Professional Documents
Culture Documents
Linear Tracking MPC For Nonlinear Systems Part I - The Model-Based Case
Linear Tracking MPC For Nonlinear Systems Part I - The Model-Based Case
Linear Tracking MPC For Nonlinear Systems Part I - The Model-Based Case
Abstract—We develop a tracking model predictive control In this paper, we propose a tracking MPC scheme for
(MPC) scheme for nonlinear systems using the linearized dynam- nonlinear systems using the linearized dynamics at the current
arXiv:2105.08560v3 [math.OC] 14 Apr 2022
ics at the current state as a prediction model. Under reasonable state as a prediction model. We consider a tracking MPC
assumptions on the linearized dynamics, we prove that the pro-
posed MPC scheme exponentially stabilizes the optimal reachable formulation to steer the system to a desired target setpoint,
equilibrium w.r.t. a desired target setpoint. Our theoretical results which may potentially change online and can be unreachable
rely on the fact that, close to the steady-state manifold, the pre- by the system dynamics and constraints, similar to [7], [8].
diction error of the linearization is small and hence, we can slide The cost function contains a tracking cost w.r.t. an artificial
along the steady-state manifold towards the optimal reachable equilibrium as well as a penalty of the deviation between this
equilibrium. The closed-loop stability properties mainly depend
on a cost matrix which allows us to trade off performance, ro- artificial equilibrium and the desired target setpoint. We prove
bustness, and the size of the region of attraction. In an application that, if the weight matrix of this penalty is chosen sufficiently
to a nonlinear continuous stirred tank reactor, we show that the small, then the linearization-based prediction model is suf-
scheme, which only requires solving a convex quadratic program ficiently accurate such that the closed loop converges to the
online, has comparable performance to a nonlinear MPC scheme optimal reachable equilibrium. The latter is the closest possible
while being computationally significantly more efficient. Further,
our results provide the basis for controlling nonlinear systems equilibrium to the target setpoint, in case this setpoint is not
based on data-dependent linear prediction models, which we an equilibrium of the system and / or does not satisfy the
explore in our companion paper. constraints. Figure 1 illustrates the main idea: At time t = 0,
the open-loop predictions reach an (artificial) equilibrium
output y s∗ (0) for the linearized system dynamics lying in
I. I NTRODUCTION a neighborhood of the nonlinear equilibrium manifold Zys .
Model predictive control (MPC) is a successful modern At time t, after repeated application of the proposed MPC
control technique, mainly due to its ability to cope with scheme, the artificial equilibrium y s∗ (t) is closer to y sr , the
nonlinear dynamics, hard input and state constraints, and optimal reachable equilibrium given the (potentially unreach-
dynamic reference trajectories [1]. The key idea of MPC able) setpoint reference y r . The model used for prediction at
centers around the repeated solution of an open-loop optimal time t relies on the linearized system dynamics at the current
control problem. While its applicability to nonlinear systems state xt . By repeated application of the proposed MPC scheme,
is one of the main advantages of MPC, this requires solving the artificial equilibrium y s∗ (t) slides along the manifold Zys
a non-convex optimization problem online. In order to cope towards y sr and eventually, the closed-loop output converges
with limited computational resources, usually only a limited to y sr .
number of iterations are performed [2], [3], [4], [5]. Further, y0
yt
most nonlinear MPC schemes require a (globally) accurate ...
model of the underlying dynamical system and, if this is not
y s∗ (0)
available, techniques such as robust nonlinear MPC need to y s∗ (t)
be employed, possibly further increasing the computational y sr
complexity [6]. Zys
yr
Funded by Deutsche Forschungsgemeinschaft (DFG, German Research
Foundation) under Germany’s Excellence Strategy - EXC 2075 - 390740016
and under grant 468094890. We acknowledge the support by the Stuttgart Fig. 1. Scheme illustrating the basic idea of our approach. The figure displays
Center for Simulation Science (SimTech). This project has received funding the output equilibrium manifold Zys , the closed-loop output and artificial
from the European Research Council (ERC) under the European Union’s equilibrium at time 0, the closed-loop output and artificial equilibrium at
Horizon 2020 research and innovation programme (grant agreement No time t, the optimal reachable equilibrium y sr , and the setpoint reference y r .
948679). The authors thank the International Max Planck Research School
for Intelligent Systems (IMPRS-IS) for supporting Julian Berberich.
The main contribution of this paper is to prove that the
1 University of Stuttgart, Institute for Systems Theory and proposed MPC scheme exponentially stabilizes y sr under
Automatic Control, 70550 Stuttgart, Germany (email:{julian.berberich, reasonable assumptions on the linearized dynamics. A key
frank.allgower}@ist.uni-stuttgart.de) advantage of our scheme if compared to nonlinear tracking
2 Institute for Dynamical Systems and Control, ETH Zurich, ZH-8092,
Switzerland (email:jkoehle@ethz.ch) MPC schemes [7], [8] lies in its computational efficiency,
3 Leibniz University Hannover, Institute of Automatic Control, 30167 since only a convex quadratic program (QP) needs to be
Hannover, Germany (e-mail:mueller@irt.uni-hannover.de) solved online. Moreover, the implementation only requires an
accurate linear model around the steady-state manifold, which
©2021 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media,
including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers
or lists, or reuse of any copyrighted component of this work in other works.
2
is easier to obtain than a complex nonlinear model. Finally, for some h0 : Rn → Rp , D ∈ Rp×m . We impose pointwise-
this paper provides the basis for the results in our companion in-time constraints on the input, i.e., ut ∈ U for all t ∈ I≥0 ,
paper [9] where we merge the idea of linearization-based where the constraint set U is assumed to be compact. Due to
tracking MPC with recent work on data-driven MPC [10], [11] the inexact prediction model, including state constraints into
to develop a data-driven MPC scheme for unknown nonlinear our framework would necessitate additional robust constraint
systems with closed-loop stability guarantees. tightening methods, which is an interesting issue for future
The real-time iteration scheme [2], [5] relies on a sequential research.
approximation of the optimal control problem, resulting in We propose a state-feedback MPC scheme to track a desired
predictions based on linear time-varying (LTV) models, and setpoint reference y r ∈ Rp with the nonlinear system (3)–
is thus conceptually similar to our work. An LTV prediction (4). In contrast to existing works with this goal such as [7],
model is also employed in [12], where a tube around the the prediction model relies on the dynamics linearized at the
previous candidate solution is used to reduce the linearization current state xt . In the remainder of this section, we state key
error. The relation of our approach to existing LTV-based assumptions that are required to prove that the MPC scheme
MPC methods for nonlinear systems will be discussed in more asymptotically tracks the desired output setpoint.
detail later in the paper. Further related approaches are [13],
[14] which estimate (approximate) LTV models of nonlinear
systems from data and employ these in MPC schemes. While A. Smoothness assumptions
these approaches lead to convex optimization problems which We assume that both f and h are continuously differentiable
can be efficiently solved, they do not provide guarantees and hence, for some x̃ ∈ Rn , we can define
on closed-loop stability. Another line of research to control
nonlinear systems via MPC with linear prediction models ∂f0
Ax̃ := , ex̃ := f0 (x̃) − Ax̃ x̃,
relies on Koopman operator theory [15]. The key differences ∂x x̃ (5)
to our approach are that we exploit local linearity properties ∂h0
Cx̃ := , rx̃ := h0 (x̃) − Cx̃ x̃.
whereas the (infinite-dimensional) Koopman operator provides ∂x x̃
a globally valid linear model, and that we provide desirable
We write fx̃ (x, u) := Ax̃ x + Bu + ex̃ for the system dynamics
closed-loop stability guarantees.
linearized at (x, u) = (x̃, 0), and hx̃ (x, u) := Cx̃ x + Du + rx̃
The paper is structured as follows. In Section II, we state
for the output linearized at (x, u) = (x̃, 0). Note that, since
the problem setting and the key assumptions required for our
System (3) is control-affine, linearizing the dynamics at time
theoretical results. Next, we present our linearization-based
t only requires the state xt and no knowledge of the input
MPC scheme for nonlinear systems in Section III, and we
ut , which is a crucial fact for the proposed MPC scheme and
prove theoretical properties such as recursive feasibility and
its theoretical analysis. If f or h are not affine in u, then it
closed-loop exponential stability. In Section IV, we apply the
can be readily enforced by defining a new, incremental input
scheme to a nonlinear continuous stirred tank reactor (CSTR).
∆uk := uk+1 − auk for some a ∈ R with |a| ≤ 1 (e.g., a = 1
Finally, Section V contains a summary and conclusion.
corresponds to a standard incremental input). In this case, the
Notation: Denote the set of all nonnegative integers by I≥0
exact “prediction model” uk+1 = auk + ∆uk still allows us
and the integers in the interval [a, b] by I[a,b] . We write k·k2 for
to enforce hard constraints on u as well as ∆u while ensuring
the 2-norm of a vector, or the induced 2-norm if the argument
that the system is control-affine. The key idea of this paper
is a matrix. Moreover, for some positive definite P = P > 0,
relies on the fact that a nonlinear system can be approximated
we define kxk2P := x> P x. The point-to-set distance w.r.t. a set
locally by its linearization, given that f is suitably smooth.
Z is defined as kxkZ := inf x0 ∈Z kx0 − xk2 . We write λmin (P )
Note that, by definition, it holds that
(λmax (P )) for the minimum (maximum) eigenvalue of P , and
we define λmin (P1 , P2 ) := min{λmin (P1 ), λmin (P2 )} for two fx (x, u) = f (x, u) and hx (x, u) = h(x, u). (6)
symmetric matrices P1 , P2 , and similarly for λmax (P1 , P2 ).
Finally, σmin (A) denotes the minimum singular value of a That is, locally at the linearization point, the linearization is
matrix A. Throughout this paper, we use the inequalities equal to the nonlinear function f . This insight is important
ka + bk2P ≤ 2kak2P + 2kbk2P , (1) for the theoretical analysis provided in the remainder of this
paper since it implies that the prediction error of the proposed
kak2P − kbk2P ≤ ka − bk2P + 2ka − bkP kbkP , (2)
MPC scheme, which uses a local linearization-based prediction
which hold for any vectors a, b and matrix P = P > 0. model, is zero in the first time step. Throughout this paper, we
assume that all vector fields involved in the system dynamics
II. P ROBLEM SETUP are twice continuously differentiable.
In this paper, we consider discrete-time nonlinear systems
Assumption 1. (Smoothness) The vector fields f0 and h0 are
of the form
twice continuously differentiable.
xt+1 = f (xt , ut ) = f0 (xt ) + But (3)
Assumption 1 implies a useful quantitative bound on the
with f0 : Rn → Rn , B ∈ Rn×m , and output difference between the nonlinear vector fields f (x, u), h(x, u)
yt = h(xt , ut ) = h0 (xt ) + Dut (4) and their linearizations fx̃ (x, u), hx̃ (x, u) for x 6= x̃.
3
Proposition 1. If Assumption 1 holds, then for any compact this minimizer is unique. Let us define the set of equilibria of
set X ⊂ Rn , there exist cX , cXh > 0 such that for any x, x̃ ∈ the linearized system at some state x̃ as
X, u ∈ U, it holds that s
(x̃) := (xs , us ) ∈ Rn × Us | xs = Ax̃ xs + Bus + ex̃ }
ZLin
kf (x, u) − fx̃ (x, u)k2 ≤ cX kx − x̃k22 , (7)
and the projection on the output as
kh(x, u) − hx̃ (x, u)k2 ≤ cXh kx − x̃k22 . (8)
s
(x̃) := y s ∈ Rp | ∃(xs , us ) ∈ ZLin
s
Zy,Lin (x̃) :
Proof. This follows directly from Taylor’s Theorem in the
y s = Cx̃ xs + Dus + rx̃ .
multivariable case, together with compactness of X and U.
The optimal reachable equilibrium of the linearized system at
For Proposition 1, it is crucial to consider compact sets
x̃ is the minimizer of
X, U, since the error bound is in general not necessarily
∗
satisfied globally. We later show that a certain (compact) Jeq,Lin (x̃) := min
s
ky s − y r k2S . (13)
y s ∈Zy,Lin (x̃)
Lyapunov function sublevel set is invariant and can hence be
used to define X. Although using uniform constants cX , cXh We denote the minimizer of (13) by yLin sr
(x̃). It follows from
in (7) and (8) over the set X may be conservative, the error S 0 that Problem (13) is strongly convex at any linearization
bound still becomes arbitrarily small if x and x̃ are sufficiently point, i.e., for any x̃ ∈ Rn , y s ∈ Zy,Lin
s
(x̃), it holds that
close to each other. Furthermore, Assumption 1 implies that
∗
f0 and h0 are locally Lipschitz continuous in x, i.e., for any ky s − y r k2S − Jeq,Lin (x̃) ≥ ky s − yLin
sr
(x̃)k2S , (14)
cLip > 0 and any compact set X there exists a constant Lf ≥ 0 compare [8, Inequality (11)].
such that for any x, x̃ ∈ X, u ∈ U satisfying kx − x̃k2 ≤ cLip
it holds that Assumption 2. (Unique steady-state) There exists σs > 0 such
that, for any x̃ ∈ Rn ,
kf (x, u) − f (x̃, u)k2 ≤ Lf kx − x̃k2 , (9)
Ax̃ − I B
σmin ≥ σs . (15)
and similarly for h. Using Assumption 1, we can derive Cx̃ D
the following bound between two linear models obtained by
linearizations at two different points x, x̃ when evaluated at Ax̃ − I B
Assumption 2 implies that has full column
two further points xa , xb : For any constant cLip > 0 and any Cx̃ D
rank, which is a standard condition in tracking, compare [1,
compact set X, there exist Lf ≥ 0 and cX > 0 such that
Lemma 1.8], [7, Remark 1]. This condition means that for any
kfx (xa , u) − fx̃ (xb , u)k2 (10) steady-state output of the linearized system, the corresponding
≤kf (xa , u) − fx (xa , u)k2 + kf (xb , u) − fx̃ (xb , u)k2 input-state pair is unique. More precisely, for any x̃ ∈ Rn ,
s s
there exists a linear map ĝx̃ : Zy,Lin (x̃) → ZLin (x̃) such that
+ kf (xa , u) − f (xb , u)k2
(7),(9) ĝx̃ (y s ) = (xs , us ), (16)
≤ cX kxa − xk22 + cX kxb − x̃k22 + Lf kxa − xb k2 ,
where (xs , us ) ∈ ZLin
s
(x̃) is the steady-state corresponding to
for all x, xa , x̃, xb ∈ X, u ∈ U with kxa − xb k2 ≤ cLip . y , i.e., y = hx̃ (x , us ). Due to the uniform lower bound (15),
s s s
with some (user-chosen) convex and compact set Us ⊆ int (U), (xsr sr sr
Lin (x̃), uLin (x̃)) := ĝx̃ (yLin (x̃)) (17)
which is required for a local controllability argument in our for the unique input-state pair corresponding to the minimizer
proofs. Further, we denote the projection of Z s on the state sr
yLin (x̃) of (13), and we write (xsr , usr ) for the input-state pair
component by Zxs and the projection of Z s on the output as corresponding to y sr , i.e.,
Zys := {y s ∈ Rp | ∃(xs , us ) ∈ Z s : y s = h(xs , us )}. (xsr , usr ) = ĝxsr (y sr ). (18)
The optimal equilibrium cost is defined as Moreover, we require that the linearized dynamics are uni-
∗ s formly controllable.
Jeq := min ky − y r k2S (12)
y ∈Zy
s s
Assumption 3. (Controllability) The pair (Ax̃ , B) is uni-
for some S 0. We denote a minimizer of (12) by y sr . In formly controllable for all x̃∈ Rn , i.e., the minimum singular
value of B . . . An−1
Section III-D, we provide sufficient conditions under which x̃ B is uniformly lower bounded.
4
Using standard arguments from linear systems theory, com- bounds on the optimal value function of the MPC problem in
pare [18, Theorem 5], it can be shown that Assumption 3 Section III-B. Further, a useful contraction property of the
implies the existence of a constant Γ > 0 such that for any Lyapunov function is stated in Section III-C. Section III-D
x̃ ∈ Rn , (xs , us ) ∈ ZLin s
(x̃), and any initial condition x0 , there contains the main result on closed-loop exponential stability
exists an input trajectory ûk ∈ Rm , k ∈ I[0,n−1] , steering of the optimal reachable equilibrium.
the linearized system from x0 to xs , i.e, x̂0 = x0 , x̂k+1 =
fx̃ (x̂k , ûk ), x̂n = xs , while satisfying A. MPC scheme
n−1
X Given the current state xt at time t as well as the lin-
kx̂k − xs k2 + kûk − us k2 ≤ Γkxs − x0 k2 . (19)
earization of the nonlinear system at xt according to (5), the
k=0
following optimal control problem will be the basis for our
Moreover, the following assumption is required for the lin- proposed MPC scheme:
earized system dynamics at any state.
N
X −1
Assumption 4. (Non-singular dynamics) There exists σ > 0 min kx̄k (t) − xs (t)k2Q + kūk (t) − us (t)k2R (20a)
x̄(t),ū(t)
such that σmin (I − Ax̃ ) ≥ σ for any x̃ ∈ Rn . xs (t),us (t) k=0
n y s (t)
Assumption 4 implies that, for any x̃ ∈ R , the matrix
I − Ax̃ has full rank and hence, for any equilibrium input us + ky s (t) − y r k2S
there exists a unique equilibrium state xs such that s.t. x̄k+1 (t) = Axt x̄k (t) + B ūk (t) + ext , (20b)
x̄0 (t) = xt , x̄N (t) = xs (t), (20c)
xs = Ax̃ xs + Bus + ex̃ .
ūk (t) ∈ U, k ∈ I[0,N −1] , (20d)
In case that U = Rm , it is straightforward to relax Assump- s s s
(x (t), u (t)) ∈ ZLin (xt ), (20e)
tion 4 by requiring that there exists a state-feedback gain
K such that AK = Ax̃ + BK satisfies the non-singularity y s (t) = Cxt xs (t) + Du (t) + rxt .
s
(20f)
condition σmin (I − AK ) ≥ σ for some σ and for all x̃ ∈ Rn . Here, x̄(t) ∈ RnN and ū(t) ∈ RmN denote the predicted state
We conjecture that it is possible to relax Assumption 4 further and input trajectory at time t, taking the value x̄k (t) ∈ Rn and
at the price of a more involved analysis. ūk (t) ∈ Rm at the k-th (predicted) time step, respectively.
Remark 1. Note that the conditions in Assumptions 2–4 are Compared to a standard linear MPC scheme with terminal
imposed for all x̃ ∈ Rn . This is mainly done for notational equality constraints (compare [1]), Problem (20) has two
convenience. As will become clear in our theoretical results, additional ingredients. First, the present scheme contains an
it actually suffices if these assumptions hold for all x̃ in artificial setpoint (xs (t), us (t)) which is optimized online and
a suitably defined compact set depending on the (positively whose distance w.r.t. the desired target setpoint y r is penalized
invariant) sublevel set of the Lyapunov function used to prove in the cost, similar to the tracking MPC formulation in [7]. If
stability. compared to classical MPC schemes with terminal equality
constraints [1], such an artificial steady-state increases the
Finally, we make an additional assumption on the steady- region of attraction and leads to recursive feasibility despite
state manifold of the linearized dynamics. online setpoint changes [7]. Further, the prediction of future
Assumption 5. (Compact steady-state manifold) There exists trajectories of the present nonlinear system is not based on the
s
a compact set B such that ZLin (x̃) ⊆ B for all x̃ ∈ Rn . full nonlinear model (3), but instead on a local linearization
around the current state xt . Therefore, also the artificial
Assumption 5 means that the union of all steady-state setpoint xs (t) is an equilibrium for the linearized dynamics
manifolds for the linearized dynamics at any point is contained according to (20e), but not necessarily for the nonlinear
in a compact set. If the input equilibrium constraints Us are system, and the artificial output setpoint y s (t) satisfies the
compact and Assumption 4 holds, then this is satisfied if linearized output equation.
the Jacobian is uniformly bounded. Assumption 5 is required We assume that the cost matrices in (20) are positive
for our theoretical results to obtain a uniform bound on the definite, i.e., Q, R 0. Under the given assumptions, it can
optimal equilibrium cost (13) and to conclude compactness of be shown analogously to [9, Proposition 1] that the optimal
certain Lyapunov function sublevel sets. The assumption can solution of (20) is unique. We denote this optimal solution
be dropped if it is known that the closed-loop trajectories lie by x̄∗ (t), ū∗ (t), xs∗ (t), us∗ (t), y s∗ (t) and the corresponding
within a compact invariant subset of the state-space. optimal cost by JN ∗
(xt ). Our results can be extended to
positive semidefinite state weightings Q 0 by invoking an
III. L INEAR TRACKING MPC FOR NONLINEAR SYSTEMS input-output-to-state stability argument (compare [19], [20]).
In this section, we propose a linear tracking MPC scheme If U, Us are polytopic, then Problem (20) is a convex QP and
to steer the nonlinear system (3)–(4) to a desired target can be solved efficiently. On the contrary, solving the non-
setpoint. The key idea is to use a local linearization-based convex problems associated with nonlinear MPC to optimality
model of the nonlinear system for prediction, and to update is in general computationally intractable. It is worth noting that
the linearization online using the current measurements. After the computational complexity of the proposed MPC approach
stating the scheme in Section III-A, we prove lower and upper is also smaller than that of alternative linearization-based
5
approaches such as the real-time iteration scheme [2], [4], Proof. (i) Lower bound
[5] since i) only the linearization w.r.t. xt instead of the Note that
previously optimal solution x̄∗ (t − 1) is needed and ii) com- ∗ ∗
V (xt ) = JN (xt ) − Jeq,Lin (xt )
parable stability guarantees of the real-time iteration scheme
∗
require a sufficiently small sampling time [5], i.e., solving ≥ kxt − xs∗ (t)k2Q + ky s∗ (t) − y r k2S − Jeq,Lin (xt )
the underlying optimization problem more frequently [4], cf. (14)
≥ kxt − xs∗ (t)k2Q + λmin (S)ky s∗ (t) − yLin
sr
(xt )k22 .
Remark 2.
Assumption 2 (i.e., the maps in (16)) implies the existence of
Algorithm 1. n-step tracking MPC Scheme
a constant ĉl > 0, which can be chosen uniformly over xt ,
1) At time t, take the current state measurement xt and such that
compute Axt , ext , Cxt , D, rxt according to (5).
2) Solve (20) and apply the first n input components ky s∗ (t) − yLin
sr
(xt )k22 ≥ ĉl kxs∗ (t) − xsr 2
Lin (xt )k2 .
ut+k = ū∗k (t), k ∈ I[0,n−1] . Hence, we obtain
3) Set t = t + n and go back to 1).
V (xt ) ≥ kxt − xs∗ (t)k2Q + ĉl λmin (S)kxs∗ (t) − xsr 2
Lin (xt )k2
(1) min{λmin (Q), ĉl λmin (S)}
In this paper, we consider the n-step MPC scheme out- ≥ kxt − xsr 2
Lin (xt )k2 .
2
lined in Algorithm 1. We employ a multi-step MPC scheme | {z }
cl :=
(compare [21], [22]) due to the joint occurrence of a model
mismatch (induced by the linearized model) and terminal (ii) Upper bound
equality constraints. More precisely, the candidate solution In the following, we construct a candidate solution to Prob-
used in our theoretical analysis is defined based on the lem (20) which will then be used to bound the optimal cost
∗
shifted previously optimal solution with an appended deadbeat JN (xt ). We choose the candidate equilibrium as xs (t) =
controller, requiring n additional time steps, i.e., a multi-step xLin (xt ), us (t) = usr
sr s sr
Lin (xt ), y (t) = yLin (xt ), i.e., the optimal
MPC scheme with at least n steps, cf. [11, Figure 1]. The same reachable equilibrium of the system linearized at xt . Using
theoretical guarantees can be given for a ν-step MPC scheme, N ≥ n and Assumption 3, there exists a trajectory of the
where ν ≤ n denotes the controllability index, provided that linearized dynamics steering the state to xs (t) within N steps
Assumption 3 is modified accordingly. while satisfying
Clearly, the prediction model in the proposed MPC scheme N −1
is not exact due to the linearization of the nonlinear system.
X
kx̄k (t) − xs (t)k2 + kūk (t) − us (t)k2 ≤ Γkxt − xs (t)k2
As we will see in the remainder of this paper, by suitably k=0
tuning the cost parameters (S needs to be small) and when
for some Γ > 0 (compare (19)). If δ is sufficiently small,
starting close to the steady-state manifold Z s of the nonlinear
then usrLin (xt ) ∈ int(U) implies that the corresponding input
system, the artificial steady-state xs (t) is always close to the
satisfies the constraints, i.e., ūk (t) ∈ U for all k ∈ I[0,N −1] .
current state xt such that the prediction error is sufficiently
Clearly, the above inequality implies
small. Then, xs (t) remains close to the nonlinear steady-
N −1
state manifold and slowly drifts towards the optimal reachable X
equilibrium xsr such that, asymptotically, the closed-loop state kx̄k (t) − xs (t)k2Q + kūk (t) − us (t)k2R
k=0
trajectory converges to xsr , compare Figure 1.
N
X −1
≤ λmax (Q, R) (kx̄k (t) − xs (t)k22 + kūk (t) − us (t)k22 )
B. Value function bound k=0
In order to prove closed-loop exponential stability, we ≤ λmax (Q, R)Γ2 kxt − xs (t)k22 .
consider a Lyapunov function candidate of the form
Thus, the following holds for the Lyapunov function candidate
∗ ∗
V (xt ) := JN (xt ) − Jeq,Lin (xt )
V (xt ) ≤λmax (Q, R)Γ2 kxt − xs (t)k22 + ky s (t) − y r k2S
∗
with Jeq,Lin (xt )
as in (13). The following result shows that V ∗
− Jeq,Lin (xt )
admits suitable quadratic lower and upper bounds, which will
be required to prove desired stability properties. = λmax (Q, R)Γ2 kxt − xsr 2
Lin (xt )k2 .
| {z }
cu :=
Lemma 1. Suppose N ≥ n and Assumptions 1, 2, and 3 hold.
For any compact set X ⊂ Rn , there exist cl , cu , δ > 0 such Lemma 1 provides bounds on the Lyapunov function can-
that didate V (xt ) that will be employed to prove closed-loop
exponential stability. As an alternative to V (xt ), one could
(i) for all xt ∈ X, the function V is lower bounded as ∗ ∗
consider the candidate JN (xt ) − Jeq , which depends on the
V (xt ) ≥ cl kxt − xsr 2
Lin (xt )k2 , (21) cost of the optimal reachable equilibrium for the nonlinear
∗
system instead of the cost Jeq,Lin (xt ) for the linearized system.
(ii) for all xt ∈ X with kxt − xsr
Lin (xt )k2 ≤ δ, the function ∗ ∗
However, deriving a useful lower bound for JN (xt ) − Jeq
V is upper bounded as
similar to (21) is difficult, which is why we consider the
V (xt ) ≤ cu kxt − xsr 2
Lin (xt )k2 . (22) proposed Lyapunov function candidate V (xt ) instead.
6
Theorem 1 should be interpreted as a qualitative result since compare Inequality (26), which may not necessarily be optimal
it does not provide explicit values of Vmax and S̄ leading to for the nonlinear dynamics. On the other hand, Assumption 4
closed-loop stability. Loosely speaking, Theorem 1 guarantees (non-singular dynamics) might possibly be relaxed, compare
stability if the initial state is sufficiently close to Zxs and the discussion after Assumption 4. Finally, we note that these
λmax (S) is sufficiently small. This is due to the fact that assumptions hold in many practical applications, e.g., for the
V (xt ) ≤ Vmax can be ensured for an arbitrarily small Vmax if CSTR example we consider in Section IV.
xt lies in a neighborhood of Zxs and S̄ is sufficiently small.
To conclude, the proposed MPC scheme based on repeatedly
In [23], it is shown for linear systems that a tracking MPC
solving the linear MPC problem (20) leads to desirable closed-
formulation recovers optimality properties of a standard MPC
loop guarantees when applied to a nonlinear system, and the
scheme if the weight on the distance between the artificial
scheme can be tuned based on a single design parameter
setpoint and the reference setpoint (i.e., the matrix S in our
S, which allows for a trade-off between the size of the
setting) is suitably large. This indicates a trade-off when
region of attraction and the convergence speed. Compared to
designing the matrix S: It needs to be suitably small such that
nonlinear tracking MPC schemes such as [7], [8], our approach
the linearization error is small and stability can be guaranteed,
has the drawback that convergence may be slower since S
but the performance deteriorates if S is chosen too small.
needs to be chosen sufficiently small. On the other hand,
Remark 2. Note that Theorem 1 guarantees closed-loop Problem (20) is a convex QP which can be solved up to global
stability of Algorithm 1 which only requires solving one convex optimality very efficiently. Further, the prediction model only
QP online. Alternative nonlinear MPC approaches based on requires an accurate description of the system close to the
convex optimization typically employ an LTI prediction model steady-state manifold, which may be simpler to obtain than a
by linearizing at the setpoint or, as in the real-time iteration globally accurate model which is required to obtain superior
scheme [2], [4], [5], an LTV prediction model by lineariz- performance with existing nonlinear MPC approaches. Finally,
ing along the candidate solution. Closed-loop guarantees of as we show in our companion paper [9], the presented idea can
such approaches require either an additional bounding of be extended in order to develop a data-driven MPC scheme to
the linearization error [12] or a sufficiently small sampling control unknown nonlinear systems by continuously updating
time [5], i.e., solving the underlying optimization problem the measured data used for prediction.
more frequently [4]. The latter ensures that the linearized
dynamics do not change too rapidly, which is analogous IV. N UMERICAL E XAMPLE
to our condition on λmax (S) being sufficiently small. The We apply the proposed MPC scheme to the CSTR from [6]
proposed approach has a large region of attraction due to the with the nonlinear system dynamics f (x, u) equal to
online optimization of the artificial steady-state. In particular, " M
#
while a standard (linearization-based) MPC only guarantees x1 + Tθs (1 − x1 ) − Ts k̄x1 e− x2
M .
stability when starting in a region around the setpoint, our x2 + Tθs (xf − x2 ) + Ts k̄x1 e− x2 − Ts αu(x2 − xc )
MPC scheme ensures stability for initial conditions far away
The states x1 and x2 are the temperature and the concentration,
from the setpoint as long as they are close to the steady-
respectively, and the control input u is the coolant flow
state manifold. This fact will be illustrated with a numerical
rate. These dynamics are obtained from the continuous-time
example in Section IV. Furthermore, as is standard in MPC,
dynamics in [6] via a simple Euler discretization with sampling
our guarantees remain true if the QP is not solved up to
time Ts = 0.2. The other parameters appearing in the vector
optimality by using a warm start and results on suboptimality
field are θ = 20, k̄ = 300, M = 5, xf = 0.3947, xc = 0.3816,
in MPC [1, Section 2.7].
α = 0.117. Our control goal is tracking of the output setpoint
Remark 3. Theorem 1 requires a number of assumptions, y r = 0.6519 for the concentration, i.e., h(x, u) = x2 , while
most of which are not too restrictive when proving closed-loop satisfying the input constraints uk ∈ U = [0.1, 2] for k ∈ I≥0 .
stability of a linearization-based MPC scheme utilizing the In order to set up the MPC, we consider the cost matrices
benefits of an artificial setpoint: Assumption 1 (smoothness) Q = I, R = 0.05, S = 100, the prediction horizon N = 40,
is clearly required for a linearization-based MPC scheme. and the input equilibrium constraints Us = [0.11, 1.99]. Since
Further, Assumption 2 is a standard condition in the liter- the dynamics are not of the control-affine form (3), we imple-
ature on tracking MPC, compare [1], [7]. If Assumption 2 ment the MPC scheme with an incremental input formulation
does not hold, we can still guarantee asymptotic stability ∆uk := uk+1 − uk and include an additional penalty k∆uk k22
of some steady-state, but not necessarily convergence to in the cost.
xsr . Assumptions on controllability (Assumption 3) are also We first investigate whether our assumptions are met by the
standard in the presence of terminal equality constraints, CSTR. For this verification, we only consider linearization
see [1]. In order to ensure that the employed bounds hold points in the relevant operating range, i.e., x̃ ∈ (0, 1]2 . It
uniformly, it is crucial to make assumptions on compactness is simple to verify that the considered system satisfies As-
of the steady-state manifold (Assumption 5). Moreover, as sumption 1 (smoothness). While Assumption 4 (non-singular
we show in Appendix C, Assumption 6 holds, in fact, as dynamics) does not hold due to the integrator dynamics
long as m = p and y r is reachable. In the absence of uk+1 = uk + ∆uk , our theoretical results still apply since
Assumption 6, we can still guarantee closed-loop stability Assumption 4 holds for the original system (without ∆u) and
of the optimal reachable steady-state of some linearization, thus, for any given (us , ∆us ) = (us , 0) there still exists a
8
0.5
0.4
to be computed, whereas the proposed MPC scheme only
0.3
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 requires the linearization at xt . While the LTV-MPC provides a
good trade-off between computational complexity and closed-
loop performance, theoretical results in the literature require
Fig. 2. State component of the steady-state manifold Z s and closed-loop either an additional bounding of the linearization error [12]
trajectory under the n-step MPC scheme (Algorithm 1) for the numerical
example in Section IV.
or sufficiently many iterations [4], [5], compare Remark 2.
Without online optimization of an artificial setpoint, all MPC
schemes considered above are initially infeasible. Furthermore,
unique steady-state xs . Assumption 2 holds in (0, 1]2 except note that the performance of the 1-step MPC scheme is
for a neighborhood of x2 = xc . Similarly, the linearized superior if compared to that of the n-step MPC scheme since
dynamics are controllable (Assumption 3) on (0, 1]2 except in the model is updated more frequently (twice as often) for the
a neighborhood of x2 = xc or x2 = 0. Further, Assumption 5 1-step scheme and hence, the influence of the prediction error
(compact steady-state manifold of the linearization) clearly is smaller. Finally, the LTI-MPC based on the linearization
holds on the set x̃ ∈ (0, 1]2 due to Assumption 4. Finally, at xsr fails to track the desired setpoint. To summarize, in
Assumption 6 holds since m = p and the setpoint y r is application to a CSTR, the presented tracking MPC scheme
reachable, i.e., Assumption 7 holds, compare Appendix C. using a linearized prediction model leads to a closed-loop
Figure 2 shows the closed-loop state trajectory under performance which is comparable to that of nonlinear tracking
Algorithm 1 when starting at the initial state x0 = MPC while being significantly more computationally efficient.
>
0.9492 0.43 . During the full closed-loop operation, the
trajectory remains close to Z s such that the prediction error V. C ONCLUSION
induced by the linearization is small and yt asymptotically In this paper, we presented a novel tracking MPC scheme
converges to y r . For comparison, we also apply the following for nonlinear systems using a prediction model based on the
MPC schemes, each with terminal equality constraints, online linearized dynamics at the current state. As a key technical
optimization of an artificial equilibrium, an incremental input contribution, we proved that the optimal reachable equilib-
penalty, and the same design parameters as above: rium is exponentially stable in closed loop under reasonable
assumptions on the underlying system. The method was suc-
1) the proposed MPC scheme in a one-step fashion (i.e.,
cessfully applied to a numerical example where it achieved
Algorithm 1 with n = 1),
good performance while being computationally more efficient
2) a one-step MPC scheme using an LTI prediction model
than a comparable nonlinear tracking MPC scheme.
obtained by linearizing the nonlinear dynamics (3) at xsr
The presented results build the basis for obtaining theoret-
(called “LTI-MPC”),
ical guarantees when using linear prediction models in MPC
3) a one-step MPC scheme using an LTV prediction model
to control nonlinear systems. In our companion paper [9],
obtained by linearizing the nonlinear dynamics (3) at
we exploit this viewpoint further to design an MPC scheme
time t along the candidate solution x̄∗1 (t − 1), x̄∗2 (t − 1),
for unknown nonlinear systems with closed-loop stability
. . . , x̄∗N (t−1) (called “LTV-MPC”), analogously to [12]
guarantees based on linear data-dependent prediction models
and comparable to the real-time iteration scheme [2], and
from behavioral systems theory [25].
4) the nonlinear tracking MPC scheme from [8].
The closed-loop state- and input-trajectories can be seen in R EFERENCES
Figure 3. First, note that, except for the LTI-MPC, all MPC
[1] J. B. Rawlings, D. Q. Mayne, and M. M. Diehl, Model Predictive
schemes achieve asymptotic tracking of the desired setpoint. Control: Theory, Computation, and Design. Nob Hill Pub, 2020, 3rd
The nonlinear tracking MPC from [8] performs better than printing.
the LTV-MPC, which in turn outperforms the proposed n-step [2] M. Diehl, R. Findeisen, F. Allgower, H. G. Bock, and J. P. Schloder,
“Nominal stability of real-time iteration scheme for nonlinear model
MPC scheme (Algorithm 1) as well as the corresponding 1- predictive control,” IEE Proceedings - Control Theory and Applications,
step MPC scheme. Table I lists the times required for setting vol. 152, no. 3, pp. 296–308, 2005.
up the QPs, including the computation of the Jacobians, and [3] M. Diehl, H. J. Ferreau, and N. Haverbeke, “Efficient numerical methods
for nonlinear MPC and moving horizon estimation,” in Lecture notes in
solving the optimization problems arising in the considered control and information sciences: Vol. 483. Nonlinear model predictive
MPC schemes (using ’quadprog’ for the QPs and CasADi [24] control. Springer, 2009, pp. 391–417.
with solver ’IPOPT’ for the nonlinear optimization problem [4] D. Liao-McPherson, M. M. Nicotra, and I. Kolmanovsky, “Time-
distributed optimization for real-time model predictive control: stability,
in [8]). Note that the LTV-MPC has slightly larger computation robustness, and constraint satisfaction,” Automatica, vol. 117, p. 108973,
times since, at each time step, N linearized dynamics need 2020.
9
1
Trans. Automat. Control, vol. 66, no. 4, pp. 1702–1717, 2021.
[12] M. Cannon, J. Buerger, B. Kouvaritakis, and S. Rakovic, “Robust tubes
in nonlinear model predictive control,” IEEE Trans. Automat. Control,
0.8 vol. 56, no. 8, pp. 1942–1947, 2011.
[13] M. Ławryńczuk and P. Tatjewski, “A computationally efficient nonlinear
predictive control algorithm with RBF neural models and its applica-
0.6
tion,” in Rough Sets and Intelligent Systems Paradigms. Springer Berlin
Heidelberg, 2007, pp. 603–612.
0.4 [14] D. Papadimitriou, U. Rosolia, and F. Borrelli, “Control of unknown
nonlinear systems with linear time-varying MPC,” in Proc. 59th IEEE
Conf. Decision and Control (CDC), 2020, pp. 2258–2263.
0.2 [15] M. Korda and I. Mezić, “Linear predictors for nonlinear dynamical sys-
0 200 400 600 800 1000 1200 1400 1600 1800 2000
tems: Koopman operator meets model predictive control,” Automatica,
vol. 93, pp. 149–160, 2018.
(a) State component x1 [16] L. Magni and R. Scattolini, “On the solution of the tracking problem
for non-linear systems with MPC,” Int. J. Systems Science, vol. 36, pp.
0.7
477–484, 2005.
[17] M. Radulescu and S. Radulescu, “Global inversion theorems and appli-
0.65 cations to differential equations,” Nonlinear Analysis, Theory, Methods
0.6
& Applications, vol. 4, no. 4, pp. 951–965, 1980.
[18] E. Sontag, Mathematical Control Theory. Springer-Verlag, New York,
0.55 1998.
0.5
[19] J. Berberich, J. Köhler, M. A. Müller, and F. Allgöwer, “Data-driven
tracking MPC for changing setpoints,” IFAC-PapersOnLine, vol. 53,
0.45 no. 2, pp. 6923–6930, 2020.
0.4 [20] C. Cai and A. R. Teel, “Input–output-to-state stability for discrete-time
systems,” Automatica, vol. 44, no. 2, pp. 326–336, 2008.
0.35 [21] L. Grüne and V. G. Palma, “Robustness of performance and stability
0 200 400 600 800 1000 1200 1400 1600 1800 2000
for multistep and updated multistep MPC schemes,” Discrete and
Continuous Dynamical Systems, vol. 35, no. 9, pp. 4385–4414, 2015.
(b) State component x2 [22] K. Worthmann, M. W. Mehrez, G. K. I. Mann, R. G. Gosine, and
J. Pannek, “Interaction of open and closed loop control in MPC,”
1
Automatica, vol. 82, pp. 243–250, 2017.
[23] A. Ferramosca, D. Limón, I. Alvarado, T. Alamo, F. Castaño, and E. F.
Camacho, “Optimal MPC for tracking of constrained linear systems,”
0.8
Int. J. Systems Science, vol. 42, no. 8, pp. 1265–1276, 2011.
[24] J. A. E. Andersson, J. Gillis, G. Horn, J. B. Rawlings, and M. Diehl,
0.6
“CasADi: a software framework for nonlinear optimization and control,”
Mathematical Programming Computation, vol. 11, no. 1, pp. 1–36, 2019.
0.4
[25] J. C. Willems, P. Rapisarda, I. Markovsky, and B. De Moor, “A note on
persistency of excitation,” Syst. Contr. Lett., vol. 54, pp. 325–329, 2005.
0.2 [26] B. Schuermann, N. Kochdumper, and M. Althoff, “Reachset model
predictive control for disturbed nonlinear systems,” in Proc. 57th IEEE
0 Conf. Decision and Control (CDC), 2018, pp. 3463–3470.
0 200 400 600 800 1000 1200 1400 1600 1800 2000
Proof. First, we define the candidate equilibrium xs0 (t + (ii.a) Bound on kx̄∗n (t) − xt+n k2
n), us0 (t + n) and the first N − n components of the candidate Using (10), which holds by Assumption 1, and (6), we obtain
input ū0 (t + n) (Part (i)). Thereafter, in Part (ii), we derive
useful bounds involving this candidate trajectory. Next, we kx̄∗n (t) − xt+n k2 (30)
show that the state of this candidate solution at time N − n is =kfxt (x̄∗n−1 (t), ū∗n−1 (t)) − fxt+n−1 (xt+n−1 , ū∗n−1 (t))k2
sufficiently close to the candidate artificial steady-state and (10)
thus, we can construct a local deadbeat controller to steer ≤ cX kx̄∗n−1 (t) − xt k22 + Lf kx̄∗n−1 (t) − xt+n−1 k2
the system to this steady-state (Part (iii)). Finally, we show (1)
≤2cX kx̄∗n−1 (t) − xs∗ (t)k22 + 2cX kxs∗ (t) − xt k22
V (xt+n ) ≤ cV1 V (xt ) in Part (iv).
Note that xt lies in the set {x ∈ Rn | V (x) ≤ Vmax }, which + Lf kx̄∗n−1 (t) − xt+n−1 k2
(29) cX
is compact due to the lower bound (21) and Assumption 5.
≤ 2 V (xt ) + Lf kx̄∗n−1 (t) − xt+n−1 k2
We define the set X as the union of the N -step reachable sets q
of the linearized and the nonlinear dynamics (compare [26, n−2
cX X
Definition 2]), starting at xt . Using that the dynamics (3) are ≤··· ≤ 2 V (xt ) Lkf ,
Lipschitz continuous and the input constraints are compact, we q
k=0
conclude that X is compact. Throughout this proof, whenever
where the summand for k = n − 1 vanishes since kx̄∗1 (t) −
we apply Inequality (10), we use the fact that all involved
xt+1 k2 = 0.
states lie in X and we use the corresponding constant cX .
(ii.b) Bound on kx̄0k (t + n) − x̄∗k+n (t)k2
(i) Definition of candidate solution for k ∈ I[0,N −n] −n
Define {ak }Nk=0 recursively in dependence of V (xt ) as
We choose the candidate equilibrium input as the old solution,
i.e., us0 (t + n) = us∗ (t). According to Assumption 4, there n−2
cX X
exists a unique equilibrium state xs0 (t + n) for the system a0 := 2 V (xt ) Lkf ,
linearized at xt+n such that q
k=0
cX
s0 s0 s0
ak := 2cX a2k−1 + Lf ak−1 + 16 V (xt )
x (t + n) = Axt+n x (t + n) + Bu (t + n) + ext+n . q
n−2
!2
The corresponding output y s0 (t + n) is computed via (20f). cX X k
+ 8cX 2 Lf V (xt )2 , k = 1, . . . , N − n.
Further, for k ∈ I[0,N −n−1] , we choose the candidate input as q
k=0
the previously optimal one, i.e., ū0k (t + n) = ū∗k+n (t). This
leads to the state trajectory candidate x̄0k (t + n), k ∈ I[0,N −n] , In the following, we prove that for any k ∈ I[0,N −n]
resulting from an open-loop application of ū0 (t+n) with initial kx̄0k (t + n) − x̄∗k+n (t)k2 ≤ ak . (31)
condition xt+n to the dynamics linearized at time t + n, i.e.,
x̄00 (t + n) = xt+n and According to (30) and using x̄00 (t+n) = xt+n , Inequality (31)
holds for k = 0. Using an induction argument over k, we have
x̄0k+1 (t + n) = fxt+n (x̄0k (t + n), ū0k (t + n))
= Axt+n x̄0k (t + n) + B ū0k (t + n) + ext+n , kx̄0k (t + n) − x̄∗k+n (t)k2
= kfxt+n (x̄0k−1 (t + n), ū∗k+n−1 (t))
for k ∈ I[0,N −n−1] . − fxt (x̄∗k+n−1 (t), ū∗k+n−1 (t))k2
(ii) Bounds on candidate solution (10)
Throughout Part (ii) of the proof, let k ∈ I[0,N −n] . Further, ≤ cX kx̄0k−1 (t + n) − xt+n k22 + cX kx̄∗k+n−1 (t) − xt k22
abbreviate q := λmin (Q), q̄ := λmax (Q), and similarly for s, + Lf kx̄0k−1 (t + n) − x̄∗k+n−1 (t)k2
s̄, r, r̄. It clearly holds that (1)
≤ 2cX kx̄0k−1 (t + n) − x̄∗k+n−1 (t)k22
N −1
X + Lf kx̄0k−1 (t + n) − x̄∗k+n−1 (t)k2
kx̄∗k (t) − xs∗ (t)k2Q (28)
k=0 + 2cX kx̄∗k+n−1 (t) − xt+n k22 + cX kx̄∗k+n−1 (t) − xt k22
N −1 (1),(31)
≤ 2cX a2k−1 + Lf ak−1 + 6cX kx̄∗k+n−1 (t) − xs∗ (t)k22
X
≤ kx̄∗k (t) − xs∗ (t)k2Q + ky s∗ (t) − y r k2S − Jeq,Lin
∗
(xt )
k=0 + 4cX kxs∗ (t) − xt+n k22 + 2cX kxs∗ (t) − xt k22
∗ ∗
≤JN (xt ) − Jeq,Lin (xt ) = V (xt ), (1),(29) cX
≤ 2cX a2k−1 + Lf ak−1 + 8 V (xt )
q
and hence,
+ 8cX kxs∗ (t) − x̄∗n (t)k22 + 8cX kx̄∗n (t) − xt+n k22
N −1 (29),(30) cX
X 1 ≤ 2cX a2k−1 + Lf ak−1 + 16 V (xt )
kx̄∗k (t) − xs∗ (t)k22 ≤ V (xt ). (29) q
q
k=0 2
n−2
!
cX X k
We bound now several expressions involving the optimal + 8cX 2 Lf V (xt )2 = ak ,
q
solution at time t and the candidate solution at time t + n. k=0
11
which proves (31). Note that ak is a polynomial in V (xt ) Moreover, using ka + b + ck22 ≤ 2kak22 + 4kbk22 + 4kck22 , which
which becomes arbitrarily small if V (xt ) is sufficiently small. holds for arbitrary a, b, c due to (1), we obtain
(ii.c) Bound on kxs∗ (t) − xs0 (t + n)k2
Note that kxs0 (t + n) − xt+n k22 ≤ 2kxs0 (t + n) − xs∗ (t)k22
+ 4kxs∗ (t) − x̄∗n (t)k22 + 4kx̄∗n (t) − xt+n k22
s∗ s∗
(I − Axt )x (t) = Bu (t) + ext , (29),(30),(33) 4
≤ 2(c1 V (xt )2 + c2 V (xt ))2 + V (xt )
(I − Axt+n )xs0 (t + n) = Bus0 (t + n) + ext+n . q
n−2
! 2
Using additionally us0 (t + n) = us∗ (t), this implies cX X
+ 4 2 V (xt ) Lkf .
q
k=0
(I − Axt+n )xs0 (t + n)
Hence, using V (xt ) ≤ Vmax , there exists c3 > 0 such that
=xs∗ (t) − xs∗ (t) + Bus0 (t + n) + ext+n
=(I − Axt )xs∗ (t) + ext+n − ext . ky s0 (t + n) − y s∗ (t)k2 ≤ c3 V (xt ). (34)
We now bound several terms on the right-hand side of (37) Inserting all of the derived bounds into (37), we arrive at
separately. The definition of the input candidate implies
∗ ∗
JN (xt+n ) − JN (xt ) (39)
N −1
X n−1
kū0k (t + n) − us0 (t + n)k2R − kū∗k (t) − us∗ (t)k2R ≤−
X
(kx̄∗k (t) − xs∗ (t)k2Q + kū∗k (t) − us∗ (t)k2R )
k=0
k=0
n−1
(36) X −n−1
NX
≤ − kū∗k (t) −u s∗
(t)k2R + 2q̄(a2k + (c1 V (xt )2 + c2 V (xt ))2 )
k=0
k=0
+ r̄ · Γ2 (aN −n + c1 V (xt )2 + c2 V (xt ))2 . p
+2 q̄V (xt )(ak + c1 V (xt )2 + c2 V (xt )
p ∗
Using ky s∗ (t) − y r kS ≤ JN (xt ), we obtain + (q̄ + r̄)Γ2 (aN −n + c1 V (xt )2 + c2 V (xt ))2
√ q
ky s0 (t + n) − y r k2S − ky s∗ (t) − y r k2S (38) + s̄c23 V (xt )2 + 2 s̄c3 V (xt ) V (xt ) + Jeq,Lin
∗ (xt ).
(2)
≤ky s0 (t + n) − y s∗ (t)k2S Note that (27) implies
+ 2ky s0 (t + n) − y s∗ (t)kS ky s∗ (t) − y r kS
n−1
≤ky s0 (t + n) − y s∗ (t)k2S
X
q − (kx̄∗k (t) − xs∗ (t)k2Q + kū∗k (t) − us∗ (t)k2R ) (40)
s0 s∗ ∗ k=0
+ 2ky (t + n) − y (t)kS V (xt ) + Jeq,Lin (xt )
n−1
√
(27) 1X ∗
(34) ≤ − (kx̄k (t) − xs∗ (t)k2Q + kū∗k (t) − us∗ (t)k2R )
q
≤ s̄c23 V (xt )2 + 2 s̄c3 V (xt ) V (xt ) + Jeq,Lin
∗ (xt ). 2
k=0
are chosen as the corresponding equilibria for the dynamics to x̂s (t + n) such that we can steer the system to x̂s (t + n) in
linearized at xt , i.e., L steps. It follows from the proof of Proposition 3 that
In the following, we bound the terms on the right-hand side for some c4 > 0, using that V (xt ) ≤ Vmax in the last
of (52). First, we obtain inequality.
(ii.b) Bound on kŷ s (t + n) − y r k2S − ky s∗ (t) − y r k2S
kxs0 (t + n − 1) − xt+n k2 (53) Similar to [8], [19], it is straightforward to exploit the con-
s0
= kAxt+n−1 (x (t + n − 1) − xt+n−1 ) vexity condition (14) in order to derive
+ B(us∗ (t) − ut+n−1 )k2
kỹ s − y r k2S − ky s∗ (t) − y r k2S (60)
s0
≤ kAxt+n−1 k2 kx (t + n − 1) − xt+n−1 k2 2 s∗ sr
≤ − (1 − λ )ky (t) − yLin (xt )k2S ,
+ kBk2 kus∗ (t) − ut+n−1 k2
where ỹ s = λy s∗ (t) + (1 − λ)yLin
sr
(xt ). Moreover, (2) implies
≤ kAxt+n−1 k2 kxs0 (t + n − 1) − xs∗ (t)k2
kŷ s (t + n) − y r k2S − kỹ s − y r k2S (61)
+ kxs∗ (t) − xt+n−1 k2 + kBk2 kus∗ (t) − ū∗n−1 (t)k2 . s
≤kŷ (t + n) − ỹ s k2S s s s
+ 2kŷ (t + n) − ỹ kS kỹ − y kS . r
(iii) Invariance of V (xt ) ≤ Vmax due to Assumption 2 and m = p. Before verifying Assump-
It follows directly from (51), (59), and (65), and using the tion 6, we first prove a technical intermediate result.
inequality a2 + b2 ≤ (a + b)2 for a, b ≥ 0, that
Lemma 2. Suppose Assumptions 1, 2, 4, and 5 hold and
∗ ∗ m = p. Then, for any compact set X with B ⊆ X × U (cf.
JN (xt+n ) − JN (xt )
c ĉ2 Assumption 5), there exists a constant cs1 > 0 such that for
X l
≤ λmax (Q, R)Γ2 4 (1 − λ)2 ky s∗ (t) − yLin
sr
(xt )k22 any steady-state (xs , us ) ∈ Z s of the nonlinear system, any
σ
state x̃ ∈ X, and any input ũ ∈ U, it holds that
√ 2
+ ĉl ((1 − λ) + cAB γ̃ γ1 )ky s∗ (t) − yLin
sr
(xt )k2 + c4 V (xt )
k(xs , us ) − (x̃, ũ)k2 (66)
− kxt − x (t)k2Q − (1 − λ2 )ky s∗ (t)
s∗
− sr
yLin (xt )k2S ≤cs1 kh(xs , us ) − h(x̃, ũ)k2 + kx̃ − f (x̃, ũ)k2 .
i.e., the numbers of inputs and outputs coincide. Further, we the desired statement. To this end, it is readily derived that
assume that the target setpoint y r is reachable, as captured in kMx̂−1 − Mx−1
sr k2 ≤ kM
−1 −1
x̂ k2 kI − Mx̂ Mxsr k2 ,
the following assumption. kI − Mx̂ Mx−1 −1
sr k2 ≤ kMxsr − Mx̂ k2 kMxsr k2 .
with
c̃X −exsr cX0
c̄eq,1 := + max kxs − x̂k2 .
σs 2 y sr − rxsr 2 σs x̂∈X,xs ∈Bx
Finally, using
kx̂ − xsr sr sr sr
Lin (x̂)k2 ≤kx̂ − x k2 + kx − xLin (x̂)k2 , Matthias A. Müller received a Diploma degree
1 in Engineering Cybernetics from the University of
we obtain the left inequality in (23) for ceq,1 := > 0. (1+c̄eq,1 )2 Stuttgart, Germany, and an M.S. in Electrical and
Proof of kx̂ − xsr k22 ≤ ceq,2 kx̂ − xsr
Lin (x̂)k 2
2
Computer Engineering from the University of Illi-
nois at Urbana-Champaign, US, both in 2009. In
Applying Lemma 2 with (xs , us ) = (xsr , usr ), (x̃, ũ) = 2014, he obtained a Ph.D. in Mechanical Engineer-
(xsr sr
Lin (x̂), uLin (x̂)), we obtain ing, also from the University of Stuttgart, Germany,
for which he received the 2015 European Ph.D.
kxsr − xsr sr sr sr
Lin (x̂)k2 ≤ cs1 ky − h(xLin (x̂), uLin (x̂))k2 (74) award on control for complex and heterogeneous
systems. Since 2019, he is director of the Institute of
+ kxLin (x̂) − f (xLin (x̂), usr
sr sr
Lin (x̂))k2 . Automatic Control and full professor at the Leibniz
University Hannover, Germany. He obtained an ERC Starting Grant in 2020
Using y sr = hx̂ (xsr sr
Lin (x̂), uLin (x̂)) due to Assumption 7 and is recipient of the inaugural Brockett-Willems Outstanding Paper Award
together with a bound of the form (10) for the vector field for the best paper published in Systems & Control Letters in the period
h, the first term is bounded as 2014-2018. His research interests include nonlinear control and estimation,
model predictive control, and data-/learning-based control, with application
khx̂ (xsr sr
Lin (x̂), uLin (x̂)) − hxsr
Lin (x̂)
(xsr sr
Lin (x̂), uLin (x̂))k2 in different fields including biomedical engineering.
for some cXh > 0. Moreover, the second term on the right-
hand side of (74) is bounded as
kxsr sr sr
Lin (x̂) − f (xLin (x̂), uLin (x̂))k2 Frank Allgöwer is professor of mechanical engi-
=kfx̂ (xsr sr
Lin (x̂), uLin (x̂)) − fxsr
Lin (x̂)
(xsr sr
Lin (x̂), uLin (x̂))k2 neering at the University of Stuttgart, Germany, and
Director of the Institute for Systems Theory and
(10)
Automatic Control (IST) there.
≤ cX kx̂ − xsr 2
Lin (x̂)k2 . Frank is active in serving the community in several
roles: Among others he has been President of the In-
Combining the above inequalities, we obtain ternational Federation of Automatic Control (IFAC)
for the years 2017-2020, Vice-president for Techni-
kx̂ − xsr k2 cal Activities of the IEEE Control Systems Society
≤kx̂ − xsr sr sr
Lin (x̂)k2 + kxLin (x̂) − x k2
for 2013/14, and Editor of the journal Automatica
from 2001 until 2015. From 2012 until 2020 Frank
≤kx̂ − xsr sr 2
Lin (x̂)k2 + cs1 (cXh + cX )kx̂ − xLin (x̂)k2 , served in addition as Vice-president for the German Research Foundation
(DFG), which is Germany’s most important research funding organization.
leading to the right inequality in (23) for His research interests include predictive control, data-based control, networked
control, cooperative control, and nonlinear control with application to a wide
ceq,2 := (1 + cs1 (cXh + cX ) maxkx̂ − xsr 2
Lin (x̂)k2 ) . range of fields including systems biology.
x̂∈X