Introduction
Model predictive control (MPC) is powerful because it combines performance optimization with explicit constraint handling. Its success in process control, robotics, reinforcement learning, and autonomous motion planning has also motivated efforts to improve its computational efficiency [1 ] . Its effectiveness, however, relies on the availability of a predictive model that is accurate over the operating region of interest [2 ] . This modeling requirement becomes a bottleneck for systems whose dynamics are high-dimensional, strongly coupled, or repeatedly reconfigured. In such settings, deriving and maintaining a first-principles model can dominate the control-design effort, whereas input-output data are often readily available.
This work was supported by the Ministry of Education, Singapore, under AcRF Tier 1 Grant RG64/23, and the NTU-VinUni Joint Lab on Robotics and Embodied AI, and in part by the National Natural Science Foundation of China under Grants U23B2059 and 62688021. (Corresponding author: Lihua Xie)
Wenjie Liu is with the School of Electrical and Electronic Engineering, Nanyang Technological University, 639798, Singapore (e-mail: wenjie.liu@ntu.edu.sg).
Gang Wang is with the State Key Lab of Autonomous Intelligent Unmanned Systems, School of Automation, Beijing Institute of Technology, Beijing 100081, China (e-mail: gangwang@bit.edu.cn).
Lihua Xie is with the NTU–VinUni Joint Research Laboratory for Robotics and Embodied AI, School of Electrical and Electronic Engineering, Nanyang Technological University, Singapore 639798 and VinUniversity, Hanoi 100000, Vietnam (e-mail: elhxie@ntu.edu.sg).
Data-driven predictive control (DDPC) has emerged as a principled way to reduce this modeling burden [3 , 4 ] . A key enabler is Willems et al.’s fundamental lemma [5 ] , which states that, under persistence of excitation, finite trajectories of an unknown linear system can be represented by linear combinations of previously measured trajectories. This observation has led to a set of MPC-like controllers with stability and robustness guarantees directly from data [6 , 7 , 8 , 9 , 10 , 11 , 12 , 13 ] . Most existing DDPC results, nevertheless, are formulated around stabilization to a fixed equilibrium or tracking of equilibrium-like setpoints. Output regulation is different: the target behavior is an exosystem-driven trajectory on a regulator manifold, and the corresponding steady-state input is characterized by regulator equations that depend on the unknown plant.
Output regulation is the natural language for many constrained control tasks: grid converters must reject periodic load variations [14 ] , vehicles must follow curvature-varying roads, and networked or robotic systems often have to synchronize with time-varying leaders. Classical solutions rely on the internal model principle and on solvability of the output regulator equations [15 , 16 , 17 , 18 ] . These results are foundational, but their implementation typically requires plant matrices and feedforward regulator gains. When the plant is unknown or only measured through noisy input-output data, identifying a model and then solving the output regulator equations introduces an additional approximation layer precisely in the part of the design that determines the steady-state behavior.
Recent data-driven approaches to output regulation are compared in Table I. Most of them are feedback-synthesis results and do not handle input constraints [19 , 20 , 21 , 22 , 23 ] , whereas DDPC formulations for repetitive tasks [24 ] lack closed-loop guarantees. Developing a constrained, data-driven predictive regulator with guarantees requires overcoming three obstacles. First, the Hankel prediction model must encode both the unknown plant response and the temporal structure induced by the known exosystem eigenstructure. Second, even with such a predictor, the usual equilibrium-centered DDPC stage cost is not aligned with the regulator manifold, since maintaining this manifold generally requires a nonzero feedforward input determined by the unknown regulator equations. Third, with noisy data, the coefficient regularization must be centered at the nonzero steady-state coefficient rather than at zero. These issues are addressed in the present paper.
We propose a predictive-control framework for data-driven output regulation of unknown linear systems driven by exogenous signals. The controller uses measured input-output trajectories, the exosystem dynamics, and online measurements; it does not identify the plant matrices and does not explicitly solve the classical output regulator equations. The analysis relies on the solvability of the output regulator equations, controllability/observability of the plant, a nonresonance condition, and incremental input-output-to-state stability of an auxiliary input-output realization. Under these standard structural conditions, the proposed receding-horizon laws steer the closed loop to the regulator manifold; under bounded output noise, they steer it to a neighborhood whose size is quantified by the noise level.
Table 1. Comparison with related output regulation approaches.
The main contributions are summarized as follows.
c1) A data-based parametrization of trajectories driven by exosignals with known dynamics and known phase, built on the decomposition in [23 ] ; it recovers the affine fundamental lemma [9 ] for S = I S=I S = I and comes with an excitation condition verifiable from the input and the exosystem basis.
c2) A constrained DDPC regulator that requires neither a plant model nor the regulator equations; for periodic exosignals, the steady-state feedforward input is removed by an incremental-input penalty.
c3) Recursive feasibility and exponential stability of the regulator manifold with noise-free data, and practical stability under noisy data and measurements, enabled by a regularization center computed offline from data with a guaranteed mismatch bound.
Notation. We denote the sets of real and complex numbers by R \mathbb{R} R and C \mathbb{C} C , respectively, and the sets of non-negative and positive integers by N \mathbb{N} N and N + \mathbb{N}_{+} N + . For any matrix M M M , M † M^\dagger M † denotes its Moore–Penrose pseudoinverse. The notation M ≻ ( ⪰ ) 0 M\succ(\succeq)0 M ≻ ( ⪰ ) 0 indicates that M M M is positive (semi-) definite, and similarly, M ≺ ( ⪯ ) 0 M\prec(\preceq)0 M ≺ ( ⪯ ) 0 denotes that M M M is negative (semi-) definite. For vector x ∈ R n x x\in\mathbb{R}^{n_x} x ∈ R n x , ∥ x ∥ \lVert x\rVert ∥ x ∥ denotes its Euclidean norm, and for a matrix M M M , ∥ M ∥ \lVert M\rVert ∥ M ∥ denotes the spectral norm. The minimum and maximum eigenvalues of matrix P P P are denoted by λ ‾ P \underline{\lambda}_P λ P and λ ‾ P \overline{\lambda}_P λ P , respectively. For a signal x : N → R n x x:\mathbb{N}\to\mathbb{R}^{n_x} x : N → R n x and T ∈ N + T\in\mathbb{N}_{+} T ∈ N + , the vectorized form of the signal x x x over the interval [ 0 , T − 1 ] [0,T-1] [ 0 , T − 1 ] is denoted by x [ 0 , T − 1 ] : = [ x ( 0 ) ′ x ( 1 ) ′ ⋯ x ( T − 1 ) ′ ] ′ x_{[0,T-1]}:=[x(0)^{\prime}\ x(1)^{\prime}\cdots x(T-1)^{\prime}]^{\prime} x [ 0 , T − 1 ] := [ x ( 0 ) ′ x ( 1 ) ′ ⋯ x ( T − 1 ) ′ ] ′ . The block-diagonal matrix with blocks A 1 , … , A N A_1,\ldots,A_N A 1 , … , A N on its main diagonal is denoted block.diag ( A 1 , … , A N ) \operatorname{block.diag}(A_1,\ldots,A_N) block.diag ( A 1 , … , A N ) . Given N , L ∈ N + N,L\in\mathbb{N}_{+} N , L ∈ N + , the Hankel matrix of order L L L of a length-N N N trajectory x [ 0 , N − 1 ] x_{[0,N-1]} x [ 0 , N − 1 ] is denoted by H L ( x [ 0 , N − 1 ] ) : = [ x 0 x 1 ⋯ x N − L x 1 x 2 ⋯ x N − L + 1 ⋮ ⋮ ⋮ x L − 1 x L ⋯ x N − 1 ] H_L(x_{[0,N-1]}):=\begin{bmatrix}x_0&x_1&\cdots&x_{N-L}\\x_1&x_2&\cdots&x_{N-L+1}\\\vdots&\vdots&&\vdots\\x_{L-1}&x_L&\cdots&x_{N-1}\end{bmatrix} H L ( x [ 0 , N − 1 ] ) := x 0 x 1 ⋮ x L − 1 x 1 x 2 ⋮ x L ⋯ ⋯ ⋯ x N − L x N − L + 1 ⋮ x N − 1 .
Fundamental Lemma for LTI Systems with Exosignals
This section establishes the data-based trajectory representation used throughout the paper. Consider the unknown linear system driven by an exosignal
w t + 1 = S w t x t + 1 = A x t + B u t + E w t y t = C x t + F w t (1) \begin{aligned}
w_{t+1} &= Sw_t \\
x_{t+1} &= Ax_t+Bu_t+Ew_t \\
y_t &= Cx_t+Fw_t \tag*{(1)}
\end{aligned} w t + 1 x t + 1 y t = S w t = A x t + B u t + E w t = C x t + F w t (1)
where x t ∈ R n x_t\in\mathbb{R}^n x t ∈ R n is the state, u t ∈ R m u_t\in\mathbb{R}^m u t ∈ R m is the control input, y t ∈ R p y_t\in\mathbb{R}^p y t ∈ R p is the measured output, and w t ∈ R n w w_t\in\mathbb{R}^{n_w} w t ∈ R n w represents disturbances or references generated by the exosystem. The plant matrices A A A , B B B , C C C , E E E , F F F and the particular trajectory w t w_t w t are unknown, whereas the exosystem matrix S S S and one input-output trajectory { u t , y t } t = 0 N − 1 \{u_t,y_t\}_{t=0}^{N-1} { u t , y t } t = 0 N − 1 are available. If w t = 0 w_t=0 w t = 0 , Willems et al’s fundamental lemma [5 ] states that, under controllability and persistence of excitation, a candidate sequence { u ˉ t , y ˉ t } t = 0 L − 1 \{\bar{u}_t,\bar{y}_t\}_{t=0}^{L-1} { u ˉ t , y ˉ t } t = 0 L − 1 is a valid trajectory if and only if there exists g ∈ R N − L + 1 g\in\mathbb{R}^{N-L+1} g ∈ R N − L + 1 satisfying
\begin{bmatrix}
H_L(u_{[0,N-1]})\\
H_L(y_{[0,N-1]})
\end{bmatrix}g=
\begin{bmatrix}
\bar{u}\\
\bar{y}
\end{bmatrix}. \tag*{(2)}\end{bmatrix}
For constant offsets, i.e., S = I S=I S = I , [9 Theorem 1] accounts for the exogenous component by imposing the affine constraint ∑ i = 0 N − L g i = 1 \sum_{i=0}^{N-L}g_i=1 ∑ i = 0 N − L g i = 1 . Dynamic exosignals require a different treatment: the contribution of w t w_t w t cannot in general be separated from the measured input-output data, but its temporal structure is known through S S S . The following decomposition converts exosystem trajectories into a finite set of known basis signals with unknown coefficients.
Lemma 1 ([23 Lemmas 3.3–3.4] ). Consider any exosignal sequence { w t } t = 0 N − 1 \{w_t\}_{t=0}^{N-1} { w t } t = 0 N − 1 generated by (1a) with initial condition w 0 ∈ R n w w_0\in\mathbb{R}^{n_w} w 0 ∈ R n w . Then, for H 1 ( w ) : = [ w 0 w 1 ⋯ w N − L ] H_1(w):=[w_0\ w_1\ \cdots\ w_{N-L}] H 1 ( w ) := [ w 0 w 1 ⋯ w N − L ] , there exist an integer n M ≤ n w n_M\le n_w n M ≤ n w , a matrix L ( w 0 ) ∈ R n M × N \mathcal{L}(w_0)\in\mathbb{R}^{n_M\times N} L ( w 0 ) ∈ R n M × N and known vectors M ( i ) ∈ R n M \mathcal{M}(i)\in\mathbb{R}^{n_M} M ( i ) ∈ R n M , i = 0 , … , N − L i=0,\ldots,N-L i = 0 , … , N − L , such that
H 1 ( w ) = L ( w 0 ) [ M ( 0 ) M ( 1 ) ⋯ M ( N − L ) ] = L ( w 0 ) H 1 ( M ) (3) H_1(w)=\mathcal{L}(w_0)[\mathcal{M}(0)\ \mathcal{M}(1)\ \cdots\ \mathcal{M}(N-L)]=\mathcal{L}(w_0)H_1(\mathcal{M}) \tag*{(3)} H 1 ( w ) = L ( w 0 ) [ M ( 0 ) M ( 1 ) ⋯ M ( N − L )] = L ( w 0 ) H 1 ( M ) (3) where H 1 ( M ) ∈ R n M × ( N − L + 1 ) H_1(\mathcal{M})\in\mathbb{R}^{n_M\times(N-L+1)} H 1 ( M ) ∈ R n M × ( N − L + 1 ) is known and of full row rank.
Remark 1: In (3) , the matrix L ( w 0 ) \mathcal{L}(w_0) L ( w 0 ) depends on the initial condition w 0 w_0 w 0 , whereas H 1 ( M ) H_1(\mathcal{M}) H 1 ( M ) is determined solely by the eigenstructure of the exosystem matrix S S S . In particular, the rows of H 1 ( M ) H_1(\mathcal{M}) H 1 ( M ) form an independent set of exosystem basis signals. The explicit construction of H 1 ( M ) H_1(\mathcal{M}) H 1 ( M ) follows from [23 Claim C.1] and is omitted here for brevity. For instance, if S S S is diagonalizable with all eigenvalues on the unit circle, as in the periodic setting of Section III-E, one can take M ( t ) \mathcal{M}(t) M ( t ) as the stack of λ i t \lambda_i^t λ i t for each distinct real eigenvalue λ i \lambda_i λ i and of ( cos ( θ j t ) , sin ( θ j t ) ) (\cos(\theta_jt),\sin(\theta_jt)) ( cos ( θ j t ) , sin ( θ j t )) for each distinct complex-conjugate pair e ± i θ j e^{\pm\mathrm{i}\theta_j} e ± i θ j ; e.g., for the exosystem in Section V, M ( t ) = [ 1 cos ( ω t ) sin ( ω t ) cos ( 2 ω t ) sin ( 2 ω t ) cos ( 3 ω t ) sin ( 3 ω t ) ] ′ \mathcal{M}(t)=[1\ \cos(\omega t)\ \sin(\omega t)\ \cos(2\omega t)\ \sin(2\omega t)\ \cos(3\omega t)\ \sin(3\omega t)]^\prime M ( t ) = [ 1 cos ( ω t ) sin ( ω t ) cos ( 2 ω t ) sin ( 2 ω t ) cos ( 3 ω t ) sin ( 3 ω t ) ] ′ . In general, Jordan blocks of S S S additionally generate the terms t k λ t t^k\lambda^t t k λ t [23 Claim C.1] .
The data-based characterization also requires excitation not only of the control channel but also of the exosystem basis. This motivates the strengthened persistence-of-excitation condition.
Definition 1: Consider system (1) . Let { x t } t = 0 N − 1 \{x_t\}_{t=0}^{N-1} { x t } t = 0 N − 1 be the state sequence corresponding to the available input-output data { u t , y t } t = 0 N − 1 \{u_t,y_t\}_{t=0}^{N-1} { u t , y t } t = 0 N − 1 . We say that the data { u t , y t } t = 0 N − 1 \{u_t,y_t\}_{t=0}^{N-1} { u t , y t } t = 0 N − 1 are persistently exciting of order L L L if
rank ( [ H 1 ( x ) H L ( u ) H 1 ( M ) ] ) = m L + n + n M . (4) \operatorname{rank}\left(\begin{bmatrix} H_1(x) \\ H_L(u) \\ H_1(\mathcal{M}) \end{bmatrix}\right)=mL+n+n_M. \tag*{(4)} rank H 1 ( x ) H L ( u ) H 1 ( M ) = m L + n + n M . (4)
Since the state sequence x x x is not measured, the rank condition in (4) cannot be checked directly from the available input-output data. To address this issue, the following proposition provides a verifiable sufficient condition involving only the applied input and the known exosystem basis.
Proposition 1. Suppose that ( A , B ) (A,B) ( A , B ) is controllable and N ≥ L + n N \ge L+n N ≥ L + n . Let M ( ⋅ ) \mathcal{M}(\cdot) M ( ⋅ ) be the known exosystem basis associated with { w t } t = 0 N − 1 \{w_t\}_{t=0}^{N-1} { w t } t = 0 N − 1 as specified in Lemma 1, and define H 1 ( M [ 0 , N − L − n ] ) : = [ M ( 0 ) M ( 1 ) ⋯ M ( N − L − n ) ] H_1(\mathcal{M}_{[0,N-L-n]}):=[\mathcal{M}(0)\ \mathcal{M}(1)\ \cdots\ \mathcal{M}(N-L-n)] H 1 ( M [ 0 , N − L − n ] ) := [ M ( 0 ) M ( 1 ) ⋯ M ( N − L − n )] . If
rank ( [ H L + n ( u ) H 1 ( M [ 0 , N − L − n ] ) ] ) = m ( L + n ) + n M , (5) \operatorname{rank}\left(\begin{bmatrix} H_{L+n}(u) \\ H_1(\mathcal{M}_{[0,N-L-n]}) \end{bmatrix}\right)=m(L+n)+n_M, \tag*{(5)} rank ( [ H L + n ( u ) H 1 ( M [ 0 , N − L − n ] ) ] ) = m ( L + n ) + n M , (5) then the data are persistently exciting of order L L L in the sense of Definition 1, i.e., (4) holds.
Proof. Let λ ∈ R 1 × n \lambda\in\mathbb{R}^{1\times n} λ ∈ R 1 × n , η = [ η 0 ⋯ η L − 1 ] ∈ R 1 × m L \eta=[\eta_0\ \cdots\ \eta_{L-1}]\in\mathbb{R}^{1\times mL} η = [ η 0 ⋯ η L − 1 ] ∈ R 1 × m L , with η k ∈ R 1 × m \eta_k\in\mathbb{R}^{1\times m} η k ∈ R 1 × m , and ρ ∈ R 1 × n M \rho\in\mathbb{R}^{1\times n_M} ρ ∈ R 1 × n M satisfy
\begin{bmatrix}\lambda&\eta&\rho\end{bmatrix}
\begin{bmatrix}H_1(x)\\H_L(u)\\H_1(\mathcal{M})\end{bmatrix}=0. \tag*{(6)}\end{bmatrix}
We prove that λ = 0 \lambda=0 λ = 0 , η = 0 \eta=0 η = 0 , and ρ = 0 \rho=0 ρ = 0 .
By Lemma 1, we have w i = L ( w 0 ) M ( i ) w_i=\mathcal{L}(w_0)\mathcal{M}(i) w i = L ( w 0 ) M ( i ) . Moreover, the explicit construction of M ( i ) \mathcal{M}(i) M ( i ) in [23 Claim C.1] implies that the retained basis signals M ( i ) \mathcal{M}(i) M ( i ) are invariant under the forward-shift operator. In other words, there exists a known matrix S M ∈ R n M × n M S_M\in\mathbb{R}^{n_M\times n_M} S M ∈ R n M × n M , determined by S S S , such that M ( i + j ) = S M j M ( i ) \mathcal{M}(i+j)=S_M^j\mathcal{M}(i) M ( i + j ) = S M j M ( i ) , i , j ≥ 0 i,j\ge0 i , j ≥ 0 . Set D M : = E L ( w 0 ) D_M:=E\mathcal{L}(w_0) D M := E L ( w 0 ) . The state dynamics in (1b) can then be written as x i + 1 = A x i + B u i + D M M ( i ) x_{i+1}=Ax_i+Bu_i+D_M\mathcal{M}(i) x i + 1 = A x i + B u i + D M M ( i ) .
For i = 0 , … , N − L − n i=0,\ldots,N-L-n i = 0 , … , N − L − n , (6) gives λ x i + ∑ k = 0 L − 1 η k u i + k + ρ M ( i ) = 0 \lambda x_i+\sum_{k=0}^{L-1}\eta_k u_{i+k}+\rho\mathcal{M}(i)=0 λ x i + ∑ k = 0 L − 1 η k u i + k + ρ M ( i ) = 0 . Applying this relation at time i + j i+j i + j and recursively substituting (1b), for j = 0 , … , n j=0,\ldots,n j = 0 , … , n , yields
0 = λ A j x i + ∑ r = 0 j − 1 λ A j − 1 − r B u i + r + ∑ k = 0 L − 1 η k u i + j + k + ( ρ S M j + ∑ r = 0 j − 1 λ A j − 1 − r D M S M r ) M ( i ) . (7) 0=\lambda A^j x_i+\sum_{r=0}^{j-1}\lambda A^{j-1-r}Bu_{i+r}+\sum_{k=0}^{L-1}\eta_k u_{i+j+k}+\left(\rho S_M^j+\sum_{r=0}^{j-1}\lambda A^{j-1-r}D_MS_M^r\right)\mathcal{M}(i). \tag*{(7)} 0 = λ A j x i + r = 0 ∑ j − 1 λ A j − 1 − r B u i + r + k = 0 ∑ L − 1 η k u i + j + k + ( ρ S M j + r = 0 ∑ j − 1 λ A j − 1 − r D M S M r ) M ( i ) . (7)
where an empty sum is understood as zero.
Let p A ( z ) = ∑ j = 0 n β j z j p_A(z)=\sum_{j=0}^{n}\beta_jz^j p A ( z ) = ∑ j = 0 n β j z j with β n = 1 \beta_n=1 β n = 1 be the characteristic polynomial of A A A . By the Cayley–Hamilton theorem, ∑ j = 0 n β j A j = 0 \sum_{j=0}^{n}\beta_jA^j=0 ∑ j = 0 n β j A j = 0 . Multiplying (7) by β j \beta_j β j and summing over j = 0 , … , n j=0,\ldots,n j = 0 , … , n therefore eliminates the state term and gives ∑ q = 0 L + n − 1 η ˉ q u i + q + ρ ˉ M ( i ) = 0 \sum_{q=0}^{L+n-1}\bar{\eta}_q u_{i+q}+\bar{\rho}\mathcal{M}(i)=0 ∑ q = 0 L + n − 1 η ˉ q u i + q + ρ ˉ M ( i ) = 0 , where
η ˉ q : = ∑ 0 ≤ j ≤ n 0 ≤ q − j ≤ L − 1 β j η q − j + ∑ j = q + 1 n β j λ A j − 1 − q B \bar{\eta}_q:=\sum_{\substack{0\le j\le n\\0\le q-j\le L-1}}\beta_j\eta_{q-j}+\sum_{j=q+1}^{n}\beta_j\lambda A^{j-1-q}B η ˉ q := 0 ≤ j ≤ n 0 ≤ q − j ≤ L − 1 ∑ β j η q − j + j = q + 1 ∑ n β j λ A j − 1 − q B
and
ρ ˉ : = ∑ j = 0 n β j ( ρ S M j + ∑ r = 0 j − 1 λ A j − 1 − r D M S M r ) . \bar{\rho}:=\sum_{j=0}^{n}\beta_j\left(\rho S_M^j+\sum_{r=0}^{j-1}\lambda A^{j-1-r}D_MS_M^r\right). ρ ˉ := j = 0 ∑ n β j ( ρ S M j + r = 0 ∑ j − 1 λ A j − 1 − r D M S M r ) .
Thus,
[ η ˉ 0 ⋯ η ˉ L + n − 1 ρ ˉ ] [ H L + n ( u ) H 1 ( M [ 0 , N − L − n ] ) ] = 0. \begin{bmatrix}\bar{\eta}_0&\cdots&\bar{\eta}_{L+n-1}&\bar{\rho}\end{bmatrix}
\begin{bmatrix}H_{L+n}(u)\\H_1(\mathcal{M}_{[0,N-L-n]})\end{bmatrix}=0. [ η ˉ 0 ⋯ η ˉ L + n − 1 ρ ˉ ] [ H L + n ( u ) H 1 ( M [ 0 , N − L − n ] ) ] = 0.
By (5) , all the coefficient blocks above are zero. For s = L − 1 , … , 0 s=L-1,\ldots,0 s = L − 1 , … , 0 , the equation η ˉ n + s = 0 \bar{\eta}_{n+s}=0 η ˉ n + s = 0 has leading term β n η s = η s \beta_n\eta_s=\eta_s β n η s = η s , while its remaining terms involve only η s + 1 , … , η L − 1 \eta_{s+1},\ldots,\eta_{L-1} η s + 1 , … , η L − 1 . Backward substitution therefore gives η 0 = ⋯ = η L − 1 = 0 \eta_0=\cdots=\eta_{L-1}=0 η 0 = ⋯ = η L − 1 = 0 . With η = 0 \eta=0 η = 0 , the equations η ˉ n − 1 = 0 , … , η ˉ 0 = 0 \bar{\eta}_{n-1}=0,\ldots,\bar{\eta}_0=0 η ˉ n − 1 = 0 , … , η ˉ 0 = 0 successively imply λ B = λ A B = ⋯ = λ A n − 1 B = 0 \lambda B=\lambda AB=\cdots=\lambda A^{n-1}B=0 λ B = λ A B = ⋯ = λ A n − 1 B = 0 . Controllability of ( A , B ) (A,B) ( A , B ) then yields λ = 0 \lambda=0 λ = 0 . Returning to (6) gives ρ H 1 ( M ) = 0 \rho H_1(\mathcal{M})=0 ρ H 1 ( M ) = 0 . Consequently, since H 1 ( M ) H_1(\mathcal{M}) H 1 ( M ) has full row rank, ρ = 0 \rho=0 ρ = 0 . Therefore, the left kernel of the matrix in (4) is trivial, which completes the proof. ■
Note that Proposition 1 does not follow from applying the fundamental lemma to the extended input ( u , M ) (u,\mathcal{M}) ( u , M ) . Since M ( ⋅ ) \mathcal{M}(\cdot) M ( ⋅ ) is generated autonomously, rank H L + n ( M ) ≤ n M \operatorname{rank}H_{L+n}(\mathcal{M})\le n_M rank H L + n ( M ) ≤ n M , so the extended input can never be persistently exciting of order L + n L+n L + n . The proof above instead exploits the shift invariance M ( i + j ) = S M j M ( i ) \mathcal{M}(i+j)=S_M^j\mathcal{M}(i) M ( i + j ) = S M j M ( i ) together with the Cayley–Hamilton theorem, and only requires excitation of the first block row H 1 ( M ) H_1(\mathcal{M}) H 1 ( M ) .
We can now state the trajectory parametrization, which plays an important role in the controller design later.
Theorem 1. Suppose the input-output data u [ 0 , N − 1 ] u_{[0,N-1]} u [ 0 , N − 1 ] , y [ 0 , N − 1 ] y_{[0,N-1]} y [ 0 , N − 1 ] are generated by (1) and are persistently exciting of order L L L in the sense of Definition 1. Let M ˉ t : = M ( t ) \bar{M}_t:=\mathcal{M}(t) M ˉ t := M ( t ) . Then a candidate sequence u ˉ [ 0 , L − 1 ] ∣ t , y ˉ [ 0 , L − 1 ] ∣ t \bar{u}_{[0,L-1]|t},\bar{y}_{[0,L-1]|t} u ˉ [ 0 , L − 1 ] ∣ t , y ˉ [ 0 , L − 1 ] ∣ t is an input-output trajectory of (1) corresponding to the exosystem basis vector M ˉ t \bar{M}_t M ˉ t , i.e., driven by the exosignal w ˉ k ∣ t = L ( w 0 ) M ( t + k ) = w t + k \bar{w}_{k|t}=\mathcal{L}(w_0)\mathcal{M}(t+k)=w_{t+k} w ˉ k ∣ t = L ( w 0 ) M ( t + k ) = w t + k , k = 0 , … , L − 1 k=0,\ldots,L-1 k = 0 , … , L − 1 , of the same exosystem realization as the data at time index t t t , if and only if there exists g t ∈ R N − L + 1 g_t\in\mathbb{R}^{N-L+1} g t ∈ R N − L + 1 such that
\begin{bmatrix}H_L(y)\\H_L(u)\\H_1(\mathcal{M})\end{bmatrix}g_t=
\begin{bmatrix}\bar{y}_{[0,L-1]|t}\\\bar{u}_{[0,L-1]|t}\\\bar{M}_t\end{bmatrix}. \tag*{(8)}\end{bmatrix}
Proof. Proof of “if”: For input-output data { u t , y t } t = 0 N − 1 \{u_t,y_t\}_{t=0}^{N-1} { u t , y t } t = 0 N − 1 generated by (1) , it holds that
y [ i , i + L − 1 ] = O L x i + T L u u [ i , i + L − 1 ] + T ^ L w L ( w 0 ) M ( i ) (9) y_{[i,i+L-1]}=\mathcal{O}_Lx_i+\mathcal{T}_L^u u_{[i,i+L-1]}+\widehat{\mathcal{T}}_L^w\mathcal{L}(w_0)\mathcal{M}(i) \tag*{(9)} y [ i , i + L − 1 ] = O L x i + T L u u [ i , i + L − 1 ] + T L w L ( w 0 ) M ( i ) (9)
where
O L = [ C C A C A 2 ⋮ C A L − 1 ] , T L u = [ 0 0 ⋯ 0 0 C B 0 ⋯ 0 0 C A B C B ⋯ 0 0 ⋮ ⋮ ⋱ ⋮ ⋮ C A L − 2 B C A L − 3 B ⋯ C B 0 ] , \mathcal{O}_L=\begin{bmatrix}C\\CA\\CA^2\\\vdots\\CA^{L-1}\end{bmatrix},\qquad
\mathcal{T}_L^u=\begin{bmatrix}
0&0&\cdots&0&0\\
CB&0&\cdots&0&0\\
CAB&CB&\cdots&0&0\\
\vdots&\vdots&\ddots&\vdots&\vdots\\
CA^{L-2}B&CA^{L-3}B&\cdots&CB&0
\end{bmatrix}, O L = C C A C A 2 ⋮ C A L − 1 , T L u = 0 C B C A B ⋮ C A L − 2 B 0 0 C B ⋮ C A L − 3 B ⋯ ⋯ ⋯ ⋱ ⋯ 0 0 0 ⋮ C B 0 0 0 ⋮ 0 ,
T ^ L w = [ C E 0 ⋯ 0 0 C A E C E ⋯ 0 0 ⋮ ⋮ ⋱ ⋮ ⋮ C A L − 2 E C C A L − 3 E ⋯ C E F ] ⏟ : = T L w [ I S 2 ⋮ S L − 1 ] . (10) \widehat{\mathcal{T}}_L^w=
\underbrace{\begin{bmatrix}
CE&0&\cdots&0&0\\
CAE&CE&\cdots&0&0\\
\vdots&\vdots&\ddots&\vdots&\vdots\\
CA^{L-2}EC&CA^{L-3}E&\cdots&CE&F
\end{bmatrix}}_{:=\mathcal{T}_L^w}
\begin{bmatrix}I\\S_2\\\vdots\\S^{L-1}\end{bmatrix}. \tag*{(10)} T L w = := T L w C E C A E ⋮ C A L − 2 E C 0 C E ⋮ C A L − 3 E ⋯ ⋯ ⋱ ⋯ 0 0 ⋮ C E 0 0 ⋮ F I S 2 ⋮ S L − 1 . (10)
Indeed, (3) gives w i = L ( w 0 ) M ( i ) w_i=\mathcal{L}(w_0)\mathcal{M}(i) w i = L ( w 0 ) M ( i ) . Hence, if (8) holds, we have that
y ˉ [ 0 , L − 1 ] ∣ t = ∑ i = 0 N − L y [ i , i + L − 1 ] g t , i = ∑ i = 0 N − L g t , i ( O L x i + T L u u [ i , i + L − 1 ] + T ^ L w L ( w 0 ) M ( i ) ) = O L ∑ i = 0 N − L g t , i x i + T L u u ˉ [ 0 , L − 1 ] ∣ t + T ^ L w L ( w 0 ) M ˉ t . \begin{aligned}
\bar{y}_{[0,L-1]|t}=\sum_{i=0}^{N-L}y_{[i,i+L-1]}g_{t,i} \\
&=\sum_{i=0}^{N-L}g_{t,i}\left(\mathcal{O}_Lx_i+\mathcal{T}_L^u u_{[i,i+L-1]}+\widehat{\mathcal{T}}_L^w\mathcal{L}(w_0)\mathcal{M}(i)\right) \\
&=\mathcal{O}_L\sum_{i=0}^{N-L}g_{t,i}x_i+\mathcal{T}_L^u\bar{u}_{[0,L-1]|t}+\widehat{\mathcal{T}}_L^w\mathcal{L}(w_0)\bar{M}_t.
\end{aligned} y ˉ [ 0 , L − 1 ] ∣ t = i = 0 ∑ N − L y [ i , i + L − 1 ] g t , i = i = 0 ∑ N − L g t , i ( O L x i + T L u u [ i , i + L − 1 ] + T L w L ( w 0 ) M ( i ) ) = O L i = 0 ∑ N − L g t , i x i + T L u u ˉ [ 0 , L − 1 ] ∣ t + T L w L ( w 0 ) M ˉ t .
This is the output response of (1) over the horizon [ 0 , L − 1 ] [0,L-1] [ 0 , L − 1 ] under the input u ˉ [ 0 , L − 1 ] ∣ t \bar{u}_{[0,L-1]|t} u ˉ [ 0 , L − 1 ] ∣ t , with initial state x ˉ 0 ∣ t : = ∑ i = 0 N − L g t , i x i \bar{x}_{0|t}:=\sum_{i=0}^{N-L}g_{t,i}x_i x ˉ 0∣ t := ∑ i = 0 N − L g t , i x i and exosignal initial value w ˉ 0 ∣ t : = L ( w 0 ) M ˉ t \bar{w}_{0|t}:=\mathcal{L}(w_0)\bar{M}_t w ˉ 0∣ t := L ( w 0 ) M ˉ t . Hence, { ( u ˉ k ∣ t , y ˉ k ∣ t ) } k = 0 L − 1 \{(\bar{u}_{k|t},\bar{y}_{k|t})\}_{k=0}^{L-1} {( u ˉ k ∣ t , y ˉ k ∣ t ) } k = 0 L − 1 is an input-output trajectory of (1) corresponding to the exosystem basis vector M ˉ t \bar{M}_t M ˉ t .
Proof of “only if”: Let u ˉ [ 0 , L − 1 ] ∣ t , y ˉ [ 0 , L − 1 ] ∣ t \bar{u}_{[0,L-1]|t},\bar{y}_{[0,L-1]|t} u ˉ [ 0 , L − 1 ] ∣ t , y ˉ [ 0 , L − 1 ] ∣ t be an input-output trajectory of (1) with initial state x t x_t x t and exosignal basis vector M ˉ t = M ( t ) \bar{M}_t=\mathcal{M}(t) M ˉ t = M ( t ) . Since the data are persistently exciting of order L L L , the matrix
[ H 1 ( x ) H L ( u ) H 1 ( M ) ] \begin{bmatrix} H_1(x) \\ H_L(u) \\ H_1(\mathcal{M}) \end{bmatrix} H 1 ( x ) H L ( u ) H 1 ( M )
has full row rank. Therefore, there exists g t ∈ R N − L + 1 g_t \in\mathbb{R}^{N-L+1} g t ∈ R N − L + 1 such that
\begin{bmatrix} H_1(x) \\ H_L(u) \\ H_1(\mathcal{M}) \end{bmatrix}g_t =
\begin{bmatrix} \bar{u}_{[0,L-1]|t}^{x_t} \\ \bar{u}_{[0,L-1]} \\ \bar{M}_t \end{bmatrix}. \tag*{(11)}\end{bmatrix}
Using the trajectory expression and (11), we obtain
y ˉ [ 0 , L − 1 ] ∣ t = O L x t + T L u u ˉ [ 0 , L − 1 ] + T ^ L w L ( w 0 ) M ( t ) \bar{y}_{[0,L-1]|t} = \mathcal{O}_L x_t + T_L^u \bar{u}_{[0,L-1]} + \hat{T}_L^w \mathcal{L}(w_0)\mathcal{M}(t) y ˉ [ 0 , L − 1 ] ∣ t = O L x t + T L u u ˉ [ 0 , L − 1 ] + T ^ L w L ( w 0 ) M ( t )
= ( 11 ) ∑ i = 0 N − L g t , i ( O L x i + T L u u [ i , i + L − 1 ] + T ^ L w L ( w 0 ) M ( i ) ) \overset{(11)}{=} \sum_{i=0}^{N-L} g_{t,i}\left(\mathcal{O}_L x_i + T_L^u u_{[i,i+L-1]} + \hat{T}_L^w\mathcal{L}(w_0)\mathcal{M}(i)\right) = ( 11 ) i = 0 ∑ N − L g t , i ( O L x i + T L u u [ i , i + L − 1 ] + T ^ L w L ( w 0 ) M ( i ) )
= ( 9 ) ∑ i = 0 N − L g t , i y [ i , i + L − 1 ] = H L ( y ) g t . \overset{(9)}{=} \sum_{i=0}^{N-L} g_{t,i}y_{[i,i+L-1]} = H_L(y)g_t. = ( 9 ) i = 0 ∑ N − L g t , i y [ i , i + L − 1 ] = H L ( y ) g t .
Together with (11), this gives (8), which completes the proof. ■
Remark 2 (Scope of Theorem 1): The matrix L ( w 0 ) \mathcal{L}(w_0) L ( w 0 ) in (3) is unknown but common to the data and to the predictions, and the constraint H 1 ( M ) g t = M ˉ t H_1(\mathcal{M})g_t = \bar{M}_t H 1 ( M ) g t = M ˉ t with M ˉ t = M ( t ) \bar{M}_t = \mathcal{M}(t) M ˉ t = M ( t ) selects the exosignal w ˉ 0 ∣ t = L ( w 0 ) M ( t ) = w t \bar{w}_{0|t} = \mathcal{L}(w_0)\mathcal{M}(t) = w_t w ˉ 0∣ t = L ( w 0 ) M ( t ) = w t . Hence, Theorem 1 characterizes the trajectories driven by the same exosystem realization as the data, evaluated at a known time index t t t . Accordingly, throughout the paper, the exosignal acting during online operation is the continuation w t = S t w 0 w_t = S^t w_0 w t = S t w 0 of the realization in the data, and the controller knows the time index t t t on the time axis of the data, i.e., the phase of the exosystem, but neither w 0 w_0 w 0 nor the samples w t w_t w t (the latter are used only in the analytical formulation (21) to evaluate the feedforward term). This setting arises in repetitive tasks in which the phase of the exogenous signal is available while its amplitude is not, e.g., path following on a known route, where the road curvature is indexed by the traveled distance; rotating machinery, where periodic disturbances are indexed by the measured shaft angle; and grid-connected converters, where the grid phase is provided by a phase-locked loop. If the online exosignal belongs to a different realization, e.g., with a different initial condition or amplitude, Theorem 1 does not apply directly; the data must then be informative for that realization, or the corresponding exosystem coefficients must be reconstructed and included in the prediction constraint.
Theorem 1 is the prediction device used in the rest of the paper. It extends the affine-system fundamental lemma in [9 Theorem 1] from constant offsets to dynamic exosignals. Indeed, when S = I S = I S = I , one has H 1 ( M ) = [ 1 ⋯ 1 ] H_1(\mathcal{M}) = [1 \cdots1] H 1 ( M ) = [ 1 ⋯ 1 ] and M ˉ t = 1 \bar{M}_t = 1 M ˉ t = 1 , so the additional constraint reduces to the usual affine normalization. Thus, [9 Theorem 1] is recovered as a special case of Theorem 1.
Nominal DDPC for Output Regulation
We first consider the ideal case in which the available input-output data and the online measurements are noise-free. The section has two roles. First, it introduces a nominal DDPC formulation whose cost regularizes the input around the steady-state feedforward action, making the connection with model-based output-regulating MPC explicit. Second, for periodic exogenous signals, it replaces this unavailable feedforward action by an incremental-input penalty, yielding the implementable nominal scheme used later in the robust design.
Consider system (1) with m = p m = p m = p and y y y denoting the regulation error. The controller may use measured input-output data, the known exosystem structure, and online measurements, but not the matrices A A A , B B B , C C C , E E E , and F F F . To describe convergence to the regulator manifold, define the extended state x t : = ( x t , w t ) ∈ X ⊆ R n + n w x_t := (x_t,w_t) \in\mathbb{X} \subseteq\mathbb{R}^{n+n_w} x t := ( x t , w t ) ∈ X ⊆ R n + n w . The corresponding extended dynamics are
x t + 1 = [ A E 0 S ] ⏟ = : A x x t + [ B 0 ] ⏟ = : B x u t (12) x_{t+1} =
\underbrace{\begin{bmatrix} A & E \\ 0 & S \end{bmatrix}}_{=:A_x}x_t+
\underbrace{\begin{bmatrix} B \\ 0 \end{bmatrix}}_{=:B_x}u_t \tag*{(12)} x t + 1 = =: A x [ A 0 E S ] x t + =: B x [ B 0 ] u t (12)
y t = [ C F ] ⏟ = : C x x t . (13) y_t = \underbrace{\begin{bmatrix} C & F \end{bmatrix}}_{=:C_x}x_t. \tag*{(13)} y t = =: C x [ C F ] x t . (13)
The following regulation assumptions are used, consistently with [25 ] . Assumption 1 is the classical solvability condition for exact regulation; Assumption 2 gives the structural properties needed for finite-memory prediction and stabilization; and Assumption 3 excludes invariant zeros at the exosystem modes.
Assumption 1 (Output regulator equations): There exist matrices ( Π , Γ ) (\Pi,\Gamma) ( Π , Γ ) satisfying
Π S = A Π + B Γ + E (14) \Pi S = A\Pi+ B\Gamma+ E \tag*{(14)} Π S = A Π + B Γ + E (14)
0 = C\Pi + F.
{#eq:13b}
Assumption 2 (Controllability and observability): The pair ( A , B ) (A,B) ( A , B ) is controllable, and the pair ( A , C ) (A,C) ( A , C ) is observable.
Assumption 3 (Nonresonance condition): The following condition holds for all λ ∈ spec ( S ) \lambda\in\operatorname{spec}(S) λ ∈ spec ( S )
rank [ A − λ I B C 0 ] = n + p . (15) \operatorname{rank}\begin{bmatrix} A-\lambda I & B \\ C & 0 \end{bmatrix} = n+p. \tag*{(15)} rank [ A − λ I C B 0 ] = n + p . (15)
Remark 3: The matrix in (14) is ( n + p ) × ( n + m ) (n+p) \times(n+m) ( n + p ) × ( n + m ) , so full row rank requires m ≥ p m \ge p m ≥ p . We restrict attention to m = p m = p m = p ; Assumptions 1 and 3 then make the regulator pair ( Π , Γ ) (\Pi,\Gamma) ( Π , Γ ) unique. This uniqueness is used later to establish detectability of the incremental-input realization. When m > p m > p m > p , an input-allocation condition is needed to exclude distinct zero-output periodic inputs [25 Sec. V-B] .
Building on these preliminaries, the data-driven constrained output regulation problem is formally stated as follows.
Problem 1: Consider (1) with unknown matrices A A A , B B B , C C C , E E E , F F F satisfying Assumptions 1–3. Given input-output data { u t , y t } t = 0 N − 1 \{u_t,y_t\}_{t=0}^{N-1} { u t , y t } t = 0 N − 1 generated by (1), and with the online exosignal and the available phase information as specified in Remark 2, design a causal control sequence u t u_t u t such that the extended state x t x_t x t converges to the regulator manifold A : = { x ∣ x = Π w } \mathcal{A} := \{x \mid x = \Pi w\} A := { x ∣ x = Π w } and y t → 0 y_t \to0 y t → 0 as t → ∞ t \to\infty t → ∞ , while satisfying the imposed input constraints.
Thus the goal is equivalent to driving the distance measure σ ( x ) : = ∥ x − Π w ∥ 2 \sigma(x) := \lVert x-\Pi w\rVert^2 σ ( x ) := ∥ x − Π w ∥ 2 to zero. This expression is used for analysis only; in the implementation below, the controller does not require the matrices Π \Pi Π and Γ \Gamma Γ .
Auxiliary State-Space System
To express regulation in terms of measured input-output histories, we recall the auxiliary realization proposed in [23 Sec. 2] . This step replaces the unmeasured plant state by a finite window of past outputs and inputs, i.e., by information available to the data-driven controller. Since ( A , C ) (A,C) ( A , C ) is observable (cf. Assumption [2 ] ), let ι \iota ι be its observability index. Then
rank ( O ι ) = [ C C A ⋮ C A ι − 1 ] = n . (16) \operatorname{rank}(\mathcal{O}_{\iota}) =
\begin{bmatrix}
C\\
CA\\
\vdots\\
CA^{\iota-1}
\end{bmatrix}
= n. \tag*{(16)} rank ( O ι ) = C C A ⋮ C A ι − 1 = n . (16)
Hence there exists matrix O ι † \mathcal{O}_{\iota}^{\dagger} O ι † such that O ι † O ι = I \mathcal{O}_{\iota}^{\dagger}\mathcal{O}_{\iota}=I O ι † O ι = I . Consider the following dynamics:
ω t + 1 = S ω t (17) \omega_{t+1}=S\omega_t \tag*{(17)} ω t + 1 = S ω t (17)
ξ t + 1 = A ι ξ t + B ι v t + L ι Z 2 ω t (18) \xi_{t+1}=A_{\iota}\xi_t+B_{\iota}v_t+L_{\iota}Z_2\omega_t \tag*{(18)} ξ t + 1 = A ι ξ t + B ι v t + L ι Z 2 ω t (18)
φ t = Z 1 ξ t + Z 2 ω t (19) \varphi_t=Z_1\xi_t+Z_2\omega_t \tag*{(19)} φ t = Z 1 ξ t + Z 2 ω t (19)
where ξ ∈ R n ξ \xi\in\mathbb{R}^{n_{\xi}} ξ ∈ R n ξ with n ξ = ι ( m + p ) n_{\xi}=\iota(m+p) n ξ = ι ( m + p ) , φ t ∈ R p \varphi_t\in\mathbb{R}^{p} φ t ∈ R p , and
S : = block.diag ( S , … , S ) , A ι : = F ι + L ι Z 1 , S:=\operatorname{block.diag}(S,\ldots,S),
\qquad
A_{\iota}:=F_{\iota}+L_{\iota}Z_1, S := block.diag ( S , … , S ) , A ι := F ι + L ι Z 1 ,
with ( ι + 1 ) (\iota+1) ( ι + 1 ) times,
F ι : = [ 0 I p 0 ⋯ 0 0 0 I p ⋯ 0 0 0 0 ⋯ I p 0 0 0 ⋯ 0 0 p ι × m ι ] [ 0 m ι × p ι 0 I m 0 ⋯ 0 0 0 I m ⋯ 0 0 0 0 ⋯ I m 0 0 0 ⋯ 0 ] , F_{\iota}:=
\begin{bmatrix}
0&I_p&0&\cdots&0\\
0&0&I_p&\cdots&0\\
0&0&0&\cdots&I_p\\
0&0&0&\cdots&0\\
0_{p\iota\times m\iota}
\end{bmatrix}
\quad
\begin{bmatrix}
0_{m\iota\times p\iota}\\
0&I_m&0&\cdots&0\\
0&0&I_m&\cdots&0\\
0&0&0&\cdots&I_m\\
0&0&0&\cdots&0
\end{bmatrix}, F ι := 0 0 0 0 0 p ι × m ι I p 0 0 0 0 I p 0 0 ⋯ ⋯ ⋯ ⋯ 0 0 I p 0 0 m ι × p ι 0 0 0 0 I m 0 0 0 0 I m 0 0 ⋯ ⋯ ⋯ ⋯ 0 0 I m 0 ,
L ι : = [ 0 0 ⋮ 0 I p 0 0 ⋮ 0 ] , B ι : = [ 0 0 ⋮ 0 0 0 ⋮ 0 I m ] , L_{\iota}:=
\begin{bmatrix}
0\\
0\\
\vdots\\
0\\
I_p\\
0\\
0\\
\vdots\\
0
\end{bmatrix},
\qquad
B_{\iota}:=
\begin{bmatrix}
0\\
0\\
\vdots\\
0\\
0\\
0\\
\vdots\\
0\\
I_m
\end{bmatrix}, L ι := 0 0 ⋮ 0 I p 0 0 ⋮ 0 , B ι := 0 0 ⋮ 0 0 0 ⋮ 0 I m ,
Z 1 : = [ C A ι O ι † C R u − C A ι O ι † T u T ] , Z_1:=\begin{bmatrix}CA^{\iota}\mathcal{O}_{\iota}^{\dagger}&CR_u-CA^{\iota}\mathcal{O}_{\iota}^{\dagger}T_u^{\mathsf{T}}\end{bmatrix}, Z 1 := [ C A ι O ι † C R u − C A ι O ι † T u T ] ,
Z 2 : = [ C R w − C A ι O ι T T w T Q ] , Z_2:=\begin{bmatrix}CR_w-CA^{\iota}\mathcal{O}_{\iota}^{\mathsf{T}}T_w^{\mathsf{T}}&Q\end{bmatrix}, Z 2 := [ C R w − C A ι O ι T T w T Q ] ,
where matrices R u R_u R u , T u T_u T u , R w R_w R w , and T w T_w T w are defined analogously to those in (10), see [23 Section 2] for details. It was shown in [23 Lemma 4.4] that (16) has the same input-output behavior as the original LTI system (1); that is, for t ≥ ι t\geq\iota t ≥ ι , with v t = u t v_t=u_t v t = u t and
ω t = [ w t − ι y [ t − ι , t − 1 ] u [ t − ι , t − 1 ] ] , \omega_t=
\begin{bmatrix}
w_{t-\iota}\\
y_{[t-\iota,t-1]}\\
u_{[t-\iota,t-1]}
\end{bmatrix}, ω t = w t − ι y [ t − ι , t − 1 ] u [ t − ι , t − 1 ] ,
it holds that
ξ t = [ y [ t − ι , t − 1 ] u [ t − ι , t − 1 ] ] \xi_t=
\begin{bmatrix}
y_{[t-\iota,t-1]}\\
u_{[t-\iota,t-1]}
\end{bmatrix} ξ t = [ y [ t − ι , t − 1 ] u [ t − ι , t − 1 ] ]
and φ t = y t \varphi_t=y_t φ t = y t .
In the rest of the paper, we use u t : = v t u_t:=v_t u t := v t , ω t : = [ w t − ι ⋯ w t ] T \omega_t:=\begin{bmatrix}w_{t-\iota}&\cdots&w_t\end{bmatrix}^{\mathsf{T}} ω t := [ w t − ι ⋯ w t ] T , y t : = φ t y_t:=\varphi_t y t := φ t , and ξ t : = [ y [ t − ι , t − 1 ] u [ t − ι , t − 1 ] ] T \xi_t:=\begin{bmatrix}y_{[t-\iota,t-1]}&u_{[t-\iota,t-1]}\end{bmatrix}^{\mathsf{T}} ξ t := [ y [ t − ι , t − 1 ] u [ t − ι , t − 1 ] ] T for simplicity. With this notation, (16) is rewritten as
ω t + 1 = S ω t (20) \omega_{t+1}=S\omega_t \tag*{(20)} ω t + 1 = S ω t (20)
ξ t + 1 = A ι ξ t + B ι u t + E ι ω t (21) \xi_{t+1}=A_{\iota}\xi_t+B_{\iota}u_t+E_{\iota}\omega_t \tag*{(21)} ξ t + 1 = A ι ξ t + B ι u t + E ι ω t (21)
y t = Z 1 ξ t + Z 2 ω t (22) y_t=Z_1\xi_t+Z_2\omega_t \tag*{(22)} y t = Z 1 ξ t + Z 2 ω t (22)
where E ι : = L ι Z 2 E_{\iota}:=L_{\iota}Z_2 E ι := L ι Z 2 . Analogously to (12), let ϱ t : = ( ξ t , ω t ) ∈ X ϱ ⊆ R n ξ + n w \varrho_t:=(\xi_t,\omega_t)\in\mathcal{X}^{\varrho}\subseteq\mathbb{R}^{n_{\xi}+n_w} ϱ t := ( ξ t , ω t ) ∈ X ϱ ⊆ R n ξ + n w be the extended state of this auxiliary system. Then, (17) can be written as
ϱ t + 1 = [ A ι L ι Z 2 0 S ] ⏟ : = A ϱ ϱ t + [ B ι 0 ] ⏟ : = B ϱ u t (23) \varrho_{t+1}=
\underbrace{\begin{bmatrix}
A_{\iota}&L_{\iota}Z_2\\
0&S
\end{bmatrix}}_{:=A_{\varrho}}
\varrho_t+
\underbrace{\begin{bmatrix}
B_{\iota}\\
0
\end{bmatrix}}_{:=B_{\varrho}}u_t \tag*{(23)} ϱ t + 1 = := A ϱ [ A ι 0 L ι Z 2 S ] ϱ t + := B ϱ [ B ι 0 ] u t (23)
y t = [ Z 1 Z 2 ] ⏟ : = C ϱ ϱ t . (24) y_t=\underbrace{\begin{bmatrix}Z_1&Z_2\end{bmatrix}}_{:=C_{\varrho}}\varrho_t. \tag*{(24)} y t = := C ϱ [ Z 1 Z 2 ] ϱ t . (24)
Nominal DDPC
For clean input-output data, the prediction model of Theorem 1 can be incorporated into a receding-horizon regulator. The resulting optimization problem is
min y ˉ , u ˉ , w ˉ , g J L ( ξ t , ω t ) : = ∑ k = 0 L − 1 ∥ y ˉ k ∣ t ∥ Q 2 + ∥ u ˉ k ∣ t − Γ ξ w ˉ k ∣ t ∥ R 2 (25) \min_{\bar{y},\bar{u},\bar{w},g} J_L(\xi_t,\omega_t)
:=\sum_{k=0}^{L-1}\left\|\bar{y}_{k|t}\right\|_{Q}^{2}
+\left\|\bar{u}_{k|t}-\Gamma_{\xi}\bar{w}_{k|t}\right\|_{R}^{2} \tag*{(25)} y ˉ , u ˉ , w ˉ , g min J L ( ξ t , ω t ) := k = 0 ∑ L − 1 y ˉ k ∣ t Q 2 + u ˉ k ∣ t − Γ ξ w ˉ k ∣ t R 2 (25)
s.t. [ H L + ν ( y ) H L + ν ( u ) H 1 ( w ) ] g t = [ y ˉ [ − ν , L − 1 ] ∣ t u ˉ [ − ν , L − 1 ] ∣ t w ˉ [ − ν , L − 1 ] ∣ t ] (26) \text{s.t.}\quad
\begin{bmatrix}
H_{L+\nu}(y)\\
H_{L+\nu}(u)\\
H_1(w)
\end{bmatrix}g_t
=
\begin{bmatrix}
\bar{y}_{[-\nu,L-1]|t}\\
\bar{u}_{[-\nu,L-1]|t}\\
\bar{w}_{[-\nu,L-1]|t}
\end{bmatrix} \tag*{(26)} s.t. H L + ν ( y ) H L + ν ( u ) H 1 ( w ) g t = y ˉ [ − ν , L − 1 ] ∣ t u ˉ [ − ν , L − 1 ] ∣ t w ˉ [ − ν , L − 1 ] ∣ t (26)
[ y ˉ [ − ν , L − 1 ] ∣ t u ˉ [ − ν , L − 1 ] ∣ t ] = [ y [ t − ν , t − 1 ] u [ t − ν , t − 1 ] ] (27) \begin{bmatrix}
\bar{y}_{[-\nu,L-1]|t}\\
\bar{u}_{[-\nu,L-1]|t}
\end{bmatrix}
=
\begin{bmatrix}
y_{[t-\nu,t-1]}\\
u_{[t-\nu,t-1]} \tag*{(27)}
\end{bmatrix} [ y ˉ [ − ν , L − 1 ] ∣ t u ˉ [ − ν , L − 1 ] ∣ t ] = [ y [ t − ν , t − 1 ] u [ t − ν , t − 1 ] ] (27)
ω ˉ 0 ∣ t = [ w t − ν ⋮ w t ] (28) \bar{\omega}_{0|t}=
\begin{bmatrix}
w_{t-\nu}\\
\vdots\\
w_t
\end{bmatrix} \tag*{(28)} ω ˉ 0∣ t = w t − ν ⋮ w t (28)
ω ˉ k + 1 ∣ t = S ω ˉ k ∣ t , k = 0 , … , L − 1 (29) \bar{\omega}_{k+1|t}=S\bar{\omega}_{k|t},\qquad k=0,\ldots,L-1 \tag*{(29)} ω ˉ k + 1∣ t = S ω ˉ k ∣ t , k = 0 , … , L − 1 (29)
u ˉ k ∣ t ∈ U , k = 0 , … , L − 1 (30) \bar{u}_{k|t}\in\mathbb{U},\qquad k=0,\ldots,L-1 \tag*{(30)} u ˉ k ∣ t ∈ U , k = 0 , … , L − 1 (30)
where L ∈ N L\in\mathbb{N} L ∈ N is the prediction horizon and U \mathbb{U} U is a compact input-constraint set. The decision variables are the predicted output, input, exosystem basis trajectory, and Hankel coefficient vector. The constraints and cost have the following interpretation.
ℓ ( ξ , ω ) : = ∥ y ∥ Q 2 + ∥ u − Γ ξ ω ∥ R 2 (31) \ell(\xi,\omega):=\lVert y\rVert_{Q}^{2}+\lVert u-\Gamma_{\xi}\omega\rVert_{R}^{2} \tag*{(31)} ℓ ( ξ , ω ) := ∥ y ∥ Q 2 + ∥ u − Γ ξ ω ∥ R 2 (31)
penalizes both the regulation error and the deviation of the predicted input from the steady-state feedforward action Γ ξ ω \Gamma_{\xi}\omega Γ ξ ω , with Q = Q T ≻ 0 Q=Q^{\mathsf{T}}\succ0 Q = Q T ≻ 0 and R = R T ≻ 0 R=R^{\mathsf{T}}\succ0 R = R T ≻ 0 .
Constraint (21b) is a data-driven prediction model supplied by Theorem 1; it replaces the state-space equations used in the model-based MPC of [25, eq. (5)].
Constraints (21c)–(21d) ensure that the internal state of the true trajectory aligns with the internal state of the predicted trajectory at time t t t .
The cost (22) also reveals the main limitation of (21): implementation requires the feedforward signal Γ ξ ω \Gamma_{\xi}\omega Γ ξ ω . This requirement is restrictive because Γ ξ \Gamma_{\xi} Γ ξ is tied to the unknown regulator equations. Therefore, (21) serves as an analytical bridge to existing output-regulating MPC results; the feedforward requirement is removed in Section III-E.
When Problem (21) is feasible at time t t t , denote its optimizer by J L ⋆ J^\star_{\mathrm{L}} J L ⋆ , ( y ˉ ⋆ , u ˉ ⋆ , ω ˉ ⋆ ) (\bar{y}^{\star},\bar{u}^{\star},\bar{\omega}^{\star}) ( y ˉ ⋆ , u ˉ ⋆ , ω ˉ ⋆ ) , and g t ⋆ g^\star_t g t ⋆ . The resulting receding-horizon implementation is summarized in Algorithm 1.
Table 1 (PDF p. 6).
The following result gives the closed-loop guarantee associated with Algorithm 1.
Closed-Loop Stability Guarantees
Before proceeding, we impose a uniform interior condition on the regulator input. It provides the input margin needed for the local feasibility argument below.
Assumption 4 (Admissibility of regulator input): For the exosignal ω t \omega_t ω t in (17a), letting B r U ( Γ ξ ω t ) : = { v ∈ R m ∣ ∥ v − Γ ξ ω t ∥ ≤ r U } \mathbb{B}_{r_{\mathbb{U}}}(\Gamma_{\xi}\omega_t):=\{v\in\mathbb{R}^{m}\mid\|v-\Gamma_{\xi}\omega_t\|\le r_{\mathbb{U}}\} B r U ( Γ ξ ω t ) := { v ∈ R m ∣ ∥ v − Γ ξ ω t ∥ ≤ r U } , there exists r U > 0 r_{\mathbb{U}}>0 r U > 0 such that B r U ( Γ ξ ω t ) ⊆ U \mathbb{B}_{r_{\mathbb{U}}}(\Gamma_{\xi}\omega_t)\subseteq\mathbb{U} B r U ( Γ ξ ω t ) ⊆ U for all t ≥ 0 t\ge0 t ≥ 0 .
We next recall two ingredients from [25, Section V.D]. Under Assumptions 1–4, the following bounds hold for the corresponding output-regulating MPC problem.
Proposition 2: Under Assumptions 2 and 4, there exist constants γ s , δ s > 0 \gamma_s,\delta_s>0 γ s , δ s > 0 , such that for any ϱ ∈ X δ s : = { ϱ ∈ X ∣ σ ( ϱ ) ≤ δ s } \varrho\in\mathbb{X}^{\delta_s}:=\{\varrho\in\mathbb{X}\mid\sigma(\varrho)\le\delta_s\} ϱ ∈ X δ s := { ϱ ∈ X ∣ σ ( ϱ ) ≤ δ s } with σ ( ϱ ) \sigma(\varrho) σ ( ϱ ) defined in (20), problem (21) is feasible and the objective function satisfies
J L ⋆ ( ξ t , ω t ) ≤ γ s ∥ ξ t − Π Ξ ω t ∥ 2 . (32) J^\star_{\mathrm{L}}(\xi_t,\omega_t)\le\gamma_s\|\xi_t-\Pi_{\Xi}\omega_t\|^2. \tag*{(32)} J L ⋆ ( ξ t , ω t ) ≤ γ s ∥ ξ t − Π Ξ ω t ∥ 2 . (32)
Proposition 3: Under Assumptions 1 and 2, system (18) is incrementally input-output-to-state stable (i-IOSS) [26 Definition 4] . In addition, there exist matrix P ξ , w = P ξ , w T ≻ 0 P_{\xi,\mathrm{w}}=P_{\xi,\mathrm{w}}^{\mathsf{T}}\succ0 P ξ , w = P ξ , w T ≻ 0 and constant ϵ 0 > 0 \epsilon_0>0 ϵ 0 > 0 such that, with W ( ϱ ) : = ∥ ξ t − Π Ξ ω t ∥ P ξ , w 2 W(\varrho):=\|\xi_t-\Pi_{\Xi}\omega_t\|^2_{P_{\xi,\mathrm{w}}} W ( ϱ ) := ∥ ξ t − Π Ξ ω t ∥ P ξ , w 2 , the following inequalities hold for all ( ϱ t , u t ) (\varrho_t,u_t) ( ϱ t , u t ) :
W ( ϱ t ) ≥ ϵ 0 ∥ ξ t − Π Ξ ω t ∥ 2 , (33) W(\varrho_t)\ge\epsilon_0\|\xi_t-\Pi_{\Xi}\omega_t\|^2, \tag*{(33)} W ( ϱ t ) ≥ ϵ 0 ∥ ξ t − Π Ξ ω t ∥ 2 , (33)
W ( ϱ t + 1 ) ≤ W ( ϱ t ) + ∥ u t − Γ ξ ω t ∥ R 2 + ∥ y t ∥ Q 2 − ϵ 0 ∥ ξ t − Π Ξ ω t ∥ 2 . (34) W(\varrho_{t+1})\le W(\varrho_t)+\|u_t-\Gamma_{\xi}\omega_t\|^2_R+\|y_t\|^2_Q-\epsilon_0\|\xi_t-\Pi_{\Xi}\omega_t\|^2. \tag*{(34)} W ( ϱ t + 1 ) ≤ W ( ϱ t ) + ∥ u t − Γ ξ ω t ∥ R 2 + ∥ y t ∥ Q 2 − ϵ 0 ∥ ξ t − Π Ξ ω t ∥ 2 . (34)
For the stability analysis of Algorithm 1, consider the Lyapunov function Y L ( ϱ t ) : = J L ⋆ ( ϱ t ) + W ( ϱ t ) Y_{\mathrm{L}}(\varrho_t):=J^\star_{\mathrm{L}}(\varrho_t)+W(\varrho_t) Y L ( ϱ t ) := J L ⋆ ( ϱ t ) + W ( ϱ t ) . The following theorem establishes the closed-loop stability of the proposed data-driven formulation. The proof combines the trajectory equivalence in Theorem 1 with the finite-horizon Lyapunov framework for output-regulating MPC in [25, Theorem 1].
Theorem (Nominal stability guarantees): Suppose Assumptions 1–4 hold and the I/O data are persistently exciting of order L + ι L+\iota L + ι in the sense of Definition 1. Then, for any Y ˉ > 0 \bar{Y}>0 Y ˉ > 0 , there exists prediction horizon L ˉ Y ˉ > 0 \bar{L}_{\bar{Y}}>0 L ˉ Y ˉ > 0 such that, for all L ≥ L ˉ Y ˉ L\ge\bar{L}_{\bar{Y}} L ≥ L ˉ Y ˉ and all initial conditions ϱ 0 ∈ X ϱ Y ˉ : = { ϱ 0 ∈ X ϱ ∣ Y L ( ϱ 0 ) ≤ Y ˉ } \varrho_0\in\mathbb{X}^{\bar{Y}}_{\varrho}:=\{\varrho_0\in\mathbb{X}^{\varrho}\mid Y_{\mathrm{L}}(\varrho_0)\le\bar{Y}\} ϱ 0 ∈ X ϱ Y ˉ := { ϱ 0 ∈ X ϱ ∣ Y L ( ϱ 0 ) ≤ Y ˉ } , the DDPC problem (21) is recursively feasible and the regulator manifold A ϱ \mathcal{A}_{\varrho} A ϱ is exponentially stable for the closed loop.
Proof. See Appendix A.
The controller (21) requires the feedforward input Γ ξ ω \Gamma_{\xi}\omega Γ ξ ω , and hence either knowledge of the regulator gain or direct access to the regulated steady-state input. This is undesirable in a data-driven setting. We now focus on periodic exosignals and replace feedforward regularization by incremental-input regularization, which is computable from past applied inputs.
Incremental Input Regularization for Periodic Exosignals
The periodic setting is captured by the following assumption.
Assumption 5 (Periodic exosystem): The exosystem matrix S S S in (1a) is known and has finite order T ∈ N + T\in\mathbb{N}_+ T ∈ N + , i.e., S T = I S^T=I S T = I .
Remark 4. Assumption 5 implies that every trajectory generated by w t + 1 = S w t w_{t+1}=Sw_t w t + 1 = S w t is T T T -periodic. Equivalently, S S S is diagonalizable over C \mathbb{C} C and all its eigenvalues are T T T -th roots of unity. Thus, for a real-valued exosystem, the eigenvalues of S S S consist of 1 1 1 , possibly − 1 -1 − 1 , and complex-conjugate pairs e ± i 2 π q / T e^{\pm i2\pi q/T} e ± i 2 π q / T , with no nontrivial Jordan blocks.
The same periodic formulation also covers the case where the particular linear matrix S S S is not known, or where the exosignal is generated by a nonlinear periodic map, provided that its period T T T is known. Indeed, for any T T T -periodic sequence { w t } \{w_t\} { w t } , define the lifted exosystem state
w t : = [ w t − T + 1 w t − T + 2 ⋮ w t ] , \mathbf{w}_t:=
\begin{bmatrix}
w_{t-T+1}\\
w_{t-T+2}\\
\vdots\\
w_t
\end{bmatrix}, w t := w t − T + 1 w t − T + 2 ⋮ w t ,
and the cyclic-shift matrix
S T : = [ 0 I n w 0 ⋯ 0 0 0 I n w ⋯ 0 ⋮ ⋮ ⋮ ⋱ ⋮ 0 0 0 ⋯ I n w I n w 0 0 ⋯ 0 ] , S s e l : = [ 0 0 ⋯ I n w ] . S_T:=
\begin{bmatrix}
0 & I_{n_w} & 0 & \cdots& 0\\
0 & 0 & I_{n_w} & \cdots& 0\\
\vdots& \vdots& \vdots& \ddots& \vdots\\
0 & 0 & 0 & \cdots& I_{n_w}\\
I_{n_w} & 0 & 0 & \cdots& 0
\end{bmatrix},
\qquad
S_{\mathrm{sel}}:=\begin{bmatrix}0 & 0 & \cdots& I_{n_w}\end{bmatrix}. S T := 0 0 ⋮ 0 I n w I n w 0 ⋮ 0 0 0 I n w ⋮ 0 0 ⋯ ⋯ ⋱ ⋯ ⋯ 0 0 ⋮ I n w 0 , S sel := [ 0 0 ⋯ I n w ] .
Then w t + 1 = S T w t \mathbf{w}_{t+1}=S_T\mathbf{w}_t w t + 1 = S T w t , S T T = I S_T^T=I S T T = I , and w t = S s e l w t w_t=S_{\mathrm{sel}}\mathbf{w}_t w t = S sel w t . Hence (1) can be equivalently written as
x t + 1 = A x t + B u t + E S s e l w t , x_{t+1}=Ax_t+Bu_t+ES_{\mathrm{sel}}\mathbf{w}_t, x t + 1 = A x t + B u t + E S sel w t ,
y t = C x t + F S s e l w t , y_t=Cx_t+FS_{\mathrm{sel}}\mathbf{w}_t, y t = C x t + F S sel w t ,
w t + 1 = S T w t . \mathbf{w}_{t+1}=S_T\mathbf{w}_t. w t + 1 = S T w t .
Therefore, for periodic exosignals, knowledge of the period is sufficient to construct a known linear periodic exosystem.
Assumption 6 (Nonresonance at the T T T -th roots of unity): Condition (15) holds for all λ ∈ C \lambda\in\mathbb{C} λ ∈ C satisfying λ T = 1 \lambda^T=1 λ T = 1 .
Remark 5: Under Assumption 5, every eigenvalue of S S S satisfies λ T = 1 \lambda^T=1 λ T = 1 , so Assumption 6 implies Assumption 3. The strengthening is needed only for the incremental-input formulation below. The input memory used to form u t − u t − T u_t-u_{t-T} u t − u t − T has all T T T -th roots of unity as eigenvalues, including those that are not eigenvalues of S S S . An input component oscillating at such a frequency is not penalized by the incremental-input cost, and it must therefore be detectable through the regulated output, cf. the proof of Proposition 4. Assumption 6 fails only if the plant has an invariant zero located exactly at a T T T -th root of unity; hence, it is generic, and it holds in the example of Section V.
Under Assumption 5, the steady-state feedforward input is periodic. At regulation one therefore has u t = u t − T u_t=u_{t-T} u t = u t − T , so the unavailable term Γ ξ ω \Gamma\xi_\omega Γ ξ ω can be replaced by the measurable periodic input increment u t − u t − T u_t-u_{t-T} u t − u t − T . The resulting data-driven counterpart of [25 eq. (18)] is
min y ˉ , u ˉ , g J L d ( ξ t , ζ t ) = ∑ k = 0 L − 1 ∥ y ˉ k ∣ t ∥ Q 2 + ∥ Δ u ˉ k ∣ t ∥ R 2 (35) \min_{\bar{y},\bar{u},g} J_L^d(\xi_t,\zeta_t)=\sum_{k=0}^{L-1}\|\bar{y}_{k|t}\|_Q^2+\|\Delta\bar{u}_{k|t}\|_R^2 \tag*{(35)} y ˉ , u ˉ , g min J L d ( ξ t , ζ t ) = k = 0 ∑ L − 1 ∥ y ˉ k ∣ t ∥ Q 2 + ∥Δ u ˉ k ∣ t ∥ R 2 (35)
s.t. [ H L + ι ( y ) H L + ι ( u ) H 1 ( M ) ] g t = [ y ˉ [ − ι , L − 1 ] ∣ t u ˉ [ − ι , L − 1 ] ∣ t M ˉ t − ι ] (36) \text{s.t.}\quad
\begin{bmatrix}
H_{L+\iota}(y)\\
H_{L+\iota}(u)\\
H_1(M)
\end{bmatrix}g_t=
\begin{bmatrix}
\bar{y}_{[-\iota,L-1]|t}\\
\bar{u}_{[-\iota,L-1]|t}\\
\bar{M}_{t-\iota}
\end{bmatrix} \tag*{(36)} s.t. H L + ι ( y ) H L + ι ( u ) H 1 ( M ) g t = y ˉ [ − ι , L − 1 ] ∣ t u ˉ [ − ι , L − 1 ] ∣ t M ˉ t − ι (36)
[ y ˉ [ − ι , − 1 ] ∣ t u ˉ [ − ι , − 1 ] ∣ t ] = [ y [ t − ι , t − 1 ] u [ t − ι , t − 1 ] ] (37) \begin{bmatrix}
\bar{y}_{[-\iota,-1]|t}\\
\bar{u}_{[-\iota,-1]|t}
\end{bmatrix}
=
\begin{bmatrix}
y_{[t-\iota,t-1]}\\
u_{[t-\iota,t-1]} \tag*{(37)}
\end{bmatrix} [ y ˉ [ − ι , − 1 ] ∣ t u ˉ [ − ι , − 1 ] ∣ t ] = [ y [ t − ι , t − 1 ] u [ t − ι , t − 1 ] ] (37)
u ˉ k ∣ t = u ˉ k − T ∣ t + Δ u ˉ k ∣ t , k = 0 , … , L − 1 (38) \bar{u}_{k|t}=\bar{u}_{k-T|t}+\Delta\bar{u}_{k|t},\qquad k=0,\ldots,L-1 \tag*{(38)} u ˉ k ∣ t = u ˉ k − T ∣ t + Δ u ˉ k ∣ t , k = 0 , … , L − 1 (38)
u ˉ k − T ∣ t = { u t + k − T , k < T , u ˉ k − T ∣ t , k ≥ T (39) \bar{u}_{k-T|t}=
\begin{cases}
u_{t+k-T}, & k<T,\\
\bar{u}_{k-T|t}, & k\ge T
\end{cases} \tag*{(39)} u ˉ k − T ∣ t = { u t + k − T , u ˉ k − T ∣ t , k < T , k ≥ T (39)
u ˉ k ∣ t ∈ U , k = 0 , … , L − 1. (40) \bar{u}_{k|t}\in\mathbb{U},\qquad k=0,\ldots,L-1. \tag*{(40)} u ˉ k ∣ t ∈ U , k = 0 , … , L − 1. (40)
The key point here is that (25) is not merely a heuristic replacement of the steady-state feedforward term. As in [25 Section IV.B] , it is equivalent to a feedforward-regularized MPC problem posed on an augmented plant whose input is the periodic increment.
Augmented plant: Define an input memory state ζ t : = [ u t − 1 ⊤ ⋯ u t − T ⊤ ] ⊤ ∈ R m T \zeta_t:=[u_{t-1}^{\top}\ \cdots\ u_{t-T}^{\top}]^{\top}\in\mathbb{R}^{mT} ζ t := [ u t − 1 ⊤ ⋯ u t − T ⊤ ] ⊤ ∈ R m T and set n a : = n ξ + m T n_a:=n_\xi+mT n a := n ξ + m T . Consider
ξ t a = ( ξ t , ζ t ) ∈ R n a , u t a = u t − u t − T ∈ R m . (41) \xi_t^a=(\xi_t,\zeta_t)\in\mathbb{R}^{n_a},\qquad u_t^a=u_t-u_{t-T}\in\mathbb{R}^m. \tag*{(41)} ξ t a = ( ξ t , ζ t ) ∈ R n a , u t a = u t − u t − T ∈ R m . (41)
Define the block cyclic permutation matrix E 0 ∈ R m T × m T E_0\in\mathbb{R}^{mT\times mT} E 0 ∈ R m T × m T and the selection matrices E 1 , E 2 ∈ R m T × m E_1,E_2\in\mathbb{R}^{mT\times m} E 1 , E 2 ∈ R m T × m :
E 0 : = [ 0 m ⋯ 0 m I m I m ⋯ 0 m 0 m 0 m ⋱ ⋮ ⋮ 0 m ⋯ I m 0 m ] , E 1 : = [ I m 0 m ⋮ 0 m ] , E 2 : = [ 0 m ⋮ 0 m I m ] . E_0:=\begin{bmatrix}
0_m&\cdots&0_m&I_m\\
I_m&\cdots&0_m&0_m\\
0_m&\ddots&\vdots&\vdots\\
0_m&\cdots&I_m&0_m
\end{bmatrix},\qquad
E_1:=\begin{bmatrix}I_m\\0_m\\\vdots\\0_m\end{bmatrix},\qquad
E_2:=\begin{bmatrix}0_m\\\vdots\\0_m\\I_m\end{bmatrix}. E 0 := 0 m I m 0 m 0 m ⋯ ⋯ ⋱ ⋯ 0 m 0 m ⋮ I m I m 0 m ⋮ 0 m , E 1 := I m 0 m ⋮ 0 m , E 2 := 0 m ⋮ 0 m I m .
This matrix satisfies E 0 ⊤ E 0 = I m T E_0^\top E_0=I_{mT} E 0 ⊤ E 0 = I m T and the eigenvalues of E 0 E_0 E 0 are λ k = e i 2 π k / T \lambda_k=e^{i2\pi k/T} λ k = e i 2 π k / T , k = 0 , … , T − 1 k=0,\ldots,T-1 k = 0 , … , T − 1 , all with a geometric and algebraic multiplicity of m m m (due to the block structure). We note that E 1 ⊤ E 0 = E 2 ⊤ E_1^\top E_0=E_2^\top E 1 ⊤ E 0 = E 2 ⊤ . The dynamics of the memory state ζ \zeta ζ and the input u u u can be compactly expressed as
ζ t + 1 = E 0 ζ t + E 1 u t a , u t = E 2 ⊤ ζ t + u t a . (42) \zeta_{t+1}=E_0\zeta_t+E_1u_t^a,\qquad u_t=E_2^\top\zeta_t+u_t^a. \tag*{(42)} ζ t + 1 = E 0 ζ t + E 1 u t a , u t = E 2 ⊤ ζ t + u t a . (42)
Thus, the augmented plant dynamics becomes
ω t + 1 = S ω t (43) \omega_{t+1}=S\omega_t \tag*{(43)} ω t + 1 = S ω t (43)
ξ t + 1 a = [ A ι B ι E 2 ⊤ 0 E 0 ] ξ t a + [ B ι E 1 ] u t a + [ E 0 ] ω t = : A a ξ t a + B a u t a + E a ω t (44) \xi_{t+1}^a=
\begin{bmatrix}
A_\iota&B_\iota E_2^\top\\
0&E_0
\end{bmatrix}\xi_t^a+
\begin{bmatrix}
B_\iota\\
E_1
\end{bmatrix}u_t^a+
\begin{bmatrix}
E\\
0
\end{bmatrix}\omega_t
=:A^a\xi_t^a+B^au_t^a+E^a\omega_t \tag*{(44)} ξ t + 1 a = [ A ι 0 B ι E 2 ⊤ E 0 ] ξ t a + [ B ι E 1 ] u t a + [ E 0 ] ω t =: A a ξ t a + B a u t a + E a ω t (44)
y t = [ Z 1 0 ] ξ t a + Z 2 ω t = : C a ξ t a + Z 2 ω t . (45) y_t=\begin{bmatrix}Z_1&0\end{bmatrix}\xi_t^a+Z_2\omega_t=:C^a\xi_t^a+Z_2\omega_t. \tag*{(45)} y t = [ Z 1 0 ] ξ t a + Z 2 ω t =: C a ξ t a + Z 2 ω t . (45)
The overall state of this augmented plant is ρ a : = ( ξ a , ω ) \rho^a:=(\xi^a,\omega) ρ a := ( ξ a , ω ) . Using the periodic input-memory augmentation motivated by [25 Lemma 1] , the feedforward-regularized DDPC can be equivalently expressed as the implementable incremental form (25).
Lemma 2: The DDPC problem (21) with the augmented state ρ a = ( ξ a , ω ) \rho^a=(\xi^a,\omega) ρ a = ( ξ a , ω ) , input u a u^a u a , the dynamics (41) –(45) , and stage cost ℓ ( ρ a , u a ) = ∥ y ∥ Q 2 + ∥ u a ∥ R 2 \ell(\rho^a,u^a)=\|y\|_Q^2+\|u^a\|_R^2 ℓ ( ρ a , u a ) = ∥ y ∥ Q 2 + ∥ u a ∥ R 2 using incremental input regularization is equivalent to the DDPC problem (25).
It remains to verify that the augmented realization retains the structural properties required for the stability analysis. These properties are established next using the PBH conditions for the periodic input-memory augmentation (cf. [25 Propositions 5 and 7] ).
Proposition 4: Let Assumptions 1–6 hold. The triplet ( A a , B a , C a ) (A^a,B^a,C^a) ( A a , B a , C a ) of the augmented system (41) –(28c) is stabilizable and detectable, and the augmented regulator equations Π a S = A a Π a + B a Γ a + E a \Pi^a S=A^a\Pi^a+B^a\Gamma^a+E^a Π a S = A a Π a + B a Γ a + E a , 0 = C a Π a + Z 2 0=C^a\Pi^a+Z_2 0 = C a Π a + Z 2 are solvable.
Proof: See Appendix B.
We can now summarize the closed-loop consequence for the implementable nominal DDPC scheme.
Corollary 1: Suppose that the plant (1) satisfies Assumptions 1–6 and the data are persistently exciting of order L + ι L+\iota L + ι in the sense of Definition 1. For any Y ˉ > 0 \bar{Y}>0 Y ˉ > 0 , there exists L ˉ Y ˉ > 0 \bar{L}_{\bar{Y}}>0 L ˉ Y ˉ > 0 such that, for all L ≥ L ˉ Y ˉ L\ge\bar{L}_{\bar{Y}} L ≥ L ˉ Y ˉ and all initial conditions ( ξ 0 , X 0 ) ∈ X × X a : = { ξ a ∈ X a ∣ J L a ( ξ a ) ≤ Y ˉ } (\xi_0,X_0)\in\mathcal{X}\times\mathcal{X}^a:=\{\xi^a\in\mathcal{X}^a\mid J_L^a(\xi^a)\le\bar{Y}\} ( ξ 0 , X 0 ) ∈ X × X a := { ξ a ∈ X a ∣ J L a ( ξ a ) ≤ Y ˉ } , the DDPC problem (25) is recursively feasible and the augmented regulator manifold A a : = { ξ a ∣ ξ a = Π a ω } \mathcal{A}^a:=\{\xi^a\mid\xi^a=\Pi^a\omega\} A a := { ξ a ∣ ξ a = Π a ω } is exponentially stable for the closed loop.
Proof: Lemma 2 identifies (25) with the DDPC problem (21) posed on the augmented plant, and Proposition 4 verifies the required structural assumptions for this augmented plant. Thus, the closed-loop properties follow from Theorem 2.
Hereafter, we refer to the implementable incremental-input formulation (25) as the nominal DDPC.
Remark 6 (Increment-based predictors): Under Assumption 5, an alternative predictor can be built from the period increments Δ T u t : = u t − u t − T \Delta_Tu_t:=u_t-u_{t-T} Δ T u t := u t − u t − T and Δ T y t : = y t − y t − T \Delta_Ty_t:=y_t-y_{t-T} Δ T y t := y t − y t − T . Since w t − w t − T = 0 w_t-w_{t-T}=0 w t − w t − T = 0 , these increments obey the disturbance-free dynamics Δ T x t + 1 = A Δ T x t + B Δ T u t \Delta_Tx_{t+1}=A\Delta_Tx_t+B\Delta_Tu_t Δ T x t + 1 = A Δ T x t + B Δ T u t , Δ T y t = C Δ T x t \Delta_Ty_t=C\Delta_Tx_t Δ T y t = C Δ T x t , so that the standard fundamental lemma applies to them, and the output prediction can be recovered from y ˉ k ∣ t = y ˉ k − T ∣ t + Δ T y ˉ k ∣ t \bar{y}_{k|t}=\bar{y}_{k-T|t}+\Delta_T\bar{y}_{k|t} y ˉ k ∣ t = y ˉ k − T ∣ t + Δ T y ˉ k ∣ t , initialized with the measured outputs of the last period; see also [24 ] for a predictive control scheme tailored to repetitive tasks. With noise-free data, both this increment-based predictor and (36) parametrize the exact trajectories of (1) ; hence, combined with the cost (35) , they lead to the same closed loop. The two predictors use different information. The increment-based predictor requires only the period T T T and no phase information, but it needs ι + T \iota+T ι + T past samples online and loses T T T samples of the offline data. The predictor (36) needs only ι \iota ι past samples, at the price of the phase information discussed in Remark 2. The robust analysis in Section IV is developed for (36) .
Robust DDPC for Output Regulation with Periodic Exosignals
Throughout this section, Assumptions 5 and 6 are in force, so the exosignal and the associated steady-state input are T T T -periodic. We consider bounded additive noise in both the offline output data and the online output measurements. These two noise sources perturb the Hankel predictor and the initial-condition constraint, respectively. Consequently, (36) no longer parametrizes exact trajectories, and the nominal DDPC problem may become infeasible or yield biased predictions. We therefore introduce slack and coefficient regularization into (25) and establish recursive feasibility and practical stability for the resulting closed loop.
Robust DDPC
Assume that both the output data used in the Hankel matrices and the online output measurements are corrupted by bounded additive noise: y ~ t = y t + ε t \tilde{y}_t = y_t + \varepsilon_t y ~ t = y t + ε t , with ∥ ε t ∥ ∞ ≤ ε ˉ \lVert\varepsilon_t\rVert_\infty\le\bar{\varepsilon} ∥ ε t ∥ ∞ ≤ ε ˉ for all t ≥ 0 t \ge0 t ≥ 0 . Given the noisy data { u t , y ~ t } t = 0 N − 1 \{u_t,\tilde{y}_t\}_{t=0}^{N-1} { u t , y ~ t } t = 0 N − 1 and the recent measurements ( u [ t − ι , t − 1 ] , y ~ [ t − ι , t − 1 ] ) (u_{[t-\iota,t-1]},\tilde{y}_{[t-\iota,t-1]}) ( u [ t − ι , t − 1 ] , y ~ [ t − ι , t − 1 ] ) , define the perturbed auxiliary state
ξ ~ t : = [ y ~ [ t − ι , t − 1 ] u [ t − ι , t − 1 ] ] = ξ t + [ ε [ t − ι , t − 1 ] 0 ] (46) \tilde{\xi}_t := \begin{bmatrix}\tilde{y}_{[t-\iota,t-1]}\\u_{[t-\iota,t-1]}\end{bmatrix} = \xi_t + \begin{bmatrix}\varepsilon_{[t-\iota,t-1]}\\0\end{bmatrix} \tag*{(46)} ξ ~ t := [ y ~ [ t − ι , t − 1 ] u [ t − ι , t − 1 ] ] = ξ t + [ ε [ t − ι , t − 1 ] 0 ] (46)
The robust DDPC problem is obtained by relaxing the output part of the Hankel equality and regularizing the coefficient vector
min y ˉ , u ˉ , g , h J L a ( ξ ~ t , ζ t ) = ∑ k = 0 L − 1 ∥ y ˉ k ∣ t ∥ Q 2 + ∥ Δ u ˉ k ∣ t ∥ R 2 + λ g ε ˉ ∥ g t − g t c ∥ 2 + λ h ε ˉ ∥ h t ∥ 2 (47) \min_{\bar{y},\bar{u},g,h} \mathcal{J}_L^a(\tilde{\xi}_t,\zeta_t) = \sum_{k=0}^{L-1}\lVert\bar{y}_{k|t}\rVert_Q^2+\lVert\Delta\bar{u}_{k|t}\rVert_R^2+\lambda_g\bar{\varepsilon}\lVert g_t-g_t^c\rVert^2+\frac{\lambda_h}{\bar{\varepsilon}}\lVert h_t\rVert^2 \tag*{(47)} y ˉ , u ˉ , g , h min J L a ( ξ ~ t , ζ t ) = k = 0 ∑ L − 1 ∥ y ˉ k ∣ t ∥ Q 2 + ∥ Δ u ˉ k ∣ t ∥ R 2 + λ g ε ˉ ∥ g t − g t c ∥ 2 + ε ˉ λ h ∥ h t ∥ 2 (47)
s.t. [ H L + ι ( y ~ ) H L + ι ( u ) H 1 ( M ) ] g t = [ y ˉ [ − ι , L − 1 ] ∣ t + h t u ˉ [ − ι , L − 1 ] ∣ t M ˉ t − ι ] (48) \text{s.t.}\quad
\begin{bmatrix}
H_{L+\iota}(\tilde{y})\\
H_{L+\iota}(u)\\
H_1(M)
\end{bmatrix}g_t
=
\begin{bmatrix}
\bar{y}_{[-\iota,L-1]|t}+h_t\\
\bar{u}_{[-\iota,L-1]|t}\\
\bar{M}_{t-\iota}
\end{bmatrix} \tag*{(48)} s.t. H L + ι ( y ~ ) H L + ι ( u ) H 1 ( M ) g t = y ˉ [ − ι , L − 1 ] ∣ t + h t u ˉ [ − ι , L − 1 ] ∣ t M ˉ t − ι (48)
[ y ˉ [ − ι , − 1 ] ( t ) u ˉ [ − ι , − 1 ] ( t ) ] = [ y ~ [ t − ι , t − 1 ] u [ t − ι , t − 1 ] ] (49) \begin{bmatrix}
\bar{y}_{[-\iota,-1]}(t)\\
\bar{u}_{[-\iota,-1]}(t)
\end{bmatrix}
=
\begin{bmatrix}
\tilde{y}_{[t-\iota,t-1]}\\
u_{[t-\iota,t-1]} \tag*{(49)}
\end{bmatrix} [ y ˉ [ − ι , − 1 ] ( t ) u ˉ [ − ι , − 1 ] ( t ) ] = [ y ~ [ t − ι , t − 1 ] u [ t − ι , t − 1 ] ] (49)
u ˉ k ∣ t = u ˉ k − T ∣ t + Δ u ˉ k ∣ t , k = 0 , ⋯ , L − 1 (50) \bar{u}_{k|t}=\bar{u}_{k-T|t}+\Delta\bar{u}_{k|t},\quad k=0,\cdots,L-1 \tag*{(50)} u ˉ k ∣ t = u ˉ k − T ∣ t + Δ u ˉ k ∣ t , k = 0 , ⋯ , L − 1 (50)
u ˉ k − T ∣ t = { u t + k − T , k < T u ˉ k − T ∣ t , k ≥ T (51) \bar{u}_{k-T|t}=
\begin{cases}
u_{t+k-T},& k<T\\
\bar{u}_{k-T|t},& k\ge T
\end{cases} \tag*{(51)} u ˉ k − T ∣ t = { u t + k − T , u ˉ k − T ∣ t , k < T k ≥ T (51)
u ˉ k ∣ t ∈ U , k = 0 , ⋯ , L − 1. (52) \bar{u}_{k|t}\in\mathbb{U},\quad k=0,\cdots,L-1. \tag*{(52)} u ˉ k ∣ t ∈ U , k = 0 , ⋯ , L − 1. (52)
Here, g t c g_t^c g t c is a regularization center, and the regularizer λ g ε ˉ ∥ g t − g t c ∥ 2 \lambda_g\bar{\varepsilon}\lVert g_t-g_t^c\rVert^2 λ g ε ˉ ∥ g t − g t c ∥ 2 limits amplification of noisy Hankel columns. Since the regulated steady state is a nonzero periodic trajectory represented by a nonzero coefficient vector, the center cannot be chosen as zero. We therefore compute the center offline from the data. Let P T P_T P T be the matrix such that P T u ˉ [ − ι , L − 1 ] ∣ t P_T\bar{u}_{[-\iota,L-1]|t} P T u ˉ [ − ι , L − 1 ] ∣ t stacks the increments u ˉ k ∣ t − u ˉ k − T ∣ t \bar{u}_{k|t}-\bar{u}_{k-T|t} u ˉ k ∣ t − u ˉ k − T ∣ t , k = T − ι , … , L − 1 k=T-\iota,\ldots,L-1 k = T − ι , … , L − 1 (the constraint involving P T P_T P T is void if L + ι ≤ T L+\iota\le T L + ι ≤ T ). For any exosystem basis vector M ˉ ∈ R n M \bar{M}\in\mathbb{R}^{n_M} M ˉ ∈ R n M and a fixed weight μ > 0 \mu>0 μ > 0 , define
g c ( M ˉ ) : = arg min g ∥ H L + ι ( y ~ ) g ∥ 2 + μ ∥ g ∥ 2 g^c(\bar{M}):=\underset{g}{\arg\min}\ \lVert H_{L+\iota}(\tilde{y})g\rVert^2+\mu\lVert g\rVert^2 g c ( M ˉ ) := g arg min ∥ H L + ι ( y ~ ) g ∥ 2 + μ ∥ g ∥ 2
s.t. Ψ g = [ M ˉ 0 ] , Ψ : = [ H 1 ( M ) P T H L + ι ( u ) ] , (53) \text{s.t.}\quad\Psi g=\begin{bmatrix}\bar{M}\\0\end{bmatrix},\quad\Psi:=\begin{bmatrix}H_1(M)\\P_TH_{L+\iota}(u)\end{bmatrix}, \tag*{(53)} s.t. Ψ g = [ M ˉ 0 ] , Ψ := [ H 1 ( M ) P T H L + ι ( u ) ] , (53)
and set g t c : = g c ( M ˉ t − ι ) g_t^c:=g^c(\bar{M}_{t-\iota}) g t c := g c ( M ˉ t − ι ) . Problem (53) seeks a coefficient vector whose predicted output is approximately zero, whose predicted input is T T T -periodic, and whose exosystem basis vector is M ˉ \bar{M} M ˉ . It is thus a data-based approximation of the regulated steady state that requires neither Γ ξ \Gamma_\xi Γ ξ nor the exosignal w w w . Since the data are persistently exciting of order L + ι L+\iota L + ι , the matrix
[ H 1 ( M ) H L + ι ( u ) ] \begin{bmatrix}H_1(M)\\H_{L+\iota}(u)\end{bmatrix} [ H 1 ( M ) H L + ι ( u ) ]
has full row rank, and so does Ψ \Psi Ψ . Hence (53) is feasible for every M ˉ \bar{M} M ˉ , and its unique minimizer is linear in M ˉ \bar{M} M ˉ :
g t c = G c M ˉ t − ι , G c : = Φ − 1 Ψ T ( Ψ Φ − 1 Ψ T ) − 1 [ I n M 0 ] , (54) g_t^c=G_c\bar{M}_{t-\iota},\quad G_c:=\Phi^{-1}\Psi^\mathsf{T}\left(\Psi\Phi^{-1}\Psi^\mathsf{T}\right)^{-1}\begin{bmatrix}I_{n_M}\\0\end{bmatrix}, \tag*{(54)} g t c = G c M ˉ t − ι , G c := Φ − 1 Ψ T ( Ψ Φ − 1 Ψ T ) − 1 [ I n M 0 ] , (54)
where Φ : = H L + ι ( y ~ ) T H L + ι ( y ~ ) + μ I ≻ 0 \Phi:=H_{L+\iota}(\tilde{y})^\mathsf{T}H_{L+\iota}(\tilde{y})+\mu I\succ0 Φ := H L + ι ( y ~ ) T H L + ι ( y ~ ) + μ I ≻ 0 . Thus, G c G_c G c is computed once offline, and only a matrix-vector product is needed online. Moreover, under Assumption 5, the basis M ( ⋅ ) \mathcal{M}(\cdot) M ( ⋅ ) is T T T -periodic, so g t c g_t^c g t c is exactly T T T -periodic and does not depend on the closed-loop optimizers. This is in contrast with the recursive choice g t c = g t − T ∗ g_t^c=g_{t-T}^* g t c = g t − T ∗ : the a priori bound on the optimizer only gives ∥ g t ∗ − g t c ∥ 2 ≤ Y ˉ / ( λ g ε ˉ ) \lVert g_t^*-g_t^c\rVert^2\le\bar{Y}/(\lambda_g\bar{\varepsilon}) ∥ g t ∗ − g t c ∥ 2 ≤ Y ˉ / ( λ g ε ˉ ) , cf. Appendix C, which deteriorates as ε ˉ → 0 \bar{\varepsilon}\to0 ε ˉ → 0 and may accumulate from one period to the next; hence, the recursive choice does not guarantee a uniformly bounded mismatch with the steady-state coefficient.
Compared with (36) , constraint (48) contains slack variable h t h_t h t that absorbs the mismatch caused by noisy output data. The penalty ( λ h / ε ˉ ) ∥ h t ∥ 2 (\lambda_h/\bar{\varepsilon})\lVert h_t\rVert^2 ( λ h / ε ˉ ) ∥ h t ∥ 2 discourages slack beyond what is justified by the noise level. With the chosen scaling, the regularizers vanish as ε ˉ → 0 \bar{\varepsilon}\to0 ε ˉ → 0 , so (30) recovers the nominal formulation. For fixed ε ˉ > 0 \bar{\varepsilon}>0 ε ˉ > 0 , the problem is a convex quadratic program with linear constraints.
When (30) is feasible at time t t t , let J L a , ∗ ( ξ ~ t , ζ t ) \mathcal{J}_L^{a,*}(\tilde{\xi}_t,\zeta_t) J L a , ∗ ( ξ ~ t , ζ t ) denote its optimal value and let ( y ˉ t ∗ , u ˉ t ∗ , g t ∗ , h t ∗ ) (\bar{y}_t^*,\bar{u}_t^*,g_t^*,h_t^*) ( y ˉ t ∗ , u ˉ t ∗ , g t ∗ , h t ∗ ) denote an optimizer. Analogously to the nominal scheme in Algorithm 1, the corresponding robust receding-horizon implementation is summarized in Algorithm 2. Its closed-loop properties are established below.
Algorithm 2 Robust DDPC for Output Regulation.
Given: Prediction horizon L ≥ ι + n L\ge\iota+n L ≥ ι + n , I/O data { u t , y ~ t } t = 0 N − 1 \{u_t,\tilde{y}_t\}_{t=0}^{N-1} { u t , y ~ t } t = 0 N − 1 that are persistently exciting of order L + ι L+\iota L + ι in the sense of Definition 1, the constraint set U \mathbb{U} U , cost matrices Q Q Q , R R R , noise bound ε ˉ > 0 \bar{\varepsilon}>0 ε ˉ > 0 , regularization weights λ g > 0 \lambda_g>0 λ g > 0 , λ h > 0 \lambda_h>0 λ h > 0 , and exosignal period T T T , and the center map G c G_c G c computed offline from (54) .
If t ≤ ι − 1 t\le\iota-1 t ≤ ι − 1 , do
Run (1) in open loop with randomly chosen input u t u_t u t , and store measurement u t , y ~ t u_t,\tilde{y}_t u t , y ~ t .
Else if t ≥ ι t\ge\iota t ≥ ι , do
Take the past ι \iota ι I/O measurements ( u [ t − ι , t − 1 ] , y ~ [ t − ι , t − 1 ] ) (u_{[t-\iota,t-1]},\tilde{y}_{[t-\iota,t-1]}) ( u [ t − ι , t − 1 ] , y ~ [ t − ι , t − 1 ] ) , set M ˉ t − ι = M ( t − ι ) \bar{M}_{t-\iota}=\mathcal{M}(t-\iota) M ˉ t − ι = M ( t − ι ) (cf. Remark 2), set g t c = G c M ˉ t − ι g_t^c=G_c\bar{M}_{t-\iota} g t c = G c M ˉ t − ι , and solve (30).
Apply the input u t = u ˉ 0 ∣ t ∗ u_t=\bar{u}_{0|t}^* u t = u ˉ 0∣ t ∗ .
Set t : = t + 1 t:=t+1 t := t + 1 and go back to 2.
For the proof, it is useful to compare (30) with a feedforward version of the same problem, paralleling the nominal analysis. Define g t s g_t^s g t s as a coefficient vector corresponding to the regulated steady-state trajectory, satisfying
\begin{bmatrix}
H_{L+\iota}(u)\\
H_{L+\iota}(u)\\
H_1(\mathcal{M})
\end{bmatrix}g_t^s
=
\begin{bmatrix}
0_{p(L+\iota)\times1}\\
\begin{bmatrix}
S^{-\iota}\\
\vdots\\
I
\end{bmatrix}\bar{w}_{[-\iota,L-1]|t}\\
\Gamma_\xi
\end{bmatrix}_{\bar{M}_{t-\iota}}, \tag*{(55)}\end{bmatrix}
w ˉ k + 1 ∣ t = S w ˉ k ∣ t , k = − ι , … , L − 2 , (56) \bar{w}_{k+1|t}=S\bar{w}_{k|t},\quad k=-\iota,\ldots,L-2, \tag*{(56)} w ˉ k + 1∣ t = S w ˉ k ∣ t , k = − ι , … , L − 2 , (56)
w ˉ − ι ∣ t = w t − ι . (57) \bar{w}_{-\iota|t}=w_{t-\iota}. \tag*{(57)} w ˉ − ι ∣ t = w t − ι . (57)
whose minimum-norm solution is denoted by g t s ∗ g_t^{s*} g t s ∗ . The sequence g t s ∗ g_t^{s*} g t s ∗ is bounded for all t ∈ N t \in\mathbb{N} t ∈ N . Under measurement noise, however, only the corrupted output y ~ \tilde{y} y ~ is available. We therefore use a noisy approximation of g t s ∗ g_t^{s*} g t s ∗ obtained from
J t s : = min g t s , h t s λ g s ∥ g t s ∥ 2 + λ h s ∥ h t s ∥ 2 (58) J_t^{s}:=\min_{g_t^s,h_t^s}\lambda_{g^s}\|g_t^s\|^2+\lambda_{h^s}\|h_t^s\|^2 \tag*{(58)} J t s := g t s , h t s min λ g s ∥ g t s ∥ 2 + λ h s ∥ h t s ∥ 2 (58)
s.t.
[ H L + ι ( y ~ ) H L + ι ( u ) H 1 ( M ) ] g t s = [ h [ − ι , L − 1 ] ∣ t s S − ι ⋮ w ˉ [ − ι , L − 1 ] ∣ t M ˉ t − ι ] (59) \begin{bmatrix}
H_{L+\iota}(\tilde{y})\\
H_{L+\iota}(u)\\
H_1(M)
\end{bmatrix}g_t^s=
\begin{bmatrix}
h^s_{[-\iota,L-1]|t}\\
S^{-\iota}\\
\vdots\\
\bar{w}_{[-\iota,L-1]|t}\\
\bar{M}_{t-\iota} \tag*{(59)}
\end{bmatrix} H L + ι ( y ~ ) H L + ι ( u ) H 1 ( M ) g t s = h [ − ι , L − 1 ] ∣ t s S − ι ⋮ w ˉ [ − ι , L − 1 ] ∣ t M ˉ t − ι (59)
w ˉ k + 1 ∣ t = S ˉ w ˉ k ∣ t , k = − ι , … , L − 2 (60) \bar{w}_{k+1|t}=\bar{S}\bar{w}_{k|t},\quad k=-\iota,\ldots,L-2 \tag*{(60)} w ˉ k + 1∣ t = S ˉ w ˉ k ∣ t , k = − ι , … , L − 2 (60)
w ˉ − ι ∣ t = w t − ι (61) \bar{w}_{-\iota|t}=w_{t-\iota} \tag*{(61)} w ˉ − ι ∣ t = w t − ι (61)
whose optimizer is denoted by ( g ~ t s , h ~ t s ) (\tilde{g}_t^s,\tilde{h}_t^s) ( g ~ t s , h ~ t s ) and whose optimal cost is denoted by J t s ∗ J_t^{s*} J t s ∗ . Then, assume that there exists J ˉ s > 0 \bar{J}^s>0 J ˉ s > 0 such that J t s ∗ ≤ J ˉ s J_t^{s*}\leq\bar{J}^s J t s ∗ ≤ J ˉ s for all t ∈ N t\in\mathbb{N} t ∈ N . This can be proved by using g ˉ t s : = g t s ∗ \bar{g}_t^s:=g_t^{s*} g ˉ t s := g t s ∗ , and h ˉ t s : = H L + ι ( ε ) g t s ∗ \bar{h}_t^s:=H_{L+\iota}(\varepsilon)g_t^{s*} h ˉ t s := H L + ι ( ε ) g t s ∗ as candidate solutions to (34). In addition, as in [9 ] , assume that there exists constant c g s > 0 c_{g^s}>0 c g s > 0 such that ∥ g t s ∗ − g ~ t s ∥ 2 ≤ c g s ε ˉ \|g_t^{s*}-\tilde{g}_t^s\|^2\leq c_{g^s}\bar{\varepsilon} ∥ g t s ∗ − g ~ t s ∥ 2 ≤ c g s ε ˉ ; this bound is used in the stability analysis below.
The analytical reference g ~ t s \tilde{g}_t^s g ~ t s is used only in the proof to bound the coefficient regularization term; it is neither computed online nor used in Algorithm 2. The next result shows that the offline center (53) stays uniformly close to this reference, so that no additional assumption on the regularization center is needed.
Proposition 5 (Bounded center mismatch): Let Assumption 5 hold, let the I/O data be persistently exciting of order L + ι L+\iota L + ι in the sense of Definition 1, and let ε ˉ ∈ ( 0 , ε max ] \bar{\varepsilon}\in(0,\varepsilon_{\max}] ε ˉ ∈ ( 0 , ε m a x ] for some ε max > 0 \varepsilon_{\max}>0 ε m a x > 0 . Then the center g t c = G c M ˉ t − ι g_t^c=G_c\bar{M}_{t-\iota} g t c = G c M ˉ t − ι in (54) satisfies
∥ g t c − g ~ t s ∥ 2 ≤ c c , ∀ t ≥ ι , (62) \|g_t^c-\tilde{g}_t^s\|^2\leq c_c,\qquad\forall t\geq\iota, \tag*{(62)} ∥ g t c − g ~ t s ∥ 2 ≤ c c , ∀ t ≥ ι , (62)
where
c c : = 2 ( 1 + c H 2 μ ) c M 2 + 2 J ˉ s λ g s , c M : = max 0 ≤ k ≤ T − 1 ∥ M ( k ) ∥ , c_c:=2\left(1+\frac{c_H^2}{\mu}\right)c_M^2+\frac{2\bar{J}^s}{\lambda_{g^s}},
\qquad
c_M:=\max_{0\leq k\leq T-1}\|M(k)\|, c c := 2 ( 1 + μ c H 2 ) c M 2 + λ g s 2 J ˉ s , c M := 0 ≤ k ≤ T − 1 max ∥ M ( k ) ∥ ,
and
c H : = ∥ H L + ι ( y ) ∥ + p ( L + ι ) ( N − L − ι + 1 ) ε max . c_H:=\|H_{L+\iota}(y)\|+\sqrt{p(L+\iota)(N-L-\iota+1)}\varepsilon_{\max}. c H := ∥ H L + ι ( y ) ∥ + p ( L + ι ) ( N − L − ι + 1 ) ε m a x .
In particular, c c c_c c c is independent of t t t and of ε ˉ ∈ ( 0 , ε max ] \bar{\varepsilon}\in(0,\varepsilon_{\max}] ε ˉ ∈ ( 0 , ε m a x ] .
Proof. Since every entry of H L + ι ( ε ) H_{L+\iota}(\varepsilon) H L + ι ( ε ) is bounded by ε ˉ \bar{\varepsilon} ε ˉ in magnitude, ∥ H L + ι ( y ~ ) ∥ ≤ ∥ H L + ι ( y ) ∥ + ∥ H L + ι ( ε ) ∥ F ≤ c H \|H_{L+\iota}(\tilde{y})\|\leq\|H_{L+\iota}(y)\|+\|H_{L+\iota}(\varepsilon)\|_F\leq c_H ∥ H L + ι ( y ~ ) ∥ ≤ ∥ H L + ι ( y ) ∥ + ∥ H L + ι ( ε ) ∥ F ≤ c H . Under Assumption 5, M ( ⋅ ) M(\cdot) M ( ⋅ ) is T T T -periodic, so that ∥ M ˉ t − ι ∥ ≤ c M \|\bar{M}_{t-\iota}\|\leq c_M ∥ M ˉ t − ι ∥ ≤ c M . The vector g 0 : = Ψ † [ M ˉ t − ι ] g_0:=\Psi^\dagger\begin{bmatrix}\bar{M}_{t-\iota}\\ \end{bmatrix} g 0 := Ψ † [ M ˉ t − ι ] is feasible for (53) because Ψ \Psi Ψ has full row rank. Optimality of g t c g_t^c g t c gives
μ ∥ g t c ∥ 2 ≤ ∥ H L + ι ( y ~ ) g 0 ∥ 2 + μ ∥ g 0 ∥ 2 ≤ ( c H 2 + μ ) ∥ Ψ † ∥ 2 c M 2 . \mu\|g_t^c\|^2\leq\|H_{L+\iota}(\tilde{y})g_0\|^2+\mu\|g_0\|^2\leq(c_H^2+\mu)\|\Psi^\dagger\|^2c_M^2. μ ∥ g t c ∥ 2 ≤ ∥ H L + ι ( y ~ ) g 0 ∥ 2 + μ ∥ g 0 ∥ 2 ≤ ( c H 2 + μ ) ∥ Ψ † ∥ 2 c M 2 .
Moreover, J t s ∗ ≤ J ˉ s J_t^{s*}\leq\bar{J}^s J t s ∗ ≤ J ˉ s yields ∥ g ~ t s ∥ 2 ≤ J ˉ s / λ g s \|\tilde{g}_t^s\|^2\leq\bar{J}^s/\lambda_{g^s} ∥ g ~ t s ∥ 2 ≤ J ˉ s / λ g s , where, by the candidate solution given below (34), J ˉ s \bar{J}^s J ˉ s can be chosen uniformly for ε ˉ ∈ ( 0 , ε max ] \bar{\varepsilon}\in(0,\varepsilon_{\max}] ε ˉ ∈ ( 0 , ε m a x ] . The claim follows from
∥ a − b ∥ 2 ≤ 2 ∥ a ∥ 2 + 2 ∥ b ∥ 2 . \|a-b\|^2\leq2\|a\|^2+2\|b\|^2. ∥ a − b ∥ 2 ≤ 2∥ a ∥ 2 + 2∥ b ∥ 2 .
Proposition 5 guarantees the mismatch bound by construction; it is not a condition to be verified along the closed loop. The constant c c c_c c c is a worst-case bound. It enters the robust analysis only through regularization terms that are multiplied by λ g s ε ˉ \lambda_{g^s}\bar{\varepsilon} λ g s ε ˉ , cf. Appendix C, and therefore does not affect the vanishing of the ultimate bound as ε ˉ → 0 \bar{\varepsilon}\to0 ε ˉ → 0 ; a center closer to g ~ t s \tilde{g}_t^s g ~ t s only reduces the corresponding constants.
With these preliminaries in place, recalling ( Π a , Γ a ) (\Pi^a,\Gamma^a) ( Π a , Γ a ) denote the regulator pair of the augmented system in Proposition 4, we now introduce the following feedforward-form representation of (30) on the augmented plant (28). This auxiliary formulation retains the noisy Hankel model, the slack variable, and the implementable regularization center g t c g_t^c g t c , while expressing the incremental-input penalty through the augmented input u a u^a u a :
min y ˉ , u ˉ , g ˉ , h , ξ ˉ , ζ ˉ , w ˉ J t f ( ξ ˉ t , ζ ˉ t , w ˉ t ) = ∑ k = 0 L − 1 ( ∥ y ˉ k ∣ t ∥ Q 2 + ∥ u ˉ k ∣ t a − Γ a w ˉ k ∣ t ∥ R 2 ) + λ g s ε ˉ ∥ g t − g t c ∥ 2 + λ h 2 ∥ h t ∥ 2 (63) \min_{\bar{y},\bar{u},\bar{g},h,\bar{\xi},\bar{\zeta},\bar{w}}
J_t^f(\bar{\xi}_t,\bar{\zeta}_t,\bar{w}_t)
=\sum_{k=0}^{L-1}\left(\|\bar{y}_{k|t}\|_Q^2+\|\bar{u}_{k|t}^a-\Gamma^a\bar{w}_{k|t}\|_R^2\right)
+\lambda_{g^s}\bar{\varepsilon}\|g_t-g_t^c\|^2+\frac{\lambda_h}{2}\|h_t\|^2 \tag*{(63)} y ˉ , u ˉ , g ˉ , h , ξ ˉ , ζ ˉ , w ˉ min J t f ( ξ ˉ t , ζ ˉ t , w ˉ t ) = k = 0 ∑ L − 1 ( ∥ y ˉ k ∣ t ∥ Q 2 + ∥ u ˉ k ∣ t a − Γ a w ˉ k ∣ t ∥ R 2 ) + λ g s ε ˉ ∥ g t − g t c ∥ 2 + 2 λ h ∥ h t ∥ 2 (63)
s.t.
\begin{bmatrix}
H_{L+\iota}(\tilde{y})\\
H_{L+\iota}(u)\\
H_1(M)
\end{bmatrix}g_t=
\begin{bmatrix}
\bar{y}_{[-\iota,L-1]|t}+h_t\\
\bar{u}_{[-\iota,L-1]|t}\\
\bar{M}_{t-\iota}
\end{bmatrix}, \tag*{(64)}\end{bmatrix}
\begin{bmatrix}
\bar{y}_{[-\iota,L-1]|t}\\
\bar{u}_{[-\iota,L-1]|t}
\end{bmatrix}
=
\begin{bmatrix}
\tilde{y}_{[t-\iota,t-1]}\\
u_{[t-\iota,t-1]}
\end{bmatrix}, \tag*{(65)}\end{bmatrix}
ζ ˉ 0 ∣ t = ζ t , (66) \bar{\zeta}_{0|t}=\zeta_t, \tag*{(66)} ζ ˉ 0∣ t = ζ t , (66)
ζ ˉ k + 1 ∣ t = E 0 ζ ˉ k ∣ t + E 1 u ˉ k ∣ t a , k = 0 , … , L − 1 , (67) \bar{\zeta}_{k+1|t}=E_0\bar{\zeta}_{k|t}+E_1\bar{u}_{k|t}^a,\quad k=0,\ldots,L-1, \tag*{(67)} ζ ˉ k + 1∣ t = E 0 ζ ˉ k ∣ t + E 1 u ˉ k ∣ t a , k = 0 , … , L − 1 , (67)
u ˉ k ∣ t = E 2 T ζ ˉ k ∣ t + u ˉ k ∣ t a , k = 0 , … , L − 1 , (68) \bar{u}_{k|t}=E_2^\mathsf{T}\bar{\zeta}_{k|t}+\bar{u}_{k|t}^a,\quad k=0,\ldots,L-1, \tag*{(68)} u ˉ k ∣ t = E 2 T ζ ˉ k ∣ t + u ˉ k ∣ t a , k = 0 , … , L − 1 , (68)
w ˉ 0 ∣ t = [ w t − ι ⋮ w t ] , w ˉ k + 1 ∣ t = S ˉ w ˉ k ∣ t , (69) \bar{w}_{0|t}=
\begin{bmatrix}
w_{t-\iota}\\
\vdots\\
w_t
\end{bmatrix},
\qquad
\bar{w}_{k+1|t}=\bar{S}\bar{w}_{k|t}, \tag*{(69)} w ˉ 0∣ t = w t − ι ⋮ w t , w ˉ k + 1∣ t = S ˉ w ˉ k ∣ t , (69)
u ˉ k ∣ t a ∈ U , k = 0 , … , L − 1. (70) \bar{u}_{k|t}^a\in\mathbb{U},\quad k=0,\ldots,L-1. \tag*{(70)} u ˉ k ∣ t a ∈ U , k = 0 , … , L − 1. (70)
Lemma 3 (Robust augmented reformulation): Let Assumptions [1 , 2 , 3 , 4 , 5 ] hold. For fixed t t t , ζ t \zeta_t ζ t , and ω t \omega_t ω t , Problems (30) and (36) are equivalent.
Proof. The equivalence follows from the same input-memory transformation as in [25 ] , Lemma 1 (cf. Lemma 2). The noisy Hankel constraint, slack variable h h h , and regularization terms are identical in (30) and (36), while u ˉ k ∣ t a = Δ u ˉ k ∣ t \bar{u}_{k|t}^a=\Delta\bar{u}_{k|t} u ˉ k ∣ t a = Δ u ˉ k ∣ t and Γ a = 0 \Gamma^a=0 Γ a = 0 . Hence, the transformation preserves feasibility and the objective value. ■
Lemma 3 allows us to establish recursive feasibility and practical stability for Problem (36). The same properties then hold for the implementable formulation (30). Proposition 5 is used only when bounding the coefficient regularization in the robust analysis below.
Main Results
For simplicity, let ξ ~ t a : = [ ξ t ζ t ] \tilde{\xi}_t^a:=\begin{bmatrix}\xi_t\\ \zeta_t\end{bmatrix} ξ ~ t a := [ ξ t ζ t ] , ζ ~ t a : = ( ξ ~ t a , ω t ) \tilde{\zeta}_t^a:=(\tilde{\xi}_t^a,\omega_t) ζ ~ t a := ( ξ ~ t a , ω t ) , and define σ a ( ζ ~ t a ) : = ∥ ξ ~ t a − Π a ω t ∥ 2 \sigma^a(\tilde{\zeta}_t^a):=\|\tilde{\xi}_t^a-\Pi^a\omega_t\|^2 σ a ( ζ ~ t a ) := ∥ ξ ~ t a − Π a ω t ∥ 2 . Noting that the output measurement noise affects only the first block of ξ ~ t a \tilde{\xi}_t^a ξ ~ t a , it follows that ∥ ξ ~ t a − ξ t ∥ ≤ p ε ˉ \|\tilde{\xi}_t^a-\xi_t\|\leq\sqrt{p\bar{\varepsilon}} ∥ ξ ~ t a − ξ t ∥ ≤ p ε ˉ . The following property, namely incremental uniform input/output-to-state stability (i-UOISS), cf. [25 ] , is stated first for the nominal augmented system and then extended to the perturbed state by continuity.
Proposition 6: Under Assumptions [1 , 2 , 3 , 4 , 5 , 6 ] , the augmented system (28) is i-UOISS. In particular, there exist a matrix P α , W = P α , W T ≻ 0 P_{\alpha,W}=P_{\alpha,W}^\mathsf{T}\succ0 P α , W = P α , W T ≻ 0 and a constant ϵ o > 0 \epsilon_o>0 ϵ o > 0 such that, with W a ( α t a ) : = ∥ ξ ~ t a − Π a ω t ∥ P α , W 2 W^a(\alpha_t^a):=\|\tilde{\xi}_t^a-\Pi^a\omega_t\|_{P_{\alpha,W}}^2 W a ( α t a ) := ∥ ξ ~ t a − Π a ω t ∥ P α , W 2 , the following inequalities hold along every trajectory of (28):
W a ( α t a ) ≥ ϵ o ∥ ξ ~ t a − Π a ω t ∥ 2 , (71) W^a(\alpha_t^a)\geq\epsilon_o\|\tilde{\xi}_t^a-\Pi^a\omega_t\|^2, \tag*{(71)} W a ( α t a ) ≥ ϵ o ∥ ξ ~ t a − Π a ω t ∥ 2 , (71)
W a ( α t + 1 a ) ≤ W a ( α t a ) + ∥ u t a − Γ a ω t ∥ R 2 + ∥ y t ∥ Q 2 − ϵ o ∥ ξ ~ t a − Π a ω t ∥ 2 . (72) W^a(\alpha_{t+1}^a)\leq W^a(\alpha_t^a)+\|u_t^a-\Gamma^a\omega_t\|_R^2+\|y_t\|_Q^2-\epsilon_o\|\tilde{\xi}_t^a-\Pi^a\omega_t\|^2. \tag*{(72)} W a ( α t + 1 a ) ≤ W a ( α t a ) + ∥ u t a − Γ a ω t ∥ R 2 + ∥ y t ∥ Q 2 − ϵ o ∥ ξ ~ t a − Π a ω t ∥ 2 . (72)
Moreover, there exists a K ∞ \mathcal{K}_\infty K ∞ function α W \alpha_W α W such that
W a ( α ~ t + 1 a ) ≤ W a ( α ~ t a ) + ∥ u t a ∥ R 2 + ∥ y t ∥ Q 2 − ϵ o σ a ( α ~ t a ) + α W ( ε ˉ ) . (73) W^a(\tilde{\alpha}_{t+1}^a)\leq W^a(\tilde{\alpha}_t^a)+\|u_t^a\|_R^2+\|y_t\|_Q^2-\epsilon_o\sigma^a(\tilde{\alpha}_t^a)+\alpha_W(\bar{\varepsilon}). \tag*{(73)} W a ( α ~ t + 1 a ) ≤ W a ( α ~ t a ) + ∥ u t a ∥ R 2 + ∥ y t ∥ Q 2 − ϵ o σ a ( α ~ t a ) + α W ( ε ˉ ) . (73)
Proof. By Proposition 4, ( A a , C a ) (A^a,C^a) ( A a , C a ) is detectable; hence, the nominal augmented error system admits a quadratic i-UOSS Lyapunov function, cf. [25 ] , Proposition 8, which yields (37) after scaling. Since Γ a = 0 \Gamma^{a}=0 Γ a = 0 and ∥ ξ ~ t a − ξ t a ∥ ≤ p ε ˉ \|\tilde{\xi}_{t}^{a}-\xi_{t}^{a}\| \leq\sqrt{p}\bar{\varepsilon} ∥ ξ ~ t a − ξ t a ∥ ≤ p ε ˉ , quadratic continuity of W a W^{a} W a and Young’s inequality give (73) for some α W ∈ K ∞ \alpha_{W}\in\mathcal{K}_{\infty} α W ∈ K ∞ .
Denoting the optimal value of Problem (36) by J L f , ∗ ( ξ ~ t , ζ t , ω t ) J_{L}^{f,*}(\tilde{\xi}_{t},\zeta_{t},\omega_{t}) J L f , ∗ ( ξ ~ t , ζ t , ω t ) , consider the following Lyapunov candidate
Y L f ( α ~ t a ) : = J L f , ∗ ( ξ ~ t , ζ t , ω t ) + W a ( α ~ t a ) . (74) Y_{L}^{f}(\tilde{\alpha}_{t}^{a}):=J_{L}^{f,*}(\tilde{\xi}_{t},\zeta_{t},\omega_{t})+W^{a}(\tilde{\alpha}_{t}^{a}). \tag*{(74)} Y L f ( α ~ t a ) := J L f , ∗ ( ξ ~ t , ζ t , ω t ) + W a ( α ~ t a ) . (74)
The next lemma is the main technical estimate for the noisy case. It shows that the Lyapunov candidate (74) retains a one-step decrease up to a perturbation term that is of class K ∞ \mathcal{K}_{\infty} K ∞ in the noise bound.
Lemma 4. Assume that Problem (36) is feasible at time t t t , the I/O data are persistently exciting of order L + ι L+\iota L + ι in the sense of Definition 1, and Y L f ( α ~ t a ) ≤ Y ˉ Y_{L}^{f}(\tilde{\alpha}_{t}^{a})\leq\bar{Y} Y L f ( α ~ t a ) ≤ Y ˉ . Then Problem (36) with horizon L − 1 L-1 L − 1 is feasible at time t + 1 t+1 t + 1 , and
Y L − 1 f ( α ~ t + 1 a ) ≤ Y L f ( α ~ t a ) − ϵ o σ a ( α ~ t a ) + α 3 ( ε ˉ ) (75) Y_{L-1}^{f}(\tilde{\alpha}_{t+1}^{a})\leq Y_{L}^{f}(\tilde{\alpha}_{t}^{a})-\epsilon_{o}\sigma^{a}(\tilde{\alpha}_{t}^{a})+\alpha_{3}(\bar{\varepsilon}) \tag*{(75)} Y L − 1 f ( α ~ t + 1 a ) ≤ Y L f ( α ~ t a ) − ϵ o σ a ( α ~ t a ) + α 3 ( ε ˉ ) (75)
where α 3 ∈ K ∞ \alpha_{3}\in\mathcal{K}_{\infty} α 3 ∈ K ∞ and ϵ o > 0 \epsilon_{o}>0 ϵ o > 0 is as in Proposition 6.
Proof. The proof is deferred to Appendix C. A noise-dependent prediction-error bound and a shifted feasible candidate at time t + 1 t+1 t + 1 are combined with a continuity estimate for the i-UIOSS Lyapunov function to obtain (75) . The perturbation term α 3 ( ε ˉ ) \alpha_{3}(\bar{\varepsilon}) α 3 ( ε ˉ ) vanishes as ε ˉ → 0 \bar{\varepsilon}\to0 ε ˉ → 0 , which is essential for the subsequent practical stability analysis.
The robust closed-loop guarantee can now be stated.
Theorem 3 (Robust stability guarantee): Under Assumptions [1 , 2 , 3 , 4 , 5 , 6 ] , suppose that the I/O data are persistently exciting of order L + ι L+\iota L + ι in the sense of Definition 1, and let the regularization center be given by (54) . If Y L f ( α ~ t a ) ≤ Y ˉ Y_{L}^{f}(\tilde{\alpha}_{t}^{a})\leq\bar{Y} Y L f ( α ~ t a ) ≤ Y ˉ , then there exist Y ~ Y ˉ ≥ 0 \tilde{Y}_{\bar{Y}}\geq0 Y ~ Y ˉ ≥ 0 and ε ^ L > 0 \hat{\varepsilon}_{L}>0 ε ^ L > 0 such that, for all L > L ~ Y ˉ L>\tilde{L}_{\bar{Y}} L > L ~ Y ˉ and all ε ˉ ≤ ε ^ L \bar{\varepsilon}\leq\hat{\varepsilon}_{L} ε ˉ ≤ ε ^ L , the robust DDPC problem (30) is feasible for all t ≥ 0 t\geq0 t ≥ 0 , the closed-loop input satisfies u t ∈ U u_{t}\in\mathbb{U} u t ∈ U , and the Lyapunov candidate Y L f = J L f , ∗ + W a Y_{L}^{f}=J_{L}^{f,*}+W^{a} Y L f = J L f , ∗ + W a satisfies
ϵ o σ a ( α ~ t a ) ≤ Y L f ( α ~ t a ) ≤ γ Y ˉ σ a ( α ~ t a ) + α L ( ε ˉ ) , (76) \epsilon_{o}\sigma^{a}(\tilde{\alpha}_{t}^{a})\leq Y_{L}^{f}(\tilde{\alpha}_{t}^{a})\leq\gamma_{\bar{Y}}\sigma^{a}(\tilde{\alpha}_{t}^{a})+\alpha_{L}(\bar{\varepsilon}), \tag*{(76)} ϵ o σ a ( α ~ t a ) ≤ Y L f ( α ~ t a ) ≤ γ Y ˉ σ a ( α ~ t a ) + α L ( ε ˉ ) , (76)
Y L f ( α ~ t + 1 a ) − Y L f ( α ~ t a ) ≤ − α ~ L σ a ( α ~ t a ) + α 8 ( ε ˉ ) , (77) Y_{L}^{f}(\tilde{\alpha}_{t+1}^{a})-Y_{L}^{f}(\tilde{\alpha}_{t}^{a})\leq-\tilde{\alpha}_{L}\sigma^{a}(\tilde{\alpha}_{t}^{a})+\alpha_{8}(\bar{\varepsilon}), \tag*{(77)} Y L f ( α ~ t + 1 a ) − Y L f ( α ~ t a ) ≤ − α ~ L σ a ( α ~ t a ) + α 8 ( ε ˉ ) , (77)
with γ Y ˉ , α ~ L > 0 \gamma_{\bar{Y}},\tilde{\alpha}_{L}>0 γ Y ˉ , α ~ L > 0 and α L , α 8 ∈ K ∞ \alpha_{L},\alpha_{8}\in\mathcal{K}_{\infty} α L , α 8 ∈ K ∞ . Consequently, the augmented regulator manifold is practically stable. Since ∥ ξ ~ t − Π ξ ω t ∥ 2 ≤ σ a ( α ~ t a ) \|\tilde{\xi}_{t}-\Pi_{\xi}\omega_{t}\|^{2}\leq\sigma^{a}(\tilde{\alpha}_{t}^{a}) ∥ ξ ~ t − Π ξ ω t ∥ 2 ≤ σ a ( α ~ t a ) , the original regulation error is practically stable as well, with an ultimate bound that vanishes as ε ˉ → 0 \bar{\varepsilon}\to0 ε ˉ → 0 .
Proof. See Appendix D.
Remark 7 (Relation to existing robust DDPC): Although the stability argument in Theorem 3 follows the general robust DDPC strategy of [7 ] , the coefficient regularization requires a different treatment because the regulated steady state is a nonzero periodic trajectory. In particular, the implementable center g t c g_{t}^{c} g t c generally differs from the analytical steady-state coefficient g ~ t s \tilde{g}_{t}^{s} g ~ t s . The offline center (53) keeps this mismatch uniformly bounded by construction (cf. Proposition 5), which allows the shifted-candidate argument to retain a vanishing noise-dependent perturbation. Theorem 3 therefore extends the robust DDPC guarantee from equilibrium stabilization to periodic output regulation.
Numerical Validation
We validate the proposed controllers on a constrained autonomous-vehicle lateral-control problem on a curvilinear road, adapted from [27 ] , [28 ] . The example is deliberately chosen as an output regulation task rather than a setpoint-tracking task: a periodic curvature signal acts as an exogenous disturbance, the regulated output is the lateral displacement error, and the steering input is constrained. The controller is designed from input-output data only. The state is x = col ( e c g , e ˙ c g , θ e , θ ˙ e ) x=\operatorname{col}(e_{cg},\dot{e}_{cg},\theta_{e},\dot{\theta}_{e}) x = col ( e c g , e ˙ c g , θ e , θ ˙ e ) , where e c g e_{cg} e c g is the lateral displacement error of the vehicle center of gravity and θ e \theta_{e} θ e is the heading error. The steering angle is denoted by u u u . The continuous-time model used to generate data and closed-loop simulations is
x ˙ ( t ) = A c x ( t ) + B c u ( t ) + E c r ( t ) , \dot{x}(t)=A_{c}x(t)+B_{c}u(t)+E_{c}r(t), x ˙ ( t ) = A c x ( t ) + B c u ( t ) + E c r ( t ) ,
where r ( t ) = v x κ ( t ) r(t)=v_{x}\kappa(t) r ( t ) = v x κ ( t ) is induced by the path curvature κ ( t ) \kappa(t) κ ( t ) and
A c = [ 0 1 0 0 0 a 1 a 2 a 3 0 0 0 1 0 a 4 a 5 a 6 ] , B c = [ 0 b 1 0 b 2 ] , E c = [ 0 a 3 − v x 0 a 6 ] . A_{c}=\begin{bmatrix}0&1&0&0\\0&a_{1}&a_{2}&a_{3}\\0&0&0&1\\0&a_{4}&a_{5}&a_{6}\end{bmatrix},\qquad
B_{c}=\begin{bmatrix}0\\b_{1}\\0\\b_{2}\end{bmatrix},\qquad
E_{c}=\begin{bmatrix}0\\a_{3}-v_{x}\\0\\a_{6}\end{bmatrix}. A c = 0 0 0 0 1 a 1 0 a 4 0 a 2 0 a 5 0 a 3 1 a 6 , B c = 0 b 1 0 b 2 , E c = 0 a 3 − v x 0 a 6 .
The numerical parameters are a 1 = − 27.280 a_{1}=-27.280 a 1 = − 27.280 , a 2 = 272.798 a_{2}=272.798 a 2 = 272.798 , a 3 = a 4 = a 5 = 0 a_{3}=a_{4}=a_{5}=0 a 3 = a 4 = a 5 = 0 , a 6 = − 29.388 a_{6}=-29.388 a 6 = − 29.388 , b 1 = 136.399 b_{1}=136.399 b 1 = 136.399 , b 2 = 126.129 b_{2}=126.129 b 2 = 126.129 , and v x = 10 v_{x}=10 v x = 10 m/s. The model is discretized by forward Euler with sampling time T s = 0.05 T_{s}=0.05 T s = 0.05 s, giving A = I + T s A c A=I+T_{s}A_{c} A = I + T s A c , B = T s B c B=T_{s}B_{c} B = T s B c , and E r = T s E c E_{r}=T_{s}E_{c} E r = T s E c . The regulated output is y t = C x t y_{t}=Cx_{t} y t = C x t with C = [ 1 0 0 0 ] C=[1\ 0\ 0\ 0] C = [ 1 0 0 0 ] , so that the output regulation objective coincides with asymptotic lane-centering.
The road curvature is generated by the seven-dimensional periodic exosystem
w t + 1 = S w t , S = block.diag ( 1 , S 1 , S 2 , S 3 ) , (78) w_{t+1}=Sw_{t},\qquad S=\operatorname{block.diag}(1,S_{1},S_{2},S_{3}), \tag*{(78)} w t + 1 = S w t , S = block.diag ( 1 , S 1 , S 2 , S 3 ) , (78)
where S i = [ cos ( i ω ) − sin ( i ω ) sin ( i ω ) cos ( i ω ) ] S_{i}=\begin{bmatrix}\cos(i\omega)&-\sin(i\omega)\\\sin(i\omega)&\cos(i\omega)\end{bmatrix} S i = [ cos ( iω ) sin ( iω ) − sin ( iω ) cos ( iω ) ] for i = 1 , 2 , 3 i=1,2,3 i = 1 , 2 , 3 , ω = 2 π / T p \omega=2\pi/T_{p} ω = 2 π / T p , and T p = 20 T_{p}=20 T p = 20 . The initial exogenous signal is w 0 = [ 1 1 0 1 0 1 0 ] T w_{0}=[1\ 1\ 0\ 1\ 0\ 1\ 0]^{\mathsf{T}} w 0 = [ 1 1 0 1 0 1 0 ] T . Thus w t w_{t} w t contains a constant component and three harmonic pairs. The curvature and yaw-rate input are κ t = [ 0.090 0 0.120 0.040 − 0.090 − 0.035 0.060 ] w t \kappa_{t}=[0.090\ 0\ 0.120\ 0.040\ -0.090\ -0.035\ 0.060]w_{t} κ t = [ 0.090 0 0.120 0.040 − 0.090 − 0.035 0.060 ] w t and r t = G w w t r_{t}=G_{w}w_{t} r t = G w w t , where G w = v x [ 0.090 0 0.120 0.040 − 0.090 − 0.035 0.060 ] G_{w}=v_{x}[0.090\ 0\ 0.120\ 0.040\ -0.090\ -0.035\ 0.060] G w = v x [ 0.090 0 0.120 0.040 − 0.090 − 0.035 0.060 ] . The resulting discrete-time system is
x t + 1 = A x t + B u t + E w t , y t = C x t , (79) x_{t+1}=Ax_{t}+Bu_{t}+Ew_{t},\qquad y_{t}=Cx_{t}, \tag*{(79)} x t + 1 = A x t + B u t + E w t , y t = C x t , (79)
where E = E r G w E=E_{r}G_{w} E = E r G w . Figure 1 shows the exogenous modes and the induced curvature. For visualization only, the reference path in Figure 2 is reconstructed by integrating the prescribed curvature, and the vehicle position is displayed by offsetting the centerline according to the simulated lateral error e c g e_{cg} e c g . The matrices in (79) are used only to generate data and simulate the true plant; they are not supplied to the DDPC optimizations.
Figure 1. Periodic exogenous signal and induced road curvature. The signal contains one constant mode and three harmonic pairs with period T p = 20 T_p = 20 T p = 20 .
Figure 2. Reconstructed curvilinear road and closed-loop vehicle trajectories. The gray corridor is centered on the curvature-generated reference path; vehicle trajectories are obtained by offsetting the centerline by the simulated lateral error e c g e_{cg} e c g .
The offline experiment consists of one input-output trajectory of length N = 800 N=800 N = 800 . During data collection, the plant is excited by u t = K x t + u ˉ t u_{t}=Kx_{t}+\bar{u}_{t} u t = K x t + u ˉ t , where u ˉ t \bar{u}_{t} u ˉ t is uniformly sampled and the steering input is saturated to [ − π / 5 , π / 5 ] [-\pi/5,\pi/5] [ − π /5 , π /5 ] . The stabilizing matrix K K K is used only to collect informative data and is not used by the online controller. The observability index is ι = 4 \iota=4 ι = 4 , and the prediction horizon is L = 40 L=40 L = 40 . The invariant zeros of ( A , B , C ) (A,B,C) ( A , B , C ) are 0.2653 ± 0.3014 i 0.2653\pm0.3014i 0.2653 ± 0.3014 i , with modulus 0.40 0.40 0.40 ; hence, Assumption 6 holds.
We first test the nominal DDPC with periodic incremental-input regularization using noise-free data. The weights are Q = 10 5 Q=10^{5} Q = 1 0 5 and R = 10 − 3 R=10^{-3} R = 1 0 − 3 , the initial condition is x 0 = [ 0.75 0 0 0 ] T x_{0}=[0.75\ 0\ 0\ 0]^{\mathsf{T}} x 0 = [ 0.75 0 0 0 ] T , and the steering constraint is u t ∈ [ − π / 5 , π / 5 ] u_{t}\in[-\pi/5,\pi/5] u t ∈ [ − π /5 , π /5 ] . Figure 3 and 4 show that the lateral error is regulated while the steering input remains within the imposed bounds. Figure 5 shows the corresponding optimal values, illustrating the decrease of the receding-horizon objective along the closed-loop trajectory. The spatial reconstruction in Figure 2 confirms that regulation of e c g e_{cg} e c g translates into path-following performance in curvilinear coordinates.
Figure 3. Closed-loop state trajectories for nominal DDPC with noise-free data (left column) and robust DDPC with noisy data and noisy online outputs (right column).
Figure 5. Optimal values returned by the nominal and robust DDPC optimization problems along the closed-loop trajectories.
We then repeat the experiment with noisy offline output data and noisy online output measurements. The additive output noise satisfies ∣ ϵ t ∣ ≤ ϵ ˉ |\epsilon_t| \le\bar{\epsilon} ∣ ϵ t ∣ ≤ ϵ ˉ , with ϵ ˉ = 5 × 10 − 3 \bar{\epsilon} = 5 \times10^{-3} ϵ ˉ = 5 × 1 0 − 3 . The robust DDPC formulation is used with Q = 10 5 Q = 10^{5} Q = 1 0 5 , R = 10 − 3 R = 10^{-3} R = 1 0 − 3 , and regularization weights λ g ϵ ˉ = 10 4 \lambda_g\bar{\epsilon} = 10^{4} λ g ϵ ˉ = 1 0 4 and λ h / ϵ ˉ = 10 15 \lambda_h/\bar{\epsilon} = 10^{15} λ h / ϵ ˉ = 1 0 15 . The regularization center g t c g_t^c g t c is calculated offline based on (53) –(54) with μ = 10 − 4 \mu= 10^{-4} μ = 1 0 − 4 . The same initial condition, horizon, and steering constraint as in the nominal test are used, so that the effect of measurement noise is isolated. As shown in Figure 2 –Figure 4 , the robust controller preserves constraint satisfaction and keeps the regulated output in a small neighborhood of zero despite corrupted data and online measurements. This behavior is consistent with Theorem 3: exact convergence is replaced by practical regulation, and the residual error is governed by the measurement-noise level.
Figure 4. Closed-loop steering input and regulated lateral error. The steering constraint is respected in both the noise-free and noisy cases.
Conclusions
This paper has developed a predictive-control approach to data-driven output regulation for unknown linear systems. The central idea is a trajectory parametrization for input-output data generated by exosignals with known dynamics and known phase, obtained by combining Willems et al.’s fundamental lemma with the known exosystem structure. This parametrization enables constrained receding-horizon regulators with exponential convergence to the regulator manifold, without a plant model or the solution of the regulator equations. For periodic exosignals, the steady-state feedforward term is removed from the online optimization by an incremental-input penalty, yielding an implementable data-driven regulator. A robust formulation with slack and coefficient regularization, whose regularization center is computed offline from data, was shown to be recursively feasible and practically stabilizing under bounded noise in both the offline data and the online measurements. The vehicle lateral-control study illustrates the main features of the theory: rejection of periodic curvature disturbances, steering-constraint satisfaction, and practical regulation when both offline and online outputs are noisy.
Future work will address exosignals whose phase is not available, non-periodic exogenous signals, data-efficient excitation conditions, and distributed output regulation settings in which multiple agents share only local input-output measurements.
Proof of Theorem 2
Proof. Theorem 1 ensures that the DDPC prediction model parametrizes the same finite-horizon input-output trajectories as the corresponding model-based formulation. Together with Propositions 2 and 3, this provides the local cost bound and i-IOSS estimate required by the finite-horizon Lyapunov argument. Consequently, Y L ( ξ t , ω t ) Y_L(\xi_t,\omega_t) Y L ( ξ t , ω t ) satisfies
ϵ 0 σ ( ϱ t ) ≤ Y L ( ϱ t ) ≤ γ Y σ ( ϱ t ) , (80) \epsilon_0\sigma(\varrho_t) \le Y_L(\varrho_t) \le\gamma_Y\sigma(\varrho_t), \tag*{(80)} ϵ 0 σ ( ϱ t ) ≤ Y L ( ϱ t ) ≤ γ Y σ ( ϱ t ) , (80)
Y L ( ϱ t + 1 ) − Y L ( ϱ t ) ≤ − α L ϵ 0 σ ( ϱ t ) , (81) Y_L(\varrho_{t+1})-Y_L(\varrho_t) \le-\alpha_L\epsilon_0\sigma(\varrho_t), \tag*{(81)} Y L ( ϱ t + 1 ) − Y L ( ϱ t ) ≤ − α L ϵ 0 σ ( ϱ t ) , (81)
for the resulting closed-loop with
α L : = 1 − γ S γ ˉ Y ϵ 0 2 ( L − 1 ) > 0. (82) \alpha_L := 1-\frac{\gamma_S\bar{\gamma}_Y}{\epsilon_0^2(L-1)}>0. \tag*{(82)} α L := 1 − ϵ 0 2 ( L − 1 ) γ S γ ˉ Y > 0. (82)
Further, the closed-loop system satisfies the following transient performance bound
α L ∑ t = 0 L − 1 ℓ ( ξ t , ω t ) ≤ Y L ( ϱ 0 ) ≤ Y ˉ , ∀ L ∈ N . (83) \alpha_L\sum_{t=0}^{L-1}\ell(\xi_t,\omega_t)\le Y_L(\varrho_0)\le\bar{Y},\quad\forall L\in\mathbb{N}. \tag*{(83)} α L t = 0 ∑ L − 1 ℓ ( ξ t , ω t ) ≤ Y L ( ϱ 0 ) ≤ Y ˉ , ∀ L ∈ N . (83)
The following proof is for completeness of the statements. We first show inequalities (80) -(81) for all ϱ t ∈ X L Y ˉ \varrho_t\in\mathbb{X}^{\bar{Y}}_L ϱ t ∈ X L Y ˉ , and then the performance bound (83) . Abbreviate ℓ k : = ℓ ( ξ k ∣ t ∗ ) \ell_k:=\ell(\xi^*_{k|t}) ℓ k := ℓ ( ξ k ∣ t ∗ ) .
Step I. The lower bound in (80) follows with ℓ 0 ≥ 0 \ell_0\ge0 ℓ 0 ≥ 0 , and
Y L ( ϱ t ) ≥ ℓ 0 + W ( ϱ t ) ≥ ϵ 0 σ ( ϱ t ) . (84) Y_L(\varrho_t)\ge\ell_0+W(\varrho_t)\ge\epsilon_0\sigma(\varrho_t). \tag*{(84)} Y L ( ϱ t ) ≥ ℓ 0 + W ( ϱ t ) ≥ ϵ 0 σ ( ϱ t ) . (84)
Recall that W ( ϱ t ) = ∥ ξ t − Π Ξ ω t ∥ P ϵ , w 2 ≤ λ max ( P ξ , w ) σ ( ϱ t ) = : γ σ σ ( ϱ t ) W(\varrho_t)=\lVert\xi_t-\Pi_{\Xi}\omega_t\rVert^2_{P_{\epsilon,w}}\le\lambda_{\max}(P_{\xi,w})\sigma(\varrho_t)=:\gamma_\sigma\sigma(\varrho_t) W ( ϱ t ) = ∥ ξ t − Π Ξ ω t ∥ P ϵ , w 2 ≤ λ m a x ( P ξ , w ) σ ( ϱ t ) =: γ σ σ ( ϱ t ) . For any ϱ t ∈ X δ Y ˉ \varrho_t\in\mathbb{X}^{\bar{Y}}_\delta ϱ t ∈ X δ Y ˉ , we directly obtain the bound Y L ( ϱ t ) ≤ ( γ S + γ σ ) σ ( ϱ t ) Y_L(\varrho_t)\le(\gamma_S+\gamma_\sigma)\sigma(\varrho_t) Y L ( ϱ t ) ≤ ( γ S + γ σ ) σ ( ϱ t ) using (32) . In addition, for ϱ t ∉ X δ Y ˉ \varrho_t\notin\mathbb{X}^{\bar{Y}}_\delta ϱ t ∈ / X δ Y ˉ , it follows from (25) that σ ( ϱ t ) ≥ Y ˉ γ δ \sigma(\varrho_t)\ge\frac{\bar{Y}}{\gamma_\delta} σ ( ϱ t ) ≥ γ δ Y ˉ . Hence, the upper bound in (44) holds with γ Y : = max { γ S + γ σ , Y ˉ γ δ } \gamma_Y:=\max\{\gamma_S+\gamma_\sigma,\frac{\bar{Y}}{\gamma_\delta}\} γ Y := max { γ S + γ σ , γ δ Y ˉ } .
Step II. The i-IOSS Lyapunov condition (34) implies
W ( ϱ L ∣ t ∗ ) − W ( ϱ t ) = ∑ k = 0 L − 1 W ( ϱ k + 1 ∣ t ∗ ) − W ( ϱ k ∣ t ∗ ) ≤ ∑ k = 0 L − 1 ϵ σ ( ϱ k ∣ t ∗ ) + ∑ k = 0 L − 1 ℓ k . (85) \begin{aligned}
W(\varrho^*_{L|t})-W(\varrho_t)=\sum_{k=0}^{L-1}W(\varrho^*_{k+1|t})-W(\varrho^*_{k|t}) \\
&\le\sum_{k=0}^{L-1}\epsilon_\sigma(\varrho^*_{k|t})+\sum_{k=0}^{L-1}\ell_k. \tag*{(85)}
\end{aligned} W ( ϱ L ∣ t ∗ ) − W ( ϱ t ) = k = 0 ∑ L − 1 W ( ϱ k + 1∣ t ∗ ) − W ( ϱ k ∣ t ∗ ) ≤ k = 0 ∑ L − 1 ϵ σ ( ϱ k ∣ t ∗ ) + k = 0 ∑ L − 1 ℓ k . (85)
Using W ( ϱ L ∣ t ∗ ) ≥ 0 W(\varrho^*_{L|t}) \ge0 W ( ϱ L ∣ t ∗ ) ≥ 0 , and ϱ t ∈ X Y ˉ ρ \varrho_t \in\mathcal{X}_{\bar{Y}}^\rho ϱ t ∈ X Y ˉ ρ , we arrive at
ϵ o ∑ k = 0 L − 1 σ ( ϱ k ∣ t ∗ ) ≤ Y L ( ϱ t ) ≤ min { Y ˉ , γ ˉ Y ˉ σ ( ϱ t ) } . (86) \epsilon_o \sum_{k=0}^{L-1} \sigma(\varrho^*_{k|t}) \le Y_L(\varrho_t) \le\min\{\bar{Y},\bar{\gamma}\bar{Y}\sigma(\varrho_t)\}. \tag*{(86)} ϵ o k = 0 ∑ L − 1 σ ( ϱ k ∣ t ∗ ) ≤ Y L ( ϱ t ) ≤ min { Y ˉ , γ ˉ Y ˉ σ ( ϱ t )} . (86)
Thus, there exists integer k ϱ ∈ { 1 , ⋯ , L − 1 } k_\varrho\in\{1,\cdots,L-1\} k ϱ ∈ { 1 , ⋯ , L − 1 } , such that
σ ( ϱ k ϱ ∣ t ∗ ) ≤ min { Y ˉ , γ ˉ Y ˉ σ ( ϱ t ) } ϵ o ( L − 1 ) . (87) \sigma(\varrho^*_{k_\varrho|t}) \le\frac{\min\{\bar{Y},\bar{\gamma}\bar{Y}\sigma(\varrho_t)\}}{\epsilon_o(L-1)}. \tag*{(87)} σ ( ϱ k ϱ ∣ t ∗ ) ≤ ϵ o ( L − 1 ) min { Y ˉ , γ ˉ Y ˉ σ ( ϱ t )} . (87)
Given L ≥ L 0 : = 1 + Y ˉ δ s ϵ o L \ge L_0 := 1+\frac{\bar{Y}}{\delta_s\epsilon_o} L ≥ L 0 := 1 + δ s ϵ o Y ˉ , this implies ϱ k ϱ ∣ t ∗ ∈ X Y ˉ ρ \varrho^*_{k_\varrho|t} \in\mathcal{X}_{\bar{Y}}^\rho ϱ k ϱ ∣ t ∗ ∈ X Y ˉ ρ . Thus, Proposition 2 ensures that starting at ϱ k ϱ ∣ t ∗ \varrho^*_{k_\varrho|t} ϱ k ϱ ∣ t ∗ there exists feasible state and input trajectory satisfying the bound (32) , which implies
J L ( ϱ t + 1 ) + ℓ 0 ≤ ∑ k = 0 k ϱ − 1 ℓ k + J L − k ϱ + 1 ( ϱ k ϱ ∣ t ∗ ) ≤ J L ( ϱ t ) + γ s σ ( ϱ k ϱ ∣ t ∗ ) ≤ J L ( ϱ t ) + γ s γ Y ˉ ϵ o ( L − 1 ) σ ( ϱ t ) . (88) \begin{aligned}
J_L(\varrho_{t+1})+\ell_0 \le\sum_{k=0}^{k_\varrho-1}\ell_k+J_{L-k_\varrho+1}(\varrho^*_{k_\varrho|t}) \\
&\le J_L(\varrho_t)+\gamma_s\sigma(\varrho^*_{k_\varrho|t}) \le J_L(\varrho_t)+\frac{\gamma_s\gamma\bar{Y}}{\epsilon_o(L-1)}\sigma(\varrho_t). \tag*{(88)}
\end{aligned} J L ( ϱ t + 1 ) + ℓ 0 ≤ k = 0 ∑ k ϱ − 1 ℓ k + J L − k ϱ + 1 ( ϱ k ϱ ∣ t ∗ ) ≤ J L ( ϱ t ) + γ s σ ( ϱ k ϱ ∣ t ∗ ) ≤ J L ( ϱ t ) + ϵ o ( L − 1 ) γ s γ Y ˉ σ ( ϱ t ) . (88)
Combining (51) and (24b), the function Y L Y_L Y L satisfies (44b). Furthermore, α L > 0 \alpha_L>0 α L > 0 follows from (45), using L > L 1 : = 1 + γ s γ Y ˉ / ϵ 2 L>L_1:=1+\gamma_s\gamma\bar{Y}/\epsilon^2 L > L 1 := 1 + γ s γ Y ˉ / ϵ 2 . All the arguments hold with L > L Y ˉ : = max { L 0 , L 1 } L>L_{\bar{Y}}:=\max\{L_0,L_1\} L > L Y ˉ := max { L 0 , L 1 } . In addition, inequality (44b) ensures Y L Y_L Y L is nonincreasing and thus ϱ t ∈ X Y ˉ ρ \varrho_t\in\mathcal{X}_{\bar{Y}}^\rho ϱ t ∈ X Y ˉ ρ for all t ≥ 0 t\ge0 t ≥ 0 .
Step III. Using Y L = J L + W ≥ J L ≥ 0 Y_L=J_L+W\ge J_L\ge0 Y L = J L + W ≥ J L ≥ 0 , α L ≤ 1 \alpha_L\le1 α L ≤ 1 and the following inequality in the telescopic sum
Y L ( ϱ t + 1 ) − Y L ( ϱ t ) − α L ( W ( A ϱ ϱ + B ϱ u ) − W ( ϱ ) ) (89) Y_L(\varrho_{t+1})-Y_L(\varrho_t)-\alpha_L\bigl(W(A_\varrho\varrho+B_\varrho u)-W(\varrho)\bigr) \tag*{(89)} Y L ( ϱ t + 1 ) − Y L ( ϱ t ) − α L ( W ( A ϱ ϱ + B ϱ u ) − W ( ϱ ) ) (89)
= ( 1 − α L ) ( Y L ( ϱ t + 1 ) − Y L ( ϱ t ) ) + α L ( J L ( ϱ t + 1 ) − J L ( ϱ t ) ) ≤ − α L ℓ ( ϱ t ) , (90) =(1-\alpha_L)\bigl(Y_L(\varrho_{t+1})-Y_L(\varrho_t)\bigr)+\alpha_L\bigl(J_L(\varrho_{t+1})-J_L(\varrho_t)\bigr)
\le-\alpha_L\ell(\varrho_t), \tag*{(90)} = ( 1 − α L ) ( Y L ( ϱ t + 1 ) − Y L ( ϱ t ) ) + α L ( J L ( ϱ t + 1 ) − J L ( ϱ t ) ) ≤ − α L ℓ ( ϱ t ) , (90)
the performance bound (83) is established.
Inequalities (44a)–(44b) directly imply lim t → ∞ σ ( ϱ t ) = 0 \lim_{t\to\infty}\sigma(\varrho_t)=0 lim t → ∞ σ ( ϱ t ) = 0 and hence asymptotic convergence. Building on this result, in the following, we show that A ϱ A_\varrho A ϱ is exponentially stable. Define the point-to-set distance ∥ ϱ ∥ A ϱ : = inf s ∈ A ϱ ∥ ϱ − s ∥ \|\varrho\|_{A_\varrho}:=\inf_{s\in A_\varrho}\|\varrho-s\| ∥ ϱ ∥ A ϱ := inf s ∈ A ϱ ∥ ϱ − s ∥ . In the following, we show that there exists constant c π > 0 c_\pi>0 c π > 0 , such that c π σ ( ϱ ) ≤ ∥ ϱ ∥ A ϱ 2 ≤ σ ( ϱ ) c_\pi\sigma(\varrho)\le\|\varrho\|^2_{A_\varrho}\le\sigma(\varrho) c π σ ( ϱ ) ≤ ∥ ϱ ∥ A ϱ 2 ≤ σ ( ϱ ) , which in combination with (44)–(46) ensures exponential stability of A ϱ A_\varrho A ϱ using standard Lyapunov arguments. For given ( ξ , ω ) (\xi,\omega) ( ξ , ω ) , denote some minimizer by ω ~ : = arg min ω ~ ∈ R ∥ ( Π ξ ω , ω ~ ) − ( ξ , ω ) ∥ \tilde{\omega}:=\arg\min_{\tilde{\omega}\in\mathbb{R}}\|(\Pi_{\xi\omega},\tilde{\omega})-(\xi,\omega)\| ω ~ := arg min ω ~ ∈ R ∥ ( Π ξ ω , ω ~ ) − ( ξ , ω ) ∥ . We can deduce that
σ ( ϱ ) = ∥ ξ − Π ξ ω ∥ 2 ≤ 2 ( ∥ ξ − Π ξ ω ~ ∥ 2 + ∥ Π ξ ω ~ − Π ξ ω ∥ 2 ) \sigma(\varrho)=\|\xi-\Pi_{\xi\omega}\|^2
\le2\bigl(\|\xi-\Pi_{\xi\tilde{\omega}}\|^2+\|\Pi_{\xi\tilde{\omega}}-\Pi_{\xi\omega}\|^2\bigr) σ ( ϱ ) = ∥ ξ − Π ξ ω ∥ 2 ≤ 2 ( ∥ ξ − Π ξ ω ~ ∥ 2 + ∥ Π ξ ω ~ − Π ξ ω ∥ 2 )
≤ 2 max { 1 , ∥ Π ξ ∥ 2 } ∥ ( ξ , ω ) − ( Π ξ ω ~ , ω ~ ) ∥ 2 = 1 c π ∥ ϱ ∥ A ϱ 2 , \le2\max\{1,\|\Pi_\xi\|^2\}\|(\xi,\omega)-(\Pi_{\xi\tilde{\omega}},\tilde{\omega})\|^2= \frac{1}{c_\pi}\|\varrho\|^2_{A_\varrho}, ≤ 2 max { 1 , ∥ Π ξ ∥ 2 } ∥ ( ξ , ω ) − ( Π ξ ω ~ , ω ~ ) ∥ 2 = c π 1 ∥ ϱ ∥ A ϱ 2 ,
where the first inequality uses ∥ a + b ∥ 2 ≤ 2 ( ∥ a ∥ 2 + ∥ b ∥ 2 ) \|a+b\|^2\le2(\|a\|^2+\|b\|^2) ∥ a + b ∥ 2 ≤ 2 ( ∥ a ∥ 2 + ∥ b ∥ 2 ) for any a , b ∈ R n a,b\in\mathbb{R}^n a , b ∈ R n . Further,
∥ ϱ ∥ A ϱ = ∥ ( Π ξ ω ~ , ω ~ ) − ( ξ , ω ) ∥ ≤ ∥ ξ − Π ξ ω ∥ = σ ( ϱ ) . (91) \|\varrho\|_{A_\varrho}=\|(\Pi_{\xi\tilde{\omega}},\tilde{\omega})-(\xi,\omega)\|\le\|\xi-\Pi_{\xi\omega}\|=\sqrt{\sigma(\varrho)}. \tag*{(91)} ∥ ϱ ∥ A ϱ = ∥ ( Π ξ ω ~ , ω ~ ) − ( ξ , ω ) ∥ ≤ ∥ ξ − Π ξ ω ∥ = σ ( ϱ ) . (91)
Hence, lim t → ∞ ∥ ϱ ∥ A ϱ = 0 \lim_{t\to\infty}\|\varrho\|_{A_\varrho}=0 lim t → ∞ ∥ ϱ ∥ A ϱ = 0 as lim t → ∞ σ ( ϱ ) = 0 \lim_{t\to\infty}\sigma(\varrho)=0 lim t → ∞ σ ( ϱ ) = 0 , which completes the proof.
Appendix B: Proof of Proposition 4
Proof: Regulator equations: Since ( Π ξ , Γ ξ ) (\Pi_\xi,\Gamma_\xi) ( Π ξ , Γ ξ ) solves the regulator equations (19) and S T = I S^T=I S T = I under Assumption 5, the augmented regulator equations are solved by
Π a = [ Π ξ Γ ξ S T − 1 ⋮ Γ ξ ] , Γ a = 0. (92) \Pi^a=
\begin{bmatrix}
\Pi_\xi\\
\Gamma_\xi S^{T-1}\\
\vdots\\
\Gamma_\xi
\end{bmatrix},
\qquad\Gamma^a=0. \tag*{(92)} Π a = Π ξ Γ ξ S T − 1 ⋮ Γ ξ , Γ a = 0. (92)
Stabilizability and detectability: Recalling that matrices ( A a , B a , C a ) (A^a,B^a,C^a) ( A a , B a , C a ) of the augmented system (28) are given by
A a = [ A ι B ι E 2 T 0 E 0 ] , B a = [ B ι E 1 ] , C a = [ Z 1 0 ] . A^a=\begin{bmatrix}A_\iota&B_\iota E_2^T\\0&E_0\end{bmatrix},
\qquad B^a=\begin{bmatrix}B_\iota\\E_1\end{bmatrix},
\qquad C^a=[\,Z_1\ 0\,]. A a = [ A ι 0 B ι E 2 T E 0 ] , B a = [ B ι E 1 ] , C a = [ Z 1 0 ] .
Stabilizability of ( A a , B a ) (A^a,B^a) ( A a , B a ) is equivalent to
rank [ A ι − λ I n ξ B ι E 2 T B ι 0 E 0 − λ I m T E 1 ] = n ξ + T m , (93) \operatorname{rank}
\begin{bmatrix}
A_\iota-\lambda I_{n_\xi}&B_\iota E_2^T&B_\iota\\
0&E_0-\lambda I_{mT}&E_1
\end{bmatrix}
=n_\xi+Tm, \tag*{(93)} rank [ A ι − λ I n ξ 0 B ι E 2 T E 0 − λ I m T B ι E 1 ] = n ξ + T m , (93)
for all λ ∈ C \lambda\in\mathbb{C} λ ∈ C satisfying ∣ λ ∣ ≥ 1 |\lambda|\ge1 ∣ λ ∣ ≥ 1 . Consider λ k = e i 2 π k / T \lambda_k=e^{i2\pi k/T} λ k = e i 2 π k / T for k = 0 , ⋯ , T − 1 k=0,\cdots,T-1 k = 0 , ⋯ , T − 1 , in which case rank ( E 0 − λ k I m T ) = m ( T − 1 ) \operatorname{rank}(E_0-\lambda_kI_{mT})=m(T-1) rank ( E 0 − λ k I m T ) = m ( T − 1 ) . It suffices to show the argument for m = 1 m=1 m = 1 , since the block case follows componentwise.
There exists an eigenvector r k = [ λ k − 1 ⋯ λ k − T ] T r^k=[\lambda_k^{-1}\ \cdots\ \lambda_k^{-T}]^T r k = [ λ k − 1 ⋯ λ k − T ] T satisfying ( E 0 − λ k I T ) r k = 0 (E_0-\lambda_kI_T)r^k=0 ( E 0 − λ k I T ) r k = 0 and E 2 T r k = 1 E_2^Tr^k=1 E 2 T r k = 1 . To verify (93) , let [ v T q T ] [v^T\ q^T] [ v T q T ] belong to the left kernel of its matrix. Then
v T ( A ι − λ k I n ξ ) = 0 , v^T(A_\iota-\lambda_kI_{n_\xi})=0, v T ( A ι − λ k I n ξ ) = 0 ,
v T B ι E 2 T + q T ( E 0 − λ k I T ) = 0 , v^TB_\iota E_2^T+q^T(E_0-\lambda_kI_T)=0, v T B ι E 2 T + q T ( E 0 − λ k I T ) = 0 ,
v T B ι + q T E 1 = 0. v^TB_\iota+q^TE_1=0. v T B ι + q T E 1 = 0.
Right multiplying the second equality by r k r^k r k yields v T B ι = 0 v^TB_\iota=0 v T B ι = 0 . Stabilizability of ( A ι , B ι ) (A_\iota,B_\iota) ( A ι , B ι ) therefore implies v = 0 v=0 v = 0 . The remaining equalities give q T ( E 0 − λ k I T ) = 0 q^T(E_0-\lambda_kI_T)=0 q T ( E 0 − λ k I T ) = 0 and q T E 1 = 0 q^TE_1=0 q T E 1 = 0 , which imply q = 0 q=0 q = 0 . Hence, (93) holds for λ = λ k \lambda=\lambda_k λ = λ k . For λ ∉ { λ k : k = 0 , ⋯ , T − 1 } \lambda\notin\{\lambda_k:k=0,\cdots,T-1\} λ ∈ / { λ k : k = 0 , ⋯ , T − 1 } , the matrix E 0 − λ I m T E_0-\lambda I_{mT} E 0 − λ I m T is nonsingular. Moreover, noting that I m − E 2 T ( E 0 − λ I m T ) − 1 E 1 = λ T λ T − 1 I m I_m-E_2^T(E_0-\lambda I_{mT})^{-1}E_1=\frac{\lambda^T}{\lambda^T-1}I_m I m − E 2 T ( E 0 − λ I m T ) − 1 E 1 = λ T − 1 λ T I m is nonsingular, block elimination reduces (93) to the PBH rank condition for ( A ι , B ι ) (A_\iota,B_\iota) ( A ι , B ι ) . Therefore, ( A a , B a ) (A^a,B^a) ( A a , B a ) is stabilizable.
The detectability of ( A a , C a ) (A^a,C^a) ( A a , C a ) follows from [25 Proposition 7] ; the argument is recalled for completeness. Detectability of ( A a , C a ) (A^a,C^a) ( A a , C a ) is equivalent to
rank [ A ι − λ I n ξ B ι E 2 T 0 E 0 − λ I m T Z 1 0 ] = n ξ + m T , (94) \operatorname{rank}
\begin{bmatrix}
A_\iota-\lambda I_{n_\xi}&B_\iota E_2^T\\
0&E_0-\lambda I_{mT}\\
Z_1&0
\end{bmatrix}
=n_\xi+mT, \tag*{(94)} rank A ι − λ I n ξ 0 Z 1 B ι E 2 T E 0 − λ I m T 0 = n ξ + m T , (94)
for all λ ∈ C \lambda\in\mathbb{C} λ ∈ C satisfying ∣ λ ∣ ≥ 1 |\lambda|\ge1 ∣ λ ∣ ≥ 1 . As in the stabilizability argument, consider λ k = e i 2 π k / T \lambda_k=e^{i2\pi k/T} λ k = e i 2 π k / T for k = 0 , ⋯ , T − 1 k=0,\cdots,T-1 k = 0 , ⋯ , T − 1 , in which case rank ( E 0 − λ k I m T ) = m ( T − 1 ) \operatorname{rank}(E_0-\lambda_kI_{mT})=m(T-1) rank ( E 0 − λ k I m T ) = m ( T − 1 ) . Again, it suffices to prove the argument for m = 1 m=1 m = 1 . There exists an eigenvector ζ 1 k = [ λ k − 1 ⋯ λ k − T ] T \zeta_1^k=[\lambda_k^{-1}\ \cdots\ \lambda_k^{-T}]^T ζ 1 k = [ λ k − 1 ⋯ λ k − T ] T satisfying ( E 0 − λ k I m T ) ζ 1 k = 0 (E_0-\lambda_kI_{mT})\zeta_1^k=0 ( E 0 − λ k I m T ) ζ 1 k = 0 . Further, we have E 2 T ζ 1 k = e − i 2 π k / T = 1 E_2^T\zeta_1^k=e^{-i2\pi k/T}=1 E 2 T ζ 1 k = e − i 2 π k / T = 1 . Thus, the rank condition (57) is equivalent to the nonresonance condition [14 ] evaluated at λ = λ k \lambda=\lambda_k λ = λ k , which holds by Assumption 6. Note that λ k \lambda_k λ k need not be an eigenvalue of S S S ; hence, Assumption 3 alone is not sufficient at this step. For λ ∉ { λ k : k = 0 , ⋯ , T − 1 } \lambda\notin\{\lambda_k:k=0,\cdots,T-1\} λ ∈ / { λ k : k = 0 , ⋯ , T − 1 } , the matrix E 0 − λ I m T E_0-\lambda I_{mT} E 0 − λ I m T is nonsingular, and (57) reduces to the PBH rank condition for ( A ι , Z 1 ) (A_\iota,Z_1) ( A ι , Z 1 ) . Hence, ( A a , C a ) (A^a,C^a) ( A a , C a ) is detectable. ■
Appendix C: Proof of Lemma 4
Proof: The proof has four steps. Step I bounds the one-step prediction error in terms of ϵ ˉ \bar{\epsilon} ϵ ˉ . Step II constructs a shifted feasible candidate of length L − 1 L-1 L − 1 at time t + 1 t+1 t + 1 . Step III bounds the cost of this candidate, and Step IV combines the resulting value-function estimate with the perturbed i-UIOSS inequality. Step I. Let ξ ^ 1 ∣ t α , ∗ \hat{\xi}^{\alpha,*}_{1|t} ξ ^ 1∣ t α , ∗ denote the one-step auxiliary state generated by the noise-free trajectory
H L + ι ( y ) g t α , ∗ = y ˉ t + h t ∗ − H L + ι ( ε ) g t α , ∗ , (95) H_{L+\iota}(y)g_t^{\alpha,*}=\bar{y}_t+h_t^*-H_{L+\iota}(\varepsilon)g_t^{\alpha,*}, \tag*{(95)} H L + ι ( y ) g t α , ∗ = y ˉ t + h t ∗ − H L + ι ( ε ) g t α , ∗ , (95)
with the optimal input and exosystem sequences at time t t t , and define δ 1 ∣ t ξ : = ξ t + 1 − ξ ^ 1 ∣ t α , ∗ \delta_{1|t}^{\xi}:=\xi_{t+1}-\hat{\xi}^{\alpha,*}_{1|t} δ 1∣ t ξ := ξ t + 1 − ξ ^ 1∣ t α , ∗ .
From Y L f ( σ ~ t a ) ≤ Y ˉ Y_L^f(\tilde{\sigma}_t^a)\leq\bar{Y} Y L f ( σ ~ t a ) ≤ Y ˉ and nonnegativity of the objective terms in (63) , we get that
∥ h t ∗ ∥ ≤ Y ˉ ε ˉ λ h , ∥ g t ∗ − g t c ∥ ≤ Y ˉ ε ˉ λ g . (96) \|h_t^*\|\leq\sqrt{\frac{\bar{Y}\bar{\varepsilon}}{\lambda_h}},\qquad
\|g_t^*-g_t^{c}\|\leq\sqrt{\frac{\bar{Y}}{\bar{\varepsilon}\lambda_g}}. \tag*{(96)} ∥ h t ∗ ∥ ≤ λ h Y ˉ ε ˉ , ∥ g t ∗ − g t c ∥ ≤ ε ˉ λ g Y ˉ . (96)
Moreover, Proposition 5 and the bounds following (34) give
∥ g t ∗ ∥ ≤ Y ˉ ε ˉ λ g + c c + J ˉ s λ g . \|g_t^*\|\leq\sqrt{\frac{\bar{Y}}{\bar{\varepsilon}\lambda_g}}+\sqrt{c_c}+\sqrt{\frac{\bar{J}_s}{\lambda_g}}. ∥ g t ∗ ∥ ≤ ε ˉ λ g Y ˉ + c c + λ g J ˉ s .
The input-history component and the exosystem prediction are exact. Hence, by shifting the output-history component in (95) , the one-step prediction error consists only of the online measurement error, the noisy-Hankel term H L + ι ( ε ) g t ∗ H_{L+\iota}(\varepsilon)g_t^* H L + ι ( ε ) g t ∗ , and the slack h t ∗ h_t^* h t ∗ . Each relevant row block of H L + ι ( ε ) H_{L+\iota}(\varepsilon) H L + ι ( ε ) is bounded by a constant multiple of ε ˉ \bar{\varepsilon} ε ˉ . Together with (96) and the auxiliary output equation, this yields functions b y , b ξ ∈ K ∞ b_y,b_\xi\in\mathcal{K}_\infty b y , b ξ ∈ K ∞ such that
∥ y t − y ˉ 0 ∣ t ∥ ≤ b y ( ε ˉ ) , ∥ δ 1 ∣ t ξ ∥ ≤ b ξ ( ε ˉ ) . (97) \|y_t-\bar{y}_{0|t}\|\leq b_y(\bar{\varepsilon}),\qquad
\|\delta_{1|t}^{\xi}\|\leq b_\xi(\bar{\varepsilon}). \tag*{(97)} ∥ y t − y ˉ 0∣ t ∥ ≤ b y ( ε ˉ ) , ∥ δ 1∣ t ξ ∥ ≤ b ξ ( ε ˉ ) . (97)
Since the input memory is generated from the applied inputs, its one-step prediction is exact. Thus, with ξ ^ 1 ∣ t α , ∗ : = [ ξ ^ 1 ∣ t α , ∗ ζ ^ 1 ∣ t α , ∗ ] \hat{\xi}_{1|t}^{\alpha,*}:=\begin{bmatrix}\hat{\xi}_{1|t}^{\alpha,*}\\ \hat{\zeta}_{1|t}^{\alpha,*}\end{bmatrix} ξ ^ 1∣ t α , ∗ := [ ξ ^ 1∣ t α , ∗ ζ ^ 1∣ t α , ∗ ] , we get that ∥ δ 1 ∣ t ξ α ∥ ≤ b ξ ( ε ˉ ) \|\delta_{1|t}^{\xi^\alpha}\|\leq b_\xi(\bar{\varepsilon}) ∥ δ 1∣ t ξ α ∥ ≤ b ξ ( ε ˉ ) , where δ 1 ∣ t ξ α : = ξ t + 1 α , ∗ − ξ 1 ∣ t α , ∗ = [ δ 1 ∣ t ξ 0 m T × 1 ] \delta_{1|t}^{\xi^\alpha}:=\xi_{t+1}^{\alpha,*}-\xi_{1|t}^{\alpha,*}=\begin{bmatrix}\delta_{1|t}^{\xi}\\0_{m_T\times1}\end{bmatrix} δ 1∣ t ξ α := ξ t + 1 α , ∗ − ξ 1∣ t α , ∗ = [ δ 1∣ t ξ 0 m T × 1 ] .
Step II. We construct a candidate of length L − 1 L-1 L − 1 at time t + 1 t+1 t + 1 by shifting the previously optimal physical-input sequence:
u ˉ [ − ι , L − 2 ] ∣ t + 1 : = u ˉ [ − ι + 1 , L − 1 ] ∣ t ∗ , (98) \bar{u}_{[-\iota,L-2]|t+1}:=\bar{u}_{[-\iota+1,L-1]|t}^*, \tag*{(98)} u ˉ [ − ι , L − 2 ] ∣ t + 1 := u ˉ [ − ι + 1 , L − 1 ] ∣ t ∗ , (98)
and set
ζ ˉ k ∣ t + 1 : = ζ ˉ k + 1 ∣ t ∗ , k = 0 , … , L − 1 , (99) \bar{\zeta}_{k|t+1}:=\bar{\zeta}_{k+1|t}^*,\qquad k=0,\ldots,L-1, \tag*{(99)} ζ ˉ k ∣ t + 1 := ζ ˉ k + 1∣ t ∗ , k = 0 , … , L − 1 , (99)
u ˉ k ∣ t + 1 : = u ˉ k + 1 ∣ t ∗ , k = 0 , … , L − 2. (100) \bar{u}_{k|t+1}:=\bar{u}_{k+1|t}^*,\qquad k=0,\ldots,L-2. \tag*{(100)} u ˉ k ∣ t + 1 := u ˉ k + 1∣ t ∗ , k = 0 , … , L − 2. (100)
Since the first optimal physical input is applied at time t t t , the shifted memory state satisfies the required initialization. The augmented dynamics and the physical-input constraint follow directly from the previous feasible solution; in particular, u ˉ k ∣ t + 1 = u ˉ k + 1 ∣ t ∗ ∈ U \bar{u}_{k|t+1}=\bar{u}_{k+1|t}^*\in\mathbb{U} u ˉ k ∣ t + 1 = u ˉ k + 1∣ t ∗ ∈ U for k = 0 , … , L − 2 k=0,\ldots,L-2 k = 0 , … , L − 2 .
It remains to construct the corresponding Hankel coefficient and slack. Let
H x u m : = [ H 1 ( x [ 0 , N − L − ι + 1 ] ) H L + ι − 1 ( u ) H 1 ( M [ 0 , N − L − ι + 1 ] ) ] . H_{xum}:=\begin{bmatrix}
H_1(x_{[0,N-L-\iota+1]})\\
H_{L+\iota-1}(u)\\
H_1(M_{[0,N-L-\iota+1]})
\end{bmatrix}. H xu m := H 1 ( x [ 0 , N − L − ι + 1 ] ) H L + ι − 1 ( u ) H 1 ( M [ 0 , N − L − ι + 1 ] ) .
The assumed persistence of excitation of order L + ι L+\iota L + ι implies that H x u m H_{xum} H xu m has full row rank at the shortened depth. Moreover, observability of (1) uniquely determines x t + 1 − ι x_{t+1-\iota} x t + 1 − ι from the last ι \iota ι noise-free I/O samples. Hence, a right inverse of H x u m H_{xum} H xu m provides
g ˉ t + 1 = H x u m † [ x t + 1 − ι u ˉ [ − ι , L − 2 ] ∣ t + 1 M ˉ t + 1 − ι ] . (101) \bar{g}_{t+1}=H_{xum}^{\dagger}
\begin{bmatrix}
x_{t+1-\iota}\\
\bar{u}_{[-\iota,L-2]|t+1}\\
\bar{M}_{t+1-\iota}
\end{bmatrix}. \tag*{(101)} g ˉ t + 1 = H xu m † x t + 1 − ι u ˉ [ − ι , L − 2 ] ∣ t + 1 M ˉ t + 1 − ι . (101)
Choose the output and slack candidates as
y ˉ [ − ι , L − 2 ] ∣ t + 1 = H L + ι − 1 ( y ) g ˉ t + 1 + [ ε [ t + 1 − ι , t ] 0 p ( L − 1 ) × 1 ] , (102) \bar{y}_{[-\iota,L-2]|t+1}=H_{L+\iota-1}(y)\bar{g}_{t+1}+
\begin{bmatrix}
\varepsilon_{[t+1-\iota,t]}\\
0_{p(L-1)\times1}
\end{bmatrix}, \tag*{(102)} y ˉ [ − ι , L − 2 ] ∣ t + 1 = H L + ι − 1 ( y ) g ˉ t + 1 + [ ε [ t + 1 − ι , t ] 0 p ( L − 1 ) × 1 ] , (102)
h ˉ [ − ι , L − 2 ] ∣ t + 1 = H L + ι − 1 ( ε ) g ˉ t + 1 − [ ε [ t + 1 − ι , t ] 0 p ( L − 1 ) × 1 ] . (103) \bar{h}_{[-\iota,L-2]|t+1}=H_{L+\iota-1}(\varepsilon)\bar{g}_{t+1}-
\begin{bmatrix}
\varepsilon_{[t+1-\iota,t]}\\
0_{p(L-1)\times1}
\end{bmatrix}. \tag*{(103)} h ˉ [ − ι , L − 2 ] ∣ t + 1 = H L + ι − 1 ( ε ) g ˉ t + 1 − [ ε [ t + 1 − ι , t ] 0 p ( L − 1 ) × 1 ] . (103)
These choices satisfy the noisy Hankel equation and the measured initial-condition constraint by construction. Together with (98) –(62), this proves feasibility of the shortened candidate.
Step III. We first compare the shifted output with the previous optimal output. Since the two predictions use the same shifted input and exosystem sequences, the auxiliary dynamics (17) give
δ k ∣ t + 1 y : = y ˉ k ∣ t + 1 − y ˉ k + 1 ∣ t ∗ = Z 1 A 1 k δ 1 ∣ t ξ α + h k + 1 ∣ t ∗ − ( H L + ι ( ε ) g ∗ ) k + 1 . \delta^y_{k|t+1}:=\bar{y}_{k|t+1}-\bar{y}_{k+1|t}^*
=Z_1A_1^k\delta_{1|t}^{\xi^\alpha}+h_{k+1|t}^*-\left(H_{L+\iota}(\varepsilon)g^*\right)_{k+1}. δ k ∣ t + 1 y := y ˉ k ∣ t + 1 − y ˉ k + 1∣ t ∗ = Z 1 A 1 k δ 1∣ t ξ α + h k + 1∣ t ∗ − ( H L + ι ( ε ) g ∗ ) k + 1 .
Combining this identity with (96) and (97) , and using the fact that the horizon is fixed, gives some b δ ∈ K ∞ b_\delta\in\mathcal{K}_\infty b δ ∈ K ∞ such that
∑ k = 0 L − 2 ∥ δ k ∣ t + 1 y ∥ Q 2 ≤ b δ ( ε ˉ ) . (104) \sum_{k=0}^{L-2}\|\delta^y_{k|t+1}\|_Q^2\leq b_\delta(\bar{\varepsilon}). \tag*{(104)} k = 0 ∑ L − 2 ∥ δ k ∣ t + 1 y ∥ Q 2 ≤ b δ ( ε ˉ ) . (104)
It remains to bound the coefficient and slack regularization terms. Partition Π α = [ Π ε α Π ζ α ] \Pi^\alpha=\begin{bmatrix}\Pi_\varepsilon^\alpha\\ \Pi_\zeta^\alpha\end{bmatrix} Π α = [ Π ε α Π ζ α ] . The regulator equations and (42) imply
u ˉ k ∣ t − Γ ξ ξ ˉ k ∣ t = E T 2 ( ξ ˉ k ∣ t − Π α ξ ˉ k ∣ t ) + u ˉ k ∣ t a . \bar{u}_{k|t}-\Gamma_\xi\bar{\xi}_{k|t}=E_T^2\left(\bar{\xi}_{k|t}-\Pi^\alpha\bar{\xi}_{k|t}\right)+\bar{u}_{k|t}^{a}. u ˉ k ∣ t − Γ ξ ξ ˉ k ∣ t = E T 2 ( ξ ˉ k ∣ t − Π α ξ ˉ k ∣ t ) + u ˉ k ∣ t a .
On the sublevel set Y L f ( σ ~ t a ) ≤ Y ˉ Y_L^f(\tilde{\sigma}_t^a)\leq\bar{Y} Y L f ( σ ~ t a ) ≤ Y ˉ , the lower bound on W a W^a W a and the objective in (63) uniformly bound the initial augmented error and the optimal augmented-input sequence. Together with the one-step error estimate from Step I, the finite-horizon augmented dynamics, and the observability reconstruction of x t + 1 − ι x_{t+1-\iota} x t + 1 − ι , this uniformly bounds the state and physical-input quantities entering the right-inverse construction (101) . Hence, Proposition 5 and ∥ g t + 1 α , ∗ − g ~ t + 1 ∥ 2 ≤ c g 2 \|g_{t+1}^{\alpha,*}-\tilde{g}_{t+1}\|^2\leq c_g^2 ∥ g t + 1 α , ∗ − g ~ t + 1 ∥ 2 ≤ c g 2 give a constant c g > 0 c_g>0 c g > 0 , independent of ε ˉ \bar{\varepsilon} ε ˉ , such that
∥ g ˉ t + 1 − g t + 1 c ∥ 2 ≤ c g 2 . (105) \|\bar{g}_{t+1}-g_{t+1}^{c}\|^2\leq c_g^2. \tag*{(105)} ∥ g ˉ t + 1 − g t + 1 c ∥ 2 ≤ c g 2 . (105)
The same estimates and the construction (103) give a constant b h > 0 b_h>0 b h > 0 , such that ∥ h ˉ t + 1 ∥ ≤ ε ˉ b h \|\bar{h}_{t+1}\|\leq\bar{\varepsilon}b_h ∥ h ˉ t + 1 ∥ ≤ ε ˉ b h .
Define ℓ 0 ∣ t α , ∗ : = ∥ y ˉ 0 ∣ t ∗ ∥ Q 2 + ∥ u ˉ 0 ∣ t a , ∗ ∥ R 2 \ell_{0|t}^{\alpha,*}:=\|\bar{y}_{0|t}^*\|_Q^2+\|\bar{u}_{0|t}^{a,*}\|_R^2 ℓ 0∣ t α , ∗ := ∥ y ˉ 0∣ t ∗ ∥ Q 2 + ∥ u ˉ 0∣ t a , ∗ ∥ R 2 . Using candidate feasibility, u ˉ a \bar{u}^{a} u ˉ a in (100) , (104) , and J L f , ∗ ≤ Y ˉ J_L^{f,*}\leq\bar{Y} J L f , ∗ ≤ Y ˉ , we obtain
J L − 1 f , ∗ ( σ ~ t + 1 a ) ≤ J L f , ∗ ( σ ~ t a ) − ℓ 0 ∣ t α , ∗ + 2 Y ˉ b δ ( ε ˉ ) + b δ ( ε ˉ ) + λ g c g 2 ε ˉ + λ h b h 2 ε ˉ = : J L f , ∗ ( σ ~ t a ) − ℓ 0 ∣ t α , ∗ + α 1 ( ε ˉ ) , (106) J_{L-1}^{f,*}(\tilde{\sigma}_{t+1}^{a})\leq J_L^{f,*}(\tilde{\sigma}_t^a)-\ell_{0|t}^{\alpha,*}
+2\sqrt{\bar{Y}b_\delta(\bar{\varepsilon})}+b_\delta(\bar{\varepsilon})+\lambda_gc_g^2\bar{\varepsilon}+\lambda_hb_h^2\bar{\varepsilon}
=:J_L^{f,*}(\tilde{\sigma}_t^a)-\ell_{0|t}^{\alpha,*}+\alpha_1(\bar{\varepsilon}), \tag*{(106)} J L − 1 f , ∗ ( σ ~ t + 1 a ) ≤ J L f , ∗ ( σ ~ t a ) − ℓ 0∣ t α , ∗ + 2 Y ˉ b δ ( ε ˉ ) + b δ ( ε ˉ ) + λ g c g 2 ε ˉ + λ h b h 2 ε ˉ =: J L f , ∗ ( σ ~ t a ) − ℓ 0∣ t α , ∗ + α 1 ( ε ˉ ) , (106)
where α 1 ∈ K ∞ \alpha_1\in\mathcal{K}_\infty α 1 ∈ K ∞ .
Step IV. The applied augmented input satisfies u t a = u 0 ∣ t a , ∗ u_t^a=u_{0|t}^{a,*} u t a = u 0∣ t a , ∗ . Moreover, (97) and the sublevel bound imply the existence of α y ∈ K ∞ \alpha_y\in\mathcal{K}_\infty α y ∈ K ∞ such that
∥ y t ∥ Q 2 ≤ ∥ y ˉ 0 ∣ t ∗ ∥ Q 2 + α y ( ε ˉ ) . (107) \|y_t\|_Q^2\leq\|\bar{y}_{0|t}^*\|_Q^2+\alpha_y(\bar{\varepsilon}). \tag*{(107)} ∥ y t ∥ Q 2 ≤ ∥ y ˉ 0∣ t ∗ ∥ Q 2 + α y ( ε ˉ ) . (107)
Applying (73) and (107) gives
W a ( σ ~ t + 1 a ) ≤ W a ( σ ~ t a ) + ∥ u 0 ∣ t a , ∗ ∥ R 2 + ∥ y ˉ 0 ∣ t ∗ ∥ Q 2 − ϵ σ a ( σ ~ t a ) + α W ( ε ˉ ) + α y ( ε ˉ ) . (108) W^a(\tilde{\sigma}_{t+1}^a)\leq W^a(\tilde{\sigma}_t^a)+\|u_{0|t}^{a,*}\|_R^2+\|\bar{y}_{0|t}^*\|_Q^2
-\epsilon_\sigma^a(\tilde{\sigma}_t^a)+\alpha_W(\bar{\varepsilon})+\alpha_y(\bar{\varepsilon}). \tag*{(108)} W a ( σ ~ t + 1 a ) ≤ W a ( σ ~ t a ) + ∥ u 0∣ t a , ∗ ∥ R 2 + ∥ y ˉ 0∣ t ∗ ∥ Q 2 − ϵ σ a ( σ ~ t a ) + α W ( ε ˉ ) + α y ( ε ˉ ) . (108)
Adding (106) and (108) proves (75) with α 3 : = α 1 + α W + α y ∈ K ∞ \alpha_3:=\alpha_1+\alpha_W+\alpha_y\in\mathcal{K}_\infty α 3 := α 1 + α W + α y ∈ K ∞ . ■
Appendix D: Proof of Theorem 3
Proof. The proof follows the same finite-horizon argument as Theorem 2; hence, we only indicate the modifications caused by the perturbation terms in Lemma 4.
Step I. The lower bound in (76) follows directly from J L f , ∗ ≥ 0 J_L^{f,*}\geq0 J L f , ∗ ≥ 0 and Proposition 6. For the upper bound, Proposition 4 and Assumption 4 give a local feedback u a , l o c = K a ( ξ q , l o c − Π a ω ) u^{a,\mathrm{loc}}=K^a(\xi^{q,\mathrm{loc}}-\Pi^a\omega) u a , loc = K a ( ξ q , loc − Π a ω ) and constants δ l o c , γ s > 0 \delta_{\mathrm{loc}},\gamma_s>0 δ loc , γ s > 0 such that, whenever σ ∞ a ( ϱ t a ) ≤ δ l o c \sigma_\infty^a(\varrho_t^a)\leq\delta_{\mathrm{loc}} σ ∞ a ( ϱ t a ) ≤ δ loc , the resulting physical inputs belong to U \mathbb{U} U and ∑ j = 0 ∞ ℓ a ( ϱ j ∣ t a , l o c , u j ∣ t a , l o c ) ≤ γ s σ ∞ a ( ϱ t a ) \sum_{j=0}^{\infty}\ell^a(\varrho_{j|t}^{a,\mathrm{loc}},u_{j|t}^{a,\mathrm{loc}})\leq\gamma_s\sigma_\infty^a(\varrho_t^a) ∑ j = 0 ∞ ℓ a ( ϱ j ∣ t a , loc , u j ∣ t a , loc ) ≤ γ s σ ∞ a ( ϱ t a ) . This is the augmented counterpart of the local construction in the nominal proof.
Since ∥ ξ ~ t a − ξ t a ∥ ≤ p ι ε ˉ \lVert\tilde{\xi}^{a}_{t}-\xi^{a}_{t}\rVert\le\sqrt{p\iota\bar{\varepsilon}} ∥ ξ ~ t a − ξ t a ∥ ≤ p ι ε ˉ , we have σ a ( ϱ t a ) ≤ 2 σ a ( ϱ ~ t a ) + 2 p ι ε ˉ 2 \sigma^{a}(\varrho^{a}_{t}) \le2\sigma^{a}(\tilde{\varrho}^{a}_{t}) + 2p\iota\bar{\varepsilon}^{2} σ a ( ϱ t a ) ≤ 2 σ a ( ϱ ~ t a ) + 2 p ι ε ˉ 2 . Hence, by choosing δ a < δ l o c / 2 \delta_{a}<\delta_{\mathrm{loc}}/2 δ a < δ loc /2 and the admissible noise bound sufficiently small, σ a ( ϱ ~ t a ) ≤ δ a \sigma^{a}(\tilde{\varrho}^{a}_{t})\le\delta_{a} σ a ( ϱ ~ t a ) ≤ δ a implies σ a ( ϱ t a ) ≤ δ l o c \sigma^{a}(\varrho^{a}_{t})\le\delta_{\mathrm{loc}} σ a ( ϱ t a ) ≤ δ loc , so the same local construction remains admissible. The trajectory parametrization in Theorem 1, together with the right-inverse and noise-compensating slack construction in the proof of Lemma 4, provides a feasible robust DDPC candidate for every finite segment of this local trajectory. Moreover, the preceding bound on σ a ( ϱ t a ) \sigma^{a}(\varrho^{a}_{t}) σ a ( ϱ t a ) and Proposition 5 imply that the state perturbation and the additional coefficient and slack regularization terms can be collected into a function in K ∞ \mathcal{K}_{\infty} K ∞ . Therefore, for some γ l o c > 0 \gamma_{\mathrm{loc}}>0 γ loc > 0 and α L ∈ K ∞ \alpha_{L}\in\mathcal{K}_{\infty} α L ∈ K ∞ , it follows that Y L f ( ϱ t a ) ≤ γ l o c σ a ( ϱ t a ) + α L ( ε ˉ ) Y^{f}_{L}(\varrho^{a}_{t})\le\gamma_{\mathrm{loc}}\sigma^{a}(\varrho^{a}_{t})+\alpha_{L}(\bar{\varepsilon}) Y L f ( ϱ t a ) ≤ γ loc σ a ( ϱ t a ) + α L ( ε ˉ ) , where σ a ( ϱ t a ) ≤ δ a \sigma^{a}(\varrho^{a}_{t})\le\delta_{a} σ a ( ϱ t a ) ≤ δ a . Combining this with Y L f ≤ Y ˉ Y^{f}_{L}\le\bar{Y} Y L f ≤ Y ˉ for σ a ( ϱ t a ) > δ a \sigma^{a}(\varrho^{a}_{t})>\delta_{a} σ a ( ϱ t a ) > δ a proves (76) with γ ˉ Y : = max { γ l o c , Y ˉ / δ a } \bar{\gamma}_{Y}:=\max\{\gamma_{\mathrm{loc}},\bar{Y}/\delta_{a}\} γ ˉ Y := max { γ loc , Y ˉ / δ a } .
Step II. Let the superscript c c c denote the shortened candidate constructed in the proof of Lemma 4, and set Y L f , c : = J L − 1 f , c + W a ( ϱ ~ t + 1 a ) Y^{f,c}_{L}:=J^{f,c}_{L-1}+W^{a}(\tilde{\varrho}^{a}_{t+1}) Y L f , c := J L − 1 f , c + W a ( ϱ ~ t + 1 a ) . The estimates used there hold for this candidate before minimization; thus,
Y L f , c ( ϱ ~ t + 1 a ) ≤ Y L f ( ϱ ~ t a ) − ϵ o σ a ( ϱ ~ t a ) + α 3 ( ε ˉ ) . (109) Y^{f,c}_{L}(\tilde{\varrho}^{a}_{t+1})\le Y^{f}_{L}(\tilde{\varrho}^{a}_{t})-\epsilon_{o}\sigma^{a}(\tilde{\varrho}^{a}_{t})+\alpha_{3}(\bar{\varepsilon}). \tag*{(109)} Y L f , c ( ϱ ~ t + 1 a ) ≤ Y L f ( ϱ ~ t a ) − ϵ o σ a ( ϱ ~ t a ) + α 3 ( ε ˉ ) . (109)
Let { ϱ ˉ k ∣ t + 1 a , c } k = 0 L − 1 \{\bar{\varrho}^{a,c}_{k|t+1}\}_{k=0}^{L-1} { ϱ ˉ k ∣ t + 1 a , c } k = 0 L − 1 denote the nominal augmented trajectory associated with this candidate. Applying (72) along this trajectory and summing from k = 0 k=0 k = 0 to L − 2 L-2 L − 2 gives
ϵ o ∑ k = 0 L − 2 σ a ( ϱ ˉ k ∣ t + 1 a , c ) ≤ W a ( ϱ t + 1 a ) + ∑ k = 0 L − 2 ℓ a ( ϱ ˉ k ∣ t + 1 a , c , u ˉ k ∣ t + 1 a , c ) , (110) \epsilon_{o}\sum_{k=0}^{L-2}\sigma^{a}(\bar{\varrho}^{a,c}_{k|t+1})\le W^{a}(\varrho^{a}_{t+1})+\sum_{k=0}^{L-2}\ell^{a}(\bar{\varrho}^{a,c}_{k|t+1},\bar{u}^{a,c}_{k|t+1}), \tag*{(110)} ϵ o k = 0 ∑ L − 2 σ a ( ϱ ˉ k ∣ t + 1 a , c ) ≤ W a ( ϱ t + 1 a ) + k = 0 ∑ L − 2 ℓ a ( ϱ ˉ k ∣ t + 1 a , c , u ˉ k ∣ t + 1 a , c ) , (110)
where the nonnegative terminal storage has been dropped.
On the sublevel set under consideration, quadratic continuity of W a W^{a} W a and ∥ ξ ~ t + 1 a − ξ t + 1 a ∥ ≤ p ι ε ˉ \lVert\tilde{\xi}^{a}_{t+1}-\xi^{a}_{t+1}\rVert\le\sqrt{p\iota\bar{\varepsilon}} ∥ ξ ~ t + 1 a − ξ t + 1 a ∥ ≤ p ι ε ˉ imply W a ( ϱ t + 1 a ) ≤ W a ( ϱ ~ t + 1 a ) + α 5 ( ε ˉ ) W^{a}(\varrho^{a}_{t+1})\le W^{a}(\tilde{\varrho}^{a}_{t+1})+\alpha_{5}(\bar{\varepsilon}) W a ( ϱ t + 1 a ) ≤ W a ( ϱ ~ t + 1 a ) + α 5 ( ε ˉ ) for some α 5 ∈ K ∞ \alpha_{5}\in\mathcal{K}_{\infty} α 5 ∈ K ∞ . Since the regularization terms are nonnegative, (109) , (110) , and the upper bound established in Step I yield
ϵ o ∑ k = 0 L − 2 σ a ( ϱ ˉ k ∣ t + 1 a , c ) ≤ c Y ˉ σ a ( ϱ ~ t a ) + α 6 ( ε ˉ ) , (111) \epsilon_{o}\sum_{k=0}^{L-2}\sigma^{a}(\bar{\varrho}^{a,c}_{k|t+1})\le c_{\bar{Y}}\sigma^{a}(\tilde{\varrho}^{a}_{t})+\alpha_{6}(\bar{\varepsilon}), \tag*{(111)} ϵ o k = 0 ∑ L − 2 σ a ( ϱ ˉ k ∣ t + 1 a , c ) ≤ c Y ˉ σ a ( ϱ ~ t a ) + α 6 ( ε ˉ ) , (111)
where c Y ˉ : = max { γ ˉ Y − ϵ o , 0 } c_{\bar{Y}}:=\max\{\bar{\gamma}_{Y}-\epsilon_{o},0\} c Y ˉ := max { γ ˉ Y − ϵ o , 0 } and α 6 : = α L + α 3 + α 5 \alpha_{6}:=\alpha_{L}+\alpha_{3}+\alpha_{5} α 6 := α L + α 3 + α 5 .
Hence, by the same averaging argument as in Step II of the proof of Theorem 2, there exists k x ∈ { 0 , … , L − 2 } k_{x}\in\{0,\ldots,L-2\} k x ∈ { 0 , … , L − 2 } such that
σ a ( ϱ ˉ k x ∣ t + 1 a , c ) ≤ c Y ˉ ϵ o ( L − 1 ) σ a ( ϱ ~ t a ) + α 6 ( ε ˉ ) ϵ o ( L − 1 ) . (112) \sigma^{a}(\bar{\varrho}^{a,c}_{k_{x}|t+1})\le\frac{c_{\bar{Y}}}{\epsilon_{o}(L-1)}\sigma^{a}(\tilde{\varrho}^{a}_{t})+\frac{\alpha_{6}(\bar{\varepsilon})}{\epsilon_{o}(L-1)}. \tag*{(112)} σ a ( ϱ ˉ k x ∣ t + 1 a , c ) ≤ ϵ o ( L − 1 ) c Y ˉ σ a ( ϱ ~ t a ) + ϵ o ( L − 1 ) α 6 ( ε ˉ ) . (112)
The same argument using only Y L f ( ϱ ~ t a ) ≤ Y ˉ Y^{f}_{L}(\tilde{\varrho}^{a}_{t})\le\bar{Y} Y L f ( ϱ ~ t a ) ≤ Y ˉ gives σ a ( ϱ ˉ k x ∣ t + 1 a , c ) ≤ Y ˉ + α 3 ( ε ˉ ) + α 5 ( ε ˉ ) ϵ o ( L − 1 ) \sigma^{a}(\bar{\varrho}^{a,c}_{k_{x}|t+1})\le\frac{\bar{Y}+\alpha_{3}(\bar{\varepsilon})+\alpha_{5}(\bar{\varepsilon})}{\epsilon_{o}(L-1)} σ a ( ϱ ˉ k x ∣ t + 1 a , c ) ≤ ϵ o ( L − 1 ) Y ˉ + α 3 ( ε ˉ ) + α 5 ( ε ˉ ) . Therefore, if L > 1 + 2 Y ˉ ϵ o δ a L>1+\frac{2\bar{Y}}{\epsilon_{o}\delta_{a}} L > 1 + ϵ o δ a 2 Y ˉ and ε ˉ \bar{\varepsilon} ε ˉ is sufficiently small such that α 3 ( ε ˉ ) + α 5 ( ε ˉ ) ≤ Y ˉ \alpha_{3}(\bar{\varepsilon})+\alpha_{5}(\bar{\varepsilon})\le\bar{Y} α 3 ( ε ˉ ) + α 5 ( ε ˉ ) ≤ Y ˉ , then σ a ( ϱ ˉ k x ∣ t + 1 a , c ) ≤ δ a \sigma^{a}(\bar{\varrho}^{a,c}_{k_{x}|t+1})\le\delta_{a} σ a ( ϱ ˉ k x ∣ t + 1 a , c ) ≤ δ a . The local controller from Step I can thus be appended at k x k_{x} k x while preserving the physical-input constraint.
Step III. Append the local feedback at k x k_{x} k x to the first k x k_{x} k x steps of the shortened candidate and propagate (28). This gives a length-L L L augmented trajectory whose physical inputs lie in U \mathbb{U} U . Persistence of excitation provides a coefficient g ˉ t + 1 f \bar{g}^{f}_{t+1} g ˉ t + 1 f for this concatenated nominal trajectory at depth L + ι L+\iota L + ι . Choose y ˉ [ t + 1 − ι , t + 1 + L − 1 ] ∣ t + 1 f = H L + ι ( y ) g ˉ t + 1 f + [ ε [ t + 1 − ι , t ] 0 p o L × 1 ] \bar{y}^{f}_{[t+1-\iota,t+1+L-1]|t+1}=H_{L+\iota}(y)\bar{g}^{f}_{t+1}+\begin{bmatrix}\varepsilon_{[t+1-\iota,t]}\\0_{p_{o}L\times1}\end{bmatrix} y ˉ [ t + 1 − ι , t + 1 + L − 1 ] ∣ t + 1 f = H L + ι ( y ) g ˉ t + 1 f + [ ε [ t + 1 − ι , t ] 0 p o L × 1 ] and h ˉ t + 1 f = H L + ι ( ε ) g ˉ t + 1 f − [ ε [ t + 1 − ι , t ] 0 p o L × 1 ] \bar{h}^{f}_{t+1}=H_{L+\iota}(\varepsilon)\bar{g}^{f}_{t+1}-\begin{bmatrix}\varepsilon_{[t+1-\iota,t]}\\0_{p_{o}L\times1}\end{bmatrix} h ˉ t + 1 f = H L + ι ( ε ) g ˉ t + 1 f − [ ε [ t + 1 − ι , t ] 0 p o L × 1 ] . Then the constraints (64) and (65) hold.
The same right-inverse estimates as in the proof of Lemma 4, now applied at depth L + ι L+\iota L + ι , together with Proposition 5, which controls g t + 1 c − g ˉ t + 1 f g^{c}_{t+1}-\bar{g}^{f}_{t+1} g t + 1 c − g ˉ t + 1 f , give constants c g f , c h f > 0 c^{f}_{g},c^{f}_{h}>0 c g f , c h f > 0 independent of ε ˉ \bar{\varepsilon} ε ˉ such that
∥ g ˉ t + 1 f − g t + 1 f ∥ ≤ c g f , ∥ h ˉ t + 1 f ∥ ≤ ε ˉ c h f . (113) \lVert\bar{g}^{f}_{t+1}-g^{f}_{t+1}\rVert\le c^{f}_{g},\qquad\lVert\bar{h}^{f}_{t+1}\rVert\le\bar{\varepsilon}c^{f}_{h}. \tag*{(113)} ∥ g ˉ t + 1 f − g t + 1 f ∥ ≤ c g f , ∥ h ˉ t + 1 f ∥ ≤ ε ˉ c h f . (113)
The augmented input is shifted exactly, and the output mismatch over the prefix is bounded as in the proof of Lemma 4, whereas the local-tail cost is bounded by γ s \gamma_{s} γ s times the estimate in Step II. Together with the regularization bounds and optimality, this yields some α J ∈ K ∞ \alpha_{J}\in\mathcal{K}_{\infty} α J ∈ K ∞ satisfying
J L f , ∗ ( ϱ ~ t + 1 a ) ≤ J L f , ∗ ( ϱ ~ t a ) − ℓ 0 ∣ t a , ∗ + γ s c Y ˉ ϵ o ( L − 1 ) σ a ( ϱ ~ t a ) + α J ( ε ˉ ) . (114) J^{f,*}_{L}(\tilde{\varrho}^{a}_{t+1})\le J^{f,*}_{L}(\tilde{\varrho}^{a}_{t})-\ell^{a,*}_{0|t}+\frac{\gamma_{s}c_{\bar{Y}}}{\epsilon_{o}(L-1)}\sigma^{a}(\tilde{\varrho}^{a}_{t})+\alpha_{J}(\bar{\varepsilon}). \tag*{(114)} J L f , ∗ ( ϱ ~ t + 1 a ) ≤ J L f , ∗ ( ϱ ~ t a ) − ℓ 0∣ t a , ∗ + ϵ o ( L − 1 ) γ s c Y ˉ σ a ( ϱ ~ t a ) + α J ( ε ˉ ) . (114)
Here α J \alpha_{J} α J collects the prefix perturbation, the term γ s α 6 ( ε ˉ ) / [ ϵ o ( L − 1 ) ] \gamma_{s}\alpha_{6}(\bar{\varepsilon})/[\epsilon_{o}(L-1)] γ s α 6 ( ε ˉ ) / [ ϵ o ( L − 1 )] , and the coefficient and slack penalties λ g ε ( c g f ) 2 \lambda_{g}\varepsilon(c^{f}_{g})^{2} λ g ε ( c g f ) 2 and λ h ε ( c h f ) 2 \lambda_{h}\varepsilon(c^{f}_{h})^{2} λ h ε ( c h f ) 2 .
Adding the perturbed storage estimate (108) proves (77) with α ˉ L : = ϵ o − γ s c Y ˉ ϵ o ( L − 1 ) > 0 \bar{\alpha}_{L}:=\epsilon_{o}-\frac{\gamma_{s}c_{\bar{Y}}}{\epsilon_{o}(L-1)}>0 α ˉ L := ϵ o − ϵ o ( L − 1 ) γ s c Y ˉ > 0 and α G : = α J + α W + α y ∈ K ∞ \alpha_{\mathcal{G}}:=\alpha_{J}+\alpha_{W}+\alpha_{y}\in\mathcal{K}_{\infty} α G := α J + α W + α y ∈ K ∞ , where positivity holds for sufficiently large L L L .
It remains to account for the additive perturbations when proving invariance. Choose L > L ˉ Y L>\bar{L}_{Y} L > L ˉ Y large enough for both the local-entry condition in Step II and α ˉ L > 0 \bar{\alpha}_{L}>0 α ˉ L > 0 , and then choose ε ^ L \hat{\varepsilon}_{L} ε ^ L such that, for ε ˉ ≤ ε ^ L \bar{\varepsilon}\le\hat{\varepsilon}_{L} ε ˉ ≤ ε ^ L , α Y ( ε ˉ ) ≤ Y ˉ \alpha_{Y}(\bar{\varepsilon})\le\bar{Y} α Y ( ε ˉ ) ≤ Y ˉ and α G ( ε ˉ ) ≤ α ˉ L γ ˉ Y ( Y ˉ − α L ( ε ˉ ) ) \alpha_{G}(\bar{\varepsilon})\le\frac{\bar{\alpha}_{L}}{\bar{\gamma}_{Y}}(\bar{Y}-\alpha_{L}(\bar{\varepsilon})) α G ( ε ˉ ) ≤ γ ˉ Y α ˉ L ( Y ˉ − α L ( ε ˉ )) . If ε ˉ ≤ ε ^ L \bar{\varepsilon}\le\hat{\varepsilon}_{L} ε ˉ ≤ ε ^ L , invariance follows directly from (77) . Otherwise, the upper bound in (76) gives σ a ( ϱ ~ t a ) ≥ Y ˉ − α G ( ε ˉ ) − α L ( ε ˉ ) γ ˉ Y \sigma^{a}(\tilde{\varrho}^{a}_{t})\ge\frac{\bar{Y}-\alpha_{G}(\bar{\varepsilon})-\alpha_{L}(\bar{\varepsilon})}{\bar{\gamma}_{Y}} σ a ( ϱ ~ t a ) ≥ γ ˉ Y Y ˉ − α G ( ε ˉ ) − α L ( ε ˉ ) . Hence, by the chosen noise bound, the decrease term dominates α G ( ε ˉ ) \alpha_{G}(\bar{\varepsilon}) α G ( ε ˉ ) and Y L f ( ϱ ~ t + 1 a ) ≤ Y L f ( ϱ ~ t a ) ≤ Y ˉ Y^{f}_{L}(\tilde{\varrho}^{a}_{t+1})\le Y^{f}_{L}(\tilde{\varrho}^{a}_{t})\le\bar{Y} Y L f ( ϱ ~ t + 1 a ) ≤ Y L f ( ϱ ~ t a ) ≤ Y ˉ . Thus, Y L f ≤ Y ˉ Y^{f}_{L}\le\bar{Y} Y L f ≤ Y ˉ is forward invariant, and recursive feasibility above proves recursive feasibility and u t ∈ U u_{t}\in\mathbb{U} u t ∈ U by induction.
Finally, (41) gives practical stability with an ultimate neighborhood that vanishes as ε ˉ → 0 \bar{\varepsilon}\to0 ε ˉ → 0 , and ∥ ξ ~ t − Π E ω t ∥ 2 ≤ σ a ( ϱ ~ t a ) \lVert\tilde{\xi}_{t}-\Pi_{\mathbb{E}}\omega_{t}\rVert^{2}\le\sigma^{a}(\tilde{\varrho}^{a}_{t}) ∥ ξ ~ t − Π E ω t ∥ 2 ≤ σ a ( ϱ ~ t a ) transfers this property to the original regularization error. Lemma 3 transfers these conclusions from (36) to (30), which completes the proof.