Optimal Control
Optimal Control
with
Aerospace Applications
by
James M. Longuski
Jose J. Guzman
John E. Prussing
10.1 Introduction
In this chapter we develop a general theory of optimal spacecraft trajectories based on
two pioneering works: Breakwell [1959] and Lawden [1963]. Lawden introduced the concept
of the primer vector, which plays a dominant role in minimum-propellant trajectories and
also in other types of optimal trajectories. A more complete discussion of the topics in this
chapter, including several example trajectories, is in Prussing [2010].
The variable Γ is the thrust acceleration vector, whose magnitude Γ is defined as the
thrust (force), T , divided by the mass of the vehicle, m. The variable u is a unit vector in
the thrust direction, and g(r) is the gravitational acceleration vector. A careful derivation
of Eq. (10.1) requires deriving the rocket equation by equating the net external force (such as
gravity) to the time rate of change of the linear momentum of the vehicle/exhaust particle
system (see sections 5.1–5.4 of Prussing and Conway [2013]).
An additional equation expresses the change in mass of the spacecraft due to the gener-
ation of thrust:
ṁ = −b, b ≥ 0 (10.2)
In Eq. (10.2) b is the (nonnegative) mass flow rate. The thrust magnitude, T , is given
by T = bc, where c is the effective exhaust velocity of the engine. The word “effective”
189
190 CHAPTER 10. GENERAL THEORY OF OPTIMAL ROCKET TRAJECTORIES
applies to high-thrust chemical engines where the exhaust gases may not be fully expanded
at the nozzle exit. In this case there exists an additional contribution to the thrust which is
incorporated into the effective exhaust velocity by defining
Ae
c ≡ ca + (pe − p∞ ) (10.3)
b
In Eq. (10.3) the subscript “e” refers to the pressure and area at the nozzle exit, ca is the
actual exhaust velocity at the exit, and p∞ is the ambient pressure. If the gases are exhausted
into the vacuum of space, p∞ = 0.
An alternative to specifying the effective exhaust velocity is to describe the engine in
terms of its specific impulse, defined to be:
(bc)∆t c
Isp ≡ = (10.4)
(b∆t)go go
where go is the standard acceleration of free fall on Earth, equal to 9.80665 m/s2 . As shown
in Eq. (10.4) the specific impulse is obtained by dividing the mechanical impulse delivered
to the vehicle by the weight of propellant consumed. The mechanical impulse provided by
the thrust force over a time ∆t is simply (bc)∆t and, in the absence of other forces acting
on the vehicle, is equal to the change in its linear momentum. The weight (on Earth) of
propellant consumed during that same time interval is (b∆t)go . We note that if instead
one divides by the mass of the propellant (which, of course, is the fundamental measure of
the amount of substance), the specific impulse would be identical to the exhaust velocity.
However, the definition in Eq. (10.4) is in standard use with the value typically expressed in
units of seconds.
drodynamic (MHD), plasma arc, and ion engines, and also for solar sails. The ratio for solar
sails is of the order of 10−5 . An electric engine requires a separate power generator such as
a radioisotope thermoelectric generator (RTG) or solar cells.
Other engine designs such as the magnetoplasmadynamic (MPD) thrusters and the vari-
able specific impulse magnetoplasma rocket (VASIMR) seek to achieve higher levels of thrust
while maintaining the high specific impulse typical of low-thrust engines.
10.4. COST FUNCTIONALS FOR ROCKET ENGINES 191
or
µ ¶ Z tf
mo
∆v = cu ln + g(r)dt (10.9)
mf to
which correctly indicates that, in the absence of gravity, the velocity change would be in the
thrust direction u. The actual velocity change achieved also depends on the gravitational
acceleration g(r) which is acting during the thrust period. The term in Eq. (10.9) involving
the gravitational acceleration g(r) is called the gravity loss. We note that there is no gravity
loss due to an (instantaneous) impulsive thrust, described in section 10.5.2.
If we ignore the gravity loss term for the time being, a cost functional representing
propellant consumed can be formulated. As will be seen, minimizing this cost functional
is equivalent to maximizing the final mass of the vehicle. Since the thrust is equal to the
product of the mass flow rate b and the exhaust velocity c, we can write:
−mΓ
ṁ = −b = (10.10)
c
192 CHAPTER 10. GENERAL THEORY OF OPTIMAL ROCKET TRAJECTORIES
dm Γ
= − dt (10.11)
m c
For the CSI case the exhaust velocity, c, is constant and Eq. (10.11) can be integrated to
yield
µ ¶ Z
mf 1 tf
ln =− Γdt (10.12)
mo c to
or
µ ¶ Z tf
mo
c ln = Γdt ≡ JCSI (10.13)
mf to
with to ≤ t1 < t2 ... < tn ≤ tf representing the times of the n thrust impulses. (See
sections 6.1–6.3 of Prussing and Conway [2013].) Using the definition of a unit impulse,
Z t+
k
δ(t − tk ) dt = 1 (10.15)
t−
k
where t±
k ≡ lim²→0 (tk ± ²), ² > 0.
Using Eq. (10.14) in Eq. (10.13) we obtain:
Z tf n
X
JCSI = Γ dt = ∆vk (10.16)
to k=1
and the total propellant cost is given by the sum of the magnitudes of the velocity changes.
The corresponding cost functional for the VSI case is obtained differently. The exhaust
power (stream or beam power) is half of the product of the thrust and the exhaust velocity:
1 1 1
P = T c = mΓc = bc2 (10.17)
2 2 2
Using Eq. (10.17) along with
µ ¶
b −ṁ d 1
2
= 2 = (10.18)
m m dt m
10.5. FIRST-ORDER NECESSARY CONDITIONS 193
results in
µ ¶
d 1 Γ2
= (10.19)
dt m 2P
which integrates to
Z tf
1 1 1 Γ2
− = dt (10.20)
mf mo 2 to P
Maximizing mf for a given value of mo regardless of whether it is optimal or not is
obtained by running the engine at maximum power P = Pmax . This conclusion is not as
obvious as it looks in Eq. (10.20), because the value of Γ might be different for different
values of P . To see that the engine should be run at maximum power we note that for a
specified trajectory, r(t), the required vector thrust acceleration is given by Eq. (10.1) as
Thus, for a given trajectory r(t) (optimal or not), the final mass in Eq. (10.20) is maximized
by running the engine at maximum power.
For this reason the VSI cost functional can be taken to be
Z
1 tf 2
JVSI = Γ dt (10.22)
2 to
This form for JVSI is also derived in Marec [1979].
To summarize, the cost functionals representing minimum-propellant expenditure are
given by
Z tf
JCSI = Γdt (10.23)
to
and
Z tf
1
JVSI = Γ2 dt (10.24)
2 to
We see from Eqs. (10.23) and (10.24) that the minimum-propellant cost can be written
in terms of the control magnitude Γ(t) rather than introducing the mass as an additional
state variable whose final value is to be maximized.
(where bmax is a constant). Note that we can also prescribe bounds on the thrust acceleration
(thrust per unit mass) Γ ≡ T /m as 0 ≤ Γ ≤ Γmax , where Γmax is achieved by running the
engine at Tmax . However, Γmax is not constant, but increases due to the decreasing mass.
One must keep track of the changing mass in order to compute Γ for a given thrust level,
but this is easy to do, especially if the thrust is held constant, e.g., at its maximum value.
However, if the propellant mass required is a small fraction of the total mass, a constant
Γmax approximation can be made.
The cost functional representing minimum-propellant consumption for the CSI case is
given in Eq. (10.13) as
Z tf
J= Γ(t)dt (10.25)
to
We note from Eq. (10.28) that F (tf ) is equal to the cost JCSI . In the constant thrust case
Γ varies according to Γ̇ = 1c Γ2 , which is consistent with the mass decreasing linearly with
time. (See Exercise 1.)
The equation of motion is
· ¸ · ¸
ṙ v
ẋ = = (10.29)
v̇ g(r) + Γu
with the initial state x(to ) specified.
The first-order necessary conditions for an optimal CSI trajectory were first derived by
Lawden [1963] using classical calculus of variations. In the derivation that follows, an optimal
control theory formulation is used, but the derivations and examples are analogous to those
of Lawden and Breakwell. One significant difference is that the mass is not considered a
state variable, but is kept track of separately which simplifies the state equations and the
adjoint equations by having fewer variables.
In order to minimize the cost in Eq. (10.25) we form the Hamiltonian using Eq. (10.29)
as
H = Γ + λTr v + λTv [g(r) + Γu] (10.30)
10.5. FIRST-ORDER NECESSARY CONDITIONS 195
T ∂H
λv = −
λ̇ = − λ Tr (10.32)
∂v
where
∂g(r)
G(r) ≡ (10.33)
∂r
is the symmetric 3 × 3 gravity gradient matrix. (See Exercise 2.)
which may describe an orbital intercept, rendezvous, etc. The boundary conditions on
Eqs. (10.31) and (10.32) are given in terms of
as
∂Φ ∂Ψ
λ Tr (tf ) = = νT (10.38)
∂r(tf ) ∂r(tf )
∂Φ ∂Ψ
λ Tv (tf ) = = νT (10.39)
∂v(tf ) ∂v(tf )
196 CHAPTER 10. GENERAL THEORY OF OPTIMAL ROCKET TRAJECTORIES
λv (t)
p(t) ≡ −λ (10.40)
The optimal thrust unit vector is then in the direction of the primer vector, specifically:
p(t)
u(t) = (10.41)
p(t)
and
λTv u = − λv = − p (10.42)
Equations (10.31), (10.32), (10.38), and (10.41) combine to yield the primer vector equation
p̈ = G(r)p (10.44)
The boundary conditions on the solution to Eq. (10.42) are obtained from Eqs. (10.36) and
(10.37):
∂Ψ
pT (tf ) = − ν T (10.45)
∂v(tf )
∂Ψ
ṗT (tf ) = ν T (10.46)
∂r(tf )
We note that in Eq. (10.43) the final value of the primer vector for an optimal intercept
(only final position specified) is the zero vector, because the terminal constraint Ψ does not
depend on the final velocity v(tf ).
Using Eqs. (10.38) and (10.41) the Hamiltonian of Eq. (10.30) can be rewritten as
H = − (p − 1) Γ + ṗT v − pT g (10.47)
To minimize the Hamiltonian over the choice of the thrust acceleration magnitude, Γ,
we note that the Hamiltonian is a linear function of Γ, and thus the minimizing value for
10.5. FIRST-ORDER NECESSARY CONDITIONS 197
Γmax
S(t)
MT MT MT
0
NT NT t
0 ≤ Γ ≤ Γmax will depend on the algebraic sign of the coefficient of Γ in Eq. (10.45). It is
convenient to define the switching function
S(t) ≡ p(t) − 1 (10.48)
The choice of the thrust acceleration magnitude, Γ, that minimizes H is then given by
the bang-bang control law:
½
Γmax for S > 0 (p > 1)
Γ= (10.49)
0 for S < 0 (p < 1)
That is, the thrust magnitude switches between its limiting values of 0 (an NT, null-thrust,
arc) and Tmax (an MT, maximum-thrust, arc) each time S(t) passes through 0 [i.e. p(t) passes
through 1] according to Eq. (10.47). Figure 10.1 shows an example switching function for
an NT-MT-NT sequence.
The possibility also exists that S(t) ≡ 0 [p(t) ≡ 1] on an interval of finite duration.
From Eq. (10.45) it is evident that in this case the thrust acceleration magnitude is not
determined by the Minimum Principle and may take on intermediate values between 0 and
Γmax . This IT (intermediate thrust arc) in Lawden [1963] is called a singular arc in optimal
control.
From the Weierstrass-Erdmann corner conditions of chapter 8, we know that, at a corner
separating thrust arcs of different types, λ r and λ v , and hence p and ṗ, are continuous.
Equation (10.46) then indicates that the switching function S(t) is also continuous.
Lawden explained to co-author Prussing the origin of the term primer vector in a personal
letter in 1990:“In regard to the term ‘primer vector’ you are quite correct in your supposition.
I served in the artillery during the war [World War II] and became familiar with the initiation
of the burning of cordite by means of a primer charge. Thus, p=1 is the signal for the rocket
motor to be ignited.”
It follows then from Eq. (10.27) that, if T = Tmax and the engine is on for a total of ∆t
time units,
Γmax (t) = eF (t)/c Tmax /mo = Tmax / (mo − bmax ∆t) (10.50)
198 CHAPTER 10. GENERAL THEORY OF OPTIMAL ROCKET TRAJECTORIES
S(t)
t1
t
NT NT
Even though the gravitational field is time-invariant, the Hamiltonian in this formulation
does not provide a first integral (constant of the motion) on an MT arc, because Γmax is
an explicit function of time as shown in Eq. (10.48). On NT and IT arcs, however, the
Hamiltonian is constant. (See Example 10.2.) From the corner conditions we note that the
Hamiltonian is continuous at a corner between arcs of different types, which is also evident
from Eq. (10.45) because S = 0 (i.e. p = 1) at those instants when Γ is discontinuous.
because p, ṗ, and g are continuous, S = 0 (p = 1), and ∆v = ∆vp. Thus ṗT p = ṗ = 0
(See Exercise 7), which implies that Ṡ = 0. So the hypothetical case shown in Fig. 10.3 does
not exist and there can be no cusp. This argument does not apply at the terminals of the
10.5. FIRST-ORDER NECESSARY CONDITIONS 199
S(t)
t1
t
NT NT
trajectory because the corner conditions do not apply at the terminals. So in general ṗ 6= 0
at the terminals.
In Lion and Handelsman [1968] a procedure is developed to iteratively improve a nonop-
timal trajectory (that violates one or more of the necessary conditions summarized below) to
converge to an optimal trajectory. As discussed in Prussing [2010], when adding a midcourse
impulse lowers the cost, the cost gradients with respect to the midcourse impulse position
and time (which are used to iterate on these variables) depend on the discontinuities in ṗ
and H at the midcourse impulse. As the midcourse impulse position and time approach
their optimal values, these discontinuities tend to zero.
The necessary conditions (NC) for an optimal impulsive trajectory, first derived by Law-
den [1963], are as follows
1. The primer vector and its first derivative are continuous everywhere.
2. The magnitude of the primer vector satisfies p(t) ≤ 1 with the impulses occurring at
those instants at which p = 1.
3. At the impulse times the primer vector is a unit vector in the optimal thrust direction.
For a linear system, Prussing [1995] shows that these NC are also sufficient conditions
(SC) for an optimal trajectory. Prussing also derives an upper bound on the number of
impulses required for an optimal solution. However, for a nonlinear system no upper bound
exists.
We note also that for a thrust impulse at time tk
p(t)
Figure 10.4: Sample primer vector history for a three-impulse trajectory. The magnitude of
the primer vector satisfies p(t) ≤ 1 with the impulses occurring at those instants at which
p = 1.
where t+ −
k and tk are times immediately after and before the impulse time, respectively.
Equation (10.27) then becomes the familiar solution to the rocket equation, Eq. (10.9), with
the interval term equal to zero.
m(t+ −
k ) = m(tk ) e
−∆vk /c
(10.54)
Figure 10.4 illustrates the primer vector magnitude for a three-impulse trajectory.
p2
āt
p̄
b̄
θ
p1
Figure 10.5: In a uniform gravity field the primer vector is restricted to a plane which means
that the thrust direction may only vary in a plane.
p = at + b (10.61)
where a and b are constant vectors. Equation (10.61) is, obviously, the equation of a straight
line. In the special case where a=0, the thrust direction never varies throughout the maneu-
ver. In the general case, the thrust direction may vary in a plane, determined by the vectors
a and b, as illustrated in Fig. 10.5. Components of the primer vector (in general) are of the
form:
p1 = a1 t + b1 (10.62a)
p2 = a2 t + b2 (10.62b)
p3 = 0 (10.62c)
p2
b̄ āt
p1
p̄
Figure 10.6: The primer vector in a uniform field may decrease in magnitude initially, reach
a minimum, and then increase.
S(t)
MT MT
NT t
Figure 10.7: At most only three subarcs can exist MT, NT, MT according to the switching
function for the primer vector in a uniform field.
are satisfied, it is possible to have a number of IT arcs. In the case where the end conditions
rule out an IT arc, the optimal trajectory will consist of no more than three NT and MT
subarcs (as shown by Leitmann [1959]). In Fig. 10.6 we illustrate possible behavior of the
primer vector. In the figure we note that p is the distance from the origin. In the general
case, we can have three phases:
1. The magnitude of the primer vector, p, starts out at some initial value,
order. We also note that for impulsive thrusts there are at most two impulses. These simple
examples demonstrate the power of Lawden’s primer vector theory.
10.7 Summary
For optimal rocket trajectories in space, we assume that the only forces acting on the
spacecraft are due to gravity and rocket thrust. The thrusters can be pointed in arbitrary
directions and the magnitude of the thrust may be limited. The cost to be minimized, J, is
the very important one of minimum-propellant consumption (which is equivalent to maxi-
mizing final mass). For a space mission, every kilogram of propellant saved by optimization
provides an additional kilogram of payload at the final destination.
For this problem, Lawden found, using the calculus of variations, that the adjoint (costate)
vector corresponding to the velocity vector provides the optimal thrust direction. Analysis
of this vector, the primer vector, indicates when null thrust (NT), maximum thrust (MT), or
intermediate thrust (IT) arcs are optimal for a constant specific impulse (CSI) engine. For
a variable specific impulse (VSI) engine the magnitude of the optimal thrust acceleration is
equal to the magnitude of the primer vector.
If the thrust magnitude in the CSI case is unbounded, we have the impulsive thrust
case and the change in velocity is instantaneous. In reality, impulsive thrust does not exist,
but the short duration of deep space thrust maneuvers compared with the time between
maneuvers can be modeled as impulsive.
Lawden neatly summarizes the necessary conditions for optimal impulsive thrust trajec-
tories in four statements involving the primer vector. These conditions also form the basis
for primer vector theory, in which the primer vector evaluated on a nonoptimal impulsive
trajectory can be used to determine an optimal trajectory for the same terminal conditions.
When Lawden’s primer vector is applied to the problem of optimal trajectories in a
uniform gravitational field with bounded thrust, we rediscover the bilinear tangent steering
law from our launch from flat-Earth problem.
10.8 Exercises
1. Show that for a constant thrust, Γ̇ = 1c Γ2 corresponds to a constant mass flow rate b.
Hint: m(t) = mo − bt.
2. Show that the gravity gradient matrix in Eq. (10.33) is symmetric for a conservative
∂V
gravitational field. Hint: write gi = − ∂ri
, where V is a (scalar) gravitational potential
∂gi
function, then calculate Gij = ∂rj .
3. Assuming a static gravitational field where g(r) is not an explicit function of time,
determine under what circumstances ġ ≡ dg dt
is nonzero.
4. Consider the gravity gradient matrix for the case of a central gravitational field:
10.8. EXERCISES 205
4a. Determine the gravity gradient matrix G(r) for the general central gravitational
field: g(r) = g(r)r. Hint: the answer depends on the variables r, g(r), and
g 0 (r) ≡ dg/dr.
µ
4b. Verify that your answer specializes to the result in Example 10.1. for g(r) = − r3
.
6a. Show that the scalar a defined by Lion and Handelsman [1968] is constant on an
NT arc between impulses, where a = 2ṗT r + pT v − 3Ht.
6b. Demonstrate that the discontinuity in the variable a at an optimal impulse is
equal to the magnitude of the velocity change ∆v.
6c. Using the result of 6b. determine a new quantity â that is both constant between
impulses and continuous at an optimal impulse. Hint: form â by adding a term
to a that depends on the mass.
7. Show that when p = 1, ṗ = ṗT p, where ṗ ≡ dp/dt 6= |ṗ| in general. Hint: differentiate
p2 = pT p.
9a. Based on the sentence following Eq. (10.46), specialize the form of the primer
vector in Eq. (10.59) for an optimal intercept.
9b. Qualitatively describe the behavior of the thrust direction.
9c. Describe the possible subarc sequences involving MT, NT, and IT arcs.
9d. For impulsive thrusts, what is the maximum number of impulses?
206 CHAPTER 10. GENERAL THEORY OF OPTIMAL ROCKET TRAJECTORIES
True False
2. In the theory of optimal rocket trajectories, when impulsive thrusts are considered,
then the switching function, S = p − 1, is never greater than zero.
True False
3. Derivation of the rules for the primer vector is based mainly on the Minimum Principle
and the Weierstrass-Erdmann corner conditions.
True False
True False
True False
True False
True False
8. No modifications of the Minimum Principle are required for the State Variable Inequal-
ity Constraint (SVIC) Problem.
True False
10.9. TRUE OR FALSE QUIZ FOR CHAPTERS 6–10 207
9. If x∗ (t) is a weak extremal, then it may be possible to find a control which provides a
lower cost.
True False
10. For a uniform gravity field, if the end conditions rule out an IT arc, then the optimal
trajectory will consist of not more than three NT and MT subarcs.
True False
11. The Minimum Principle always tells us everything we need to know to find optimal
trajectories.
True False
12. The main difference between the Pontryagin Minimum Principle and the Weierstrass
condition is the class of functions of the admissible controls.
True False
13. In general, the Legendre-Clebsch condition applies to a broader class of controls than
does the Pontryagin Minimum Principle.
True False
True False
True False
Solution: 1T, 2T, 3T, 4T, 5F, 6F, 7F, 8F, 9T, 10T, 11F, 12T, 13F, 14F, 15F.
208 CHAPTER 10. GENERAL THEORY OF OPTIMAL ROCKET TRAJECTORIES
10.10 References
J.V. Breakwell, The Optimization of Trajectories. Journal of the Society for Industrial
and Applied Mathematics, 7(2):215-247, June 1959.
D.F. Lawden, Optimal Trajectories for Space Navigation. Butterworths, London, 1963.
J.P. Marec, Optimal Space Trajectories. Elsevier Scientific Publishing Company, New
York, 1979.
J.E. Prussing and B.A. Conway, Orbital Mechanics. Oxford University Press, New
York, second edition, 2013.
J.E. Prussing, Optimal Impulsive Linear Systems: Sufficient Conditions and Maximum
Number of Impulses. Journal of the Astronautical Sciences, 43(2):195-206, April-June
1995.
J.E. Prussing and S.L. Sandrik, Second-Order Necessary Conditions and Sufficient
Conditions Applied to Continuous-Thrust Trajectories. Journal of Guidance, Control,
and Dynamics, Engineering Note, 28(4):812-816, 2005.
J.E. Prussing, Chapter 2: Primer Vector Theory and Applications. Spacecraft Tra-
jectory Optimization. B.A. Conway, editor. Cambridge University Press, New York,
2010.