Automatica 70 (2016) 128–139

Contents lists available at ScienceDirect

Automatica
journal homepage: www.elsevier.com/locate/automatica

Economic model predictive control without terminal constraints for

optimal periodic behavior✩
Matthias A. Müller a , Lars Grüne b
a
Institute for Systems Theory and Automatic Control, University of Stuttgart, 70550 Stuttgart, Germany
b
Mathematical Institute, University of Bayreuth, 95440 Bayreuth, Germany

article info abstract

Article history: In this paper, we analyze economic model predictive control schemes without terminal constraints, where
Received 8 June 2015 the optimal operating behavior is not steady-state operation, but periodic behavior. We first show by
Received in revised form means of a counterexample, that a classical receding horizon control scheme does not necessarily result
21 January 2016
in an optimal closed-loop behavior. Instead, a multi-step MPC scheme may be needed in order to establish
Accepted 26 February 2016
Available online 15 April 2016
near optimal performance of the closed-loop system. This behavior is analyzed in detail, and we show that
under suitable dissipativity and controllability conditions, desired closed-loop performance guarantees as
Keywords:
well as convergence to the optimal periodic orbit can be established.
Economic model predictive control © 2016 Elsevier Ltd. All rights reserved.
Optimal periodic operation
Nonlinear systems

1. Introduction
In recent years, the study of economic model predictive control
(MPC) schemes has received a significant amount of attention.
In contrast to standard stabilizing MPC, the control objective is
the minimization of some general performance criterion, which
need not be related to any specific steady-state to be stabilized.
These type of control problems arise in many different fields
of application, ranging, e.g., from the process industry over
building climate control or the control of wind turbines to the
development of sustainable climate policies (see, e.g., Amrit,
Rawlings, & Biegler, 2013; Chu, Duncan, Papachristodoulou, &
Hepburn, 2012; Gros, 2013; Heidarinejad, Liu, & Christofides, 2012;
Hovgaard, Larsen, Edlund, & Jørgensen, 2012; Mendoza-Serrano &
Chmielewski, 2014). In the literature, various different economic
MPC schemes have been developed for which desired closed-
loop properties such as performance estimates or convergence can
be guaranteed. These include schemes with additional terminal
equality or terminal region constraints (Amrit, Rawlings, & Angeli,
✩ The work of Matthias A. Müller was supported by DFG Grant MU3929/1-1. The
work of Lars Grüne was supported by DFG Grant GR1569/13-1. The material in this
paper was partially presented at the 54th IEEE Conference on Decision and Control,
December 15–18, 2015, Osaka, Japan. This paper was recommended for publication
in revised form by Associate Editor Alessandro Astolfi under the direction of Editor
Andrew R. Teel.
E-mail addresses: matthias.mueller@ist.uni-stuttgart.de (M.A. Müller),
lars.gruene@uni-bayreuth.de (L. Grüne).
http://dx.doi.org/10.1016/j.automatica.2016.03.024
0005-1098/© 2016 Elsevier Ltd. All rights reserved.
2011; Angeli, Amrit, & Rawlings, 2012), with generalized terminal
constraints (Fagiano & Teel, 2013; Müller, Angeli, & Allgöwer,
2013), without terminal constraints (Grüne, 2013), and Lyapunov-
based schemes (Heidarinejad et al., 2012) (see also the recent
survey article Ellis, Durand, & Christofides, 2014).
A distinctive feature of economic MPC is the fact that the closed-
loop trajectories are not necessarily convergent to a steady-state,
but can exhibit more complex, e.g., periodic, behavior. In particu-
lar, the optimal operating behavior for a given system depends on
its dynamics, the considered performance criterion and the con-
straints which need to be satisfied. The case where steady-state
operation is optimal is by now fairly well understood, and vari-
ous closed-loop guarantees have been established in this case. For
example, a certain dissipativity property is both sufficient (Angeli
et al., 2012) and (under a mild controllability condition) neces-
sary (Müller, Angeli, & Allgöwer, 2015) for a system to be opti-
mally operated at steady-state. The same dissipativity condition
(strengthened to strict dissipativity) was used in Amrit et al. (2011)
and Angeli et al. (2012) to prove asymptotic stability of the opti-
mal steady-state for the resulting closed-loop system with the help
of suitable terminal constraints. Similar (practical) stability results
were established in Grüne (2013) and Grüne and Stieler (2014)
without such terminal constraints.
On the other hand, the picture is still much less complete
in case that the optimal operating behavior is non-stationary.
This situation occurs in many cases of practical interest, such
as in certain chemical reactors (see, e.g. Amrit et al., 2013;
Angeli et al., 2012; Ellis et al., 2014) or in applications with
time varying (energy) prices or demand (see, e.g., Limon, Pereira,
 M.A. Müller, L. Grüne / Automatica 70 (2016) 128–139
Muñoz de la Peña, Alamo, & Grosso, 2014; Mendoza-Serrano &
Chmielewski, 2014). For such cases, in Angeli et al. (2012) it was
shown that when using some periodic orbit as (periodic) terminal
constraint within the economic MPC problem formulation, then
the resulting closed-loop system will have an asymptotic average
performance which is at least as good as the average cost of
the periodic orbit. Convergence to the optimal periodic orbit was
established in1 Huang et al. (2011) and Zanon, Gros, and Diehl
(2013) using similar terminal constraints, and in Limon et al.
(2014) for linear systems and convex cost functions using less
restrictive generalized periodic terminal constraints. Furthermore,
dissipativity conditions which are suited as sufficient conditions
such that the optimal operating behavior of a system is some
periodic orbit were recently proposed in Grüne and Zanon (2014).
In this paper, we consider economic MPC without terminal con-
straints for the case where periodic operation is optimal. Using no
terminal constraints is in particular desirable in this case as the op-
timal periodic orbit then need not be known a priori (i.e., for im-
plementing the economic MPC scheme). Furthermore, the online
computational burden might be lower and a larger feasible region
is in general obtained. We first show by means of a counterexample
(see Section 3), that the classical receding horizon control scheme,
consisting of applying the first step of the optimal predicted input
sequence to the system at each time, does not necessarily result in
an optimal closed-loop performance. We then prove in Section 4
that this undesirable behavior can be resolved by possibly using a
multi-step MPC scheme instead. In particular, we show that the re-
sulting closed-loop system has an asymptotic average performance
which is equal to the average cost of the optimal periodic orbit
(up to an error term which vanishes as the prediction horizon in-
creases). This recovers the results of Angeli et al. (2012), where
periodic terminal constraints were used as discussed above. In
Section 5 we derive checkable sufficient conditions based on dissi-
pativity and controllability in order to apply the results of Section 4.
Furthermore, in Section 6 we show that under the same conditions,
also (practical) convergence of the closed-loop system to the opti-
mal periodic orbit can be established.
We close this section by noting that our analysis partly builds
on the one in Grüne (2013), where closed-loop performance
guarantees and convergence results for economic MPC without
terminal constraints were established for the case where the
optimal operating behavior is steady-state operation. However,
while some of the employed concepts and ideas are similar to
those in Grüne (2013), various properties of predicted and closed-
loop sequences are different in the periodic case considered in
this paper, and hence also different analysis methods are required.
Finally, we note that a preliminary version of some of the results of
this paper have appeared in the conference paper (Müller & Grüne,
2015). Compared to Müller and Grüne (2015), the main novelties
of the present paper are a more comprehensive exposition of
the subject including various additional remarks and examples
as well as all proofs (which were partly missing in Müller and
Grüne (2015)), the development of our results using a more
general dissipativity condition, and the establishment of closed-
loop (practical) convergence to the optimal periodic orbit.
2. Preliminaries and setup
Let I[a,b] denote the set of integers in the interval [a, b] ⊆ R,
and I≥a the set of integers greater than or equal to a. For a set
1 In Huang, Harinath, and Biegler (2011), however, again the standard
assumption as in stabilizing MPC, i.e., positive definiteness of the cost function, was
imposed, which means that no general performance criterion as in economic MPC
could be considered.
129
A ⊆ I≥0 , #A denotes its cardinality (i.e., the number of elements).
For a ∈ R, ⌊a⌋ is defined as the largest integer smaller than or equal
to a. The distance of a point x ∈ Rn to a set A ⊆ Rn is defined
as |x|A := infa∈A |x − a|. For a set A ⊆ Rn and ε > 0, denote
by Bε (A) := {x ∈ Rn : |x|A ≤ ε}. By L we denote the set of
functions ϕ : R≥0 → R≥0 which are continuous, nonincreasing
and satisfy limk→∞ ϕ(k) = 0. Furthermore, by KL we denote the
set of functions γ : R≥0 × R≥0 → R≥0 such that for each ϕ ∈ L,
the function γ (k) := γ (ϕ(k), k) satisfies γ ∈ L. Note that the
definition of a KL-function requires weaker properties than those
for classical KL-functions, i.e., each KL-function is also a KL-
function (but the converse does not hold).
We consider nonlinear discrete-time systems of the form
x(k + 1) = f (x(k), u(k)), x(0) = x (1)
n m n
with k ∈ I≥0 and f : R × R → R . System (1) is subject to
pointwise-in-time state and input constraints x(k) ∈ X ⊆ Rn and
u(k) ∈ U ⊆ Rm for all k ∈ I≥0 . For a given control sequence
u = (u(0), . . . , u(K )) ∈ UK +1 (or u = (u(0), . . .) ∈ U∞ ), denote
by xu (k, x) the corresponding solution of system (1) with initial
condition xu (0, x) = x. For a given x ∈ X, the set of all feasible
control sequences of length N is denoted by UN (x), where a feasible
control sequence is such that u(k) ∈ U for all k ∈ I[0,N −1] and
xu (k, x) ∈ X for all k ∈ I[0,N ] . Similarly, the set of all feasible control
sequences of infinite length is denoted by U∞ (x). In the following,
we assume for simplicity that U∞ (x) ̸= ∅ for all x ∈ X. This means
that for all initial conditions x ∈ X, there exists a trajectory which
stays in X for all times, i.e., the set X is control invariant under
controls in U. This assumption might be restrictive in general, and
it can be relaxed if desired, using, e.g., methods similar to those
in Grüne and Pannek (2011, Chap. 8) or Faulwasser and Bonvin
(2015). However, the technical details of such an extension are
beyond the scope of this paper.
Remark 1. For ease of presentation, we use decoupled state and
input constraint sets X and U in the statement of our results.
Nevertheless, all results in this paper are also valid for possibly
coupled state and input constraints, i.e., (x(k), u(k)) ∈ Z for all
k ∈ I≥0 and some Z ⊆ Rn × Rm , which will also be used in the
examples. 
System (1) is equipped with a stage cost function ℓ : X × U → R,
which is assumed to be bounded from below on X × U, i.e., ℓmin :=
infx∈X,u∈U ℓ(x, u) is finite. Note that this is, e.g., the case if X × U
is compact and ℓ is continuous. Without loss of generality, in the
following we assume that ℓmin ≥ 0. We then define the following
finite horizon cost functional
N −1

JN (x, u) := ℓ(xu (k, x), u(k)) (2)
k=0
and the corresponding optimal value function
VN (x) := inf JN (x, u). (3)
u∈UN (x)
In the following, we assume that for each x ∈ X, a control sequence
u∗N ,x ∈ UN (x) exists such that the infimum in (3) is attained, i.e., u∗N ,x
satisfies VN (x) = JN (x, u∗N ,x ). Since we assume that U∞ (x) ̸= ∅
for all x ∈ X, this is, e.g., satisfied if f and ℓ are continuous and U
is compact. A standard MPC scheme without additional terminal
cost and terminal constraints then consists of minimizing, at each
time instant k ∈ I≥0 with current system state x = x(k), the
cost functional (2) with respect to u ∈ UN (x) and applying the
first element of the resulting optimal input sequence u∗N ,x to the
system. This means that the resulting receding horizon control
input to system (1) is given by uMPC (k) := u∗N ,xu (k,x) (0), where
MPC
xuMPC (·, x) denotes the corresponding closed-loop state sequence.
130 M.A. Müller, L. Grüne / Automatica 70 (2016) 128–139

For T ∈ I≥0 , the cost along T steps of these closed-loop sequences

is denoted by JTcl (x, uMPC ) = k=0 ℓ(xuMPC (k, x), uMPC (k)), and the
 T −1

av (x, uMPC ) :=
cl
resulting asymptotic average cost is given by J∞,
lim supT →∞ (1/T )JTcl (x, uMPC ).
In Grüne (2013), it was shown that if system (1) is optimally op-
erated at some steady-state (x∗ , u∗ ) with cost ℓ0 := ℓ(x∗ , u∗ ), then Fig. 1. Illustration of the states x (nodes) and feasible transitions (edges) with
corresponding input u and cost ℓ in Example 4.
under suitable conditions the asymptotic average performance of
av , equals ℓ0 (up to an error term which
cl
the closed-loop system, J∞,
3. Motivating example
vanishes as N → ∞), and the closed-loop system converges to
x∗ (again up to an error term which vanishes as N → ∞). In this
paper, we consider the more general case where the optimal oper- Example 4. Consider the one-dimensional system x(k + 1) = u(k)
ating behavior of system (1) is not necessarily stationary but given with state and input constraint set Z = {(−1, −1), (−1, 0), (0, 1),
by some periodic orbit with period P ∈ I≥1 . To this end, consider (1, 0)} consisting of four elements only and cost ℓ(x, u) defined as
the following definitions. ℓ(−1, −1) = 1, ℓ(−1, 0) = 1,
p p p
ℓ(0, 1) = 1 − 2ε, ℓ(1, 0) = 1 + ε
Definition 2. A set of state/input pairs Π = {(x0 , u0 ), . . . , (xP −1 ,
p for some constant ε > 0, see also Fig. 1. The system is optimally
uP −1 )} with P ∈ I≥1 is called a feasible P-periodic orbit of system (1),
p p p p p operated at the two-periodic orbit given by Π = {(0, 1), (1, 0)},
if xk ∈ X and uk ∈ U for all k ∈ I[0,P −1] , xk+1 = f (xk , uk ) for p p
and with average cost ℓ0 := (1/2) k=0 ℓ(xk , uk ) = 1 − ε/2. For
1
p p p
all k ∈ I[0,P −2] , and x0 = f (xP −1 , uP −1 ). It is called a minimal P- initial condition x0 = −1, it follows that for any even prediction
p p
periodic orbit if xk1 ̸= xk2 for all k1 , k2 ∈ I[0,P −1] with k1 ̸= k2 .  horizon N ∈ I≥2 , the optimal open-loop input sequence u∗N ,x0
is such that xu∗ (1, x0 ) = 0 and then xu∗ (·, x0 ) stays on ΠX .
In the following, denote by ΠX the projection of Π on X, i.e., ΠX := N ,x0 N , x0
This means that also the closed-loop system converges to the set
{xp0 , . . . , xpP −1 }.
ΠX and J∞, av (−1, uMPC ) = ℓ0 . On the other hand, for any odd
cl

prediction horizon N ∈ I≥3 , the optimal open-loop input sequence

Definition 3. System (1) is optimally operated at a periodic orbit Π u∗N ,x0 is such that xu∗ (1, x0 ) = −1, xu∗ (2, x0 ) = 0, and
N ,x0 N ,x0
if for each x ∈ X and each u ∈ U∞ (x) the following inequality then xu∗ (·, x0 ) stays on ΠX . But this means that the closed-loop
N , x0
holds:
system stays at x = −1 for all times, i.e., xuMPC (k, x0 ) = −1 for all
av (−1, uMPC ) = 1 > 1 − ε/2 = ℓ0 .
T −1 cl
k ∈ I≥0 , and hence J∞, 
ℓ(xu (k, x), u(k))

P −1
k=0 1 The above example shows that the ‘‘phase’’ on the periodic orbit
lim inf ≥ ℓ(xpk , upk ).
T →∞ T P k=0 is decisive, i.e., what is the optimal time to converge to the
periodic orbit. Similar examples can also be constructed where
the cost along the optimal periodic orbit is constant (see Müller
Definition 3 means that each feasible solution will result in an & Grüne, 2015, Example 5). The above shows that in general,
asymptotic average performance which is at most as good as the one cannot guarantee that for all sufficiently large prediction
average performance of the periodic orbit Π . Furthermore, for P = horizons N, the closed-loop asymptotic average performance
1 the notion of optimal steady-state operation (Angeli et al., 2012;
 P −1 p p
av (x, uMPC ) = (1/P ) k=0 ℓ(xk , uk ) (plus some error
cl
satisfies J∞,
Müller et al., 2015) is recovered. Note that if system (1) is optimally term which vanishes as N → ∞), as could be established in Grüne
p p p p
operated at some periodic orbit Π = {(x̄0 , ū0 ), . . . , (x̄P −1 , ūP −1 )}, (2013) for the case of optimal steady-state operation, i.e., P = 1. On
then Π is necessarily an optimal periodic orbit for system (1), the other hand, one observes in the above example that if the MPC
i.e. we have scheme is modified in such a way that not only the first value of
the optimal control sequence is applied to the system, but the first
P −1 P −1
1 p p 1 two values, then the closed-loop system converges to the optimal
ℓ(x̄k , ūk ) = inf ℓ(xpk , upk ), (4) P −1 p p
av (x, uMPC ) = (1/P ) k=0 ℓ(xk , uk ),
cl
P k=0 P ∈I≥1 , Π ∈SΠ
P P k=0 periodic orbit and hence J∞,
for all prediction horizons N ∈ I≥2 . In the following, we will
P
where SΠ denotes the set of all feasible P-periodic orbits. Note see that this behavior can be rigorously established under suitable
that in general, there can exist multiple (minimal) optimal periodic assumptions on the problem.
orbits for system (1). On the other hand, if the assumption of strict
dissipativity as required in Section 5 is satisfied for some minimal Remark 5. We note that the above observations do not depend on
the fact that the constraint set Z only consists of a finite number of
periodic orbit Π (cf. Assumption 9), then Π is necessarily the
points. While for clarity of presentation, we chose the above simple
unique minimal optimal periodic orbit.
motivating example, it is not difficult to find examples where the
In case that a system is optimally operated at a periodic orbit Π , same effects occur and Z has non-empty interior. 
the closed-loop system resulting from application of the economic
av (x, uMPC ) =
cl
MPC scheme exhibits optimal performance if J∞, 4. Closed-loop performance guarantees
(1/P ) ℓ(xpk , upk ). As discussed in Section 1, in Angeli et al.
P −1
k=0
(2012) it was shown that this can be achieved in case that ΠX is As mentioned above, in the following we consider a multi-step
used as a periodic terminal constraint. When using no terminal MPC scheme where for some P ∈ I≥1 , an optimal input sequence
constraints, this equality may in general not be achieved, as we u∗N ,x is only calculated every P time instants, and then the first P
show in the following section by means of a counterexample. elements of this sequence are applied to system (1). This means
Nevertheless, optimal performance and (practical) convergence to that the control input to system (1) at time k is given by
the optimal periodic orbit can still be guaranteed also without uMPC (k) = u∗N ,x′ ([k]), (5)
terminal constraints in case a multi-step MPC scheme is used, as
will be shown in Sections 4–6. where x = xuMPC (P ⌊k/P ⌋, x) and [k] := k mod P.
′
 M.A. Müller, L. Grüne / Automatica 70 (2016) 128–139 131

Remark 6. The fact that the P-step MPC scheme operates in open
loop for P time steps will in general reduce robustness with respect
to perturbations compared to a classical MPC scheme in which
the loop is closed in each time step. However, this problem can
be resolved by the following variant, for which the subsequent
results still apply. In this variant, an optimal input sequence is
computed at each time and only the first element is applied to
the system as in standard MPC, but the prediction horizon is
periodically time-varying, i.e., N in (2) is replaced by N − [k]. By
the dynamic programming principle, in absence of perturbations
the closed-loop sequences resulting from application of these
two schemes are the same. However, the second will in general
exhibit better robustness properties in case of uncertainties and
disturbances (see Grüne & Palma, 2015). 
In the following, we establish closed-loop performance guarantees
for the P-step MPC scheme as defined via (5). We first derive a
preparatory result (Proposition 7), which is a generalization of
Proposition 4.1 in Grüne (2013). This result will later be applied to
ensure the desired closed-loop average performance guarantees.
JN′ −P (x(iP )) − δ1 (N − P ). Furthermore, by optimality of VN we
get VN (x(iP )) ≤ JN (x(iP ), ūN −P ,x(iP ) ). Combining the above and
defining I := {k1N −P ,x(iP ) , . . . , kPN −P ,x(iP ) }, from condition (ii) of the
proposition and the definitions of JN and JN′ we obtain
VN (x(iP )) − VN −P (x(iP ))
≤ JN (x(iP ), ūN −P ,x(iP ) ) − JN′ −P (x(iP )) + δ1 (N − P )

= ℓ(xūN −P ,x(iP ) (k, x(iP )), ūN −P ,x(iP ) (k)) + δ1 (N − P )
k∈I
≤ P ℓ0 + P δ2 (N − P ) + δ1 (N − P ). (9)
Recalling that x(0) = x and inserting (9) into (8) for i = 1, . . . , K −
cl
1 yields JKP (x, uMPC ) ≤ VN (x) − VN −P (x(KP )) + (K − 1)(P ℓ0 +
P δ2 (N − P ) + δ1 (N − P )). Moreover, using (9) for i = K yields
−VN −P (x(KP )) ≤ −VN (x(KP )) + P ℓ0 + P δ2 (N − P ) + δ1 (N − P ).
Together with the above, this results in (6). Finally, (7) follows
from (6) by dividing by KP and letting K → ∞, due to the fact
that ℓ(x, u) ≥ ℓmin ≥ 0 for all (x, u) ∈ X × U and VN (x(KP )) ≥
N ℓmin ≥ 0. 

Proposition 7. Assume there exist N , P ∈ I≥1 and δ1 , δ2 ∈ L such As is the case in most MPC proofs, Proposition 7 compares the cost
that for each x ∈ X and each N ∈ I≥N there exists a control sequence along different input sequences. In particular, existence of an input
ūN ,x ∈ UN +P (x) and time instants k1N ,x , . . . , kPN ,x ∈ I[0,...,N +P −1] sequence ūN ,x of length N + P is required such that when summing
satisfying the following conditions. up the stage cost along this sequence for all but P time instants
k1N ,x , . . . , kPN ,x , the resulting cost JN′ (x) is close to the optimal cost
(i) The inequality JN′ (x) ≤ VN (x) + δ1 (N ) holds for VN (x) of problem (3) with horizon N (see condition (i)). If the
N +P −1 sum of the stage cost along these P time instants k1N ,x , . . . , kPN ,x
is close to some value ℓ0 (see condition (ii)), then (7) guarantees

JN′ (x) := ℓ(xūN ,x (k, x), ūN ,x (k)).
k=0 that the closed-loop average performance is upper-bounded by
k̸∈{k1 ,...,kP }
N ,x N ,x this constant ℓ0 (up to an error term which vanishes as N → ∞).
(ii) There exists ℓ0 ∈ R such that for all x ∈ X the following inequality In the following, we construct control sequences ūN ,x such that
P −1 p p
is satisfied: Proposition 7 can be applied with ℓ0 = (1/P ) k=0 ℓ(xk , uk ) for
some P-periodic orbit Π . Then, inequality (7) yields the desired
1 
ℓ(xūN ,x (k, x), ūN ,x (k)) ≤ ℓ0 + δ2 (N ). property that the asymptotic average performance of the closed-
P loop system resulting from application of the P-step MPC scheme
k∈{k1N ,x ,...,kPN ,x }
is less than or equal to the average performance of the periodic
orbit Π (up to an error term which vanishes as N → ∞). As
Then the inequalities
discussed above, this approximately recovers asymptotic average
cl
JKP (x, uMPC ) ≤ VN (x) − VN (xuMPC (KP , x)) + KP ℓ0 performance results obtained in MPC schemes with (periodic)
+ K δ1 (N − P ) + KP δ2 (N − P ) (6) terminal constraints (Angeli et al., 2012). The following theorem
gives sufficient conditions under which the control sequences ūN ,x
and required in Proposition 7 can be constructed. As will be seen
in the proof, for these sequences the time instants k1N ,x , . . . , kPN ,x
av (x, uMPC ) ≤ ℓ0 + δ1 (N − P )/P + δ2 (N − P )
cl
J∞, (7)
are consecutive time instants. However, we note that this is not
hold for all x ∈ X, all N ∈ I≥N +P and all K ∈ I≥0 .  necessarily needed in order to apply Proposition 7.
Proof. Fix x ∈ X and N ∈ I≥N +P . Using the abbreviation
x(k) = xuMPC (k, x), from the dynamic programming principle and Theorem 8. Assume that there exist constants ℓ0 ≥ 0, δ̄ > 0, and
the definition of the multi-step MPC control input in (5), we obtain P ∈ I≥1 and a set Y ⊆ X such that the following properties hold.
that for all i ∈ I≥0
(a) There exists γℓ ∈ K∞ such that for all δ ∈ (0, δ̄] and all
P −1
 x ∈ Bδ (Y) ∩ X there exists a control sequence ux ∈ UP (x) such
ℓ(x(iP + k), uMPC (iP + k)) = VN (x(iP )) − VN −P (x((i + 1)P )). that the inequality (1/P ) k=0 ℓ(xux (k, x), ux (k)) ≤ ℓ0 + γℓ (δ)
 P −1
k=0 holds.
Summing up for i = 0, . . . , K − 1 then yields (b) There exist N0 ∈ I≥1 and a function γV ∈ KL such that for all δ ∈
K −1 
P −1
(0, δ̄], all N ∈ I≥N0 , all x ∈ Bδ (Y) ∩ X and the control sequence
ux ∈ UP from (a) the inequality |VN (x) − VN (xux (P , x))| ≤

cl
JKP (x, uMPC ) = ℓ(x(iP + k), uMPC (iP + k))
i=0 k=0 γV (δ, N ) holds.
= VN (x(0)) − VN −P (x(KP )) (c) There exist σ ∈ L and N1 ∈ I≥N0 with N0 from (b) such that
for all x ∈ X and all N ∈ I≥N1 , each optimal trajectory xu∗ (·, x)
K −1 N ,x
satisfies |xu∗ (kx , x)|Y ≤ σ (N ) for some kx ∈ I[0,N −N0 ] .

VN (x(iP )) − VN −P (x(iP )) .
 
+ (8) N ,x

i =1
Then the conditions of Proposition 7 are satisfied. 
Now consider the summands in (8). Condition (i) of the proposition
with N − P in place of N and x = x(iP ) implies that VN −P (x(iP )) ≥ Proof. See Appendix A. 
132 M.A. Müller, L. Grüne / Automatica 70 (2016) 128–139

In the proof of Theorem 8, the main idea for defining the control Now choose W ⊆ X × U large enough and δ̄ > 0 small enough
sequence ūN ,x ∈ UN +P (x) required in Proposition 7 is the follow- such that Bmax{δ̄,ηP (δ̄)} (Π ) ⊆ W; note that this is possible since
ing: given an optimal input sequence u∗N ,x ∈ UN (x), we insert P
f
Π ⊆ int(X × U). Considering the above, we obtain
additional control values (those from (a)) starting at time kx (from
(c)), where the optimal trajectory xu∗ (·, x) is close to the set Y. |xux (1, x) − xp[j+1] | = |f (x, upj ) − f (xpj , upj )|
N ,x
The corresponding P (consecutive) time instants are then the time ≤ ηf (|(x, upj ) − (xpj , upj )|) = ηf (|x − xpj |) ≤ ηf (δ).
instants k1N ,x , . . . , kPN ,x in Proposition 7 (see Appendix A for details).
Theorem 8 uses similar conditions as Theorem 4.2 in Grüne (2013), Using this argument recursively results in
which were shown to hold in case of optimal steady-state opera- |xux (k, x) − xp[j+k] | ≤ ηfk (δ) (11)
tion. However, there are some crucial differences. Namely, Grüne
(2013, Theorem 4.2) requires that2 |VN (x) − VN (y)| ≤ γV (δ) has for all k ∈ I[1,P ] . Furthermore,
to hold for all y ∈ Y and all x ∈ Bδ (Y) with γV ∈ K∞ , which in P −1 P −1
1 1 
particular implies that VN (x) = VN (y) for all x, y ∈ Y, i.e., the opti- ℓ(xux (k, x), ux (k)) ≤ ℓ(xp[j+k] , up[j+k] )
mal value function is constant on Y. In case that Y = ΠX for some P k =0 P k=0
periodic orbit Π , this can in general not be satisfied, as is the case 
in our motivating examples in Section 3. In Theorem 8, condition + ηℓ (|(xux (k, x), up[j+k] ) − (xp[j+k] , up[j+k] )|)
(b) instead only requires that |VN (x) − VN (xux (P , x))| ≤ γV (δ, N )
holds for all x ∈ Bδ (Y) ∩ X, where ux is the control sequence from P −1
1
condition (a). Furthermore, γV may depend on N, and in particular ≤ ℓ0 + ηℓ (ηfk (δ)) =: ℓ0 + γℓ (δ). (12)
for fixed N, |VN (x) − VN (xux (P , x))| need not go to zero as δ → 0, P k=0
but we only require that γV (δ, N ) → 0 if both N → ∞ and δ → 0.
Since Bmax{δ̄,ηP (δ̄)} (Π ) ⊆ W ⊆ X × U, it follows that xux (k, x) ∈ X
These relaxations are crucial such that Theorem 8 can be applied f

with Y = ΠX for some periodic orbit Π , as shown in the following. for all k ∈ I[1,P ] and all x ∈ Bδ̄ (Y), i.e., ux ∈ UP (x). Hence
condition (a) of Theorem 8 is satisfied with γℓ given by (12) and
ℓ0 = (1/P ) kP=−01 ℓ(xpk , upk ).

5. Checkable sufficient conditions based on dissipativity and
controllability
5.1. Turnpike behavior with respect to periodic orbits
It is easy to verify that the motivating example satisfies the
conditions of Theorem 8 with Y = ΠX , which explains the We now turn our attention to condition (c) of Theorem 8, which
fact that a 2-step MPC scheme results in optimal closed-loop requires that each optimal solution is close to the set Y for at
performance, as observed in Section 3. In general, however, the least one time instant in the interval [0, N − N0 ]. To this end,
conditions of Theorem 8 might be difficult to check since they for system (1) we define the corresponding P-step system with
involve properties of optimal trajectories and the optimal value state x̃ = (x0 , . . . , xP −1 ) ∈ XP , input ũ = (u0 , . . . , uP −1 ) ∈
function. The goal of this section is to provide checkable sufficient UP , dynamics x̃(k + 1) = f P (x̃(k), ũ(k)) and initial condition3
conditions for conditions (a)–(c) of Theorem 8 for the case where xP −1 (0) = x ∈ X, where
Y = ΠX for some periodic orbit Π of system (1). First, we
xũ (1, xP −1 ) f (xP −1 , u0 )
   
briefly discuss that condition (a) follows in a straightforward
way from continuity of f and ℓ. Second, we show that a certain f (x̃, ũ) :=
P
... = f (f (xP −1 , u0 ), u1 ) . (13)
dissipativity condition results in a turnpike behavior of the system xũ (P , xP −1 ) ...
with respect to the optimal periodic orbit, from which together
For a given control sequence ũ ∈ UKP with K ∈ I≥1 , the
with suitable controllability assumptions condition (c) follows (see
corresponding solution of system (13) is denoted by x̃ũ (k, x) for
Section 5.1). Third, we discuss in Section 5.2 how condition (b)
k ∈ I[1,K ] . This means that for a given control sequence u ∈ UKP
can be established under the same dissipativity and controllability
with K ∈ I≥1 , partitioned into ũ(k) = (u(kP ), . . . , u((k + 1)P − 1))
assumptions. Finally, in Section 5.3 we combine all these aspects to
for all k ∈ I[0,K −1] , we have that
obtain the main result concerning closed-loop performance under
dissipativity and controllability. x̃ũ (k, x) = (xu ((k − 1)P + 1, x), . . . , xu (kP , x))
Before turning our attention to conditions (b) and (c) of
for all k ∈ I[1,K ] . For the P-step system (13), define the cost
Theorem 8, we briefly discuss how for the case that Y = ΠX
function ℓ̃(x̃, ũ) ℓ(xũ (j, xP −1 ), uj ). Furthermore, for
 P −1
for some P-periodic orbit Π ⊆ int(X × U), condition (a) with
:= j =0

ℓ0 = (1/P ) Pk=−01 ℓ(xpk , upk ) follows if f and ℓ are continuous. In

 (x̃, ũ) ∈ XP × UP and a P-periodic orbit Π , define |(x̃, ũ)|Π :=
j=0 |(xũ (j, xP −1 ), uj )|Π and |x̃|ΠX := j=0 |xũ (j, xP −1 )|ΠX . We
P −1 P −1
this case, for each x ∈ Bδ (Y) for some δ ∈ (0, δ̄], by definition of
p
Y it holds that x ∈ Bδ (xj ) for some j ∈ I[0,P −1] . Then, if f and ℓ are can now make the following assumption of strict dissipativity for
the P-step system (13).
continuous, the control sequence ux ∈ UP in condition (a) can be
chosen as Assumption 9 (Strict Dissipativity). The P-step system (13) is
p p p p
ux = (uj , . . . , uP −1 , u0 , . . . , uj−1 ), (10) strictly dissipative with respect to a periodic orbit Π and the
 P −1 p p
supply rate s(x̃, ũ) = ℓ̃(x̃, ũ) − k=0 ℓ(xk , uk ), i.e., there exists a
and the function γℓ can be computed as follows. As f and ℓ are con-
storage function λ̃ : XP → R and a function αℓ ∈ K∞ such that
tinuous, for each compact set W ⊆ X × U there exist ηf , ηℓ ∈ K∞
such that |f (x, u) − f (x′ , u′ )| ≤ ηf (|(x, u) − (x′ , u′ )|) and |ℓ(x, u) − P −1

ℓ(x′ , u′ )| ≤ ηℓ (|(x, u) − (x′ , u′ )|) for all (x, u), (x′ , u′ ) ∈ W. λ̃(f P (x̃, ũ)) − λ̃(x̃) ≤ ℓ̃(x̃, ũ) − ℓ(xpk , upk ) − αℓ (|(x̃, ũ)|Π ) (14)
k=0

2 We note that in Grüne (2013), all results were stated in terms of averaged
cost functionals, i.e., the right hand side of (2) multiplied by 1/N; this is why the 3 Initial conditions for the first P − 1 components of x̃, i.e., x (0), . . . , x
0 P −2 (0), can
condition in Grüne (2013, Theorem 4.2) in fact reads |VN (x) − VN (y)| ≤ γV (δ)/N. be arbitrary.
 M.A. Müller, L. Grüne / Automatica 70 (2016) 128–139 133

for all x̃ ∈ XP and all ũ ∈ UP . Furthermore, the storage function λ̃ as specified in the theorem we obtain that
is bounded on XP .  P −1

J̃N (x, u) = JN (x, u) − ⌊N /P ⌋ ℓ(xpk , upk )
As was discussed in Grüne and Zanon (2014), Assumption 9 is k=0
a sufficient condition for system (1) to be optimally operated at
the periodic orbit Π . In fact, as shown in Müller, Grüne, and All- + λ̃(x̃ũ (0, x)) − λ̃(x̃ũ (⌊N /P ⌋, x))
göwer (2015), Assumption 9 is sufficient and (under an additional
  P −1
N
controllability condition) also necessary for a slightly stronger ≤ JN (x, u) − −1 ℓ(xpk , upk ) + C ′
P k=0
property than optimal periodic operation, namely suboptimal op-
eration off the periodic orbit Π , meaning that the unique optimal P −1
N 
operating behavior is the periodic orbit Π . Furthermore, note that = JN (x, u) − ℓ(xpk , upk ) + C
P k=0
for P = 1, we recover the standard dissipativity condition which is
typically used in economic MPC in case that steady-state operation ≤ ν + C. (15)
is optimal (see, e.g., Angeli et al., 2012; Müller et al., 2015). Finally, Now assume for contradiction that Qε < N −P +1−P (ν+C )/αℓ (ε).
we remark that our dissipativity condition here is weaker than the Then there exists a set N ⊆ I[0,N −1] of N − Qε > P − 1 +
one used in the preliminary conference version (Müller & Grüne, P (ν + C )/αℓ (ε) time instants such that |(xu (k, x), u(k))|Π > ε for
2015). all k ∈ N . Hence there exists a set N  ⊆ I[0,⌊N /P ⌋−1] of at least
Besides the above dissipativity assumption, we impose two (N − Qε − P + 1)/P > (ν + C )/αℓ (ε) time instants such that
controllability conditions on system (1), namely local controllabil- |(x̃ũ (k, x), ũ(k))|Π > ε for all k ∈ N
. By the assumption of strict
ity in a neighborhood of the periodic orbit Π and finite time con- dissipativity, this implies that
trollability into this neighborhood. /P ⌋−1
⌊N
J̃N (x, u) ≥ αℓ (|(x̃ũ (k, x), ũ(k))|Π )
Assumption 10 (Local Controllability on Bκ (ΠX )). There exists κ > k=0
0, M ′ ∈ I≥0 and ρ ∈ K∞ such that for all z ∈ ΠX and all ν+C
′ > αℓ (ε) = ν + C ,
x, y ∈ Bκ (z ) ∩ X there exists a control sequence u ∈ UM (x) such αℓ (ε)
that xu (M , x) = y and
′
which contradicts (15) and hence proves the theorem. 
|(xu (k, x), u(k))|Π ≤ ρ(max{|x|ΠX , |y|ΠX }) Theorem 12 gives a lower bound Qε for the number of time
holds for all k ∈ I[0,M ′ −1] . 
instants where the considered trajectory is ‘‘close’’ to the periodic
orbit Π . This turnpike result can now be used together with the
controllability conditions specified by Assumptions 10 and 11 to
Assumption 11 (Finite Time Controllability into Bκ (ΠX )). For κ > 0 conclude condition (c) of Theorem 8, as shown in the following.
from Assumption 10 there exists M ′′ ∈ I≥0 such that for each
x ∈ X there exists k ∈ I[0,M ′′ ] and u ∈ Uk (x) such that xu (k, x) ∈ Theorem 13. Suppose that Assumptions 9–11 hold and ℓ is bounded
Bκ (ΠX ).  on X × U. Then condition (c ) of Theorem 8 holds for Y = ΠX .
Proof. From Assumptions 10 and 11, it follows that for each x ∈ X
We are now in a position to state the following theorem which
there exists a control sequence u such that the system is steered
establishes a turnpike property (Dorfman, Samuelson, & Solow,
to a point on ΠX in at most M ′ + M ′′ steps and then stays on the
1987; Trélat & Zuazua, 2015; Zaslavski, 2006) for system (1)
periodic orbit Π for an arbitrary number of time steps. Hence for
with respect to a periodic orbit Π . Turnpike properties with each N ∈ I≥1 we have for some j ∈ I[0,P −1]
respect to an optimal steady-state have recently been studied
N −1
in the context of economic MPC both in discrete-time (Damm, 
Grüne, Stieler, & Worthmann, 2014; Grüne, 2013) and continuous- VN (x) ≤ JN (x, u) = ℓ(xp[k+j] , up[k+j] )
k=0
time (Faulwasser, Korda, Jones, & Bonvin, 2014). The following
result can be seen as a generalization to the case of time-varying min{N ,M ′ +M ′′ }−1

periodic turnpikes. + ℓ(xu (k, x), u(k)) − ℓ(xp[k+j] , up[k+j] ). (16)
k=0

Theorem 12. Suppose that Assumption 9 is satisfied. Then there ex- Now define C ′′ := (P − 1) max(x,u)∈Π ℓ(x, u) and consider the first
ists C > 0 such that for each x ∈ X, each ν > 0, each control sequence sum in (16). We obtain
 P −1 p p
u ∈ UN (x) satisfying JN (x, u) ≤ (N /P ) k=0 ℓ(xk , uk ) + ν , and each N −1 ⌊N /
P ⌋ P −1 N −1
ε > 0 the value Qε := #{k ∈ I[0,N −1] : |(xu (k, x), u(k))|Π ≤ ε}
 
ℓ(xp[k] , up[k] ) = ℓ(xpk , upk ) + ℓ(xp[k] , up[k] )
satisfies the inequality Qε ≥ N − P + 1 − P (ν + C )/αℓ (ε). k=0 k=0 k=⌊N /P ⌋P

P −1
Proof. Let C ′ := 2 supx̃∈XP |λ̃(x̃)|, where x̃ is the state of the P-step N 
 P −1 p p ≤ ℓ(xpk , upk ) + C ′′ .
system (13), and C := C ′ + k=0 ℓ(xk , uk ). Define the rotated cost P k=0
function L̃(x̃, ũ) := ℓ̃(x̃, ũ) − ℓ(xpk , upk ) + λ̃(x̃) − λ̃(f P (x̃, ũ)),
 P −1
k=0 Furthermore, each summand in the second sum of inequal-
and note that from Assumption 9, it follows that L̃(x̃, ũ) ≥ ity (16) can be upper bounded by Ĉ := supx∈X,u∈U ℓ(x, u) −
αℓ (|(x̃, ũ)|Π ). Now for a given x ∈ X and u ∈ UN (x), consider the min(x,u)∈Π ℓ(x, u) < ∞. Hence (16) yields VN (x) ≤ (N /P )
P −1
⌊N /P ⌋−1 k=0
modified cost functional J̃N (x, u) := L̃(x̃ũ (k, x), ũ(k)) + p p
k=0 ℓ( , ) + ν with ν := C ′′ + (M ′ + M ′′ )Ĉ . Now choose N1 :=
xk uk
ℓ( ( , ), ( )) ( ) ( (kP ), . . . , u((k +
N −1
k=⌊N /P ⌋P x u k x u k , where ũ k = u N0 + P and define σ (N ) arbitrary for N ∈ I[0,N1 −1] and σ (N ) :=
1)P − 1)) for all k ∈ I[0,⌊N /P ⌋−1] . From the definition of L̃ and the αℓ−1 (P (ν + C )/(N − N0 − P + 1)) for N ∈ I≥N1 , with C as de-
fact that ⌊N /P ⌋ ≥ N /P − 1, for each control sequence u ∈ UN (x) fined in the proof of Theorem 12. From the above considerations,
134 M.A. Müller, L. Grüne / Automatica 70 (2016) 128–139
it follows that for each x ∈ X and each N ∈ I≥N1 , Theorem 12
can be applied with control sequence u∗N and ε = σ (N ), result-
ing in Qσ (N ) ≥ N − P + 1 − P (ν + C )/αℓ (σ (N )) = N0 . This
means that there are at least N0 time instants k ∈ I[0,N −1] such
that |(xu∗ (k, x), u∗N (k))|Π ≤ σ (N ), and hence also |xu∗ (k, x)|ΠX ≤
N N
σ (N ). As there are at least N0 such time instants k, at least one of
these k must satisfy k ∈ I[0,N −N0 ] , i.e., condition (c) of Theorem 8
holds with kx equal to this k and Y = ΠX . 
Remark 14. Assumption 9 is slightly stronger than the usual defi-
nition of strict dissipativity. Namely, in Assumption 9 ‘‘strictness’’
both with respect to x and u is considered (via the function αℓ in
(14)), while typically this is only required with respect to x. In The-
orem 12, this results in a turnpike property where both the states
and inputs are close to the periodic orbit Π . In fact, the preced-
ing results would still hold in a similar fashion if αℓ (|(x̃, ũ)|Π ) in
(14) was replaced by αℓ (|x̃|ΠX ). In Theorem 12, the definition of Qε
would then need to be slightly changed to Qε := #{k ∈ I[0,N −1] :
|xu (k, x)|ΠX ≤ ε}, which would constitute a more classical turn-
pike property involving only the states (but not the inputs). While
this would still be sufficient for establishing Theorem 13, strict dis-
sipativity as in Assumption 9 (i.e.,using αℓ (|(x̃, ũ)|Π ) in (14)) will
be needed for the results in Section 5.2. 
5.2. Local optimal value function properties
Next, we turn our attention to condition (b) of Theorem 8 and
derive checkable sufficient conditions for it for the case where Π
is a minimal periodic orbit of system (1). In this case, all state and
control sequences satisfying (xu (k, x), u(k)) ∈ Π for k ∈ I[a,b]
with a, b ∈ I≥0 must necessarily follow the unique P-periodic
orbit specified by Π during this time interval,4 i.e., there exists
p p
j ∈ I[0,P −1] such that xu (k, x) = x[k+j] and u(k) = u[k+j] for all
k ∈ I[a,b] . The following auxiliary result shows that also all state
and control sequences staying in a sufficiently small neighborhood
of Π during some time interval must necessarily approximately
follow the unique P-periodic orbit specified by Π during this time
interval.
Lemma 15. Let Π be a minimal P-periodic orbit for system (1), and
assume that the function f in (1) is continuous. Then there exists
ε̄ > 0 such that for all 0 ≤ ε < ε̄ and each state and control
sequence satisfying (xu (k, x), u(k)) ∈ Bε (Π ) for all k ∈ I[a,b] with
a, b ∈ I≥0 , there exists j ∈ I[0,P −1] such that (xu (k, x), u(k)) ∈
p p
Bε ((x[k+j] , u[k+j] )) for all k ∈ I[a,b] .
Proof. See Appendix B. 
With the help of the above, we can now prove the following result.
Theorem 16. Suppose that Assumptions 9 and 10 are satisfied for
some minimal P-periodic orbit Π ⊆ int(X × U) of system (1) and
with M ′ = iP for some i ∈ I≥1 . Furthermore, assume that f and ℓ are
continuous and that the control sequence ux in condition (a) of Theo-
rem 8 is chosen according to (10). Then condition (b) of Theorem 8 is
satisfied for Y = ΠX .
Proof. See Appendix C.  prediction horizon N = 4i for arbitrary i ∈ I≥0 . The (unique) op-
Remark 17. It is worth noting the following interesting and
fundamental difference in the behavior of the optimal state
sequences xu∗ (·, x) in case of optimal periodic operation (as is
N ,x
4 If Π is not minimal, this is not necessarily the case, but different solutions
staying inside Π for all times might exist.
Fig. 2. Illustration of the states x (nodes) and feasible transitions (edges) with
corresponding input u and cost ℓ in Example 18.
considered in this paper) compared to the case of optimal steady-
state operation as in Grüne (2013). Namely, for the latter case it
was established in Grüne (2013) that each optimal state sequence,
which starts close to the optimal steady-state, will stay close to
the optimal steady-state for at least the first N /2 time instants.
This is not necessarily the case anymore when considering periodic
orbits as in this paper, i.e., an optimal state sequence starting
close to the optimal periodic orbit does not necessarily stay close
to the periodic orbit at the beginning (see also Example 18 for
an illustration of this fact). However, as shown in the proof of
Theorem 16, each optimal state sequence will be sufficiently close
to the optimal periodic orbit for at least 2M ′ + 1 consecutive time
instants at some point in the interval [0, N − 1] (but not necessarily
at the beginning), which can be exploited to ensure condition (b)
of Theorem 8. This observation also implies that in general only
convergence of the closed loop to the optimal periodic orbit can be
expected, but not asymptotic stability of the set ΠX . This will be
shown in Section 6. 
Example 18. With this example, we illustrate that even if the
initial condition x is close to ΠX , this does not necessarily imply
that |(xu∗ (k, x), u∗N ,x (k))|Π is small for the first 2M ′ + 1 time
N ,x
instants, but only for 2M ′ + 1 consecutive time instants at some
point in the interval [0, N − 1]. Namely, consider the system
x(k + 1) = u(k) with state and input constraint set Z =
{(1, 2), (2, 3), (2, 4), (4, 2), (3, ε), (ε, 1), (ε, ε) : 0 ≤ ε ≤ ε̄} for
some 0 < ε̄ < 1 and cost function ℓ defined by ℓ(1, 2) = ℓ(2, 3) =
1, ℓ(2, 4) = −10, ℓ(4, 2) = 30, ℓ(3, ε) = ℓ(ε, 1) = 1 + ε and
ℓ(ε, ε) = 2 + ε̄−ε , see also Fig. 2. The system is optimally operated
at the minimal 4-periodic orbit Π = {(0, 1), (1, 2), (2, 3), (3, 0)}
p p
with average cost ℓ0 := k=0 ℓ(xk , uk )/4 = 1. Furthermore, it
3
is straightforward to verify that the Assumptions of Theorem 16
are satisfied5 ; in particular, the corresponding P-step system is
strictly dissipative with respect to Π and the supply rate s(x̃, ũ) =
P −1 p p
ℓ̃(x̃, ũ) − k=0 ℓ(xk , uk ), i.e., (14) is satisfied with αℓ (r ) = r and
storage function

0 0 ≤ x P −1 ≤ 3
λ̃(x̃) = (17)
−20(xP −1 − 3) 3 < xP −1 ≤ 4.
Now consider an initial condition x0 = δ with 0 ≤ δ ≤ ε̄ and
timal input sequence u∗N ,x0 is such that the system stays at the ini-
tial state x0 for one step, then follows the optimal periodic orbit
5 Note that ℓ can easily be extended such that it is continuous on R2 . Furthermore,
Π is not contained in the interior of the constraint set as required in Theorem 16.
However, this assumption is only needed in the proof to ensure that the control
sequence ux defined by (10) is feasible. Here, this is still the case and hence the
conclusions of Theorem 16 are valid.
 M.A. Müller, L. Grüne / Automatica 70 (2016) 128–139
until it moves to x = 4 in the last step, i.e., u∗N ,x0 = (u∗N ,x0 (0),
. . . , u∗N ,x0 (N − 1)) = (δ, 1, 2, 3, 0, 1, . . . , 2, 4) and xu∗ =
(xu∗N ,x (0, x0 ), . . . , xu∗N ,x (N , x0 )) = (δ, δ, 1, 2, 3, 0, 1, . . . , 2, 4).
0 0
However, this means that |(xu∗ (0, x), u∗N ,x (0))|Π = |(−δ, 1 −
N ,x
δ)| ≥ 1/2 independent δ and N, i.e., the optimal solution is
not close to Π for the first time instant. On the other hand,
|(xu∗N ,x (k, x), u∗N ,x (k))|Π = 0 for all k ∈ I[1,N −2] . 
5.3. Closed-loop performance under dissipativity and controllability
Combining all the above, under the assumptions of strict
dissipativity of the P-step system with respect to a periodic
orbit Π , local controllability on a neighborhood of Π and finite
time controllability into this neighborhood of Π , it follows that
the closed-loop asymptotic average performance is near optimal,
i.e., equals the average cost of the periodic orbit Π up to an error
term which vanishes as N → ∞. This is summarized in the
following corollary.
Corollary 19. Consider the P-step MPC scheme as defined via (5) and
suppose that Assumptions 9–11 are satisfied for some minimal P-
periodic orbit Π ⊆ int(X × U) of system (1) and with M ′ = iP
for some i ∈ I≥1 . Furthermore, assume that f and ℓ are continuous
and ℓ is bounded on X × U. Then system (1) is optimally operated
at the periodic orbit Π and there exist δ1 , δ2 ∈ L and N ∈
I≥1 such that for the resulting closed-loop system, the performance
P −1 p p
estimates (6) and (7) with ℓ0 = (1/P ) k=0 ℓ(xk , uk ) are satisfied
for all x ∈ X, all N ∈ I≥N +P and all K ∈ I≥0 . 
Remark 20. Corollary 19 gives closed-loop performance guaran-
tees for the P-step MPC scheme as defined via (5), where P is the
period of the corresponding minimal periodic orbit Π . Neverthe-
less, we note that in certain cases, Theorem 8 can possibly also be
applied with ℓ0 equal to the average cost of the optimal periodic
orbit, but for some value of P which is less than the period of the
optimal periodic orbit. This is illustrated with the following exam-
ple. 
Example 21. Consider the one-dimensional system x(k + 1) =
u(k) with state and input constraint set Z = {(0, 1), (1, 2), (2, 3),
(3, 0)} consisting of four elements only and cost ℓ(x, u) defined as
ℓ(0, 1) = ℓ(2, 3) = 1, ℓ(1, 2) = ℓ(3, 0) = 1 + ε,
for some constant ε > 0. This means that the only feasible state
and input sequence follows the periodic orbit Π = Z for all times.
It is easy to see that the conditions of Theorem 8 can be satisfied
p p
with ℓ0 = k=0 ℓ(xk , uk ) = 1 + ε/2 and P = 2k for all k ∈ I≥1 ,
3
and hence the performance estimates (6) and (7) also hold if, e.g., a
2-step MPC scheme is applied. 
Remark 22. Some comments on implications of the presented
results on practical implementation aspects of the proposed MPC
scheme are in order. First, note that a priori knowledge of the
period length P of the optimal periodic orbit is needed for
implementing the corresponding P-step MPC scheme. On the other
hand, the optimal periodic orbit itself needs not be known for
implementing the proposed MPC method, in contrast to the case
when using additional terminal constraints as in Angeli et al.
(2012) and Zanon et al. (2013). In our setting, the optimal periodic
orbit only needs to be known for a theoretical validation of the
proposed method, in particular for verifying Assumptions 9–11. In
practical applications, the period length P of the optimal periodic
orbit could, e.g., be determined via (offline) simulations of the
system. 
135
6. Convergence to the optimal periodic orbit
N ,x0
In the following, we show that under the same conditions
as needed to establish near optimal performance, the closed-
loop system also asymptotically converges into a neighborhood of
the optimal periodic orbit Π , i.e., the closed-loop system ‘‘finds’’
the optimal operating behavior. This neighborhood can be made
arbitrarily small by choosing N large enough.
Theorem 23. Suppose that the conditions of Corollary 19 are
satisfied and that the storage function λ̃ is continuous on XP . Then
there exists N̂ ∈ I≥1 such that for all N ∈ I≥N̂ , the closed-loop system
resulting from application of the P-step MPC controller (5) practically
asymptotically converges to the optimal periodic orbit Π , i.e., there
exists ν ∈ L and for each x ∈ X some k̂ ∈ I≥0 and j ∈ I[0,P −1] such
p p
that (xuMPC (k, x), uMPC (k)) ∈ Bν(N ) (x[k+j] , u[k+j] ) for all k ∈ I≥k̂ . 
Proof. As shown above, the conditions of Corollary 19 imply that
the conditions of Theorem 8 and hence also those of Proposition 7
are satisfied. Now let δ̂(N ) := δ1 (N − P ) + P δ2 (N − P ) with δ1 and
δ2 from Proposition 7, fix β > 0 and define a := αℓ−1 ((1 + β)δ̂(N ))
with αℓ from (14). Since f and λ̃ are assumed to be continuous,
there exist ηf , ηλ ∈ K∞ such that |f (x, u)−f (x′ , u′ )| ≤ ηf (|(x, u)−
(x′ , u′ )|) for all (x, u), (x′ , u′ ) ∈ Ba (Π ) and |λ̃(x̃) − λ̃(ỹ)| ≤
X := {(xpk , . . . , xpP −1 , xp0 , . . . , xpk−1 ) : k ∈
ηλ (|x̃ − ỹ|) for all x̃ ∈ Π
I[0,P −1] } and all ỹ such that |x̃ − ỹ| ≤ P max{a, ηf (a)}. Next, let
a′ := δ̂(N ) + 2γV (a, N ) + 2ηλ ((P − 1)a + ηf (a)) with γV from
Theorem 8 and a′′ := αℓ−1 (a′ + δ̂(N )). Then, along the closed-loop
sequence, for each i ∈ I≥0 define the values V̂N (xuMPC (iP , x)) :=
VN (xuMPC (iP , x)) + λ̃(x̃(i)) with6 x̃(i) := (xuMPC ((i − 1)P +
1, x), . . . , xuMPC (iP , x)). Furthermore, for each j ∈ I[0,P −1] , let
p p p p p p p
V̂N (xj ) := VN (xj ) + λ̃(x̃j ) with x̃j := (xj+1 , . . . , xP −1 , x0 , . . . , xj ).
p
Finally, choose N̂ ∈ I≥1 large enough such that the following four
conditions are satisfied7 for all N ∈ I≥N̂ : (i) max{a, a′′ } ≤ ε̄ with ε̄
from Lemma 15, (ii) ηf (a) ≤ max{δ̄, ηfP (δ̄)} with δ̄ chosen as in the
p
proof of Theorem 16, (iii) a′ < mini,j∈I p p V̂ (xi ) −
[0,P −1] ,V̂ (x )−V̂ (x )>0
i j
p
V̂ (xj ), and (iv) N̂ ≥ N + P with N from Corollary 20. In the
following, consider an arbitrary N ∈ I≥N̂ .
Applying (6) with K = 1, we obtain that for all i ∈ I≥0
VN (xuMPC ((i + 1)P , x)) − VN (xuMPC (iP , x))
(i+
1)P −1
≤− ℓ(xuMPC (k, x), uMPC (k)) + P ℓ0 + δ̂(N ). (18)
k=iP
Using Assumption 9, this results in
V̂N (xuMPC ((i + 1)P , x)) − V̂N (xuMPC (iP , x))
= VN (xuMPC ((i + 1)P , x)) − VN (xuMPC (iP , x))
+ λ̃(x̃(i + 1)) − λ̃(x(i))
(i+
1)P −1
(18), (14)
≤ − αℓ (|(xuMPC (k, x), uMPC (k))|Π ) + δ̂(N ). (19)
k=iP
Since V̂ is bounded from below (which follows from the fact that
both ℓ and λ̃ are bounded from below), from (19) it follows that for
6 For i = 0, the values x
uMPC (−P + 1, x), . . . , xuMPC (−1, x) can be arbitrary.
7 Note that this is possible since a, a′ , and a′′ are decreasing to zero for
p p
increasing N. Furthermore, condition (iii) does not apply if V̂ (xi ) = V̂ (xj ) for all
i, j ∈ I[0,P −1] .
136 M.A. Müller, L. Grüne / Automatica 70 (2016) 128–139

each initial condition x ∈ X, there exists some i ∈ I≥0 such that Since V̂N (xuMPC ((i + 1)P , x)) − V̂N (xuMPC (iP , x)) ≤ 0 for all i ∈
(i+
1)P −1 I[ī1 ,i2 −1] by definition of ī1 and i2 and V̂N (xuMPC ((i2 + 1)P , x)) −
αℓ (|(xuMPC (k, x), uMPC (k))|Π ) ≤ (1 + β)δ̂(N ); (20) V̂N (xuMPC (i2 P , x)) ≤ δ̂(N ) by (19), from (19) and (25) it follows that
k=iP
(i+
1)P −1
denote the smallest such i by i1 . In particular, from (20) it follows αℓ (|(xuMPC (k, x), uMPC (k))|Π ) ≤ a′ + δ̂(N )
that |(xuMPC (k, x), uMPC (k))|Π ≤ αℓ−1 ((1 + β)δ̂(N )) = a for all k=iP
k ∈ I[i1 P ,(i1 +1)P −1] . for all i ∈ I[ī1 ,i2 −1] , and hence in particular |(xuMPC (k, x), uMPC (k))|Π
Now define ī1 := inf{i ∈ I≥i1 +1 :
≤ αℓ−1 (a′ + δ̂(N )) = a′′ for all k ∈ I[ī1 P ,i2 P −1] .
p
(i+
1)P −1 Now for both the cases where (a) V̂ (xji ) = V̂ (xpji ) and
αℓ (|(xuMPC (k, x), uMPC (k))|Π ) > (1 + β)δ̂(N )}, (21) p p
2 1

k=iP
and consider the following. If ī1 = ∞, we have |(xuMPC (k, x), uMPC
(k))|Π ≤ a for all k ∈ I≥i1 P , and hence the statement of the
theorem with ν(N ) = a and k̂ = i1 P follows from application of
Lemma 15 (considering the fact that a ≤ ε̄ ). On the other hand,
in case that ī1 < ∞, we have |(xuMPC (k, x), uMPC (k))|Π ≤ a for all
k ∈ I[i1 P ,ī1 P −1] . Hence we can again apply Lemma 15 to conclude
that there exists ji1 ∈ I[0,P −1] such that (xuMPC (k, x), uMPC (k)) ∈
p p
Ba (x[k+ji ] , u[k+ji ] ) for all k ∈ I[i1 P ,ī1 P −1] . By continuity of f , it then
1 1
p
follows that |xuMPC (ī1 P , x) − x[ī P +j ] | ≤ ηf (a). Since [i1 P + ji1 ] =
1 i1
[ī1 P + ji1 ] = ji1 and due to the fact that ηf (a) ≤ max{δ̄, ηfP (δ̄)}, the
proof of Theorem 16 shows that
|V (xuMPC (ī1 P , x)) − V (xpji )| ≤ γV (a, N ),
1
and hence, by definition of V̂ ,
(b) V̂ (xji ) < V̂ (xji ), we can go back to (21) and repeat all the
2 1
steps from there (replacing i1 by i2 etc.). Doing this recursively, it
follows that case (b) can occur at most P − 1 times (since Π only
contains P elements). Denote by ir the last value of the sequence
i1 , i2 , . . . obtained in this way such that case (b) occurs, i.e., such
p p
that V̂ (xji ) < V̂ (xji ). Combining the above, for all k ∈ I≥ir we
r r −1
have
|(xuMPC (k, x), uMPC (k))|Π ≤ max{a, a′′ }.
The conclusion of the theorem with k̂ := ir P and ν(N ) :=
max{a, a′′ } then follows from application of Lemma 15. Noting that
ν ∈ L as required then concludes the proof of Theorem 23. 
Remark 24. A close inspection of the proof reveals that it can
(22) be modified such that requiring continuity of λ̃ only on Π X :=
{(xpk , . . . , xpP −1 , xp0 , . . . , xpk−1 ) : k ∈ I[0,P −1] } (instead of XP ) is
enough. 

|V̂ (xuMPC (ī1 P , x)) − V̂ (xpji )| ≤ γV (a, N ) + |λ̃(x̃(ī1 )) − λ̃(x̃pji )| Remark 25. One might wonder why from (19), one cannot use
1 1
standard Lyapunov and invariance arguments which are typically
≤ γV (a, N ) + ηλ ((P − 1)a + ηf (a)). (23) employed when establishing practical asymptotic stability (or
convergence). The reason for this is that while both the inequalities
Furthermore, by the same argument as above, it follows from (19)
(19) and V̂N (xuMPC (iP , x)) ≥ αℓ (|xuMPC (iP , x)|ΠX ) hold, in general
that there exists some i ∈ I≥ī1 such that (20) is satisfied; denote the
smallest such i by i2 . This means that |(xuMPC (k, x), uMPC (k))|Π ≤ a there is no upper bound on V̂ of the form V̂N (xuMPC (iP , x)) ≤
for all k ∈ I[i2 P ,(i2 +1)P −1] . Using again Lemma 15, this implies α2 (|xuMPC (iP , x)|ΠX ) for some α2 ∈ K∞ . This results in the fact that
that there exists ji2 ∈ I[0,P −1] such that (xuMPC (k, x), uMPC (k)) ∈ the closed-loop state and input need not stay in a neighborhood
p p
Ba (x[k+ji ] , u[k+ji ] ) for all k ∈ I[i2 P ,(i2 +1)P −1] and |xuMPC ((i2 + of the optimal periodic orbit once the closed-loop state is close
2
p
2 to ΠX . This was, e.g., the case in Example 18, where we had
1)P , x) − x[ji ] | ≤ ηf (a). Using the same reasoning as in the |(xuMPC (0, x), uMPC (0))|Π ≥ 1/2 for arbitrarily small |x|ΠX . In
2
derivation of (23), we can conclude that the proof of Theorem 23, practical asymptotic convergence of the
closed-loop system could nevertheless be established by showing
|V̂ (xuMPC ((i2 + 1)P , x)) − V̂ (xpji )| that such a situation can only occur a finite number of times. 
2

≤ γV (a, N ) + ηλ ((P − 1)a + ηf (a)). (24) Remark 26. As already mentioned above, in case of optimal
steady-state operation, not only (practical) asymptotic conver-
Combining (23)–(24) with the fact that V̂N (xuMPC ((i + 1)P , x)) −
gence but (practical) asymptotic stability of the optimal steady-
V̂N (xuMPC (iP , x)) ≤ 0 for all i ∈ I[ī1 ,i2 −1] by definition of ī1 and i2 state can be established. In both cases with and without additional
and V̂N (xuMPC ((i2 + 1)P , x)) − V̂N (xuMPC (i2 P , x)) ≤ δ̂(N ) by (19), it terminal constraints, this is typically done by using as a Lyapunov
follows that function the optimal value function of a modified optimization
p p problem, where J̃N (x, u) as defined in the proof of Theorem 12 is
V̂ (xji ) ≤ V̂ (xji ) + δ̂(N ) + 2γV (a, N ) used as objective function. This is possible since the optimal in-
2 1
put sequences resulting from this modified optimization problem
+ 2ηλ ((P − 1)a + ηf (a)) = V̂ (xpji ) + a′ . are the same (when using terminal constraints) or, at least initially,
1
sufficiently close (when using no terminal constraints) to the ones
But then, since N is chosen such that a′ < mini,j∈I p p
[0,P −1] ,V̂ (xi )−V̂ (xj )>0 of the original optimization problem (see, e.g., Amrit et al., 2011;
p p p p
V̂ (xi ) − V̂ (xj ), it follows that V̂ (xji ) ≤ V̂ (xji ). Angeli et al., 2012; Grüne, 2013; Grüne & Stieler, 2014). On the
2 1
p p other hand, in case of optimal periodic operation, this is not neces-
Now in case that V̂ (xji ) = V̂ (xji ), consider the following. sarily the case anymore (such as, e.g., in Example 18), and hence the
2 1
p p
Combining (23)–(24) with the fact that V̂ (xji ) = V̂ (xji ), we obtain optimal value function of the modified problem cannot be used as
2 1 a Lyapunov function. Instead, in the proof of Theorem 23 we em-
V̂ (xuMPC ((i2 + 1)P , x)) − V̂ (xuMPC (ī1 P , x)) ploy V̂N as a Lyapunov-like function in order to establish practi-
cal asymptotic convergence of the closed-loop system, using the
≥ −2γV (a, N ) − 2ηλ ((P − 1)a + ηf (a)) = −a′ + δ̂(N ). (25) more involved argument as discussed in Remark 25. We note that
 M.A. Müller, L. Grüne / Automatica 70 (2016) 128–139
a similar Lyapunov function was also used in Faulwasser and Bon-
vin (2015) considering the case of optimal steady-state operation
in the context of continuous-time systems. 
7. Conclusions
In this paper, we investigated economic MPC without terminal
constraints for the case where the optimal operating behavior
is not stationary but periodic. While near optimal average
performance and convergence to the optimal periodic orbit in
general cannot be achieved for a classical receding horizon control
scheme, this could be established using a P-step MPC scheme
under suitable dissipativity and controllability conditions, with P
being the period length of the optimal periodic orbit. This means
that the only information about the optimal behavior of the system
which is needed a priori (i.e., for implementing the economic MPC
scheme) is the period length P, but the optimal periodic orbit need
not be known. Furthermore, we highlighted our findings as well as
the differences in both the qualitative behavior and the required
analysis methods compared to the case of optimal steady-state
operation with several illustrative examples. Interesting topics for
future research include a suitable definition of transient optimality
and a quantification of the resulting transient performance in the
considered setting of optimal periodic operation. Furthermore, it
would be of interest to examine whether closed-loop guarantees
can be obtained for a classical one-step MPC scheme under
additional conditions. Finally, extending our results to the general
case of time-varying systems, cost functions, and constraint sets
would be desirable.
Appendix A. Proof of Theorem 8
Choose N ∈ I≥N1 such that σ (N ) ≤ δ̄ holds with σ from
condition (c) of the theorem. Fix N ≥ N and consider an arbitrary
x ∈ X together with a corresponding optimal control sequence
u∗N ,x ∈ UN (x). Let kx be the time index from (c), abbreviate x′ :=
xu∗ (kx , x) and denote by ux′ ∈ UP the control sequence from
N ,x
condition (a) with x = x′ . Let x′′ := xux′ (P , x′ ) and let u∗N −k ,x′′ be
x ≤ ηf (|(xu (k, x), u(k)) − (xp[k+j] , up[k+j] )|) ≤ ηf (ε) <
an optimal control sequence for the initial condition x = x′′ and
horizon N − kx . Using the above, we define the control sequence
ūN ,x ∈ UN +P (x) by ūN ,x (k) := u∗N ,x (k) for k ∈ I[0,kx −1] , ūN ,x (k) :=
ux′ (k − kx ) for k ∈ I[kx ,kx +P −1] and ūN ,x (k) := u∗N −k ,x′′ (k − kx − P )
x
for k ∈ I[kx +P ,N +P −1] . This means that xūN ,x (k, x) = xu∗ (k, x) for
N ,x
k ∈ I[0,kx −1] and xūN ,x (k, x) = xux′ (k − kx , x′ ) for k ∈ I[kx ,kx +P −1] .
Furthermore, by condition (c) we have |x′ |Y ≤ σ (N ), and from
condition (a) it follows that
k +P −1
1 x
ℓ(xūN ,x (k, x), ūN ,x (k))
P k=kx
P −1
1
= ℓ(xux′ (i, x′ ), ux′ (i)) ≤ ℓ0 + γℓ (σ (N )).
P i =0
Moreover, condition (b) of the theorem implies that for all K ∈
I≥N0 , we have
VK (x′′ ) = VK (xux′ (P , x′ )) ≤ VK (x′ ) + γV (σ (N ), N ).
Now distinguish two cases. First, in case that N − kx ≥ 1, since
N − kx ∈ I≥N0 by condition (c) we can use (A.2) with K = N − kx
to conclude that
N +P −1

ℓ(xūN ,x (k, x), ūN ,x (k)) = JN −kx (x′′ , u∗N −kx ,x′′ )
k=kx +P
(A.2)
= VN −kx (x′′ ) ≤ VN −kx (x′ ) + γV (σ (N ), N ).
137
Setting kiN ,x := kx + i − 1 in Proposition 7 for i ∈ I[1,P ] , we obtain
kx −1 N +P −1
 
JN′ (x) = ℓ(xu∗N ,x (k, x), u∗N ,x (k)) + ℓ(xūN ,x (k, x), ūN ,x (k))
k=0 k=kx +P
≤ VN (x) − VN −kx (x′ ) + VN −kx (x′ ) + γV (σ (N ), N )
= VN (x) + γV (σ (N ), N ),
where the above inequality follows from the dynamic program-
ming principle and (A.3). Hence condition (i) of Proposition 7 is sat-
isfied with δ1 (N ) = γV (σ (N ), N ); note that δ1 ∈ L as required due
to the fact that σ ∈ L and γV ∈ KL. Second, if N − kx = 0, then
J ′ (x) = VN (x) and hence condition (i) of Proposition 7 is satisfied
for arbitrary δ1 (N ). Finally, by (A.1) we have that condition (ii) of
Proposition 7 is satisfied with δ2 (N ) = γℓ (σ (N )), which concludes
the proof of Theorem 8. 
Appendix B. Proof of Lemma 15
Let dmin := minx,y∈ΠX |x − y| and note that dmin > 0 as
Π is a minimal P-periodic orbit. As discussed previously, from
the fact that f is continuous and Bdmin /2 (Π ) is compact, it follows
that there exists ηf ∈ K∞ such that |f (x, u) − f (x′ , u′ )| ≤
ηf (|(x, u) − (x′ , u′ )|) for all (x, u), (x′ , u′ ) ∈ Bdmin /2 (Π ). Now
define ε̄ := min{ηf−1 (dmin /2), dmin /2}, and consider an arbitrary
state and control sequence pair satisfying (xu (k, x), u(k)) ∈ Bε (Π )
for all k ∈ I[a,b] with a, b ∈ I≥0 and some 0 ≤ ε < ε̄ .
In the following, we consider the case b ≥ a, as otherwise
there is nothing to show, and establish the statement of the
lemma by induction. Since (xu (a, x), u(a)) ∈ Bε (Π ), we have
(xu (a, x), u(a)) ∈ Bε ((xp[a+j] , up[a+j] )) for some j ∈ I[0,P −1] . Now
p p
assume that (xu (k, x), u(k)) ∈ Bε ((x[k+j] , u[k+j] )) for some j ∈
I[0,P −1] and some k ∈ I[a,b−1] . Considering the above, we obtain
|xu (k + 1, x) − xp[k+1+j] |
= |f (xu (k, x), u(k)) − f (xp[k+j] , up[k+j] )|
dmin
.
2
But then, from the fact that (xu (k + 1, x), u(k + 1)) ∈ Bε (Π ), the
definition of dmin , and the fact that ε < ε̄ ≤ dmin /2, it follows that
p
xu (k + 1, x) ̸∈ Bε (xi ) for all i ∈ I[0,P −1] with i ̸= [k + 1 + j]. Hence
p p
necessarily (xu (k + 1, x), u(k + 1)) ∈ Bε ((x[k+1+j] , u[k+1+j] )), which
closes the proof of the lemma. 
Appendix C. Proof of Theorem 16
As discussed previously, for each x ∈ Bδ (ΠX ) for some δ ∈
(0, δ̄], by definition of ΠX it holds that x ∈ Bδ (xpj ) for some j ∈
p
I[0,P −1] . Furthermore, (11) yields that |xux (P , x) − x[j+P ] | ≤ ηfP (δ),
p
where ux is the control sequence defined by (10). As x[j+P ] = xj ,
p
(A.1)
this implies that both x and xux (P , x) are contained in the set
p
Bmax{δ,ηP (δ)} (xj ). Hence a sufficient condition for condition (b) of
f
Theorem 8 to be satisfied is that the inequality |VN (x) − VN (x′ )| ≤
γV (δ, N ) holds for all N ∈ I≥N0 , all y ∈ ΠX , all δ ∈ (0, δ̄], and all
(A.2) x, x′ ∈ Bmax{δ,ηP (δ)} (y). This will be shown in the following.
f
Choose δ̄ small enough such that Bmax{δ̄,ηP (δ̄)} (ΠX ) ⊆ X, such
f
that max{δ̄, ηfP (δ̄)} ≤ κ and such that ρ(max{δ̄, ηfP (δ̄)}) ≤ ε̄
with κ and ρ from Assumption 10 and ε̄ from Lemma 15. Now
consider arbitrary y ∈ ΠX and x ∈ Bmax{δ̄,ηP (δ̄)} (y). For each such x,
f
by Assumption 10 there exists a control sequence u such that the
system is steered to a point on ΠX in M ′ steps and then stays on
(A.3) the periodic orbit Π for an arbitrary number of time steps. Using
138 M.A. Müller, L. Grüne / Automatica 70 (2016) 128–139

the same argument as in the proof of Theorem 13, this results in

the fact that for each N ∈ I≥1 , we have VN (x) ≤ JN (x, u) ≤
(N /P ) Pk=−01 ℓ(xpk , upk ) + a with a := C ′′ + M ′ ηℓ (ρ(max{δ̄, ηfP (δ̄)}))


and C ′′
:= (P − 1) max(x,u)∈Π ℓ(x, u). Now choose N0 ∈
I≥P (2M ′ +1)+1 large enough such that αℓ−1 ((a + C )(2M ′ + 1)P /(N0 −
P (2M ′ + 1))) ≤ min{ε̄, κ} and such that ρ(αℓ−1 ((a + C )(2M ′ +
Fig. C.1. Exemplary illustration of state sequences xu∗ (red solid) and xū (blue
1)P /(N0 − P (2M ′ + 1)))) ≤ ε̄ , with ε̄ from Lemma 15 and κ N ,x
dashed) in the proof of Theorem 16.
and ρ from Assumption 10. In the following, consider an arbitrary
N ∈ I≥N0 . The above inequality for VN (x) implies that we can apply
construction of ū, we obtain xū (k, x′ ) = xu1 (k, x′ ) for k ∈ I[0,M ′ −1]
Theorem 12 with ν = a and ε = αℓ−1 ((a + C )(2M ′ + 1)P /(N − and xū (M ′ , x′ ) = x, xū (k, x′ ) = xu∗ (k − M ′ , x) for k ∈ I[M ′ ,k′ +M ′ −1] ,
P (2M ′ + 1))) to conclude that |(xu∗ (k, x), u∗N ,x (k))|Π ≤ ε for N ,x
N ,x xū (k, x′ ) = xu2 (k − k′ − M ′ , xu∗ (k′ , x)) for k ∈ I[k′ +M ′ ,k′ +2M ′ −1] ,
Qε ≥ N − P + 1 − P (ν + C )/αℓ (ε) = N + 1 − N /(2M + 1) ′ N ,x

time instants k ∈ I[0,N −1] , and thus |(xu∗ (k, x), u∗N ,x (k))|Π > ε and xū (k, x′ ) = xu∗ (k, x) for k ∈ I[k′ +2M ′ ,N −1] , see also Fig. C.1.
N ,x
N ,x
This yields
for at most ⌊N /(2M ′ + 1) − 1⌋ time instants k ∈ I[0,N −1] . But this
′
implies that there are at least 2M + 1 consecutive time instants

(C.2), (C.3)
VN (x′ ) ≤ JN (x′ , ū) iP ηℓ (ρ(δ̂)) + ηℓ (ρ(ε))
 
k ∈ I[0,N −1] such that |(xu∗ (k, x), u∗N ,x (k))|Π ≤ ε ; denote these ≤
N ,x
time instants by k′ , . . . , k′ + 2M ′ (compare Fig. C.1).
P −1 k′ −1
By the above choice of N0 and the fact that N ∈ I≥N0 , Lemma 15  
+ 2i ℓ(xpk , upk ) + ℓ(xu∗N ,x (k, x), u∗N ,x (k))
can now be used with ε = αℓ−1 ((a + C )(2M ′ + 1)P /(N − P (2M ′ +
k=0 k=0
1))) to conclude from the above that there exists j ∈ I[0,P −1] such
p p N −1
that (xu∗ (k, x), u∗N ,x (k)) ∈ Bε ((x[k+j] , u[k+j] )) for all k ∈ I[k′ ,k′ +2M ′ ] .  
N ,x + ℓ(xu∗N ,x (k, x), u∗N ,x (k))
By continuity of ℓ and the fact that M ′ = iP for some i ∈ I≥1 , this k=k′ +2M ′
implies that
(C.1)
≤ VN (x) + iP ηℓ (ρ(δ̂)) + ηℓ (ρ(ε)) + 2ηℓ (ε) .
 
k′ +2M ′ −1

ℓ(xu∗N ,x (k, x), u∗N ,x (k))
Defining γV (δ, N ) := iP (ηℓ (ρ(δ̂)) + ηℓ (ρ(ε)) + 2ηℓ (ε)) with
k=k′
P −1 δ̂ = max{δ, ηfP (δ)} and ε = αℓ−1 ((a + C )(2M ′ + 1)/(N − 2M ′ − 1))
results in VN (x′ ) ≤ VN (x) + γV (δ, N ). Exchanging x and x′ yields

≥ −2iP ηℓ (ε) + 2i ℓ(xpk , upk ). (C.1)
k =0 the converse inequality VN (x′ ) ≥ VN (x) − γV (δ, N ) and hence
Furthermore, again due to the fact that M ′ = iP for some i ∈
|VN (x′ ) − VN (x)| ≤ γV (δ, N ). Noting that γV (δ, N ) ∈ KL as
required then concludes the proof of Theorem 16. 
I≥1 , there exists j ∈ I[0,P −1] such that (xu∗N ,x (k, x), u∗N ,x (k)) ∈
p p
Bε ((xj , uj )) for k ∈ {k′ , k′ + M ′ , k′ + 2M ′ }, and hence also References
p
xu∗ (k, x) ∈ Bε (xj ) for k ∈ {k′ , k′ + M ′ , k′ + 2M ′ }. We can then
N ,x
Amrit, R., Rawlings, J. B., & Angeli, D. (2011). Economic optimization using model
use Assumption 10 to conclude that there exists a control sequence
′ predictive control with a terminal cost. Annual Reviews in Control, 35(2),
u2 ∈ UM such that xu2 (M ′ , xu∗ (k′ , x)) = xu∗ (k′ + 2M ′ , x) and 178–186.
N ,x N ,x
Amrit, Rishi, Rawlings, James B., & Biegler, Lorenz T. (2013). Optimizing process
|(xu2 (k, xu∗N ,x (k′ , x)), u2 (k))|Π ≤ ρ(ε) for all k ∈ I[0,M ′ −1] . By choice economics online using model predictive control. Computers & Chemical
of N0 and the fact that N ∈ I≥N0 , we have ρ(ε) ≤ ε̄ . Hence we can Engineering, 58, 334–343.
Angeli, D., Amrit, R., & Rawlings, J. B. (2012). On average performance and stability
again apply Lemma 15 to conclude that there exists j ∈ I[0,P −1] of economic model predictive control. IEEE Transactions on Automatic Control,
p p
such that (xu2 (k, xu∗ (k′ , x)), u2 (k)) ∈ Bρ(ε) ((x[k+j] , u[k+j] )) for all 57(7), 1615–1626.
N ,x
Chu, B., Duncan, S., Papachristodoulou, A., & Hepburn, C. (2012). Using economic
k ∈ I[0,M ′ −1] . Then, using again continuity of ℓ and the fact that model predictive control to design sustainable policies for mitigating climate
M ′ = iP for some i ∈ I≥1 , it follows that change. In Proceedings of the 51st IEEE conference on decision and control
(pp. 406–411).
M ′ −1
 Damm, T., Grüne, L., Stieler, M., & Worthmann, K. (2014). An exponential turnpike
ℓ(xu2 (k, xu∗N ,x (k′ , x)), u2 (k)) theorem for dissipative discrete time optimal control problems. SIAM Journal
k=0
on Control and Optimization, 52(3), 1935–1957.
Dorfman, R., Samuelson, P. A., & Solow, R. M. (1987). Linear programming and
P −1
 economic analysis. New York: Dover Publications. Reprint of the 1958 original.
≤ iP ηℓ (ρ(ε)) + i ℓ(xpk , upk ). (C.2) Ellis, M., Durand, H., & Christofides, P. D. (2014). A tutorial review of economic
model predictive control methods. Journal of Process Control, 24(8), 1156–1178.
k=0
Fagiano, L., & Teel, A. R. (2013). Generalized terminal state constraint for model
Now for given δ ∈ (0, δ̄] and x ∈ Bδ̂ (y) with y ∈ ΠX and δ̂ := predictive control. Automatica, 49(9), 2622–2631.
Faulwasser, T., & Bonvin, D. (2015). On the design of economic NMPC based on
max{δ, ηfP (δ)}, consider an arbitrary x′ ∈ Bδ̂ (y). By Assumption 10, approximate turnpike properties. In Proceedings of the 54th IEEE conference on
there exists a control sequence u1 such that xu1 (M ′ , x′ ) = x and decision and control (pp. 4964–4970).
Faulwasser, T., Korda, M., Jones, C.N., & Bonvin, D. (2014). Turnpike and dissipativity
|(xu1 (k, x′ ), u1 (k))|Π ≤ ρ(δ̂) for all k ∈ I[0,M ′ −1] . As above, we can properties in dynamic real-time optimization and economic MPC. In Proceedings
use Lemma 15 as well as continuity of ℓ and the fact that M ′ = iP of the 53rd IEEE conference on decision and control (pp. 2734–2739).
Gros, S. (2013). An economic NMPC formulation for wind turbine control. In
for some i ∈ I≥1 to conclude that Proceedings of the 52nd IEEE conference on decision and control (pp. 1001–1006).
Grüne, L. (2013). Economic receding horizon control without terminal constraints.
M ′ −1
 P −1
 Automatica, 49(3), 725–734.
ℓ(xu1 (k, x′ ), u1 (k)) ≤ iP ηℓ (ρ(δ̂)) + i ℓ(xpk , upk ). (C.3) Grüne, L., & Palma, V. G. (2015). Robustness of performance and stability for
k=0 k=0
multistep and updated multistep MPC schemes. Discrete and Continuous
Dynamical Systems, 35(9), 4385–4414.
Combining the above, we now define the following control Grüne, L., & Pannek, J. (2011). Nonlinear model predictive control. Theory and
sequence ū ∈ UN via ū(k) = u1 (k) for k ∈ I[0,M ′ −1] , ū(k) = algorithms. London: Springer.
Grüne, L., & Stieler, M. (2014). Asymptotic stability and transient optimality of
u∗N ,x (k − M ′ ) for k ∈ I[M ′ ,k′ +M ′ −1] , ū(k) = u2 (k − k′ − M ′ ) for economic MPC without terminal constraints. Journal of Process Control, 24(8),
k ∈ I[k′ +M ′ ,k′ +2M ′ −1] , and ū(k) = u∗N ,x (k) for k ∈ I[k′ +2M ′ ,N −1] . By 1187–1196.
 M.A. Müller, L. Grüne / Automatica 70 (2016) 128–139
Grüne, L., & Zanon, M. (2014). Periodic optimal control, dissipativity and MPC.
In Proceedings of the 21st international symposium on mathematical theory of
networks and systems (pp. 1804–1807).
Heidarinejad, M., Liu, J., & Christofides, P. D. (2012). Economic model predictive
control of nonlinear process systems using Lyapunov techniques. AIChE Journal,
58(3), 855–870.
Hovgaard, T. G., Larsen, L. F. S., Edlund, K., & Jørgensen, J. B. (2012). Model
predictive control technologies for efficient and flexible power consumption
in refrigeration systems. Energy, 44(1), 105–116.
Huang, R., Harinath, E., & Biegler, L. T. (2011). Lyapunov stability of economically
oriented NMPC for cyclic processes. Journal of Process Control, 21(4), 501–509.
Limon, D., Pereira, M., Muñoz de la Peña, D., Alamo, T., & Grosso, J. M. (2014).
Single-layer economic model predictive control for periodic operation. Journal
of Process Control, 24(8), 1207–1224.
Mendoza-Serrano, David I., & Chmielewski, Donald J. (2014). Smart grid coordina-
tion in building HVAC systems: EMPC and the impact of forecasting. Journal of
Process Control, 24(8), 1301–1310.
Müller, M. A., Angeli, D., & Allgöwer, F. (2013). Economic model predictive control
with self-tuning terminal cost. European Journal of Control, 19(5), 408–416.
Müller, M. A., Angeli, D., & Allgöwer, F. (2015). On necessity and robustness
of dissipativity in economic model predictive control. IEEE Transactions on
Automatic Control, 60(6), 1671–1676.
Müller, M.A., & Grüne, L. (2015). Economic model predictive control without
terminal constraints: optimal periodic operation. In Proceedings of the 54th IEEE
conference on decision and control (pp. 4946–4951).
Müller, M.A., Grüne, L., & Allgöwer, F. (2015). On the role of dissipativity in economic
model predictive control. In Proceedings of the 5th IFAC conference on nonlinear
model predictive control (pp. 110–116).
Trélat, E., & Zuazua, E. (2015). The turnpike property in finite-dimensional nonlinear
optimal control. Journal of Differential Equations, 258(1), 81–114.
Zanon, M., Gros, S., & Diehl, M. (2013). A Lyapunov function for periodic economic
optimizing model predictive control. In Proceedings of the 52nd IEEE conference
on decision and control (pp. 5107–5112).
Zaslavski, A. J. (2006). Turnpike properties in the calculus of variations and optimal
control. New York: Springer.
139
Matthias A. Müller received a diploma degree in en-
gineering cybernetics from the University of Stuttgart,
Germany, and an M.S. in electrical and computer engineer-
ing from the University of Illinois at Urbana-Champaign,
US, both in 2009. In 2014, he obtained a Ph.D. in me-
chanical engineering, also from the University of Stuttgart,
Germany. He is currently working as a lecturer and post-
doctoral researcher (Akademischer Rat) at the Institute for
Systems Theory and Automatic Control at the University of
Stuttgart, Germany. In 2012, he was a visiting researcher
at the Imperial College, London, UK. He is a member of the
Eliteprogram for PostDocs of the Baden-Württemberg Foundation and was a semi-
plenary speaker at the 5th IFAC Conference on Nonlinear Model Predictive Control
2015. His research interests include nonlinear control theory, model predictive con-
trol, distributed control and switched systems.
Lars Grüne has been a professor for applied mathematics
at the University of Bayreuth, Germany, since 2002. He
received his diploma and Ph.D. in mathematics in 1994
and 1996, respectively, from the University of Augsburg
and his habilitation from the J.W. Goethe University in
Frankfurt/M in 2001. He held visiting positions at the
Universities of Rome ‘Sapienza’ (Italy), Padova (Italy),
Melbourne (Australia), Paris IX — Dauphine (France) and
Newcastle (Australia). Prof. Grüne is the editor-in-chief of
the journal Mathematics of Control, Signals and Systems
(MCSS), an associate editor of the Journal of Optimization
Theory and Applications (JOTA) and of the Journal of Applied Mathematics and
Mechanics (ZAMM), and a member of the Managing Board of the International
Association of Applied Mathematics and Mechanics (GAMM). His research interests
lie in the area of mathematical systems and control theory with a focus on numerical
and optimization-based methods for nonlinear systems.