International Journal of Control

ISSN: (Print) (Online) Journal homepage: https://www.tandfonline.com/loi/tcon20

Model predictive control for singular differential-

algebraic equations

Achim Ilchmann, Jonas Witschel & Karl Worthmann

To cite this article: Achim Ilchmann, Jonas Witschel & Karl Worthmann (2021): Model predictive
control for singular differential-algebraic equations, International Journal of Control, DOI:
10.1080/00207179.2021.1900604

To link to this article: https://doi.org/10.1080/00207179.2021.1900604

Published online: 23 Mar 2021.

Submit your article to this journal

Article views: 40

View related articles

View Crossmark data

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=tcon20
INTERNATIONAL JOURNAL OF CONTROL
https://doi.org/10.1080/00207179.2021.1900604

Model predictive control for singular differential-algebraic equations

Achim Ilchmann, Jonas Witschel and Karl Worthmann
TU Ilmenau, Ilmenau, Germany

ABSTRACT ARTICLE HISTORY

We study model predictive control for singular differential-algebraic equations with higher index. This is Received 13 August 2019
a novelty when compared to the literature where only regular differential-algebraic equations with addi- Accepted 12 February 2021
tional assumptions on the index and/or controllability are considered. By regularisation techniques, we KEYWORDS
are able to derive an equivalent optimal control problem for an ordinary differential equation to which Differential-algebraic
well-known model predictive control techniques can be applied. This allows the construction of terminal equations; optimal control;
constraints and costs such that the origin is asymptotically stable w.r.t. the resulting closed-loop system. model predictive control;
regularisation

1. Introduction Most approaches to MPC for DAEs do not explicitly exploit

Differential-algebraic equations (DAEs) play an important role the structure of the DAE, but treat it as an additional constraint
in the modelling of electrical networks, mechanical multi-body of the optimal control problem (OCP) (Bock et al., 2007; Diehl
systems or chemical plants (S. Campbell et al., 2019). To derive et al., 2002). Exploiting the structure of the DAE before discreti-
the system model, often automatic modelling techniques are sation is done by Yonchev et al. (2004), Sjöberg et al. (2007), and
employed (Riaza, 2008). This may lead to singular differential- Ilchmann et al. (2019). However, these results are only applica-
algebraic systems with higher index which pose special chal- ble to regular DAEs, and in case of Yonchev et al. (2004) and
lenges w.r.t. control. Sjöberg et al. (2007) some further controllability assumptions
We want to control DAEs using model predictive con- are imposed.
trol (MPC): this is a control technique widely used to control In the present paper, we consider MPC for arbitrary linear
systems under state and input constraints (Kouvaritakis & Can- time-invariant DAEs without additional regularity or controlla-
non, 2016; Rawlings et al., 2017). To this end, the current state bility assumptions. We follow the scheme presented in Figure 1:
of the system is measured in order to predict and optimise the in order to be able to exploit the structure of the DAE-OCP
future system behaviour on a given (finite) prediction horizon. and to use well-known results from the MPC theory, we reduce
The optimal solution on the first portion of the considered time the DAE-OCP to an ODE-OCP by using numerically advan-
interval is then implemented as a control input at the plant tageous regularisation techniques and transforming the cost
before the whole process is repeated after obtaining a new state functional of the nominal DAE-OCP.
measurement. On the basis of this equivalent ODE-OCP, we construct sta-
While there are innumerable results regarding the stability bilising terminal constraints and costs for the ODE-OCP. These
and robustness of MPC schemes for ordinary differential equa- terminal constraints and costs can then be transformed into ter-
tions (ODEs), few results are known for systems governed by minal constraints and costs for the original DAE-OCP. While
DAEs. The main challenge for this class of systems is the fact that the preceding steps of regularisation and construction of the ter-
input and state cannot be considered separately, so approaches minal ingredients for the ODE-OCP are only adapted by us, the
known from ODE systems based on calculating a stabilising last step is a novelty. Its advantage is that it allows to use numer-
state feedback do not work in general. ically feasible schemes to solve the DAE-OCP, while retaining
There has been a lot of research in the related field of the stability guarantees which are offered by the appropriate
optimal control for DAEs, both in an analytical context using construction of the terminal ingredients.
Riccati (S. L. Campbell et al., 2012; Cobb, 1983; Kunkel The paper is structured as follows: in Section 2, we define
& Mehrmann, 2008; Lamour et al., 2013) or Lur’e qua- the problem and propose an MPC scheme as a solution.
tions (Bankmann, 2016; Reis & Voigt, 2019) as well as in In Section 3, we construct suitable terminal constraints for
a numerical context (Gerdts, 2011). However, the analytical the MPC scheme. This is achieved by regularising the DAE as
results do not encompass state or input constraints and to our described in Theorem 3.5, which allows to derive an equiva-
knowledge, none of these results has been extended yet to the lent ODE-OCP in Section 3.2. Our main contribution is the
stability analysis of MPC schemes. The latter requires either new structural approach and the results in Section 4; this
the construction of stabilising terminal constraints and costs or allows to prove asymptotic stability of the MPC scheme for
additional controllability assumptions like cost controllability, the nominal DAE. We conclude with an illustrative example
see, e.g. Coron et al. (2020) and the references therein. in Section 5.

CONTACT Jonas Witschel [email protected] TU Ilmenau, Ehrenbergstraße 29, Ilmenau 98693, Germany
© 2021 Informa UK Limited, trading as Taylor & Francis Group
2 A. ILCHMANN ET AL.

Figure 1. Scheme of the proposed approach for MPC for DAEs followed in this paper.

⎡ ⎤
Notation: L1loc (I, Rp ), p ∈ N, denotes the space of Lebesgue- sEU − AU 0 0 0
measurable functions defined on the interval I ⊆ R that are ⎢ 0 sEJ − AJ 0 0 ⎥
⎢
=⎣ ⎥ , (3)
locally absolutely integrable. In this context, we use the abbrevi- 0 0 sEN − AN 0 ⎦
1,1
ations ae for almost everywhere and aa for almost all. Wloc (I, Rp ) 0 0 0 sEO − AO
denotes the Sobolev space of weakly differentiable functions f :
I → Rp such that f , ḟ ∈ L1loc (I, Rp ). GLn (R) ⊂ Rn×n denotes where
the space of invertible real matrices. For M ∈ Rn×n , M > 0
(M ≥ 0) means that M is positive (semi-)definite. 1 is the all- • sEU − AU ∈ R[s]U ×nU , 0 ≤ U < nU , ∀ λ ∈ C : rk(λEU −
ones vector. For x ∈ Rn , x ≤ 1 means that xi ≤ 1 holds for all AU ) = U ,
i ∈ {1, . . . , n}. • sEJ − AJ ∈ R[s]nJ ×nJ , rk(EJ ) = nJ ,
• sEN − AN ∈ R[s]nN ×nN , ∀ λ ∈ C : rk(λEN − AN ) = nN , EN
nilpotent,
2. Problem formulation: MPC for DAEs • sEO − AO ∈ R[s]O ×nO , O > nO ≥ 0, ∀ λ ∈ C : rk(λEO −
We consider the differential-algebraic equation system AO ) = nO .
[E, A, B] ∈  := R×n × R×n × R×m associated to
The block sizes U , nU , nJ , nN , O , and nO are uniquely
d determined.
(Ex(t)) = Ax(t) + Bu(t). (1) The index of a pencil sE−A is given by
dt
i
The system [E, A, B] is called regular if, and only if, the linear ind(sE − A) := nil ind(EN ) := min{i ∈ N | EN = 0}.
matrix pencil sE−A is regular, i.e.  = n and there exists λ ∈
C such that det(λE − A) = 0; otherwise, the system [E, A, B] is It can be shown that the index does not depend on the choice of
called singular. The behaviour transformation into quasi Kronecker form. We also define the
index of the system [E, A, B] ∈  as the index of the correspond-
 
1 n+m 
 ing pencil sE−A.
B[E,A,B] := (x, u) ∈ Lloc (R≥0 , R ) Let the mixed state and control constraints be given by

1,1
Ex ∈ Wloc (R≥0 , R ), x(t) ae
[F G] ≤1 (4)
(1) holds for almost all t ≥ 0 u(t)

of the system [E, A, B] ∈  is the set of all solutions of (1). An with F ∈ Rp×n and G ∈ Rp×m . The constraints are only
initial value x0 ∈ Rn is called weakly consistent if, and only if, a required to be fulfilled almost everywhere because x need not be
solution (x, u) ∈ B[E,A,B] exists with continuous. If F = F E holds for some F ∈ Rp× , then the con-
straints hold everywhere since Ex is continuous by definition
(Ex)(0) = Ex0 . (2) of B[E,A,B] .

The space of weakly consistent initial values is denoted by

2.1 Singular DAEs do not allow stabilisation by state
C[E,A,B] ⊆ R . n feedback
Consider the singular system (1) given by E = (0, 1), A = (1, 0),
For every pencil sE − A ∈ R[s]×n , there exist T ∈ GL (R), and B = 1. Its behaviour is given by
U ∈ GLn (R) that transform it into a standard form, the so- ⎧⎛ ⎞  ⎫
called quasi Kronecker form (Berger & Trenn, 2012): ⎨ x1  ⎬

B[E,A,B] = ⎝x2 ⎠ ∈ L1loc (R≥0 , R3 ) ẋ2 = x1 + u .
ae
⎩  ⎭
T(sE − A)U u
 INTERNATIONAL JOURNAL OF CONTROL 3

Here, x1 is a free variable that can be potentially unbounded. Algorithm: Input param-
It cannot be influenced by the control input u, which shows eters: [E, A, B] ∈ , F ∈ Rp×n , G ∈ Rp×m ,
that for singular DAEs, prescribing u is generally not sufficient T > 0, S ∈ R(n+m)×(n+m) , Xf ⊆ C[E,A,B] ,
to achieve convergence limt→∞ x(t) = 0 of the state: for all Vf : Xf → R≥0 . Set k = 0.
choices u ∈ L1loc (R≥0 , R), there exist solutions (x, u) ∈ B[E,A,B]
with unbounded x, e.g. Step 1: Measure x̂k := (Ex)(kδ).
T    
 t Step 2: Minimise 0 ūx̄kk(s) S ūx̄kk(s)
(s) (s) ds + Vf (x̄k (T)) s.t.
x1 (t) = et , x2 (t) = x1 (τ ) + u(τ ) dτ , t ≥ 0. d
0 • dt (Ex̄k )(t) = Ax̄k (t) + Būk (t)

Hence, in general it is necessary to have control over both the • F x̄k (s) + Gūk (s) ≤ 1 for almost all s ∈ [0, T]
input u as well as the (free variable part of the) state x for sin-
• x̄k (T) ∈ Xf , (Ex̄k )(0) = x̂k .
gular DAEs to achieve stabilisation of the state. Moreover, in
Step 3: Implement first piece (x̄k∗ (·), ū∗k (·))|[0,δ) of optimal solu-
contrast to ODEs, it is generally impossible to construct a sta-
tion for system [E, A, B] to obtain (x(·), u(·))|[kδ,(k+1)δ] ,
bilising linear state feedback: Let u = kx for some arbitrary but
set k := k + 1, go to Step 1.
fixed k ∈ R1×2 . Then the closed-loop system has the form
d To construct suitable terminal constraints and terminal costs,
(0, 1)x(t) = (1 + k1 , k2 )x(t), we will consider the optimal control problem from Step 2 of
dt
the algorithm with infinite optimisation horizon T = ∞ and
which is still singular. We obtain that without the mixed state and control constraints (4). Using regu-
  larisation techniques, we will transform the DAE constraint (1)
x1 1

2 
B[(0,1),(1+k1 ,k2 ),0] = ∈ Lloc (R≥0 , R ) into an equivalent ODE constraint so that we can solve the OCP
x2
by well-known Riccati theory. The optimal value of this OCP

x1 arbitrary, will then serve as the terminal cost Vf : Xf → R≥0 , which fulfils
ae .
ẋ2 (t) = k2 x2 (t) + (1 + k1 )x1 (t) the decrease condition

Therefore, the closed-loop system admits unbounded solutions ∀ δ > 0 ∀ x̂ ∈ C[E,A,B] : Vf (x(δ))
no matter how the state feedback is chosen.  δ
x(t) x(t)
≤ Vf (x̂) − S dt (5)
0 u(t) u(t)
2.2 Model predictive control by virtue of the Belman equation. To construct a controlled
Our goal is to construct a feedback law such that the origin is forward invariant terminal region Xf ⊆ C[E,A,B] , a sub-level set
asymptotically stable w.r.t. the closed-loop system while valid- of Vf where the state and control constraints (4) are fulfilled
ity of the constraints (4) is maintained. To this end, we employ can be chosen. Together with the decrease condition (5), asymp-
the following MPC scheme. In every step, we measure the cur- totic stability of the origin w.r.t. the MPC closed loop can be
rent state Ex of the system (1) and solve a quadratic optimal shown analogously to the classical ODE case, see, e.g. Mayne
control problem on the optimisation/prediction horizon T > 0. et al. (2000).
The OCP is constrained by the system (1) with the current
system state as an initial value (2) and the mixed state and 3. Construction of terminal ingredients
control constraints (4). Clearly, solving the OCP on an infinite-
time horizon, i.e. T = ∞, would be desirable. However, due To construct the terminal region and terminal costs, we reduce
to the (mixed) state and control constraints, this is, in general, the DAE to an equivalent ODE by reducing its index to 1.
computationally intractable. In order to do so, we transform the DAE using the methods
We obtain a predicted optimal solution (x̄0∗ (·), ū∗0 (·))|[0,T] , explained in the next subsection so that it fulfils the following
indicated by ¯·, of which we implement only the first piece algebraic characterisation.
(x̄0∗ (·), ū∗0 (·))|[0,δ) up to a time shift δ ∈ (0, T) at the plant. Note Proposition 3.1 (Berger & Reis, 2013, Equation (3.4)): The
that the resulting closed-loop solution only coincides with the system [E, A, B] ∈  has index at most one if, and only if,
predicted one during the first δ time units but may differ after-
wards. imA ⊆ imE + A ker E.
After the time δ has passed, we repeat the procedure with the
new system state. If the DAE has index at most one, it can be transformed to
This scheme alone would not necessarily allow for a sys- an ODE with a state transformation.
tem that is asymptotically stable w.r.t. the origin, see Rawlings Proposition 3.2 (Benner et al., 2015, Thm. 8.1): The sys-
et al. (2017, Sec. 1.3.4). To guarantee stability, we incorpo- tem [E, A, B] ∈  is regular with index at most one if, and only
rate additional terminal constraints Xf ⊆ C[E,A,B] and terminal if, there are transformation matrices Sr , Tr ∈ GLn (R) such that
costs Vf : Xf → R≥0 into the basic MPC scheme to obtain the      
following algorithm. These additional degrees of freedom guar- I 0 A A12 B
Sr ETr = n̂ , Sr ATr = 11 , Sr B = 1 ,
antee asymptotic stability if chosen suitably, as we will show in 0 0 A21 A22 B2
this paper. (6)
4 A. ILCHMANN ET AL.

with n̂ ≤ n, A22 ∈ GLn−n̂ (R). 3.1 Proof of Theorem thm3.5

We use two different approaches to regularise the system (1):
A remark to Proposition 3.2 is warranted. If the system fulfils a certain rank condition, we use the method
by Bunse-Gerstner et al. (1992, 1994) to find a regularising feed-
Remark 3.3: E can be transformed to the form as in (6) using back. If it is not, we use the more general approach proposed
a singular value decomposition: Choose orthogonal matrices by Berger and Van Dooren (2015). While the latter approach is
U ∈ Rn×n and V ∈ Rn×n such that E = Udiag(σ1 , . . . , σn̂ , 0, also applicable to the case where the rank condition is fulfilled,
. . . , 0)V for σ1 , . . . , σn̂ > 0. Then Sr := U , Tr := Vdiag the former is in this case numerically more attractive.
(σ1−1 , . . . , σn̂−1 , 1, . . . , 1) leads to (6).
Definition 3.6 (Berger & Reis, 2013, Def. 2.1): [E, A, B] ∈  is
called impulse controllable if, and only if,
If [E, A, B] ∈  is regular and has index at most one, then
the DAE (1) can be transformed into an ODE with an explicit ∀ x0 ∈ Rn ∃ (x, u) ∈ B[E,A,B] : Ex0 = Ex(0).
representation of the remaining states. This is made precise
in the following lemma, which is an immediate consequence Proposition 3.7 (Berger & Reis, 2013, Rem. 4.2): If [E, A, B] ∈
of Proposition 3.2.  is regular, then it is impulse controllable if, and only if,

Lemma 3.4: Let [E, A, B] ∈  be regular with index at most one rk[E, AZ, B] = n for some Z ∈ Rn×(n−rkE) , imZ = ker E.
and consider the transformation into (6). Then it holds that (7)

x Next, we present a well-known technique to regularise the

∈ B[E,A,B] ⇐⇒ x = Tr z system (1).
u

Proposition 3.8: [E, A, B] ∈  fulfils (7) if, and only if, there
with z = (z1 , z2 ), where z2 = −A−1 −1
22 A21 z1 − A22 B2 u and z1 exists a feedback matrix K ∈ Rn×m such that [E, A + BK, B] is
solves
regular and has index at most one. Moreover, the behaviours are
linked by
ż1 = (A11 − A12 A−1 −1
22 A21 )z1 + (B1 − A12 A22 B2 )u
x x
for some u ∈ L1loc (R≥0 , Rm ). ∈ B[E,A,B] ⇐⇒ ∈ B[E,A+BK,B] ,
u v
where u = Kx + v.
Lemma 3.4 essentially allows to reformulate the DAE-OCP
as an ODE-OCP, provided that the DAE system is regular and Proof: Sufficiency of the first assertion is proved in Bunse-
has index at most one. Using the regularisation techniques Gerstner et al. (1992, Theorem 6), while necessity is shown
shown in the following theorem allows to transform every DAE in Bunse-Gerstner et al. (1994, Theorem 4). The second state-
system (1) into the form form (6). ment is immediate. 

Theorem 3.5: Consider [E, A, B] ∈ . Then there exist 

T∈ The behaviour of the nominal system (1) and the regularised
GLn+m (R) and unimodular U(s) = sU1 + U0 ∈ R[s]× , U0 , system is coupled by an input transformation.
U1 ∈ R× , such that If [E, A, B] ∈  does not fulfil (7), we will conduct the reg-
⎡ ⎤ ularisation by special unimodular transformations. A polyno-
0 0 0 mial matrix U(s) ∈ R[s]n×n is called unimodular if, and only
T = U(s) ⎣sIn̂ − A11
[sE − A, −B] −A12 −B1 ⎦ , if, U −1 (s) ∈ R[s]n×n exists such that U(s) U −1 (s) = In . Clearly,
−A22 −A22 −B2 U(s) is unimodular if, and only if, its determinant is a constant
nonzero polynomial.
where n̂ ≤ n, A22 ∈ GLn−n̂ (R). In the next proposition, we show that the nominal system (1)
If [E, A, B] additionally fulfils (7), then 
T and U(s) can be can be regularised by an input transformation.
chosen as
  Proposition 3.9: For any system [E, A, B] ∈  there exist 
T∈
 Tr 0 GLn+m (R) and unimodular U(s) = sU1 + U0 ∈ R[s]× , U0 ,
T= , U(s) = S−1
KTr Im r U1 ∈ R× , such that
 
for some K ∈ Rm×n , Sr , Tr ∈ GLn (R).  0 0
[sE − A, −B]T = U(s) , (8)
sEr − Ar −Br

Proof: This will be proved in the following Section 3.1: The first where sEr − Ar ∈ R[s]r×r , r ≤  is regular and has index at most
part of the assertion follows immediately from Proposition 3.9 one, Br ∈ R×(n+m−r) . Then, the following implications hold
in combination with Proposition 3.2, while the second part is a
x x
direct consequence of Propositions 3.8 and 3.2.  (i) u ∈ B[E,A,B] ⇒ 
T −1 u ∈ B[Er ,Ar ,Br ] .
 INTERNATIONAL JOURNAL OF CONTROL 5

x
(ii) If ∈ B[Er ,Ar ,Br ] satisfies
u In the following example a singular DAE (with over- and
    underdetermined blocks) is considered.
0 0 0 0 x 1,1
U0 − U1 ∈ Wloc ,
Er 0 Ar Br u Example 3.11: Consider the singular system
  ⎡ ⎤ ⎡ ⎤ ⎛ ⎞
then 
T ux ∈ B[E,A,B] holds. d ⎣
0 1 0 1 0 0 0
0 0 0⎦ x(t) = ⎣0 0 1⎦ x(t) + ⎝0⎠ u(t). (10)
dt 0 0 1
Proof: The first part of the proposition is shown in Berger 0 0 0 1
and Van Dooren (2015). Note that it also follows from Berger
and Van Dooren (2015) that U(s), Er and Ar are chosen such Using Proposition 3.9, the transformation
that no quadratic term occurs on the right-hand side of (8). ⎛ ⎞ ⎡ ⎤⎛ ⎞
Assertion (ii) can be immediately concluded from y1 0 0 1 0 x1
⎜y2 ⎟ ⎢0 1 0 0⎥ ⎜x2 ⎟
    ⎜ ⎟ := ⎢ ⎥⎜ ⎟
0 0 0 0 ⎝y3 ⎠ ⎣0 0 0 1⎦ ⎝x3 ⎠
[sE − A, −B] T = s U0 − U1 v 1 0 0 0 u
Er 0 Ar Br
 
0 0 yields the regular index 1 system
− U0 .
Ar Br
⎡ ⎤ ⎡ ⎤ ⎛ ⎞
  d ⎣
0 0 0 1 0 0 0
To show Assertion (i), let ux ∈ B[E,A,B] be arbitrary, so by 0 1 0⎦ y(t) = ⎣0 0 0 ⎦ y(t) + ⎝1⎠ v(t).
definition Ex ∈ Wloc1,1
(R≥0 , R ). We need to show that [Er , dt 0 0 0 0 0 −1 0
 
0r×(n+m−r) ]T −1 ux ∈ Wloc 1,1
(R≥0 , Rr ). By following the proof
outlined in Berger and Van Dooren (2015), we obtain Hence the equivalent ODE is given by
  ż1 (t) = v(t), z2,1 (t) = z2,2 (t) = 0,
 0 0
ker[E, 0×m ]T ⊆ ker .
Er 0
where z1 = y2 , z2,1 = y1 , z2,2 = y3 . We see that in the singular
Hence system (10), x1 can be chosen freely and is therefore more of an
  input than a state, which is reflected by the regularised system
0 0 −1 x where v = x1 = z1 . In turn, the ‘input’ u is 0 almost everywhere,
T
Er 0 u hence it is a state u = y3 = z2,2 of the regularised system.
 
x 0 0 x
= [E, 0] + − [E, 0]
T 
T −1 ,
u Er 0 u 3.2 Optimal control for DAEs without constraints
!" # !" #
1,1
∈Wloc (R≥0 ,R ) =0
We consider the optimal control problem
which proves Assertion (i).   T
x(t) x(t)
Min S dt subject to (1) (11)
0 u(t) u(t)
The next example illustrates Proposition 3.9 for a regular
index two DAE. with optimisation horizon T ∈ R ∪ {∞}, S=S ∈
R(n+m)×(n+m) , and cost functional
Example 3.10: Consider the regular DAE
    JT : B[E,A,B] → R ∪ {±∞},
d 0 1 1 0
x(t) = x(t). (9)  T
dt 0 0 0 1 x(t) x(t)
(x, u) → S dt.
0 u(t) u(t)
Its behaviour is given by (x, u) ∈ L1loc (R≥0 , R2 ) satisfying x2 ∈
1,1 ae
Wloc (R≥0 , R), x1 = 0, and x2 = 0. Setting The value function for the OCP without state and control con-
  straints is defined by
 −1 s
T = I2 , U(s) = , VT : C[E,A,B] → R ∪ {±∞}
0 −1
&
x0 → inf JT (x, u) | (x, u) ∈ B[E,A,B] (12)
the system (9) can be transformed, as described in Proposition
'
3.9, into the regular index 1 system [02×2 , I2 ]. Its behaviour is with (Ex)(0) = Ex0 .
  
x  ae
∈ L1loc (R≥0 , R2 ) x1 = 0, x2 = 0 . In order to solve this OCP, we transform it into an equiva-
ae
u lent OCP that is constrained by an ODE instead of the nominal
DAE with the help of Theorem 3.5. To this end, define for
Note that the component x2 is required to be smoother in the [E, A, B] ∈ ,
nominal DAE than in the regularised DAE, as it needs to be 0
everywhere instead of almost everywhere. 
A := A11 − A12 A−1 B := B1 − A12 A−1

22 A21 , 22 B2 , (13)
6 A. ILCHMANN ET AL.

⎡ ⎤
where A11 , A12 , A21 , A22 , B1 and B2 are defined as in Theo- In̂ 0
z1
rem 3.5. ×
T S
T ⎣−A−1
22 A21 −A−1
22 B2
⎦
Then the ODE–OCP we want to consider is given by the cost v
0 Im
functional
z1  z

JT : B[I → R ∪ {±∞}, = S 1 ,
n̂ ,Â,B̂] v v
 T
z1 z1 (t)  z (t) therefore
→ S 1 dt,
v 0 v(t) v(t)
x z1
JT =
JT .
where, for 
T as in (8), u v
⎡ ⎤
In̂ 0 Now following the same argument as Ilchmann et al. (2019,

S := X −1 SX, X := 
T ⎣−A−1 −A−1 ⎦ (14) Thm. 1), we conclude that
22 A21 22 B2
0 Im
T (x0 ),
∀ x0 ∈ C[E,A,B] : VT (x0 ) = V
using the notation from Theorem 3.5. Its value function is
given by which proves the assertion.
If instead the additional condition (ii) holds, we assume with-
T : C[E,A,B] → R ∪ {±∞}
V out loss of generality that [E, A] is given in Kronecker form,
⎧  ⎫ i.e.
⎨  (z1 , v) ∈ B
 [In̂ ,Â,B̂] with ⎬
x0 → inf JT (z1 , v)  −1 {x } ⎭ .
0
s[E, 0] − [A, B]
⎩  z1 (0) ∈ [In̂ , 0]X Rm ⎡ ⎤
(15) diag(Kn1 , . . . , Knu ) 0 0
= s⎣ 0 InJ 0 0 ⎦
We obtain that the value function (12) of the DAE-OCP and its 0 0 N
counterpart (15) of the ODE-OCP coincide, as detailed in the ⎡ ⎤
diag(Ln1 , . . . , Lnu ) 0 0
following theorem.
−⎣ 0 J 0 B ⎦
Theorem 3.12: Consider [E, A, B] ∈  and the optimal control 0 0 InN
problems defined by (12) and (15). Then ∈ R[s]×(+u+m) ,
T .
VT ≡ V where
If, in addition, one of the following conditions holds: ⎡ ⎤
s 1 0
⎢ .. .. ⎥ (n −1)×ni
sKni − Lni = ⎣ . . ⎦∈R i ,
(i) [E, A, B] ∈  is regular or
(ii) the extended system [[E, 0], [A, B]] does not contain an 0 s 1
overdetermined part, i.e. sEO − AO is void in any transfor- i = 1, . . . , u and N nilpotent.
mation of [[E, 0], [A, B]] into quasi Kronecker form (3),
Writing
then, for 
T and X as in (8) and (14), resp.,
  ei := (01×(i−1) , 1, 01×(n1 +···+nu −i) ) ∈ Rn1 +···+nu ,
{x0 }  −1 x
0
[In̂ , 0]X −1 = [I , 0]T , (16) i = 1, 2, . . . , n1 + · · · + nu
Rm n̂
0  
  I 0
Er := n̂ ∈ R× , n̂ :=  − nN ,
i.e. we can replace the initial constraint z1 (0) ∈ [In̂ , 0]X −1 {x0 }
0 0nN ×nN
 0 Rm
⎡ ⎤
by an initial condition z1 (0) = [In̂ , 0]
T −1 x . 0 diag(Nn1 −1 , . . . , Nnu −1 ) 0 0
Ar := ⎣ 0 J 0 ⎦ ∈ R× ,
Proof: Consider (x, u) ∈ B[E,A,B] . By Theorem 3.5 and 0 0 InN
Lemma 3.4, it follows for  
⎡ ⎤−1 −en1 −1 . . . −e(n1 −1)+···+(nu −1)
Br := B
z1
In̂ 0 0(nJ +nN −u)×1 . . . 0(nJ +nN −u)×1
−1 ⎦ −1 x x
:= ⎣−A−1 22 A21 −A22 B2 T = X −1
v
0 Im
u u ∈ R×(u+m) ,
⎡
z  x diag(Kn1 , . . . , Knu ) 0 en1 . . .
that v1 ∈ B[I ,Â,B̂] . Let u ∈ B[E,A,B] be arbitrary. Substitut- 
n̂ T := ⎣ 0 InJ +nN 0 ...
ing this in (11) yields 0 0 0 ...
⎡ ⎤ ⎤
In̂ 0 en1 +···+nu 0
x x z1 ⎣−A−1 A21 −A−1 B2 ⎦
S = 22 22 0 0 ⎦ ∈ R(+u+m)×(+u+m) ,
u u v
0 Im 0 Im
 INTERNATIONAL JOURNAL OF CONTROL 7

⎡ ⎤
In1 +···+nu −u 0 0 we impose some standard assumptions on the ODE–OCP
U(s) := ⎣ 0 InJ 0 ⎦ ∈ R[s]× (Lancaster & Rodman, 1995): for the system [E, A, B] ∈ ,
0 0 −(sN − InN ) define

A := A11 − A12 A−1 B := B1 − A12 A−1

yields 22 A21 , 22 B2 (17)
[sE − A, −B]T where Aij , Bi , i ∈ {1, 2} are given by (6). Furthermore partition
S,
⎡ ⎤ defined in (14), as
I(n1 −1)+···+(nu −1) 0 0
 
= s⎣ 0 InJ 0 0 0 ⎦ 
Q H 

S :=  , Q ∈ Rn̂×n̂ . (18)
0 0 N H 
R
⎡
diag(Nn1 −1 , . . . , Nnu −1 ) 0 0 Then we require the following assumption.
−⎣ 0 J 0
0 0 InN Assumption 3.13: Assume that the following properties hold
⎤ for  B, 
A,  S as defined in (17) and (18).
en1 −1 . . . e(n1 −1)+···+(nu −1)
0(nJ −u)×1 . . . 0(nJ −u)×1 −B ⎦ 
• S ≥ 0,
0nN ×1 ... 0nN ×1
⎛ ⎡ ⎤ • the pair (
A, 
B) is stabilisable,
I(n1 −1)+···+(nu −1) 0 0 • R = [0, In−n̂ ]
 S[0, In−n̂ ] > 0,
= U(s) ⎝s ⎣ 0 InJ 0 0 0 ⎦ • (
A, 
Q) is observable,
0 0 0nN ×nN • rk
S = rk( Q + R).
⎡
diag(Nn1 −1 , . . . , Nnu −1 ) 0 0
Proposition 3.14 (Lancaster & Rodman, 1995, Prop. 16.2.8):
−⎣ 0 J 0
Assume that Assumption 3.13 holds, and consider for  B, 
A,  ,
Q, H
0 0 InN 
⎤⎞ R as defined in (17) and (18) the algebraic Riccati equation
en1 −1 . . . e(n1 −1)+···+(nu −1)

A  P
P +A+
Q − (
P R−1 (
)
B+H P ) = 0.
B+H (19)
0(nJ −u)×1 . . . 0(nJ −u)×1 −B ⎦⎠
0nN ×1 ... 0nN ×1
Then this equation has a unique solution 
P =
P ∈ Rn̂×n̂ and
= U(s)[sEr − Ar , −Br ]. this solution satisfies 
P > 0.

The system [Er , Ar , Br ] ∈ R× × R× × R×(m+u) is regular We obtain the desired existence and uniqueness result on the
with index 1. Since  T −1 = T , it follows that optimal control.
⎡ ⎤−1
(14)
In̂ 0 Proposition 3.15: Consider the system [E, A, B] ∈ , and
[In̂ , 0]X −1 = [In̂ , 0n̂×(u+m) ] ⎣0nN ×n̂ ∗ ⎦  T −1 assume that Assumption 3.13 holds. Then
0 Im+u
  x∗
∀ x0 ∈ C[E,A,B] ∃ unique ∈ B[E,A,B] : (Ex∗ )(0)
= [In̂ , 0n̂×(u+m) ]
In̂ 0n̂×nN 0 −1
T u∗
0 0(m+u)×nN Im+u
( ) = Ex0 ∧ JT (x∗ , u∗ ) = VT (x0 ),
= In̂ 0n̂×(nN +m+u)  T
( ) Proof: Consider the ODE–OCP
= diag(Kn1 , . . . , Knu ) 0n̂×(nJ +nN +m) .
 T
z1 (t)  z (t)
Therefore, Minimise S 1 dt (20)
0 v(t) v(t)
x0 x0
∀ x0 ∈ Rn ∀ u0 ∈ Rm : [In̂ , 0]X −1 = [In̂ , 0]X −1 , s.t. ż1 (t) = 
Az1 (t) + 
Bv(t),
u0 0
 {x0 }
which proves (16). z1 (0) ∈ [In̂ , 0] X −1 . (21)
Rm
Conjecture: We strongly believe – although we could not prove
For T < ∞, this optimal control problem has a unique
it – that (16) holds without any of the assumptions (i) or (ii) in
solution (z∗ , u∗ ) ∈ C ∞ (R≥0 , Rn × Rm ) according to Lancaster
Theorem 3.12.
and Rodman (1995, Theorem 16.4.2); for T = ∞, the same
We have transformed the DAE-OCP into an equiva- result follows from Lancaster and Rodman (1995, Theorem  ∗
lent ODE-OCP. In order to ensure existence and uniqueness of 16.3.3). According to Proposition 3.9, it follows that T̂ uz ∗ ∈
an optimal control trajectory, i.e. B[E,A,B] , hence

x∗ T (x0 ) =  x∗ z∗
∀ x0 ∈ C[E,A,B] ∃ unique ∈ B[E,A,B] ∩ C ∞ (R≥0 , Rn+m ) : VT (x0 ) = V JT = JT T .
u∗ u∗ u∗

(Ex∗ )(0) = Ex0 ∧ JT (x∗ , u∗ ) = VT (x0 ), This proves the assertion. 

8 A. ILCHMANN ET AL.

Furthermore, we can prove the Bellman equation for For (x̃, ũ), it follows that (Ex̃)(0) = Ex0 and for t ≥ 0,
the DAE–OCP:
F x̃(t) + Gũ(t)∞
Proposition 3.16: Consider *   *
 system [E, A, B] ∈ .
 x∗the Let x0 ∈ *( ) I *
∗
C[E,A,B] be arbitrary and u∗ ∈ B[E,A,B] with (Ex )(0) = x0 *
=* F G  T −1    x̂(t)*
n̂
 ∗  −R (B P + H ) *
be an optimal trajectory, i.e. J∞ ux∗ = V∞ (x0 ). Then for all ∞
*  *
T > 0, it holds that *( ) In̂ *
≤* * F G T −
 *
) * x̂(t)2
R−1 (
B  P+H
x∗ x∗ (T) ∞
V∞ (x0 ) = JT
u∗
+ V∞
u∗ (T)
. (22) * * +
*( ) *
* x̂(t)
Px̂(t)
≤* * F G T* λmin ( P)
Proof: This follows as in Ilchmann et al. (2018, Th. 9).  ∞
≤ 1.
4. MPC: asymptotic stability of the origin This shows that Xf is controlled forward invariant.
To prove asymptotic stability of the origin w.r.t. the MPC scheme Satisfaction of the decrease condition (5) follows immedi-
from Section 2.2, we employ the equivalent ODE constructed ately for the optimal solution of the DAE–OCP (11) (guar-
in Lemma 3.4. anteed to exist by Proposition 3.15). In light of the Bellman
equation (22), the terminal cost is simply given by the optimal
Definition 4.1: The set Xf ⊆ Rn is called controlled forward cost V∞ . 
invariant w.r.t. the system [E, A, B] ∈  if, and only if,
5. Example
∀ x0 ∈ Xf ∃ (x, u) ∈ B[E,A,B]
aa Minimise the cost functional
∀ t ≥ 0 : x(t) ∈ Xf ∧ Ex0 = Ex(0).  T
x(t)2 + u(t)2 dt
The following theorem states that the optimal solution fulfils 0
the condition in the preceding definition.
subject to the singular DAE
⎡ ⎤
Theorem 4.2: Consider [E, A, B] ∈  with constraints (4). Let 0 0 0 0 0
the transformation matrix T be defined as in (8), and  , 
B, H R be ⎢ 1 0 0 0 0 ⎥
defined by (17) and (18). Denote by 
P the solution of the algebraic d ⎢⎢ 0 0 0 0
⎥
⎥ x(t)
0
Riccati equation (19). Define dt ⎢
⎣ 0 0 1 0 0
⎥
⎦
*  *−2 0 0 0 0 1
* In̂ * ⎡ ⎤ ⎛ ⎞
ρ := λmin (P) *[F G]T −1    *
 *  > 0.
−R (B P + H ) *∞ 1 0 0 0 0 0
⎢ 0 0 0 0 0 ⎥ ⎜0⎟
⎢ ⎥ ⎜ ⎟
=⎢ ⎥ x(t) + ⎜0⎟ u(t)
Then the set ⎢ 0 1 0 0 0 ⎥ ⎜ ⎟
     ⎣ 0 0 −1 0 0 ⎦ ⎝0⎠
In̂ 
 
Xf := [In , 0]T −1    x̂ x̂ ∈ R ∧ x̂ 
n̂
Px̂ ≤ ρ 0 0 0 1 0 1
−R (B P + H )
(23) and the initial condition

is controlled forward invariant. Moreover, the optimal solution (Ex)(0) = Ex0 .

consisting of x∗ and u∗ satisfies the decrease condition (5) almost
everywhere. The ODE obtained from regularisation is given by
   
−1 0 0 0
Proof: Let x0 ∈ Xf be arbitrary, and let (x, u) ∈ B[E,A,B] be any ż1 (t) = z1 (t) + v(t), z2 (t) = 0, (24)
0 0 1 1
solution with (Ex)(0) = x0 . Consider the solution of the ODE
where z1,1 = x3 , z1,2 = x5 , z2,1 = x1 , z2,2 = −x2 , v1 = x4 , v2 =
˙ = [ x0
x̂(t) A −R−1 (
B B  )]x̂(t),
P+H x̂(0) = X −1 . u, and the equivalent OCP is
u(0)
 T
By the Bellman equation, it holds that Minimise z1 (t)2 + v(t)2 dt
0
∀ t ≥ 0 : x̂(t) 
Px̂(t) ≤ x̂(0) 
Px̂(0) ≤ ρ. x30
s.t. (24) with z1 (0) = .
x50
Note that x̂ ∈ C ∞ (R≥0 , Rr ), therefore by Proposition 3.8 and
Lemma 3.4, it holds that Obviously Assumption 3.13 is satisfied. √The solution of the alge-
  braic Riccati equation (19) is diag( 12 , 2). For the constraints
x̃ I
= T −1 n̂   x ∈ B[E,A,B] .
ũ −R (B P + H ) −1 ≤ xi (t) ≤ 1, i ∈ {1, . . . , 5}, −1 ≤ u(t) ≤ 1,
 INTERNATIONAL JOURNAL OF CONTROL 9

Figure 2. Closed-loop performance of the MPC scheme and predicted solution of the optimal control problem with added terminal constraints and costs from Step 2 of
the MPC algorithm at time t = kδ = 0.

written in matrix form as MPC closed-loop solution coincides with the (unconstrained)
⎡ ⎤ infinite-horizon optimal solution by construction once the ter-
I5 0 minal region is reached.
⎢−I5 0⎥
⎢ ⎥ x(t) ≤ 1,
⎣ 0 1 ⎦ u(t)
6. Conclusion and open problem
0 −1
In this paper, we describe a way to obtain an MPC scheme for
we obtain a DAE with state and input constraints that guarantees the sta-
*⎡ ⎤*−2 bility of the closed loop w.r.t. the origin. This is achieved by
* 0
* 0 * * regularising the DAE to obtain an ODE optimal control prob-
*⎢0 0 ⎥ *
1 *⎢⎢1
⎥*
⎥* = 1 ,
lem for which a terminal region and costs can be constructed.
ρ= * ⎢ 0
√ ⎥*
2* 4 These terminal ingredients can then be expressed in terms of
*⎣0
* 2⎦** the nominal DAE by a state transformation.
* 0 1 *∞ In the future, we want to investigate whether it is possible to
⎧⎡ ⎤  ⎫
⎪ 0 0  ⎪
achieve similar results without having to resort to a transforma-
⎪
⎪  ⎪
⎪
⎪ ⎢0
⎨ 0 ⎥  ⎪
⎬
tion to an equivalent ODE. This would allow to express Assump-
⎢ ⎥  1 √ 1
Xf = ⎢ ⎥ x̂ x̂ ∈ R ∧ x̂ + 2x̂ ≤
2 2 2 tion 3.13 directly in terms of the DAE: for example, it can
⎪ ⎢1 0
√ ⎥  2 1 2
4⎪
,
⎪ ⎣0
⎪
⎪ 2⎦  ⎪
⎪
⎪
be easily seen that the stabilisability of the equivalent ODE
⎩  ⎭ in Assumption 3.13 is equivalent to the behavioural stabilisabil-
0 1
ity of the nominal DAE (1). We would like to obtain similar
1 √ results for the rest of the assumptions where the situation is
Vf (x̂) = x̂32 + 2x̂52 , x̂ ∈ Xf .
2 much less obvious.
One way to work directly with the DAE is to adapt the
The constructed terminal region Xf and the performance for
approach by Reis and Voigt (2019) to model predictive control:
an MPC scheme with step size δ = 0.1 and prediction hori-
their results allow to characterise the optimal value and optimal
zon T = 3δ are depicted in Figure 2. The states x1 and x2 are
ae solution using so-called Lur’e equations for the DAE. In order
omitted as it follows from the DAE that x1 = 0, x2 = 0. Start- to use these findings for MPC, it is necessary to characterise the
ing with an initial value of (Ex)(0) = (0, 0, 0 − 0.9, −0.55), the positive definiteness of the optimal value in terms of the DAE-
closed-loop solution for x converges to the origin. In addition, OCP. Using these results, a construction similar to Theorem 4.2
the figure also depicts the solution of the optimal control prob- yields a terminal region that, together with the optimal value
lem with added terminal constraints and costs from Step 2 of as terminal costs, guarantees asymptotic stability of the MPC
the MPC algorithm at time t = kδ = 0. It can be seen that the scheme w.r.t. the origin.
state reaches the boundary of the terminal region within the
prediction horizon T = 3δ, since that is mandated by the ter-
Acknowledgments
minal constraint. On the other hand, the actual closed-loop
MPC solution does not reach the terminal region within this The authors are indebted to the German Research Foundation (DFG)
(grants IL25/10-1, WO2056/2-1, RE2917/4-1, and WO2056/6-1) and the
time due to the receding-horizon nature of the MPC scheme. Studienstiftung des Deutschen Volkes for their support. Furthermore we
Once the interior of the terminal region is reached by the thank our colleague Thomas Berger (Paderborn) for constructive sugges-
closed-loop solution, it is never left again. This follows since the tions and discussions.
10 A. ILCHMANN ET AL.

Disclosure statement Cobb, J. D. (1983). Descriptor variable systems and optimal state regu-
lation. IEEE Transactions on Automatic Control, 28, 601–611. https://
No potential conflict of interest was reported by the author(s).
doi.org/10.1109/TAC.1983.1103283
Coron, J. M., Grüne, L., & Worthmann, K. (2020). Model Predictive Con-
Funding trol, Cost Controllability, and Homogeneity. SIAM Journal on Control
and Optimization, 58(5), 2979–2996. https://doi.org/10.1137/19M1265
The authors are indebted to the German Research Foundation (Deutsche 995
Forschungsgemeinschaft (DFG)) (grants IL25/10-1, WO2056/2-1, Diehl, M., Bock, H. G., Schlöder, J. P., Findeisen, R., Nagy, Z., & Allgöwer, F.
RE2917/4-1, and WO2056/6-1) and the Studienstiftung des Deutschen (2002). Real-time optimization and nonlinear model predictive control
Volkes for their support. of processes governed by differential-algebraic equations. Journal of Pro-
cess Control, 12, 577–585. https://doi.org/10.1016/S0959-1524(01)000
23-3
References
Gerdts, M. (2011). Optimal control of ODEs and DAEs. Walter de Gruyter.
Bankmann, D. (2016). On linear-quadratic control theory of implicit Ilchmann, A., Leben, L., Witschel, J., & Worthmann, K. (2019). Optimal
difference equations [Master’s thesis, GermanyFakultät II – control of differential-algebraic equations from an ordinary differential
Mathematik und Naturwissenschaften, Technische Universität equation perspective. Optimal Control Applications and Methods, 40(2),
Berlin]. 351–366. https://doi.org/10.1002/oca.v40.2
Benner, P., Losse, P., Mehrmann, V., & Voigt, M. (2015). Numerical linear Ilchmann, A., Witschel, J., & Worthmann, K. (2018). Model predictive
algebra methods for linear differential-algebraic equations. In A. Ilch- control for linear differential-algebraic equations. IFAC PapersOnLine,
mann & T. Reis (Eds.), Surveys in differential-algebraic equations III (pp. 51, 98–103. https://doi.org/10.1016/j.ifacol.2018.10.1816th IFAC Con-
117–175). Springer-Verlag. ference on Nonlinear Model Predictive Control, Madison, Wisconsin,
Berger, T., & Reis, T. (2013). Controllability of linear differential-algebraic USA.
systems – a survey. In A. Ilchmann & T. Reis (Eds.), Surveys in Kouvaritakis, B., & Cannon, M. (2016). Model predictive control. Springer.
differential-algebraic equations I (pp. 1–61). Springer-Verlag. Kunkel, P., & Mehrmann, V. (2008). Optimal control for unstructured non-
Berger, T., & Trenn, S. (2012). The quasi-Kronecker form for matrix pen- linear differential-algebraic equations of arbitrary index. Mathematics of
cils. SIAM Journal on Matrix Analysis and Applications, 33(2), 336–368. Control, Signals, and Systems, 20, 227–269. https://doi.org/10.1007/s00
https://doi.org/10.1137/110826278 498-008-0032-1
Berger, T., & Van Dooren, P. (2015). Computing the regularization of a lin- Lamour, R., März, R., & Tischendorf, C. (2013). Differential-algebraic equa-
ear differential-algebraic system. Systems & Control Letters, 86, 48–53. tions: A projector based analysis. Springer-Verlag.
https://doi.org/10.1016/j.sysconle.2015.10.003 Lancaster, P., & Rodman, L. (1995). Algebraic Riccati equations. Clarendon
Bock, H. G., Diehl, M., Kostina, E., & Schlöder, J. P. (2007). Constrained Press.
optimal feedback control of systems governed by large differential alge- Mayne, D. Q., Rawlings, J. B., Rao, C. V., & Scokaert, P. O. (2000). Con-
braic equations. In L. T. Biegler, O. Ghattas, M. Heinkenschloss, D. strained model predictive control: Stability and optimality. Automatica,
Keyes, & B. van Bloemen Waanders (Eds.), Real-time PDE-constrained 36(6), 789–814. https://doi.org/10.1016/S0005-1098(99)00214-9
optimization (pp. 3–24). SIAM. Rawlings, J. B., Mayne, D. Q., & Diehl, M. M. (2017). Model predictive
Bunse-Gerstner, A., Mehrmann, V., & Nichols, N. K. (1992). Regulariza- control: Theory, computation, and design (2nd ed.). Nob Hill Publishing.
tion of descriptor systems by derivative and proportional state feed- Reis, T., & Voigt, M. (2019). Linear-quadratic optimal control of
back. SIAM Journal on Matrix Analysis and Applications, 13(1), 46–67. differential-algebraic systems: The infinite time horizon problem with
https://doi.org/10.1137/0613007 zero terminal state. SIAM Journal on Control and Optimization, 57(3),
Bunse-Gerstner, A., Mehrmann, V., & Nichols, N. K. (1994). Regular- 1567–1596. https://doi.org/10.1137/18M1189609
ization of Descriptor Systems by Output Feedback. IEEE Transactions Riaza, R. (2008). Differential-algebraic systems: Analytical aspects and circuit
on Automatic Control, 39(8), 1742–1748. https://doi.org/10.1109/9.310 applications. World Scientific Publishing.
065 Sjöberg, J., Findeisen, R., & Allgöwer, F. (2007). Model Predictive Control of
Campbell, S., Ilchmann, A., Mehrmann, V., & Reis, T. (Eds.). (2019). Appli- Continuous Time Nonlinear Differential Algebraic Systems. IFAC Pro-
cations of differential-algebraic equations: Examples and benchmarks. ceedings Volumes, 40(12), 48–53. https://doi.org/10.3182/20070822-3-
Springer. ZA-2920.00008
Campbell, S. L., Kunkel, P., & Mehrmann, V. (2012). Regularization of lin- Yonchev, A., Findeisen, R., Ebenbauer, C., & Allgöwer, F. (2004). Model
ear and nonlinear descriptor systems. In L. T. Biegler, S. L. Campbell, & predictive control of linear continuous time singular systems subject to
V. Mehrmann (Eds,), Control and optimization with differential-algebraic input constraints. In Proceedings of the 43rd IEEE conference decision
constraints (pp. 17–36). SIAM. control (CDC) (Vol. 2, pp. 2047–2052). IEEE.