projectivization of T ∗ ( Q × R ) T^*(Q\times\mathbb{R}) T ∗ ( Q × R ) , thereby treating normal and abnormal extremals uniformly by contact-geometric means. What we add is only the observation that, read through the reduction theory of this paper, the extension is itself an instance of that theory: translation in the cost variable is a symmetry whose momentum map is the multiplier p 0 p_0 p 0 , so that the normalization is a choice of level set rather than a convention.
Let
P ^ = P × R , C ^ = C × R ⟶ P ^ , \widehat{P}=P\times\mathbb{R},\qquad\widehat{C}=C\times\mathbb{R}\longrightarrow\widehat{P}, P = P × R , C = C × R ⟶ P ,
with the extra coordinate y y y recording the accumulated cost, and let
Γ ^ ( x , y , u ) = ( Γ ( x , u ) , L ( x , u ) ) , L ^ = 0. (131) \widehat{\Gamma}(x,y,u)=\bigl(\Gamma(x,u),L(x,u)\bigr),\qquad\widehat{L}=0. \tag*{(131)} Γ ( x , y , u ) = ( Γ ( x , u ) , L ( x , u ) ) , L = 0. (131)
Minimizing the terminal objective Φ ( x , y ) = y \Phi(x,y)=y Φ ( x , y ) = y evaluated at t = T t=T t = T , subject to (131) with y ( 0 ) = 0 y(0)=0 y ( 0 ) = 0 and x ( 0 ) = x 0 x(0)=x_0 x ( 0 ) = x 0 , x ( T ) = x T x(T)=x_T x ( T ) = x T , is the Mayer form of the original problem. Its Pontryagin Hamiltonian, with costate ( p , p 0 ) (p,p_0) ( p , p 0 ) , is
H ^ ( x , y , p , p 0 , u ) = ⟨ p , Γ ( x , u ) ⟩ + p 0 L ( x , u ) , (132) \widehat{H}(x,y,p,p_0,u)=\langle p,\Gamma(x,u)\rangle+p_0L(x,u), \tag*{(132)} H ( x , y , p , p 0 , u ) = ⟨ p , Γ ( x , u )⟩ + p 0 L ( x , u ) , (132)
which does not depend on y y y .
Proposition 12. The group G ^ = R \widehat{G}=\mathbb{R} G = R acting on P ^ \widehat{P} P by y ↦ y + s y\mapsto y+s y ↦ y + s , and trivially on controls, is a symmetry group of the extended optimal control problem in the sense of Definition 7. Its momentum map is J = p 0 J=p_0 J = p 0 , the bundle P ^ → P ^ / G ^ ≅ P \widehat{P}\to\widehat{P}/\widehat{G}\cong P P → P / G ≅ P is trivial with flat product connection, and g ^ \widehat{\mathfrak{g}} g is abelian. Consequently the reduced equations (26) for the extended problem read
x ˙ = ∂ H ^ ∂ p , p ˙ = − ∂ H ^ ∂ x , p ˙ 0 = 0 , ξ = ∂ H ^ ∂ p 0 = L , 0 = ∂ H ^ ∂ u , (133) \dot{x}=\frac{\partial\widehat{H}}{\partial p},\qquad
\dot{p}=-\frac{\partial\widehat{H}}{\partial x},\qquad
\dot{p}_0=0,\qquad
\xi=\frac{\partial\widehat{H}}{\partial p_0}=L,\qquad
0=\frac{\partial\widehat{H}}{\partial u}, \tag*{(133)} x ˙ = ∂ p ∂ H , p ˙ = − ∂ x ∂ H , p ˙ 0 = 0 , ξ = ∂ p 0 ∂ H = L , 0 = ∂ u ∂ H , (133) and the framework of Sections 2–4 is the restriction of (133) to the leaf p 0 = − 1 p_0=-1 p 0 = − 1 .
One point about the extension deserves to be made explicitly, because it concerns the boundary conditions and not the interior equations. Translation in y y y is a symmetry of the dynamics (131) and of L ^ = 0 \widehat{L}=0 L = 0 , hence of the interior variational problem; it is not a symmetry of the Mayer boundary problem, whose terminal objective Φ = y \Phi=y Φ = y and whose initial condition y ( 0 ) = 0 y(0)=0 y ( 0 ) = 0 both single out a value of y y y . That is as it should be: the PMP transversality condition for the terminal cost Φ \Phi Φ , namely p 0 ( T ) = − ∂ Φ / ∂ y = − 1 p_0(T)=-\partial\Phi/\partial y=-1 p 0 ( T ) = − ∂ Φ/ ∂ y = − 1 in the normal case, is precisely what selects the leaf p 0 = − 1 p_0=-1 p 0 = − 1 and fixes the sign and normalization of the multiplier, and no transversality condition is imposed on p ( T ) p(T) p ( T ) because x ( T ) x(T) x ( T ) is prescribed. The abnormal case p 0 = 0 p_0=0 p 0 = 0 is the alternative permitted when the nontriviality requirement ( p , p 0 ) ≠ 0 (p,p_0)\ne0 ( p , p 0 ) = 0 is met by p p p alone. Conservation of p 0 p_0 p 0 along the flow, which is the content of (133), is a consequence of the interior symmetry and is unaffected by this remark.
Proof. Invariance is immediate from (131), since neither Γ \Gamma Γ nor L L L depends on y y y and L ^ = 0 \widehat{L}=0 L = 0 . The fundamental vector field of ζ ∈ g ^ = R \zeta\in\widehat{\mathfrak{g}}=\mathbb{R} ζ ∈ g = R is ζ ∂ y \zeta\partial_y ζ ∂ y , so ⟨ J , ζ ⟩ = ⟨ ( p , p 0 ) , ζ ∂ y ⟩ = ζ p 0 \langle J,\zeta\rangle=\langle(p,p_0),\zeta\partial_y\rangle=\zeta p_0 ⟨ J , ζ ⟩ = ⟨( p , p 0 ) , ζ ∂ y ⟩ = ζ p 0 . Both the curvature and the coadjoint terms in (26) vanish, the first because the product connection on P × R P\times\mathbb{R} P × R is flat and the second because g ^ \widehat{\mathfrak{g}} g is abelian, which gives (133). Restricting to p 0 = − 1 p_0=-1 p 0 = − 1 turns (132) into ⟨ p , Γ ⟩ − L = H P \langle p,\Gamma\rangle-L=H_P ⟨ p , Γ ⟩ − L = H P . □
Thus reducing the Mayer problem by the translation symmetry of the cost variable is what converts it to Lagrange form; the reconstruction variable ξ = L \xi=L ξ = L is the running cost, and reconstruction recovers the accumulated cost. Because the reduction used here is of Poisson rather than symplectic type, all levels of p 0 p_0 p 0 are described simultaneously, the level appearing as a coordinate on the reduced space. In particular:
Definition 13. The abnormal extremals of ( P , C , Γ , L ) (P,C,\Gamma,L) ( P , C , Γ , L ) are the solutions of (133) lying on the leaf p 0 = 0 p_0=0 p 0 = 0 , with p ≠ 0 p\ne0 p = 0 .
On that leaf H ^ = ⟨ p , Γ ⟩ \widehat{H}=\langle p,\Gamma\rangle H = ⟨ p , Γ ⟩ carries no cost term, the Hessian in the controls is ⟨ p , ∂ 2 Γ / ∂ u 2 ⟩ \langle p,\partial^2\Gamma/\partial u^2\rangle ⟨ p , ∂ 2 Γ/ ∂ u 2 ⟩ , and this vanishes identically whenever Γ \Gamma Γ is affine in u u u . For control-affine systems, abnormal extremals are therefore always singular in the sense of Section 2, and it is exactly here that the algorithm of Section 7.2 is required. The restriction is necessary: for a Γ \Gamma Γ depending nonlinearly on u u u an abnormal extremal may be regular. With P = R P=\mathbb{R} P = R and Γ ( x , u ) = ( 1 + u 2 ) ∂ x \Gamma(x,u)=(1+u^2)\partial_x Γ ( x , u ) = ( 1 + u 2 ) ∂ x one has H ^ = p ( 1 + u 2 ) \widehat{H}=p(1+u^2) H = p ( 1 + u 2 ) on p 0 = 0 p_0=0 p 0 = 0 , so H ^ u = 2 p u \widehat{H}_u=2pu H u = 2 p u vanishes at u = 0 u=0 u = 0 while H ^ u u = 2 p ≠ 0 \widehat{H}_{uu}=2p\ne0 H uu = 2 p = 0 ; the curve x ˙ = 1 \dot{x}=1 x ˙ = 1 with p p p a nonzero constant is an abnormal extremal through a regular point of M 1 M_1 M 1 . All the systems treated below, namely (134), are control-linear, so the hypothesis is in force throughout the rest of this section.
Control-linear systems: abnormality and curvature
We now specialize to the situation in which the admissible velocities form a distribution, and show that the constraint algorithm is governed by the curvature of the connection.
Let ρ : P → B = P / G \rho:P\to B=P/G ρ : P → B = P / G be a principal G G G -bundle with connection A A A and curvature F A F_A F A , let
D = ker A ⊂ T P \mathcal{D}=\ker A\subset TP D = ker A ⊂ T P
be the horizontal distribution, and consider the driftless control-linear system obtained by taking C = ρ ∗ U ^ C=\rho^*\widehat{U} C = ρ ∗ U with U ^ ≅ R dim B \widehat{U}\cong\mathbb{R}^{\dim B} U ≅ R d i m B and
Γ ( x , u ) = ∑ a u a X a h ( x ) , (134) \Gamma(x,u)=\sum_a u^aX_a^h(x), \tag*{(134)} Γ ( x , u ) = a ∑ u a X a h ( x ) , (134)
where X a h X_a^h X a h are the horizontal lifts of a local frame X a X_a X a of T B TB T B . The sub-Riemannian problem of Section 6 is of this form, and so is the Martinet problem below.
Convention 1. Unless the contrary is stated, admissible curves of (134) are assumed immersed, that is, u ( t ) ≠ 0 u(t) \ne0 u ( t ) = 0 for every t t t ; whenever a metric is present they are assumed parametrized by arclength. Constant curves, for which every covector in D ∘ D^\circ D ∘ is trivially an abnormal multiplier, are thereby excluded, and so are reparametrizations that stop. The convention is used at three places below — the constraint chain of § 8.6 , Corollary 2 and Proposition 15 — and we recall it at each. Theorem 6, by contrast, is deliberately stated without it, so that the reader can see exactly where the hypothesis is needed.
Proposition 13. For the system (134) the abnormal primary constraint manifold is the annihilator of the horizontal distribution,
M 1 ∣ p 0 = 0 = D ∘ × P C , \left.M_1\right|_{p_0=0}=D^\circ\times_P C, M 1 ∣ p 0 = 0 = D ∘ × P C , and under the identification (18) determined by A A A this is precisely the leaf
of the reduced Pontryagin bundle: the base costate of an abnormal extremal vanishes identically, and the extremal is described by the coadjoint variable μ \mu μ alone.
Proof. On p 0 = 0 p_0=0 p 0 = 0 we have H ^ = ∑ a u a ⟨ p , X a h ⟩ \widehat{H}=\sum_a u^a\langle p,X_a^h\rangle H = ∑ a u a ⟨ p , X a h ⟩ , so the primary constraints ∂ H ^ / ∂ u a = 0 \partial\widehat{H}/\partial u^a=0 ∂ H / ∂ u a = 0 read ⟨ p , X a h ⟩ = 0 \langle p,X_a^h\rangle=0 ⟨ p , X a h ⟩ = 0 for all a a a , that is, p ∈ D ∘ p\in D^\circ p ∈ D ∘ . In the splitting (18) the component p z p_z p z is the restriction of p p p to the horizontal subspace, so p ∈ D ∘ p\in D^\circ p ∈ D ∘ is exactly p z = 0 p_z=0 p z = 0 , and μ \mu μ , which is determined by p = A x ∗ μ p=A_x^*\mu p = A x ∗ μ , that is, by ⟨ p , v ⟩ = ⟨ μ , A x ( v ) ⟩ \langle p,v\rangle=\langle\mu,A_x(v)\rangle ⟨ p , v ⟩ = ⟨ μ , A x ( v )⟩ for v ∈ T x P v\in T_xP v ∈ T x P , is unconstrained. (Note the direction: A x : T x P → g A_x:T_xP\to\mathfrak{g} A x : T x P → g , so A x ∗ : g ∗ → T x ∗ P A_x^*:\mathfrak{g}^*\to T_x^*P A x ∗ : g ∗ → T x ∗ P , and it is p p p that is the pullback of μ \mu μ , not the reverse. Equivalently μ = J ( p ) \mu=J(p) μ = J ( p ) is the value of the momentum map.) The identification uses D = ker A D=\ker A D = ker A , so that a covector annihilating D D D is determined by its values on the vertical subspace.
Lemma 4. For horizontal lifts, A ( [ X h , Y h ] ) = − F A ( X , Y ) A([X^h,Y^h])=-F_A(X,Y) A ([ X h , Y h ]) = − F A ( X , Y ) . Consequently, if p ∈ D ∘ p\in D^\circ p ∈ D ∘ corresponds to μ \mu μ under (18),
⟨ p , [ X h , Y h ] ⟩ = − ⟨ μ , F A ( X , Y ) ⟩ . (135) \langle p,[X^h,Y^h]\rangle=-\langle\mu,F_A(X,Y)\rangle. \tag*{(135)} ⟨ p , [ X h , Y h ]⟩ = − ⟨ μ , F A ( X , Y )⟩ . (135)
Proof. Since A ( X h ) = A ( Y h ) = 0 A(X^h)=A(Y^h)=0 A ( X h ) = A ( Y h ) = 0 ,
F A ( X h , Y h ) = d A ( X h , Y h ) + [ A ( X h ) , A ( Y h ) ] = − A ( [ X h , Y h ] ) , F_A(X^h,Y^h)=dA(X^h,Y^h)+[A(X^h),A(Y^h)]=-A([X^h,Y^h]), F A ( X h , Y h ) = d A ( X h , Y h ) + [ A ( X h ) , A ( Y h )] = − A ([ X h , Y h ]) ,
and F A F_A F A is basic, so F A ( X h , Y h ) = F A ( X , Y ) F_A(X^h,Y^h)=F_A(X,Y) F A ( X h , Y h ) = F A ( X , Y ) . For p ∈ D ∘ p\in D^\circ p ∈ D ∘ one has ⟨ p , v ⟩ = ⟨ μ , A ( v ) ⟩ \langle p,v\rangle=\langle\mu,A(v)\rangle ⟨ p , v ⟩ = ⟨ μ , A ( v )⟩ for every v v v , whence (135) .
Theorem 6. For the system (134) , write
C μ ( z ) = ⟨ μ , F A ( ⋅ , ⋅ ) ⟩ (136) C_\mu(z)=\langle\mu,F_A(\mathord\cdot,\mathord\cdot)\rangle \tag*{(136)} C μ ( z ) = ⟨ μ , F A ( ⋅ , ⋅ )⟩ (136) a skew-symmetric bilinear form on T z B T_zB T z B . The first secondary constraint of the algorithm on the abnormal leaf is that the control take values in the kernel of C μ C_\mu C μ :
u ( t ) ∈ ker C μ ( t ) ( z ( t ) ) , that is, C μ ( t ) ( u ( t ) , ⋅ ) = 0. (137) u(t)\in\ker C_{\mu(t)}(z(t)),\qquad\text{that is,}\qquad C_{\mu(t)}(u(t),\mathord\cdot)=0. \tag*{(137)} u ( t ) ∈ ker C μ ( t ) ( z ( t )) , that is, C μ ( t ) ( u ( t ) , ⋅ ) = 0. (137) At any time at which u ( t ) ≠ 0 u(t)\ne0 u ( t ) = 0 , and only then, (137) forces C μ ( t ) C_{\mu(t)} C μ ( t ) to be degenerate. In particular:
if C μ C_\mu C μ is nondegenerate for every μ ≠ 0 \mu\ne0 μ = 0 at a point, then no abnormal extremal with nonvanishing control — equivalently, with nonvanishing speed — passes through that point;
if dim B = 2 \dim B=2 dim B = 2 then, along an immersed curve, (137) reduces to ⟨ μ , F A ⟩ = 0 \langle\mu,F_A\rangle=0 ⟨ μ , F A ⟩ = 0 , which is the classical Goh condition; if moreover dim g = 1 \dim\mathfrak{g}=1 dim g = 1 , every immersed abnormal extremal projects into the locus { F A = 0 } \{F_A=0\} { F A = 0 } where the curvature vanishes.
Proof. On the abnormal leaf H ^ = ∑ a u a h a \widehat{H}=\sum_a u^a h_a H = ∑ a u a h a with h a = ⟨ p , X a h ⟩ h_a=\langle p,X_a^h\rangle h a = ⟨ p , X a h ⟩ , and M 1 = { h a = 0 } M_1=\{h_a=0\} M 1 = { h a = 0 } by Proposition 13. Tangency of the dynamics to M 1 M_1 M 1 requires h ˙ a = ∑ b u b { h a , h b } = 0 \dot h_a=\sum_bu^b\{h_a,h_b\}=0 h ˙ a = ∑ b u b { h a , h b } = 0 for each a a a , and { h a , h b } = − ⟨ p , [ X a h , X b h ] ⟩ \{h_a,h_b\}=-\langle p,[X_a^h,X_b^h]\rangle { h a , h b } = − ⟨ p , [ X a h , X b h ]⟩ , which by Lemma 4 equals ⟨ μ , F A ( X a , X b ) ⟩ \langle\mu,F_A(X_a,X_b)\rangle ⟨ μ , F A ( X a , X b )⟩ . Hence tangency reads C μ ( u , ⋅ ) = 0 C_\mu(u,\mathord\cdot)=0 C μ ( u , ⋅ ) = 0 , that is, (137) ; a nonvanishing control satisfying this occurs only where C μ C_\mu C μ is degenerate, which is Assertion (i) in contrapositive form. Nothing is asserted at instants where u ( t ) = 0 u(t)=0 u ( t ) = 0 , and indeed a constant curve with u = 0 u=0 u = 0 and any $0\ne p\in D^\circs a t i s f i e s ( 133 ) o n satisfies (133) on s a t i s f i es ( 133 ) o n p_0=0 vacuously. For (ii), a skew form on a two-dimensional space is a multiple of the area form, so it is degenerate if and only if it vanishes; and if in addition g = R \mathfrak{g}=\mathbb{R} g = R then C μ = μ F A C_\mu=\mu F_A C μ = μ F A with μ ≠ 0 \mu\ne0 μ = 0 on a nontrivial abnormal extremal, since μ = 0 \mu=0 μ = 0 together with p z = 0 p_z=0 p z = 0 would give p = 0 p=0 p = 0 .
Remark 17. Four comments place Theorem 6 in context. It concerns abnormal extremals, not strictly abnormal ones in the sense of Definition 14: an abnormal extremal may carry a normal lift as well, as in the Martinet problem below, and deciding which abnormal extremals are strict requires more than the curvature and is taken up in § 8.8 . Condition (i) is the classical criterion of fatness, that
a distribution for which C μ C_\mu C μ is nondegenerate for all μ ≠ 0 \mu\ne0 μ = 0 admits no nontrivial abnormal extremals [2 , 33 ] , recovered here as the nondegeneracy of the curvature pairing appearing in (26) . Intrinsic criteria valid for an arbitrary distribution are available, notably that of Jóźwikowski and Respondek [25 Thm. 5, Cor. 1] , which characterizes the abnormal sub-Riemannian extremals as the curves along which the smallest flow-invariant distribution containing D D D has rank less than dim P \dim P dim P , a condition on the iterated brackets ad X k ( Γ ( D ) ) \operatorname{ad}_{\mathcal{X}}^k(\Gamma(D)) ad X k ( Γ ( D )) for smooth fields; for D = ker A D = \ker A D = ker A the iterated brackets are generated by the curvature and the criterion becomes the degeneracy of C μ C_\mu C μ , so the content here is not a new criterion but the identification of the obstruction with the curvature term of (26) . Finally, (137) is the form taken along the extremal by the classical Goh condition [2 , 5 ] , with which it coincides when dim B = 2 \dim B = 2 dim B = 2 ; in general the Goh condition requires C μ C_\mu C μ to vanish on all of T z B T_zB T z B , of which (137) retains the part enforced by tangency, namely the vanishing of the term ⟨ μ , F A ( z ˙ , ⋅ ) ⟩ \langle\mu,F_A(\dot z,\cdot)\rangle ⟨ μ , F A ( z ˙ , ⋅ )⟩ to which the reduced costate equation (26) reduces on the abnormal leaf, where p z ≡ 0 p_z \equiv0 p z ≡ 0 and h ≡ 0 h \equiv0 h ≡ 0 . The curvature term of the reduced Pontryagin Maximum Principle is thus simultaneously the term that distinguishes it from the unreduced principle and the term that obstructs abnormality.
Corollary 1. The sub-Riemannian problem on the Heisenberg group of Section 6 has no nontrivial abnormal extremals.
Proof. There g = R \mathfrak{g} = \mathbb{R} g = R and F A = − d x ∧ d y F_A = -dx \wedge dy F A = − d x ∧ d y is nowhere zero, so Theorem 6(i) applies with empty locus. □
Example: the Martinet problem
The complementary example is the Martinet distribution, for which the curvature degenerates on a hypersurface. Let P = R 3 P = \mathbb{R}^3 P = R 3 with coordinates ( x , y , z ) (x,y,z) ( x , y , z ) , let
A = d z − x 2 2 d y , F A = d A = − x d x ∧ d y , (138) A = dz - \frac{x^2}{2}dy,\qquad F_A = dA = -x\,dx\wedge dy, \tag*{(138)} A = d z − 2 x 2 d y , F A = d A = − x d x ∧ d y , (138)
and let G = R G = \mathbb{R} G = R act by translation in z z z , so that B = R 2 B = \mathbb{R}^2 B = R 2 and μ = p z \mu= p_z μ = p z . The horizontal lifts of ∂ x , ∂ y \partial_x,\partial_y ∂ x , ∂ y are
X = ∂ ∂ x , Y = ∂ ∂ y + x 2 2 ∂ ∂ z , [ X , Y ] = x ∂ ∂ z , X = \frac{\partial}{\partial x},\qquad Y = \frac{\partial}{\partial y} + \frac{x^2}{2}\frac{\partial}{\partial z},\qquad[X,Y] = x\frac{\partial}{\partial z}, X = ∂ x ∂ , Y = ∂ y ∂ + 2 x 2 ∂ z ∂ , [ X , Y ] = x ∂ z ∂ ,
and A ( [ X , Y ] ) = x = − F A ( ∂ x , ∂ y ) A([X,Y]) = x = -F_A(\partial_x,\partial_y) A ([ X , Y ]) = x = − F A ( ∂ x , ∂ y ) , in agreement with Lemma 4. The corresponding optimal control problem, with Γ = u 1 X + u 2 Y \Gamma= u_1X + u_2Y Γ = u 1 X + u 2 Y and L = 1 2 ∣ u ∣ 2 L = \frac{1}{2}|u|^2 L = 2 1 ∣ u ∣ 2 , is the Martinet sub-Riemannian problem [33 , 1 ] .
Remark 18. Conventions differ. We place the Martinet surface at { x = 0 } \{x = 0\} { x = 0 } , whereas [1 ] use ω = d z − 1 2 y 2 d x \omega= dz - \frac{1}{2}y^2dx ω = d z − 2 1 y 2 d x and place it at { y = 0 } \{y = 0\} { y = 0 } . The transposition σ ( x , y , z ) = ( y , x , z ) \sigma(x,y,z) = (y,x,z) σ ( x , y , z ) = ( y , x , z ) satisfies σ ∗ ω = A \sigma^*\omega= A σ ∗ ω = A and is an isometry of the flat metric d x 2 + d y 2 dx^2+dy^2 d x 2 + d y 2 , so the two models are the same problem; being orientation-reversing on the base it changes the sign of F A F_A F A , which accounts for the sign discrepancies between the formulas below and those of [1 ] .
Unlike the Heisenberg case, F A F_A F A vanishes on the Martinet surface { x = 0 } \{x = 0\} { x = 0 } , and Theorem 6 predicts that abnormal extremals are confined to it. The algorithm confirms this. On the abnormal leaf p 0 = 0 p_0 = 0 p 0 = 0 ,
h X = p x , h Y = p y + x 2 2 p z , H ^ = u 1 h X + u 2 h Y , h_X = p_x,\qquad h_Y = p_y + \frac{x^2}{2}p_z,\qquad\widehat{H} = u_1h_X + u_2h_Y, h X = p x , h Y = p y + 2 x 2 p z , H = u 1 h X + u 2 h Y ,
so that
M 1 = { h X = h Y = 0 } , { h X , h Y } = − x p z , M_1 = \{h_X = h_Y = 0\},\qquad\{h_X,h_Y\} = -xp_z, M 1 = { h X = h Y = 0 } , { h X , h Y } = − x p z ,
and tangency gives h ˙ X = − u 2 x p z \dot h_X = -u_2xp_z h ˙ X = − u 2 x p z , h ˙ Y = u 1 x p z \dot h_Y = u_1xp_z h ˙ Y = u 1 x p z . The secondary constraint is therefore
M 2 = { u 1 x p z = u 2 x p z = 0 } ∩ M 1 , M_2 = \{u_1xp_z = u_2xp_z = 0\}\cap M_1, M 2 = { u 1 x p z = u 2 x p z = 0 } ∩ M 1 ,
which is (137) for this connection. As a subset of M M M this is the union { x p z = 0 } ∪ { u = 0 } \{xp_z = 0\}\cup\{u = 0\} { x p z = 0 } ∪ { u = 0 } , which is not a submanifold; by Convention 1 we work on the immersed branch u ≠ 0 u \ne0 u = 0 , where it is
M 2 = { x p z = 0 } ∩ M 1 . M_2 = \{xp_z = 0\}\cap M_1. M 2 = { x p z = 0 } ∩ M 1 .
On M 1 M_1 M 1 the vanishing of p z p_z p z would force p y = 0 p_y = 0 p y = 0 and hence p = 0 p = 0 p = 0 , so p z ≠ 0 p_z \ne0 p z = 0 and
Tangency of { x = 0 } \{x = 0\} { x = 0 } requires x ˙ = u 1 = 0 \dot x = u_1 = 0 x ˙ = u 1 = 0 , after which p ˙ x = − u 2 x p z = 0 \dot p_x = -u_2xp_z = 0 p ˙ x = − u 2 x p z = 0 and z ˙ = u 2 x 2 / 2 = 0 \dot z = u_2x^2/2 = 0 z ˙ = u 2 x 2 /2 = 0 , and the algorithm stabilizes. Note that the control constraint u 1 = 0 u_1 = 0 u 1 = 0 is part of the final constraint manifold, which lives in M = T ∗ P × P C M = T^*P\times_P C M = T ∗ P × P C and therefore carries u u u among its coordinates. The abnormal extremal is
x ≡ 0 , p x = p y ≡ 0 , p z = const ≠ 0 , z = const , x \equiv0,\qquad p_x = p_y \equiv0,\qquad p_z = \operatorname{const}\ne0,\qquad z = \operatorname{const}, x ≡ 0 , p x = p y ≡ 0 , p z = const = 0 , z = const ,
with y y y traversed at the arbitrary rate u 2 u_2 u 2 : the Martinet line, recovered here as the final constraint manifold of the algorithm.
Since all the data are invariant under translation in z z z , and the whole chain was computed from μ = p z \mu=p_z μ = p z and F A F_A F A alone, the computation may equally be performed on M / G M/G M / G ; by Theorem 5 the two answers agree, the reduced final constraint manifold being { x = 0 , p x = p y = 0 , u 1 = 0 } ⊂ M / G \{x=0,\ p_x=p_y=0,\ u_1=0\}\subset M/G { x = 0 , p x = p y = 0 , u 1 = 0 } ⊂ M / G , with μ \mu μ a nonzero parameter and u 2 u_2 u 2 a free nonzero parameter, by Convention 1.
The Martinet normal geodesics
The abnormal leaf occupied us above; we now treat the normal one, p 0 = − 1 p_0=-1 p 0 = − 1 . This is the first place in the paper where the curvature term of (26) is neither zero nor constant, and it is what makes the reduced equations interesting: they turn out to be those of a pendulum.
Since the controlled vector field is horizontal, γ ~ = 0 \tilde{\gamma}=0 γ ~ = 0 and the reduced Pontryagin Hamiltonian is
h ( z , p z , μ , u ) = ⟨ p z , u ⟩ − 1 2 ∣ u ∣ 2 , (139) h(z,p_z,\mu,u)=\langle p_z,u\rangle-\frac{1}{2}|u|^2, \tag*{(139)} h ( z , p z , μ , u ) = ⟨ p z , u ⟩ − 2 1 ∣ u ∣ 2 , (139)
independent of z z z and of μ \mu μ , exactly as in § 7 . The stationarity condition gives u = p z u=p_z u = p z , and (26) becomes, using F A = − x d x ∧ d y F_A=-x\,dx\wedge dy F A = − x d x ∧ d y ,
{ x ˙ = p x , p ˙ x = − μ x p y , y ˙ = p y , p ˙ y = + μ x p x , μ ˙ = 0 , ξ = 0. (140) \begin{cases}
\dot{x}=p_x,\quad\dot{p}_x=-\mu x p_y,\\
\dot{y}=p_y,\quad\dot{p}_y=+\mu x p_x,\\
\dot{\mu}=0,\quad\xi=0. \tag*{(140)}
\end{cases} ⎩ ⎨ ⎧ x ˙ = p x , p ˙ x = − μx p y , y ˙ = p y , p ˙ y = + μx p x , μ ˙ = 0 , ξ = 0. (140)
The contrast with the Heisenberg equations (118) is the whole point: there the momentum p z p_z p z rotated at the constant rate μ \mu μ , here at the rate μ x \mu x μx . The curvature is no longer a constant multiple of the area form, and the reduced system ceases to be linear.
Proposition 14. Along solutions of (140) both
∣ p z ∣ = ∣ u ∣ and p y − 1 2 μ x 2 |p_z|=|u| \qquad\text{and}\qquad p_y-\frac{1}{2}\mu x^2 ∣ p z ∣ = ∣ u ∣ and p y − 2 1 μ x 2 are constant. The first states that normal extremals are parametrized proportionally to arclength. The second is the unreduced costate λ y \lambda_y λ y , conserved because the Martinet structure is invariant under translation in y y y as well as in z z z .
Proof. d d t ( p x 2 + p y 2 ) = 2 p x ( − μ x p y ) + 2 p y ( μ x p x ) = 0 \frac{d}{dt}(p_x^2+p_y^2)=2p_x(-\mu xp_y)+2p_y(\mu xp_x)=0 d t d ( p x 2 + p y 2 ) = 2 p x ( − μx p y ) + 2 p y ( μx p x ) = 0 , and d d t ( p y − 1 2 μ x 2 ) = μ x p x − μ x x ˙ = 0 \frac{d}{dt}\left(p_y-\frac{1}{2}\mu x^2\right)=\mu xp_x-\mu x\dot{x}=0 d t d ( p y − 2 1 μ x 2 ) = μx p x − μx x ˙ = 0 . For the identification, recall from Section 7.6 that p y = λ y + 1 2 x 2 λ z p_y=\lambda_y+\frac{1}{2}x^2\lambda_z p y = λ y + 2 1 x 2 λ z and μ = λ z \mu=\lambda_z μ = λ z , whence p y − 1 2 μ x 2 = λ y p_y-\frac{1}{2}\mu x^2=\lambda_y p y − 2 1 μ x 2 = λ y . □
The second conserved quantity is worth a comment. In the variables ( z , p z , μ ) (z,p_z,\mu) ( z , p z , μ ) adapted to the R z R_z R z -reduction it is not manifest; the term 1 2 μ x 2 \frac{1}{2}\mu x^2 2 1 μ x 2 is precisely the bookkeeping of the splitting (18) . A symmetry that has not been quotiented is present in the reduced equations, but only in a form that the connection makes visible.
Theorem 7. Write p z = r ( cos θ , sin θ ) p_z=r(\cos\theta,\sin\theta) p z = r ( cos θ , sin θ ) with r = ∣ u ∣ r=|u| r = ∣ u ∣ constant. Then along solutions of (140)
θ ˙ = μ x , and hence θ ¨ = μ r cos θ . (141) \dot{\theta}=\mu x,\qquad\text{and hence}\qquad\ddot{\theta}=\mu r\cos\theta. \tag*{(141)} θ ˙ = μx , and hence θ ¨ = μ r cos θ . (141) Equivalently, setting φ = θ − π / 2 \varphi=\theta-\pi/2 φ = θ − π /2 and ω 2 = μ r \omega^2=\mu r ω 2 = μ r ,
φ ¨ = − ω 2 sin φ , (142) \ddot{\varphi}=-\omega^2\sin\varphi, \tag*{(142)} φ ¨ = − ω 2 sin φ , (142) the equation of a simple pendulum. The base curve is recovered from x ˙ = r cos θ \dot{x}=r\cos\theta x ˙ = r cos θ , y ˙ = r sin θ \dot{y}=r\sin\theta y ˙ = r sin θ , and the fiber coordinate by the quadrature z ˙ = 1 2 x 2 y ˙ \dot{z}=\frac{1}{2}x^2\dot{y} z ˙ = 2 1 x 2 y ˙ furnished by ξ = 0 \xi=0 ξ = 0 .
Proof. Differentiating p x = r cos θ p_x=r\cos\theta p x = r cos θ gives − r sin θ θ ˙ = − μ x p y = − μ x r sin θ -r\sin\theta\,\dot{\theta}=-\mu xp_y=-\mu xr\sin\theta − r sin θ θ ˙ = − μx p y = − μx r sin θ , so θ ˙ = μ x \dot{\theta}=\mu x θ ˙ = μx ; the same follows from the equation for p y p_y p y . Then θ ¨ = μ x ˙ = μ p x = μ r cos θ \ddot{\theta}=\mu\dot{x}=\mu p_x=\mu r\cos\theta θ ¨ = μ x ˙ = μ p x = μ r cos θ . The substitution φ = θ − π / 2 \varphi=\theta-\pi/2 φ = θ − π /2 turns cos θ \cos\theta cos θ into − sin φ -\sin\varphi − sin φ . The reconstruction equation is A ( x ˙ ) = ξ = 0 A(\dot{x})=\xi=0 A ( x ˙ ) = ξ = 0 , that is, z ˙ = 1 2 x 2 y ˙ \dot{z}=\frac{1}{2}x^2\dot{y} z ˙ = 2 1 x 2 y ˙ . □
Corollary 2. Let r > 0 r>0 r > 0 and μ > 0 \mu>0 μ > 0 . For a librating solution of (142) with modulus k < 1 k<1 k < 1 ,
x ( t ) = 2 k r μ cn ( μ r ( t − t 0 ) , k ) , (143) x(t)=2k\sqrt{\frac{r}{\mu}}\,\operatorname{cn}\left(\sqrt{\mu r}(t-t_0),k\right), \tag*{(143)} x ( t ) = 2 k μ r cn ( μ r ( t − t 0 ) , k ) , (143) periodic with period 4 K ( k ) / μ r 4K(k)/\sqrt{\mu r} 4 K ( k ) / μ r , where K K K is the complete elliptic integral of the first kind. One recovers in this way the classical description of the Martinet geodesics by elliptic functions [33 , 1 ] : with the arclength normalization r = 1 r=1 r = 1 the amplitude 2 k r / μ 2k\sqrt{r/\mu} 2 k r / μ and the period 4 K ( k ) / μ r 4K(k)/\sqrt{\mu r} 4 K ( k ) / μ r are those of [1 §4.5] , where the same solution has − 2 k r / μ cn ( K ( k ) + t λ , k ) -2k\sqrt{r/\mu}\operatorname{cn}(K(k)+t\sqrt{\lambda},k) − 2 k r / μ cn ( K ( k ) + t λ , k ) with λ = p z \lambda=p_z λ = p z , the difference in sign and phase being the transposition of Remark 18 together with a choice of origin.
The two excluded signs of μ \mu μ behave differently. For μ < 0 \mu<0 μ < 0 the roles of the two equilibria of (142) are exchanged: the libration takes place about φ = π \varphi=\pi φ = π , and (143) is replaced by
x ( t ) = − 2 k r ∣ μ ∣ cn ( ∣ μ ∣ r ( t − t 0 ) , k ) x(t)=-2k\sqrt{\frac{r}{|\mu|}}\,\operatorname{cn}\left(\sqrt{|\mu|r}(t-t_0),k\right) x ( t ) = − 2 k ∣ μ ∣ r cn ( ∣ μ ∣ r ( t − t 0 ) , k )
of period 4 K ( k ) / ∣ μ ∣ r 4K(k)/\sqrt{|\mu|r} 4 K ( k ) / ∣ μ ∣ r ; the expression (143) itself is complex there. For μ = 0 \mu= 0 μ = 0 the pendulum degenerates, since ω 2 = μ r = 0 \omega^2 = \mu r = 0 ω 2 = μ r = 0 and the substitution x = θ ˙ / μ x = \dot{\theta}/\mu x = θ ˙ / μ used in the proof is unavailable: the direction θ \theta θ of the momentum is constant, and the base curve is a straight line. The case r = 0 r = 0 r = 0 is excluded by Convention 1, since r = ∣ u ∣ r = |u| r = ∣ u ∣ .
Proof. Since x = θ ˙ / μ = φ ˙ / μ x = \dot{\theta}/\mu= \dot{\varphi}/\mu x = θ ˙ / μ = φ ˙ / μ and the librating pendulum has φ ˙ = 2 k ω cn ( ω ( t − t 0 ) , k ) \dot{\varphi} = 2k\omega\operatorname{cn}(\omega(t-t_0),k) φ ˙ = 2 k ω cn ( ω ( t − t 0 ) , k ) with ω = μ r \omega= \sqrt{\mu r} ω = μ r , the amplitude is 2 k ω / μ = 2 k r / μ 2k\omega/\mu= 2k\sqrt{r/\mu} 2 k ω / μ = 2 k r / μ and the period is that of cn \operatorname{cn} cn .
The abnormal line seen from the normal leaf. The set { x = 0 , p x = 0 } \{x = 0, p_x = 0\} { x = 0 , p x = 0 } is invariant under (140), since x ˙ = p x \dot{x} = p_x x ˙ = p x and p ˙ x = − μ x p y \dot{p}_x = -\mu x p_y p ˙ x = − μx p y both vanish on it. On that set p z = ( 0 , ± r ) p_z = (0,\pm r) p z = ( 0 , ± r ) and the base curve is the y y y -axis: it is the Martinet line of § 8.6 .
Proposition 15. The abnormal extremal of § 8.6 , traversed at constant speed, is also a normal extremal; the Martinet line is therefore not strictly abnormal as an unparametrized curve. The parametrization qualifier is needed: the abnormal computation of § 8.6 left the rate u 2 ( t ) u_2(t) u 2 ( t ) along the Martinet line entirely arbitrary, whereas a normal lift of the quadratic cost has u = p z u = p_z u = p z with ∣ p z ∣ |p_z| ∣ p z ∣ constant by Proposition 14, so an abnormal traversal at non-constant speed carries no normal lift. This is the arclength convention of Convention 1 and of Theorem 8. In the pendulum (142) it appears as the two equilibria φ = 0 \varphi= 0 φ = 0 and φ = π \varphi= \pi φ = π , corresponding to the line traversed in the directions + y +y + y and − y -y − y respectively. For fixed μ > 0 \mu> 0 μ > 0 the first is a center and the second a saddle, so the two directions along the Martinet line are not equivalent.
Proof. Both equilibria lie in the invariant set { x = 0 , p x = 0 } \{x = 0, p_x = 0\} { x = 0 , p x = 0 } and satisfy (140) with p 0 = − 1 p_0 = -1 p 0 = − 1 , so the corresponding curves are normal extremals; by § 8.6 the same base curve carries an abnormal extremal. Linearizing (142) gives δ φ ¨ = − ω 2 δ φ \ddot{\delta\varphi} = -\omega^2\delta\varphi δ φ ¨ = − ω 2 δ φ at φ = 0 \varphi= 0 φ = 0 and δ φ ¨ = + ω 2 δ φ \ddot{\delta\varphi} = +\omega^2\delta\varphi δ φ ¨ = + ω 2 δ φ at φ = π \varphi= \pi φ = π .
The first assertion is Lemma 2.7 of [1 ] , obtained there by exhibiting the solution directly; here it falls out of the reduced equations, as the statement that the invariant set { x = 0 , p x = 0 } \{x = 0, p_x = 0\} { x = 0 , p x = 0 } of (140) projects onto the final constraint manifold of the algorithm of § 8.6 . The reduction is what makes the two computations comparable: both take place on M / G M/G M / G , with μ \mu μ a parameter.
The asymmetry is not an artifact: the reflection y ↦ − y y \mapsto-y y ↦ − y preserves the Martinet distribution only together with z ↦ − z z \mapsto-z z ↦ − z , hence together with μ ↦ − μ \mu\mapsto-\mu μ ↦ − μ , so for a fixed value of the reduced coadjoint variable the two directions along the abnormal line are genuinely different. The separatrix of the saddle is the family of normal geodesics asymptotic to the abnormal line, and it is this structure that is responsible for the delicate behavior of the Martinet sphere. Two of its consequences are recorded in [1 ] . First, the period 4 K ( k ) / μ r 4K(k)/\sqrt{\mu r} 4 K ( k ) / μ r of Corollary 2 diverges logarithmically as k → 1 − k \to1^{-} k → 1 − , that is, as the librating geodesics approach the separatrix cycle joining the saddle to its translate by 2 π 2\pi 2 π ; the resulting logarithmic term is what makes the Martinet sphere and the sub-Riemannian distance fail to be sub-analytic [1 Thm. 1.2] . Second, the cut locus acquires there a singularity like that of a Poincaré–Dulac return mapping computed for a section to a separatrix connecting two saddles [1 ] . Both are statements about the phase portrait of (142), which is to say about the reduced system; the parameter μ \mu μ that organizes the family is the reduced coadjoint variable.
Remark 19. The curvature criterion is stable under deformation of the model, and the deformation of [1 ] is instructive about what the criterion does and does not see. Agrachev, Bonnard, Chyba and Kupka replace the flat structure by
D ε = span { ( 1 + ε y ) ∂ ∂ x + y 2 2 ∂ ∂ z , ∂ ∂ y } . D_\varepsilon= \operatorname{span}\left\{(1+\varepsilon y)\frac{\partial}{\partial x}+\frac{y^2}{2}\frac{\partial}{\partial z},\frac{\partial}{\partial y}\right\}. D ε = span { ( 1 + ε y ) ∂ x ∂ + 2 y 2 ∂ z ∂ , ∂ y ∂ } .
the two fields being declared orthonormal. Distribution and cost are again invariant under translation in z z z , so Definition 7 applies, with
A ε = d z − y 2 2 ( 1 + ε y ) d x , F A ε = y ( 2 + ε y ) 2 ( 1 + ε y ) 2 d x ∧ d y . A_\varepsilon= dz-\frac{y^2}{2(1+\varepsilon y)}\,dx,\qquad F_{A_\varepsilon} = \frac{y(2+\varepsilon y)}{2(1+\varepsilon y)^2}\,dx\wedge dy. A ε = d z − 2 ( 1 + ε y ) y 2 d x , F A ε = 2 ( 1 + ε y ) 2 y ( 2 + ε y ) d x ∧ d y .
All of this is local in y y y , and the correct domain is
∣ ε y ∣ < 1 , (144) |\varepsilon y| < 1, \tag*{(144)} ∣ ε y ∣ < 1 , (144)
or any connected neighborhood of { y = 0 } \{y = 0\} { y = 0 } avoiding the hypersurface { 1 + ε y = 0 } \{1+\varepsilon y = 0\} { 1 + ε y = 0 } , on which A ε A_\varepsilon A ε , F A ε F_{A_\varepsilon} F A ε and the reduced metric (154) below are all singular and the first generator of D ε D_\varepsilon D ε degenerates to 1 2 y 2 ∂ z \frac{1}{2}y^2\partial_z 2 1 y 2 ∂ z , which is tangent to the G G G -orbits and therefore no longer horizontal. On (144) the numerator y ( 2 + ε y ) y(2+\varepsilon y) y ( 2 + ε y ) vanishes only at y = 0 y = 0 y = 0 , since the other root y = − 2 / ε y = -2/\varepsilon y = − 2/ ε is excluded, so the curvature vanishes exactly on { y = 0 } \{y = 0\} { y = 0 } and Theorem 6 confines the abnormal extremals to that hypersurface; they are the lines y = 0 y = 0 y = 0 , z = z 0 z = z_0 z = z 0 found in [1 ] . (The domain ∣ y ∣ < 2 / ∣ ε ∣ |y| < 2/|\varepsilon| ∣ y ∣ < 2/∣ ε ∣ , which one might be tempted to read off from the numerator alone, will not do: it contains y = − 1 / ε y=-1/\varepsilon y = − 1/ ε .) For ε = 0 \varepsilon=0 ε = 0 the condition (144) is empty and the domain is all of R 3 \mathbb{R}^{3} R 3 , as it should be. What ε \varepsilon ε changes is not the location of the abnormal leaf but its position relative to the normal one: for ε = 0 \varepsilon=0 ε = 0 the abnormal extremal is also normal (Proposition 15 ), while for ε ≠ 0 \varepsilon\ne0 ε = 0 it is strictly abnormal [1, Thm. 1.3. Strictness is a statement about the intersection of the two leaves, and is invisible to a criterion that sees only the abnormal one.
Reducing by the full symmetry group. The Martinet structure is invariant under translations in y y y as well as in z z z , so one may instead take G = R 2 G=\mathbb{R}^{2} G = R 2 , acting by ( y , z ) ↦ ( y + s 1 , z + s 2 ) (y,z)\mapsto(y+s_{1},z+s_{2}) ( y , z ) ↦ ( y + s 1 , z + s 2 ) . The hypotheses of Definition 7 hold as before, the base is now B = R B=\mathbb{R} B = R with coordinate x x x , and the natural connection is
A = ( d y , d z ) , A=(dy,dz), A = ( d y , d z ) ,
which is flat: F A = 0 F_{A}=0 F A = 0 . The whole of the dynamics moves into the vertical part,
Γ ~ ( x , u ) = u 1 ∂ ∂ x , γ ~ ( x , u ) = ( u 2 , 1 2 u 2 x 2 ) , \widetilde{\Gamma}(x,u)=u_{1}\frac{\partial}{\partial x},\qquad\widetilde{\gamma}(x,u)=\left(u_{2},\frac{1}{2}u_{2}x^{2}\right), Γ ( x , u ) = u 1 ∂ x ∂ , γ ( x , u ) = ( u 2 , 2 1 u 2 x 2 ) ,
so that h = p x u 1 + μ 1 u 2 + 1 2 μ 2 u 2 x 2 − 1 2 ∣ u ∣ 2 h=p_{x}u_{1}+\mu_{1}u_{2}+\frac{1}{2}\mu_{2}u_{2}x^{2}-\frac{1}{2}|u|^{2} h = p x u 1 + μ 1 u 2 + 2 1 μ 2 u 2 x 2 − 2 1 ∣ u ∣ 2 . Stationarity gives u 1 = p x u_{1}=p_{x} u 1 = p x and u 2 = μ 1 + 1 2 μ 2 x 2 u_{2}=\mu_{1}+\frac{1}{2}\mu_{2}x^{2} u 2 = μ 1 + 2 1 μ 2 x 2 , the coadjoint variables are constant because g = R 2 \mathfrak{g}=\mathbb{R}^{2} g = R 2 is abelian, and (26) reduces to the single second-order equation
x ¨ = − μ 2 x ( μ 1 + 1 2 μ 2 x 2 ) , (145) \ddot{x}=-\mu_{2}x\left(\mu_{1}+\frac{1}{2}\mu_{2}x^{2}\right), \tag*{(145)} x ¨ = − μ 2 x ( μ 1 + 2 1 μ 2 x 2 ) , (145)
with first integral
1 2 x ˙ 2 + 1 2 ( μ 1 + 1 2 μ 2 x 2 ) 2 = 1 2 r 2 . (146) \frac{1}{2}\dot{x}^{2}+\frac{1}{2}\left(\mu_{1}+\frac{1}{2}\mu_{2}x^{2}\right)^{2}=\frac{1}{2}r^{2}. \tag*{(146)} 2 1 x ˙ 2 + 2 1 ( μ 1 + 2 1 μ 2 x 2 ) 2 = 2 1 r 2 . (146)
Here μ 1 = λ y \mu_{1}=\lambda_{y} μ 1 = λ y and μ 2 = λ z \mu_{2}=\lambda_{z} μ 2 = λ z , so that (145) is exactly what (140) gives after eliminating p y p_{y} p y by Proposition 14; the two reductions produce the same base curve. Equation (145) is an oscillator in the quartic potential V ( x ) = 1 2 μ 1 μ 2 x 2 + 1 8 μ 2 2 x 4 V(x)=\frac{1}{2}\mu_{1}\mu_{2}x^{2}+\frac{1}{8}\mu_{2}^{2}x^{4} V ( x ) = 2 1 μ 1 μ 2 x 2 + 8 1 μ 2 2 x 4 , again integrable by elliptic functions.
Remark 20. The two reductions of the same problem are instructive about the structure of (26). Reducing by R z \mathbb{R}_{z} R z leaves a curved connection and a horizontal controlled field, so that γ ~ = 0 \widetilde{\gamma}=0 γ = 0 and the entire nonlinearity is carried by the curvature term. Reducing by R 2 \mathbb{R}^{2} R 2 leaves a flat connection, so that the curvature term disappears and the entire nonlinearity is carried by γ ~ \widetilde{\gamma} γ . The reduced dynamics is the same. The division of the reduced equations into a curvature term and a vertical term is therefore a feature of the chosen splitting rather than of the problem, in the same way that Theorem 2 shows the division to be unaffected by a change of connection for a fixed group. A systematic treatment of the relation between the two reductions belongs to the theory of reduction by stages, which we do not develop here.
Strict and non-strict abnormality
Theorem 6 locates the abnormal extremals using the curvature alone, and Proposition 15 showed that in the Martinet problem the curve so located also carries a normal extremal. That cannot be read off from the curvature: the abnormal leaf does not depend on the cost at all, whereas the normal leaf does. This subsection makes the difference precise, and reduces the question to a linear problem in the fiber g ∗ \mathfrak{g}^{*} g ∗ .
Definition 14. An admissible curve is abnormal if it is the projection of an extremal with p 0 = 0 p_{0}=0 p 0 = 0 and nonvanishing multiplier, normal if it is the projection of an extremal with p 0 = − 1 p_{0}=-1 p 0 = − 1 , and strictly abnormal if it is abnormal and not normal.
The terminology is that of [1 , 30 ] . It is worth a warning that the adjective strict is used by other authors for a different property, namely that the abnormal multiplier be unique up to scale [11, Def. 3.2; the two notions are independent.
The two leaves share an adjoint equation. Fix an admissible reduced curve, that is, a curve t ↦ ( z ( t ) , ξ ( t ) ) t\mapsto(z(t),\xi(t)) t ↦ ( z ( t ) , ξ ( t )) together with a control u ( t ) u(t) u ( t ) satisfying the first and fourth equations of (26). A lift of it is a pair ( p z , μ ) (p_{z},\mu) ( p z , μ ) along the curve satisfying the remaining equations of (26) for the given u u u .
Lemma 5. Suppose the controlled field is linear in the control, [ Γ ] ( z , u ) = ∑ a u a ( Γ ~ a ( z ) , γ ~ a ( z ) ) [\Gamma](z,u)=\sum_{a}u^{a}(\widetilde{\Gamma}_{a}(z),\widetilde{\gamma}_{a}(z)) [ Γ ] ( z , u ) = ∑ a u a ( Γ a ( z ) , γ a ( z )) , and the cost is L = 1 2 ∣ u ∣ 2 L=\frac{1}{2}|u|^{2} L = 2 1 ∣ u ∣ 2 with ∣ ⋅ ∣ |\cdot| ∣ ⋅ ∣ a fixed G G G -invariant norm on U U U , so that L L L does not depend on the base point. Then along a fixed admissible reduced curve the equations for ( p z , μ ) (p_{z},\mu) ( p z , μ ) in (26) form one and the same linear homogeneous system on both leaves; the leaves differ only in the algebraic condition imposed by ∂ h / ∂ u = 0 \partial h/\partial u=0 ∂ h / ∂ u = 0 , namely
h a = 0 ( p 0 = 0 ) , versus h a = u a ( p 0 = − 1 ) , h_{a}=0\quad(p_{0}=0),\qquad\text{versus}\qquad h_{a}=u^{a}\quad(p_{0}=-1), h a = 0 ( p 0 = 0 ) , versus h a = u a ( p 0 = − 1 ) , where h a = ⟨ p z , Γ ~ a ⟩ + ⟨ μ , γ ~ a ⟩ h_{a}=\langle p_{z},\widetilde{\Gamma}_{a}\rangle+\langle\mu,\widetilde{\gamma}_{a}\rangle h a = ⟨ p z , Γ a ⟩ + ⟨ μ , γ a ⟩ .
Proof. h = p 0 L ( u ) + ∑ a u a h a h=p_0L(u)+\sum_a u^a h_a h = p 0 L ( u ) + ∑ a u a h a with h a h_a h a linear in ( p z , μ ) (p_z,\mu) ( p z , μ ) and L L L independent of z z z . Hence ∂ h / ∂ z = ∑ a u a ∂ h a / ∂ z \partial h/\partial z=\sum_a u^a\partial h_a/\partial z ∂ h / ∂ z = ∑ a u a ∂ h a / ∂ z carries no p 0 p_0 p 0 , and neither do the covariant derivative terms, which come from the splitting alone. The condition ∂ h / ∂ u a = 0 \partial h/\partial u^a=0 ∂ h / ∂ u a = 0 reads p 0 u a + h a = 0 p_0u^a+h_a=0 p 0 u a + h a = 0 . □
Proposition 16. For a fixed admissible reduced curve, the set of normal lifts is either empty or an affine space over the vector space of abnormal lifts. In particular an abnormal curve is non-strictly abnormal precisely when the inhomogeneous system h a = u a h_a=u^a h a = u a is consistent within the solution space of the common adjoint equation.
Proof. Immediate from Lemma 5: the solutions of the linear homogeneous system form a vector space V \mathcal V V , the map ev : V → C ∞ ( [ 0 , T ] , R m ) \operatorname{ev}:\mathcal V\to C^\infty([0,T],\mathbb{R}^m) ev : V → C ∞ ([ 0 , T ] , R m ) , λ ↦ ( h a ( λ ( ⋅ ) ) ) a \lambda\mapsto(h_a(\lambda(\cdot)))_a λ ↦ ( h a ( λ ( ⋅ )) ) a , is linear, the abnormal lifts are ker ev \ker\operatorname{ev} ker ev and the normal lifts are ev − 1 ( u ) \operatorname{ev}^{-1}(u) ev − 1 ( u ) . □
The unreduced counterpart of Proposition 16 is [30 Rem. 2] . What the reduction contributes is that V \mathcal V V and ev \operatorname{ev} ev are computable from reduced data, which we now do.
The criterion. Take the horizontal system (134) with U ≅ R dim B U\cong\mathbb{R}^{\dim B} U ≅ R d i m B and L = 1 2 ∣ u ∣ 2 L=\frac{1}{2}|u|^2 L = 2 1 ∣ u ∣ 2 , so that the X a X_a X a are the horizontal lifts of a frame X a X_a X a of T B TB T B that we now take orthonormal for a Riemannian metric g g g on B B B ; equivalently, D = ker A \mathcal D=\ker A D = ker A carries the pullback of g g g . Let ∇ \nabla ∇ be the Levi-Civita connection of g g g , let ∇ A \nabla^A ∇ A be the connection induced by A A A on g ∗ \mathfrak{g}^* g ∗ , and write ♯ \sharp ♯ and ♭ \flat ♭ for the musical isomorphisms of g g g . Along a curve z ( ⋅ ) z(\cdot) z ( ⋅ ) in B B B define the curvature pairing
C z ˙ : g z ∗ ⟶ T z B , ⟨ C z ˙ μ , v ⟩ = ⟨ μ , F A ( z ˙ , v ) ⟩ , (147) C_{\dot z}:\mathfrak{g}^*_z\longrightarrow T_zB,\qquad\langle C_{\dot z}\mu,v\rangle=\langle\mu,F_A(\dot z,v)\rangle, \tag*{(147)} C z ˙ : g z ∗ ⟶ T z B , ⟨ C z ˙ μ , v ⟩ = ⟨ μ , F A ( z ˙ , v )⟩ , (147)
and let
K z ˙ = { v ∈ T z B : F A ( z ˙ , v ) = 0 } ⊇ ⟨ z ˙ ⟩ , im C z ˙ = K z ˙ ∘ . K_{\dot z}=\{v\in T_zB:F_A(\dot z,v)=0\}\supseteq\langle\dot z\rangle,\qquad\operatorname{im}C_{\dot z}=K_{\dot z}^{\circ}. K z ˙ = { v ∈ T z B : F A ( z ˙ , v ) = 0 } ⊇ ⟨ z ˙ ⟩ , im C z ˙ = K z ˙ ∘ .
Lemma 6. For the system (134) with the above cost, ξ ≡ 0 \xi\equiv0 ξ ≡ 0 and (26) becomes
p 0 = − 1 : ∇ z ˙ z ˙ = − ( C z ˙ μ ) ♯ , ∇ z ˙ A μ = 0 , p z = z ˙ ♭ ; (148) p_0=-1:\qquad\nabla_{\dot z}\dot z=-(C_{\dot z}\mu)^\sharp,\qquad\nabla^A_{\dot z}\mu=0,\qquad p_z=\dot z^\flat; \tag*{(148)} p 0 = − 1 : ∇ z ˙ z ˙ = − ( C z ˙ μ ) ♯ , ∇ z ˙ A μ = 0 , p z = z ˙ ♭ ; (148)
p 0 = 0 : p z ≡ 0 , C z ˙ μ = 0 , ∇ z ˙ A μ = 0 , μ ≠ 0. (149) p_0=0:\qquad p_z\equiv0,\qquad C_{\dot z}\mu=0,\qquad\nabla^A_{\dot z}\mu=0,\qquad\mu\ne0. \tag*{(149)} p 0 = 0 : p z ≡ 0 , C z ˙ μ = 0 , ∇ z ˙ A μ = 0 , μ = 0. (149)
Proof. Since the controlled field is horizontal, γ ~ = 0 \widetilde\gamma=0 γ = 0 and ξ = ∂ h / ∂ μ = 0 \xi=\partial h/\partial\mu=0 ξ = ∂ h / ∂ μ = 0 , so the third equation of (26) gives ∇ z ˙ A μ = 0 \nabla^A_{\dot z}\mu=0 ∇ z ˙ A μ = 0 in both cases. On p 0 = − 1 p_0=-1 p 0 = − 1 stationarity gives u a = h a u^a=h_a u a = h a , so p z = z ˙ ♭ p_z=\dot z^\flat p z = z ˙ ♭ , and the second equation of (26) , whose curvature term is ⟨ μ , F A ( ⋅ , z ˙ ) ⟩ = − C z ˙ μ \langle\mu,F_A(\cdot,\dot z)\rangle=-C_{\dot z}\mu ⟨ μ , F A ( ⋅ , z ˙ )⟩ = − C z ˙ μ , becomes (148) . The last step deserves its computation, since it is what produces Wong’s equation. The auxiliary connection of Section 3.2 is here the Levi-Civita connection of g g g , so D p z / D t = ∇ z ˙ p z D p_z/Dt=\nabla_{\dot z}p_z D p z / D t = ∇ z ˙ p z by definition; and on the constraint manifold h = 1 2 ∣ p z ∣ g ∗ 2 h=\frac12|p_z|_{g^*}^2 h = 2 1 ∣ p z ∣ g ∗ 2 , whose covariant partial derivative in z z z vanishes: parallel-translating p z p_z p z ,
( ∂ h ∂ z ) k = 1 2 ( ∂ k g i j ) p i p j + g i j Γ k i l p l p j = ( − g l j Γ k l i + g l j Γ k l i ) p i p j = 0 , \left(\frac{\partial h}{\partial z}\right)_k=\frac12\left(\partial_k g^{ij}\right)p_i p_j+g^{ij}\Gamma^l_{ki}p_l p_j=\left(-g^{lj}\Gamma^i_{kl}+g^{lj}\Gamma^i_{kl}\right)p_i p_j=0, ( ∂ z ∂ h ) k = 2 1 ( ∂ k g ij ) p i p j + g ij Γ k i l p l p j = ( − g l j Γ k l i + g l j Γ k l i ) p i p j = 0 ,
the first term having been rewritten with ∂ k g i j = − Γ k l i g l j − Γ k l j g i l \partial_k g^{ij}=-\Gamma^i_{kl}g^{lj}-\Gamma^j_{kl}g^{il} ∂ k g ij = − Γ k l i g l j − Γ k l j g i l , which is ∇ g − 1 = 0 \nabla g^{-1}=0 ∇ g − 1 = 0 , and the second by relabeling. The second equation of (26) therefore reads ∇ z ˙ p z = − C z ˙ μ \nabla_{\dot z}p_z=-C_{\dot z}\mu ∇ z ˙ p z = − C z ˙ μ , which with p z = z ˙ ♭ p_z=\dot z^\flat p z = z ˙ ♭ and ∇ g = 0 \nabla g=0 ∇ g = 0 is (148) . On p 0 = 0 p_0=0 p 0 = 0 , Proposition 13 gives p z = 0 p_z=0 p z = 0 , and the same equation reduces to C z ˙ μ = 0 C_{\dot z}\mu=0 C z ˙ μ = 0 , which is Theorem 6 evaluated at the initial point. □
Equation (148) is Wong’s equation for a classical particle of charge μ \mu μ in the Yang–Mills field F A F_A F A ; that reduced normal sub-Riemannian geodesics on a principal bundle are magnetic geodesics on the base is classical [33 ] . The point of writing the two leaves side by side is the following statement, which is immediate from them.
Theorem 8. Let z ( ⋅ ) z(\cdot) z ( ⋅ ) be the base projection of an abnormal extremal of (134) , parametrized by arclength. Then z ( ⋅ ) z(\cdot) z ( ⋅ ) is non-strictly abnormal if and only if there exists a ∇ A \nabla^A ∇ A -parallel section μ n \mu^n μ n of g ∗ \mathfrak{g}^* g ∗ along z ( ⋅ ) z(\cdot) z ( ⋅ ) with
∇ z ˙ z ˙ = − ( C z ˙ μ n ) ♯ . (150) \nabla_{\dot z}\dot z=-(C_{\dot z}\mu^n)^\sharp. \tag*{(150)} ∇ z ˙ z ˙ = − ( C z ˙ μ n ) ♯ . (150) The set of such μ n \mu^n μ n is then an affine space over the parallel sections of ker C z ˙ \ker C_{\dot z} ker C z ˙ , which contains the abnormal multiplier μ a \mu^a μ a .
Proof. By Lemma 6 a normal lift of the same curve exists if and only if (148) holds for some parallel μ n \mu^n μ n , and its p z p_z p z component is then forced to be z ˙ ♭ \dot z^\flat z ˙ ♭ , which is the same curve. The affine structure is Proposition 16, the abnormal lifts being the parallel sections of ker C z ˙ \ker C_{\dot z} ker C z ˙ by (149) . □
Two consequences are worth isolating. The first is pointwise and costs nothing; the second is the case in which it decides everything.
Corollary 3. A necessary condition for non-strictness is
g ( ∇ z ˙ z ˙ , v ) = 0 for every v ∈ K z ˙ . (151) g(\nabla_{\dot z}\dot z,v)=0 \qquad\text{for every }v\in K_{\dot z}. \tag*{(151)} g ( ∇ z ˙ z ˙ , v ) = 0 for every v ∈ K z ˙ . (151) Since z ˙ ∈ K z ˙ \dot z\in K_{\dot z} z ˙ ∈ K z ˙ and g ( ∇ z ˙ z ˙ , z ˙ ) = 0 g(\nabla_{\dot z}\dot z,\dot z)=0 g ( ∇ z ˙ z ˙ , z ˙ ) = 0 for an arclength parametrization, the obstruction (151) takes values in ( K z ˙ / ⟨ z ˙ ⟩ ) ∗ (K_{\dot z}/\langle\dot z\rangle)^* ( K z ˙ / ⟨ z ˙ ⟩ ) ∗ .
Proof. im C z ˙ = K z ˙ ∘ \operatorname{im}C_{\dot z}=K_{\dot z}^{\circ} im C z ˙ = K z ˙ ∘ , so (150) is solvable pointwise for μ n \mu^{\mathrm n} μ n if and only if ( ∇ z ˙ z ˙ ) ♭ (\nabla_{\dot z}\dot z)^\flat ( ∇ z ˙ z ˙ ) ♭ annihilates K z ˙ K_{\dot z} K z ˙ .
Corollary 4. Suppose dim g = 1 \dim\mathfrak g=1 dim g = 1 . Then C z ˙ ≡ 0 C_{\dot z}\equiv0 C z ˙ ≡ 0 along an abnormal extremal, and z ( ⋅ ) z(\cdot) z ( ⋅ ) is non-strictly abnormal if and only if it is a geodesic of ( B , g ) (B,g) ( B , g ) . In that case every parallel μ n \mu^{\mathrm n} μ n is admissible, so the normal lifts form a full line.
Proof. By (149) there is μ a b ≠ 0 \mu^{\mathrm{ab}}\ne0 μ ab = 0 with C z ˙ μ a b = 0 C_{\dot z}\mu^{\mathrm{ab}}=0 C z ˙ μ ab = 0 ; as g ∗ \mathfrak g^* g ∗ is a line bundle this forces C z ˙ = 0 C_{\dot z}=0 C z ˙ = 0 , whence K z ˙ = T z B K_{\dot z}=T_zB K z ˙ = T z B and (150) reads ∇ z ˙ z ˙ = 0 \nabla_{\dot z}\dot z=0 ∇ z ˙ z ˙ = 0 , with no condition on μ n \mu^{\mathrm n} μ n ; parallel sections of a line bundle along a curve exist and are unique up to scale.
Corollary 4 is not new in the abelian rank-one case: for B = R 2 B=\mathbb{R}^2 B = R 2 with the flat metric and G = R G=\mathbb{R} G = R it is essentially Theorem 3.1 of [31 ] , which states that the horizontal lift of the zero locus of a magnetic field fails to be a normal geodesic as soon as its planar curvature is nonzero, and for Martinet-type distributions in R 3 \mathbb{R}^3 R 3 it is Proposition 2.4 of [35 ] , where the same conclusion is reached in a normal form. What Theorem 8 adds is the general principal-bundle statement, with g \mathfrak g g of arbitrary dimension and possibly nonabelian and with μ \mu μ a ∇ A \nabla^A ∇ A -parallel section rather than a constant; in that generality Corollary 4 fails, as the rank-two example below shows. A criterion of a comparable kind for Engel structures, phrased through a structure function of the canonical frame, is Theorem 4 of [7 ] .
Remark 21. The reason Theorem 6 cannot decide strictness is now visible. Abnormal extremals are determined by D = ker A D=\ker A D = ker A , hence by A A A and F A F_A F A , and are independent of the cost; strictness compares them with the normal extremals, which depend on the cost through ∇ \nabla ∇ . Every object in (151) and (150) except ∇ z ˙ z ˙ \nabla_{\dot z}\dot z ∇ z ˙ z ˙ is metric-free, and ∇ z ˙ z ˙ \nabla_{\dot z}\dot z ∇ z ˙ z ˙ carries the entire dependence on the cost. Generic perturbation of the metric therefore makes abnormal extremals strictly abnormal without moving them [16 ] .
The parallelism obstruction. Condition (151) is only necessary: a pointwise solution μ n ( t ) \mu^{\mathrm n}(t) μ n ( t ) of (150) need not be parallel. The residual obstruction is again resolved by a constraint algorithm, this time run in the fiber g ∗ \mathfrak g^* g ∗ rather than on M M M .
Trivialize g ∗ \mathfrak g^* g ∗ along z ( ⋅ ) z(\cdot) z ( ⋅ ) by ∇ A \nabla^A ∇ A -parallel transport and T ∗ B T^*B T ∗ B by ∇ \nabla ∇ -parallel transport, and write Φ ( t ) \Phi(t) Φ ( t ) for C z ˙ ( t ) C_{\dot z(t)} C z ˙ ( t ) and b ( t ) b(t) b ( t ) for − ( ∇ z ˙ z ˙ ) ♭ -(\nabla_{\dot z}\dot z)^\flat − ( ∇ z ˙ z ˙ ) ♭ in these trivializations. Then (150) together with parallelism reads
Φ ( t ) μ = b ( t ) for all t , μ constant . (152) \Phi(t)\mu=b(t)\qquad\text{for all }t,\qquad\mu\text{ constant}. \tag*{(152)} Φ ( t ) μ = b ( t ) for all t , μ constant . (152)
Differentiating repeatedly, set
W k ( t ) = { μ : Φ ( j ) ( t ) μ = b ( j ) ( t ) , j = 0 , … , k } , \mathcal W_k(t)=\{\mu:\Phi^{(j)}(t)\mu=b^{(j)}(t),\ j=0,\ldots,k\}, W k ( t ) = { μ : Φ ( j ) ( t ) μ = b ( j ) ( t ) , j = 0 , … , k } ,
a decreasing sequence of affine subspaces of g ∗ \mathfrak g^* g ∗ .
Proposition 17. For each fixed t t t , the decreasing sequence W k ( t ) \mathcal W_k(t) W k ( t ) of affine subspaces is eventually constant. If the data are real-analytic, then z ( ⋅ ) z(\cdot) z ( ⋅ ) is non-strictly abnormal if and only if ⋂ k ≥ 0 W k ( t 0 ) ≠ ∅ \bigcap_{k\geq0}\mathcal W_k(t_0)\ne\varnothing ⋂ k ≥ 0 W k ( t 0 ) = ∅ for one, equivalently every, t 0 t_0 t 0 ; the set of normal multipliers is then that intersection, which is independent of t 0 t_0 t 0 . Moreover, there is a discrete exceptional set Σ ⊂ [ 0 , T ] \Sigma\subset[0,T] Σ ⊂ [ 0 , T ] such that for t 0 ∉ Σ t_0\notin\Sigma t 0 ∈ / Σ the intersection is already attained at order N = dim g N=\dim\mathfrak g N = dim g ,
⋂ k ≥ 0 W k ( t 0 ) = W N ( t 0 ) , \bigcap_{k\geq0}\mathcal W_k(t_0)=\mathcal W_N(t_0), k ≥ 0 ⋂ W k ( t 0 ) = W N ( t 0 ) , so that, writing Φ [ N ] = ( Φ ; Φ ′ ; … ; Φ ( N ) ) \Phi_{[N]}=(\Phi;\Phi';\ldots;\Phi^{(N)}) Φ [ N ] = ( Φ ; Φ ′ ; … ; Φ ( N ) ) and b [ N ] = ( b ; b ′ ; … ; b ( N ) ) b_{[N]}=(b;b';\ldots;b^{(N)}) b [ N ] = ( b ; b ′ ; … ; b ( N ) ) , non-strictness holds if and only if
rank Φ [ N ] ( t 0 ) = rank ( Φ [ N ] ( t 0 ) ∣ b [ N ] ( t 0 ) ) , t 0 ∉ Σ . (153) \operatorname{rank}\Phi_{[N]}(t_0)=\operatorname{rank}\bigl(\Phi_{[N]}(t_0)\mid b_{[N]}(t_0)\bigr),\qquad t_0\notin\Sigma. \tag*{(153)} rank Φ [ N ] ( t 0 ) = rank ( Φ [ N ] ( t 0 ) ∣ b [ N ] ( t 0 ) ) , t 0 ∈ / Σ. (153) Proof. For fixed t t t , each strict inclusion W k + 1 ( t ) ⊊ W k ( t ) \mathcal W_{k+1}(t)\subsetneq\mathcal W_k(t) W k + 1 ( t ) ⊊ W k ( t ) lowers the dimension by at least one or produces the empty set, so at most dim g + 1 \dim\mathfrak g+1 dim g + 1 of the inclusions are strict and the sequence is eventually constant. The index at which it stabilizes is, however, not bounded by dim g \dim\mathfrak g dim g at every point: with dim g = 1 \dim\mathfrak g=1 dim g = 1 , Φ ≡ 0 \Phi\equiv0 Φ ≡ 0 and b ( t ) = ( t − t 0 ) 2 b(t)=(t-t_0)^2 b ( t ) = ( t − t 0 ) 2 one has W 0 ( t 0 ) = W 1 ( t 0 ) = g ∗ \mathcal W_0(t_0)=\mathcal W_1(t_0)=\mathfrak g^* W 0 ( t 0 ) = W 1 ( t 0 ) = g ∗ while W 2 ( t 0 ) = ∅ \mathcal W_2(t_0)=\varnothing W 2 ( t 0 ) = ∅ . This is the reason for the exceptional set Σ \Sigma Σ .
If μ ∈ W k ( t 0 ) \mu\in\mathcal W_k(t_0) μ ∈ W k ( t 0 ) for every k k k then all derivatives of t ↦ Φ ( t ) μ − b ( t ) t\mapsto\Phi(t)\mu-b(t) t ↦ Φ ( t ) μ − b ( t ) vanish at t 0 t_0 t 0 , so by analyticity the function vanishes identically and μ \mu μ is a normal multiplier; the converse is differentiation. Hence ⋂ k W k ( t 0 ) = { μ : Φ μ ≡ b } \bigcap_k\mathcal W_k(t_0)=\{\mu:\Phi\mu\equiv b\} ⋂ k W k ( t 0 ) = { μ : Φ μ ≡ b } , which proves the second assertion and the independence of t 0 t_0 t 0 .
For the last part, let N k \mathcal{N}_k N k denote the span, over the field of functions meromorphic on [ 0 , T ] [0,T] [ 0 , T ] , of the rows of the augmented matrices ( Φ ( j ) ∣ b ( j ) ) (\Phi^{(j)} \mid b^{(j)}) ( Φ ( j ) ∣ b ( j ) ) with j ≤ k j \le k j ≤ k . If N k + 1 = N k \mathcal{N}_{k+1}=\mathcal{N}_k N k + 1 = N k , then every row of order k + 1 k+1 k + 1 is a meromorphic combination of rows of order at most k k k ; differentiating the combination expresses the rows of order k + 2 k+2 k + 2 through rows of order at most k + 1 k+1 k + 1 , so the flag is constant from k k k on. Since dim N k ≤ dim g + 1 \dim\mathcal{N}_k \le\dim\mathfrak{g}+1 dim N k ≤ dim g + 1 and each strict step increases the dimension by at least one, N k = N N \mathcal{N}_k=\mathcal{N}_N N k = N N for every k ≥ N = dim g k\ge N=\dim\mathfrak{g} k ≥ N = dim g ; iterating the relations, every row of ( Φ ( k ) ∣ b ( k ) ) (\Phi^{(k)} \mid b^{(k)}) ( Φ ( k ) ∣ b ( k ) ) with k > N k>N k > N is a meromorphic combination of the rows of ( Φ [ N ] ∣ b [ N ] ) (\Phi^{[N]} \mid b^{[N]}) ( Φ [ N ] ∣ b [ N ] ) , with coefficients whose poles form a discrete set. Let Σ \Sigma Σ consist of those poles together with the (discrete) zero sets of the finitely many minors realizing the generic ranks of Φ [ N ] \Phi_{[N]} Φ [ N ] and of ( Φ [ N ] ∣ b [ N ] ) (\Phi_{[N]} \mid b_{[N]}) ( Φ [ N ] ∣ b [ N ] ) . For t 0 ∉ Σ t_0\notin\Sigma t 0 ∈ / Σ the equations defining W k ( t 0 ) \mathcal{W}_k(t_0) W k ( t 0 ) with k > N k>N k > N are consequences, by evaluation at t 0 t_0 t 0 , of those defining W N ( t 0 ) \mathcal{W}_N(t_0) W N ( t 0 ) , so W k ( t 0 ) = W N ( t 0 ) \mathcal{W}_k(t_0)=\mathcal{W}_N(t_0) W k ( t 0 ) = W N ( t 0 ) for every k ≥ N k\ge N k ≥ N ; and (153) is the consistency criterion for the linear system Φ [ N ] μ = b [ N ] \Phi_{[N]}\mu=b_{[N]} Φ [ N ] μ = b [ N ] .
The primary constraint of this algorithm, that W 0 ( t ) \mathcal{W}_0(t) W 0 ( t ) be nonempty, is exactly (151); the later stages are the parallelism. Note that Φ [ N ] ( t 0 ) μ a b = 0 \Phi_{[N]}(t_0)\mu^{ab}=0 Φ [ N ] ( t 0 ) μ ab = 0 , so Φ [ N ] \Phi_{[N]} Φ [ N ] is never injective on an abnormal, in agreement with Theorem 8. When dim g = 1 \dim\mathfrak{g}=1 dim g = 1 one has Φ ≡ 0 \Phi\equiv0 Φ ≡ 0 along an abnormal extremal, so the test reduces to the identical vanishing of b b b , that is, to the geodesic condition of Corollary 4; generically in t 0 t_0 t 0 the algorithm halts at k = 0 k=0 k = 0 .
The three rank-one examples. For the Heisenberg problem of Section 6 there is nothing to test: F A F_A F A is nowhere zero, so by Corollary 1 there are no abnormal extremals.
For the flat Martinet problem of Section 7.6, B = R 2 B=\mathbb{R}^2 B = R 2 carries the Euclidean metric and the abnormal extremal projects to the line { x = 0 } \{x=0\} { x = 0 } , which is a geodesic. Corollary 4 therefore says it is non-strictly abnormal, with a line of normal lifts parametrized by μ \mu μ . This is Proposition 15, obtained there from the explicit pendulum and here from the metric alone.
The deformation of Remark 19 is the interesting case, and the criterion says exactly what the deformation does. The reduced metric is read off from the declared orthonormal frame: the projections of ( 1 + ε y ) ∂ x + 1 2 y 2 ∂ z (1+\varepsilon y)\partial_x+\frac{1}{2}y^2\partial_z ( 1 + ε y ) ∂ x + 2 1 y 2 ∂ z and ∂ y \partial_y ∂ y are ( 1 + ε y ) ∂ x (1+\varepsilon y)\partial_x ( 1 + ε y ) ∂ x and ∂ y \partial_y ∂ y , so
g ε = d x 2 ( 1 + ε y ) 2 + d y 2 , Γ x x y = ε ( 1 + ε y ) 3 , (154) g_\varepsilon=\frac{dx^2}{(1+\varepsilon y)^2}+dy^2,\qquad\Gamma^y_{xx}=\frac{\varepsilon}{(1+\varepsilon y)^3}, \tag*{(154)} g ε = ( 1 + ε y ) 2 d x 2 + d y 2 , Γ xx y = ( 1 + ε y ) 3 ε , (154)
both on the domain (144), off which they are singular. The abnormal locus is { y = 0 } \{y=0\} { y = 0 } for every ε \varepsilon ε , and the abnormal curve is the x x x -axis, which is arclength-parametrized by x ˙ = 1 \dot{x}=1 x ˙ = 1 there. Hence
( ∇ z ˙ z ˙ ) y ∣ y = 0 = Γ x x y ∣ y = 0 = ε , (155) \left.(\nabla_{\dot z}\dot z)^y\right|_{y=0}=\left.\Gamma^y_{xx}\right|_{y=0}=\varepsilon, \tag*{(155)} ( ∇ z ˙ z ˙ ) y ∣ y = 0 = Γ xx y ∣ y = 0 = ε , (155)
and Corollary 4 gives:
Proposition 18. The abnormal extremal of the deformed Martinet problem of Remark 19 is non-strictly abnormal if and only if ε = 0 \varepsilon=0 ε = 0 .
This is the content of [1 Thm. 1.3] in its first assertion, recovered from (154) by one Christoffel symbol. The reading it supplies is that the deformation does not act on the abnormal extremals at all: on the domain (144) the curvature still vanishes exactly on { y = 0 } \{y=0\} { y = 0 } and the abnormal curves are still the lines y = 0 y=0 y = 0 , z = z 0 z=z_0 z = z 0 . What the deformation destroys is the property of those lines of being geodesics of the reduced metric, and with it the normal lift.
A higher-rank phenomenon. When dim g ≥ 2 \dim\mathfrak{g}\ge2 dim g ≥ 2 the first-order test (151) may be vacuous while strictness still fails, so that Corollary 4 has no analogue and the algorithm of Proposition 17 is needed. The following example is the smallest we know.
Let B = R 2 B=\mathbb{R}^2 B = R 2 with the Euclidean metric, G = R 2 G=\mathbb{R}^2 G = R 2 , P = B × G P=B\times G P = B × G with coordinates ( x , y , z 1 , z 2 ) (x,y,z_1,z_2) ( x , y , z 1 , z 2 ) and
A = ( d z 1 + a 1 ( x , y ) d y , d z 2 + ( 2 x + 1 2 x 2 ) d y ) , F A = ( ( x 2 + y 2 − 1 ) 2 , 2 + x ) d x ∧ d y , (156) A=\left(dz_1+a_1(x,y)\,dy,\ dz_2+\left(2x+\frac{1}{2}x^2\right)dy\right),\qquad
F_A=\left((x^2+y^2-1)^2,\ 2+x\right)dx\wedge dy, \tag*{(156)} A = ( d z 1 + a 1 ( x , y ) d y , d z 2 + ( 2 x + 2 1 x 2 ) d y ) , F A = ( ( x 2 + y 2 − 1 ) 2 , 2 + x ) d x ∧ d y , (156)
where
a 1 ( x , y ) = 1 5 x 5 + 2 3 x 3 ( y 2 − 1 ) + x ( y 2 − 1 ) 2 , ∂ a 1 ∂ x = ( x 2 + y 2 − 1 ) 2 , a_1(x,y)=\frac{1}{5}x^5+\frac{2}{3}x^3(y^2-1)+x(y^2-1)^2,\qquad
\frac{\partial a_1}{\partial x}=(x^2+y^2-1)^2, a 1 ( x , y ) = 5 1 x 5 + 3 2 x 3 ( y 2 − 1 ) + x ( y 2 − 1 ) 2 , ∂ x ∂ a 1 = ( x 2 + y 2 − 1 ) 2 ,
so that F A = d A F_A=dA F A = d A , the group being abelian. Consider the unit circle z ( s ) = ( cos s , sin s ) z(s)=(\cos s,\sin s) z ( s ) = ( cos s , sin s ) , of geodesic curvature 1 1 1 . Since the first component of F A F_A F A vanishes on it, μ a b = ( 1 , 0 ) \mu^{ab}=(1,0) μ ab = ( 1 , 0 ) is a constant multiplier with C z μ a b = 0 C_z\mu^{ab}=0 C z μ ab = 0 , so the circle is abnormal.
The second component 2 + cos s 2+\cos s 2 + cos s is nowhere zero on the circle, so C z ≠ 0 C_z\ne0 C z = 0 at every point, hence K z = ⟨ z ˙ ⟩ K_z=\langle\dot z\rangle K z = ⟨ z ˙ ⟩ everywhere and (151) holds automatically: the first-order test sees nothing. The full criterion does. Since ( d x ∧ d y ) ( z ˙ , v ) = − ⟨ z ˙ , v ⟩ (dx\wedge dy)(\dot z,v)=-\langle\dot z,v\rangle ( d x ∧ d y ) ( z ˙ , v ) = − ⟨ z ˙ , v ⟩ and ∇ z ˙ z ˙ = − z \nabla_{\dot z}\dot z=-z ∇ z ˙ z ˙ = − z , in the ∇ \nabla ∇ -parallel trivialization of (152) one has
Φ ( s ) μ = − μ 2 ( 2 + cos s ) z ′ ( s ) , b ( s ) = z ′ ( s ) , \Phi(s)\mu=-\mu_2(2+\cos s)z'(s),\qquad b(s)=z'(s), Φ ( s ) μ = − μ 2 ( 2 + cos s ) z ′ ( s ) , b ( s ) = z ′ ( s ) ,
so that the residual is
Φ ( s ) μ − b ( s ) = − r ( s ) z ♭ ( s ) , r ( s ) = 1 + μ 2 2 ( 2 + cos s ) . (157) \Phi(s)\mu-b(s)=-r(s)z^{\flat}(s),\qquad r(s)=1+\mu_{2}^{2}(2+\cos s). \tag*{(157)} Φ ( s ) μ − b ( s ) = − r ( s ) z ♭ ( s ) , r ( s ) = 1 + μ 2 2 ( 2 + cos s ) . (157)
The covector z ♭ ( s ) z^{\flat}(s) z ♭ ( s ) is nowhere zero, so (152) reads
μ 2 2 ( 2 + cos s ) = − 1 for all s , \mu_{2}^{2}(2+\cos s)=-1\qquad\text{for all }s, μ 2 2 ( 2 + cos s ) = − 1 for all s ,
which has no constant solution: the circle is strictly abnormal.
Because z ♭ z^{\flat} z ♭ never vanishes, (157) also reduces the whole algorithm of Proposition 17 to the scalar function r r r : if r r r and its first k − 1 k-1 k − 1 derivatives vanish at s s s , then the k k k -th derivative of the residual is − r ( k ) ( s ) z ♭ ( s ) -r^{(k)}(s)z^{\flat}(s) − r ( k ) ( s ) z ♭ ( s ) , so W k ( s ) = { μ : r ( j ) ( s ) = 0 , j ≤ k } W_k(s)=\{\mu:r^{(j)}(s)=0,\ j\leq k\} W k ( s ) = { μ : r ( j ) ( s ) = 0 , j ≤ k } . Now W 0 ( s ) = { μ 2 2 = − 1 / ( 2 + cos s ) } W_0(s)=\{\mu_2^2=-1/(2+\cos s)\} W 0 ( s ) = { μ 2 2 = − 1/ ( 2 + cos s )} is nonempty for every s s s , and on it μ 2 2 ≠ 0 \mu_2^2\ne0 μ 2 2 = 0 , so r ′ ( s ) = − μ 2 2 sin s r'(s)=-\mu_2^2\sin s r ′ ( s ) = − μ 2 2 sin s vanishes only at s ∈ { 0 , π } s\in\{0,\pi\} s ∈ { 0 , π } : W 1 ( s ) = ∅ W_1(s)=\varnothing W 1 ( s ) = ∅ off those two points. At them W 1 = W 0 W_1=W_0 W 1 = W 0 , while r ′ ′ ( s ) = − μ 2 2 cos s ≠ 0 r''(s)=-\mu_2^2\cos s\ne0 r ′′ ( s ) = − μ 2 2 cos s = 0 there, so W 2 = ∅ W_2=\varnothing W 2 = ∅ . Thus ⋂ k W k = W 2 = ∅ \bigcap_k W_k=W_2=\varnothing ⋂ k W k = W 2 = ∅ at every s s s , in accordance with (153) with N = dim g = 2 N=\dim\mathfrak{g}=2 N = dim g = 2 and with no exceptional set needed; and the obstruction lies beyond the first stage of the algorithm, since W 0 W_0 W 0 is nowhere empty.
Replacing the second component of F A F_A F A in (156) by the constant 2 2 2 , that is, deleting the term 1 2 x 2 \frac{1}{2}x^2 2 1 x 2 from the second component of A A A , changes the conclusion: then μ 2 = − 1 2 \mu_2=-\frac{1}{2} μ 2 = − 2 1 solves (152) and the same circle is non-strictly abnormal, with μ 1 \mu_1 μ 1 free. The two cases have the same abnormal extremal and the same first-order data, and are distinguished only by (153).
Remark 22. Both components of (156) have been chosen to vanish to a prescribed order on the circle, and it is worth saying which property each one carries. The first is what makes the circle abnormal, and its double zero on { x 2 + y 2 = 1 } \{x^2+y^2=1\} { x 2 + y 2 = 1 } makes ∇ ⟨ μ a b , F A ⟩ \nabla\langle\mu^{\mathrm{ab}},F_A\rangle ∇ ⟨ μ ab , F A ⟩ vanish there as well; this is used in Remark 26 below, and a simple zero would not do. The second is bounded away from zero, which is what makes the first-order test (151) vacuous at every point, and nonconstant, which is what makes the parallelism obstruction bite.
Remark 23. We have assumed D = ker A \mathcal{D}=\ker A D = ker A , so that the reduced problem is Riemannian on B B B , and this hypothesis is used in an essential way rather than for convenience. It enters through Proposition 13, whose conclusion p z = 0 p_z=0 p z = 0 on the abnormal leaf follows from p ∈ D ∘ p\in\mathcal{D}^{\circ} p ∈ D ∘ only because D \mathcal{D} D is the full horizontal space. If D \mathcal{D} D is a proper subbundle of ker A \ker A ker A , then p ∈ D ∘ p\in\mathcal{D}^{\circ} p ∈ D ∘ gives merely p z ∈ ( ρ ∗ D ) ∘ p_z\in(\rho_*\mathcal{D})^{\circ} p z ∈ ( ρ ∗ D ) ∘ , so the abnormal base costate need not vanish; (149) is then not the pair of equations C z μ = 0 C_z\mu=0 C z μ = 0 , ∇ z A μ = 0 \nabla^{A}_{z}\mu=0 ∇ z A μ = 0 alone, and the identification p z = z ˙ ♭ p_z=\dot z^{\flat} p z = z ˙ ♭ on the normal leaf, on which the proof of Theorem 8 rests, likewise fails: only the restriction of p z p_z p z to ρ ∗ D \rho_*\mathcal{D} ρ ∗ D is determined. Strictness then becomes a coupled system in ( p z , μ ) (p_z,\mu) ( p z , μ ) rather than a single acceleration equation for μ \mu μ , and neither Theorem 8 nor Corollary 4 extends as stated. One expects the replacement of (150) to say that the projected curve is a magnetic sub-Riemannian geodesic for a parallel charge, with ∇ z ˙ z ˙ \nabla_{\dot z}\dot z ∇ z ˙ z ˙ replaced by the normal sub-Riemannian geodesic operator of the reduced structure on B B B ; but that is a conjecture and not a corollary of what is proved here, and it would in any case require the full reduced adjoint system together with its own normal and abnormal branches. See Remark 27 for an explicit case in which the hypothesis fails.
The reduced index form, and a trichotomy
Section 7.8 decided whether an abnormal extremal carries a normal lift. It did not decide whether it minimizes, and Proposition 4 disclaimed any second-order statement. We now supply one, in the same reduced variables, and the outcome is a clean division of labor: the second-order condition turns out to be metric -free, so that it is independent of the criterion of Section 7.8 rather than a refinement of it.
Throughout, the hypotheses are those of Theorem 8: the horizontal system (134) with D = ker A \mathcal{D}=\ker A D = ker A carrying the pullback of a Riemannian metric g g g on B B B , and z ( ⋅ ) z(\cdot) z ( ⋅ ) the base projection of an abnormal extremal with ∇ A \nabla^{A} ∇ A -parallel multiplier μ ≠ 0 \mu\ne0 μ = 0 , so that C z μ = 0 C_z\mu=0 C z μ = 0 by (149).
The endpoint map, reduced.
Lemma 7. Fix x 0 ∈ P x_0\in P x 0 ∈ P over z 0 ∈ B z_0\in B z 0 ∈ B and let E : u ( ⋅ ) ↦ x ( T ) E:u(\cdot)\mapsto x(T) E : u ( ⋅ ) ↦ x ( T ) be the endpoint map of (134). Then E E E factors through the base curve: x ( ⋅ ) x(\cdot) x ( ⋅ ) is the horizontal lift of z ( ⋅ ) z(\cdot) z ( ⋅ ) through x 0 x_0 x 0 , and
E ( u ) = ( z ( T ) , hol A ( z ∣ [ 0 , T ] ) ) E(u)=\bigl(z(T),\operatorname{hol}_A(z|_{[0,T]})\bigr) E ( u ) = ( z ( T ) , hol A ( z ∣ [ 0 , T ] ) ) in a local trivialization. If G G G is abelian, with A = θ − a A=\theta-a A = θ − a for a ∈ Ω 1 ( B ; g ) a\in\Omega^1(B;\mathfrak{g}) a ∈ Ω 1 ( B ; g ) locally and F A = − d a F_A=-da F A = − d a , the fiber component is
J [ z ] = ∫ 0 T a ( z ˙ ) d t . (158) J[z]=\int_0^T a(\dot z)\,dt. \tag*{(158)} J [ z ] = ∫ 0 T a ( z ˙ ) d t . (158) An abnormal multiplier is a μ ∈ g ∗ ∖ { 0 } \mu\in\mathfrak{g}^*\setminus\{0\} μ ∈ g ∗ ∖ { 0 } annihilating the image of d E dE d E on variations fixing z ( 0 ) z(0) z ( 0 ) and z ( T ) z(T) z ( T ) , and δ ( μ ⋅ J ) [ v ] = ∫ 0 T ⟨ μ , F A ( z ˙ , v ) ⟩ d t \delta(\mu\cdot J)[v]=\int_0^T\langle\mu,F_A(\dot z,v)\rangle\,dt δ ( μ ⋅ J ) [ v ] = ∫ 0 T ⟨ μ , F A ( z ˙ , v )⟩ d t , so criticality is (149).
Proof. The controlled field is horizontal, so A ( x ˙ ) = 0 A(\dot{x})=0 A ( x ˙ ) = 0 and x x x is the horizontal lift; the fiber displacement of a horizontal lift is the holonomy, which in the abelian case is (158). For the first variation, with J = ∫ a i ( z ˙ ) z ˙ i d t J=\int a_i(\dot{z})\dot{z}^{i}\,dt J = ∫ a i ( z ˙ ) z ˙ i d t and v v v vanishing at the endpoints, an integration by parts gives δ J = ∫ ( ∂ j a i − ∂ i a j ) v j z ˙ i d t = ∫ ( d a ) ( v , z ˙ ) d t = ∫ ⟨ μ , F A ( z ˙ , v ) ⟩ d t \delta J=\int(\partial_j a_i-\partial_i a_j)v^j\dot{z}^{i}\,dt=\int(da)(v,\dot{z})\,dt=\int\langle\mu,F_A(\dot{z},v)\rangle\,dt δ J = ∫ ( ∂ j a i − ∂ i a j ) v j z ˙ i d t = ∫ ( d a ) ( v , z ˙ ) d t = ∫ ⟨ μ , F A ( z ˙ , v )⟩ d t after pairing with μ \mu μ .
Theorem 9. Let G G G be abelian and let z ( ⋅ ) z(\cdot) z ( ⋅ ) be an abnormal extremal with multiplier μ \mu μ . The Hessian of μ ⋅ E \mu\cdot E μ ⋅ E on variations v v v of the base curve vanishing at the endpoints is
Q μ ( v ) = ∫ 0 T ⟨ μ , F A ( ∇ z ˙ v , v ) + ( ∇ v F A ) ( z ˙ , v ) ⟩ d t . (159) Q_\mu(v)=\int_0^T\left\langle\mu,F_A(\nabla_{\dot{z}}v,v)+(\nabla_vF_A)(\dot{z},v)\right\rangle\,dt. \tag*{(159)} Q μ ( v ) = ∫ 0 T ⟨ μ , F A ( ∇ z ˙ v , v ) + ( ∇ v F A ) ( z ˙ , v ) ⟩ d t . (159)
Here ∇ z ˙ v \nabla_{\dot{z}}v ∇ z ˙ v is the covariant derivative of the variation field along z ( ⋅ ) z(\cdot) z ( ⋅ ) with respect to the Levi-Civita connection ∇ \nabla ∇ of g g g , and ∇ F A \nabla F_A ∇ F A is the covariant derivative of the section F A ∈ Γ ( Λ 2 T ∗ B ⊗ g ) F_A\in\Gamma(\Lambda^2T^*B\otimes\mathfrak{g}) F A ∈ Γ ( Λ 2 T ∗ B ⊗ g ) with respect to the connection induced on Λ 2 T ∗ B ⊗ g \Lambda^2T^*B\otimes\mathfrak{g} Λ 2 T ∗ B ⊗ g by ∇ \nabla ∇ on the T ∗ B T^*B T ∗ B -factors and by ∇ A \nabla^A ∇ A on the g \mathfrak{g} g -factor, that is,
( ∇ v F A ) ( w 1 , w 2 ) = ∇ v A ( F A ( W 1 , W 2 ) ) − F A ( ∇ v W 1 , w 2 ) − F A ( w 1 , ∇ v W 2 ) (\nabla_vF_A)(w_1,w_2)=\nabla^A_v(F_A(W_1,W_2))-F_A(\nabla_vW_1,w_2)-F_A(w_1,\nabla_vW_2) ( ∇ v F A ) ( w 1 , w 2 ) = ∇ v A ( F A ( W 1 , W 2 )) − F A ( ∇ v W 1 , w 2 ) − F A ( w 1 , ∇ v W 2 )
for vector fields W 1 , W 2 W_1,W_2 W 1 , W 2 extending w 1 , w 2 w_1,w_2 w 1 , w 2 . The right-hand side of (159) does not change if ∇ \nabla ∇ is replaced by any other torsion-free connection on B B B ; in particular Q μ Q_\mu Q μ does not depend on the metric g g g , but only on A A A and F A F_A F A .
Two remarks on the domain of Q μ Q_\mu Q μ . Since abnormality is the pointwise identity ⟨ μ , F A ( z ˙ , ⋅ ) ⟩ = 0 \langle\mu,F_A(\dot{z},\cdot)\rangle=0 ⟨ μ , F A ( z ˙ , ⋅ )⟩ = 0 , the first variation δ ( μ ⋅ J ) [ v ] = ∫ 0 T ⟨ μ , F A ( z ˙ , v ) ⟩ d t \delta(\mu\cdot J)[v]=\int_0^T\langle\mu,F_A(\dot{z},v)\rangle\,dt δ ( μ ⋅ J ) [ v ] = ∫ 0 T ⟨ μ , F A ( z ˙ , v )⟩ d t vanishes for every v v v with v ( 0 ) = v ( T ) = 0 v(0)=v(T)=0 v ( 0 ) = v ( T ) = 0 , so the Hessian is already well defined, independently of a choice of complement, on that whole space; this is why (159) may be stated there. The index form of Agrachev and Sarychev is, however, the restriction of Q μ Q_\mu Q μ to the intrinsic domain
ker d E = { v : v ( 0 ) = v ( T ) = 0 , ∫ 0 T F A ( z ˙ , v ) d t = 0 ∈ g } . (160) \ker dE=\left\{v:v(0)=v(T)=0,\ \int_0^T F_A(\dot{z},v)\,dt=0\in\mathfrak{g}\right\}. \tag*{(160)} ker d E = { v : v ( 0 ) = v ( T ) = 0 , ∫ 0 T F A ( z ˙ , v ) d t = 0 ∈ g } . (160)
the variations preserving the full endpoint to first order, and it is on (160) that the index and nullity of [3 §4] are computed. The two spaces differ only when dim g ≥ 2 \dim\mathfrak{g}\ge2 dim g ≥ 2 , and definiteness on the larger one is the stronger hypothesis; we shall use only the pointwise consequence (164), for which the distinction is immaterial.
Proof. Take G G G abelian and expand (158) at z + ε v z+\varepsilon v z + ε v . The coefficient of ε 2 \varepsilon^2 ε 2 is
∫ 0 T [ 1 2 ∂ k ∂ j a i v k v j z ˙ i + ∂ j a i v j v ˙ i ] d t , \int_0^T\left[\frac{1}{2}\partial_k\partial_j a_i v^k v^j\dot{z}^{i}+\partial_j a_i v^j\dot{v}^{i}\right]dt, ∫ 0 T [ 2 1 ∂ k ∂ j a i v k v j z ˙ i + ∂ j a i v j v ˙ i ] d t ,
and, writing ω = d a \omega=da ω = d a and using ∂ k ∂ j a i − ∂ i ∂ j a k = ∂ j ω k i \partial_k\partial_j a_i-\partial_i\partial_j a_k=\partial_j\omega_{ki} ∂ k ∂ j a i − ∂ i ∂ j a k = ∂ j ω k i , this equals
1 2 ∫ 0 T [ ω ( v , v ˙ ) + ( ∂ v ω ) ( v , z ˙ ) ] d t + 1 2 [ ∂ j a i v i v j ] 0 T , \frac{1}{2}\int_0^T\left[\omega(v,\dot{v})+(\partial_v\omega)(v,\dot{z})\right]dt+\frac{1}{2}\left[\partial_j a_i v^i v^j\right]_0^T, 2 1 ∫ 0 T [ ω ( v , v ˙ ) + ( ∂ v ω ) ( v , z ˙ ) ] d t + 2 1 [ ∂ j a i v i v j ] 0 T ,
the boundary term vanishing. Since ω = − F A \omega=-F_A ω = − F A , twice this is ∫ [ ⟨ μ , F A ( v ˙ , v ) + ( ∂ v F A ) ( z ˙ , v ) ⟩ ] d t \int[\langle\mu,F_A(\dot{v},v)+(\partial_vF_A)(\dot{z},v)\rangle]dt ∫ [⟨ μ , F A ( v ˙ , v ) + ( ∂ v F A ) ( z ˙ , v )⟩] d t . Replacing the coordinate derivatives by covariant ones with a torsion-free ∇ \nabla ∇ changes the integrand by − ⟨ μ , F A ( z ˙ , Γ ( v , v ) ) ⟩ -\langle\mu,F_A(\dot{z},\Gamma(v,v))\rangle − ⟨ μ , F A ( z ˙ , Γ ( v , v ))⟩ alone, the terms in Γ ( z ˙ , v ) \Gamma(\dot{z},v) Γ ( z ˙ , v ) canceling by symmetry of Γ \Gamma Γ ; and that remainder vanishes because ⟨ μ , F A ( z ˙ , ⋅ ) ⟩ = 0 \langle\mu,F_A(\dot{z},\cdot)\rangle=0 ⟨ μ , F A ( z ˙ , ⋅ )⟩ = 0 along an abnormal extremal. The same cancellation proves the independence of the torsion-free connection: if ∇ ′ = ∇ + S \nabla'=\nabla+S ∇ ′ = ∇ + S with S S S a symmetric ( 1 , 2 ) (1,2) ( 1 , 2 ) -tensor field, then
F A ( ∇ z ˙ ′ v , v ) − F A ( ∇ z ˙ v , v ) = F A ( S ( z ˙ , v ) , v ) , F_A(\nabla'_{\dot{z}}v,v)-F_A(\nabla_{\dot{z}}v,v)=F_A(S(\dot{z},v),v), F A ( ∇ z ˙ ′ v , v ) − F A ( ∇ z ˙ v , v ) = F A ( S ( z ˙ , v ) , v ) ,
( ∇ v ′ F A ) ( z ˙ , v ) − ( ∇ v F A ) ( z ˙ , v ) = − F A ( S ( v , z ˙ ) , v ) − F A ( z ˙ , S ( v , v ) ) ; (\nabla'_vF_A)(\dot{z},v)-(\nabla_vF_A)(\dot{z},v)=-F_A(S(v,\dot{z}),v)-F_A(\dot{z},S(v,v)); ( ∇ v ′ F A ) ( z ˙ , v ) − ( ∇ v F A ) ( z ˙ , v ) = − F A ( S ( v , z ˙ ) , v ) − F A ( z ˙ , S ( v , v )) ;
the first two changes cancel by the symmetry of S S S , and the last is annihilated by μ \mu μ along an abnormal extremal, since C z μ = 0 C_z\mu=0 C z μ = 0 . The g \mathfrak{g} g -factor of the connection on Λ 2 T ∗ B ⊗ g \Lambda^2T^*B\otimes\mathfrak{g} Λ 2 T ∗ B ⊗ g is ∇ A \nabla^A ∇ A in every case, fixed by A A A , so the metric enters only through the T B TB T B -factor, whose choice has just been shown to be immaterial. □
Remark 24. The abelian hypothesis in Theorem 9 is not an artifact of the proof. For nonabelian G G G the fiber component of the endpoint map is the path-ordered exponential Pexp ( ∫ 0 T α ( z ˙ ) d t ) \operatorname{Pexp}\left(\int_0^T\alpha(\dot{z})\,dt\right) Pexp ( ∫ 0 T α ( z ˙ ) d t ) , which reduces to (158) only in the abelian case, and the second-order term of its Magnus–Chen expansion contains, besides the terms of (159), the time-ordered double integral
⟨ μ 0 , ∫ 0 T ∫ 0 t [ ν ( σ ) , ν ( t ) ] d σ d t ⟩ , ν ( t ) = Ad U ( t ) ( F A ( z ˙ , v ) ( t ) ) , (161) \left\langle\mu_0,\int_0^T\int_0^t[\nu(\sigma),\nu(t)]\,d\sigma\,dt\right\rangle,\qquad\nu(t)=\operatorname{Ad}_{U(t)}(F_A(\dot{z},v)(t)), \tag*{(161)} ⟨ μ 0 , ∫ 0 T ∫ 0 t [ ν ( σ ) , ν ( t )] d σ d t ⟩ , ν ( t ) = Ad U ( t ) ( F A ( z ˙ , v ) ( t )) , (161)
where U ( t ) U(t) U ( t ) is the holonomy up to time t t t and μ 0 \mu_0 μ 0 the multiplier at the initial point. Abnormality gives ⟨ μ 0 , ν ( t ) ⟩ ≡ 0 \langle\mu_0,\nu(t)\rangle\equiv0 ⟨ μ 0 , ν ( t )⟩ ≡ 0 , that is, ν ( t ) ∈ ker μ 0 \nu(t)\in\ker\mu_0 ν ( t ) ∈ ker μ 0 , which does not annihilate (161): the term vanishes identically precisely when ⟨ μ 0 , [ ker μ 0 , ker μ 0 ] ⟩ = 0 \langle\mu_0,[\ker\mu_0,\ker\mu_0]\rangle=0 ⟨ μ 0 , [ ker μ 0 , ker μ 0 ]⟩ = 0 , in particular whenever g \mathfrak{g} g is abelian. Nor is it removed by the Goh condition. It is not negligible either: for B = R 2 B=\mathbb{R}^2 B = R 2 , G = S O ( 3 ) G=\mathrm{SO}(3) G = SO ( 3 ) , P = B × G P=B\times G P = B × G and the constant local connection form a = c e 3 d x + e 1 d y a=c\,e_3\,dx+e_1\,dy a = c e 3 d x + e 1 d y , so that F A = 1 2 [ a , a ] = c e 2 d x ∧ d y F_A=\frac{1}{2}[a,a]=c\,e_2\,dx\wedge dy F A = 2 1 [ a , a ] = c e 2 d x ∧ d y , the line z ( t ) = ( t , 0 ) z(t)=(t,0) z ( t ) = ( t , 0 ) with μ 0 = e 3 ∗ \mu_0=e^*_3 μ 0 = e 3 ∗ is abnormal (note that ad e 3 ∗ e 3 ∗ = 0 \operatorname{ad}^*_{e_3}e^*_3=0 ad e 3 ∗ e 3 ∗ = 0 , so μ 0 \mu_0 μ 0 is indeed ∇ A \nabla^A ∇ A -parallel), satisfies the full Goh condition C μ ≡ 0 C_\mu\equiv0 C μ ≡ 0 , and has L μ ( v ) = c v 2 \mathcal{L}_\mu(v)=cv^2 L μ ( v ) = c v 2 . Take c > 0 c>0 c > 0 . Formula (159) would then give Q μ ( v ) = c ∫ 0 T y 2 d t > 0 Q_\mu(v)=c\int_0^T y^2\,dt>0 Q μ ( v ) = c ∫ 0 T y 2 d t > 0 for every T T T , whereas the true second variation acquires from (161) a Lévy-area functional whose kernel is proportional to sinc ( c ( t − σ ) ) \operatorname{sinc}(c(t-\sigma)) sinc ( c ( t − σ )) and whose smallest eigenvalue becomes negative for c T cT c T large. The pointwise consequence (164) survives for general G G G , since under a concentration of width ε \varepsilon ε the term (161) is O ( ε 2 ) O(\varepsilon^2) O ( ε 2 ) against the O ( ε ) O(\varepsilon) O ( ε ) of R L μ R_{L_\mu} R L μ ; what needs G G G abelian is the global conclusion of Corollary 5. Deriving the nonabelian index form, and identifying the hypotheses under which (161) cancels, we leave open.
Lemma 8. Let V , Z V,Z V , Z be vector fields on B B B and let λ = μ ∘ A ∈ D ∘ \lambda=\mu\circ A\in D^\circ λ = μ ∘ A ∈ D ∘ correspond to μ \mu μ under (18). At a point at which ∇ V = ∇ Z = 0 \nabla V=\nabla Z=0 ∇ V = ∇ Z = 0 ,
λ ( [ V h , [ Z h , V h ] ] ) = − ⟨ μ , ( ∇ V F A ) ( Z , V ) ⟩ . (162) \lambda\left([V^h,[Z^h,V^h]]\right)=-\langle\mu,(\nabla_VF_A)(Z,V)\rangle. \tag*{(162)} λ ( [ V h , [ Z h , V h ]] ) = − ⟨ μ , ( ∇ V F A ) ( Z , V )⟩ . (162)
Proof. By Lemma 4, [ Z h , V h ] = [ Z , V ] h − ( F A ( Z , V ) ) † [Z^h,V^h]=[Z,V]^h-(F_A(Z,V))^\dagger [ Z h , V h ] = [ Z , V ] h − ( F A ( Z , V ) ) † , where σ † \sigma^\dagger σ † denotes the vertical field of a section σ \sigma σ of g ~ \widetilde{\mathfrak{g}} g . Since [ V h , σ † ] = ( ∇ V A σ ) † [V^h,\sigma^\dagger]=(\nabla^A_V\sigma)^\dagger [ V h , σ † ] = ( ∇ V A σ ) † ,
A ( [ V h , [ Z h , V h ] ] ) = − F A ( V , [ Z , V ] ) − ∇ V A ( F A ( Z , V ) ) . A([V^h,[Z^h,V^h]])=-F_A(V,[Z,V])-\nabla^A_V(F_A(Z,V)). A ([ V h , [ Z h , V h ]]) = − F A ( V , [ Z , V ]) − ∇ V A ( F A ( Z , V )) .
Expanding ∇ V A ( F A ( Z , V ) ) = ( ∇ V F A ) ( Z , V ) + F A ( ∇ V Z , V ) + F A ( Z , ∇ V V ) \nabla^A_V(F_A(Z,V))=(\nabla_VF_A)(Z,V)+F_A(\nabla_VZ,V)+F_A(Z,\nabla_VV) ∇ V A ( F A ( Z , V )) = ( ∇ V F A ) ( Z , V ) + F A ( ∇ V Z , V ) + F A ( Z , ∇ V V ) and [ Z , V ] = ∇ Z V − ∇ V Z [Z,V]=\nabla_ZV-\nabla_VZ [ Z , V ] = ∇ Z V − ∇ V Z , and evaluating where ∇ V = ∇ Z = 0 \nabla V=\nabla Z=0 ∇ V = ∇ Z = 0 , leaves − ( ∇ V F A ) ( Z , V ) -(\nabla_VF_A)(Z,V) − ( ∇ V F A ) ( Z , V ) ; the term F A ( Z , ∇ V V ) F_A(Z,\nabla_VV) F A ( Z , ∇ V V ) is in any case killed by μ \mu μ on an abnormal. □
Identity (162) is the translation announced above: the quadratic form v ↦ λ ( [ v , [ γ ˙ , v ] ] ) v\mapsto\lambda([v,[\dot\gamma,v]]) v ↦ λ ([ v , [ γ ˙ , v ]]) of Agrachev and Sarychev [3, (4.15)], which is the generalized Legendre form of an abnormal extremal, becomes on a principal bundle a form built from the covariant derivative of the curvature. We give it a name.
Definition 15. The reduced Legendre form of an abnormal extremal at time t t t is the quadratic form
L μ ( v ) = ⟨ μ , ( ∇ v F A ) ( z ˙ , v ) ⟩ , v ∈ T z B . (163) \mathcal{L}_\mu(v)=\langle\mu,(\nabla_vF_A)(\dot z,v)\rangle,\qquad v\in T_zB. \tag*{(163)} L μ ( v ) = ⟨ μ , ( ∇ v F A ) ( z ˙ , v )⟩ , v ∈ T z B . (163)
Its domain is the full tangent space of the base, modulo the velocity, and not the kernel K z ˙ K_{\dot z} K z ˙ of the curvature pairing. This is worth stating explicitly, because K z ˙ K_{\dot z} K z ˙ is the natural domain of the first-order obstruction (151) and the analogy is misleading. The reason is that the generalized Legendre form of [3 (4.15)] is tested on all control directions: taking v = φ ( t ) e ( t ) v=\varphi(t)e(t) v = φ ( t ) e ( t ) with e e e a ∇ \nabla ∇ -parallel field and φ \varphi φ a bump function makes the first term of (159) vanish, since F A ( ∇ z ˙ v , v ) = φ ′ φ F A ( e , e ) = 0 F_A(\nabla_{\dot z}v,v)=\varphi'\varphi F_A(e,e)=0 F A ( ∇ z ˙ v , v ) = φ ′ φ F A ( e , e ) = 0 , and leaves Q μ ( φ e ) = ∫ φ 2 L μ ( e ) d t Q_\mu(\varphi e)=\int\varphi^2\mathcal{L}_\mu(e)\,dt Q μ ( φ e ) = ∫ φ 2 L μ ( e ) d t , so that L μ ( e ) ≥ 0 \mathcal{L}_\mu(e)\geq0 L μ ( e ) ≥ 0 is forced for every e ∈ T z B e\in T_zB e ∈ T z B . Restricting the test to K z ˙ K_{\dot z} K z ˙ would make it vacuous whenever K z ˙ = ⟨ z ˙ ⟩ K_{\dot z}=\langle\dot z\rangle K z ˙ = ⟨ z ˙ ⟩ , that is, in exactly the situations the test is meant to decide.
Two properties justify passing to the quotient by z ˙ \dot z z ˙ . First, L μ ( z ˙ ) = 0 \mathcal{L}_\mu(\dot z)=0 L μ ( z ˙ ) = 0 by antisymmetry of F A F_A F A . Second, and less immediately, L μ \mathcal{L}_\mu L μ descends to T z B / ⟨ z ˙ ⟩ T_zB/\langle\dot z\rangle T z B / ⟨ z ˙ ⟩ under the full Goh condition C μ = 0 C_\mu=0 C μ = 0 : polarizing (163) and differentiating the abnormality relation ⟨ μ , F A ( z ˙ , V ) ⟩ = 0 \langle\mu,F_A(\dot z,V)\rangle=0 ⟨ μ , F A ( z ˙ , V )⟩ = 0 along z ( ⋅ ) z(\cdot) z ( ⋅ ) , using ∇ z ˙ A μ = 0 \nabla^A_{\dot z}\mu=0 ∇ z ˙ A μ = 0 , gives
2 L μ ( v , z ˙ ) = ⟨ μ , ( ∇ z ˙ F A ) ( z ˙ , v ) ⟩ = − ⟨ μ , F A ( ∇ z ˙ z ˙ , v ) ⟩ = − C μ ( ∇ z ˙ z ˙ , v ) , 2\mathcal{L}_\mu(v,\dot z)=\langle\mu,(\nabla_{\dot z}F_A)(\dot z,v)\rangle=-\langle\mu,F_A(\nabla_{\dot z}\dot z,v)\rangle=-C_\mu(\nabla_{\dot z}\dot z,v), 2 L μ ( v , z ˙ ) = ⟨ μ , ( ∇ z ˙ F A ) ( z ˙ , v )⟩ = − ⟨ μ , F A ( ∇ z ˙ z ˙ , v )⟩ = − C μ ( ∇ z ˙ z ˙ , v ) ,
which vanishes when C μ = 0 C_\mu=0 C μ = 0 . Note that this is a further reason why K z ˙ K_{\dot z} K z ˙ is not the right domain: L μ \mathcal{L}_\mu L μ need not descend to K z ˙ / ⟨ z ˙ ⟩ K_{\dot z}/\langle\dot z\rangle K z ˙ / ⟨ z ˙ ⟩ in the absence of Goh either.
Corollary 5. Along an abnormal extremal the generalized Legendre–Clebsch condition reads
L μ ≥ 0 on T z B . (164) \mathcal{L}_\mu\geq0\quad\text{on }T_zB. \tag*{(164)} L μ ≥ 0 on T z B . (164)
after normalizing the sign of μ \mu μ ; this holds for every structure group, by the localization argument in the proof below together with Remark 24. Suppose in addition that G G G is abelian, that the Goh condition holds in its full form, C μ = 0 C_\mu=0 C μ = 0 along the extremal, and that L μ \mathcal{L}_\mu L μ is uniformly definite on T z B / ⟨ z ˙ ⟩ T_zB/\langle\dot z\rangle T z B / ⟨ z ˙ ⟩ . Then the extremal is W 1 , ∞ W^{1,\infty} W 1 , ∞ -rigid, hence a local length minimizer.
The full Goh condition C μ = 0 C_\mu=0 C μ = 0 is a genuine additional hypothesis and is not supplied by the constraint algorithm. What Theorem 6 produces as a secondary constraint is only the tangency relation (137), that is, C μ ( z ˙ , ⋅ ) = 0 C_\mu(\dot z,\cdot)=0 C μ ( z ˙ , ⋅ ) = 0 ; the two coincide when dim B = 2 \dim B=2 dim B = 2 , where a degenerate skew form on a plane vanishes, and they differ as soon as dim B > 2 \dim B>2 dim B > 2 . This is the distinction already recorded after Theorem 6. The full condition holds in all the examples below.
Proof. For (164), take v = φ e v=\varphi e v = φ e as in the discussion following Definition 15; the correction needed to place v v v in ker d E \ker dE ker d E is of second order in the width of φ \varphi φ and does not affect the sign. This argument is local and survives for nonabelian G G G , since by Remark 24 the extra holonomy term is of second order in that width against the first order of ∫ L μ \int\mathcal{L}_{\mu} ∫ L μ .
For the second assertion, C μ = 0 C_{\mu}=0 C μ = 0 annihilates the first term of (159), so that
Q μ ( v ) = ∫ 0 T L μ ( v ) d t (165) Q_{\mu}(v)=\int_{0}^{T}\mathcal{L}_{\mu}(v)\,\mathrm{d}t \tag*{(165)} Q μ ( v ) = ∫ 0 T L μ ( v ) d t (165)
exactly, with no nonlocal quadratic remainder. This is where the abelian hypothesis is used, and it is the whole of the argument: Remark 24 exhibits the term that would otherwise appear on the right of (165). Uniform definiteness means L μ ( v ) ≥ β ∣ v ˉ ∣ 2 \mathcal{L}_{\mu}(v)\geq\beta\lvert\bar{v}\rvert^{2} L μ ( v ) ≥ β ∣ v ˉ ∣ 2 , with v ˉ \bar{v} v ˉ the class of v v v in T z B / ⟨ z ˙ ⟩ T_{z}B/\langle\dot{z}\rangle T z B / ⟨ z ˙ ⟩ and β > 0 \beta>0 β > 0 independent of t t t ; by (165),
Q μ ( v ) ≥ β ∥ v ˉ ∥ L 2 2 (166) Q_{\mu}(v)\geq\beta\lVert\bar{v}\rVert_{L^{2}}^{2} \tag*{(166)} Q μ ( v ) ≥ β ∥ v ˉ ∥ L 2 2 (166)
for every variation v v v vanishing at the endpoints. Writing w = v ˙ w=\dot{v} w = v ˙ for the corresponding control variation, so that v v v is its primitive vanishing at t = 0 t=0 t = 0 , the right-hand side of (166) is β \beta β times the squared H − 1 H^{-1} H − 1 norm of w w w in the convention of [3 §4] , in which the duality is taken against test functions not required to vanish at t = T t=T t = T , so that
∥ v ˙ ∥ H − 1 = sup { ∫ 0 T v ψ d t : ∥ ψ ∥ L 2 ≤ 1 } = ∥ v ∥ L 2 \lVert\dot{v}\rVert_{H^{-1}}=\sup\left\{\int_{0}^{T}v\psi\,\mathrm{d}t:\lVert\psi\rVert_{L^{2}}\leq1\right\}=\lVert v\rVert_{L^{2}} ∥ v ˙ ∥ H − 1 = sup { ∫ 0 T v ψ d t : ∥ ψ ∥ L 2 ≤ 1 } = ∥ v ∥ L 2
for v ( 0 ) = 0 v(0)=0 v ( 0 ) = 0 ; the estimate is taken on the quotient by the reparametrization direction z ˙ \dot{z} z ˙ along which the unparametrized curve does not move. Estimate (166) is therefore the weak positive definiteness of the second variation hypothesized in [3 Thm. 4.8] , whose conclusion is W 1 , ∞ W^{1,\infty} W 1 , ∞ -rigidity of the whole arc, that is, isolation up to reparametrization. □
Remark 25 (attribution and terminology). The ingredients are classical and are used as such. The index form of an abnormal extremal, its metric independence, the Goh and generalized Legendre conditions and the rigidity theorem are due to Agrachev and Sarychev [3 ] , in the Lie-bracket form λ ( [ v , [ γ ˙ , v ] ] ) \lambda([v,[\dot{\gamma},v]]) λ ([ v , [ γ ˙ , v ]]) on the total space; the description of the abnormal extremals of a bundle sub-Riemannian structure as the horizontal lifts of the curves with ⟨ μ , F A ( z ˙ , ⋅ ) ⟩ = 0 \langle\mu,F_{A}(\dot{z},\cdot)\rangle=0 ⟨ μ , F A ( z ˙ , ⋅ )⟩ = 0 and ∇ z ˙ A μ = 0 \nabla^{A}_{\dot{z}}\mu=0 ∇ z ˙ A μ = 0 is Montgomery’s [32 §3.2] ; conditions that rule out strictly abnormal minimizers using the distribution alone (medium fatness) are in [4 ] . What is done here is to write the second-order theory in the same reduced variables as the first-order theory, so that all three rows of Theorem 10 below are computations with F A F_{A} F A , ∇ F A \nabla F_{A} ∇ F A and g g g on the base rather than with iterated brackets upstairs. Three points of interpretation follow. First, the hypotheses of the second assertion of Corollary 5 should not be weakened casually: Theorem 4.8 of [3 ] assumes weak positive definiteness of the full second variation, and what pointwise Goh together with the strong Legendre condition gives in general is the finite-codimension positivity of [3 Prop. 4.5] , from which rigidity follows for sufficiently short subarcs [3 Thm. 5.5] but not for the whole arc; what rescues the global conclusion in the abelian case is the exactness of Q μ = ∫ L μ Q_{\mu}=\int\mathcal{L}_{\mu} Q μ = ∫ L μ under full Goh, displayed in the proof, whereas for nonabelian G G G the remainder (161) is present and, by the example of Remark 24, genuinely destroys positivity on long arcs. Second, the constraint algorithm of Section 7 produces, as its secondary constraints, the equations of Goh and of Legendre–Clebsch [5 ] , as Theorem 6 records for the first of them, but it is insensitive to the replacement of μ \mu μ by − μ -\mu − μ and so cannot produce a sign; the index form (159) supplies exactly that missing datum, and (164) is its first consequence. Third, rigid is applied by Hsu [24 Prop. 8] and by Bryant and Hsu [12 ] to an integral curve admitting no compactly supported variations, and by Agrachev and Sarychev [3 Thm. 4.8] to a curve isolated, up to reparametrization, in the W 1 , ∞ W^{1,\infty} W 1 , ∞ topology; Corollary 5 asserts the second. The C 0 C^{0} C 0 -isolation among geodesics of bounded length in [1 Thm. 1.3] is a further, distinct statement, and we do not assert that it is equivalent to W 1 , ∞ W^{1,\infty} W 1 , ∞ -rigidity in the sense of [3 ] : the admissible classes and the topologies differ, and the two conclusions should be read separately.
The trichotomy.
Theorem 10. For the horizontal system (134) the following three questions about an admissible curve are answered by three reduced tests, of which the first two involve only the connection and its curvature and the third involves the metric:
Table 1.
The first is a precondition for the other two, and the two remaining tests are independent of one another: the outcome of one determines neither the outcome nor the applicability of the other.
Three qualifications belong with the table. The second row is a sufficient certificate and not a criterion: by Corollary 5 and Remark 25 it certifies rigidity of the whole arc for abelian G G G , and of sufficiently short subarcs in general, while the corresponding necessary condition is the weaker L μ ⪰ 0 L_\mu\succeq0 L μ ⪰ 0 of (164) . The first and third rows require the connection A A A itself and not only its curvature, since both involve ∇ A \nabla_A ∇ A -parallel transport of μ \mu μ , and F A F_A F A does not determine A A A . And the metric enters the third row only, through ∇ z ˙ z ˙ \nabla_{\dot z}\dot z ∇ z ˙ z ˙ , which is the content of Remark 21.
Proof. The three tests are (149) , Corollary 5 and Theorem 8; the assertions about the data used are Theorem 9 and the definition of ∇ \nabla ∇ . For the independence, the family of Remark 19, worked out below, is rigid for every ε \varepsilon ε and changes strictness at ε = 0 \varepsilon= 0 ε = 0 , so rigidity determines nothing about strictness; and the rank-two example of Section 7.8 is strictly abnormal while its reduced Legendre form vanishes identically, so that the second-order test is silent about its rigidity, by Remark 26. The strictness test can therefore be decisive where the second-order test is not. □
The examples. For the flat Martinet problem of Section 7.6, F A = − x d x ∧ d y F_A=-x\,dx\wedge dy F A = − x d x ∧ d y and the abnormal is x ≡ 0 x \equiv0 x ≡ 0 with z ˙ = ( 0 , 1 ) \dot z=(0,1) z ˙ = ( 0 , 1 ) , so
L μ ( v ) = μ v x 2 (167) L_\mu(v)=\mu v_x^2 \tag*{(167)} L μ ( v ) = μ v x 2 (167)
definite on T z B / ⟨ z ˙ ⟩ T_zB/\langle\dot z\rangle T z B / ⟨ z ˙ ⟩ , which is here the line spanned by the class of ∂ x \partial_x ∂ x , and uniformly so since μ \mu μ is a nonzero constant. The group is R \mathbb{R} R , hence abelian, and C μ = 0 C_\mu=0 C μ = 0 because F A F_A F A vanishes on the abnormal, so Corollary 5 applies to the whole arc: the abnormal is rigid and locally minimizing. This is consistent with, and independent of, [1 Thm. 1.2] , where minimality is obtained from the explicit description of the sphere: there the abnormal geodesic is removed from the cut locus, so it minimizes for all time.
For the deformation of Remark 19, with F A ε F_{A_\varepsilon} F A ε as displayed there, the metric (154) , the domain (144) and the abnormal y ≡ 0 y \equiv0 y ≡ 0 , z ˙ = ( 1 , 0 ) \dot z=(1,0) z ˙ = ( 1 , 0 ) ,
L μ ( v ) = μ v y 2 for every ε , (168) L_\mu(v)=\mu v_y^2 \qquad\text{for every }\varepsilon, \tag*{(168)} L μ ( v ) = μ v y 2 for every ε , (168)
using f ε ′ ( 0 ) = 1 f_\varepsilon'(0)=1 f ε ′ ( 0 ) = 1 for f ε ( y ) = y ( 2 + ε y ) 2 ( 1 + ε y ) 2 f_\varepsilon(y)=\frac{y(2+\varepsilon y)}{2(1+\varepsilon y)^2} f ε ( y ) = 2 ( 1 + ε y ) 2 y ( 2 + ε y ) ; again definite on T z B / ⟨ z ˙ ⟩ T_zB/\langle\dot z\rangle T z B / ⟨ z ˙ ⟩ , and again with R \mathbb{R} R abelian and C μ = 0 C_\mu=0 C μ = 0 on the abnormal, so the abnormal is rigid for every ε \varepsilon ε . This is the counterpart, for the whole family and not only for ε ≠ 0 \varepsilon\ne0 ε = 0 , of the C 0 C^0 C 0 -isolation of [1 Thm. 1.3] ; the two statements are not the same, and we do not identify them (see Remark 25). Together with Proposition 18 this is the promised independence: the location of the abnormal and its rigidity are the same for all ε \varepsilon ε , and only strictness changes. In the language of Theorem 10, the deformation moves nothing in the first two rows and everything in the third.
Remark 26. The rank-two example of Section 7.8 shows the other degeneracy. There μ a b , F A = ( x 2 + y 2 − 1 ) d x ∧ d y \mu^{\mathrm{ab}},F_A=(x^2+y^2-1)\,dx\wedge dy μ ab , F A = ( x 2 + y 2 − 1 ) d x ∧ d y , and since ∇ ( x 2 + y 2 − 1 ) 2 = 4 ( x 2 + y 2 − 1 ) ∇ \nabla(x^2+y^2-1)^2=4(x^2+y^2-1)\nabla ∇ ( x 2 + y 2 − 1 ) 2 = 4 ( x 2 + y 2 − 1 ) ∇ vanishes on the unit circle,
L μ a b ≡ 0 along the circle. L_{\mu^{\mathrm{ab}}}\equiv0 \qquad\text{along the circle.} L μ ab ≡ 0 along the circle.
The Goh condition holds in its full form there, C μ a b = 0 C_{\mu^{\mathrm{ab}}}=0 C μ ab = 0 , since ( μ a b , F A ) (\mu^{\mathrm{ab}},F_A) ( μ ab , F A ) itself vanishes on the circle. So (164) holds, but only with equality: the reduced Legendre form is degenerate, Corollary 5 does not apply, and the second-order test says nothing about rigidity. Strictness, by contrast, was decided there. This is the promised complementarity: the two tests can each be decisive where the other is not.
The double zero of the first curvature component is what produces the degeneracy, as noted in Remark 22. Had that component been x 2 + y 2 − 1 x^2+y^2-1 x 2 + y 2 − 1 , with a simple zero, one would have obtained L μ a b ( v ) = − 2 ⟨ z , v ⟩ 2 L_{\mu^{\mathrm{ab}}}(v)=-2\langle z,v\rangle^2 L μ ab ( v ) = − 2 ⟨ z , v ⟩ 2 , whose kernel is exactly ⟨ z ˙ ⟩ \langle\dot z\rangle ⟨ z ˙ ⟩ ; on the one-dimensional quotient T z B / ⟨ z ˙ ⟩ T_zB/\langle\dot z\rangle T z B / ⟨ z ˙ ⟩ , spanned by the radial class, that form is definite after normalizing μ a b \mu^{\mathrm{ab}} μ ab to ( − 1 , 0 ) (-1,0) ( − 1 , 0 ) , so Corollary 5 would certify the circle as rigid, hence locally minimizing — a strictly abnormal local minimizer, decided by both tests. Both variants are instructive, and it is the domain T z B / ⟨ z ˙ ⟩ T_zB/\langle\dot z\rangle T z B / ⟨ z ˙ ⟩ of Definition 15, rather than K z / ⟨ z ˙ ⟩ K_z/\langle\dot z\rangle K z / ⟨ z ˙ ⟩ , that makes the second computation meaningful: on K z K_z K z the test would be empty in both variants, for the accidental reason that K z = ⟨ z ˙ ⟩ K_z=\langle\dot z\rangle K z = ⟨ z ˙ ⟩ there, and would therefore distinguish nothing.
Remark 27. The standing hypothesis D = ker A \mathcal{D}=\ker A D = ker A of this subsection and of Section 7.8 is essential rather than convenient, and the example of [24 ] shows why. The Hilbert–Cartan system on R 5 \mathbb{R}^5 R 5 , generated by
θ 1 = d y − p d x , θ 2 = d p − q d x , θ 3 = d z − q 2 d x , \theta^1=dy-p\,dx,\qquad\theta^2=dp-q\,dx,\qquad\theta^3=dz-q^2\,dx, θ 1 = d y − p d x , θ 2 = d p − q d x , θ 3 = d z − q 2 d x ,
is invariant under translation in z z z , so that G = R G=\mathbb{R} G = R , B = R 4 B=\mathbb{R}^4 B = R 4 with coordinates ( x , y , p , q ) (x,y,p,q) ( x , y , p , q ) , and
A = d z − q 2 d x , F A = 2 q d x ∧ d q . A=dz-q^2\,dx,\qquad F_A=2q\,dx\wedge dq. A = d z − q 2 d x , F A = 2 q d x ∧ d q .
But D \mathcal{D} D has rank two while ker A \ker A ker A has rank four, and the conclusion of Theorem 6 fails accordingly. The family R \mathcal{R} R of [24 (2.16)] consists of the integral curves with q ′ ′ = 0 q^{\prime\prime}=0 q ′′ = 0 , that is, q ( x ) = a 2 + a 3 x q(x)=a_2+a_3x q ( x ) = a 2 + a 3 x with y y y the corresponding cubic in x x x ; it carries the five parameters ( a 0 , a 1 , a 2 , a 3 , b 0 ) (a_0,a_1,a_2,a_3,b_0) ( a 0 , a 1 , a 2 , a 3 , b 0 ) . Along such a curve, with x ˙ = 1 \dot{x}=1 x ˙ = 1 , the reduced distribution contains ∂ q \partial_q ∂ q , and F A ( z ˙ , ∂ q ) = 2 q ( x ) F_A(\dot{z},\partial_q)=2q(x) F A ( z ˙ , ∂ q ) = 2 q ( x ) , which vanishes identically along the curve only on the three-parameter subfamily q ≡ 0 q\equiv0 q ≡ 0 . The curvature pairing therefore does not detect the generic member of R \mathcal{R} R . What goes wrong is that when D \mathcal{D} D is a proper subbundle of ker A \ker A ker A the brackets of D \mathcal{D} D are not generated by F A F_A F A alone, and the reduced problem is sub-Riemannian rather than Riemannian on the base, as anticipated in Remark 23.
Departamento de Matemáticas, Universidad Carlos III de Madrid, Avda. de la Universidad 30, 28911 Leganés, Madrid, Spain
Email address: albertoi@math.uc3m.es
Department of Mathematics, University of California, San Diego, 9500 Gilman Drive, La Jolla, CA 92093-0112, USA
Email address: mleok@ucsd.edu
Departamento de Matemáticas, Universidad Carlos III de Madrid, Avda. de la Universidad 30, 28911 Leganés, Madrid, Spain
Departamento de Matemática Aplicada, Universidad Politécnica de Madrid, Avenida Juan de Herrera 6, 28040 Madrid, Spain
Email address: mariacristina.sardon@upm.es