Optimal shape of sensors or actuators for heat and wave equations - PowerPoint PPT Presentation

Optimal shape of sensors or actuators for heat and wave equations with random initial data Yannick Privat , Emmanuel Tr´ elat and Enrique Zuazua CNRS, LJLL, Univ. Paris 6 june 2015 Yannick Privat (LJLL) From Open to Closed Loop Control - Graz june 2015 1 / 27

Motivations What is the best shape and placement of sensors ? - Reduce the cost of instruments. - Maximize the efficiency of reconstruction and estimations. Yannick Privat (LJLL) From Open to Closed Loop Control - Graz june 2015 2 / 27

Outlines of this talk Introduction and motivation 1 Modeling of the problem : a randomized criterion 2 Optimal observability for wave and heat equations 3 Yannick Privat (LJLL) From Open to Closed Loop Control - Graz june 2015 3 / 27

Introduction and motivation N-D wave/heat equation → Ω open bounded connected subset of R n such that ∂ Ω � = ∅ ֒ ֒ → T > 0 fixed ֒ → ω ⊂ Ω subset of positive measure N-D wave equation N-D heat equation ∂ tt y − △ y = 0 in (0 , T ) × Ω ∂ t y − △ y = 0 in (0 , T ) × Ω y | ∂ Ω = 0 y | ∂ Ω = 0 y (0 , · ) = y 0 , ∂ t y (0 , · ) = y 1 in Ω y (0 , · ) = y 0 in Ω → Assume that ∂ Ω is C 2 (for simplicity, but can be → Generalization to many other boun- ֒ ֒ dary conditions (Neumann or mixed Dirichlet-Neumann or easily weakened, e.g. when Ω is a convex polygon) . → ∀ y 0 ∈ H 1 0 ∩ H 2 (Ω), there exists a ֒ Robin on ∂ Ω) → ∀ ( y 0 , y 1 ) ∈ L 2 (Ω) × H − 1 (Ω), there ∈ C 0 ([0 , T ] , H 1 0 ∩ H 2 (Ω)) ∩ ֒ unique y exists a unique y ∈ C 0 ([0 , T ] , L 2 (Ω)) ∩ C 1 ([0 , T ] , L 2 (Ω)) C 1 ([0 , T ] , H − 1 (Ω)). Observable variable ( ω ⊂ Ω of positive measure) � y ( t , x ) if x ∈ ω z ( t , x ) = χ ω ( x ) y ( t , x ) = 0 else. Yannick Privat (LJLL) From Open to Closed Loop Control - Graz june 2015 4 / 27

Introduction and motivation Observability of the N-D wave/heat equation Observability inequality (wave equation) The time T being chosen large enough, how to choose ω ⊂ Ω to ensure that � T � ∃ C > 0 | ∀ ( y 0 , y 1 ) ∈ H 1 0 (Ω) × L 2 (Ω) , C � ( y 0 , y 1 ) � 2 z ( t , x ) 2 dxdt ? 0 (Ω) × L 2 (Ω) ≤ H 1 0 Ω Microlocal Analysis. Bardos, Lebeau and Rauch proved that, roughly in the class of C ∞ domains, the observability inequality holds iff ( ω, T ) satisfies the GCC. � T ω y ( t , x ) 2 dxdt � C wave 0 Observability constant : = inf . T � ( y 0 , y 1 ) � 2 y solution of the wave eq. L 2 (Ω) × H − 1 (Ω) ( y 0 , y 1 ) ∈ L 2 (Ω) × H − 1 (Ω) Yannick Privat (LJLL) From Open to Closed Loop Control - Graz june 2015 5 / 27

Introduction and motivation Observability of the N-D wave/heat equation Observability inequality (heat equation) The time T being fixed, how to choose ω ⊂ Ω to ensure that � T � y ( t , x ) 2 dxdt , C � y ( T , · ) � 2 ∃ C > 0 | L 2 (Ω) ≤ 0 ω for every solution of the heat equation such that y (0 , · ) ∈ H 1 0 ∩ H 2 (Ω) ? ֒ → this ineq. holds for every open subset ω of Ω ; → related to the inverse problem of recovering its final data from the L 2 -observation of ֒ its solution on the set ω during a time T . � T ω | y ( t , x ) | 2 dx dt � C heat 0 ( χ ω ) = inf . T � y ( T , · ) � 2 y solution of the heat eq. L 2 (Ω) y (0 , · ) ∈ H 1 0 ∩ H 2 (Ω) Yannick Privat (LJLL) From Open to Closed Loop Control - Graz june 2015 5 / 27

Modeling of the problem : a randomized criterion Toward a new shape optimization problem The randomization procedure A first natural idea of modeling We investigate the problem of maximizing the quantity C T ( χ ω ) over all possible subsets ω ⊂ Ω of Lebesgue measure L | Ω | . BUT, two difficulties Theoretical difficulties 1 The model is not relevant w.r.t. practical expectation 2 The usual observability constant is deterministic and gives an account for the worst case. It is pessimistic. ֒ → In practice : many experiments, many measures. → Objective : optimize the sensor shape and location in average . → randomized observability constant. Yannick Privat (LJLL) From Open to Closed Loop Control - Graz june 2015 6 / 27

Modeling of the problem : a randomized criterion Toward a new shape optimization problem The randomization procedure Introduce ( − λ 2 j , φ j ), the j -th eigenpair of the Laplace-Dirichlet operator on Ω. ֒ → Spectral expansion of the solution y of the wave equation + ∞ � ∀ t ∈ (0 , T ) , y ( t , · ) = ( a j cos( λ j t ) + b j sin( λ j t )) φ j , j =1 where a j , b j are determined by the initial conditions. → Spectral expansion of the solution y of the heat equation ֒ + ∞ a j e − λ 2 � j t φ j , ∀ t ∈ (0 , T ) , y ( t , · ) = j =1 where a j is determined by the initial condition. Yannick Privat (LJLL) From Open to Closed Loop Control - Graz june 2015 6 / 27

Modeling of the problem : a randomized criterion Toward a new shape optimization problem The randomization procedure ֒ → Random selection of the initial data (see Burq - Tzvetkov, Invent. Math. 2008) : + ∞ 1 , j a j e i λ j t + β ν � 2 , j b j e − i λ j t � y ν ( t , x ) = � β ν φ j ( x ) , j =1 where ( β ν 1 , j ) j ∈ N ∗ and ( β ν 2 , j ) j ∈ N ∗ are two sequences of i.i.d random variables (Bernoulli/Gaussian) on a probability space ( X , A , P ) of mean 0. ֒ → Effects of the randomization Gaussian randomization : the map ν ∈ X �→ ( y ν (0 , · ) , y ν t (0 , · ) “generates” a full measure set in H s × H s − 1 (Ω) for almost every initial data ( y (0 , · ) , y t (0 , · )) ∈ H s × H s − 1 (Ω). Bernoulli randomization : keeps the H s × H s − 1 (Ω)-norm of the original function. No regularization effect : the map ν �→ y ν is measurable and [ ν �→ ( y ν , ∂ t y ν )] ∈ L 2 ( X , L 2 × H − 1 (Ω)) . ∈ H s + ε × H s − 1+ ε (Ω), then ( y ν , ∂ t y ν ) / ∈ H s + ε × H s − 1+ ε (Ω) almost If ( y (0 , · ) , y t (0 , · ) / surely. Yannick Privat (LJLL) From Open to Closed Loop Control - Graz june 2015 6 / 27

Modeling of the problem : a randomized criterion A randomized observability constant ֒ → Wave eq. : we consider the randomized observability inequality �� T � � y ν ( t , x ) 2 dxdt C T , rand ( χ ω ) � ( y 0 , y 1 ) � 2 0 × L 2 ≤ E , H 1 0 ω for all y 0 ( · ) ∈ L 2 (Ω) and y 1 ( · ) ∈ H − 1 (Ω), where y ν denotes the solution of the wave eq. with random initial data y 0 ,ν and y 1 ,ν . ֒ → Heat eq. : we consider the randomized observability inequality �� T � � y ν ( t , x ) 2 dx dt C T , rand ( χ ω ) � y ( T , · ) � 2 L 2 ≤ E , 0 ω 0 ∩ H 2 (Ω), where y ν denotes the solution of the heat equation with for all y (0 , · ) ∈ H 1 the random intial data y 0 ,ν . Proposition For every measurable set ω ⊂ Ω,  1 for the wave eq, � 2 φ j ( x ) 2 dx  e 2 λ 2 C T , rand ( χ ω ) = T inf j ∈ N ∗ γ j where γ j = j T − 1 for the heat eq. ω 2 λ 2  j There holds C T , rand ( χ ω ) ≥ C T ( χ ω ). There are examples where the inequality is strict. Yannick Privat (LJLL) From Open to Closed Loop Control - Graz june 2015 7 / 27

Optimal observability for wave and heat equations Optimal observability with respect to the domain Question What is the “best possible” observation domain ω of given measure ? Optimal design problem (energy concentration criterion) We investigate the problem of maximizing C T , rand ( χ ω ) � φ j ( x ) 2 dx . = inf j ∈ N ∗ γ j T ω over all possible subset ω ⊂ Ω of Lebesgue measure L | Ω | . Yannick Privat (LJLL) From Open to Closed Loop Control - Graz june 2015 8 / 27

Optimal observability for wave and heat equations Related problems Optimal design for control/stabilization problems What is the ”best domain” for achieving HUM optimal control ? 1 y tt − ∆ y = χ ω u What is the ”best domain” domain for stabilization (with localized damping) ? 2 y tt − ∆ y = − k χ ω y t See works by - P. H´ ebrard, A. Henrot : theoretical and numerical results in 1D for optimal stabilization (for all initial data). - A. M¨ unch, P. Pedregal, F. Periago : numerical investigations of the optimal domain (for one fixed initial data). Study of the relaxed problem. - S. Cox, P. Freitas, F. Fahroo, K. Ito, ... : variational formulations and numerics. - M.I. Frecker, C.S. Kubrusly, H. Malebranche, S. Kumar, J.H. Seinfeld, ... : numerical investigations (among a finite number of possible initial data). - K. Morris, S.L. Padula, O. Sigmund, M. Van de Wal, ... : numerical investigations for actuator placements (predefined set of possible candidates), Riccati approaches. - and many others. . . Yannick Privat (LJLL) From Open to Closed Loop Control - Graz june 2015 9 / 27

Optimal observability for wave and heat equations Additional remark Let A > 0 fixed. If we restrict the search to { ω ⊂ Ω | | ω | = L | Ω | and P Ω ( ω ) ≤ A } (perimeter) or { ω ⊂ Ω | | ω | = L | Ω | and � χ ω � BV (Ω) ≤ A } (total variation) or { ω ⊂ Ω | | ω | = L | Ω | and ω satisfies the 1 / A -cone property } or ω ranges over some finite-dimensional (or ”compact”) prescribed set... then there always exists (at least) one optimal set ω . → but then... - the complexity of ω may increase with A - we want to know if there is a ”very best” set (over all possible measurable) Yannick Privat (LJLL) From Open to Closed Loop Control - Graz june 2015 10 / 27