Duality on geodesics of Cartan distributions and sub-Riemannian pseudo-product structures

Given a five dimensional space endowed with a Cartan distribution, the abnormal geodesics form another five dimensional space with a cone structure. Then it is shown, if the cone structure is regarded as a control system, then, the space of abnormal geodesics of the cone structure is naturally identified with the original space. In this paper, we provide an exposition on the duality by abnormal geodesics in a wider framework, namely, in terms of quotients of control systems and sub-Riemannian pseudo-product structures. Also we consider the controllability of cone structures and describe the constrained Hamiltonian equations on normal and abnormal geodesics.

Let D ⊂ T Y be a Cartan distribution on Y . In [15], it is shown that abnormal geodesics (or abnormal extremals, singular curves) of D form another five dimensional space X with a cone structure C ⊂ T X, and if the cone structure is regarded as a control system, the space of "abnormal geodesics" of the cone structure is naturally identified with the original space Y . In fact, the duality is studied via the Cartan prolongation (Z, E) of the given Cartan distribution (Y, D) and the double fibration Then we obtain the decomposition E = L ⊕ K by integrable subbundles L = Ker(π Y * ), K = Ker(π X * ) of E. Note that abnormal geodesics are called singular paths in [15].
For a space with Cartan distribution, the construction of the space of abnormal geodesics and the existence of natural double fibration was mentioned for the first time in unpublished lecture notes by Bryant [9]. Nurowski [23] constructed indefinite conformal structures of signature (3,2) on such spaces. Then, as an alternative construction, the same cone structure as in [15] was explicitly used in [4] before [15]. In fact, in [4], the cone structure was given by a foliation on the space P ((D 2 ) ⊥ ) ⊂ P (T * Y ), while in [15] we describe it from an essentially same foliation in the space P (D) ⊂ P (T Y ), which is canonically identified with P (D 2 ) ⊥ .
The construction of double fibrations and cone structures reminds us the general notion of pseudo-product structures in the sense of Tanaka ([26][29] [28], see also [16]).
A pseudo-product structure on a manifold M is a distribution E ⊂ T M on a manifold M with a decomposition E = L ⊕ K into integrable subbundles. Then locally we have a double fibration The abnormal geodesics for the Cartan distribution with the maximal symmetry, namely with G 2 -symmetry, was calculated in [19]. The double fibration Y ← Z → X with G 2 -symmetry and the canonical geometric structures E ⊂ T Z, D ⊂ T Y and C ⊂ T X have been explicitly constructed in [18]. Then in [15] we determined abnormal geodesics, by direct calculations in G 2case for control systems E : E ֒→ T Z → Z, D : D ֒→ T Y → Y and C = E/π X : E → T X → X. Theorems 4.1 and 4.3 show, from the viewpoint of geometric control theory the duality and asymmetry hold, not only for G 2 -case, but also for general Cartan distributions. Note that double fibrations and cone structures arise also in works by Doubrov and Zelenko [12][13] [14]. In Sato and Yamaguchi's paper [25], a kind of geometric structure is naturally constructed on a space of singular curves (abnormal geodesics).
In §2, we recall the basic notions in geometric control theory, equivalence, admissible controls, controllability, etc., and, in particular, we make clear the notion of "quotients" of control systems. In §3, we introduce the notion of abnormal geodesics, or singular paths, of a given control system and show the constrained Hamiltonian system describing abnormal geodesics. The duality and asymmetry on abnormal geodesics (singular paths) for Cartan distributions ( [15]) is explained in §4. Then the exposition is performed along the general notion of "pseudo-product structures" in §5. In fact we show that a pseudo-product structure induces locally but naturally a double fibration and cone-structures, and we discuss the controllability of cone-structures briefly. After recalling the optimal control problems and associated normal and abnormal extremals in §6, we give the constrained Hamiltonian systems describing both normal and abnormal extremals (geodesics) for pseudo-product structures with sub-Riemannian metrics and for associated conestructures with metrics in §7. Distributions can be regarded as cone structures of special types. Then naturally we hope the basic constructions given in this paper can be helpful to treat cone structures arising from pseudo-product structures. In §8, we apply the general theory presented in the previous sections to the case of Cartan distributions. We do not treat in this paper on the classification problem of double fibrations, cone-structures nor pseudo-product structures. Nevertheless we should mention that we naturally suppose that the exact classification result of Cartan distributions by Zhitomirskii [32] can be interpret in terms of the classification of double fibrations via Cartan prolongations which have canonical pseudo-product structures. Also we should mention that the notion of Cartan connections plays the central role in the classification problem of geometric structures. Morimoto [22] constructed the canonical Cartan connections associated with sub-Riemannian structures. Then we naturally ask for the construction of Cartan connections on sub-Riemannian pseudo-product structures.
All manifolds and mappings are assumed to be of class C ∞ unless otherwise stated. For an interval I, we say that an assertion holds for almost every t ∈ I if it holds outside of some measure zero set of I.

Control systems
Let M be a finite dimensional C ∞ manifold. A control system on M is given by a locally trivial fibration π U : U → M over M and a C ∞ mapping F : ). Then the control system is simply written as U If a local triviality U | V ∼ = V × U, V ⊂ M on M is given, then the control system is given by a family of vector fields {f u } u∈U , f u (z) := F (z, u), and the equationż(t) = f u(t) (z(t)).
− −− → M be a control system over a manifold M and π N : M → N a fibration. Suppose also the composition π N • π U : U → N is a fibration. Then we have a control system C/π N over N by We call the control system C/π N the quotient of C by π N .
Any distribution D ⊂ T M , namely, a vector subbundle of the tangent bundle T M , is regarded as a control system D : naturally via the fibration π = π T M | D and the inclusion F : D ֒→ T M .
Two control systems C : for some diffeomorphisms ψ and ϕ. The pair (ψ, ϕ) of diffeomorphisms is called an equivalence of the control systems C and C ′ . In this paper we suppose the fibre of π : U → M is an open subset of R r . Given a control system C : U (x(t)), is called a C-admissible control or simply an admissible control if the differential equatioṅ is satisfied. Here we suppose that x(t) = π U • c(t) is absolutely continuous so that the above equation has a meaning. Then the curve π U • c : [a, b] → M turns to be a Lipschitz curve. We call the curve π U • c the C-trajectory associated to the admissible control c. We use the term "path" for a smooth (C ∞ ) immersive trajectory regarded up to parametrisation as in [15].
If the fibration π U : U → M is trivial with a fibre U , namely if π U : U = M × U → M is the projection to the first factor, then the trajectory x(t) associated to an admissible control c : [a, b] → M × U , c(t) = (x(t), u(t)) is uniquely determined by u(t) and the initial point x(a) ∈ M .
A control system C on a manifold M is called controllable (resp. locally controllable at z 0 ∈ M ) if any points z 0 , z 1 ∈ M are joined by a C-trajectory. Lemma 2.1 For the quotient C/π N of a control system C over a manifold M by a fibration π N : M → N , we have (1) If z : I → M is a C-trajectory, then π N • z : I → N is a C/π N -trajectory.
(2) If C is controllable, then C/π N is controllable.
(2) If C is locally controllable at a point z 0 ∈ M , then C/π N is locally controllable at π N (z 0 ) ∈ N .
(3) : Let V ⊂ N be any open neighbourhood of x 0 = π N (z 0 ). Since C is locally controllable at a point z 0 , there exists an open neighbourhood W ′ ⊂ π −1 N (V ) of z 0 such that any points in W ′ are connected by a C| π −1 Then the totality C of admissible controls c : [a, b] → U with a given initial point z 0 ∈ M , namely, with the condition π U • c(a) = z 0 form a Banach manifold. The endpoint mapping End : C → M is defined by The control system C is controllable if the endpoint mapping is surjective, for any z 0 . We are concerning with the infinitesimal behaviour of the endpoint mapping. An admissible control c : [a, b] → U with the initial point π U (c(a)) = z 0 is called a singular control or an abnormal control, if c ∈ C is a singular point of End, namely if the differential End * : T c C → T End(c) M is not surjective ( [7]). If c is a singular control, then the trajectory π U • c is called a singular trajectory, singular curve or an abnormal extremal.
A control system C : is the zero map. Then the trivial system is not controllable at all, provided M = pt. Given a control system C : that is the fibre product of π U : U → M and π T * M : Note that π U ′ (ψ(z, u)) = ϕ(x). Moreover note that there exists the commutative diagram: We define the Hamiltonian function We will write H(z, p, u) instead of H((z, u), (z, p)) for brevity. If (ψ, ϕ) is an equivalence of control systems C and C ′ , then we have In fact, An admissible control c : I → U , c(t) = (z(t), u(t)) of the control system C is an abnormal control or a singular control if and only if there exists an absolutely continuous curve β : I → U × M T * M , β(t) = ((z(t), u(t)), (z(t), p(t))) satisfying, for any local triviality, the following constrained Hamiltonian system Note that abnormal (singular) controls are defined as "singular points" of the end-point mapping which is one of principal objects to study in control theory. The above characterisation for them is a consequence of the calculation on the differential of the endpoint mapping (see for example [3] pp.306-307).
A curve β : I → U × M T * M satisfying the above constrained Hamiltonian system is called an abnormal bi-extremal and its projection z : I → M to M is called an abnormal extremal or a singular trajectory.
We call a C ∞ immersed abnormal extremal considered up to C ∞ parametrisations an abnormal geodesic or a singular path for the control system C.
For an equivalence (ψ, ϕ) of control systems C and C ′ , the abnormal C-geodesics are mapped to the abnormal C ′ -geodesics by the diffeomorphism ϕ. The class of abnormal geodesics is one of important invariants on the equivalence classes of control systems.

Cartan distributions
Let Y be a 5-dimensional manifold and D ⊂ T Y a distribution of rank 2. Then D is called a Cartan distribution if it has growth (2, 3, 5), namely, if rank(D (2) ) = 3 and rank(D (3) ) = 5, where [10]). Here D means the sheaf of sections to D and [ , ] means the Lie bracket.
Let D ⊂ T Y be a Cartan distribution. Consider the control system Then it is known that, for any point y of Y and for any direction ℓ ⊂ D y , there exists an abnormal D-geodesic (an abnormal geodesic or a singular path for D) through y with the given direction ℓ ([20] [21]). Thus the abnormal D-geodesics form another five dimensional manifold X.
be the space of tangential lines in D. Then we have dim(Z) = 6. Then Z is naturally foliated by the liftings of abnormal D-geodesics, and we have locally double fibrations: Note that, for each y ∈ Y , we have π −1 Y (y) = P (D y ), and, for each x ∈ X, π −1 X (x) is the locus in Z of the lifted abnormal D-geodesics corresponding to x.
Then it is known that E is a distribution of rank 2 with growth (2, 3, 4, 5, 6). If we put L = Ker(π Y * ), K = Ker(π X * ), then we have a decomposition E = L ⊕ K by integrable subbundles. Note that the decomposition is intrinsically obtained from the distribution D ⊂ T Y .
A cone field C ⊂ T X is defined by setting, for each x ∈ X, Thus, so far, we have distributions D, E, L, K and the cone field C from the double fibrations: We regard the cone field C ⊂ T X as a control system over X: Then we have the following result: Theorem 4.1 (Duality [15]) Abnormal geodesics, namely singular paths, of the control system C are given by π X -images of π Y -fibres. Therefore, for any x ∈ X and for any direction ℓ ⊂ C x , there exists uniquely an abnormal geodesic for the control system C passing through x with the direction ℓ. The original space Y is identified with the space of singular paths on the space X, the space of singular paths on Y .
Note that abnormal geodesics (singular paths) of D : D ֒→ T Y → Y are given by π Y -images of π X -fibres. Remark 4.2 A different kind of duality was studied in [17] between a contact structure on a 3-dimensional space and an indefinite conformal structure of signature (2, 1) on another 3dimensional space via an Engel structure. Nevertheless, the duality via singular geodesics does not hold in that case, since contact structures do not have any abnormal geodesic, while the cone structure in that case actually has abnormal geodesics induced by the Engel line field.
We are naturally led to suppose that there exists a kind of symmetry on π Y -fibres and π Xfibres. However the double fibration Y π Y ← − − Z π X − − → X is asymmetric in the following sense. An abnormal extremal x(t) for E ⊂ T Z is called regular abnormal if it is associated with an abnormal bi-extremal (x(t), p(t), u(t)) such that p(t) ∈ E (2)⊥ \ E (3)⊥ ⊂ T * Z. An abnormal extremal x(t) for E ⊂ T Z is called totally irregular abnormal if any associated abnormal biextremal satisfies that p(t) ∈ E (3)⊥ ⊂ T * Z. Then we have Theorem 4.3 (Asymmetry [15]) An abnormal extremal, i.e. a singular trajectory for E ֒→ T Z → Z is either a (parametrisation of) π Y -fibre or a (parametrisation of) π X -fibre. Each π Y -fibre is regular abnormal, while each π X -fibre is totally irregular abnormal.
The main idea of the proof of Theorem 4.1 is show that any abnormal geodesic of the cone structure regarded as a control system on X, lifts to an abnormal geodesic of via π X : Z → X. Then we apply the first half of Theorem 4.3 to show that the lifted abnormal geodesic is either a π X -fibre or a π Y -fibre and then we conclude that the lifted abnormal geodesic must be a π Y -fibre.
We are going to present some of the results in [15] in the following sections under a wider framework.

Pseudo-product structures and control systems
Recall that a pseudo-product structure on a manifold M is a distribution E ⊂ T M on a manifold M with a decomposition E = L ⊕ K into integrable subbundles. Then locally we have a double fibration P to the leaf space P of L and N of K respectively. Then K = Ker(π N * ), L = Ker(π P * ). Suppose π N • π E : E → N, π N • π L : L → N, π P • π E : E → P, π P • π K : K → P are all fibrations. Then we are led to consider the following five control systems: which is a control system over M , Note that control systems L/π P and K/π N are reduced to be trivial.
Proof : By Chow-Rashevskii theorem [21], E is locally controllable at z 0 . Then, by Lemma 2.1, E/π N (resp. E/π P ) is locally controllable at π N (z 0 ) (resp. at π P (z 0 )). To show the local controllability of L/π N at π N (z 0 ), take a sufficiently small neighbourhood V ⊂ N of π N (z 0 ) such that π N • π E : E → N and π N : M → V are trivial over V respectively. Moreover take a sufficiently small neighbourhood V ′ ⊂ M of z 0 such that π N : V ′ → V is a fibration and L, K are trivial over V ′ . Then there is an open neighbourhood W ⊂ V of π N (z 0 ) such that any pair of points x 1 , x 2 ∈ W is joined by an E/π N -trajectory x : [a, b] → V . We take an associated . Then solve the differential equation d dt z(t) = ℓ(t) on curves z : [a, b] → V ′ with an initial point z(a). Then d dt (π N • z)(t) = π N * (ℓ(t)) =ẋ(t). Then, by the uniqueness of the solution, we have π N • z = x. Therefore x is an (L/π N )| V -trajectory connecting x 1 and x 2 . Hence L/π N is locally controllable at π N (z 0 ). Similarly we have that K/π P is locally controllable at π P (z 0 ).  (x 1 , u 1 ). We express simply by the pair (C, e) the optimal control problem. We call e the cost function of the optimal control problem.
Consider two control systems and two cost functions e : U → R, e ′ : U ′ → R respectively. Then two optimal control problems (C, e) and (C ′ , e ′ ) are called equivalent if the diagram commutes for some diffeomorphisms ψ and ϕ. The pair (ψ, ϕ) of diffeomorphisms is called an equivalence of the optimal control problems (C, e) and (C ′ , e ′ ).
The Hamiltonian function Here U × N T * N means the fibre product of π U : U → N and π T * N : T * N → N , namely, and H C is the Hamiltonian function of the control system C as in §3.
By the Pontryagin maximum principle, any solution (which is called a optimal control, a minimal control or minimiser) c(t) = (x(t), u(t)) of the optimal control problem is associated with a bi-extremal (x(t), p(t), u(t), p 0 ) which is a solution of the constrained Hamiltonian system,

See [24][3].
A bi-extremal is called normal (resp. abnormal) if p 0 < 0 (resp. p 0 = 0). A curve x : I → N (resp. a control c = (x, u) : I → U ) is called a normal extremal (resp. an abnormal extremal) if it possesses a normal bi-extremal (resp. an abnormal bi-extremal) lift (x(t), p(t), u(t), p 0 ). If a normal extremal (resp. an abnormal extremal) is regarded up to C ∞ parametrisations, then it is called a normal geodesic (resp. an abnormal geodesic). Therefore we do not mean by a "geodesic" an optimum but just a "stationary point" of the cost functional in general.
For an equivalence (ψ, ϕ) of optimal control problems (C, e) and (C ′ , e ′ ), the C-geodesics are mapped to the C ′ -geodesics by the diffeomorphism ϕ. Thus class of normal geodesics is a natural invariant on the equivalence of optimal control problems. However it is difficult to describe all normal geodesics in general. Example 6.1 ("generalised" sub-Riemannian geodesics). Let X 1 , . . . , X r be vector fields over a manifold N . Consider the control systeṁ Here π : U = N × R r → N is given by π(x, u) = x, and F : U → T N is given by F (x, u) = r i=1 u i X i (x). Moreover consider the optimal control problem to minimise the energy functional It is known that the problem is equivalent to minimising the length: If X 1 , . . . , X r are linearly independent everywhere, then the problem is exactly to minimise the Carnot-Carathéodory distances in sub-Riemannian geometry ([21] [6]).
The Hamiltonian function of the optimal control problem is given by Then the constraint ∂H ∂u = 0 is equivalent to that For a normal extremal, we have p 0 < 0. Then we have Then From the linearity of Hamiltonian function on (p, p 0 ), we can normalise p 0 to −1, so that Thus our formulation coincides with the ordinary normal Hamiltonian formalism in sub-Riemannian geometry ( [30][20] [21]).
For an abnormal extremal, the equation reduces to

Sub-Riemannian pseudo-product structures
Let E = L ⊕ K ⊂ T M be a pseudo-product structure on M . Then we have associated with it, locally, a double fibration P Moreover we have associated five control systems to the pseudo-product structures in §5. We are going to treat several sub-Riemannian optimal control problems related to pseudo-product structures. Given Riemann metrics g L and g K on the integrable subbundles L and K respectively, we give the product metric g E of g L and g K on E: Example 7.1 Let (E, g) be a sub-Riemannian structure on a distribution E ⊂ T M and L ⊂ E be an integrable subbundle of E. Then we set K = L ⊥ in E. If K is integrable, then E = L ⊕ K and we have a sub-Riemannian pseudo-product structure. In particular, let E be a sub-Riemannian Engel distribution and L ⊂ E be the Engel line bundle ( [21]). Set K := L ⊥ ⊂ E. Then E = K ⊕L is a sub-Riemannian pseudo-product structure, K, L being of rank one (integrable).
We study the sub-Riemannian optimal control problems on the three control systems Optimal control problems on E/π P and L/π P are studied similarly.
For the control system E, the Hamiltonian H is given by and a = (a 1 , . . . , a ℓ ) (resp. b = (b 1 , . . . , b k )) are the fibre coordinates of L (resp. K).
We set as the sub-Riemannian energy functions on (L, g L ), (K, g K ) respectively and as the total sub-Riemannian energy function on (E, g E ).
Then the Hamiltonian H of the sub-Riemannian optimal control problem (E, e E ) is given by with variables x, w, p, ψ and a non-positive parameter p 0 . The constrained Hamiltonian system is given bẏ with (p(t), ψ(t), p 0 ) = 0.
Consider the case p 0 = 0, i.e. the normal case. From the linearity of the equation on (p, ψ, p 0 ), we can normalise p 0 = −1: If the equation is solved by (p(t), ψ(t), p 0 ), then it is solved by ( 1 |p 0 | p(t), 1 |p 0 | ψ(t), −1). Then we have The optimal control problem (E/π, e E ) over N is described as follows: The system of local coordinates of N (resp. T * N , E) is given by x = (x 1 , . . . , x n ), (resp. (x; p) = (x; p 1 , . . . , p n ), (x, w; a, b) = ( x, w; a 1 , . . . , a ℓ , b 1 , . . . , b k )). Then the system of local coordinates of E × N T * N is given by (x, w; p; a, b). Note that dim(E × N T * N ) = 2n + 2k + ℓ. In this case, the mapping F : E → T N for the control system E/π is locally given by and the Hamiltonian function of the control system E/π and that for the optimal control problem (E/π, e E ) are given respectively by Here the control parameters are given by w, a, b. Note that H E/π is independent of b, while H (E/π,e E ) does depend on b. The constrained Hamiltonian system for extremals is given bẏ After normalising p 0 = −1, the constraints implies that Suppose π X * : L \ {0} → T X is an embedding. Then a normal extremal of (E/π, e E ) lifts to a normal extremal of (E, e E ) which is contained in a π P -fibre, by setting ψ = 0. Then the lifted geodesic for (E, e E ) contained in a π P -fibre is a geodesic for the metric g L . Thus we have Proposition 7.2 Suppose π X * : L \ {0} → T X is an embedding. Then the normal geodesics for the optimal control problem (E/π N , e E ) are obtained as π N -images of Riemannian geodesics for the Riemannian metric g L in L-leaves.
Similarly we have Proposition 7.3 Suppose π Y * : K \ {0} → T Y is an embedding. Then the normal geodesics for the optimal control problem (E/π P , e E ) are obtained as π P -images of Riemannian geodesics for the Riemannian metric g K in K-leaves.
Next we consider the optimal control problem (E/π N , e L ). The Hamiltonian function is given by Then we have the following constrained Hamiltonian system: After normalisation p 0 = −1, we have a i = 1≤j≤n c ij p j .
We consider the optimal control problem (L/π N , e L ). The Hamiltonian function is given by Then we have the following constrained Hamiltonian system: We observe there there is no relation between normal geodesics for (E, e E ) and those for (E/π N , e L ) (resp. (L/π N , e L ), because the latter problem has no information on g K . Now consider abnormal geodesics (p 0 = 0). Then we have: Proof of Lemma 7.4: Let β(t) = (x(t), w(t); p(t); a(t), b(t)) be an abnormal bi-extremal for E/π N . Note that b(t) can be taken as arbitrary L ∞ function. Suppose β(t) is Lipschitz. Then in particular the control w(t) is Lipschitz. Replace µ(t) byẇ(t), which is of class L ∞ , and take ψ(t) = 0. We set β(t) = (x(t), w(t); p(t), 0; a(t),ẇ(t)). Then β is an abnormal bi-extremal for E and γ(t) = (x(t), w(t)) is a lift of γ(t) = x(t), π • γ = γ. ✷ The projection π L : E = L ⊕ K → L induces the projection ρ : E × N T * N → L × N T * N . Then we have Lemma 7.5 ([15]) A curve β : I → E × N T * N is an abnormal bi-extremal (resp. a Lipschitz abnormal bi-extremal) for E/π if and only if ρ • β : I → L × N T * N is an abnormal bi-extremal (resp. a Lipschitz abnormal bi-extremal) for L. Moreover any abnormal bi-extremal (resp. a Lipschitz abnormal bi-extremal) β : I → L × N T * N for L is written as ρ • β by an abnormal bi-extremal (resp. a Lipschitz abnormal bi-extremal) β : I → E × N T * N for E/π. Proof : The system of local coordinates of L × N T * N is given (x, w; p; a). The constrained Hamiltonian system for L is of the same form for E/π if the in-efficient component b is deleted.
We have that β(t) = (x(t), w(t); p(t); a(t), bt)) is an abnormal bi-extremal (resp. a Lipschitz abnormal bi-extremal) for E/π if and only if ρ • β(t) = (x(t), w(t); p(t); a(t)) is an abnormal biextremal (resp. a Lipschitz abnormal bi-extremal) for L. Therefore we have the first assertion. The second assertion is clear. ✷ These Lemmata are applied also to the systems E/π P and K/π P . Using these general results, we have shown our main Theorem 4.1 in [15].

Sub-Riemannian Cartan distributions
We apply the general theory presented in the previous section to the case of Cartan distributions.
Let D ⊂ T Y be a Cartan distribution and g D be a Riemannian metric on D. Then we have a sub-Riemannian Cartan distribution (D, g D ). Let E ⊂ T Z be the Cartan prolongation of D. Then we have the intrinsic pseudo-product structure E = L ⊕ K, where L and K are of rank one ( §4). Moreover there are induced natural metrics g L on L and g K on K. In fact, for each z ∈ Z, π Y * : T z Z → T π Y (z) Y induces the injective linear map K z → D z . Therefore the metric on D z induces a metric on K z . Thus we obtain the metric g K from g D . The metric g L is obtained from the following general construction.
Let V be a finite dimensional metric vector space of positive definite, namely, a Hilbert space with dim(V ) < ∞. Then we naturally define a Riemannian metric on the projective space P V as follows. Let the unit sphere, which is a submanifold of codimension one. Then we have P V = SV / ∼ by identifying the pair of antipodal points. Let π : SV → P V denote the double covering. Take [x] ∈ P V and u, v ∈ T [x] P V . Define for u = π * ( u), v = π * ( v), u, v ∈ T x SV . Then the metric g P V is well-defined because holds.
In our case, the fibre (π Y ) −1 (y) of π Y : Z → Y over y ∈ Y is identified with P (D y ). Therefore (π Y ) −1 (y) possesses the natural Riemannian metric, and, for each point z ∈ Z, L z = T z ((π Y ) −1 (π Y (z))) has the natural metric. We endow E with the product metric g E of g L and g K .
Let η 1 , η 2 be an orthonormal frame of (D, g D ). Take a system of local coordinates y 1 , y 2 , y 3 , y 4 , y 5 on Y . Then we have a system of local coordinates y 1 , y 2 , y 3 , y 4 , y 5 , v on Z such that π Y (y, v) = y, v is the arc-length parameter of the fibres of π Y . Then ξ = ∂ ∂v is a unit frame of L, η = (cos v)η 1 + (sin v)η 2 + ρ ∂ ∂v is a unit frame of K, for some function ρ on Z, and then ξ, η form the orthonormal frame of E. Write The Hamiltonian function of the optimal control problem (E, e E ) is given by H (E,e E ) (y, v; q, ϕ; λ, µ; q 0 ) = λϕ + µ Moreover we have H (K/π Y ,e K ) (y, v; q; µ; q 0 ) = H (E/π Y ,e K ) (y, v; q; λ, µ; q 0 ) and the constrained Hamiltonian equation for (K/π Y , e K ) which is of the same form that for (E/π Y , e K ).