Introduction
Harvesting removes prey and leaves less food for predators, while diffusion spreads both effects in space [2 , 3 , 4 , 5 , 37 ] . At position x x x and time t t t , we write v 1 ( x , t ) v_1(x,t) v 1 ( x , t ) and v 2 ( x , t ) v_2(x,t) v 2 ( x , t ) for predator and prey densities, and u ( x , t ) u(x,t) u ( x , t ) for fishing effort. The objective balances population tracking, effort costs and harvest reward. Discretizing this problem means choosing how to approximate both the dynamics and the reward [32 , 31 , 29 , 34 , 30 ] .
Implicit–explicit (IMEX) Euler treats diffusion implicitly and reactions explicitly, preserving nonnegativity under a time-step restriction [6 , 7 ] . A Patankar production–destruction step preserves positivity for every time step [8 ] . How, then, should harvest be counted over a step? To separate the effects of the state scheme and the reward, we compare three discrete problems. I use IMEX and the prey density at the start of the step. P 0 P_0 P 0 and P 1 P_1 P 1 share the same triangular positive scheme but use prey density at the start and end of the step, respectively. These are the old-state and new-state rules. Comparing I I I with P 0 P_0 P 0 isolates the state scheme, and comparing P 0 P_0 P 0 with P 1 P_1 P 1 isolates harvest quadrature.
The new-state rule rewards exactly the removal flux in the population balance. In the model with harvesting alone, the old-state rule overestimates removal. That extra reward can create negative curvature in directions that redistribute effort between time steps, add a strict local minimum, and favor extreme allocations at fixed total effort. For the full discrete problem, we derive state bounds uniform in the control and mesh, exact discrete adjoints, and Hessian-vector products that require no assembled Hessian matrix [10 , 11 ] . Because the reward depends on the final prey state, its derivative enters the terminal adjoint condition.
For the positive scheme with diffusion in one dimension, we bound the gradient and Hessian differences uniformly in the mesh. Under coercivity, these bounds control the change in the optimal control. On Ω = ( 0 , 1 ) \Omega=(0,1) Ω = ( 0 , 1 ) , with nonnegative bounded H 1 H^1 H 1 initial data, all three problems have the same weak-* variational limit. Discrete controls whose costs approach the discrete global minimum have subsequences converging strongly in L 2 L^2 L 2 to continuous global minimizers, and discrete local minimizers approximate every strict L 2 L^2 L 2 local minimizer.
The numerical tests check the derivatives and compare four optimizers on a common reference. Parameter continuation and grid refinement then trace the effects of quadrature on temporal switching, sensitivity to the ini tial control and reduced curvature. For relaxed forward–backward sweep (FBS), we give an exact local contraction criterion: under strict complementarity and a positive definite Hessian on the free coordinates, the criterion depends on that Hessian’s smallest and largest eigenvalues.
This analysis builds on consistent discrete adjoints and second derivatives [12 , 13 , 14 , 15 ] , together with the positivity, conservation and accuracy of Patankar schemes [8 , 16 , 9 ] . Cacace et al. study production–destruction control by dynamic programming with a modified Patankar semi-Lagrangian scheme [17 ] . Here we examine how harvest quadrature changes curvature, derivatives and controls.
Continuous problem
We write R \mathbb{R} R for the real numbers and N : = { 1 , 2 , … } \mathbb{N}:=\{1,2,\ldots\} N := { 1 , 2 , … } . The symbol ≔ \coloneqq : = introduces a definition. Let Ω ⊂ R d \Omega\subset\mathbb{R}^{d} Ω ⊂ R d be a bounded Lipschitz domain, d ∈ N d\in\mathbb{N} d ∈ N , and let T > 0 T>0 T > 0 be the final time. On the space–time domain Q : = Ω × ( 0 , T ) Q:=\Omega\times(0,T) Q := Ω × ( 0 , T ) , write v 1 , v 2 : Q → R v_{1},v_{2}:Q\to\mathbb{R} v 1 , v 2 : Q → R for the densities and u : Q → R u:Q\to\mathbb{R} u : Q → R for effort. In the following system, the evolution equations hold in Q Q Q , the boundary conditions on ∂ Ω × ( 0 , T ) \partial\Omega\times(0,T) ∂ Ω × ( 0 , T ) , and the initial conditions in Ω \Omega Ω for i = 1 , 2 i=1,2 i = 1 , 2 :
{ ∂ t v 1 − μ 1 Δ v 1 = v 1 ( − m + α v 2 ) , ∂ t v 2 − μ 2 Δ v 2 = v 2 ( r − β v 1 ) − q u v 2 , ∂ n v 1 = ∂ n v 2 = 0 , v i ( 0 ) = v i 0 . (1) \begin{cases}
\partial_{t}v_{1}-\mu_{1}\Delta v_{1}=v_{1}(-m+\alpha v_{2}),\\
\partial_{t}v_{2}-\mu_{2}\Delta v_{2}=v_{2}(r-\beta v_{1})-quv_{2},\\
\partial_{n}v_{1}=\partial_{n}v_{2}=0,\\
v_{i}(0)=v_{i}^{0}. \tag*{(1)}
\end{cases} ⎩ ⎨ ⎧ ∂ t v 1 − μ 1 Δ v 1 = v 1 ( − m + α v 2 ) , ∂ t v 2 − μ 2 Δ v 2 = v 2 ( r − β v 1 ) − q u v 2 , ∂ n v 1 = ∂ n v 2 = 0 , v i ( 0 ) = v i 0 . (1)
Here ∂ Ω \partial\Omega ∂ Ω is the boundary, ∂ t \partial_{t} ∂ t is the time derivative, Δ \Delta Δ is the spatial Laplacian and ∂ n \partial_{n} ∂ n the outward normal derivative. We write v i ( t ) v_{i}(t) v i ( t ) for the spatial profile v i ( ⋅ , t ) v_{i}(\cdot,t) v i ( ⋅ , t ) and v i 0 v_{i}^{0} v i 0 for its initial value. The model has positive diffusivities μ i \mu_{i} μ i , mortality m m m , growth rate r r r , interaction coefficients α , β \alpha,\beta α , β , catchability q q q , and effort bound U max U_{\max} U m a x . We assume μ i , m , r , α , β , q , U max > 0 \mu_{i},m,r,\alpha,\beta,q,U_{\max}>0 μ i , m , r , α , β , q , U m a x > 0 . The weights γ i ≥ 0 \gamma_{i}\geq0 γ i ≥ 0 , λ > 0 \lambda>0 λ > 0 , and η ≥ 0 \eta\geq0 η ≥ 0 determine the penalties for tracking and effort, and the reward for harvest. For a domain D D D , L p ( D ) L^{p}(D) L p ( D ) , 1 ≤ p < ∞ 1\leq p<\infty 1 ≤ p < ∞ , consists of measurable functions with finite norm
∥ f ∥ L p ( D ) : = ( ∫ D ∣ f ∣ p ) 1 / p . \|f\|_{L^{p}(D)}:=\left(\int_{D}|f|^{p}\right)^{1/p}. ∥ f ∥ L p ( D ) := ( ∫ D ∣ f ∣ p ) 1/ p .
Thus L 1 L^{1} L 1 means integrable and L 2 L^{2} L 2 square-integrable. The space L ∞ ( D ) L^{\infty}(D) L ∞ ( D ) contains essentially bounded functions, with norm equal to the essential supremum of ∣ f ∣ |f| ∣ f ∣ . Functions equal almost everywhere are identified. The abbreviation “a.e.” means outside a set of measure zero. We write ∣ D ∣ |D| ∣ D ∣ for the measure of D D D . We constrain effort to the admissible set
U a d : = { u ∈ L ∞ ( Q ) : 0 ≤ u ≤ U max a.e. } . \mathcal{U}_{\mathrm{ad}}:=\{u\in L^{\infty}(Q):0\leq u\leq U_{\max}\ \text{a.e.}\}. U ad := { u ∈ L ∞ ( Q ) : 0 ≤ u ≤ U m a x a.e. } .
For target profiles v i d ∈ L 2 ( Q ) v_{i}^{d}\in L^{2}(Q) v i d ∈ L 2 ( Q ) , write v : = ( v 1 , v 2 ) v:=(v_{1},v_{2}) v := ( v 1 , v 2 ) and define the cost functional
J : L 2 ( Q ) 2 × L 2 ( Q ) ⟶ R J:L^{2}(Q)^{2}\times L^{2}(Q)\longrightarrow\mathbb{R} J : L 2 ( Q ) 2 × L 2 ( Q ) ⟶ R
( v , u ) ⟼ 1 2 ∑ i = 1 2 γ i ∥ v i − v i d ∥ L 2 ( Q ) 2 + λ 2 ∥ u ∥ L 2 ( Q ) 2 − η q ∫ Q u v 2 . (v,u)\longmapsto\frac{1}{2}\sum_{i=1}^{2}\gamma_{i}\|v_{i}-v_{i}^{d}\|_{L^{2}(Q)}^{2}
+\frac{\lambda}{2}\|u\|_{L^{2}(Q)}^{2}-\eta q\int_{Q}uv_{2}. ( v , u ) ⟼ 2 1 i = 1 ∑ 2 γ i ∥ v i − v i d ∥ L 2 ( Q ) 2 + 2 λ ∥ u ∥ L 2 ( Q ) 2 − η q ∫ Q u v 2 .
Spaces, bounds, and existence
All spaces are real. The Sobolev space H 1 ( D ) H^{1}(D) H 1 ( D ) consists of L 2 ( D ) L^{2}(D) L 2 ( D ) functions whose first weak derivatives also lie in L 2 ( D ) L^{2}(D) L 2 ( D ) , with
∥ f ∥ H 1 ( D ) 2 = ∥ f ∥ L 2 ( D ) 2 + ∥ ∇ f ∥ L 2 ( D ) 2 . \|f\|_{H^{1}(D)}^{2}=\|f\|_{L^{2}(D)}^{2}+\|\nabla f\|_{L^{2}(D)}^{2}. ∥ f ∥ H 1 ( D ) 2 = ∥ f ∥ L 2 ( D ) 2 + ∥∇ f ∥ L 2 ( D ) 2 .
Here ∇ f \nabla f ∇ f is the spatial gradient. Weak derivatives are defined by integration by parts against smooth test functions. For a Banach space X X X , L p ( 0 , T ; X ) L^{p}(0,T;X) L p ( 0 , T ; X ) denotes strongly measurable X X X -valued functions whose norm lies in scalar L p ( 0 , T ) L^{p}(0,T) L p ( 0 , T ) . The space H 1 ( 0 , T ; X ) H^{1}(0,T;X) H 1 ( 0 , T ; X ) also requires a weak time derivative in L 2 ( 0 , T ; X ) L^{2}(0,T;X) L 2 ( 0 , T ; X ) . The continuous dual X ∗ X^{*} X ∗ consists of continuous linear functionals, with pairing ⟨ ℓ , x ⟩ X ∗ , X = ℓ ( x ) \langle\ell,x\rangle_{X^{*},X}=\ell(x) ⟨ ℓ , x ⟩ X ∗ , X = ℓ ( x ) . We write X m X^{m} X m for the product of m m m copies of X X X . Set
W : = L 2 ( 0 , T ; H 1 ( Ω ) ) ∩ H 1 ( 0 , T ; H 1 ( Ω ) ∗ ) , \mathcal{W}:=L^{2}(0,T;H^{1}(\Omega))\cap H^{1}(0,T;H^{1}(\Omega)^{*}), W := L 2 ( 0 , T ; H 1 ( Ω )) ∩ H 1 ( 0 , T ; H 1 ( Ω ) ∗ ) ,
Y : = W ∩ L ∞ ( Q ) , \mathcal{Y}:=\mathcal{W}\cap L^{\infty}(Q), Y := W ∩ L ∞ ( Q ) ,
with norm
∥ y ∥ Y : = ∥ y ∥ L 2 ( 0 , T ; H 1 ) + ∥ ∂ t y ∥ L 2 ( 0 , T ; H 1 ∗ ) + ∥ y ∥ L ∞ ( Q ) . \|y\|_{\mathcal{Y}}:=\|y\|_{L^{2}(0,T;H^{1})}+\|\partial_{t}y\|_{L^{2}(0,T;H^{1*})}+\|y\|_{L^{\infty}(Q)}. ∥ y ∥ Y := ∥ y ∥ L 2 ( 0 , T ; H 1 ) + ∥ ∂ t y ∥ L 2 ( 0 , T ; H 1 ∗ ) + ∥ y ∥ L ∞ ( Q ) .
The control-to-state map and reduced objective are
S : U a d ⟶ Y S:\mathcal{U}_{\mathrm{ad}}\longrightarrow\mathcal{Y} S : U ad ⟶ Y
u\longmapsto v\text{ solving \({xref}`(1)`}
and
j : U a d ⟶ R j:\mathcal{U}_{\mathrm{ad}}\longrightarrow\mathbb{R} j : U ad ⟶ R
u ⟼ J ( S ( u ) , u ) . u\longmapsto J(S(u),u). u ⟼ J ( S ( u ) , u ) .
We write S i ( u ) S_{i}(u) S i ( u ) for the i i i th component of S ( u ) S(u) S ( u ) . The space C ( [ 0 , T ] ; X ) C([0,T];X) C ([ 0 , T ] ; X ) consists of functions continuous in the norm of X X X . C ( D ) C(D) C ( D ) denotes continuous scalar functions. The symbol ↪ \hookrightarrow ↪ denotes a continuous embedding. Thus W ↪ C ( [ 0 , T ] ; L 2 ( Ω ) ) \mathcal{W}\hookrightarrow C([0,T];L^{2}(\Omega)) W ↪ C ([ 0 , T ] ; L 2 ( Ω )) gives an L 2 ( Ω ) L^{2}(\Omega) L 2 ( Ω ) -continuous representative and well-defined values at each time. We use the standard product norm on products of Hilbert spaces and the sum norm on Y 2 \mathcal{Y}^{2} Y 2 . Write ( a , b ) D : = ∫ D a b (a,b)_{D}:=\int_{D}ab ( a , b ) D := ∫ D ab , using the Euclidean dot product for vectors. An ordinary arrow denotes convergence in norm. Thus f k → L 2 ( D ) f f_{k}\xrightarrow{L^{2}(D)}f f k L 2 ( D ) f means ∥ f k − f ∥ L 2 ( D ) → 0 \|f_{k}-f\|_{L^{2}(D)}\to0 ∥ f k − f ∥ L 2 ( D ) → 0 . The half-arrow denotes weak convergence: f k ⇀ L 2 ( D ) f f_{k}\xrightharpoonup{L^{2}(D)}f f k L 2 ( D ) f means that, for every φ ∈ L 2 ( D ) \varphi\in L^{2}(D) φ ∈ L 2 ( D ) , we have ∫ D f k φ → ∫ D f φ \int_{D}f_{k}\varphi\to\int_{D}f\varphi ∫ D f k φ → ∫ D f φ . In a general normed space, weak convergence is tested against every continuous linear functional. In L ∞ ( D ) L^\infty(D) L ∞ ( D ) , f k ⇀ ∗ f f_k \rightharpoonup^* f f k ⇀ ∗ f denotes weak-∗ \ast ∗ convergence: for every ϕ ∈ L 1 ( D ) \phi\in L^1(D) ϕ ∈ L 1 ( D ) , we have ∫ D f k ϕ → ∫ D f ϕ \int_D f_k\phi\to\int_D f\phi ∫ D f k ϕ → ∫ D f ϕ . The symbols lim inf \liminf lim inf and lim sup \limsup lim sup denote lower and upper limits. For ϕ ∈ H 1 ( Ω ) \phi\in H^1(\Omega) ϕ ∈ H 1 ( Ω ) and almost every t ∈ ( 0 , T ) t \in(0,T) t ∈ ( 0 , T ) , the weak equation is
⟨ ∂ t v i , ϕ ⟩ H ∗ 1 , H 1 + μ i ( ∇ v i , ∇ ϕ ) Ω = ( f i ( v , u ) , ϕ ) Ω . (2) \langle\partial_t v_i,\phi\rangle_{H^1_*,H^1}+\mu_i(\nabla v_i,\nabla\phi)_\Omega=(f_i(v,u),\phi)_\Omega. \tag*{(2)} ⟨ ∂ t v i , ϕ ⟩ H ∗ 1 , H 1 + μ i ( ∇ v i , ∇ ϕ ) Ω = ( f i ( v , u ) , ϕ ) Ω . (2)
where the reaction map f : = ( f 1 , f 2 ) f := (f_1,f_2) f := ( f 1 , f 2 ) is
f : R 2 × R ⟶ R 2 ( v , u ) ⟼ ( v 1 ( − m + α v 2 ) v 2 ( r − β v 1 − q u ) ) . \begin{aligned}
f:\mathbb{R}^2\times\mathbb{R}&\longrightarrow\mathbb{R}^2\\
(v,u)&\longmapsto
\begin{pmatrix}
v_1(-m+\alpha v_2)\\
v_2(r-\beta v_1-qu)
\end{pmatrix}.
\end{aligned} f : R 2 × R ( v , u ) ⟶ R 2 ⟼ ( v 1 ( − m + α v 2 ) v 2 ( r − β v 1 − q u ) ) .
This weak equation includes the homogeneous Neumann conditions. We omit the domain in ∥ ⋅ ∥ p \|\cdot\|_p ∥ ⋅ ∥ p when the context fixes it, and allow constants C C C to change between estimates. Each result states what these constants depend on. For b ε > 0 b_\varepsilon>0 b ε > 0 , as ε → 0 \varepsilon\to0 ε → 0 , O ( b ε ) O(b_\varepsilon) O ( b ε ) denotes a quantity bounded in the relevant norm by C b ε Cb_\varepsilon C b ε , and o ( b ε ) o(b_\varepsilon) o ( b ε ) one whose norm divided by b ε b_\varepsilon b ε tends to zero. Set V i 0 : = ∥ v i 0 ∥ L ∞ ( Ω ) V_i^0:=\|v_i^0\|_{L^\infty(\Omega)} V i 0 := ∥ v i 0 ∥ L ∞ ( Ω ) .
Theorem 2.1 (State bounds). For u ∈ U a d u\in\mathcal{U}_{ad} u ∈ U a d and 0 ≤ v i 0 ∈ L ∞ ( Ω ) 0\le v_i^0\in L^\infty(\Omega) 0 ≤ v i 0 ∈ L ∞ ( Ω ) , (1) has a unique weak solution in Y 2 \mathcal{Y}^2 Y 2 . For almost every ( x , t ) ∈ Ω × ( 0 , T ) (x,t)\in\Omega\times(0,T) ( x , t ) ∈ Ω × ( 0 , T ) , we have
0 ≤ v 2 ( x , t ) ≤ V 2 0 e r t , 0\le v_2(x,t)\le V_2^0e^{rt}, 0 ≤ v 2 ( x , t ) ≤ V 2 0 e r t , and
0 ≤ v 1 ( x , t ) ≤ V 1 0 exp ( − m t + α V 2 0 r ( e r t − 1 ) ) . 0\le v_1(x,t)\le V_1^0\exp\left(-mt+\frac{\alpha V_2^0}{r}(e^{rt}-1)\right). 0 ≤ v 1 ( x , t ) ≤ V 1 0 exp ( − m t + r α V 2 0 ( e r t − 1 ) ) .
Define the weighted population by
M ( t ) : = β ∫ Ω v 1 ( t ) + α ∫ Ω v 2 ( t ) . \mathcal{M}(t):=\beta\int_\Omega v_1(t)+\alpha\int_\Omega v_2(t). M ( t ) := β ∫ Ω v 1 ( t ) + α ∫ Ω v 2 ( t ) .
For almost every t ∈ ( 0 , T ) t\in(0,T) t ∈ ( 0 , T ) , it satisfies
M ′ ( t ) = − β m ∫ Ω v 1 + α r ∫ Ω v 2 − α q ∫ Ω u v 2 . \mathcal{M}'(t)=-\beta m\int_\Omega v_1+\alpha r\int_\Omega v_2-\alpha q\int_\Omega uv_2. M ′ ( t ) = − β m ∫ Ω v 1 + α r ∫ Ω v 2 − α q ∫ Ω u v 2 .
Thus, for every t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] ,
M ( t ) ≤ e r t M ( 0 ) . \mathcal{M}(t)\le e^{rt}\mathcal{M}(0). M ( t ) ≤ e r t M ( 0 ) .
If μ 1 = μ 2 \mu_1=\mu_2 μ 1 = μ 2 , then for almost every ( x , t ) ∈ Ω × ( 0 , T ) (x,t)\in\Omega\times(0,T) ( x , t ) ∈ Ω × ( 0 , T ) , we also have
β v 1 + α v 2 ≤ e r t ∥ β v 1 0 + α v 2 0 ∥ ∞ . (3) \beta v_1+\alpha v_2\le e^{rt}\|\beta v_1^0+\alpha v_2^0\|_\infty. \tag*{(3)} β v 1 + α v 2 ≤ e r t ∥ β v 1 0 + α v 2 0 ∥ ∞ . (3)
Theorem 2.2 (Existence of an optimal control). Under the assumptions above, min u ∈ U a d j ( u ) \min_{u\in\mathcal{U}_{ad}}j(u) min u ∈ U a d j ( u ) has a solution.
Appendix A proves both results.
First and second derivatives
For a differentiable map F F F , D F ( u ) [ d ] DF(u)[d] D F ( u ) [ d ] is its derivative at u u u applied to direction d d d , and D 2 F ( u ) [ d , e ] D^2F(u)[d,e] D 2 F ( u ) [ d , e ] its second derivative in directions d , e d,e d , e . We also use F ′ ( u ) F'(u) F ′ ( u ) and F ′ ′ ( u ) F''(u) F ′′ ( u ) . Subscripts specify the variable being differentiated. The notation C k C^k C k means that derivatives through order k k k are continuous, and C ∞ C^\infty C ∞ that derivatives of every order are continuous. For maps between Banach spaces, derivatives are understood in the Fréchet sense. The state map has a C 2 C^2 C 2 extension S : L ∞ ( Q ) → Y 2 S:L^\infty(Q)\to\mathcal{Y}^2 S : L ∞ ( Q ) → Y 2 (Appendix A). The same formula extends j j j to this space. With the adjoint p : = ( p 1 , p 2 ) ∈ W 2 p:=(p_1,p_2)\in\mathcal{W}^2 p := ( p 1 , p 2 ) ∈ W 2 defined by (40) in Appendix A.3, the gradient map
g : L ∞ ( Q ) ⟶ L 2 ( Q ) u ⟼ λ u − q S 2 ( u ) ( η + p 2 ( u ) ) \begin{aligned}
g:L^\infty(Q)&\longrightarrow L^2(Q)\\
u&\longmapsto\lambda u-qS_2(u)(\eta+p_2(u))
\end{aligned} g : L ∞ ( Q ) u ⟶ L 2 ( Q ) ⟼ λ u − q S 2 ( u ) ( η + p 2 ( u ))
gives
j ′ ( u ) : L ∞ ( Q ) ⟶ R ξ ⟼ ∫ Q g ( u ) ξ . (4) \begin{aligned}
j'(u):L^\infty(Q)&\longrightarrow\mathbb{R}\\
\xi&\longmapsto\int_Q g(u)\xi. \tag*{(4)}
\end{aligned} j ′ ( u ) : L ∞ ( Q ) ξ ⟶ R ⟼ ∫ Q g ( u ) ξ . (4)
The scalar projection
Π [ 0 , U max ] : R ⟶ [ 0 , U max ] s ⟼ min { U max , max { 0 , s } } \begin{aligned}
\Pi_{[0,U_{\max}]}:\mathbb{R}&\longrightarrow[0,U_{\max}]\\
s&\longmapsto\min\{U_{\max},\max\{0,s\}\}
\end{aligned} Π [ 0 , U m a x ] : R s ⟶ [ 0 , U m a x ] ⟼ min { U m a x , max { 0 , s }}
acts pointwise on functions and componentwise on arrays. At a local minimizer u ∗ u^* u ∗ , for every u ∈ U a d u\in\mathcal{U}_{ad} u ∈ U a d ,
∫ Q g ( u ∗ ) ( u − u ∗ ) ≥ 0. \int_Q g(u^*)(u-u^*)\ge0. ∫ Q g ( u ∗ ) ( u − u ∗ ) ≥ 0.
Equivalently, for almost every ( x , t ) ∈ Ω × ( 0 , T ) (x,t)\in\Omega\times(0,T) ( x , t ) ∈ Ω × ( 0 , T ) ,
u ∗ = Π [ 0 , U max ] ( q λ S 2 ( u ∗ ) ( η + p 2 ( u ∗ ) ) ) . u^*=\Pi_{[0,U_{\max}]}\left(\frac{q}{\lambda}S_2(u^*)(\eta+p_2(u^*))\right). u ∗ = Π [ 0 , U m a x ] ( λ q S 2 ( u ∗ ) ( η + p 2 ( u ∗ )) ) .
For second-order sufficient conditions in PDE-constrained optimization, see [18 , 35 ] .
Two state schemes and three discrete objectives
Spatial operator and quadrature
For the convergence analysis and experiments, take Ω = ( 0 , 1 ) \Omega=(0,1) Ω = ( 0 , 1 ) . For M , N ∈ N M,N\in\mathbb{N} M , N ∈ N , set h : = 1 / M h:=1/M h := 1/ M and τ : = T / N \tau:=T/N τ := T / N . For 0 ≤ j ≤ M 0\le j\le M 0 ≤ j ≤ M and 0 ≤ n ≤ N 0\le n\le N 0 ≤ n ≤ N , define x j : = j h x_j:=jh x j := j h and t n : = n τ t_n:=n\tau t n := n τ , respectively. We identify arrays with their coordinate vectors. The entry v i , j n v^n_{i,j} v i , j n is the discrete density of species i i i at node x j x_j x j and time t n t_n t n . v i n v_i^n v i n collects its nodal values, and u j n u_j^n u j n is the fishers' effort. State trajectories lie in Y h , τ = R 2 ( N + 1 ) ( M + 1 ) \mathcal{Y}_{h,\tau}=\mathbb{R}^{2(N+1)(M+1)} Y h , τ = R 2 ( N + 1 ) ( M + 1 ) , the adjoints ( p i n ) 1 ≤ n ≤ N (p_i^n)_{1\le n\le N} ( p i n ) 1 ≤ n ≤ N in R 2 N ( M + 1 ) \mathbb{R}^{2N(M+1)} R 2 N ( M + 1 ) , and the controls in R N ( M + 1 ) \mathbb{R}^{N(M+1)} R N ( M + 1 ) . Define the discrete Neumann Laplacian by
L h : R M + 1 ⟶ R M + 1 y ⟼ h − 2 ( 2 ( y 1 − y 0 ) ( y j − 1 − 2 y j + y j + 1 ) 1 ≤ j < M 2 ( y M − 1 − y M ) ) . \begin{aligned}
L_h:\mathbb{R}^{M+1}&\longrightarrow\mathbb{R}^{M+1}\\
y&\longmapsto h^{-2}
\begin{pmatrix}
2(y_1-y_0)\\
(y_{j-1}-2y_j+y_{j+1})_{1\le j<M}\\
2(y_{M-1}-y_M)
\end{pmatrix}.
\end{aligned} L h : R M + 1 y ⟶ R M + 1 ⟼ h − 2 2 ( y 1 − y 0 ) ( y j − 1 − 2 y j + y j + 1 ) 1 ≤ j < M 2 ( y M − 1 − y M ) .
The dual cells are C 0 : = [ 0 , h / 2 ) C_0:=[0,h/2) C 0 := [ 0 , h /2 ) , C M : = ( 1 − h / 2 , 1 ] C_M:=(1-h/2,1] C M := ( 1 − h /2 , 1 ] , and for 1 ≤ j ≤ M 1\le j\le M 1 ≤ j ≤ M ,
C j : = ( x j − h / 2 , x j + h / 2 ) . C_j:=(x_j-h/2,x_j+h/2). C j := ( x j − h /2 , x j + h /2 ) .
Their lengths are w 0 : = w M : = h / 2 w_0 := w_M := h/2 w 0 := w M := h /2 , w j : = h w_j := h w j := h otherwise. For a , b ∈ R M + 1 a,b \in\mathbb{R}^{M+1} a , b ∈ R M + 1 and a , b ∈ R N ( M + 1 ) a,b \in\mathbb{R}^{N(M+1)} a , b ∈ R N ( M + 1 ) , respectively,
⟨ a , b ⟩ h : = ∑ j = 0 M w j a j b j , ⟨ a , b ⟩ h , τ : = τ ∑ n = 0 N − 1 ⟨ a n , b n ⟩ h . \langle a,b\rangle_h := \sum_{j=0}^{M} w_j a_j b_j,
\qquad
\langle a,b\rangle_{h,\tau} := \tau\sum_{n=0}^{N-1} \langle a^n,b^n\rangle_h. ⟨ a , b ⟩ h := j = 0 ∑ M w j a j b j , ⟨ a , b ⟩ h , τ := τ n = 0 ∑ N − 1 ⟨ a n , b n ⟩ h .
These inner products give the norms ∥ a ∥ h 2 : = ⟨ a , a ⟩ h \lVert a\rVert_h^2 := \langle a,a\rangle_h ∥ a ∥ h 2 := ⟨ a , a ⟩ h and ∥ a ∥ h , τ 2 : = ⟨ a , a ⟩ h , τ \lVert a\rVert_{h,\tau}^2 := \langle a,a\rangle_{h,\tau} ∥ a ∥ h , τ 2 := ⟨ a , a ⟩ h , τ . For arrays, ∥ ⋅ ∥ ∞ \lVert\cdot\rVert_\infty ∥ ⋅ ∥ ∞ denotes the largest absolute entry and ∥ ⋅ ∥ 2 \lVert\cdot\rVert_2 ∥ ⋅ ∥ 2 the Euclidean norm. Set
W h : = diag ( w 0 , … , w M ) ∈ R ( M + 1 ) × ( M + 1 ) , W_h := \operatorname{diag}(w_0,\ldots,w_M) \in\mathbb{R}^{(M+1)\times(M+1)}, W h := diag ( w 0 , … , w M ) ∈ R ( M + 1 ) × ( M + 1 ) ,
W : = τ I N ⊗ W h ∈ R N ( M + 1 ) × N ( M + 1 ) . W := \tau I_N \otimes W_h \in\mathbb{R}^{N(M+1)\times N(M+1)}. W := τ I N ⊗ W h ∈ R N ( M + 1 ) × N ( M + 1 ) .
Here diag ( a ) \operatorname{diag}(a) diag ( a ) is the diagonal matrix with entries from a a a , I s I_s I s the s × s s\times s s × s identity (I I I when its size is clear), ⊗ \otimes ⊗ the Kronecker product, 1 ∈ R M + 1 \boldsymbol{1}\in\mathbb{R}^{M+1} 1 ∈ R M + 1 the vector of ones, the superscript T T T matrix transposition, and ∘ \circ ∘ componentwise multiplication. Inequalities between arrays or matrices are componentwise. Then W h L h = L h T W h W_hL_h=L_h^TW_h W h L h = L h T W h , L h 1 = 0 L_h\boldsymbol{1}=0 L h 1 = 0 , and
⟨ − L h y , y ⟩ h = h − 1 ∑ j = 0 M − 1 ( y j + 1 − y j ) 2 . \langle-L_hy,y\rangle_h=h^{-1}\sum_{j=0}^{M-1}(y_{j+1}-y_j)^2. ⟨ − L h y , y ⟩ h = h − 1 j = 0 ∑ M − 1 ( y j + 1 − y j ) 2 .
The control box is
U a d , h , τ : = [ 0 , U max ] N ( M + 1 ) ⊂ R N ( M + 1 ) . \mathcal{U}_{\mathrm{ad},h,\tau}:=[0,U_{\max}]^{N(M+1)}\subset\mathbb{R}^{N(M+1)}. U ad , h , τ := [ 0 , U m a x ] N ( M + 1 ) ⊂ R N ( M + 1 ) .
Let V h ⊂ H 1 ( 0 , 1 ) V_h\subset H^1(0,1) V h ⊂ H 1 ( 0 , 1 ) be the space of continuous functions that are affine on each [ x j , x j + 1 ] [x_j,x_{j+1}] [ x j , x j + 1 ] , j = 0 , … , M − 1 j=0,\ldots,M-1 j = 0 , … , M − 1 . For x ∈ [ x j , x j + 1 ] x\in[x_j,x_{j+1}] x ∈ [ x j , x j + 1 ] , define the affine reconstruction
I h : R M + 1 ⟶ V h \mathcal{I}_h:\mathbb{R}^{M+1}\longrightarrow V_h I h : R M + 1 ⟶ V h
y ⟼ [ x ⟼ x j + 1 − x h y j + x − x j h y j + 1 ] y\longmapsto\left[x\longmapsto\frac{x_{j+1}-x}{h}y_j+\frac{x-x_j}{h}y_{j+1}\right] y ⟼ [ x ⟼ h x j + 1 − x y j + h x − x j y j + 1 ]
and the piecewise-constant reconstruction
Q h : R M + 1 ⟶ L 2 ( 0 , 1 ) \mathcal{Q}_h:\mathbb{R}^{M+1}\longrightarrow L^2(0,1) Q h : R M + 1 ⟶ L 2 ( 0 , 1 )
y ⟼ ∑ j = 0 M y j 1 C j . y\longmapsto\sum_{j=0}^{M}y_j\mathbf{1}_{C_j}. y ⟼ j = 0 ∑ M y j 1 C j .
Here 1 E \mathbf{1}_E 1 E is the indicator of E E E , equal to one on E E E and zero outside. Define
a h : R M + 1 × R M + 1 ⟶ R a_h:\mathbb{R}^{M+1}\times\mathbb{R}^{M+1}\longrightarrow\mathbb{R} a h : R M + 1 × R M + 1 ⟶ R
( y , z ) ⟼ h − 1 ∑ j = 0 M − 1 ( y j + 1 − y j ) ( z j + 1 − z j ) = ⟨ − L h y , z ⟩ h \begin{aligned}
(y,z)\longmapsto h^{-1}\sum_{j=0}^{M-1}(y_{j+1}-y_j)(z_{j+1}-z_j) \\
&=\langle-L_hy,z\rangle_h
\end{aligned} ( y , z ) ⟼ h − 1 j = 0 ∑ M − 1 ( y j + 1 − y j ) ( z j + 1 − z j ) = ⟨ − L h y , z ⟩ h
with seminorm
∣ y ∣ 1 , h : = a h ( y , y ) 1 / 2 = ∥ ∂ x I h y ∥ L 2 ( 0 , 1 ) . \lvert y\rvert_{1,h}:=a_h(y,y)^{1/2}=\lVert\partial_x\mathcal{I}_hy\rVert_{L^2(0,1)}. ∣ y ∣ 1 , h := a h ( y , y ) 1/2 = ∥ ∂ x I h y ∥ L 2 ( 0 , 1 ) .
The space–time reconstruction is
Q h , τ : R N ( M + 1 ) ⟶ L 2 ( Q ) \mathcal{Q}_{h,\tau}:\mathbb{R}^{N(M+1)}\longrightarrow L^2(Q) Q h , τ : R N ( M + 1 ) ⟶ L 2 ( Q )
u ⟼ ∑ n = 0 N − 1 ∑ j = 0 M u j n 1 C j × [ t n , t n + 1 ) . u\longmapsto\sum_{n=0}^{N-1}\sum_{j=0}^{M}u_j^n\mathbf{1}_{C_j\times[t_n,t_{n+1})}. u ⟼ n = 0 ∑ N − 1 j = 0 ∑ M u j n 1 C j × [ t n , t n + 1 ) .
It is an isometry: ∥ u ∥ h , τ = ∥ Q h , τ u ∥ L 2 ( Q ) \lVert u\rVert_{h,\tau}=\lVert\mathcal{Q}_{h,\tau}u\rVert_{L^2(Q)} ∥ u ∥ h , τ = ∥ Q h , τ u ∥ L 2 ( Q ) . The averaging map is
A h , τ : L 2 ( Q ) ⟶ R N ( M + 1 ) \mathcal{A}_{h,\tau}:L^2(Q)\longrightarrow\mathbb{R}^{N(M+1)} A h , τ : L 2 ( Q ) ⟶ R N ( M + 1 )
f ⟼ ( 1 τ w j ∫ t n t n + 1 ∫ C j f ( x , t ) d x d t ) 0 ≤ n < N 0 ≤ j ≤ M . f\longmapsto\left(\frac{1}{\tau w_j}\int_{t_n}^{t_{n+1}}\int_{C_j}f(x,t)\,\mathrm{d}x\,\mathrm{d}t\right)_{\substack{0\le n<N\\0\le j\le M}}. f ⟼ ( τ w j 1 ∫ t n t n + 1 ∫ C j f ( x , t ) d x d t ) 0 ≤ n < N 0 ≤ j ≤ M .
Averaging preserves the control box. It also gives
Q h , τ A h , τ f → h , τ → 0 L 2 ( Q ) f . \mathcal{Q}_{h,\tau}\mathcal{A}_{h,\tau}f\xrightarrow[h,\tau\to0]{L^2(Q)}f. Q h , τ A h , τ f L 2 ( Q ) h , τ → 0 f .
Choose target arrays v i d , n ∈ R M + 1 v_i^{d,n}\in\mathbb{R}^{M+1} v i d , n ∈ R M + 1 whose reconstructions converge strongly to v i d v_i^d v i d in L 2 ( Q ) L^2(Q) L 2 ( Q ) . Cell averages provide such arrays. In the experiments, we instead sample the smooth stationary targets at the grid nodes. We also write v i 0 v_i^0 v i 0 for the initial grid vector and specify its discretization when needed.
IMEX Euler step
For n = 0 , … , N − 1 n=0,\ldots,N-1 n = 0 , … , N − 1 , set A i : = I − τ μ i L h ∈ R ( M + 1 ) × ( M + 1 ) A_i:=I-\tau\mu_iL_h\in\mathbb{R}^{(M+1)\times(M+1)} A i := I − τ μ i L h ∈ R ( M + 1 ) × ( M + 1 ) and solve
{ A 1 v 1 n + 1 = v 1 n + τ v 1 n ∘ ( − m + α v 2 n ) , A 2 v 2 n + 1 = v 2 n + τ v 2 n ∘ ( r − β v 1 n − q u n ) . (5) \begin{cases}
A_1v_1^{n+1}=v_1^n+\tau v_1^n\circ(-m+\alpha v_2^n),\\
A_2v_2^{n+1}=v_2^n+\tau v_2^n\circ(r-\beta v_1^n-qu^n). \tag*{(5)}
\end{cases} { A 1 v 1 n + 1 = v 1 n + τ v 1 n ∘ ( − m + α v 2 n ) , A 2 v 2 n + 1 = v 2 n + τ v 2 n ∘ ( r − β v 1 n − q u n ) . (5)
Since A i − 1 ≥ 0 A_i^{-1}\ge0 A i − 1 ≥ 0 componentwise, positivity at each step follows from
min 0 ≤ j ≤ M { 1 + τ ( − m + α v 2 , j n ) } ≥ 0 , \min_{0\le j\le M}\{1+\tau(-m+\alpha v_{2,j}^n)\}\ge0, 0 ≤ j ≤ M min { 1 + τ ( − m + α v 2 , j n )} ≥ 0 ,
and
min 0 ≤ j ≤ M { 1 + τ ( r − β v 1 , j n − q u j n ) } ≥ 0. \min_{0\le j\le M}\{1+\tau(r-\beta v_{1,j}^n-qu_j^n)\}\ge0. 0 ≤ j ≤ M min { 1 + τ ( r − β v 1 , j n − q u j n )} ≥ 0.
Positive transfer-implicit step
Define the loss matrices
B : = I − τ μ 1 L h + τ m I , B:=I-\tau\mu_1L_h+\tau mI, B := I − τ μ 1 L h + τ m I ,
C n : = I − τ μ 2 L h + τ diag ( β v 1 n + q u n ) . C^n:=I-\tau\mu_2L_h+\tau\operatorname{diag}(\beta v_1^n+qu^n). C n := I − τ μ 2 L h + τ diag ( β v 1 n + q u n ) .
Solve first for the prey, then for the predator:
{ C n v 2 n + 1 = ( 1 + τ r ) v 2 n , B v 1 n + 1 = v 1 n + τ α v 1 n ∘ v 2 n + 1 . (6) \begin{cases}
C^n v_2^{n+1}=(1+\tau r)v_2^n,\\
Bv_1^{n+1}=v_1^n+\tau\alpha v_1^n\circ v_2^{n+1}. \tag*{(6)}
\end{cases} { C n v 2 n + 1 = ( 1 + τ r ) v 2 n , B v 1 n + 1 = v 1 n + τ α v 1 n ∘ v 2 n + 1 . (6)
Both equations use the transfer v 1 n ∘ v 2 n + 1 v_1^n\circ v_2^{n+1} v 1 n ∘ v 2 n + 1 , which cancels in the weighted balance.
Theorem 3.1 (Positivity and exact balance). For u ∈ U a d , h , τ u \in\mathcal{U}_{\mathrm{ad},h,\tau} u ∈ U ad , h , τ , nonnegative initial vectors, and any τ > 0 \tau> 0 τ > 0 , (6) has a unique nonnegative trajectory. With
M h n : = β ⟨ v 1 n , 1 ⟩ h + α ⟨ v 2 n , 1 ⟩ h , \mathcal{M}_h^n := \beta\langle v_1^n,\mathbf{1}\rangle_h+\alpha\langle v_2^n,\mathbf{1}\rangle_h, M h n := β ⟨ v 1 n , 1 ⟩ h + α ⟨ v 2 n , 1 ⟩ h , it satisfies
M h n + 1 − M h n τ = − β m ⟨ v 1 n + 1 , 1 ⟩ h + α r ⟨ v 2 n , 1 ⟩ h − α q ⟨ u n , v 2 n + 1 ⟩ h , (7) \frac{\mathcal{M}_h^{n+1}-\mathcal{M}_h^n}{\tau}
=-\beta m\langle v_1^{n+1},\mathbf{1}\rangle_h+\alpha r\langle v_2^n,\mathbf{1}\rangle_h
-\alpha q\langle u^n,v_2^{n+1}\rangle_h, \tag*{(7)} τ M h n + 1 − M h n = − β m ⟨ v 1 n + 1 , 1 ⟩ h + α r ⟨ v 2 n , 1 ⟩ h − α q ⟨ u n , v 2 n + 1 ⟩ h , (7) and
M h n ≤ ( 1 + τ r ) n M h 0 ≤ e r t n M h 0 . \mathcal{M}_h^n\leq(1+\tau r)^n\mathcal{M}_h^0\leq e^{rt_n}\mathcal{M}_h^0. M h n ≤ ( 1 + τ r ) n M h 0 ≤ e r t n M h 0 . See Appendix § 16 .
Proposition 3.2 (Bounds uniform in the control and mesh). Let K 2 : = ∥ v 2 0 ∥ ∞ e r T K_2:=\|v_2^0\|_{\infty}e^{rT} K 2 := ∥ v 2 0 ∥ ∞ e r T , K 1 : = ∥ v 1 0 ∥ ∞ e α K 2 T K_1:=\|v_1^0\|_{\infty}e^{\alpha K_2T} K 1 := ∥ v 1 0 ∥ ∞ e α K 2 T , and
τ 0 : = min { m − 1 , ( β K 1 + q U max ) − 1 } . \tau_0:=\min\{m^{-1},(\beta K_1+qU_{\max})^{-1}\}. τ 0 := min { m − 1 , ( β K 1 + q U m a x ) − 1 } . For every τ > 0 \tau>0 τ > 0 with N τ = T N\tau=T N τ = T , the positive scheme satisfies 0 ≤ v i n ≤ K i 1 0\leq v_i^n\leq K_i\mathbf{1} 0 ≤ v i n ≤ K i 1 . For τ ≤ τ 0 \tau\leq\tau_0 τ ≤ τ 0 , IMEX satisfies the same bounds independently of h h h . If μ 1 = μ 2 \mu_1=\mu_2 μ 1 = μ 2 , both nonnegative schemes satisfy
β v 1 n + α v 2 n ≤ ( 1 + τ r ) n ∥ β v 1 0 + α v 2 0 ∥ ∞ 1. (8) \beta v_1^n+\alpha v_2^n\leq(1+\tau r)^n\|\beta v_1^0+\alpha v_2^0\|_\infty\mathbf{1}. \tag*{(8)} β v 1 n + α v 2 n ≤ ( 1 + τ r ) n ∥ β v 1 0 + α v 2 0 ∥ ∞ 1 . (8) Appendix § 16 proves these bounds.
Remark 3.3 (Other spatial operators). Positivity, balance, and adjoint identities hold for L ∈ R s × s L\in\mathbb{R}^{s\times s} L ∈ R s × s and positive diagonal W h ∈ R s × s W_h\in\mathbb{R}^{s\times s} W h ∈ R s × s satisfying L i j ≥ 0 L_{ij}\geq0 L ij ≥ 0 for i ≠ j i\ne j i = j , L 1 = 0 L\mathbf{1}=0 L 1 = 0 , W h L = L T W h W_hL=L^{T}W_h W h L = L T W h . The shifted loss matrices are nonsingular M M M -matrices. We use the one-dimensional grid for the convergence analysis.
Harvest quadrature is part of the discrete problem
The switch ζ ∈ { 0 , 1 } \zeta\in\{0,1\} ζ ∈ { 0 , 1 } selects old-state reward when ζ = 0 \zeta=0 ζ = 0 and new-state reward when ζ = 1 \zeta=1 ζ = 1 . Define
J h , τ ζ : Y h , τ × R N ( M + 1 ) ⟶ R J_{h,\tau}^{\zeta}:\mathcal{Y}_{h,\tau}\times\mathbb{R}^{N(M+1)}\longrightarrow\mathbb{R} J h , τ ζ : Y h , τ × R N ( M + 1 ) ⟶ R
( v , u ) ⟼ τ 2 ∑ n = 0 N − 1 ∑ i = 1 2 γ i ∥ v i n − v i d , n ∥ h 2 + λ 2 ∥ u ∥ h , τ 2 − η q τ ∑ n = 0 N − 1 ⟨ u n , ( 1 − ζ ) v 2 n + ζ v 2 n + 1 ⟩ h . (9) (v,u)\longmapsto\frac{\tau}{2}\sum_{n=0}^{N-1}\sum_{i=1}^{2}\gamma_i\|v_i^n-v_i^{d,n}\|_h^2+\frac{\lambda}{2}\|u\|_{h,\tau}^2
-\eta q\tau\sum_{n=0}^{N-1}\langle u^n,(1-\zeta)v_2^n+\zeta v_2^{n+1}\rangle_h. \tag*{(9)} ( v , u ) ⟼ 2 τ n = 0 ∑ N − 1 i = 1 ∑ 2 γ i ∥ v i n − v i d , n ∥ h 2 + 2 λ ∥ u ∥ h , τ 2 − η q τ n = 0 ∑ N − 1 ⟨ u n , ( 1 − ζ ) v 2 n + ζ v 2 n + 1 ⟩ h . (9)
The state maps are
S h , τ I : U a d , h , τ ⟶ Y h , τ S_{h,\tau}^{I}:\mathcal{U}_{\mathrm{ad},h,\tau}\longrightarrow\mathcal{Y}_{h,\tau} S h , τ I : U ad , h , τ ⟶ Y h , τ
u ⟼ v solving x r e f ‘ ( 5 ) ‘ u\longmapsto v\ \text{solving }{xref}`(5)` u ⟼ v solving x r e f ‘ ( 5 ) ‘
and
S h , τ + : U a d , h , τ ⟶ Y h , τ S_{h,\tau}^{+}:\mathcal{U}_{\mathrm{ad},h,\tau}\longrightarrow\mathcal{Y}_{h,\tau} S h , τ + : U ad , h , τ ⟶ Y h , τ
u ⟼ v solving x r e f ‘ ( 6 ) ‘ u\longmapsto v\ \text{solving }{xref}`(6)` u ⟼ v solving x r e f ‘ ( 6 ) ‘
Their reduced costs are
j h , τ I : U a d , h , τ ⟶ R j_{h,\tau}^{I}:\mathcal{U}_{\mathrm{ad},h,\tau}\longrightarrow\mathbb{R} j h , τ I : U ad , h , τ ⟶ R
u ⟼ J h , τ 0 ( S h , τ I ( u ) , u ) u\longmapsto J_{h,\tau}^{0}(S_{h,\tau}^{I}(u),u) u ⟼ J h , τ 0 ( S h , τ I ( u ) , u )
and
j h , τ + , ζ : U a d , h , τ ⟶ R j_{h,\tau}^{+,\zeta}:\mathcal{U}_{\mathrm{ad},h,\tau}\longrightarrow\mathbb{R} j h , τ + , ζ : U ad , h , τ ⟶ R
u ⟼ J h , τ ζ ( S h , τ + ( u ) , u ) . u\longmapsto J_{h,\tau}^{\zeta}(S_{h,\tau}^{+}(u),u). u ⟼ J h , τ ζ ( S h , τ + ( u ) , u ) .
The two costs for the positive scheme share one trajectory, but P 1 P_1 P 1 rewards exactly the removal flux q τ ⟨ u n , v 2 n + 1 ⟩ h q\tau\langle u^n,v_2^{n+1}\rangle_h q τ ⟨ u n , v 2 n + 1 ⟩ h in (7).
Exact discrete adjoints and Hessian actions
For any of the three discrete costs j j j , the state map and cost are smooth on an open neighborhood O h , τ ⊂ R N ( M + 1 ) \mathcal{O}_{h,\tau}\subset\mathbb{R}^{N(M+1)} O h , τ ⊂ R N ( M + 1 ) of U a d , h , τ \mathcal{U}_{\mathrm{ad},h,\tau} U ad , h , τ (Appendix § 17 ). Write ∇ E j \nabla_Ej ∇ E j for the Euclidean gradient. The weighted gradient is
g : O h , τ ⟶ R N ( M + 1 ) g:\mathcal{O}_{h,\tau}\longrightarrow\mathbb{R}^{N(M+1)} g : O h , τ ⟶ R N ( M + 1 )
u ⟼ W − 1 ∇ E j ( u ) u\longmapsto W^{-1}\nabla_Ej(u) u ⟼ W − 1 ∇ E j ( u )
and the weighted Hessian action is
H ( u ) : R N ( M + 1 ) ⟶ R N ( M + 1 ) H(u):\mathbb{R}^{N(M+1)}\longrightarrow\mathbb{R}^{N(M+1)} H ( u ) : R N ( M + 1 ) ⟶ R N ( M + 1 )
d ⟼ D g ( u ) d , d\longmapsto Dg(u)d, d ⟼ D g ( u ) d ,
where D g ( u ) d Dg(u)d D g ( u ) d is the directional derivative of g g g . Equivalently,
( ∇ E j ) n , j = τ w j g n , j , ∇ E j = W g . (\nabla_Ej)_{n,j}=\tau w_jg_{n,j},\qquad\nabla_Ej=Wg. ( ∇ E j ) n , j = τ w j g n , j , ∇ E j = W g .
Write L ( X ) \mathcal{L}(X) L ( X ) for bounded linear maps from a normed space X X X into itself. In R N ( M + 1 ) \mathbb{R}^{N(M+1)} R N ( M + 1 ) , every linear map is bounded. The state and adjoint maps are
S : O h , τ ⟶ Y h , τ S:\mathcal{O}_{h,\tau}\longrightarrow\mathcal{Y}_{h,\tau} S : O h , τ ⟶ Y h , τ
u ⟼ S h , τ I ( u ) or S h , τ + ( u ) u\longmapsto S_{h,\tau}^{I}(u)\ \text{or}\ S_{h,\tau}^{+}(u) u ⟼ S h , τ I ( u ) or S h , τ + ( u )
and
p : O h , τ ⟶ R 2 N ( M + 1 ) p:\mathcal{O}_{h,\tau}\longrightarrow\mathbb{R}^{2N(M+1)} p : O h , τ ⟶ R 2 N ( M + 1 )
u ⟼ ( p i n ( u ) ) i = 1 , 2 1 ≤ n ≤ N . u\longmapsto\left(p_i^n(u)\right)_{\substack{i=1,2\\1\leq n\leq N}}. u ⟼ ( p i n ( u ) ) i = 1 , 2 1 ≤ n ≤ N .
The adjoints solve (49) for IMEX and (10) –(11) for the positive scheme. Their derivatives are
D S ( u ) : R N ( M + 1 ) ⟶ Y h , τ DS(u):\mathbb{R}^{N(M+1)}\longrightarrow\mathcal{Y}_{h,\tau} D S ( u ) : R N ( M + 1 ) ⟶ Y h , τ
and
D p ( u ) : R N ( M + 1 ) ⟶ R 2 N ( M + 1 ) Dp(u):\mathbb{R}^{N(M+1)}\longrightarrow\mathbb{R}^{2N(M+1)} D p ( u ) : R N ( M + 1 ) ⟶ R 2 N ( M + 1 )
Algorithms that use Euclidean gradients take W g Wg W g as input. Appendix § 17 .§ 1 gives the IMEX adjoint, gradient and Hessian action.
Positive scheme: terminal and backward equations
For P ζ P_{\zeta} P ζ , the terminal conditions are
p 1 N = 0 , C N − 1 p 2 N = − τ ζ η q u N − 1 . (10) p_1^N=0,\qquad C^{N-1}p_2^N=-\tau\zeta\eta q u^{N-1}. \tag*{(10)} p 1 N = 0 , C N − 1 p 2 N = − τ ζ η q u N − 1 . (10)
For k = N − 1 , … , 1 k=N-1,\ldots,1 k = N − 1 , … , 1 , solve p 1 k p_1^k p 1 k before p 2 k p_2^k p 2 k :
{ B p 1 k = p 1 k + 1 + τ [ γ 1 ( v 1 k − v 1 d , k ) + v 2 k + 1 ∘ ( α p 1 k + 1 − β p 2 k + 1 ) ] , C k − 1 p 2 k = ( 1 + τ r ) p 2 k + 1 + τ [ γ 2 ( v 2 k − v 2 d , k ) + α v 1 k − 1 ∘ p 1 k − η q ( ( 1 − ζ ) u k + ζ u k − 1 ) ] . (11) \begin{cases}
Bp_1^k=p_1^{k+1}+\tau\left[\gamma_1(v_1^k-v_1^{d,k})+v_2^{k+1}\circ(\alpha p_1^{k+1}-\beta p_2^{k+1})\right],\\
C^{k-1}p_2^k=(1+\tau r)p_2^{k+1}+\tau\left[\gamma_2(v_2^k-v_2^{d,k})+\alpha v_1^{k-1}\circ p_1^k-\eta q((1-\zeta)u^k+\zeta u^{k-1})\right]. \tag*{(11)}
\end{cases} ⎩ ⎨ ⎧ B p 1 k = p 1 k + 1 + τ [ γ 1 ( v 1 k − v 1 d , k ) + v 2 k + 1 ∘ ( α p 1 k + 1 − β p 2 k + 1 ) ] , C k − 1 p 2 k = ( 1 + τ r ) p 2 k + 1 + τ [ γ 2 ( v 2 k − v 2 d , k ) + α v 1 k − 1 ∘ p 1 k − η q (( 1 − ζ ) u k + ζ u k − 1 ) ] . (11)
The gradient is
g + , ζ , n = λ u n − η q [ ( 1 − ζ ) v 2 n + ζ v 2 n + 1 ] − q v 2 n + 1 ∘ p 2 n + 1 . (12) g^{+,\zeta,n}=\lambda u^n-\eta q\left[(1-\zeta)v_2^n+\zeta v_2^{n+1}\right]-qv_2^{n+1}\circ p_2^{n+1}. \tag*{(12)} g + , ζ , n = λ u n − η q [ ( 1 − ζ ) v 2 n + ζ v 2 n + 1 ] − q v 2 n + 1 ∘ p 2 n + 1 . (12)
For φ ∈ R M + 1 \varphi\in\mathbb{R}^{M+1} φ ∈ R M + 1 , differentiating the terminal harvest term gives
D v 2 N J τ ζ ( v , u ) [ φ ] = − τ ζ η q ⟨ u N − 1 , φ ⟩ h , D_{v_2^N}J_{\tau}^{\zeta}(v,u)[\varphi]=-\tau\zeta\eta q\langle u^{N-1},\varphi\rangle_h, D v 2 N J τ ζ ( v , u ) [ φ ] = − τ ζ η q ⟨ u N − 1 , φ ⟩ h ,
hence (10) . The initial state is fixed, so the adjoint stops at n = 1 n=1 n = 1 . Appendix C.3 gives the positive tangent, tangent-adjoint and Hessian formulas.
Proposition 4.1 (Exactness and weighted symmetry). For each of I I I , P 0 P_0 P 0 , and P 1 P_1 P 1 , the formulas in this section and Appendix C satisfy
D j ( u ) d = ⟨ g ( u ) , d ⟩ h , τ , Dj(u)d=\langle g(u),d\rangle_{h,\tau}, D j ( u ) d = ⟨ g ( u ) , d ⟩ h , τ , and
D 2 j ( u ) [ d , e ] = ⟨ H ( u ) d , e ⟩ h , τ = ⟨ d , H ( u ) e ⟩ h , τ . D^2j(u)[d,e]=\langle H(u)d,e\rangle_{h,\tau}=\langle d,H(u)e\rangle_{h,\tau}. D 2 j ( u ) [ d , e ] = ⟨ H ( u ) d , e ⟩ h , τ = ⟨ d , H ( u ) e ⟩ h , τ . With the states and adjoints stored, one tangent-state sweep and one tangent-adjoint sweep give a Hessian action. Both use tridiagonal spatial solves. The total cost is O ( N M ) O(NM) O ( N M ) . Appendix C derives the formulas.
Convergence of the discrete control problems
Throughout this section, Ω = ( 0 , 1 ) \Omega=(0,1) Ω = ( 0 , 1 ) and
0 ≤ v i 0 ∈ H 1 ( 0 , 1 ) ∩ L ∞ ( 0 , 1 ) , v i d ∈ L 2 ( Q ) . (13) 0\leq v_i^0\in H^1(0,1)\cap L^\infty(0,1),\qquad v_i^d\in L^2(Q). \tag*{(13)} 0 ≤ v i 0 ∈ H 1 ( 0 , 1 ) ∩ L ∞ ( 0 , 1 ) , v i d ∈ L 2 ( Q ) . (13)
We interpolate the initial data at the nodes and choose target reconstructions that converge strongly in L 2 ( Q ) L^2(Q) L 2 ( Q ) . Let h k , τ k ⟶ 0 h_k,\tau_k\longrightarrow0 h k , τ k ⟶ 0 independently. For IMEX, the step restriction τ k ≤ τ 0 \tau_k\leq\tau_0 τ k ≤ τ 0 holds for all sufficiently large k k k by Proposition 3.2. Reconstructions indexed by k k k use ( h k , τ k ) (h_k,\tau_k) ( h k , τ k ) .
Uniform state estimates and compactness
On [ t n , t n + 1 ) [t_n,t_{n+1}) [ t n , t n + 1 ) , define the forward time difference
D τ v i n : = ( v i n + 1 − v i n ) / τ ∈ R M + 1 , D_\tau v_i^n:=(v_i^{n+1}-v_i^n)/\tau\in\mathbb{R}^{M+1}, D τ v i n := ( v i n + 1 − v i n ) / τ ∈ R M + 1 ,
and the old-state and new-state reconstructions
V i − : = Q h v i n , V i + : = Q h v i n + 1 , V_i^-:=\mathcal{Q}_h v_i^n,\qquad V_i^+:=\mathcal{Q}_h v_i^{n+1}, V i − := Q h v i n , V i + := Q h v i n + 1 ,
with linear time interpolation
V ^ i ( x , t ) : = t n + 1 − t τ Q h v i n ( x ) + t − t n τ Q h v i n + 1 ( x ) . \widehat V_i(x,t):=\frac{t_{n+1}-t}{\tau}\mathcal{Q}_h v_i^n(x)+\frac{t-t_n}{\tau}\mathcal{Q}_h v_i^{n+1}(x). V i ( x , t ) := τ t n + 1 − t Q h v i n ( x ) + τ t − t n Q h v i n + 1 ( x ) .
On each grid, V i ± ∈ L ∞ ( Q ) V_i^\pm\in L^\infty(Q) V i ± ∈ L ∞ ( Q ) and V ^ i ∈ H 1 ( 0 , T ; V h ) \widehat V_i\in H^1(0,T;V_h) V i ∈ H 1 ( 0 , T ; V h ) .
Lemma 5.1 (Uniform discrete energy and time differences). For either scheme, a constant C C C independent of h h h , τ \tau τ , and u ∈ U a d , h , τ u\in\mathcal{U}_{\mathrm{ad},h,\tau} u ∈ U ad , h , τ gives
max 0 ≤ n ≤ N ∥ S h v i n ∥ H 1 ( 0 , 1 ) 2 + τ ∑ n = 0 N − 1 ∥ D τ v i n ∥ h 2 ≤ C . (14) \max_{0\leq n\leq N}\|\mathcal{S}_h v_i^n\|_{H^1(0,1)}^2+\tau\sum_{n=0}^{N-1}\|D_\tau v_i^n\|_h^2\leq C. \tag*{(14)} 0 ≤ n ≤ N max ∥ S h v i n ∥ H 1 ( 0 , 1 ) 2 + τ n = 0 ∑ N − 1 ∥ D τ v i n ∥ h 2 ≤ C . (14) In particular,
∥ V i + − V i − ∥ L 2 ( Q ) ≤ C τ , (15) \|V_i^+-V_i^-\|_{L^2(Q)}\leq C\tau, \tag*{(15)} ∥ V i + − V i − ∥ L 2 ( Q ) ≤ C τ , (15) and
∥ Q h v i n − S h v i n ∥ L 2 ( 0 , 1 ) ≤ C h ∥ ∂ x S h v i n ∥ L 2 . (16) \|\mathcal{Q}_h v_i^n-\mathcal{S}_h v_i^n\|_{L^2(0,1)}
\leq Ch\|\partial_x\mathcal{S}_h v_i^n\|_{L^2}. \tag*{(16)} ∥ Q h v i n − S h v i n ∥ L 2 ( 0 , 1 ) ≤ C h ∥ ∂ x S h v i n ∥ L 2 . (16)
Theorem 5.2 (State consistency for varying controls). Let u k ∈ U a d , h k , τ k u_k\in\mathcal{U}_{\mathrm{ad},h_k,\tau_k} u k ∈ U ad , h k , τ k and set U k : = Q h , τ u k U_k:=\mathcal{Q}_{h,\tau}u_k U k := Q h , τ u k . Suppose that in L ∞ ( Q ) L^\infty(Q) L ∞ ( Q ) we have U k ⇀ ∗ u U_k\rightharpoonup^* u U k ⇀ ∗ u . Then, for either state scheme, both reconstructions converge strongly in L 2 ( Q ) L^2(Q) L 2 ( Q ) :
V i , k − ⟶ S i ( u ) , V i , k + ⟶ S i ( u ) . (17) V_{i,k}^-\longrightarrow S_i(u),\qquad V_{i,k}^+\longrightarrow S_i(u). \tag*{(17)} V i , k − ⟶ S i ( u ) , V i , k + ⟶ S i ( u ) . (17) The linear interpolants also converge strongly in L 2 ( Q ) L^2(Q) L 2 ( Q ) :
V ^ i , k ⟶ S i ( u ) . \widehat V_{i,k}\longrightarrow S_i(u). V i , k ⟶ S i ( u ) . This limit is the nonnegative solution of (1) *with initial data v i 0 v_i^0 v i 0 .
Appendix D proves that the sequence ( V ^ i , k ) k (\widehat V_{i,k})_k ( V i , k ) k is bounded in L ∞ ( 0 , T ; H 1 ( 0 , 1 ) ) ∩ H 1 ( 0 , T ; L 2 ( 0 , 1 ) ) L^\infty(0,T;H^1(0,1))\cap H^1(0,T;L^2(0,1)) L ∞ ( 0 , T ; H 1 ( 0 , 1 )) ∩ H 1 ( 0 , T ; L 2 ( 0 , 1 )) , then identifies the limit using compactness [19 ] and products with one weakly convergent factor and one strongly convergent factor.
The two harvest objectives have the same limit
Proposition 5.3 (Uniform harvest-quadrature defect). For the positive state scheme and data (13) ,
sup u ∈ U a d , h , τ ∣ j h , τ + , 1 ( u ) − j h , τ + , 0 ( u ) ∣ ≤ C τ . \sup_{u\in\mathcal{U}_{\mathrm{ad},h,\tau}}\left|j_{h,\tau}^{+,1}(u)-j_{h,\tau}^{+,0}(u)\right|\leq C\tau. u ∈ U ad , h , τ sup j h , τ + , 1 ( u ) − j h , τ + , 0 ( u ) ≤ C τ . The same bound holds for the difference between their global minimum values.
See Appendix D.6.
Remark 5.4 (Control comparison under quadratic growth). Write j ζ : = j h , τ + ζ j_{\zeta}:=j_{h,\tau}^{+\zeta} j ζ := j h , τ + ζ , δ : = sup u ∈ U a d , h , τ ∣ j 1 ( u ) − j 0 ( u ) ∣ \delta:=\sup_{u\in\mathcal{U}_{\mathrm{ad},h,\tau}}|j_1(u)-j_0(u)| δ := sup u ∈ U ad , h , τ ∣ j 1 ( u ) − j 0 ( u ) ∣ . For u 0 ∈ U a d , h , τ u_0\in\mathcal{U}_{\mathrm{ad},h,\tau} u 0 ∈ U ad , h , τ , r > 0 r>0 r > 0 , set
K r : = { u ∈ U a d , h , τ : ∥ u − u 0 ∥ h , τ ≤ r } . K_r:=\{u\in\mathcal{U}_{\mathrm{ad},h,\tau}:\lVert u-u_0\rVert_{h,\tau}\le r\}. K r := { u ∈ U ad , h , τ : ∥ u − u 0 ∥ h , τ ≤ r } .
Suppose that for some c > 0 c>0 c > 0 and every u ∈ K r u\in K_r u ∈ K r ,
j 0 ( u ) ≥ j 0 ( u 0 ) + c ∥ u − u 0 ∥ h , τ 2 . j_0(u)\ge j_0(u_0)+c\lVert u-u_0\rVert_{h,\tau}^2. j 0 ( u ) ≥ j 0 ( u 0 ) + c ∥ u − u 0 ∥ h , τ 2 .
If u 1 ∈ K r u_1\in K_r u 1 ∈ K r and j 1 ( u 1 ) ≤ j 1 ( u 0 ) + ε j_1(u_1)\le j_1(u_0)+\varepsilon j 1 ( u 1 ) ≤ j 1 ( u 0 ) + ε , with ε ≥ 0 \varepsilon\ge0 ε ≥ 0 , then
∥ u 1 − u 0 ∥ h , τ ≤ 2 δ + ε c . \lVert u_1-u_0\rVert_{h,\tau}\le\sqrt{\frac{2\delta+\varepsilon}{c}}. ∥ u 1 − u 0 ∥ h , τ ≤ c 2 δ + ε .
If 2 δ < c r 2 2\delta<cr^2 2 δ < c r 2 , minimizing j 1 j_1 j 1 on K r K_r K r gives a local minimizer on U a d , h , τ \mathcal{U}_{\mathrm{ad},h,\tau} U ad , h , τ (Appendix D.7). Since δ ≤ C τ \delta\le C\tau δ ≤ C τ by Proposition 5.3, a mesh-independent positive lower bound for c c c and ε = O ( τ ) \varepsilon=O(\tau) ε = O ( τ ) give an O ( τ 1 / 2 ) O(\tau^{1/2}) O ( τ 1/2 ) bound for u 1 ∈ K r u_1\in K_r u 1 ∈ K r .
Variational convergence and strong control convergence
For I, P0 _0 0 , or P1 _1 1 , denote the reduced cost on grid k k k by j k j_k j k and define
F k : U a d ⟶ R ∪ { + ∞ } \mathcal{F}_k:\mathcal{U}_{\mathrm{ad}}\longrightarrow\mathbb{R}\cup\{+\infty\} F k : U ad ⟶ R ∪ { + ∞ }
U ⟼ { j k ( u ) , U = Q h k , τ k u , u ∈ U a d , h k , τ k , + ∞ , otherwise . U\longmapsto
\begin{cases}
j_k(u), & U=\mathcal{Q}_{h_k,\tau_k}u,\quad u\in\mathcal{U}_{\mathrm{ad},h_k,\tau_k},\\
+\infty, & \text{otherwise}.
\end{cases} U ⟼ { j k ( u ) , + ∞ , U = Q h k , τ k u , u ∈ U ad , h k , τ k , otherwise .
We write arg min K f \operatorname*{arg\,min}_K f arg min K f for the set of points in K K K where f f f attains its minimum.
Theorem 5.5 (Variational consistency and global minimizers). Assume [13 ] . For each of I, P0 _0 0 , and P1 _1 1 :
If, in L ∞ ( Q ) L^\infty(Q) L ∞ ( Q ) , U k ⇀ ∗ u U_k\rightharpoonup^\ast u U k ⇀ ∗ u , then
j ( u ) ≤ lim inf k → ∞ F k ( U k ) . (18) j(u)\le\liminf_{k\to\infty}\mathcal{F}_k(U_k). \tag*{(18)} j ( u ) ≤ k → ∞ lim inf F k ( U k ) . (18) For every u ∈ U a d u\in\mathcal{U}_{\mathrm{ad}} u ∈ U ad , u k : = A h k , τ k u u_k:=\mathcal{A}_{h_k,\tau_k}u u k := A h k , τ k u is a recovery sequence:
Q h k , τ k u k → k → ∞ L 2 ( Q ) u , \mathcal{Q}_{h_k,\tau_k}u_k\xrightarrow[k\to\infty]{L^2(Q)}u, Q h k , τ k u k L 2 ( Q ) k → ∞ u , j k ( u k ) → k → ∞ j ( u ) . (19) j_k(u_k)\xrightarrow[k\to\infty]{}j(u). \tag*{(19)} j k ( u k ) k → ∞ j ( u ) . (19) The minima converge:
min U a d , h k , τ k j k → k → ∞ min U a d j . \min_{\mathcal{U}_{\mathrm{ad},h_k,\tau_k}}j_k\xrightarrow[k\to\infty]{}\min_{\mathcal{U}_{\mathrm{ad}}}j. U ad , h k , τ k min j k k → ∞ U ad min j . If ε k → 0 \varepsilon_k\to0 ε k → 0 and
j k ( u k ) ≤ min j k + ε k , (20) j_k(u_k)\le\min j_k+\varepsilon_k, \tag*{(20)} j k ( u k ) ≤ min j k + ε k , (20) then, along a subsequence, the reconstructed controls converge strongly in L 2 ( Q ) L^2(Q) L 2 ( Q ) to some u ∗ ∈ arg min U a d j u^\ast\in\operatorname*{arg\,min}_{\mathcal{U}_{\mathrm{ad}}}j u ∗ ∈ arg min U ad j :
Q h k , τ k u k → k → ∞ u ∗ . \mathcal{Q}_{h_k,\tau_k}u_k\xrightarrow[k\to\infty]{}u^\ast. Q h k , τ k u k k → ∞ u ∗ . The states converge as in (17) .
The liminf bound and recovery-sequence property in the first two statements define sequential Γ \Gamma Γ -convergence in weak-∗ \ast ∗ L ∞ ( Q ) L^\infty(Q) L ∞ ( Q ) . Appendix D proves the theorem.
Corollary 5.6 (Comparison of global minimizers). If the continuous global minimizer is unique, every sequence satisfying (20) with ε k → 0 \varepsilon_k\to0 ε k → 0 converges strongly to it, so reconstructed controls from different schemes approach each other in L 2 ( Q ) L^2(Q) L 2 ( Q ) . When there are several continuous minimizers, the controls converge to their set.
Corollary 5.7 (Approximation of a strict local minimizer). Under Theorem 5.5, let u ∗ ∈ U a d u^\ast\in\mathcal{U}_{\mathrm{ad}} u ∗ ∈ U ad be a strict L 2 ( Q ) L^2(Q) L 2 ( Q ) local minimizer: for some r > 0 r>0 r > 0 and every u ∈ U a d u\in\mathcal{U}_{\mathrm{ad}} u ∈ U ad with 0 < ∥ u − u ∗ ∥ L 2 ( Q ) ≤ r 0<\lVert u-u^\ast\rVert_{L^2(Q)}\le r 0 < ∥ u − u ∗ ∥ L 2 ( Q ) ≤ r , j ( u ) > j ( u ∗ ) j(u)>j(u^\ast) j ( u ) > j ( u ∗ ) . For each scheme, there are local minimizers on all sufficiently fine grids whose reconstructions converge strongly to u ∗ u^\ast u ∗ in L 2 ( Q ) L^2(Q) L 2 ( Q ) .
Appendix D constructs these controls by minimizing over the closed L 2 ( Q ) L^2(Q) L 2 ( Q ) ball of radius r r r around u ∗ u^\ast u ∗ . With the regularity and sampling required by Proposition E.1, either scheme has O ( τ + h 2 ) O(\tau+h^2) O ( τ + h 2 ) state error (Appendix E.1).
What the harvest quadrature changes
For ζ ∈ { 0 , 1 } \zeta\in\{0,1\} ζ ∈ { 0 , 1 } , write j ζ : = j h , τ + ζ j_\zeta:=j_{h,\tau}^{+\zeta} j ζ := j h , τ + ζ . The quadrature defect is
D : U a d , h , τ ⟶ R D:\mathcal{U}_{\mathrm{ad},h,\tau}\longrightarrow\mathbb{R} D : U ad , h , τ ⟶ R
u ⟼ j 1 ( u ) − j 0 ( u ) . u\longmapsto j_1(u)-j_0(u). u ⟼ j 1 ( u ) − j 0 ( u ) .
Both costs use the same state, so
D ( u ) = − η q τ ∑ n = 0 N − 1 ⟨ u 2 n , u 2 n + 1 − u 2 n ⟩ h . (21) D(u)=-\eta q\tau\sum_{n=0}^{N-1}\langle u_2^n,u_2^{n+1}-u_2^n\rangle_h. \tag*{(21)} D ( u ) = − η q τ n = 0 ∑ N − 1 ⟨ u 2 n , u 2 n + 1 − u 2 n ⟩ h . (21)
Where the curvature is small, even a small cost difference can change the local minima.
A reduced model
On R N \mathbb{R}^N R N , use the weighted inner product and norm
⟨ ξ , χ ⟩ τ : = τ ∑ n = 0 N − 1 ξ n χ n , ∥ ξ ∥ τ 2 : = ⟨ ξ , ξ ⟩ τ . \langle\xi,\chi\rangle_\tau:=\tau\sum_{n=0}^{N-1}\xi^n\chi^n,\qquad\lVert\xi\rVert_\tau^2:=\langle\xi,\xi\rangle_\tau. ⟨ ξ , χ ⟩ τ := τ n = 0 ∑ N − 1 ξ n χ n , ∥ ξ ∥ τ 2 := ⟨ ξ , ξ ⟩ τ .
For u ∈ [ 0 , U max ] N u\in[0,U_{\max}]^N u ∈ [ 0 , U m a x ] N and initial population y 0 > 0 y^0>0 y 0 > 0 , let y n y^n y n denote the population at time t n t_n t n in the pure-harvesting model. For 0 ≤ n < N 0\le n<N 0 ≤ n < N ,
y n + 1 = y n 1 + τ q u n . (22) y^{n+1}=\frac{y^n}{1+\tau q u^n}. \tag*{(22)} y n + 1 = 1 + τ q u n y n . (22)
The trajectory map and costs are
y : [ 0 , U max ] N ⟶ ( 0 , ∞ ) N + 1 y:[0,U_{\max}]^N\longrightarrow(0,\infty)^{N+1} y : [ 0 , U m a x ] N ⟶ ( 0 , ∞ ) N + 1
u ⟼ ( y 0 ∏ n = 0 k − 1 ( 1 + τ q u n ) − 1 ) k = 0 N , u\longmapsto\left(y^0\prod_{n=0}^{k-1}(1+\tau q u^n)^{-1}\right)_{k=0}^{N}, u ⟼ ( y 0 n = 0 ∏ k − 1 ( 1 + τ q u n ) − 1 ) k = 0 N ,
j 0 : [ 0 , U max ] N ⟶ R j_0 : [0,U_{\max}]^N \longrightarrow\mathbb{R} j 0 : [ 0 , U m a x ] N ⟶ R
u ⟼ λ τ 2 ∑ n = 0 N − 1 ( u n ) 2 − η τ q ∑ n = 0 N − 1 u n y n u \longmapsto\frac{\lambda\tau}{2}\sum_{n=0}^{N-1}(u^n)^2-\eta\tau q\sum_{n=0}^{N-1}u^ny^n u ⟼ 2 λ τ n = 0 ∑ N − 1 ( u n ) 2 − η τ q n = 0 ∑ N − 1 u n y n
and
j 1 : [ 0 , U max ] N ⟶ R j_1 : [0,U_{\max}]^N \longrightarrow\mathbb{R} j 1 : [ 0 , U m a x ] N ⟶ R
u ⟼ λ τ 2 ∑ n = 0 N − 1 ( u n ) 2 − η τ q ∑ n = 0 N − 1 u n y n + 1 . u \longmapsto\frac{\lambda\tau}{2}\sum_{n=0}^{N-1}(u^n)^2-\eta\tau q\sum_{n=0}^{N-1}u^ny^{n+1}. u ⟼ 2 λ τ n = 0 ∑ N − 1 ( u n ) 2 − η τ q n = 0 ∑ N − 1 u n y n + 1 .
Taking the empty product as 1 and summing the population removed at each step gives
j 1 ( u ) = λ τ 2 ∑ n = 0 N − 1 ( u n ) 2 − η τ q ∑ n = 0 N − 1 u n y n + 1 = λ τ 2 ∑ n = 0 N − 1 ( u n ) 2 − η ∑ n = 0 N − 1 ( y n − y n + 1 ) = λ τ 2 ∑ n = 0 N − 1 ( u n ) 2 − η ( y 0 − y N ) . (23) \begin{aligned}
j_1(u)=\frac{\lambda\tau}{2}\sum_{n=0}^{N-1}(u^n)^2-\eta\tau q\sum_{n=0}^{N-1}u^ny^{n+1} \\
&=\frac{\lambda\tau}{2}\sum_{n=0}^{N-1}(u^n)^2-\eta\sum_{n=0}^{N-1}(y^n-y^{n+1}) \\
&=\frac{\lambda\tau}{2}\sum_{n=0}^{N-1}(u^n)^2-\eta(y^0-y^N). \tag*{(23)}
\end{aligned} j 1 ( u ) = 2 λ τ n = 0 ∑ N − 1 ( u n ) 2 − η τ q n = 0 ∑ N − 1 u n y n + 1 = 2 λ τ n = 0 ∑ N − 1 ( u n ) 2 − η n = 0 ∑ N − 1 ( y n − y n + 1 ) = 2 λ τ n = 0 ∑ N − 1 ( u n ) 2 − η ( y 0 − y N ) . (23)
Old-state quadrature adds a reward term:
j 0 ( u ) = j 1 ( u ) − η τ q ∑ n = 0 N − 1 u n ( y n − y n + 1 ) = j 1 ( u ) − η τ 2 q 2 ∑ n = 0 N − 1 ( u n ) 2 y n + 1 . (24) \begin{aligned}
j_0(u)=j_1(u)-\eta\tau q\sum_{n=0}^{N-1}u^n(y^n-y^{n+1}) \\
&=j_1(u)-\eta\tau^2q^2\sum_{n=0}^{N-1}(u^n)^2y^{n+1}. \tag*{(24)}
\end{aligned} j 0 ( u ) = j 1 ( u ) − η τ q n = 0 ∑ N − 1 u n ( y n − y n + 1 ) = j 1 ( u ) − η τ 2 q 2 n = 0 ∑ N − 1 ( u n ) 2 y n + 1 . (24)
This term rewards population already removed (Appendix F.1). For u ∈ [ 0 , U max ] N u\in[0,U_{\max}]^N u ∈ [ 0 , U m a x ] N and ζ ∈ { 0 , 1 } \zeta\in\{0,1\} ζ ∈ { 0 , 1 } , the weighted Hessian action is
H ζ ( u ) : R N ⟶ R N H_\zeta(u):\mathbb{R}^N\longrightarrow\mathbb{R}^N H ζ ( u ) : R N ⟶ R N
ξ ⟼ τ − 1 ∇ E 2 j ζ ( u ) ξ \xi\longmapsto\tau^{-1}\nabla_E^2j_\zeta(u)\xi ξ ⟼ τ − 1 ∇ E 2 j ζ ( u ) ξ
where ∇ E 2 \nabla_E^2 ∇ E 2 is the Euclidean Hessian of the cost’s smooth rational extension.
Proposition 6.1 (Convexity and a quadrature-induced negative mode). For every N N N , λ > 0 \lambda>0 λ > 0 , and η ≥ 0 \eta\geq0 η ≥ 0 , j 1 j_1 j 1 is strictly convex on [ 0 , U max ] N [0,U_{\max}]^N [ 0 , U m a x ] N . Its unique minimizer is constant in time. If it lies strictly inside the control box, its value s s s is the unique solution of
λ s = η q y 0 ( 1 + τ q s ) − N − 1 . \lambda s=\eta q y^0(1+\tau q s)^{-N-1}. λ s = η q y 0 ( 1 + τ q s ) − N − 1 . The operators H ζ ( u ) ∈ L ( R N ) H_\zeta(u)\in\mathcal{L}(\mathbb{R}^N) H ζ ( u ) ∈ L ( R N ) satisfy
D 2 j ζ ( u ) [ ξ , χ ] = ⟨ H ζ ( u ) ξ , χ ⟩ τ D^2j_\zeta(u)[\xi,\chi]=\langle H_\zeta(u)\xi,\chi\rangle_\tau D 2 j ζ ( u ) [ ξ , χ ] = ⟨ H ζ ( u ) ξ , χ ⟩ τ and
H 1 ( 0 ) = λ I + η τ q 2 y 0 ( 11 T + I ) , H_1(0)=\lambda I+\eta\tau q^2y^0(11^{\mathsf T}+I), H 1 ( 0 ) = λ I + η τ q 2 y 0 ( 1 1 T + I ) , H 0 ( 0 ) = λ I + η τ q 2 y 0 ( 11 T − I ) , H_0(0)=\lambda I+\eta\tau q^2y^0(11^{\mathsf T}-I), H 0 ( 0 ) = λ I + η τ q 2 y 0 ( 1 1 T − I ) , where I ∈ R N × N I\in\mathbb{R}^{N\times N} I ∈ R N × N and 1 ∈ R N 1\in\mathbb{R}^N 1 ∈ R N . For every nonzero direction whose components sum to zero, the Rayleigh quotient is λ + η τ q 2 y 0 \lambda+\eta\tau q^2y^0 λ + η τ q 2 y 0 for H 1 ( 0 ) H_1(0) H 1 ( 0 ) and λ − η τ q 2 y 0 \lambda-\eta\tau q^2y^0 λ − η τ q 2 y 0 for H 0 ( 0 ) H_0(0) H 0 ( 0 ) . For N ≥ 2 N\geq2 N ≥ 2 and λ < η τ q 2 y 0 \lambda<\eta\tau q^2y^0 λ < η τ q 2 y 0 , j 0 j_0 j 0 is nonconvex.
The proof is in Appendix F.2. For even N N N , the direction ( 1 , − 1 , … , 1 , − 1 ) (1,-1,\ldots,1,-1) ( 1 , − 1 , … , 1 , − 1 ) shifts effort between adjacent steps while keeping the total fixed. At zero effort, the old-state curvature along this direction changes sign at λ = η τ q 2 y 0 \lambda=\eta\tau q^2y^0 λ = η τ q 2 y 0 . This gives a mechanism for temporal oscillations. The refinement estimate is
∥ H 1 ( u ) − H 0 ( u ) ∥ L ( R N ) ≤ η τ q 2 y 0 ( 2 + 4 q U max T + 2 q 2 U max 2 T 2 ) . \lVert H_1(u)-H_0(u)\rVert_{\mathcal{L}(\mathbb{R}^N)}
\leq\eta\tau q^2y^0(2+4qU_{\max}T+2q^2U_{\max}^2T^2). ∥ H 1 ( u ) − H 0 ( u ) ∥ L ( R N ) ≤ η τ q 2 y 0 ( 2 + 4 q U m a x T + 2 q 2 U m a x 2 T 2 ) .
Here the operator norm is induced by ∥ ⋅ ∥ τ \lVert\cdot\rVert_\tau ∥ ⋅ ∥ τ . Once the bound falls below λ \lambda λ , j 0 j_0 j 0 is strictly convex (Appendix F.3).
Proposition 6.2 (A quadrature-induced local minimum that is not global). For N = 2 N=2 N = 2 , define x : = τ q u 0 x:=\tau qu^0 x := τ q u 0 and z : = τ q u 1 z:=\tau qu^1 z := τ q u 1 . If η τ q 2 y 0 = 2 λ \eta\tau q^2y^0=2\lambda η τ q 2 y 0 = 2 λ and τ q U max > 2 \tau qU_{\max}>2 τ q U m a x > 2 , then ( x , z ) = ( 0 , 2 ) (x,z)=(0,2) ( x , z ) = ( 0 , 2 ) is a strict constrained local minimum of j 0 j_0 j 0 , but not a global one. The flux cost has a unique global minimum.
The proof is in Appendix F.4. Both the flux and continuous costs are strictly convex. For u ∈ L ∞ ( 0 , T ) u\in L^\infty(0,T) u ∈ L ∞ ( 0 , T ) with 0 ≤ u ≤ U max 0\leq u\leq U_{\max} 0 ≤ u ≤ U m a x a.e., the continuous cost is
j c ( u ) = λ 2 ∫ 0 T u 2 d t − η y 0 + η y 0 exp ( − q ∫ 0 T u d t ) . j_c(u)=\frac{\lambda}{2}\int_0^T u^2\,\mathrm{d}t-\eta y^0+\eta y^0\exp\left(-q\int_0^T u\,\mathrm{d}t\right). j c ( u ) = 2 λ ∫ 0 T u 2 d t − η y 0 + η y 0 exp ( − q ∫ 0 T u d t ) .
Its smooth extension to L 2 ( 0 , T ) L^2(0,T) L 2 ( 0 , T ) satisfies, for ξ ∈ L 2 ( 0 , T ) \xi\in L^2(0,T) ξ ∈ L 2 ( 0 , T ) ,
D 2 j c ( u ) [ ξ , ξ ] = λ ∥ ξ ∥ L 2 ( 0 , T ) 2 + η q 2 y 0 e − q ∫ 0 T u d t ( ∫ 0 T ξ d t ) 2 ≥ λ ∥ ξ ∥ L 2 ( 0 , T ) 2 . D^2j_c(u)[\xi,\xi]=\lambda\lVert\xi\rVert_{L^2(0,T)}^2+\eta q^2y^0e^{-q\int_0^T u\,\mathrm{d}t}\left(\int_0^T\xi\,\mathrm{d}t\right)^2
\geq\lambda\lVert\xi\rVert_{L^2(0,T)}^2. D 2 j c ( u ) [ ξ , ξ ] = λ ∥ ξ ∥ L 2 ( 0 , T ) 2 + η q 2 y 0 e − q ∫ 0 T u d t ( ∫ 0 T ξ d t ) 2 ≥ λ ∥ ξ ∥ L 2 ( 0 , T ) 2 .
Quadrature-induced local minima can therefore occur even with a positive state step.
Optimal allocation at fixed total effort
For two time steps with u 0 + u 1 = E u^0+u^1=E u 0 + u 1 = E , 0 < E ≤ U max 0<E\leq U_{\max} 0 < E ≤ U m a x , set a : = u 0 ∈ [ 0 , E ] a:=u^0\in[0,E] a := u 0 ∈ [ 0 , E ] , u 1 = E − a u^1=E-a u 1 = E − a . The restricted costs are
f 0 : [ 0 , E ] ⟶ R f_0:[0,E]\longrightarrow\mathbb{R} f 0 : [ 0 , E ] ⟶ R
a ⟼ λ τ 2 [ a 2 + ( E − a ) 2 ] − η τ q y 0 E + η τ 2 q 2 y 0 a ( E − a ) 1 + τ q a a\longmapsto\frac{\lambda\tau}{2}[a^2+(E-a)^2]-\eta\tau qy^0E+\eta\tau^2q^2y^0\frac{a(E-a)}{1+\tau qa} a ⟼ 2 λ τ [ a 2 + ( E − a ) 2 ] − η τ q y 0 E + η τ 2 q 2 y 0 1 + τ q a a ( E − a )
and
f 1 : [ 0 , E ] ⟶ R f_1:[0,E]\longrightarrow\mathbb{R} f 1 : [ 0 , E ] ⟶ R
a ⟼ λ τ 2 [ a 2 + ( E − a ) 2 ] − η y 0 + η y 0 ( 1 + τ q a ) ( 1 + τ q ( E − a ) ) . a\longmapsto\frac{\lambda\tau}{2}[a^2+(E-a)^2]-\eta y^0+\frac{\eta y^0}{(1+\tau qa)(1+\tau q(E-a))}. a ⟼ 2 λ τ [ a 2 + ( E − a ) 2 ] − η y 0 + ( 1 + τ q a ) ( 1 + τ q ( E − a )) η y 0 .
Proposition 6.3 (Quadrature-induced extreme allocations). For λ > 0 \lambda> 0 λ > 0 , f 1 f_1 f 1 has the unique global minimizer a = E / 2 a = E/2 a = E /2 . If
λ τ < η τ 2 q 2 y 0 ( 1 + τ q E ) 2 , \lambda\tau< \frac{\eta\tau^2 q^2 y^0}{(1+\tau qE)^2}, λ τ < ( 1 + τ q E ) 2 η τ 2 q 2 y 0 ,
then f 0 f_0 f 0 is strictly concave on [ 0 , E ] [0,E] [ 0 , E ] , with exactly two global minimizers a = 0 , E a=0,E a = 0 , E . The exact continuous removal model, with effort constant on each interval and the same total effort, has the unique minimizer a = E / 2 a=E/2 a = E /2 .
See Appendix F.5.
The sign of the coupled defect
Return to j ζ : = j h , τ + , ζ j_\zeta:= j_{h,\tau}^{+,\zeta} j ζ := j h , τ + , ζ and D : = j 1 − j 0 D := j_1-j_0 D := j 1 − j 0 . Then
D ( u ) = ( 21 ) − η q τ ∑ n = 0 N − 1 ⟨ u n , v 2 n + 1 − v 2 n ⟩ h D(u) \overset{(21)}{=} -\eta q\tau\sum_{n=0}^{N-1}\langle u^n,v_2^{n+1}-v_2^n\rangle_h D ( u ) = ( 21 ) − η q τ n = 0 ∑ N − 1 ⟨ u n , v 2 n + 1 − v 2 n ⟩ h
= ( 6 ) η q τ 2 ∑ n = 0 N − 1 [ q ⟨ ( u n ) 2 , v 2 n + 1 ⟩ h + β ⟨ u n v 1 n , v 2 n + 1 ⟩ h − r ⟨ u n , v 2 n ⟩ h − μ 2 ⟨ u n , L h v 2 n + 1 ⟩ h ] . (25) \overset{(6)}{=} \eta q\tau^2\sum_{n=0}^{N-1}\left[q\langle(u^n)^2,v_2^{n+1}\rangle_h+\beta\langle u^n v_1^n,v_2^{n+1}\rangle_h-r\langle u^n,v_2^n\rangle_h-\mu_2\langle u^n,L_hv_2^{n+1}\rangle_h\right]. \tag*{(25)} = ( 6 ) η q τ 2 n = 0 ∑ N − 1 [ q ⟨( u n ) 2 , v 2 n + 1 ⟩ h + β ⟨ u n v 1 n , v 2 n + 1 ⟩ h − r ⟨ u n , v 2 n ⟩ h − μ 2 ⟨ u n , L h v 2 n + 1 ⟩ h ] . (25)
Growth, predation and diffusion determine the sign: the defect can be positive or negative (Appendix F.6). Differentiating it also requires the state’s response to u u u .
Uniform derivative comparison and control stability
On R N ( M + 1 ) \mathbb{R}^{N(M+1)} R N ( M + 1 ) , equipped with ∥ ⋅ ∥ h , τ \|\cdot\|_{h,\tau} ∥ ⋅ ∥ h , τ , use the operator norm
∥ A ∥ L ( R N ( M + 1 ) ) : = sup a ∈ R N ( M + 1 ) ∖ { 0 } ∥ A a ∥ h , τ ∥ a ∥ h , τ . \|A\|_{\mathcal{L}(\mathbb{R}^{N(M+1)})}:=\sup_{a\in\mathbb{R}^{N(M+1)}\setminus\{0\}}\frac{\|Aa\|_{h,\tau}}{\|a\|_{h,\tau}}. ∥ A ∥ L ( R N ( M + 1 ) ) := a ∈ R N ( M + 1 ) ∖ { 0 } sup ∥ a ∥ h , τ ∥ A a ∥ h , τ .
For w = ( w n ) n = 0 N ∈ R ( N + 1 ) ( M + 1 ) w=(w^n)_{n=0}^{N}\in\mathbb{R}^{(N+1)(M+1)} w = ( w n ) n = 0 N ∈ R ( N + 1 ) ( M + 1 ) , define
w − : = ( w n ) n = 0 N − 1 ∈ R N ( M + 1 ) , w + : = ( w n + 1 ) n = 0 N − 1 ∈ R N ( M + 1 ) . w^-:=(w^n)_{n=0}^{N-1}\in\mathbb{R}^{N(M+1)},\qquad
w^+:=(w^{n+1})_{n=0}^{N-1}\in\mathbb{R}^{N(M+1)}. w − := ( w n ) n = 0 N − 1 ∈ R N ( M + 1 ) , w + := ( w n + 1 ) n = 0 N − 1 ∈ R N ( M + 1 ) .
Set
D τ w n : = w n + 1 − w n τ , ∥ D τ w ∥ h , τ 2 = τ ∑ n = 0 N − 1 ∥ D τ w n ∥ h 2 . D_\tau w^n:=\frac{w^{n+1}-w^n}{\tau},\qquad
\|D_\tau w\|_{h,\tau}^2=\tau\sum_{n=0}^{N-1}\|D_\tau w^n\|_h^2. D τ w n := τ w n + 1 − w n , ∥ D τ w ∥ h , τ 2 = τ n = 0 ∑ N − 1 ∥ D τ w n ∥ h 2 .
Theorem 7.1 (A uniform C 2 C^2 C 2 quadrature estimate in one dimension). On Ω = ( 0 , 1 ) \Omega=(0,1) Ω = ( 0 , 1 ) , fix T T T , the model parameters with μ 1 , μ 2 > 0 \mu_1,\mu_2>0 μ 1 , μ 2 > 0 , and U max U_{\max} U m a x . For i = 1 , 2 i=1,2 i = 1 , 2 , assume 0 ≤ v i 0 ∈ R M + 1 0\le v_i^0\in\mathbb{R}^{M+1} 0 ≤ v i 0 ∈ R M + 1 and
sup h = 1 / M M ∈ N ( ∥ v i 0 ∥ ∞ + ∥ v i 0 ∥ h + ∣ v i 0 ∣ 1 , h ) < ∞ . \sup_{\substack{h=1/M\\M\in\mathbb{N}}}\left(\|v_i^0\|_\infty+\|v_i^0\|_h+|v_i^0|_{1,h}\right)<\infty. h = 1/ M M ∈ N sup ( ∥ v i 0 ∥ ∞ + ∥ v i 0 ∥ h + ∣ v i 0 ∣ 1 , h ) < ∞.
For (6) and identical tracking terms in j 0 , j 1 j_0,j_1 j 0 , j 1 , there exist τ ˉ > 0 \bar{\tau}>0 τ ˉ > 0 , C < ∞ C<\infty C < ∞ , independent of h h h , N N N , and u ∈ U a d , h , τ u\in\mathcal{U}_{ad,h,\tau} u ∈ U a d , h , τ , such that for τ ≤ τ ˉ \tau\le\bar{\tau} τ ≤ τ ˉ ,
∣ j 1 ( u ) − j 0 ( u ) ∣ ≤ C τ , (26) |j_1(u)-j_0(u)|\le C\tau, \tag*{(26)} ∣ j 1 ( u ) − j 0 ( u ) ∣ ≤ C τ , (26)
and
∥ g 1 ( u ) − g 0 ( u ) ∥ h , τ ≤ C τ , (27) \|g_1(u)-g_0(u)\|_{h,\tau}\le C\tau, \tag*{(27)} ∥ g 1 ( u ) − g 0 ( u ) ∥ h , τ ≤ C τ , (27)
and
∥ H 1 ( u ) − H 0 ( u ) ∥ L ( R N ( M + 1 ) ) ≤ C τ . (28) \|H_1(u)-H_0(u)\|_{\mathcal{L}(\mathbb{R}^{N(M+1)})}\le C\tau. \tag*{(28)} ∥ H 1 ( u ) − H 0 ( u ) ∥ L ( R N ( M + 1 ) ) ≤ C τ . (28)
The proof in Appendix G.1 uses the one-dimensional H 1 H^1 H 1 embedding to bound the second variation. Remark G.1 in Appendix G.2 defines the constants C z , C s C_z,C_s C z , C s and the step bound τ ˉ \bar{\tau} τ ˉ used below.
Corollary 7.2 (Global uniqueness for a sufficiently large effort penalty). Under Theorem 7.1, also assume ∥ v i d ∥ h , τ ≤ D i \|v_i^d\|_{h,\tau}\le D_i ∥ v i d ∥ h , τ ≤ D i uniformly. Define
B : = C s ∑ i = 1 2 γ i ( K i T + D i ) + 2 η q C z + η q U max T C s . B:=C_s\sum_{i=1}^{2}\gamma_i(K_i\sqrt{T}+D_i)+2\eta q C_z+\eta q U_{\max}\sqrt{T}C_s. B := C s i = 1 ∑ 2 γ i ( K i T + D i ) + 2 η q C z + η q U m a x T C s .
If λ > B \lambda>B λ > B and τ ≤ τ ˉ \tau\le\bar{\tau} τ ≤ τ ˉ , both objectives for the positive scheme are strongly convex on U a d , h , τ \mathcal{U}_{ad,h,\tau} U a d , h , τ , with modulus at least λ − B \lambda-B λ − B uniformly in the mesh. Each has a unique global minimizer.
The bound B B B is independent of λ \lambda λ and u u u (Appendix G.3).
Corollary 7.3 (Control stability under uniform coercivity). Under Theorem 7.1, take τ ≤ τ ˉ \tau\le\bar{\tau} τ ≤ τ ˉ , a common constant C C C in (27)–(28), and u 0 ∈ U a d , h , τ u_0\in\mathcal{U}_{ad,h,\tau} u 0 ∈ U a d , h , τ such that, for every u ∈ U a d , h , τ u\in\mathcal{U}_{ad,h,\tau} u ∈ U a d , h , τ , ⟨ g 0 ( u 0 ) , u − u 0 ⟩ h , τ ≥ 0 \langle g_0(u_0),u-u_0\rangle_{h,\tau}\ge0 ⟨ g 0 ( u 0 ) , u − u 0 ⟩ h , τ ≥ 0 . For r > 0 r>0 r > 0 , set
K r : = { u ∈ U a d , h , τ : ∥ u − u 0 ∥ h , τ ≤ r } . K_r:=\{u\in\mathcal{U}_{ad,h,\tau}:\|u-u_0\|_{h,\tau}\le r\}. K r := { u ∈ U a d , h , τ : ∥ u − u 0 ∥ h , τ ≤ r } .
Suppose that for some κ > 0 \kappa>0 κ > 0 , every u ∈ K r u\in K_r u ∈ K r and every a ∈ R N ( M + 1 ) a\in\mathbb{R}^{N(M+1)} a ∈ R N ( M + 1 ) ,
⟨ H 0 ( u ) a , a ⟩ h , τ ≥ κ ∥ a ∥ h , τ 2 . \langle H_0(u)a,a\rangle_{h,\tau}\ge\kappa\|a\|_{h,\tau}^2. ⟨ H 0 ( u ) a , a ⟩ h , τ ≥ κ ∥ a ∥ h , τ 2 .
If C τ < κ min { 1 , r } C\tau<\kappa\min\{1,r\} C τ < κ min { 1 , r } , the flux problem has a unique stationary point u 1 u_1 u 1 in K r K_r K r , which is a strict local minimizer and satisfies
∥ u 1 − u 0 ∥ h , τ ≤ C τ κ . (29) \|u_1-u_0\|_{h,\tau}\le\frac{C\tau}{\kappa}. \tag*{(29)} ∥ u 1 − u 0 ∥ h , τ ≤ κ C τ . (29)
Stationarity means the box variational inequality, and the estimate is uniform in the mesh for fixed r , κ r,\kappa r , κ . With coercivity on the whole box, the minimizers are unique and global, with the same bound.
See Appendix G.4. At a common control and on a common subspace, corresponding ordered Hessian eigenvalues differ by at most C τ C\tau C τ (Appendix G.5).
Second-order conditions and local sweep theory
Fix a grid and one discrete objective.
Stationarity and the critical cone
Define the stationarity residual by
R : U a d , h , τ ⟶ R N ( M + 1 ) R:\mathcal{U}_{\mathrm{ad},h,\tau}\longrightarrow\mathbb{R}^{N(M+1)} R : U ad , h , τ ⟶ R N ( M + 1 )
u ⟼ u − Π [ 0 , U max ] ( u − g ( u ) ) . u\longmapsto u-\Pi_{[0,U_{\max}]}(u-g(u)). u ⟼ u − Π [ 0 , U m a x ] ( u − g ( u )) .
Then R ( u ∗ ) = 0 R(u^*)=0 R ( u ∗ ) = 0 is equivalent to the following inequality for every u ∈ U a d , h , τ u\in\mathcal{U}_{\mathrm{ad},h,\tau} u ∈ U ad , h , τ :
⟨ g ( u ∗ ) , u − u ∗ ⟩ h , τ ≥ 0. \langle g(u^*),u-u^*\rangle_{h,\tau}\ge0. ⟨ g ( u ∗ ) , u − u ∗ ⟩ h , τ ≥ 0.
Write the control array as a vector with coordinates i = 1 , … , N ( M + 1 ) i=1,\ldots,N(M+1) i = 1 , … , N ( M + 1 ) . Coordinates at their bounds form the active sets :
A 0 : = { i : u i ∗ = 0 } , A U : = { i : u i ∗ = U max } , A_0:=\{i:u_i^*=0\},\qquad A_U:=\{i:u_i^*=U_{\max}\}, A 0 := { i : u i ∗ = 0 } , A U := { i : u i ∗ = U m a x } ,
The other coordinates form the free set :
F : = { i : 0 < u i ∗ < U max } , A : = A 0 ∪ A U . F:=\{i:0<u_i^*<U_{\max}\},\qquad A:=A_0\cup A_U. F := { i : 0 < u i ∗ < U m a x } , A := A 0 ∪ A U .
The tangent cone and critical cone in R N ( M + 1 ) \mathbb{R}^{N(M+1)} R N ( M + 1 ) are
T ( u ∗ ) : = { d : d i ≥ 0 on A 0 , d i ≤ 0 on A U } , T(u^*):=\{d:d_i\ge0\text{ on }A_0,\ d_i\le0\text{ on }A_U\}, T ( u ∗ ) := { d : d i ≥ 0 on A 0 , d i ≤ 0 on A U } ,
and
C ( u ∗ ) : = { d ∈ T ( u ∗ ) : ⟨ g ( u ∗ ) , d ⟩ h , τ = 0 } . \mathcal{C}(u^*):=\{d\in T(u^*):\langle g(u^*),d\rangle_{h,\tau}=0\}. C ( u ∗ ) := { d ∈ T ( u ∗ ) : ⟨ g ( u ∗ ) , d ⟩ h , τ = 0 } .
Strict complementarity consists of the following two conditions, the first for every i ∈ A 0 i\in A_0 i ∈ A 0 and the second for every j ∈ A U j\in A_U j ∈ A U :
g i ( u ∗ ) > 0 , g j ( u ∗ ) < 0 , g_i(u^*)>0,\qquad g_j(u^*)<0, g i ( u ∗ ) > 0 , g j ( u ∗ ) < 0 ,
and, at a stationary control, gives
C ( u ∗ ) = { d ∈ R N ( M + 1 ) : d i = 0 on A } . \mathcal{C}(u^*)=\{d\in\mathbb{R}^{N(M+1)}:d_i=0\text{ on }A\}. C ( u ∗ ) = { d ∈ R N ( M + 1 ) : d i = 0 on A } .
Theorem 8.1 (Fixed-grid second-order conditions ). At a local minimizer, for every d ∈ C ( u ∗ ) d\in\mathcal{C}(u^*) d ∈ C ( u ∗ ) ,
⟨ H ( u ∗ ) d , d ⟩ h , τ ≥ 0. \langle H(u^*)d,d\rangle_{h,\tau}\ge0. ⟨ H ( u ∗ ) d , d ⟩ h , τ ≥ 0. Conversely, if u ∗ u^* u ∗ is stationary and, for some κ > 0 \kappa>0 κ > 0 and every d ∈ C ( u ∗ ) d\in\mathcal{C}(u^*) d ∈ C ( u ∗ ) ,
⟨ H ( u ∗ ) d , d ⟩ h , τ ≥ κ ∥ d ∥ h , τ 2 , \langle H(u^*)d,d\rangle_{h,\tau}\ge\kappa\lVert d\rVert_{h,\tau}^{2}, ⟨ H ( u ∗ ) d , d ⟩ h , τ ≥ κ ∥ d ∥ h , τ 2 , then, for some c > 0 c>0 c > 0 and all nearby feasible u u u , we have
j ( u ) ≥ j ( u ∗ ) + c ∥ u − u ∗ ∥ h , τ 2 . j(u)\ge j(u^*)+c\lVert u-u^*\rVert_{h,\tau}^{2}. j ( u ) ≥ j ( u ∗ ) + c ∥ u − u ∗ ∥ h , τ 2 .
See Appendix E.
Relaxation inside the projection
For θ > 0 \theta>0 θ > 0 , the projected sweep is
Φ θ : U a d , h , τ ⟶ U a d , h , τ \Phi_\theta:\mathcal{U}_{\mathrm{ad},h,\tau}\longrightarrow\mathcal{U}_{\mathrm{ad},h,\tau} Φ θ : U ad , h , τ ⟶ U ad , h , τ
u ⟼ Π [ 0 , U max ] ( u − θ λ g ( u ) ) . (30) u\longmapsto\Pi_{[0,U_{\max}]}\left(u-\frac{\theta}{\lambda}g(u)\right). \tag*{(30)} u ⟼ Π [ 0 , U m a x ] ( u − λ θ g ( u ) ) . (30)
Extend Φ θ \Phi_\theta Φ θ to O h , τ \mathcal{O}_{h,\tau} O h , τ by the same formula. For IMEX, the unprojected update is
u n − g n ( u ) λ = g n ( u ) λ v 2 n ∘ ( η + p 2 n + 1 ) . u^n-\frac{g^n(u)}{\lambda}=\frac{g^n(u)}{\lambda}v_2^n\circ(\eta+p_2^{n+1}). u n − λ g n ( u ) = λ g n ( u ) v 2 n ∘ ( η + p 2 n + 1 ) .
FBS is projected gradient descent with step θ / λ \theta/\lambda θ / λ , with relaxation applied before projection. Reversing this order changes the update on active coordinates. The choice θ ≤ 1 \theta\le1 θ ≤ 1 gives under-relaxation. For finite index sets, vertical bars denote the number of elements. For F = { i 1 < ⋯ < i ∣ F ∣ } ≠ ∅ F=\{i_1<\cdots<i_{\lvert F\rvert}\}\ne\varnothing F = { i 1 < ⋯ < i ∣ F ∣ } = ∅ , define insertion and projection by
E F : R ∣ F ∣ ⟶ R N ( M + 1 ) E_F:\mathbb{R}^{\lvert F\rvert}\longrightarrow\mathbb{R}^{N(M+1)} E F : R ∣ F ∣ ⟶ R N ( M + 1 )
x ⟼ ∑ k = 1 ∣ F ∣ x k e i k , x\longmapsto\sum_{k=1}^{\lvert F\rvert}x_ke_{i_k}, x ⟼ k = 1 ∑ ∣ F ∣ x k e i k ,
and
P F : R N ( M + 1 ) ⟶ R N ( M + 1 ) P_F:\mathbb{R}^{N(M+1)}\longrightarrow\mathbb{R}^{N(M+1)} P F : R N ( M + 1 ) ⟶ R N ( M + 1 )
d ⟼ E F E F T d , d\longmapsto E_FE_F^Td, d ⟼ E F E F T d ,
where ( e i ) i = 1 N ( M + 1 ) (e_i)_{i=1}^{N(M+1)} ( e i ) i = 1 N ( M + 1 ) is the coordinate basis of R N ( M + 1 ) \mathbb{R}^{N(M+1)} R N ( M + 1 ) . Set
W F : = E F T W E F ∈ R ∣ F ∣ × ∣ F ∣ , W_F:=E_F^TWE_F\in\mathbb{R}^{\lvert F\rvert\times\lvert F\rvert}, W F := E F T W E F ∈ R ∣ F ∣ × ∣ F ∣ ,
and
H F : = P F H ( u ∗ ) P F ∣ F ≃ E F T H ( u ∗ ) E F ∈ R ∣ F ∣ × ∣ F ∣ . H_F:=P_FH(u^*)P_F\big|_F\simeq E_F^TH(u^*)E_F\in\mathbb{R}^{\lvert F\rvert\times\lvert F\rvert}. H F := P F H ( u ∗ ) P F F ≃ E F T H ( u ∗ ) E F ∈ R ∣ F ∣ × ∣ F ∣ .
Here ∣ F \big|_F F means restriction to the free-coordinate subspace. Use ∥ x ∥ W F : = ( x T W F x ) 1 / 2 \lVert x\rVert_{W_F}:=(x^TW_Fx)^{1/2} ∥ x ∥ W F := ( x T W F x ) 1/2 and set P F : = 0 P_F:=0 P F := 0 if F = ∅ F=\varnothing F = ∅ .
Theorem 8.2 (Exact local FBS criterion ). Suppose u ∗ u^* u ∗ is stationary and strictly complementary. Then Φ θ \Phi_\theta Φ θ is differentiable at u ∗ u^* u ∗ and
D Φ θ ( u ∗ ) = P F ( I − θ λ H ( u ∗ ) ) . D\Phi_\theta(u^*)=P_F\left(I-\frac{\theta}{\lambda}H(u^*)\right). D Φ θ ( u ∗ ) = P F ( I − λ θ H ( u ∗ ) ) . Near u ∗ u^* u ∗ , one step fixes all active coordinates at their bounds [36 ] . If F = ∅ F=\varnothing F = ∅ , the map is locally constant. Otherwise, on the free affine space,
D Φ θ ( u ∗ ) ∣ F = I F − θ λ H F . D\Phi_\theta(u^*)\big|_F=I_F-\frac{\theta}{\lambda}H_F. D Φ θ ( u ∗ ) F = I F − λ θ H F . If H F H_F H F is positive definite, with smallest and largest weighted eigenvalues 0 < μ min ≤ μ max 0<\mu_{\min}\le\mu_{\max} 0 < μ m i n ≤ μ m a x , the linear iteration factor on the free coordinates is
ρ θ : = max { ∣ 1 − θ μ min λ ∣ , ∣ 1 − θ μ max λ ∣ } . (31) \rho_\theta:=\max\left\{\left|1-\frac{\theta\mu_{\min}}{\lambda}\right|,\left|1-\frac{\theta\mu_{\max}}{\lambda}\right|\right\}. \tag*{(31)} ρ θ := max { 1 − λ θ μ m i n , 1 − λ θ μ m a x } . (31) It is below one exactly when
0 < θ < θ max : = 2 λ μ max . 0<\theta<\theta_{\max}:=\frac{2\lambda}{\mu_{\max}}. 0 < θ < θ m a x := μ m a x 2 λ . For 0 < θ < θ max 0<\theta<\theta_{\max} 0 < θ < θ m a x , convergence is locally q q q -linear once the active set is identified. The optimal positive relaxation is
θ R : = 2 λ μ min + μ max , ρ θ R = μ max − μ min μ max + μ min . \theta_R:=\frac{2\lambda}{\mu_{\min}+\mu_{\max}},\qquad
\rho_{\theta_R}=\frac{\mu_{\max}-\mu_{\min}}{\mu_{\max}+\mu_{\min}}. θ R := μ m i n + μ m a x 2 λ , ρ θ R = μ m a x + μ m i n μ m a x − μ m i n . For under-relaxation, use θ o p t : = min { 1 , θ R } \theta_{\mathrm{opt}}:=\min\{1,\theta_R\} θ opt := min { 1 , θ R } in (31) . For θ > θ max \theta>\theta_{\max} θ > θ m a x , the fixed point is linearly unstable.
Here q q q -linear convergence means that each error norm is eventually bounded by a fixed factor smaller than one times the preceding error norm. Locally, FBS is Richardson iteration on the free coordinates (Appendix § 19 ). Global FBS convergence is studied in [20 ] .
Matrix-free spectral diagnostics
Set
H ~ F : = W F 1 / 2 H F W F − 1 / 2 , \widetilde{H}_{F} := W_{F}^{1/2} H_{F} W_{F}^{-1/2}, H F := W F 1/2 H F W F − 1/2 ,
which is Euclidean symmetric by weighted symmetry. For y ∈ R ∣ F ∣ y \in\mathbb{R}^{|F|} y ∈ R ∣ F ∣ , its action is
H ~ F : R ∣ F ∣ ⟶ R ∣ F ∣ y ⟼ W F 1 / 2 E F T H ( u ∗ ) E F W F − 1 / 2 y . \widetilde{H}_{F} : \mathbb{R}^{|F|} \longrightarrow\mathbb{R}^{|F|} \\
y \longmapsto W_{F}^{1/2} E_{F}^{T} H(u^{*}) E_{F} W_{F}^{-1/2} y. H F : R ∣ F ∣ ⟶ R ∣ F ∣ y ⟼ W F 1/2 E F T H ( u ∗ ) E F W F − 1/2 y .
For a matrix A A A , σ ( A ) \sigma(A) σ ( A ) denotes its set of eigenvalues and ∥ A ∥ 2 \lVert A\rVert_{2} ∥ A ∥ 2 the operator norm induced by the Euclidean vector norm. Positive diagonal matrix powers, such as W F 1 / 2 W_{F}^{1/2} W F 1/2 , are taken entrywise on the diagonal. We estimate the smallest and largest eigenvalues with Lanczos and mark the estimates with a hat. For nonempty F F F and an approximate eigenpair ( μ ^ , y ^ ) (\widehat{\mu}, \widehat{y}) ( μ , y ) with ∥ y ^ ∥ 2 = 1 \lVert\widehat{y}\rVert_{2}=1 ∥ y ∥ 2 = 1 , the Ritz residual bounds the distance to the spectrum σ ( H ~ F ) \sigma(\widetilde{H}_{F}) σ ( H F ) :
min μ ∈ σ ( H ~ F ) ∣ μ ^ − μ ∣ ≤ ∥ H ~ F y ^ − μ ^ y ^ ∥ 2 . \min_{\mu\in\sigma(\widetilde{H}_{F})} \left|\widehat{\mu}-\mu\right| \le\left\lVert\widetilde{H}_{F}\widehat{y}-\widehat{\mu}\widehat{y}\right\rVert_{2}. μ ∈ σ ( H F ) min ∣ μ − μ ∣ ≤ H F y − μ y 2 .
See Appendix § 19 .§ 5 .
Algorithms and reproducibility
We accept a run when
∥ R ( u ) ∥ ∞ ≤ 10 − 6 . (32) \lVert R(u)\rVert_{\infty} \le10^{-6}. \tag*{(32)} ∥ R ( u ) ∥ ∞ ≤ 1 0 − 6 . (32)
The four methods are FBS, Anderson-accelerated FBS, spectral projected gradient (SPG), and the limited-memory BFGS method with bound constraints (L-BFGS-B). Appendix § 23 .§ 1 gives their iteration formulas and parameters. To compare their work across hardware, we count state and adjoint solves. The code and run instructions are in [1 ] .
Numerical results
Reference data and comparison protocol
We use Ω : = ( 0 , 1 ) \Omega:= (0,1) Ω := ( 0 , 1 ) , ( T , M , N ) : = ( 4 , 80 , 120 ) (T,M,N):=(4,80,120) ( T , M , N ) := ( 4 , 80 , 120 ) , ( m , r , α , β , q ) : = ( 1 , 1.2 , 1 , 1 , 1 ) (m,r,\alpha,\beta,q):=(1,1.2,1,1,1) ( m , r , α , β , q ) := ( 1 , 1.2 , 1 , 1 , 1 ) , μ 1 : = μ 2 : = 0.002 \mu_{1}:=\mu_{2}:=0.002 μ 1 := μ 2 := 0.002 and
( U max , γ 1 , γ 2 , λ , η ) : = ( 2 , 0.5 , 0.5 , 0.05 , 0.8 ) , (U_{\max},\gamma_{1},\gamma_{2},\lambda,\eta):=(2,0.5,0.5,0.05,0.8), ( U m a x , γ 1 , γ 2 , λ , η ) := ( 2 , 0.5 , 0.5 , 0.05 , 0.8 ) ,
with initial profiles
v 1 0 ( x ) : = 1.3 e − ( x − 0.30 ) 2 / 0.02 + 0.05 , v_{1}^{0}(x):=1.3e^{-(x-0.30)^{2}/0.02}+0.05, v 1 0 ( x ) := 1.3 e − ( x − 0.30 ) 2 /0.02 + 0.05 ,
v 2 0 ( x ) : = 1.4 e − ( x − 0.70 ) 2 / 0.02 + 0.05 , v_{2}^{0}(x):=1.4e^{-(x-0.70)^{2}/0.02}+0.05, v 2 0 ( x ) := 1.4 e − ( x − 0.70 ) 2 /0.02 + 0.05 ,
and targets
v 1 d ( x ) : = r / β + 0.15 cos ( 2 π x ) , v_{1}^{d}(x):=r/\beta+0.15\cos(2\pi x), v 1 d ( x ) := r / β + 0.15 cos ( 2 π x ) ,
v 2 d ( x ) : = m / α − 0.15 cos ( 2 π x ) . v_{2}^{d}(x):=m/\alpha-0.15\cos(2\pi x). v 2 d ( x ) := m / α − 0.15 cos ( 2 π x ) .
Unless stated otherwise, all runs start from zero. For baseline validation, we use IMEX with M = 80 M=80 M = 80 , N = 1920 N=1920 N = 1920 . For each tested N ∣ 1920 N\mid1920 N ∣ 1920 , copy each control value onto its corresponding finer time intervals. The common objective is
J h , T / 1920 I ( ( u j ⌊ k N / 1920 ⌋ ) k = 0 1919 ) . J_{h,T/1920}^{\mathrm{I}}\left(\left(u_{j}^{\lfloor kN/1920\rfloor}\right)_{k=0}^{1919}\right). J h , T /1920 I ( ( u j ⌊ k N /1920 ⌋ ) k = 0 1919 ) .
Here ⌊ s ⌋ \lfloor s\rfloor ⌊ s ⌋ is the largest integer not exceeding s s s . The derivative and Hessian tests recover the expected Taylor orders. Appendix § 23 .§ 2 gives these results alongside the weighted symmetry defects and positivity tests.
How harvest quadrature changes the computed control
Table 1 and Figure 1 show the controls reached from zero. A coordinate i i i is counted as free when 10 − 8 < u i < U max − 10 − 8 10^{-8}<u_{i}<U_{\max}-10^{-8} 1 0 − 8 < u i < U m a x − 1 0 − 8 . Of the 9720 coordinates, 18 are free for P 0 P_{0} P 0 , compared with 3127 for I and 3076 for P 1 P_{1} P 1 . To measure switching, define the temporal variation by
Table 1. Zero-start reference controls. Common validation uses the same IMEX objective with N = 1920 N = 1920 N = 1920 .
Figure 1. Reference controls on a common color scale: I (left), P 0 P_0 P 0 (center), P 1 P_1 P 1 (right). The positive P 0 P_0 P 0 scheme produces temporal switching.
TV t ( u ) : = ∑ n = 0 N − 2 ∑ j = 0 M w j ∣ u j n + 1 − u j n ∣ . \operatorname{TV}_{t}(u):=\sum_{n=0}^{N-2}\sum_{j=0}^{M}w_{j}\left|u_{j}^{n+1}-u_{j}^{n}\right|. TV t ( u ) := n = 0 ∑ N − 2 j = 0 ∑ M w j u j n + 1 − u j n .
P 0 P_{0} P 0 switches sharply in time and has the lowest own-grid cost. On the common validation grid, I has the lowest cost, followed by P 1 P_{1} P 1 , then P 0 P_{0} P 0 .
Free Hessians and the local threshold
The matrix-free Ritz estimates in Table 2 put the relaxation threshold near 0.01 for I and P 1 P_{1} P 1 . IMEX has unrelaxed free modes as low as about − 207.3 -207.3 − 207.3 . For P 0 P_{0} P 0 , the threshold exceeds one, but the small μ ^ min \widehat{\mu}_{\min} μ m i n gives ρ ^ ≈ 0.9995 \widehat{\rho}\approx0.9995 ρ ≈ 0.9995 at the shared θ = 0.0093 \theta=0.0093 θ = 0.0093 by (31) . Thus the local sweep is predicted to converge slowly near this stationary control. For the largest computed mode d d d , the local directional test compares
Table 2. Ritz estimates of free spectra and relaxation thresholds. The last column gives the Euclidean residual for the smallest computed eigenpair of H ^ F \widehat{H}_F H F .
Φ θ ( u + ε d ) − Φ θ ( u ) ε \frac{\Phi_{\theta}(u+\varepsilon d)-\Phi_{\theta}(u)}{\varepsilon} ε Φ θ ( u + ε d ) − Φ θ ( u )
with d d d . At θ = 0.95 θ ^ max \theta=0.95\widehat{\theta}_{\max} θ = 0.95 θ m a x and 1.05 θ ^ max 1.05\widehat{\theta}_{\max} 1.05 θ m a x , the predicted factors are − 0.9 -0.9 − 0.9 and − 1.1 -1.1 − 1.1 , with magnitudes 0.9 and 1.1. All three problems match these predictions within 3 × 10 − 4 3\times10^{-4} 3 × 1 0 − 4 . For P 0 P_{0} P 0 , both tests use extrapolated steps, θ > 1 \theta>1 θ > 1 .
Algorithm comparison and rejected runs
All four methods pass the acceptance test for I and P 1 P_{1} P 1 , with L-BFGS-B requiring the fewest state and adjoint solves. The two sweeps reach the iteration limit for P 0 P_0 P 0 . SPG and L-BFGS-B reach stationary controls, but with different costs. Appendix I.3 gives the parameters, final residuals and solve counts for every run.
Figure 3 compares the residual histories against elapsed time on the reference grid M = 80 M = 80 M = 80 , N = 120 N = 120 N = 120 , starting from zero. We repeat each method–problem pair five times, with all runs executed sequentially and thread limits set to one. A monotonic clock measures each solver call, including final diagnostics. JIT compilation is warmed up beforehand. We perform the independent residual check after timing. Each curve is the run with median total runtime, rather than an average of trajectories.
Figure 3. Stationarity residual against elapsed time for the reference problem with M = 80 M = 80 M = 80 , N = 120 N = 120 N = 120 and zero initial control. Each curve shows the run with median total runtime among five sequential repetitions. Horizontal scales differ between panels. The dashed horizontal line marks the 10 − 6 10^{-6} 1 0 − 6 acceptance threshold. Crosses mark runs that reached the iteration limit without satisfying this threshold.
Table 3 reports total runtimes under each method’s stopping rules from Appendix I.1. These durations need not equal the first time the residual crosses the common 10 − 6 10^{-6} 1 0 − 6 acceptance threshold. The runs used ARM64 macOS 27.2.1, Python 3.12.14, NumPy 2.3.5, SciPy 1.17.0 and Numba 0.65.1. SPG has the shortest median runtime for I. L-BFGS-B is fastest for P 1 P_1 P 1 and among the accepted P 0 P_0 P 0 runs. The two accepted P 0 P_0 P 0 methods reach different objective values, as reported in Appendix I.3.
Table 3. Total runtimes in seconds over five repetitions, shown as median [minimum, maximum]. A dagger marks runs that reached 12000 iterations without passing the acceptance test.
Control differences under temporal refinement
With M = 80 M = 80 M = 80 fixed, we refine through N = 30 , 60 , 120 , 240 , 480 , 960 N = 30, 60, 120, 240, 480, 960 N = 30 , 60 , 120 , 240 , 480 , 960 . The first run starts from zero at N = 30 N = 30 N = 30 . Later runs start from the previous control, copied to the two corresponding finer intervals. At N = 120 N = 120 N = 120 , we replace that control with the reference control computed from zero before continuing the refinement. For candidates u N I , u N P ∈ R N ( M + 1 ) u_N^{I}, u_N^{P} \in\mathbb{R}^{N(M+1)} u N I , u N P ∈ R N ( M + 1 ) on the same grid, set
D ζ ( N ) : = ∥ u N I − u N P ∥ h , τ . D_{\zeta}(N) := \lVert u_N^{I} - u_N^{P} \rVert_{h,\tau}. D ζ ( N ) := ∥ u N I − u N P ∥ h , τ .
All controls pass the residual check. The P 1 P_1 P 1 distance decreases with ratios tending to two. The P 0 P_0 P 0 distance follows a different path: it stays near one through N = 120 N = 120 N = 120 , then drops on finer grids (Table 4 ). For state self-convergence at a fixed control, Appendix I.4 separates time and space refinement.
Table 4. Temporal refinement at M = 80 M = 80 M = 80 . All candidates pass the independent residual check.
Coarse-grid optimization and multistart
At N = 4 N = 4 N = 4 , the accepted IMEX control gives negative states. Both positive problems keep their states nonnegative, yet their common validation costs fall on opposite sides of I: higher for P 0 P_0 P 0 , lower for P 1 P_1 P 1 (Table 5 ). The ordering changes with N N N . The five initial guesses are 0, 1, U max U_{\max} U m a x , independent uniform samples on [ 0 , U max ] [0,U_{\max}] [ 0 , U m a x ] , and U max 1 [ 0 , T / 2 ] ( t ) U_{\max}\mathbf{1}_{[0,T/2]}(t) U m a x 1 [ 0 , T /2 ] ( t ) . Every control passes the residual check. For I and P 1 P_1 P 1 , the costs differ only by rounding, and weighted distances to the zero-start control stay below 3.1 × 10 − 7 3.1 \times10^{-7} 3.1 × 1 0 − 7 . The P 0 P_0 P 0 runs reach distinct controls, with a spread of about 8.02 × 10 − 4 8.02 \times10^{-4} 8.02 × 1 0 − 4 in cost and 1.425 in distance (Table 6 ).
Table 5. Coarse-grid candidates and common fine-grid validation. All candidates meet the residual tolerance.
Table 6. Results from five initial guesses. Distances are weighted L 2 L^2 L 2 distances to the zero-start control.
Targeted tests of the quadrature mechanism
The next tests vary one parameter at a time. They also reverse the direction of continuation and separate time refinement from space refinement.
Figure 2. Minimum state at fixed maximal effort (left), predicted free linear factor against normalized relaxation (right).
Exact global verification in the removal model
Take τ = q = y 0 = λ = 1 \tau=q=y^0=\lambda=1 τ = q = y 0 = λ = 1 , η = 2 \eta=2 η = 2 , U max = 3 U_{\max}=3 U m a x = 3 , and ( x , z ) : = ( u 0 , u 1 ) ∈ [ 0 , 3 ] 2 (x,z):=(u^0,u^1)\in[0,3]^2 ( x , z ) := ( u 0 , u 1 ) ∈ [ 0 , 3 ] 2 . The two costs are
F 0 : [ 0 , 3 ] 2 ⟶ R F_0:[0,3]^2\longrightarrow\mathbb{R} F 0 : [ 0 , 3 ] 2 ⟶ R
( x , z ) ⟼ 1 2 ( x 2 + z 2 ) − 2 ( x + z 1 + x ) . (x,z)\longmapsto\frac{1}{2}(x^2+z^2)-2\left(x+\frac{z}{1+x}\right). ( x , z ) ⟼ 2 1 ( x 2 + z 2 ) − 2 ( x + 1 + x z ) .
and
F 1 : [ 0 , 3 ] 2 ⟶ R F_1:[0,3]^2\longrightarrow\mathbb{R} F 1 : [ 0 , 3 ] 2 ⟶ R
( x , z ) ⟼ 1 2 ( x 2 + z 2 ) + 2 ( 1 + x ) ( 1 + z ) − 2. (x,z)\longmapsto\frac{1}{2}(x^2+z^2)+\frac{2}{(1+x)(1+z)}-2. ( x , z ) ⟼ 2 1 ( x 2 + z 2 ) + ( 1 + x ) ( 1 + z ) 2 − 2.
The old-state global minimizer is near
( 1.822 , 0.7087 ) , (1.822, 0.7087), ( 1.822 , 0.7087 ) ,
with cost − 2.235 -2.235 − 2.235 . The extra strict local minimum ( 0 , 2 ) (0,2) ( 0 , 2 ) has cost − 2 -2 − 2 (Appendix § 20 .§ 7 ). The unique flux minimizer is ( t , t ) (t,t) ( t , t ) , t ≈ 0.5437 t\approx0.5437 t ≈ 0.5437 , with cost − 0.8651 -0.8651 − 0.8651 . For T = 2 T=2 T = 2 and K c : = { u ∈ L 2 ( 0 , 2 ) : 0 ≤ u ≤ 3 a . e . } K_c:=\{u\in L^2(0,2):0\leq u\leq3\ \mathrm{a.e.}\} K c := { u ∈ L 2 ( 0 , 2 ) : 0 ≤ u ≤ 3 a.e. } , the common exact continuous cost is
j c : K c ⟶ R j_c:K_c\longrightarrow\mathbb{R} j c : K c ⟶ R
u ⟼ 1 2 ∫ 0 2 u ( t ) 2 d t − 2 ( 1 − e − ∫ 0 2 u ( t ) d t ) . u\longmapsto\frac{1}{2}\int_0^2u(t)^2\,\mathrm{d}t-2\left(1-\mathrm{e}^{-\int_0^2u(t)\,\mathrm{d}t}\right). u ⟼ 2 1 ∫ 0 2 u ( t ) 2 d t − 2 ( 1 − e − ∫ 0 2 u ( t ) d t ) .
The continuous cost has the unique minimizer u ≡ s u\equiv s u ≡ s , s ≈ 0.6011 s\approx0.6011 s ≈ 0.6011 , with value − 1.038 -1.038 − 1.038 . For two-step controls,
j c ( x , z ) = 1 2 ( x 2 + z 2 ) − 2 + 2 e − ( x + z ) ; j_c(x,z)=\frac{1}{2}(x^2+z^2)-2+2\mathrm{e}^{-(x+z)}; j c ( x , z ) = 2 1 ( x 2 + z 2 ) − 2 + 2 e − ( x + z ) ;
it takes values 0.07021 and − 1.030 -1.030 − 1.030 at the certified old-state and flux global minimizers. The quadrature bias is visible in these values.
With η = 10 \eta=10 η = 10 , u 0 + u 1 = 2 u^0+u^1=2 u 0 + u 1 = 2 and the other coefficients unchanged, Proposition 6.3 gives old-state minimizers ( 0 , 2 ) , ( 2 , 0 ) (0,2),(2,0) ( 0 , 2 ) , ( 2 , 0 ) , while the flux and continuous costs share the unique allocation ( 1 , 1 ) (1,1) ( 1 , 1 ) (Figure 5 ).
Figure 5. Pure removal. Left: two-step costs at η = 2 \eta= 2 η = 2 . Red markers identify old-state global and extra local minima and the unique flux minimum. Right: at η = 10 \eta= 10 η = 10 and fixed total effort, old-state quadrature favors extreme allocations, while the flux and continuous costs favor equal allocation.
A common reference and acceptance protocol
For u ∈ U a d , h , τ u\in\mathcal{U}_{\mathrm{ad},h,\tau} u ∈ U ad , h , τ , spatial interpolation gives U r e f ( t ) ∈ R M r e f + 1 U_{\mathrm{ref}}(t)\in\mathbb{R}^{M_{\mathrm{ref}}+1} U ref ( t ) ∈ R M ref + 1 , constant on each original time interval. Set h r e f : = 1 / M r e f h_{\mathrm{ref}}:=1/M_{\mathrm{ref}} h ref := 1/ M ref , M r e f : = 80 M_{\mathrm{ref}}:=80 M ref := 80 , and integrate
V ˙ 1 = μ 1 L h r e f V 1 + ( − m 1 + α V 2 ) ∘ V 1 , \dot{V}_1=\mu_1L_{h_{\mathrm{ref}}}V_1+(-m_1+\alpha V_2)\circ V_1, V ˙ 1 = μ 1 L h ref V 1 + ( − m 1 + α V 2 ) ∘ V 1 ,
V ˙ 2 = μ 2 L h r e f V 2 + ( r 1 − β V 1 − q U r e f ) ∘ V 2 . \dot{V}_2=\mu_2L_{h_{\mathrm{ref}}}V_2+(r_1-\beta V_1-qU_{\mathrm{ref}})\circ V_2. V ˙ 2 = μ 2 L h ref V 2 + ( r 1 − β V 1 − q U ref ) ∘ V 2 .
where V i : [ 0 , T ] ⟶ R M r e f + 1 V_i:[0,T]\longrightarrow\mathbb{R}^{M_{\mathrm{ref}}+1} V i : [ 0 , T ] ⟶ R M ref + 1 and V ˙ i \dot{V}_i V ˙ i denotes its time derivative. We sample the analytic initial and target profiles at the reference nodes, then integrate the four cost terms alongside the state equations. DOP853 restarts at every control jump. With relative and absolute tolerances of 10 − 9 10^{-9} 1 0 − 9 and 10 − 11 10^{-11} 1 0 − 11 , this gives a continuous-time reference on the shared spatial grid. We check the solver against the exact exponential solution for zero predators and piecewise constant harvesting. Of 94 final targeted controls, 93 satisfy {(citep}(32)). We retry failed runs with L-BFGS memory eight instead of 38.
Curvature at common controls
Both positive problems produce the same state at a fixed control. On M = 32 M = 32 M = 32 , take
u n ( x ) : = 0.7 + 0.2 cos ( 2 π x ) cos ( 2 π t n / T ) . u^n(x) := 0.7 + 0.2 \cos(2\pi x)\cos(2\pi t_n/T). u n ( x ) := 0.7 + 0.2 cos ( 2 π x ) cos ( 2 π t n / T ) .
The alternating direction d ∈ R N ( M + 1 ) d \in\mathbb{R}^{N(M+1)} d ∈ R N ( M + 1 ) , with entries d j n : = ( − 1 ) n / T d_j^n := (-1)^n/\sqrt{T} d j n := ( − 1 ) n / T , has ∥ d ∥ h , τ = 1 \lVert d\rVert_{h,\tau} = 1 ∥ d ∥ h , τ = 1 because ∑ j = 0 M w j = 1 \sum_{j=0}^{M} w_j = 1 ∑ j = 0 M w j = 1 . Exact Hessian actions give
Along this direction, P 0 P_0 P 0 has negative curvature on coarse grids, while P 1 P_1 P 1 has positive curvature. Refinement brings both towards λ = 0.05 \lambda= 0.05 λ = 0.05 . At these controls, measure ∥ g 1 ( u ) − g 0 ( u ) ∥ h , τ \lVert g_1(u) - g_0(u)\rVert_{h,\tau} ∥ g 1 ( u ) − g 0 ( u ) ∥ h , τ and set
B ( u ) : = W 1 / 2 ( H 0 ( u ) − H 1 ( u ) ) W − 1 / 2 ∈ R N ( M + 1 ) × N ( M + 1 ) . B(u) := W^{1/2}(H_0(u) - H_1(u))W^{-1/2}
\in\mathbb{R}^{N(M+1)\times N(M+1)}. B ( u ) := W 1/2 ( H 0 ( u ) − H 1 ( u )) W − 1/2 ∈ R N ( M + 1 ) × N ( M + 1 ) .
For symmetric B ( u ) B(u) B ( u ) , let ν ^ \hat{\nu} ν ^ be the Ritz value with largest magnitude. We estimate the norm
∥ B ( u ) ∥ 2 = ∥ H 1 ( u ) − H 0 ( u ) ∥ L ( R N ( M + 1 ) ) \lVert B(u)\rVert_2 = \lVert H_1(u) - H_0(u)\rVert_{\mathcal{L}(\mathbb{R}^{N(M+1)})} ∥ B ( u ) ∥ 2 = ∥ H 1 ( u ) − H 0 ( u ) ∥ L ( R N ( M + 1 ) )
by ∣ ν ^ ∣ \lvert\hat{\nu}\rvert ∣ ν ^ ∣ and, for ∥ y ∥ 2 = 1 \lVert y\rVert_2 = 1 ∥ y ∥ 2 = 1 , record the residual ∥ B ( u ) y − ν ^ y ∥ 2 \lVert B(u)y - \hat{\nu}y\rVert_2 ∥ B ( u ) y − ν ^ y ∥ 2 . The tests cover M = 8 , 16 , 32 M = 8, 16, 32 M = 8 , 16 , 32 , N = 16 , 32 , … , 1024 N = 16, 32, \ldots, 1024 N = 16 , 32 , … , 1024 , using both the smooth control and the spatially constant alternating control u n = 0.7 + 0.2 ( − 1 ) n u^n = 0.7 + 0.2(-1)^n u n = 0.7 + 0.2 ( − 1 ) n .
At M = 32 M = 32 M = 32 , a fit over N = 256 , 512 , 1024 N = 256, 512, 1024 N = 256 , 512 , 1024 gives gradient-difference rates 0.981, 0.988 and Hessian-estimate rates 0.997, 1.004, respectively. At N = 1024 N = 1024 N = 1024 , division by τ \tau τ gives gradient differences 1.888, 1.987 and Hessian estimates 9.444, 8.320. The measured rates agree with Theorem 7.1.
Figure 4. Same-grid control differences from I (left), common N = 1920 N = 1920 N = 1920 validation (right). The differences shrink on the finer time grids.
Continuation and parameter changes
On M = 20 M = 20 M = 20 , N = 60 N = 60 N = 60 , we increase η \eta η through 0, 0.2, . . . , 1.2, then run a separate sequence in reverse. Each sequence starts from zero and passes its accepted control to the next run. The P 0 P_0 P 0 controls from opposite directions have L 2 L^2 L 2 distances of 1.558–1.576 at η = 0.4 , . . . , 1.2 \eta= 0.4, . . . , 1.2 η = 0.4 , ... , 1.2 , but agree within about 1.1 × 10 − 6 1.1 \times10^{-6} 1.1 × 1 0 − 6 at η = 0 , 0.2 \eta= 0, 0.2 η = 0 , 0.2 . All P 1 P_1 P 1 distances stay below 4.9 × 10 − 7 4.9 \times10^{-7} 4.9 × 1 0 − 7 . For the old-state rule, the direction of continuation changes the control reached. At η = 0.8 \eta= 0.8 η = 0.8 , we start from zero, one, the upper bound, a seeded uniform random control, and the upper bound until 0.45 T 0.45T 0.45 T followed by zero. All ten runs pass the residual test. The P 0 P_0 P 0 runs reach distinct controls: costs range from − 1.653 -1.653 − 1.653 to − 1.648 -1.648 − 1.648 and temporal variation from 6.501 to 19.61. The P 1 P_1 P 1 runs reach the same control, with costs agreeing within 9 × 10 − 16 9 \times10^{-16} 9 × 1 0 − 16 at − 1.484 -1.484 − 1.484 and variation about 2.199 2.199 2.199 . Separate series vary η = 0 , 0.2 , 0.4 , 0.8 , 1.2 \eta= 0, 0.2, 0.4, 0.8, 1.2 η = 0 , 0.2 , 0.4 , 0.8 , 1.2 , λ = 0.02 , 0.05 , 0.1 , 0.2 \lambda= 0.02, 0.05, 0.1, 0.2 λ = 0.02 , 0.05 , 0.1 , 0.2 and T = 1 , 2 , 4 , 6 T = 1, 2, 4, 6 T = 1 , 2 , 4 , 6 . In the horizon series, N = 15 T N = 15T N = 15 T keeps τ = 1 / 15 \tau= 1/15 τ = 1/15 fixed. The rewards agree at η = 0 \eta= 0 η = 0 . With positive η \eta η and a zero initial guess, P 1 P_1 P 1 has the lower common-reference cost in every case, by 0.000502 0.000502 0.000502 –0.07664 0.07664 0.07664 . The finer time grids below reverse this ranking. At T = 6 T = 6 T = 6 , temporal variation is 24.57 24.57 24.57 for P 0 P_0 P 0 and 2.341 2.341 2.341 for P 1 P_1 P 1 .
Figure 6. Quadrature perturbations at common controls, M = 32 M = 32 M = 32 . Left and middle: weighted gradient differences and largest absolute Ritz values of the mass-scaled Hessian difference. Right: normalized alternating-direction curvature at the smooth control.
Separate temporal and spatial refinement
With M = 20 M = 20 M = 20 fixed, we refine through N = 30 , 60 , 120 , 240 , 480 , 960 N = 30, 60, 120, 240, 480, 960 N = 30 , 60 , 120 , 240 , 480 , 960 , starting each run from the previous accepted control. The P 0 P_0 P 0 temporal variations are
5.400 , 9.100 , 19.40 , 12.02 , 2.302 , 2.261 , 5.400,\quad9.100,\quad19.40,\quad12.02,\quad2.302,\quad2.261, 5.400 , 9.100 , 19.40 , 12.02 , 2.302 , 2.261 ,
while the variation for P 1 P_1 P 1 moves gradually from $2.164t o to t o 2.225. Their L 2 L^2 L 2 distances are 0.949 , 1.011 , 1.029 , 0.513 , 0.0963 , 0.0455 0.949, 1.011, 1.029, 0.513, 0.0963, 0.0455 0.949 , 1.011 , 1.029 , 0.513 , 0.0963 , 0.0455 : the gap first grows, then falls sharply. At N = 960 N = 960 N = 960 , the common-reference costs are approximately − 1.55388 -1.55388 − 1.55388 for P 0 P_0 P 0 , − 1.55384 -1.55384 − 1.55384 for P 1 P_1 P 1 and − 1.55389 -1.55389 − 1.55389 for IMEX. IMEX comes first, followed by P 0 P_0 P 0 and P 1 P_1 P 1 .
Space refinement gives a different picture. With N = 120 N = 120 N = 120 fixed, we refine through M = 10 , 20 , 40 , 80 M = 10, 20, 40, 80 M = 10 , 20 , 40 , 80 , and old-state variation grows from 16.48 16.48 16.48 to 27.87 27.87 27.87 . At M = 80 M = 80 M = 80 , restarting with L-BFGS memory eight reduces the residual from 0.01886 0.01886 0.01886 to 1.161 × 10 − 6 1.161 \times10^{-6} 1.161 × 1 0 − 6 , still above the acceptance threshold. All flux controls are accepted, with variation 2.221 2.221 2.221 –2.232 2.232 2.232 . Refining space leaves the temporal switching in place at this coarse time step.
For these final controls, tightening the common-reference relative tolerance from 10 − 8 10^{-8} 1 0 − 8 to 10 − 10 10^{-10} 1 0 − 10 changes costs by at most about 8.3 × 10 − 12 8.3 \times10^{-12} 8.3 × 1 0 − 12 . Increasing M r e f M_{\mathrm{ref}} M ref from 80 to 160 has a much larger effect: about 2.22 × 10 − 4 2.22 \times10^{-4} 2.22 × 1 0 − 4 for IMEX, 1.03 × 10 − 3 1.03 \times10^{-3} 1.03 × 1 0 − 3 for rejected P 0 P_0 P 0 and 2.48 × 10 − 4 2.48 \times10^{-4} 2.48 × 1 0 − 4 for P 1 P_1 P 1 . Here the costs respond much more to spatial resolution than to the time-integration tolerance.
Extension to production–destruction networks
A single flux can account for both donor loss and receiver gain in a network [33 ] . With rates fixed within each step, this shared-flux scheme preserves positivity and records harvest exactly. Let s , d ∈ N s,d \in\mathbb{N} s , d ∈ N count species and spatial nodes, with indices 1 ≤ i , j ≤ s 1 \le i,j \le s 1 ≤ i , j ≤ s , 1 ≤ a ≤ d 1 \le a \le d 1 ≤ a ≤ d . For positive weights c i , w a c_i,w_a c i , w a , set
Figure 7. Opposite continuation directions on M = 20 M = 20 M = 20 , N = 60 N = 60 N = 60 , each starting from zero: temporal variation, same-parameter control distance and common-reference cost difference.
Figure 8. Temporal refinement at M = 20 M = 20 M = 20 , with each run starting from the previous accepted control. Right: weighted L 2 L^2 L 2 distance between consecutive controls, after copying each coarse value to its two finer intervals. The old-state distance first grows, then drops sharply for N ≥ 480 N \ge480 N ≥ 480 . All controls pass the residual test.
y : = ( y i , a ) ∈ R s d , ρ : = ( c i w a ) ∈ ( 0 , ∞ ) s d . y := (y_{i,a}) \in\mathbb{R}^{sd}, \qquad\rho:= (c_iw_a) \in(0,\infty)^{sd}. y := ( y i , a ) ∈ R s d , ρ := ( c i w a ) ∈ ( 0 , ∞ ) s d .
Define the weighted mass by
M : R s d ⟶ R \mathcal{M} : \mathbb{R}^{sd} \longrightarrow\mathbb{R} M : R s d ⟶ R
y ⟼ ρ T y . y \longmapsto\rho^{T}y. y ⟼ ρ T y .
Here ∣ y ∣ |y| ∣ y ∣ is the componentwise absolute value, and R + s d : = [ 0 , ∞ ) s d \mathbb{R}^{sd}_{+} := [0,\infty)^{sd} R + s d := [ 0 , ∞ ) s d is the set of nonnegative vectors. Let V ⊂ R d u \mathcal{V} \subset\mathbb{R}^{du} V ⊂ R d u , d u ∈ N d_u \in\mathbb{N} d u ∈ N , contain the admissible controls. At step 0 ≤ n < N 0 \le n < N 0 ≤ n < N , take y n ∈ R + s d y^n \in\mathbb{R}^{sd}_{+} y n ∈ R + s d , u n ∈ V u^n \in\mathcal{V} u n ∈ V and finite rates k i j , a n ≥ 0 k^n_{ij,a} \ge0 k ij , a n ≥ 0 for transfer from j j j to i ≠ j i \ne j i = j , evaluated at ( y n , u n ) (y^n,u^n) ( y n , u n ) . The donor loses and the receiver gains, respectively k i j , a n y j , a n + 1 k^n_{ij,a}y^{n+1}_{j,a} k ij , a n y j , a n + 1 and ( c j / c i ) k i j , a n y j , a n + 1 (c_j/c_i)k^n_{ij,a}y^{n+1}_{j,a} ( c j / c i ) k ij , a n y j , a n + 1 . Weighting by c j w a c_jw_a c j w a and c i w a c_iw_a c i w a makes the loss and gain equal. Define K n ∈ R s d × s d K^n \in\mathbb{R}^{sd \times sd} K n ∈ R s d × s d as follows. For i ≠ j i \ne j i = j , set
K ( i , a ) , ( j , a ) n : = ( c j / c i ) k i j , a n , K^n_{(i,a),(j,a)} := (c_j/c_i)k^n_{ij,a}, K ( i , a ) , ( j , a ) n := ( c j / c i ) k ij , a n ,
and
K ( j , a ) , ( j , a ) n : = − ∑ i = 1 i ≠ j s k i j , a n . K^n_{(j,a),(j,a)} := -\sum_{\substack{i=1 \\ i\ne j}}^{s} k^n_{ij,a}. K ( j , a ) , ( j , a ) n := − i = 1 i = j ∑ s k ij , a n .
with zero entries between different spatial nodes, so ρ T K n = 0 \rho^T K^n = 0 ρ T K n = 0 . Let L h ∈ R d × d L_h \in\mathbb{R}^{d \times d} L h ∈ R d × d have nonnegative off-diagonal entries and w T L h = 0 w^T L_h = 0 w T L h = 0 , w : = ( w a ) ∈ ( 0 , ∞ ) d w := (w_a) \in(0,\infty)^d w := ( w a ) ∈ ( 0 , ∞ ) d . For μ i ≥ 0 \mu_i \ge0 μ i ≥ 0 , set
G n : = blockdiag i = 1 s ( μ i L h ) + K n . G^n := \operatorname{blockdiag}_{i=1}^{s}(\mu_iL_h) + K^n. G n := blockdiag i = 1 s ( μ i L h ) + K n .
Here blockdiag places the listed matrices on the diagonal and zeros elsewhere. Then G n G^n G n is Metzler (nonnegative off-diagonal entries), with ρ T G n = 0 \rho^T G^n = 0 ρ T G n = 0 . For mortality, harvesting and production m n , h n , p n ∈ R + s d m^n,h^n,p^n \in\mathbb{R}^{sd}_{+} m n , h n , p n ∈ R + s d , use
A n y n + 1 = y n + τ p n , (33) A^n y^{n+1} = y^n + \tau p^n, \tag*{(33)} A n y n + 1 = y n + τ p n , (33)
where
A n : = I s d − τ G n + τ diag ( m n + h n ) ∈ R s d × s d . A^n := I_{sd} - \tau G^n + \tau\operatorname{diag}(m^n+h^n) \in\mathbb{R}^{sd \times sd}. A n := I s d − τ G n + τ diag ( m n + h n ) ∈ R s d × s d .
The rates may depend on ( y n , u n ) (y^n,u^n) ( y n , u n ) . We require them to remain finite even when a state component is zero.
Theorem 12.1 (Positive transfer network and harvest accounting). For every τ > 0 \tau> 0 τ > 0 , nonnegative y n , p n , m n , h n y^n,p^n,m^n,h^n y n , p n , m n , h n and finite nonnegative transfer rates, (33) has a unique nonnegative solution. It satisfies the exact mass balance
M ( y n + 1 ) − M ( y n ) \mathcal{M}(y^{n+1})-\mathcal{M}(y^n) M ( y n + 1 ) − M ( y n ) = τ ρ T p n − τ ρ T diag ( m n + h n ) y n + 1 . (34) = \tau\rho^T p^n-\tau\rho^T\operatorname{diag}(m^n+h^n)y^{n+1}. \tag*{(34)} = τ ρ T p n − τ ρ T diag ( m n + h n ) y n + 1 . (34)
With A n A^n A n fixed at these rates, for every z ∈ R s d z \in\mathbb{R}^{sd} z ∈ R s d ,
ρ T ∣ ( A n ) − 1 z ∣ ≤ ρ T ∣ z ∣ . (35) \rho^T\lvert(A^n)^{-1}z\rvert\le\rho^T\lvert z\rvert. \tag*{(35)} ρ T ∣( A n ) − 1 z ∣ ≤ ρ T ∣ z ∣ . (35)
If ρ T p n ≤ σ M ( y n ) + b n \rho^T p^n \le\sigma\mathcal{M}(y^n)+b_n ρ T p n ≤ σ M ( y n ) + b n , with σ , b n ≥ 0 \sigma,b_n\ge0 σ , b n ≥ 0 , then, for t n = n τ t_n=n\tau t n = n τ ,
M ( y n ) ≤ e σ t n ( M ( y 0 ) + τ ∑ r = 0 n − 1 b r ) . \mathcal{M}(y^n)\le e^{\sigma t_n}\left(\mathcal{M}(y^0)+\tau\sum_{r=0}^{n-1}b_r\right). M ( y n ) ≤ e σ t n ( M ( y 0 ) + τ r = 0 ∑ n − 1 b r ) .
If p n = 0 p^n=0 p n = 0 at every step, then
τ ∑ n = 0 N − 1 ρ T diag ( h n ) y n + 1 ≤ M ( y 0 ) . \tau\sum_{n=0}^{N-1}\rho^T\operatorname{diag}(h^n)y^{n+1}\le\mathcal{M}(y^0). τ n = 0 ∑ N − 1 ρ T diag ( h n ) y n + 1 ≤ M ( y 0 ) .
The proof is in Appendix H.2. For Lotka–Volterra, s = 2 s=2 s = 2 , d = M + 1 d=M+1 d = M + 1 , c 1 = β c_1=\beta c 1 = β , c 2 = α c_2=\alpha c 2 = α , and k 12 , a n : = β v 1 , a n k_{12,a}^n:=\beta v_{1,a}^n k 12 , a n := β v 1 , a n and k 21 , a n : = 0 k_{21,a}^n:=0 k 21 , a n := 0 . We also set m n : = ( m 1 , 0 ) m^n:=(m_1,0) m n := ( m 1 , 0 ) , h n : = ( 0 , q u n ) h^n:=(0,qu^n) h n := ( 0 , q u n ) , and p n : = ( 0 , r v n ) p^n:=(0,rv^n) p n := ( 0 , r v n ) . These choices turn (33) into (6). An acyclic transfer graph can be ordered into a block triangular system. Cycles require a coupled solve.
Reward accounting
For economic weights ℓ ∈ R + s d \ell\in\mathbb{R}^{sd}_{+} ℓ ∈ R + s d and ζ ∈ { 0 , 1 } \zeta\in\{0,1\} ζ ∈ { 0 , 1 } , define
R ζ : V N ⟶ R \mathcal{R}_{\zeta}:\mathcal{V}^{N}\longrightarrow\mathbb{R} R ζ : V N ⟶ R
u ⟼ τ ∑ n = 0 N − 1 ℓ T diag ( h n ) ⋅ ( ( 1 − ζ ) y n + ζ y n + 1 ) . (36) u\longmapsto\tau\sum_{n=0}^{N-1}\ell^T\operatorname{diag}(h^n)\cdot\left((1-\zeta)y^n+\zeta y^{n+1}\right). \tag*{(36)} u ⟼ τ n = 0 ∑ N − 1 ℓ T diag ( h n ) ⋅ ( ( 1 − ζ ) y n + ζ y n + 1 ) . (36)
For ℓ = ρ \ell=\rho ℓ = ρ , R 1 \mathcal{R}_1 R 1 is the controlled mass loss in (34). Other weights assign different prices to those removals. For a common cost C : V N → R \mathcal{C}:\mathcal{V}^{N}\to\mathbb{R} C : V N → R , set
j ζ : V N ⟶ R j_{\zeta}:\mathcal{V}^{N}\longrightarrow\mathbb{R} j ζ : V N ⟶ R
u ⟼ C ( u ) − η R ζ ( u ) . u\longmapsto\mathcal{C}(u)-\eta\mathcal{R}_{\zeta}(u). u ⟼ C ( u ) − η R ζ ( u ) .
Then
j 0 ( u ) − j 1 ( u ) = η ( R 1 ( u ) − R 0 ( u ) ) = η τ ∑ n = 0 N − 1 ℓ T diag ( h n ) ( y n + 1 − y n ) . (37) \begin{aligned}
j_0(u)-j_1(u)=\eta\left(\mathcal{R}_1(u)-\mathcal{R}_0(u)\right) \\
&=\eta\tau\sum_{n=0}^{N-1}\ell^T\operatorname{diag}(h^n)(y^{n+1}-y^n). \tag*{(37)}
\end{aligned} j 0 ( u ) − j 1 ( u ) = η ( R 1 ( u ) − R 0 ( u ) ) = η τ n = 0 ∑ N − 1 ℓ T diag ( h n ) ( y n + 1 − y n ) . (37)
Production and incoming transfers can change its sign. For one donor, y , p , h ≥ 0 y,p,h\ge0 y , p , h ≥ 0 , y + = ( y + τ p ) / ( 1 + τ h ) y^+=(y+\tau p)/(1+\tau h) y + = ( y + τ p ) / ( 1 + τ h ) ,
τ h y − τ h y + = τ h ( y − y + τ p 1 + τ h ) = τ 2 h ( h y − p ) 1 + τ h . \tau hy-\tau hy^+=\tau h\left(y-\frac{y+\tau p}{1+\tau h}\right)=\frac{\tau^2h(hy-p)}{1+\tau h}. τ h y − τ h y + = τ h ( y − 1 + τ h y + τ p ) = 1 + τ h τ 2 h ( h y − p ) .
When p = 0 p=0 p = 0 , τ h y + = y − y + ≤ y \tau hy^+=y-y^+\le y τ h y + = y − y + ≤ y , whereas τ h y > y \tau hy>y τ h y > y if τ h > 1 \tau h>1 τ h > 1 and y > 0 y>0 y > 0 . For h > 0 h>0 h > 0 , this difference is negative when p > h y p>hy p > h y . We test three species in a directed cycle at a fixed control, with positive spatial diffusion and an initially empty receiver. For τ = 0.001 , 0.1 , 2 , 20 \tau=0.001,0.1,2,20 τ = 0.001 , 0.1 , 2 , 20 , the positivity, mass and resolvent contraction checks pass to roundoff, with mass errors below 1.8 × 10 − 15 1.8\times10^{-15} 1.8 × 1 0 − 15 . Without production, eight steps at τ = 20 \tau=20 τ = 20 harvest about 2.969 from an initial mass of 3.844. The old-state rule rewards 46.31. If the coefficients are C 2 C^2 C 2 on an open neighborhood of the computed state–control pairs, the fixed-grid state map is C 2 C^2 C 2 near the computed control. For a C 2 C^2 C 2 cost, Appendix H.3 gives the exact tangent, adjoint and Hessian formulas. Although the continuous problem has no terminal cost, the final discrete removal contributes to the terminal adjoint.
Relation to existing work
Patankar positivity and conservative transfers [8 , 16 ] extend to multistep methods [25 ] . Accuracy and oscillations are studied in [9 , 26 ] . Cacace, Oliviero and Pezzella [17 ] use dynamic programming for production–destruction control with a modified Patankar semi-Lagrangian Hamilton–Jacobi–Bellman approximation. Our two reward rules share a positive state solver, so removal quadrature accounts for their curvature difference. We obtain derivatives directly from the discrete problem [12 , 13 , 14 , 15 ] .
Conclusion
Harvest quadrature changes the cost and its curvature, even when the positive state scheme stays fixed. In the pure-removal model, the flux cost is strictly convex. The old-state rule can instead create negative curvature in alternating directions, favor extreme allocations and add a strict local minimum that is not global. Rational calculations certify the global solutions for two time steps.
For the positive scheme in one dimension, nonnegative bounded H 1 H^1 H 1 initial data and positive diffusivities give O ( τ ) O(\tau) O ( τ ) differences in costs, weighted gradients and Hessians, uniformly in the mesh. Coercivity in all directions on a common neighborhood gives the same order for the distance between controls. A large enough effort penalty makes each discrete minimizer unique and global.
The experiments trace the effects of old-state quadrature on coarse-grid curvature and on sensitivity to the initial guess and the direction of continuation. The measured control differences shrink on finer time grids, and a common reference with adaptive time stepping lets us compare their costs with small time-integration error. The same shared transfer fluxes extend positivity and exact harvest accounting to weighted production–destruction systems.
Acknowledgment
The author used AI tools to edit the language, organize the manuscript, and help write and review the code used in the numerical experiments. These tools also assisted with formalizing and checking proofs in Lean (Lean).
Continuous proofs
State bounds and well-posedness
We use ∥ ⋅ ∥ 2 \lVert\cdot\rVert_2 ∥ ⋅ ∥ 2 for the L 2 ( Ω ) L^2(\Omega) L 2 ( Ω ) norm and the Euclidean product norm for vectors. To truncate the reactions, fix R > 0 R>0 R > 0 and set
T R : R ⟶ [ 0 , R ] T_R:\mathbb{R}\longrightarrow[0,R] T R : R ⟶ [ 0 , R ]
s ⟼ min { R , max { 0 , s } } . s\longmapsto\min\{R,\max\{0,s\}\}. s ⟼ min { R , max { 0 , s }} .
The truncated reactions are
f R , 1 ( v , u ) = − m v 1 + α T R ( v 1 ) T R ( v 2 ) , f R , 2 ( v , u ) = ( r − q u ) v 2 − β T R ( v 1 ) T R ( v 2 ) . f_{R,1}(v,u)=-mv_1+\alpha T_R(v_1)T_R(v_2),\qquad f_{R,2}(v,u)=(r-qu)v_2-\beta T_R(v_1)T_R(v_2). f R , 1 ( v , u ) = − m v 1 + α T R ( v 1 ) T R ( v 2 ) , f R , 2 ( v , u ) = ( r − q u ) v 2 − β T R ( v 1 ) T R ( v 2 ) .
Since T R T_R T R is 1-Lipschitz and 0 ≤ T R ≤ R 0\leq T_R\leq R 0 ≤ T R ≤ R ,
∣ T R ( a 1 ) T R ( a 2 ) − T R ( b 1 ) T R ( b 2 ) ∣ ≤ R ∣ a 1 − b 1 ∣ + R ∣ a 2 − b 2 ∣ , \lvert T_R(a_1)T_R(a_2)-T_R(b_1)T_R(b_2)\rvert\leq R\lvert a_1-b_1\rvert+R\lvert a_2-b_2\rvert, ∣ T R ( a 1 ) T R ( a 2 ) − T R ( b 1 ) T R ( b 2 )∣ ≤ R ∣ a 1 − b 1 ∣ + R ∣ a 2 − b 2 ∣ ,
and
∣ f R ( a , u ) − f R ( b , u ) ∣ ≤ C R ∣ a − b ∣ , ∣ f R ( a , u ) ∣ ≤ C R ( 1 + ∣ a ∣ ) . \lvert f_R(a,u)-f_R(b,u)\rvert\leq C_R\lvert a-b\rvert,\qquad\lvert f_R(a,u)\rvert\leq C_R(1+\lvert a\rvert). ∣ f R ( a , u ) − f R ( b , u )∣ ≤ C R ∣ a − b ∣ , ∣ f R ( a , u )∣ ≤ C R ( 1 + ∣ a ∣) .
Here C R C_R C R is uniform for 0 ≤ u ≤ U max 0\leq u\leq U_{\max} 0 ≤ u ≤ U m a x . To construct the solution, take an L 2 ( Ω ) L^2(\Omega) L 2 ( Ω ) -orthonormal Neumann eigenbasis ( e ℓ ) ℓ ≥ 1 (e_\ell)_{\ell\geq1} ( e ℓ ) ℓ ≥ 1 and let P K P_K P K be the projection onto its first K K K elements. Write
v i , K ( t ) = ∑ ℓ = 1 K c i , ℓ ( t ) e ℓ , v i , K ( 0 ) = P K v i 0 . v_{i,K}(t)=\sum_{\ell=1}^{K}c_{i,\ell}(t)e_\ell,\qquad v_{i,K}(0)=P_Kv_i^0. v i , K ( t ) = ℓ = 1 ∑ K c i , ℓ ( t ) e ℓ , v i , K ( 0 ) = P K v i 0 .
The Galerkin equations, for 1 ≤ ℓ ≤ K 1\leq\ell\leq K 1 ≤ ℓ ≤ K , are
c i , ℓ ′ ( t ) + μ i ∑ k = 1 K c i , k ( t ) ( ∇ e k , ∇ e ℓ ) Ω = ( f R , i ( v K , u ) , e ℓ ) Ω . c'_{i,\ell}(t)+\mu_i\sum_{k=1}^{K}c_{i,k}(t)(\nabla e_k,\nabla e_\ell)_\Omega=(f_{R,i}(v_K,u),e_\ell)_\Omega. c i , ℓ ′ ( t ) + μ i k = 1 ∑ K c i , k ( t ) ( ∇ e k , ∇ e ℓ ) Ω = ( f R , i ( v K , u ) , e ℓ ) Ω .
Their coefficients are measurable in time and globally Lipschitz in the coefficient vector. Multiplying by c i , ℓ c_{i,\ell} c i , ℓ and summing gives
1 2 d d t ∑ i = 1 2 ∥ v i , K ∥ 2 2 + ∑ i = 1 2 μ i ∥ ∇ v i , K ∥ 2 2 = ∑ i = 1 2 ( f R , i ( v K , u ) , v i , K ) Ω ≤ C R ( 1 + ∑ i = 1 2 ∥ v i , K ∥ 2 2 ) . \frac{1}{2}\frac{\mathrm{d}}{\mathrm{d}t}\sum_{i=1}^{2}\lVert v_{i,K}\rVert_2^2+\sum_{i=1}^{2}\mu_i\lVert\nabla v_{i,K}\rVert_2^2=\sum_{i=1}^{2}(f_{R,i}(v_K,u),v_{i,K})_\Omega\leq C_R\left(1+\sum_{i=1}^{2}\lVert v_{i,K}\rVert_2^2\right). 2 1 d t d i = 1 ∑ 2 ∥ v i , K ∥ 2 2 + i = 1 ∑ 2 μ i ∥ ∇ v i , K ∥ 2 2 = i = 1 ∑ 2 ( f R , i ( v K , u ) , v i , K ) Ω ≤ C R ( 1 + i = 1 ∑ 2 ∥ v i , K ∥ 2 2 ) .
Gronwall and integration in time yield bounds independent of K K K :
sup 0 ≤ t ≤ T ∑ i = 1 2 ∥ v i , K ( t ) ∥ 2 2 + ∑ i = 1 2 ∫ 0 T ∥ ∇ v i , K ∥ 2 2 d t ≤ C R . \sup_{0\leq t\leq T}\sum_{i=1}^{2}\lVert v_{i,K}(t)\rVert_2^2+\sum_{i=1}^{2}\int_0^T\lVert\nabla v_{i,K}\rVert_2^2\,\mathrm{d}t\leq C_R. 0 ≤ t ≤ T sup i = 1 ∑ 2 ∥ v i , K ( t ) ∥ 2 2 + i = 1 ∑ 2 ∫ 0 T ∥ ∇ v i , K ∥ 2 2 d t ≤ C R .
The projections P K P_K P K are bounded in H 1 ( Ω ) H^1(\Omega) H 1 ( Ω ) . For ϕ ∈ H 1 ( Ω ) \phi\in H^1(\Omega) ϕ ∈ H 1 ( Ω ) , the Galerkin equation therefore gives
∣ ⟨ ∂ t v i , K , ϕ ⟩ ∣ ≤ ( μ i ∥ ∇ v i , K ∥ 2 + ∥ f R , i ( v K , u ) ∥ 2 ) ∥ ϕ ∥ H 1 , \lvert\langle\partial_t v_{i,K},\phi\rangle\rvert\leq\left(\mu_i\lVert\nabla v_{i,K}\rVert_2+\lVert f_{R,i}(v_K,u)\rVert_2\right)\lVert\phi\rVert_{H^1}, ∣⟨ ∂ t v i , K , ϕ ⟩∣ ≤ ( μ i ∥ ∇ v i , K ∥ 2 + ∥ f R , i ( v K , u ) ∥ 2 ) ∥ ϕ ∥ H 1 ,
and hence ∥ ∂ t v i , K ∥ L 2 ( 0 , T ; H 1 ∗ ) ≤ C R \lVert\partial_t v_{i,K}\rVert_{L^2(0,T;H^{1*})}\leq C_R ∥ ∂ t v i , K ∥ L 2 ( 0 , T ; H 1 ∗ ) ≤ C R . Compactness gives a subsequence with
v K → K → ∞ L 2 ( Q ) 2 v , ∥ f R ( v K , u ) − f R ( v , u ) ∥ L 2 ( Q ) 2 2 ≤ C R ∥ v K − v ∥ L 2 ( Q ) 2 2 → K → ∞ 0. v_K\xrightarrow[K\to\infty]{L^2(Q)^2}v,\qquad\lVert f_R(v_K,u)-f_R(v,u)\rVert_{L^2(Q)^2}^2\leq C_R\lVert v_K-v\rVert_{L^2(Q)^2}^2\xrightarrow[K\to\infty]{}0. v K L 2 ( Q ) 2 K → ∞ v , ∥ f R ( v K , u ) − f R ( v , u ) ∥ L 2 ( Q ) 2 2 ≤ C R ∥ v K − v ∥ L 2 ( Q ) 2 2 K → ∞ 0.
Passing to the limit in the time-integrated Galerkin equations gives the weak Neumann problem and the initial trace v i ( 0 ) = v i 0 v_i(0)=v_i^0 v i ( 0 ) = v i 0 [27 , 28 ] . Set s − : = max { − s , 0 } s^-:=\max\{-s,0\} s − := max { − s , 0 } and test component i i i with − v i − ∈ H 1 ( Ω ) -v_i^-\in H^1(\Omega) − v i − ∈ H 1 ( Ω ) . The interaction terms vanish because T R ( v i ) = 0 T_R(v_i)=0 T R ( v i ) = 0 where v i < 0 v_i<0 v i < 0 . For almost every t ∈ ( 0 , T ) t\in(0,T) t ∈ ( 0 , T ) ,
1 2 d d t ∑ i = 1 2 ∥ v i − ∥ 2 2 + ∑ i = 1 2 μ i ∥ ∇ v i − ∥ 2 2 ≤ C ∑ i = 1 2 ∥ v i − ∥ 2 2 . \frac{1}{2}\frac{\mathrm{d}}{\mathrm{d}t}\sum_{i=1}^{2}\lVert v_i^-\rVert_2^2+\sum_{i=1}^{2}\mu_i\lVert\nabla v_i^-\rVert_2^2\leq C\sum_{i=1}^{2}\lVert v_i^-\rVert_2^2. 2 1 d t d i = 1 ∑ 2 ∥ v i − ∥ 2 2 + i = 1 ∑ 2 μ i ∥ ∇ v i − ∥ 2 2 ≤ C i = 1 ∑ 2 ∥ v i − ∥ 2 2 .
The two reaction terms on the negative sets are
∫ Ω f R , 1 ( v , u ) ( − v 1 − ) = − m ∥ v 1 − ∥ 2 2 , ∫ Ω f R , 2 ( v , u ) ( − v 2 − ) = ∫ Ω ( r − q u ) ( v 2 − ) 2 ≤ r ∥ v 2 − ∥ 2 2 . \int_{\Omega} f_{R,1}(v,u)(-v_1^{-})=-m\lVert v_1^{-}\rVert_2^2,\qquad\int_{\Omega} f_{R,2}(v,u)(-v_2^{-})=\int_{\Omega}(r-qu)(v_2^{-})^2\le r\lVert v_2^{-}\rVert_2^2. ∫ Ω f R , 1 ( v , u ) ( − v 1 − ) = − m ∥ v 1 − ∥ 2 2 , ∫ Ω f R , 2 ( v , u ) ( − v 2 − ) = ∫ Ω ( r − q u ) ( v 2 − ) 2 ≤ r ∥ v 2 − ∥ 2 2 .
Thus
∑ i = 1 2 ∥ v i ( t ) ∥ 2 2 ≤ e 2 r t ∑ i = 1 2 ∥ ( v i 0 ) − ∥ 2 2 = 0. \sum_{i=1}^{2}\lVert v_i(t)\rVert_2^2\le e^{2rt}\sum_{i=1}^{2}\lVert(v_i^0)^{-}\rVert_2^2=0. i = 1 ∑ 2 ∥ v i ( t ) ∥ 2 2 ≤ e 2 r t i = 1 ∑ 2 ∥( v i 0 ) − ∥ 2 2 = 0.
Since 0 ≤ T R ( v i ) ≤ v i 0\le T_R(v_i)\le v_i 0 ≤ T R ( v i ) ≤ v i , in Ω × ( 0 , T ) \Omega\times(0,T) Ω × ( 0 , T ) we have
∂ t v 2 − μ 2 Δ v 2 ≤ r v 2 . \partial_t v_2-\mu_2\Delta v_2\le rv_2. ∂ t v 2 − μ 2 Δ v 2 ≤ r v 2 .
For the comparison, set b 2 ( t ) = V 2 0 e r t b_2(t)=V_2^0e^{rt} b 2 ( t ) = V 2 0 e r t and test the equation for v 2 − b 2 v_2-b_2 v 2 − b 2 with ( v 2 − b 2 ) + (v_2-b_2)^+ ( v 2 − b 2 ) + . Since b 2 ′ = r b 2 b_2'=rb_2 b 2 ′ = r b 2 and ( v 2 0 − b 2 ( 0 ) ) + = 0 (v_2^0-b_2(0))^+=0 ( v 2 0 − b 2 ( 0 ) ) + = 0 ,
1 2 d d t ∥ ( v 2 − b 2 ) + ∥ 2 2 + μ 2 ∥ ∇ ( v 2 − b 2 ) + ∥ 2 2 ≤ r ∥ ( v 2 − b 2 ) + ∥ 2 2 . \frac{1}{2}\frac{\mathrm{d}}{\mathrm{d}t}\lVert(v_2-b_2)^+\rVert_2^2+\mu_2\lVert\nabla(v_2-b_2)^+\rVert_2^2\le r\lVert(v_2-b_2)^+\rVert_2^2. 2 1 d t d ∥( v 2 − b 2 ) + ∥ 2 2 + μ 2 ∥ ∇ ( v 2 − b 2 ) + ∥ 2 2 ≤ r ∥( v 2 − b 2 ) + ∥ 2 2 .
Gronwall gives, for almost every ( x , t ) ∈ Ω × ( 0 , T ) (x,t)\in\Omega\times(0,T) ( x , t ) ∈ Ω × ( 0 , T ) ,
v 2 ( x , t ) ≤ V 2 0 e r t . v_2(x,t)\le V_2^0e^{rt}. v 2 ( x , t ) ≤ V 2 0 e r t .
In Ω × ( 0 , T ) \Omega\times(0,T) Ω × ( 0 , T ) , we then have
∂ t v 1 − μ 1 Δ v 1 ≤ ( − m + α V 2 0 e r t ) v 1 . \partial_t v_1-\mu_1\Delta v_1\le(-m+\alpha V_2^0e^{rt})v_1. ∂ t v 1 − μ 1 Δ v 1 ≤ ( − m + α V 2 0 e r t ) v 1 .
Compare v 1 v_1 v 1 with the solution of the ODE on ( 0 , T ) (0,T) ( 0 , T ) , with initial value V 1 0 V_1^0 V 1 0 :
y ′ ( t ) = ( − m + α V 2 0 e r t ) y ( t ) , y ( 0 ) = V 1 0 . y'(t)=(-m+\alpha V_2^0e^{rt})y(t),\qquad y(0)=V_1^0. y ′ ( t ) = ( − m + α V 2 0 e r t ) y ( t ) , y ( 0 ) = V 1 0 .
Its solution is
y ( t ) = V 1 0 exp ( ∫ 0 t [ − m + α V 2 0 e r s ] d s ) = V 1 0 exp ( − m t + α V 2 0 r ( e r t − 1 ) ) . y(t)=V_1^0\exp\left(\int_0^t[-m+\alpha V_2^0e^{rs}]\,\mathrm{d}s\right)=V_1^0\exp\left(-mt+\frac{\alpha V_2^0}{r}(e^{rt}-1)\right). y ( t ) = V 1 0 exp ( ∫ 0 t [ − m + α V 2 0 e r s ] d s ) = V 1 0 exp ( − m t + r α V 2 0 ( e r t − 1 ) ) .
Testing the inequality for v 1 − y v_1-y v 1 − y with ( v 1 − y ) + (v_1-y)^+ ( v 1 − y ) + gives
1 2 d d t ∥ ( v 1 − y ) + ∥ 2 2 + μ 1 ∥ ∇ ( v 1 − y ) + ∥ 2 2 ≤ ( − m + α V 2 0 e r t ) ∥ ( v 1 − y ) + ∥ 2 2 . \frac{1}{2}\frac{\mathrm{d}}{\mathrm{d}t}\lVert(v_1-y)^+\rVert_2^2+\mu_1\lVert\nabla(v_1-y)^+\rVert_2^2\le(-m+\alpha V_2^0e^{rt})\lVert(v_1-y)^+\rVert_2^2. 2 1 d t d ∥( v 1 − y ) + ∥ 2 2 + μ 1 ∥ ∇ ( v 1 − y ) + ∥ 2 2 ≤ ( − m + α V 2 0 e r t ) ∥( v 1 − y ) + ∥ 2 2 .
The initial positive part is zero, so v 1 ≤ y v_1\le y v 1 ≤ y . Once R R R exceeds both bounds on [ 0 , T ] [0,T] [ 0 , T ] , the truncation disappears: the solution satisfies (1) . To prove uniqueness, take two solutions v , v ~ ∈ Y 2 v,\tilde v\in\mathcal{Y}^2 v , v ~ ∈ Y 2 and set e : = v − v ~ ∈ W 2 e:=v-\tilde v\in\mathcal{W}^2 e := v − v ~ ∈ W 2 . Then
v 1 v 2 − v ~ 1 v ~ 2 = e 1 v 2 + v ~ 1 e 2 . (38) v_1v_2-\tilde v_1\tilde v_2=e_1v_2+\tilde v_1e_2. \tag*{(38)} v 1 v 2 − v ~ 1 v ~ 2 = e 1 v 2 + v ~ 1 e 2 . (38)
The difference equations are
{ ∂ t e 1 − μ 1 Δ e 1 = ( − m + α v 2 ) e 1 + α v ~ 1 e 2 , ∂ t e 2 − μ 2 Δ e 2 = − β v 2 e 1 + ( r − q u − β v ~ 1 ) e 2 . \begin{cases}
\partial_t e_1-\mu_1\Delta e_1=(-m+\alpha v_2)e_1+\alpha\tilde v_1e_2,\\
\partial_t e_2-\mu_2\Delta e_2=-\beta v_2e_1+(r-qu-\beta\tilde v_1)e_2.
\end{cases} { ∂ t e 1 − μ 1 Δ e 1 = ( − m + α v 2 ) e 1 + α v ~ 1 e 2 , ∂ t e 2 − μ 2 Δ e 2 = − β v 2 e 1 + ( r − q u − β v ~ 1 ) e 2 .
All coefficients are bounded. In particular,
∣ ∫ Ω ( α v ~ 1 − β v 2 ) e 1 e 2 ∣ ≤ 1 2 ∥ α v ~ 1 − β v 2 ∥ ∞ ( ∥ e 1 ∥ 2 2 + ∥ e 2 ∥ 2 2 ) , \left|\int_{\Omega}(\alpha\tilde v_1-\beta v_2)e_1e_2\right|\le\frac{1}{2}\lVert\alpha\tilde v_1-\beta v_2\rVert_{\infty}(\lVert e_1\rVert_2^2+\lVert e_2\rVert_2^2), ∫ Ω ( α v ~ 1 − β v 2 ) e 1 e 2 ≤ 2 1 ∥ α v ~ 1 − β v 2 ∥ ∞ (∥ e 1 ∥ 2 2 + ∥ e 2 ∥ 2 2 ) ,
so ∣ f ( v , u ) − f ( v ~ , u ) ∣ ≤ C ∣ e ∣ |f(v,u)-f(\tilde v,u)|\le C|e| ∣ f ( v , u ) − f ( v ~ , u ) ∣ ≤ C ∣ e ∣ . For almost every t ∈ ( 0 , T ) t\in(0,T) t ∈ ( 0 , T ) , testing the difference equation with e e e yields
1 2 d d t ∥ e ∥ 2 2 + ∑ i = 1 2 μ i ∥ ∇ e i ∥ 2 2 ≤ C ∥ e ∥ 2 2 . \frac{1}{2}\frac{\mathrm{d}}{\mathrm{d}t}\lVert e\rVert_2^2+\sum_{i=1}^{2}\mu_i\lVert\nabla e_i\rVert_2^2\le C\lVert e\rVert_2^2. 2 1 d t d ∥ e ∥ 2 2 + i = 1 ∑ 2 μ i ∥ ∇ e i ∥ 2 2 ≤ C ∥ e ∥ 2 2 .
Since e ( 0 ) = 0 e(0)=0 e ( 0 ) = 0 , Gronwall gives e = 0 e=0 e = 0 . Testing with ϕ = 1 ∈ H 1 ( Ω ) \phi=1\in H^{1}(\Omega) ϕ = 1 ∈ H 1 ( Ω ) eliminates diffusion. For almost every t ∈ ( 0 , T ) t\in(0,T) t ∈ ( 0 , T ) ,
M ′ ( 2 ) = β ∫ Ω ( − m v 1 + α v 1 v 2 ) + α ∫ Ω ( r v 2 − β v 1 v 2 − q u v 2 ) = − β m ∫ Ω v 1 ⏟ ≤ 0 + α r ∫ Ω v 2 ⏟ ≤ r M + ( − α ) q ∫ Ω u v 2 ⏟ ≤ 0 ≤ r M . \mathcal{M}'^{(2)}=\beta\int_{\Omega}(-mv_{1}+\alpha v_{1}v_{2})+\alpha\int_{\Omega}(rv_{2}-\beta v_{1}v_{2}-quv_{2})=-\beta m\underbrace{\int_{\Omega}v_{1}}_{\leq0}+\alpha r\underbrace{\int_{\Omega}v_{2}}_{\leq r\mathcal{M}}+(-\alpha)q\underbrace{\int_{\Omega}uv_{2}}_{\leq0}\leq r\mathcal{M}. M ′ ( 2 ) = β ∫ Ω ( − m v 1 + α v 1 v 2 ) + α ∫ Ω ( r v 2 − β v 1 v 2 − q u v 2 ) = − β m ≤ 0 ∫ Ω v 1 + α r ≤ r M ∫ Ω v 2 + ( − α ) q ≤ 0 ∫ Ω u v 2 ≤ r M .
Equivalently,
d d t ( e − r t M ( t ) ) = − e − r t [ β ( m + r ) ∫ Ω v 1 + α q ∫ Ω u v 2 ] ≤ 0. \frac{\mathrm{d}}{\mathrm{d}t}\left(e^{-rt}\mathcal{M}(t)\right)=-e^{-rt}\left[\beta(m+r)\int_{\Omega}v_{1}+\alpha q\int_{\Omega}uv_{2}\right]\leq0. d t d ( e − r t M ( t ) ) = − e − r t [ β ( m + r ) ∫ Ω v 1 + α q ∫ Ω u v 2 ] ≤ 0.
Integration from 0 0 0 to t t t gives M ( t ) ≤ e r t M ( 0 ) \mathcal{M}(t)\leq e^{rt}\mathcal{M}(0) M ( t ) ≤ e r t M ( 0 ) . For μ 1 = μ 2 = μ \mu_{1}=\mu_{2}=\mu μ 1 = μ 2 = μ , set w : = β v 1 + α v 2 ∈ Y w:=\beta v_{1}+\alpha v_{2}\in\mathcal{Y} w := β v 1 + α v 2 ∈ Y . In Ω × ( 0 , T ) \Omega\times(0,T) Ω × ( 0 , T ) ,
∂ t w − μ Δ w = − β m v 1 + α r v 2 − α q u v 2 ≤ r w , \partial_{t}w-\mu\Delta w=-\beta mv_{1}+\alpha rv_{2}-\alpha quv_{2}\leq rw, ∂ t w − μ Δ w = − β m v 1 + α r v 2 − α q u v 2 ≤ r w ,
and comparison with e r t ∥ w ( 0 ) ∥ ∞ e^{rt}\|w(0)\|_{\infty} e r t ∥ w ( 0 ) ∥ ∞ proves (3) .
Existence of a continuous minimizer
The bound j ( u ) ≥ − η q U max K 2 ∣ Q ∣ j(u)\geq-\eta qU_{\max}K_{2}|Q| j ( u ) ≥ − η q U m a x K 2 ∣ Q ∣ , with K 2 : = V 2 0 e r T K_{2}:=V_{2}^{0}e^{rT} K 2 := V 2 0 e r T , gives a finite infimum. Take a minimizing sequence ( u k ) ⊂ U a d (u_{k})\subset\mathcal{U}_{\mathrm{ad}} ( u k ) ⊂ U ad . Because Q Q Q has finite measure and L 1 ( Q ) L^{1}(Q) L 1 ( Q ) is separable, we can extract a subsequence such that, in L ∞ ( Q ) L^{\infty}(Q) L ∞ ( Q ) , u k ⇀ ∗ u ∗ u_{k}\overset{*}{\rightharpoonup}u^{*} u k ⇀ ∗ u ∗ .
For 0 ≤ ϕ ∈ L 1 ( Q ) 0\leq\phi\in L^{1}(Q) 0 ≤ ϕ ∈ L 1 ( Q ) ,
0 ≤ ∫ Q u ∗ ϕ = lim k ∫ Q u k ϕ ≤ U max ∫ Q ϕ , 0\leq\int_{Q}u^{*}\phi=\lim_{k}\int_{Q}u_{k}\phi\leq U_{\max}\int_{Q}\phi, 0 ≤ ∫ Q u ∗ ϕ = k lim ∫ Q u k ϕ ≤ U m a x ∫ Q ϕ ,
so u ∗ ∈ U a d u^{*}\in\mathcal{U}_{\mathrm{ad}} u ∗ ∈ U ad . Set v k : = S ( u k ) ∈ Y 2 v_{k}:=S(u_{k})\in\mathcal{Y}^{2} v k := S ( u k ) ∈ Y 2 , K 1 : = V 1 0 max 0 ≤ t ≤ T exp ( − m t + α V 2 0 r ( e r t − 1 ) ) K_{1}:=V_{1}^{0}\max_{0\leq t\leq T}\exp\left(-mt+\frac{\alpha V_{2}^{0}}{r}(e^{rt}-1)\right) K 1 := V 1 0 max 0 ≤ t ≤ T exp ( − m t + r α V 2 0 ( e r t − 1 ) ) . Theorem 2.1 gives ∥ v i , k ∥ L ∞ ( Q ) ≤ K i \|v_{i,k}\|_{L^{\infty}(Q)}\leq K_{i} ∥ v i , k ∥ L ∞ ( Q ) ≤ K i uniformly in k k k . For almost every t ∈ ( 0 , T ) t\in(0,T) t ∈ ( 0 , T ) , testing by v i , k v_{i,k} v i , k gives
1 2 d d t ∥ v i , k ∥ L 2 ( Ω ) 2 + μ i ∥ ∇ v i , k ∥ L 2 ( Ω ) 2 = ( f i ( v k , u k ) , v i , k ) Ω . \frac{1}{2}\frac{\mathrm{d}}{\mathrm{d}t}\|v_{i,k}\|_{L^{2}(\Omega)}^{2}+\mu_{i}\|\nabla v_{i,k}\|_{L^{2}(\Omega)}^{2}=(f_{i}(v_{k},u_{k}),v_{i,k})_{\Omega}. 2 1 d t d ∥ v i , k ∥ L 2 ( Ω ) 2 + μ i ∥∇ v i , k ∥ L 2 ( Ω ) 2 = ( f i ( v k , u k ) , v i , k ) Ω .
The right side is uniformly bounded, so integration in time, together with
∥ ∂ t v i , k ∥ H 1 ∗ ≤ μ i ∥ ∇ v i , k ∥ L 2 ( Ω ) + ∥ f i ( v k , u k ) ∥ L 2 ( Ω ) \|\partial_{t}v_{i,k}\|_{H^{1*}}\leq\mu_{i}\|\nabla v_{i,k}\|_{L^{2}(\Omega)}+\|f_{i}(v_{k},u_{k})\|_{L^{2}(\Omega)} ∥ ∂ t v i , k ∥ H 1 ∗ ≤ μ i ∥∇ v i , k ∥ L 2 ( Ω ) + ∥ f i ( v k , u k ) ∥ L 2 ( Ω )
gives
∥ v i , k ∥ L 2 ( 0 , T ; H 1 ) + ∥ ∂ t v i , k ∥ L 2 ( 0 , T ; H 1 ∗ ) ≤ C . \|v_{i,k}\|_{L^{2}(0,T;H^{1})}+\|\partial_{t}v_{i,k}\|_{L^{2}(0,T;H^{1*})}\leq C. ∥ v i , k ∥ L 2 ( 0 , T ; H 1 ) + ∥ ∂ t v i , k ∥ L 2 ( 0 , T ; H 1 ∗ ) ≤ C .
The compact embedding H 1 ( Ω ) ↪ L 2 ( Ω ) H^{1}(\Omega)\hookrightarrow L^{2}(\Omega) H 1 ( Ω ) ↪ L 2 ( Ω ) and time compactness [19 ] give a subsequence converging almost everywhere in Q Q Q , with
v i , k ⇀ ( W k → ∞ v i ∗ , v i , k ⟶ ( L 2 ( Q ) k → ∞ v i ∗ . v_{i,k}\mathrel{\mathop{\rightharpoonup}^{\mathstrut W}}_{k\to\infty}v_{i}^{*},\qquad v_{i,k}\mathrel{\mathop{\longrightarrow}^{\mathstrut L^{2}(Q)}}_{k\to\infty}v_{i}^{*}. v i , k ⇀ ( W k → ∞ v i ∗ , v i , k ⟶ ( L 2 ( Q ) k → ∞ v i ∗ .
Thus 0 ≤ v i ∗ ≤ K i 0\leq v_{i}^{*}\leq K_{i} 0 ≤ v i ∗ ≤ K i and v ∗ ∈ Y 2 v^{*}\in\mathcal{Y}^{2} v ∗ ∈ Y 2 . Then
∥ v 1 , k v 2 , k − v 1 ∗ v 2 ∗ ∥ 2 ≤ K 1 ∥ v 2 , k − v 2 ∗ ∥ 2 + K 2 ∥ v 1 , k − v 1 ∗ ∥ 2 ⟶ k → ∞ 0. (39) \|v_{1,k}v_{2,k}-v_{1}^{*}v_{2}^{*}\|_{2}\leq K_{1}\|v_{2,k}-v_{2}^{*}\|_{2}+K_{2}\|v_{1,k}-v_{1}^{*}\|_{2}\longrightarrow_{k\to\infty}0. \tag*{(39)} ∥ v 1 , k v 2 , k − v 1 ∗ v 2 ∗ ∥ 2 ≤ K 1 ∥ v 2 , k − v 2 ∗ ∥ 2 + K 2 ∥ v 1 , k − v 1 ∗ ∥ 2 ⟶ k → ∞ 0. (39)
For ϕ ∈ L 2 ( Q ) \phi\in L^{2}(Q) ϕ ∈ L 2 ( Q ) ,
∫ Q ( u k v 2 , k − u ∗ v 2 ∗ ) ϕ = ∫ Q u k ( v 2 , k − v 2 ∗ ) ϕ + ∫ Q ( u k − u ∗ ) v 2 ∗ ϕ . \int_{Q}(u_{k}v_{2,k}-u^{*}v_{2}^{*})\phi=\int_{Q}u_{k}(v_{2,k}-v_{2}^{*})\phi+\int_{Q}(u_{k}-u^{*})v_{2}^{*}\phi. ∫ Q ( u k v 2 , k − u ∗ v 2 ∗ ) ϕ = ∫ Q u k ( v 2 , k − v 2 ∗ ) ϕ + ∫ Q ( u k − u ∗ ) v 2 ∗ ϕ .
Then
∣ ∫ Q u k ( v 2 , k − v 2 ∗ ) ϕ ∣ ≤ U max ∥ v 2 , k − v 2 ∗ ∥ 2 ∥ ϕ ∥ 2 ⟶ k → ∞ 0. \left|\int_{Q}u_{k}(v_{2,k}-v_{2}^{*})\phi\right|\leq U_{\max}\|v_{2,k}-v_{2}^{*}\|_{2}\|\phi\|_{2}\longrightarrow_{k\to\infty}0. ∫ Q u k ( v 2 , k − v 2 ∗ ) ϕ ≤ U m a x ∥ v 2 , k − v 2 ∗ ∥ 2 ∥ ϕ ∥ 2 ⟶ k → ∞ 0.
The second term tends to zero by weak-* convergence, since v 2 ∗ ϕ ∈ L 1 ( Q ) v_2^*\phi\in L^1(Q) v 2 ∗ ϕ ∈ L 1 ( Q ) . Hence u k v 2 , k ⟶ L 2 ( Q ) u ∗ v 2 ∗ u_kv_{2,k} \overset{L^2(Q)}{\longrightarrow} u^*v_2^* u k v 2 , k ⟶ L 2 ( Q ) u ∗ v 2 ∗ . Test the weak equations with ϕ ( x ) χ ( t ) \phi(x)\chi(t) ϕ ( x ) χ ( t ) , where ϕ ∈ H 1 ( Ω ) \phi\in H^1(\Omega) ϕ ∈ H 1 ( Ω ) , χ ∈ C 1 ( [ 0 , T ] ) \chi\in C^1([0,T]) χ ∈ C 1 ([ 0 , T ]) and χ ( T ) = 0 \chi(T)=0 χ ( T ) = 0 . We pass to the limit in every term, including the fixed initial trace. Uniqueness identifies v ∗ = S ( u ∗ ) v^*=S(u^*) v ∗ = S ( u ∗ ) . For each tracking term, strong convergence gives
∣ ∥ v i , k − v i d ∥ 2 2 − ∥ v i ∗ − v i d ∥ 2 2 ∣ = ∣ ∫ Q ( v i , k − v i ∗ ) ( v i , k + v i ∗ − 2 v i d ) ∣ ≤ ∥ v i , k − v i ∗ ∥ 2 ( ∥ v i , k − v i d ∥ 2 + ∥ v i ∗ − v i d ∥ 2 ) → k → ∞ 0. \left|\|v_{i,k}-v_i^d\|_2^2-\|v_i^*-v_i^d\|_2^2\right|=\left|\int_Q(v_{i,k}-v_i^*)(v_{i,k}+v_i^*-2v_i^d)\right|\leq\|v_{i,k}-v_i^*\|_2\left(\|v_{i,k}-v_i^d\|_2+\|v_i^*-v_i^d\|_2\right)\xrightarrow{k\to\infty}0. ∥ v i , k − v i d ∥ 2 2 − ∥ v i ∗ − v i d ∥ 2 2 = ∫ Q ( v i , k − v i ∗ ) ( v i , k + v i ∗ − 2 v i d ) ≤ ∥ v i , k − v i ∗ ∥ 2 ( ∥ v i , k − v i d ∥ 2 + ∥ v i ∗ − v i d ∥ 2 ) k → ∞ 0.
For the harvest term,
∣ ∫ Q u k v 2 , k − ∫ Q u ∗ v 2 ∗ ∣ ≤ U max ∣ Q ∣ 1 / 2 ∥ v 2 , k − v 2 ∗ ∥ 2 + ∣ ∫ Q ( u k − u ∗ ) v 2 ∗ ∣ ⟶ k → ∞ 0. \left|\int_Q u_kv_{2,k}-\int_Q u^*v_2^*\right|\leq U_{\max}|Q|^{1/2}\|v_{2,k}-v_2^*\|_2+\left|\int_Q(u_k-u^*)v_2^*\right|\longrightarrow_{k\to\infty}0. ∫ Q u k v 2 , k − ∫ Q u ∗ v 2 ∗ ≤ U m a x ∣ Q ∣ 1/2 ∥ v 2 , k − v 2 ∗ ∥ 2 + ∫ Q ( u k − u ∗ ) v 2 ∗ ⟶ k → ∞ 0.
For the effort cost, weak convergence u k ⇀ L 2 ( Q ) u ∗ u_k\overset{L^2(Q)}{\rightharpoonup}u^* u k ⇀ L 2 ( Q ) u ∗ gives
∥ u ∗ ∥ 2 2 ≤ lim inf k ∥ u k ∥ 2 2 ⟹ j ( u ∗ ) ≤ lim inf k j ( u k ) . \|u^*\|_2^2\leq\liminf_k\|u_k\|_2^2\quad\Longrightarrow\quad j(u^*)\leq\liminf_k j(u_k). ∥ u ∗ ∥ 2 2 ≤ k lim inf ∥ u k ∥ 2 2 ⟹ j ( u ∗ ) ≤ k lim inf j ( u k ) .
Thus u ∗ u^* u ∗ attains the minimum.
For u , ξ ∈ L ∞ ( Q ) u,\xi\in L^\infty(Q) u , ξ ∈ L ∞ ( Q ) , the tangent z ξ = S ′ ( u ) ξ z^\xi=S'(u)\xi z ξ = S ′ ( u ) ξ satisfies the following system. The evolution equations hold in Q Q Q , the Neumann conditions on ∂ Ω × ( 0 , T ) \partial\Omega\times(0,T) ∂ Ω × ( 0 , T ) , and the initial conditions in Ω \Omega Ω for i = 1 , 2 i=1,2 i = 1 , 2 :
{ ∂ t z 1 − μ 1 Δ z 1 = ( − m + α v 2 ) z 1 + α v 1 z 2 , ∂ t z 2 − μ 2 Δ z 2 = − β v 2 z 1 + ( r − β v 1 − q u ) z 2 − q v 2 ξ , ∂ n z i = 0 , z i ( 0 ) = 0. \begin{cases}
\partial_tz_1-\mu_1\Delta z_1=(-m+\alpha v_2)z_1+\alpha v_1z_2,\\
\partial_tz_2-\mu_2\Delta z_2=-\beta v_2z_1+(r-\beta v_1-qu)z_2-qv_2\xi,\\
\partial_nz_i=0,\qquad z_i(0)=0.
\end{cases} ⎩ ⎨ ⎧ ∂ t z 1 − μ 1 Δ z 1 = ( − m + α v 2 ) z 1 + α v 1 z 2 , ∂ t z 2 − μ 2 Δ z 2 = − β v 2 z 1 + ( r − β v 1 − q u ) z 2 − q v 2 ξ , ∂ n z i = 0 , z i ( 0 ) = 0.
The adjoint p : = ( p 1 , p 2 ) ∈ W 2 p:=(p_1,p_2)\in\mathcal{W}^2 p := ( p 1 , p 2 ) ∈ W 2 satisfies the following system. The evolution equations hold in Q Q Q , the Neumann conditions on ∂ Ω × ( 0 , T ) \partial\Omega\times(0,T) ∂ Ω × ( 0 , T ) , and the terminal conditions in Ω \Omega Ω for i = 1 , 2 i=1,2 i = 1 , 2 :
{ − ∂ t p 1 − μ 1 Δ p 1 = γ 1 ( v 1 − v 1 d ) + ( − m + α v 2 ) p 1 − β v 2 p 2 , − ∂ t p 2 − μ 2 Δ p 2 = γ 2 ( v 2 − v 2 d ) − η q u + α v 1 p 1 + ( r − β v 1 − q u ) p 2 , ∂ n p i = 0 , p i ( T ) = 0. (40) \begin{cases}
-\partial_tp_1-\mu_1\Delta p_1=\gamma_1(v_1-v_1^d)+(-m+\alpha v_2)p_1-\beta v_2p_2,\\
-\partial_tp_2-\mu_2\Delta p_2=\gamma_2(v_2-v_2^d)-\eta qu+\alpha v_1p_1+(r-\beta v_1-qu)p_2,\\
\partial_np_i=0,\qquad p_i(T)=0. \tag*{(40)}
\end{cases} ⎩ ⎨ ⎧ − ∂ t p 1 − μ 1 Δ p 1 = γ 1 ( v 1 − v 1 d ) + ( − m + α v 2 ) p 1 − β v 2 p 2 , − ∂ t p 2 − μ 2 Δ p 2 = γ 2 ( v 2 − v 2 d ) − η q u + α v 1 p 1 + ( r − β v 1 − q u ) p 2 , ∂ n p i = 0 , p i ( T ) = 0. (40)
For z ξ : = S ′ ( u ) ξ z^\xi:=S'(u)\xi z ξ := S ′ ( u ) ξ , z ζ : = S ′ ( u ) ζ z^\zeta:=S'(u)\zeta z ζ := S ′ ( u ) ζ , with ξ , ζ ∈ L ∞ ( Q ) \xi,\zeta\in L^\infty(Q) ξ , ζ ∈ L ∞ ( Q ) ,
j ′ ′ ( u ) : L ∞ ( Q ) × L ∞ ( Q ) ⟶ R j''(u):L^\infty(Q)\times L^\infty(Q)\longrightarrow\mathbb{R} j ′′ ( u ) : L ∞ ( Q ) × L ∞ ( Q ) ⟶ R
( ξ , ζ ) ⟼ λ ∫ Q ξ ζ + ∑ i = 1 2 γ i ∫ Q z i ξ z i ζ + ∫ Q ( α p 1 − β p 2 ) ⋅ ( z 1 ξ z 2 ζ + z 1 ζ z 2 ξ ) − q ∫ Q ( η + p 2 ) ( ξ z 2 ζ + ζ z 2 ξ ) . (41) (\xi,\zeta)\longmapsto\lambda\int_Q\xi\zeta+\sum_{i=1}^{2}\gamma_i\int_Q z_i^\xi z_i^\zeta+\int_Q(\alpha p_1-\beta p_2)\cdot(z_1^\xi z_2^\zeta+z_1^\zeta z_2^\xi)-q\int_Q(\eta+p_2)(\xi z_2^\zeta+\zeta z_2^\xi). \tag*{(41)} ( ξ , ζ ) ⟼ λ ∫ Q ξ ζ + i = 1 ∑ 2 γ i ∫ Q z i ξ z i ζ + ∫ Q ( α p 1 − β p 2 ) ⋅ ( z 1 ξ z 2 ζ + z 1 ζ z 2 ξ ) − q ∫ Q ( η + p 2 ) ( ξ z 2 ζ + ζ z 2 ξ ) . (41)
Differentiability and second variation
Fix u ∈ L ∞ ( Q ) u\in L^\infty(Q) u ∈ L ∞ ( Q ) and R > ∥ u ∥ ∞ R>\|u\|_\infty R > ∥ u ∥ ∞ , and set v = S ( u ) v=S(u) v = S ( u ) . Consider increments ζ ∈ L ∞ ( Q ) \zeta\in L^\infty(Q) ζ ∈ L ∞ ( Q ) such that ∥ u + ζ ∥ ∞ ≤ R \|u+\zeta\|_\infty\leq R ∥ u + ζ ∥ ∞ ≤ R . Nonnegativity holds for signed controls too. Replacing r r r by r + q R r+qR r + q R gives the prey bound. Thus both the states and the reaction Jacobian
A ( v , u ) : = ( − m + α v 2 α v 1 − β v 2 r − β v 1 − q u ) ∈ L ∞ ( Q ; R 2 × 2 ) A(v,u):=\begin{pmatrix}
-m+\alpha v_2 & \alpha v_1\\
-\beta v_2 & r-\beta v_1-qu
\end{pmatrix}\in L^\infty(Q;\mathbb{R}^{2\times2}) A ( v , u ) := ( − m + α v 2 − β v 2 α v 1 r − β v 1 − q u ) ∈ L ∞ ( Q ; R 2 × 2 )
are uniformly bounded. Define
L v : Y 2 ⟶ L 2 ( 0 , T ; H 1 ( Ω ) ∗ ) 2 L_v:\mathcal{Y}^2\longrightarrow L^2(0,T;H^1(\Omega)^*)^2 L v : Y 2 ⟶ L 2 ( 0 , T ; H 1 ( Ω ) ∗ ) 2
y ⟼ ∂ t y − diag ( μ 1 , μ 2 ) Δ y − A ( v , u ) y y\longmapsto\partial_ty-\operatorname{diag}(\mu_1,\mu_2)\Delta y-A(v,u)y y ⟼ ∂ t y − diag ( μ 1 , μ 2 ) Δ y − A ( v , u ) y
with weak Neumann conditions. Write ∥ y ( t ) ∥ p : = ∥ y ( t ) ∥ L p ( Ω ) 2 \|y(t)\|_p := \|y(t)\|_{L^p(\Omega)^2} ∥ y ( t ) ∥ p := ∥ y ( t ) ∥ L p ( Ω ) 2 , and likewise for F F F . If L v y = F ∈ L ∞ ( Q ) 2 \mathcal{L}_v y = F \in L^\infty(Q)^2 L v y = F ∈ L ∞ ( Q ) 2 and y ( 0 ) = 0 y(0)=0 y ( 0 ) = 0 , variation of constants for the Neumann heat semigroups gives
∥ y ( t ) ∥ ∞ ≤ C ∫ 0 t ( ∥ y ( s ) ∥ ∞ + ∥ F ( s ) ∥ ∞ ) d s . \|y(t)\|_\infty\le C\int_0^t\left(\|y(s)\|_\infty+\|F(s)\|_\infty\right)\,\mathrm{d}s. ∥ y ( t ) ∥ ∞ ≤ C ∫ 0 t ( ∥ y ( s ) ∥ ∞ + ∥ F ( s ) ∥ ∞ ) d s .
Let C A C_A C A bound max i , j = 1 2 ∣ A i j ∣ \max_{i,j=1}^{2}|A_{ij}| max i , j = 1 2 ∣ A ij ∣ on this control ball. The componentwise Neumann heat semigroup is contractive in the maximum norm. Its variation-of-constants formula and Gronwall give
max i ∥ y i ( t ) ∥ L ∞ ( Ω ) ≤ ∫ 0 t e C A ( t − s ) max i ∥ F i ( s ) ∥ L ∞ ( Ω ) d s ≤ T e C A T max i ∥ F i ∥ L ∞ ( Q ) . \max_i\|y_i(t)\|_{L^\infty(\Omega)}\le\int_0^t e^{C_A(t-s)}\max_i\|F_i(s)\|_{L^\infty(\Omega)}\,\mathrm{d}s\le T e^{C_A T}\max_i\|F_i\|_{L^\infty(Q)}. i max ∥ y i ( t ) ∥ L ∞ ( Ω ) ≤ ∫ 0 t e C A ( t − s ) i max ∥ F i ( s ) ∥ L ∞ ( Ω ) d s ≤ T e C A T i max ∥ F i ∥ L ∞ ( Q ) .
The inequalities max i ∣ a i ∣ ≤ ∣ a ∣ R 2 ≤ 2 max i ∣ a i ∣ \max_i|a_i|\le|a|_{\mathbb{R}^2}\le\sqrt{2}\max_i|a_i| max i ∣ a i ∣ ≤ ∣ a ∣ R 2 ≤ 2 max i ∣ a i ∣ give the corresponding bound in the Euclidean product norm. For almost every t ∈ ( 0 , T ) t\in(0,T) t ∈ ( 0 , T ) , testing by y y y gives
1 2 d d t ∥ y ∥ 2 2 + ∑ i = 1 2 μ i ∥ ∇ y i ∥ 2 2 = ( A y , y ) Ω + ( F , y ) Ω ≤ C ∥ y ∥ 2 2 + 1 2 ∥ F ∥ 2 2 . \frac{1}{2}\frac{\mathrm{d}}{\mathrm{d}t}\|y\|_2^2+\sum_{i=1}^{2}\mu_i\|\nabla y_i\|_2^2=(Ay,y)_\Omega+(F,y)_\Omega\le C\|y\|_2^2+\frac{1}{2}\|F\|_2^2. 2 1 d t d ∥ y ∥ 2 2 + i = 1 ∑ 2 μ i ∥∇ y i ∥ 2 2 = ( A y , y ) Ω + ( F , y ) Ω ≤ C ∥ y ∥ 2 2 + 2 1 ∥ F ∥ 2 2 .
Integration and Gronwall give
sup 0 ≤ t ≤ T ∥ y ( t ) ∥ 2 2 + ∑ i = 1 2 μ i ∥ ∇ y i ∥ L 2 ( Q ) 2 ≤ C R ∥ F ∥ L 2 ( Q ) 2 2 . \sup_{0\le t\le T}\|y(t)\|_2^2+\sum_{i=1}^{2}\mu_i\|\nabla y_i\|_{L^2(Q)}^2\le C_R\|F\|_{L^2(Q)^2}^2. 0 ≤ t ≤ T sup ∥ y ( t ) ∥ 2 2 + i = 1 ∑ 2 μ i ∥∇ y i ∥ L 2 ( Q ) 2 ≤ C R ∥ F ∥ L 2 ( Q ) 2 2 .
For ϕ ∈ H 1 ( Ω ) 2 \phi\in H^1(\Omega)^2 ϕ ∈ H 1 ( Ω ) 2 , the equation also gives
∣ ⟨ ∂ t y , ϕ ⟩ ∣ ≤ C R ( ∥ ∇ y ∥ 2 + ∥ y ∥ 2 + ∥ F ∥ 2 ) ∥ ϕ ∥ H 1 ( Ω ) 2 . |\langle\partial_t y,\phi\rangle|\le C_R(\|\nabla y\|_2+\|y\|_2+\|F\|_2)\|\phi\|_{H^1(\Omega)^2}. ∣ ⟨ ∂ t y , ϕ ⟩ ∣ ≤ C R ( ∥∇ y ∥ 2 + ∥ y ∥ 2 + ∥ F ∥ 2 ) ∥ ϕ ∥ H 1 ( Ω ) 2 .
Taking the dual norm and then integrating in time bounds ∂ t y \partial_t y ∂ t y in L 2 ( 0 , T ; H ∗ 1 ) 2 L^2(0,T;H^1_*)^2 L 2 ( 0 , T ; H ∗ 1 ) 2 . Together, these estimates give
∥ y ∥ Y 2 ≤ C R ∥ F ∥ L ∞ ( Q ) 2 . (42) \|y\|_{Y^2}\le C_R\|F\|_{L^\infty(Q)^2}. \tag*{(42)} ∥ y ∥ Y 2 ≤ C R ∥ F ∥ L ∞ ( Q ) 2 . (42)
The constant C R C_R C R depends only on the data, T T T and R R R . For ξ ∈ L ∞ ( Q ) \xi\in L^\infty(Q) ξ ∈ L ∞ ( Q ) , the tangent satisfies the following evolution equation in Q Q Q , initial condition in Ω \Omega Ω , and norm bound:
L v z ξ = ( 0 − q v 2 ξ ) , z ξ ( 0 ) = 0 , ∥ z ξ ∥ Y 2 ≤ C R ∥ ξ ∥ ∞ . (43) \mathcal{L}_v z^\xi=\begin{pmatrix}0\\-qv_2\xi\end{pmatrix},\qquad z^\xi(0)=0,\qquad\|z^\xi\|_{Y^2}\le C_R\|\xi\|_\infty. \tag*{(43)} L v z ξ = ( 0 − q v 2 ξ ) , z ξ ( 0 ) = 0 , ∥ z ξ ∥ Y 2 ≤ C R ∥ ξ ∥ ∞ . (43)
For ζ ∈ L ∞ ( Q ) \zeta\in L^\infty(Q) ζ ∈ L ∞ ( Q ) , set δ v : = S ( u + ζ ) − S ( u ) \delta v:=S(u+\zeta)-S(u) δ v := S ( u + ζ ) − S ( u ) . The state difference solves an equation with bounded coefficients and source ( 0 , − q ζ v 2 ( u + ζ ) ) T (0,-q\zeta v_2(u+\zeta))^T ( 0 , − q ζ v 2 ( u + ζ ) ) T , so ∥ δ v ∥ Y 2 ≤ C R ∥ ζ ∥ ∞ \|\delta v\|_{Y^2}\le C_R\|\zeta\|_\infty ∥ δ v ∥ Y 2 ≤ C R ∥ ζ ∥ ∞ . Subtracting the tangent equation gives the following evolution equation in Q Q Q and initial condition in Ω \Omega Ω :
L v ( δ v − z ζ ) = ( α δ v 1 δ v 2 − β δ v 1 δ v 2 − q ζ δ v 2 ) , ( δ v − z ζ ) ( 0 ) = 0. \mathcal{L}_v(\delta v-z^\zeta)=\begin{pmatrix}\alpha\delta v_1\delta v_2\\-\beta\delta v_1\delta v_2-q\zeta\delta v_2\end{pmatrix},\qquad(\delta v-z^\zeta)(0)=0. L v ( δ v − z ζ ) = ( α δ v 1 δ v 2 − β δ v 1 δ v 2 − q ζ δ v 2 ) , ( δ v − z ζ ) ( 0 ) = 0.
This source is O ( ∥ ζ ∥ ∞ 2 ) O(\|\zeta\|_\infty^2) O ( ∥ ζ ∥ ∞ 2 ) in L ∞ ( Q ) 2 L^\infty(Q)^2 L ∞ ( Q ) 2 , so
∥ S ( u + ζ ) − S ( u ) − z ζ ∥ Y 2 ≤ C R ∥ ζ ∥ ∞ 2 . \|S(u+\zeta)-S(u)-z^\zeta\|_{Y^2}\le C_R\|\zeta\|_\infty^2. ∥ S ( u + ζ ) − S ( u ) − z ζ ∥ Y 2 ≤ C R ∥ ζ ∥ ∞ 2 .
This proves S ′ ( u ) ξ = z ξ S'(u)\xi=z^\xi S ′ ( u ) ξ = z ξ . For ξ , ζ ∈ L ∞ ( Q ) \xi,\zeta\in L^\infty(Q) ξ , ζ ∈ L ∞ ( Q ) , define B ξ ζ , D ξ ζ ∈ L ∞ ( Q ) B_{\xi\zeta},D_{\xi\zeta}\in L^\infty(Q) B ξ ζ , D ξ ζ ∈ L ∞ ( Q ) by
B ξ ζ : = z 1 ξ z 2 ζ + z 1 ζ z 2 ξ , D ξ ζ : = ξ z 2 ζ + ζ z 2 ξ . B_{\xi\zeta}:=z_1^\xi z_2^\zeta+z_1^\zeta z_2^\xi,\qquad D_{\xi\zeta}:=\xi z_2^\zeta+\zeta z_2^\xi. B ξ ζ := z 1 ξ z 2 ζ + z 1 ζ z 2 ξ , D ξ ζ := ξ z 2 ζ + ζ z 2 ξ .
Then w : = S ′ ′ ( u ) [ ξ , ζ ] ∈ Y 2 w:=S''(u)[\xi,\zeta]\in Y^2 w := S ′′ ( u ) [ ξ , ζ ] ∈ Y 2 satisfies the following system. The evolution equations hold in Q Q Q , the Neumann conditions on ∂ Ω × ( 0 , T ) \partial\Omega\times(0,T) ∂ Ω × ( 0 , T ) , and the initial conditions in Ω \Omega Ω for i = 1 , 2 i=1,2 i = 1 , 2 :
∂ t w 1 − μ 1 Δ w 1 = ( − m + α v 2 ) w 1 + α v 1 w 2 + α B ξ ζ , ∂ t w 2 − μ 2 Δ w 2 = − β v 2 w 1 + ( r − β v 1 − q u ) w 2 − β B ξ ζ − q D ξ ζ , ∂ n w i = 0 , w i ( 0 ) = 0. (44) \begin{aligned}
\partial_t w_1-\mu_1\Delta w_1&=(-m+\alpha v_2)w_1+\alpha v_1w_2+\alpha B_{\xi\zeta},\\
\partial_t w_2-\mu_2\Delta w_2&=-\beta v_2w_1+(r-\beta v_1-qu)w_2-\beta B_{\xi\zeta}-qD_{\xi\zeta},\\
\partial_n w_i&=0,\qquad w_i(0)=0. \tag*{(44)}
\end{aligned} ∂ t w 1 − μ 1 Δ w 1 ∂ t w 2 − μ 2 Δ w 2 ∂ n w i = ( − m + α v 2 ) w 1 + α v 1 w 2 + α B ξ ζ , = − β v 2 w 1 + ( r − β v 1 − q u ) w 2 − β B ξ ζ − q D ξ ζ , = 0 , w i ( 0 ) = 0. (44)
with ∥ w ∥ Y 2 ≤ C R ∥ ξ ∥ ∞ ∥ ζ ∥ ∞ \lVert w\rVert_{Y^2} \le C_R\lVert\xi\rVert_\infty\lVert\zeta\rVert_\infty ∥ w ∥ Y 2 ≤ C R ∥ ξ ∥ ∞ ∥ ζ ∥ ∞ . To identify this as the second derivative, set z ~ ξ : = S ′ ′ ( u + ζ ) ξ \widetilde{z}^{\xi}:=S''(u+\zeta)\xi z ξ := S ′′ ( u + ζ ) ξ , δ z ξ : = z ~ ξ − z ξ \delta z^{\xi}:=\widetilde{z}^{\xi}-z^{\xi} δ z ξ := z ξ − z ξ . Since
δ A : = ( α δ v 2 α δ v 1 − β δ v 2 − β δ v 1 − q ζ ) , \delta A:=\begin{pmatrix} \alpha\delta v_2 & \alpha\delta v_1 \\ -\beta\delta v_2 & -\beta\delta v_1-q\zeta\end{pmatrix}, δ A := ( α δ v 2 − β δ v 2 α δ v 1 − β δ v 1 − q ζ ) ,
subtraction gives, in Ω × ( 0 , T ) \Omega\times(0,T) Ω × ( 0 , T ) ,
L v δ z ξ = δ A z ~ ξ + ( 0 − q δ v 2 ξ ) . \mathcal{L}_v\delta z^{\xi}=\delta A\widetilde{z}^{\xi}+\begin{pmatrix} 0 \\ -q\delta v_2\xi\end{pmatrix}. L v δ z ξ = δ A z ξ + ( 0 − q δ v 2 ξ ) .
thus
∥ δ z ξ ∥ Y 2 ≤ C R ∥ ξ ∥ ∞ ∥ ζ ∥ ∞ . \lVert\delta z^{\xi}\rVert_{Y^2}\le C_R\lVert\xi\rVert_\infty\lVert\zeta\rVert_\infty. ∥ δ z ξ ∥ Y 2 ≤ C R ∥ ξ ∥ ∞ ∥ ζ ∥ ∞ .
Set r ζ = δ v − z ζ r^\zeta=\delta v-z^\zeta r ζ = δ v − z ζ . Subtracting (44) gives the remainder equation
L v ( δ z ξ − w ) = δ A δ z ξ + ( α ( r 2 ζ ξ 1 ξ + r 1 ζ ξ 2 ξ ) − β ( r 2 ζ ξ 1 ξ + r 1 ζ ξ 2 ξ ) − q ξ r 2 ζ ) , ( δ z ξ − w ) ( 0 ) = 0. \mathcal{L}_v(\delta z^\xi-w)=\delta A\delta z^\xi+\begin{pmatrix} \alpha(r_2^\zeta\xi_1^\xi+r_1^\zeta\xi_2^\xi) \\ -\beta(r_2^\zeta\xi_1^\xi+r_1^\zeta\xi_2^\xi)-q\xi r_2^\zeta\end{pmatrix},\qquad(\delta z^\xi-w)(0)=0. L v ( δ z ξ − w ) = δ A δ z ξ + ( α ( r 2 ζ ξ 1 ξ + r 1 ζ ξ 2 ξ ) − β ( r 2 ζ ξ 1 ξ + r 1 ζ ξ 2 ξ ) − q ξ r 2 ζ ) , ( δ z ξ − w ) ( 0 ) = 0.
The two parts of the source satisfy
∥ δ A δ z ξ ∥ ∞ ≤ C R ∥ ζ ∥ ∞ ∥ δ z ξ ∥ ∞ ≤ C R ∥ ξ ∥ ∞ ∥ ζ ∥ ∞ 2 , ∥ r ζ ∥ ∞ ( ∥ z ζ ∥ ∞ + ∥ ξ ∥ ∞ ) ≤ C R ∥ ξ ∥ ∞ ∥ ζ ∥ ∞ 2 . \lVert\delta A\delta z^\xi\rVert_\infty\le C_R\lVert\zeta\rVert_\infty\lVert\delta z^\xi\rVert_\infty\le C_R\lVert\xi\rVert_\infty\lVert\zeta\rVert_\infty^2,\qquad\lVert r^\zeta\rVert_\infty(\lVert z^\zeta\rVert_\infty+\lVert\xi\rVert_\infty)\le C_R\lVert\xi\rVert_\infty\lVert\zeta\rVert_\infty^2. ∥ δ A δ z ξ ∥ ∞ ≤ C R ∥ ζ ∥ ∞ ∥ δ z ξ ∥ ∞ ≤ C R ∥ ξ ∥ ∞ ∥ ζ ∥ ∞ 2 , ∥ r ζ ∥ ∞ (∥ z ζ ∥ ∞ + ∥ ξ ∥ ∞ ) ≤ C R ∥ ξ ∥ ∞ ∥ ζ ∥ ∞ 2 .
Consequently,
\lVert\widetilde{z}^{\xi}-z^\xi-w\rVert_{Y^2}\le C_R\lVert\xi\rVert_\infty\lVert\zeta\rVert_\infty^2. {#eq:40}
For two base controls u , u ~ u,\widetilde{u} u , u in this ball, subtracting their tangent equations and applying (42) gives
∥ S ′ ( u ~ ) ξ − S ′ ( u ) ξ ∥ Y 2 ≤ C R ∥ u ~ − u ∥ ∞ ∥ ξ ∥ ∞ . \lVert S'(\widetilde{u})\xi-S'(u)\xi\rVert_{Y^2}\le C_R\lVert\widetilde{u}-u\rVert_\infty\lVert\xi\rVert_\infty. ∥ S ′ ( u ) ξ − S ′ ( u ) ξ ∥ Y 2 ≤ C R ∥ u − u ∥ ∞ ∥ ξ ∥ ∞ .
In the second-variation source, differences of products are expanded as
z ~ 1 ξ z ~ 2 ζ − z 1 ξ z 2 ζ = ( z ~ 1 ξ − z 1 ξ ) z ~ 2 ζ + z 1 ξ ( z ~ 2 ζ − z 2 ζ ) . \widetilde{z}_1^\xi\widetilde{z}_2^\zeta-z_1^\xi z_2^\zeta=(\widetilde{z}_1^\xi-z_1^\xi)\widetilde{z}_2^\zeta+z_1^\xi(\widetilde{z}_2^\zeta-z_2^\zeta). z 1 ξ z 2 ζ − z 1 ξ z 2 ζ = ( z 1 ξ − z 1 ξ ) z 2 ζ + z 1 ξ ( z 2 ζ − z 2 ζ ) .
Each term is bounded by C R ∥ u ~ − u ∥ ∞ ∥ ξ ∥ ∞ ∥ ζ ∥ ∞ C_R\lVert\widetilde{u}-u\rVert_\infty\lVert\xi\rVert_\infty\lVert\zeta\rVert_\infty C R ∥ u − u ∥ ∞ ∥ ξ ∥ ∞ ∥ ζ ∥ ∞ . The change in the linear operator contributes ( A ( S ( u ~ ) , u ~ ) − A ( S ( u ) , u ) ) S ′ ′ ( u ~ ) [ ξ , ζ ] (A(S(\widetilde{u}),\widetilde{u})-A(S(u),u))S''(\widetilde{u})[\xi,\zeta] ( A ( S ( u ) , u ) − A ( S ( u ) , u )) S ′′ ( u ) [ ξ , ζ ] , with the same bound. Hence
∥ S ′ ′ ( u ~ ) [ ξ , ζ ] − S ′ ′ ( u ) [ ξ , ζ ] ∥ Y 2 ≤ C R ∥ u ~ − u ∥ ∞ ∥ ξ ∥ ∞ ∥ ζ ∥ ∞ . \lVert S''(\widetilde{u})[\xi,\zeta]-S''(u)[\xi,\zeta]\rVert_{Y^2}\le C_R\lVert\widetilde{u}-u\rVert_\infty\lVert\xi\rVert_\infty\lVert\zeta\rVert_\infty. ∥ S ′′ ( u ) [ ξ , ζ ] − S ′′ ( u ) [ ξ , ζ ] ∥ Y 2 ≤ C R ∥ u − u ∥ ∞ ∥ ξ ∥ ∞ ∥ ζ ∥ ∞ .
Taking suprema over unit directions proves continuity of both derivatives in operator norm. Thus S : L ∞ ( Q ) → Y 2 S:L^\infty(Q)\to Y^2 S : L ∞ ( Q ) → Y 2 is C 2 C^2 C 2 on bounded control sets. Set
ℓ v : = ( γ 1 ( v 1 − v 1 d ) γ 2 ( v 2 − v 2 d ) − η q u ) ∈ L 2 ( Q ) 2 . \ell_v:=\begin{pmatrix} \gamma_1(v_1-v_1^d) \\ \gamma_2(v_2-v_2^d)-\eta qu \end{pmatrix}\in L^2(Q)^2. ℓ v := ( γ 1 ( v 1 − v 1 d ) γ 2 ( v 2 − v 2 d ) − η q u ) ∈ L 2 ( Q ) 2 .
The backward linear problem (40) has a weak Neumann solution p ∈ W 2 p\in\mathcal{W}^2 p ∈ W 2 . For y = z ξ y=z^\xi y = z ξ or w w w , integration by parts gives
\int_Q p\cdot\mathcal{L}_v y=\int_0^T\left[\langle\partial_t y,p\rangle+\sum_{i=1}^{2}\mu_i(\nabla y_i,\nabla p_i)_\Omega-(Ay,p)_\Omega\right]\mathrm{d}t=[(p,y)_\Omega]_0^T+\int_Q\ell_v\cdot y. {#eq:43}
Here ⟨ ⋅ , ⋅ ⟩ \langle\cdot,\cdot\rangle ⟨ ⋅ , ⋅ ⟩ is the componentwise dual pairing between ( H 1 ( Ω ) ∗ ) 2 (H^1(\Omega)^*)^2 ( H 1 ( Ω ) ∗ ) 2 and H 1 ( Ω ) 2 H^1(\Omega)^2 H 1 ( Ω ) 2 . The trace term vanishes because p ( T ) = y ( 0 ) = 0 p(T)=y(0)=0 p ( T ) = y ( 0 ) = 0 , and the chain rule gives
j ′ ( u ) ξ = ∫ Q ℓ v ⋅ z ξ + ∫ Q ( λ u − η q v 2 ) ξ = ∫ Q p ⋅ L v z ξ + ∫ Q ( λ u − η q v 2 ) ξ = ∫ Q [ λ u − q v 2 ( η + p 2 ) ] ξ , j'(u)\xi=\int_Q\ell_v\cdot z^\xi+\int_Q(\lambda u-\eta qv_2)\xi=\int_Qp\cdot\mathcal{L}_v z^\xi+\int_Q(\lambda u-\eta qv_2)\xi=\int_Q[\lambda u-qv_2(\eta+p_2)]\xi, j ′ ( u ) ξ = ∫ Q ℓ v ⋅ z ξ + ∫ Q ( λ u − η q v 2 ) ξ = ∫ Q p ⋅ L v z ξ + ∫ Q ( λ u − η q v 2 ) ξ = ∫ Q [ λ u − q v 2 ( η + p 2 )] ξ ,
proving (4) . At second order,
j ′ ′ ( u ) [ ξ , ζ ] = λ ∫ Q ξ ζ + ∑ i = 1 2 γ i ∫ Q z i ξ z i ζ − η q ∫ Q D ξ ζ + ∫ Q ℓ v ⋅ w , j''(u)[\xi,\zeta]=\lambda\int_Q\xi\zeta+\sum_{i=1}^{2}\gamma_i\int_Q z_i^\xi z_i^\zeta-\eta q\int_Q D\xi\zeta+\int_Q\ell_v\cdot w, j ′′ ( u ) [ ξ , ζ ] = λ ∫ Q ξ ζ + i = 1 ∑ 2 γ i ∫ Q z i ξ z i ζ − η q ∫ Q D ξ ζ + ∫ Q ℓ v ⋅ w ,
and
∫ Q ℓ v ⋅ w = ∫ Q p ⋅ L v , w = ∫ Q ( α p 1 − β p 2 ) B ξ ζ − q ∫ Q p 2 D ξ ζ . \int_Q \ell_v\cdot w = \int_Q p\cdot\mathcal{L}_{v,w} = \int_Q(\alpha p_1-\beta p_2)B_{\xi\zeta}-q\int_Qp_2D_{\xi\zeta}. ∫ Q ℓ v ⋅ w = ∫ Q p ⋅ L v , w = ∫ Q ( α p 1 − β p 2 ) B ξ ζ − q ∫ Q p 2 D ξ ζ .
Substitution gives (41) . The bounds ξ , ζ ∈ L ∞ ( Q ) \xi,\zeta\in L^\infty(Q) ξ , ζ ∈ L ∞ ( Q ) , z ξ , z ζ ∈ L ∞ ( Q ) 2 z^\xi,z^\zeta\in L^\infty(Q)^2 z ξ , z ζ ∈ L ∞ ( Q ) 2 and p ∈ L 2 ( Q ) 2 p\in L^2(Q)^2 p ∈ L 2 ( Q ) 2 ensure that every term is integrable.
Variational inequality and pointwise projection
Let u ∗ u^* u ∗ be a local minimizer and take u ∈ U a d u\in\mathcal{U}_{\mathrm{ad}} u ∈ U ad . The segment u ∗ + t ( u − u ∗ ) u^*+t(u-u^*) u ∗ + t ( u − u ∗ ) is admissible for 0 ≤ t ≤ 1 0\leq t\leq1 0 ≤ t ≤ 1 . Thus
0 ≤ lim t → 0 + j ( u ∗ + t ( u − u ∗ ) ) − j ( u ∗ ) t = ∫ Q g ( u ∗ ) ( u − u ∗ ) . 0\leq\lim_{t\to0^+}\frac{j(u^*+t(u-u^*))-j(u^*)}{t}=\int_Q g(u^*)(u-u^*). 0 ≤ t → 0 + lim t j ( u ∗ + t ( u − u ∗ )) − j ( u ∗ ) = ∫ Q g ( u ∗ ) ( u − u ∗ ) .
Set a = q S 2 ( u ∗ ) ( η + p 2 ( u ∗ ) ) / λ ∈ L 2 ( Q ) a=qS_2(u^*)(\eta+p_2(u^*))/\lambda\in L^2(Q) a = q S 2 ( u ∗ ) ( η + p 2 ( u ∗ )) / λ ∈ L 2 ( Q ) and w = Π [ 0 , U max ] ( a ) ∈ U a d w=\Pi_{[0,U_{\max}]}(a)\in\mathcal{U}_{\mathrm{ad}} w = Π [ 0 , U m a x ] ( a ) ∈ U ad . The scalar projection satisfies, for every b ∈ [ 0 , U max ] b\in[0,U_{\max}] b ∈ [ 0 , U m a x ] ,
( w − a ) ( b − w ) ≥ 0. (w-a)(b-w)\geq0. ( w − a ) ( b − w ) ≥ 0.
Apply this pointwise with b = u ∗ b=u^* b = u ∗ and use g ( u ∗ ) = λ ( u ∗ − a ) g(u^*)=\lambda(u^*-a) g ( u ∗ ) = λ ( u ∗ − a ) . Testing the variational inequality with w w w gives
0 ≤ ∫ Q ( u ∗ − a ) ( w − u ∗ ) = − ∥ u ∗ − w ∥ L 2 ( Q ) 2 + ∫ Q ( w − a ) ( w − u ∗ ) ⏟ ≤ 0 ≤ − ∥ u ∗ − w ∥ L 2 ( Q ) 2 . 0\leq\int_Q(u^*-a)(w-u^*)=-\lVert u^*-w\rVert_{L^2(Q)}^2+\underbrace{\int_Q(w-a)(w-u^*)}_{\leq0}\leq-\lVert u^*-w\rVert_{L^2(Q)}^2. 0 ≤ ∫ Q ( u ∗ − a ) ( w − u ∗ ) = − ∥ u ∗ − w ∥ L 2 ( Q ) 2 + ≤ 0 ∫ Q ( w − a ) ( w − u ∗ ) ≤ − ∥ u ∗ − w ∥ L 2 ( Q ) 2 .
Therefore u ∗ = w u^*=w u ∗ = w , which is the pointwise projection formula. Conversely, the scalar projection inequality integrates to the variational inequality. The pointwise conditions are
{ g ( u ∗ ) ≥ 0 where u ∗ = 0 , g ( u ∗ ) = 0 where 0 < u ∗ < U max , g ( u ∗ ) ≤ 0 where u ∗ = U max . \begin{cases}
g(u^*)\geq0 & \text{where }u^*=0,\\
g(u^*)=0 & \text{where }0<u^*<U_{\max},\\
g(u^*)\leq0 & \text{where }u^*=U_{\max}.
\end{cases} ⎩ ⎨ ⎧ g ( u ∗ ) ≥ 0 g ( u ∗ ) = 0 g ( u ∗ ) ≤ 0 where u ∗ = 0 , where 0 < u ∗ < U m a x , where u ∗ = U m a x .
Discrete bounds
Positivity and population balance
Proof. For D : = diag ( d 0 , … , d M ) ∈ R ( M + 1 ) × ( M + 1 ) D:=\operatorname{diag}(d_0,\ldots,d_M)\in\mathbb{R}^{(M+1)\times(M+1)} D := diag ( d 0 , … , d M ) ∈ R ( M + 1 ) × ( M + 1 ) , d j ≥ 0 d_j\geq0 d j ≥ 0 , μ > 0 \mu>0 μ > 0 and τ > 0 \tau>0 τ > 0 , set
K : = I − τ μ L h + τ D . K:=I-\tau\mu L_h+\tau D. K := I − τ μ L h + τ D .
Its off-diagonal entries are nonpositive, and L h 1 = 0 L_h\mathbf{1}=0 L h 1 = 0 gives
K j j − ∑ 0 ≤ k ≤ M k ≠ j ∣ K j k ∣ = 1 + τ d j > 0. K_{jj}-\sum_{\substack{0\leq k\leq M\\k\neq j}}\lvert K_{jk}\rvert=1+\tau d_j>0. K j j − 0 ≤ k ≤ M k = j ∑ ∣ K j k ∣ = 1 + τ d j > 0.
If y ∈ R M + 1 y\in\mathbb{R}^{M+1} y ∈ R M + 1 , K y ≥ 0 Ky\geq0 K y ≥ 0 , y j = min 0 ≤ k ≤ M y k < 0 y_j=\min_{0\leq k\leq M}y_k<0 y j = min 0 ≤ k ≤ M y k < 0 , then
( L h y ) j = ∑ 0 ≤ k ≤ M k ≠ j ( L h ) j k ( y k − y j ) ≥ 0 , (L_hy)_j=\sum_{\substack{0\leq k\leq M\\k\neq j}}(L_h)_{jk}(y_k-y_j)\geq0, ( L h y ) j = 0 ≤ k ≤ M k = j ∑ ( L h ) j k ( y k − y j ) ≥ 0 ,
so
( K y ) j = ( 1 + τ d j ) y j − τ μ ( L h y ) j < 0 , (Ky)_j=(1+\tau d_j)y_j-\tau\mu(L_hy)_j<0, ( K y ) j = ( 1 + τ d j ) y j − τ μ ( L h y ) j < 0 ,
a contradiction. Applying the argument to y y y and − y -y − y when K y = 0 Ky=0 K y = 0 proves invertibility. Taking y = K − 1 b y=K^{-1}b y = K − 1 b with b ≥ 0 b\ge0 b ≥ 0 then gives K − 1 ≥ 0 K^{-1}\ge0 K − 1 ≥ 0 . We obtain
K − 1 ≥ 0 , K 1 ≥ 1 , ∥ K − 1 ∥ ∞ = ∥ K − 1 1 ∥ ∞ ≤ 1. (45) K^{-1}\ge0,\qquad K\mathbf{1}\ge\mathbf{1},\qquad\lVert K^{-1}\rVert_{\infty}=\lVert K^{-1}\mathbf{1}\rVert_{\infty}\le1. \tag*{(45)} K − 1 ≥ 0 , K 1 ≥ 1 , ∥ K − 1 ∥ ∞ = ∥ K − 1 1 ∥ ∞ ≤ 1. (45)
The same inverse is contractive in the weighted L 2 L^2 L 2 norm. Indeed, if K y = b Ky=b K y = b , weighted summation by parts gives
∥ y ∥ h 2 + τ μ ∣ y ∣ 1 , h 2 + τ ∑ j = 0 M w j d j y j 2 = ⟨ b , y ⟩ h ≤ ∥ b ∥ h ∥ y ∥ h ⟹ ∥ K − 1 b ∥ h ≤ ∥ b ∥ h . \lVert y\rVert_h^2+\tau\mu\lvert y\rvert_{1,h}^2+\tau\sum_{j=0}^{M}w_jd_jy_j^2=\langle b,y\rangle_h\le\lVert b\rVert_h\lVert y\rVert_h\quad\Longrightarrow\quad\lVert K^{-1}b\rVert_h\le\lVert b\rVert_h. ∥ y ∥ h 2 + τ μ ∣ y ∣ 1 , h 2 + τ j = 0 ∑ M w j d j y j 2 = ⟨ b , y ⟩ h ≤ ∥ b ∥ h ∥ y ∥ h ⟹ ∥ K − 1 b ∥ h ≤ ∥ b ∥ h .
Also B 1 = ( 1 + τ m ) 1 B\mathbf{1}=(1+\tau m)\mathbf{1} B 1 = ( 1 + τ m ) 1 gives ∥ B − 1 ∥ ∞ = ( 1 + τ m ) − 1 \lVert B^{-1}\rVert_{\infty}=(1+\tau m)^{-1} ∥ B − 1 ∥ ∞ = ( 1 + τ m ) − 1 , so solving either linear system in (6) preserves nonnegativity. Weighted symmetry gives ⟨ L h y , 1 ⟩ h = ⟨ y , L h 1 ⟩ h = 0 \langle L_hy,\mathbf{1}\rangle_h=\langle y,L_h\mathbf{1}\rangle_h=0 ⟨ L h y , 1 ⟩ h = ⟨ y , L h 1 ⟩ h = 0 . Testing with 1 \mathbf{1} 1 yields
⟨ v 1 n + 1 − v 1 n , 1 ⟩ h = τ ⟨ − m v 1 n + 1 + α v 1 n ∘ v 2 n + 1 , 1 ⟩ h , ⟨ v 2 n + 1 − v 2 n , 1 ⟩ h = τ ⟨ r v 2 n − ( β v 1 n + q u n ) ∘ v 2 n + 1 , 1 ⟩ h . \langle v_1^{n+1}-v_1^n,\mathbf{1}\rangle_h=\tau\langle-mv_1^{n+1}+\alpha v_1^n\circ v_2^{n+1},\mathbf{1}\rangle_h,\qquad\langle v_2^{n+1}-v_2^n,\mathbf{1}\rangle_h=\tau\langle rv_2^n-(\beta v_1^n+qu^n)\circ v_2^{n+1},\mathbf{1}\rangle_h. ⟨ v 1 n + 1 − v 1 n , 1 ⟩ h = τ ⟨ − m v 1 n + 1 + α v 1 n ∘ v 2 n + 1 , 1 ⟩ h , ⟨ v 2 n + 1 − v 2 n , 1 ⟩ h = τ ⟨ r v 2 n − ( β v 1 n + q u n ) ∘ v 2 n + 1 , 1 ⟩ h .
Multiplication by β \beta β and α \alpha α gives the transfer cancellation
τ α β ⟨ v 1 n + 1 ∘ v 2 n + 1 , 1 ⟩ h − τ α β ⟨ v 1 n ∘ v 2 n + 1 , 1 ⟩ h = 0. \tau\alpha\beta\langle v_1^{n+1}\circ v_2^{n+1},\mathbf{1}\rangle_h-\tau\alpha\beta\langle v_1^n\circ v_2^{n+1},\mathbf{1}\rangle_h=0. τ α β ⟨ v 1 n + 1 ∘ v 2 n + 1 , 1 ⟩ h − τ α β ⟨ v 1 n ∘ v 2 n + 1 , 1 ⟩ h = 0.
Consequently,
M h n + 1 − M h n = − τ β m ⟨ v 1 n + 1 , 1 ⟩ h + τ α r ⟨ v 2 n , 1 ⟩ h − τ α q ⟨ u n , v 2 n + 1 ⟩ h , \mathcal{M}_h^{n+1}-\mathcal{M}_h^n=-\tau\beta m\langle v_1^{n+1},\mathbf{1}\rangle_h+\tau\alpha r\langle v_2^n,\mathbf{1}\rangle_h-\tau\alpha q\langle u^n,v_2^{n+1}\rangle_h, M h n + 1 − M h n = − τ β m ⟨ v 1 n + 1 , 1 ⟩ h + τ α r ⟨ v 2 n , 1 ⟩ h − τ α q ⟨ u n , v 2 n + 1 ⟩ h ,
and
M h N + τ β m ∑ n = 0 N − 1 ⟨ v 1 n + 1 , 1 ⟩ h + τ α q ∑ n = 0 N − 1 ⟨ u n , v 2 n + 1 ⟩ h = M h 0 + τ α r ∑ n = 0 N − 1 ⟨ v 2 n , 1 ⟩ h . \mathcal{M}_h^N+\tau\beta m\sum_{n=0}^{N-1}\langle v_1^{n+1},\mathbf{1}\rangle_h+\tau\alpha q\sum_{n=0}^{N-1}\langle u^n,v_2^{n+1}\rangle_h=\mathcal{M}_h^0+\tau\alpha r\sum_{n=0}^{N-1}\langle v_2^n,\mathbf{1}\rangle_h. M h N + τ β m n = 0 ∑ N − 1 ⟨ v 1 n + 1 , 1 ⟩ h + τ α q n = 0 ∑ N − 1 ⟨ u n , v 2 n + 1 ⟩ h = M h 0 + τ α r n = 0 ∑ N − 1 ⟨ v 2 n , 1 ⟩ h .
Dropping the nonpositive terms in the one-step identity gives
M h n + 1 ≤ M h n + τ α r ⟨ v 2 n , 1 ⟩ h ≤ ( 1 + τ r ) M h n . (46) \mathcal{M}_h^{n+1}\le\mathcal{M}_h^n+\tau\alpha r\langle v_2^n,\mathbf{1}\rangle_h\le(1+\tau r)\mathcal{M}_h^n. \tag*{(46)} M h n + 1 ≤ M h n + τ α r ⟨ v 2 n , 1 ⟩ h ≤ ( 1 + τ r ) M h n . (46)
Uniform bounds
For y , z ∈ R M + 1 y,z\in\mathbb{R}^{M+1} y , z ∈ R M + 1 , weighted summation by parts and the definition of I h \mathcal{I}_h I h give
a h ( y , z ) = 1 h ∑ j = 0 M − 1 ( y j + 1 − y j ) ( z j + 1 − z j ) = ∫ 0 1 ∂ x I h y ∂ x I h z d x = − ⟨ L h y , z ⟩ h . (47) a_h(y,z)=\frac{1}{h}\sum_{j=0}^{M-1}(y_{j+1}-y_j)(z_{j+1}-z_j)=\int_0^1\partial_x\mathcal{I}_hy\,\partial_x\mathcal{I}_hz\,\mathrm{d}x=-\langle L_hy,z\rangle_h. \tag*{(47)} a h ( y , z ) = h 1 j = 0 ∑ M − 1 ( y j + 1 − y j ) ( z j + 1 − z j ) = ∫ 0 1 ∂ x I h y ∂ x I h z d x = − ⟨ L h y , z ⟩ h . (47)
To see the role of the endpoint weights, expand the left side:
− ⟨ L h y , z ⟩ h = 1 h [ ( y 0 − y 1 ) z 0 + ∑ j = 1 M − 1 ( 2 y j − y j − 1 − y j + 1 ) z j + ( y M − y M − 1 ) z M ] = 1 h ∑ j = 0 M − 1 ( y j + 1 − y j ) ( z j + 1 − z j ) . -\langle L_hy,z\rangle_h=\frac{1}{h}\left[(y_0-y_1)z_0+\sum_{j=1}^{M-1}(2y_j-y_{j-1}-y_{j+1})z_j+(y_M-y_{M-1})z_M\right]=\frac{1}{h}\sum_{j=0}^{M-1}(y_{j+1}-y_j)(z_{j+1}-z_j). − ⟨ L h y , z ⟩ h = h 1 [ ( y 0 − y 1 ) z 0 + j = 1 ∑ M − 1 ( 2 y j − y j − 1 − y j + 1 ) z j + ( y M − y M − 1 ) z M ] = h 1 j = 0 ∑ M − 1 ( y j + 1 − y j ) ( z j + 1 − z j ) .
On [ x j , x j + 1 ] [x_j,x_{j+1}] [ x j , x j + 1 ] , the reconstructed derivative is ( y j + 1 − y j ) / h (y_{j+1}-y_j)/h ( y j + 1 − y j ) / h . Thus
∫ x j x j + 1 ( I h y ) ′ ( I h z ) ′ d x = ( y j + 1 − y j ) ( z j + 1 − z j ) h . \int_{x_j}^{x_{j+1}}(\mathcal{I}_hy)'(\mathcal{I}_hz)'\,\mathrm{d}x=\frac{(y_{j+1}-y_j)(z_{j+1}-z_j)}{h}. ∫ x j x j + 1 ( I h y ) ′ ( I h z ) ′ d x = h ( y j + 1 − y j ) ( z j + 1 − z j ) .
Summing over the intervals proves (47) . Its symmetry and nonnegativity give
W h L h = L h T W h , − y T W h L h y = ∣ y ∣ 1 , h 2 ≥ 0. W_hL_h=L_h^{T}W_h,\qquad-y^{T}W_hL_hy=\lvert y\rvert_{1,h}^{2}\ge0. W h L h = L h T W h , − y T W h L h y = ∣ y ∣ 1 , h 2 ≥ 0.
For the positive scheme, define comparison sequences by
P 2 n + 1 : = ( 1 + τ r ) P 2 n , P 1 n + 1 : = 1 + τ α P 2 n + 1 1 + τ m P 1 n . (48) P_2^{n+1} := (1+\tau r)P_2^n,\qquad P_1^{n+1} := \frac{1+\tau\alpha P_2^{n+1}}{1+\tau m}P_1^n. \tag*{(48)} P 2 n + 1 := ( 1 + τ r ) P 2 n , P 1 n + 1 := 1 + τ m 1 + τ α P 2 n + 1 P 1 n . (48)
with P i 0 : = ∥ v i 0 ∥ ∞ P_i^0 := \lVert v_i^0\rVert_\infty P i 0 := ∥ v i 0 ∥ ∞ . If 0 ≤ v i n ≤ P i n 1 0\le v_i^n\le P_i^n\mathbf{1} 0 ≤ v i n ≤ P i n 1 , the inverse bounds give
v 2 n + 1 ≤ ( 1 + τ r ) P 2 n ( C n ) − 1 1 ≤ P 2 n + 1 1 , v 1 n + 1 ≤ P 1 n ( 1 + τ α P 2 n + 1 ) B − 1 1 ≤ P 1 n + 1 1. v_2^{n+1}\le(1+\tau r)P_2^n(C^n)^{-1}\mathbf{1}\le P_2^{n+1}\mathbf{1},\qquad v_1^{n+1}\le P_1^n(1+\tau\alpha P_2^{n+1})B^{-1}\mathbf{1}\le P_1^{n+1}\mathbf{1}. v 2 n + 1 ≤ ( 1 + τ r ) P 2 n ( C n ) − 1 1 ≤ P 2 n + 1 1 , v 1 n + 1 ≤ P 1 n ( 1 + τ α P 2 n + 1 ) B − 1 1 ≤ P 1 n + 1 1 .
For s ≥ 0 s\ge0 s ≥ 0 , we have 1 + s ≤ e s 1+s\le e^s 1 + s ≤ e s . Induction therefore gives
P 2 n ≤ P 2 0 ( 1 + τ r ) n ≤ P 2 0 e r t n ≤ K 2 , P_2^n\le P_2^0(1+\tau r)^n\le P_2^0e^{rtn}\le K_2, P 2 n ≤ P 2 0 ( 1 + τ r ) n ≤ P 2 0 e r t n ≤ K 2 ,
and
P 1 n ≤ P 1 0 ∏ ℓ = 0 n − 1 ( 1 + τ α P 2 ℓ + 1 ) ≤ P 1 0 ( 1 + τ α K 2 ) n ≤ P 1 0 e α K 2 T = K 1 . P_1^n\le P_1^0\prod_{\ell=0}^{n-1}(1+\tau\alpha P_2^{\ell+1})\le P_1^0(1+\tau\alpha K_2)^n\le P_1^0e^{\alpha K_2T}=K_1. P 1 n ≤ P 1 0 ℓ = 0 ∏ n − 1 ( 1 + τ α P 2 ℓ + 1 ) ≤ P 1 0 ( 1 + τ α K 2 ) n ≤ P 1 0 e α K 2 T = K 1 .
For IMEX, assume the bounds hold at time n n n and τ ≤ τ 0 \tau\le\tau_0 τ ≤ τ 0 . The old-state factors satisfy
1 + τ ( − m + α v 2 , j n ) ≥ 1 − τ m ≥ 0 , 1 + τ ( r − β v 1 , j n − q u j n ) ≥ 1 − τ ( β K 1 + q U max ) ≥ 0. 1+\tau(-m+\alpha v_{2,j}^n)\ge1-\tau m\ge0,\qquad1+\tau(r-\beta v_{1,j}^n-qu_j^n)\ge1-\tau(\beta K_1+qU_{\max})\ge0. 1 + τ ( − m + α v 2 , j n ) ≥ 1 − τ m ≥ 0 , 1 + τ ( r − β v 1 , j n − q u j n ) ≥ 1 − τ ( β K 1 + q U m a x ) ≥ 0.
Since ∥ A i − 1 ∥ ∞ ≤ 1 \lVert A_i^{-1}\rVert_\infty\le1 ∥ A i − 1 ∥ ∞ ≤ 1 ,
0 ≤ v 2 n + 1 ≤ ( 1 + τ r ) ∥ v 2 n ∥ ∞ 1 , 0 ≤ v 1 n + 1 ≤ ( 1 + τ α K 2 ) ∥ v 1 n ∥ ∞ 1. 0\le v_2^{n+1}\le(1+\tau r)\lVert v_2^n\rVert_\infty\mathbf{1},\qquad0\le v_1^{n+1}\le(1+\tau\alpha K_2)\lVert v_1^n\rVert_\infty\mathbf{1}. 0 ≤ v 2 n + 1 ≤ ( 1 + τ r ) ∥ v 2 n ∥ ∞ 1 , 0 ≤ v 1 n + 1 ≤ ( 1 + τ α K 2 ) ∥ v 1 n ∥ ∞ 1 .
This closes the induction. Under (13) , the embedding H 1 ( 0 , 1 ) ↪ C ( [ 0 , 1 ] ) H^1(0,1)\hookrightarrow C([0,1]) H 1 ( 0 , 1 ) ↪ C ([ 0 , 1 ]) gives P i 0 ≤ ∥ v i 0 ∥ L ∞ ( 0 , 1 ) P_i^0\le\lVert v_i^0\rVert_{L^\infty(0,1)} P i 0 ≤ ∥ v i 0 ∥ L ∞ ( 0 , 1 ) , so neither K i K_i K i nor τ 0 \tau_0 τ 0 depends on h h h . For equal diffusivities, let A : = I − τ μ L h A:=I-\tau\mu L_h A := I − τ μ L h and w n : = β v 1 n + α v 2 n ∈ R M + 1 w^n:=\beta v_1^n+\alpha v_2^n\in\mathbb{R}^{M+1} w n := β v 1 n + α v 2 n ∈ R M + 1 . The positive equations give
A w n + 1 = w n − τ β m v 1 n + 1 + τ α r v 2 n − τ α q u n ∘ v 2 n ≤ ( 1 + τ r ) w n . Aw^{n+1}=w^n-\tau\beta m v_1^{n+1}+\tau\alpha r v_2^n-\tau\alpha q u^n\circ v_2^n\le(1+\tau r)w^n. A w n + 1 = w n − τ β m v 1 n + 1 + τ α r v 2 n − τ α q u n ∘ v 2 n ≤ ( 1 + τ r ) w n .
Using A − 1 ≥ 0 A^{-1}\ge0 A − 1 ≥ 0 and A − 1 1 = 1 A^{-1}\mathbf{1}=\mathbf{1} A − 1 1 = 1 ,
w n + 1 ≤ ( 1 + τ r ) ∥ w n ∥ ∞ 1 ≤ ( 1 + τ r ) n + 1 ∥ w 0 ∥ ∞ 1. w^{n+1}\le(1+\tau r)\lVert w^n\rVert_\infty\mathbf{1}\le(1+\tau r)^{n+1}\lVert w^0\rVert_\infty\mathbf{1}. w n + 1 ≤ ( 1 + τ r ) ∥ w n ∥ ∞ 1 ≤ ( 1 + τ r ) n + 1 ∥ w 0 ∥ ∞ 1 .
For nonnegative IMEX trajectories, cancellation instead gives
A w n + 1 = w n − τ β m v 1 n + τ α r v 2 n − τ α q u n ∘ v 2 n ≤ ( 1 + τ r ) w n . Aw^{n+1}=w^n-\tau\beta m v_1^n+\tau\alpha r v_2^n-\tau\alpha q u^n\circ v_2^n\le(1+\tau r)w^n. A w n + 1 = w n − τ β m v 1 n + τ α r v 2 n − τ α q u n ∘ v 2 n ≤ ( 1 + τ r ) w n .
The same induction proves (8) .
Discrete Lagrangians and derivative identities
IMEX
Set p i N = 0 p_i^N=0 p i N = 0 . For n = N − 1 , … , 1 n=N-1,\ldots,1 n = N − 1 , … , 1 ,
{ A 1 p 1 n = p 1 n + 1 + τ [ γ 1 ( v 1 n − v 1 d , n ) + ( − m + α v 2 n ) ∘ p 1 n + 1 − β v 2 n ∘ p 2 n + 1 ] , A 2 p 2 n = p 2 n + 1 + τ [ γ 2 ( v 2 n − v 2 d , n ) − η q u n + α v 1 n ∘ p 1 n + 1 + ( r − β v 1 n − q u n ) ∘ p 2 n + 1 ] . (49) \begin{cases}
A_1p_1^n=p_1^{n+1}+\tau\left[\gamma_1(v_1^n-v_1^{d,n})+(-m+\alpha v_2^n)\circ p_1^{n+1}-\beta v_2^n\circ p_2^{n+1}\right],\\
A_2p_2^n=p_2^{n+1}+\tau\left[\gamma_2(v_2^n-v_2^{d,n})-\eta q u^n+\alpha v_1^n\circ p_1^{n+1}+(r-\beta v_1^n-q u^n)\circ p_2^{n+1}\right]. \tag*{(49)}
\end{cases} ⎩ ⎨ ⎧ A 1 p 1 n = p 1 n + 1 + τ [ γ 1 ( v 1 n − v 1 d , n ) + ( − m + α v 2 n ) ∘ p 1 n + 1 − β v 2 n ∘ p 2 n + 1 ] , A 2 p 2 n = p 2 n + 1 + τ [ γ 2 ( v 2 n − v 2 d , n ) − η q u n + α v 1 n ∘ p 1 n + 1 + ( r − β v 1 n − q u n ) ∘ p 2 n + 1 ] . (49)
The gradient is
g I , n = λ u n − η q v 2 n − q v 2 n ∘ p 2 n + 1 . (50) g^{I,n}=\lambda u^n-\eta qv_2^n-qv_2^n\circ p_2^{n+1}. \tag*{(50)} g I , n = λ u n − η q v 2 n − q v 2 n ∘ p 2 n + 1 . (50)
For d ∈ R N ( M + 1 ) d\in\mathbb{R}^{N(M+1)} d ∈ R N ( M + 1 ) , z : = D S h , τ I ( u ) d z:=DS_{h,\tau}^{I}(u)d z := D S h , τ I ( u ) d satisfies z i 0 = 0 z_i^0=0 z i 0 = 0 and
{ A 1 z 1 n + 1 = z 1 n + τ [ ( − m + α v 2 n ) ∘ z 1 n + α v 1 n ∘ z 2 n ] , A 2 z 2 n + 1 = z 2 n + τ [ − β v 2 n ∘ z 1 n + ( r − β v 1 n − q u n ) ∘ z 2 n − q v 2 n ∘ d n ] . (51) \begin{cases}
A_1z_1^{n+1}=z_1^n+\tau\left[(-m+\alpha v_2^n)\circ z_1^n+\alpha v_1^n\circ z_2^n\right],\\
A_2z_2^{n+1}=z_2^n+\tau\left[-\beta v_2^n\circ z_1^n+(r-\beta v_1^n-q u^n)\circ z_2^n-qv_2^n\circ d^n\right]. \tag*{(51)}
\end{cases} { A 1 z 1 n + 1 = z 1 n + τ [ ( − m + α v 2 n ) ∘ z 1 n + α v 1 n ∘ z 2 n ] , A 2 z 2 n + 1 = z 2 n + τ [ − β v 2 n ∘ z 1 n + ( r − β v 1 n − q u n ) ∘ z 2 n − q v 2 n ∘ d n ] . (51)
With ξ : = D p ( u ) d \xi:= Dp(u)d ξ := D p ( u ) d obtained from the differentiated adjoint (53) , the Hessian action is
( H I d ) n = λ d n − η q z 2 n − q ( z 2 n ∘ p 2 n + 1 + v 2 n ∘ ξ 2 n + 1 ) . (52) (H^I d)^n = \lambda d^n - \eta q z_2^n - q(z_2^n \circ p_2^{n+1} + v_2^n \circ\xi_2^{n+1}). \tag*{(52)} ( H I d ) n = λ d n − η q z 2 n − q ( z 2 n ∘ p 2 n + 1 + v 2 n ∘ ξ 2 n + 1 ) . (52)
To derive these formulas, fix the initial vectors and define the residuals F i n ∈ R M + 1 F_i^n \in\mathbb{R}^{M+1} F i n ∈ R M + 1 by
F 1 n : = A 1 v 1 n + 1 − v 1 n − τ v 1 n ∘ ( − m + α v 2 n ) , F 2 n : = A 2 v 2 n + 1 − v 2 n − τ v 2 n ∘ ( r − β v 1 n − q u n ) . F_1^n := A_1v_1^{n+1} - v_1^n - \tau v_1^n \circ(-m + \alpha v_2^n), \qquad F_2^n := A_2v_2^{n+1} - v_2^n - \tau v_2^n \circ(r - \beta v_1^n - qu^n). F 1 n := A 1 v 1 n + 1 − v 1 n − τ v 1 n ∘ ( − m + α v 2 n ) , F 2 n := A 2 v 2 n + 1 − v 2 n − τ v 2 n ∘ ( r − β v 1 n − q u n ) .
The Lagrangian is
L I : Y h , τ × R N ( M + 1 ) × R 2 N ( M + 1 ) ⟶ R \mathcal{L}^I : \mathcal{Y}_{h,\tau} \times\mathbb{R}^{N(M+1)} \times\mathbb{R}^{2N(M+1)} \longrightarrow\mathbb{R} L I : Y h , τ × R N ( M + 1 ) × R 2 N ( M + 1 ) ⟶ R
( v , u , p ) ⟼ J h , τ 0 ( v , u ) − ∑ n = 0 N − 1 ∑ i = 1 2 ⟨ p i n + 1 , F i n ⟩ h . (v,u,p) \longmapsto J_{h,\tau}^0(v,u) - \sum_{n=0}^{N-1}\sum_{i=1}^{2}\langle p_i^{n+1},F_i^n\rangle_h. ( v , u , p ) ⟼ J h , τ 0 ( v , u ) − n = 0 ∑ N − 1 i = 1 ∑ 2 ⟨ p i n + 1 , F i n ⟩ h .
For arbitrary state variations δ v \delta v δ v and a control variation d d d , differentiation gives
D F 1 n [ δ v , d ] = A 1 δ v 1 n + 1 − δ v 1 n − τ [ ( − m + α v 2 n ) ∘ δ v 1 n + α v 1 n ∘ δ v 2 n ] , DF_1^n[\delta v,d] = A_1\delta v_1^{n+1} - \delta v_1^n - \tau\left[(-m+\alpha v_2^n)\circ\delta v_1^n+\alpha v_1^n\circ\delta v_2^n\right], D F 1 n [ δ v , d ] = A 1 δ v 1 n + 1 − δ v 1 n − τ [ ( − m + α v 2 n ) ∘ δ v 1 n + α v 1 n ∘ δ v 2 n ] ,
D F 2 n [ δ v , d ] = A 2 δ v 2 n + 1 − δ v 2 n − τ [ − β v 2 n ∘ δ v 1 n + ( r − β v 1 n − q u n ) ∘ δ v 2 n − q v 2 n ∘ d n ] . DF_2^n[\delta v,d] = A_2\delta v_2^{n+1} - \delta v_2^n - \tau\left[-\beta v_2^n\circ\delta v_1^n+(r-\beta v_1^n-qu^n)\circ\delta v_2^n-qv_2^n\circ d^n\right]. D F 2 n [ δ v , d ] = A 2 δ v 2 n + 1 − δ v 2 n − τ [ − β v 2 n ∘ δ v 1 n + ( r − β v 1 n − q u n ) ∘ δ v 2 n − q v 2 n ∘ d n ] .
Weighted symmetry and componentwise multiplication give
⟨ p i n + 1 , A i δ v i n + 1 ⟩ h = ⟨ A i p i n + 1 , δ v i n + 1 ⟩ h , ⟨ p , a ∘ δ v ⟩ h = ⟨ a ∘ p , δ v ⟩ h . \langle p_i^{n+1},A_i\delta v_i^{n+1}\rangle_h=\langle A_ip_i^{n+1},\delta v_i^{n+1}\rangle_h,\qquad\langle p,a\circ\delta v\rangle_h=\langle a\circ p,\delta v\rangle_h. ⟨ p i n + 1 , A i δ v i n + 1 ⟩ h = ⟨ A i p i n + 1 , δ v i n + 1 ⟩ h , ⟨ p , a ∘ δ v ⟩ h = ⟨ a ∘ p , δ v ⟩ h .
The time index is shifted by
∑ n = 0 N − 1 ⟨ A i p i n + 1 , δ v i n + 1 ⟩ h = ∑ k = 1 N ⟨ A i p i k , δ v i k ⟩ h . \sum_{n=0}^{N-1}\langle A_ip_i^{n+1},\delta v_i^{n+1}\rangle_h=\sum_{k=1}^{N}\langle A_ip_i^k,\delta v_i^k\rangle_h. n = 0 ∑ N − 1 ⟨ A i p i n + 1 , δ v i n + 1 ⟩ h = k = 1 ∑ N ⟨ A i p i k , δ v i k ⟩ h .
With v i 0 v_i^0 v i 0 fixed, write D v L I [ δ v ] = ∑ i = 1 2 ∑ k = 1 N ⟨ c i k , δ v i k ⟩ h D_v\mathcal{L}^I[\delta v]=\sum_{i=1}^{2}\sum_{k=1}^{N}\langle c_i^k,\delta v_i^k\rangle_h D v L I [ δ v ] = ∑ i = 1 2 ∑ k = 1 N ⟨ c i k , δ v i k ⟩ h , δ v i k ∈ R M + 1 \delta v_i^k\in\mathbb{R}^{M+1} δ v i k ∈ R M + 1 . At k = N k=N k = N ,
c i N = − A i p i N = 0 ⟹ p i N = 0. c_i^N=-A_ip_i^N=0\quad\Longrightarrow\quad p_i^N=0. c i N = − A i p i N = 0 ⟹ p i N = 0.
For 1 ≤ k < N 1\leq k<N 1 ≤ k < N , the cost and the residuals at steps k − 1 , k k-1,k k − 1 , k give
c 1 k = τ γ 1 ( v 1 k − v 1 d , k ) − A 1 p 1 k + p 1 k + 1 + τ ( − m + α v 2 k ) ∘ p 1 k + 1 − τ β v 2 k ∘ p 2 k + 1 , c_1^k=\tau\gamma_1(v_1^k-v_1^{d,k})-A_1p_1^k+p_1^{k+1}+\tau(-m+\alpha v_2^k)\circ p_1^{k+1}-\tau\beta v_2^k\circ p_2^{k+1}, c 1 k = τ γ 1 ( v 1 k − v 1 d , k ) − A 1 p 1 k + p 1 k + 1 + τ ( − m + α v 2 k ) ∘ p 1 k + 1 − τ β v 2 k ∘ p 2 k + 1 ,
and
c 2 k = τ [ γ 2 ( v 2 k − v 2 d , k ) − η q u k ] − A 2 p 2 k + p 2 k + 1 + τ α v 1 k ∘ p 1 k + 1 + τ ( r − β v 1 k − q u k ) ∘ p 2 k + 1 . c_2^k=\tau[\gamma_2(v_2^k-v_2^{d,k})-\eta qu^k]-A_2p_2^k+p_2^{k+1}+\tau\alpha v_1^k\circ p_1^{k+1}+\tau(r-\beta v_1^k-qu^k)\circ p_2^{k+1}. c 2 k = τ [ γ 2 ( v 2 k − v 2 d , k ) − η q u k ] − A 2 p 2 k + p 2 k + 1 + τ α v 1 k ∘ p 1 k + 1 + τ ( r − β v 1 k − q u k ) ∘ p 2 k + 1 .
Setting c i k = 0 c_i^k=0 c i k = 0 gives (49) . With states fixed,
D u F 1 n [ d n ] = 0 , D u F 2 n [ d n ] = τ q v 2 n ∘ d n . D_uF_1^n[d^n]=0,\qquad D_uF_2^n[d^n]=\tau qv_2^n\circ d^n. D u F 1 n [ d n ] = 0 , D u F 2 n [ d n ] = τ q v 2 n ∘ d n .
Thus
D u L I [ d ] = τ ∑ n = 0 N − 1 ⟨ λ u n − η q v 2 n , d n ⟩ h − τ q ∑ n = 0 N − 1 ⟨ p 2 n + 1 , v 2 n ∘ d n ⟩ h = τ ∑ n = 0 N − 1 ⟨ λ u n − η q v 2 n − q v 2 n ∘ p 2 n + 1 , d n ⟩ h . D_u\mathcal{L}^I[d]=\tau\sum_{n=0}^{N-1}\langle\lambda u^n-\eta qv_2^n,d^n\rangle_h-\tau q\sum_{n=0}^{N-1}\langle p_2^{n+1},v_2^n\circ d^n\rangle_h=\tau\sum_{n=0}^{N-1}\langle\lambda u^n-\eta qv_2^n-qv_2^n\circ p_2^{n+1},d^n\rangle_h. D u L I [ d ] = τ n = 0 ∑ N − 1 ⟨ λ u n − η q v 2 n , d n ⟩ h − τ q n = 0 ∑ N − 1 ⟨ p 2 n + 1 , v 2 n ∘ d n ⟩ h = τ n = 0 ∑ N − 1 ⟨ λ u n − η q v 2 n − q v 2 n ∘ p 2 n + 1 , d n ⟩ h .
This gives (50) . Differentiating F i n ( v ( u ) , u ) = 0 F_i^n(v(u),u)=0 F i n ( v ( u ) , u ) = 0 gives (51) , while differentiating c i n = 0 c_i^n=0 c i n = 0 gives ξ i N = 0 \xi_i^N=0 ξ i N = 0 and
A 1 ξ 1 n = ξ 1 n + 1 + τ [ γ 1 z 1 n + α z 2 n ∘ p 1 n + 1 + ( − m + α v 2 n ) ∘ ξ 1 n + 1 − β z 2 n ∘ p 2 n + 1 − β v 2 n ∘ ξ 2 n + 1 ] , A 2 ξ 2 n = ξ 2 n + 1 + τ [ γ 2 z 2 n − η q d n + α z 1 n ∘ p 1 n + 1 + α v 1 n ∘ ξ 1 n + 1 + ( − β z 1 n − q d n ) ∘ p 2 n + 1 + ( r − β v 1 n − q u n ) ∘ ξ 2 n + 1 ] . (53) \begin{aligned}
A_1\xi_1^n&=\xi_1^{n+1}+\tau\left[\gamma_1z_1^n+\alpha z_2^n\circ p_1^{n+1}+(-m+\alpha v_2^n)\circ\xi_1^{n+1}-\beta z_2^n\circ p_2^{n+1}-\beta v_2^n\circ\xi_2^{n+1}\right],\\
A_2\xi_2^n&=\xi_2^{n+1}+\tau\left[\gamma_2z_2^n-\eta qd^n+\alpha z_1^n\circ p_1^{n+1}+\alpha v_1^n\circ\xi_1^{n+1}\right.\\
&\qquad\left.+(-\beta z_1^n-qd^n)\circ p_2^{n+1}+(r-\beta v_1^n-qu^n)\circ\xi_2^{n+1}\right]. \tag*{(53)}
\end{aligned} A 1 ξ 1 n A 2 ξ 2 n = ξ 1 n + 1 + τ [ γ 1 z 1 n + α z 2 n ∘ p 1 n + 1 + ( − m + α v 2 n ) ∘ ξ 1 n + 1 − β z 2 n ∘ p 2 n + 1 − β v 2 n ∘ ξ 2 n + 1 ] , = ξ 2 n + 1 + τ [ γ 2 z 2 n − η q d n + α z 1 n ∘ p 1 n + 1 + α v 1 n ∘ ξ 1 n + 1 + ( − β z 1 n − q d n ) ∘ p 2 n + 1 + ( r − β v 1 n − q u n ) ∘ ξ 2 n + 1 ] . (53)
Then
D ( v 2 n ∘ p 2 n + 1 ) [ d ] = z 2 n ∘ p 2 n + 1 + v 2 n ∘ ξ 2 n + 1 , D(v_2^n\circ p_2^{n+1})[d]=z_2^n\circ p_2^{n+1}+v_2^n\circ\xi_2^{n+1}, D ( v 2 n ∘ p 2 n + 1 ) [ d ] = z 2 n ∘ p 2 n + 1 + v 2 n ∘ ξ 2 n + 1 ,
so (50) gives (52) .
Positive step and endpoint-dependent harvest
Define G i n ∈ R M + 1 G_i^n \in\mathbb{R}^{M+1} G i n ∈ R M + 1 by
G 2 n : = C v 2 n + 1 − ( 1 + τ r ) v 2 n , G 1 n : = B v 1 n + 1 − v 1 n − τ α v 1 n ∘ v 2 n + 1 . G_2^n := C v_2^{n+1} - (1+\tau r)v_2^n,\qquad G_1^n := Bv_1^{n+1} - v_1^n - \tau\alpha v_1^n \circ v_2^{n+1}. G 2 n := C v 2 n + 1 − ( 1 + τ r ) v 2 n , G 1 n := B v 1 n + 1 − v 1 n − τ α v 1 n ∘ v 2 n + 1 .
and set
L + , ζ : Y h , τ × R N ( M + 1 ) × R 2 N ( M + 1 ) ⟶ R \mathcal{L}^{+,\zeta}:\mathcal{Y}_{h,\tau}\times\mathbb{R}^{N(M+1)}\times\mathbb{R}^{2N(M+1)}\longrightarrow\mathbb{R} L + , ζ : Y h , τ × R N ( M + 1 ) × R 2 N ( M + 1 ) ⟶ R
( v , u , p ) ⟼ J h , τ ζ ( v , u ) − ∑ n = 0 N − 1 ∑ i = 1 2 ⟨ p i n + 1 , G i n ⟩ h . (v,u,p)\longmapsto J_{h,\tau}^{\zeta}(v,u)-\sum_{n=0}^{N-1}\sum_{i=1}^{2}\langle p_i^{n+1},G_i^n\rangle_h. ( v , u , p ) ⟼ J h , τ ζ ( v , u ) − n = 0 ∑ N − 1 i = 1 ∑ 2 ⟨ p i n + 1 , G i n ⟩ h .
For arbitrary variations, the positive residuals give
D G 2 n [ δ v , d ] = C δ v 2 n + 1 − ( 1 + τ r ) δ v 2 n + τ ( β δ v 1 n + q d n ) ∘ v 2 n + 1 , DG_2^n[\delta v,d]=C\delta v_2^{n+1}-(1+\tau r)\delta v_2^n+\tau(\beta\delta v_1^n+qd^n)\circ v_2^{n+1}, D G 2 n [ δ v , d ] = C δ v 2 n + 1 − ( 1 + τ r ) δ v 2 n + τ ( β δ v 1 n + q d n ) ∘ v 2 n + 1 ,
D G 1 n [ δ v , d ] = B δ v 1 n + 1 − δ v 1 n − τ α ( δ v 1 n ∘ v 2 n + 1 + v 1 n ∘ δ v 2 n + 1 ) . DG_1^n[\delta v,d]=B\delta v_1^{n+1}-\delta v_1^n-\tau\alpha(\delta v_1^n\circ v_2^{n+1}+v_1^n\circ\delta v_2^{n+1}). D G 1 n [ δ v , d ] = B δ v 1 n + 1 − δ v 1 n − τ α ( δ v 1 n ∘ v 2 n + 1 + v 1 n ∘ δ v 2 n + 1 ) .
The harvest variation with respect to the prey trajectory is
− η q τ ∑ n = 0 N − 1 ⟨ u n , ( 1 − ζ ) δ v 2 n + ζ δ v 2 n + 1 ⟩ h = − η q τ ∑ k = 1 N − 1 ⟨ ( 1 − ζ ) u k + ζ u k − 1 , δ v 2 k ⟩ h − η q τ ζ ⟨ u N − 1 , δ v 2 N ⟩ h , -\eta q\tau\sum_{n=0}^{N-1}\langle u^n,(1-\zeta)\delta v_2^n+\zeta\delta v_2^{n+1}\rangle_h
=-\eta q\tau\sum_{k=1}^{N-1}\langle(1-\zeta)u^k+\zeta u^{k-1},\delta v_2^k\rangle_h-\eta q\tau\zeta\langle u^{N-1},\delta v_2^N\rangle_h, − η q τ n = 0 ∑ N − 1 ⟨ u n , ( 1 − ζ ) δ v 2 n + ζ δ v 2 n + 1 ⟩ h = − η q τ k = 1 ∑ N − 1 ⟨( 1 − ζ ) u k + ζ u k − 1 , δ v 2 k ⟩ h − η q τ ζ ⟨ u N − 1 , δ v 2 N ⟩ h ,
because δ v 2 0 = 0 \delta v_2^0=0 δ v 2 0 = 0 . This identity accounts for both time levels in the adjoint source and for the terminal term. The coefficients of the terminal variations are
c 1 N = − B p 1 N , c 2 N = − τ ζ η q u N − 1 − C N − 1 p 2 N + τ α v 1 N − 1 ∘ p 1 N . c_1^N=-Bp_1^N,\qquad c_2^N=-\tau\zeta\eta q u^{N-1}-C^{N-1}p_2^N+\tau\alpha v_1^{N-1}\circ p_1^N. c 1 N = − B p 1 N , c 2 N = − τ ζ η q u N − 1 − C N − 1 p 2 N + τ α v 1 N − 1 ∘ p 1 N .
Setting these to zero gives (10) . For 1 ≤ k < N 1\leq k<N 1 ≤ k < N ,
c 1 k = τ γ 1 ( v 1 k − v 1 d , k ) − B p 1 k + p 1 k + 1 + τ v 2 k + 1 ∘ ( α p 1 k + 1 − β p 2 k + 1 ) , c_1^k=\tau\gamma_1(v_1^k-v_1^{d,k})-Bp_1^k+p_1^{k+1}+\tau v_2^{k+1}\circ(\alpha p_1^{k+1}-\beta p_2^{k+1}), c 1 k = τ γ 1 ( v 1 k − v 1 d , k ) − B p 1 k + p 1 k + 1 + τ v 2 k + 1 ∘ ( α p 1 k + 1 − β p 2 k + 1 ) ,
and
c 2 k = τ γ 2 ( v 2 k − v 2 d , k ) − τ η q [ ( 1 − ζ ) u k + ζ u k − 1 ] − C k − 1 p 2 k + ( 1 + τ r ) p 2 k + 1 + τ α v 1 k − 1 ∘ p 1 k . c_2^k=\tau\gamma_2(v_2^k-v_2^{d,k})-\tau\eta q[(1-\zeta)u^k+\zeta u^{k-1}]-C^{k-1}p_2^k+(1+\tau r)p_2^{k+1}+\tau\alpha v_1^{k-1}\circ p_1^k. c 2 k = τ γ 2 ( v 2 k − v 2 d , k ) − τ η q [( 1 − ζ ) u k + ζ u k − 1 ] − C k − 1 p 2 k + ( 1 + τ r ) p 2 k + 1 + τ α v 1 k − 1 ∘ p 1 k .
The coefficient c 2 k c_2^k c 2 k includes harvest at both steps k k k and k − 1 k-1 k − 1 . Setting c i k = 0 c_i^k=0 c i k = 0 gives (11) . Holding the states fixed, we have
D u G 2 n [ d n ] = τ q d n ∘ v 2 n + 1 , D u G 1 n [ d n ] = 0 , D_uG_2^n[d^n]=\tau qd^n\circ v_2^{n+1},\qquad D_uG_1^n[d^n]=0, D u G 2 n [ d n ] = τ q d n ∘ v 2 n + 1 , D u G 1 n [ d n ] = 0 ,
so
D u L + , ζ [ d ] = τ ∑ n = 0 N − 1 ⟨ λ u n − η q [ ( 1 − ζ ) v 2 n + ζ v 2 n + 1 ] , d n ⟩ h − τ q ∑ n = 0 N − 1 ⟨ v 2 n + 1 ∘ p 2 n + 1 , d n ⟩ h . D_u\mathcal{L}^{+,\zeta}[d]=\tau\sum_{n=0}^{N-1}\langle\lambda u^n-\eta q[(1-\zeta)v_2^n+\zeta v_2^{n+1}],d^n\rangle_h-\tau q\sum_{n=0}^{N-1}\langle v_2^{n+1}\circ p_2^{n+1},d^n\rangle_h. D u L + , ζ [ d ] = τ n = 0 ∑ N − 1 ⟨ λ u n − η q [( 1 − ζ ) v 2 n + ζ v 2 n + 1 ] , d n ⟩ h − τ q n = 0 ∑ N − 1 ⟨ v 2 n + 1 ∘ p 2 n + 1 , d n ⟩ h .
This proves (12) . For d ∈ R N ( M + 1 ) d\in\mathbb{R}^{N(M+1)} d ∈ R N ( M + 1 ) , the product rule gives
D ( C n v 2 n + 1 ) [ d ] = C n z 2 n + 1 + τ ( β z 1 n + q d n ) ∘ v 2 n + 1 , D(C^n v_2^{n+1})[d]=C^n z_2^{n+1}+\tau(\beta z_1^n+qd^n)\circ v_2^{n+1}, D ( C n v 2 n + 1 ) [ d ] = C n z 2 n + 1 + τ ( β z 1 n + q d n ) ∘ v 2 n + 1 ,
and
D ( C k − 1 p 2 k ) [ d ] = C k − 1 ξ 2 k + τ ( β z 1 k − 1 + q d k − 1 ) ∘ p 2 k , D(C^{k-1}p_2^k)[d]=C^{k-1}\xi_2^k+\tau(\beta z_1^{k-1}+qd^{k-1})\circ p_2^k, D ( C k − 1 p 2 k ) [ d ] = C k − 1 ξ 2 k + τ ( β z 1 k − 1 + q d k − 1 ) ∘ p 2 k ,
and
D ( v 1 n ∘ v 2 n + 1 ) [ d ] = z 1 n ∘ v 2 n + 1 + v 1 n ∘ z 2 n + 1 . D(v_1^n\circ v_2^{n+1})[d]=z_1^n\circ v_2^{n+1}+v_1^n\circ z_2^{n+1}. D ( v 1 n ∘ v 2 n + 1 ) [ d ] = z 1 n ∘ v 2 n + 1 + v 1 n ∘ z 2 n + 1 .
These identities yield (54) and (55) –(56) . Differentiating (12) gives (57) .
Positive tangent and tangent-adjoint equations
For d ∈ R N ( M + 1 ) d \in\mathbb{R}^{N(M+1)} d ∈ R N ( M + 1 ) , set z i 0 = 0 z_i^0=0 z i 0 = 0 . Since
D C n ( u ) [ d ] = τ diag ( β z 1 n + q d n ) , DC^n(u)[d]=\tau\operatorname{diag}(\beta z_1^n+qd^n), D C n ( u ) [ d ] = τ diag ( β z 1 n + q d n ) ,
the tangent satisfies
{ C 2 n z 2 n + 1 = ( 1 + τ r ) z 2 n − τ ( β z 1 n + q d n ) ∘ v 2 n + 1 , B z 1 n + 1 = z 1 n + τ α ( z 1 n ∘ v 2 n + 1 + v 1 n ∘ z 2 n + 1 ) . (54) \begin{cases}
C_2^n z_2^{n+1}=(1+\tau r)z_2^n-\tau(\beta z_1^n+qd^n)\circ v_2^{n+1},\\
Bz_1^{n+1}=z_1^n+\tau\alpha(z_1^n\circ v_2^{n+1}+v_1^n\circ z_2^{n+1}). \tag*{(54)}
\end{cases} { C 2 n z 2 n + 1 = ( 1 + τ r ) z 2 n − τ ( β z 1 n + q d n ) ∘ v 2 n + 1 , B z 1 n + 1 = z 1 n + τ α ( z 1 n ∘ v 2 n + 1 + v 1 n ∘ z 2 n + 1 ) . (54)
Differentiating the terminal equation gives
ξ 1 N = 0 , C 2 N − 1 ξ 2 N = − τ ζ η q d N − 1 − τ ( β z 1 N − 1 + q d N − 1 ) ∘ p 2 N . (55) \xi_1^N=0,\qquad C_2^{N-1}\xi_2^N=-\tau\zeta\eta qd^{N-1}-\tau(\beta z_1^{N-1}+qd^{N-1})\circ p_2^N. \tag*{(55)} ξ 1 N = 0 , C 2 N − 1 ξ 2 N = − τ ζ η q d N − 1 − τ ( β z 1 N − 1 + q d N − 1 ) ∘ p 2 N . (55)
The backward derivative is
{ B ξ 1 k = ξ 1 k + 1 + τ [ γ 1 z 1 k + z 2 k + 1 ∘ ( α p 1 k + 1 − β p 2 k + 1 ) + v 2 k + 1 ∘ ( α ξ 1 k + 1 − β ξ 2 k + 1 ) ] , C k − 1 ξ 2 k = ( 1 + τ r ) ξ 2 k + 1 + τ [ γ 2 z 2 k − η q ( ( 1 − ζ ) d k + ζ d k − 1 ) + α z 1 k − 1 ∘ p 1 k + α v 1 k − 1 ∘ ξ 1 k − ( β z 1 k − 1 + q d k − 1 ) ∘ p 2 k ] . (56) \begin{cases}
B\xi_1^k=\xi_1^{k+1}+\tau\left[\gamma_1z_1^k+z_2^{k+1}\circ(\alpha p_1^{k+1}-\beta p_2^{k+1})+v_2^{k+1}\circ(\alpha\xi_1^{k+1}-\beta\xi_2^{k+1})\right],\\
C^{k-1}\xi_2^k=(1+\tau r)\xi_2^{k+1}+\tau\left[\gamma_2z_2^k-\eta q((1-\zeta)d^k+\zeta d^{k-1})+\alpha z_1^{k-1}\circ p_1^k+\alpha v_1^{k-1}\circ\xi_1^k-(\beta z_1^{k-1}+qd^{k-1})\circ p_2^k\right]. \tag*{(56)}
\end{cases} { B ξ 1 k = ξ 1 k + 1 + τ [ γ 1 z 1 k + z 2 k + 1 ∘ ( α p 1 k + 1 − β p 2 k + 1 ) + v 2 k + 1 ∘ ( α ξ 1 k + 1 − β ξ 2 k + 1 ) ] , C k − 1 ξ 2 k = ( 1 + τ r ) ξ 2 k + 1 + τ [ γ 2 z 2 k − η q (( 1 − ζ ) d k + ζ d k − 1 ) + α z 1 k − 1 ∘ p 1 k + α v 1 k − 1 ∘ ξ 1 k − ( β z 1 k − 1 + q d k − 1 ) ∘ p 2 k ] . (56)
Then
( H + ∗ ζ d ) n = λ d n − η q [ ( 1 − ζ ) z 2 n + ζ z 2 n + 1 ] − q ( z 2 n + 1 ∘ p 2 n + 1 + v 2 n + 1 ∘ ξ 2 n + 1 ) . (57) (H_+^\ast\zeta d)^n=\lambda d^n-\eta q[(1-\zeta)z_2^n+\zeta z_2^{n+1}]-q(z_2^{n+1}\circ p_2^{n+1}+v_2^{n+1}\circ\xi_2^{n+1}). \tag*{(57)} ( H + ∗ ζ d ) n = λ d n − η q [( 1 − ζ ) z 2 n + ζ z 2 n + 1 ] − q ( z 2 n + 1 ∘ p 2 n + 1 + v 2 n + 1 ∘ ξ 2 n + 1 ) . (57)
Second variations of the discrete costs
Fix either state scheme and take control directions d , e ∈ R N ( M + 1 ) d,e\in\mathbb{R}^{N(M+1)} d , e ∈ R N ( M + 1 ) . Write z [ d ] = D S ( u ) [ d ] z[d]=DS(u)[d] z [ d ] = D S ( u ) [ d ] , z [ e ] = D S ( u ) [ e ] z[e]=DS(u)[e] z [ e ] = D S ( u ) [ e ] and s [ d , e ] = D 2 S ( u ) [ d , e ] s[d,e]=D^2S(u)[d,e] s [ d , e ] = D 2 S ( u ) [ d , e ] . All three variations have zero initial value. For IMEX, differentiating (51) in direction e e e gives
{ A 1 n s 1 n + 1 = s 1 n + τ [ ( − m + α v 2 n ) ∘ s 1 n + α v 1 n ∘ s 2 n ] + τ α ( z 1 [ d ] n ∘ z 2 [ e ] n + z 1 [ e ] n ∘ z 2 [ d ] n ) , A 2 n s 2 n + 1 = s 2 n + τ [ − β v 2 n ∘ s 1 n + ( r − β v 1 n − q u n ) ∘ s 2 n ] − τ β ( z 1 [ d ] n ∘ z 2 [ e ] n + z 1 [ e ] n ∘ z 2 [ d ] n ) − τ q ( d n ∘ z 2 [ e ] n + e n ∘ z 2 [ d ] n ) . \begin{cases}
A_1^n s_1^{n+1}=s_1^n+\tau[(-m+\alpha v_2^n)\circ s_1^n+\alpha v_1^n\circ s_2^n]+\tau\alpha(z_1[d]^n\circ z_2[e]^n+z_1[e]^n\circ z_2[d]^n),\\
A_2^n s_2^{n+1}=s_2^n+\tau[-\beta v_2^n\circ s_1^n+(r-\beta v_1^n-qu^n)\circ s_2^n]\\
\qquad-\tau\beta(z_1[d]^n\circ z_2[e]^n+z_1[e]^n\circ z_2[d]^n)-\tau q(d^n\circ z_2[e]^n+e^n\circ z_2[d]^n).
\end{cases} ⎩ ⎨ ⎧ A 1 n s 1 n + 1 = s 1 n + τ [( − m + α v 2 n ) ∘ s 1 n + α v 1 n ∘ s 2 n ] + τ α ( z 1 [ d ] n ∘ z 2 [ e ] n + z 1 [ e ] n ∘ z 2 [ d ] n ) , A 2 n s 2 n + 1 = s 2 n + τ [ − β v 2 n ∘ s 1 n + ( r − β v 1 n − q u n ) ∘ s 2 n ] − τ β ( z 1 [ d ] n ∘ z 2 [ e ] n + z 1 [ e ] n ∘ z 2 [ d ] n ) − τ q ( d n ∘ z 2 [ e ] n + e n ∘ z 2 [ d ] n ) .
For the positive scheme, differentiating (54) instead gives
{ C 2 n s 2 n + 1 = ( 1 + τ r ) s 2 n − τ β s 1 n ∘ v 2 n + 1 − τ ( β z 1 [ d ] n + q d n ) ∘ z 2 [ e ] n + 1 − τ ( β z 1 [ e ] n + q e n ) ∘ z 2 [ d ] n + 1 , B s 1 n + 1 = s 1 n + τ α ( s 1 n ∘ v 2 n + 1 + v 1 n ∘ s 2 n + 1 ) + τ α ( z 1 [ d ] n ∘ z 2 [ e ] n + 1 + z 1 [ e ] n ∘ z 2 [ d ] n + 1 ) . \begin{cases}
C_2^n s_2^{n+1}=(1+\tau r)s_2^n-\tau\beta s_1^n\circ v_2^{n+1}-\tau(\beta z_1[d]^n+qd^n)\circ z_2[e]^{n+1}-\tau(\beta z_1[e]^n+qe^n)\circ z_2[d]^{n+1},\\
Bs_1^{n+1}=s_1^n+\tau\alpha(s_1^n\circ v_2^{n+1}+v_1^n\circ s_2^{n+1})+\tau\alpha(z_1[d]^n\circ z_2[e]^{n+1}+z_1[e]^n\circ z_2[d]^{n+1}).
\end{cases} { C 2 n s 2 n + 1 = ( 1 + τ r ) s 2 n − τ β s 1 n ∘ v 2 n + 1 − τ ( β z 1 [ d ] n + q d n ) ∘ z 2 [ e ] n + 1 − τ ( β z 1 [ e ] n + q e n ) ∘ z 2 [ d ] n + 1 , B s 1 n + 1 = s 1 n + τ α ( s 1 n ∘ v 2 n + 1 + v 1 n ∘ s 2 n + 1 ) + τ α ( z 1 [ d ] n ∘ z 2 [ e ] n + 1 + z 1 [ e ] n ∘ z 2 [ d ] n + 1 ) .
The cost derivatives follow without differentiating the adjoint. With ζ = 0 \zeta=0 ζ = 0 for IMEX, the first derivative is
D j ( u ) [ d ] = λ ⟨ u , d ⟩ h , τ + τ ∑ n = 0 N − 1 ∑ i = 1 2 γ i ⟨ v i n − v i d , n , z i [ d ] n ⟩ h − η q τ ∑ n = 0 N − 1 [ ⟨ d n , ( 1 − ζ ) v 2 n + ζ v 2 n + 1 ⟩ h + ⟨ u n , ( 1 − ζ ) z 2 [ d ] n + ζ z 2 [ d ] n + 1 ⟩ h ] . Dj(u)[d]=\lambda\langle u,d\rangle_{h,\tau}+\tau\sum_{n=0}^{N-1}\sum_{i=1}^{2}\gamma_i\langle v_i^n-v_i^{d,n},z_i[d]^n\rangle_h-\eta q\tau\sum_{n=0}^{N-1}\left[\langle d^n,(1-\zeta)v_2^n+\zeta v_2^{n+1}\rangle_h+\langle u^n,(1-\zeta)z_2[d]^n+\zeta z_2[d]^{n+1}\rangle_h\right]. D j ( u ) [ d ] = λ ⟨ u , d ⟩ h , τ + τ n = 0 ∑ N − 1 i = 1 ∑ 2 γ i ⟨ v i n − v i d , n , z i [ d ] n ⟩ h − η q τ n = 0 ∑ N − 1 [ ⟨ d n , ( 1 − ζ ) v 2 n + ζ v 2 n + 1 ⟩ h + ⟨ u n , ( 1 − ζ ) z 2 [ d ] n + ζ z 2 [ d ] n + 1 ⟩ h ] .
Differentiating each factor gives
D 2 j ( u ) [ d , e ] = λ ⟨ d , e ⟩ h , τ + τ ∑ n = 0 N − 1 ∑ i = 1 2 γ i [ ⟨ z i [ d ] n , z i [ e ] n ⟩ h + ⟨ v i n − v i d , n , s i [ d , e ] n ⟩ h ] − η q τ ∑ n = 0 N − 1 [ ⟨ d n , ( 1 − ζ ) z 2 [ e ] n + ζ z 2 [ e ] n + 1 ⟩ h + ⟨ e n , ( 1 − ζ ) z 2 [ d ] n + ζ z 2 [ d ] n + 1 ⟩ h + ⟨ u n , ( 1 − ζ ) s 2 [ d , e ] n + ζ s 2 [ d , e ] n + 1 ⟩ h ] . \begin{aligned}
D^2j(u)[d,e]={}&\lambda\langle d,e\rangle_{h,\tau}+\tau\sum_{n=0}^{N-1}\sum_{i=1}^{2}\gamma_i\left[\langle z_i[d]^n,z_i[e]^n\rangle_h+\langle v_i^n-v_i^{d,n},s_i[d,e]^n\rangle_h\right]\\
&-\eta q\tau\sum_{n=0}^{N-1}\left[\langle d^n,(1-\zeta)z_2[e]^n+\zeta z_2[e]^{n+1}\rangle_h\right.\\
&\qquad+\langle e^n,(1-\zeta)z_2[d]^n+\zeta z_2[d]^{n+1}\rangle_h\\
&\qquad\left.+\langle u^n,(1-\zeta)s_2[d,e]^n+\zeta s_2[d,e]^{n+1}\rangle_h\right].
\end{aligned} D 2 j ( u ) [ d , e ] = λ ⟨ d , e ⟩ h , τ + τ n = 0 ∑ N − 1 i = 1 ∑ 2 γ i [ ⟨ z i [ d ] n , z i [ e ] n ⟩ h + ⟨ v i n − v i d , n , s i [ d , e ] n ⟩ h ] − η q τ n = 0 ∑ N − 1 [ ⟨ d n , ( 1 − ζ ) z 2 [ e ] n + ζ z 2 [ e ] n + 1 ⟩ h + ⟨ e n , ( 1 − ζ ) z 2 [ d ] n + ζ z 2 [ d ] n + 1 ⟩ h + ⟨ u n , ( 1 − ζ ) s 2 [ d , e ] n + ζ s 2 [ d , e ] n + 1 ⟩ h ] .
Both second-variation systems are symmetric in d , e d,e d , e , so uniqueness gives s [ d , e ] = s [ e , d ] s[d,e]=s[e,d] s [ d , e ] = s [ e , d ] . Thus the bilinear form is symmetric. Differentiating D j ( u ) [ e ] = ⟨ g ( u ) , e ⟩ h , τ Dj(u)[e]=\langle g(u),e\rangle_{h,\tau} D j ( u ) [ e ] = ⟨ g ( u ) , e ⟩ h , τ also gives
D 2 j ( u ) [ d , e ] = ⟨ D g ( u ) [ d ] , e ⟩ h , τ = ⟨ H ( u ) d , e ⟩ h , τ , D^2j(u)[d,e]=\langle Dg(u)[d],e\rangle_{h,\tau}=\langle H(u)d,e\rangle_{h,\tau}, D 2 j ( u ) [ d , e ] = ⟨ D g ( u ) [ d ] , e ⟩ h , τ = ⟨ H ( u ) d , e ⟩ h , τ ,
which identifies the same Hessian as (50) and (55).
Symmetry and local smoothness
For an invertible matrix A ∈ R s × s A\in\mathbb{R}^{s\times s} A ∈ R s × s and perturbations E , F ∈ R s × s E,F\in\mathbb{R}^{s\times s} E , F ∈ R s × s , differentiating A A − 1 = I s AA^{-1}=I_s A A − 1 = I s gives
D ( A − 1 ) [ E ] = − A − 1 E A − 1 , D 2 ( A − 1 ) [ E , F ] = A − 1 E A − 1 F A − 1 + A − 1 F A − 1 E A − 1 . D(A^{-1})[E]=-A^{-1}EA^{-1},\qquad D^2(A^{-1})[E,F]=A^{-1}EA^{-1}FA^{-1}+A^{-1}FA^{-1}EA^{-1}. D ( A − 1 ) [ E ] = − A − 1 E A − 1 , D 2 ( A − 1 ) [ E , F ] = A − 1 E A − 1 F A − 1 + A − 1 F A − 1 E A − 1 .
Each IMEX step combines polynomials with fixed matrix inverses. For the positive scheme, the loss matrices remain invertible near the admissible box, where their inverses depend smoothly on the entries. All three reduced costs are therefore smooth near U a d , h , τ \mathcal{U}_{\mathrm{ad},h,\tau} U ad , h , τ on each fixed grid. The weighted gradient and the Euclidean derivatives are related by
D j ( u ) [ d ] = g ( u ) T W d = ∇ j ( u ) T d , ∇ j ( u ) = W g ( u ) , ∇ 2 j ( u ) = W H ( u ) . Dj(u)[d]=g(u)^T Wd=\nabla j(u)^T d,\qquad\nabla j(u)=Wg(u),\qquad\nabla^2j(u)=WH(u). D j ( u ) [ d ] = g ( u ) T W d = ∇ j ( u ) T d , ∇ j ( u ) = W g ( u ) , ∇ 2 j ( u ) = W H ( u ) .
Since W W W is constant,
D 2 j ( u ) [ d , e ] = e T W D g ( u ) d = e T W H ( u ) d = D 2 j ( u ) [ e , d ] . D^2j(u)[d,e]=e^TWDg(u)d=e^TWH(u)d=D^2j(u)[e,d]. D 2 j ( u ) [ d , e ] = e T W D g ( u ) d = e T W H ( u ) d = D 2 j ( u ) [ e , d ] .
Thus
H ( u ) T W = W H ( u ) , ( W 1 / 2 H ( u ) W − 1 / 2 ) T = W 1 / 2 H ( u ) W − 1 / 2 . H(u)^T W=WH(u),\qquad\left(W^{1/2}H(u)W^{-1/2}\right)^T=W^{1/2}H(u)W^{-1/2}. H ( u ) T W = W H ( u ) , ( W 1/2 H ( u ) W − 1/2 ) T = W 1/2 H ( u ) W − 1/2 .
These identities show that H H H is self-adjoint in the weighted inner product and W 1 / 2 H W − 1 / 2 W^{1/2}HW^{-1/2} W 1/2 H W − 1/2 is symmetric. This proves Proposition 4.1.
Compactness and variational convergence proofs
Proof of Lemma 5.1
Both schemes have the form
D τ v i n − μ i L h v i n + 1 = f i n . (58) D_\tau v_i^n-\mu_iL_hv_i^{n+1}=f_i^n. \tag*{(58)} D τ v i n − μ i L h v i n + 1 = f i n . (58)
For IMEX, f i n : = f i ( v n , u n ) ∈ R M + 1 f_i^n:=f_i(v^n,u^n)\in\mathbb{R}^{M+1} f i n := f i ( v n , u n ) ∈ R M + 1 . For the positive scheme,
f 1 n : = − m v 1 n + 1 + α v 1 n ∘ v 2 n + 1 , f 2 n : = r v 2 n − ( β v 1 n + q u n ) ∘ v 2 n + 1 . f_1^n:=-mv_1^{n+1}+\alpha v_1^n\circ v_2^{n+1},\qquad f_2^n:=rv_2^n-(\beta v_1^n+qu^n)\circ v_2^{n+1}. f 1 n := − m v 1 n + 1 + α v 1 n ∘ v 2 n + 1 , f 2 n := r v 2 n − ( β v 1 n + q u n ) ∘ v 2 n + 1 .
For both schemes, Proposition 3.2 gives
∥ f 1 n ∥ ∞ ≤ m K 1 + α K 1 K 2 , ∥ f 2 n ∥ ∞ ≤ r K 2 + ( β K 1 + q U max ) K 2 . \|f_1^n\|_\infty\le mK_1+\alpha K_1K_2,\qquad\|f_2^n\|_\infty\le rK_2+(\beta K_1+qU_{\max})K_2. ∥ f 1 n ∥ ∞ ≤ m K 1 + α K 1 K 2 , ∥ f 2 n ∥ ∞ ≤ r K 2 + ( β K 1 + q U m a x ) K 2 .
Since ∑ j = 0 M w j = 1 \sum_{j=0}^{M}w_j=1 ∑ j = 0 M w j = 1 , we have
max i ∈ { 1 , 2 } 0 ≤ n < N ∥ f i n ∥ ∞ ≤ C , τ ∑ n = 0 N − 1 ∥ f i n ∥ h 2 ≤ C T . \max_{\substack{i\in\{1,2\}\\0\le n<N}}\|f_i^n\|_\infty\le C,\qquad\tau\sum_{n=0}^{N-1}\|f_i^n\|_h^2\le CT. i ∈ { 1 , 2 } 0 ≤ n < N max ∥ f i n ∥ ∞ ≤ C , τ n = 0 ∑ N − 1 ∥ f i n ∥ h 2 ≤ C T .
These estimates use only the bounds on u u u . By symmetry of a h a_h a h and the definition of D τ D_\tau D τ ,
2 τ a h ( v i n + 1 , D τ v i n ) = ∣ v i n + 1 ∣ 1 , h 2 − ∣ v i n ∣ 1 , h 2 + τ 2 ∣ D τ v i n ∣ 1 , h 2 . (59) 2\tau a_h(v_i^{n+1},D_\tau v_i^n)=|v_i^{n+1}|_{1,h}^2-|v_i^n|_{1,h}^2+\tau^2|D_\tau v_i^n|_{1,h}^2. \tag*{(59)} 2 τ a h ( v i n + 1 , D τ v i n ) = ∣ v i n + 1 ∣ 1 , h 2 − ∣ v i n ∣ 1 , h 2 + τ 2 ∣ D τ v i n ∣ 1 , h 2 . (59)
Test (58) with 2 τ D τ v i n 2\tau D_\tau v_i^n 2 τ D τ v i n . Young’s inequality gives
2 τ ∥ D τ v i n ∥ h 2 + 2 τ μ i a h ( v i n + 1 , D τ v i n ) = 2 τ ⟨ f i n , D τ v i n ⟩ h ≤ 2 τ ∥ f i n ∥ h ∥ D τ v i n ∥ h ≤ τ ∥ f i n ∥ h 2 + τ ∥ D τ v i n ∥ h 2 . 2\tau\lVert D_\tau v_i^n\rVert_h^2+2\tau\mu_i a_h(v_i^{n+1},D_\tau v_i^n)=2\tau\langle f_i^n,D_\tau v_i^n\rangle_h
\leq2\tau\lVert f_i^n\rVert_h\lVert D_\tau v_i^n\rVert_h
\leq\tau\lVert f_i^n\rVert_h^2+\tau\lVert D_\tau v_i^n\rVert_h^2. 2 τ ∥ D τ v i n ∥ h 2 + 2 τ μ i a h ( v i n + 1 , D τ v i n ) = 2 τ ⟨ f i n , D τ v i n ⟩ h ≤ 2 τ ∥ f i n ∥ h ∥ D τ v i n ∥ h ≤ τ ∥ f i n ∥ h 2 + τ ∥ D τ v i n ∥ h 2 .
Subtract τ ∥ D τ v i n ∥ h 2 \tau\lVert D_\tau v_i^n\rVert_h^2 τ ∥ D τ v i n ∥ h 2 from the resulting inequality, substitute (59) , and sum over 0 ≤ n < ℓ 0\leq n<\ell 0 ≤ n < ℓ . The spatial energies telescope:
τ ∑ n = 0 ℓ − 1 ∥ D τ v i n ∥ h 2 + μ i ∥ v i ℓ ∥ 1 , h 2 + μ i τ 2 ∑ n = 0 ℓ − 1 ∣ D τ v i n ∣ 1 , h 2 ≤ μ i ∥ v i 0 ∥ 1 , h 2 + τ ∑ n = 0 ℓ − 1 ∥ f i n ∥ h 2 . (60) \tau\sum_{n=0}^{\ell-1}\lVert D_\tau v_i^n\rVert_h^2+\mu_i\lVert v_i^\ell\rVert_{1,h}^2+\mu_i\tau^2\sum_{n=0}^{\ell-1}\lvert D_\tau v_i^n\rvert_{1,h}^2\leq\mu_i\lVert v_i^0\rVert_{1,h}^2+\tau\sum_{n=0}^{\ell-1}\lVert f_i^n\rVert_h^2. \tag*{(60)} τ n = 0 ∑ ℓ − 1 ∥ D τ v i n ∥ h 2 + μ i ∥ v i ℓ ∥ 1 , h 2 + μ i τ 2 n = 0 ∑ ℓ − 1 ∣ D τ v i n ∣ 1 , h 2 ≤ μ i ∥ v i 0 ∥ 1 , h 2 + τ n = 0 ∑ ℓ − 1 ∥ f i n ∥ h 2 . (60)
For nodal initial data, Cauchy–Schwarz on each grid interval gives
∣ v i 0 ∣ 1 , h 2 = ∑ j = 0 M − 1 h ∣ 1 h ∫ x j x j + 1 ∂ x v i 0 d x ∣ 2 ≤ ∑ j = 0 M − 1 ∫ x j x j + 1 ∣ ∂ x v i 0 ∣ 2 d x = ∥ ∂ x v i 0 ∥ 2 2 . \lvert v_i^0\rvert_{1,h}^2=\sum_{j=0}^{M-1}h\left\lvert\frac{1}{h}\int_{x_j}^{x_{j+1}}\partial_xv_i^0\,\mathrm{d}x\right\rvert^2\leq\sum_{j=0}^{M-1}\int_{x_j}^{x_{j+1}}\lvert\partial_xv_i^0\rvert^2\,\mathrm{d}x=\lVert\partial_xv_i^0\rVert_2^2. ∣ v i 0 ∣ 1 , h 2 = j = 0 ∑ M − 1 h h 1 ∫ x j x j + 1 ∂ x v i 0 d x 2 ≤ j = 0 ∑ M − 1 ∫ x j x j + 1 ∣ ∂ x v i 0 ∣ 2 d x = ∥ ∂ x v i 0 ∥ 2 2 .
With ∥ I h v i 0 ∥ 2 ≤ K i \lVert\mathcal I_h v_i^0\rVert_2\leq K_i ∥ I h v i 0 ∥ 2 ≤ K i , this gives (14) . The initial-data bound and the forcing estimate give
τ ∑ n = 0 N − 1 ∥ D τ v i n ∥ h 2 ≤ C , max 0 ≤ ℓ ≤ N ∣ v i ℓ ∣ 1 , h 2 ≤ C . \tau\sum_{n=0}^{N-1}\lVert D_\tau v_i^n\rVert_h^2\leq C,\qquad\max_{0\leq\ell\leq N}\lvert v_i^\ell\rvert_{1,h}^2\leq C. τ n = 0 ∑ N − 1 ∥ D τ v i n ∥ h 2 ≤ C , 0 ≤ ℓ ≤ N max ∣ v i ℓ ∣ 1 , h 2 ≤ C .
The constants depend on the initial H 1 H^1 H 1 norm, the uniform reaction bounds, T T T and μ i \mu_i μ i , with no factor involving h − 1 h^{-1} h − 1 or τ − 1 \tau^{-1} τ − 1 . The difference between the time reconstructions satisfies
∥ V i + − V i − ∥ L 2 ( Q ) 2 = τ ∑ n = 0 N − 1 ∥ v i n + 1 − v i n ∥ h 2 = τ 3 ∑ n = 0 N − 1 ∥ D τ v i n ∥ h 2 ≤ C τ 2 . \lVert V_i^+-V_i^-\rVert_{L^2(Q)}^2=\tau\sum_{n=0}^{N-1}\lVert v_i^{n+1}-v_i^n\rVert_h^2=\tau^3\sum_{n=0}^{N-1}\lVert D_\tau v_i^n\rVert_h^2\leq C\tau^2. ∥ V i + − V i − ∥ L 2 ( Q ) 2 = τ n = 0 ∑ N − 1 ∥ v i n + 1 − v i n ∥ h 2 = τ 3 n = 0 ∑ N − 1 ∥ D τ v i n ∥ h 2 ≤ C τ 2 .
For the spatial reconstructions, integrate over both halves of each grid interval:
∫ 0 h / 2 s 2 d s + ∫ h / 2 h ( h − s ) 2 d s = h 3 12 . \int_0^{h/2}s^2\,\mathrm{d}s+\int_{h/2}^h(h-s)^2\,\mathrm{d}s=\frac{h^3}{12}. ∫ 0 h /2 s 2 d s + ∫ h /2 h ( h − s ) 2 d s = 12 h 3 .
Thus
∥ Q h y − I h y ∥ 2 2 = ∑ j = 0 M − 1 ( y j + 1 − y j ) 2 h 2 h 3 12 = h 12 ∑ j = 0 M − 1 ( y j + 1 − y j ) 2 = (45) h 2 12 a h ( y , y ) , \lVert\mathcal Q_hy-\mathcal I_hy\rVert_2^2=\sum_{j=0}^{M-1}\frac{(y_{j+1}-y_j)^2}{h^2}\frac{h^3}{12}=\frac{h}{12}\sum_{j=0}^{M-1}(y_{j+1}-y_j)^2\mathrel{\overset{\text{\scriptsize(45)}}{=}}\frac{h^2}{12}a_h(y,y), ∥ Q h y − I h y ∥ 2 2 = j = 0 ∑ M − 1 h 2 ( y j + 1 − y j ) 2 12 h 3 = 12 h j = 0 ∑ M − 1 ( y j + 1 − y j ) 2 = (45) 12 h 2 a h ( y , y ) ,
which proves (16) .
Proof of Theorem 5.2
For y ∈ R M + 1 y\in\mathbb{R}^{M+1} y ∈ R M + 1 ,
∥ I h y ∥ 2 2 = ∑ j = 0 M − 1 h 3 ( y j 2 + y j y j + 1 + y j + 1 2 ) = (45) ∥ y ∥ h 2 − h 2 6 a h ( y , y ) . \lVert\mathcal I_hy\rVert_2^2=\sum_{j=0}^{M-1}\frac{h}{3}(y_j^2+y_jy_{j+1}+y_{j+1}^2)\mathrel{\overset{\text{\scriptsize(45)}}{=}}\lVert y\rVert_h^2-\frac{h^2}{6}a_h(y,y). ∥ I h y ∥ 2 2 = j = 0 ∑ M − 1 3 h ( y j 2 + y j y j + 1 + y j + 1 2 ) = (45) ∥ y ∥ h 2 − 6 h 2 a h ( y , y ) .
Since y j 2 + y j y j + 1 + y j + 1 2 ≥ 1 2 ( y j 2 + y j + 1 2 ) y_j^2+y_jy_{j+1}+y_{j+1}^2\geq\frac{1}{2}(y_j^2+y_{j+1}^2) y j 2 + y j y j + 1 + y j + 1 2 ≥ 2 1 ( y j 2 + y j + 1 2 ) , we have
1 3 ∥ y ∥ h 2 ≤ ∥ I h y ∥ L 2 ( 0 , 1 ) 2 ≤ ∥ y ∥ h 2 . \frac{1}{3}\lVert y\rVert_h^2\leq\lVert\mathcal I_hy\rVert_{L^2(0,1)}^2\leq\lVert y\rVert_h^2. 3 1 ∥ y ∥ h 2 ≤ ∥ I h y ∥ L 2 ( 0 , 1 ) 2 ≤ ∥ y ∥ h 2 .
The energy estimate gives
∥ ∂ t V ^ i ∥ L 2 ( Q ) 2 = τ ∑ n = 0 N − 1 ∥ I h D τ v i n ∥ 2 2 ≤ (58) C , ∥ V ^ i ∥ L ∞ ( 0 , T ; H 1 ) ≤ C . \lVert\partial_t\widehat V_i\rVert_{L^2(Q)}^2=\tau\sum_{n=0}^{N-1}\lVert\mathcal I_hD_\tau v_i^n\rVert_2^2\mathrel{\overset{\text{\scriptsize(58)}}{\leq}} C,\qquad\lVert\widehat V_i\rVert_{L^\infty(0,T;H^1)}\leq C. ∥ ∂ t V i ∥ L 2 ( Q ) 2 = τ n = 0 ∑ N − 1 ∥ I h D τ v i n ∥ 2 2 ≤ (58) C , ∥ V i ∥ L ∞ ( 0 , T ; H 1 ) ≤ C .
The compact embedding H 1 ( 0 , 1 ) ↪ L 2 ( 0 , 1 ) H^1(0,1) \hookrightarrow L^2(0,1) H 1 ( 0 , 1 ) ↪ L 2 ( 0 , 1 ) and time compactness [19 ] give a subsequence that converges strongly in L 2 ( Q ) L^2(Q) L 2 ( Q ) : V ^ i , k → k → ∞ v i \widehat{V}_{i,k} \xrightarrow[k\to\infty]{} v_i V i , k k → ∞ v i . On [ t n , t n + 1 ) [t_n,t_{n+1}) [ t n , t n + 1 ) , set V ~ i − ( t ) : = I h v i n \widetilde{V}_i^-(t):=\mathcal{I}_{h}v_i^n V i − ( t ) := I h v i n , V ~ i + ( t ) : = I h v i n + 1 \widetilde{V}_i^+(t):=\mathcal{I}_{h}v_i^{n+1} V i + ( t ) := I h v i n + 1 . These functions satisfy V ~ i ± ∈ L ∞ ( 0 , T ; V h ) \widetilde{V}_i^\pm\in L^\infty(0,T;V_h) V i ± ∈ L ∞ ( 0 , T ; V h ) and
∥ V ^ i − V ~ i − ∥ L 2 ( Q ) 2 = τ 3 3 ∑ n = 0 N − 1 ∥ I h D τ v i n ∥ 2 2 ≤ C τ 2 , ∥ V ~ i ± − V i ± ∥ L 2 ( Q ) ≤ C h . \left\|\widehat{V}_i-\widetilde{V}_i^-\right\|_{L^2(Q)}^2=\frac{\tau^3}{3}\sum_{n=0}^{N-1}\left\|\mathcal{I}_{h}D_\tau v_i^n\right\|_2^2\le C\tau^2,\qquad\left\|\widetilde{V}_i^\pm-V_i^\pm\right\|_{L^2(Q)}\le Ch. V i − V i − L 2 ( Q ) 2 = 3 τ 3 n = 0 ∑ N − 1 ∥ I h D τ v i n ∥ 2 2 ≤ C τ 2 , V i ± − V i ± L 2 ( Q ) ≤ C h .
The same time bound holds for V ~ i + \widetilde{V}_i^+ V i + . Thus
∥ V i , k ± − V ^ i , k ∥ L 2 ( Q ) ≤ ∥ V i , k ± − V ~ i , k ± ∥ L 2 ( Q ) ⏟ ≤ C h k + ∥ V ~ i , k ± − V ^ i , k ∥ L 2 ( Q ) ⏟ ≤ C τ k ≤ C ( h k + τ k ) . \left\|V_{i,k}^\pm-\widehat{V}_{i,k}\right\|_{L^2(Q)}
\le\underbrace{\left\|V_{i,k}^\pm-\widetilde{V}_{i,k}^\pm\right\|_{L^2(Q)}}_{\le Ch_k}
+\underbrace{\left\|\widetilde{V}_{i,k}^\pm-\widehat{V}_{i,k}\right\|_{L^2(Q)}}_{\le C\tau_k}
\le C(h_k+\tau_k). V i , k ± − V i , k L 2 ( Q ) ≤ ≤ C h k V i , k ± − V i , k ± L 2 ( Q ) + ≤ C τ k V i , k ± − V i , k L 2 ( Q ) ≤ C ( h k + τ k ) .
Every reconstruction therefore has the same strong limit v i v_i v i , and V ~ i , k ± ⇀ v i \widetilde{V}_{i,k}^\pm\rightharpoonup v_i V i , k ± ⇀ v i in L 2 ( 0 , T ; H 1 ( 0 , 1 ) ) L^2(0,T;H^1(0,1)) L 2 ( 0 , T ; H 1 ( 0 , 1 )) . For ϕ ∈ C ∞ ( Q ‾ ) \phi\in C^\infty(\overline{Q}) ϕ ∈ C ∞ ( Q ) with ϕ ( ⋅ , T ) = 0 \phi(\cdot,T)=0 ϕ ( ⋅ , T ) = 0 , set ϕ j n : = ϕ ( x j , t n ) \phi_j^n:=\phi(x_j,t_n) ϕ j n := ϕ ( x j , t n ) . Summation by parts gives
τ ∑ n = 0 N − 1 ⟨ D τ v i n , ϕ n + 1 ⟩ h = − ⟨ v i 0 , ϕ 0 ⟩ h − τ ∑ n = 0 N − 1 ⟨ v i n , D τ ϕ n ⟩ h . (61) \tau\sum_{n=0}^{N-1}\langle D_\tau v_i^n,\phi^{n+1}\rangle_h
=-\langle v_i^0,\phi^0\rangle_h-\tau\sum_{n=0}^{N-1}\langle v_i^n,D_\tau\phi^n\rangle_h. \tag*{(61)} τ n = 0 ∑ N − 1 ⟨ D τ v i n , ϕ n + 1 ⟩ h = − ⟨ v i 0 , ϕ 0 ⟩ h − τ n = 0 ∑ N − 1 ⟨ v i n , D τ ϕ n ⟩ h . (61)
Testing (58) with ϕ n + 1 \phi^{n+1} ϕ n + 1 therefore yields
− τ ∑ n = 0 N − 1 ⟨ v i n , D τ ϕ n ⟩ h + μ i τ ∑ n = 0 N − 1 a h ( v i n + 1 , ϕ n + 1 ) = τ ∑ n = 0 N − 1 ⟨ f i n , ϕ n + 1 ⟩ h + ⟨ v i 0 , ϕ 0 ⟩ h . -\tau\sum_{n=0}^{N-1}\langle v_i^n,D_\tau\phi^n\rangle_h+\mu_i\tau\sum_{n=0}^{N-1}a_h(v_i^{n+1},\phi^{n+1})
=\tau\sum_{n=0}^{N-1}\langle f_i^n,\phi^{n+1}\rangle_h+\langle v_i^0,\phi^0\rangle_h. − τ n = 0 ∑ N − 1 ⟨ v i n , D τ ϕ n ⟩ h + μ i τ n = 0 ∑ N − 1 a h ( v i n + 1 , ϕ n + 1 ) = τ n = 0 ∑ N − 1 ⟨ f i n , ϕ n + 1 ⟩ h + ⟨ v i 0 , ϕ 0 ⟩ h .
The mass-lumped time term is
τ k ∑ n = 0 N k − 1 ⟨ v i , k n , D τ k ϕ n ⟩ h k = ∫ Q V ~ i , k Q h k , τ k ( D τ k ϕ ) . \tau_k\sum_{n=0}^{N_k-1}\langle v_{i,k}^n,D_{\tau_k}\phi^n\rangle_{h_k}
=\int_Q \widetilde{V}_{i,k}\mathcal{Q}_{h_k,\tau_k}(D_{\tau_k}\phi). τ k n = 0 ∑ N k − 1 ⟨ v i , k n , D τ k ϕ n ⟩ h k = ∫ Q V i , k Q h k , τ k ( D τ k ϕ ) .
where N k = T / τ k N_k=T/\tau_k N k = T / τ k is the number of time steps. Smoothness of ϕ \phi ϕ gives uniform convergence of the reconstructed difference quotient to ∂ t ϕ \partial_t\phi ∂ t ϕ . Therefore
∣ ∫ Q V ~ i , k Q h k , τ k ( D τ k ϕ ) − ∫ Q v i ∂ t ϕ ∣ ≤ ∥ V ~ i , k − v i ∥ 2 ∥ Q h k , τ k ( D τ k ϕ ) ∥ 2 + ∥ v i ∥ 2 ∥ Q h k , τ k ( D τ k ϕ ) − ∂ t ϕ ∥ 2 → k → ∞ 0. \left|\int_Q\widetilde{V}_{i,k}\mathcal{Q}_{h_k,\tau_k}(D_{\tau_k}\phi)-\int_Qv_i\partial_t\phi\right|
\le\left\|\widetilde{V}_{i,k}-v_i\right\|_2\left\|\mathcal{Q}_{h_k,\tau_k}(D_{\tau_k}\phi)\right\|_2+\left\|v_i\right\|_2\left\|\mathcal{Q}_{h_k,\tau_k}(D_{\tau_k}\phi)-\partial_t\phi\right\|_2
\xrightarrow[k\to\infty]{}0. ∫ Q V i , k Q h k , τ k ( D τ k ϕ ) − ∫ Q v i ∂ t ϕ ≤ V i , k − v i 2 ∥ Q h k , τ k ( D τ k ϕ ) ∥ 2 + ∥ v i ∥ 2 ∥ Q h k , τ k ( D τ k ϕ ) − ∂ t ϕ ∥ 2 k → ∞ 0.
For the diffusion term, let ϕ ~ k + ( t ) = I h k ϕ n + 1 \widetilde{\phi}_k^+(t)=\mathcal{I}_{h_k}\phi^{n+1} ϕ k + ( t ) = I h k ϕ n + 1 on [ t n , t n + 1 ) [t_n,t_{n+1}) [ t n , t n + 1 ) . Then
τ k ∑ n = 0 N k − 1 a h k ( v i , k n + 1 , ϕ n + 1 ) = ∫ Q ∂ x V ~ i , k + ∂ x ϕ ~ k + . \tau_k\sum_{n=0}^{N_k-1}a_{h_k}(v_{i,k}^{n+1},\phi^{n+1})
=\int_Q\partial_x\widetilde{V}_{i,k}^+\partial_x\widetilde{\phi}_k^+. τ k n = 0 ∑ N k − 1 a h k ( v i , k n + 1 , ϕ n + 1 ) = ∫ Q ∂ x V i , k + ∂ x ϕ k + .
Since ∂ x V ~ i , k + \partial_x\widetilde{V}_{i,k}^+ ∂ x V i , k + converges weakly in L 2 ( Q ) L^2(Q) L 2 ( Q ) and ∂ x ϕ ~ k + \partial_x\widetilde{\phi}_k^+ ∂ x ϕ k + converges uniformly to ∂ x ϕ \partial_x\phi ∂ x ϕ ,
∫ Q ∂ x V ~ i , k + ∂ x ϕ ~ k + − ∫ Q ∂ x v i ∂ x ϕ = ∫ Q ∂ x V ~ i , k + ( ∂ x ϕ ~ k + − ∂ x ϕ ) + ∫ Q ( ∂ x V ~ i , k + − ∂ x v i ) ∂ x ϕ → k → ∞ 0. \int_Q\partial_x\widetilde{V}_{i,k}^+\partial_x\widetilde{\phi}_k^+-\int_Q\partial_xv_i\partial_x\phi
=\int_Q\partial_x\widetilde{V}_{i,k}^+(\partial_x\widetilde{\phi}_k^+-\partial_x\phi)+\int_Q(\partial_x\widetilde{V}_{i,k}^+-\partial_xv_i)\partial_x\phi
\xrightarrow[k\to\infty]{}0. ∫ Q ∂ x V i , k + ∂ x ϕ k + − ∫ Q ∂ x v i ∂ x ϕ = ∫ Q ∂ x V i , k + ( ∂ x ϕ k + − ∂ x ϕ ) + ∫ Q ( ∂ x V i , k + − ∂ x v i ) ∂ x ϕ k → ∞ 0.
The initial term is
⟨ v i 0 , ϕ 0 ⟩ h k = ∫ 0 1 Q h k v i 0 Q h k ϕ 0 → k → ∞ ∫ 0 1 v i 0 ( x ) ϕ ( x , 0 ) d x , \langle v_i^0,\phi^0\rangle_{h_k}=\int_0^1\mathcal{Q}_{h_k}v_i^0\mathcal{Q}_{h_k}\phi^0\xrightarrow[k\to\infty]{}\int_0^1v_i^0(x)\phi(x,0)\,dx, ⟨ v i 0 , ϕ 0 ⟩ h k = ∫ 0 1 Q h k v i 0 Q h k ϕ 0 k → ∞ ∫ 0 1 v i 0 ( x ) ϕ ( x , 0 ) d x ,
because the nodal reconstruction of the fixed H 1 ( 0 , 1 ) H^1(0,1) H 1 ( 0 , 1 ) initial datum converges in L 2 ( 0 , 1 ) L^2(0,1) L 2 ( 0 , 1 ) . The nonlinear reaction products converge strongly, since
∥ V ~ 1 , k − V ~ 2 , k + − v 1 v 2 ∥ 2 ≤ K 1 ∥ V ~ 2 , k + − v 2 ∥ 2 + K 2 ∥ V ~ 1 , k − − v 1 ∥ 2 → k → ∞ 0. \left\|\widetilde{V}_{1,k}^-\widetilde{V}_{2,k}^+-v_1v_2\right\|_2
\le K_1\left\|\widetilde{V}_{2,k}^+-v_2\right\|_2+K_2\left\|\widetilde{V}_{1,k}^--v_1\right\|_2
\xrightarrow[k\to\infty]{}0. V 1 , k − V 2 , k + − v 1 v 2 2 ≤ K 1 V 2 , k + − v 2 2 + K 2 V 1 , k − − v 1 2 k → ∞ 0.
For U k = Q h k , τ k u k U_k=\mathcal{Q}_{h_k,\tau_k}u_k U k = Q h k , τ k u k and ψ ∈ L 2 ( Q ) \psi\in L^2(Q) ψ ∈ L 2 ( Q ) ,
∫ Q ( U k V 2 , k + − u v 2 ) ψ = ∫ Q U k ( V 2 , k + − v 2 ) ψ + ∫ Q ( U k − u ) v 2 ψ → k → ∞ 0. \int_Q (U_kV_{2,k}^{+}-uv_2)\psi=\int_Q U_k(V_{2,k}^{+}-v_2)\psi+\int_Q(U_k-u)v_2\psi\xrightarrow{k\to\infty}0. ∫ Q ( U k V 2 , k + − u v 2 ) ψ = ∫ Q U k ( V 2 , k + − v 2 ) ψ + ∫ Q ( U k − u ) v 2 ψ k → ∞ 0.
For the first term,
∣ ∫ Q U k ( V 2 , k + − v 2 ) ψ ∣ ≤ U max ∥ V 2 , k + − v 2 ∥ 2 ∥ ψ ∥ 2 → k → ∞ 0. \left|\int_Q U_k(V_{2,k}^{+}-v_2)\psi\right|\le U_{\max}\lVert V_{2,k}^{+}-v_2\rVert_2\lVert\psi\rVert_2\xrightarrow{k\to\infty}0. ∫ Q U k ( V 2 , k + − v 2 ) ψ ≤ U m a x ∥ V 2 , k + − v 2 ∥ 2 ∥ ψ ∥ 2 k → ∞ 0.
and v 2 ψ ∈ L 1 ( Q ) v_2\psi\in L^1(Q) v 2 ψ ∈ L 1 ( Q ) gives the second limit by weak-∗ * ∗ convergence. By (15) , using the state at either endpoint of each time step gives the same weak identity:
− ∫ Q v i ∂ t ϕ + μ i ∫ Q ∂ x v i ∂ x ϕ = ∫ Q f i ( v , u ) ϕ + ∫ 0 1 v i 0 ϕ ( ⋅ , 0 ) . -\int_Q v_i\partial_t\phi+\mu_i\int_Q\partial_xv_i\partial_x\phi=\int_Q f_i(v,u)\phi+\int_0^1v_i^0\phi(\cdot,0). − ∫ Q v i ∂ t ϕ + μ i ∫ Q ∂ x v i ∂ x ϕ = ∫ Q f i ( v , u ) ϕ + ∫ 0 1 v i 0 ϕ ( ⋅ , 0 ) .
The nonnegative limit solves (1) with the Neumann conditions and prescribed initial data, so uniqueness gives v = S ( u ) v=S(u) v = S ( u ) . Every convergent subsequence has this limit. Relative compactness then gives convergence of the whole sequence.
Recovery controls and the liminf inequality
On each space–time cell D j n : = C j × [ t n , t n + 1 ) D_j^n:=C_j\times[t_n,t_{n+1}) D j n := C j × [ t n , t n + 1 ) , Jensen’s inequality gives
∥ Q h , τ A h , τ f ∥ 2 2 = ∑ n = 0 N − 1 ∑ j = 0 M ∣ D j n ∣ ∣ 1 ∣ D j n ∣ ∫ D j n f ∣ 2 ≤ ∑ n = 0 N − 1 ∑ j = 0 M ∫ D j n ∣ f ∣ 2 = ∥ f ∥ 2 2 . \lVert\mathcal{Q}_{h,\tau}\mathcal{A}_{h,\tau}f\rVert_2^2=\sum_{n=0}^{N-1}\sum_{j=0}^{M}\lvert D_j^n\rvert\left|\frac{1}{\lvert D_j^n\rvert}\int_{D_j^n}f\right|^2\le\sum_{n=0}^{N-1}\sum_{j=0}^{M}\int_{D_j^n}\lvert f\rvert^2=\lVert f\rVert_2^2. ∥ Q h , τ A h , τ f ∥ 2 2 = n = 0 ∑ N − 1 j = 0 ∑ M ∣ D j n ∣ ∣ D j n ∣ 1 ∫ D j n f 2 ≤ n = 0 ∑ N − 1 j = 0 ∑ M ∫ D j n ∣ f ∣ 2 = ∥ f ∥ 2 2 .
Cell averages preserve [ 0 , U max ] [0,U_{\max}] [ 0 , U m a x ] . For f ∈ L 2 ( Q ) f\in L^2(Q) f ∈ L 2 ( Q ) , choose f ε ∈ C ( Q ‾ ) f_\varepsilon\in C(\overline{Q}) f ε ∈ C ( Q ) with ∥ f − f ε ∥ 2 ≤ ε \lVert f-f_\varepsilon\rVert_2\le\varepsilon ∥ f − f ε ∥ 2 ≤ ε . Then
∥ Q h , τ A h , τ f − f ∥ 2 ≤ 2 ∥ f − f ε ∥ 2 + ∥ Q h , τ A h , τ f ε − f ε ∥ 2 . \lVert\mathcal{Q}_{h,\tau}\mathcal{A}_{h,\tau}f-f\rVert_2\le2\lVert f-f_\varepsilon\rVert_2+\lVert\mathcal{Q}_{h,\tau}\mathcal{A}_{h,\tau}f_\varepsilon-f_\varepsilon\rVert_2. ∥ Q h , τ A h , τ f − f ∥ 2 ≤ 2 ∥ f − f ε ∥ 2 + ∥ Q h , τ A h , τ f ε − f ε ∥ 2 .
Uniform continuity bounds the last term by
∣ Q ∣ 1 / 2 sup z , z ′ ∈ Q ‾ ∣ z − z ′ ∣ ≤ h 2 + τ 2 ∣ f ε ( z ) − f ε ( z ′ ) ∣ → h , τ → 0 0. \lvert Q\rvert^{1/2}\sup_{\substack{z,z'\in\overline{Q}\\\lvert z-z'\rvert\le\sqrt{h^2+\tau^2}}}\lvert f_\varepsilon(z)-f_\varepsilon(z')\rvert\xrightarrow{h,\tau\to0}0. ∣ Q ∣ 1/2 z , z ′ ∈ Q ∣ z − z ′ ∣ ≤ h 2 + τ 2 sup ∣ f ε ( z ) − f ε ( z ′ )∣ h , τ → 0 0.
Let h , τ → 0 h,\tau\to0 h , τ → 0 , then ε → 0 \varepsilon\to0 ε → 0 . For bounded controls, strong L 2 L^2 L 2 convergence also implies weak-∗ * ∗ convergence. Indeed, for ψ ∈ L 1 ( Q ) \psi\in L^1(Q) ψ ∈ L 1 ( Q ) choose ψ ε ∈ L 2 ( Q ) \psi_\varepsilon\in L^2(Q) ψ ε ∈ L 2 ( Q ) with ∥ ψ − ψ ε ∥ 1 ≤ ε \lVert\psi-\psi_\varepsilon\rVert_1\le\varepsilon ∥ ψ − ψ ε ∥ 1 ≤ ε . This gives
∣ ∫ Q ( U k − u ) ψ ∣ ≤ 2 U max ∥ ψ − ψ ε ∥ 1 + ∥ U k − u ∥ 2 ∥ ψ ε ∥ 2 . \left|\int_Q(U_k-u)\psi\right|\le2U_{\max}\lVert\psi-\psi_\varepsilon\rVert_1+\lVert U_k-u\rVert_2\lVert\psi_\varepsilon\rVert_2. ∫ Q ( U k − u ) ψ ≤ 2 U m a x ∥ ψ − ψ ε ∥ 1 + ∥ U k − u ∥ 2 ∥ ψ ε ∥ 2 .
Let k → ∞ k\to\infty k → ∞ , then ε → 0 \varepsilon\to0 ε → 0 . Here N = T / τ k N=T/\tau_k N = T / τ k and M = 1 / h k M=1/h_k M = 1/ h k . With target array ( v i d , n ) n < N ∈ R N ( M + 1 ) (v_i^{d,n})_{n<N}\in\mathbb{R}^{N(M+1)} ( v i d , n ) n < N ∈ R N ( M + 1 ) and ζ = 0 \zeta=0 ζ = 0 for IMEX, separate the effort term from the cost:
A k ( u k ) : = j k ( u k ) − λ 2 ∥ U k ∥ 2 2 = 1 2 ∑ i = 1 2 γ i ∥ V i , k − − Q h k , τ k ( ( v i d , n ) n < N ) ∥ 2 2 − η q ∫ Q U k [ ( 1 − ζ ) V 2 , k − + ζ V 2 , k + ] . A_k(u_k):=j_k(u_k)-\frac{\lambda}{2}\lVert U_k\rVert_2^2=\frac{1}{2}\sum_{i=1}^{2}\gamma_i\lVert V_{i,k}^{-}-\mathcal{Q}_{h_k,\tau_k}((v_i^{d,n})_{n<N})\rVert_2^2-\eta q\int_Q U_k\left[(1-\zeta)V_{2,k}^{-}+\zeta V_{2,k}^{+}\right]. A k ( u k ) := j k ( u k ) − 2 λ ∥ U k ∥ 2 2 = 2 1 i = 1 ∑ 2 γ i ∥ V i , k − − Q h k , τ k (( v i d , n ) n < N ) ∥ 2 2 − η q ∫ Q U k [ ( 1 − ζ ) V 2 , k − + ζ V 2 , k + ] .
Define
A : U a d ⟶ R A:\mathcal{U}_{\mathrm{ad}}\longrightarrow\mathbb{R} A : U ad ⟶ R
u ⟼ j ( u ) − λ ∥ u ∥ 2 2 / 2. u\longmapsto j(u)-\lambda\lVert u\rVert_2^2/2. u ⟼ j ( u ) − λ ∥ u ∥ 2 2 /2.
For the tracking terms, write V i , k d = Q h k , τ k ( ( v i d , n ) n < N ) V_{i,k}^{d}=\mathcal{Q}_{h_k,\tau_k}((v_i^{d,n})_{n<N}) V i , k d = Q h k , τ k (( v i d , n ) n < N ) . Then
∥ ( V i , k − − V i , k d ) − ( v i − v i d ) ∥ 2 ≤ ∥ V i , k − − v i ∥ 2 + ∥ V i , k d − v i d ∥ 2 → k → ∞ 0. \lVert(V_{i,k}^{-}-V_{i,k}^{d})-(v_i-v_i^d)\rVert_2\le\lVert V_{i,k}^{-}-v_i\rVert_2+\lVert V_{i,k}^{d}-v_i^d\rVert_2\xrightarrow{k\to\infty}0. ∥( V i , k − − V i , k d ) − ( v i − v i d ) ∥ 2 ≤ ∥ V i , k − − v i ∥ 2 + ∥ V i , k d − v i d ∥ 2 k → ∞ 0.
For either endpoint of the harvest quadrature,
∣ ∫ Q U k v 2 , k ± − ∫ Q u v 2 ∣ ≤ U max ∣ Q ∣ 1 / 2 ∥ v 2 , k ± − v 2 ∥ 2 + ∣ ∫ Q ( U k − u ) v 2 ∣ → k → ∞ 0. \left|\int_Q U_kv_{2,k}^{\pm}-\int_Q uv_2\right|\le U_{\max}|Q|^{1/2}\|v_{2,k}^{\pm}-v_2\|_2+\left|\int_Q(U_k-u)v_2\right|\xrightarrow{k\to\infty}0. ∫ Q U k v 2 , k ± − ∫ Q u v 2 ≤ U m a x ∣ Q ∣ 1/2 ∥ v 2 , k ± − v 2 ∥ 2 + ∫ Q ( U k − u ) v 2 k → ∞ 0.
Thus the state, target and harvest terms converge, and whenever U k ⇀ ∗ u U_k\rightharpoonup^* u U k ⇀ ∗ u ,
A k ( u k ) → k → ∞ A ( u ) . (62) A_k(u_k)\xrightarrow{k\to\infty}A(u). \tag*{(62)} A k ( u k ) k → ∞ A ( u ) . (62)
Lower semicontinuity of the effort term then gives
lim inf k j k ( u k ) ≥ ( 60 ) A ( u ) + λ 2 lim inf k ∥ U k ∥ 2 2 ≥ A ( u ) + λ 2 ∥ u ∥ 2 2 = j ( u ) . \liminf_k j_k(u_k)\overset{(60)}{\ge} A(u)+\frac{\lambda}{2}\liminf_k\|U_k\|_2^2\ge A(u)+\frac{\lambda}{2}\|u\|_2^2=j(u). k lim inf j k ( u k ) ≥ ( 60 ) A ( u ) + 2 λ k lim inf ∥ U k ∥ 2 2 ≥ A ( u ) + 2 λ ∥ u ∥ 2 2 = j ( u ) .
For the extended cost, apply the argument to a subsequence attaining the liminf if it is finite. Otherwise, (18) follows. For u k = A h k , τ k u u_k=\mathcal{A}_{h_k,\tau_k}u u k = A h k , τ k u , strong L 2 L^2 L 2 convergence gives weak-∗ * ∗ convergence and convergence of the effort term. Then (60) gives
j k ( u k ) = A k ( u k ) + λ 2 ∥ U k ∥ 2 2 → k → ∞ j ( u ) , j_k(u_k)=A_k(u_k)+\frac{\lambda}{2}\|U_k\|_2^2\xrightarrow{k\to\infty}j(u), j k ( u k ) = A k ( u k ) + 2 λ ∥ U k ∥ 2 2 k → ∞ j ( u ) ,
which proves (19).
Minimum values and strong convergence of minimizers
Compactness of the discrete box ensures that m k : = min u ∈ U a d , h k , τ k j k ∈ R m_k:=\min_{u\in\mathcal{U}_{\mathrm{ad},h_k,\tau_k}}j_k\in\mathbb{R} m k := min u ∈ U ad , h k , τ k j k ∈ R is attained. Set m ⋆ : = min u ∈ U a d j m_\star:=\min_{u\in\mathcal{U}_{\mathrm{ad}}}j m ⋆ := min u ∈ U ad j . Recovery controls for a global minimizer and the liminf inequality give the two bounds
lim sup k m k ≤ m ⋆ ≤ lim inf k m k ⟹ m k → k → ∞ m ⋆ . \limsup_k m_k\le m_\star\le\liminf_k m_k\quad\Longrightarrow\quad m_k\xrightarrow{k\to\infty}m_\star. k lim sup m k ≤ m ⋆ ≤ k lim inf m k ⟹ m k k → ∞ m ⋆ .
For (20),
m k ≤ j k ( u k ) ≤ m k + ε k ⟹ j k ( u k ) → k → ∞ m ⋆ . m_k\le j_k(u_k)\le m_k+\varepsilon_k\quad\Longrightarrow\quad j_k(u_k)\xrightarrow{k\to\infty}m_\star. m k ≤ j k ( u k ) ≤ m k + ε k ⟹ j k ( u k ) k → ∞ m ⋆ .
Take a subsequence U k ⇀ ∗ u ∗ U_k\rightharpoonup^* u^* U k ⇀ ∗ u ∗ . The liminf inequality gives j ( u ∗ ) = m ⋆ j(u^*)=m_\star j ( u ∗ ) = m ⋆ , and (60) gives
λ 2 ∥ U k ∥ 2 2 = j k ( u k ) − A k ( u k ) → k → ∞ ( 60 ) j ( u ∗ ) − A ( u ∗ ) = λ 2 ∥ u ∗ ∥ 2 2 . \frac{\lambda}{2}\|U_k\|_2^2=j_k(u_k)-A_k(u_k)\xrightarrow[k\to\infty]{(60)}j(u^*)-A(u^*)=\frac{\lambda}{2}\|u^*\|_2^2. 2 λ ∥ U k ∥ 2 2 = j k ( u k ) − A k ( u k ) ( 60 ) k → ∞ j ( u ∗ ) − A ( u ∗ ) = 2 λ ∥ u ∗ ∥ 2 2 .
Since λ > 0 \lambda>0 λ > 0 and U k → k → ∞ L 2 ( Q ) u ∗ U_k\xrightarrow[k\to\infty]{L^2(Q)}u^* U k L 2 ( Q ) k → ∞ u ∗ ,
∥ U k − u ∗ ∥ 2 2 = ∥ U k ∥ 2 2 + ∥ u ∗ ∥ 2 2 − 2 ( U k , u ∗ ) Q → k → ∞ 0. \|U_k-u^*\|_2^2=\|U_k\|_2^2+\|u^*\|_2^2-2(U_k,u^*)_Q\xrightarrow{k\to\infty}0. ∥ U k − u ∗ ∥ 2 2 = ∥ U k ∥ 2 2 + ∥ u ∗ ∥ 2 2 − 2 ( U k , u ∗ ) Q k → ∞ 0.
This proves Theorem 5.5. Every subsequence has a further subsequence converging strongly to a global minimizer, so
inf w ∈ arg min u ∈ U a d j ∥ U k − w ∥ L 2 ( Q ) → k → ∞ 0. \inf_{w\in\operatorname{arg\,min}_{u\in\mathcal{U}_{\mathrm{ad}}}j}\|U_k-w\|_{L^2(Q)}\xrightarrow{k\to\infty}0. w ∈ arg min u ∈ U ad j inf ∥ U k − w ∥ L 2 ( Q ) k → ∞ 0.
If the global minimizer u ∗ u^* u ∗ is unique, the full sequence converges to it. For two asymptotically minimizing reconstructions U k , U ~ k U_k,\widetilde{U}_k U k , U k ,
∥ U k − U ~ k ∥ 2 ≤ ∥ U k − u ∗ ∥ 2 + ∥ U ~ k − u ∗ ∥ 2 → k → ∞ 0. \|U_k-\widetilde{U}_k\|_2\le\|U_k-u^*\|_2+\|\widetilde{U}_k-u^*\|_2\xrightarrow{k\to\infty}0. ∥ U k − U k ∥ 2 ≤ ∥ U k − u ∗ ∥ 2 + ∥ U k − u ∗ ∥ 2 k → ∞ 0.
which proves Corollary 5.6.
Proof of Corollary 5.7
For fine enough grids, minimize j k j_k j k over
K k : = { u ∈ U a d , h k , τ k : ∥ Q h k , τ k u − u ∗ ∥ 2 ≤ r } . K_k := \{u \in\mathcal{U}_{\mathrm{ad},h_k,\tau_k} : \|\mathcal{Q}_{h_k,\tau_k}u-u^*\|_2 \le r\}. K k := { u ∈ U ad , h k , τ k : ∥ Q h k , τ k u − u ∗ ∥ 2 ≤ r } .
The recovery controls of u ∗ u^* u ∗ belong to K k K_k K k for large k k k . This compact set is therefore nonempty. Choose a minimizer u k u_k u k and write U k = Q h k , τ k u k U_k=\mathcal{Q}_{h_k,\tau_k}u_k U k = Q h k , τ k u k . From every subsequence we can extract a further subsequence such that, in L ∞ ( Q ) L^\infty(Q) L ∞ ( Q ) , U k ⇀ ∗ k → ∞ u U_k \mathrel{\rightharpoonup^*}_{k\to\infty} u U k ⇀ ∗ k → ∞ u . The limit is admissible. Weak lower semicontinuity keeps it in the closed L 2 L^2 L 2 ball of radius r r r , while the liminf inequality and recovery controls give
j ( u ) ≤ lim inf k j k ( u k ) ≤ lim sup k j k ( u k ) ≤ j ( u ∗ ) . j(u) \le\liminf_{k} j_k(u_k) \le\limsup_{k} j_k(u_k) \le j(u^*). j ( u ) ≤ k lim inf j k ( u k ) ≤ k lim sup j k ( u k ) ≤ j ( u ∗ ) .
Strict local minimality gives u = u ∗ u=u^* u = u ∗ , j k ( u k ) ⟶ k → ∞ j ( u ∗ ) j_k(u_k)\mathrel{\longrightarrow}_{k\to\infty}j(u^*) j k ( u k ) ⟶ k → ∞ j ( u ∗ ) . Then
λ 2 ∥ U k ∥ 2 2 = j k ( u k ) − A k ( u k ) ⟶ k → ∞ j ( u ∗ ) − A ( u ∗ ) = λ 2 ∥ u ∗ ∥ 2 2 . (63) \frac{\lambda}{2}\|U_k\|_2^2=j_k(u_k)-A_k(u_k)\mathrel{\longrightarrow}_{k\to\infty}j(u^*)-A(u^*)=\frac{\lambda}{2}\|u^*\|_2^2. \tag*{(63)} 2 λ ∥ U k ∥ 2 2 = j k ( u k ) − A k ( u k ) ⟶ k → ∞ j ( u ∗ ) − A ( u ∗ ) = 2 λ ∥ u ∗ ∥ 2 2 . (63)
Since λ > 0 \lambda>0 λ > 0 , the norms converge, turning weak convergence into strong convergence. The subsequence argument gives U k ⟶ k → ∞ u ∗ U_k\mathrel{\longrightarrow}_{k\to\infty}u^* U k ⟶ k → ∞ u ∗ for the full sequence. Eventually ∥ U k − u ∗ ∥ 2 < r \|U_k-u^*\|_2<r ∥ U k − u ∗ ∥ 2 < r : the ball constraint is locally inactive, and u k u_k u k is a local minimizer over the admissible box.
Proof of Proposition 5.3
Proof. Subtracting the two costs and using (15) gives
∣ j h , τ + 1 ( u ) − j h , τ + 0 ( u ) ∣ = η q ∣ ∫ Q ( Q h , τ u ) ( V 2 + − V 2 − ) ∣ ≤ η q ∥ Q h , τ u ∥ L 2 ( Q ) ∥ V 2 + − V 2 − ∥ L 2 ( Q ) ≤ C τ . |j_{h,\tau}^{+1}(u)-j_{h,\tau}^{+0}(u)|=\eta q\left|\int_Q(\mathcal{Q}_{h,\tau}u)(V_2^+-V_2^-)\right|
\le\eta q\|\mathcal{Q}_{h,\tau}u\|_{L^2(Q)}\|V_2^+-V_2^-\|_{L^2(Q)}
\le C\tau. ∣ j h , τ + 1 ( u ) − j h , τ + 0 ( u ) ∣ = η q ∫ Q ( Q h , τ u ) ( V 2 + − V 2 − ) ≤ η q ∥ Q h , τ u ∥ L 2 ( Q ) ∥ V 2 + − V 2 − ∥ L 2 ( Q ) ≤ C τ .
The bound is uniform in u u u , so taking infima gives the same bound for the minimum values. □
Proof. Quadratic growth and the cost comparison give
c ∥ u 1 − u 0 ∥ h , τ 2 ≤ j 0 ( u 1 ) − j 0 ( u 0 ) = j 0 ( u 1 ) − j 1 ( u 1 ) + j 1 ( u 1 ) − j 1 ( u 0 ) + j 1 ( u 0 ) − j 0 ( u 0 ) ≤ ε + 2 δ . c\|u_1-u_0\|_{h,\tau}^2\le j_0(u_1)-j_0(u_0)=j_0(u_1)-j_1(u_1)+j_1(u_1)-j_1(u_0)+j_1(u_0)-j_0(u_0)\le\varepsilon+2\delta. c ∥ u 1 − u 0 ∥ h , τ 2 ≤ j 0 ( u 1 ) − j 0 ( u 0 ) = j 0 ( u 1 ) − j 1 ( u 1 ) + j 1 ( u 1 ) − j 1 ( u 0 ) + j 1 ( u 0 ) − j 0 ( u 0 ) ≤ ε + 2 δ .
If 2 δ < c r 2 2\delta<cr^2 2 δ < c r 2 , a minimizer of j 1 j_1 j 1 on K r K_r K r satisfies
∥ u 1 − u 0 ∥ h , τ 2 ≤ 2 δ / c < r 2 . \|u_1-u_0\|_{h,\tau}^2\le2\delta/c<r^2. ∥ u 1 − u 0 ∥ h , τ 2 ≤ 2 δ / c < r 2 .
Thus the ball constraint is inactive and u 1 u_1 u 1 is a local minimizer on U a d , h , τ \mathcal{U}_{\mathrm{ad},h,\tau} U ad , h , τ . □
Accuracy and local optimization proofs
Accuracy for smooth states
The space W 1 , ∞ ( 0 , T ; X ) W^{1,\infty}(0,T;X) W 1 , ∞ ( 0 , T ; X ) consists of X X X -valued functions whose norm and weak time-derivative norm are essentially bounded.
Proposition E.1 (Accuracy for compatible smooth data). Let u ∈ U a d ∩ W 1 , ∞ ( 0 , T ; C ( [ 0 , 1 ] ) ) u\in\mathcal{U}_{\mathrm{ad}}\cap W^{1,\infty}(0,T;C([0,1])) u ∈ U ad ∩ W 1 , ∞ ( 0 , T ; C ([ 0 , 1 ])) . Assume the states, their second time derivatives, and their fourth spatial derivatives are continuous and uniformly bounded. At x = 0 , 1 x=0,1 x = 0 , 1 , assume that for every t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] ,
∂ x v i = ∂ x 3 v i = 0. (64) \partial_xv_i=\partial_x^3v_i=0. \tag*{(64)} ∂ x v i = ∂ x 3 v i = 0. (64)
For u j n : = u ( x j , t n ) u_j^n := u(x_j,t_n) u j n := u ( x j , t n ) and nonnegative initial nodal vectors with O ( h 2 ) O(h^2) O ( h 2 ) maximum-norm error, either scheme satisfies, for small h , τ h,\tau h , τ ,
max 0 ≤ n ≤ N ∑ i = 1 2 ∥ v i ( t n ) − Φ h v i n ∥ L 2 ( 0 , 1 ) ≤ C ( τ + h 2 ) . (65) \max_{0\le n\le N}\sum_{i=1}^{2}\|v_i(t_n)-\Phi_h v_i^n\|_{L^2(0,1)}\le C(\tau+h^2). \tag*{(65)} 0 ≤ n ≤ N max i = 1 ∑ 2 ∥ v i ( t n ) − Φ h v i n ∥ L 2 ( 0 , 1 ) ≤ C ( τ + h 2 ) . (65)
Condition (64) removes the first-order endpoint error.
Proof of Proposition E.1
Set V i n : = ( v i ( x j , t n ) ) j = 0 M ∈ R M + 1 V_i^n := (v_i(x_j,t_n))_{j=0}^M\in\mathbb{R}^{M+1} V i n := ( v i ( x j , t n ) ) j = 0 M ∈ R M + 1 . Inserting V V V into the scheme, subtracting the right side from the left and dividing by τ \tau τ gives the residual r i n ∈ R M + 1 r_i^n\in\mathbb{R}^{M+1} r i n ∈ R M + 1 . Taylor’s formula yields
v i ( x , t + τ ) − v i ( x , t ) τ = ∂ t v i ( x , t ) + 1 τ ∫ t t + τ ( t + τ − s ) ∂ t 2 v i ( x , s ) d s . \frac{v_i(x,t+\tau)-v_i(x,t)}{\tau}=\partial_t v_i(x,t)+\frac{1}{\tau}\int_t^{t+\tau}(t+\tau-s)\partial_t^2v_i(x,s)\,\mathrm{d}s. τ v i ( x , t + τ ) − v i ( x , t ) = ∂ t v i ( x , t ) + τ 1 ∫ t t + τ ( t + τ − s ) ∂ t 2 v i ( x , s ) d s .
Its remainder is bounded by τ 2 ∥ ∂ t 2 v i ∥ ∞ \frac{\tau}{2}\|\partial_t^2v_i\|_\infty 2 τ ∥ ∂ t 2 v i ∥ ∞ . At an interior node,
∣ v i ( x − h , t ) − 2 v i ( x , t ) + v i ( x + h , t ) h 2 − ∂ x 2 v i ( x , t ) ∣ ≤ h 2 12 ∥ ∂ x 4 v i ∥ ∞ . \left|\frac{v_i(x-h,t)-2v_i(x,t)+v_i(x+h,t)}{h^2}-\partial_x^2v_i(x,t)\right|\le\frac{h^2}{12}\|\partial_x^4v_i\|_\infty. h 2 v i ( x − h , t ) − 2 v i ( x , t ) + v i ( x + h , t ) − ∂ x 2 v i ( x , t ) ≤ 12 h 2 ∥ ∂ x 4 v i ∥ ∞ .
The derivative norms are over [ 0 , 1 ] × [ 0 , T ] [0,1]\times[0,T] [ 0 , 1 ] × [ 0 , T ] . At the endpoints,
2 ( v i ( h , t ) − v i ( 0 , t ) ) h 2 = 2 h ∂ x v i ( 0 , t ) + ∂ x 2 v i ( 0 , t ) + h 3 ∂ x 3 v i ( 0 , t ) + O ( h 2 ) , \frac{2(v_i(h,t)-v_i(0,t))}{h^2}=\frac{2}{h}\partial_xv_i(0,t)+\partial_x^2v_i(0,t)+\frac{h}{3}\partial_x^3v_i(0,t)+O(h^2), h 2 2 ( v i ( h , t ) − v i ( 0 , t )) = h 2 ∂ x v i ( 0 , t ) + ∂ x 2 v i ( 0 , t ) + 3 h ∂ x 3 v i ( 0 , t ) + O ( h 2 ) ,
and
2 ( v i ( 1 − h , t ) − v i ( 1 , t ) ) h 2 = − 2 h ∂ x v i ( 1 , t ) + ∂ x 2 v i ( 1 , t ) − h 3 ∂ x 3 v i ( 1 , t ) + O ( h 2 ) . \frac{2(v_i(1-h,t)-v_i(1,t))}{h^2}=-\frac{2}{h}\partial_xv_i(1,t)+\partial_x^2v_i(1,t)-\frac{h}{3}\partial_x^3v_i(1,t)+O(h^2). h 2 2 ( v i ( 1 − h , t ) − v i ( 1 , t )) = − h 2 ∂ x v i ( 1 , t ) + ∂ x 2 v i ( 1 , t ) − 3 h ∂ x 3 v i ( 1 , t ) + O ( h 2 ) .
The Neumann conditions cancel the h − 1 h^{-1} h − 1 terms, and (64) cancels the h h h terms. Since ∣ V i n + 1 − V i n ∣ ≤ τ ∥ ∂ t v i ∥ ∞ \lvert V_i^{n+1}-V_i^n\rvert\le\tau\|\partial_t v_i\|_\infty ∣ V i n + 1 − V i n ∣ ≤ τ ∥ ∂ t v i ∥ ∞ , the reaction quadratures differ by O ( τ ) O(\tau) O ( τ ) . For diffusion at the new time, differentiate (1) in time in the distributional sense:
μ i ∂ t ∂ x 2 v i = ∂ t 2 v i − ∂ t f i ( v , u ) , \mu_i\partial_t\partial_x^2v_i=\partial_t^2v_i-\partial_t f_i(v,u), μ i ∂ t ∂ x 2 v i = ∂ t 2 v i − ∂ t f i ( v , u ) ,
where
∂ t f 1 = ( − m + α v 2 ) ∂ t v 1 + α v 1 ∂ t v 2 , ∂ t f 2 = ( r − β v 1 − q u ) ∂ t v 2 − β v 2 ∂ t v 1 − q v 2 ∂ t u . \partial_t f_1=(-m+\alpha v_2)\partial_t v_1+\alpha v_1\partial_t v_2,\qquad\partial_t f_2=(r-\beta v_1-qu)\partial_t v_2-\beta v_2\partial_t v_1-qv_2\partial_t u. ∂ t f 1 = ( − m + α v 2 ) ∂ t v 1 + α v 1 ∂ t v 2 , ∂ t f 2 = ( r − β v 1 − q u ) ∂ t v 2 − β v 2 ∂ t v 1 − q v 2 ∂ t u .
The regularity assumptions bound every term. Hence ∥ ∂ x 2 v i ( t + τ ) − ∂ x 2 v i ( t ) ∥ ∞ ≤ C τ \|\partial_x^2v_i(t+\tau)-\partial_x^2v_i(t)\|_\infty\le C\tau ∥ ∂ x 2 v i ( t + τ ) − ∂ x 2 v i ( t ) ∥ ∞ ≤ C τ and
max i ∈ { 1 , 2 } 0 ≤ n < N ∥ r i n ∥ ∞ ≤ C ( τ + h 2 ) . \max_{\substack{i\in\{1,2\}\\0\le n<N}}\|r_i^n\|_\infty\le C(\tau+h^2). i ∈ { 1 , 2 } 0 ≤ n < N max ∥ r i n ∥ ∞ ≤ C ( τ + h 2 ) .
Set e i n : = V i n − v i n ∈ R M + 1 e_i^n:=V_i^n-v_i^n\in\mathbb{R}^{M+1} e i n := V i n − v i n ∈ R M + 1 and E n : = ∥ e 1 n ∥ ∞ + ∥ e 2 n ∥ ∞ E_n:=\|e_1^n\|_\infty+\|e_2^n\|_\infty E n := ∥ e 1 n ∥ ∞ + ∥ e 2 n ∥ ∞ . For IMEX,
A i e i n + 1 = e i n + τ [ f i ( V n , u n ) − f i ( v n , u n ) ] + τ r i n . A_i e_i^{n+1}=e_i^n+\tau[f_i(V^n,u^n)-f_i(v^n,u^n)]+\tau r_i^n. A i e i n + 1 = e i n + τ [ f i ( V n , u n ) − f i ( v n , u n )] + τ r i n .
Lipschitz continuity on the bounded range of V n , v n V^n,v^n V n , v n and ∥ A i − 1 ∥ ∞ ≤ 1 \|A_i^{-1}\|_\infty\le1 ∥ A i − 1 ∥ ∞ ≤ 1 give
E n + 1 ≤ ( 1 + C τ ) E n + C τ ( τ + h 2 ) . (66) E_{n+1}\le(1+C\tau)E_n+C\tau(\tau+h^2). \tag*{(66)} E n + 1 ≤ ( 1 + C τ ) E n + C τ ( τ + h 2 ) . (66)
For the positive scheme, subtract the equations using the computed matrix C n C^n C n to get
C n e 2 n + 1 = ( 1 + τ r ) e 2 n − τ β e 1 n ∘ V 2 n + 1 + τ r 2 n , B e 1 n + 1 = e 1 n + τ α e 1 n ∘ V 2 n + 1 + τ α v 1 n ∘ e 2 n + 1 + τ r 1 n . C^n e_2^{n+1}=(1+\tau r)e_2^n-\tau\beta e_1^n\circ V_2^{n+1}+\tau r_2^n,\qquad Be_1^{n+1}=e_1^n+\tau\alpha e_1^n\circ V_2^{n+1}+\tau\alpha v_1^n\circ e_2^{n+1}+\tau r_1^n. C n e 2 n + 1 = ( 1 + τ r ) e 2 n − τ β e 1 n ∘ V 2 n + 1 + τ r 2 n , B e 1 n + 1 = e 1 n + τ α e 1 n ∘ V 2 n + 1 + τ α v 1 n ∘ e 2 n + 1 + τ r 1 n .
The inverse bounds give
∥ e 2 n + 1 ∥ ∞ ≤ ( 1 + C τ ) ∥ e 2 n ∥ ∞ + C τ ∥ e 1 n ∥ ∞ + C τ ( τ + h 2 ) , ∥ e 1 n + 1 ∥ ∞ ≤ ( 1 + C τ ) ∥ e 1 n ∥ ∞ + C τ ∥ e 2 n + 1 ∥ ∞ + C τ ( τ + h 2 ) . \|e_2^{n+1}\|_\infty\le(1+C\tau)\|e_2^n\|_\infty+C\tau\|e_1^n\|_\infty+C\tau(\tau+h^2),\qquad\|e_1^{n+1}\|_\infty\le(1+C\tau)\|e_1^n\|_\infty+C\tau\|e_2^{n+1}\|_\infty+C\tau(\tau+h^2). ∥ e 2 n + 1 ∥ ∞ ≤ ( 1 + C τ ) ∥ e 2 n ∥ ∞ + C τ ∥ e 1 n ∥ ∞ + C τ ( τ + h 2 ) , ∥ e 1 n + 1 ∥ ∞ ≤ ( 1 + C τ ) ∥ e 1 n ∥ ∞ + C τ ∥ e 2 n + 1 ∥ ∞ + C τ ( τ + h 2 ) .
Substituting the first bound into the second and using τ 2 ≤ T τ \tau^2 \le T\tau τ 2 ≤ T τ recovers (63). Iteration gives
E n ≤ ( 1 + C τ ) n E 0 + C τ ( τ + h 2 ) ∑ ℓ = 0 n − 1 ( 1 + C τ ) ℓ ≤ e C T [ E 0 + C T ( τ + h 2 ) ] ≤ C ( τ + h 2 ) . (67) E_n \le(1+C\tau)^n E_0+C\tau(\tau+h^2)\sum_{\ell=0}^{n-1}(1+C\tau)^\ell\le e^{CT}[E_0+CT(\tau+h^2)]\le C(\tau+h^2). \tag*{(67)} E n ≤ ( 1 + C τ ) n E 0 + C τ ( τ + h 2 ) ℓ = 0 ∑ n − 1 ( 1 + C τ ) ℓ ≤ e C T [ E 0 + C T ( τ + h 2 )] ≤ C ( τ + h 2 ) . (67)
The interpolation error on each interval follows from the linear interpolation remainder:
∣ v i ( x , t n ) − I h V i n ( x ) ∣ ≤ 1 2 ( x − x j ) ( x j + 1 − x ) ∥ ∂ x 2 v i ( t n ) ∥ ∞ ≤ h 2 8 ∥ ∂ x 2 v i ( t n ) ∥ ∞ . |v_i(x,t_n)-\mathcal{I}_hV_i^n(x)|\le\frac{1}{2}(x-x_j)(x_{j+1}-x)\|\partial_x^2v_i(t_n)\|_\infty\le\frac{h^2}{8}\|\partial_x^2v_i(t_n)\|_\infty. ∣ v i ( x , t n ) − I h V i n ( x ) ∣ ≤ 2 1 ( x − x j ) ( x j + 1 − x ) ∥ ∂ x 2 v i ( t n ) ∥ ∞ ≤ 8 h 2 ∥ ∂ x 2 v i ( t n ) ∥ ∞ .
Also, I h e i n \mathcal{I}_he_i^n I h e i n is a convex combination of its two endpoint values, so ∥ I h e i n ∥ ∞ ≤ ∥ e i n ∥ ∞ \|\mathcal{I}_he_i^n\|_\infty\le\|e_i^n\|_\infty ∥ I h e i n ∥ ∞ ≤ ∥ e i n ∥ ∞ . Therefore
∥ v i ( t n ) − I h v i n ∥ 2 ≤ ∥ v i ( t n ) − I h V i n ∥ 2 + ∥ I h e i n ∥ 2 ≤ C h 2 + ∥ e i n ∥ ∞ . \|v_i(t_n)-\mathcal{I}_hv_i^n\|_2\le\|v_i(t_n)-\mathcal{I}_hV_i^n\|_2+\|\mathcal{I}_he_i^n\|_2\le Ch^2+\|e_i^n\|_\infty. ∥ v i ( t n ) − I h v i n ∥ 2 ≤ ∥ v i ( t n ) − I h V i n ∥ 2 + ∥ I h e i n ∥ 2 ≤ C h 2 + ∥ e i n ∥ ∞ .
This proves (65) .
Proof of Theorem 8.1
For d ∈ C ( u ∗ ) ⊂ R N ( M + 1 ) d\in\mathcal{C}(u^*)\subset\mathbb{R}^{N(M+1)} d ∈ C ( u ∗ ) ⊂ R N ( M + 1 ) , u ∗ + t d u^*+td u ∗ + t d is feasible whenever
0 ≤ t < min ( { U max − u i ∗ d i : d i > 0 } ∪ { u i ∗ − d i : d i < 0 } ) , 0\le t<\min\left(\left\{\frac{U_{\max}-u_i^*}{d_i}:d_i>0\right\}\cup\left\{\frac{u_i^*}{-d_i}:d_i<0\right\}\right), 0 ≤ t < min ( { d i U m a x − u i ∗ : d i > 0 } ∪ { − d i u i ∗ : d i < 0 } ) ,
with min ∅ : = + ∞ \min\varnothing:=+\infty min ∅ := + ∞ and all ratios positive. The fundamental theorem of calculus applied twice gives
j ( u ∗ + t d ) − j ( u ∗ ) = t ⟨ g ( u ∗ ) , d ⟩ h , τ + t 2 ∫ 0 1 ( 1 − s ) ⟨ H ( u ∗ + s t d ) d , d ⟩ h , τ d s . j(u^*+td)-j(u^*)=t\langle g(u^*),d\rangle_{h,\tau}+t^2\int_0^1(1-s)\langle H(u^*+std)d,d\rangle_{h,\tau}\,ds. j ( u ∗ + t d ) − j ( u ∗ ) = t ⟨ g ( u ∗ ) , d ⟩ h , τ + t 2 ∫ 0 1 ( 1 − s ) ⟨ H ( u ∗ + s t d ) d , d ⟩ h , τ d s .
Continuity of H H H makes the difference between this integral and 1 2 ⟨ H ( u ∗ ) d , d ⟩ h , τ \frac{1}{2}\langle H(u^*)d,d\rangle_{h,\tau} 2 1 ⟨ H ( u ∗ ) d , d ⟩ h , τ tend to zero. Since d d d is critical, Taylor expansion gives
j ( u ∗ + t d ) − j ( u ∗ ) = t ⟨ g ( u ∗ ) , d ⟩ h , τ + t 2 2 ⟨ H ( u ∗ ) d , d ⟩ h , τ + o ( t 2 ) = t 2 2 ⟨ H ( u ∗ ) d , d ⟩ h , τ + o ( t 2 ) . j(u^*+td)-j(u^*)=t\langle g(u^*),d\rangle_{h,\tau}+\frac{t^2}{2}\langle H(u^*)d,d\rangle_{h,\tau}+o(t^2)=\frac{t^2}{2}\langle H(u^*)d,d\rangle_{h,\tau}+o(t^2). j ( u ∗ + t d ) − j ( u ∗ ) = t ⟨ g ( u ∗ ) , d ⟩ h , τ + 2 t 2 ⟨ H ( u ∗ ) d , d ⟩ h , τ + o ( t 2 ) = 2 t 2 ⟨ H ( u ∗ ) d , d ⟩ h , τ + o ( t 2 ) .
At a local minimizer the left side is nonnegative, which proves necessity. For sufficiency, suppose quadratic growth fails. Then feasible u k ⟶ u ∗ u_k\longrightarrow u^* u k ⟶ u ∗ , u k ≠ u ∗ u_k\ne u^* u k = u ∗ , satisfy
j ( u k ) < j ( u ∗ ) + k − 1 ∥ u k − u ∗ ∥ h , τ 2 . j(u_k)<j(u^*)+k^{-1}\|u_k-u^*\|_{h,\tau}^2. j ( u k ) < j ( u ∗ ) + k − 1 ∥ u k − u ∗ ∥ h , τ 2 .
Set s k : = ∥ u k − u ∗ ∥ h , τ > 0 s_k:=\|u_k-u^*\|_{h,\tau}>0 s k := ∥ u k − u ∗ ∥ h , τ > 0 , d k : = ( u k − u ∗ ) / s k ∈ R N ( M + 1 ) d_k:=(u_k-u^*)/s_k\in\mathbb{R}^{N(M+1)} d k := ( u k − u ∗ ) / s k ∈ R N ( M + 1 ) . On the fixed grid, compactness gives a subsequence with
d k ⟶ d ∈ T ( u ∗ ) , ∥ d ∥ h , τ = 1. d_k\longrightarrow d\in T(u^*),\qquad\|d\|_{h,\tau}=1. d k ⟶ d ∈ T ( u ∗ ) , ∥ d ∥ h , τ = 1.
Stationarity gives ⟨ g ( u ∗ ) , d k ⟩ h , τ ≥ 0 \langle g(u^*),d_k\rangle_{h,\tau}\ge0 ⟨ g ( u ∗ ) , d k ⟩ h , τ ≥ 0 , while Taylor expansion and boundedness of H ( u ∗ ) H(u^*) H ( u ∗ ) give
⟨ g ( u ∗ ) , d k ⟩ h , τ < s k [ 1 k − 1 2 ⟨ H ( u ∗ ) d k , d k ⟩ h , τ + o ( 1 ) ] ≤ C s k ⟶ 0 , \langle g(u^*),d_k\rangle_{h,\tau}<s_k\left[\frac{1}{k}-\frac{1}{2}\langle H(u^*)d_k,d_k\rangle_{h,\tau}+o(1)\right]\le Cs_k\longrightarrow0, ⟨ g ( u ∗ ) , d k ⟩ h , τ < s k [ k 1 − 2 1 ⟨ H ( u ∗ ) d k , d k ⟩ h , τ + o ( 1 ) ] ≤ C s k ⟶ 0 ,
so d ∈ C ( u ∗ ) d\in\mathcal{C}(u^*) d ∈ C ( u ∗ ) . Dividing by s k 2 s_k^2 s k 2 ,
1 k > ⟨ g ( u ∗ ) , d k ⟩ h , τ s k ⏟ ≥ 0 + 1 2 ⟨ H ( u ∗ ) d k , d k ⟩ h , τ + o ( 1 ) . \frac{1}{k}>\underbrace{\frac{\langle g(u^*),d_k\rangle_{h,\tau}}{s_k}}_{\ge0}+\frac{1}{2}\langle H(u^*)d_k,d_k\rangle_{h,\tau}+o(1). k 1 > ≥ 0 s k ⟨ g ( u ∗ ) , d k ⟩ h , τ + 2 1 ⟨ H ( u ∗ ) d k , d k ⟩ h , τ + o ( 1 ) .
The first term is nonnegative, so taking limits gives
0 ≥ 1 2 ⟨ H ( u ∗ ) d , d ⟩ h , τ ≥ κ 2 ∥ d ∥ h , τ 2 = κ 2 > 0 , 0\ge\frac{1}{2}\langle H(u^*)d,d\rangle_{h,\tau}\ge\frac{\kappa}{2}\|d\|_{h,\tau}^2=\frac{\kappa}{2}>0, 0 ≥ 2 1 ⟨ H ( u ∗ ) d , d ⟩ h , τ ≥ 2 κ ∥ d ∥ h , τ 2 = 2 κ > 0 ,
a contradiction. For stationary u ∗ u^* u ∗ , g i ( u ∗ ) = 0 g_i(u^*)=0 g i ( u ∗ ) = 0 on F F F . Under strict complementarity,
⟨ g ( u ∗ ) , d ⟩ h , τ = ∑ i ∈ A 0 ∪ A U W i i g i ( u ∗ ) d i . \langle g(u^*),d\rangle_{h,\tau}=\sum_{i\in A_0\cup A_U}W_{ii}g_i(u^*)d_i. ⟨ g ( u ∗ ) , d ⟩ h , τ = i ∈ A 0 ∪ A U ∑ W ii g i ( u ∗ ) d i .
Every summand is nonnegative and vanishes only for d i = 0 d_i=0 d i = 0 , so the critical cone is the free subspace.
Proof of Theorem 8.2
For θ > 0 \theta>0 θ > 0 , set y ∗ : = u ∗ − θ λ g ( u ∗ ) ∈ R N ( M + 1 ) y^*:=u^*-\frac{\theta}{\lambda}g(u^*)\in\mathbb{R}^{N(M+1)} y ∗ := u ∗ − λ θ g ( u ∗ ) ∈ R N ( M + 1 ) . Stationarity on F F F and strict complementarity on A A A imply
{ y i ∗ < 0 , if i ∈ A 0 , y i ∗ > U max , if i ∈ A U , y i ∗ = u i ∗ ∈ ( 0 , U max ) , if i ∈ F . \begin{cases}
y_i^*<0, & \text{if }i\in A_0,\\
y_i^*>U_{\max}, & \text{if }i\in A_U,\\
y_i^*=u_i^*\in(0,U_{\max}), & \text{if }i\in F.
\end{cases} ⎩ ⎨ ⎧ y i ∗ < 0 , y i ∗ > U m a x , y i ∗ = u i ∗ ∈ ( 0 , U m a x ) , if i ∈ A 0 , if i ∈ A U , if i ∈ F .
On the fixed grid, all strict inequalities have a positive margin. Choose δ > 0 \delta>0 δ > 0 smaller than every number in
{ − y i ∗ : i ∈ A 0 } ∪ { y i ∗ − U max : i ∈ A U } ∪ { u i ∗ , U max − u i ∗ : i ∈ F } . \{-y_i^*:i\in A_0\}\cup\{y_i^*-U_{\max}:i\in A_U\}\cup\{u_i^*,U_{\max}-u_i^*:i\in F\}. { − y i ∗ : i ∈ A 0 } ∪ { y i ∗ − U m a x : i ∈ A U } ∪ { u i ∗ , U m a x − u i ∗ : i ∈ F } .
Continuity gives a neighborhood of u ∗ u^* u ∗ on which
∥ u − θ λ g ( u ) − y ∗ ∥ ∞ < δ . \left\|u-\frac{\theta}{\lambda}g(u)-y^*\right\|_\infty<\delta. u − λ θ g ( u ) − y ∗ ∞ < δ .
Projection therefore fixes active coordinates at their bounds and leaves free coordinates unchanged throughout this neighborhood. Its derivative is P F P_F P F , giving
D Φ θ ( u ∗ ) = P F ( I − θ λ H ( u ∗ ) ) . D\Phi_\theta(u^*)=P_F\left(I-\frac{\theta}{\lambda}H(u^*)\right). D Φ θ ( u ∗ ) = P F ( I − λ θ H ( u ∗ ) ) .
The map is locally constant when F = ∅ F=\varnothing F = ∅ . Otherwise, write A = { i 1 A < ⋯ < i ∣ A ∣ A } A=\{i_1^A<\cdots<i_{|A|}^A\} A = { i 1 A < ⋯ < i ∣ A ∣ A } and define
E A : R ∣ A ∣ ⟶ R N ( M + 1 ) E_A:\mathbb{R}^{|A|}\longrightarrow\mathbb{R}^{N(M+1)} E A : R ∣ A ∣ ⟶ R N ( M + 1 )
x ⟼ ∑ k = 1 ∣ A ∣ x k e i k A . x\longmapsto\sum_{k=1}^{|A|}x_ke_{i_k^A}. x ⟼ k = 1 ∑ ∣ A ∣ x k e i k A .
Ordering coordinates as ( F , A ) (F,A) ( F , A ) , with E F T H ( u ∗ ) E A ∈ R ∣ F ∣ × ∣ A ∣ E_F^TH(u^*)E_A\in\mathbb{R}^{|F|\times|A|} E F T H ( u ∗ ) E A ∈ R ∣ F ∣ × ∣ A ∣ ,
D Φ θ ( u ∗ ) ≃ ( I F − θ λ H F − θ λ E F T H ( u ∗ ) E A 0 0 ) . D\Phi_\theta(u^*)\simeq
\begin{pmatrix}
I_F-\frac{\theta}{\lambda}H_F & -\frac{\theta}{\lambda}E_F^TH(u^*)E_A\\
0 & 0
\end{pmatrix}. D Φ θ ( u ∗ ) ≃ ( I F − λ θ H F 0 − λ θ E F T H ( u ∗ ) E A 0 ) .
Starting close enough to u ∗ u^* u ∗ , one sweep identifies the active coordinates. On u A = u A ∗ u_A=u_A^* u A = u A ∗ , write u = u ∗ + E F x u=u^*+E_Fx u = u ∗ + E F x and choose an open neighborhood O F ⊂ R ∣ F ∣ O_F\subset\mathbb{R}^{|F|} O F ⊂ R ∣ F ∣ of zero where u ∗ + E F x ∈ Θ h , τ u^*+E_Fx\in\Theta_{h,\tau} u ∗ + E F x ∈ Θ h , τ and projection leaves free coordinates unchanged. The free coordinates then follow
Ψ θ : O F ⟶ R ∣ F ∣ \Psi_\theta:O_F\longrightarrow\mathbb{R}^{|F|} Ψ θ : O F ⟶ R ∣ F ∣
x ⟼ x − θ λ E F T g ( u ∗ + E F x ) . x\longmapsto x-\frac{\theta}{\lambda}E_F^Tg(u^*+E_Fx). x ⟼ x − λ θ E F T g ( u ∗ + E F x ) .
with Ψ θ ( 0 ) = 0 \Psi_\theta(0)=0 Ψ θ ( 0 ) = 0 and D Ψ θ ( 0 ) = I F − θ λ H F D\Psi_\theta(0)=I_F-\frac{\theta}{\lambda}H_F D Ψ θ ( 0 ) = I F − λ θ H F . Since H F T W F = W F H F H_F^TW_F=W_FH_F H F T W F = W F H F and W F W_F W F is positive diagonal, there is a W F W_F W F -orthonormal eigenbasis ( q ℓ ) ℓ = 1 ∣ F ∣ ⊂ R ∣ F ∣ (q_\ell)_{\ell=1}^{|F|}\subset\mathbb{R}^{|F|} ( q ℓ ) ℓ = 1 ∣ F ∣ ⊂ R ∣ F ∣ with
H F q ℓ = μ ℓ q ℓ , q ℓ T W F q m = δ ℓ m , H_Fq_\ell=\mu_\ell q_\ell,\qquad q_\ell^TW_Fq_m=\delta_{\ell m}, H F q ℓ = μ ℓ q ℓ , q ℓ T W F q m = δ ℓ m ,
and hence
( I F − θ λ H F ) q ℓ = ( 1 − θ λ μ ℓ ) q ℓ . \left(I_F-\frac{\theta}{\lambda}H_F\right)q_\ell=\left(1-\frac{\theta}{\lambda}\mu_\ell\right)q_\ell. ( I F − λ θ H F ) q ℓ = ( 1 − λ θ μ ℓ ) q ℓ .
where δ ℓ m \delta_{\ell m} δ ℓ m is the Kronecker delta, equal to one if ℓ = m \ell=m ℓ = m and zero otherwise. For e = ∑ ℓ = 1 ∣ F ∣ c ℓ q ℓ ∈ R ∣ F ∣ e=\sum_{\ell=1}^{|F|}c_\ell q_\ell\in\mathbb{R}^{|F|} e = ∑ ℓ = 1 ∣ F ∣ c ℓ q ℓ ∈ R ∣ F ∣ , c ℓ ∈ R c_\ell\in\mathbb{R} c ℓ ∈ R ,
∥ ( I F − θ λ H F ) e ∥ W F 2 = ∑ ℓ = 1 ∣ F ∣ ∣ 1 − θ λ μ ℓ ∣ 2 ∣ c ℓ ∣ 2 . \left\|\left(I_F-\frac{\theta}{\lambda}H_F\right)e\right\|_{W_F}^2=\sum_{\ell=1}^{|F|}\left|1-\frac{\theta}{\lambda}\mu_\ell\right|^2|c_\ell|^2. ( I F − λ θ H F ) e W F 2 = ℓ = 1 ∑ ∣ F ∣ 1 − λ θ μ ℓ 2 ∣ c ℓ ∣ 2 .
Thus the free operator norm is (31). The condition ρ θ < 1 \rho_\theta<1 ρ θ < 1 is equivalent to requiring that, for every μ ∈ σ ( H F ) \mu\in\sigma(H_F) μ ∈ σ ( H F ) , we have − 1 < 1 − θ μ / λ < 1 -1<1-\theta\mu/\lambda<1 − 1 < 1 − θ μ / λ < 1 . Here σ ( H F ) \sigma(H_F) σ ( H F ) is the spectrum. Since μ min > 0 \mu_{\min}>0 μ m i n > 0 , this gives
ρ θ < 1 ⟺ 0 < θ < 2 λ μ max . \rho_\theta<1\quad\Longleftrightarrow\quad0<\theta<\frac{2\lambda}{\mu_{\max}}. ρ θ < 1 ⟺ 0 < θ < μ m a x 2 λ .
For ρ θ < ρ ′ < 1 \rho_\theta<\rho'<1 ρ θ < ρ ′ < 1 , continuity gives ε > 0 \varepsilon>0 ε > 0 such that, on the ball ∥ x ∥ W F < ε \|x\|_{W_F}<\varepsilon ∥ x ∥ W F < ε , we have ∥ D Ψ θ ( x ) ∥ W F ≤ ρ ′ \|D\Psi_\theta(x)\|_{W_F}\leq\rho' ∥ D Ψ θ ( x ) ∥ W F ≤ ρ ′ . For such x x x ,
∥ Ψ θ ( x ) ∥ W F = ∥ ∫ 0 1 D Ψ θ ( t x ) x d t ∥ W F ≤ ρ ′ ∥ x ∥ W F . \|\Psi_\theta(x)\|_{W_F}=\left\|\int_0^1D\Psi_\theta(tx)x\,dt\right\|_{W_F}\leq\rho'\|x\|_{W_F}. ∥ Ψ θ ( x ) ∥ W F = ∫ 0 1 D Ψ θ ( t x ) x d t W F ≤ ρ ′ ∥ x ∥ W F .
The update keeps this ball invariant. If the starting control is close enough to u ∗ u^* u ∗ , continuity puts the first iterate in the ball, and all later iterates satisfy
∥ u k + 1 − u ∗ ∥ h , τ ≤ ρ ′ ∥ u k − u ∗ ∥ h , τ , \|u_{k+1}-u^*\|_{h,\tau}\leq\rho'\|u_k-u^*\|_{h,\tau}, ∥ u k + 1 − u ∗ ∥ h , τ ≤ ρ ′ ∥ u k − u ∗ ∥ h , τ ,
which is local q q q -linear convergence. For θ > 2 λ / μ max \theta>2\lambda/\mu_{\max} θ > 2 λ / μ m a x , the mode with the largest eigenvalue has factor 1 − θ μ max / λ < − 1 1-\theta\mu_{\max}/\lambda<-1 1 − θ μ m a x / λ < − 1 . At equality, the factor is − 1 -1 − 1 and higher-order terms determine stability. The smallest spectral factor comes from balancing the factors at the two extreme eigenvalues:
1 − θ μ min λ = − 1 + θ μ max λ ⟹ θ R = 2 λ μ min + μ max . 1-\frac{\theta\mu_{\min}}{\lambda}=-1+\frac{\theta\mu_{\max}}{\lambda}
\quad\Longrightarrow\quad
\theta_R=\frac{2\lambda}{\mu_{\min}+\mu_{\max}}. 1 − λ θ μ m i n = − 1 + λ θ μ m a x ⟹ θ R = μ m i n + μ m a x 2 λ .
Substitution gives ρ θ R = ( μ max − μ min ) / ( μ max + μ min ) \rho_{\theta_R}=(\mu_{\max}-\mu_{\min})/(\mu_{\max}+\mu_{\min}) ρ θ R = ( μ m a x − μ m i n ) / ( μ m a x + μ m i n ) . Restricting θ ≤ 1 \theta\leq1 θ ≤ 1 gives min { 1 , θ R } \min\{1,\theta_R\} min { 1 , θ R } .
Ritz residual and spectral distance
Proof. Choose an orthonormal eigenbasis ( e ℓ ) ℓ = 1 ∣ F ∣ (e_\ell)_{\ell=1}^{|F|} ( e ℓ ) ℓ = 1 ∣ F ∣ of H ~ F \widetilde{H}_F H F , with eigenvalues μ ℓ \mu_\ell μ ℓ . For y = ∑ ℓ = 1 ∣ F ∣ a ℓ e ℓ y=\sum_{\ell=1}^{|F|}a_\ell e_\ell y = ∑ ℓ = 1 ∣ F ∣ a ℓ e ℓ , ∑ ℓ = 1 ∣ F ∣ a ℓ 2 = 1 \sum_{\ell=1}^{|F|}a_\ell^2=1 ∑ ℓ = 1 ∣ F ∣ a ℓ 2 = 1 ,
∥ H ~ F y − μ ^ y ∥ 2 2 = ∑ ℓ = 1 ∣ F ∣ a ℓ 2 ( μ ℓ − μ ^ ) 2 ≥ min μ ∈ σ ( H ~ F ) ∣ μ ^ − μ ∣ 2 . \|\widetilde{H}_Fy-\widehat{\mu}y\|_2^2=\sum_{\ell=1}^{|F|}a_\ell^2(\mu_\ell-\widehat{\mu})^2\geq\min_{\mu\in\sigma(\widetilde{H}_F)}|\widehat{\mu}-\mu|^2. ∥ H F y − μ y ∥ 2 2 = ℓ = 1 ∑ ∣ F ∣ a ℓ 2 ( μ ℓ − μ ) 2 ≥ μ ∈ σ ( H F ) min ∣ μ − μ ∣ 2 .
Thus
min μ ∈ σ ( H ~ F ) ∣ μ ^ − μ ∣ ≤ ∥ H ~ F y − μ ^ y ∥ 2 . \min_{\mu\in\sigma(\widetilde{H}_F)}|\widehat{\mu}-\mu|\leq\|\widetilde{H}_Fy-\widehat{\mu}y\|_2. μ ∈ σ ( H F ) min ∣ μ − μ ∣ ≤ ∥ H F y − μ y ∥ 2 .
Harvest-quadrature proofs
Removal identities
Proof. The removal step gives
y n − y n + 1 = τ q n y n + 1 , y k = y 0 ∏ n = 0 k − 1 ( 1 + τ q n ) − 1 . (68) y^n-y^{n+1}=\tau q^n y^{n+1},\qquad y^k=y^0\prod_{n=0}^{k-1}(1+\tau q^n)^{-1}. \tag*{(68)} y n − y n + 1 = τ q n y n + 1 , y k = y 0 n = 0 ∏ k − 1 ( 1 + τ q n ) − 1 . (68)
Summing over time gives
τ q ∑ n = 0 N − 1 y n y n + 1 = ∑ n = 0 N − 1 ( y n − y n + 1 ) = y 0 − y N , \tau q\sum_{n=0}^{N-1}y^n y^{n+1}=\sum_{n=0}^{N-1}(y^n-y^{n+1})=y^0-y^N, τ q n = 0 ∑ N − 1 y n y n + 1 = n = 0 ∑ N − 1 ( y n − y n + 1 ) = y 0 − y N ,
and
j 1 ( u ) − j 0 ( u ) = η τ q ∑ n = 0 N − 1 u n ( y n − y n + 1 ) = η τ 2 q 2 ∑ n = 0 N − 1 ( u n ) 2 y n + 1 . j_1(u)-j_0(u)=\eta\tau q\sum_{n=0}^{N-1}u^n(y^n-y^{n+1})=\eta\tau^2q^2\sum_{n=0}^{N-1}(u^n)^2y^{n+1}. j 1 ( u ) − j 0 ( u ) = η τ q n = 0 ∑ N − 1 u n ( y n − y n + 1 ) = η τ 2 q 2 n = 0 ∑ N − 1 ( u n ) 2 y n + 1 .
These are (23) and (24) .
Proof of Proposition 6.1
Proof. Differentiating the product formula for y N y^N y N gives, for 0 ≤ m , n < N 0\le m,n<N 0 ≤ m , n < N ,
∂ u n y N = − τ q y N 1 + τ q u n , ∂ u m 1 1 + τ q u n = − τ q δ m n ( 1 + τ q u n ) 2 . \partial_{u_n}y^N=-\frac{\tau qy^N}{1+\tau qu^n},\qquad\partial_{u_m}\frac{1}{1+\tau qu^n}=-\frac{\tau q\delta_{mn}}{(1+\tau qu^n)^2}. ∂ u n y N = − 1 + τ q u n τ q y N , ∂ u m 1 + τ q u n 1 = − ( 1 + τ q u n ) 2 τ q δ mn .
Here δ m n \delta_{mn} δ mn is the Kronecker delta. Applying the product rule once more gives
∂ u m ∂ u n y N = − τ q ( ∂ u m y N 1 + τ q u n − τ q δ m n y N ( 1 + τ q u n ) 2 ) = τ 2 q 2 y N ( 1 + δ m n ) ( 1 + τ q u m ) ( 1 + τ q u n ) . \partial_{u_m}\partial_{u_n}y^N=-\tau q\left(\frac{\partial_{u_m}y^N}{1+\tau qu^n}-\frac{\tau q\delta_{mn}y^N}{(1+\tau qu^n)^2}\right)=\frac{\tau^2q^2y^N(1+\delta_{mn})}{(1+\tau qu^m)(1+\tau qu^n)}. ∂ u m ∂ u n y N = − τ q ( 1 + τ q u n ∂ u m y N − ( 1 + τ q u n ) 2 τ q δ mn y N ) = ( 1 + τ q u m ) ( 1 + τ q u n ) τ 2 q 2 y N ( 1 + δ mn ) .
The weighted gradient is g 1 ( u ) n = τ − 1 ∂ u n j 1 ( u ) g_1(u)_n=\tau^{-1}\partial_{u_n}j_1(u) g 1 ( u ) n = τ − 1 ∂ u n j 1 ( u ) . Thus (23) gives
g 1 ( u ) n = λ u n − η q y N 1 + τ q u n , ( H 1 ( u ) ) m n = λ δ m n + η τ q 2 y N ( 1 + δ m n ) ( 1 + τ q u m ) ( 1 + τ q u n ) . g_1(u)_n=\lambda u^n-\frac{\eta qy^N}{1+\tau qu^n},\qquad(H_1(u))_{mn}=\lambda\delta_{mn}+\frac{\eta\tau q^2y^N(1+\delta_{mn})}{(1+\tau qu^m)(1+\tau qu^n)}. g 1 ( u ) n = λ u n − 1 + τ q u n η q y N , ( H 1 ( u ) ) mn = λ δ mn + ( 1 + τ q u m ) ( 1 + τ q u n ) η τ q 2 y N ( 1 + δ mn ) .
For ξ ∈ R N \xi\in\mathbb{R}^N ξ ∈ R N , the two terms in 1 + δ m n 1+\delta_{mn} 1 + δ mn yield
⟨ H 1 ( u ) ξ , ξ ⟩ τ = λ ∥ ξ ∥ τ 2 + η τ 2 q 2 y N [ ( ∑ n = 0 N − 1 ξ n 1 + τ q u n ) 2 + ∑ n = 0 N − 1 ( ξ n 1 + τ q u n ) 2 ] ≥ λ ∥ ξ ∥ τ 2 . \langle H_1(u)\xi,\xi\rangle_\tau=\lambda\|\xi\|_\tau^2+\eta\tau^2q^2y^N\left[\left(\sum_{n=0}^{N-1}\frac{\xi^n}{1+\tau qu^n}\right)^2+\sum_{n=0}^{N-1}\left(\frac{\xi^n}{1+\tau qu^n}\right)^2\right]\ge\lambda\|\xi\|_\tau^2. ⟨ H 1 ( u ) ξ , ξ ⟩ τ = λ ∥ ξ ∥ τ 2 + η τ 2 q 2 y N ( n = 0 ∑ N − 1 1 + τ q u n ξ n ) 2 + n = 0 ∑ N − 1 ( 1 + τ q u n ξ n ) 2 ≥ λ ∥ ξ ∥ τ 2 .
Permuting the entries leaves the unique minimizer unchanged: u = s 1 u=s\mathbf{1} u = s 1 . Along this line,
d d s j 1 ( s 1 ) = N τ [ λ s − η q y 0 ( 1 + τ q s ) − N − 1 ] . \frac{\mathrm{d}}{\mathrm{d}s}j_1(s\mathbf{1})=N\tau\left[\lambda s-\eta qy^0(1+\tau qs)^{-N-1}\right]. d s d j 1 ( s 1 ) = N τ [ λ s − η q y 0 ( 1 + τ q s ) − N − 1 ] .
The bracketed expression is strictly increasing. Its nonnegative root determines s s s , capped at U max U_{\max} U m a x . When η = 0 \eta=0 η = 0 , this gives s = 0 s=0 s = 0 . To compute the old-state Hessian at zero, use that y n y^n y n depends only on u 0 , … , u n − 1 u^0,\ldots,u^{n-1} u 0 , … , u n − 1 :
∂ u m y n ( 0 ) = { − τ q y 0 , m < n , 0 , m ≥ n . \partial_{u_m}y^n(0)=
\begin{cases}
-\tau qy^0,&m<n,\\
0,&m\ge n.
\end{cases} ∂ u m y n ( 0 ) = { − τ q y 0 , 0 , m < n , m ≥ n .
Twice differentiating − η τ q ∑ n = 0 N − 1 u n y n -\eta\tau q\sum_{n=0}^{N-1}u^ny^n − η τ q ∑ n = 0 N − 1 u n y n and evaluating at zero gives, for m ≠ ℓ m\ne\ell m = ℓ ,
∂ u m ∂ u ℓ j 0 ( 0 ) = − η τ q ( ∂ u ℓ y m ( 0 ) + ∂ u m y ℓ ( 0 ) ) = η τ 2 q 2 y 0 . \partial_{u_m}\partial_{u_\ell}j_0(0)=-\eta\tau q\left(\partial_{u_\ell}y^m(0)+\partial_{u_m}y^\ell(0)\right)=\eta\tau^2q^2y^0. ∂ u m ∂ u ℓ j 0 ( 0 ) = − η τ q ( ∂ u ℓ y m ( 0 ) + ∂ u m y ℓ ( 0 ) ) = η τ 2 q 2 y 0 .
For m = ℓ m=\ell m = ℓ , the reward contributes zero and the effort contributes λ τ \lambda\tau λ τ . Dividing by the time weight τ \tau τ yields
H 0 ( 0 ) = λ I + η τ q 2 y 0 ( 11 T − I ) , H 1 ( 0 ) = λ I + η τ q 2 y 0 ( 11 T + I ) . H_0(0)=\lambda I+\eta\tau q^2y^0(\mathbf{1}\mathbf{1}^{\mathsf T}-I),\qquad H_1(0)=\lambda I+\eta\tau q^2y^0(\mathbf{1}\mathbf{1}^{\mathsf T}+I). H 0 ( 0 ) = λ I + η τ q 2 y 0 ( 1 1 T − I ) , H 1 ( 0 ) = λ I + η τ q 2 y 0 ( 1 1 T + I ) .
Thus, at zero,
H 1 ( 0 ) − H 0 ( 0 ) = 2 η τ q 2 y 0 I . H_1(0)-H_0(0)=2\eta\tau q^2y^0I. H 1 ( 0 ) − H 0 ( 0 ) = 2 η τ q 2 y 0 I .
For ξ ≠ 0 \xi\ne0 ξ = 0 with 1 T ξ = 0 \mathbf{1}^{\mathsf T}\xi=0 1 T ξ = 0 , multiplication by the rank-one matrix gives 11 T ξ = 0 \mathbf{1}\mathbf{1}^{\mathsf T}\xi=0 1 1 T ξ = 0 . Hence
⟨ H 0 ( 0 ) ξ , ξ ⟩ τ ∥ ξ ∥ τ 2 = λ − η τ q 2 y 0 , ⟨ H 1 ( 0 ) ξ , ξ ⟩ τ ∥ ξ ∥ τ 2 = λ + η τ q 2 y 0 . \frac{\langle H_0(0)\xi,\xi\rangle_\tau}{\|\xi\|_\tau^2}=\lambda-\eta\tau q^2y^0,\qquad\frac{\langle H_1(0)\xi,\xi\rangle_\tau}{\|\xi\|_\tau^2}=\lambda+\eta\tau q^2y^0. ∥ ξ ∥ τ 2 ⟨ H 0 ( 0 ) ξ , ξ ⟩ τ = λ − η τ q 2 y 0 , ∥ ξ ∥ τ 2 ⟨ H 1 ( 0 ) ξ , ξ ⟩ τ = λ + η τ q 2 y 0 .
Negative curvature persists by continuity at u = s 1 u=s\mathbf{1} u = s 1 for small 0 < s < U max 0<s<U_{\max} 0 < s < U m a x . Both u ± ε ξ u\pm\varepsilon\xi u ± ε ξ are then feasible for small ε > 0 \varepsilon>0 ε > 0 , so j 0 j_0 j 0 is nonconvex on [ 0 , U max ] N [0,U_{\max}]^N [ 0 , U m a x ] N .
Refinement bound in the pure-removal model
Proof. For ξ , χ ∈ R N \xi,\chi\in\mathbb{R}^{N} ξ , χ ∈ R N , set z [ ξ ] k : = D y k ( u ) [ ξ ] z[\xi]^k:=D y^k(u)[\xi] z [ ξ ] k := D y k ( u ) [ ξ ] , s [ ξ , χ ] k : = D 2 y k ( u ) [ ξ , χ ] s[\xi,\chi]^k:=D^2y^k(u)[\xi,\chi] s [ ξ , χ ] k := D 2 y k ( u ) [ ξ , χ ] . Differentiating the product formula for y k y^k y k gives
z [ ξ ] k = − τ q y k ∑ n = 0 k − 1 ξ n 1 + τ q u n . z[\xi]^k=-\tau q y^k\sum_{n=0}^{k-1}\frac{\xi^n}{1+\tau q u^n}. z [ ξ ] k = − τ q y k n = 0 ∑ k − 1 1 + τ q u n ξ n .
and
s [ ξ , χ ] k = τ 2 q 2 y k [ ( ∑ n = 0 k − 1 ξ n 1 + τ q u n ) ( ∑ n = 0 k − 1 χ n 1 + τ q u n ) + ∑ n = 0 k − 1 ξ n χ n ( 1 + τ q u n ) 2 ] . s[\xi,\chi]^k=\tau^2q^2y^k\left[\left(\sum_{n=0}^{k-1}\frac{\xi^n}{1+\tau q u^n}\right)\left(\sum_{n=0}^{k-1}\frac{\chi^n}{1+\tau q u^n}\right)+\sum_{n=0}^{k-1}\frac{\xi^n\chi^n}{(1+\tau q u^n)^2}\right]. s [ ξ , χ ] k = τ 2 q 2 y k [ ( n = 0 ∑ k − 1 1 + τ q u n ξ n ) ( n = 0 ∑ k − 1 1 + τ q u n χ n ) + n = 0 ∑ k − 1 ( 1 + τ q u n ) 2 ξ n χ n ] .
Since u n ≥ 0 u^n\geq0 u n ≥ 0 and k τ ≤ T k\tau\leq T k τ ≤ T , Cauchy–Schwarz gives
τ ∑ n = 0 k − 1 ∣ ξ n ∣ 1 + τ q u n ≤ k τ ( τ ∑ n = 0 k − 1 ∣ ξ n ∣ 2 ) 1 / 2 ≤ T ∥ ξ ∥ τ , τ 2 ∑ n = 0 k − 1 ∣ ξ n χ n ∣ ( 1 + τ q u n ) 2 ≤ τ ∥ ξ ∥ τ ∥ χ ∥ τ . \tau\sum_{n=0}^{k-1}\frac{|\xi^n|}{1+\tau q u^n}\leq\sqrt{k\tau}\left(\tau\sum_{n=0}^{k-1}|\xi^n|^2\right)^{1/2}\leq\sqrt{T}\|\xi\|_\tau,\qquad\tau^2\sum_{n=0}^{k-1}\frac{|\xi^n\chi^n|}{(1+\tau q u^n)^2}\leq\tau\|\xi\|_\tau\|\chi\|_\tau. τ n = 0 ∑ k − 1 1 + τ q u n ∣ ξ n ∣ ≤ k τ ( τ n = 0 ∑ k − 1 ∣ ξ n ∣ 2 ) 1/2 ≤ T ∥ ξ ∥ τ , τ 2 n = 0 ∑ k − 1 ( 1 + τ q u n ) 2 ∣ ξ n χ n ∣ ≤ τ ∥ ξ ∥ τ ∥ χ ∥ τ .
Combining these estimates with 0 < y k ≤ y 0 0<y^k\leq y^0 0 < y k ≤ y 0 gives
max 0 ≤ k ≤ N ∣ z [ ξ ] k ∣ ≤ q y 0 T ∥ ξ ∥ τ , (69) \max_{0\leq k\leq N}|z[\xi]^k|\leq qy^0\sqrt{T}\|\xi\|_\tau, \tag*{(69)} 0 ≤ k ≤ N max ∣ z [ ξ ] k ∣ ≤ q y 0 T ∥ ξ ∥ τ , (69)
and
max 0 ≤ k ≤ N ∣ s [ ξ , χ ] k ∣ ≤ q 2 y 0 ( T + τ ) ∥ ξ ∥ τ ∥ χ ∥ τ . (70) \max_{0\leq k\leq N}|s[\xi,\chi]^k|\leq q^2y^0(T+\tau)\|\xi\|_\tau\|\chi\|_\tau. \tag*{(70)} 0 ≤ k ≤ N max ∣ s [ ξ , χ ] k ∣ ≤ q 2 y 0 ( T + τ ) ∥ ξ ∥ τ ∥ χ ∥ τ . (70)
Differentiating each factor in the defect twice gives
D 2 ( j 1 − j 0 ) ( u ) [ ξ , χ ] = {xref}‘(24)‘ η τ q 2 ∑ n = 0 N − 1 [ 2 ξ n χ n y n + 1 + 2 u n ξ n z [ χ ] n + 1 + 2 u n χ n z [ ξ ] n + 1 + ( u n ) 2 s [ ξ , χ ] n + 1 ] . D^2(j_1-j_0)(u)[\xi,\chi]\overset{\text{\{xref\}`(24)`}}{=}\eta\tau q^2\sum_{n=0}^{N-1}\left[2\xi^n\chi^n y^{n+1}+2u^n\xi^n z[\chi]^{n+1}+2u^n\chi^n z[\xi]^{n+1}+(u^n)^2s[\xi,\chi]^{n+1}\right]. D 2 ( j 1 − j 0 ) ( u ) [ ξ , χ ] = {xref}‘(24)‘ η τ q 2 n = 0 ∑ N − 1 [ 2 ξ n χ n y n + 1 + 2 u n ξ n z [ χ ] n + 1 + 2 u n χ n z [ ξ ] n + 1 + ( u n ) 2 s [ ξ , χ ] n + 1 ] .
Use u n ≤ U max u^n\leq U_{\max} u n ≤ U m a x , τ ∑ n = 0 N − 1 ∣ ξ n ∣ ≤ T ∥ ξ ∥ τ \tau\sum_{n=0}^{N-1}|\xi^n|\leq\sqrt{T}\|\xi\|_\tau τ ∑ n = 0 N − 1 ∣ ξ n ∣ ≤ T ∥ ξ ∥ τ , and (69) –(70) . The first term, the two cross terms and the last term give, respectively,
∣ D 2 ( j 1 − j 0 ) ( u ) [ ξ , χ ] ∣ ≤ η τ q 2 y 0 [ 2 + 4 q U max T + q 2 U max 2 T ( T + τ ) ] ∥ ξ ∥ τ ∥ χ ∥ τ . |D^2(j_1-j_0)(u)[\xi,\chi]|\leq\eta\tau q^2y^0\left[2+4qU_{\max}T+q^2U_{\max}^2T(T+\tau)\right]\|\xi\|_\tau\|\chi\|_\tau. ∣ D 2 ( j 1 − j 0 ) ( u ) [ ξ , χ ] ∣ ≤ η τ q 2 y 0 [ 2 + 4 q U m a x T + q 2 U m a x 2 T ( T + τ ) ] ∥ ξ ∥ τ ∥ χ ∥ τ .
Taking the supremum over unit directions and using τ ≤ T \tau\leq T τ ≤ T gives
∥ H 1 ( u ) − H 0 ( u ) ∥ L ( R N ) ≤ η τ q 2 y 0 ( 2 + 4 q U max T + 2 q 2 U max 2 T 2 ) . \|H_1(u)-H_0(u)\|_{\mathcal{L}(\mathbb{R}^N)}\leq\eta\tau q^2y^0\left(2+4qU_{\max}T+2q^2U_{\max}^2T^2\right). ∥ H 1 ( u ) − H 0 ( u ) ∥ L ( R N ) ≤ η τ q 2 y 0 ( 2 + 4 q U m a x T + 2 q 2 U m a x 2 T 2 ) .
For fixed parameters, the bound falls below λ \lambda λ on all sufficiently fine time grids, making j 0 j_0 j 0 strictly convex.
Proof of Proposition 6.2
Proof. Multiplying the cost by τ q 2 / λ > 0 \tau q^2/\lambda>0 τ q 2 / λ > 0 gives, on [ 0 , τ q U max ] 2 [0,\tau qU_{\max}]^2 [ 0 , τ q U m a x ] 2 ,
F 0 ( x , z ) = 1 2 ( x 2 + z 2 ) − 2 ( x + z 1 + x ) . F_0(x,z)=\frac{1}{2}(x^2+z^2)-2\left(x+\frac{z}{1+x}\right). F 0 ( x , z ) = 2 1 ( x 2 + z 2 ) − 2 ( x + 1 + x z ) .
Its derivatives are
∂ x F 0 = x − 2 + 2 z ( 1 + x ) 2 , ∂ z F 0 = z − 2 1 + x , ∂ z z F 0 = 1. \partial_xF_0=x-2+\frac{2z}{(1+x)^2},\qquad\partial_zF_0=z-\frac{2}{1+x},\qquad\partial_{zz}F_0=1. ∂ x F 0 = x − 2 + ( 1 + x ) 2 2 z , ∂ z F 0 = z − 1 + x 2 , ∂ z z F 0 = 1.
At ( 0 , 2 ) (0,2) ( 0 , 2 ) , ∂ x F 0 = 2 \partial_xF_0=2 ∂ x F 0 = 2 and ∂ z F 0 = 0 \partial_zF_0=0 ∂ z F 0 = 0 . By continuity, choose a feasible neighborhood in which ∂ x F 0 ≥ 1 \partial_xF_0\geq1 ∂ x F 0 ≥ 1 . For points in a smaller rectangular neighborhood, the segment from ( 0 , z ) (0,z) ( 0 , z ) to ( x , z ) (x,z) ( x , z ) stays in that set, and
F 0 ( x , z ) − F 0 ( 0 , z ) = ∫ 0 x ∂ x F 0 ( s , z ) d s ≥ x . F_0(x,z)-F_0(0,z)=\int_0^x\partial_xF_0(s,z)\,\mathrm{d}s\geq x. F 0 ( x , z ) − F 0 ( 0 , z ) = ∫ 0 x ∂ x F 0 ( s , z ) d s ≥ x .
Since F 0 ( 0 , z ) = 1 2 z 2 − 2 z = − 2 + 1 2 ( z − 2 ) 2 F_0(0,z)=\frac{1}{2}z^2-2z=-2+\frac{1}{2}(z-2)^2 F 0 ( 0 , z ) = 2 1 z 2 − 2 z = − 2 + 2 1 ( z − 2 ) 2 , we obtain
F 0 ( x , z ) − F 0 ( 0 , 2 ) ≥ x + 1 2 ( z − 2 ) 2 > 0 F_0(x,z)-F_0(0,2)\ge x+\frac{1}{2}(z-2)^2>0 F 0 ( x , z ) − F 0 ( 0 , 2 ) ≥ x + 2 1 ( z − 2 ) 2 > 0
for every nearby feasible ( x , z ) ≠ ( 0 , 2 ) (x,z)\ne(0,2) ( x , z ) = ( 0 , 2 ) . In particular,
F 0 ( x , z ) ≥ F 0 ( 0 , z ) = − 2 + 1 2 ( z − 2 ) 2 , F_0(x,z)\ge F_0(0,z)=-2+\frac{1}{2}(z-2)^2, F 0 ( x , z ) ≥ F 0 ( 0 , z ) = − 2 + 2 1 ( z − 2 ) 2 ,
with equality in the first inequality only for x = 0 x=0 x = 0 . Thus ( 0 , 2 ) (0,2) ( 0 , 2 ) is a strict local minimum, but the feasible point ( 2 , 2 / 3 ) (2,2/3) ( 2 , 2/3 ) has lower cost:
F 0 ( 2 , 2 / 3 ) = − 20 / 9 < − 2 = F 0 ( 0 , 2 ) . F_0(2,2/3)=-20/9<-2=F_0(0,2). F 0 ( 2 , 2/3 ) = − 20/9 < − 2 = F 0 ( 0 , 2 ) .
Proposition 6.1 gives strict convexity for the flux cost.
Proof of Proposition 6.3
Proof. Rewrite the old-state reward using
a + E − a 1 + τ q a = E − τ q a ( E − a ) 1 + τ q a . a+\frac{E-a}{1+\tau qa}=E-\frac{\tau qa(E-a)}{1+\tau qa}. a + 1 + τ q a E − a = E − 1 + τ q a τ q a ( E − a ) .
The quotient rule gives
d d a a ( E − a ) 1 + τ q a = ( E − 2 a ) ( 1 + τ q a ) − τ q a ( E − a ) ( 1 + τ q a ) 2 = E − 2 a − τ q a 2 ( 1 + τ q a ) 2 . \frac{\mathrm{d}}{\mathrm{d}a}\frac{a(E-a)}{1+\tau qa}
=\frac{(E-2a)(1+\tau qa)-\tau qa(E-a)}{(1+\tau qa)^2}
=\frac{E-2a-\tau qa^2}{(1+\tau qa)^2}. d a d 1 + τ q a a ( E − a ) = ( 1 + τ q a ) 2 ( E − 2 a ) ( 1 + τ q a ) − τ q a ( E − a ) = ( 1 + τ q a ) 2 E − 2 a − τ q a 2 .
Differentiating again, and using a ≤ E a\le E a ≤ E and the assumed bound on λ \lambda λ , yields
f 0 ′ ′ ( a ) = 2 λ τ − 2 η τ 2 q 2 y 0 ( 1 + τ q E ) ( 1 + τ q a ) 3 ≤ 2 λ τ − 2 η τ 2 q 2 y 0 ( 1 + τ q E ) 2 < 0. f_0''(a)=2\lambda\tau-\frac{2\eta\tau^2q^2y^0(1+\tau qE)}{(1+\tau qa)^3}
\le2\lambda\tau-\frac{2\eta\tau^2q^2y^0}{(1+\tau qE)^2}<0. f 0 ′′ ( a ) = 2 λ τ − ( 1 + τ q a ) 3 2 η τ 2 q 2 y 0 ( 1 + τ q E ) ≤ 2 λ τ − ( 1 + τ q E ) 2 2 η τ 2 q 2 y 0 < 0.
For the flux reward, expand the denominator as
( 1 + τ q a ) ( 1 + τ q ( E − a ) ) = 1 + τ q E + τ 2 q 2 a ( E − a ) . (1+\tau qa)(1+\tau q(E-a))=1+\tau qE+\tau^2q^2a(E-a). ( 1 + τ q a ) ( 1 + τ q ( E − a )) = 1 + τ q E + τ 2 q 2 a ( E − a ) .
Its first and second derivatives are τ 2 q 2 ( E − 2 a ) \tau^2q^2(E-2a) τ 2 q 2 ( E − 2 a ) and − 2 τ 2 q 2 -2\tau^2q^2 − 2 τ 2 q 2 . The reciprocal rule therefore gives
f 1 ′ ′ ( a ) = 2 λ τ + 2 η τ 2 q 2 y 0 [ 1 + τ q E + τ 2 q 2 a ( E − a ) ] 2 + 2 η τ 4 q 4 y 0 ( E − 2 a ) 2 [ 1 + τ q E + τ 2 q 2 a ( E − a ) ] 3 > 0. f_1''(a)=2\lambda\tau+
\frac{2\eta\tau^2q^2y^0}{[1+\tau qE+\tau^2q^2a(E-a)]^2}
+\frac{2\eta\tau^4q^4y^0(E-2a)^2}{[1+\tau qE+\tau^2q^2a(E-a)]^3}>0. f 1 ′′ ( a ) = 2 λ τ + [ 1 + τ q E + τ 2 q 2 a ( E − a ) ] 2 2 η τ 2 q 2 y 0 + [ 1 + τ q E + τ 2 q 2 a ( E − a ) ] 3 2 η τ 4 q 4 y 0 ( E − 2 a ) 2 > 0.
Since f 1 ( a ) = f 1 ( E − a ) f_1(a)=f_1(E-a) f 1 ( a ) = f 1 ( E − a ) , strict convexity puts its minimum at a = E / 2 a=E/2 a = E /2 . The strictly concave function f 0 f_0 f 0 has equal minima at the two endpoints. For the continuous control u a ( t ) : = a u_a(t):=a u a ( t ) := a on [ 0 , τ ) [0,\tau) [ 0 , τ ) , u a ( t ) : = E − a u_a(t):=E-a u a ( t ) := E − a on [ τ , 2 τ ] [\tau,2\tau] [ τ , 2 τ ] ,
∫ 0 2 τ u a d t = τ E , j c ( u a ) = λ τ 2 [ a 2 + ( E − a ) 2 ] − η y 0 ( 1 − e − τ q E ) . \int_0^{2\tau}u_a\,\mathrm{d}t=\tau E,\qquad
j_c(u_a)=\frac{\lambda\tau}{2}[a^2+(E-a)^2]-\eta y^0(1-e^{-\tau qE}). ∫ 0 2 τ u a d t = τ E , j c ( u a ) = 2 λ τ [ a 2 + ( E − a ) 2 ] − η y 0 ( 1 − e − τ q E ) .
The effort term alone depends on a a a , with its minimum at $E/2.
Sign of the coupled defect
Proof. In (25) , the harvest and predation terms are nonnegative, the growth term is nonpositive, and
− μ 2 ⟨ u n , L h v 2 n + 1 ⟩ h = μ 2 h ∑ j = 0 M − 1 ( u j + 1 n − u j n ) ( v 2 , j + 1 n + 1 − v 2 , j n + 1 ) -\mu_2\langle u^n,L_hv_2^{n+1}\rangle_h
=\frac{\mu_2}{h}\sum_{j=0}^{M-1}(u_{j+1}^n-u_j^n)(v_{2,j+1}^{n+1}-v_{2,j}^{n+1}) − μ 2 ⟨ u n , L h v 2 n + 1 ⟩ h = h μ 2 j = 0 ∑ M − 1 ( u j + 1 n − u j n ) ( v 2 , j + 1 n + 1 − v 2 , j n + 1 )
can have either sign, as can the total defect. To see this, take one step with constant data v 1 0 = a 1 v_1^0=a\mathbf{1} v 1 0 = a 1 , v 2 0 = b 1 v_2^0=b\mathbf{1} v 2 0 = b 1 , u 0 = c 1 u^0=c\mathbf{1} u 0 = c 1 , where a ≥ 0 a\ge0 a ≥ 0 , b > 0 b>0 b > 0 , 0 < c ≤ U max 0<c\le U_{\max} 0 < c ≤ U m a x . Since L h 1 = 0 L_h\mathbf{1}=0 L h 1 = 0 , ⟨ 1 , 1 ⟩ h = 1 \langle\mathbf{1},\mathbf{1}\rangle_h=1 ⟨ 1 , 1 ⟩ h = 1 , the prey step gives
v 2 1 = ( 1 + τ r ) b 1 + τ ( β a + q c ) 1 (71) v_2^1=\frac{(1+\tau r)b}{1+\tau(\beta a+qc)}\mathbf{1} \tag*{(71)} v 2 1 = 1 + τ ( β a + q c ) ( 1 + τ r ) b 1 (71)
so
D ( u ) = η q τ c ( b − ( 1 + τ r ) b 1 + τ ( β a + q c ) ) = η q τ 2 c b ( β a + q c − r ) 1 + τ ( β a + q c ) . D(u)=\eta q\tau c\left(b-\frac{(1+\tau r)b}{1+\tau(\beta a+qc)}\right)=\frac{\eta q\tau^{2}cb(\beta a+qc-r)}{1+\tau(\beta a+qc)}. D ( u ) = η q τ c ( b − 1 + τ ( β a + q c ) ( 1 + τ r ) b ) = 1 + τ ( β a + q c ) η q τ 2 c b ( β a + q c − r ) .
For η > 0 \eta>0 η > 0 , choose 0 < c < min { U max , r / q } 0<c<\min\{U_{\max},r/q\} 0 < c < min { U m a x , r / q } . Then D ( u ) < 0 D(u)<0 D ( u ) < 0 when a = 0 a=0 a = 0 , while D ( u ) > 0 D(u)>0 D ( u ) > 0 when a > r / β a>r/\beta a > r / β . For η = 0 \eta=0 η = 0 , D = 0 D=0 D = 0 .
Exact global certificate for the two-step problem
Proof. For F 0 F_0 F 0 in § 11.1 , ∂ z F 0 = z − 2 / ( 1 + x ) \partial_zF_0=z-2/(1+x) ∂ z F 0 = z − 2/ ( 1 + x ) and ∂ z z F 0 = 1 \partial_{zz}F_0=1 ∂ z z F 0 = 1 . Thus, at each x ∈ [ 0 , 3 ] x\in[0,3] x ∈ [ 0 , 3 ] , the unique minimizing value is z = 2 / ( 1 + x ) ∈ [ 1 / 2 , 2 ] z=2/(1+x)\in[1/2,2] z = 2/ ( 1 + x ) ∈ [ 1/2 , 2 ] . Substitution reduces the problem to
φ 0 : [ 0 , 3 ] ⟶ R \varphi_0:[0,3]\longrightarrow\mathbb{R} φ 0 : [ 0 , 3 ] ⟶ R
x ⟼ F 0 ( x , 2 1 + x ) = 1 2 x 2 − 2 x − 2 ( 1 + x ) 2 . x\longmapsto F_0\left(x,\frac{2}{1+x}\right)=\frac{1}{2}x^2-2x-\frac{2}{(1+x)^2}. x ⟼ F 0 ( x , 1 + x 2 ) = 2 1 x 2 − 2 x − ( 1 + x ) 2 2 .
Set
P : R ⟶ R P:\mathbb{R}\longrightarrow\mathbb{R} P : R ⟶ R
x ⟼ x 4 + x 3 − 3 x 2 − 5 x + 2. x\longmapsto x^4+x^3-3x^2-5x+2. x ⟼ x 4 + x 3 − 3 x 2 − 5 x + 2.
Then
φ 0 ′ ( x ) = x − 2 + 4 ( 1 + x ) 3 = P ( x ) ( 1 + x ) 3 . \varphi_0'(x)=x-2+\frac{4}{(1+x)^3}=\frac{P(x)}{(1+x)^3}. φ 0 ′ ( x ) = x − 2 + ( 1 + x ) 3 4 = ( 1 + x ) 3 P ( x ) .
The rational Sturm sequence P 0 : = P P_0:=P P 0 := P , P 1 : = P ′ P_1:=P' P 1 := P ′ , P k + 1 : = − rem ( P k − 1 , P k ) P_{k+1}:=-\operatorname{rem}(P_{k-1},P_k) P k + 1 := − rem ( P k − 1 , P k ) , where rem ( A , B ) \operatorname{rem}(A,B) rem ( A , B ) is the remainder of polynomial division of A A A by B B B , is
P 1 ( x ) = 4 x 3 + 3 x 2 − 6 x − 5 , P 2 ( x ) = 27 x 2 + 54 x − 37 16 , P 3 ( x ) = 320 − 256 x 27 , P 4 ( x ) = − 1163 256 . P_1(x)=4x^3+3x^2-6x-5,\qquad P_2(x)=\frac{27x^2+54x-37}{16},\qquad P_3(x)=\frac{320-256x}{27},\qquad P_4(x)=-\frac{1163}{256}. P 1 ( x ) = 4 x 3 + 3 x 2 − 6 x − 5 , P 2 ( x ) = 16 27 x 2 + 54 x − 37 , P 3 ( x ) = 27 320 − 256 x , P 4 ( x ) = − 256 1163 .
Define
V : R ⟶ { 0 , 1 , 2 , 3 , 4 } V:\mathbb{R}\longrightarrow\{0,1,2,3,4\} V : R ⟶ { 0 , 1 , 2 , 3 , 4 }
a ⟼ sign changes in ( P 0 ( a ) , … , P 4 ( a ) ) ′ a\longmapsto\text{sign changes in }(P_0(a),\ldots,P_4(a))' a ⟼ sign changes in ( P 0 ( a ) , … , P 4 ( a ) ) ′
with zeros omitted. We have V ( 0 ) = 3 V(0)=3 V ( 0 ) = 3 and V ( 3 ) = 1 V(3)=1 V ( 3 ) = 1 , so there are exactly two interior stationary points. Rational root isolation places one in each interval
I 1 : = [ 0.3408975428038 , 0.3408975428039 ] , I 2 : = [ 1.8220156900313 , 1.8220156900314 ] . I_1:=[0.3408975428038,0.3408975428039],\qquad I_2:=[1.8220156900313,1.8220156900314]. I 1 := [ 0.3408975428038 , 0.3408975428039 ] , I 2 := [ 1.8220156900313 , 1.8220156900314 ] .
For 0 ≤ a ≤ x ≤ b 0\leq a\leq x\leq b 0 ≤ a ≤ x ≤ b , monotonicity of each term gives
a 2 2 − 2 b − 2 ( 1 + a ) 2 ≤ φ 0 ( x ) ≤ b 2 2 − 2 a − 2 ( 1 + b ) 2 . \frac{a^2}{2}-2b-\frac{2}{(1+a)^2}\leq\varphi_0(x)\leq\frac{b^2}{2}-2a-\frac{2}{(1+b)^2}. 2 a 2 − 2 b − ( 1 + a ) 2 2 ≤ φ 0 ( x ) ≤ 2 b 2 − 2 a − ( 1 + b ) 2 2 .
On I 1 , I 2 I_1,I_2 I 1 , I 2 , these bounds give
φ 0 ( I 1 ) ⊂ [ − 1.7360327224 , − 1.7360327223 ] , φ 0 ( I 2 ) ⊂ [ − 2.2352980504 , − 2.2352980503 ] . \varphi_0(I_1)\subset[-1.7360327224,-1.7360327223],\qquad\varphi_0(I_2)\subset[-2.2352980504,-2.2352980503]. φ 0 ( I 1 ) ⊂ [ − 1.7360327224 , − 1.7360327223 ] , φ 0 ( I 2 ) ⊂ [ − 2.2352980504 , − 2.2352980503 ] .
All interval endpoints are exact rationals. The Sturm variations at the isolating endpoints satisfy
V ( 0.3408975428038 ) = 3 , V ( 0.3408975428039 ) = 2 , V ( 1.8220156900313 ) = 2 , V ( 1.8220156900314 ) = 1. V(0.3408975428038)=3,\qquad V(0.3408975428039)=2,\qquad V(1.8220156900313)=2,\qquad V(1.8220156900314)=1. V ( 0.3408975428038 ) = 3 , V ( 0.3408975428039 ) = 2 , V ( 1.8220156900313 ) = 2 , V ( 1.8220156900314 ) = 1.
Each interval therefore contains exactly one root, and V ( 0 ) − V ( 3 ) = 2 V(0)-V(3)=2 V ( 0 ) − V ( 3 ) = 2 excludes any other interior critical point. Comparing these bounds with φ 0 ( 0 ) = − 2 \varphi_0(0)=-2 φ 0 ( 0 ) = − 2 and φ 0 ( 3 ) = − 13 / 8 \varphi_0(3)=-13/8 φ 0 ( 3 ) = − 13/8 proves that the stationary point in I 2 I_2 I 2 is the unique global minimizer. Its control is approximately ( 1.822 , 0.7087 ) (1.822,0.7087) ( 1.822 , 0.7087 ) , with F 0 ≈ − 2.235 F_0\approx-2.235 F 0 ≈ − 2.235 . Appendix § 20 .§ 4 proves that ( 0 , 2 ) (0,2) ( 0 , 2 ) is a strict local minimum. For F 1 F_1 F 1 , strict convexity and symmetry give x = z = t x=z=t x = z = t . Then
t − 2 ( 1 + t ) 3 = 0 ⟺ t ( 1 + t ) 3 = 2. (72) t-\frac{2}{(1+t)^3}=0\quad\Longleftrightarrow\quad t(1+t)^3=2. \tag*{(72)} t − ( 1 + t ) 3 2 = 0 ⟺ t ( 1 + t ) 3 = 2. (72)
Rational root isolation gives
t ∈ [ 0.5436890126920 , 0.5436890126921 ] , t \in[0.5436890126920, 0.5436890126921], t ∈ [ 0.5436890126920 , 0.5436890126921 ] ,
so t ≈ 0.5437 t \approx0.5437 t ≈ 0.5437 and F 1 ≈ − 0.8651 F_1 \approx-0.8651 F 1 ≈ − 0.8651 . For the continuous cost j c j_c j c on K c K_c K c from § 11.1 , take u ∈ K c u \in K_c u ∈ K c and set s : = 1 2 ∫ 0 2 u s := \frac{1}{2}\int_0^2 u s := 2 1 ∫ 0 2 u .
Jensen’s inequality gives 1 2 ∫ 0 2 u 2 ≥ s 2 \frac{1}{2}\int_0^2 u^2 \ge s^2 2 1 ∫ 0 2 u 2 ≥ s 2 , with equality only for u ≡ s u \equiv s u ≡ s . Define
f c : [ 0 , 3 ] ⟶ R s ⟼ s 2 − 2 + 2 e − 2 s . \begin{aligned}
f_c : [0,3] &\longrightarrow\mathbb{R}\\
s &\longmapsto s^2-2+2e^{-2s}.
\end{aligned} f c : [ 0 , 3 ] s ⟶ R ⟼ s 2 − 2 + 2 e − 2 s .
Then f c ′ ( s ) = 2 s − 4 e − 2 s f_c'(s)=2s-4e^{-2s} f c ′ ( s ) = 2 s − 4 e − 2 s , f c ′ ′ ( s ) = 2 + 8 e − 2 s > 0 f_c''(s)=2+8e^{-2s}>0 f c ′′ ( s ) = 2 + 8 e − 2 s > 0 . The unique minimizer is s ≈ 0.6011 s\approx0.6011 s ≈ 0.6011 , with cost approximately − 1.038 -1.038 − 1.038 . For two-step controls, j c ( x , z ) = 1 2 ( x 2 + z 2 ) − 2 + 2 e − ( x + z ) j_c(x,z)=\frac{1}{2}(x^2+z^2)-2+2e^{-(x+z)} j c ( x , z ) = 2 1 ( x 2 + z 2 ) − 2 + 2 e − ( x + z ) . The old-state and flux global controls give continuous costs of about 0.07021 0.07021 0.07021 and − 1.030 -1.030 − 1.030 , respectively.
Uniform derivative estimates and stability proofs
Proof of Theorem 7.1
Proof. On a fixed grid, the state map S + S^+ S + is smooth near U a d , h , τ \mathcal{U}_{\mathrm{ad},h,\tau} U ad , h , τ . For a , b ∈ R N ( M + 1 ) a,b\in\mathbb{R}^{N(M+1)} a , b ∈ R N ( M + 1 ) , write its first and second variations as
z [ a ] : = D S + ( u ) [ a ] , s [ a , b ] : = D 2 S + ( u ) [ a , b ] . z[a]:=DS^+(u)[a],\qquad s[a,b]:=D^2S^+(u)[a,b]. z [ a ] := D S + ( u ) [ a ] , s [ a , b ] := D 2 S + ( u ) [ a , b ] .
Both variations lie in R 2 ( N + 1 ) ( M + 1 ) \mathbb{R}^{2(N+1)(M+1)} R 2 ( N + 1 ) ( M + 1 ) and start at zero. With spatial products taken componentwise, the tangent equations are
D τ z 2 n − μ 2 L h z 2 n + 1 = r z 2 n − ( β v 1 n + q u 2 n ) z 2 n + 1 − β v 2 n + 1 z 1 n − q v 2 n + 1 a n D_\tau z_2^n-\mu_2L_hz_2^{n+1}
=rz_2^n-(\beta v_1^n+qu_2^n)z_2^{n+1}-\beta v_2^{n+1}z_1^n-qv_2^{n+1}a^n D τ z 2 n − μ 2 L h z 2 n + 1 = r z 2 n − ( β v 1 n + q u 2 n ) z 2 n + 1 − β v 2 n + 1 z 1 n − q v 2 n + 1 a n
D τ z 1 n − μ 1 L h z 1 n + 1 = − m z 1 n + 1 + α v 2 n + 1 z 1 n + α v 1 n z 2 n + 1 . D_\tau z_1^n-\mu_1L_hz_1^{n+1}
=-mz_1^{n+1}+\alpha v_2^{n+1}z_1^n+\alpha v_1^nz_2^{n+1}. D τ z 1 n − μ 1 L h z 1 n + 1 = − m z 1 n + 1 + α v 2 n + 1 z 1 n + α v 1 n z 2 n + 1 .
Write the common linear part as
D τ w n − blockdiag ( μ 1 L h , μ 2 L h ) w n + 1 = P n w n + Q n w n + 1 + F n . D_\tau w^n-\operatorname{blockdiag}(\mu_1L_h,\mu_2L_h)w^{n+1}
=P_nw^n+Q_nw^{n+1}+F^n. D τ w n − blockdiag ( μ 1 L h , μ 2 L h ) w n + 1 = P n w n + Q n w n + 1 + F n .
where w n , F n ∈ R 2 ( M + 1 ) w^n,F^n\in\mathbb{R}^{2(M+1)} w n , F n ∈ R 2 ( M + 1 ) . In the component order ( w 1 , w 2 ) (w_1,w_2) ( w 1 , w 2 ) , the block matrices are
P n = ( α diag ( v 2 n + 1 ) 0 − β diag ( v 2 n + 1 ) r I ) , Q n = ( − m I α diag ( v 1 n ) 0 − diag ( β v 1 n + q u n ) ) . P_n=
\begin{pmatrix}
\alpha\operatorname{diag}(v_2^{n+1}) & 0\\
-\beta\operatorname{diag}(v_2^{n+1}) & rI
\end{pmatrix},
\qquad
Q_n=
\begin{pmatrix}
-mI & \alpha\operatorname{diag}(v_1^n)\\
0 & -\operatorname{diag}(\beta v_1^n+qu^n)
\end{pmatrix}. P n = ( α diag ( v 2 n + 1 ) − β diag ( v 2 n + 1 ) 0 r I ) , Q n = ( − m I 0 α diag ( v 1 n ) − diag ( β v 1 n + q u n ) ) .
Each diagonal block acts by componentwise multiplication. Since 0 ≤ v i n ≤ K i 1 0\le v_i^n\le K_i\mathbf{1} 0 ≤ v i n ≤ K i 1 and 0 ≤ u n ≤ U max 1 0\le u^n\le U_{\max}\mathbf{1} 0 ≤ u n ≤ U m a x 1 , the weighted operator norms satisfy
∥ P n ∥ ≤ ( α + β ) K 2 + r ≤ A , ∥ Q n ∥ ≤ m + ( α + β ) K 1 + q U max ≤ A , \lVert P_n\rVert\le(\alpha+\beta)K_2+r\le A,\qquad
\lVert Q_n\rVert\le m+(\alpha+\beta)K_1+qU_{\max}\le A, ∥ P n ∥ ≤ ( α + β ) K 2 + r ≤ A , ∥ Q n ∥ ≤ m + ( α + β ) K 1 + q U m a x ≤ A ,
with A A A from Remark G.1. Set E n : = ∥ w 1 n ∥ h 2 + ∥ w 2 n ∥ h 2 E_n:=\lVert w_1^n\rVert_h^2+\lVert w_2^n\rVert_h^2 E n := ∥ w 1 n ∥ h 2 + ∥ w 2 n ∥ h 2 . Test with 2 τ w n + 1 2\tau w^{n+1} 2 τ w n + 1 and use 2 ⟨ x − y , x ⟩ = ∥ x ∥ 2 − ∥ y ∥ 2 + ∥ x − y ∥ 2 2\langle x-y,x\rangle=\lVert x\rVert^2-\lVert y\rVert^2+\lVert x-y\rVert^2 2 ⟨ x − y , x ⟩ = ∥ x ∥ 2 − ∥ y ∥ 2 + ∥ x − y ∥ 2 . The energy identity is
E n + 1 − E n + ∥ w n + 1 − w n ∥ h 2 + 2 τ ∑ i = 1 2 μ i ∣ w i n + 1 ∣ 1 , h 2 = 2 τ ⟨ P n w n , w n + 1 ⟩ h + 2 τ ⟨ Q n w n + 1 , w n + 1 ⟩ h + 2 τ ⟨ F n , w n + 1 ⟩ h ≤ 2 τ A E n E n + 1 + 2 τ A E n + 1 + τ ( ∥ F n ∥ h 2 + E n + 1 ) ≤ τ A ( E n + E n + 1 ) + 2 τ A E n + 1 + τ ( ∥ F n ∥ h 2 + E n + 1 ) . \begin{aligned}
E_{n+1}-E_n+\lVert w^{n+1}-w^n\rVert_h^2
+2\tau\sum_{i=1}^{2}\mu_i\lvert w_i^{n+1}\rvert_{1,h}^2
&=2\tau\langle P_nw^n,w^{n+1}\rangle_h
+2\tau\langle Q_nw^{n+1},w^{n+1}\rangle_h
+2\tau\langle F^n,w^{n+1}\rangle_h\\
&\le2\tau A\sqrt{E_nE_{n+1}}
+2\tau AE_{n+1}
+\tau(\lVert F^n\rVert_h^2+E_{n+1})\\
&\le\tau A(E_n+E_{n+1})+2\tau AE_{n+1}
+\tau(\lVert F^n\rVert_h^2+E_{n+1}).
\end{aligned} E n + 1 − E n + ∥ w n + 1 − w n ∥ h 2 + 2 τ i = 1 ∑ 2 μ i ∣ w i n + 1 ∣ 1 , h 2 = 2 τ ⟨ P n w n , w n + 1 ⟩ h + 2 τ ⟨ Q n w n + 1 , w n + 1 ⟩ h + 2 τ ⟨ F n , w n + 1 ⟩ h ≤ 2 τ A E n E n + 1 + 2 τ A E n + 1 + τ (∥ F n ∥ h 2 + E n + 1 ) ≤ τ A ( E n + E n + 1 ) + 2 τ A E n + 1 + τ (∥ F n ∥ h 2 + E n + 1 ) .
Dropping the nonnegative terms on the left gives
E n + 1 − E n ≤ τ [ A E n + ( 3 A + 1 ) E n + 1 + ∥ F n ∥ h 2 ] . E_{n+1}-E_n\le\tau\left[AE_n+(3A+1)E_{n+1}+\lVert F^n\rVert_h^2\right]. E n + 1 − E n ≤ τ [ A E n + ( 3 A + 1 ) E n + 1 + ∥ F n ∥ h 2 ] .
Rearranging this inequality gives
[ 1 − ( 3 A + 1 ) τ ] E n + 1 ≤ ( 1 + A τ ) E n + τ ∥ F n ∥ h 2 . [1-(3A+1)\tau]E_{n+1}\le(1+A\tau)E_n+\tau\lVert F^n\rVert_h^2. [ 1 − ( 3 A + 1 ) τ ] E n + 1 ≤ ( 1 + A τ ) E n + τ ∥ F n ∥ h 2 .
For τ ≤ [ 2 ( 3 A + 1 ) ] − 1 \tau\le[2(3A+1)]^{-1} τ ≤ [ 2 ( 3 A + 1 ) ] − 1 , the denominator is at least 1 / 2 1/2 1/2 , and
1 + A τ 1 − ( 3 A + 1 ) τ = 1 + ( 4 A + 1 ) τ 1 − ( 3 A + 1 ) τ ≤ 1 + 2 ( 4 A + 1 ) τ . \frac{1+A\tau}{1-(3A+1)\tau}=1+\frac{(4A+1)\tau}{1-(3A+1)\tau}\le1+2(4A+1)\tau. 1 − ( 3 A + 1 ) τ 1 + A τ = 1 + 1 − ( 3 A + 1 ) τ ( 4 A + 1 ) τ ≤ 1 + 2 ( 4 A + 1 ) τ .
Consequently,
E n + 1 ≤ [ 1 + 2 ( 4 A + 1 ) τ ] E n + 2 τ ∥ F n ∥ h 2 . E_{n+1}\le[1+2(4A+1)\tau]E_n+2\tau\|F^n\|_h^2. E n + 1 ≤ [ 1 + 2 ( 4 A + 1 ) τ ] E n + 2 τ ∥ F n ∥ h 2 .
Iteration bounds max 0 ≤ n ≤ N E n \max_{0\le n\le N}E_n max 0 ≤ n ≤ N E n uniformly in terms of E 0 E_0 E 0 and ∥ F ∥ h , τ 2 \|F\|_{h,\tau}^2 ∥ F ∥ h , τ 2 , which also controls P n w n + Q n w n + 1 + F n P_nw^n+Q_nw^{n+1}+F^n P n w n + Q n w n + 1 + F n in the space–time norm. For a scalar heat step D τ w n − μ L h w n + 1 = f n D_\tau w^n-\mu L_hw^{n+1}=f^n D τ w n − μ L h w n + 1 = f n , test with 2 τ D τ w n 2\tau D_\tau w^n 2 τ D τ w n . Summation by parts gives
2 τ ∥ D τ w n ∥ h 2 + μ ( ∣ w n + 1 ∣ 1 , h 2 − ∣ w n ∣ 1 , h 2 + ∣ w n + 1 − w n ∣ 1 , h 2 ) = 2 τ ⟨ f n , D τ w n ⟩ h 2\tau\|D_\tau w^n\|_h^2+\mu\left(|w^{n+1}|_{1,h}^2-|w^n|_{1,h}^2+|w^{n+1}-w^n|_{1,h}^2\right)=2\tau\langle f^n,D_\tau w^n\rangle_h 2 τ ∥ D τ w n ∥ h 2 + μ ( ∣ w n + 1 ∣ 1 , h 2 − ∣ w n ∣ 1 , h 2 + ∣ w n + 1 − w n ∣ 1 , h 2 ) = 2 τ ⟨ f n , D τ w n ⟩ h
≤ 2 τ ∥ f n ∥ h ∥ D τ w n ∥ h ≤ τ ∥ f n ∥ h 2 + τ ∥ D τ w n ∥ h 2 . \le2\tau\|f^n\|_h\|D_\tau w^n\|_h
\le\tau\|f^n\|_h^2+\tau\|D_\tau w^n\|_h^2. ≤ 2 τ ∥ f n ∥ h ∥ D τ w n ∥ h ≤ τ ∥ f n ∥ h 2 + τ ∥ D τ w n ∥ h 2 .
Thus, for 1 ≤ k ≤ N 1\le k\le N 1 ≤ k ≤ N ,
τ ∑ n = 0 k − 1 ∥ D τ w n ∥ h 2 + μ ∣ w k ∣ 1 , h 2 + μ ∑ n = 0 k − 1 ∣ w n + 1 − w n ∣ 1 , h 2 ≤ τ ∑ n = 0 k − 1 ∥ f n ∥ h 2 + μ ∣ w 0 ∣ 1 , h 2 . (73) \tau\sum_{n=0}^{k-1}\|D_\tau w^n\|_h^2+\mu|w^k|_{1,h}^2+\mu\sum_{n=0}^{k-1}|w^{n+1}-w^n|_{1,h}^2\le\tau\sum_{n=0}^{k-1}\|f^n\|_h^2+\mu|w^0|_{1,h}^2. \tag*{(73)} τ n = 0 ∑ k − 1 ∥ D τ w n ∥ h 2 + μ ∣ w k ∣ 1 , h 2 + μ n = 0 ∑ k − 1 ∣ w n + 1 − w n ∣ 1 , h 2 ≤ τ n = 0 ∑ k − 1 ∥ f n ∥ h 2 + μ ∣ w 0 ∣ 1 , h 2 . (73)
Combine the time-derivative bound at k = N k=N k = N with the gradient bound at each k k k to get
τ ∑ n = 0 N − 1 ∥ D τ w n ∥ h 2 + μ max 0 ≤ n ≤ N ∣ w n ∣ 1 , h 2 ≤ 2 ( τ ∑ n = 0 N − 1 ∥ f n ∥ h 2 + μ ∣ w 0 ∣ 1 , h 2 ) . \tau\sum_{n=0}^{N-1}\|D_\tau w^n\|_h^2+\mu\max_{0\le n\le N}|w^n|_{1,h}^2\le2\left(\tau\sum_{n=0}^{N-1}\|f^n\|_h^2+\mu|w^0|_{1,h}^2\right). τ n = 0 ∑ N − 1 ∥ D τ w n ∥ h 2 + μ 0 ≤ n ≤ N max ∣ w n ∣ 1 , h 2 ≤ 2 ( τ n = 0 ∑ N − 1 ∥ f n ∥ h 2 + μ ∣ w 0 ∣ 1 , h 2 ) .
For any f ∈ H 1 ( 0 , 1 ) f\in H^1(0,1) f ∈ H 1 ( 0 , 1 ) and x ∈ [ 0 , 1 ] x\in[0,1] x ∈ [ 0 , 1 ] , the continuous representative satisfies
f ( x ) = ∫ 0 1 f ( y ) d y + ∫ 0 1 ( f ( x ) − f ( y ) ) d y , ∣ f ( x ) − f ( y ) ∣ ≤ ∫ min ( x , y ) max ( x , y ) ∣ f ′ ( s ) ∣ d s . f(x)=\int_0^1 f(y)\,dy+\int_0^1(f(x)-f(y))\,dy,\qquad|f(x)-f(y)|\le\int_{\min(x,y)}^{\max(x,y)}|f'(s)|\,ds. f ( x ) = ∫ 0 1 f ( y ) d y + ∫ 0 1 ( f ( x ) − f ( y )) d y , ∣ f ( x ) − f ( y ) ∣ ≤ ∫ m i n ( x , y ) m a x ( x , y ) ∣ f ′ ( s ) ∣ d s .
Thus ∣ f ( x ) ∣ ≤ ∥ f ∥ L 1 + ∥ f ′ ∥ L 1 ≤ ∥ f ∥ L 2 + ∥ f ′ ∥ L 2 |f(x)|\le\|f\|_{L^1}+\|f'\|_{L^1}\le\|f\|_{L^2}+\|f'\|_{L^2} ∣ f ( x ) ∣ ≤ ∥ f ∥ L 1 + ∥ f ′ ∥ L 1 ≤ ∥ f ∥ L 2 + ∥ f ′ ∥ L 2 . Applying this to the nodal interpolant I h w ∈ H 1 ( 0 , 1 ) I_hw\in H^1(0,1) I h w ∈ H 1 ( 0 , 1 ) gives
∥ w ∥ ∞ = ∥ I h w ∥ L ∞ ≤ ∥ I h w ∥ L 2 + ∥ ( I h w ) ′ ∥ L 2 ≤ ∥ w ∥ h + ∣ w ∣ 1 , h . \|w\|_\infty=\|I_hw\|_{L^\infty}\le\|I_hw\|_{L^2}+\|(I_hw)'\|_{L^2}\le\|w\|_h+|w|_{1,h}. ∥ w ∥ ∞ = ∥ I h w ∥ L ∞ ≤ ∥ I h w ∥ L 2 + ∥ ( I h w ) ′ ∥ L 2 ≤ ∥ w ∥ h + ∣ w ∣ 1 , h .
These mesh-independent estimates apply first to the bounded state reactions, then to the tangent forcing F 2 n = − q v 2 n + 1 a n F_2^n=-qv_2^{n+1}a^n F 2 n = − q v 2 n + 1 a n , F 1 n = 0 F_1^n=0 F 1 n = 0 , giving
∥ D τ v ∥ h , τ ≤ C , (74) \|D_\tau v\|_{h,\tau}\le C, \tag*{(74)} ∥ D τ v ∥ h , τ ≤ C , (74)
and
∥ D τ z [ a ] ∥ h , τ + max 0 ≤ n ≤ N ∥ z [ a ] n ∥ ∞ ≤ C ∥ a ∥ h , τ . (75) \|D_\tau z[a]\|_{h,\tau}+\max_{0\le n\le N}\|z[a]^n\|_\infty\le C\|a\|_{h,\tau}. \tag*{(75)} ∥ D τ z [ a ] ∥ h , τ + 0 ≤ n ≤ N max ∥ z [ a ] n ∥ ∞ ≤ C ∥ a ∥ h , τ . (75)
The second variation has the same linear part, with forcing
F 2 n : = − ( β z 1 [ a ] n + q a n ) z 2 [ b ] n + 1 − ( β z 1 [ b ] n + q b n ) z 2 [ a ] n + 1 , F 1 n : = α ( z 1 [ a ] n z 2 [ b ] n + 1 + z 1 [ b ] n z 2 [ a ] n + 1 ) . F_2^n:=-(\beta z_1[a]^n+qa^n)z_2[b]^{n+1}-(\beta z_1[b]^n+qb^n)z_2[a]^{n+1},\qquad F_1^n:=\alpha\left(z_1[a]^nz_2[b]^{n+1}+z_1[b]^nz_2[a]^{n+1}\right). F 2 n := − ( β z 1 [ a ] n + q a n ) z 2 [ b ] n + 1 − ( β z 1 [ b ] n + q b n ) z 2 [ a ] n + 1 , F 1 n := α ( z 1 [ a ] n z 2 [ b ] n + 1 + z 1 [ b ] n z 2 [ a ] n + 1 ) .
The product inequality ∥ x y ∥ h ≤ ∥ x ∥ ∞ ∥ y ∥ h \|xy\|_h\le\|x\|_\infty\|y\|_h ∥ x y ∥ h ≤ ∥ x ∥ ∞ ∥ y ∥ h gives, for example,
∥ a ∘ z 2 [ b ] + ∥ h , τ ≤ max 0 ≤ n < N ∥ z 2 [ b ] n + 1 ∥ ∞ ∥ a ∥ h , τ ≤ C ∥ a ∥ h , τ ∥ b ∥ h , τ . \|a\circ z_2[b]^+\|_{h,\tau}\le\max_{0\le n<N}\|z_2[b]^{n+1}\|_\infty\|a\|_{h,\tau}\le C\|a\|_{h,\tau}\|b\|_{h,\tau}. ∥ a ∘ z 2 [ b ] + ∥ h , τ ≤ 0 ≤ n < N max ∥ z 2 [ b ] n + 1 ∥ ∞ ∥ a ∥ h , τ ≤ C ∥ a ∥ h , τ ∥ b ∥ h , τ .
Applying this inequality to each product gives ∥ F ∥ h , τ ≤ C ∥ a ∥ h , τ ∥ b ∥ h , τ \|F\|_{h,\tau}\le C\|a\|_{h,\tau}\|b\|_{h,\tau} ∥ F ∥ h , τ ≤ C ∥ a ∥ h , τ ∥ b ∥ h , τ , so
∥ D τ s [ a , b ] ∥ h , τ ≤ C ∥ a ∥ h , τ ∥ b ∥ h , τ . (76) \|D_\tau s[a,b]\|_{h,\tau}\le C\|a\|_{h,\tau}\|b\|_{h,\tau}. \tag*{(76)} ∥ D τ s [ a , b ] ∥ h , τ ≤ C ∥ a ∥ h , τ ∥ b ∥ h , τ . (76)
Only the stated data enter the constants. The directions enter through their weighted L 2 L^2 L 2 norms. Differentiating (21) gives
D ( j 1 − j 0 ) ( u ) [ a ] = − η q τ 2 ∑ n = 0 N − 1 ⟨ a n , D τ v 2 n ⟩ h − η q τ 2 ∑ n = 0 N − 1 ⟨ u n , D τ z 2 [ a ] n ⟩ h , D(j_1-j_0)(u)[a]=-\eta q\tau^2\sum_{n=0}^{N-1}\langle a^n,D_\tau v_2^n\rangle_h-\eta q\tau^2\sum_{n=0}^{N-1}\langle u^n,D_\tau z_2[a]^n\rangle_h, D ( j 1 − j 0 ) ( u ) [ a ] = − η q τ 2 n = 0 ∑ N − 1 ⟨ a n , D τ v 2 n ⟩ h − η q τ 2 n = 0 ∑ N − 1 ⟨ u n , D τ z 2 [ a ] n ⟩ h ,
and
D 2 ( j 1 − j 0 ) ( u ) [ a , b ] = − η q τ 2 ∑ n = 0 N − 1 ⟨ a n , D τ z 2 [ b n ] ⟩ h − η q τ 2 ∑ n = 0 N − 1 ⟨ b n , D τ z 2 [ a n ] ⟩ h − η q τ 2 ∑ n = 0 N − 1 ⟨ u n , D τ s 2 [ a , b ] n ⟩ h . D^2(j_1-j_0)(u)[a,b] = -\eta q\tau^2 \sum_{n=0}^{N-1}\langle a^n,D_\tau z_2[b^n]\rangle_h-\eta q\tau^2\sum_{n=0}^{N-1}\langle b^n,D_\tau z_2[a^n]\rangle_h-\eta q\tau^2\sum_{n=0}^{N-1}\langle u^n,D_\tau s_2[a,b]^n\rangle_h. D 2 ( j 1 − j 0 ) ( u ) [ a , b ] = − η q τ 2 n = 0 ∑ N − 1 ⟨ a n , D τ z 2 [ b n ] ⟩ h − η q τ 2 n = 0 ∑ N − 1 ⟨ b n , D τ z 2 [ a n ] ⟩ h − η q τ 2 n = 0 ∑ N − 1 ⟨ u n , D τ s 2 [ a , b ] n ⟩ h .
Since ∥ u ∥ h , τ ≤ U max T \lVert u\rVert_{h,\tau}\le U_{\max}\sqrt{T} ∥ u ∥ h , τ ≤ U m a x T , Cauchy–Schwarz gives
∣ D ( u ) ∣ ≤ η q τ ∥ u ∥ h , τ ∥ D τ v 2 ∥ h , τ ≤ C τ , (77) \lvert D(u)\rvert\le\eta q\tau\lVert u\rVert_{h,\tau}\lVert D_\tau v_2\rVert_{h,\tau}\le C\tau, \tag*{(77)} ∣ D ( u )∣ ≤ η q τ ∥ u ∥ h , τ ∥ D τ v 2 ∥ h , τ ≤ C τ , (77)
and
∣ D ( j 1 − j 0 ) ( u ) [ a ] ∣ ≤ η q τ ( ∥ a ∥ h , τ ∥ D τ v 2 ∥ h , τ + ∥ u ∥ h , τ ∥ D τ z 2 [ a ] ∥ h , τ ) ≤ C τ ∥ a ∥ h , τ , (78) \lvert D(j_1-j_0)(u)[a]\rvert\le\eta q\tau\left(\lVert a\rVert_{h,\tau}\lVert D_\tau v_2\rVert_{h,\tau}+\lVert u\rVert_{h,\tau}\lVert D_\tau z_2[a]\rVert_{h,\tau}\right)\le C\tau\lVert a\rVert_{h,\tau}, \tag*{(78)} ∣ D ( j 1 − j 0 ) ( u ) [ a ]∣ ≤ η q τ ( ∥ a ∥ h , τ ∥ D τ v 2 ∥ h , τ + ∥ u ∥ h , τ ∥ D τ z 2 [ a ] ∥ h , τ ) ≤ C τ ∥ a ∥ h , τ , (78)
and
∣ D 2 ( j 1 − j 0 ) ( u ) [ a , b ] ∣ ≤ η q τ ( ∥ a ∥ h , τ ∥ D τ z 2 [ b ] ∥ h , τ + ∥ b ∥ h , τ ∥ D τ z 2 [ a ] ∥ h , τ + ∥ u ∥ h , τ ∥ D τ s 2 [ a , b ] ∥ h , τ ) ≤ C τ ∥ a ∥ h , τ ∥ b ∥ h , τ . (79) \lvert D^2(j_1-j_0)(u)[a,b]\rvert\le\eta q\tau\left(\lVert a\rVert_{h,\tau}\lVert D_\tau z_2[b]\rVert_{h,\tau}+\lVert b\rVert_{h,\tau}\lVert D_\tau z_2[a]\rVert_{h,\tau}+\lVert u\rVert_{h,\tau}\lVert D_\tau s_2[a,b]\rVert_{h,\tau}\right)\le C\tau\lVert a\rVert_{h,\tau}\lVert b\rVert_{h,\tau}. \tag*{(79)} ∣ D 2 ( j 1 − j 0 ) ( u ) [ a , b ]∣ ≤ η q τ ( ∥ a ∥ h , τ ∥ D τ z 2 [ b ] ∥ h , τ + ∥ b ∥ h , τ ∥ D τ z 2 [ a ] ∥ h , τ + ∥ u ∥ h , τ ∥ D τ s 2 [ a , b ] ∥ h , τ ) ≤ C τ ∥ a ∥ h , τ ∥ b ∥ h , τ . (79)
The bound on D D D gives (26) , and Riesz representation turns the derivative bounds into (27)–(28). □
Explicit constants
Remark G.1 (Explicit constants). Let K i K_i K i be the uniform state bounds. Set
A : = m + r + ( α + β ) ( K 1 + K 2 ) + q U max , R : = 2 exp ( ( 4 A + 1 ) T ) . A:=m+r+(\alpha+\beta)(K_1+K_2)+qU_{\max},\qquad R:=\sqrt{2}\exp((4A+1)T). A := m + r + ( α + β ) ( K 1 + K 2 ) + q U m a x , R := 2 exp (( 4 A + 1 ) T ) .
Assume τ ≤ τ ˉ : = [ 2 ( 3 A + 1 ) ] − 1 \tau\le\bar{\tau}:=[2(3A+1)]^{-1} τ ≤ τ ˉ := [ 2 ( 3 A + 1 ) ] − 1 . The first variation is controlled in the discrete L 2 L^2 L 2 and maximum norms by
C z : = T R q K 2 , C z , ∞ : = ( R + 1 + 2 A T R min { μ 1 , μ 2 } ) q K 2 . C_z:=\sqrt{T}RqK_2,\qquad C_{z,\infty}:=\left(R+\frac{1+2A\sqrt{T}R}{\sqrt{\min\{\mu_1,\mu_2\}}}\right)qK_2. C z := T R q K 2 , C z , ∞ := ( R + min { μ 1 , μ 2 } 1 + 2 A T R ) q K 2 .
For the second variation in the discrete L 2 L^2 L 2 norm, use
C s : = 2 T R C z , ∞ ( ( α + β ) C z + q ) . C_s:=2\sqrt{T}RC_{z,\infty}\left((\alpha+\beta)C_z+q\right). C s := 2 T R C z , ∞ ( ( α + β ) C z + q ) .
The proof below gives these sensitivity bounds and
∥ H 1 ( u ) − H 0 ( u ) ∥ L ( R N ( M + 1 ) ) ≤ η q ( 1 + 2 A T R ) ( 2 q K 2 + U max C s R ) τ . \lVert H_1(u)-H_0(u)\rVert_{\mathcal{L}(\mathbb{R}^{N(M+1)})}\le\eta q(1+2A\sqrt{T}R)\left(2qK_2+\frac{U_{\max}C_s}{R}\right)\tau. ∥ H 1 ( u ) − H 0 ( u ) ∥ L ( R N ( M + 1 ) ) ≤ η q ( 1 + 2 A T R ) ( 2 q K 2 + R U m a x C s ) τ .
Proof. For the same variation system with w 0 = 0 w^0=0 w 0 = 0 , iterate the energy recurrence. For 1 ≤ k ≤ N 1\le k\le N 1 ≤ k ≤ N ,
E k ≤ 2 τ ∑ n = 0 k − 1 [ 1 + 2 ( 4 A + 1 ) τ ] k − 1 − n ∥ F n ∥ h 2 ≤ 2 e 2 ( 4 A + 1 ) T ∥ F ∥ h , τ 2 = R 2 ∥ F ∥ h , τ 2 . E_k\le2\tau\sum_{n=0}^{k-1}[1+2(4A+1)\tau]^{k-1-n}\lVert F^n\rVert_h^2\le2e^{2(4A+1)T}\lVert F\rVert_{h,\tau}^2=R^2\lVert F\rVert_{h,\tau}^2. E k ≤ 2 τ n = 0 ∑ k − 1 [ 1 + 2 ( 4 A + 1 ) τ ] k − 1 − n ∥ F n ∥ h 2 ≤ 2 e 2 ( 4 A + 1 ) T ∥ F ∥ h , τ 2 = R 2 ∥ F ∥ h , τ 2 .
Hence max 0 ≤ n ≤ N ∥ w n ∥ h ≤ R ∥ F ∥ h , τ \max_{0\le n\le N}\lVert w^n\rVert_h\le R\lVert F\rVert_{h,\tau} max 0 ≤ n ≤ N ∥ w n ∥ h ≤ R ∥ F ∥ h , τ . The heat-equation forcing is f n = P n w n + Q n w n + 1 + F n f^n=P_nw^n+Q_nw^{n+1}+F^n f n = P n w n + Q n w n + 1 + F n . It satisfies
∥ f ∥ h , τ ≤ A ( ∥ w − ∥ h , τ + ∥ w + ∥ h , τ ) + ∥ F ∥ h , τ ≤ ( 1 + 2 A T R ) ∥ F ∥ h , τ . \lVert f\rVert_{h,\tau}\le A\left(\lVert w^-\rVert_{h,\tau}+\lVert w^+\rVert_{h,\tau}\right)+\lVert F\rVert_{h,\tau}\le(1+2A\sqrt{T}R)\lVert F\rVert_{h,\tau}. ∥ f ∥ h , τ ≤ A ( ∥ w − ∥ h , τ + ∥ w + ∥ h , τ ) + ∥ F ∥ h , τ ≤ ( 1 + 2 A T R ) ∥ F ∥ h , τ .
Applying (73) to each component gives
∥ D τ w ∥ h , τ ≤ ( 1 + 2 A T R ) ∥ F ∥ h , τ , max 0 ≤ n ≤ N ∣ w n ∣ 1 , h ≤ 1 + 2 A T R min { μ 1 , μ 2 } ∥ F ∥ h , τ . \lVert D_\tau w\rVert_{h,\tau}\le(1+2A\sqrt{T}R)\lVert F\rVert_{h,\tau},\qquad\max_{0\le n\le N}\lvert w^n\rvert_{1,h}\le\frac{1+2A\sqrt{T}R}{\sqrt{\min\{\mu_1,\mu_2\}}}\lVert F\rVert_{h,\tau}. ∥ D τ w ∥ h , τ ≤ ( 1 + 2 A T R ) ∥ F ∥ h , τ , 0 ≤ n ≤ N max ∣ w n ∣ 1 , h ≤ min { μ 1 , μ 2 } 1 + 2 A T R ∥ F ∥ h , τ .
For pairs, ∣ w ∣ 1 , h 2 = ∣ w 1 ∣ 1 , h 2 + ∣ w 2 ∣ 1 , h 2 |w|_{1,h}^{2}=|w_{1}|_{1,h}^{2}+|w_{2}|_{1,h}^{2} ∣ w ∣ 1 , h 2 = ∣ w 1 ∣ 1 , h 2 + ∣ w 2 ∣ 1 , h 2 . The tangent forcing has norm at most q K 2 ∥ a ∥ h , τ qK_{2}\|a\|_{h,\tau} q K 2 ∥ a ∥ h , τ . Thus, for either time reconstruction,
∥ z i [ a ] ± ∥ h , τ ≤ C z ∥ a ∥ h , τ , max 0 ≤ n ≤ N ∥ z i [ a ] n ∥ ∞ ≤ C z , ∞ ∥ a ∥ h , τ . \|z_i[a]^{\pm}\|_{h,\tau}\le C_z\|a\|_{h,\tau},\qquad\max_{0\le n\le N}\|z_i[a]^n\|_{\infty}\le C_{z,\infty}\|a\|_{h,\tau}. ∥ z i [ a ] ± ∥ h , τ ≤ C z ∥ a ∥ h , τ , 0 ≤ n ≤ N max ∥ z i [ a ] n ∥ ∞ ≤ C z , ∞ ∥ a ∥ h , τ .
For the second-variation forcing, bound each product by the spatial supremum norm of one factor and the space–time norm of the other. The two components satisfy
∥ F 1 ∥ h , τ ≤ 2 α C z , ∞ C z ∥ a ∥ h , τ ∥ b ∥ h , τ , ∥ F 2 ∥ h , τ ≤ 2 C z , ∞ ( β C z + q ) ∥ a ∥ h , τ ∥ b ∥ h , τ . \|F_1\|_{h,\tau}\le2\alpha C_{z,\infty}C_z\|a\|_{h,\tau}\|b\|_{h,\tau},\qquad\|F_2\|_{h,\tau}\le2C_{z,\infty}(\beta C_z+q)\|a\|_{h,\tau}\|b\|_{h,\tau}. ∥ F 1 ∥ h , τ ≤ 2 α C z , ∞ C z ∥ a ∥ h , τ ∥ b ∥ h , τ , ∥ F 2 ∥ h , τ ≤ 2 C z , ∞ ( β C z + q ) ∥ a ∥ h , τ ∥ b ∥ h , τ .
Since the norm of the pair is at most the sum of the component norms,
∥ F ∥ h , τ ≤ 2 C z , ∞ ( ( α + β ) C z + q ) ∥ a ∥ h , τ ∥ b ∥ h , τ . \|F\|_{h,\tau}\le2C_{z,\infty}\bigl((\alpha+\beta)C_z+q\bigr)\|a\|_{h,\tau}\|b\|_{h,\tau}. ∥ F ∥ h , τ ≤ 2 C z , ∞ ( ( α + β ) C z + q ) ∥ a ∥ h , τ ∥ b ∥ h , τ .
The energy bound applied to the second variation then gives
∥ s i [ a , b ] ± ∥ h , τ ≤ 2 T R C z , ∞ ( ( α + β ) C z + q ) ∥ a ∥ h , τ ∥ b ∥ h , τ = C s ∥ a ∥ h , τ ∥ b ∥ h , τ . \|s_i[a,b]^{\pm}\|_{h,\tau}\le2\sqrt{T}RC_{z,\infty}\bigl((\alpha+\beta)C_z+q\bigr)\|a\|_{h,\tau}\|b\|_{h,\tau}=C_s\|a\|_{h,\tau}\|b\|_{h,\tau}. ∥ s i [ a , b ] ± ∥ h , τ ≤ 2 T R C z , ∞ ( ( α + β ) C z + q ) ∥ a ∥ h , τ ∥ b ∥ h , τ = C s ∥ a ∥ h , τ ∥ b ∥ h , τ .
Applying the time-derivative estimate to the tangent and second-variation equations gives
∥ D τ z [ a ] ∥ h , τ ≤ ( 1 + 2 A T R ) q K 2 ∥ a ∥ h , τ , ∥ D τ s [ a , b ] ∥ h , τ ≤ ( 1 + 2 A T R ) C s T R ∥ a ∥ h , τ ∥ b ∥ h , τ . \|D_\tau z[a]\|_{h,\tau}\le(1+2A\sqrt{TR})qK_2\|a\|_{h,\tau},\qquad\|D_\tau s[a,b]\|_{h,\tau}\le(1+2A\sqrt{TR})\frac{C_s}{\sqrt{TR}}\|a\|_{h,\tau}\|b\|_{h,\tau}. ∥ D τ z [ a ] ∥ h , τ ≤ ( 1 + 2 A T R ) q K 2 ∥ a ∥ h , τ , ∥ D τ s [ a , b ] ∥ h , τ ≤ ( 1 + 2 A T R ) T R C s ∥ a ∥ h , τ ∥ b ∥ h , τ .
Substitute these bounds into the formula for D 2 ( j 1 − j 0 ) D^2(j_1-j_0) D 2 ( j 1 − j 0 ) in Appendix § 21 .§ 1 and use ∥ u ∥ h , τ ≤ U max T \|u\|_{h,\tau}\le U_{\max}\sqrt{T} ∥ u ∥ h , τ ≤ U m a x T :
∣ D 2 ( j 1 − j 0 ) ( u ) [ a , b ] ∣ ≤ η q τ ( ∥ a ∥ h , τ ∥ D τ z 2 [ b ] ∥ h , τ + ∥ b ∥ h , τ ∥ D τ z 2 [ a ] ∥ h , τ + ∥ u ∥ h , τ ∥ D τ s 2 [ a , b ] ∥ h , τ ) ≤ η q ( 1 + 2 A T R ) ( 2 q K 2 + U max C s R ) τ ∥ a ∥ h , τ ∥ b ∥ h , τ . \begin{aligned}
|D^2(j_1-j_0)(u)[a,b]|&\le\eta q\tau\bigl(\|a\|_{h,\tau}\|D_\tau z_2[b]\|_{h,\tau}+\|b\|_{h,\tau}\|D_\tau z_2[a]\|_{h,\tau}+\|u\|_{h,\tau}\|D_\tau s_2[a,b]\|_{h,\tau}\bigr)\\
&\le\eta q(1+2A\sqrt{TR})\left(2qK_2+\frac{U_{\max}C_s}{R}\right)\tau\|a\|_{h,\tau}\|b\|_{h,\tau}.
\end{aligned} ∣ D 2 ( j 1 − j 0 ) ( u ) [ a , b ] ∣ ≤ η q τ ( ∥ a ∥ h , τ ∥ D τ z 2 [ b ] ∥ h , τ + ∥ b ∥ h , τ ∥ D τ z 2 [ a ] ∥ h , τ + ∥ u ∥ h , τ ∥ D τ s 2 [ a , b ] ∥ h , τ ) ≤ η q ( 1 + 2 A T R ) ( 2 q K 2 + R U m a x C s ) τ ∥ a ∥ h , τ ∥ b ∥ h , τ .
For a linear operator L L L on the weighted control space,
∥ L ∥ h , τ = sup ∥ a ∥ h , τ = 1 ∥ L a ∥ h , τ = sup ∥ a ∥ h , τ = ∥ b ∥ h , τ = 1 ∣ ⟨ L a , b ⟩ h , τ ∣ . \|L\|_{h,\tau}=\sup_{\|a\|_{h,\tau}=1}\|La\|_{h,\tau}=\sup_{\|a\|_{h,\tau}=\|b\|_{h,\tau}=1}|\langle La,b\rangle_{h,\tau}|. ∥ L ∥ h , τ = ∥ a ∥ h , τ = 1 sup ∥ L a ∥ h , τ = ∥ a ∥ h , τ = ∥ b ∥ h , τ = 1 sup ∣ ⟨ L a , b ⟩ h , τ ∣.
With L = H 1 ( u ) − H 0 ( u ) L=H_1(u)-H_0(u) L = H 1 ( u ) − H 0 ( u ) , the bilinear form is D 2 ( j 1 − j 0 ) ( u ) [ a , b ] D^2(j_1-j_0)(u)[a,b] D 2 ( j 1 − j 0 ) ( u ) [ a , b ] . Taking the supremum proves the Hessian bound.
Proof of Corollary 7.2
Proof. For a ∈ R N ( M + 1 ) a\in\mathbb{R}^{N(M+1)} a ∈ R N ( M + 1 ) , use γ i ≥ 0 \gamma_i\ge0 γ i ≥ 0 to bound the tracking contribution:
γ i ∥ z i [ a ] ∥ h , τ 2 + γ i ⟨ v i − v i d , s i [ a , a ] − ⟩ h , τ ≥ − γ i ( K i T + D i ) C s ∥ a ∥ h , τ 2 . \gamma_i\|z_i[a]\|_{h,\tau}^{2}+\gamma_i\langle v_i-v_i^d,s_i[a,a]^{-}\rangle_{h,\tau}\ge-\gamma_i(K_i\sqrt{T}+D_i)C_s\|a\|_{h,\tau}^{2}. γ i ∥ z i [ a ] ∥ h , τ 2 + γ i ⟨ v i − v i d , s i [ a , a ] − ⟩ h , τ ≥ − γ i ( K i T + D i ) C s ∥ a ∥ h , τ 2 .
For ζ ∈ { 0 , 1 } \zeta\in\{0,1\} ζ ∈ { 0 , 1 } , the harvest contribution satisfies
∣ − 2 η q ⟨ a , ( 1 − ζ ) z 2 [ a ] − + ζ z 2 [ a ] + ⟩ h , τ − η q ⟨ u , ( 1 − ζ ) s 2 [ a , a ] − + ζ s 2 [ a , a ] + ⟩ h , τ ∣ ≤ 2 η q ∣ ⟨ a , ( 1 − ζ ) z 2 [ a ] − + ζ z 2 [ a ] + ⟩ h , τ ∣ + η q ∣ ⟨ u , ( 1 − ζ ) s 2 [ a , a ] − + ζ s 2 [ a , a ] + ⟩ h , τ ∣ ≤ ( 2 η q C z + η q U max T C s ) ∥ a ∥ h , τ 2 . \begin{aligned}
&\left|-2\eta q\langle a,(1-\zeta)z_2[a]^{-}+\zeta z_2[a]^{+}\rangle_{h,\tau}-\eta q\langle u,(1-\zeta)s_2[a,a]^{-}+\zeta s_2[a,a]^{+}\rangle_{h,\tau}\right|\\
&\le2\eta q|\langle a,(1-\zeta)z_2[a]^{-}+\zeta z_2[a]^{+}\rangle_{h,\tau}|+\eta q|\langle u,(1-\zeta)s_2[a,a]^{-}+\zeta s_2[a,a]^{+}\rangle_{h,\tau}|\\
&\le(2\eta qC_z+\eta qU_{\max}\sqrt{T}C_s)\|a\|_{h,\tau}^{2}.
\end{aligned} − 2 η q ⟨ a , ( 1 − ζ ) z 2 [ a ] − + ζ z 2 [ a ] + ⟩ h , τ − η q ⟨ u , ( 1 − ζ ) s 2 [ a , a ] − + ζ s 2 [ a , a ] + ⟩ h , τ ≤ 2 η q ∣ ⟨ a , ( 1 − ζ ) z 2 [ a ] − + ζ z 2 [ a ] + ⟩ h , τ ∣ + η q ∣ ⟨ u , ( 1 − ζ ) s 2 [ a , a ] − + ζ s 2 [ a , a ] + ⟩ h , τ ∣ ≤ ( 2 η q C z + η q U m a x T C s ) ∥ a ∥ h , τ 2 .
Adding the effort contribution gives
⟨ H ζ ( u ) a , a ⟩ h , τ ≥ ( λ − B ) ∥ a ∥ h , τ 2 . \langle H_\zeta(u)a,a\rangle_{h,\tau}\ge(\lambda-B)\|a\|_{h,\tau}^{2}. ⟨ H ζ ( u ) a , a ⟩ h , τ ≥ ( λ − B ) ∥ a ∥ h , τ 2 .
For u , v ∈ U a d , h , τ u,v\in\mathcal{U}_{\mathrm{ad},h,\tau} u , v ∈ U ad , h , τ , integration along the feasible segment gives
j ζ ( v ) − j ζ ( u ) − ⟨ g ζ ( u ) , v − u ⟩ h , τ = ∫ 0 1 ( 1 − t ) ⟨ H ζ ( u + t ( v − u ) ) ( v − u ) , v − u ⟩ h , τ d t ≥ λ − B 2 ∥ v − u ∥ h , τ 2 . j_\zeta(v)-j_\zeta(u)-\langle g_\zeta(u),v-u\rangle_{h,\tau}
=\int_0^1(1-t)\langle H_\zeta(u+t(v-u))(v-u),v-u\rangle_{h,\tau}\,dt
\ge\frac{\lambda-B}{2}\|v-u\|_{h,\tau}^{2}. j ζ ( v ) − j ζ ( u ) − ⟨ g ζ ( u ) , v − u ⟩ h , τ = ∫ 0 1 ( 1 − t ) ⟨ H ζ ( u + t ( v − u )) ( v − u ) , v − u ⟩ h , τ d t ≥ 2 λ − B ∥ v − u ∥ h , τ 2 .
A minimizer exists by compactness of U a d , h , τ \mathcal{U}_{\mathrm{ad},h,\tau} U ad , h , τ . If u u u and v v v were two minimizers, their variational inequalities would give
0 ≥ ⟨ g ζ ( v ) − g ζ ( u ) , v − u ⟩ h , τ ≥ ( λ − B ) ∥ v − u ∥ h , τ 2 . 0 \ge\langle g_{\zeta}(v)-g_{\zeta}(u),v-u\rangle_{h,\tau} \ge(\lambda-B)\|v-u\|_{h,\tau}^{2}. 0 ≥ ⟨ g ζ ( v ) − g ζ ( u ) , v − u ⟩ h , τ ≥ ( λ − B ) ∥ v − u ∥ h , τ 2 .
Since λ > B \lambda>B λ > B , this forces u = v u=v u = v .
Proof of Corollary 7.3
Proof. The set K r K_r K r is nonempty, convex and compact. For every u ∈ K r u\in K_r u ∈ K r and a ∈ R N ( M + 1 ) a\in\mathbb{R}^{N(M+1)} a ∈ R N ( M + 1 ) ,
⟨ H 1 ( u ) a , a ⟩ h , τ ≥ ( κ − C τ ) ∥ a ∥ h , τ 2 (80) \langle H_1(u)a,a\rangle_{h,\tau} \ge(\kappa-C\tau)\|a\|_{h,\tau}^{2} \tag*{(80)} ⟨ H 1 ( u ) a , a ⟩ h , τ ≥ ( κ − C τ ) ∥ a ∥ h , τ 2 (80)
Because κ − C τ > 0 \kappa-C\tau>0 κ − C τ > 0 , j 1 j_1 j 1 has a unique minimizer u 1 u_1 u 1 on K r K_r K r . Set d : = u 1 − u 0 d:=u_1-u_0 d := u 1 − u 0 . The segment joining the two controls lies in K r K_r K r , so
⟨ g 0 ( u 1 ) − g 0 ( u 0 ) , d ⟩ h , τ = ∫ 0 1 ⟨ H 0 ( u 0 + t d ) d , d ⟩ h , τ d t ≥ κ ∥ d ∥ h , τ 2 . \langle g_0(u_1)-g_0(u_0),d\rangle_{h,\tau}=\int_0^1\langle H_0(u_0+td)d,d\rangle_{h,\tau}\,\mathrm{d}t\ge\kappa\|d\|_{h,\tau}^{2}. ⟨ g 0 ( u 1 ) − g 0 ( u 0 ) , d ⟩ h , τ = ∫ 0 1 ⟨ H 0 ( u 0 + t d ) d , d ⟩ h , τ d t ≥ κ ∥ d ∥ h , τ 2 .
Testing each variational inequality with the other control gives ⟨ g 0 ( u 0 ) , d ⟩ h , τ ≥ 0 \langle g_0(u_0),d\rangle_{h,\tau}\ge0 ⟨ g 0 ( u 0 ) , d ⟩ h , τ ≥ 0 , ⟨ g 1 ( u 1 ) , d ⟩ h , τ ≤ 0 \langle g_1(u_1),d\rangle_{h,\tau}\le0 ⟨ g 1 ( u 1 ) , d ⟩ h , τ ≤ 0 . Thus
κ ∥ d ∥ h , τ 2 ≤ ⟨ g 0 ( u 1 ) − g 0 ( u 0 ) , d ⟩ h , τ = ⟨ g 0 ( u 1 ) − g 1 ( u 1 ) , d ⟩ h , τ + ⟨ g 1 ( u 1 ) , d ⟩ h , τ − ⟨ g 0 ( u 0 ) , d ⟩ h , τ ≤ ⟨ g 0 ( u 1 ) − g 1 ( u 1 ) , d ⟩ h , τ ≤ C τ ∥ d ∥ h , τ . (81) \begin{aligned}
\kappa\|d\|_{h,\tau}^{2}
&\le\langle g_0(u_1)-g_0(u_0),d\rangle_{h,\tau}\\
&=\langle g_0(u_1)-g_1(u_1),d\rangle_{h,\tau}+\langle g_1(u_1),d\rangle_{h,\tau}-\langle g_0(u_0),d\rangle_{h,\tau}\\
&\le\langle g_0(u_1)-g_1(u_1),d\rangle_{h,\tau}\\
&\le C\tau\|d\|_{h,\tau}. \tag*{(81)}
\end{aligned} κ ∥ d ∥ h , τ 2 ≤ ⟨ g 0 ( u 1 ) − g 0 ( u 0 ) , d ⟩ h , τ = ⟨ g 0 ( u 1 ) − g 1 ( u 1 ) , d ⟩ h , τ + ⟨ g 1 ( u 1 ) , d ⟩ h , τ − ⟨ g 0 ( u 0 ) , d ⟩ h , τ ≤ ⟨ g 0 ( u 1 ) − g 1 ( u 1 ) , d ⟩ h , τ ≤ C τ ∥ d ∥ h , τ . (81)
This proves (29) and gives ∥ d ∥ h , τ < r \|d\|_{h,\tau}<r ∥ d ∥ h , τ < r . For any u ∈ U a d , h , τ u\in\mathcal{U}_{\mathrm{ad},h,\tau} u ∈ U ad , h , τ , the point u 1 + t ( u − u 1 ) u_1+t(u-u_1) u 1 + t ( u − u 1 ) stays in K r K_r K r for small t > 0 t>0 t > 0 . Testing along this segment and dividing by t t t gives the full box inequality at u 1 u_1 u 1 . Strong convexity then makes u 1 u_1 u 1 a strict local minimizer and the only box-stationary point in K r K_r K r . Apply the same argument on the whole box to prove the global statement.
Spectral perturbation at a common control
Proof. For any a ∈ R N ( M + 1 ) a\in\mathbb{R}^{N(M+1)} a ∈ R N ( M + 1 ) ,
∣ ⟨ ( H 1 ( u ) − H 0 ( u ) ) a , a ⟩ h , τ ∣ ≤ C τ ∥ a ∥ h , τ 2 (82) \left|\langle(H_1(u)-H_0(u))a,a\rangle_{h,\tau}\right|\le C\tau\|a\|_{h,\tau}^{2} \tag*{(82)} ∣ ⟨( H 1 ( u ) − H 0 ( u )) a , a ⟩ h , τ ∣ ≤ C τ ∥ a ∥ h , τ 2 (82)
For a ≠ 0 a\ne0 a = 0 , division by ∥ a ∥ h , τ 2 \|a\|_{h,\tau}^{2} ∥ a ∥ h , τ 2 gives the Rayleigh-quotient bounds
⟨ H 0 ( u ) a , a ⟩ h , τ ∥ a ∥ h , τ 2 − C τ ≤ ⟨ H 1 ( u ) a , a ⟩ h , τ ∥ a ∥ h , τ 2 ≤ ⟨ H 0 ( u ) a , a ⟩ h , τ ∥ a ∥ h , τ 2 + C τ . \frac{\langle H_0(u)a,a\rangle_{h,\tau}}{\|a\|_{h,\tau}^{2}}-C\tau
\le
\frac{\langle H_1(u)a,a\rangle_{h,\tau}}{\|a\|_{h,\tau}^{2}}
\le
\frac{\langle H_0(u)a,a\rangle_{h,\tau}}{\|a\|_{h,\tau}^{2}}+C\tau. ∥ a ∥ h , τ 2 ⟨ H 0 ( u ) a , a ⟩ h , τ − C τ ≤ ∥ a ∥ h , τ 2 ⟨ H 1 ( u ) a , a ⟩ h , τ ≤ ∥ a ∥ h , τ 2 ⟨ H 0 ( u ) a , a ⟩ h , τ + C τ .
The matrices H ~ ζ = W 1 / 2 H ζ W − 1 / 2 \widetilde{H}_{\zeta}=W^{1/2}H_{\zeta}W^{-1/2} H ζ = W 1/2 H ζ W − 1/2 are Euclidean symmetric. Moreover, for x = W 1 / 2 a x=W^{1/2}a x = W 1/2 a ,
x T H ~ ζ x x T x = ⟨ H ζ a , a ⟩ h , τ ∥ a ∥ h , τ 2 . \frac{x^{T}\widetilde{H}_{\zeta}x}{x^{T}x}=\frac{\langle H_{\zeta}a,a\rangle_{h,\tau}}{\|a\|_{h,\tau}^{2}}. x T x x T H ζ x = ∥ a ∥ h , τ 2 ⟨ H ζ a , a ⟩ h , τ .
For the eigenvalues ordered increasingly, the min–max formula is
λ k ( H ζ ) = min E ⊂ R N ( M + 1 ) dim E = k max a ∈ E ∖ { 0 } ⟨ H ζ a , a ⟩ h , τ ∥ a ∥ h , τ 2 . \lambda_k(H_{\zeta})=\min_{\substack{E\subset\mathbb{R}^{N(M+1)}\\ \dim E=k}}\ \max_{a\in E\setminus\{0\}}\frac{\langle H_{\zeta}a,a\rangle_{h,\tau}}{\|a\|_{h,\tau}^{2}}. λ k ( H ζ ) = E ⊂ R N ( M + 1 ) d i m E = k min a ∈ E ∖ { 0 } max ∥ a ∥ h , τ 2 ⟨ H ζ a , a ⟩ h , τ .
Taking the maximum and then the minimum in the two Rayleigh inequalities yields
λ k ( H 0 ) − C τ ≤ λ k ( H 1 ) ≤ λ k ( H 0 ) + C τ . \lambda_k(H_0)-C\tau\le\lambda_k(H_1)\le\lambda_k(H_0)+C\tau. λ k ( H 0 ) − C τ ≤ λ k ( H 1 ) ≤ λ k ( H 0 ) + C τ .
For a fixed subspace with an orthonormal basis matrix E E E , its Euclidean compressions satisfy
∥ E T ( H ~ 1 − H ~ 0 ) E ∥ 2 ≤ ∥ H ~ 1 − H ~ 0 ∥ 2 ≤ C τ , \lVert E^{T}(\widetilde{H}_{1}-\widetilde{H}_{0})E\rVert_{2}\leq\lVert\widetilde{H}_{1}-\widetilde{H}_{0}\rVert_{2}\leq C\tau, ∥ E T ( H 1 − H 0 ) E ∥ 2 ≤ ∥ H 1 − H 0 ∥ 2 ≤ C τ ,
so the same argument applies to the compressed operators.
Production–destruction proofs
Weighted transfer cancellation
Proof. The weighted column sum is
∑ i = 1 s c i w a K ( i , a ) , ( j , a ) n = c j w a ∑ 1 ≤ i ≤ s i ≠ j k i j , a n − c j w a ∑ 1 ≤ i ≤ s i ≠ j k i j , a n = 0. \sum_{i=1}^{s}c_{i}w_{a}K_{(i,a),(j,a)}^{n}
=c_{j}w_{a}\sum_{\substack{1\leq i\leq s\\i\ne j}}k_{ij,a}^{n}
-c_{j}w_{a}\sum_{\substack{1\leq i\leq s\\i\ne j}}k_{ij,a}^{n}=0. i = 1 ∑ s c i w a K ( i , a ) , ( j , a ) n = c j w a 1 ≤ i ≤ s i = j ∑ k ij , a n − c j w a 1 ≤ i ≤ s i = j ∑ k ij , a n = 0.
Proof of Theorem 12.1
Proof. Set Q : = diag ( ρ ) − 1 ( A n ) T diag ( ρ ) Q:=\operatorname{diag}(\rho)^{-1}(A^{n})^{T}\operatorname{diag}(\rho) Q := diag ( ρ ) − 1 ( A n ) T diag ( ρ ) . Index the coordinates by 1 ≤ κ , ν ≤ s d 1\leq\kappa,\nu\leq sd 1 ≤ κ , ν ≤ s d . For κ ≠ ν \kappa\ne\nu κ = ν , we have Q κ ν ≤ 0 Q_{\kappa\nu}\leq0 Q κ ν ≤ 0 . For each row κ \kappa κ ,
∑ ν = 1 s d Q κ ν = 1 + τ ( m κ n + h κ n ) > 0. \sum_{\nu=1}^{sd}Q_{\kappa\nu}=1+\tau(m_{\kappa}^{n}+h_{\kappa}^{n})>0. ν = 1 ∑ s d Q κ ν = 1 + τ ( m κ n + h κ n ) > 0.
Thus Q Q Q is a strictly row-diagonally dominant Z-matrix, meaning that its off-diagonal entries are nonpositive. If Q x = f ≥ 0 Qx=f\geq0 Q x = f ≥ 0 and x κ = min 1 ≤ ν ≤ s d x ν < 0 x_{\kappa}=\min_{1\leq\nu\leq sd}x_{\nu}<0 x κ = min 1 ≤ ν ≤ s d x ν < 0 , then
f κ = Q κ κ x κ + ∑ 1 ≤ ν ≤ s d ν ≠ κ Q κ ν x ν ≤ ( ∑ ν = 1 s d Q κ ν ) x κ < 0. f_{\kappa}=Q_{\kappa\kappa}x_{\kappa}+\sum_{\substack{1\leq\nu\leq sd\\\nu\ne\kappa}}Q_{\kappa\nu}x_{\nu}
\leq\left(\sum_{\nu=1}^{sd}Q_{\kappa\nu}\right)x_{\kappa}<0. f κ = Q κκ x κ + 1 ≤ ν ≤ s d ν = κ ∑ Q κ ν x ν ≤ ( ν = 1 ∑ s d Q κ ν ) x κ < 0.
This contradicts the assumption. Applying the minimum and maximum arguments to Q x = 0 Qx=0 Q x = 0 gives x = 0 x=0 x = 0 , so Q Q Q is invertible with Q − 1 ≥ 0 Q^{-1}\geq0 Q − 1 ≥ 0 . Transpose and undo the positive diagonal scaling to get ( A n ) − 1 ≥ 0 (A^{n})^{-1}\geq0 ( A n ) − 1 ≥ 0 : the step has a unique nonnegative solution. Multiplying (33) by ρ T \rho^{T} ρ T gives (34) . Since
ρ T A n = ρ T + τ ρ T diag ( m n + h n ) , \rho^{T}A^{n}=\rho^{T}+\tau\rho^{T}\operatorname{diag}(m^{n}+h^{n}), ρ T A n = ρ T + τ ρ T diag ( m n + h n ) ,
right multiplication by ( A n ) − 1 ≥ 0 (A^{n})^{-1}\geq0 ( A n ) − 1 ≥ 0 gives ρ T ( A n ) − 1 ≤ ρ T \rho^{T}(A^{n})^{-1}\leq\rho^{T} ρ T ( A n ) − 1 ≤ ρ T . Thus
ρ T ∣ ( A n ) − 1 z ∣ ≤ ρ T ( A n ) − 1 ∣ z ∣ ≤ ρ T ∣ z ∣ . \rho^{T}|(A^{n})^{-1}z|\leq\rho^{T}(A^{n})^{-1}|z|\leq\rho^{T}|z|. ρ T ∣ ( A n ) − 1 z ∣ ≤ ρ T ( A n ) − 1 ∣ z ∣ ≤ ρ T ∣ z ∣.
This proves (35) . The production bound gives
M ( y n + 1 ) ≤ xref‘(34)‘ ( 1 + τ σ ) M ( y n ) + τ b n , \mathcal{M}(y^{n+1})\overset{\text{\scriptsize{xref}`(34)`}}{\leq}(1+\tau\sigma)\mathcal{M}(y^{n})+\tau b_{n}, M ( y n + 1 ) ≤ xref ‘(34)‘ ( 1 + τ σ ) M ( y n ) + τ b n ,
and iteration yields
M ( y n ) ≤ ( 1 + τ σ ) n M ( y 0 ) + τ ∑ r = 0 n − 1 ( 1 + τ σ ) n − 1 − r b r ≤ e σ t n ( M ( y 0 ) + τ ∑ r = 0 n − 1 b r ) . \mathcal{M}(y^{n})\leq(1+\tau\sigma)^{n}\mathcal{M}(y^{0})+\tau\sum_{r=0}^{n-1}(1+\tau\sigma)^{n-1-r}b_{r}
\leq e^{\sigma t_{n}}\left(\mathcal{M}(y^{0})+\tau\sum_{r=0}^{n-1}b_{r}\right). M ( y n ) ≤ ( 1 + τ σ ) n M ( y 0 ) + τ r = 0 ∑ n − 1 ( 1 + τ σ ) n − 1 − r b r ≤ e σ t n ( M ( y 0 ) + τ r = 0 ∑ n − 1 b r ) .
For p n = 0 p^{n}=0 p n = 0 , summation gives
M ( y N ) ⏟ ≥ 0 + τ ∑ n = 0 N − 1 ρ T diag ( m n ) y n + 1 ⏟ ≥ 0 + τ ∑ n = 0 N − 1 ρ T diag ( h n ) y n + 1 ≤ xref‘(34)‘ M ( y 0 ) . \underbrace{\mathcal{M}(y^{N})}_{\geq0}
+\tau\underbrace{\sum_{n=0}^{N-1}\rho^{T}\operatorname{diag}(m^{n})y^{n+1}}_{\geq0}
+\tau\sum_{n=0}^{N-1}\rho^{T}\operatorname{diag}(h^{n})y^{n+1}
\overset{\text{\scriptsize{xref}`(34)`}}{\leq}\mathcal{M}(y^{0}). ≥ 0 M ( y N ) + τ ≥ 0 n = 0 ∑ N − 1 ρ T diag ( m n ) y n + 1 + τ n = 0 ∑ N − 1 ρ T diag ( h n ) y n + 1 ≤ xref ‘(34)‘ M ( y 0 ) .
Remove the nonnegative final mass and mortality terms to obtain the harvest bound.
Exact fixed-grid derivatives
Assume G , p , m , h G,p,m,h G , p , m , h are C 2 C^{2} C 2 on an open neighborhood N ⊂ R s d × R d u \mathcal{N}\subset\mathbb{R}^{sd}\times\mathbb{R}^{du} N ⊂ R s d × R d u of the computed pairs of states and controls, with p , m , h p,m,h p , m , h taking values in R s d \mathbb{R}^{sd} R s d and
G : N ⟶ R s d × s d ( y , u ) ⟼ blockdiag i = 1 s ( μ i L h ) + K ( y , u ) . \begin{aligned}
G:\quad\mathcal{N}&\longrightarrow\mathbb{R}^{sd\times sd}\\
(y,u)&\longmapsto\operatorname{blockdiag}_{i=1}^{s}(\mu_iL_h)+K(y,u).
\end{aligned} G : N ( y , u ) ⟶ R s d × s d ⟼ blockdiag i = 1 s ( μ i L h ) + K ( y , u ) .
Define
F n : { ( y ˉ , y + , u ) : ( y ˉ , u ) ∈ N , y + ∈ R s d } ⟶ R s d ( y ˉ , y + , u ) ⟼ A ( y ˉ , u ) y + − y ˉ − τ p ( y ˉ , u ) . \begin{aligned}
F_n:\{(\bar y,y^{+},u):(\bar y,u)\in\mathcal{N},\ y^{+}\in\mathbb{R}^{sd}\}&\longrightarrow\mathbb{R}^{sd}\\
(\bar y,y^{+},u)&\longmapsto A(\bar y,u)y^{+}-\bar y-\tau p(\bar y,u).
\end{aligned} F n : {( y ˉ , y + , u ) : ( y ˉ , u ) ∈ N , y + ∈ R s d } ( y ˉ , y + , u ) ⟶ R s d ⟼ A ( y ˉ , u ) y + − y ˉ − τ p ( y ˉ , u ) .
At y − = y n y^{-}=y^{n} y − = y n , y + = y n + 1 y^{+}=y^{n+1} y + = y n + 1 , u = u n u=u^{n} u = u n , the matrix A ( y n , u n ) = A n A(y^{n},u^{n})=A^{n} A ( y n , u n ) = A n is invertible, so the implicit function theorem gives a C 2 C^{2} C 2 state map on an open neighborhood O n e t ⊂ ( R d u ) N \mathcal{O}_{\mathrm{net}}\subset(\mathbb{R}^{du})^{N} O net ⊂ ( R d u ) N of the computed control:
S n e t : O n e t ⟶ ( R s d ) N + 1 u ⟼ ( y n ) n = 0 N solving xref‘(33)’ . \begin{aligned}
S_{\mathrm{net}}:\quad\mathcal{O}_{\mathrm{net}}&\longrightarrow(\mathbb{R}^{sd})^{N+1}\\
u&\longmapsto(y^{n})_{n=0}^{N}\text{ solving }\text{{xref}`(33)'}.
\end{aligned} S net : O net u ⟶ ( R s d ) N + 1 ⟼ ( y n ) n = 0 N solving xref ‘(33)’ .
Write the residual derivatives as
∂ y + F n = A n , C n : = ∂ y − F n ∈ R s d × s d , B n : = ∂ u F n ∈ R s d × d u . \partial_{y^{+}}F_n=A^{n},\qquad C_n:=\partial_{y^{-}}F_n\in\mathbb{R}^{sd\times sd},\qquad B_n:=\partial_{u}F_n\in\mathbb{R}^{sd\times du}. ∂ y + F n = A n , C n := ∂ y − F n ∈ R s d × s d , B n := ∂ u F n ∈ R s d × d u .
For directions d , e ∈ ( R d u ) N d,e\in(\mathbb{R}^{du})^{N} d , e ∈ ( R d u ) N , the tangents z [ d ] , z [ e ] ∈ ( R s d ) N + 1 z[d],z[e]\in(\mathbb{R}^{sd})^{N+1} z [ d ] , z [ e ] ∈ ( R s d ) N + 1 and second variation w [ d , e ] w[d,e] w [ d , e ] satisfy
A n z [ d ] n + 1 = − C n z [ d ] n − B n d n , z [ d ] 0 = 0 , A^{n}z[d]^{n+1}=-C_nz[d]^{n}-B_nd^{n},\qquad z[d]^{0}=0, A n z [ d ] n + 1 = − C n z [ d ] n − B n d n , z [ d ] 0 = 0 ,
and
A n w [ d , e ] n + 1 = − C n w [ d , e ] n − D 2 F n [ ( z [ d ] n , z [ d ] n + 1 , d n ) , ( z [ e ] n , z [ e ] n + 1 , e n ) ] , w [ d , e ] 0 = 0. A^{n}w[d,e]^{n+1}=-C_nw[d,e]^{n}-D^{2}F_n\left[(z[d]^{n},z[d]^{n+1},d^{n}),(z[e]^{n},z[e]^{n+1},e^{n})\right],\qquad w[d,e]^{0}=0. A n w [ d , e ] n + 1 = − C n w [ d , e ] n − D 2 F n [ ( z [ d ] n , z [ d ] n + 1 , d n ) , ( z [ e ] n , z [ e ] n + 1 , e n ) ] , w [ d , e ] 0 = 0.
Take a C 2 C^{2} C 2 cost J : O J → R J:\mathcal{O}_{J}\to\mathbb{R} J : O J → R on an open neighborhood O J ⊂ ( R s d ) N + 1 × ( R d u ) N \mathcal{O}_{J}\subset(\mathbb{R}^{sd})^{N+1}\times(\mathbb{R}^{du})^{N} O J ⊂ ( R s d ) N + 1 × ( R d u ) N of the trajectory–control pair. Choose the neighborhoods small enough that ( y n , u n ) ∈ N (y^{n},u^{n})\in\mathcal{N} ( y n , u n ) ∈ N for n = 0 , … , N − 1 n=0,\ldots,N-1 n = 0 , … , N − 1 and ( S n e t ( u ) , u ) ∈ O J (S_{\mathrm{net}}(u),u)\in\mathcal{O}_{J} ( S net ( u ) , u ) ∈ O J for u ∈ O n e t u\in\mathcal{O}_{\mathrm{net}} u ∈ O net . Define
j : O n e t ⟶ R u ⟼ J ( S n e t ( u ) , u ) . \begin{aligned}
j:\quad\mathcal{O}_{\mathrm{net}}&\longrightarrow\mathbb{R}\\
u&\longmapsto J(S_{\mathrm{net}}(u),u).
\end{aligned} j : O net u ⟶ R ⟼ J ( S net ( u ) , u ) .
The Euclidean adjoints a n + 1 ∈ R s d a^{n+1}\in\mathbb{R}^{sd} a n + 1 ∈ R s d satisfy
( A N − 1 ) T a N = − ∂ y N J (83) (A^{N-1})^{T}a^{N}=-\partial_{y^{N}}J \tag*{(83)} ( A N − 1 ) T a N = − ∂ y N J (83)
and, for 1 ≤ k < N 1\leq k<N 1 ≤ k < N ,
( A k − 1 ) T a k = − ∂ y k J − C k T a k + 1 . (A^{k-1})^{T}a^{k}=-\partial_{y^{k}}J-C_{k}^{T}a^{k+1}. ( A k − 1 ) T a k = − ∂ y k J − C k T a k + 1 .
The Euclidean reduced gradient blocks and Hessian bilinear form are
∂ u n j = ∂ u n J + B n T a n + 1 , \partial_{u^{n}}j=\partial_{u^{n}}J+B_{n}^{T}a^{n+1}, ∂ u n j = ∂ u n J + B n T a n + 1 ,
and
D 2 j ( u ) [ d , e ] = D 2 J [ ( z [ d ] , d ) , ( z [ e ] , e ) ] + ∑ n = 0 N − 1 ( a n + 1 ) T D 2 F n [ ( z [ d ] n , z [ d ] n + 1 , d n ) , ( z [ e ] n , z [ e ] n + 1 , e n ) ] . D^{2}j(u)[d,e]=D^{2}J[(z[d],d),(z[e],e)]+\sum_{n=0}^{N-1}(a^{n+1})^{T}D^{2}F_n[(z[d]^{n},z[d]^{n+1},d^{n}),(z[e]^{n},z[e]^{n+1},e^{n})]. D 2 j ( u ) [ d , e ] = D 2 J [( z [ d ] , d ) , ( z [ e ] , e )] + n = 0 ∑ N − 1 ( a n + 1 ) T D 2 F n [( z [ d ] n , z [ d ] n + 1 , d n ) , ( z [ e ] n , z [ e ] n + 1 , e n )] .
Appendix H.4 derives these identities. If only the harvest reward depends on y N y^{N} y N , then
∂ y N J = − ζ η τ diag ( h N − 1 ) ℓ . \partial_{y^{N}}J=-\zeta\eta\tau\operatorname{diag}(h^{N-1})\ell. ∂ y N J = − ζ η τ diag ( h N − 1 ) ℓ .
Thus a N = 0 a^{N}=0 a N = 0 for the old-state rule, while the flux reward gives
( A N − 1 ) T a N = η τ diag ( h N − 1 ) ℓ . (A^{N-1})^{T}a^{N}=\eta\tau\operatorname{diag}(h^{N-1})\ell. ( A N − 1 ) T a N = η τ diag ( h N − 1 ) ℓ .
Residual derivatives and adjoint identities
Proof. The step matrix is
A ( y , u ) = I − τ G ( y , u ) + τ diag ( m ( y , u ) + h ( y , u ) ) . A(y,u)=I-\tau G(y,u)+\tau\operatorname{diag}(m(y,u)+h(y,u)). A ( y , u ) = I − τ G ( y , u ) + τ diag ( m ( y , u ) + h ( y , u )) .
For directions δ = ( ξ , b ) \delta=(\xi,b) δ = ( ξ , b ) and ε = ( χ , c ) \varepsilon=(\chi,c) ε = ( χ , c ) in the state–control variables, its derivatives are
D A ( y , u ) [ δ ] = − τ D G ( y , u ) [ δ ] + τ diag ( D m ( y , u ) [ δ ] + D h ( y , u ) [ δ ] ) , DA(y,u)[\delta]=-\tau DG(y,u)[\delta]+\tau\operatorname{diag}(Dm(y,u)[\delta]+Dh(y,u)[\delta]), D A ( y , u ) [ δ ] = − τ D G ( y , u ) [ δ ] + τ diag ( D m ( y , u ) [ δ ] + D h ( y , u ) [ δ ]) ,
D 2 A ( y , u ) [ δ , ε ] = − τ D 2 G ( y , u ) [ δ , ε ] + τ diag ( D 2 m ( y , u ) [ δ , ε ] + D 2 h ( y , u ) [ δ , ε ] ) . D^{2}A(y,u)[\delta,\varepsilon]=-\tau D^{2}G(y,u)[\delta,\varepsilon]+\tau\operatorname{diag}(D^{2}m(y,u)[\delta,\varepsilon]+D^{2}h(y,u)[\delta,\varepsilon]). D 2 A ( y , u ) [ δ , ε ] = − τ D 2 G ( y , u ) [ δ , ε ] + τ diag ( D 2 m ( y , u ) [ δ , ε ] + D 2 h ( y , u ) [ δ , ε ]) .
The diffusion block is constant, so only the transfer matrix K ( y , u ) K(y,u) K ( y , u ) contributes to D G DG D G and D 2 G D^{2}G D 2 G . Differentiating
F n = A ( y − , u ) y + − y − − τ p ( y − , u ) F_n=A(y^{-},u)y^{+}-y^{-}-\tau p(y^{-},u) F n = A ( y − , u ) y + − y − − τ p ( y − , u )
once gives
D F n [ ξ − , ξ + , b ] = A n ξ + + D A ( y n , u n ) [ ξ − , b ] y n + 1 − ξ − − τ D p ( y n , u n ) [ ξ − , b ] . DF_n[\xi^{-},\xi^{+},b]=A^n\xi^{+}+DA(y^n,u^n)[\xi^{-},b]y^{n+1}-\xi^{-}-\tau Dp(y^n,u^n)[\xi^{-},b]. D F n [ ξ − , ξ + , b ] = A n ξ + + D A ( y n , u n ) [ ξ − , b ] y n + 1 − ξ − − τ D p ( y n , u n ) [ ξ − , b ] .
The residual derivatives with respect to the old state and control are
C n : R s d ⟶ R s d C_n:\mathbb{R}^{sd}\longrightarrow\mathbb{R}^{sd} C n : R s d ⟶ R s d
ξ ⟼ D y A ( y n , u n ) [ ξ ] y n + 1 − ξ − τ D y p ( y n , u n ) [ ξ ] , \xi\longmapsto D_yA(y^n,u^n)[\xi]y^{n+1}-\xi-\tau D_yp(y^n,u^n)[\xi], ξ ⟼ D y A ( y n , u n ) [ ξ ] y n + 1 − ξ − τ D y p ( y n , u n ) [ ξ ] ,
and
B n : R d u ⟶ R s d B_n:\mathbb{R}^{du}\longrightarrow\mathbb{R}^{sd} B n : R d u ⟶ R s d
b ⟼ D u A ( y n , u n ) [ b ] y n + 1 − τ D u p ( y n , u n ) [ b ] . b\longmapsto D_uA(y^n,u^n)[b]y^{n+1}-\tau D_up(y^n,u^n)[b]. b ⟼ D u A ( y n , u n ) [ b ] y n + 1 − τ D u p ( y n , u n ) [ b ] .
For directions ( ξ − , ξ + , b ) , ( χ − , χ + , c ) ∈ R s d × R s d × R d u (\xi^{-},\xi^{+},b),(\chi^{-},\chi^{+},c)\in\mathbb{R}^{sd}\times\mathbb{R}^{sd}\times\mathbb{R}^{du} ( ξ − , ξ + , b ) , ( χ − , χ + , c ) ∈ R s d × R s d × R d u , evaluating the derivatives of A , p A,p A , p at ( y n , u n ) (y^n,u^n) ( y n , u n ) gives
D 2 F n [ ( ξ − , ξ + , b ) , ( χ − , χ + , c ) ] = D 2 A [ ( ξ − , b ) , ( χ − , c ) ] y n + 1 + D A [ ( ξ − , b ) ] χ + + D A [ ( χ − , c ) ] ξ + − τ D 2 p [ ( ξ − , b ) , ( χ − , c ) ] . \begin{aligned}
D^{2}F_n[(\xi^{-},\xi^{+},b),(\chi^{-},\chi^{+},c)]
={}&D^{2}A[(\xi^{-},b),(\chi^{-},c)]y^{n+1}\\
&+DA[(\xi^{-},b)]\chi^{+}+DA[(\chi^{-},c)]\xi^{+}-\tau D^{2}p[(\xi^{-},b),(\chi^{-},c)].
\end{aligned} D 2 F n [( ξ − , ξ + , b ) , ( χ − , χ + , c )] = D 2 A [( ξ − , b ) , ( χ − , c )] y n + 1 + D A [( ξ − , b )] χ + + D A [( χ − , c )] ξ + − τ D 2 p [( ξ − , b ) , ( χ − , c )] .
For the cost and adjoints of Appendix § 22 .§ 3 , set
L : O J × ( R s d ) N ⟶ R \mathcal{L}:\mathcal{O}_J\times(\mathbb{R}^{sd})^N\longrightarrow\mathbb{R} L : O J × ( R s d ) N ⟶ R
( y , u , a ) ⟼ J ( y , u ) + ∑ n = 0 N − 1 ( a n + 1 ) T F n . (y,u,a)\longmapsto J(y,u)+\sum_{n=0}^{N-1}(a^{n+1})^{T}F_n. ( y , u , a ) ⟼ J ( y , u ) + n = 0 ∑ N − 1 ( a n + 1 ) T F n .
The terminal state appears only in J J J and the final residual, so
∂ y N L = ∂ y N J + ( A N − 1 ) T a N . \partial_{y^N}\mathcal{L}=\partial_{y^N}J+(A^{N-1})^{T}a^N. ∂ y N L = ∂ y N J + ( A N − 1 ) T a N .
For 1 ≤ k < N 1\leq k<N 1 ≤ k < N , the state y k y^k y k appears in the two adjacent residuals. Hence
∂ y k L = ∂ y k J + ( A k − 1 ) T a k + C k T a k + 1 . \partial_{y^k}\mathcal{L}=\partial_{y^k}J+(A^{k-1})^{T}a^k+C_k^{T}a^{k+1}. ∂ y k L = ∂ y k J + ( A k − 1 ) T a k + C k T a k + 1 .
Setting these derivatives to zero gives (83) and the backward adjoint equation. For 0 ≤ n < N 0\leq n<N 0 ≤ n < N , the control derivative is
∂ u n L = ∂ u n J + B n T a n + 1 . \partial_{u^n}\mathcal{L}=\partial_{u^n}J+B_n^{T}a^{n+1}. ∂ u n L = ∂ u n J + B n T a n + 1 .
Fix the adjoints at u u u . For every nearby control u ~ \widetilde{u} u , the state equations give j ( u ~ ) = L ( S n e t ( u ~ ) , u ~ , a ~ ) j(\widetilde{u})=\mathcal{L}(S_{\mathrm{net}}(\widetilde{u}),\widetilde{u},\widetilde{a}) j ( u ) = L ( S net ( u ) , u , a ) . The chain rule therefore gives
D j ( u ) [ d ] = D u L [ d ] + D y L [ z [ d ] ] = ∑ n = 0 N − 1 ( ∂ u n J + B n T a n + 1 ) T d n . Dj(u)[d]=D_u\mathcal{L}[d]+D_y\mathcal{L}[z[d]]
=\sum_{n=0}^{N-1}(\partial_{u^n}J+B_n^{T}a^{n+1})^{T}d^n. D j ( u ) [ d ] = D u L [ d ] + D y L [ z [ d ]] = n = 0 ∑ N − 1 ( ∂ u n J + B n T a n + 1 ) T d n .
The state term vanishes because ∂ y k L = 0 \partial_{y^k}\mathcal{L}=0 ∂ y k L = 0 for 1 ≤ k ≤ N 1\leq k\leq N 1 ≤ k ≤ N and z [ d ] 0 = 0 z[d]^0=0 z [ d ] 0 = 0 . A second application of the chain rule, with the adjoints still fixed, gives
D 2 j ( u ) [ d , e ] = D ( y , u ) 2 L [ ( z [ d ] , d ) , ( z [ e ] , e ) ] + D y L [ w [ d , e ] ] . D^{2}j(u)[d,e]=D^{2}_{(y,u)}\mathcal{L}[(z[d],d),(z[e],e)]+D_y\mathcal{L}[w[d,e]]. D 2 j ( u ) [ d , e ] = D ( y , u ) 2 L [( z [ d ] , d ) , ( z [ e ] , e )] + D y L [ w [ d , e ]] .
The last term is zero, since w [ d , e ] 0 = 0 w[d,e]^0=0 w [ d , e ] 0 = 0 . Expanding the remaining second derivative of L \mathcal{L} L gives the Hessian bilinear form.
Numerical methods and supplementary checks
Algorithm details
At iteration k = 0 , 1 , … k=0,1,\ldots k = 0 , 1 , … , write u k ∈ U a d , h , τ u_k\in\mathcal{U}_{\mathrm{ad},h,\tau} u k ∈ U ad , h , τ for the control, g k : = g ( u k ) g_k:=g(u_k) g k := g ( u k ) for its weighted gradient and d k ∈ R N ( M + 1 ) d_k\in\mathbb{R}^{N(M+1)} d k ∈ R N ( M + 1 ) for the search direction.
FBS and Anderson. FBS uses (30) . For damped, regularized Type-II Anderson [21 ] , set f k : = Φ θ ( u k ) − u k f_k:=\Phi_\theta(u_k)-u_k f k := Φ θ ( u k ) − u k , m k : = min { k , 10 } m_k:=\min\{k,10\} m k := min { k , 10 } , and
Δ U k : = [ u k − m k + 1 − u k − m k , … , u k − u k − 1 ] , Δ F k : = [ f k − m k + 1 − f k − m k , … , f k − f k − 1 ] . \Delta U_k:=[u_{k-m_k+1}-u_{k-m_k},\ldots,u_k-u_{k-1}],\qquad\Delta F_k:=[f_{k-m_k+1}-f_{k-m_k},\ldots,f_k-f_{k-1}]. Δ U k := [ u k − m k + 1 − u k − m k , … , u k − u k − 1 ] , Δ F k := [ f k − m k + 1 − f k − m k , … , f k − f k − 1 ] .
Both matrices lie in R N ( M + 1 ) × m k \mathbb{R}^{N(M+1)\times m_k} R N ( M + 1 ) × m k , and we solve for a k ∈ R m k a_k\in\mathbb{R}^{m_k} a k ∈ R m k :
a k : = arg min a ∈ R m k { ∥ f k − Δ F k a ∥ 2 2 + δ ∥ a ∥ 2 2 } . a_k:=\underset{a\in\mathbb{R}^{m_k}}{\arg\min}\{\lVert f_k-\Delta F_k a\rVert_2^2+\delta\lVert a\rVert_2^2\}. a k := a ∈ R m k arg min {∥ f k − Δ F k a ∥ 2 2 + δ ∥ a ∥ 2 2 } .
For m k > 0 m_k>0 m k > 0 , differentiation of the least-squares objective gives the normal equations
( Δ F k T Δ F k + δ I m k ) a k = Δ F k T f k . (\Delta F_k^{T}\Delta F_k+\delta I_{m_k})a_k=\Delta F_k^{T}f_k. ( Δ F k T Δ F k + δ I m k ) a k = Δ F k T f k .
For b ≠ 0 b\ne0 b = 0 , b T ( Δ F k T Δ F k + δ I ) b = ∥ Δ F k b ∥ 2 2 + δ ∥ b ∥ 2 2 > 0 b^T(\Delta F_k^T\Delta F_k+\delta I)b=\lVert\Delta F_kb\rVert_2^2+\delta\lVert b\rVert_2^2>0 b T ( Δ F k T Δ F k + δ I ) b = ∥ Δ F k b ∥ 2 2 + δ ∥ b ∥ 2 2 > 0 , so a k a_k a k is uniquely defined. Omit the correction when m k = 0 m_k=0 m k = 0 :
u k + 1 : = Π [ 0 , U max ] ( Φ θ ( u k ) − ω ( Δ U k + Δ F k ) a k ) . u_{k+1}:=\Pi_{[0,U_{\max}]}\bigl(\Phi_\theta(u_k)-\omega(\Delta U_k+\Delta F_k)a_k\bigr). u k + 1 := Π [ 0 , U m a x ] ( Φ θ ( u k ) − ω ( Δ U k + Δ F k ) a k ) .
We use ω = 0.8 \omega=0.8 ω = 0.8 , δ = 10 − 10 \delta=10^{-10} δ = 1 0 − 10 and no safeguard.
SPG. Spectral projected gradient [22 ] uses a scalar step a k > 0 a_k>0 a k > 0 and
d k : = Π [ 0 , U max ] ( u k − a k g k ) − u k , u k + 1 : = u k + σ k d k . d_k:=\Pi_{[0,U_{\max}]}(u_k-a_kg_k)-u_k,\qquad u_{k+1}:=u_k+\sigma_kd_k. d k := Π [ 0 , U m a x ] ( u k − a k g k ) − u k , u k + 1 := u k + σ k d k .
Because W W W is positive diagonal, coordinate clipping is also the weighted projection onto the box. Its variational inequality, with u k u_k u k as the comparison point, gives
⟨ d k + a k g k , − d k ⟩ h , τ ≥ 0 ⟹ ⟨ g k , d k ⟩ h , τ ≤ − 1 a k ∥ d k ∥ h , τ 2 . \langle d_k+a_kg_k,-d_k\rangle_{h,\tau}\ge0\quad\Longrightarrow\quad\langle g_k,d_k\rangle_{h,\tau}\le-\frac{1}{a_k}\lVert d_k\rVert_{h,\tau}^2. ⟨ d k + a k g k , − d k ⟩ h , τ ≥ 0 ⟹ ⟨ g k , d k ⟩ h , τ ≤ − a k 1 ∥ d k ∥ h , τ 2 .
Thus d k ≠ 0 d_k\ne0 d k = 0 is a descent direction. Moreover, u k + σ d k u_k+\sigma d_k u k + σ d k stays in the box for 0 ≤ σ ≤ 1 0\le\sigma\le1 0 ≤ σ ≤ 1 . The expansion
j ( u k + σ d k ) = j ( u k ) + σ ⟨ g k , d k ⟩ h , τ + o ( σ ) j(u_k+\sigma d_k)=j(u_k)+\sigma\langle g_k,d_k\rangle_{h,\tau}+o(\sigma) j ( u k + σ d k ) = j ( u k ) + σ ⟨ g k , d k ⟩ h , τ + o ( σ )
ensures acceptance for sufficiently small positive σ \sigma σ . Take the first σ k \sigma_k σ k in 1 , 1 / 2 , 1 / 4 , … 1,1/2,1/4,\ldots 1 , 1/2 , 1/4 , … such that
j ( u k + σ k d k ) ≤ max max { 0 , k − 19 } ≤ i ≤ k j ( u i ) + 10 − 4 σ k ⟨ g k , d k ⟩ h , τ . j(u_k+\sigma_kd_k)\le\max_{\max\{0,k-19\}\le i\le k}j(u_i)+10^{-4}\sigma_k\langle g_k,d_k\rangle_{h,\tau}. j ( u k + σ k d k ) ≤ m a x { 0 , k − 19 } ≤ i ≤ k max j ( u i ) + 1 0 − 4 σ k ⟨ g k , d k ⟩ h , τ .
For s k : = u k + 1 − u k s_k:=u_{k+1}-u_k s k := u k + 1 − u k , y k : = g k + 1 − g k ∈ R N ( M + 1 ) y_k:=g_{k+1}-g_k\in\mathbb{R}^{N(M+1)} y k := g k + 1 − g k ∈ R N ( M + 1 ) , the next step a k + 1 a_{k+1} a k + 1 alternates
⟨ s k , s k ⟩ h , τ ⟨ s k , y k ⟩ h , τ , ⟨ s k , y k ⟩ h , τ ⟨ y k , y k ⟩ h , τ , \frac{\langle s_k,s_k\rangle_{h,\tau}}{\langle s_k,y_k\rangle_{h,\tau}},\qquad\frac{\langle s_k,y_k\rangle_{h,\tau}}{\langle y_k,y_k\rangle_{h,\tau}}, ⟨ s k , y k ⟩ h , τ ⟨ s k , s k ⟩ h , τ , ⟨ y k , y k ⟩ h , τ ⟨ s k , y k ⟩ h , τ ,
clipped to [ 10 − 8 , 10 4 ] [10^{-8},10^4] [ 1 0 − 8 , 1 0 4 ] , with the first ratio used for even k k k . When ⟨ s k , y k ⟩ h , τ > 0 \langle s_k,y_k\rangle_{h,\tau}>0 ⟨ s k , y k ⟩ h , τ > 0 , these ratios follow from the two scalar secant fits
arg min b ∈ R ∥ b s k − y k ∥ h , τ 2 = ⟨ s k , y k ⟩ h , τ ⟨ s k , s k ⟩ h , τ , arg min a ∈ R ∥ s k − a y k ∥ h , τ 2 = ⟨ s k , y k ⟩ h , τ ⟨ y k , y k ⟩ h , τ . \underset{b\in\mathbb{R}}{\arg\min}\ \lVert bs_k-y_k\rVert_{h,\tau}^2=\frac{\langle s_k,y_k\rangle_{h,\tau}}{\langle s_k,s_k\rangle_{h,\tau}},\qquad\underset{a\in\mathbb{R}}{\arg\min}\ \lVert s_k-ay_k\rVert_{h,\tau}^2=\frac{\langle s_k,y_k\rangle_{h,\tau}}{\langle y_k,y_k\rangle_{h,\tau}}. b ∈ R arg min ∥ b s k − y k ∥ h , τ 2 = ⟨ s k , s k ⟩ h , τ ⟨ s k , y k ⟩ h , τ , a ∈ R arg min ∥ s k − a y k ∥ h , τ 2 = ⟨ y k , y k ⟩ h , τ ⟨ s k , y k ⟩ h , τ .
The first step is the reciprocal of the fitted Hessian scalar b b b , while the second fits the inverse directly. We start with a 0 = 0.5 a_0=0.5 a 0 = 0.5 and reset to 0.5 0.5 0.5 whenever ⟨ s k , y k ⟩ h , τ ≤ 0 \langle s_k,y_k\rangle_{h,\tau}\le0 ⟨ s k , y k ⟩ h , τ ≤ 0 . Rejected trials need only a state solve.
L-BFGS-B [23 , 24 ] uses W g Wg W g , a memory of 38, ftol=0, at most 60 line-search trials, and a Euclidean projected-gradient tolerance of 10 − 7 min 0 ≤ n < N , 0 ≤ j ≤ M ( τ w j ) 10^{-7}\min_{0\le n<N,\ 0\le j\le M}(\tau w_j) 1 0 − 7 min 0 ≤ n < N , 0 ≤ j ≤ M ( τ w j ) . The iteration limits are 1500 for reference runs and 3000 for refinement and multistart runs.
Derivative, Hessian, and balance tests
Use the reference data of § 10 . At u ≡ 0.7 u \equiv0.7 u ≡ 0.7 , take d ~ , d , e ∈ R N ( M + 1 ) \tilde{d},d,e \in\mathbb{R}^{N(M+1)} d ~ , d , e ∈ R N ( M + 1 ) with
d ~ j : = 0.35 cos ( 2 π x j ) cos ( π t n / T ) + 0.1 ξ j n , d : = d ~ / ∥ d ~ ∥ ∞ . \tilde{d}_j := 0.35\cos(2\pi x_j)\cos(\pi t_n/T) + 0.1\xi_j^n,\qquad d := \tilde{d}/\lVert\tilde{d}\rVert_\infty. d ~ j := 0.35 cos ( 2 π x j ) cos ( π t n / T ) + 0.1 ξ j n , d := d ~ / ∥ d ~ ∥ ∞ .
where ξ j n \xi_j^n ξ j n are independent standard normal samples. For e e e , draw a second independent standard normal array and normalize it in the infinity norm. Set
E 1 ( ε ) : = ∣ j ( u + ε d ) − j ( u ) − ε ⟨ g ( u ) , d ⟩ h , τ ∣ , E g ( ε ) : = ∥ g ( u + ε d ) − g ( u ) − ε H ( u ) d ∥ h , τ , E_1(\varepsilon) := \lvert j(u+\varepsilon d)-j(u)-\varepsilon\langle g(u),d\rangle_{h,\tau}\rvert,\qquad E_g(\varepsilon) := \lVert g(u+\varepsilon d)-g(u)-\varepsilon H(u)d\rVert_{h,\tau}, E 1 ( ε ) := ∣ j ( u + ε d ) − j ( u ) − ε ⟨ g ( u ) , d ⟩ h , τ ∣ , E g ( ε ) := ∥ g ( u + ε d ) − g ( u ) − ε H ( u ) d ∥ h , τ ,
and
E 2 ( ε ) : = ∣ j ( u + ε d ) − j ( u ) − ε ⟨ g ( u ) , d ⟩ h , τ − ε 2 2 ⟨ H ( u ) d , d ⟩ h , τ ∣ . E_2(\varepsilon) := \left\lvert j(u+\varepsilon d)-j(u)-\varepsilon\langle g(u),d\rangle_{h,\tau}-\frac{\varepsilon^2}{2}\langle H(u)d,d\rangle_{h,\tau}\right\rvert. E 2 ( ε ) := j ( u + ε d ) − j ( u ) − ε ⟨ g ( u ) , d ⟩ h , τ − 2 ε 2 ⟨ H ( u ) d , d ⟩ h , τ .
The state maps and reduced costs are smooth near u u u . For small ε \varepsilon ε , the fundamental theorem of calculus gives
j ( u + ε d ) − j ( u ) − ε ⟨ g ( u ) , d ⟩ h , τ = ε 2 ∫ 0 1 ( 1 − t ) ⟨ H ( u + t ε d ) d , d ⟩ h , τ d t , j(u+\varepsilon d)-j(u)-\varepsilon\langle g(u),d\rangle_{h,\tau}=\varepsilon^2\int_0^1(1-t)\langle H(u+t\varepsilon d)d,d\rangle_{h,\tau}\,\mathrm{d}t, j ( u + ε d ) − j ( u ) − ε ⟨ g ( u ) , d ⟩ h , τ = ε 2 ∫ 0 1 ( 1 − t ) ⟨ H ( u + tε d ) d , d ⟩ h , τ d t ,
g ( u + ε d ) − g ( u ) − ε H ( u ) d = ε ∫ 0 1 [ H ( u + t ε d ) − H ( u ) ] d d t . g(u+\varepsilon d)-g(u)-\varepsilon H(u)d=\varepsilon\int_0^1[H(u+t\varepsilon d)-H(u)]d\,\mathrm{d}t. g ( u + ε d ) − g ( u ) − ε H ( u ) d = ε ∫ 0 1 [ H ( u + tε d ) − H ( u )] d d t .
Subtracting the second-order term in the first identity gives
E 2 ( ε ) = ε 2 ∣ ∫ 0 1 ( 1 − t ) ⟨ [ H ( u + t ε d ) − H ( u ) ] d , d ⟩ h , τ d t ∣ . E_2(\varepsilon)=\varepsilon^2\left\lvert\int_0^1(1-t)\langle[H(u+t\varepsilon d)-H(u)]d,d\rangle_{h,\tau}\,\mathrm{d}t\right\rvert. E 2 ( ε ) = ε 2 ∫ 0 1 ( 1 − t ) ⟨[ H ( u + tε d ) − H ( u )] d , d ⟩ h , τ d t .
Let L H L_H L H be a Lipschitz constant for H H H on this neighborhood in the weighted operator norm. Then
E 1 ( ε ) ≤ ε 2 2 sup 0 ≤ t ≤ 1 ∥ H ( u + t ε d ) ∥ h , τ ∥ d ∥ h , τ 2 , E g ( ε ) ≤ L H ε 2 2 ∥ d ∥ h , τ 2 , E 2 ( ε ) ≤ L H ∣ ε ∣ 3 6 ∥ d ∥ h , τ 2 . E_1(\varepsilon)\leq\frac{\varepsilon^2}{2}\sup_{0\leq t\leq1}\lVert H(u+t\varepsilon d)\rVert_{h,\tau}\lVert d\rVert_{h,\tau}^2,\qquad E_g(\varepsilon)\leq\frac{L_H\varepsilon^2}{2}\lVert d\rVert_{h,\tau}^2,\qquad E_2(\varepsilon)\leq\frac{L_H\lvert\varepsilon\rvert^3}{6}\lVert d\rVert_{h,\tau}^2. E 1 ( ε ) ≤ 2 ε 2 0 ≤ t ≤ 1 sup ∥ H ( u + tε d ) ∥ h , τ ∥ d ∥ h , τ 2 , E g ( ε ) ≤ 2 L H ε 2 ∥ d ∥ h , τ 2 , E 2 ( ε ) ≤ 6 L H ∣ ε ∣ 3 ∥ d ∥ h , τ 2 .
These identities predict orders 2, 2 and 3 when the leading terms do not vanish. The measured orders in Table 7 come from fitting log E \log E log E against log ε \log\varepsilon log ε on [ 10 − 3 , 10 − 1 ] [10^{-3},10^{-1}] [ 1 0 − 3 , 1 0 − 1 ] . Smaller steps enter the cancellation regime. The symmetry defect is
Table 7. Derivative checks: E g E_g E g tests the full Hessian action, E 2 E_2 E 2 the second variation.
∣ ⟨ H ( u ) d , e ⟩ h , τ − ⟨ d , H ( u ) e ⟩ h , τ ∣ max { 1 , ∣ ⟨ H ( u ) d , e ⟩ h , τ ∣ , ∣ ⟨ d , H ( u ) e ⟩ h , τ ∣ } . \frac{\lvert\langle H(u)d,e\rangle_{h,\tau}-\langle d,H(u)e\rangle_{h,\tau}\rvert}{\max\{1,\lvert\langle H(u)d,e\rangle_{h,\tau}\rvert,\lvert\langle d,H(u)e\rangle_{h,\tau}\rvert\}}. max { 1 , ∣⟨ H ( u ) d , e ⟩ h , τ ∣ , ∣⟨ d , H ( u ) e ⟩ h , τ ∣} ∣⟨ H ( u ) d , e ⟩ h , τ − ⟨ d , H ( u ) e ⟩ h , τ ∣ .
With u = 2 u=2 u = 2 , the positive solver keeps the states nonnegative on every tested time grid, down to N = 4 N=4 N = 4 , whereas IMEX gives negative states at N = 4 , 6 , 8 N=4,6,8 N = 4 , 6 , 8 . Balance is checked on the raw states, without clipping.
Algorithm comparison and rejected runs
All methods start from zero. FBS uses θ = 0.0093 \theta=0.0093 θ = 0.0093 and Anderson-FBS θ = 0.01 \theta=0.01 θ = 0.01 , both with a 12000-iteration limit. SPG allows 10000 iterations. We use the same relaxation parameters for all three problems. Anderson changes the iteration map and, for IMEX, converges slightly above the unaccelerated threshold from Theorem 8.2 . Table 8 reports final residuals.
Table 8. Algorithm comparison. Acceptance requires the independent check ∥ R ∥ ∞ ≤ 10 − 6 \lVert R\rVert_\infty\le10^{-6} ∥ R ∥ ∞ ≤ 1 0 − 6 . Counts include terminal verification.
Figure 9. Hessian validation: order 2 for E g E_g E g (left), order 3 for E 2 E_2 E 2 before cancellation (right). The flux test includes the nonzero terminal adjoint.
Figure 10. Residual histories on separate iteration scales. Dashed lines mark the common tolerance. Failed P 0 P_0 P 0 runs are included.
State self-convergence
Use the reference data of § 10 . For state self-convergence, fix
u ( x , t ) : = 0.7 + 0.25 sin ( π x ) cos ( 2 π t / T ) . u(x,t) := 0.7 + 0.25\sin(\pi x)\cos(2\pi t/T). u ( x , t ) := 0.7 + 0.25 sin ( π x ) cos ( 2 π t / T ) .
Temporal tests compare terminal states at N N N , 2 N 2N 2 N with M = 160 M = 160 M = 160 . Spatial tests use N = 960 N = 960 N = 960 and restrict the fine states to coarse nodes. In both cases, measure the difference on the coarse grid:
E v : = ( ∑ i = 1 2 ∥ v i , c o a r s e N − v i , f i n e N ∥ h 2 ) 1 / 2 . E_v := \left(\sum_{i=1}^{2}\left\lVert v_{i,\mathrm{coarse}}^N-v_{i,\mathrm{fine}}^N\right\rVert_h^2\right)^{1/2}. E v := ( i = 1 ∑ 2 v i , coarse N − v i , fine N h 2 ) 1/2 .
Each terminal index refers to the final time of its trajectory. The time-refinement difference is
E v t i m e ( N ) = ( ∑ i = 1 2 ∥ v i N , ( M , N ) − v i 2 N , ( M , 2 N ) ∥ h 2 ) 1 / 2 , E_v^{\mathrm{time}}(N) = \left(\sum_{i=1}^{2}\left\lVert v_i^{N,(M,N)}-v_i^{2N,(M,2N)}\right\rVert_h^2\right)^{1/2}, E v time ( N ) = ( i = 1 ∑ 2 v i N , ( M , N ) − v i 2 N , ( M , 2 N ) h 2 ) 1/2 ,
where ( M , N ) (M,N) ( M , N ) records the spatial and temporal resolutions. For spatial refinement, restricting the fine grid to the coarse nodes gives
E v s p a c e ( M ) = ( ∑ i = 1 2 ∑ j = 0 M ω j ∣ v i , j N , ( M , N ) − v i , 2 j N , ( 2 M , N ) ∣ 2 ) 1 / 2 . E_v^{\mathrm{space}}(M) = \left(\sum_{i=1}^{2}\sum_{j=0}^{M}\omega_j\left|v_{i,j}^{N,(M,N)}-v_{i,2j}^{N,(2M,N)}\right|^2\right)^{1/2}. E v space ( M ) = ( i = 1 ∑ 2 j = 0 ∑ M ω j v i , j N , ( M , N ) − v i , 2 j N , ( 2 M , N ) 2 ) 1/2 .
If the measured difference has the asymptotic form E v ( Δ ) = c Δ p + o ( Δ p ) E_v(\Delta)=c\Delta^p+o(\Delta^p) E v ( Δ ) = c Δ p + o ( Δ p ) with c > 0 c>0 c > 0 , then
E v ( Δ ) E v ( Δ / 2 ) = c Δ p + o ( Δ p ) c ( Δ / 2 ) p + o ( Δ p ) ⟶ 2 p , log 2 E v ( Δ ) E v ( Δ / 2 ) ⟶ p . \frac{E_v(\Delta)}{E_v(\Delta/2)}=\frac{c\Delta^p+o(\Delta^p)}{c(\Delta/2)^p+o(\Delta^p)}\longrightarrow2^p,\qquad\log_2\frac{E_v(\Delta)}{E_v(\Delta/2)}\longrightarrow p. E v ( Δ/2 ) E v ( Δ ) = c ( Δ/2 ) p + o ( Δ p ) c Δ p + o ( Δ p ) ⟶ 2 p , log 2 E v ( Δ/2 ) E v ( Δ ) ⟶ p .
The orders log 2 ( E v ( Δ ) / E v ( Δ / 2 ) ) \log_2(E_v(\Delta)/E_v(\Delta/2)) log 2 ( E v ( Δ ) / E v ( Δ/2 )) , with Δ = τ \Delta=\tau Δ = τ or h h h , approach one in time and two in space (Table 9 ). Since P 0 P_0 P 0 and P 1 P_1 P 1 share the positive state solver, they have the same state errors.
Table 9. State self-convergence at fixed control, with separate time and space refinement.