betaFeedback

WorksMathematics

Sparse optimal control for a class of infinite-dimensional bilinear systems: optimality conditions and uniqueness

Abstract

We study optimal control of an evolution equation in which a scalar control multiplies a bounded operator acting on the state, with free dynamics given by a C0C_0-semigroup. The cost combines tracking terms, a quadratic control cost and an L1L^1 term promoting sparsity in time. We prove existence, derive optimality conditions, and show that optimal controls are thresholded functions of a switching function whose derivative is governed by the commutator of the drift and control operators. Consequently, commuting problems with a terminal cost admit no temporally sparse controls. The L1L^1 term also matters for uniqueness: two distinct controls satisfying the optimality conditions differ on a set of measure bounded from below, contained in the union of their supports, which yields, for the first time, uniqueness for every horizon in stable problems steered to zero. Second-order sufficient conditions and experiments on a bilinear heat equation complete the study.

Introduction

In a bilinear control system the control multiplies the state instead of entering as an additive input. Such systems arise whenever an actuator changes a coefficient of the dynamics: a reaction rate in a diffusion process, the kill rate of a drug in a pharmacokinetic model, the strength of a dipole coupling in quantum mechanics, or the drift of a Fokker–Planck equation [20]. For bounded control operators the reachable set has dense complement in infinite dimensions [4], so exact steering is out of reach and optimal control is the natural way to formulate the task.

We consider the abstract bilinear system

y′(t)=Ay(t)+u(t)By(t),t∈(0,T),y(0)=y0,(1)y'(t) = Ay(t) + u(t)By(t), \qquad t \in(0,T), \qquad y(0) = y_0, \tag*{(1)}

where AA generates a C0C_0-semigroup on a Hilbert space HH, BB is a bounded operator on HH, and the scalar control uu takes values in an interval [a,b][a,b] containing 0. Writing yuy_u for the state associated with uu, we minimize

J(u)=γ12∥yu(T)−yT∥2+γ22∫0T∥yu(t)−yd(t)∥2 dt+α2∥u∥L2(0,T)2+β∥u∥L1(0,T)(2)J(u) = \frac{\gamma_1}{2}\lVert y_u(T) - y_T\rVert^2 + \frac{\gamma_2}{2}\int_0^T \lVert y_u(t) - y_d(t)\rVert^2\,\mathrm{d}t + \frac{\alpha}{2}\lVert u\rVert_{L^2(0,T)}^2 + \beta\lVert u\rVert_{L^1(0,T)} \tag*{(2)}

E-mail addresses: [email protected] (R. El Mezegueldy, corresponding author), [email protected] (Z. Dardour).

over the admissible set UadU_{\mathrm{ad}} of controls in L2(0,T)L^2(0,T) with values in [a,b][a,b], and call this problem

min⁡u∈UadJ(u).(3)\min_{u\in U_{\mathrm{ad}}} J(u). \tag*{(3)}

For β=0\beta=0 this is the classical problem with a quadratic control cost. For β>0\beta>0 the L1L^1 term charges the controller for every instant at which it acts, and optimal controls switch off on whole time intervals. This is the desired behaviour when actuation is expensive or harmful, as with drug doses or radio-frequency pulses, and it raises two questions: when does sparsity in time occur, and how does the L1L^1 term change the properties of the problem?

Sparse controls are preferred in engineering, but they make the mathematical problem harder. The L1L^1 norm is not differentiable, and it fails to be differentiable precisely at the controls of interest, those that vanish on whole intervals. The first-order condition then involves the subdifferential of the L1L^1 norm, the pointwise control law acquires kinks at the thresholds ±β\pm\beta, second-order conditions must take the kink of the cost at u=0u=0 into account, and Newton-type methods have to be replaced by nonsmooth ones. It is therefore not clear in advance which properties of the smooth problem survive.

We show that the classical theory carries over, and that the nonsmooth term also brings something that the smooth cost does not provide. The uniqueness results available for bilinear problems are local in time: they require the horizon TT to be small relative to the initial and desired data and to the regularization parameter. With the L1L^1 term, an exponentially stable system steered to zero has a unique optimal control for every horizon, provided α\alpha exceeds a threshold that depends neither on the horizon nor on the control bounds. For the smooth cost, our argument gives such a result only when the control bounds are small compared with the decay rate of the semigroup. This gain is somewhat surprising, because the L1L^1 term does not convexify the problem (Example 6.8). It comes instead from a bound on the L1L^1 norm of optimal controls that does not depend on the horizon, combined with the decay of state and adjoint.

Optimal control of bilinear distributed systems is well developed. The classical results give existence of optimal controls, a first-order optimality system, and uniqueness under a smallness condition on the final time [9, 7] or on the initial state [1]; bounded and unbounded control sets are treated by Zerrik et al. [38]. When the cost is linear in the control, Aronna et al. [2] derived second-order conditions for semigroup models through the Goh transform, and Casas et al. [16] analysed bang-bang bilinear control of elliptic equations.

Sparse optimal control was introduced in PDE-constrained optimization by Stadler [31] and has since been developed in many directions [37, 12, 23, 13, 24]. Closest to our setting are results on sparsity in time: for the stabilization of semilinear parabolic equations, an L1L^1-type cost makes the optimal control vanish after a finite time [14], and infinite-horizon sparse control of finite-dimensional systems is studied by Kalise et al. [25]. In control engineering, maximum hands-off control computes the sparsest control in time by L1L^1 minimization [28, 17], and L1L^1-minimal controls of control-affine systems are analysed by Vossen and Maurer [36]. For bilinear systems, L1L^1 costs have mostly been used in quantum control, where they produce sparse pulses [18] or fields supported on few frequencies [21], and where the L1L^1 norm of the field measures its effect on the state [8]. In biomedical models, objectives that are linear in the control are argued to describe drug cost and toxicity better than quadratic ones [26].

Two further lines of work are relevant here. The structure of an optimal control is read off its switching function σ=⟨By,p⟩\sigma=\langle By,p\rangle, where pp is the adjoint state; for control-affine ordinary differential equations, the derivative of the switching function is given by a Lie bracket of the drift and the control vector field [30]. Long horizons with zero targets lead to the turnpike phenomenon, in which optimal controls for different horizons nearly coincide away from the final time [33, 22]; infinite-horizon bilinear problems are treated by Breiten et al. [10].

To our knowledge, the abstract bilinear problem with a combined L2L^2–L1L^1 cost has not been analysed, and the effect of the L1L^1 term on uniqueness has not been examined. We state every result for β≥0\beta\ge0, so that each theorem shows what already holds for the smooth cost and what the L1L^1 term adds.

Our first contribution is the basic theory of the problem. We prove that optimal controls exist and derive an optimality system in which the optimal control is a projected soft-thresholding of the switching function, with explicit critical weights beyond which the zero control satisfies this optimality system or is the unique solution. Second, we establish an identity for the derivative of the switching function, the counterpart of the Lie-bracket formula, which shows that sparsity in time is driven by the commutator BA−ABBA - AB. When AA and BB commute and only the final state is penalized, every control in UadU_{\mathrm{ad}} that satisfies the optimality system is constant in time, so no sparsity in time can occur.

Our third and main contribution concerns uniqueness. We prove a separation estimate: two distinct controls in UadU_{\mathrm{ad}} that satisfy the optimality system differ on a set of measure at least α/Lσ\alpha/L_\sigma, where LσL_\sigma is the Lipschitz constant of the switching function with respect to the control in the L1L^1 norm, and this set lies in the union of their supports. For β>0\beta> 0 the supports are the sets where ∣σ∣|\sigma| exceeds β\beta, so short supports force uniqueness. For exponentially stable systems steered to zero, this yields a unique optimal control for every horizon, with constants independent of the horizon and of the control bounds. To our knowledge, this is the first uniqueness result for bilinear optimal control problems that holds on every time interval [0,T][0,T], whereas the earlier results [9, 1, 7] are local in time.

Fourth, we give second-order sufficient conditions together with an a posteriori test in which the measure of the support replaces the length of the horizon. Finally, we propose a positivity-preserving discretization with an exact discrete adjoint and report experiments on a bilinear heat equation that confirm these properties, including mesh independence and a direct check of the second-order condition.

The paper is organized as follows. § 2 states the standing assumptions and collects the properties of the state equation used throughout: well-posedness, Lipschitz dependence of the state on the control in the L1L^1 norm, continuity from weak to strong convergence, differentiability, and a duality identity between state and adjoint. § 3 proves the existence of optimal controls. § 4 derives the optimality system, the soft-thresholding law and the critical weights for the zero control. § 5 establishes the switching identity and treats the commuting case. § 6 contains the separation estimate and the uniqueness results, including uniqueness for every horizon in stable problems steered to zero, together with a scalar example of non-uniqueness. § 7 gives second-order sufficient conditions and the a posteriori test. § 8 reports numerical experiments on a bilinear heat equation, and § 9 concludes.

Setting and preliminaries

The analysis rests on four properties of the control-to-state map: an a priori bound, Lipschitz dependence on the control in the L1L^1 norm, continuity from weak to strong convergence, and smoothness. A duality identity between state and adjoint completes the toolbox. These facts are classical, and most of them go back to the analysis of distributed bilinear systems by Ball et al. [4]; we state them in the form used later, refer to the literature for their proofs, and give arguments only where explicit constants are needed in § 6 and § 7. For background on semigroups we refer to Engel and Nagel [19] and Pazy [29], and for optimal control of evolution equations to Li and Yong (1995) and Tröltzsch [34].

Throughout the paper, HH is a real Hilbert space with inner product ⟨⋅,⋅⟩\langle\cdot,\cdot\rangle and norm ∥⋅∥\|\cdot\|, and the operator A:D(A)⊂H→HA:D(A)\subset H\to H generates a C0C_0-semigroup (S(t))t≥0(S(t))_{t\geq0} on HH. We set MT:=sup⁡0≤t≤T∥S(t)∥M_T:=\sup_{0\leq t\leq T}\|S(t)\| and recall that the adjoint operators S(t)∗S(t)^* form a C0C_0-semigroup generated by A∗A^* [29 Section 1.10]. The control operator satisfies B∈L(H)B\in\mathcal{L}(H), the space of bounded linear operators on HH. The admissible set is Uad:={u∈L2(0,T):a≤u(t)≤b for a.e. t}U_{\mathrm{ad}}:=\{u\in L^2(0,T):a\leq u(t)\leq b\text{ for a.e. }t\}, where a≤0≤ba\leq0\leq b and a<ba<b, and we set ρ:=max⁡{−a,b}\rho:=\max\{-a,b\}. The data satisfy y0,yT∈Hy_0,y_T\in H and yd∈L2(0,T;H)y_d\in L^2(0,T;H), and the weights satisfy γ1,γ2≥0\gamma_1,\gamma_2\geq0, γ1+γ2>0\gamma_1+\gamma_2>0, α>0\alpha>0 and β≥0\beta\geq0. These standing assumptions are used in all results without further mention; additional hypotheses are stated where they are needed.

A mild solution of the state equation (1) is a function y∈C([0,T];H)y\in C([0,T];H) such that

y(t)=S(t)y0+∫0tS(t−s)u(s)By(s) ds,t∈[0,T].(4)y(t)=S(t)y_0+\int_0^t S(t-s)u(s)By(s)\,\mathrm{d}s,\qquad t\in[0,T]. \tag*{(4)}

We write J=F+βjJ=F+\beta j, where j(u):=∥u∥L1(0,T)j(u):=\|u\|_{L^1(0,T)} and FF collects the other three terms. The norm of C([0,T];H)C([0,T];H) or C([0,T])C([0,T]) is ∥⋅∥C\|\cdot\|_C, and ∣E∣|E| is the Lebesgue measure of a set E⊂[0,T]E\subset[0,T]. The constants

ET:=eMT∥B∥ρT,Ky:=MT∥y0∥ET,Ly:=MT∥B∥KyET(5)E_T:=e^{M_T\|B\|\rho T},\qquad K_y:=M_T\|y_0\|E_T,\qquad L_y:=M_T\|B\|K_yE_T \tag*{(5)}

appear throughout.

We first show that the state equation is well posed for every integrable control.

Lemma 2.1. For every u∈L1(0,T)u\in L^1(0,T), (1) has a unique mild solution yu∈C([0,T];H)y_u\in C([0,T];H), and

∥yu(t)∥≤MT∥y0∥exp⁡(MT∥B∥∫0t∣u(s)∣ ds),t∈[0,T].\|y_u(t)\|\leq M_T\|y_0\|\exp\left(M_T\|B\|\int_0^t|u(s)|\,\mathrm{d}s\right),\qquad t\in[0,T].

In particular ∥yu∥C≤Ky\|y_u\|_C\leq K_y for all u∈Uadu\in U_{\mathrm{ad}}.

Proof. For u∈L1(0,T)u\in L^1(0,T), existence and uniqueness of the mild solution in C([0,T];H)C([0,T];H) is a special case of Ball et al. [4], which allows controls in L1(0,T)L^1(0,T) and covers every bounded linear operator BB. For bounded controls, such as those in UadU_{\mathrm{ad}}, it is also a special case of the existence theory for semilinear evolution equations (Li and Yong, 1995, Chapter 2, Proposition 5.3). The bound follows from ∥yu(t)∥≤MT∥y0∥+MT∥B∥∫0t∣u(s)∣∥yu(s)∥ ds\|y_u(t)\|\leq M_T\|y_0\|+M_T\|B\|\int_0^t|u(s)|\|y_u(s)\|\,\mathrm{d}s and Gronwall’s lemma.

The state depends Lipschitz continuously on the control, measured in the L1L^1 norm. This is a quantitative form of the continuity of u↦yuu\mapsto y_u from L1(0,T)L^1(0,T) into C([0,T];H)C([0,T];H) contained in Ball et al. [4]; we give the short proof because the constant LyL_y is used in § 6.

References

  1. [1]Addou, A., & Benbrik, A. (2002). Existence and uniqueness of optimal control for a distributed-parameter bilinear system. Journal of Dynamical and Control Systems, 8, 141–152.DOI
  2. [2]Aronna, M. S., Bonnans, J. F., & Kröner, A. (2018). Optimal control of infinite dimensional bilinear systems: Application to the heat and wave equations. Mathematical Programming, 168, 717–757.
  3. [3]Ball, J. M. (1977). Strongly continuous semigroups, weak solutions, and the variation of constants formula. Proceedings of the American Mathematical Society, 63, 370–373.
  4. [4]Ball, J. M., Marsden, J. E., & Slemrod, M. (1982). Controllability for distributed bilinear systems. SIAM Journal on Control and Optimization, 20, 575–597.DOI
  5. [5]Bauschke, H. H., & Combettes, P. L. (2017). Convex analysis and monotone operator theory in Hilbert spaces (2nd ed.). Springer.
  6. [6]Beck, A., & Teboulle, M. (2009). A fast iterative shrinkage-thresholding algorithm for linear inverse problems. SIAM Journal on Imaging Sciences, 2, 183–202.
  7. [7]Bouhamed, A., El Kabouss, A., & Bouzahir, H. (2022). Bilinear boundary optimal control of a Kirchhoff plate equation. Results in Control and Optimization, 9, 100179.
  8. [8]Boussaïd, N., Caponigro, M., & Chambrion, T. (2013). Which notion of energy for bilinear quantum systems? arXiv preprint arXiv:1302.0815.
  9. [9]Bradley, M. E., & Lenhart, S. (1997). Bilinear optimal control of a Kirchhoff plate to a desired profile. Optimal Control Applications and Methods, 18, 217–226.
  10. [10]Breiten, T., Kunisch, K., & Pfeiffer, L. (2018). Infinite-horizon bilinear optimal control problems: Sensitivity analysis and polynomial feedback laws. SIAM Journal on Control and Optimization, 56, 3184–3214.
  11. [11]Byrd, R. H., Lu, P., Nocedal, J., & Zhu, C. (1995). A limited memory algorithm for bound constrained optimization. SIAM Journal on Scientific Computing, 16, 1190–1208.
  12. [12]Casas, E., Herzog, R., & Wachsmuth, G. (2012). Optimality conditions and error analysis of semilinear elliptic control problems with L¹ cost functional. SIAM Journal on Optimization, 22, 795–820.
  13. [13]Casas, E., Herzog, R., & Wachsmuth, G. (2017). Analysis of spatio-temporally sparse optimal control problems of semilinear parabolic equations. ESAIM: Control, Optimisation and Calculus of Variations, 23, 263–295.
  14. [14]Casas, E., & Kunisch, K. (2017). Stabilization by sparse controls for a class of semilinear parabolic equations. SIAM Journal on Control and Optimization, 55, 512–532.
  15. [15]Casas, E., & Tröltzsch, F. (2012). Second order analysis for optimal control problems: Improving results expected from abstract theory. SIAM Journal on Optimization, 22, 261–279.
  16. [16]Casas, E., Wachsmuth, D., & Wachsmuth, G. (2018). Second-order analysis and numerical approximation for bang-bang bilinear control problems. SIAM Journal on Control and Optimization, 56, 4203–4227.
  17. [17]Chatterjee, D., Nagahara, M., Quevedo, D. E., & Rao, K. S. M. (2016). Characterization of maximum hands-off control. Systems & Control Letters, 94, 31–36.DOI
  18. [18]Ciaramella, G., & Borzì, A. (2016). Quantum optimal control problems with a sparsity cost functional. Numerical Functional Analysis and Optimization, 37, 938–965.
  19. [19]Engel, K.-J., & Nagel, R. (2000). One-parameter semigroups for linear evolution equations. Springer.
  20. [20]Fleig, A., & Guglielmi, R. (2017). Optimal control of the Fokker–Planck equation with space-dependent controls. Journal of Optimization Theory and Applications, 174, 408–427.
  21. [21]Friescke, G., Henneke, F., & Kunisch, K. (2018). Frequency-sparse optimal quantum control. Mathematical Control and Related Fields, 8, 155–176.
  22. [22]Grüne, L., Schaller, M., & Schiela, A. (2020). Exponential sensitivity and turnpike analysis for linear quadratic optimal control of general evolution equations. Journal of Differential Equations, 268, 7311–7341.
  23. [23]Herzog, R., Stadler, G., & Wachsmuth, G. (2012). Directional sparsity in optimal control of partial differential equations. SIAM Journal on Control and Optimization, 50, 943–963.DOI
  24. [24]Hoppe, F., & Neitzel, I. (2022). Purely time-dependent optimal control of quasilinear parabolic PDEs with sparsity enforcing penalization. ESAIM: Control, Optimisation and Calculus of Variations, 28, 68.DOI
  25. [25]Kalise, D., Kunisch, K., & Rao, Z. (2017). Infinite horizon sparse optimal control. Journal of Optimization Theory and Applications, 172, 481–517.DOI
  26. [26]Ledzewicz, U., & Schä​ttler, H. (2020). On the role of the objective in the optimization of compartmental models for biomedical therapies. Journal of Optimization Theory and Applications, 187, 305–335.DOI
  27. [28]Nagahara, M., Quevedo, D. E., & Nešić, D. (2016). Maximum hands-off control: A paradigm of control effort minimization. IEEE Transactions on Automatic Control, 61, 735–747.DOI
  28. [29]Pazy, A. (1983). Semigroups of linear operators and applications to partial differential equations. Springer.DOI
  29. [30]Schättler, H., & Ledzewicz, U. (2012). Geometric optimal control: Theory, methods and examples. Springer.
  30. [31]Stadler, G. (2009). Elliptic optimal control problems with L¹-control cost and applications for the placement of control devices. Computational Optimization and Applications, 44, 159–181.
  31. [32]Tikhonov, A. N., Leonov, A. S., & Yagola, A. G. (1998). Nonlinear ill-posed problems (Vols. 1–2). Chapman & Hall.
  32. [33]Trélat, E., & Zuazua, E. (2015). The turnpike property in finite-dimensional nonlinear optimal control. Journal of Differential Equations, 258, 81–114.arxiv.org/abs/1402.3263
  33. [34]Tröltzsch, F. (2010). Optimal control of partial differential equations: Theory, methods and applications. American Mathematical Society.
  34. [35]Ulbrich, M. (2011). Semismooth Newton methods for variational inequalities and constrained optimization problems in function spaces. SIAM.
  35. [36]Vossen, G., & Maurer, H. (2006). On L¹-minimization in optimal control and applications to robotics. Optimal Control Applications and Methods, 27, 301–321.
  36. [37]Wachsmuth, G., & Wachsmuth, D. (2011). Convergence and regularization results for optimal control problems with sparsity functional. ESAIM: Control, Optimisation and Calculus of Variations, 17, 858–886.DOI
  37. [38]Zerrik, E. H., Ait Aadi, A., & Ouhafsa, M. (2022). Bounded and unbounded optimal control for infinite dimensional bilinear systems. International Journal of Control, 95, 3336–3343.DOI

Paper details

vlnc:W_cds53aqz
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents