Universität Konstanz
Trust-Region POD using A-Posteriori Error Estimation for Semilinear Parabolic Optimal
Control Problems
Sabrina Rogg Stefan Trenz Stefan Volkwein
Konstanzer Schriften in Mathematik Nr. 359, März 2017
ISSN 1430-3558
© Fachbereich Mathematik und Statistik Universität Konstanz
Fach D 197, 78457 Konstanz, Germany
Konstanzer Online-Publikations-System (KOPS) URL: http://nbn-resolving.de/urn:nbn:de:bsz:352-0-401106
Trust-Region POD using A-Posteriori Error Estimation for Semilinear Parabolic Optimal Control Problems
Sabrina Rogg, Stefan Trenz, and Stefan Volkwein
AbstractAn optimal control problem governed by a semilinear heat equation is solved using a globalized inexact Newton method. To reduce the computational ef- fort a model order reduction approach based on proper orthogonal decomposition (POD) is applied. Within a trust region framework we guarantee that the reduced- order models are sufficiently accurate by ensuring gradient accuracy. The gradient error is successfully monitored by an a-posteriori error estimate. Numerical results are presented and discussed.
1 Introduction
Optimal control problems for partial differential equations are often hard to tackle numerically because their discretization leads to very large scale optimization prob- lems. Therefore different techniques of model reduction were developed to approx- imate these problems by smaller ones that are tractable with less effort. Recently the application ofreduced-order modelsto linear time varying and nonlinear sys- tems, in particular to nonlinear control systems, has received an increasing amount of attention. The reduced-order approach is based on projecting the dynamical sys- tem onto subspaces consisting of basis elements that contain characteristics of the expected solution. This is in contrast to, e.g., finite element techniques, where the basis elements of the subspaces do not relate to the physical properties of the system that they approximate. Thereduced basis(RB) method, as developed in [26] and [35, 17], is one such reduced-order method, where the basis elements correspond to the dynamics of expected control regimes. Let us refer to the [11, 23, 32, 34] for the successful use of reduced basis method in PDE constrained optimization problems.
CurrentlyProper orthogonal decomposition(POD) is probably the mostly used and Sabrina Rogg, Stefan Trenz and Stefan Volkwein
Universit¨at Konstanz, Fachbereich Mathematik und Statistik, Universit¨atsstraße 10, 78457 Kon- stanz, Germany, e-mail:{Sabrina.Rogg, Stefan.Trenz, Stefan.Volkwein}@uni-konstanz.de
1
most successful model reduction technique for nonlinear optimal control problems, where the basis functions contain information from the solutions of the dynamical system at pre-specified time-instances, so-called snapshots. Due to a possible linear dependence or almost linear dependence the snapshots themselves are not appropri- ate as a basis. Hence a singular value decomposition is carried out and the leading generalized eigenfunctions are chosen as a basis, referred to as the POD basis. POD is successfully used in a variety of fields including fluid dynamics, coherent struc- tures [2, 4] and inverse problems [7]. Moreover in [6] POD is successfully applied to compute reduced-order controllers. The relationship between POD and balancing was considered in [30, 39, 48]. An error analysis for non-linear dynamical systems in finite dimensions were carried out in [37] and a missing point estimation in mod- els described by POD was studied in [5]. Let us also mention that POD and the reduced basis method are successfully combined by variants of the POD greedy algorithm; see [21, 22], for instance.
In this paper we consider an optimal control problem governed by a semilin- ear parabolic equation together with control constraints. For the numerical solution we apply a POD Galerkin approximation. However, to obtain the state data under- lying the POD model, it is necessary to solve once the full state system using a reference control. Consequently, the POD approximations depend on the chosen reference control, so that the choice of a reference control turned out to be essential for the computation of a POD basis for the optimal control problem. To overcome this problem we investigate theTrust-Region POD method[4, 13, 38, 40, 41] as a basis update strategies for improving the POD basis. In this strategy the POD basis is changed in the optimization method in order to ensure convergence and a certain accuracy for the obtained controls. For that reason we apply an a-posteriori error estimation of the POD approximation. Let us also refer to the papers [49] and [36], where the trust-region optimization is efficiently combined with the reduced-basis method for linear elliptic and parabolic problems without control constraints. More- over,optimality system PODas a further basis update strategy has been introduced in [29] and numerically tested in [15, 18, 46].
The paper is organized as follows: In Section 2 the optimal control problem is introduced. First and second-order necessary optimality conditions are studied in Section 3. The finite element Galerkin approximation is explained in Section 4. In Section 5 the reduced-order modeling is described. The a-posteriori error analysis for the POD approximation is presented and illustrated by numerical tests in Sec- tion 6. Finally, the TR-POD method is studied in Section 6, where also numerical experiments are shown.
2 The optimal control problem
Suppose thatΩ⊂Rd,d∈ {1,2,3}, is a bounded domain with Lipschitz-continuous boundary Γ =∂ Ω. For T >0 we set Q= (0,T)×Ω andΣ = (0,T)×Γ. Let b1, . . . ,bNu ∈L∞(Ω) be given nonnegative shape functions with Nu∈N. We set
V =H1(Ω)andH=L2(Ω), whereV is endowed with the standard inner product hϕ,ϕi˜ V =
Z
Ω
ϕϕ˜+∇ϕ·∇ϕ˜dx forϕ,ϕ˜∈V.
For the definition of Sobolev spaces we refer, e.g., to [1, 12]. The Bochner space L2(0,T;V)denotes the space of (equivalence classes) of measurable abstract func- tionsϕ:[0,T]→V, which are square integrable, i.e.,
Z T 0
kϕ(t)kV2dt<∞.
Whentis fixed, the expressionϕ(t)stands for the functionϕ(t,·)considered as a function inΩ only. Recall that
W(0,T) =
ϕ∈L2(0,T;V):|ϕt∈L2(0,T;V0)
is a Hilbert space supplied with its common inner product. In particular, we have Z T
0
hϕt(t),χ(t)iV0,V+hχt(t),ϕ(t)iV0,Vdt= Z T
0
d
dthϕ(t),χ(t)iHdt
=hϕ(T),χ(T)iH− hϕ(0),χ(0)iH
for everyϕ,χ∈W(0,T), whereh·,·iV0,Vdenotes the dual pairing betweenV and its dual spaceV0. For the details we refer the reader to [10, pp. 472-479], for instance.
Let us consider the following optimal control problem:
min
(y,u)J(y,u) =1
2ky(T)−yΩk2H+κ
2kuk2U (1a)
subject to the semilinear evolution problem cpyt(t,x)−∆y(t,x) +cny(t,x)3=
Nu
k=1
∑
uk(t)bk(x) +f(t,x), (t,x)∈Q,
∂y
∂n(t,s) =0, (t,s)∈Σ, y(0,x) =y◦(x), x∈Ω,
(1b)
and to bilateral control constraints u∈Uad=
˜
u∈U:ua≤u˜≤ubalmost everywhere (a.e.) in[0,T] (1c) withU=L2(0,T;RNu). In (1a) we assume that the desired states satisfyyΩ∈L∞(Ω) andκ>0. In (1b) let f ∈Lr(Q)withr>d/2+1,y◦∈L∞(Ω)andcp>0,cn>0.
In (1c),ua= (uai)1≤i≤Nu,ub= (ubi)1≤i≤Nu∈L∞(0,T;RNu)satisfyua(t)≤ub(t)for almost all (f.a.a.)t∈[0,T]and ’≤’ is interpreted componentwise. In particular, we haveUad⊂L∞(0,T;RNu).
The functiony∈Y =W(0,T)∩L∞(Q)is called a weak solution to (1b) provided y(0) =y◦holds inHand
cp d
dthy(t),ϕiH+ Z
Ω
∇y(t)·∇ϕ+cny(t)3ϕdx
= Z
Ω
f(t) +
Nu
∑
i=1
ui(t)bi
ϕdx for allϕ∈V and f.a.a.t∈(0,T].
(2)
It follows from [43, Theorem 5.5] that for any controlu∈Uadthere exists a unique statey=y(u)∈Y solving (2). Hence, we can introduce thereduced cost functional by
J(u) =ˆ J(y(u),u) foru∈Uad.
Instead of (1) we consider now thereduced optimal control problem
min ˆJ(u) subject to (s.t.) u∈Uad. (P)ˆ The next result is proved in [43, Theorem 5.7], for instance.
Theorem 1.Problem(P)ˆ admits a (global) solution.
Sincecn>0 holds, (P) is a nonconvex programming problem. Therefore, differentˆ local minima might occur. Numerical methods will deliver a local minimum close to their starting point. Throughout this paper we will assume that a fixed reference solution ¯u∈Uad is given satisfying first-and second-order optimality conditions (ensuring local optimality of the solution).
3 First- and second-order optimality conditions
Using well-known arguments as in [43, Sections 5.5 and 5.7] it can be shown that the mappingUad3u7→J(u)ˆ is twice continuously Fr´echet-differentiable and the second derivative is locally Lipschitz-continuous. Moreover, for any u∈Uad the reduced gradient∇J(u)ˆ ∈Ucan be expressed by
∇J(u) =ˆ κu(·) +
Z
Ω
p(·,x)bi(x)dx
1≤i≤Nu a.e. in[0,T], where the adjoint variablep=p(t,x)solvesp(T) =y(T)−yΩ inHand
−cp d
dthp(t),ϕiH+ Z
Ω
∇p(t)·∇ϕ+3cny(t)2p(t)ϕdx=0 for allϕ∈Vand f.a.a.t∈[0,T).
(3)
Further,y=y(u)denotes the solution to (2). For the second derivative at a point u∈Uwe derive that
∇2J(u)uˆ δ =
κuδ(·) + Z
Ω
pδ(·,x)bi(x)dx
1≤i≤Nu
a.e. in[0,T] for everyuδ = (uδ1, . . . ,uδN
u)∈U. Here the statey=y(u)is the solution to (2), the dualp=p(u)solves (3), the linearized dualpδ =pδ(uδ)satisfies
−cp d
dthpδ(t),ϕiH+ Z
Ω
∇pδ(t)·∇ϕ+3cny(t)2pδ(t)ϕdx
= Z
Ω
1−6cny(t)p(t)
yδ(t)ϕdt for allϕ∈Vand f.a.a.t∈[0,T) together withpδ(T) =yδ(T)inHand the linearized stateyδ =yδ(uδ)solves
cp d
dthyδ(t),ϕiH+ Z
Ω
∇yδ(t)·∇ϕ+3cny(t)2yδ(t)ϕdx
= Z
Ω
f(t) +
Nu
∑
i=1
uδi(t)bi
ϕdx for allϕ∈Vand f.a.a.t∈(0,T] withyδ(0) =0 inH.
Theorem 2 (First-order conditions). Suppose thatu¯∈Uad is a local solution to (P)ˆ and y¯=y(u)¯ denotes the associated state solving (2) for u=u. Then, there¯ exists a Lagrange multiplier p¯∈Y satisfying(3)for y=y. Moreover, we have the¯ variational inequality
Nu
∑
i=1 Z T
0
κu¯i(t) +
Z
Ω
p(t,¯ x)bi(x)dx
ui(t)−u¯i(t)
dt≥0 for all u∈Uad. (4)
Proof.The claim is shown in [43, Theorem 5.12].
Suppose that we have computed a solution to the variational inequality (4). The next theorem gives sufficient conditions that we have found a local minimum to (2). For a proof we refer to [43, Theorem 5.17].
Theorem 3 (Second-order conditions).Letu¯∈Uadbe an admissible control solv- ing together with the associated statey¯6=0and dualp¯6=0the variational inequality (4). Assume that there exist positive constantsγandτsuch that the hessian∇2J(ˆu)¯ satisfies the second-order sufficient optimality condition
h∇2J(ˆu)u¯ δ,uδiU≥γkuδkU2 (5) for every uδ = (uδ1, . . . ,uδN
u)∈U belonging to theτ-critical cone, i.e., uδ satisfies
uδi(t)
=0 if t∈Aτi,
≥0 ifu¯i(t) =uai(t)and t6∈Aτi,
≤0 ifu¯i(t) =ubi(t)and t6∈Aτi
for i=1, . . . ,Nu,
where now for i=1, . . . ,Nuthe sets of strongly active constraintsAτi are given by Aτi =n
t∈[0,T]: κu¯i(t) +
Z
Ω
¯
p(t,x)bi(x)dx >τ
o .
Then, there are positive constantsεandσsuch that the quadratic growth condition J(u)ˆ ≥J(ˆu) +¯ σku−uk¯ U2 for all u∈Uadwithku−uk¯ L∞(0,T;RNu)≤ε holds. In particular,u is a strict local minimum of¯ (P).ˆ
Remark 1.Condition (5) can be ensured provided the function Q3(t,x)7→1− 6cny(t,¯ x)p(t,¯ x)is nonnegative f.a.a.(t,x)∈Q; see [43, Sections 5.5 and 5.7] and [14, Section 2.3], for instance. This can be ensured if the residual termsαQky¯− yQkLr(Q)andαΩky(T¯ )−yΩkL∞(Ω)are small enough; cf. [45, Proposition 9]. In this case we are able to boundkpk¯ L∞(Q)by 1/(6cnkyk¯ L∞(Q)). This implies that 1−6 ¯yp¯≥
0 holds inQa.e. ♦
4 The finite element (FE) Galerkin approximation
The FE space.ForNx∈Nthe functionsϕ1, . . . ,ϕNxdenoteNxlinearly independent nodal piecewise linear finite element (FE) ansatz functions. Then, we define the Nx-dimensional subspace
Vh=span
ϕ1, . . . ,ϕNx ⊂V
endowed with the topology inV. Moreover, we introduce the FE projectionPh: H→Vh as follows: For anyw∈H the elementwh=Phw∈Vhis given as the solution to the linear system
hwh,ϕihiH=hw,ϕihiH fori=1, . . . ,Nx. (6) Letwh=∑Nj=1x whjϕhj ∈Vh. Then, the coefficient vector wh= (whi)1≤i≤Nx∈RNx is uniquely determined as the solution to the linear system
Mhwh=bh
with the mass matrix Mh= ((hϕhj,ϕihiH))∈RNx×Nx and the right-hand side bh= (hw,ϕihiH)1≤i≤Nx∈RNx. From (6) we infer that
hPhy◦,ϕhiH=hy◦,ϕhiH for allϕh∈Vh. Thus,
kPhy◦k2H=hPhy◦,Phy◦iH=hPhy◦,y◦iH≤ kPhy◦kHky◦kH
which implieskPhy◦k ≤ ky◦kH.
FE approximation for(P).ˆ First we apply a standard Galerkin scheme for the state equation. Thus, we look for a function
yh(t) =
Nx
∑
i=1yhi(t)ϕi∈Vh f.a.a.t∈[0,T]
satisfying the initial conditionyh(0) =Phy◦∈Vhand the variational equation cp d
dthyh(t),ϕhiH+ Z
Ω
∇yh(t)·∇ϕh+cnyh(t)3ϕhdx
= Z
Ω
f(t) +
Nu
i=1
∑
ui(t)bi
ϕhdx for allϕh∈Vhand f.a.a.t∈(0,T].
(7)
Lemma 1.Let y◦∈H, f ∈L2(0,T;H), u∈U and b1, . . . ,bNu ∈L∞(Ω). Then a solution yhto(7)satisfies
kyhk2L∞(0,T;H)+kyhk2L2(0,T;V)≤C ky◦k2H+kfk2L2(0,T;H)+kukU2
(8) with a constant C>0.
Proof.Choosingϕh=yh(t)∈Vhf.a.a.t∈[0,T]and using cp
Z
Ω
yh(t)4dx≥0 f.a.a.t∈[0,T], we obtain from (7)
cp 2
d
dtkyh(t)k2H+ Z
Ω
|∇yh(t)|2dx≤ kf(t)kH+√
cb|u(t)|2
kyh(t)kH
≤ kf(t)k2H+cb|u(t)|22+1
2kyh(t)k2H, (9)
where| · |2stands for the Euclidean norm (here inRNu) andcb=∑Ni=1n kbik2H. By the Gronwall lemma [10, p. 559] it follows that
kyh(t)k2H≤ et cp
kyh(0)k2H+ Z t
0
kf(s)k2H+cb|u(s)|22ds
≤ et
cp kPhy◦k2H+kfk2L2(0,T;H)+cbkukU2
≤cH ky◦k2H+kfk2L2(0,T;H)+kuk2U
(10)
for the time-independent constantcH=eTmax{1,cb}/cp. Integrating (9) over[0,T] and using (10) we find that
kyhk2L2(0,T;V)≤cp
2kyh(0)k2H+kfk2L2(0,T;H)+cbkukU2+3
2kyk2L2(0,T;H)
≤cV ky◦k2H+kfk2L2(0,T;H)+kuk2U
withcV =max{cp/2,1,cb}+3cHT/2. Now the claim follows with the constant
C=cH+cV.
Remark 2.By using Lemma 1 it can be shown by standard arguments that (7) admits a unique solution which even belongs toH1(0,T;Vh),→W(0,T). Throughout we
also assume thatyh∈L∞(Q)holds. ♦
We suppose that (7) admits a unique FE solutionyh=yh(u)∈Y for anyu∈Uad. Thus, we define the FE discretization of the reduced functional ˆJby
Jˆh(u) =J(yh(u),u) foru∈Uad. The FE approximation for (P) reads as follows:ˆ
min ˆJh(u) s.t. u∈Uad. (Pˆh) The FE optimality conditions.To characterize a local solution to (Pˆh) we introduce the FE adjoint stateph(t) =∑Ni=1x pi(t)ϕi∈Vhwhich solvesph(T) =yh(T)−PhyΩ together with
−cp d
dthph(t),ϕiH+ Z
Ω
∇ph(t)·∇ϕ+3cnyh(t)2ph(t)ϕdx
= Z
Ω
yh(t)−yQ(t)
ϕdx for allϕ∈Vhand f.a.a.t∈[0,T).
(11)
Hence, the FE reduced gradient is given by
∇Jˆh(u) = κu(·) +
Z
Ω
ph(·,x)bi(x)dx
1≤i≤Nu a.e. in[0,T].
Consequently, the FE approximation for the variational inequality (4) is given as
Nu
∑
i=1 Z T
0
κu¯hi(t) + Z
Ω
¯
ph(t,x)bi(x)dx
ui(t)−u¯hi(t)
dt≥0 (12)
for all u∈Uad, where ¯uh= (u¯hi)1≤i≤Nu ∈Uad is a local solution to (Pˆh) and ¯ph solves (11) for the optimal FE state ¯yh=yh(u¯h). Analogously to Theorem 3 the second-order sufficient optimality condition reads: There exist positive constantsγ andτsuch that the hessian∇2Jˆh(u¯h)satisfies the second-order sufficient optimality condition
h∇2Jˆh(u¯h)uδ,uδiU≥γkuδk2U for everyuδ = (uδ1, . . . ,uδN
u)∈Ubelonging to theτ-critical cone, i.e.,uδ satisfies
uδi(t)
=0 ift∈Ah,τi ,
≥0 if ¯uhi(t) =uai(t)andt6∈Ah,τi ,
≤0 if ¯uhi(t) =ubi(t)andt6∈Ah,τi
fori=1, . . . ,Nu,
where fori=1, . . . ,Nuthe sets of strongly active constraintsAh,τi are given by Ah,τi =n
t∈[0,T]:
κu¯hi(t) + Z
Ω
¯
ph(t,x)bi(x)dx >τ
o .
5 Reduced order modelling
Computation of the POD basis.Suppose that for an admissibleu∈Uadthe trajec- toriesyh(t),ph(t)∈Vh,t∈[0,T], are the FE solutions to (7) and (11), respectively.
We introduce the snapshot space Vh=span
yh(t),ph(t):t∈[0,T] ⊂Vh⊂V
anddh=dimVh≤Nx. Notice that we do not include the time derivatives into the snapshot space; cf. [20, 28]. For any`∈ {1, . . . ,dh}we construct a low-dimensional orthonormal basis by solving the optimization problem
min Z T
0
yh(t)−
`
∑
i=1
hyh(t),ψihiVψih
2 V+
ph(t)−
`
∑
i=1
hph(t),ψihiVψih
2 Vdt s.t.{ψih}`i=1⊂Vhandhψih,ψhji
X=δi jfor 1≤i,j≤`.
(13)
The solution to (13) is presented in [20, 24], for instance. Let us define the linear, bounded, finite-rank, nonnegative and selfadjoint operatorRh:V →Vhby
Rhψ= Z T
0
hyh(t),ψiVyh(t) +hph(t),ψiVph(t)dt forψ∈V.
Now the solution to (13) is given by the eigenvectors corresponding to the`largest (positive) eigenvaluesλ1h≥. . .≥λh
dh>0 solving the symmetric eigenvalue problem Rhψih=λihψih fori=1, . . . ,dh.
We can quantify the POD approximation error as follows Z T
0
yh(t)−
` i=1
∑
hyh(t),ψihiVψih
2 V
+ ph(t)−
` i=1
∑
hph(t),ψihiVψih
2 Vdt=
dh
∑
i=`+1
λih.
Next, we introduce a POD Galerkin scheme for (1b). Suppose that we have deter- mined the POD basis{ψih}`i=1⊂Vhof rank`Nx. We define
Vh,`=span
ψ1h, . . . ,ψ`h ⊂Vh⊂V
endowed with the topology inV. We introduce the projection operator Ph,`:H→Vh,`, vh,`=Ph,`ϕsolves min
wh,`∈Vh,`
kϕ−wh,`kHforϕ∈H. (14) Lemma 2.The projectionPh,`is given as
Ph,`ϕ=
` i=1
∑
ϕh,`i ψi forϕ∈H, (15) where the vectorϕh,`= (ϕh,`1 , . . . ,ϕh,`` )>∈R`solves the linear system
`
∑
j=1hψj,ψiiHϕh,`j =hϕ,ψiiH for1≤i≤`. (16)
In particular,Ph,`is linear, continuous and orthogonal in H.
Proof.The first-order optimality conditions for (14) reads: for anyϕ∈Hthe element ϕh,`=Ph,`ϕ∈Vh,`solves
hϕh,`,ψihiH=hϕ,ψihiH for 1≤i≤`. (17) Writing ϕh,`in the form ϕh,`=∑`j=1ϕh,`j ψhj we derive from (17) that the vector ϕh,`= (ϕh,`i )1≤i≤`satisfies (16). Thus, we have shown (15). Utilizing (16) and (17) the continuity ofPh,`follows from
kPh,`ϕk2H=
` i=1
∑
ϕh,`i ψih
2
H
=
` i=1
∑
`
∑
j=1ϕh,`i ϕh,`j hψih,ψhji
H
=
`
∑
j=1
ϕh,`j `
∑
i=1
ϕh,`i hψih,ψhji
H
=
`
∑
j=1
ϕh,`j hϕh,`,ψhji
H
=
`
∑
j=1
ϕh,`j hϕ,ψhji
H=hϕ,Ph,`ϕiH≤ kϕkHkPh,`ϕkH
for everyϕ∈H. For the specific choiceϕ=∑`i=1ϕh,`i ψi∈Hwith arbitrarily chosen ϕh,`1 , . . . ,ϕh,`` we even find thatPh,`ϕ=ϕ. Thus, the operator norm ofPh,`is equal
to one.
POD approximations for(Pˆh).Next, we look for a reduced-order state solution yh,`(t) =
`
∑
i=1
yh,`i (t)ψi∈Vh,` f.a.a.t∈[0,T]
satisfyingyh,`(0) =Ph,`y◦∈Vh,`and cp d
dthyh,`(t),ψhiH+ Z
Ω
∇yh,`(t)·∇ψh+cnyh,`(t)3ψhdx
= Z
Ω
f(t) +
Nu
∑
i=1
ui(t)bi
ψhdx for allψh∈Vh,`and a.e. in(0,T].
(18)
Utilizing similar arguments as in [43, Theorem 5.5] it can be shown that (18) admits a unique solutionyh,`=yh,`(u)∈Y. Thus, we define the POD discretization of the reduced functional ˆJby
Jˆh,`(u) =J(yh,`(u),u) foru∈Uad. Thus, POD approximation for (P) reads as follows:ˆ
min ˆJh,`(u) s.t. u∈Uad. (Pˆh,`) Remark 3 (A-priori analysis).For a givenu∈Uadsuppose thatyh=yh(u)andyh,`= yh,`(u)denote the solutions to (7) and (18), respectivly. To derive an a-priori error estimate for the term
Z T 0
kyh(t)−yh,`(t)kV2dt one makes use of the decomposition
yh(t)−yh,`(t) =yh(t)−Ph,`yh(t)−Ph,`yh(t)−yh,`(t) =ρh,`(t) +ϑh,`(t) withρh,`(t) =yh(t)−Ph,`yh(t)∈Vhandϑh,`(t) =Ph,`yh(t)−yh,`(t)∈Vh,`f.a.a.
t∈[0,T]. From [42, Theorem 5.2] we have Z T
0
kρh,`(t)kV2dt= Z T
0
kyh(t)−Ph,`yh(t)k2Vdt
=
dh
∑
i=`+1
λihkψih−Ph,`ψihkV2.
(19)
In [47, Section 3.4] an estimate forϑh,`is derived for a linear evolution problem.
Since our nonlinearity is monotone, we can apply similar arguments here. It follows that there exists a constantC>0 satisfying
Z T 0
kϑh,`(t)k2Vdt≤C
kϑh,`(0)k2H+
dh i=`+1
∑
λihkψih−Ph,`ψihkV2
. (20)
From (17),ψih∈Vhfor 1≤i≤`and (6) we infer that
hPh,`y◦,ψihiH=hy◦,ψihiH=hPhy◦,ψihiH for 1≤i≤`.
Let Id denotes the identity operator. Sinceϑh,`(0) =Ph,`(Ph−Id)y◦∈Vh,`holds, we have
kϑh,`(0)k2H=kPh,`(Ph−Id)y◦k2H=hPh,`(Ph−Id)y◦,ϑh,`(0)iH
=hPh(Ph−Id)y◦,ϑh,`(0)iH=h(Ph−Ph)y◦,ϑh,`(0)iH=0, i,e,ϑh,`(0) =0 is valid inH. Combining (19) and (20) we have
kyh−yh,`k2L2(0,T;V)≤2 Z T
0
kρh,`(t)k2Hdt+2 Z T
0
kϑh,`(t)k2Hdt
≤C˜
dh
∑
i=`+1
λihkψih−Ph,`ψihk2V
for the positive constant ˜C=2(1+C). ♦
The POD optimality conditions.Next we define the reduced-order dual variable ph,`=
`
∑
i=1
ph,`i (t)ψih∈Vh,` f.a.a.t∈[0,T].
We suppose that the terminal condition ph,`(T) =yh,`(T)−Ph,`yΩ and the varia- tional equation
− hcpph,`t (t),ψiV0,V+ Z
Ω
∇ph,`(t)·∇ψ+3cnyh,`(t)2ph,`(t)ψdx
= Z
Ω
yh,`(t)−yQ(t)
ψdx for allψ∈Vh,`and a.e.[0,T)
(21)
hold. Finally, the POD Galerkin approximation of the reduced gradient is given by
∇Jˆh,`(u) = κu(·) +
Z
Ω
ph,`(·,x)bi(x)dx
1≤i≤Nu a.e. in[0,T].
Proceeding formally as in Section 4 we obtain that a local solution ¯uh,`∈Uad to (Pˆh,`) is characterized by the following reduced-order variational inequality
Nu
∑
i=1 Z T
0
κu¯h,`i (t) + Z
Ω
¯
ph,`(t,x)bi(x)dx
ui(t)−u¯h,`i (t)
dt≥0 (22) for allu∈Uad.
6 A-posteriori error estimation for the reduced gradient
A-posteriori error estimation for the state equation. For an arbitrarily chosen controlu∈Uadletyh=yh(t)∈Vhandyh,`=yh,`(t)∈Vh,`be the unique solutions to (7) and (18), respectively. The goal is to estimate the error
epr(t) =yh(t)−yh,`(t)∈Vh fort∈[0,T] (23) between the FE and the POD solutions. We infer from (7), (18) and (23) that
cp
d
dthepr(t),ϕiH+ Z
Ω
∇epr(t)·∇ϕ+cn yh(t)3−yh,`(t)3 ϕdx
=D
cp yht −yh,`t (t),ϕ
E
V0,V+ Z
Ω
∇ yh−yh,`
(t)·∇ϕ+cn yh(t)3−yh,`(t)3 ϕdx
=−hcpyh,`t (t),ϕiV0,V− Z
Ω
∇yh,`(t)·∇ϕ−
cnyh,`(t)3−f(t)−
Nu
i=1
∑
ui(t)bi
ϕdx
=−hRpr(t),ϕiV0,V for allϕ∈Vhand f.a.a.t∈(0,T] with the primal residual
Rpr(t) =cpyh,`t (t)−∆yh,`(t) +cnyh,`(t)3−f(t)−
m
∑
i=1
ui(t)bi∈(Vh)0. Thus, we conclude that
cp
d
dthepr(t),ϕiH+ Z
Ω
∇epr(t)·∇ϕ+cn yh(t)3−yh,`(t)3 ϕdx
=−hRpr(t),ϕiV0,V for allϕ∈Vhand f.a.a.t∈(0,T].
(24)
For fixedt∈[0,T]we define the function Ψ(s;t) =cn yh,`(t) +sepr(t)3
=cn yh,`(t) +s(yh(t)−yh,`(t))3
fors∈[0,1], so that
Ψ(0;t) =cnyh,`(t)3,Ψ(1;t) =cnyh(t)3andΨ0(s;t) =3cn yh,`(t) +sepr(t)2
epr(t) holds f.a.a.t∈[0,T]. This implies
cn yh(t)3−yh,`(t)3
=Ψ(1;t)−Ψ(0;t) = Z 1
0
Ψ0(s;t)ds
=3cn Z 1
0
yh,`(t) +sepr(t)2
ds
epr(t)
(25)
f.a.a.t∈[0,T]. Choosingϕ=epr(t)∈Vhin (24) and using (25) we obtain
cp
2 d
dtkepr(t)k2H+ Z
Ω
∇epr(t)
2dx+
Z
Ω
cn yh(t)3−yh,`(t)3 epr(t)dx
=cp 2
d
dtkepr(t)k2H+kepr(t)k2V− kepr(t)k2H +3cn
Z 1 0
Z
Ω
yh,`(t) +sepr(t)2 epr(t)
2dxds f.a.a.t∈[0,T]. From
Z 1 0
Z
Ω
yh,`(t) +sepr(t)2 epr(t)
2dxds≥0 for allt∈[0,T]we derive that
cp
2 d
dtkepr(t)k2H+ Z
Ω
∇epr(t)
2dx+
Z
Ω
cn yh(t)3−yh,`(t)3 epr(t)dx
≥cp
2 d
dtkepr(t)k2H+kepr(t)k2V− kepr(t)k2H
f.a.a.t∈[0,T]. . Utilizing (24) and Young’s inequality [3, p. 52]
ab= √ εa
b
√ ε
≤εa2 2 +b2
2ε for alla,b∈Randε>0 witha=kRpr(t)k(Vh)0,b=kepr(t)kV, andε=1 we find that
cp 2
d
dtkepr(t)k2H+kepr(t)k2V≤ kepr(t)k2H+kRpr(t)k(Vh)0kepr(t)kV
≤ kepr(t)k2H+1
2kRpr(t)k2(Vh)0+1
2kepr(t)k2V, which implies
d
dtkepr(t)k2H+kepr(t)kV2 ≤ 2 cp
kepr(t)k2H+ 1 cp
kRpr(t)k2(Vh)0 f.a.a.t∈[0,T]. (26) Recall thatepr(0) =yh(0)−yh,`(0) = (Ph−Ph,`)y◦holds. Thus, from Gronwall’s lemma [10, p. 559] it follows that
kepr(t)k2H≤e2t/cp
k(Ph−Ph,`)y◦k2H+ 1 cp
Z t 0
kRpr(s)k2(Vh)0ds
(27) f.a.a.t∈[0,T]. Integrating (26) over[0,t]⊂[0,T]and applying (27) we derive that
kepr(t)k2H+ Z t
0
kepr(s)k2Vds
≤ kepr(0)k2H+ 2 cp
Z t 0
kepr(s)k2Hds+ 1 cp
Z t 0
kRpr(s)k2(Vh)0ds
≤ kepr(0)k2H+ 2 cp
Z t 0
e2s/cp
kepr(0)k2H+ 1 cp
Z s 0
kRpr(τ)k2(Vh)0dτ
ds + 1
cp Z t
0
kRpr(s)k2(Vh)0ds Consequently,
kepr(t)k2H+ Z t
0
kepr(s)kV2ds
≤ kepr(0)k2H+ e2t/cp−1
kepr(0)k2H+ 1 cp
Z t 0
kRpr(τ)k2(Vh)0dτ
+ 1 cp
Z t 0
kRpr(s)k2(Vh)0ds
≤e2t/cpkepr(0)k2H+
e2t/cp+cp+1 cp
Z t 0
kRpr(s)k2(Vh)0ds
Proposition 1.For an arbitrarily chosen control u∈Uad let yh=yh(t)∈Vhand yh,`=yh,`(t)∈Vh,`be the unique solutions to(7)and(18), respectively. Then,
kepr(t)k2H=kyh(t)−yh,`(t)k2H
≤e2t/cp
k(Ph−Ph,`)y◦k2H+ 1 cp
Z t 0
kRpr(s)k2(Vh)0ds
f.a.a. t∈[0,T] and
Z t 0
kyh(s)−yh,`(s)kV2ds≤e2t/cpk(Ph−Ph,`)y◦k2H + e2t/cp+c˜pZ t
0
kRpr(s)k2(Vh)0ds
f.a.a. t∈[0,T] withc˜p= (1+cp)/cp.
A-posteriori error estimation for the dual equation.Suppose thatu∈Uad is an arbitrary admissible control. Letyhandyh,`denote the associated unique solutions to (7) and (18), respectively. By phand ph,` we denote the unique corresponding solutions to (11) and (21), respectively. We estimate the dual erroredu(t) =ph(t)− ph,`(t)∈Vh. From (11) and (21) we infer for allϕ∈Vhand f.a.a.t∈[0,T)
−cp d
dthedu(t),ϕiH+ Z
Ω
∇edu(t)·∇ϕ+3cn yh(t)2ph(t)−yh,`(t)2ph,`(t) ϕdx
=−cp d
dthph(t),ϕiH+ Z
Ω
∇ph(t)·∇ϕ+3cnyh(t)2ph(t)ϕdx +cp d
dthph,`(t),ϕiH− Z
Ω
∇ph,`(t)·∇ϕ−3cnyh,`(t)2ph,`(t)ϕdx
= Z
Ω
yh(t)−yQ(t)
ϕdx− hRdu(t),ϕiV0,V+ Z
Ω
yh,`(t)−yQ(t) ϕdx
=−hRdu(t),ϕiV0,V− hepr(t),ϕiH,
where we define the dual residual as
Rdu(t;u) =−cpph,`t (t)−∆ph,`+3cnyh,`(t)2ph,`(t)−yh,`(t) +yQ(t)∈(Vh)0. Thus, we have
−cp
d
dthedu(t),ϕiH+ Z
Ω
∇edu(t)·∇ϕ+3cn yh(t)2ph(t)−yh,`(t)2ph,`(t) ϕdx
≤ kRdu(t)k(Vh)0kϕkV+kepr(t)kHkϕkH
(28) for allϕ∈Vhand f.a.a.t∈[0,T]. Notice that
Z
Ω
yh(t)2ph(t)−yh,`(t)2ph,`(t)
edu(t)dx
= Z
Ω
yh(t)2edu(t)2+ yh(t)2−yh,`(t)2
ph,`(t)edu(t)dx
≥ Z
Ω
yh(t)2−yh,`(t)2
ph,`(t)edu(t)dx f.a.a.t∈[0,T]. Thus, choosingϕ=edu(t)∈Vhwe have
−cp d
dtkedu(t)k2H+ Z
Ω
∇edu(t)
2+3cn yh(t)2ph(t)−yh,`(t)2ph,`(t)
edu(t)dx
≥ −cp
2 d
dtkedu(t)k2H+kedu(t)k2V− kedu(t)k2H +3cn
Z
Ω
yh(t)2edu(t)2+ yh(t)2−yh,`(t)2
ph,`(t)dx
≥ −cp 2
d
dtkedu(t)k2H+kedu(t)k2V− kedu(t)k2H +3cn
Z
Ω
yh(t)2−yh,`(t)2
ph,`(t)edu(t)dx f.a.a.t∈[0,T]. It follows that