Geometric and Numeric Techniques in Optimal

Orbital Transfer using Low Propulsion

B. Bonnard, J.-B. Caillau, G. Picot
October 5, 2009

Dedicated to John Baillieul on the occasion of his 65th birthday

Abstract The objective of this article is to present the geometric and

numeric techniques developed to study the obit transfert between Keplerian
elliptic orbits in the two-body problem or between quasi-Keplerian orbits in the
Earth-Moon transfer when low-propulsion is used. We concentrate our study
to the energy minimization problem and from Pontryagin Maximum Principle
the optimal solution can be found solving a shooting equation for a smooth
Hamiltonian dynamics. A rst step in the analysis is to nd in the Kepler case
an analytical solution for the averaged Hamiltonian, which corresponds to a
Riemannian metric. This will allow to compute the solution for the original
Kepler problem, using a numerical continuation method, where the smoothness
of the path is related to the conjugate point condition. Similarly, the solution of
the Earth-Moon transfer is computed using a numerical continuation technique.

1 Introduction
In this article we consider the orbit transfer in the two and three bodies problem,
using low propulsion. In the rst case, the model is given by Kepler equation
q u
q̈ = − 3 + (1)
|q| m
when m represents the mass of the satellite, subjected to ṁ = −δu and the
control satises the constraint |u| ≤ ² where ² is a small parameter.
The physical optimal control amounts to maximise the nal mass and this leads
to
Z tf
Minu(.) |u|dt
0
where tf is the transfer time. From the Pontryagin's Maximum Principle [18],
xing the boundary conditions, e.g a transfer from a low eccentric to a geo-
stationary orbit, an optimal solution can be numerically computed, using a
shooting algorithm. This leads to a complicated numeric problem. In [14], the
following numerical scheme is proposed : compute the optimal solution using
the convex homotopy.
Z tf
Minu(.) (λ|u|2 + (1 − λ)|u|)dt,λ ∈ [0, 1]
0

1
wich amounts to regularize the L1 -minimization problem into a L2 -problem.
This was the starting point of the use of continuation methods in orbital trans-
fer, when low propulsion is applied, see also [9] for the use of the continuation
method in the time minimal control problem, where the homotopy parameter
is the bound of the maximal amplitude of the trust.

If we examine this approach from the mathematical point of view, this

amounts to replace the original Hamiltonian dynamics associated to the optimal
ow by another Hamiltonian dynamics and making a continuation to solve the
shooting equation.

The rst motivation of this article is to present a neat geometric result

from [7] : neglecting the main variation, restricting to coplanar transfer (the
inclination being considered as an homotopy parameter) and replacing the L1 -
problem by an averaged L2 -problem one can replace the Hamiltonian vector
eld dened by the maximum principle by

1 5 − 4e2 2
H= 5 [18n2 p2n + 5(1 − e2 )p2e + pθ ] (2)
8n 3 e2
where n is the mean motion, e is the eccentricity and θ the angle of the pericenter
where the singularity e = 0 corresponds to circular orbits. Moreover (n, e, θ)
are orthogonal coordinates for the Riemannian metric associated to H :
5 5
dn2 2n 3 2n 3
g= 1 + 2
de2 + dθ2 .
9n 3 5(1 − e ) 5 − 4e2

Such a metric is isometric to

r2
g = dr2 + (dφ2 + G(φ)2 dθ2 )
c2
where :
2 5
r= n 6 , φ = arcsine
5
and
r
2 5 sin2 (φ)
c= , G(Φ) =
5 1 + 4 cos2 (φ)

2 5
r= n 6 , φ = arcsine
5
→
−
Such an Hamiltonian ow H is Liouville integrable and the metric is the above
normal form captures the main properties of the averaged orbital transfer. In-
deed, one can extract from g the following 2D-Riemannian metrics :

• g1 = dr2 +r2 dψ 2 which is associated to the orbital transfer where θ is kept

xed, that is towards circular orbits. Such a metric is at and geodesics
are straight-lines in suitables coordinates.

2
 • g2 = dφ2 + G(φ)dθ2 wich represents the restrition of the metric to r2 = c2
and describes by homogeneity the orbit transfer, in the general case.
A generalization of the results of [21] will allow for the metric g2 to compute
the conjugate and cut loci and to get a global optimality solution for the aver-
aged optimal control problem. This is the starting point to analyse the original
optimal control problem, using a continuation method.

A second motivation of this article is to present some results from geometric

control theory connected to our analysis with adapted numerical codes devel-
oped to compute the solutions. First of all, the maximum principle is only a
necessary optimality condition. In order to get necessary sucient optimality
conditions ( under generic assumptions) one must dene the concept of con-
jugate point, associated to the energy minimization point. This concept was
already introduced in the standard litterature of calculus of variations [2]. If
the Hamiltonian optimal dynamics is described by a smoth Hamiltonian vector
→
−
eld H , conjugate points are the image of the singularities of the exponential
−
→
mapping : expq(0) : p(0) −→ Π(exp tf H (q(0), p(0))) where Π : (q, p) −→ q is
the standard projection. Such point can be numerically computed using the
Cotcot code [11]. An important remark, in view of the use of the (smooth)
continuation method in optimal control is to observe that the shooting equation
is precisely to nd p(0) such that expq(0) (p(0)) = q1 where q1 is the terminal
condition and the derivative can be generated using the variational equation of
→
−
H . This will lead to convergence results for the smooth continuation method
in optimal control, related to estimates of conjugate points

The third section is devoted to the Earth-Moon transfer, using low-propulsion.

The model is the standard circular restricted model [15] where the two primaries
are xed in a rotating frame. Up to a normalization the system can be written
in the hamiltonian form
−→ −→ −→
ż = H0 (z) + u1 H1 (z) + u2 H2 (z)
−→
where z = (q, p) ∈ R4 and the drift H0 is given by
1−µ µ
H(z) = 12 (p21 + p22 ) + p1 q2 − p2 q1 − %1 − %2 ,

q being the position of the spacecraft, %1 represents the distance to the earth
with mass 1 − µ located at (−µ, 0) and %2 the distance to the moon with mass
µ and located at (1 − µ, 0), while µ ' 0.012153 is a small parameter. The
Hamiltonian elds associated to the control are given by :

Hi (z) = −qi , for i=1,2.

and the control bound is |u| ≤ ². The parameter µ is small and this remark was
−→
used by Poincaré to study the dynamics of the free motion described by H0 , by
making a deformation of the case µ = 0 which corresponds to Kepler equation in
rotating coordinates [16]. Inspired by this approach, and using our preliminary
geometric analysis, we propose a simply solution to the Earth-Moon transfer,
using low propulsion, for the energy minimization problem.

3
2 Geometric and numeric methods
2.1 Maximum principle
R tf 2
We consider the energy minimization problem
Pm Minu(.) 0 |u| dt, for a smooth
control system of the form ẋ = F0 (x(t))+ i=1 ui (t)Fi (x(t)) = F (x(t), u(t)), x ∈
X , the set of admissible controls is the subset U of L∞ [0, tf ] of measurable
mappings u(.) with corresponding trajectory q(.) dened on the whole [0, tf ].
The Pontryagin maximum principle [18] tells us the following
Proposition 2.1. If (x, u) is an optimal pair on [0, tf ], then there exists an
absolutely continuous adjoint vector p ∈ Tx∗ X such that on [0, tf ] we have
 ∂H
 ẋ = ∂p (x, p, u)
(3)

ṗ = − ∂H
∂x (x, p, u)

and

H(x, p, u) = Maxv∈Rm H(x, p, v)

P
where H(x, p, u) = p0 m
i=1 ui + < p, F (x, u) >, p ≤ 0.
2 0

Denition 2.1. The mapping H from Tx∗ X × Rm to R is called the pseudo-

Hamiltonian. A triple (x, p, u) solution of (1) and (2) is called an extremal
trajectory.

2.2 Computation of the extremal

∂H
From the maximization condition (2), one deduces ∂v = 0 and we have two
types of extremals.

• Abnormal case They correspond to the situation p0 = 0 and they are

implicitely dened by the relations Hi = 0, i = 1 . . . m where Hi =<
p, F i(x) > is the Hamiltonian lift.

• Normal case If p0 < 0, the by homogeneity it can be normalized to − 12 .

From ∂H∂v = 0, one deduces ui = Hi for i = 1 . . . m and plugging
P such Hi
into H it denes a true smooth Hamiltonian Hn = H0 + 12 Hi2 where
solutions are the normal extremals.

2.3 The concept of conjugate point

Denition 2.2. Let z = (x, p) be a normal reference extremal dened on [0, tf ].
The variational equation
→
−
δ̇z(t) = d H (z(t))δz(t)

is called the Jacobi equation. A Jacobi eld is a non-trivial solution δz =

(δx, δp). It is said to be vertical if δx(t) = dΠ(z(t)).δz(t) = 0 where Π is the
projection (x, p) −→ x.

4
 The following standard geometric result is crucial.
−→
Proposition 2.2. Let L0 be the ber Tx∗0 X and Lt = expt(Hn )(L0 ) be the
image by the one-parameter subgroup generated by Hn . Then Lt is a Lagrangian
manifold whose tangent space at z(t) is generated by the Jacobi elds wich are
vertical at t = 0.
Denition 2.3. We x x0 = x(0) and we dene for t ∈ [0, tf ] the exponential
mapping

expx0 ,t (p(0)) = Π(z(t, z(0)))

where z(t, z(0)), with z(0) = (x(0), p(0)), denotes the normal extremal starting
at t = 0 from z(0).

Denition 2.4. Let z=(x,p) be the reference normal extremal. The time tc ∈
[0, tf ] is called conjugate if the mapping expx0 ,tc is not an immersion at p(0).
The associated point x(t) is said to be conjugate to x0 . We denote by t1,c the
rst conjugate point and C(x0 ) the conjugate locus formed by the set of rst
conjugate points, at time tf , when we consider all the normal extremal starting
from x0 .

Let us note that the conjugate time notion admit the following generaliza-
tion.

Denition 2.5. Let M1 be a regular submanifold of M and denote M1⊥ =

{(x, p) x ∈ M1 , p ∈ Tx M1 }. So tf oc ∈ [0, tf ] is called focal time if there exists a
Jacobi eld J = (δx, δp) such that δx(tf oc ) = 0 and J(tf oc ) is tangent to M1⊥ .

Remark . The concept of conjugate point is related to the necessary and su-
cient optimality conditions, under generic assumptions, see for instance [6].

2.4 Conjugate points and the smooth continuation method

in optimal control
The smooth continuation method is a general numerical method to solve a
system of equations F (x) = 0 where F : Rn −→ Rn is a smooth mapping,
see [1]. The principle is to construct an homotopy path H(x, λ) such that
H(x, 0) = G(x) and H(x, 1) = F (x) where G(x) is a map having known zeros
or where the zeros can be easily numerically computed using a Newton type al-
gorithm. The zeros along the path can be computed with dierent algorithms,
the simplest being a discretization on 0=λ0 < λ1 < · · · < λn =1 of the homotopy
parameter where at step i + 1 the zero computed at step i is used to initialize
the Newton algorithm.

It is a quite general approach which has to be adapted to optimal control

problems : the shooting equation comes from the projection of a symplectic
mapping, the Jacobian can be computed using Jacobi elds and one must con-
sider the central extremal elds associated to the problem, see [8] for the details
of the geometric concepts. A short description of the method is given below in
our case study.

5
2.4.1 Shooting equation
One considers a family Hλ , λ ∈ [0, 1] of smooth Hamiltonians on T ∗ X associated
to normal extremals of an energy minimization problem and we x the boundary
conditions x0 , x1 and the transfer times. This leads to the family expλx0 ,tf (p0 )
of exponential mappings. Using the notation E λ : p0 −→ expλx0 ,tf (p0 ). One
must solve the shooting equation E λ (p0 ) = x1 . The following result follows
immediately :
Proposition 2.3. For each λ, the shooting equation is of maximal rank if and
only if the point x1 is not conjugate to x0 for the corresponding λ. Moreover, in
this case, the solutions of the shooting equation contain a smooth branch, wich
can be parametrized by λ and the derivative E 0λ can be generated integrating the
Jacobi equation.
From the above proposition, to ensure the convergence of the method one
must control :
• The distance to the conjugate loci.
• The branch has to be dened on the whole interval [0,1].
This second point is related to two standard problems in optimal control : exis-
tence of Lipschitzian minimizers and hence solutions of the maximum principle,
see [22] and compactness of the domain of the exponential mapping.

Next we present a nice geometric situation for which convergence of the

method is ensured. First of all recall the geometric concept :
−→
Denition 2.6. Consider the normal extremal eld Hn associated to an energy
minimization problem with xed transfer time tf . Fixing the initial condition to
x0 , the separating locus L(x0 ) is the set of separating points where two distincts
normal extremals curves intersect with the same cost. The cut point along a
normal extremal is the rst point where it ceases to be optimal and the cut locus
Cut(x0 ) is the set of such points when we consider all the extremals, initiating
from x0 and optimality is lost exactly at time tf .

2.4.2 Convergence of the continuation method in the Riemannian

case
First of all, recall that the Riemannian problem can be at least locally reset in
the following framework.

Let F1 , . . . , Fn be a set of n-smooth vector elds on a manifold X and assume

that they are linearly independant. One can dene a Riemannian metric on
X by setting that {F1 , . . . , Fn } form an orthonormal frame. Introducing the
Pn
control system dx(t) dt = i=1 ui (t)Fi (x(t)), le length of the curve x(t), t ∈ [0, T ]
R T Pn 2 1
is l(x) = 0 i=1 (ui (t)) . From Maupertuis principle, minimizing the length
2
R T Pn 2
is equivalent to minimize the energy 0 i=1 ui (t). There exists only normal
Pn
extremals and Hn is given by 2 i=1 Hi . Fixing the level set Hn = 12 amounts
1 2

to parametrize by arc-length and for the energy minimization problem, the

transfer time can be arbitrarly xed e.g to 1. In this framework we have the
following

6
Theorem 2.1. Let gλ , λ ∈ [0, 1] be a smooth family of complete Riemannian
metrics on M . Let us x the initial point x0 . Denote iλ (x0 ) the distance from
x0 to the cut locus Cut
λ
(x0 ) and by iλ = infx(0) iλ (x0 ) the injectivity radius of
the corresponding metric. Then :
• For length shorter than infλ iλ (x0 ), the continuation method with initial
condition of the shooting equation at x0 converges.
• For length shorter than infλ iλ the continuation method converges for every
initial condition for the shooting equation.
Remark : in the Riemannian case, the situation is neat ; completness leads to
existence of smooth normal minimizers, the domain of the exponential mapping
is the sphere and estimates of the injectivity radius are related to the curva-
ture concept. In the general case such estimates are a dicult problem and a
pragmatic point of view is to have numeric estimates. This leads to the next
section.

Numeric methods
They are implemented in the Cotcot (Conditions of Order Two, COnjugate
Times) whose aim is to provide the numerical tool :
−→
• to integrate the smooth Hamiltonian vector eld Hn
• to solve the associated shooting equation
• to compute the corresponding Jacobi elds along the extremals
• to evaluate the resulting conjugate points
The code is written in Fortran language, while automatic dierentiation is used
to generate the Hamiltonian dierential equation and the variational one. For
the users the language is Matlab, see [6] for a precise description of the Cotcot
code, with the underlying algorithms : ODE integrators, Newton method solver.
The conjugate point test consists in checking a rank condition wich is based on
two methods : evaluating zeros of a determinant or a SVD ( Singular Value
Decomposition). The continuation method will be included in the code which
can be a simple discretization of the homotopy path, with a Newton method at
each step or a numeric continuation of the path, using the implicit ODE, where
the Newton method is used only at the rst step to initialize the continuation
and at the nal step, to solve the nal shooting equation with arbitrary accuracy.

3 The energy minimization problem in orbital

transfer with low thrust
3.1 Preliminaries
Neglecting the mass variation and restricting to the coplanar case, the system
is represented in Cartesian coordinates by
q
q̈ = − +u
|q|3

7
where q = (q1 , q2 ) is the position and the state of the system is x = (q, q̇) ∈ R4 .
1
We denote by H0 (q, q̇) = 12 q̇ 2 − |q| the Hamiltonian of the free motion. We have
the following rst integrals :
• C = q ∧ q̇ (momentum)
q
• L = − |q| + q̇ ∧ C (Laplace integral)
We have the following.
Proposition 3.1. The domain Σe = {(q, q̇); H < 0, C 6= 0} called the elliptic
domain is lled by elliptic orbits and to each orbit (C, L) corresponds an unique
(oriented) ellipse
To represent the space of ellipses, one introduce the following geometric
coordinates :
• θ : angle representing the argument of the pericenter
• e : eccentricity. The eccentricity vector being −
→
e = (ex , ey ), ex = e cos(θ)
ey = e sin(θ).
• a : semi-major axis of the ellipse related to the semi-latus rectum by the
P
relation a = √1−e 2
.

To represent the position of the satellite we use the longitude l ∈ S1 , while l ∈ R

takes into account the rotation number and is called the cumulated longitude.
Observe that e = 0 corresponds to circular orbits and the boundary e = 1
corresponds to parabolic orbits. The control u can be decomposed into moving
frames atached to the satellite, the two-standard such frames being :
q ∂
• The radial/ orthoradial frame {Fr , For } where Fr = |q| ∂ q̇

q̇ ∂
• The tangential/normal frame {Ft , Fn } where Ft = |q̇| ∂ q̇

3.2 Riemannian Metric of the Averaged Control Kepler

Equation
Preliminaries Let X be a n-dimensional smooth manifold and let Fi (x, l), i =
1, · · · , n be smooth vector elds parameterized by l ∈ S 1 . We consider the
control system
dx
Pm
dt = i=1 ui (t)Fi (x, l))
dl
dt = g0 (x, l)
where g0 is a smooth 2π -periodic function with respect to l and g0 > 0 and the
minimum energy problem
Z T X m
min ( u2i (t)dt) .
u(·) 0 i=1

The control is rescaled with u = εv to introduce the small parameter ε and the
trajectories parameterized by l are solutions of
Pm
dx vi Fi (x, l)
= ε i=1
dl g0 (x, l)

8
whenever the cost takes the form
Z l(T ) X
m
vi2 (l)
ε2 dl .
l(0) i=1 g0 (x, l)

We assume that l(0) and l(T ) are xed. The cost extended system takes the
form Pn Pm 2
dx i=1 vi Fi (x, l) d c v (l)Fi (x, l)
=ε , = ε i=1 i
dl g0 (x, l) dt ε g0 (x, l)
and we rescale c into c̄ = εc .
The associated pseudo-Hamiltonian is
Xm
ε
H̃(x, p, l, v) = (p0 |v|2 + vi Hi (x, p, l))
g0 (x, l) i=1

where Hi (x, p, l) = hp, Fi (x, l)i, i = 1, · · · , n are the 2π -periodic Hamiltonian

lifts and p0 ≤ 0 is a constant. We consider the normal case p0 < 0, which can
be normalized to p0 = − 12 and the Hamiltonian takes the form
m
ε 1 2 X
H̃(x, p, l, v) = (− |v| + vi Hi (x, p, l)) .
g0 (x, l) 2 i=1

Since v is valued in the whole Rm , the maximum principle gives ∂∂v

H̃
= 0 and we
get vi = Hi . Plugging such v into H̃ , we obtain the true Hamiltonian
X m
1
Hn (x, p, l) = H 2 (x, p, l)
2g0 (x, l) i=1 i

where ε is omitted to simplify the notations. We observe that since g0 is positive,

H can be written as a sum of squares.
Lemma 3.1. The function Hn is a non-negative quadratic form in p which is
denoted w(x, l).
Denition 3.1. The averaged Hamiltonian is
Z 2π
1
H(x, p) = H(x, p, l)dl .
2 0

The following result is clear.

Lemma 3.2. The averaged Hamiltonian denes a non-negative quadratic form
in p denoted w̄(x). Moreover

ker w̄(x) = ∩l∈S 1 ker w(x, l) .

Remark . According to this lemma, the rank of w̄(x) is not smaller than m if
the Fi0 s are m independent vector eld and we can only expect it to increase.
The geometric interpretation is straightforward. From the maximum principle,
an extremal control is computed as mapping of the form u(x, p, l) which is 2π -
periodic with respect to l. Hence, the corresponding trajectory is solution of the

9
averaged dierential equation. The oscillations induced by l which act as a fast
variable generate new control directions, namely Lie brackets in the linear space
E1 (t) = Span{adk F0 · Fi ; k ≥ 0, i = 1, · · · , m}. Moreover, generically we can
expect to generate all the Lie brackets in E1 (t) to provide an averaged system
of full rank.
Denition 3.2. The averaged system is said to be regular if the rank of w̄(x)
is a constant k .
In this case, there exists an orthogonal matrix R(x) such that if P = R(x)p
then w̄(x) is written as a sum of squares
k
1X
λi (x)Pi2
2 i=1

where λ1 , · · · , λk are the non-negative eigenvalues of the symmetric matrix S(x)

dened by
1
w̄(x) = t pS(x)p .
2
Hence, we can write
k k
1X p 1X
w̄(x) = ( λi Pi )2 = hp, Fi i2
2 i=1 2 i=1

where the Fi 's are smooth vector elds on X . We deduce the following propo-
sition.
Proposition 3.2. If the averaged system is regular of rank k, the averaged
Hamiltonian H can be written as a sum of squares
k
1X 2
H= H , Hi = hp, Fi i
2 i=1 i

and is the Hamiltonian of the SR-problem

k
X Z T k
X
ẋ = ui Fi (x), min u2i (t)dt
u(·) 0
i=1 i=1

where k is not smaller than n. If k = n = dim X then H is the Hamiltonian of

a Riemannian problem.
For this new optimal control problem, the extremal controls are not related
to the previous ones, but still the true extremal control u(x, p, l) can be approx-
imated by u(x̄, p̄, l), where x̄ and p̄ are the averaged values. Moreover, if we
apply standard approximation arguments, we deduce:
Proposition 3.3. The extremals of the averaged Hamiltonian systems are ap-
proximations of the true extremal trajectories of order o(ε) for a length of order
o(1/ε) and the cost of the SR-problem is an approximation of the true cost up
to order o(ε2 ).
A direct consequence is the following:

10
Corollary 3.1. If tF is the transfer time lF − l0 then tF ε → M constant as
ε → 0.
Proof. If we consider the SR-problem, it is equivalent
Pk to a time-minimal control
problem where the controls satisfy the bounds i=1 u2i (t) = 1 which amounts
to x the level set of the Hamiltonian to 1/2. By homogeneity, rescaling u into
εu, rescales the transfer time from t to t/ε. Therefore, the relation tF ε → M is
exact in the SR-case and the same is true for the original problem asymptotically
as ε → 0 since it is approximated by the SR-problem.

3.3 Computation of the Averaged System in Coplanar Or-

bital Transfer
Preliminaries To simplify the computation, the control is decomposed in the
radial-orthoradial frame. Applying the previous process, the true Hamiltonian
in the normal case is Hn = 21 (H12 + H22 ) where

P 5/4
H1 = W (pex sin l − pey cos l)
P 5/4 2P ex +cos l ey +sin l
H2 = W [pp W + pex (cos l + W ) + pey (sin l + W )]

with W = 1 + ex cos l + ey sin l.

The computation of the averaged system requires evaluations of integrals of
the form Z 2π
Q(cos l sin l)
dl
0 Wk
where Q is a polynomial and k is an integer between 2 and 4. Such integrals
are computed by means of the residue theorem. Using the complex notation
~e = ex + iey , the poles are
√
−1 ± 1 − e2
z± =
ē
and only z+ belongs to the unit disk.
An inspection of the Hamiltonian shows that the following averages are re-
quired.
√
Lemma 3.3. With δ = 1/ 1 − e2
• 1/W 2 = δ 3
• cos l/W 3 = −(3/2)ex δ 5 , sin l/W 3 = −(3/2)ey δ 5

• cos2 l/W 3 = 1/2(δ 3 + 3e2x δ 5 ), sin2 l/W 3 = 1/2(δ 3 + 3e2y δ 5 )

• cos l sin l/W 3 = 3/2ex ey δ 5

• 1/W 4 = 1/2(2 + 3e2 )δ 7

• cos l/W 4 = (−1/2)ex (4 + |e|2 )δ 7 , sin l/W 4 = (−1/2)ey (4 + |e|2 )δ 7

• cos2 l/W 4 = 1/2(δ 5 + 5e2x δ 7 ), sin2 l/W 4 = 1/2(δ 5 + 5e2y δ 7 )

• cos l sin l/W 4 = 5/2ex ey δ 7

11
 Substituting these expressions, we obtain the averaged Hamiltonian.
Proposition 3.4. We have
P 5/2
H(x, p) = [4p2 P 2 (−3 + 5(1 − e2 )−1 ) +
4(1 − e2 )5/2 p
p2ex (5(1 − e2 ) + e2y ) + p2ey (5(1 − e2 ) + e2x ) − 20Pp p2ex − 20Pp p2ey − 2pex pey ex ey ] .

3.4 The Analysis of the Averaged System

At this point, to identify the metric, H has to be written as a sum of squares.
More precisely, we make the following change of variables
1 − e2
P = , ex = e cos θ, ey = e sin θ
n2/3
where n is the so-called mean motion related to the semi-major axis by n =
a−3/2 . Such a transformation is singular for circular orbits. On the Hamiltonian,
this amounts to the Mathieu transformation: x = φ(y), p = q ∂φ ∂y where q is the
new adjoint variable. In the new coordinates, we have:
Proposition 3.5. In the coordinates (n, e, θ), the averaged Hamiltonian is
1 2 2 2 2 5 − 4e2 2
H= [18n pn + 5(1 − e )pe + pθ ]
8n5/3 e2
where the singularity e = 0 corresponds to circular orbits. In particular, (n, e, θ)
are orthogonal coordinates for the Riemannian metric associated to H
dn2 2n5/3 de2 2n5/3 dθ2
g= + + .
9n1/3 5(1 − e2 ) 5 − 4e2
The main step in the analysis is to use further normalization to obtain a
geometric interpretation.
Proposition 3.6. In the elliptic domain, we set
2 5/6
r= n , φ = arcsin e
5
and the metric is isometric to
r2
g = dr2 + (dφ2 + G(φ)dθ2 )
c2
p 5 sin2 φ
with c = 2/5 and G(φ) = 1+4 cos2 φ .

Geometric interpretation This normal form captures the main properties

of the averaged orbital transfer. Indeed, we extract from g two 2D-Riemannian
metric
g1 = dr2 + r2 dψ 2
with ψ = φ/c which is associated to orbital transfer where θ is kept xed and
the metric
g2 = dφ2 + G(φ)dθ2
which represents the restriction to r2 = c2 .
We next make a complete analysis of g1 and g2 .

12
 L1 1 n L3

2 3

L2 ex

Figure 1: Geodesics of the metric g2 in (n, ex ) and at coordinates

3.4.1 Analysis of g1
θ is a cyclic coordinate and pθ a rst integral. If pθ = 0 then θ is constant. The
corresponding extremals are geodesics of the 2D-Riemannian problem dened
by dθ = 0. We extend the elliptic domain to

Σ0 = {n > 0, e ∈] − 1, +1[, e = ex , ey }

and in polar coordinates (r, ψ), Σ0 is dened by {r > 0, ψ ∈] − π/2c, π/2c[}.

This extension allows to go through the singularity corresponding to circular
orbits.
Geometrically, this describes transfer where the angle of the semi-major axis
is kept xed and pθ = 0 corresponds to the transversality condition. Such a
policy is clearly associated of steering the system towards circular orbits where
the angle θ of the pericenter is not prescribed. An important physical subcase
is when the nal orbit is geostationary.
In particular in the domain Σ0 , the metric ḡ1 = dr2 +r2 dψ 2 is a polar metric
isometric to the at metric dx2 + dz 2 if we set x = r sin ψ and z = r cos ψ .
We deduce the following proposition.

Proposition 3.7. The extremals of the averaged coplanar transfer are straight
lines in the domain Σ0 in suitable coordinates, namely

23/2 5/6 1 23/2 5/6 1

x= n sin( arcsin e), z = n cos( arcsin e)
5 c 5 c
p
with c = 2/5. Since c < 1, the domain is not convex and the metric g1 is not
complete.
Proof. The extremals are represented in Figs. 1 in the physical coordinates
(n, ex ) (ey is xed to 0) and in the at coordinates.
The axis ex = 0 corresponds to circular orbits. Among the extremals, we
have two types: complete curves of type 1 and non-complete curves of type 2
when meeting the boundary of the domain. The domain is not geodesically con-
vex and in the domain II, the existence theorem fails. For each initial condition,
there exists a separatrix S which corresponds to a segment line in the orbital
coordinates which is meeting n = 0 in nite time. Its length gives the bound
for a sphere to be compact.

13
 In order to complete the analysis of g and to understand the role of g2 , we
present now the integration algorithm.

3.4.2 Integrability of the Extremal Flow

The integrability property is a consequence of the normal form only

g = dr2 + r2 (dφ2 + G(φ)dθ2 )

and the associated Hamiltonian is decomposed into

1 2 1 1 p2
H= pr + 2 H 0 , H 0 = (p2φ + θ ) .
2 r 2 G(φ)
~ admits three rst integrals in
Lemma 3.4. The Hamiltonian vector eld H
involution: H , H 0 and pθ and is Liouville integrable.
To get a complete parameterization, we proceed as follows. We use the
(e, n, θ) coordinates and we write
1
H= [18n2 p2n + H 00 ]
4n5/3
5−4e2 2
with H 00 = 5(1 − e2 )p2e + e2 pθ .

Lemma 3.5. Let s = n5/3 then s(t) is a polynomial of degree 2: s(t) = c1 t2 +

25
ṡ(0)t + s(0) with s(0) = n5/3 (0), ṡ(0) = 15n(0)pn (0) and c1 = 2 H.

Lemma 3.6. Let dT = dt/4n5/3 then if H 00 (0) 6= 0, T (t) = 2√1|∆| [arctan L(s)]t0
where L(t) = 2at+b
√ , a = c1 , b = ṡ(0) and ∆ = − 25
2 H (0) is the discriminant of
00
|∆|
s(t).
This allows to make the integration. Indeed if H 00 = 0, pe = pθ = 0 and
the trajectories are straight lines (the line S in Figs. 1). Otherwise, we observe
that n5/3 (t) is known and depends only upon n(0), pn (0) and H which can be
xed to 1/2 by parameterizing by arc-length. Hence, it is sucient to integrate
the ow associated to H 00 using the parameter dT = 4ndt5/3 where T is given by
the previous lemma.
We proceed as follows. Let H 00 = c23 and pθ = c2 . Using pe = ė/10(1 − e2 ),
we obtain
20(1 − e2 )
ė2 = [c3 e2 − (5 − 4e2 )c22 ] .
e2
To integrate, we set for e ∈]0, 1[, w = 1 − e2 and the equation takes the form
dw
= Q(w)
dT
where
Q(w) = 80w[(c23 − c22 ) − (c23 + 4c22 )w]
with positive discriminant. Hence the solution is
1 c23 − c22 √ q
w= [1 + sin(4/ 5 c23 + 4c22 )T + K] ,
2 c23 + 4c22

14
K being a constant. We deduce that
Z T
1 + 4w(s)
θ(T ) = θ(0) + 2c2 ds
0 1 − w(s)
where θ(0) can be set to 0 by symmetry. To conclude, we must compute
R T 1+4w(s) p
0 1−w(s)
ds with w = K1 (1 + sin x) and x = √45 c23 + 4c22 s + K . Therefore,
we must evaluate an integral of the form
Z
A + B sin x
dx
C + D sin x
which is a standard exercise. More precisely, the formula is
Z Z
A + B sin x B dx
dx = x + AD − BC
C + D sin x D C + D sin x
with Z
dx 2 C tan(x/2) + D
=√ arctan( √ )
C + D sin x C 2 − D2 C 2 − D2
for C 2 − D2 > 0 in our case. The previous lemmas and computations give:
Proposition 3.8. For H 00 6= 0, the solutions of H ~ can be computed using
elementary functions and we get
n(t) = ( 25 2
2 Ht + 15n(0pn (0)t + n
5/3
(0))3/5
1
e(t) = (1 − 2 K1 (1 + sin K2 (t)))1/2
K (t)
θ(t) = θ(0) + 2|p pθ
θ|
K3 [−4x + K 10
3
arctan( (1−K1 ) tan(x/2)−K
K3 )]K2
2
1/2−p2 p
with K = arcsin( 1−e(0) K1
−1
), K1 = 21 1/2+4pθ2 , K2 (t) = √45 1/2 + 4p2θ (T (t) + K)
r θ

5p2θ
and K3 = 1/2+4p 2 . For H
00
= 0, they are straight lines.
θ

3.4.3 Geometric Properties of g2

The previous integration algorithm shows that the extremals of this metric
describe the evolution of the angular variables θ and φ, parameterized by dT =
dt 2
r(t)2 where r(t) is a second order polynomial whose coecients depend only
upon the energy level H xed to 1/2, r(0) and pr (0). We next describe some
basic properties of g2 .
Lemma 3.7. The metric g2 can be extended to an analytic metric on the whole
S 2 , where θ and φ are spherical coordinates with two polar singularities at φ =
0, π corresponding to e = 0 and the equator to e = 1; θ is an angle of revolution.
The meridians are projections on S 2 of the extremals of g1 .
Lemma 3.8. The metric is isometric for the two transformations (φ, θ) 7→
(φ, −θ) and (φ, θ) 7→ (π − φ, θ). This induces the following symmetries for the
extremal ow.
• If pθ 7→ −pθ then we have two extremals with the same length symmetric
with respect to the meridian.
• If pφ 7→ −pφ then we have two extremals of same length intersecting on
the antipodal parallel φ = π − φ(0).
Such properties are shown by the following one-parameter family of metrics.

15
 π−φ0

φ
φ0

0 θ θ

Figure 2: Action of the symmetry group on the extremals

Metrics induced by the at metric on oblate ellipsoid of revolution

We consider the at metric of R3 : g = dx2 + dy 2 + dz 2 restricted to the ellipsoid
dened by
x = sin φ cos θ, y = sin φ sin θ, z = µ cos φ
where µ ∈]0, 1[. A simple computation leads to

g2 = Eµ (φ)dφ2 + sin2 φdθ2

where Eµ (φ) = µ2 + (1 − µ2 ) cos2 φ. Computing for g2 = dφ2 + G(φ)dθ2 ,

5 sin2 φ
G(φ) = 1+4 cos2 φ we can write

1
g2 = (Eµ (φ)dφ2 + sin2 φdθ2 )
Eµ (φ)
√
where µ = 1/ 5. We deduce the following lemma:

Lemma 3.9. The metric g2 is conformal to the at √ metric restricted to an

oblate ellipsoid of revolution with parameter µ = 1/ 5.

3.4.4 A Global Optimality Result with Application to Orbital Trans-

fer
In this section, we consider an analytic metric on R+ × S 2

g = dr2 + (dφ2 + G(φ)dθ2 ) (4)

and let H be the associated Hamiltonian. We x the parameterization to arc-

length by restricting to the level set H = 1/2. Let x1 , x2 be two extremal curves
starting from the same initial point x0 and intersecting at some positive t̄. We
get the relations

r1 (t̄) = r2 (t̄), φ1 (t̄) = φ2 (t̄), θ1 (t̄) = θ2 (t̄)

and from lemma 3.5, we deduce the following lemma.

Lemma 3.10. Both extremals x1 and x2 share the same pr (0) and for each t,
r1 (t) = r2 (t).

16
 0
If we consider now the integral curves of H 0 where H = 12 p2r + H
r 2 on the
dt
xed induced level and parameterizing these curves using dT = r2 , we deduce
the following characterization.

Proposition 3.9. The following conditions are necessary and sucient to char-
acterize extremals of H 00 6= 0 intersecting with the same length

φ1 (T̄ ) = φ2 (T̄ ), θ1 (T̄ ) = θ2 (T̄ )

with the compatibility condition

Z t̄
dt 2
T̄ = = [ √ arctan L(t)]t̄t=0 .
0 r2 (t) ∆
Theorem 3.1. A necessary global optimality condition for an analytic metric
on R+ × S 1 normalized to

g = dr2 + r2 (dφ2 + G(φ)dθ2 )

is that the injectivity radius be greater than or equal to π on the sphere r = 1,

the bound being reached by the at metric in spherical coordinates.
Proof. We observe that in the at case, the compatibility condition cannot be
satised. Moreover, the injectivity radius on S 2 is π corresponding to the half-
length of a great circle. Let us now complete the proof. For the analytic metric
on S 2 , the injectivity radius is the length of the conjugate point at minimum
distance of the half-length of a closed geodesic. The conjugate point is, in
addition, a limit point of the separating line. Hence, if the injectivity radius is
smaller than π , we have two minimizers for the restriction of the metric on S 2
which intersects with a length smaller than π . We shall show that it corresponds
to a projection of two extremals x1 and x2 which intersect with the same length.
For such extremals r(0) = 1, we set pr (0) = ε, H = 1/2 and we get

p2θ (0) p
2H 0 = p2φ (0) + = λ2 (ε), λ(ε) = 1 − ε2 .
G(φ(0))

If t1 is the injectivity radius on the level set H 0 = 1/2 which corresponds to

2
pr (0) = ε = 0. For H 0 = λ 2(ε) and pr (0) = ε, it is rescaled as T1 = t1 /λ(ε).
The compatibility relation for T̄ = T1 gives

t̄ + ε ε
T1 = arctan[ ] − arctan[ ].
λ(ε) λ(ε)

Clearly, the maximum of the right member is π , taking ε < 0, |ε| → 1. Hence,
it can be satised since t1 < π . The at case shows that it is the sharpest
bound.

By homogeneity, we deduce the following corollary.

r2
Corollary 3.2. If the metric is normalized to dr2 + c2 (dφ
2
+ G(φ)dθ2 ) then
the bound for the injectivity radius on r = c is cπ .2 2

17
3.4.5 Riemann Curvature and Injectivity Radius in Orbital Transfer
Using standard formulae from Riemannian geometry [11], we have the following
proposition.

Proposition 3.10. Let g be a smooth metric of the form dr2 +r2 (dφ2 +G(φ)dθ2 )
with x = (x1 , x2 , x3 ) = (r, θ, φ) the coordinates. Then the only non-zero compo-
nent of the Riemann tensor is

G00 (φ) G0 (φ)2

R2323 = r2 [− − G(φ) + ]
2 4G(φ)

which takes the form R2323 = −r2 F (F 00 + F ) if we set G(φ) = F 2 (φ). We have
therefore R2323 = 0 if and only if F (φ) = A sin(φ + φ0 ) which is induced by the
at case in spherical coordinates.
Hence, the main non-zero sectional curvature of the metric is
R2323
K= ∂ ∂ 2
| ∂θ∧ ∂φ |

and computing this term in the case of orbital transfer, we get:

Lemma 3.11. The sectional curvature in the plane (φ, θ) is given by

(1 − 24 cos2 φ − 16 cos4 φ)
KV =
r2 (1 + 4 cos2 φ)2

and KV → 0 as r → +∞.
Proposition 3.11. The Gauss curvature of the metric on S 2 , g2 = dφ2 +
5 sin2 φ
G(φ)dθ2 with G(φ) = 1+4 cos2 φ is

5(1 − 8 cos2 φ)
KV = .
(1 + 4 cos2 φ)2

Theorem 3.2. The Gauss curvature of g2 is negative near the poles and
√ max-
imum (constant equal to 5) at the equator. The injectivity radius is π/ 5 and
is reached by the shortest conjugate point along the equator.
Proof. Clearly K is maximum and constant equal to 5 along the equator which
is an extremal solution. Hence a direct computation √ gives that the shortest
conjugate point is along the equator with length π/ 5. It corresponds to the
injectivity
√ radius if the half-length of a shortest periodic extremal is greater
than π/ 5. Simple closed extremals are computed in [Bonnard,Caillau, Forum]
using the integrability property but a simple reasoning gives that the shortest
corresponds to meridians with length 2π . Hence the result is proved.
√ p
Corollary 3.3. Since π/ 5 < π 2/5, the necessary optimality condition of
theorem 3.2 is not satised in orbital transfer for the extension of the metric to
R+ × S 2 .

18
 π−φ0

φ
φ0

Figure 3: Conjugate and cut loci in averaged orbital transfer

3.4.6 Cut Locus on S 2 and Global Optimality Results in Orbital

Transfer
From the previous section, the computation of the injectivity radius for the
metric on S 2 is not sucient to conclude about global optimality. A more
complete analysis is necessary to evaluate the cut locus. This analysis requires
numerical simulations. The explicit analytic representation of the extremal ows
is given in [10]. The main results of this analysis are:
Proposition 3.12. For the metric g2 on S 2 , they are exactly ve simple closed
extremals modulo rotations around the poles, the shortest being
√ a meridian with
length 2π and the longest being the equator with length 2π 5.
Theorem 3.3. • 1) Except for poles, the conjugate locus is a deformation
of a standard astroid with axial symmetry and two cusps located on the
antipodal parallel.
• 2) Except for poles, the cut locus is a simple segment, located on the an-
tipodal parallel with axial symmetry and whose extremities are cusps points
of the conjugate locus.
• 3) For a pole, the cut locus is reduced to the antipodal pole.
Proof. The proof is made by direct analysis of the extremal curves. The main
problem is to prove that the separating line is given by points on the antipodal
parallel, where due to the isometry φ → π − φ, two extremals curves with same
length intersect. This property cannot occur before. The results are represented
in Fig 3.

Geometric interpretation and comments The metric is conformed to the

restriction of the at metric to an oblate ellipsoid of revolution. For such a
metric, the cut locus is known since Jacobi and is similar to the one represented
on Fig. 3. It is a remarkable property that there is no bifurcation of the cut locus
when the metric is deformed by the factor Eµ (φ) although the properties of the
metric are quite dierent. For instance, in orbital transfer, the Gauss curvature
is not positive. The mathematical proof requires a thorough analysis of the
extremal ow. A similar result can be obtained with numerical simulations.
Indeed on S 2 , relations between the conjugate and cut loci allow to deduce the

19
cut locus from the conjugate locus. Also a domain bounded by two intersecting
minimizing curves must contain a conjugate point.
In this case, the conjugate locus can be easily computed using the Cotcot
code. In such a situation, it can also be deduced by inspecting the extremal ow
only, the conjugate locus being an envelope. Finally, we observe that in order to
have intersecting minimizers, we must cross the equator φ = π for which e = 1.
The same is true for conjugate points. Hence we deduce:
Theorem 3.4. Conjugate loci and separating lines of the averaged Kepler met-
ric in the spaces of ellipses for which e ∈ [0, 1[ are always empty.

3.5 The Averaged System in the Tangential Case

An interesting question is to analyze if the averaged system in the tangential
case where the control is oriented along Ft conserves similar properties. The
rst step is to compute the corresponding averaged system.
Proposition 3.13. If the control is oriented along Ft only, the averaged Hamil-
tonian associated to the energy minimization is

1 2 2 4(1 − e2 )3/2 2 4(1 − e2 ) p2θ

Ht = [9n pn + √ pe + √ ]
2n5/3 1 + 1 − e2 1 + 1 − e2 e2
which corresponds to the Riemannian metric
√ √
dn2 n5/3 1 + 1 − e2 2 1 + 1 − e2 2 2
gt = 1/3 + ( de + e dθ )
9n 4 (1 − e2 )3/2 (1 − e2 )

where (n, e, θ) are orthogonal coordinates.

3.5.1 Construction of the Normal Form

We proceed as in Section 3.4. We set
2 5/6 p
r= n , e = sin φ 1 + cos2 φ .
5
The metric takes the form
r2 2
gt = dr2 + 2
(dφ2 + G(φ)dθ2 ), c2 = < 1
c 5
and
1 − (1/2) sin2 φ 2
G(φ) = sin2 φ( ) .
1 − sin2 φ
Hence the normal form is similar to the full control case. We introduce the
metrics
g1 = dr2 + r2 dψ 2 , ψ = φ/c
and
g2 = dφ2 + G(φ)dθ2 .
We next make the analysis by comparing with the full control case. The main
dierence will concern the singularities of G.

20
 π−φ
0

φ
φ0 φ0

Figure 4: Conjugate and cut loci in averaged orbital transfer (tangential case)

3.5.2 The Metric g1

The metric corresponds again to transfer to circular orbits and is the polar form
of the at metric dx2 + dz 2 , if x = r sin ψ and z = r cos ψ .

3.5.3 The Metric g2

The normal form reveals the same homogeneity property between the full control
and the tangential case, the metric g2 can be used to make a similar optimality
analysis, evaluating the conjugate and cut locus. But the metric g2 cannot be
interpreted as a smooth metric on S 2 . This can be seen by computing the Gauss
curvature.
Proposition 3.14. The Gauss curvature of g2 is given by
(3 + cos2 φ)(cos2 φ − 2)
K= .
(1 + cos2 φ) cos2 φ
In particular K → −∞ when φ → π/2 since K < 0 and the conjugate locus of
a point is empty.
Nevertheless, the extremals can be smoothly extended through the singular
boundary of the domain where φ = π/2 and we get a similar picture than for
the full transfer case represented in Fig. 4.

3.5.4 The Integration of the Extremal Flow

The algorithm based on the normal form is similar to the bi-input case, but we
compare the respective transcendence. The Hamiltonian is written as
1
H= [18n2 p2n + H 00 ]
4n5/3
where H 00 takes now the form
8(1 − e2 )3/2 2 8(1 − e2 ) p2θ
H 00 = √ pe + √ .
1 + 1 − e2 1 + 1 − e2 e2
√
2
We set H 00 = c23 , pθ = c2 and from pe = 4n5/3 e(1+ 1−e )
16(1−e2 )3/2
, we obtain with
√
w = 1−e 2
dw 2 Q(w)
( ) =
dT (1 + w)2

21
 3 3

2.5 2.5

2 2
φ

φ
1.5 1.5

1 1

0.5 0.5

0 0
0 0.5 1 1.5 2 2.5 3 0 0.5 1 1.5 2 2.5 3
θ θ

Figure 5: Extremal ow of g2 in the full control and tangential cases, in the
(φ, θ) coordinates, starting from φ = π/6

where T is as in the bi-input case and Q is the fourth-order polynomial

Q(w) = 32w[c23 (1 − w2 )(1 + w) − 8c22 w2 ] .

Hence, the integration requires the computation of the elliptic integral

Z
dw(1 + w)
p
Q(w)

which is an additional complexity.

3.5.5 Conclusion in Both Cases

The previous analysis shows that the full control case and the tangential one ad-
mit an uniform representation in the coordinates (Φ, θ). In particular, it allows
in such coordinates to make a continuation between the respective Hamiltoni-
ans, i.e., between the respective G(Φ). A correction has to be made between
orbit elements e which are respectively dened by

e = sin φ

and p
e = sin φ 1 + cos2 φ .
The ows in the two cases are presented on Fig. 5 and reveal the similar
structure and the conjuagte locus reached after having crossed the equator.

3.6 Preliminary Continuation Results

On Fig, we present a rst continuation result, in the tangential, case showing
the convergence of the method from the averaged to the non averaged trajectory.
(The boundary conditions are GTO toward GO orbits).

22
3.7 Conjugate and cut loci on a two-sphere of revolution
The problem of computing the conjugate and cut loci is connected to a very old
geometric problem wich goes back to Jacobi wich is briey introduced next.

Denition 3.3. The two-sphere S2 encowed with a smooth metric of the form
dφ2 + G2 (φ)dθ2 in spherical coordinates is called two-sphere of revolution.
Many of them can be realized as 2D-Riemannian surfaces of revolution em-
bedded in R3 by rotating on smooth curve homeomorphic to an half-circle.
The classical examples are the following :

• Round sphere S2 . It is constructed restricting the euclidian metric to S2

and given by dφ2 + sin2 (φ)dθ2 .

• Oblate ellipsoid of revolution O(µ). If we restrict the euclidian metric to

the surface :

x = sin φ cos θ, y = sin φ sin θ, z = µ cos φ

with µ < 1, the metric takes the form

(1 − (1 − µ2 ) sin2 φ)dφ2 + sin2 (φ)dθ2 .

wich can be set in the form dφ2 + G2 (φ)dθ2 using a quadrature.

Proposition 3.15. On an oblate ellipsoid of revolution we have :

1. The Gauss curvature is monotone increasing from the North pole to the
Equator.
2. The cut-locus of a point wich is not a pole is a sub of the antipodal parallel.
3. The conjugate locus of a point wich is not a pole is has a standard astroid
shape with four cusps.
The simple structure of the cut-locus is a consequence of [21].

Theorem 3.5. Let dφ2 + sin2 (φ)dθ2 be a metric on a two-sphere of revolution.

we assume :
1. The transformation φ −→ π − φ is an isometry i.e G(π − φ) = G(φ)
2. The Gauss curvature K is monotone non decreasing along a meridian
from the North to the equator.
Then the cut locus of a point not a pole is a simple branch located on the antiposal
parallel.

Application Let gλ be the family of analytic metrics on S2 dened by

(1 + λ)2 sin2 φ
gλ = dφ2 + G2λ (φ)dθ2 , G2λ (φ) = , λ≥0
(1 + λ cos2 φ)

23
The Gauss curvature is given by :

1 ∂2G (1 + λ)(1 − 2λ cos2 φ)

Kλ = − 2
=
G∂ φ (1 + λ cos2 φ)2

Here if 0 < λ ≤ 2, then Kλ is monotone non decreasing from the North pole
to the Equator and the previous theorem asserts that the one parameter has a
cutlocus reduced to a simple branch for λ ∈]0, 2].

If λ > 2 the Gaussian curvature Kλ is not monotone and the theorem can-
not be applied. In particular the orbit transfer with full control corresponds to
λ = 4, while at the limit case λ = +∞ a singularity appears : The Riemannian
metric has a pole at the equator , this situation being similar to the one occuring
when the thrust is only tangential. Hence the theorem 3.5 has to be rened to
deal with such situations. the nal result is coming from [5] .

We consider a metric of the form g = φ2 + G(φ)dθ2 where G is non zero on

π
]0,2[and G(π − φ) = G(φ). We introduce the following.

Denition 3.4. The rst return mapping to the equator is the map :
π
R : pθ ∈]0, G( )[−→ ∆θ(pθ )
2
where ∆θ is the θ-variation of the extremal parametrized by arc-length and
associated to the adjoint pθ . The extremal ow is called tame is the rst return
mapping is monotone non-increasing for pθ ∈]0, G( π2 )[.

The generalizing theorem is :

Theorem 3.6. In the tame case, the cut locus of a point dierent frame a
pole is a suset of the antipodal parallel. If moreover R0 (pθ ) < 0 < R00 (pθ ) on
]0, G( Π2 )[ then the conjugate locus of such a point has exactly four cusps.
Remark . This result can be extended to the singular case.

4 The energy minimization problem in the Earth-

Moon space mission with low thrust
4.1 Mathematical model and presentation of the problem.
In this section, we follow mainly [15], see also [17] and [23]

4.1.1 The N-bodies problem

Considr N point masses m1 , . . . , mN moving in a Galilean reference system R3
where the only forces acting being their mutual attraction. If q = (q1 , . . . , qN ) ∈
R3N is the state and p = (p1 , . . . , pN ) being the momentum vector, the equations
of the motion are :
∂H ∂H
q̇ = , ṗ = −
∂p ∂q

24
where the Hamiltonian is :
XN N
X
k pi k2 Gmi mj
H= − U , U (q) = .
i=1
2mi k qi − qj k
1≤i<j≤N

A restricting case being the coplanar situation where the N masses are in a plane
R2 . In this case the Galilean reference frame can be replaced by a rotating frame
dened by
µ ¶ µ ¶
0 1 cos ωt sin ωt
K= , exp(ωtK) = ,
−1 0 − sin ωt cos ωt

and introducing a set of coordinates wich uniformly rotates with frequency ω

dened the symplectic transformation :

ui = exp(ωtK)qi , vi = exp(ωtK).

and a standard computation gives the Hamiltonian of the N -body problem in

rotating coordinates :

XN N N
k v k2 X t X Gmi mj
H= − w ui Kvi −
i=1
2mi i=1
k qi − qj k
1≤i<j≤N

In particular, the Kepler problem in rotating coordinates up to a normalization

has the following Hamiltonian :
k p k2 t 1
H= − qKp − .
2 kqk

4.2 The circular restricted 3-body problem in Jacobi co-

ordinates
Recall the following representation of the Earth-Moon problem. In the rotating
frame, the Earth wich is the biggest primary planet with mass 1 − µ is located
at (−µ, 0) while the Moon with mass µ, is located at (1 − µ, 0) with the small
parameter µ ' 0.012153. We note z = x + iy the position of the spacecraft,
ρ1 , ρ2 are the distances to the primaries :
q¡ ¢
%1 = (x + µ)2 + y 2
q¡ ¢
%2 = (x − 1 + µ)2 + y 2

The equation of the motion takes the form :

z+µ z−1+µ
z̈ + 2iż − z = −(1 − µ) −µ
%31 %32
wich can be written:
∂V
ẍ − 2ẏ − x =
∂x
∂V
ÿ + 2ẋ − y =
∂y

25
 1−µ µ
where −V is the potential of the system dened by V = %31
+ %32
. The system
can be written using Hamiltonian formalism setting :
q1 = x, q2 = y , p1 = ẋ − y , q2 = ẏ + x,
and the Hamiltonian describing the motion takes the form
1 2 1−µ µ
H0 (q1 , q2 , p1 , p2 ) = (p + p22 ) + p1 q2 − p2 q1 − − ,
2 1 %1 %2

4.3 Jacobi Integral and Hill rgions

The Jacobi integral using Hamiltonian formalism is simply the Hamiltonian H0
wich gives :
ẋ2 + ẏ 2
H(x, y, ẋ − y, ẏ + x) = − Ω(x, y)
2
where
1 2 1−µ µ
Ω(x, y) = (x + y 2 ) + + .
2 %1 %2
Hence solutions are conned on the level set
ẋ2 + ẏ 2
− Ω(x, y) = h
2
where h is a constant.The Hill domain for the value h is the region where the
motion can occur, that is {(x, y) ∈ R2 , Ω(x, y) + h ≥ 0}

4.4 Equilibrium points

The equilibrium points of the problem are well known. They split in two dierent
types :
• Euler points: They are the collinear points denoted L1 ,L2 et L3 located
on the line y = 0 dened by the primairies. For the Earth-Moon problem
they are given by :
x1 ' −1.0051, x2 ' 0.8369, x3 ' 1.1557.

• Lagrange points : L4 et L5 form with the two primaries an equilateral

triangle.
Some important informations about the stability of the equilibrium points are
provided by the eigenvalues of the linearized system :
    
δ ẋ 0 0 1 0 δx
 δ ẏ   0 0 0 1   δy 
    
 δ ẍ  =  Ωxx (a, b) Ωxy (a, b) 0 2   δ ẋ 
δ ÿ Ωxy (a, b) Ωyy (a, b) −2 0 δ ẏ
The linearized matrix evaluated at points L1 ,L2 or L3 admits two real eigen-
values wich one is positive and two imaginary eigenvalues. The collinear points
are consequently non stable. When it is evaluated at L4 où√L5 , the linearized
matrix has two imaginary eigenvalues when µ < µ1 = 21 (1 − 969 ). So the points
L4 and L5 are stable since µ < µ1 .

26
4.5 The continuation method
The (mathematical) continuation method in the restricted circular problem is
omnipresent in Poincaré work, in particular for the continuation of circular or-
bits. Geometrically, it is simply a continuation of trajectories of Kepler problem
into trajectories of the 3-Body problem. It amounts to consider µ as a small
parameter, the limit case µ = 0 being Kepler-problem in the rotating frame,
writing :
k p k2 t 1
H0 = − qKp − + o(µ).
2 kqk
and the approximation for µ is valid, a neighborhood of the primaries being
excluded. In the Earth-Moon problem, since µ is very small, the Kepler-problem
is clearly a good approximation of the motion in a large neighborhood of the
Earth. This point of view is important in our analysis, as indicated by the
status report of the SMART-1 mission since most of the time mission is under
the inuence only of the Earth attraction, see [19] and [20].

4.5.1 The control problem

The control system in the rotating frame is deduced from the previous model
and can be written in the Hamiltonian form :
dz − → −
→ −
→
= H 0 (z) + u1 H 1 (z) + u2 H 2 (z)+
dt
−
→ −
→ − → →
−
where z = (q, p), H 0 is the free motion and H 1 , H 2 are given by H i =
−qi ,i =1,2. As for the Kepler problem, the mass variation of the satellite can be
introduced in the model dividing ui by m(t) and adding the equation ṁ = −δ|u|.
Again, it will be not taken into account in the problem. Moreover R t we still re-
strict out analysis to the energy minimization problem : Minu(.) t0f u21 + u22 dt,
where the transfer time tf is xed and the control valued in R2 . The physical
problem wich corresponds to maximize the nal mass can be analyses using a
(numeric) continuation method.

It is worth to point out that a lunar mission using low-propulsion called

SMART-1 was realized by ESA and the practical details of the mission, in par-
ticular the description of the trajectory, are reported in [19] and [20].Next we
present a trajetory analysis based on our geometric and numeric techniques.
For simplicity, we have xed the boundary conditions to circular orbits, the one
around the Earth corresponding to the stationary one. But everything can be
applied to other boundary conditions, like the GTO ellipse as the initial orbit,
as described in the report status of the mission SMART-1.

Our analysis is based on a numerical continuation, taking into account

second-order optimality condition, where µ is the parameter of the continu-
ation. The averaged system is nally applied to get an approximate energy
minimizing trajectory for the phase of the mission, starting from the circular
Earth orbit with the nal orbit is the quasi-elliptic orbit where the apogee is
about the maximum value of 338000 km. In the restricted problem, the eect
of the inclination which is about 30◦ in phase 2 of the mission is not taken into
account.

27
4.5.2 Boundary conditions and shooting equation the Earth-L2 trans-
fer
As a rst approach we choose to simulate the Earth-L2 transfer in the restricted
three-Body problem. Indeed, at the limit case µ = 0, the Moon and the point L2
are identical. Moreover, in the Earth-Moon system, the point L2 and the Moon
are located very closely. As a result, the rst phase of a Earth-Moon transfer is
comparable to a Earth-L2 transfer. Solving the shooting function associated to
the Earth-L2 transfer is consequently usefull to provide a good approximation
of the solution of the Earth-Moon transfer shooting function.

Denoting x = (q, q̇) ∈ R4 , the control system in the rotating frame associated
to µ ∈ [0, 21 ] is equivalent to

ẋ = F0 (x) + F1 (x)u1 + F2 (x)u2

where
 
x3
 x4 
 
F0 (x) =  x1 +µ x1 −1+µ
 2x4 + x1 − (1 − µ) ((x +µ)2 +x2 ) 32 − µ ((x −1+µ)2 +x2 ) 32


 1
x2
2 1
x2
2 
−2q3 + x2 − (1 − µ) 2 2
3 − µ
2 2
3
((x1 +µ) +x2 ) 2 ((x1 −1+µ) +x2 ) 2

   
0 0
 0   0 
F1 (x) =   
 1  , F2 (x) = 
.
0 
0 1

Using the circular orbit around the Earth to approximate the stationary one,
we set as initial condition x0 = (1 − µ + 0.1099, 0, 0, 2.8792). The L2 point is
located on (xL2 , 0) with xL2 solution of the equation :
(1 − µ)(x + µ) µ(x − 1 + µ)
x− − = 0.
%31 %32
Since we want to reach L2 with a speed of zero, we x the target xtf =
(xL2 , 0, 0, 0). The energy minimization problem with xed time transfer tf can
consequently be written


 ẋ = F0 (x) + F1 (x)u1 + F2 (x)u2 ,


 Rt
(Pµ ) Minu(.) t0f u21 + u22 dt





x(0) = x0 , x(tf ) = xf

The maximum principle is applied which asserts that in the normal case p0 < 0,
extremals of the problem are solutions of the Hamiltonian system
½ −
→
ż = H (z)
(5)
z(0) = (x0 , p0 )

28
 P2 P2
with H(x, p) = i=1 Hi2 (x, p) = i=1 (< p, Fi (x)> )2 . This leads to solve the
shooting equation S(p0 ) = 0 where S is the shooting equation dened by

S : R4 −→ R4
p0 −→ expx0 ,tf (p0 ) − xf

By making the parameter µ vary in [0,0.012153], one build up a family (Sµ )µ of

shooting functions which connects the Kepler and the three Body problem. The
numerical continuation method can be applied to deduce low-thrust extremal
trajectories of the Earth-L2 transfer from the Kepler ones.

4.5.3 The numerical continuation method for the Earth-L2 transfer

According to the rapport status of ESA, we xed the time transfer to 121 time
units of the restricted 3-body problem which approximatively corresponds to
the time transfert from the Earth to the point L2 while the SMART-1 mission
(about 17 months). In addition we considered a constant spacecraft mass of
350kg, see [19] and [20]. Setting µ = 0, we computed an extremal using the sim-
ple shooting method. Then we made the parameter µ vary from 0 to 0.012153
with an initial step of 10−3 that could be automatically reduced by the contin-
uation algorithm if necessary.

At each step, the rst conjugate time tc,1 along the extremal have been
computed to ensure the necessary condition of convergence of the continuation
method tc,1 > tf . The conjugate time test follows the next principle : since
the time transfer is xed, the domain of expx0 ,tc is locally dieomorphic to
R4 . One has to compute numerically the Jacobi elds Ji (t) = (δxi (t), δpi (t)),
i = 1,..,4 corresponding to the initial condition δxi (0) = 0 and δpi (0) = ei ,
where {ei , i = 1,..,4} represents the canonical basis of R4 and to compute rank
(δx1 (t), .., δx4 (t)) which is equal to 4 outside a conjugate and being lower than
or equal to 3 at a conjugate time.

Additionally, the euclidian norm of the extremal control have been plotted to
stand a comparison between the control bound and the maximal thrust allowed
by electro-ionic engines used while the SMART-1 mission.

Fig. 7 to Fig.14 present the computed spacececraft trajectories in the both

rotating and xed frames, as well as the rst conjugate time and the norm of
control along trajectories in the both Kepler case and Earth-Moon system.

29
 Figure 6: Earth-L2 trajectory in rotating frame, µ = 0.

0.8

0.6

0.4

0.2

−0.2

−0.4

−0.6

−0.8

−1 −0.5 0 0.5 1

Figure 7: Earth-L2 trajectory in xed frame, µ = 0.

30
 100

80

arcsh det(δ x) 60

40

20

0
0 100 200 300 400 500 600 700
t

10

6
σn

0
0 100 200 300 400 500 600 700
t

Figure 8: First conjugated time, Earth-L2 transfer, µ = 0.

Figure 9: Norm of extremal control, Earth-L2 transfer, µ = 0.

31
Figure 10: Earth-L2 trajectory in rotating frame, µ = 0.012153.

0.8

0.6

0.4

0.2

−0.2

−0.4

−0.6

−0.8

−1 −0.5 0 0.5 1

Figure 11: Earth-L2 trajectories in xed frame, µ = 0.012153.

32
 Figure 12: First conjugate time, Earth-L2 transfer, µ = 0.012153.

Figure 13: Norm of extremal control, Earth-L2 transfer, µ = 0.012153.

33
 The numerical continuation method considering µ as a small parameter led
to deduce an extremal trajectory of the energy minimization Earth-L2 trans-
fer problem from one corresponding to the Kepler case, in accordance with the
Poincaré Work. We checked that the rst conjugate time along each extremal
is higher than the time transfer, that was needed for the convergence of the
method. Moreover the second order optimality conditions ensure that the com-
puted extremals are locally energy minimizing in L∞ ([0, tf ]). We also note that
in both cases µ = 0 and µ = 0.012153, the maximum value reached by the
norm of the extremal control is highly inferior to the bound k u k≤ 0.08 which
corresponds to the maximal thrust of the Smart-1 electro-ionic engine (0.073
N). As one can see, we actually found a maximal bound of the norm of extremal
control twice lower than the one associated to Smart-1, the time transfer being
the same.

4.5.4 Boundary conditions and shooting function for the the Earth-
Moon transfer.
The second part of our trajectories analysis has been devoted to the Earth-Moon
transfer. We used the same dynamics and initial condition as previously. In this
case, the circular orbit around the Moon, denoted OL and is dened by :
 
 (x1 − 1 + µ)2 + x22 = 0.0017 
OL = (x1 , x2 , x3 , x4 ) ∈ R4 , x23 + x24 = 0.2946
 
< (x1 − 1 + µ, x2 ), (x3 , x4 ) >= 0.

is chosen as the target orbit. Setting

h : R4 −→ R3
 
(x1 − 1 + µ)2 + x22 − 0.0017
x −→  x23 + x24 = 0.2946 ,
< (x1 − 1 + µ, x2 ), (x3 , x4 ) >

the energy minimization problem with xed time transfer tf becomes



 ẋ = F0 (x) + F1 (x)u1 + F2 (x)u2 ,


0
 Rt
(Pµ ) Minu(.) t0f u21 + u22 dt





x(0) = x0 , h(x(tf )) = 0

In addition to the former necessary conditions, the maximum principle provides

the transversality condition p(tf ) ⊥ Tx(tf ) OL . The mapping h being a submer-
sion at each point of OL , it is equivalent to p(tf ) ∈ [Ker h0 (xf )]⊥ which is a
one-dimensionnal vector space.
Using the standard projections Πx and Πp and introducing wz(0) such that
[Ker h0 (Πx (zµ (tf , z0 )(p(0)))]⊥ = Span(wz(0) ), the shooting function becomes

S : R4 −→ R4
µ ¶
h ◦ Πx (z(tf , z(0))
p0 −→
h0 (Πx (z(tf , z(0))).Πx (z(tf , z(0))

34
4.5.5 Numerical continuation method for the Earth-Moon transfer.
The time transfer is xed to 124 time units of the restricted 3-body problem and
the spacecraft mass remains 350 kg, see [19] and [20]. The extremal trajectory
corresponding to µ = 0 was computed using the simple shooting method and
initializing p0 with the initial costate vector associated to the Earth-L2 transfer.
The step of the variation parameter µ is 10−3 .

Since the target is a manifold of codimension one, the concept of conju-

gate point is replaced by the concept of focal point, see the Def. 2.5 .At each
step, the rst focal time tf oc along generated extremal have been computed
to ensure the necessary condition of convergence of the continuation method
tf oc > tf . The focal time test follows the next principle : A retrograd integra-
tion is proceeding. Let us consider n-dimension vector space spanned by Jacobi
⊥
eld Ji (t) = (δxi , δpi ) for i = 1,..,4 and checking the condition Ji (0) ∈ Tz(tf ) OL .
At time t we compute the rank of (δx1 (−t), .., δx4 (−t)) which is lower than or
equal to 3 at focal time and 4 outside a focal time .

The Earth-Moon trajectories in both rotating and xed frames, the rst fo-
cal time and the norm of extremal control from Fig. 14 to Fig.21 for µ = 0 and
µ = 0012153.

Once again, we computed extremal trajectory of the energy minimization

Earth-Moon transfer thanks to the continuation method. In both cases µ = 0
and µ = 0.012153, the rst focal time along extremals tf oc,1 is higher than
3/2tf , ensuring local optimality. The maximal bound of the norm of extremal
control is 0.045, that approximatively corresponds to the half of the maximal
thrust allowed during the mission Smart-One.

It is interesting to notice that the Earth-L2 Keplerian trajectory greatly

diers from the Earth-Moon Keplerian trajectory. This dierence illustrates
the restricting role of the transversality condition provided by the maximum
principle when the target is a submanifold. On the contrary, for µ = 0.012153
the rst phase of the Earth-Moon transfer matches the Earth-L2 transfer. It
underlines the crucial role of the neighborhood of the point L2 which acts as
the way from Earth to Lunar attraction for µ 6= 0.

35
 0.8

0.6

0.4

0.2

−0.2

−0.4

−0.6

−0.8

−0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1 1.2

Figure 14: Earth-Moon trajectory in rotating frame, µ = 0.

0.8

0.6

0.4

0.2

−0.2

−0.4

−0.6

−0.8

−1 −0.5 0 0.5 1

Figure 15: Earth-Moon trajectory in xed frame, µ = 0.

36
 50

arcsh det(δ x)
0

−50
−200 −180 −160 −140 −120 −100 −80 −60 −40 −20 0
t

0.8

0.6
σn

0.4

0.2

0
−200 −180 −160 −140 −120 −100 −80 −60 −40 −20 0
t

Figure 16: First focal time and norm of extremal control, Earth-Moon transfer,
µ = 0.

Figure 17: Norm of extremal control, Earth-Moon transfer, µ = 0.

37
 0.6

0.4

0.2

−0.2

−0.4

−0.6

−0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1

Figure 18: Earth-Moon trajectory in rotating frame, µ = 0.012153.

0.8

0.6

0.4

0.2

−0.2

−0.4

−0.6

−0.8

−1 −0.5 0 0.5 1

Figure 19: Earth-Moon trajectory in xed frame, µ = 0.012153.

38
 50

arcsh det(δ x)
0

−50
−250 −200 −150 −100 −50 0
t

0.1

0.08

0.06
n
σ

0.04

0.02

0
−250 −200 −150 −100 −50 0
t

Figure 20: First focal time, Earth-Moon transfer, µ = 0.012153.

Figure 21: Norm of extremal control, Earth-Moon transfer, µ = 0.012153.

39
References
[1] Allgower, Eugene L.; Georg, Kurt. Numerical continuation methods. An in-
troduction. Springer Series in Computationnal Mathematics, 13. Springer-
Verlag, Berlin, 1990. xiv+388 pp. ISBN: 3-540-12760-7
[2] Bliss, Gilbert A. Lectures on the Calculus of Variations. University of
Chicago Press, Chicago, Ill., 1946. ix+296 pp.
[3] Bombrum, Alex; Chetboun, Jonathan; Pomet, Jean-Baptiste. Transfert
Terre-Lune en pousse faible par contrôle feedback-La mission SMART-1,
Rapport de recherche INRIA (2006), no.5955, 1-27.
[4] Bonnard, Bernard; Caillau, Jean-Baptiste. Riemannian metric of the aver-
aged energy minimization problem in orbital transfer with low thrust Ann.
Inst. H. Poincaré Anal. Non Linéaire 24 (2007), no. 3, 395-411.
[5] Bonnard, Bernard; Sinclair, Robert; Tanaka, Minoru. Conjugate and cut
loci of a two-sphere of revolution with application to optimal control Ann.
Inst. H. Poincaré Anal. Non Linéaire 26 (2009), no. 4, 1081-1098.
[6] Bonnard, Bernard; Caillau, Jean-Baptiste; Trélat, Emmanuel. Second order
optimality conditions in the smooth case and applications in optimal control
ESAIM Control Optim. and Calc. Var. 13 (2007), no. 2, 207-236.
[7] Bonnard, Bernard; Caillau, Jean-Baptiste; Dujol, Romain. Energy mini-
mization of single-input orbit transfer by averaging and continuation Bull.
Sci. Math. 130 (2006), no. 8, 707-719.
[8] Bonnard, Bernard; Shcherbakova, Natalia; Sugny, Dominique. The smooth
continuation method in optimal control with an application to control sys-
tems, submitted to COCV-ESAIM, 2009.
[9] Bonnard, Bernard; Caillau, Jean-Baptiste. Geodesic ow of the averaged
controlled Kepler equation. Forum math., vol 21, pp. 797-814, 2009.
[10] Caillau, Jean-Baptiste. Contribution à l'étude du contrôle en temps mini-
mal des transferts orbitaux PhD thesis, Toulouse Univ., 2000.
[11] Caillau, Jean-Baptiste. http://apo.enseeiht.fr/cotoct/
[12] do Carmo, Manfredo Perdigaão. Riemannian geometry. translated from
the second Portuguese edition by Francis Flaherty. Mathematics: Theory
& Applications, Birkhäuser Boston, Inc., Boston, MA, 1992. xiv+300pp.
[13] Guerra, Manuel; Sarychev, Andrey. Existence and Lipschitzian regularity
for relaxed minimizers. Mathematical control theory and nance, 231-250,
Springer, Berlin, 2008.
[14] Gergaud, Joseph; Haberkorn, Thomas. Homotopy method for minimum
consumption orbit transfer problem. ESAIM Control Optim. Calc. Var. 12
(2006), no. 2, 294310 (electronic).
[15] Meyer, Kenneth R.; Hall, Glen R. Introduction to Hamiltonian dynami-
cal systems and the N -body problem. Applied Mathematical Sciences, 90.
Springer-Verlag, New York, 1992. xii+292 pp.

40
[16] Poincaré, Henri. Oeuvres Complètes, Gauthier-Villars.

[17] Pollard, Harry. Mathematical introduction to celestial mechanics. Prentice-

Hall, Inc., Englewood Clis, N.J. 1966 x+111 pp.

[18] Pontryagin, L. S.; Boltyanskii, V. G.; Gamkrelidze, R. V.; Mishchenko, E.

F. The mathematical theory of optimal processes. Translated from the Rus-
sian by K. N. Trirogo; edited by L. W. Neustadt Interscience Publishers
John Wiley & Sons, Inc.\, New York-London 1962 viii+360 pp.

[19] Racca, Giuseppe D.; Foing, Bernard H.; Coradini, Marcello. SMART-1 :
the rst time of Europe to the Moon. Earth, Moon and planets 85-86 (2001)
379-390.

[20] Racca ,G.D & al. SMART-1 mission description and development status.
Planetary and space science 50 (2002), 1323-1337.
[21] Sinclair, Robert; Tanaka, Minoru The cut locus of a two-sphere of revolu-
tion and Toponogov's comparison theorem. Tohoku Math. J. (2) 59 (2007),
no. 3, 379399.

[22]

[23] Szebehely, Victor, Theory of orbits : the restricted problem of three bodies,

41