Low-Gain Stability of Projected Integral Control for Input-Constrained

Discrete-Time Nonlinear Systems (Extended Version)

John W. Simpson-Porco

Abstract— We consider the problem of zeroing an error loop, and that the closed-loop stability will be guaranteed if
output of a nonlinear discrete-time system in the presence the integral gain is sufficiently low; a famous and widely-
of constant exogenous disturbances, subject to hard convex deployed example of this design philosophy is the tuning of
arXiv:2103.06836v1 [math.OC] 11 Mar 2021

constraints on the input signal. The design specification is

formulated as a variational inequality, and we adapt a forward- automatic generation control in power systems [11].
backward splitting algorithm to act as an integral controller For finite-dimensional multi-input multi-output (MIMO)
which ensures that the input constraints are met at each time linear time-invariant (LTI) systems, the fundamental stability
step. We establish a low-gain stability result for the closed- result for this approach is due to Davison [12, Lemma 3]; see
loop system when the plant is exponentially stable, generalizing also [13, Theorem 3] and [14, Lemma 1, A.2, A.3]. While
previously known results for integral control of discrete-time
systems. Specifically, it is shown that if the composition of the [12] is in continuous-time, the key result is identical in the
plant equilibrium input-output map and the integral feedback discrete-time case [15], [16]. Consider the plant model
gain is strongly monotone, then the closed-loop system is
exponentially stable for all sufficiently small integral gains. The xk+1 = Axk + Buk + Bw w
(1)
method is illustrated via a simple numerical example. ek = Cxk + Duk + Dw w
I. INTRODUCTION with state x ∈ Rn , control input u ∈ Rm , constant distur-
It is a well-known principle of control engineering that bance/reference signal w ∈ Rnw , and error output e ∈ Rp ; we
regulation of an error signal to zero can be achieved ro- associate a sampling period Ts > 0 with (1). Assume that A
bustly in the presence of model uncertainty and constant is Schur stable, and let G(z) = C(zIn − A)−1 B + D denote
references/disturbances only through integral feedback con- the transfer matrix of (1) from u to e. One interconnects the
trol [1]. The presence of control input constraints however system (1) with the integral controller
presents challenges to traditional integral controller designs; Ts
ηk+1 = ηk − Ti ek , uk = Kηk , (2)
sufficient actuator authority may not be available to achieve
exact regulation for all references/disturbances, and dynamic where K ∈ Rm×p is a gain matrix and Ti > 0 is the integral
performance is sometimes degraded through the so-called time constant. Davison’s low-gain stability result states that
wind-up phenomenon [2]. if −G(1)K is Hurwitz stable, then there exists Ti? > 0 such
There are two broad approaches for accomodating lim- that the closed-loop system is exponentially stable for all
ited actuator authority. The explicit approach is to directly Ti ∈ (Ti? , ∞). A substantial literature exists on extensions
include input constraints into the design, as done in receding- of this core result to infinite-dimensional systems, including
horizon/model-predictive control [3], bounded integral control static nonlinearities, sampled-data implementations, and anti-
[4], and in other nonlinear/adaptive approaches [5]. A more windup compensation; see [17]–[19] and the references
traditional implicit approach is to proceed by first designing therein. Initial extensions to the continuous-time nonlinear
ignoring the actuator limits, and then to augment or retro-fit case were given in [20]. In [21] the author further generalized
the design in order to improve performance in the presence these conditions via contraction theory, and provided a LMI-
of saturation; this category would include both classic and based procedure to design low-gain integral controllers for
modern anti-windup design [6]–[8], and reference/command continuous-time nonlinear systems.
modification [9], [10].
Contributions: In this paper we further contribute to the
Returning now to the fundamentals of integral control, a
study of constrained and low-gain integral control. We begin
commonly encountered case in practice is that the system
by formulating the error regulation criteria in the presence
one wishes to control is complex, and limited dynamic model
of input constraints as a variational inequality [22], which
information is available, but it is however known that the
leads us to adopt a version of the projection or forward-
system is stable (possibly achieved via a stabilizing controller
backward algorithm [23] as a constrained integral controller.
design). A general and well-established design philosophy
The design is in discrete-time, and is therefore immediately
is that asymptotic tracking and disturbance rejection can
appropriate for digital control implementations. While this
be guaranteed by adding a supplementary integral control
design explicitly enforces input constraints at each time
instant, it has the following commonality with the more
J. W. Simpson-Porco is with the Department of Electrical and Computer implicit anti-windup approaches: if the input constraints
Engineering, University of Toronto, 10 King’s College Road, Toronto,
ON, M5S 3G4, Canada. Email: jwsimpson@ece.utoronto.ca. Work are not encountered during operation, the scheme reduces
supported by NSERC Discovery Grant RGPIN-2017-04008. to the classical integral controller (2). Our main stability
result (Theorem 3.1) establishes that the “low-gain integral B. Constrained Error-Zeroing Specification
control stability principle” described above also holds for this Let C ⊆ U be a closed non-empty convex set which
projected integral controller, which extends the main result describes actuator limits. Following (2), the control signal
of [21] to discrete-time nonlinear systems. uk from our new integral controller will be generated as
Notation: Given two vectors x and y, col(x, y) denotes
uk = Kηk (6)
their vertical concatenation. The matrix In is the n×n identity
matrix. If P is a symmetric matrix λmin (P ) and λmax (P ) where K ∈ Rm×p is a gain matrix to be designed and
denote its minimum and maximum eigenvalues. If P  0, the ηk ∈ Rp is the controller state. It follows that the preimage
inner product induced by P is denoted by hx, yiP = xT P y.
Γ := {η ∈ Rp | Kη ∈ C}
A function f : X → Rn is Lipschitz continuous on X ⊆ Rn
if there exists L > 0 such that kf (x) − f (y)k ≤ Lkx − yk is also a closed and non-empty convex set. For example, in
for all x, y ∈ X. applications C is often polyhedral, in which case so is Γ.
Our ideal design objective would be to ensure that for
I I . P R O B L E M F O R M U L AT I O N
any disturbance w ∈ W, the error signal ek is asymptotically
A. Plant Model and Assumptions driven to zero, and that the input constraint uk ∈ C is satisfied
We consider a plant described by a finite-dimensional at all times. As one might expect however, input constraints
nonlinear time-invariant state-space model may prevent us from exactly zeroing the steady-state error
xk+1 = f (xk , uk , w) (3a) ē = π(ū, w) for at least some disturbances w ∈ W. We
therefore relax the design objective, and instead seek an
ek = h(xk , uk , w) (3b) equilibrium value η̄ ∈ Γ for the controller state such that
n m
where xk ∈ R is the state, uk ∈ R is the control input, and
hē, η − η̄iP = hπ(K η̄, w), η − η̄iP ≥ 0, ∀η ∈ Γ, (7)
w ∈ Rnw is a vector of exogenous signals (reference signals
and/or disturbances). The signal ek ∈ Rp with p ≤ m is an where hx, yiP = x P y is the inner product on Rp induced
T

error output to be driven to zero. The model (3) would most by some positive definite matrix P  0. The inequality (7) is
commonly arise by discretizing a continuous-time model; we called a variational inequality [22], and we notate a solution
therefore associate a sampling period Ts > 0 to (3). of the inequality as η̄ ∈ VIP (Γ, π ◦ K). Note that if η̄ lies
For any fixed w, the possible equilibrium state-input-error in the interior of the set Γ, then there exists τ > 0 such
triplets (x̄, ū, ē) are determined by the algebraic equations that η = η̄ − τ ē ∈ Γ. The inequality (7) then implies that
−ēT ē ≥ 0, implying that ē = 0. In other words, if the input
x̄ = f (x̄, ū, w), ē = h(x̄, ū, w).
constraints are strictly feasible, then (7) is an exact error-
To capture the steady-state and dynamic behaviour of (3), we zeroing design specification. A geometric interpretation of
assume that there exist convex sets X , U, W such that (7) uses the normal cone of the set Γ at η̄ ∈ Γ, defined as
(A1) f and h are continuous in all arguments on X × U × W,
NΓP (η̄) := {d ∈ Rp | hd, η − η̄iP ≤ 0 for all η ∈ Γ}.
f is continuously differentiable with respect to x and u,
and f , h, ∂f ∂f
∂x , and ∂u are all Lipschitz continuous on
With this, (7) can be equivalently expressed as
X × U uniformly in w ∈ W; −ē = −π(K η̄, w) ∈ NΓP (η̄), as illustrated in Figure
(A2) there is a class C 1 map πx : U × W → X which is 1. The interpretation of Figure 1 is that from the point η̄, any
Lipschitz continuous on U × W such that further attempt to adjust in the direction −ē = −π(K η̄, w)
will result in constraint violation.
πx (u, w) = f (πx (u, w), u, w), (u, w) ∈ U × W.
(A3) the equilibrium x̄ = πx (u, w) is exponentially stable,
uniformly in (u, w) ∈ U × W.
Assumptions (A1)–(A3) capture the idea that the plant
model is sufficiently smooth, and converges exponentially
to a locally unique equilibrium when subject to reasonable
constant inputs u and w. We call the map
π : U × W → Rm , π(ū, w) := h(πx (ū, w), ū, w) (4) Fig. 1: Illustration of constrained error-zeroing specification.

the equilibrium input-to-error map, which produces the equi-

librium error ē = π(ū, w) associated with the constant control Remark 2.1 (Minimization Interpretation): To see how
input ū and disturbance w. When applied to the LTI system else the design specification (7) could arise, suppose that
(1), (A1)–(A3) simply reduce to A being Schur stable, and k = h (xk , uk , w) is a measured tracking error of interest
the mapping π becomes for the system (3), with associated equilibrium mapping
¯ = π (ū, w) defined similar to (4). Consider the steady-state
π(ū, w) = G(1)ū + Gw (1)w, (5)
minimization problem
−1
where Gw (z) = C(zIn −A) Bw +Dw is the transfer matrix
from w to e. minimize J(¯
) ⇔ minimize J(π (K η̄, w)) (8)
ū∈C η̄∈Γ
where J : Rp → R is a class C 1 convex and positive is a contraction mapping on Γ, satisfying
definite function. Critical points of this (generally, non-
convex) problem are determined by the inclusion kΦd (η) − Φd (η 0 )kP ≤ cdfb kη − η 0 kP .
∂π
− KT (K η̄, w)T ∇J(π (K η̄, w)) ∈ NΓ (η̄). (9) for all η, η 0 ∈ Γ, where cdfb = 1 − λ(1 − cfb ) ∈ (0, 1).
∂ ū
Proof: The existence/uniqueness statement is [22,
If we define the error signal ek := K T ∂π T
∂ ū (uk , w) ∇J(k ),

Theorem 2.3.3], and the proof of (i) requires only minor
then the inclusion (9) is precisely the variational inequality modifications of the proof of [23, Theorem 12.1.2]; (ii) then
(7). Thus, one could interpret the specification (7) as arising follows immediately from (i).
from a steady-state optimization problem where the goal
is to minimize a function of the tracking error. This per- I I I . DA M P E D P RO J E C T E D I N T E G R A L C O N T RO L
spective connects our approach directly with recent ideas in A N D L O W - G A I N S TA B I L I T Y R E S U LT
autonomous and feedback-based optimization; see [24]–[27]
for recent contributions.  A. Damped Projected Integral Control
C. The Projection (Forward-Backward) Algorithm We propose adapting the forward-backward splitting algo-
The error-zeroing specification (7) is equivalent to the so- rithm (12) as an integral feedback controller for enforcing
called natural equation [22, Chp. 1.5] the error-zeroing specification (7). Specifically, we propose
the damped projected integral (DP-I) controller
η̄ = ProjP
Γ (η̄ − απ(K η̄, w)) (10)
ηk+1 = (1 − λ)ηk + λ ProjP
Γ (ηk −
Ts
Ti ek ) (14a)
for any α > 0, where ProjPΓ : R
m
→ Γ is the projection
operator uk = Kηk (14b)
ProjP
Γ (η) = argmin kη − νkP (11)
ν∈Γ where Ti > 0 is the integral time constant. We make several
observations regarding (14):
√ the closest point to η in Γ measured in the norm
which yields
kxkP = xT P x induced by P  0. The equation (10) leads (i) Constrained Error-Zeroing: If (14) is in equilibrium
immediately to the classic projection or forward-backward with the plant (3), then it is immediate from (10) that
splitting algorithm [28, Section 25.3]. η̄ ∈ VIP (Γ, π ◦ K).
(ii) Input Constraint Satisfaction & Windup: If ηk ∈ Γ, then
ηk+1 = (1 − λ)ηk + λ ProjP
Γ (ηk − απ(Kηk , w)) (12) ηk+1 ∈ Γ, since by (14) ηk+1 is a convex combination
for solving the variational inequality VIP (Γ, π ◦ K), where of two points in Γ. Therefore, uk = Kηk ∈ C at all
λ ∈ (0, 1) is a damping parameter. We summarize some well- points in time. As a result, (14) will never suffer from
known conditions which ensure exponential stability of the traditional integrator windup.
iteration (12) to a unique equilibrium satisfying (10). (iii) Reduction to Classical Integral Control: If ηk ∈ Γ and
ηk − TTsi ek ∈ Γ, then the update (14) reduces to
Proposition 2.1: (Equilibrium and Contraction Proper-
ties of Forward-Backward Algorithm) Let w ∈ W. If Ts
ηk+1 = ηk − Ti0 ek , uk = Kηk (15)
η 7→ π(Kη, w) is µ-strongly monotone on Γ with respect to
h·, ·iP , i.e., if
where Ti0 = λ/Ti . Thus, when constraints are not
0 0
hπ(Kη, w) − π(Kη , w), η − η iP ≥ µkη − η 0 k2P encountered, (14) reduces to the integral controller (2).
(iv) Computation of Projection: The projection in (14) re-
for some µ > 0 and all η, η 0 ∈ Γ, then (12) possesses a unique
quires the solution of the convex optimization prob-
equilibrium point η̄ ∈ Γ satisfying (10). If η 7→ π(Kη, w)
lem (11), but need only be computed at step k if
is additionally L-Lipschitz continuous on Γ with respect to
ηk − TTsi ek ∈
/ Γ. Projections onto many simple types
the norm k · kP , and α is selected such that α ∈ (0, 2µ/L2 ),
of constraint sets are computable in closed-form; see,
then the following statements hold:
e.g., [29, App. B].
(i) the foward-backward operator (v) Alternative Controller: The controller (14) is based on
Φ : Γ → Γ, Φ(η) = ProjP the natural equation associated with the inequality (7).
Γ (η − απ(Kη, w)) (13)
If one instead uses a skewed natural equation (see [22,
is a contraction mapping on Γ, satisfying Chp. 1.5]), one can arrive at the alternative update law
kΦ(η) − Φ(η 0 )kP ≤ cfb kη − η 0 kP , η, η 0 ∈ Γ, ηk+1 = (1 − λ)ηk + λ ProjIΓr (ηk − Ts −1
p Ti P ek ),
where cfb = 1 − 2αµ + α2 L2 ∈ [0, 1).
(ii) the damped forward-backward operator where the projection is now with respect to the standard
Euclidean norm. In what follows though, we proceed
Φd : Γ → Γ, Φd (η) = (1 − λ)η + λΦ(η) with the formulation (14), mostly due to point (iii) above.
B. Low-Gain Stability with DP-I Control point (x̄, η̄) ∈ X × Γ with ē = h(x̄, K η̄, w) = π(K η̄, w) and
The closed-loop system consists of the interconnection of control ū = K η̄ ∈ C. Consider the change of state variable
the plant (3) and the controller (14); we can now state our ξk := xk − πx (Kηk , w).
main stability result.
With this, the dynamics (3),(14) become
Theorem 3.1 (Low-Gain Stability with DP-I Control):
Consider the plant (3) under Assumptions (A1)–(A3) with the ξk+1 = f (ξk + πx (Kηk , w), Kηk , w) − πx (Kηk+1 , w)
DP-I controller (14). Suppose that there exists a matrix P  0 ek = h(ξk + πx (Kηk , w), Kηk , w)
and constants µ, L > 0 such that η 7→ π(Kη, w) is µ-strongly
ηk+1 = (1 − λ)ηk + λ ProjP
Γ (ηk − αek ),
monotone and L-Lipschitz continuous on Γ with respect (18)
to h·, ·iP , uniformly in w ∈ W. Define Ti? := Ts L2 /2µ. and the equilibrium point of interest is (ξ, η) = (0, η̄).
Then for any Ti ∈ (Ti? , ∞), there exists λ? ∈ (0, 1) such
that for any λ ∈ (0, λ? ) and any w ∈ W, the closed-loop Step #2 — Analyzing the Slow Dynamics: Let
system possesses an exponentially stable equilibrium point Vs (η) = kη − η̄k2P . Using Φ and Φd from Proposition 2.1,
(x̄, η̄) ∈ X × Γ and the pair (ē, η̄) = (π(K η̄, w), η̄) satisfies we compute that
the error-zeroing specification (7). 1
Vs (ηk+1 ) 2 = k(1 − λ)ηk + λ ProjP
Γ (ηk − αek ) − η̄kP
To interpret the conditions in Theorem 3.1, consider again
the LTI case (5). The condition for strong monotonicity = k(1 − λ)ηk + λΦ(ηk ) − η̄
requires that there exist P  0 satisfying + λ(ProjP
Γ (ηk − αek ) − Φ(ηk ))kP
= kΦd (ηk ) − Φd (η̄)
(G(1)K)T P + P G(1)K  0
+ λ(ProjP
Γ (ηk − αek ) − Φ(ηk ))kP
which is equivalent to the matrix −G(1)K being Hurwitz ≤ cdfb kηk − η̄kP + λkδkP
stable; this is precisely Davison’s classical condition, as
described in Section I. The main condition required in where δ = ProjP
Γ (ηk − αek ) − Φ(ηk ). To bound kδkP we
Theorem 3.1 is that of strong monotonicity of the mapping compute that
η 7→ π(Kη, w); as shown in [21], the same condition is
kδk2P = k ProjP P 2
Γ (ηk − αek ) − ProjΓ (ηk − απ(Kηk , w))kP
sufficient for stability of low-gain integral control applied to
continuous-time nonlinear systems. In [21], it was further ≤ α2 kek − π(Kηk , w)k2P
shown that if π admits a linear fractional representation, = α2 kh(ξk + πx (Kηk , w), Kηk , w)
methods from robust control and semidefinite programming − h(πx (Kηk , w), Kηk , w)k2P
can be used to certify the monotonicity condition. The same 2
≤α λmax (P )L2h kξk k22
methods can in fact be used to synthesize controller gains K
which achieve robust performance for the reduced dynamics where Lh is the Lipschitz constant of h. Combining the
(12). While we omit the details here due to space limitations, above, with ∆Vs = Vs (ηk+1 ) − Vs (ηk ), one finds that along
the interested reader will have no trouble adapting the analysis trajectories of (18) it holds that
and synthesis results from [21] to the current context.
∆Vs ≤ (c2dfb − 1)kηk − η̄k2P + λmax (P )L2h α2 λ2 kξk k22
1
Proof of Theorem 3.1: The proof is based on a composite + 2λmax (P ) 2 αLh λcdfb kηk − η̄kP kξk k2
Lyapunov construction, and is divided into five steps.
Step #1 — Equilibrium and Error Equations: Let w ∈ W = ζkT Qs ζk
and set α := Ts /Ti . Equilibria (x̄, η̄) are characterized by where ζk = col(kξk k2 , kηk − η̄kP ) and
η̄ = ProjP
Γ (η̄ − αē) q1 = λmax (P )α2 L2h
 2
x̄ = f (x̄, ū, w),

q1 λ q2 λ
(16) Qs = 2 ,
ē = h(x̄, η̄, w), ū = K η̄. q2 λ cdfb − 1 q2 = αλmax (P )1/2 Lh cdfb .
If such an equilibrium exists, then necessarily η̄ ∈ Γ, and
Step #3 — Bounding kηk+1 − ηk kP : We compute using
hence ū = K η̄ ∈ C. Given any such ū, it follows from (A2)
the triangle inequality that
that the first equation in (16) can be solved for x̄ = πx (ū, w);
together, (A2)/(A3) imply that x̄ is isolated. Eliminating x̄ kηk+1 − ηk kP ≤ kηk+1 − Φd (ηk )kP
(19)
and ē, we obtain the reduced equilibrium equation + kΦd (ηk ) − ηk kP .
η̄ = ProjP
Γ (η̄ − απ(K η̄, w)) = Φ(η̄) = Φd (η̄) (17) Using our previous calculations, the first term in (19) can be
bounded as
which is equivalent to the error-zeroing specification (7).
kηk+1 − Φd (ηk )kP = λk ProjP
Γ (ηk − αek ) − Φ(ηk )kP
Since η 7→ π(Kη, w) is µ-strongly monotone on Γ uniformly
in w, and Γ is closed, convex, and non-empty, VIP (Γ, π ◦ K) = λkδkP
admits a unique solution [22, Theorem 2.3.3]. We conclude ≤ λαLh λmax (P )1/2 kξk k2 .
that the closed-loop system possess a unique equilibrium (20)
To bound the second term in (19), it follows from Proposition Since πx (Kη, w) = f (πx (Kη, w), Kη, w) and f is Lipschitz
2.1 and the triangle inequality that continuous in (x, u) uniformly in w, we have

kΦd (ηk ) − ηk kP = λkΦ(ηk ) − ηk kP kf (ξk + πx (Kηk , w), Kηk , w) − πx (Kηk+1 , w)k2

 
ξk + πx (Kηk , w) − πx (Kηk+1 , w)
= λk(ηk − η̄) − (Φ(ηk ) − η̄)kP (21) ≤ Lf
K(ηk − ηk+1 )
≤ λ(1 + cfb )kηk − η̄kP . 2
1
≤ Lf kξk k2 + Lf (1 + Lπx )kKk2 λmin (P )− 2 kηk+1 − ηk kP
Putting things together we obtain (28)
where we used that kzk2 ≤ kzk1 for z ∈ Rn . Similarly, we
kηk+1 − ηk kP ≤ λLh αλmax (P )1/2 kξk k2
(22) compute that
+ λ(1 + cfb )kηk − η̄kP .
kf (ξk + πx (Kηk , w), Kηk , w) − πx (Kηk , w)k2 ≤ Lf kξk k2 .
(29)
Step #4 — Analyzing the Fast Dynamics: Define the
Substituting (27)–(29) back into (26), we obtain
deviation vector field g : Rn × U × W → Rn by
|∆Vf1 | ≤ c3 λmin (P )−1 Lf Lπx (1 + Lπx )kKk22 kηk+1 − ηk k2P
g(ξ, u, w) = f (ξ + πx (u, w), u, w) − f (πx (u, w), u, w) 1
+ 2c3 Lπx Lf kKk2 λmin (P )− 2 kξk k2 kηk+1 − ηk kP .
= f (ξ + πx (u, w), u, w) − πx (u, w). (30)
Under Assumptions (A1)–(A3), the conditions of a converse Substituting (22) into (30) and collecting terms, we finally
Lyapunov theorem in the appendix are met: there exists a obtain the bound
k λ2 + k2 λ k3 λ2 + k4 λ
 
set Z containing the origin in its interior, positive constants
|∆Vf1 | ≤ ζkT Q1 ζk = ζkT 1 2 ζk
c1 , c2 , c3 , c4 > 0, ρf ∈ [0, 1), and a continuous function k3 λ + k4 λ k 5 λ2
(31)
Vf : Z × U × W → R≥0 , (ξ, u, w) 7→ Vf (ξ, u, w) where the explicit expressions for the coefficients are
1
satisfying the Lyapunov conditions k1 = c3 Lf Lπx (1 + Lπx )L2h kKk22 α2 λmin (P )− 2
1 1
c1 kξk22 ≤ Vf (ξ, u, w) ≤ c2 kξk22 (23a) k2 = 2c3 Lπx Lf kKk2 αλmax (P ) 2 λmin (P )− 2
Vf (g(ξ, u, w), u, w) − Vf (ξ, u, w) ≤ −ρf kξk22 (23b) k3 = c3 Lf Lπx (1 + Lπx )Lh kKk22 α(1 + cfb )
1
|Vf (ξ, u, w) − Vf (ξ 0 , u, w)| ≤ c3 (kξk2 + kξ 0 k2 )kξ − ξ 0 k2 · λmin (P )−1 λmax (P ) 2
(23c) 1
0 0 k4 = c3 Lπx Lf kKk2 (1 + cfb )λmin (P )− 2
|Vf (ξ, u, w) − Vf (ξ, u , w)| ≤ c4 kξk22 ku − u k2 (23d)
k5 = c3 Lf Lπx (1 + Lπx )kKk22 (1 + cfb )2 λmin (P )−1 .
0 0
for all ξ, ξ ∈ Z, all u, u ∈ U, and all w ∈ W. Let
To bound |∆Vf2 | we may apply (23d) to obtain
∆Vf = Vf (ξk+1 , ηk+1 , w) − Vf (ξk , ηk , w) |∆Vf2 | ≤ c4 kf (ξk + πx (Kηk , w), Kηk , w) − πx (Kηk , w)k22
· kK(ηk+1 − ηk )k2
denote the increment of Vf along trajectories of (18). Substi-
1
tuting in the dynamics and adding and subtracting terms, we ≤ c4 L2f kKk22 λmin (P )− 2 kξk k22 kηk+1 − ηk kP ,
can express the overall change as
where in the second line we used (29). Since 0 is
∆Vf = ∆Vf1 + ∆Vf2 + ∆Vf3 , (24) an interior point of Z, there exists r > 0 such that
Br (0) := {ξ ∈ Rn | kξk2 < r} ⊂ Z. For ξk ∈ Br (0)
where the summands are defined in (25). we therefore have the further bound
We now bound each term in (24) individually. Applying 1
|∆Vf2 | ≤ c4 L2f kKk22 λmin (P )− 2 rkξk k2 kηk+1 − ηk kP .
(23c) to |∆Vf1 | we obtain
Substituting (22) into this, we finally obtain
|∆Vf1 | ≤ c3 kπx (Kηk+1 , w) − πx (Kηk , w)k2  
k λ k7 λ
|∆Vf2 | ≤ ζkT Q2 ζk = ζkT 6

· kf (ξk + πx (Kηk , w), Kηk , w) − πx (Kηk+1 , w)k2 ζk (32)
k7 λ 0

+ kf (ξk + πx (Kηk , w), Kηk , w) − πx (Kηk , w)k2 where
(26) 1 1
k6 = c4 L2f kKk22 λmin (P )− 2 rLh αλmax (P ) 2
Since πx is Lipschitz continuous we have 1
k7 = 12 c4 L2f kKk22 λmin (P )− 2 r(1 + cfb ).
kπx (Kηk+1 , w) − πx (Kηk , w)k2 ≤ Lπx kKk2 kηk+1 − ηk k2
1 To bound ∆Vf3 we apply (23b) to obtain
≤ λmin (P )− 2 Lπx kKk2 kηk+1 − ηk kP .
(27) ∆Vf3 ≤ −ρf kξk k22 . (33)
 ∆Vf1 := Vf (f (ξk + πx (Kηk , w), Kηk , w) − πx (Kηk+1 , w), Kηk+1 , w)
− Vf (f (ξk + πx (Kηk , w), Kηk , w) − πx (Kηk , w), Kηk+1 , w)
∆Vf2 := Vf (f (ξk + πx (Kηk , w), Kηk , w) − πx (Kηk , w), Kηk+1 , w) (25)
− Vf (f (ξk + πx (Kηk , w), Kηk , w) − πx (Kηk , w), Kηk , w)
3
∆Vf := Vf (f (ξk + πx (Kηk , w), Kηk , w) − πx (Kηk , w), Kηk , w) − Vf (ξk , Kηk , w).

Substituting the individual bounds for |∆Vf1 |, |∆Vf2 |, and parameters are selected as Ti = 30, and λ = 0.9. The inputs
∆Vf3 from (31), (32), and (33) back into (24), we obtain the uk ∈ R3 are constrained to the convex set
overall bound
C = {u ∈ R3 | |u1 | + |u2 | ≤ 1, |u3 | ≤ 2}.
∆Vf ≤ ζk Qf ζk
For comparison purposes, we simulate (i) the response of
where Qf := Q1 + Q2 + Q3 evaluates to
the closed-loop system without constraints with the basic
−ρf + k1 λ2 + (k2 + k6 )λ k3 λ2 + (k4 + k7 )λ
 
integral controller (2), and (ii) the response of the closed-
Qf = .
k3 λ2 + (k4 + k7 )λ k5 λ2 loop system with constraints with the DP-I controller (14).
Step #5 – Putting the Pieces Together: Define the com- Figure 2 shows the response under several step changes to the
posite Lyapunov candidate V (ξ, η, w) = Vs (η) + Vf (ξ, η, w). two reference signals. In the unconstrained case, as expected,
Along trajectories of (18), we combine the previous inequal- the reference signals are tracked after the initial transient
ities to compute that response. In the constrained case, sufficient actuator capacity
is not available, and exact tracking is not achieved. Figure 3
∆V = V (ξk+1 , ηk+1 , w) − V (ξk , ηk , w) ≤ ζkT Qζk shows the control signal, projected onto the u1 –u2 plane; the
two responses are identical when the projection operation is
holds for all (ξk , ηk , w) ∈ Br (0) × Γ × W, where
inactive (see Section III-A), but as expected uk ∈ C at all
−ρf + (k1 + q1 )λ2 + k̃2 λ k3 λ2 + k̃4 λ
 
times in the constrained case.
Q=
k3 λ2 + k̃4 λ −(1 − c2dfb ) + k5 λ2
and where for compactness we set k̃2 = k2 + k6 and
k̃4 = k4 + k7 + q2 . Note that the (1, 1) element of Q is
negative and O(1) as λ → 01 . From Proposition 2.1, we
have that
1 − c2dfb = 2λ(1 − cfb ) − λ2 (1 − cfb )2 ,
with cfb ∈ (0, 1), and therefore the (2, 2) element of Q
is negative and O(λ) as λ → 0. Since the off diagonal
elements are O(λ) as λ → 0, it is straightforward to
argue that there exists some λ? > 0 such that Q ≺ 0
for all λ ∈ (0, λ? ). Using (23a), there therefore exists
ε > 0 such that ∆V (ξk , ηk , w) ≤ −εV (ξk , ηk , w) for all
(ξk , ηk , w) ∈ Br (0) × Γ × W. Standard arguments (e.g., [30,
Thm. 13.2] now complete the proof.  Fig. 2: Tracking response of closed-loop system; dashed black lines
denote reference set-points.
I V. S I M U L AT I O N E X A M P L E
We now illustrate the action of the controller via an V. C O N C L U S I O N S
academic example. For simplicity, we generate a random
We have formulated an approximate tracking specification
internally stable discrete-time LTI system with sampling
as a variational inequality, and designed projected integral
period Ts = 1, n = 5 states, r = 2 measurements, m = 3
controller to meet this specification while maintaining arbi-
control inputs, and nw = 2 exogenous reference signals, and
trary convex constraints on the input signal at all times. In the
consider a standard reference-tracking control arrangement
absence of constraints, the approximate tracking specification
where ek = yk − wk . For this system, the DC gain matrix
reduces to an exact tracking specification, and the projected
G(1) ∈ Rr×m has full row rank. For the DP-I controller
integral controller reduces to a classical integral controller.
(14), the gain K is selected as K = G(1)† , and hence the
The controller inherits what is perhaps the most important
monotonicity and Lipschitz conditions in Theorem 3.1 are
stability property of traditional integral control; under mild
satisfied with P = I2 and µ = L = 1; the remaining
monotonicity conditions on the plant equilibrium mapping,
1 For a function g : R → R which is positive definite with respect to 0, a closed-loop stability can be guaranteed when the plant is
function f : R → R is O(g(λ)) as λ → 0 if limλ→0 |f (λ)|/g(λ) < ∞. exponentially stable and the integral gain is sufficiently low.
 Proof of Theorem A1: Let φk (z; u) denote the solution of
(35) from initial condition z0 = z and with input u ∈ U. By
uniform exponential stability, there exist constants M ≥ 1,
ρ ∈ (0, 1) and δ > 0 such that
kzk2 < δ =⇒ kφk (z; u)k2 ≤ M ρk kzk2 (37)
for all k ≥ 0 and for any u ∈ U. In particular
then, kφk (z; u)k2 ≤ M δ :=  for all k ≥ 0. Let
Bδ (0) = {ξ ∈ Rn | kξk2 < δ} and define
Z = {ζ ∈ Rn | ∃k ≥ 0, z ∈ Bδ (0), u ∈ U s.t. φk (z; u) = ζ}.
In other words, Z is the image of all solutions with initial
conditions in Bδ (0) and inputs in U. Due to uniform
exponential stability, we have that Z ⊆ B (0), so Z is a
bounded set. Moreover, by definition, Z is positively invariant.
Finally, note that Bδ (0) ⊆ Z, so 0 is in the interior of Z.
With this refined notation, we have that if z ∈ Bδ (0), then
Fig. 3: Control inputs; dashed black lines denote constraints. φk (z; u) ∈ Z for any k ≥ 0 and any u ∈ U.
Next, define the function V : Z × U → R by
XN −1
Future work will consider applications of this control scheme, V (z, u) = φk (z; u)T φk (z; u)
k=0
the extension of this scheme to a projected PID controller, and
the extension to more general discrete-time output-regulating for some N ≥ 1 to be determined. Since g is Lipschitz
controllers which admit a representation in incremental form. continuous in both arguments, so is φk , so V is continuous
in both arguments. We first observe that
APPENDIX XN −1
V (z, u) = φ0 (z; u)T φ0 (z; u) + φk (z; u)T φk (z; u)
Consider the discrete-time nonlinear system k=1
≥ φ0 (z; u)T φ0 (z; u)
xk+1 = f (xk , u) (34)
= kzk22 .
where f : X × U → Rn , with X ⊆ Rn and U ⊆ Rm being By uniform exponential stability, it also holds that
convex sets. Assume that XN −1
(i) f is continuously differentiable on X × U, V (z, u) ≤ M 2 ρ2k kzk22 ,
k=0
(ii) f, ∂f ∂f
∂x , ∂u are Lipschitz continuous in x on the set X , and henceP(36a) holds for all (z, u) ∈ Z × U with c1 = 1
uniformly in the inputs u ∈ U, N −1
and c2 = k=0 M 2 ρ2k . Next, note that since Z is forward
(iii) there exists a continuously differentiable map
invariant, we have that g(z, u) ∈ Z for any z ∈ Z and any
πx : U → X which is Lipschitz continuous on U and
u ∈ U. We may therefore compute for z ∈ Z that
satisfies πx (u) = f (πx (u), u) for all u ∈ U.
V (g(z, u), u) − V (z, u)
Defining the new state variable z = x − πx (u), we obtain
N −1
the deviation model X
= kφk (g(z, u); u)k22 − kφk (z; u)k22
zk+1 = f (zk + πx (u), u) − πx (u) := g(zk , u), (35) k=0

By the semi-group property of solutions, we have

where now z = 0 is an equilibrium point of (35) for all
u ∈ U. φk (g(z, u); u) = φk+1 (z; u).
Theorem A1: Consider the dynamics (35) under the previ- The sum therefore telescopes, and we find that
ous assumptions. If the origin z = 0 is locally exponentially
V (g(z, u)) − V (z, u) = kφN (z; u)k22 − kzk22
stable uniformly in u ∈ U, then there exists a set Z containing
≤ − 1 − M 2 ρ2N kzk22 ,


0 in its interior, a continuous function V : Z × U → R, and
constants c1 , c2 , c3 , c4 > 0 and ρf ∈ (0, 1) such that where we have used (37). Selecting N large enough such
that M 2 ρ2N < 1, we define ρf = 1 − M 2 ρ2N ∈ (0, 1) and
c1 kzk22 ≤ V (z, u) ≤ c2 kzk22 (36a)
we obtain (36b) for all (z, u) ∈ Z × U. To show (36c) we
V (g(z, u), u) − V (z, u) ≤ −ρf kzk22 (36b) first compute for z, z 0 ∈ Z that
0 0 0
|V (z, u) − V (z , u)| ≤ c3 (kzk2 + kz k2 )kz − z k2 (36c)
kg(z, u) − g(z 0 , u)k2
|V (z, u) − V (z, u0 )| ≤ c4 kzk22 ku − u0 k2 (36d)
≤ kf (z + πx (u), u) − f (z 0 + πx (u), u)k
0
for all z, z ∈ Z and all u, u ∈ U.0
≤ L1 kz − z 0 k2
where L1 > 0 is the Lipschitz constant of f . It quickly We are now ready to compute that
follows then that kφk (z; u) − φk (z 0 ; u)k2 ≤ Lk1 kz − z 0 k2 .  XN −1
|V (z, u) − V (z, u0 )| =  φk (z; u)T φk (z; u)

We may now compute that k=0
 XN −1 XN −1 
|V (z, u) − V (z 0 , u)| =  φk (z; u)T φk (z; u) − φk (z; u0 )T φk (z; u0 )
 
k=0 k=0
XN −1 XN −1
φk (z; u)T [φk (z; u) − φk (z; u0 )]

− φk (z 0 ; u)T φk (z 0 ; u)
 =
k=0 k=0
 XN −1 XN −1 
φk (z; u)T (φk (z; u) − φk (z 0 ; u)) − φk (z; u0 )T [φk (z; u0 ) − φk (z; u)]

=

k=0 k=0
XN −1 XN −1
(kφk (z; u)k2 + kφk (z; u0 )k2 )k∆k k2

− φk (z 0 ; u)T (φk (z 0 ; u) − φk (z; u))
 ≤
k=0 k=0
XN −1
N
X −1 ≤ κkzk22 ku − u0 k2 2M ρk
≤ Lk1 (kφk (z; u)k2 + kφk (z 0 ; u)k2 )kz − z 0 k2 k=0
PN −1
k=0 which completes the proof with c5 = k=0 2κM ρk . 
N
X −1
≤ Lk1 M ρk (kzk2 + kz 0 k2 )kz − z 0 k2 REFERENCES
k=0
[1] B. A. Francis and W. Wonham, “The internal model principle of control
PN −1 k k theory,” Automatica, vol. 12, no. 5, pp. 457–465, 1976.
so (36c) holds with c4 = k=0 L1 M ρ . To show (36d), [2] K. J. Astrom and L. Rundqwist, “Integrator windup and how to avoid
begin by computing (35) that it,” in Proc. ACC, Pittsburgh, PA, USA, 1989, pp. 1693–1698.
[3] J. B. Rawlings and D. Q. Mayne, Model Predictive Control: Theory
∂g ∂f ∂πx
∂u (z, u) = ∂x (z + πx (u), u) ∂u (u) and design. Nob Hill Publishing, 2009.
∂f
(38) [4] G. C. Konstantopoulos, Q.-C. Zhong, B. Ren, and M. Krstic, “Bounded
+ ∂u (z + πx (u), u) − ∂π
∂u (u).
x
integral control of input-to-state practically stable nonlinear systems to
guarantee closed-loop stability,” IEEE Trans. Autom. Control, vol. 61,
Next, since πx (u) = f (πx (u), u) for any u ∈ U, we have by no. 12, pp. 4196–4202, 2016.
continuous differentiability that [5] D. E. Miller and E. J. Davison, “An adaptive tracking problem with
a control input constraint,” Automatica, vol. 29, no. 4, pp. 877–887,
∂πx ∂f ∂πx ∂f 1993.
∂u (u) = ∂x (πx (u), u) ∂u (u) + ∂u (πx (u), u). (39) [6] G. Grimm, J. Hatfield, I. Postlethwaite, A. R. Teel, M. C. Turner,
and L. Zaccarian, “Antiwindup for stable linear systems with input
Inserting (39) into (38) we find that saturation: an LMI-based synthesis,” IEEE Trans. Autom. Control,
h i vol. 48, no. 9, pp. 1509–1525, 2003.
∂g ∂f ∂f ∂πx
∂u (z, u) = ∂x (z + π x (u), u) − ∂x (π x (u), u) ∂u (u)
[7] M. C. Turner, G. Herrmann, and I. Postlethwaite, “Discrete-time anti-
h i windup: Part 1 – stability and performance,” in Proc. ECC, Cambridge,
∂f ∂f UK, 2003, pp. 473–478.
+ ∂u (z + πx (u), u) − ∂u (πx (u), u) . [8] S. Tarbouriech and M. Turner, “Anti-windup design: an overview of
some recent advances and open problems,” IET Control Theory &
∂f ∂f Applications, vol. 3, pp. 1–19(18), 2009.
Using Lipschitz continuity of ∂x , ∂u , and πx , we therefore
[9] I. Kolmanovsky, E. Garone, and S. Di Cairano, “Reference and com-
obtain the bound mand governors: A tutorial on their theory and automotive applications,”
∂g in Proc. ACC, Portland, OR, USA, 2014, pp. 226–241.
k ∂u (z, u)k2 ≤ `kzk2 . [10] P. Schwerdtner, S. A. Bortoff, C. Danielson, and S. Di Cairano,
“Projection-based anti-windup for multivariable control with heat pump
for some ` > 0. By convexity of U and the multivariable application,” in Proc. ECC, Naples, Italy, 2019, pp. 1281–1287.
mean-value theorem, this immediately implies that [11] J. W. Simpson-Porco and N. Monshizadeh, “Diagonal stability of
systems with rank-1 interconnections and application to automatic
generation control in power systems,” IEEE Trans. Control Net. Syst.,
kg(z, u) − g(z, u0 )k2 ≤ `kzk2 ku − u0 k2 . (40) Sep. 2020, submitted.
[12] E. J. Davison, “Multivariable tuning regulators: The feedforward and
Let ∆k = φk (z; u) − φk (z; u0 ) denote the difference between robust control of a general servomechanism problem,” IEEE Trans.
the flows given different inputs. We bound this as Autom. Control, vol. 21, no. 1, pp. 35–47, 1976.
[13] M. Morari, “Robust stability of systems with integral control,” IEEE
k∆k k2 = kg(φk−1 (z; u), u) − g(φk−1 (z; u0 ), u0 )k2 Trans. Autom. Control, vol. 30, no. 6, pp. 574–577, 1985.
[14] D. E. Miller and E. J. Davison, “The self-tuning robust servomechanism
= kg(φk−1 (z; u), u) − g(φk−1 (z; u0 ), u) problem,” IEEE Trans. Autom. Control, vol. 34, no. 5, pp. 511–523,
1989.
− [g(φk−1 (z; u0 ), u0 ) − g(φk−1 (z; u0 ), u)]k2 [15] A. Peltomaa and H. N. Koivo, “Tuning of a multivariable discrete time
≤ L1 k∆k−1 k2 + L1 `kφk−1 (z, u0 )k2 ku − u0 k2 pi controller for unknown systems,” Int. J. Control, vol. 38, no. 4, pp.
735–745, 1983.
≤ L1 k∆k−1 k2 + L1 `M ρk−1 kzk2 ku − u0 k2 [16] H. Logemann and S. Townley, “Discrete-time low-gain control of
uncertain infinite-dimensional systems,” IEEE Trans. Autom. Control,
where we added and subtracted terms and then bounded using vol. 42, no. 1, pp. 22–37, 1997.
(40) and (37). Using this bound recursively and noting that [17] C. Guiver, H. Logemann, and S. Townley, “Low-gain integral control
for multi-input multioutput linear systems with input nonlinearities,”
∆0 = 0, we obtain IEEE Trans. Autom. Control, vol. 62, no. 9, pp. 4776–4783, Sep. 2017.
Xk−1 [18] A. Terrand-Jeanne, V. Andrieu, V. Dos Santos Martins, and C. Xu,
k∆k k2 ≤ kzk2 ku − u0 k2 Lk−m `M ρm . “Adding integral action for open-loop exponentially stable semigroups
m=0 1 and application to boundary control of pde systems,” IEEE Trans.
| {z }
:=κ Autom. Control, vol. 65, no. 11, pp. 4481–4492, 2019.
[19] M. E. Gilmore, C. Guiver, and H. Logemann, “Sampled-data integral
control of multivariable linear infinite-dimensional systems with input
nonlinearities,” Mathematical Control & Related Fields, vol. 0, 2021,
to appear.
[20] C. Desoer and C.-A. Lin, “Tracking and disturbance rejection of MIMO
nonlinear systems with PI controller,” IEEE Trans. Autom. Control,
vol. 30, no. 9, pp. 861–867, 1985.
[21] J. W. Simpson-Porco, “Analysis and synthesis of low-gain integral
controllers for nonlinear systems,” IEEE Trans. Autom. Control, 2020,
to appear.
[22] F. Facchinei and J.-S. Pang, Finite-Dimensional Variational Inequalities
and Complementarity Problems, Vol. 1. Springer Series in Operations
Research and Financial Engineering, 2003.
[23] ——, Finite-Dimensional Variational Inequalities and Complementar-
ity Problems, Vol. 2. Springer Series in Operations Research and
Financial Engineering, 2003.
[24] L. S. P. Lawrence, J. W. Simpson-Porco, and E. Mallada, “Linear-
convex optimal steady-state control,” IEEE Trans. Autom. Control,
2020, to appear.
[25] M. Colombino, J. W. Simpson-Porco, and A. Bernstein, “Towards
robustness guarantees for feedback-based optimization,” in Proc. IEEE
CDC, Nice, France, Dec. 2019, pp. 6207–6214.
[26] A. Hauswirth, S. Bolognani, G. Hug, and F. Dörfler, “Timescale
separation in autonomous optimization,” IEEE Trans. Autom. Control,
vol. 66, no. 2, pp. 611–624, 2021.
[27] G. Bianchin, J. Cortés, J. I. Poveda, and E. Dall?Anese, “Time-varying
optimization of lti systems via projected primal-dual gradient flows,”
2021, https://arxiv.org/pdf/2101.01799.pdf.
[28] H. H. Bauschke and P. L. Combettes, Convex Analysis and Monotone
Operator Theory in Hilbert Spaces, 1st ed. Springer Verlag, 2011.
[29] A. Beck, First-Order Methods in Optimization, 1st ed. SIAM, 2017.
[30] W. M. Haddad and V. Chellaboina, Nonlinear Dynamical Systems and
Control: A Lyapunov-Based Approach. Princeton Univ Press, 2008.