Introduction
We consider the optimal control of switched linear systems (SLS) of the form
Σ σ { E σ ( t ) x ˙ ( t ) = A σ ( t ) x ( t ) + B σ ( t ) u ( t ) , x ( 0 ) = 0 , x ( t + ) = K σ ( t + ) , σ ( t − ) x ( t − ) , y ( t ) = C σ ( t ) x ( t ) , (1) \Sigma_{\sigma}
\left\{
\begin{aligned}
E_{\sigma(t)}\dot{x}(t) &= A_{\sigma(t)}x(t) + B_{\sigma(t)}u(t), \quad x(0)=0,\\
x(t^+) &= K_{\sigma(t^+),\sigma(t^-)}x(t^-),\\
y(t) &= C_{\sigma(t)}x(t),
\end{aligned}
\right. \tag*{(1)} Σ σ ⎩ ⎨ ⎧ E σ ( t ) x ˙ ( t ) x ( t + ) y ( t ) = A σ ( t ) x ( t ) + B σ ( t ) u ( t ) , x ( 0 ) = 0 , = K σ ( t + ) , σ ( t − ) x ( t − ) , = C σ ( t ) x ( t ) , (1)
where σ : R → K : = { 1 , … , L } \sigma:\mathbb{R}\to\mathcal{K}:=\{1,\ldots,L\} σ : R → K := { 1 , … , L } is the external switching signal, which we assume to be an element of the set
P : = { σ : R → K ∣ σ is left-continuous with locally finitely many jumps } . (2) \mathcal{P}:=\{\sigma:\mathbb{R}\to\mathcal{K}\mid\sigma\text{ is left-continuous with locally finitely many jumps}\}. \tag*{(2)} P := { σ : R → K ∣ σ is left-continuous with locally finitely many jumps } . (2)
The symbols x ( t ) ∈ R n x(t)\in\mathbb{R}^{n} x ( t ) ∈ R n , u ( t ) ∈ R m u(t)\in\mathbb{R}^{m} u ( t ) ∈ R m , and y ( t ) ∈ R p y(t)\in\mathbb{R}^{p} y ( t ) ∈ R p denote the state, control (or input ), and measured output , respectively. We use x ( u ) x(u) x ( u ) and y ( u ) y(u) y ( u ) to emphasize their dependence on the control u ∈ L 2 ( 0 , T ; R m ) u\in L^{2}(0,T;\mathbb{R}^{m}) u ∈ L 2 ( 0 , T ; R m ) with terminal time T > 0 T>0 T > 0 . For each mode j ∈ K j\in\mathcal{K} j ∈ K , the matrices E j , A j ∈ R n × n E_j,A_j\in\mathbb{R}^{n\times n} E j , A j ∈ R n × n , B j ∈ R n × m B_j\in\mathbb{R}^{n\times m} B j ∈ R n × m , and C j ∈ R p × n C_j\in\mathbb{R}^{p\times n} C j ∈ R p × n define the active subsystem and E j E_j E j is assumed to be invertible. The matrices K i , j ∈ R n × n K_{i,j}\in\mathbb{R}^{n\times n} K i , j ∈ R n × n describe state transitions from mode j j j to mode i i i at the switching times, with K i , i = I n K_{i,i}=I_n K i , i = I n . As defined in § 1.3 , σ ( t − ) \sigma(t^-) σ ( t − ) and σ ( t + ) \sigma(t^+) σ ( t + ) denote the left and right limits, respectively. We refer to (1.1) as the full-order model (FOM). For each fixed σ ∈ P \sigma\in\mathcal{P} σ ∈ P , let 0 = t 0 < ⋯ < t N + 1 = T 0 = t_0 < \cdots< t_{N+1} = T 0 = t 0 < ⋯ < t N + 1 = T be the partition induced by the switching times, i.e., the discontinuities of σ \sigma σ in the interval ( 0 , T ) (0,T) ( 0 , T ) . The state and output satisfy x ∣ ( t k , t k + 1 ) ∈ H 1 ( t k , t k + 1 ; R n ) ↪ C ( [ t k , t k + 1 ] ; R n ) \mathbf{x}|_{(t_k,t_{k+1})} \in H^1(t_k,t_{k+1};\mathbb{R}^n) \hookrightarrow C([t_k,t_{k+1}];\mathbb{R}^n) x ∣ ( t k , t k + 1 ) ∈ H 1 ( t k , t k + 1 ; R n ) ↪ C ([ t k , t k + 1 ] ; R n ) and y ∣ ( t k , t k + 1 ) ∈ C ( [ t k , t k + 1 ] ; R p ) y|_{(t_k,t_{k+1})} \in C([t_k,t_{k+1}];\mathbb{R}^p) y ∣ ( t k , t k + 1 ) ∈ C ([ t k , t k + 1 ] ; R p ) for k = 0 , … , N k = 0, \ldots, N k = 0 , … , N .
Given a cost parameter λ > 0 \lambda> 0 λ > 0 , we consider the finite-horizon optimal control of (1) with respect to the quadratic cost function
J ( y , u ) : = ∫ 0 T ( 1 2 ∥ y ( t ) − y d ( t ) ∥ R p 2 + λ 2 ∥ u ( t ) − u d ( t ) ∥ R m 2 ) d t . (3) J(\mathbf{y},\mathbf{u}) := \int_0^T \left(\frac{1}{2}\lVert y(t)-y_{\mathrm{d}}(t)\rVert_{\mathbb{R}^p}^2 + \frac{\lambda}{2}\lVert u(t)-u_{\mathrm{d}}(t)\rVert_{\mathbb{R}^m}^2\right)\,\mathrm{d}t. \tag*{(3)} J ( y , u ) := ∫ 0 T ( 2 1 ∥ y ( t ) − y d ( t ) ∥ R p 2 + 2 λ ∥ u ( t ) − u d ( t ) ∥ R m 2 ) d t . (3)
where u d ∈ L 2 ( 0 , T ; R m ) u_{\mathrm{d}} \in L^2(0,T;\mathbb{R}^m) u d ∈ L 2 ( 0 , T ; R m ) and y d ∈ L 2 ( 0 , T ; R p ) y_{\mathrm{d}} \in L^2(0,T;\mathbb{R}^p) y d ∈ L 2 ( 0 , T ; R p ) are prescribed references. Let U a d ⊆ R m U_{\mathrm{ad}} \subseteq\mathbb{R}^m U ad ⊆ R m and Y a d ⊆ R p Y_{\mathrm{ad}} \subseteq\mathbb{R}^p Y ad ⊆ R p be nonempty, closed, convex sets of admissible inputs and outputs, respectively, and define U a d : = L 2 ( 0 , T ; U a d ) \mathcal{U}_{\mathrm{ad}} := L^2(0,T;U_{\mathrm{ad}}) U ad := L 2 ( 0 , T ; U ad ) . The pointwise output-constrained full-order optimal control problem (FOM-OCP) is
min J ( y , u ) s.t. ( y , u ) solves x r e f ‘ ( 1.1 ) ‘ , u ∈ U a d , y ( t ) ∈ Y a d for all t ∈ [ 0 , T ] . (4) \min J(\mathbf{y},\mathbf{u}) \quad\text{s.t.}\quad(\mathbf{y},\mathbf{u})\text{ solves }{xref}`(1.1)`,\quad\mathbf{u}\in\mathcal{U}_{\mathrm{ad}},\quad y(t)\in Y_{\mathrm{ad}}\text{ for all }t\in[0,T]. \tag*{(4)} min J ( y , u ) s.t. ( y , u ) solves x r e f ‘ ( 1.1 ) ‘ , u ∈ U ad , y ( t ) ∈ Y ad for all t ∈ [ 0 , T ] . (4)
By piecewise continuity and closedness of Y a d Y_{\mathrm{ad}} Y ad , the output constraints in (4) are equivalent to y ( t ) ∈ Y a d y(t) \in Y_{\mathrm{ad}} y ( t ) ∈ Y ad for almost every t ∈ ( 0 , T ) t \in(0,T) t ∈ ( 0 , T ) , in particular, both one-sided limits at each interior switching time belong to Y a d Y_{\mathrm{ad}} Y ad . A natural setting for multiplier theory is the product of continuous-function spaces on the closed switching intervals, in which Lagrange multipliers consist of tuples of finite vector-valued measures (cf., e.g., [35, Sec. 6.2.1]). To obtain computable multipliers, one can consider a Moreau–Yosida (MY) regularization in L 2 L^2 L 2 (cf. [6 ] , [19, Sec. 2.7.2.1]). To this end, define the nonempty, closed, convex set
Y a d : = { v ∈ L 2 ( 0 , T ; R p ) ∣ v ( t ) ∈ Y a d for almost every t ∈ ( 0 , T ) } . \mathcal{Y}_{\mathrm{ad}} := \left\{\mathbf{v}\in L^2(0,T;\mathbb{R}^p)\mid v(t)\in Y_{\mathrm{ad}}\text{ for almost every }t\in(0,T)\right\}. Y ad := { v ∈ L 2 ( 0 , T ; R p ) ∣ v ( t ) ∈ Y ad for almost every t ∈ ( 0 , T ) } .
For a regularization parameter γ > 0 \gamma> 0 γ > 0 , the MY approximation is based on the metric projection
P Y a d : L 2 ( 0 , T ; R p ) → Y a d , P Y a d ( y ) : = arg min v ∈ Y a d ∥ y − v ∥ L 2 ( 0 , T ; R p ) . P_{\mathcal{Y}_{\mathrm{ad}}}:L^2(0,T;\mathbb{R}^p)\to\mathcal{Y}_{\mathrm{ad}},\qquad P_{\mathcal{Y}_{\mathrm{ad}}}(\mathbf{y}):=\operatorname*{arg\,min}_{\mathbf{v}\in\mathcal{Y}_{\mathrm{ad}}}\lVert\mathbf{y}-\mathbf{v}\rVert_{L^2(0,T;\mathbb{R}^p)}. P Y ad : L 2 ( 0 , T ; R p ) → Y ad , P Y ad ( y ) := v ∈ Y ad arg min ∥ y − v ∥ L 2 ( 0 , T ; R p ) .
With the regularized cost functional J γ ( y , u ) : = J ( y , u ) + γ 2 ∥ y − P Y a d ( y ) ∥ L 2 ( 0 , T ; R p ) 2 J_\gamma(\mathbf{y},\mathbf{u}):=J(\mathbf{y},\mathbf{u})+\frac{\gamma}{2}\lVert\mathbf{y}-P_{\mathcal{Y}_{\mathrm{ad}}}(\mathbf{y})\rVert_{L^2(0,T;\mathbb{R}^p)}^2 J γ ( y , u ) := J ( y , u ) + 2 γ ∥ y − P Y ad ( y ) ∥ L 2 ( 0 , T ; R p ) 2 we obtain the MY-regularized FOM-OCP
min J γ ( y , u ) s.t. ( y , u ) solves x r e f ‘ ( 1.1 ) ‘ and u ∈ U a d . (5) \min J_\gamma(\mathbf{y},\mathbf{u}) \quad\text{s.t.}\quad(\mathbf{y},\mathbf{u})\text{ solves }{xref}`(1.1)`\text{ and }\mathbf{u}\in\mathcal{U}_{\mathrm{ad}}. \tag*{(5)} min J γ ( y , u ) s.t. ( y , u ) solves x r e f ‘ ( 1.1 ) ‘ and u ∈ U ad . (5)
We denote the unique solution of (5) by ( y ⋆ , u ⋆ ) (\mathbf{y}^\star,\mathbf{u}^\star) ( y ⋆ , u ⋆ ) ; see Proposition 2.1. The convergence of solutions of (5) towards the solution of (4) as γ → ∞ \gamma\to\infty γ → ∞ is investigated, e.g., in [21]. Approximating the solution of (5) numerically requires repeated evaluations of (1) for different inputs u ∈ U a d \mathbf{u}\in\mathcal{U}_{\mathrm{ad}} u ∈ U ad . These evaluations become computationally expensive when n n n is large, particularly if the OCP must be solved for many configurations of the parameter tuple
μ : = ( σ , y d , u d , U a d , Y a d , λ , γ , T ) . (6) \mu:= (\sigma,y_{\mathrm{d}},u_{\mathrm{d}},U_{\mathrm{ad}},Y_{\mathrm{ad}},\lambda,\gamma,T). \tag*{(6)} μ := ( σ , y d , u d , U ad , Y ad , λ , γ , T ) . (6)
Our objective is, therefore, to develop a rapidly computable approximation of the OCP solution operator
G : μ ↦ ( y ⋆ ( μ ) , u ⋆ ( μ ) ) . (7) \mathcal{G}:\mu\mapsto(\mathbf{y}^\star(\mu),\mathbf{u}^\star(\mu)). \tag*{(7)} G : μ ↦ ( y ⋆ ( μ ) , u ⋆ ( μ )) . (7)
To construct this approximation, we apply projection-based model-order reduction (MOR) and replace (1) by the reduced-order model (ROM)
Σ ~ σ { E ~ σ ( t ) x ~ ˙ ( t ) = A ~ σ ( t ) x ~ ( t ) + B ~ σ ( t ) u ( t ) , x ~ ( 0 ) = 0 , x ~ ( t + ) = K ~ σ ( t + ) , σ ( t − ) x ~ ( t − ) , y ~ ( t ) = C ~ σ ( t ) x ~ ( t ) , (8) \widetilde{\Sigma}_{\sigma}\left\{\begin{aligned}
\widetilde{E}_{\sigma(t)}\dot{\widetilde{x}}(t) &= \widetilde{A}_{\sigma(t)}\widetilde{x}(t) + \widetilde{B}_{\sigma(t)}u(t), \qquad\widetilde{x}(0)=0,\\
\widetilde{x}(t^{+}) &= \widetilde{K}_{\sigma(t^{+}),\sigma(t^{-})}\widetilde{x}(t^{-}),\\
\widetilde{y}(t) &= \widetilde{C}_{\sigma(t)}\widetilde{x}(t),
\end{aligned}\right. \tag*{(8)} Σ σ ⎩ ⎨ ⎧ E σ ( t ) x ˙ ( t ) x ( t + ) y ( t ) = A σ ( t ) x ( t ) + B σ ( t ) u ( t ) , x ( 0 ) = 0 , = K σ ( t + ) , σ ( t − ) x ( t − ) , = C σ ( t ) x ( t ) , (8)
with the reduced matrices E ~ j , A ~ j , K ~ i , j ∈ R r × r \widetilde{E}_j,\widetilde{A}_j,\widetilde{K}_{i,j}\in\mathbb{R}^{r\times r} E j , A j , K i , j ∈ R r × r , B ~ j ∈ R r × m \widetilde{B}_j\in\mathbb{R}^{r\times m} B j ∈ R r × m , C ~ j ∈ R p × r \widetilde{C}_j\in\mathbb{R}^{p\times r} C j ∈ R p × r , and r ≪ n r\ll n r ≪ n . The reduced system matrices are obtained via Petrov–Galerkin projection. We choose full-column-rank matrices V j , W j ∈ R n × r V_j,W_j\in\mathbb{R}^{n\times r} V j , W j ∈ R n × r such that W j T E j V j W_j^{\mathsf{T}}E_jV_j W j T E j V j is invertible for each mode j j j , and define
E ~ j : = W j T E j V j , A ~ j : = W j T A j V j , B ~ j : = W j T B j , C ~ j : = C j V j , K ~ i , j : = E ~ i − 1 W i T E i K i , j V j . (9) \begin{aligned}
\widetilde{E}_j &:= W_j^{\mathsf{T}}E_jV_j, & \widetilde{A}_j &:= W_j^{\mathsf{T}}A_jV_j, & \widetilde{B}_j &:= W_j^{\mathsf{T}}B_j,\\
\widetilde{C}_j &:= C_jV_j, & \widetilde{K}_{i,j} &:= \widetilde{E}_i^{-1}W_i^{\mathsf{T}}E_iK_{i,j}V_j. \tag*{(9)}
\end{aligned} E j C j := W j T E j V j , := C j V j , A j K i , j := W j T A j V j , := E i − 1 W i T E i K i , j V j . B j := W j T B j , (9)
The transition formula follows by imposing W i T E i ( V i x ~ ( t + ) − K i , j V j x ~ ( t − ) ) = 0 W_i^{\mathsf{T}}E_i(V_i\widetilde{x}(t^{+})-K_{i,j}V_j\widetilde{x}(t^{-}))=0 W i T E i ( V i x ( t + ) − K i , j V j x ( t − )) = 0 at a switch from mode j j j to mode i i i . By construction, we obtain K ~ i , i = I r \widetilde{K}_{i,i}=I_r K i , i = I r . The MY-regularized ROM-OCP reads as
min J γ ( y ~ , u ) s.t. ( y ~ , u ) solves x r e f ‘ ( 1.8 ) ‘ and u ∈ U a d , (10) \min J_\gamma(\widetilde{y},u) \quad\text{s.t. }(\widetilde{y},u)\text{ solves }{xref}`(1.8)`\text{ and }u\in U_{\mathrm{ad}}, \tag*{(10)} min J γ ( y , u ) s.t. ( y , u ) solves x r e f ‘ ( 1.8 ) ‘ and u ∈ U ad , (10)
and, for each configuration μ \mu μ , (10) has the unique solution ( y ~ ⋆ ( μ ) , u ~ ⋆ ( μ ) ) (\widetilde{y}^{\star}(\mu),\widetilde{u}^{\star}(\mu)) ( y ⋆ ( μ ) , u ⋆ ( μ )) . We do not study convergence as γ → ∞ \gamma\to\infty γ → ∞ nor do we quantify the regularization error compared to (4) . Instead, for each prescribed γ > 0 \gamma>0 γ > 0 , we certify the MOR error between the FOM-OCP and the ROM-OCP for a fixed γ \gamma γ . Thus, the goal of certified MOR is to seek projection matrices W j W_j W j and V j V_j V j for which the ROM-OCP (10) approximates the FOM-OCP (5) accurately for all configurations μ \mu μ , and a rigorous, possibly online-efficient error bound Δ y ⋆ , u ⋆ ( y ~ ⋆ ( μ ) , u ~ ⋆ ( μ ) ) \Delta_{y^\star,u^\star}(\widetilde{y}^{\star}(\mu),\widetilde{u}^{\star}(\mu)) Δ y ⋆ , u ⋆ ( y ⋆ ( μ ) , u ⋆ ( μ )) satisfying
∥ ( y ~ ⋆ ( μ ) , u ~ ⋆ ( μ ) ) − ( y ⋆ ( μ ) , u ⋆ ( μ ) ) ∥ ≤ Δ y ⋆ , u ⋆ ( y ~ ⋆ ( μ ) , u ~ ⋆ ( μ ) ) for all μ , (11) \left\|(\widetilde{y}^{\star}(\mu),\widetilde{u}^{\star}(\mu))-(y^{\star}(\mu),u^{\star}(\mu))\right\|\leq\Delta_{y^\star,u^\star}(\widetilde{y}^{\star}(\mu),\widetilde{u}^{\star}(\mu))\qquad\text{for all }\mu, \tag*{(11)} ∥ ( y ⋆ ( μ ) , u ⋆ ( μ )) − ( y ⋆ ( μ ) , u ⋆ ( μ )) ∥ ≤ Δ y ⋆ , u ⋆ ( y ⋆ ( μ ) , u ⋆ ( μ )) for all μ , (11)
where ∥ ⋅ ∥ \|\cdot\| ∥ ⋅ ∥ denotes a suitable norm. By online-efficient , we mean that the computational cost required to evaluate the bound is independent of the full-order dimension n n n . Expressing the optimal input and output as functions of μ \mu μ highlights the requirement that μ \mu μ can be varied.
Systems of the form (1) arise in engineering applications that operate in several modes or regimes. Optimal control must ensure efficient, safe, and reliable operation while respecting input and output constraints. The combination of switching dynamics, constraints, and real-time requirements makes the resulting problems computationally demanding and motivates our study.
Literature review
MOR is a highly active research field that has seen tremendous development in recent years. For an introduction and overview, we refer to the monographs and collections [8 , 9 ] [17, 31]. Regarding MOR for SLS, one straightforward strategy is to construct a separate ROM for each subsystem [27, 28]. Because it ignores transitions between modes, however, this strategy can produce arbitrarily large output errors; see [33, Sec. 2]. Coupled linear matrix inequalities (LMIs) offer an alternative [37], but become computationally infeasible for large-scale systems. Related matrix-inequality techniques yield balancing-based ROMs with theoretical guarantees of quadratic stability and an associated error bound [29]. The approaches in [30, 33] instead reinterpret the switching signal as an input and reformulate the SLS as a non-switched system. The method in [33] embeds the dynamics into a larger linear time-invariant (LTI) system, whereas [30] uses a bilinear reformulation and the corresponding balanced-truncation method. Both approaches provide error guarantees, but only under restrictive assumptions. The piecewise balanced reduction (PBR) method for SLS introduced in [24], and subsequently extended to handle switched systems of linear differential-algebraic equations [23], relaxes these assumptions by constructing projection matrices from approximate solutions of suitably defined generalized Lyapunov equations (GLEs). It provides an a posteriori error estimator under general assumptions, although the efficiency of its evaluation depends on the given SLS. Because PBR is the specific MOR method we chose to efficiently approximate the OCP, we review it in more detail in the forthcoming § 3 . Interpolation-based alternatives for hybrid and switched systems are discussed in [7 ] [32].
In the context of optimal control, MOR is used, e.g., in [1 , 5 ] [15, 16, 20, 22, 26, 34, 36]. Balanced truncation (BT) and related balancing-based methods have been used for controller reduction and feedback design; see, e.g., [10 , 11 , 13 ] [25]. Certified reduced basis (RB) approaches to parametrized control-constrained OCPs are studied in, e.g., [4 , 12 ] . The contributions [3 ] [14] also discuss regularization-based approaches to state constraints and a posteriori error estimation with RB and proper orthogonal decomposition (POD) models, respectively.
Outline and main contributions
We combine output-based a posteriori error analysis with PBR to approximate the solution operator (7) for MY-regularized optimal control of SLS. Our results quantify the model-reduction error at a prescribed regularization parameter and consist of three contributions.
First, for piecewise Petrov–Galerkin ROMs, § 2 establishes a joint error bound for the optimal control, output, and approximate multiplier (Theorem 2.4 ) and two-sided bounds for the difference of the optimal value functions (Theorem 2.5 ). These estimates use errors in the forward and adjoint outputs, which allows them to be combined with input-output error bounds from balancing-based reduction. They require no full-order optimization; their direct evaluation uses forward and adjoint FOM simulations driven by reduced optimal quantities. We therefore call them full-order-evaluation bounds .
Second, building on the PBR method of [24], we prove that reduction commutes with the derivation of the optimality system (Theorem 3.3 ). For asymptotically stable modes and uniquely solvable GLEs (see Assumptions 3.1 and 3.2), the reachability and observability matrices of the forward system and its time-reversed adjoint are interchanged. The adjoint of the reduced forward system consequently coincides with the system obtained by reducing the full-order adjoint. Thus, the forward balancing construction also determines the reduced adjoint, without a separate balancing computation.
Third, Corollary 3.5 combines certified forward and adjoint PBR error estimates with the output-based bounds to certify the optimal control, output, approximate multiplier, and value function. § 3.4 uses these estimates to select the reduced dimension adaptively from a precomputed hierarchy. Online efficiency requires a certified offline–online decomposition of the estimator, e.g., (51) gives an explicit realization when the singular-value-tail bound (38) applies. When this bound holds uniformly over the bounded parameter set specified in Corollary 3.7 , that corollary proves uniform approximation of the solution operator. For fixed system matrices, the projection bases can thus be reused as the regularization parameters, cost weights, constraint sets, switching signal, and reference trajectories vary. Finally, Section 4 studies a synthetic switched system and a spatially discretized two-room heating problem. The numerical experiments display approximation errors, estimator sharpness, online computational costs, and computed control and output trajectories.
Notation
We define the left and right limits as σ ( t − ) : = lim s ↑ t σ ( s ) \sigma(t^{-}) := \lim_{s\uparrow t}\sigma(s) σ ( t − ) := lim s ↑ t σ ( s ) and σ ( t + ) : = lim s ↓ t σ ( s ) \sigma(t^{+}) := \lim_{s\downarrow t}\sigma(s) σ ( t + ) := lim s ↓ t σ ( s ) , if they exist. If it is clear from the context, we abbreviate, e.g., L 2 ( 0 , T ; R m ) L^{2}(0,T;\mathbb{R}^{m}) L 2 ( 0 , T ; R m ) or L 2 ( 0 , T ; R p ) L^{2}(0,T;\mathbb{R}^{p}) L 2 ( 0 , T ; R p ) as L 2 L^{2} L 2 . For x , p ∈ L 2 ( 0 , T ; R n ) \mathbf{x},\mathbf{p}\in L^{2}(0,T;\mathbb{R}^{n}) x , p ∈ L 2 ( 0 , T ; R n ) , we abbreviate, e.g., ⟨ A σ x , p ⟩ R n ( t ) : = ⟨ A σ ( t ) x ( t ) , p ( t ) ⟩ R n \langle A_{\sigma}\mathbf{x},\mathbf{p}\rangle_{\mathbb{R}^{n}}(t):=\langle A_{\sigma(t)}\mathbf{x}(t),\mathbf{p}(t)\rangle_{\mathbb{R}^{n}} ⟨ A σ x , p ⟩ R n ( t ) := ⟨ A σ ( t ) x ( t ) , p ( t ) ⟩ R n for almost all t ∈ ( 0 , T ) t\in(0,T) t ∈ ( 0 , T ) .
Error estimates for the optimal control, output, and value function
Throughout this section, we fix a configuration μ \mu μ as in (6) and compare the optimal solutions of the OCPs (5) and (10) . We first introduce the adjoint system and the regularized multiplier, then state the optimality conditions on which the error analysis is based. The gradient of the MY penalty term in (5) is the approximate multiplier ξ ( y ) = γ ( y − P Y a d ( y ) ) ∈ L 2 ( 0 , T ; R p ) \xi(\mathbf{y})=\gamma(\mathbf{y}-P_{\mathcal{Y}_{\mathrm{ad}}}(\mathbf{y}))\in L^{2}(0,T;\mathbb{R}^{p}) ξ ( y ) = γ ( y − P Y ad ( y )) ∈ L 2 ( 0 , T ; R p ) . For arbitrary y , ξ ∈ L 2 ( 0 , T ; R p ) \mathbf{y},\boldsymbol{\xi}\in L^{2}(0,T;\mathbb{R}^{p}) y , ξ ∈ L 2 ( 0 , T ; R p ) , we denote by p ( y , ξ ) \mathbf{p}(\mathbf{y},\boldsymbol{\xi}) p ( y , ξ ) and q ( y , ξ ) \mathbf{q}(\mathbf{y},\boldsymbol{\xi}) q ( y , ξ ) the state and output of
Σ σ T { − E σ ( t ) T p ˙ ( t ) = A σ ( t ) T p ( t ) + C σ ( t ) T ( y ( t ) − y d ( t ) + ξ ( t ) ) , p ( T ) = 0 , E σ ( t − ) T p ( t − ) = K σ ( t + ) , σ ( t − ) T E σ ( t + ) T p ( t + ) , q ( t ) = B σ ( t ) T p ( t ) . (12) \Sigma_{\sigma}^{\mathrm{T}}
\left\{
\begin{aligned}
-\mathrm{E}_{\sigma(t)}^{\mathrm{T}}\dot{\mathbf{p}}(t)&=\mathrm{A}_{\sigma(t)}^{\mathrm{T}}\mathbf{p}(t)+\mathrm{C}_{\sigma(t)}^{\mathrm{T}}(\mathbf{y}(t)-\mathbf{y}_{\mathrm{d}}(t)+\boldsymbol{\xi}(t)), && \mathbf{p}(T)=0,\\
\mathrm{E}_{\sigma(t^-)}^{\mathrm{T}}\mathbf{p}(t^-)&=\mathrm{K}_{\sigma(t^+),\sigma(t^-)}^{\mathrm{T}}\mathrm{E}_{\sigma(t^+)}^{\mathrm{T}}\mathbf{p}(t^+),\\
\mathbf{q}(t)&=\mathrm{B}_{\sigma(t)}^{\mathrm{T}}\mathbf{p}(t).
\end{aligned}
\right. \tag*{(12)} Σ σ T ⎩ ⎨ ⎧ − E σ ( t ) T p ˙ ( t ) E σ ( t − ) T p ( t − ) q ( t ) = A σ ( t ) T p ( t ) + C σ ( t ) T ( y ( t ) − y d ( t ) + ξ ( t )) , = K σ ( t + ) , σ ( t − ) T E σ ( t + ) T p ( t + ) , = B σ ( t ) T p ( t ) . p ( T ) = 0 , (12)
which we refer to as the adjoint system associated with (1) . As in [22], integration by parts on the switching intervals yields the terminal and transition conditions. The full and reduced optimality systems are stated in the next two propositions.
Proposition 2.1 (MY-regularized FOM-OCP optimality system). The MY-regularized FOM-OCP (5) admits a unique solution ( y ⋆ , u ⋆ ) (\mathbf{y}^{\star},\mathbf{u}^{\star}) ( y ⋆ , u ⋆ ) . Together with the optimal state x ⋆ \mathbf{x}^{\star} x ⋆ , optimal adjoint p ⋆ \mathbf{p}^{\star} p ⋆ , optimal adjoint output q ⋆ \mathbf{q}^{\star} q ⋆ , and the approximate Lagrange multiplier ξ ⋆ ∈ L 2 ( 0 , T ; R p ) \boldsymbol{\xi}^{\star}\in L^{2}(0,T;\mathbb{R}^{p}) ξ ⋆ ∈ L 2 ( 0 , T ; R p ) it satisfies
E σ ( t ) x ˙ ⋆ ( t ) = A σ ( t ) x ⋆ ( t ) + B σ ( t ) u ⋆ ( t ) , x ⋆ ( 0 ) = 0 , x ⋆ ( t + ) = K σ ( t + ) , σ ( t − ) x ⋆ ( t − ) , y ⋆ ( t ) = C σ ( t ) x ⋆ ( t ) , − E σ ( t ) T p ˙ ⋆ ( t ) = A σ ( t ) T p ⋆ ( t ) + C σ ( t ) T ( y ⋆ ( t ) − y d ( t ) + ξ ⋆ ( t ) ) , p ⋆ ( T ) = 0 , E σ ( t − ) T p ⋆ ( t − ) = K σ ( t + ) , σ ( t − ) T E σ ( t + ) T p ⋆ ( t + ) , q ⋆ ( t ) = B σ ( t ) T p ⋆ ( t ) , ξ ⋆ ( t ) = γ ( y ⋆ − P Y a d ( y ⋆ ) ) ( t ) , 0 ≤ ⟨ λ ( u ⋆ − u d ) + q ⋆ , v − u ⋆ ⟩ L 2 ( 0 , T ; R m ) ∀ v ∈ U a d . (13) \begin{aligned}
\mathrm{E}_{\sigma(t)}\dot{\mathbf{x}}^{\star}(t)&=\mathrm{A}_{\sigma(t)}\mathbf{x}^{\star}(t)+\mathrm{B}_{\sigma(t)}\mathbf{u}^{\star}(t), && \mathbf{x}^{\star}(0)=0,\\
\mathbf{x}^{\star}(t^+)&=\mathrm{K}_{\sigma(t^+),\sigma(t^-)}\mathbf{x}^{\star}(t^-),\\
\mathbf{y}^{\star}(t)&=\mathrm{C}_{\sigma(t)}\mathbf{x}^{\star}(t),\\
-\mathrm{E}_{\sigma(t)}^{\mathrm{T}}\dot{\mathbf{p}}^{\star}(t)&=\mathrm{A}_{\sigma(t)}^{\mathrm{T}}\mathbf{p}^{\star}(t)+\mathrm{C}_{\sigma(t)}^{\mathrm{T}}(\mathbf{y}^{\star}(t)-\mathbf{y}_{\mathrm{d}}(t)+\boldsymbol{\xi}^{\star}(t)), && \mathbf{p}^{\star}(T)=0,\\
\mathrm{E}_{\sigma(t^-)}^{\mathrm{T}}\mathbf{p}^{\star}(t^-)&=\mathrm{K}_{\sigma(t^+),\sigma(t^-)}^{\mathrm{T}}\mathrm{E}_{\sigma(t^+)}^{\mathrm{T}}\mathbf{p}^{\star}(t^+),\\
\mathbf{q}^{\star}(t)&=\mathrm{B}_{\sigma(t)}^{\mathrm{T}}\mathbf{p}^{\star}(t),\\
\boldsymbol{\xi}^{\star}(t)&=\gamma(\mathbf{y}^{\star}-P_{\mathcal{Y}_{\mathrm{ad}}}(\mathbf{y}^{\star}))(t),\\
0&\leq\langle\lambda(\mathbf{u}^{\star}-\mathbf{u}_{\mathrm{d}})+\mathbf{q}^{\star},\mathbf{v}-\mathbf{u}^{\star}\rangle_{L^{2}(0,T;\mathbb{R}^{m})}\quad\forall\mathbf{v}\in U_{\mathrm{ad}}. \tag*{(13)}
\end{aligned} E σ ( t ) x ˙ ⋆ ( t ) x ⋆ ( t + ) y ⋆ ( t ) − E σ ( t ) T p ˙ ⋆ ( t ) E σ ( t − ) T p ⋆ ( t − ) q ⋆ ( t ) ξ ⋆ ( t ) 0 = A σ ( t ) x ⋆ ( t ) + B σ ( t ) u ⋆ ( t ) , = K σ ( t + ) , σ ( t − ) x ⋆ ( t − ) , = C σ ( t ) x ⋆ ( t ) , = A σ ( t ) T p ⋆ ( t ) + C σ ( t ) T ( y ⋆ ( t ) − y d ( t ) + ξ ⋆ ( t )) , = K σ ( t + ) , σ ( t − ) T E σ ( t + ) T p ⋆ ( t + ) , = B σ ( t ) T p ⋆ ( t ) , = γ ( y ⋆ − P Y ad ( y ⋆ )) ( t ) , ≤ ⟨ λ ( u ⋆ − u d ) + q ⋆ , v − u ⋆ ⟩ L 2 ( 0 , T ; R m ) ∀ v ∈ U ad . x ⋆ ( 0 ) = 0 , p ⋆ ( T ) = 0 , (13)
Proposition 2.2 (ROM-OCP optimality system). The MY-regularized ROM-OCP (10) admits a unique solution ( y ~ , u ~ ) (\widetilde{\mathbf{y}},\widetilde{\mathbf{u}}) ( y , u ) . Together with the optimal state x ~ ⋆ \widetilde{\mathbf{x}}^{\star} x ⋆ , optimal adjoint p ~ ⋆ \widetilde{\mathbf{p}}^{\star} p ⋆ , optimal adjoint output q ~ ⋆ \widetilde{\mathbf{q}}^{\star} q ⋆ , and the approximate Lagrange multiplier
ξ ~ ⋆ ∈ L 2 ( 0 , T ; R p ) \widetilde{\xi}^{\star} \in L^2(0,T;\mathbb{R}^p) ξ ⋆ ∈ L 2 ( 0 , T ; R p ) it satisfies
E ~ σ ( t ) x ~ ˙ ⋆ ( t ) = A ~ σ ( t ) x ~ ⋆ ( t ) + B ~ σ ( t ) u ~ ⋆ ( t ) , x ~ ⋆ ( 0 ) = 0 , (14) \widetilde{E}_{\sigma(t)}\dot{\widetilde{x}}^{\star}(t)=\widetilde{A}_{\sigma(t)}\widetilde{x}^{\star}(t)+\widetilde{B}_{\sigma(t)}\widetilde{u}^{\star}(t),\qquad\widetilde{x}^{\star}(0)=0, \tag*{(14)} E σ ( t ) x ˙ ⋆ ( t ) = A σ ( t ) x ⋆ ( t ) + B σ ( t ) u ⋆ ( t ) , x ⋆ ( 0 ) = 0 , (14)
x ~ ⋆ ( t + ) = K ~ σ ( t + ) , σ ( t − ) x ~ ⋆ ( t − ) , (15) \widetilde{x}^{\star}(t^+)=\widetilde{K}_{\sigma(t^+),\sigma(t^-)}\widetilde{x}^{\star}(t^-), \tag*{(15)} x ⋆ ( t + ) = K σ ( t + ) , σ ( t − ) x ⋆ ( t − ) , (15)
y ~ ⋆ ( t ) = C ~ σ ( t ) x ~ ⋆ ( t ) , (16) \widetilde{y}^{\star}(t)=\widetilde{C}_{\sigma(t)}\widetilde{x}^{\star}(t), \tag*{(16)} y ⋆ ( t ) = C σ ( t ) x ⋆ ( t ) , (16)
− E ~ σ ( t ) T p ~ ˙ ⋆ ( t ) = A ~ σ ( t ) T p ~ ⋆ ( t ) + C ~ σ ( t ) T ( y ~ ⋆ ( t ) − y d ( t ) + ξ ~ ⋆ ( t ) ) , p ~ ⋆ ( T ) = 0 , (17) -\widetilde{E}_{\sigma(t)}^{\mathsf{T}}\dot{\widetilde{p}}^{\star}(t)=\widetilde{A}_{\sigma(t)}^{\mathsf{T}}\widetilde{p}^{\star}(t)+\widetilde{C}_{\sigma(t)}^{\mathsf{T}}\bigl(\widetilde{y}^{\star}(t)-y_{\mathrm{d}}(t)+\widetilde{\xi}^{\star}(t)\bigr),\qquad\widetilde{p}^{\star}(T)=0, \tag*{(17)} − E σ ( t ) T p ˙ ⋆ ( t ) = A σ ( t ) T p ⋆ ( t ) + C σ ( t ) T ( y ⋆ ( t ) − y d ( t ) + ξ ⋆ ( t ) ) , p ⋆ ( T ) = 0 , (17)
E ~ σ ( t − ) T p ~ ⋆ ( t − ) = K ~ σ ( t + ) , σ ( t − ) T E ~ σ ( t + ) T p ~ ⋆ ( t + ) , (18) \widetilde{E}_{\sigma(t^-)}^{\mathsf{T}}\widetilde{p}^{\star}(t^-)=\widetilde{K}_{\sigma(t^+),\sigma(t^-)}^{\mathsf{T}}\widetilde{E}_{\sigma(t^+)}^{\mathsf{T}}\widetilde{p}^{\star}(t^+), \tag*{(18)} E σ ( t − ) T p ⋆ ( t − ) = K σ ( t + ) , σ ( t − ) T E σ ( t + ) T p ⋆ ( t + ) , (18)
q ~ ⋆ ( t ) = B ~ σ ( t ) T p ~ ⋆ ( t ) , (19) \widetilde{q}^{\star}(t)=\widetilde{B}_{\sigma(t)}^{\mathsf{T}}\widetilde{p}^{\star}(t), \tag*{(19)} q ⋆ ( t ) = B σ ( t ) T p ⋆ ( t ) , (19)
ξ ~ ⋆ ( t ) = γ ( y ~ ⋆ − P Y a d ( y ~ ⋆ ) ) ( t ) , (20) \widetilde{\xi}^{\star}(t)=\gamma\bigl(\widetilde{y}^{\star}-P_{Y_{\mathrm{ad}}}(\widetilde{y}^{\star})\bigr)(t), \tag*{(20)} ξ ⋆ ( t ) = γ ( y ⋆ − P Y ad ( y ⋆ ) ) ( t ) , (20)
0 ≤ ⟨ λ ( u ~ ⋆ − u d ) + q ~ ⋆ , v − u ~ ⋆ ⟩ L 2 ( 0 , T ; R m ) ∀ v ∈ U a d . (21) 0\leq\langle\lambda(\widetilde{u}^{\star}-u_{\mathrm{d}})+\widetilde{q}^{\star},v-\widetilde{u}^{\star}\rangle_{L^2(0,T;\mathbb{R}^m)}\qquad\forall v\in U_{\mathrm{ad}}. \tag*{(21)} 0 ≤ ⟨ λ ( u ⋆ − u d ) + q ⋆ , v − u ⋆ ⟩ L 2 ( 0 , T ; R m ) ∀ v ∈ U ad . (21)
To establish an output-based a posteriori error estimator, we first collect properties of the regularized multiplier.
Proposition 2.3 (Properties of the approximate Lagrange multiplier). Let y ∈ L 2 ( 0 , T ; R p ) \mathbf{y}\in L^2(0,T;\mathbb{R}^p) y ∈ L 2 ( 0 , T ; R p ) and define ξ ( y ) = γ ( y − P Y a d ( y ) ) \boldsymbol{\xi}(\mathbf{y})=\gamma(\mathbf{y}-P_{Y_{\mathrm{ad}}}(\mathbf{y})) ξ ( y ) = γ ( y − P Y ad ( y )) . Then the following properties hold:
(i) ⟨ ξ ( y ) , v − P Y a d ( y ) ⟩ L 2 ≤ 0 \langle\boldsymbol{\xi}(\mathbf{y}),\mathbf{v}-P_{Y_{\mathrm{ad}}}(\mathbf{y})\rangle_{L^2}\leq0 ⟨ ξ ( y ) , v − P Y ad ( y ) ⟩ L 2 ≤ 0 for all v ∈ Y a d \mathbf{v}\in Y_{\mathrm{ad}} v ∈ Y ad .
(ii) ⟨ ξ ( y ) − ξ ( y ^ ) , P Y a d ( y ^ ) − P Y a d ( y ) ⟩ L 2 ≤ 0 \langle\boldsymbol{\xi}(\mathbf{y})-\boldsymbol{\xi}(\widehat{\mathbf{y}}),P_{Y_{\mathrm{ad}}}(\widehat{\mathbf{y}})-P_{Y_{\mathrm{ad}}}(\mathbf{y})\rangle_{L^2}\leq0 ⟨ ξ ( y ) − ξ ( y ) , P Y ad ( y ) − P Y ad ( y ) ⟩ L 2 ≤ 0 for all y , y ^ ∈ L 2 ( 0 , T ; R p ) \mathbf{y},\widehat{\mathbf{y}}\in L^2(0,T;\mathbb{R}^p) y , y ∈ L 2 ( 0 , T ; R p ) .
Proof. Property (i) follows immediately from the characterization of the metric projection and the definition of ξ \boldsymbol{\xi} ξ . To prove (ii), apply this characterization first to y \mathbf{y} y with v = P Y a d ( y ^ ) \mathbf{v}=P_{Y_{\mathrm{ad}}}(\widehat{\mathbf{y}}) v = P Y ad ( y ) and then to y ^ \widehat{\mathbf{y}} y with v = P Y a d ( y ) \mathbf{v}=P_{Y_{\mathrm{ad}}}(\mathbf{y}) v = P Y ad ( y ) . Combining the resulting inequalities using γ > 0 \gamma>0 γ > 0 yields (ii).
Now, we estimate the error in the optimal control and output.
Theorem 2.4 (Output-based a posteriori error estimate for u ⋆ , y ⋆ \mathbf{u}^{\star},\mathbf{y}^{\star} u ⋆ , y ⋆ ). Let ( y ⋆ , u ⋆ , ξ ⋆ ) (\mathbf{y}^{\star},\mathbf{u}^{\star},\boldsymbol{\xi}^{\star}) ( y ⋆ , u ⋆ , ξ ⋆ ) and ( y ~ ⋆ , u ~ ⋆ , ξ ~ ⋆ ) (\widetilde{\mathbf{y}}^{\star},\widetilde{\mathbf{u}}^{\star},\widetilde{\boldsymbol{\xi}}^{\star}) ( y ⋆ , u ⋆ , ξ ⋆ ) be the optimal output, input, and approximate Lagrange multiplier of the FOM-OCP (1.5) and the ROM-OCP (1.10), respectively. Define the error
e y , u ⋆ : = ( ∥ y ~ ⋆ − y ⋆ ∥ L 2 2 + λ ∥ u ~ ⋆ − u ⋆ ∥ L 2 2 + 1 γ ∥ ξ ~ ⋆ − ξ ⋆ ∥ L 2 2 ) 1 / 2 . (22) e_{\mathbf{y},\mathbf{u}}^{\star}:=\left(\|\widetilde{\mathbf{y}}^{\star}-\mathbf{y}^{\star}\|_{L^2}^2+\lambda\|\widetilde{\mathbf{u}}^{\star}-\mathbf{u}^{\star}\|_{L^2}^2+\frac{1}{\gamma}\|\widetilde{\boldsymbol{\xi}}^{\star}-\boldsymbol{\xi}^{\star}\|_{L^2}^2\right)^{1/2}. \tag*{(22)} e y , u ⋆ := ( ∥ y ⋆ − y ⋆ ∥ L 2 2 + λ ∥ u ⋆ − u ⋆ ∥ L 2 2 + γ 1 ∥ ξ ⋆ − ξ ⋆ ∥ L 2 2 ) 1/2 . (22) Then e y , u ⋆ ≤ Δ y , u ⋆ ( y ~ ⋆ , u ~ ⋆ , ξ ~ ⋆ ) e_{\mathbf{y},\mathbf{u}}^{\star}\leq\Delta_{\mathbf{y},\mathbf{u}}^{\star}(\widetilde{\mathbf{y}}^{\star},\widetilde{\mathbf{u}}^{\star},\widetilde{\boldsymbol{\xi}}^{\star}) e y , u ⋆ ≤ Δ y , u ⋆ ( y ⋆ , u ⋆ , ξ ⋆ ) with
Δ y , u ⋆ ( y ~ ⋆ , u ~ ⋆ , ξ ~ ⋆ ) : = ( 1 λ ∥ q ( y ~ ⋆ , ξ ~ ⋆ ) − q ~ ⋆ ∥ L 2 2 + ( 1 + γ ) ∥ y ( u ~ ⋆ ) − y ~ ⋆ ∥ L 2 2 ) 1 / 2 . (23) \Delta_{\mathbf{y},\mathbf{u}}^{\star}(\widetilde{\mathbf{y}}^{\star},\widetilde{\mathbf{u}}^{\star},\widetilde{\boldsymbol{\xi}}^{\star}):=\left(\frac{1}{\lambda}\|\mathbf{q}(\widetilde{\mathbf{y}}^{\star},\widetilde{\boldsymbol{\xi}}^{\star})-\widetilde{\mathbf{q}}^{\star}\|_{L^2}^2+(1+\gamma)\|\mathbf{y}(\widetilde{\mathbf{u}}^{\star})-\widetilde{\mathbf{y}}^{\star}\|_{L^2}^2\right)^{1/2}. \tag*{(23)} Δ y , u ⋆ ( y ⋆ , u ⋆ , ξ ⋆ ) := ( λ 1 ∥ q ( y ⋆ , ξ ⋆ ) − q ⋆ ∥ L 2 2 + ( 1 + γ ) ∥ y ( u ⋆ ) − y ⋆ ∥ L 2 2 ) 1/2 . (23) Proof. We adapt the proof from [22] to SLS with state-transition maps, Petrov–Galerkin ROMs, and MY-regularized output constraints. Choose v = u ~ ⋆ \mathbf{v}=\widetilde{\mathbf{u}}^{\star} v = u ⋆ in (2.2h) and v = u ⋆ \mathbf{v}=\mathbf{u}^{\star} v = u ⋆ in (2.3h). Adding the resulting inequalities gives
λ ∥ u ~ ⋆ − u ⋆ ∥ L 2 2 ≤ ⟨ q ~ ⋆ − q ⋆ , u ⋆ − u ~ ⋆ ⟩ L 2 . \lambda\|\widetilde{\mathbf{u}}^{\star}-\mathbf{u}^{\star}\|_{L^2}^2\leq\langle\widetilde{\mathbf{q}}^{\star}-\mathbf{q}^{\star},\mathbf{u}^{\star}-\widetilde{\mathbf{u}}^{\star}\rangle_{L^2}. λ ∥ u ⋆ − u ⋆ ∥ L 2 2 ≤ ⟨ q ⋆ − q ⋆ , u ⋆ − u ⋆ ⟩ L 2 . Adding and subtracting q ( y ~ ⋆ , ξ ~ ⋆ ) \mathbf{q}(\widetilde{\mathbf{y}}^{\star},\widetilde{\boldsymbol{\xi}}^{\star}) q ( y ⋆ , ξ ⋆ ) , which is the output of the FOM adjoint system (2.1) with inputs y ~ ⋆ , ξ ~ ⋆ \widetilde{\mathbf{y}}^{\star},\widetilde{\boldsymbol{\xi}}^{\star} y ⋆ , ξ ⋆ , leads to
λ ∥ u ~ ⋆ − u ⋆ ∥ L 2 2 ≤ ⟨ q ~ ⋆ − q ( y ~ ⋆ , ξ ~ ⋆ ) , u ⋆ − u ~ ⋆ ⟩ L 2 + ⟨ q ( y ~ ⋆ , ξ ~ ⋆ ) − q ⋆ , u ⋆ − u ~ ⋆ ⟩ L 2 . (24) \lambda\|\widetilde{\mathbf{u}}^{\star}-\mathbf{u}^{\star}\|_{L^2}^2\leq\langle\widetilde{\mathbf{q}}^{\star}-\mathbf{q}(\widetilde{\mathbf{y}}^{\star},\widetilde{\boldsymbol{\xi}}^{\star}),\mathbf{u}^{\star}-\widetilde{\mathbf{u}}^{\star}\rangle_{L^2}+\langle\mathbf{q}(\widetilde{\mathbf{y}}^{\star},\widetilde{\boldsymbol{\xi}}^{\star})-\mathbf{q}^{\star},\mathbf{u}^{\star}-\widetilde{\mathbf{u}}^{\star}\rangle_{L^2}. \tag*{(24)} λ ∥ u ⋆ − u ⋆ ∥ L 2 2 ≤ ⟨ q ⋆ − q ( y ⋆ , ξ ⋆ ) , u ⋆ − u ⋆ ⟩ L 2 + ⟨ q ( y ⋆ , ξ ⋆ ) − q ⋆ , u ⋆ − u ⋆ ⟩ L 2 . (24)
By the definition of the adjoint output, the second term on the right-hand side is given as
⟨ q ( y ~ ⋆ , ξ ~ ⋆ ) − q ⋆ , u ⋆ − u ~ ⋆ ⟩ L 2 = ∫ 0 T ⟨ p ( y ~ ⋆ , ξ ~ ⋆ ) − p ⋆ , B σ ( u ⋆ − u ~ ⋆ ) ⟩ R n ( t ) d t \left\langle q(\tilde{y}^{\star},\tilde{\xi}^{\star})-q^{\star},u^{\star}-\tilde{u}^{\star}\right\rangle_{L^2}
=\int_{0}^{T}\left\langle p(\tilde{y}^{\star},\tilde{\xi}^{\star})-p^{\star},B_{\sigma}(u^{\star}-\tilde{u}^{\star})\right\rangle_{\mathbb{R}^{n}}(t)\,\mathrm{d}t ⟨ q ( y ~ ⋆ , ξ ~ ⋆ ) − q ⋆ , u ⋆ − u ~ ⋆ ⟩ L 2 = ∫ 0 T ⟨ p ( y ~ ⋆ , ξ ~ ⋆ ) − p ⋆ , B σ ( u ⋆ − u ~ ⋆ ) ⟩ R n ( t ) d t
Using the FOM (1) with input u ⋆ − u ~ ⋆ u^{\star}-\tilde{u}^{\star} u ⋆ − u ~ ⋆ and integrating by parts on each switching interval leads to
⟨ q ( y ~ ⋆ , ξ ~ ⋆ ) − q ⋆ , u ⋆ − u ~ ⋆ ⟩ L 2 = ∫ 0 T ⟨ p ( y ~ ⋆ , ξ ~ ⋆ ) − p ⋆ , E σ d d t ( x ⋆ − x ( u ~ ⋆ ) ) − A σ ( x ⋆ − x ( u ~ ⋆ ) ) ⟩ R n ( t ) d t = ∫ 0 T ⟨ − E σ T d d t ( p ( y ~ ⋆ , ξ ~ ⋆ ) − p ⋆ ) − A σ T ( p ( y ~ ⋆ , ξ ~ ⋆ ) − p ⋆ ) , x ⋆ − x ( u ~ ⋆ ) ⟩ R n ( t ) d t + ∑ k = 0 N [ ⟨ E σ T ( p ( y ~ ⋆ , ξ ~ ⋆ ) − p ⋆ ) , x ⋆ − x ( u ~ ⋆ ) ⟩ R n ( t ) ] t = t k − t = t k + 1 − . (25) \begin{aligned}
\left\langle q(\tilde{y}^{\star},\tilde{\xi}^{\star})-q^{\star},u^{\star}-\tilde{u}^{\star}\right\rangle_{L^2}
={}&\int_{0}^{T}\left\langle p(\tilde{y}^{\star},\tilde{\xi}^{\star})-p^{\star},E_{\sigma}\frac{\mathrm{d}}{\mathrm{d}t}\left(x^{\star}-x(\tilde{u}^{\star})\right)-A_{\sigma}\left(x^{\star}-x(\tilde{u}^{\star})\right)\right\rangle_{\mathbb{R}^{n}}(t)\,\mathrm{d}t \\
={}&\int_{0}^{T}\left\langle-E_{\sigma}^{\mathsf{T}}\frac{\mathrm{d}}{\mathrm{d}t}\left(p(\tilde{y}^{\star},\tilde{\xi}^{\star})-p^{\star}\right)-A_{\sigma}^{\mathsf{T}}\left(p(\tilde{y}^{\star},\tilde{\xi}^{\star})-p^{\star}\right),x^{\star}-x(\tilde{u}^{\star})\right\rangle_{\mathbb{R}^{n}}(t)\,\mathrm{d}t \\
&+\sum_{k=0}^{N}\left[\left\langle E_{\sigma}^{\mathsf{T}}\left(p(\tilde{y}^{\star},\tilde{\xi}^{\star})-p^{\star}\right),x^{\star}-x(\tilde{u}^{\star})\right\rangle_{\mathbb{R}^{n}}(t)\right]_{t=t_{k}^{-}}^{t=t_{k+1}^{-}}. \tag*{(25)}
\end{aligned} ⟨ q ( y ~ ⋆ , ξ ~ ⋆ ) − q ⋆ , u ⋆ − u ~ ⋆ ⟩ L 2 = = ∫ 0 T ⟨ p ( y ~ ⋆ , ξ ~ ⋆ ) − p ⋆ , E σ d t d ( x ⋆ − x ( u ~ ⋆ ) ) − A σ ( x ⋆ − x ( u ~ ⋆ ) ) ⟩ R n ( t ) d t ∫ 0 T ⟨ − E σ T d t d ( p ( y ~ ⋆ , ξ ~ ⋆ ) − p ⋆ ) − A σ T ( p ( y ~ ⋆ , ξ ~ ⋆ ) − p ⋆ ) , x ⋆ − x ( u ~ ⋆ ) ⟩ R n ( t ) d t + k = 0 ∑ N [ ⟨ E σ T ( p ( y ~ ⋆ , ξ ~ ⋆ ) − p ⋆ ) , x ⋆ − x ( u ~ ⋆ ) ⟩ R n ( t ) ] t = t k − t = t k + 1 − . (25)
The final-time term vanishes by the terminal conditions in (2.1) and (2.2d), and the initial-time term vanishes since ( x ⋆ − x ( u ~ ⋆ ) ) ( 0 + ) = 0 \left(x^{\star}-x(\tilde{u}^{\star})\right)(0^{+})=0 ( x ⋆ − x ( u ~ ⋆ ) ) ( 0 + ) = 0 (see (1) ). The switching terms cancel as both full-order adjoint states satisfy the transition conditions in (12) and (2.2e), respectively. Applying the FOM adjoint equation (12) to p ( y ~ ⋆ , ξ ~ ⋆ ) − p ⋆ p(\tilde{y}^{\star},\tilde{\xi}^{\star})-p^{\star} p ( y ~ ⋆ , ξ ~ ⋆ ) − p ⋆ yields
⟨ q ( y ~ ⋆ , ξ ~ ⋆ ) − q ⋆ , u ⋆ − u ~ ⋆ ⟩ L 2 = ⟨ y ~ ⋆ − y ⋆ , y ⋆ − y ( u ~ ⋆ ) ⟩ L 2 2 + ⟨ ξ ~ ⋆ − ξ ⋆ , y ⋆ − y ( u ~ ⋆ ) ⟩ L 2 2 . (26) \left\langle q(\tilde{y}^{\star},\tilde{\xi}^{\star})-q^{\star},u^{\star}-\tilde{u}^{\star}\right\rangle_{L^2}
=\left\langle\tilde{y}^{\star}-y^{\star},y^{\star}-y(\tilde{u}^{\star})\right\rangle_{L^2}^{2}
+\left\langle\tilde{\xi}^{\star}-\xi^{\star},y^{\star}-y(\tilde{u}^{\star})\right\rangle_{L^2}^{2}. \tag*{(26)} ⟨ q ( y ~ ⋆ , ξ ~ ⋆ ) − q ⋆ , u ⋆ − u ~ ⋆ ⟩ L 2 = ⟨ y ~ ⋆ − y ⋆ , y ⋆ − y ( u ~ ⋆ ) ⟩ L 2 2 + ⟨ ξ ~ ⋆ − ξ ⋆ , y ⋆ − y ( u ~ ⋆ ) ⟩ L 2 2 . (26)
The first term is estimated via the Cauchy–Schwarz and Young inequalities, for ε 1 > 0 \varepsilon_{1}>0 ε 1 > 0 , as
⟨ y ~ ⋆ − y ⋆ , y ⋆ − y ( u ~ ⋆ ) ⟩ L 2 2 = − ∥ y ~ ⋆ − y ⋆ ∥ L 2 2 + ⟨ y ~ ⋆ − y ⋆ , y ~ ⋆ − y ( u ~ ⋆ ) ⟩ L 2 2 ≤ ( ε 1 2 − 1 ) ∥ y ~ ⋆ − y ⋆ ∥ L 2 2 + 1 2 ε 1 ∥ y ~ ⋆ − y ( u ~ ⋆ ) ∥ L 2 2 . \begin{aligned}
\left\langle\tilde{y}^{\star}-y^{\star},y^{\star}-y(\tilde{u}^{\star})\right\rangle_{L^2}^{2}
&=-\left\|\tilde{y}^{\star}-y^{\star}\right\|_{L^2}^{2}
+\left\langle\tilde{y}^{\star}-y^{\star},\tilde{y}^{\star}-y(\tilde{u}^{\star})\right\rangle_{L^2}^{2} \\
&\leq\left(\frac{\varepsilon_{1}}{2}-1\right)\left\|\tilde{y}^{\star}-y^{\star}\right\|_{L^2}^{2}
+\frac{1}{2\varepsilon_{1}}\left\|\tilde{y}^{\star}-y(\tilde{u}^{\star})\right\|_{L^2}^{2}.
\end{aligned} ⟨ y ~ ⋆ − y ⋆ , y ⋆ − y ( u ~ ⋆ ) ⟩ L 2 2 = − ∥ y ~ ⋆ − y ⋆ ∥ L 2 2 + ⟨ y ~ ⋆ − y ⋆ , y ~ ⋆ − y ( u ~ ⋆ ) ⟩ L 2 2 ≤ ( 2 ε 1 − 1 ) ∥ y ~ ⋆ − y ⋆ ∥ L 2 2 + 2 ε 1 1 ∥ y ~ ⋆ − y ( u ~ ⋆ ) ∥ L 2 2 .
Using Proposition 2.3 (ii) and Young’s inequality, the second term is estimated for ε 2 > 0 \varepsilon_{2}>0 ε 2 > 0 as
⟨ ξ ~ ⋆ − ξ ⋆ , y ⋆ − y ( u ~ ⋆ ) ⟩ L 2 2 = ⟨ ξ ~ ⋆ − ξ ⋆ , y ⋆ − y ~ ⋆ ⟩ L 2 2 + ⟨ ξ ~ ⋆ − ξ ⋆ , y ~ ⋆ − y ( u ~ ⋆ ) ⟩ L 2 2 = − 1 γ ∥ ξ ~ ⋆ − ξ ⋆ ∥ L 2 2 + ⟨ ξ ~ ⋆ − ξ ⋆ , P Y a d ( y ⋆ ) − P Y a d ( y ~ ⋆ ) ⟩ L 2 2 + ⟨ ξ ~ ⋆ − ξ ⋆ , y ~ ⋆ − y ( u ~ ⋆ ) ⟩ L 2 2 ≤ ( ε 2 2 − 1 γ ) ∥ ξ ~ ⋆ − ξ ⋆ ∥ L 2 2 + 1 2 ε 2 ∥ y ~ ⋆ − y ( u ~ ⋆ ) ∥ L 2 2 . \begin{aligned}
\left\langle\tilde{\xi}^{\star}-\xi^{\star},y^{\star}-y(\tilde{u}^{\star})\right\rangle_{L^2}^{2}
={}&\left\langle\tilde{\xi}^{\star}-\xi^{\star},y^{\star}-\tilde{y}^{\star}\right\rangle_{L^2}^{2}
+\left\langle\tilde{\xi}^{\star}-\xi^{\star},\tilde{y}^{\star}-y(\tilde{u}^{\star})\right\rangle_{L^2}^{2} \\
={}&-\frac{1}{\gamma}\left\|\tilde{\xi}^{\star}-\xi^{\star}\right\|_{L^2}^{2}
+\left\langle\tilde{\xi}^{\star}-\xi^{\star},P_{Y_{\mathrm{ad}}}(y^{\star})-P_{Y_{\mathrm{ad}}}(\tilde{y}^{\star})\right\rangle_{L^2}^{2} \\
&+\left\langle\tilde{\xi}^{\star}-\xi^{\star},\tilde{y}^{\star}-y(\tilde{u}^{\star})\right\rangle_{L^2}^{2} \\
\leq{}&\left(\frac{\varepsilon_{2}}{2}-\frac{1}{\gamma}\right)\left\|\tilde{\xi}^{\star}-\xi^{\star}\right\|_{L^2}^{2}
+\frac{1}{2\varepsilon_{2}}\left\|\tilde{y}^{\star}-y(\tilde{u}^{\star})\right\|_{L^2}^{2}.
\end{aligned} ⟨ ξ ~ ⋆ − ξ ⋆ , y ⋆ − y ( u ~ ⋆ ) ⟩ L 2 2 = = ≤ ⟨ ξ ~ ⋆ − ξ ⋆ , y ⋆ − y ~ ⋆ ⟩ L 2 2 + ⟨ ξ ~ ⋆ − ξ ⋆ , y ~ ⋆ − y ( u ~ ⋆ ) ⟩ L 2 2 − γ 1 ξ ~ ⋆ − ξ ⋆ L 2 2 + ⟨ ξ ~ ⋆ − ξ ⋆ , P Y ad ( y ⋆ ) − P Y ad ( y ~ ⋆ ) ⟩ L 2 2 + ⟨ ξ ~ ⋆ − ξ ⋆ , y ~ ⋆ − y ( u ~ ⋆ ) ⟩ L 2 2 ( 2 ε 2 − γ 1 ) ξ ~ ⋆ − ξ ⋆ L 2 2 + 2 ε 2 1 ∥ y ~ ⋆ − y ( u ~ ⋆ ) ∥ L 2 2 .
Combining both estimates leads to
⟨ q ( y ~ ⋆ , ξ ~ ⋆ ) − q ⋆ , u ⋆ − u ~ ⋆ ⟩ L 2 ≤ ( 1 2 ε 1 + 1 2 ε 2 ) ∥ y ~ ⋆ − y ( u ~ ⋆ ) ∥ L 2 2 + ( ε 2 2 − 1 γ ) ∥ ξ ~ ⋆ − ξ ⋆ ∥ L 2 2 + ( ε 1 2 − 1 ) ∥ y ~ ⋆ − y ⋆ ∥ L 2 2 . \begin{aligned}
\left\langle q(\tilde{y}^{\star},\tilde{\xi}^{\star})-q^{\star},u^{\star}-\tilde{u}^{\star}\right\rangle_{L^2}
\leq{}&\left(\frac{1}{2\varepsilon_{1}}+\frac{1}{2\varepsilon_{2}}\right)\left\|\tilde{y}^{\star}-y(\tilde{u}^{\star})\right\|_{L^2}^{2}
+\left(\frac{\varepsilon_{2}}{2}-\frac{1}{\gamma}\right)\left\|\tilde{\xi}^{\star}-\xi^{\star}\right\|_{L^2}^{2} \\
&+\left(\frac{\varepsilon_{1}}{2}-1\right)\left\|\tilde{y}^{\star}-y^{\star}\right\|_{L^2}^{2}.
\end{aligned} ⟨ q ( y ~ ⋆ , ξ ~ ⋆ ) − q ⋆ , u ⋆ − u ~ ⋆ ⟩ L 2 ≤ ( 2 ε 1 1 + 2 ε 2 1 ) ∥ y ~ ⋆ − y ( u ~ ⋆ ) ∥ L 2 2 + ( 2 ε 2 − γ 1 ) ξ ~ ⋆ − ξ ⋆ L 2 2 + ( 2 ε 1 − 1 ) ∥ y ~ ⋆ − y ⋆ ∥ L 2 2 .
For the first term on the right-hand side in (24) we use the weighted Young’s inequality with ε 3 > 0 \varepsilon_{3}>0 ε 3 > 0 to obtain
⟨ q ⋆ − q ( y ~ ⋆ , ξ ~ ⋆ ) , u ⋆ − u ~ ⋆ ⟩ L 2 ≤ 1 2 ε 3 ∥ q ⋆ − q ( y ~ ⋆ , ξ ~ ⋆ ) ∥ L 2 2 + ε 3 2 ∥ u ⋆ − u ~ ⋆ ∥ L 2 2 . \left\langle q^{\star}-q(\tilde{y}^{\star},\tilde{\xi}^{\star}),u^{\star}-\tilde{u}^{\star}\right\rangle_{L^2}
\leq\frac{1}{2\varepsilon_{3}}\left\|q^{\star}-q(\tilde{y}^{\star},\tilde{\xi}^{\star})\right\|_{L^2}^{2}
+\frac{\varepsilon_{3}}{2}\left\|u^{\star}-\tilde{u}^{\star}\right\|_{L^2}^{2}. ⟨ q ⋆ − q ( y ~ ⋆ , ξ ~ ⋆ ) , u ⋆ − u ~ ⋆ ⟩ L 2 ≤ 2 ε 3 1 q ⋆ − q ( y ~ ⋆ , ξ ~ ⋆ ) L 2 2 + 2 ε 3 ∥ u ⋆ − u ~ ⋆ ∥ L 2 2 .
Combining all inequalities yields the estimate
( 1 − ε 1 2 ) ∥ y ⋆ − y ~ ⋆ ∥ L 2 2 + ( 1 γ − ε 2 2 ) ∥ ξ ~ ⋆ − ξ ⋆ ∥ L 2 2 + ( λ − ε 3 2 ) ∥ u ~ ⋆ − u ⋆ ∥ L 2 2 ≤ ( 1 2 ε 1 + 1 2 ε 2 ) ∥ y ⋆ − y ( u ~ ⋆ ) ∥ L 2 2 + 1 2 ε 3 ∥ q ~ ⋆ − q ( y ~ ⋆ , ξ ~ ⋆ ) ∥ L 2 2 . (27) \left(1-\frac{\varepsilon_1}{2}\right)\lVert\mathbf{y}^\star-\widetilde{\mathbf{y}}^\star\rVert_{L^2}^2+\left(\frac{1}{\gamma}-\frac{\varepsilon_2}{2}\right)\lVert\widetilde{\boldsymbol{\xi}}^\star-\boldsymbol{\xi}^\star\rVert_{L^2}^2+\left(\lambda-\frac{\varepsilon_3}{2}\right)\lVert\widetilde{\mathbf{u}}^\star-\mathbf{u}^\star\rVert_{L^2}^2
\leq\left(\frac{1}{2\varepsilon_1}+\frac{1}{2\varepsilon_2}\right)\lVert\mathbf{y}^\star-\mathbf{y}(\widetilde{\mathbf{u}}^\star)\rVert_{L^2}^2+\frac{1}{2\varepsilon_3}\lVert\widetilde{\mathbf{q}}^\star-\mathbf{q}(\widetilde{\mathbf{y}}^\star,\widetilde{\boldsymbol{\xi}}^\star)\rVert_{L^2}^2. \tag*{(27)} ( 1 − 2 ε 1 ) ∥ y ⋆ − y ⋆ ∥ L 2 2 + ( γ 1 − 2 ε 2 ) ∥ ξ ⋆ − ξ ⋆ ∥ L 2 2 + ( λ − 2 ε 3 ) ∥ u ⋆ − u ⋆ ∥ L 2 2 ≤ ( 2 ε 1 1 + 2 ε 2 1 ) ∥ y ⋆ − y ( u ⋆ ) ∥ L 2 2 + 2 ε 3 1 ∥ q ⋆ − q ( y ⋆ , ξ ⋆ ) ∥ L 2 2 . (27)
Choosing ε 1 = 1 \varepsilon_1=1 ε 1 = 1 , ε 2 = 1 / γ \varepsilon_2=1/\gamma ε 2 = 1/ γ , and ε 3 = λ \varepsilon_3=\lambda ε 3 = λ , multiplying by two, and taking square roots proves the claim.
Next, we derive an error estimate for the optimal value function.
Theorem 2.5 (A posteriori error estimate for the value function). In the situation of Theorem 2.4, we have
η − ≤ J γ ( y ⋆ , u ⋆ ) − J γ ( y ~ ⋆ , u ~ ⋆ ) ≤ η + , (28) \eta_-\leq J_\gamma(\mathbf{y}^\star,\mathbf{u}^\star)-J_\gamma(\widetilde{\mathbf{y}}^\star,\widetilde{\mathbf{u}}^\star)\leq\eta_+, \tag*{(28)} η − ≤ J γ ( y ⋆ , u ⋆ ) − J γ ( y ⋆ , u ⋆ ) ≤ η + , (28) where
η ± : = ∥ y ~ ⋆ − y d + ξ ~ ⋆ , y ( u ~ ⋆ ) − y ~ ⋆ ∥ L 2 2 + { − 1 2 λ ∥ q ( y ~ ⋆ , ξ ~ ⋆ ) − q ~ ⋆ ∥ L 2 2 , for η − , 1 + γ 2 ∥ y ( u ~ ⋆ ) − y ~ ⋆ ∥ L 2 2 , for η + . (29) \eta_\pm:=\left\lVert\widetilde{\mathbf{y}}^\star-\mathbf{y}_d+\widetilde{\boldsymbol{\xi}}^\star,\mathbf{y}(\widetilde{\mathbf{u}}^\star)-\widetilde{\mathbf{y}}^\star\right\rVert_{L^2}^2+
\begin{cases}
-\frac{1}{2\lambda}\lVert\mathbf{q}(\widetilde{\mathbf{y}}^\star,\widetilde{\boldsymbol{\xi}}^\star)-\widetilde{\mathbf{q}}^\star\rVert_{L^2}^2, & \text{for }\eta_-,\\
\frac{1+\gamma}{2}\lVert\mathbf{y}(\widetilde{\mathbf{u}}^\star)-\widetilde{\mathbf{y}}^\star\rVert_{L^2}^2, & \text{for }\eta_+.
\end{cases} \tag*{(29)} η ± := y ⋆ − y d + ξ ⋆ , y ( u ⋆ ) − y ⋆ L 2 2 + { − 2 λ 1 ∥ q ( y ⋆ , ξ ⋆ ) − q ⋆ ∥ L 2 2 , 2 1 + γ ∥ y ( u ⋆ ) − y ⋆ ∥ L 2 2 , for η − , for η + . (29) Consequently, the value function error
e J γ : = ∣ J γ ( y ⋆ , u ⋆ ) − J γ ( y ~ ⋆ , u ~ ⋆ ) ∣ (30) e_{J_\gamma}:=\left|J_\gamma(\mathbf{y}^\star,\mathbf{u}^\star)-J_\gamma(\widetilde{\mathbf{y}}^\star,\widetilde{\mathbf{u}}^\star)\right| \tag*{(30)} e J γ := ∣ J γ ( y ⋆ , u ⋆ ) − J γ ( y ⋆ , u ⋆ ) ∣ (30) satisfies
e J γ ≤ Δ J γ ( y ~ ⋆ , u ~ ⋆ , ξ ~ ⋆ ) : = max { ∣ η − ∣ , ∣ η + ∣ } . (31) e_{J_\gamma}\leq\Delta_{J_\gamma}(\widetilde{\mathbf{y}}^\star,\widetilde{\mathbf{u}}^\star,\widetilde{\boldsymbol{\xi}}^\star):=\max\{|\eta_-|,|\eta_+|\}. \tag*{(31)} e J γ ≤ Δ J γ ( y ⋆ , u ⋆ , ξ ⋆ ) := max { ∣ η − ∣ , ∣ η + ∣ } . (31)
Proof. For fixed u \mathbf{u} u , the output-dependent part of J γ J_\gamma J γ is convex, with gradient y − y d + ξ ( y ) \mathbf{y}-\mathbf{y}_d+\boldsymbol{\xi}(\mathbf{y}) y − y d + ξ ( y ) . This gradient is ( 1 + γ ) (1+\gamma) ( 1 + γ ) -Lipschitz, since y ↦ y − P Y a d ( y ) \mathbf{y}\mapsto\mathbf{y}-P_{\mathcal{Y}_{\mathrm{ad}}}(\mathbf{y}) y ↦ y − P Y ad ( y ) is nonexpansive. Turning to the lower bound, for every v ∈ U a d \mathbf{v}\in\mathcal{U}_{\mathrm{ad}} v ∈ U ad , convexity in the output and the quadratic control cost give
J γ ( y ( v ) , v ) − J γ ( y ~ ⋆ , u ~ ⋆ ) ≥ ⟨ y ~ ⋆ − y d + ξ ~ ⋆ , y ( v ) − y ~ ⋆ ⟩ L 2 + λ ⟨ u ~ ⋆ − u d , v − u ~ ⋆ ⟩ L 2 + λ 2 ∥ v − u ~ ⋆ ∥ L 2 2 . J_\gamma(\mathbf{y}(\mathbf{v}),\mathbf{v})-J_\gamma(\widetilde{\mathbf{y}}^\star,\widetilde{\mathbf{u}}^\star)\geq\langle\widetilde{\mathbf{y}}^\star-\mathbf{y}_d+\widetilde{\boldsymbol{\xi}}^\star,\mathbf{y}(\mathbf{v})-\widetilde{\mathbf{y}}^\star\rangle_{L^2}+\lambda\langle\widetilde{\mathbf{u}}^\star-\mathbf{u}_d,\mathbf{v}-\widetilde{\mathbf{u}}^\star\rangle_{L^2}+\frac{\lambda}{2}\lVert\mathbf{v}-\widetilde{\mathbf{u}}^\star\rVert_{L^2}^2. J γ ( y ( v ) , v ) − J γ ( y ⋆ , u ⋆ ) ≥ ⟨ y ⋆ − y d + ξ ⋆ , y ( v ) − y ⋆ ⟩ L 2 + λ ⟨ u ⋆ − u d , v − u ⋆ ⟩ L 2 + 2 λ ∥ v − u ⋆ ∥ L 2 2 .
Split y ( v ) − y ~ ⋆ \mathbf{y}(\mathbf{v})-\widetilde{\mathbf{y}}^\star y ( v ) − y ⋆ into y ( v ) − y ( u ~ ⋆ ) \mathbf{y}(\mathbf{v})-\mathbf{y}(\widetilde{\mathbf{u}}^\star) y ( v ) − y ( u ⋆ ) and y ( u ~ ⋆ ) − y ~ ⋆ \mathbf{y}(\widetilde{\mathbf{u}}^\star)-\widetilde{\mathbf{y}}^\star y ( u ⋆ ) − y ⋆ . Linearity of the FOM input-to-output map, the adjoint identity as in (26) , and the reduced optimality condition (2.3h) then imply
J γ ( y ( v ) , v ) − J γ ( y ~ ⋆ , u ~ ⋆ ) ≥ ⟨ y ~ ⋆ − y d + ξ ~ ⋆ , y ( u ~ ⋆ ) − y ~ ⋆ ⟩ L 2 + ⟨ q ( y ~ ⋆ , ξ ~ ⋆ ) − q ~ ⋆ , v − u ~ ⋆ ⟩ L 2 + λ 2 ∥ v − u ~ ⋆ ∥ L 2 2 ≥ ⟨ y ~ ⋆ − y d + ξ ~ ⋆ , y ( u ~ ⋆ ) − y ~ ⋆ ⟩ L 2 − 1 2 λ ∥ q ( y ~ ⋆ , ξ ~ ⋆ ) − q ~ ⋆ ∥ L 2 2 = η − , \begin{aligned}
J_\gamma(\mathbf{y}(\mathbf{v}),\mathbf{v})-J_\gamma(\widetilde{\mathbf{y}}^\star,\widetilde{\mathbf{u}}^\star)
&\geq\langle\widetilde{\mathbf{y}}^\star-\mathbf{y}_d+\widetilde{\boldsymbol{\xi}}^\star,\mathbf{y}(\widetilde{\mathbf{u}}^\star)-\widetilde{\mathbf{y}}^\star\rangle_{L^2}\\
&\quad+\langle\mathbf{q}(\widetilde{\mathbf{y}}^\star,\widetilde{\boldsymbol{\xi}}^\star)-\widetilde{\mathbf{q}}^\star,\mathbf{v}-\widetilde{\mathbf{u}}^\star\rangle_{L^2}+\frac{\lambda}{2}\lVert\mathbf{v}-\widetilde{\mathbf{u}}^\star\rVert_{L^2}^2\\
&\geq\langle\widetilde{\mathbf{y}}^\star-\mathbf{y}_d+\widetilde{\boldsymbol{\xi}}^\star,\mathbf{y}(\widetilde{\mathbf{u}}^\star)-\widetilde{\mathbf{y}}^\star\rangle_{L^2}\\
&\quad-\frac{1}{2\lambda}\lVert\mathbf{q}(\widetilde{\mathbf{y}}^\star,\widetilde{\boldsymbol{\xi}}^\star)-\widetilde{\mathbf{q}}^\star\rVert_{L^2}^2=\eta_-,
\end{aligned} J γ ( y ( v ) , v ) − J γ ( y ⋆ , u ⋆ ) ≥ ⟨ y ⋆ − y d + ξ ⋆ , y ( u ⋆ ) − y ⋆ ⟩ L 2 + ⟨ q ( y ⋆ , ξ ⋆ ) − q ⋆ , v − u ⋆ ⟩ L 2 + 2 λ ∥ v − u ⋆ ∥ L 2 2 ≥ ⟨ y ⋆ − y d + ξ ⋆ , y ( u ⋆ ) − y ⋆ ⟩ L 2 − 2 λ 1 ∥ q ( y ⋆ , ξ ⋆ ) − q ⋆ ∥ L 2 2 = η − ,
where the last step is Young’s inequality. Taking v = u ⋆ \mathbf{v}=\mathbf{u}^\star v = u ⋆ proves the lower bound in (2.10). For the upper bound, optimality of u ⋆ \mathbf{u}^\star u ⋆ , feasibility of u ~ ⋆ ∈ U a d \widetilde{\mathbf{u}}^\star\in\mathcal{U}_{\mathrm{ad}} u ⋆ ∈ U ad , and the Lipschitz continuity of the output gradient give
J γ ( y ⋆ , u ⋆ ) − J γ ( y ~ ⋆ , u ~ ⋆ ) ≤ J γ ( y ( u ~ ⋆ ) , u ~ ⋆ ) − J γ ( y ~ ⋆ , u ~ ⋆ ) ≤ ⟨ y ~ ⋆ − y d + ξ ~ ⋆ , y ( u ~ ⋆ ) − y ~ ⋆ ⟩ L 2 + 1 + γ 2 ∥ y ( u ~ ⋆ ) − y ~ ⋆ ∥ L 2 2 = η + . \begin{aligned}
J_\gamma(\mathbf{y}^\star,\mathbf{u}^\star)-J_\gamma(\widetilde{\mathbf{y}}^\star,\widetilde{\mathbf{u}}^\star)
&\leq J_\gamma(\mathbf{y}(\widetilde{\mathbf{u}}^\star),\widetilde{\mathbf{u}}^\star)-J_\gamma(\widetilde{\mathbf{y}}^\star,\widetilde{\mathbf{u}}^\star)\\
&\leq\langle\widetilde{\mathbf{y}}^\star-\mathbf{y}_d+\widetilde{\boldsymbol{\xi}}^\star,\mathbf{y}(\widetilde{\mathbf{u}}^\star)-\widetilde{\mathbf{y}}^\star\rangle_{L^2}+\frac{1+\gamma}{2}\lVert\mathbf{y}(\widetilde{\mathbf{u}}^\star)-\widetilde{\mathbf{y}}^\star\rVert_{L^2}^2=\eta_+.
\end{aligned} J γ ( y ⋆ , u ⋆ ) − J γ ( y ⋆ , u ⋆ ) ≤ J γ ( y ( u ⋆ ) , u ⋆ ) − J γ ( y ⋆ , u ⋆ ) ≤ ⟨ y ⋆ − y d + ξ ⋆ , y ( u ⋆ ) − y ⋆ ⟩ L 2 + 2 1 + γ ∥ y ( u ⋆ ) − y ⋆ ∥ L 2 2 = η + .
This proves the upper bound in (2.10). Taking the maximum of the absolute values of the two endpoints yields (2.13).
To estimate how the reduced control performs in the FOM cost, note that the map u ↦ J γ ( y ( u ) , u ) \mathbf{u} \mapsto J_\gamma(\mathbf{y}(\mathbf{u}),\mathbf{u}) u ↦ J γ ( y ( u ) , u ) is λ \lambda λ -strongly convex, with gradient λ ( u − u d ) + q ( y ( u ) , ξ ( y ( u ) ) ) \lambda(\mathbf{u}-\mathbf{u}_d)+\mathbf{q}(\mathbf{y}(\mathbf{u}),\boldsymbol{\xi}(\mathbf{y}(\mathbf{u}))) λ ( u − u d ) + q ( y ( u ) , ξ ( y ( u ))) . Thus, using the reduced optimality condition (2.3h) with v = u ⋆ \mathbf{v}=\mathbf{u}^{\star} v = u ⋆ yields
0 ≤ J γ ( y ( u ~ ⋆ ) , u ~ ⋆ ) − J γ ( y ⋆ , u ⋆ ) ≤ ⟨ λ ( u ~ ⋆ − u d ) + q ( y ( u ~ ⋆ ) , ξ ( y ( u ~ ⋆ ) ) ) , u ~ ⋆ − u ⋆ ⟩ L 2 − λ 2 ∥ u ~ ⋆ − u ⋆ ∥ L 2 2 ≤ ⟨ q ( y ( u ~ ⋆ ) , ξ ( y ( u ~ ⋆ ) ) ) − q ⋆ , u ~ ⋆ − u ⋆ ⟩ L 2 − λ 2 ∥ u ~ ⋆ − u ⋆ ∥ L 2 2 . \begin{aligned}
0 &\le J_\gamma(\mathbf{y}(\widetilde{\mathbf{u}}^{\star}),\widetilde{\mathbf{u}}^{\star})-J_\gamma(\mathbf{y}^{\star},\mathbf{u}^{\star}) \\
&\le\left\langle\lambda(\widetilde{\mathbf{u}}^{\star}-\mathbf{u}_d)+\mathbf{q}(\mathbf{y}(\widetilde{\mathbf{u}}^{\star}),\boldsymbol{\xi}(\mathbf{y}(\widetilde{\mathbf{u}}^{\star}))),\widetilde{\mathbf{u}}^{\star}-\mathbf{u}^{\star}\right\rangle_{L^2}-\frac{\lambda}{2}\lVert\widetilde{\mathbf{u}}^{\star}-\mathbf{u}^{\star}\rVert_{L^2}^{2} \\
&\le\left\langle\mathbf{q}(\mathbf{y}(\widetilde{\mathbf{u}}^{\star}),\boldsymbol{\xi}(\mathbf{y}(\widetilde{\mathbf{u}}^{\star})))-\mathbf{q}^{\star},\widetilde{\mathbf{u}}^{\star}-\mathbf{u}^{\star}\right\rangle_{L^2}-\frac{\lambda}{2}\lVert\widetilde{\mathbf{u}}^{\star}-\mathbf{u}^{\star}\rVert_{L^2}^{2}.
\end{aligned} 0 ≤ J γ ( y ( u ⋆ ) , u ⋆ ) − J γ ( y ⋆ , u ⋆ ) ≤ ⟨ λ ( u ⋆ − u d ) + q ( y ( u ⋆ ) , ξ ( y ( u ⋆ ))) , u ⋆ − u ⋆ ⟩ L 2 − 2 λ ∥ u ⋆ − u ⋆ ∥ L 2 2 ≤ ⟨ q ( y ( u ⋆ ) , ξ ( y ( u ⋆ ))) − q ⋆ , u ⋆ − u ⋆ ⟩ L 2 − 2 λ ∥ u ⋆ − u ⋆ ∥ L 2 2 .
The bounds in Theorem 2.4 and Theorem 2.5 can be evaluated without computing the FOM optimal control. Their direct evaluation requires one forward FOM simulation driven by u ~ ⋆ \widetilde{\mathbf{u}}^{\star} u ⋆ , to compute y ( u ~ ⋆ ) \mathbf{y}(\widetilde{\mathbf{u}}^{\star}) y ( u ⋆ ) , and one backward FOM adjoint simulation driven by, e.g., ( y ~ ⋆ , ξ ~ ⋆ ) (\widetilde{\mathbf{y}}^{\star},\widetilde{\boldsymbol{\xi}}^{\star}) ( y ⋆ , ξ ⋆ ) , to compute q ( y ~ ⋆ , ξ ~ ⋆ ) \mathbf{q}(\widetilde{\mathbf{y}}^{\star},\widetilde{\boldsymbol{\xi}}^{\star}) q ( y ⋆ , ξ ⋆ ) . This direct evaluation uses the full-order forward and adjoint models without requiring full-order optimization, which motivates the term full-order-evaluation bounds. The second numerical experiment compares these bounds with an online-efficient PBR approximation.
Remark 2.6 (Augmented-Lagrangian regularization). Our error estimation framework can also be adapted to augmented Lagrangian methods (see, e.g., [19]). In this case, for a fixed shift s ∈ L 2 ( 0 , T ; R p ) \mathbf{s}\in L^{2}(0,T;\mathbb{R}^{p}) s ∈ L 2 ( 0 , T ; R p ) , the MY penalty in (1.5) is replaced by
γ 2 dist L 2 2 ( y + s γ , Y a d ) − 1 2 γ ∥ s ∥ L 2 2 . \frac{\gamma}{2}\operatorname{dist}_{L^2}^{2}\left(\mathbf{y}+\frac{\mathbf{s}}{\gamma},Y_{\mathrm{ad}}\right)-\frac{1}{2\gamma}\lVert\mathbf{s}\rVert_{L^2}^{2}. 2 γ dist L 2 2 ( y + γ s , Y ad ) − 2 γ 1 ∥ s ∥ L 2 2 .
The effective multiplier is ξ γ , s ( y ) = γ ( y + s / γ − P Y a d ( y + s / γ ) ) \boldsymbol{\xi}_{\gamma,\mathbf{s}}(\mathbf{y})=\gamma\left(\mathbf{y}+\mathbf{s}/\gamma-P_{Y_{\mathrm{ad}}}(\mathbf{y}+\mathbf{s}/\gamma)\right) ξ γ , s ( y ) = γ ( y + s / γ − P Y ad ( y + s / γ ) ) . Then, the estimates in Theorems 2.4 and 2.5 extend unchanged, using this multiplier in the adjoint input.
Piecewise balanced reduction for the optimal control problem
The analysis from Section 2 can be applied to any Petrov–Galerkin projection-based ROM (1.8) that approximates the forward and adjoint input-output maps. We show in Section 3.2 that the PBR from [24] is a well-suited tool for reducing the OCP, since the duality between observability and controllability makes PBR commute with the derivation of the adjoint system: reducing the FOM first yields the same adjoint system as deriving the full-order adjoint first and then applying PBR.
Throughout this section, we impose the following assumption.
Assumption 3.1. For each j = 1 , … , L j=1,\ldots,L j = 1 , … , L , all eigenvalues of E j − 1 A j E_j^{-1}A_j E j − 1 A j lie in the open left half-plane.
Assumption 3.1 ensures that each mode of the SLS is asymptotically stable. Although this condition alone does not guarantee asymptotic stability for the SLS under arbitrary switching signals, it remains useful for introducing the notion of generalized Gramians defined through Lyapunov-type equations, which are fundamental to the PBR method, which we recall in Section 3.1.
The PBR for SLS
Given the SLS (1.1), we consider for each i ∈ K i\in\mathcal{K} i ∈ K the pair of GLEs
A i P i , x E i T + E i P i , x A i T + ∑ j = 1 L ( N i , j , x P i , x N i , j , x T + B i , j , x B i , j , x T ) = 0. (32) A_i\mathcal{P}_{i,x}E_i^{\mathsf{T}}+E_i\mathcal{P}_{i,x}A_i^{\mathsf{T}}+\sum_{j=1}^{L}\left(N_{i,j,x}\mathcal{P}_{i,x}N_{i,j,x}^{\mathsf{T}}+B_{i,j,x}B_{i,j,x}^{\mathsf{T}}\right)=0. \tag*{(32)} A i P i , x E i T + E i P i , x A i T + j = 1 ∑ L ( N i , j , x P i , x N i , j , x T + B i , j , x B i , j , x T ) = 0. (32)
A i T Q i , x E i + E i T Q i , x A i + ∑ j = 1 L ( N i , j , x T Q i , x N i , j , x + C i , j , x T C i , j , x ) = 0. (33) A_i^{\mathsf{T}}Q_{i,x}E_i+E_i^{\mathsf{T}}Q_{i,x}A_i+\sum_{j=1}^{L}\left(N_{i,j,x}^{\mathsf{T}}Q_{i,x}N_{i,j,x}+C_{i,j,x}^{\mathsf{T}}C_{i,j,x}\right)=0. \tag*{(33)} A i T Q i , x E i + E i T Q i , x A i + j = 1 ∑ L ( N i , j , x T Q i , x N i , j , x + C i , j , x T C i , j , x ) = 0. (33)
where for all i , j ∈ K i,j\in\mathcal{K} i , j ∈ K
B i , j , x : = E i K i , j E j − 1 B j , C i , j , x : = C j K j , i , N i , j , x : = E i K i , j E j − 1 A j K j , i − A i . B_{i,j,x}:=E_iK_{i,j}E_j^{-1}B_j,\qquad C_{i,j,x}:=C_jK_{j,i},\qquad N_{i,j,x}:=E_iK_{i,j}E_j^{-1}A_jK_{j,i}-A_i. B i , j , x := E i K i , j E j − 1 B j , C i , j , x := C j K j , i , N i , j , x := E i K i , j E j − 1 A j K j , i − A i .
Note that we use the subscript x x x to emphasize that the matrices are computed for the forward system (1) . In contrast, if we compute the matrices for the adjoint system, we write P i , p \mathcal{P}_{i,\mathrm{p}} P i , p and similarly for the other matrices. The existence and uniqueness of the solutions to (32) , and analogously to (33) , can be studied using Kronecker algebra. To this end, define L i , Π i , x ∈ R n 2 × n 2 \mathcal{L}_i,\Pi_{i,\mathrm{x}} \in\mathbb{R}^{n^2 \times n^2} L i , Π i , x ∈ R n 2 × n 2 via
L i : = E i ⊗ A i + A i ⊗ E i and Π i , x : = ∑ j = 1 L N i , j , x ⊗ N i , j , x . (34) \mathcal{L}_i := \mathbf{E}_i \otimes\mathbf{A}_i + \mathbf{A}_i \otimes\mathbf{E}_i
\qquad\text{and}\qquad
\Pi_{i,\mathrm{x}} := \sum_{j=1}^{L}\mathbf{N}_{i,j,\mathrm{x}} \otimes\mathbf{N}_{i,j,\mathrm{x}}. \tag*{(34)} L i := E i ⊗ A i + A i ⊗ E i and Π i , x := j = 1 ∑ L N i , j , x ⊗ N i , j , x . (34)
The vectorized form of (32) is then ( L i + Π i , x ) vec ( P i , x ) = − vec ( ∑ j = 1 L B i , j , x B i , j , x T ) (\mathcal{L}_i+\Pi_{i,\mathrm{x}})\operatorname{vec}(\mathcal{P}_{i,\mathrm{x}})=-\operatorname{vec}(\sum_{j=1}^{L}\mathbf{B}_{i,j,\mathrm{x}}\mathbf{B}_{i,j,\mathrm{x}}^{\mathsf{T}}) ( L i + Π i , x ) vec ( P i , x ) = − vec ( ∑ j = 1 L B i , j , x B i , j , x T ) . Hence, (32) is uniquely solvable if and only if L i + Π i , x \mathcal{L}_i+\Pi_{i,\mathrm{x}} L i + Π i , x is nonsingular.
Assumption 3.2. The GLEs in (3.1) are uniquely solvable for all i ∈ K i \in\mathcal{K} i ∈ K .
The solutions P i , x \mathcal{P}_{i,\mathrm{x}} P i , x and Q i , x \mathcal{Q}_{i,\mathrm{x}} Q i , x are symmetric positive semidefinite due to Assumption 3.1 and satisfy R = range ( P i , x ) \mathcal{R}=\operatorname{range}(\mathcal{P}_{i,\mathrm{x}}) R = range ( P i , x ) and O = range ( E i T Q i , x E i ) \mathcal{O}=\operatorname{range}(\mathbf{E}_i^{\mathsf{T}}\mathcal{Q}_{i,\mathrm{x}}\mathbf{E}_i) O = range ( E i T Q i , x E i ) for i = 1 , … , L i=1,\ldots,L i = 1 , … , L , where R \mathcal{R} R and O \mathcal{O} O are the reachable and observable sets of the SLS, respectively (see [24, Sec. 2.1] and [30, Thm. 3]). We therefore refer to the L L L pairs ( P i , x , Q i , x ) (\mathcal{P}_{i,\mathrm{x}},\mathcal{Q}_{i,\mathrm{x}}) ( P i , x , Q i , x ) as the reachability and observability matrices of the SLS.
Let S i , x \mathbf{S}_{i,\mathrm{x}} S i , x and R i , x \mathbf{R}_{i,\mathrm{x}} R i , x be Cholesky factors satisfying P i , x = S i , x S i , x T \mathcal{P}_{i,\mathrm{x}}=\mathbf{S}_{i,\mathrm{x}}\mathbf{S}_{i,\mathrm{x}}^{\mathsf{T}} P i , x = S i , x S i , x T and Q i , x = R i , x R i , x T \mathcal{Q}_{i,\mathrm{x}}=\mathbf{R}_{i,\mathrm{x}}\mathbf{R}_{i,\mathrm{x}}^{\mathsf{T}} Q i , x = R i , x R i , x T . Since the observability matrix of the normalized subsystem is E i T Q i , x E i \mathbf{E}_i^{\mathsf{T}}\mathcal{Q}_{i,\mathrm{x}}\mathbf{E}_i E i T Q i , x E i with associated Cholesky factor E i T R i , x \mathbf{E}_i^{\mathsf{T}}\mathbf{R}_{i,\mathrm{x}} E i T R i , x , we compute the mass-weighted singular value decomposition
S i , x T E i T R i , x = [ U i , x , 1 , U i , x , 2 ] [ Σ i , x , 1 0 0 Σ i , x , 2 ] [ Ψ i , x , 1 , Ψ i , x , 2 ] T , (35) \mathbf{S}_{i,\mathrm{x}}^{\mathsf{T}}\mathbf{E}_i^{\mathsf{T}}\mathbf{R}_{i,\mathrm{x}}
=
[\mathbf{U}_{i,\mathrm{x},1},\mathbf{U}_{i,\mathrm{x},2}]
\begin{bmatrix}
\boldsymbol{\Sigma}_{i,\mathrm{x},1} & 0\\
0 & \boldsymbol{\Sigma}_{i,\mathrm{x},2}
\end{bmatrix}
[\mathbf{\Psi}_{i,\mathrm{x},1},\mathbf{\Psi}_{i,\mathrm{x},2}]^{\mathsf{T}}, \tag*{(35)} S i , x T E i T R i , x = [ U i , x , 1 , U i , x , 2 ] [ Σ i , x , 1 0 0 Σ i , x , 2 ] [ Ψ i , x , 1 , Ψ i , x , 2 ] T , (35)
where the first block contains the r r r retained positive singular values. The projection matrices V i , x \mathbf{V}_{i,\mathrm{x}} V i , x and W i , x \mathbf{W}_{i,\mathrm{x}} W i , x are then given by
V i , x = S i , x U i , x , 1 Σ i , x , 1 − 1 / 2 and W i , x = R i , x Ψ i , x , 1 Σ i , x , 1 − 1 / 2 . (36) \mathbf{V}_{i,\mathrm{x}}=\mathbf{S}_{i,\mathrm{x}}\mathbf{U}_{i,\mathrm{x},1}\boldsymbol{\Sigma}_{i,\mathrm{x},1}^{-1/2}
\qquad\text{and}\qquad
\mathbf{W}_{i,\mathrm{x}}=\mathbf{R}_{i,\mathrm{x}}\mathbf{\Psi}_{i,\mathrm{x},1}\boldsymbol{\Sigma}_{i,\mathrm{x},1}^{-1/2}. \tag*{(36)} V i , x = S i , x U i , x , 1 Σ i , x , 1 − 1/2 and W i , x = R i , x Ψ i , x , 1 Σ i , x , 1 − 1/2 . (36)
These matrices satisfy
W i , x T E i V i , x = Σ i , x , 1 − 1 / 2 Ψ i , x , 1 T ( R i , x T E i S i , x ) U i , x , 1 Σ i , x , 1 − 1 / 2 = I r . \mathbf{W}_{i,\mathrm{x}}^{\mathsf{T}}\mathbf{E}_i\mathbf{V}_{i,\mathrm{x}}
=
\boldsymbol{\Sigma}_{i,\mathrm{x},1}^{-1/2}\mathbf{\Psi}_{i,\mathrm{x},1}^{\mathsf{T}}
(\mathbf{R}_{i,\mathrm{x}}^{\mathsf{T}}\mathbf{E}_i\mathbf{S}_{i,\mathrm{x}})
\mathbf{U}_{i,\mathrm{x},1}\boldsymbol{\Sigma}_{i,\mathrm{x},1}^{-1/2}
=
\mathbf{I}_r. W i , x T E i V i , x = Σ i , x , 1 − 1/2 Ψ i , x , 1 T ( R i , x T E i S i , x ) U i , x , 1 Σ i , x , 1 − 1/2 = I r .
Thus, the reduced descriptor matrices satisfy E ~ i = I r \widetilde{\mathbf{E}}_i=\mathbf{I}_r E i = I r , even when E i ≠ I n \mathbf{E}_i\ne\mathbf{I}_n E i = I n . Applying these projection matrices in (9) yields the ROM introduced in (8) with transition matrices K ~ i , j = W i , x T E i K i , j V j , x \widetilde{\mathbf{K}}_{i,j}=\mathbf{W}_{i,\mathrm{x}}^{\mathsf{T}}\mathbf{E}_i\mathbf{K}_{i,j}\mathbf{V}_{j,\mathrm{x}} K i , j = W i , x T E i K i , j V j , x . This procedure is called PBR for SLS. Its stability properties are discussed in [24, Sec. 4.2]. Moreover, PBR provides an a posteriori estimator for the L 2 L^2 L 2 -norm of the output error [24, Thm. 4.3]:
∥ y ( u ) − y ~ ( u ) ∥ L 2 ( 0 , T ; R p ) ≤ Δ ( Σ x , 2 , x ~ , λ 1 , u ) , where \lVert\mathbf{y}(u)-\widetilde{\mathbf{y}}(u)\rVert_{L^2(0,T;\mathbb{R}^p)}
\le\Delta(\boldsymbol{\Sigma}_{\mathrm{x},2},\widetilde{x},\lambda_1,u),
\qquad\text{where} ∥ y ( u ) − y ( u ) ∥ L 2 ( 0 , T ; R p ) ≤ Δ ( Σ x , 2 , x , λ 1 , u ) , where
Δ ( Σ x , 2 , x ~ , λ 1 , u ) : = ∑ i = r + 1 n ( 2 ∥ u ∥ L 2 f 1 , i ( σ 1 , … , σ L ) + f 2 , i ( x ~ − , x ~ + ) + f 3 , i ( λ 1 ( M c ) , λ 1 ( M o ) ) ) . (37) \Delta(\boldsymbol{\Sigma}_{\mathrm{x},2},\widetilde{x},\lambda_1,u)
:=
\sum_{i=r+1}^{n}
\left(
2\lVert u\rVert_{L^2}f_{1,i}(\sigma_1,\ldots,\sigma_L)
+f_{2,i}(\widetilde{x}_{-},\widetilde{x}_{+})
+f_{3,i}(\lambda_1(\mathbf{M}_{\mathrm{c}}),\lambda_1(\mathbf{M}_{\mathrm{o}}))
\right). \tag*{(37)} Δ ( Σ x , 2 , x , λ 1 , u ) := i = r + 1 ∑ n ( 2 ∥ u ∥ L 2 f 1 , i ( σ 1 , … , σ L ) + f 2 , i ( x − , x + ) + f 3 , i ( λ 1 ( M c ) , λ 1 ( M o )) ) . (37)
and Δ ( Σ x , 2 , x ~ , λ 1 , u ) \Delta(\boldsymbol{\Sigma}_{\mathrm{x},2},\widetilde{x},\lambda_1,u) Δ ( Σ x , 2 , x , λ 1 , u ) depends on the sum of three main contributions. The terms f 1 , i f_{1,i} f 1 , i account for the neglected singular values in (35) . Through f 2 , i f_{2,i} f 2 , i , the effect of discontinuities in the reduced state at the switching instants is quantified. Finally, the terms involving f 3 , i f_{3,i} f 3 , i account for violations of the relevant linear matrix inequalities as measured by the largest positive eigenvalue. The derivation of the bound, the precise definitions of all quantities involved, and a detailed discussion of its properties can be found in [24, Sec. 4]. We note that the error estimator Δ ( Σ x , 2 , x ~ , λ 1 , u ) \Delta(\Sigma_{x,2},\tilde{x},\lambda_1,u) Δ ( Σ x , 2 , x ~ , λ 1 , u ) in (37) is not computationally efficient to evaluate, as it requires an a posteriori computation of trajectories in the full-order space of dimension n n n . In practical implementations, the summation up to n n n is therefore replaced by a truncated sum up to r ~ \tilde{r} r ~ , where r < r ~ ≪ n r < \tilde{r} \ll n r < r ~ ≪ n . Provided that the series terms decay sufficiently rapidly to zero, or converge to a limiting value (noting that some contributions may stagnate as the dimension increases), the resulting quantity serves as a reliable error indicator while remaining computationally tractable.
The estimator simplifies for certain problem classes, for example when the matrices E i − 1 A i \mathbf{E}_i^{-1}\mathbf{A}_i E i − 1 A i are well-conditioned and have negative-definite Hermitian parts (see [24, Sec. 3.4]). In particular, if the SLS has identity state transitions and admits common balancing bases whose Gramians satisfy the required LMIs for all modes, the switching and LMI-defect corrections can be omitted. In this case, we recover the classical BT error bound
∥ y ( u ) − y ~ ( u ) ∥ L 2 ≤ c r ∥ u ∥ L 2 , c r : = 2 tr ( Σ 1 , x , 2 ) . (38) \lVert y(u)-\tilde{y}(u)\rVert_{L_2} \le c_r\lVert u\rVert_{L_2}, \qquad c_r := 2\operatorname{tr}(\Sigma_{1,x,2}). \tag*{(38)} ∥ y ( u ) − y ~ ( u ) ∥ L 2 ≤ c r ∥ u ∥ L 2 , c r := 2 tr ( Σ 1 , x , 2 ) . (38)
Since c r c_r c r can be precomputed offline, this bound can be evaluated online without trajectory-dependent correction terms; see [24, Thm. 2.4]. In the general case, a rigorous evaluation of the estimator can be computationally demanding, and the resulting bound may substantially overestimate the true reduction error; consequently, it is often preferable to trade strict rigor for improved computational efficiency.
Duality of the PBR for the optimal control problem
So far, we obtained the optimality system in Proposition 2.2 and the resulting error estimators presented in Theorems 2.4 and 2.5 by the first-reduce-then-optimize approach, illustrated by the dashed arrows in Figure 1 . In this approach, we first compute the ROM Σ ~ σ \widetilde{\Sigma}_{\sigma} Σ σ and then derive its optimality system, which includes the adjoint system Σ ~ σ T \widetilde{\Sigma}_{\sigma}^{\mathsf{T}} Σ σ T associated with the ROM. Alternatively, we first derive the optimality system for the FOM, yielding the adjoint system Σ σ T \Sigma_{\sigma}^{\mathsf{T}} Σ σ T , and then apply MOR to the adjoint system to obtain the reduced adjoint system Σ ~ σ T \widetilde{\Sigma}_{\sigma}^{\mathsf{T}} Σ σ T . This approach, called first-optimize-then-reduce , is illustrated by the solid arrows in Figure 1 . A natural question is whether these two approaches commute, that is, whether Σ ~ σ T ≡ Σ ~ σ T \widetilde{\Sigma}_{\sigma}^{\mathsf{T}} \equiv\widetilde{\Sigma}_{\sigma}^{\mathsf{T}} Σ σ T ≡ Σ σ T . The following theorem establishes that this is indeed the case for PBR.
Figure 1. First-optimize-then-reduce (solid path) vs. first-reduce-then-optimize (dashed path).
Theorem 3.3 (Commutativity of the OCP-PBR and PBR-OCP). Suppose that Assumptions 3.1 and 3.2 hold.
(i) Let ( P i , x , Q i , x ) i = 1 L (P_{i,x}, Q_{i,x})_{i=1}^{L} ( P i , x , Q i , x ) i = 1 L and ( P i , p , Q i , p ) i = 1 L (P_{i,p}, Q_{i,p})_{i=1}^{L} ( P i , p , Q i , p ) i = 1 L denote the reachability and observability matrices of (1) and (12) , respectively. Then
P i , p = Q i , x and Q i , p = P i , x for i = 1 , … , L . (39) P_{i,p}=Q_{i,x}\qquad\text{and}\qquad Q_{i,p}=P_{i,x}\qquad\text{for }i=1,\ldots,L. \tag*{(39)} P i , p = Q i , x and Q i , p = P i , x for i = 1 , … , L . (39) (ii) Let Σ ~ σ T \widetilde{\Sigma}_{\sigma}^{\mathrm{T}} Σ σ T denote the adjoint system corresponding to the PBR-ROM Σ ~ σ \widetilde{\Sigma}_{\sigma} Σ σ and let Σ ~ σ T \widetilde{\Sigma}_{\sigma}^{\mathrm{T}} Σ σ T denote the PBR-ROM for the adjoint system Σ σ T \Sigma_{\sigma}^{\mathrm{T}} Σ σ T . Then Σ ~ σ T ≡ Σ ~ σ T \widetilde{\Sigma}_{\sigma}^{\mathrm{T}}\equiv\widetilde{\Sigma}_{\sigma}^{\mathrm{T}} Σ σ T ≡ Σ σ T .
Proof. After reversing time in the adjoint system (12) , its L L L pairs of GLEs read
A i T P i , p E i + E i T P i , p A i + ∑ j = 1 L ( N i , j , p P i , p N i , j , p T + C i , j , p T C i , j , p ) = 0 , (40) A_i^{\mathrm{T}}P_{i,p}E_i+E_i^{\mathrm{T}}P_{i,p}A_i+\sum_{j=1}^{L}\left(N_{i,j,p}P_{i,p}N_{i,j,p}^{\mathrm{T}}+C_{i,j,p}^{\mathrm{T}}C_{i,j,p}\right)=0, \tag*{(40)} A i T P i , p E i + E i T P i , p A i + j = 1 ∑ L ( N i , j , p P i , p N i , j , p T + C i , j , p T C i , j , p ) = 0 , (40)
A i Q i , p E i T + E i Q i , p A i T + ∑ j = 1 L ( N i , j , p T Q i , p N i , j , p + B i , j , p B i , j , p T ) = 0 , (41) A_iQ_{i,p}E_i^{\mathrm{T}}+E_iQ_{i,p}A_i^{\mathrm{T}}+\sum_{j=1}^{L}\left(N_{i,j,p}^{\mathrm{T}}Q_{i,p}N_{i,j,p}+B_{i,j,p}B_{i,j,p}^{\mathrm{T}}\right)=0, \tag*{(41)} A i Q i , p E i T + E i Q i , p A i T + j = 1 ∑ L ( N i , j , p T Q i , p N i , j , p + B i , j , p B i , j , p T ) = 0 , (41)
where K j , i a d : = E j − T K i , j T E i T K_{j,i}^{\mathrm{ad}}:=E_j^{-\mathrm{T}}K_{i,j}^{\mathrm{T}}E_i^{\mathrm{T}} K j , i ad := E j − T K i , j T E i T denotes the full-order adjoint transition from mode i i i to mode j j j in reversed time, and for all i , j ∈ K i,j\in\mathcal{K} i , j ∈ K
B i , j , p : = ( K j , i a d ) T B j , B_{i,j,p}:=(K_{j,i}^{\mathrm{ad}})^{\mathrm{T}}B_j, B i , j , p := ( K j , i ad ) T B j ,
C i , j , p : = C j E j − 1 ( K i , j a d ) E i , (42) C_{i,j,p}:=C_jE_j^{-1}(K_{i,j}^{\mathrm{ad}})E_i, \tag*{(42)} C i , j , p := C j E j − 1 ( K i , j ad ) E i , (42)
N i , j , p : = E i T K i , j a d E j − T A j K j , i a d − A i T . N_{i,j,p}:=E_i^{\mathrm{T}}K_{i,j}^{\mathrm{ad}}E_j^{-\mathrm{T}}A_jK_{j,i}^{\mathrm{ad}}-A_i^{\mathrm{T}}. N i , j , p := E i T K i , j ad E j − T A j K j , i ad − A i T .
Substituting the adjoint transition matrices K j , i a d K_{j,i}^{\mathrm{ad}} K j , i ad into (42) , we find
B i , j , p = E i K i , j E j − 1 B j = B i , j , x , C i , j , p = C j E j − 1 E j K j , i E i − 1 E i = C i , j , x , B_{i,j,p}=E_iK_{i,j}E_j^{-1}B_j=B_{i,j,x},\qquad C_{i,j,p}=C_jE_j^{-1}E_jK_{j,i}E_i^{-1}E_i=C_{i,j,x}, B i , j , p = E i K i , j E j − 1 B j = B i , j , x , C i , j , p = C j E j − 1 E j K j , i E i − 1 E i = C i , j , x ,
N i , j , p = K j , i T A j E j − T K i , j E i T − A i T = ( E i K i , j E j − 1 A j K j , i − A i ) T = N i , j , x T . (43) N_{i,j,p}=K_{j,i}^{\mathrm{T}}A_jE_j^{-\mathrm{T}}K_{i,j}E_i^{\mathrm{T}}-A_i^{\mathrm{T}}=(E_iK_{i,j}E_j^{-1}A_jK_{j,i}-A_i)^{\mathrm{T}}=N_{i,j,x}^{\mathrm{T}}. \tag*{(43)} N i , j , p = K j , i T A j E j − T K i , j E i T − A i T = ( E i K i , j E j − 1 A j K j , i − A i ) T = N i , j , x T . (43)
Substituting the identities in (43) into (3.8) and invoking the uniqueness assumption (Assumption 3.2) yields (39) .
For the second statement, we observe that (i) allows us to choose S i , p = R i , x S_{i,p}=R_{i,x} S i , p = R i , x and R i , p = S i , x R_{i,p}=S_{i,x} R i , p = S i , x . Since the adjoint mass matrix is E i T E_i^{\mathrm{T}} E i T , its mass-weighted factor product satisfies
S i , p T E i R i , p = R i , x T E i S i , x = ( S i , x T E i T R i , x ) T . S_{i,p}^{\mathrm{T}}E_iR_{i,p}=R_{i,x}^{\mathrm{T}}E_iS_{i,x}=(S_{i,x}^{\mathrm{T}}E_i^{\mathrm{T}}R_{i,x})^{\mathrm{T}}. S i , p T E i R i , p = R i , x T E i S i , x = ( S i , x T E i T R i , x ) T .
Thus, choosing the singular vectors and retained subspaces consistently in (35) and (36) , we get
V i , x = S i , x U i , x , 1 Σ i , x , 1 − 1 / 2 = R i , p Ψ i , p , 1 Σ i , p , 1 − 1 / 2 = W i , p , V_{i,x}=S_{i,x}U_{i,x,1}\Sigma_{i,x,1}^{-1/2}=R_{i,p}\Psi_{i,p,1}\Sigma_{i,p,1}^{-1/2}=W_{i,p}, V i , x = S i , x U i , x , 1 Σ i , x , 1 − 1/2 = R i , p Ψ i , p , 1 Σ i , p , 1 − 1/2 = W i , p ,
W i , x = R i , x Ψ i , x , 1 Σ i , x , 1 − 1 / 2 = S i , p U i , p , 1 Σ i , p , 1 − 1 / 2 = V i , p . (44) W_{i,x}=R_{i,x}\Psi_{i,x,1}\Sigma_{i,x,1}^{-1/2}=S_{i,p}U_{i,p,1}\Sigma_{i,p,1}^{-1/2}=V_{i,p}. \tag*{(44)} W i , x = R i , x Ψ i , x , 1 Σ i , x , 1 − 1/2 = S i , p U i , p , 1 Σ i , p , 1 − 1/2 = V i , p . (44)
for i = 1 , … , L i=1,\ldots,L i = 1 , … , L . Now, let us state system Σ ~ σ T \widetilde{\Sigma}_{\sigma}^{\mathrm{T}} Σ σ T , i.e., the adjoint system of the ROM (8) , which reads as
Σ ~ σ T { − E ~ σ ( t ) T p ~ ˙ ( t ) = A ~ σ ( t ) T p ~ ( t ) + C ~ σ ( t ) T ( y ~ ( t ) − y d ( t ) + ξ ~ ( t ) ) , p ~ ( T ) = 0 , E ~ σ ( t − ) T p ~ ( t − ) = K ~ σ ( t + ) , σ ( t − ) T E ~ σ ( t + ) T p ~ ( t + ) , q ~ ( t ) = B ~ σ ( t ) T p ~ ( t ) . \widetilde{\Sigma}_{\sigma}^{\mathrm{T}}
\left\{
\begin{aligned}
-\widetilde{E}_{\sigma(t)}^{\mathrm{T}}\dot{\widetilde{p}}(t)&=\widetilde{A}_{\sigma(t)}^{\mathrm{T}}\widetilde{p}(t)+\widetilde{C}_{\sigma(t)}^{\mathrm{T}}\left(\widetilde{y}(t)-y_d(t)+\widetilde{\xi}(t)\right),\qquad\widetilde{p}(T)=0,\\
\widetilde{E}_{\sigma(t^-)}^{\mathrm{T}}\widetilde{p}(t^-)&=\widetilde{K}_{\sigma(t^+),\sigma(t^-)}^{\mathrm{T}}\widetilde{E}_{\sigma(t^+)}^{\mathrm{T}}\widetilde{p}(t^+),\\
\widetilde{q}(t)&=\widetilde{B}_{\sigma(t)}^{\mathrm{T}}\widetilde{p}(t).
\end{aligned}
\right. Σ σ T ⎩ ⎨ ⎧ − E σ ( t ) T p ˙ ( t ) E σ ( t − ) T p ( t − ) q ( t ) = A σ ( t ) T p ( t ) + C σ ( t ) T ( y ( t ) − y d ( t ) + ξ ( t ) ) , p ( T ) = 0 , = K σ ( t + ) , σ ( t − ) T E σ ( t + ) T p ( t + ) , = B σ ( t ) T p ( t ) .
with E ~ i \widetilde{E}_i E i , A ~ i \widetilde{A}_i A i , K ~ i , j \widetilde{K}_{i,j} K i , j , B ~ i \widetilde{B}_i B i , and C ~ i \widetilde{C}_i C i defined as in (1.9) via the projection matrices of (3.4). The ROM obtained via PBR for the adjoint system (2.1) reads as
Σ ~ T ⊤ { − E ^ σ ( t ) p ^ ˙ ( t ) = A ^ σ ( t ) p ^ ( t ) + C ^ σ ( t ) ( y ~ ( t ) − y d ( t ) + ξ ~ ( t ) ) , p ^ ( T ) = 0 , p ^ ( t − ) = K ^ σ ( t − ) , σ ( t + ) p ^ ( t + ) , q ^ ( t ) = B ^ σ ( t ) p ^ ( t ) , (45) \widetilde{\Sigma}_{\mathcal{T}}^{\top}
\left\{
\begin{aligned}
-\widehat{E}_{\sigma(t)}\dot{\widehat{p}}(t)&=\widehat{A}_{\sigma(t)}\widehat{p}(t)+\widehat{C}_{\sigma(t)}\left(\widetilde{y}(t)-y_d(t)+\widetilde{\xi}(t)\right),\quad\widehat{p}(T)=0,\\
\widehat{p}(t^-)&=\widehat{K}_{\sigma(t^-),\sigma(t^+)}\widehat{p}(t^+),\\
\widehat{q}(t)&=\widehat{B}_{\sigma(t)}\widehat{p}(t),
\end{aligned}
\right. \tag*{(45)} Σ T ⊤ ⎩ ⎨ ⎧ − E σ ( t ) p ˙ ( t ) p ( t − ) q ( t ) = A σ ( t ) p ( t ) + C σ ( t ) ( y ( t ) − y d ( t ) + ξ ( t ) ) , p ( T ) = 0 , = K σ ( t − ) , σ ( t + ) p ( t + ) , = B σ ( t ) p ( t ) , (45)
where for i , j = 1 , … , L i,j=1,\ldots,L i , j = 1 , … , L we have
E ^ j = W j , p ⊤ E j ⊤ V j , p , C ^ j = W j , p ⊤ C j ⊤ , A ^ j = W j , p ⊤ A j ⊤ V j , p , B ^ j = B j ⊤ V j , p , \widehat{E}_j=\mathbf{W}_{j,p}^{\top}E_j^{\top}V_{j,p},\quad\widehat{C}_j=\mathbf{W}_{j,p}^{\top}C_j^{\top},\quad\widehat{A}_j=\mathbf{W}_{j,p}^{\top}A_j^{\top}V_{j,p},\quad\widehat{B}_j=B_j^{\top}V_{j,p}, E j = W j , p ⊤ E j ⊤ V j , p , C j = W j , p ⊤ C j ⊤ , A j = W j , p ⊤ A j ⊤ V j , p , B j = B j ⊤ V j , p ,
and
K ^ j , i = E ^ j − 1 W j , p ⊤ E j ⊤ K j , i a d V i , p = E ^ j − 1 W j , p ⊤ K i , j ⊤ E i ⊤ V i , p . \widehat{K}_{j,i}=\widehat{E}_j^{-1}\mathbf{W}_{j,p}^{\top}E_j^{\top}K_{j,i}^{\mathrm{ad}}V_{i,p}=\widehat{E}_j^{-1}\mathbf{W}_{j,p}^{\top}K_{i,j}^{\top}E_i^{\top}V_{i,p}. K j , i = E j − 1 W j , p ⊤ E j ⊤ K j , i ad V i , p = E j − 1 W j , p ⊤ K i , j ⊤ E i ⊤ V i , p .
Finally, using (3.11) we get
E ~ j ⊤ = V j , x ⊤ E j ⊤ W j , x = W j , p ⊤ E j ⊤ V j , p = E ^ j , B ~ j ⊤ = B j ⊤ W j , x = B j ⊤ V j , p = B ^ j , \widetilde{E}_j^{\top}=V_{j,x}^{\top}E_j^{\top}W_{j,x}=\mathbf{W}_{j,p}^{\top}E_j^{\top}V_{j,p}=\widehat{E}_j,\quad\widetilde{B}_j^{\top}=B_j^{\top}W_{j,x}=B_j^{\top}V_{j,p}=\widehat{B}_j, E j ⊤ = V j , x ⊤ E j ⊤ W j , x = W j , p ⊤ E j ⊤ V j , p = E j , B j ⊤ = B j ⊤ W j , x = B j ⊤ V j , p = B j ,
A ~ j ⊤ = V j , x ⊤ A j ⊤ W j , x = W j , p ⊤ A j ⊤ V j , p = A ^ j , C ~ j ⊤ = V j , x ⊤ C j ⊤ = W j , p ⊤ C j ⊤ = C ^ j , \widetilde{A}_j^{\top}=V_{j,x}^{\top}A_j^{\top}W_{j,x}=\mathbf{W}_{j,p}^{\top}A_j^{\top}V_{j,p}=\widehat{A}_j,\quad\widetilde{C}_j^{\top}=V_{j,x}^{\top}C_j^{\top}=\mathbf{W}_{j,p}^{\top}C_j^{\top}=\widehat{C}_j, A j ⊤ = V j , x ⊤ A j ⊤ W j , x = W j , p ⊤ A j ⊤ V j , p = A j , C j ⊤ = V j , x ⊤ C j ⊤ = W j , p ⊤ C j ⊤ = C j ,
and
E ~ j − ⊤ K ~ i , j ⊤ E ~ i ⊤ = E ~ j − ⊤ ( E ~ i − 1 W i , x ⊤ E i K i , j V j , x ) ⊤ E ~ i ⊤ = E ^ j − 1 W j , p ⊤ K i , j ⊤ E i ⊤ V i , p = K ^ j , i , \widetilde{E}_j^{-\top}\widetilde{K}_{i,j}^{\top}\widetilde{E}_i^{\top}=\widetilde{E}_j^{-\top}\left(\widetilde{E}_i^{-1}\mathbf{W}_{i,x}^{\top}E_iK_{i,j}V_{j,x}\right)^{\top}\widetilde{E}_i^{\top}=\widehat{E}_j^{-1}\mathbf{W}_{j,p}^{\top}K_{i,j}^{\top}E_i^{\top}V_{i,p}=\widehat{K}_{j,i}, E j − ⊤ K i , j ⊤ E i ⊤ = E j − ⊤ ( E i − 1 W i , x ⊤ E i K i , j V j , x ) ⊤ E i ⊤ = E j − 1 W j , p ⊤ K i , j ⊤ E i ⊤ V i , p = K j , i ,
where we used V j , x = W j , p V_{j,x}=W_{j,p} V j , x = W j , p , W i , x = V i , p W_{i,x}=V_{i,p} W i , x = V i , p , and E ~ j ⊤ = E ^ j \widetilde{E}_j^{\top}=\widehat{E}_j E j ⊤ = E j in the second equality. This concludes the proof of the second statement.
Remark 3.4. For LTI systems in control form, Theorem 3.3 holds by replacing the PBR method with standard BT. For duality in the FOM, see, e.g., [18, Sec. 6.2.5].
Application to the OCP error bounds
A significant implication of Theorem 3.3 is that the projection matrices obtained by applying PBR to the forward SLS also reduce its adjoint system. Thus, no separate balancing computation is required for the adjoint, which is why reducing the OCP is as costly as reducing the SLS. Particularly, we can use the PBR error bound for the output of the FOM and its adjoint to bound their respective outputs in Theorems 2.4 and 2.5 .
Corollary 3.5 (PBR error bounds). In the setting of Theorems 2.4 and 2.5 , suppose that (3.5) applies to the forward and time-reversed adjoint systems. Define the primal and dual PBR estimators
Δ x : = Δ ( Σ x , 2 , x ~ ⋆ , λ 1 , u ~ ⋆ ) , Δ p : = Δ ( Σ p , 2 , p ~ ⋆ , λ 1 , p , y ~ ⋆ − y d + ξ ~ ⋆ ) . (46) \Delta_x:=\Delta(\Sigma_{x,2},\widetilde{x}^{\star},\lambda_1,\widetilde{u}^{\star}),\qquad\Delta_p:=\Delta(\Sigma_{p,2},\widetilde{p}^{\star},\lambda_{1,p},\widetilde{y}^{\star}-y_d+\widetilde{\xi}^{\star}). \tag*{(46)} Δ x := Δ ( Σ x , 2 , x ⋆ , λ 1 , u ⋆ ) , Δ p := Δ ( Σ p , 2 , p ⋆ , λ 1 , p , y ⋆ − y d + ξ ⋆ ) . (46) Then,
η ~ − ≤ η − ≤ J γ ( y ⋆ , u ⋆ ) − J γ ( y ~ ⋆ , u ~ ⋆ ) ≤ η + ≤ η ~ + , (47) \widetilde{\eta}_{-}\leq\eta_{-}\leq J_{\gamma}(y^{\star},u^{\star})-J_{\gamma}(\widetilde{y}^{\star},\widetilde{u}^{\star})\leq\eta_{+}\leq\widetilde{\eta}_{+}, \tag*{(47)} η − ≤ η − ≤ J γ ( y ⋆ , u ⋆ ) − J γ ( y ⋆ , u ⋆ ) ≤ η + ≤ η + , (47) where η − \eta_{-} η − and η + \eta_{+} η + are defined in (2.11) and
η ~ ± : = ± ∥ y ~ ⋆ − y d + ξ ~ ⋆ ∥ L 2 Δ x + { − 1 2 λ Δ p 2 , for η ~ − , 1 + γ 2 Δ x 2 , for η ~ + . (48) \widetilde{\eta}_{\pm}:=\pm\left\|\widetilde{y}^{\star}-y_d+\widetilde{\xi}^{\star}\right\|_{L^2}\Delta_x+
\begin{cases}
-\frac{1}{2\lambda}\Delta_p^2,&\text{for }\widetilde{\eta}_{-},\\
\frac{1+\gamma}{2}\Delta_x^2,&\text{for }\widetilde{\eta}_{+}.
\end{cases} \tag*{(48)} η ± := ± y ⋆ − y d + ξ ⋆ L 2 Δ x + { − 2 λ 1 Δ p 2 , 2 1 + γ Δ x 2 , for η − , for η + . (48) Moreover, the errors defined in Theorems 2.4 and 2.5 satisfy
e y ⋆ , u ⋆ ≤ Δ y ⋆ , u ⋆ ≤ Δ ~ y ⋆ , u ⋆ ( y ~ ⋆ , u ~ ⋆ , ξ ~ ⋆ ) : = ( 1 λ Δ p 2 + ( 1 + γ ) Δ x 2 ) 1 / 2 , e_{y^{\star},u^{\star}}\leq\Delta_{y^{\star},u^{\star}}\leq\widetilde{\Delta}_{y^{\star},u^{\star}}(\widetilde{y}^{\star},\widetilde{u}^{\star},\widetilde{\xi}^{\star}):=\left(\frac{1}{\lambda}\Delta_p^2+(1+\gamma)\Delta_x^2\right)^{1/2}, e y ⋆ , u ⋆ ≤ Δ y ⋆ , u ⋆ ≤ Δ y ⋆ , u ⋆ ( y ⋆ , u ⋆ , ξ ⋆ ) := ( λ 1 Δ p 2 + ( 1 + γ ) Δ x 2 ) 1/2 , e J γ ≤ Δ J γ ≤ Δ ~ J γ ( y ~ ⋆ , u ~ ⋆ , ξ ~ ⋆ ) : = max { ∣ η ~ − ∣ , ∣ η ~ + ∣ } , (49) e_{J_{\gamma}}\leq\Delta_{J_{\gamma}}\leq\widetilde{\Delta}_{J_{\gamma}}(\widetilde{y}^{\star},\widetilde{u}^{\star},\widetilde{\xi}^{\star}):=\max\{|\widetilde{\eta}_{-}|,|\widetilde{\eta}_{+}|\}, \tag*{(49)} e J γ ≤ Δ J γ ≤ Δ J γ ( y ⋆ , u ⋆ , ξ ⋆ ) := max { ∣ η − ∣ , ∣ η + ∣ } , (49) where the estimators Δ y ⋆ , u ⋆ \Delta_{y^{\star},u^{\star}} Δ y ⋆ , u ⋆ and Δ J γ \Delta_{J_{\gamma}} Δ J γ are defined in (2.5) and (2.13), respectively.
Proof. Applying (37) to the time-reversed adjoint system gives, for any y , ξ ∈ L 2 ( 0 , T ; R p ) \mathbf{y},\boldsymbol{\xi} \in L^2(0,T;\mathbb{R}^p) y , ξ ∈ L 2 ( 0 , T ; R p ) ,
∥ q ( y , ξ ) − q ~ ( y , ξ ) ∥ L 2 ≤ Δ ( Σ p , 2 , p ~ ( y , ξ ) , λ 1 , p , y − y d + ξ ) (50) \|\mathbf{q}(\mathbf{y},\boldsymbol{\xi})-\widetilde{\mathbf{q}}(\mathbf{y},\boldsymbol{\xi})\|_{L^2} \leq\Delta\left(\boldsymbol{\Sigma}_{p,2},\widetilde{\mathbf{p}}(\mathbf{y},\boldsymbol{\xi}),\lambda_{1,p},\mathbf{y}-\mathbf{y}_d+\boldsymbol{\xi}\right) \tag*{(50)} ∥ q ( y , ξ ) − q ( y , ξ ) ∥ L 2 ≤ Δ ( Σ p , 2 , p ( y , ξ ) , λ 1 , p , y − y d + ξ ) (50)
By Theorem 3.3 and (2.3d)–(2.3f), the reduced adjoint driven by y ~ ⋆ − y d + ξ ~ ⋆ \widetilde{\mathbf{y}}^\star-\mathbf{y}_d+\widetilde{\boldsymbol{\xi}}^\star y ⋆ − y d + ξ ⋆ has state p ~ ⋆ \widetilde{\mathbf{p}}^\star p ⋆ and output q ~ ⋆ \widetilde{\mathbf{q}}^\star q ⋆ . Thus,
∥ y ( u ~ ⋆ ) − y ~ ⋆ ∥ L 2 ≤ Δ x , ∥ q ( y ~ ⋆ , ξ ~ ⋆ ) − q ~ ⋆ ∥ L 2 ≤ Δ p . \|\mathbf{y}(\widetilde{\mathbf{u}}^\star)-\widetilde{\mathbf{y}}^\star\|_{L^2}\leq\Delta_x,\qquad\|\mathbf{q}(\widetilde{\mathbf{y}}^\star,\widetilde{\boldsymbol{\xi}}^\star)-\widetilde{\mathbf{q}}^\star\|_{L^2}\leq\Delta_p. ∥ y ( u ⋆ ) − y ⋆ ∥ L 2 ≤ Δ x , ∥ q ( y ⋆ , ξ ⋆ ) − q ⋆ ∥ L 2 ≤ Δ p .
Inserting these inequalities into (23) proves the control and output estimate by Theorem 2.4 . For the lower bound in (48) , the Cauchy–Schwarz inequality and the primal PBR estimate give
⟨ y ~ ⋆ − y d + ξ ~ ⋆ , y ( u ~ ⋆ ) − y ~ ⋆ ⟩ L 2 ≥ − ∥ y ~ ⋆ − y d + ξ ~ ⋆ ∥ L 2 Δ x . \langle\widetilde{\mathbf{y}}^\star-\mathbf{y}_d+\widetilde{\boldsymbol{\xi}}^\star,\mathbf{y}(\widetilde{\mathbf{u}}^\star)-\widetilde{\mathbf{y}}^\star\rangle_{L^2}\geq-\|\widetilde{\mathbf{y}}^\star-\mathbf{y}_d+\widetilde{\boldsymbol{\xi}}^\star\|_{L^2}\Delta_x. ⟨ y ⋆ − y d + ξ ⋆ , y ( u ⋆ ) − y ⋆ ⟩ L 2 ≥ − ∥ y ⋆ − y d + ξ ⋆ ∥ L 2 Δ x .
Together with the dual PBR estimate, this gives η ~ − ≤ η − \widetilde{\eta}_{-}\leq\eta_{-} η − ≤ η − by (29) , and Theorem 2.5 proves the lower bound in (47) . Turning to the upper bound, Cauchy–Schwarz bounds the same inner product from above by ∥ y ~ ⋆ − y d + ξ ~ ⋆ ∥ L 2 Δ x \|\widetilde{\mathbf{y}}^\star-\mathbf{y}_d+\widetilde{\boldsymbol{\xi}}^\star\|_{L^2}\Delta_x ∥ y ⋆ − y d + ξ ⋆ ∥ L 2 Δ x . Together with the primal PBR estimate, this gives η + ≤ η ~ + \eta_{+}\leq\widetilde{\eta}_{+} η + ≤ η + and proves the upper bound.
Remark 3.6. Corollary 3.5 is based on the a posteriori error bound (37) , which is associated with the specific MOR procedure described in § 3.1 . We emphasize that, more generally, any reduction methodology that is commutative with the optimality conditions for SLS, i.e., for which Theorem 3.3 (ii) holds and that is equipped with a certified error estimator can be used to establish corollaries of this type. Clearly, the computational cost of evaluating such an error bound should be favorable compared with the cost of evaluating the FOM at the optimal input provided by the ROM-OCP.
Computational procedure and uniform operator approximation
The offline–online computational procedure can be summarized as follows.
Offline phase
Solve the GLEs in (3.1), construct the balancing bases via (36) , build a hierarchy of ROMs for selected dimensions r ≤ n r\leq n r ≤ n , and prepare the data for the error estimator (37) . By Theorem 3.3 , the forward bases also determine the reduced adjoint, so no separate balancing computation is required. The discarded singular values coincide for both systems, but the adjoint trajectory and LMI-defect contributions must be treated separately. In the special case of (38) , precompute c r c_r c r . For fixed system matrices, the projection bases and, in this special case, c r c_r c r are independent of the parameter tuple μ \boldsymbol{\mu} μ .
Online phase
For the current parameter tuple μ \boldsymbol{\mu} μ , solve the ROM-OCP and evaluate the estimators in Corollary 3.5 . If the certificate exceeds the prescribed tolerance, increase r r r and repeat the solve with a larger precomputed ROM. In the special case of (38) , we have the online-efficient choice
Δ ~ y ⋆ , u ⋆ ( y ~ ⋆ , u ~ ⋆ , ξ ~ ⋆ ) = c r ( 1 λ ∥ y ~ ⋆ − y d + ξ ~ ⋆ ∥ L 2 2 + ( 1 + γ ) ∥ u ~ ⋆ ∥ L 2 2 ) 1 / 2 . (51) \widetilde{\Delta}_{y^\star,u^\star}(\widetilde{\mathbf{y}}^\star,\widetilde{\mathbf{u}}^\star,\widetilde{\boldsymbol{\xi}}^\star)=c_r\left(\frac{1}{\lambda}\|\widetilde{\mathbf{y}}^\star-\mathbf{y}_d+\widetilde{\boldsymbol{\xi}}^\star\|_{L^2}^2+(1+\gamma)\|\widetilde{\mathbf{u}}^\star\|_{L^2}^2\right)^{1/2}. \tag*{(51)} Δ y ⋆ , u ⋆ ( y ⋆ , u ⋆ , ξ ⋆ ) = c r ( λ 1 ∥ y ⋆ − y d + ξ ⋆ ∥ L 2 2 + ( 1 + γ ) ∥ u ⋆ ∥ L 2 2 ) 1/2 . (51)
With this bound, or a certified offline–online decomposition in the general PBR case, the procedure is online-adaptive and online-efficient. These are online-efficient PBR-based bounds, in contrast to the full-order-evaluation bounds in Theorems 2.4 and 2.5 . In this case, the neglected singular-value tail controls the uniform approximation of (7) .
Corollary 3.7 (Uniform approximation of the solution operator). Let U ‾ ⊆ R m \overline{U} \subseteq\mathbb{R}^{m} U ⊆ R m and Y ‾ ⊆ R p \overline{Y} \subseteq\mathbb{R}^{p} Y ⊆ R p be nonempty and compact, 0 < λ min ≤ λ max 0 < \lambda_{\min} \le\lambda_{\max} 0 < λ m i n ≤ λ m a x , 0 < γ min ≤ γ max 0 < \gamma_{\min} \le\gamma_{\max} 0 < γ m i n ≤ γ m a x , and T ‾ > 0 \overline{T} > 0 T > 0 . In addition to the standing assumptions on U a d U_{\mathrm{ad}} U ad and Y a d Y_{\mathrm{ad}} Y ad , define M \mathcal{M} M as the set of parameter tuples μ \mu μ from (6) satisfying σ ∈ P \sigma\in\mathcal{P} σ ∈ P , 0 < T ≤ T ‾ 0 < T \le\overline{T} 0 < T ≤ T , λ ∈ [ λ min , λ max ] \lambda\in[\lambda_{\min}, \lambda_{\max}] λ ∈ [ λ m i n , λ m a x ] , γ ∈ [ γ min , γ max ] \gamma\in[\gamma_{\min}, \gamma_{\max}] γ ∈ [ γ m i n , γ m a x ] , y d ∈ L 2 ( 0 , T ; Y ‾ ) y_{\mathrm{d}} \in L^{2}(0,T;\overline{Y}) y d ∈ L 2 ( 0 , T ; Y ) , u d ∈ L 2 ( 0 , T ; U ‾ ) u_{\mathrm{d}} \in L^{2}(0,T;\overline{U}) u d ∈ L 2 ( 0 , T ; U ) , and U a d ⊆ U ‾ U_{\mathrm{ad}} \subseteq\overline{U} U ad ⊆ U , Y a d ⊆ Y ‾ Y_{\mathrm{ad}} \subseteq\overline{Y} Y ad ⊆ Y . Further, assume that (38) holds uniformly, with c r = 2 tr ( Σ 1 , x , 2 ) c_{r} = 2\operatorname{tr}(\Sigma_{1,x,2}) c r = 2 tr ( Σ 1 , x , 2 ) independent of μ ∈ M \mu\in\mathcal{M} μ ∈ M . Then,
sup μ ∈ M Δ ~ y , u ⋆ ( μ ) ≤ C c r , \sup_{\mu\in\mathcal{M}} \widetilde{\Delta}_{y,u}^{\star}(\mu) \le C c_{r}, μ ∈ M sup Δ y , u ⋆ ( μ ) ≤ C c r , where C C C is independent of μ \mu μ and r r r . Hence, with G r ( μ ) : = ( y ~ ⋆ ( μ ) , u ~ ⋆ ( μ ) ) \mathcal{G}_{r}(\mu) := (\widetilde{y}^{\star}(\mu),\widetilde{u}^{\star}(\mu)) G r ( μ ) := ( y ⋆ ( μ ) , u ⋆ ( μ )) ,
sup μ ∈ M ∥ G ( μ ) − G r ( μ ) ∥ L 2 × L 2 ≤ C c r → 0 for r → n . \sup_{\mu\in\mathcal{M}} \left\|\mathcal{G}(\mu)-\mathcal{G}_{r}(\mu)\right\|_{L^{2}\times L^{2}} \le Cc_{r}\to0 \qquad\text{for }r\to n. μ ∈ M sup ∥ G ( μ ) − G r ( μ ) ∥ L 2 × L 2 ≤ C c r → 0 for r → n .
Proof. The uniformly bounded time horizons, admissible input sets and output references, and uniformly bounded reduced input-to-output operators imply
∥ u ~ ⋆ ( μ ) ∥ L 2 2 ≤ M u , ∥ y ~ ⋆ ( μ ) − y d + ξ ~ ⋆ ( μ ) ∥ L 2 2 ≤ M y . \left\|\widetilde{u}^{\star}(\mu)\right\|_{L^{2}}^{2}\le M_{u}, \qquad\left\|\widetilde{y}^{\star}(\mu)-y_{\mathrm{d}}+\widetilde{\xi}^{\star}(\mu)\right\|_{L^{2}}^{2}\le M_{y}. ∥ u ⋆ ( μ ) ∥ L 2 2 ≤ M u , y ⋆ ( μ ) − y d + ξ ⋆ ( μ ) L 2 2 ≤ M y .
Inserting these bounds into (51) gives
Δ ~ y , u ⋆ ( μ ) ≤ C 0 c r , C 0 2 = M y 2 λ min + ( 1 + γ max ) M u 2 . \widetilde{\Delta}_{y,u}^{\star}(\mu)\le C_{0}c_{r}, \qquad C_{0}^{2}=\frac{M_{y}^{2}}{\lambda_{\min}}+(1+\gamma_{\max})M_{u}^{2}. Δ y , u ⋆ ( μ ) ≤ C 0 c r , C 0 2 = λ m i n M y 2 + ( 1 + γ m a x ) M u 2 .
Using (22) , λ ≥ λ min \lambda\ge\lambda_{\min} λ ≥ λ m i n , and Corollary 3.5 it follows
∥ G ( μ ) − G r ( μ ) ∥ L 2 × L 2 2 = ∥ y ⋆ − y ~ ⋆ ∥ L 2 2 + ∥ u ⋆ − u ~ ⋆ ∥ L 2 2 ≤ max { 1 , λ min − 1 } C 0 2 c r 2 . \left\|\mathcal{G}(\mu)-\mathcal{G}_{r}(\mu)\right\|_{L^{2}\times L^{2}}^{2}=\left\|y^{\star}-\widetilde{y}^{\star}\right\|_{L^{2}}^{2}+\left\|u^{\star}-\widetilde{u}^{\star}\right\|_{L^{2}}^{2}\le\max\{1,\lambda_{\min}^{-1}\}C_{0}^{2}c_{r}^{2}. ∥ G ( μ ) − G r ( μ ) ∥ L 2 × L 2 2 = ∥ y ⋆ − y ⋆ ∥ L 2 2 + ∥ u ⋆ − u ⋆ ∥ L 2 2 ≤ max { 1 , λ m i n − 1 } C 0 2 c r 2 .
Thus, both claims hold with C = max { 1 , λ min − 1 / 2 } C 0 C=\max\{1,\lambda_{\min}^{-1/2}\}C_{0} C = max { 1 , λ m i n − 1/2 } C 0 , which is independent of μ \mu μ and r r r . □
Numerical Experiments
The purpose of this section is to provide a numerical validation of the a posteriori bounds established in § 2 and § 3 , and to evaluate both the approximation accuracy and the computational speed-ups achieved. All simulations were performed with MATLAB R2024b on a 2020 MacBook Pro with a 2.3 GHz quad-core Intel Core i7 processor and 16 GB of RAM. The source code is accessible here5 .
Computational details
We choose the time interval [ 0 , T ] = [ 0 , 1 ] [0,T]=[0,1] [ 0 , T ] = [ 0 , 1 ] and discretize it using the implicit Euler method with K = 500 K=500 K = 500 time steps, whereas the cost functional is discretized by a right-hand rectangle rule, making time discretization and optimization commute. As a solution method for the OCP, we use MY continuation with increased penalties γ k = min ( 2 k − 1 , γ max ) \gamma_{k}=\min(2^{k-1},\gamma_{\max}) γ k = min ( 2 k − 1 , γ m a x ) . Increasing γ k \gamma_{k} γ k progressively strengthens the penalization of violations of the output constraints. As an inner solver for fixed γ k \gamma_{k} γ k , we use a Barzilai–Borwein proximal gradient method [2 ] . The estimators are only evaluated once, after the last inner ROM OCP solve for the last γ k \gamma_{k} γ k . The parameters and stopping tolerances are specified below. We impose the pointwise box constraints u a ( t ) ≤ u ( t ) ≤ u b ( t ) u_{\mathrm{a}}(t)\le u(t)\le u_{\mathrm{b}}(t) u a ( t ) ≤ u ( t ) ≤ u b ( t ) and y a ( t ) ≤ y ( t ) ≤ y b ( t ) y_{\mathrm{a}}(t)\le y(t)\le y_{\mathrm{b}}(t) y a ( t ) ≤ y ( t ) ≤ y b ( t ) . The reported online speed-up is the FOM solution time divided by the sum of the ROM solution and estimator evaluation times, averaged over 5 repetitions.
Synthetic example
Consider a two-mode SLS of dimension n = 10 4 n = 10^{4} n = 1 0 4 with state matrices A 1 , A 2 ∈ R n × n \mathbf{A}_1,\mathbf{A}_2 \in\mathbb{R}^{n \times n} A 1 , A 2 ∈ R n × n defined by
A 1 ( i , j ) : = { − 1 , if i = j , 1 2 , if i − 1 = j , 0 , otherwise, A 2 ( i , j ) : = { − 2 , if i = j , 4 5 , if i − 1 = j , − 1 5 , if i + 1 = j , 0 , otherwise. (52) \mathbf{A}_1(i,j) := \begin{cases} -1, & \text{if } i=j, \\ \frac{1}{2}, & \text{if } i-1=j, \\ 0, & \text{otherwise,} \end{cases}
\qquad
\mathbf{A}_2(i,j) := \begin{cases} -2, & \text{if } i=j, \\ \frac{4}{5}, & \text{if } i-1=j, \\ -\frac{1}{5}, & \text{if } i+1=j, \\ 0, & \text{otherwise.} \end{cases} \tag*{(52)} A 1 ( i , j ) := ⎩ ⎨ ⎧ − 1 , 2 1 , 0 , if i = j , if i − 1 = j , otherwise, A 2 ( i , j ) := ⎩ ⎨ ⎧ − 2 , 5 4 , − 5 1 , 0 , if i = j , if i − 1 = j , if i + 1 = j , otherwise. (52)
The remaining matrices are E 1 = E 2 = I n \mathbf{E}_1=\mathbf{E}_2=\mathbf{I}_n E 1 = E 2 = I n , B 1 T = C 1 = [ 0 , … , 0 , 1 ] \mathbf{B}_1^\mathsf{T}=\mathbf{C}_1=[0,\ldots,0,1] B 1 T = C 1 = [ 0 , … , 0 , 1 ] , B 2 T = C 2 = [ 1 , 0 , … , 0 ] \mathbf{B}_2^\mathsf{T}=\mathbf{C}_2=[1,0,\ldots,0] B 2 T = C 2 = [ 1 , 0 , … , 0 ] , and K i , j = I n × n \mathbf{K}_{i,j}=\mathbf{I}_{n\times n} K i , j = I n × n . The system matrices are taken from [30]. We set u d ≡ 0 u_{\mathrm{d}}\equiv0 u d ≡ 0 and to construct the BT-ROM, we use the stopping tolerance ε = 10 − 11 \varepsilon=10^{-11} ε = 1 0 − 11 to compute the solution of GLEs, and precompute the ROM offline for r = 2 , … , 16 r=2,\ldots,16 r = 2 , … , 16 . Online, we treat the following two set-ups of μ \boldsymbol{\mu} μ introduced in (6) :
Configuration μ 1 \boldsymbol{\mu}_1 μ 1 : we choose the target y d = y ( u ~ ) \mathbf{y}_{\mathrm{d}}=\mathbf{y}(\tilde{\mathbf{u}}) y d = y ( u ~ ) with u ~ ≡ 1 \tilde{\mathbf{u}}\equiv1 u ~ ≡ 1 and set λ = 10 − 1 \lambda=10^{-1} λ = 1 0 − 1 . The bounds are u a ≡ 0.1 u_{\mathrm{a}}\equiv0.1 u a ≡ 0.1 , u b ≡ 0.6 u_{\mathrm{b}}\equiv0.6 u b ≡ 0.6 , y a ≡ 0 y_{\mathrm{a}}\equiv0 y a ≡ 0 , and y b ≡ 0.1 y_{\mathrm{b}}\equiv0.1 y b ≡ 0.1 . We use N = 5 N=5 N = 5 randomly chosen switches starting in mode 1, and 10 − 14 10^{-14} 1 0 − 14 as optimization tolerance.
Configuration μ 2 \boldsymbol{\mu}_2 μ 2 : we choose the target y d ( t ) = 2 + sin ( 2 π t ) \mathbf{y}_{\mathrm{d}}(t)=2+\sin(2\pi t) y d ( t ) = 2 + sin ( 2 π t ) and set λ = 10 − 3 \lambda=10^{-3} λ = 1 0 − 3 . The bounds are u a ≡ 0.01 u_{\mathrm{a}}\equiv0.01 u a ≡ 0.01 , u b ≡ 20 u_{\mathrm{b}}\equiv20 u b ≡ 20 , y a ≡ 0 y_{\mathrm{a}}\equiv0 y a ≡ 0 , and y b ≡ 2 y_{\mathrm{b}}\equiv2 y b ≡ 2 . We use N = 20 N=20 N = 20 randomly chosen switches starting in mode 2. The inner optimization tolerance is chosen as 10 − 13 10^{-13} 1 0 − 13 .
The results for μ 1 \boldsymbol{\mu}_1 μ 1 and μ 2 \boldsymbol{\mu}_2 μ 2 are shown in Figure 2 and Figure 3 , respectively. As r r r increases from 2 to 16, the combined error decreases from 8.6 ⋅ 10 − 4 8.6\cdot10^{-4} 8.6 ⋅ 1 0 − 4 to 7.8 ⋅ 10 − 14 7.8\cdot10^{-14} 7.8 ⋅ 1 0 − 14 for μ 1 \boldsymbol{\mu}_1 μ 1 , and from 1.2 ⋅ 10 − 2 1.2\cdot10^{-2} 1.2 ⋅ 1 0 − 2 to 4.9 ⋅ 10 − 13 4.9\cdot10^{-13} 4.9 ⋅ 1 0 − 13 for μ 2 \boldsymbol{\mu}_2 μ 2 . Figure 2 a and Figure 3 a compare the computed errors e γ e_{\gamma} e γ and e y μ ⋆ e_{\mathbf{y}^{\star}_{\boldsymbol{\mu}}} e y μ ⋆ with the full-order evaluation estimators from Theorems 2.4 and 2.5 and the BT-based indicator obtained from the expressions in Corollary 3.5 , using the simplified error estimate (38) , as this is applicable for this problem, as shown in [24, Sec. 5.1]. The FOM output-constraint violation ∥ y ⋆ − P Y a d y ⋆ ∥ L 2 \lVert\mathbf{y}^{\star}-P_{\mathcal{Y}_{\mathrm{ad}}}\mathbf{y}^{\star}\rVert_{L^2} ∥ y ⋆ − P Y ad y ⋆ ∥ L 2 is approximately 3.80 ⋅ 10 − 4 3.80\cdot10^{-4} 3.80 ⋅ 1 0 − 4 for μ 1 \boldsymbol{\mu}_1 μ 1 and 1.85 ⋅ 10 − 3 1.85\cdot10^{-3} 1.85 ⋅ 1 0 − 3 for μ 2 \boldsymbol{\mu}_2 μ 2 , both at γ = 250 \gamma=250 γ = 250 , which is negligible and, correspondingly, not discernible in Figure 2 d and Figure 3 d.
Figure 2. Synthetic example, configuration μ 1 \mu_1 μ 1 with MY regularization γ = 250 \gamma= 250 γ = 250 . The lower panels are for r = 8 r = 8 r = 8 . Gray dotted lines in the control panel mark the switching times.
Figure 3. (a) Errors and associated estimators. (b) Averaged speed-up including solution of the ROM-OCP and evaluation of the error estimators. (c) Optimal control and associated constraints. Switching times are indicated by dotted lines. (d) Optimal output and associated constraints. Note that the jumps in the output are precisely at the switching times. Figure 3. Synthetic example, configuration μ 2 \mu_2 μ 2 with MY regularization γ = 250 \gamma= 250 γ = 250 . The lower panels are for r = 8 r = 8 r = 8 . Gray dotted lines in the control panel mark the switching times.
Figure 2 a and Figure 3 a also show that the full-order-evaluation estimator for the combined error exceeds the computed error by factors of 5.3–11.1 for μ 1 \boldsymbol{\mu}_1 μ 1 and 16.9–27.6 for μ 2 \boldsymbol{\mu}_2 μ 2 . The value-function estimator almost coincides with the computed cost error, except at r = 2 r=2 r = 2 for μ 2 \boldsymbol{\mu}_2 μ 2 , where the ratio is 2.63. At the smallest cost errors, some computed estimates fall below the measured errors by less than 10 − 16 10^{-16} 1 0 − 16 in absolute terms, indicating the limit of numerical resolution. The BT-based estimators are larger than the corresponding errors by factors ranging from approximately 4.5 ⋅ 10 2 4.5\cdot10^{2} 4.5 ⋅ 1 0 2 to 4.9 ⋅ 10 5 4.9\cdot10^{5} 4.9 ⋅ 1 0 5 across both configurations. However, the ratios between the mean time needed for the evaluation of the FOM-based estimator and the mean time needed for the evaluation of the BT-based estimator, averaged over all dimensions and timing repetitions, are approximately 3000 for both configurations. Thus, while the BT-based error estimate is less sharp than its FOM-based counterpart, it is substantially faster to evaluate. It should be noted that, when the solution time of the ROM-OCP is included, the resulting computational advantage is substantially reduced. This is because the time required to solve the ROM-OCP dominates the online computational cost and, consequently, also exceeds the time needed to evaluate both the FOM-based and BT-based error estimators, particularly for μ 2 \boldsymbol{\mu}_2 μ 2 ; see Figure 2 b and Figure 3 b. Increasing the FOM size would, as expected, raise the computational burden of the FOM-based estimator, whereas it would essentially not affect the computational requirements of the BT-based estimator. The nonmonotone speed-up curve for μ 2 \mu_2 μ 2 reflects the varying numbers of Barzilai–Borwein iterations across reduced dimensions.
Sparse economic optimal control of the advection-diffusion-reaction equation
We consider the two-room heating example from [22, Sec. 2.3 and 2.5], which is inspired by [33]. The rectangular domain Ω ⊆ R 2 \Omega\subseteq\mathbb{R}^2 Ω ⊆ R 2 consists of the rooms Ω 1 \Omega_1 Ω 1 and Ω 3 \Omega_3 Ω 3 , the door Ω 2 \Omega_2 Ω 2 , and the walls Ω 4 \Omega_4 Ω 4 (see [22, Fig. 1] for details). Its boundary is partitioned into two disjoint sets Γ c \Gamma_{\mathrm{c}} Γ c and Γ o \Gamma_{\mathrm{o}} Γ o . The control u : [ 0 , 1 ) → R m \mathbf{u}\colon[0,1) \to\mathbb{R}^m u : [ 0 , 1 ) → R m acts on Γ c \Gamma_{\mathrm{c}} Γ c through m = 10 m = 10 m = 10 actuators. The switching signal σ : [ 0 , 1 ) → { 1 , 2 } \sigma\colon[0,1) \to\{1,2\} σ : [ 0 , 1 ) → { 1 , 2 } indicates whether the door is closed (σ = 1 \sigma= 1 σ = 1 ) or open (σ = 2 \sigma= 2 σ = 2 ). We fix two switches at t = 0.028 t = 0.028 t = 0.028 and t = 0.436 t = 0.436 t = 0.436 , with mode sequence 2 → 1 → 2 2 \to1 \to2 2 → 1 → 2 . The goal is to control the following switched heat equation
ζ ( x , σ ( t ) ) ∂ t θ ( t , x ) + A ( x , σ ( t ) ) θ ( t , x ) = 0 , ( t , x ) ∈ ( 0 , 1 ) × Ω , (53) \zeta(x,\sigma(t))\partial_t\theta(t,x)+\mathcal{A}(x,\sigma(t))\theta(t,x)=0,\qquad(t,x)\in(0,1)\times\Omega, \tag*{(53)} ζ ( x , σ ( t )) ∂ t θ ( t , x ) + A ( x , σ ( t )) θ ( t , x ) = 0 , ( t , x ) ∈ ( 0 , 1 ) × Ω , (53)
κ ( s , σ ( t ) ) ∂ ∂ n θ ( t , s ) + γ o θ ( t , s ) = 0 , ( t , s ) ∈ ( 0 , 1 ) × Γ o (54) \kappa(s,\sigma(t))\frac{\partial}{\partial n}\theta(t,s)+\gamma_o\theta(t,s)=0,\qquad(t,s)\in(0,1)\times\Gamma_o \tag*{(54)} κ ( s , σ ( t )) ∂ n ∂ θ ( t , s ) + γ o θ ( t , s ) = 0 , ( t , s ) ∈ ( 0 , 1 ) × Γ o (54)
κ ( s , σ ( t ) ) ∂ ∂ n θ ( t , s ) − ∑ i = 1 m χ i ( s ) u i ( t ) = 0 , ( t , s ) ∈ ( 0 , 1 ) × Γ c , (55) \kappa(s,\sigma(t))\frac{\partial}{\partial n}\theta(t,s)-\sum_{i=1}^{m}\chi_i(s)u_i(t)=0,\qquad(t,s)\in(0,1)\times\Gamma_c, \tag*{(55)} κ ( s , σ ( t )) ∂ n ∂ θ ( t , s ) − i = 1 ∑ m χ i ( s ) u i ( t ) = 0 , ( t , s ) ∈ ( 0 , 1 ) × Γ c , (55)
θ ( 0 , x ) = θ 0 ( x ) , x ∈ Ω , (56) \theta(0,x)=\theta_0(x),\qquad x\in\Omega, \tag*{(56)} θ ( 0 , x ) = θ 0 ( x ) , x ∈ Ω , (56)
with the differential operator
A ( x , σ ) θ ( t , x ) : = − ∇ ⋅ κ ( x , σ ) ∇ θ ( t , x ) + b ( x ) ⋅ ∇ θ ( t , x ) + c ( x ) θ ( t , x ) . \mathcal{A}(x,\sigma)\theta(t,x):=-\nabla\cdot\kappa(x,\sigma)\nabla\theta(t,x)+b(x)\cdot\nabla\theta(t,x)+c(x)\theta(t,x). A ( x , σ ) θ ( t , x ) := − ∇ ⋅ κ ( x , σ ) ∇ θ ( t , x ) + b ( x ) ⋅ ∇ θ ( t , x ) + c ( x ) θ ( t , x ) .
The coefficients ζ \zeta ζ and κ \kappa κ are piecewise constant, and switching occurs only on the door Ω 2 \Omega_2 Ω 2 (see [22, Sec. 4.2] for the precise definitions and parameter values). The output consists of the average temperatures in the two rooms Ω 1 \Omega_1 Ω 1 and Ω 3 \Omega_3 Ω 3 . Accordingly, we set p = 2 p=2 p = 2 and define C : L 2 ( Ω ) → R p C:L^2(\Omega)\to\mathbb{R}^p C : L 2 ( Ω ) → R p by φ ↦ y = ( y 1 , y 2 ) = C φ \varphi\mapsto y=(y_1,y_2)=C\varphi φ ↦ y = ( y 1 , y 2 ) = C φ , where y i : = ∣ D i ∣ − 1 ∫ D i φ ( x ) d x y_i:=|D_i|^{-1}\int_{D_i}\varphi(x)\,\mathrm{d}x y i := ∣ D i ∣ − 1 ∫ D i φ ( x ) d x for D 1 = Ω 1 D_1=\Omega_1 D 1 = Ω 1 and D 2 = Ω 3 D_2=\Omega_3 D 2 = Ω 3 . To promote sparse controls, we add an ℓ 1 \ell^1 ℓ 1 -term to the objective. For c 0 ≥ 0 c_0\geq0 c 0 ≥ 0 , c 1 > 0 c_1>0 c 1 > 0 , and a fixed shift s s s (cf. Remark 2.6), the computed objective is
J γ , s ( y , u ) = c 0 2 ∥ y − y d ∥ L 2 2 + λ 2 ∥ u ∥ L 2 2 + c 1 ∫ 0 1 ∥ u ( t ) ∥ ℓ 1 d t + γ 2 ∥ y + s / γ − P Y a d ( y + s / γ ) ∥ L 2 2 . J_{\gamma,s}(y,\mathbf{u})=\frac{c_0}{2}\lVert y-y_d\rVert_{L^2}^2+\frac{\lambda}{2}\lVert\mathbf{u}\rVert_{L^2}^2+c_1\int_0^1\lVert\mathbf{u}(t)\rVert_{\ell^1}\,\mathrm{d}t+\frac{\gamma}{2}\lVert y+s/\gamma-P_{Y_{\mathrm{ad}}}(y+s/\gamma)\rVert_{L^2}^2. J γ , s ( y , u ) = 2 c 0 ∥ y − y d ∥ L 2 2 + 2 λ ∥ u ∥ L 2 2 + c 1 ∫ 0 1 ∥ u ( t ) ∥ ℓ 1 d t + 2 γ ∥ y + s / γ − P Y ad ( y + s / γ ) ∥ L 2 2 .
The argument for convex control regularization in the proof of [22, Thm. 3] extends the error analysis from Theorem 2.4 to the additional ℓ 1 \ell^1 ℓ 1 -term. The value-function interval in Theorem 2.5 is enlarged by the Lipschitz constant c 1 m c_1\sqrt{m} c 1 m of the ℓ 1 \ell^1 ℓ 1 -norm term multiplied by the control-error bound, since T = 1 T=1 T = 1 .
Piecewise linear finite elements yield a discrete system of dimension n = 2170 n=2170 n = 2170 with nonsingular mass matrices E j \mathbf{E}_j E j and identity transition matrices K i , j \mathbf{K}_{i,j} K i , j . We compute the PBR models from Section 3 as follows: The generalized Lyapunov equations are solved by a low-rank iteration with LR-ADI inner solves with a tolerance of 10 − 10 10^{-10} 1 0 − 10 for the residual norm. The PBR-ROMs for each r = 10 , … , 60 r=10,\ldots,60 r = 10 , … , 60 are precomputed offline, and tested online for the following configurations:
Configuration μ 3 \mu_3 μ 3 : a staggered room-heating schedule with θ 0 = 0 \theta_0=0 θ 0 = 0 . Room 1 is brought to its operating temperature first, and room 2 is heated for later use after the door reopens. For S a , b ( t ) = 3 z 2 − 2 z 3 S_{a,b}(t)=3z^2-2z^3 S a , b ( t ) = 3 z 2 − 2 z 3 , where z = min { 1 , max { 0 , ( t − a ) / ( b − a ) } } z=\min\{1,\max\{0,(t-a)/(b-a)\}\} z = min { 1 , max { 0 , ( t − a ) / ( b − a )}} , we prescribe y d , 1 ( t ) = 9 10 S 0 , 0.15 ( t ) y_{d,1}(t)=\frac{9}{10}S_{0,0.15}(t) y d , 1 ( t ) = 10 9 S 0 , 0.15 ( t ) and y d , 2 ( t ) = S 0.45 , 0.65 ( t ) y_{d,2}(t)=S_{0.45,0.65}(t) y d , 2 ( t ) = S 0.45 , 0.65 ( t ) . We use c 0 = 100 c_0=100 c 0 = 100 , λ = 10 − 1 \lambda=10^{-1} λ = 1 0 − 1 , and c 1 = 10 − 4 c_1=10^{-4} c 1 = 1 0 − 4 . The boundary heaters satisfy 0 ≤ u i ≤ 3 0\leq u_i\leq3 0 ≤ u i ≤ 3 , and the output bounds are y a = ( 0 , 0 ) T y_a=(0,0)^{\mathsf{T}} y a = ( 0 , 0 ) T and y b = ( 0.95 , 1.1 ) T y_b=(0.95,1.1)^{\mathsf{T}} y b = ( 0.95 , 1.1 ) T .
Configuration μ 4 \mu_4 μ 4 : economic heating with staggered occupancy, θ 0 ≡ 3 \theta_0\equiv3 θ 0 ≡ 3 , c 0 = 0 c_0=0 c 0 = 0 , λ = 10 − 1 \lambda=10^{-1} λ = 1 0 − 1 , and c 1 = 10 − 4 c_1=10^{-4} c 1 = 1 0 − 4 . We minimize heating effort subject to 0 ≤ u i ≤ 3 0\leq u_i\leq3 0 ≤ u i ≤ 3 , y b ≡ ( 3 , 3 ) T y_b\equiv(3,3)^{\mathsf{T}} y b ≡ ( 3 , 3 ) T , and the time-dependent lower bounds y a , 1 ( t ) = 29 10 − 4 10 S 0.60 , 0.80 ( t ) y_{a,1}(t)=\frac{29}{10}-\frac{4}{10}S_{0.60,0.80}(t) y a , 1 ( t ) = 10 29 − 10 4 S 0.60 , 0.80 ( t ) and y a , 2 ( t ) = 23 10 + 4 10 S 0.50 , 0.75 ( t ) y_{a,2}(t)=\frac{23}{10}+\frac{4}{10}S_{0.50,0.75}(t) y a , 2 ( t ) = 10 23 + 10 4 S 0.50 , 0.75 ( t ) . The initial-state solution is subtracted from the bounds to obtain a zero-initial-state optimization problem and added back for plotting.
For each configuration, an augmented-Lagrangian iteration (see Remark 2.6) on a fixed r = 20 r=20 r = 20 ROM determines s s s , starting from the zero control and multiplier and using γ k = min ( 2 k − 1 , 1000 ) \gamma_k=\min(2^{k-1},1000) γ k = min ( 2 k − 1 , 1000 ) . We retain the shift used in its final subproblem. Every FOM and ROM comparison then solves one shifted MY problem with this same fixed s s s and γ = 1000 \gamma=1000 γ = 1000 , as in Remark 2.6.
For each optimal ROM control, the PBR approximation solves one enriched primal and adjoint ROM of dimension r ~ = 81 \tilde r=81 r ~ = 81 and computes their output differences to the optimization ROM. Insertion into the OCP estimates, including the ℓ 1 \ell^1 ℓ 1 -norm cost correction, gives Δ y ⋆ , u ⋆ P B R \Delta^{\mathrm{PBR}}_{y_\star,u_\star} Δ y ⋆ , u ⋆ PBR and Δ J P B R \Delta^{\mathrm{PBR}}_{J} Δ J PBR . The online calculation uses no FOM solve and has a state dimension independent of n n n . Again, we remark that approximating the trajectories and truncating the hierarchy at r ˉ = 90 < n \bar r=90<n r ˉ = 90 < n makes this an error indicator rather than a certified bound. The optimization tolerance is 10 − 10 10^{-10} 1 0 − 10 .
Figure 4 and Figure 5 display errors, estimators, speed-ups, and the r = 60 r=60 r = 60 trajectories, analogously to the synthetic example. A single FOM solution per configuration is the reference for all reduced dimensions. For μ 4 \mu_4 μ 4 with c 0 = 0 c_0=0 c 0 = 0 , the combined error contains only the control and multiplier errors, with no direct output-error contribution.
Figure 4. (a) Errors and associated estimators. (b) Averaged speed-up including solution of the ROM-OCP and evaluation of the error estimators. (c) Optimal control and associated constraints. Switching times are indicated by dotted lines. (d) Optimal output and associated constraints. Figure 4. Two-room heating example with n = 2170. Configuration μ₃: staggered temperature targets, nonnegative bounded heating, and an overheating limit in room 1. The FOM and ROM curves solve shifted MY problems at γ = 1000 with a common fixed shift from an r = 20 ROM. Gray dotted lines in the control panel mark the switching times.
Figure 5. Two-room heating example with n = 2170 n = 2170 n = 2170 . Configuration μ 4 \mu_4 μ 4 : minimum heating effort with staggered occupancy, time-dependent comfort bounds, and stored initial heat. The FOM and ROM curves solve shifted MY problems at γ = 1000 \gamma= 1000 γ = 1000 with a common fixed shift from an r = 20 r = 20 r = 20 ROM. Gray dotted lines in the control panel mark the switching times.
For μ 3 \mu_3 μ 3 , the combined error decreases from 6.43 ⋅ 10 − 3 6.43\cdot10^{-3} 6.43 ⋅ 1 0 − 3 at r = 10 r=10 r = 10 to 1.60 ⋅ 10 − 5 1.60\cdot10^{-5} 1.60 ⋅ 1 0 − 5 at r = 60 r=60 r = 60 . The FOM-evaluated estimates are 1.97 ⋅ 10 − 2 1.97\cdot10^{-2} 1.97 ⋅ 1 0 − 2 and 9.01 ⋅ 10 − 5 9.01\cdot10^{-5} 9.01 ⋅ 1 0 − 5 , while the PBR approximations are 1.97 ⋅ 10 − 2 1.97 \cdot10^{-2} 1.97 ⋅ 1 0 − 2 and 1.43 ⋅ 10 − 4 1.43 \cdot10^{-4} 1.43 ⋅ 1 0 − 4 . For μ 4 \mu_4 μ 4 , the combined error decreases from 1.23 ⋅ 10 − 3 1.23 \cdot10^{-3} 1.23 ⋅ 1 0 − 3 at r = 10 r = 10 r = 10 to 5.97 ⋅ 10 − 6 5.97 \cdot10^{-6} 5.97 ⋅ 1 0 − 6 at r = 60 r = 60 r = 60 . The FOM-evaluated estimates range from 7.12 ⋅ 10 − 3 7.12 \cdot10^{-3} 7.12 ⋅ 1 0 − 3 to 2.15 ⋅ 10 − 5 2.15 \cdot10^{-5} 2.15 ⋅ 1 0 − 5 , while the PBR approximations range from 7.14 ⋅ 10 − 3 7.14 \cdot10^{-3} 7.14 ⋅ 1 0 − 3 to 4.87 ⋅ 10 − 5 4.87 \cdot10^{-5} 4.87 ⋅ 1 0 − 5 . The PBR approximation retains a visible tail floor at the larger reduced dimensions. Both estimators exceed the measured combined and cost errors at every tested dimension.
The evaluation of the PBR error indicator takes about 16.8 ms for μ 3 \mu_3 μ 3 and 15.3 ms for μ 4 \mu_4 μ 4 , versus 368.9 ms and 347.4 ms for the error estimate based on the FOM evaluation with the optimal control and optimal output derived from the PBR-ROM. Including the ROM solve, the online speed-ups are 27.8–34.2 (PBR-based error indicator) versus 23.7–28.2 (FOM-based error estimate) for μ 3 \mu_3 μ 3 , and 27.1–37.2 versus 21.3–27.6 for μ 4 \mu_4 μ 4 . These online times exclude system projection, PBR preparation, shift generation, and the common Gramians computation, which takes 60.9 s with four computation threads. The control norms are ∥ u ⋆ ∥ L 2 = 5.3447 \lVert u^{\star}\rVert_{L^2} = 5.3447 ∥ u ⋆ ∥ L 2 = 5.3447 for μ 3 \mu_3 μ 3 and 4.1925 for μ 4 \mu_4 μ 4 . With activity threshold ∣ ξ i ⋆ ∣ > 10 − 8 |\xi_i^{\star}| > 10^{-8} ∣ ξ i ⋆ ∣ > 1 0 − 8 , the first-output MY multiplier in μ 3 \mu_3 μ 3 is active in 23.2% of time steps, while heaters reach their upper limit in 14.2% of actuator-time samples, and ∥ y ⋆ − y d ∥ L 2 = 0.1388 \lVert y^{\star} - y_{\mathrm{d}}\rVert_{L^2} = 0.1388 ∥ y ⋆ − y d ∥ L 2 = 0.1388 . In μ 4 \mu_4 μ 4 , the lower-bound MY multipliers are active in (10.6%, 5.0%) of time steps and the first-room upper-bound multiplier in 3.6%, while the heater upper limit remains inactive. Both hard-constrained discretized OCPs are feasible: zero control suffices for μ 3 \mu_3 μ 3 . An independent μ 4 \mu_4 μ 4 witness satisfies the input bounds and has output margin at least 2.45 ⋅ 10 − 3 2.45 \cdot10^{-3} 2.45 ⋅ 1 0 − 3 at all positive grid times, verified using the original finite-element matrices and the MATLAB FOM; the upper bound is attained at t = 0 t = 0 t = 0 . The shifted MY-FOM optima have maximum output-constraint violations 5.27 ⋅ 10 − 3 5.27 \cdot10^{-3} 5.27 ⋅ 1 0 − 3 and 7.05 ⋅ 10 − 3 7.05 \cdot10^{-3} 7.05 ⋅ 1 0 − 3 , respectively.
Conclusion
Conclusion. We developed a balancing-based ROM for MY-regularized, output-input-constrained optimal control of SLS with an unknown switching signal in the set (1.2) and potential state transition jumps at the switching times. By combining a posteriori error bounds for the optimal control, optimal output, and optimal value function, namely Theorems 2.4 and 2.5 , we obtain a certified approximation of the regularized OCP solution.
Subsequently, we propose employing the PBR-MOR framework as a ROM surrogate for the SLS. The resulting procedure incorporates an error estimate that can be leveraged to derive computationally efficient error indicators for the OCP; see § 3.1 and Corollary 3.5 . A key ingredient that allows this is a duality (or commutativity) result, which ensures that the reduced forward and adjoint systems can be obtained via the same balancing construction; see Theorem 3.3 .
Overall, by exploiting the offline–online decomposition of the error estimators and indicators, the proposed approach yields an online-efficient procedure for the reliable solution of the ROM-OCP.