Introduction
Uniqueness and observability are closely related, but they are not equivalent. The absence of nonzero invisible solutions does not prevent sequences of unit-energy states from becoming arbitrarily weakly observed. On bounded domains, such sequences can often be ruled out by a compactness–uniqueness argument. Once an estimate has controlled high-frequency losses, compactness produces an invisible limit of nonzero energy, and unique continuation forces that limit to vanish [2 ] .
On unbounded domains, this argument faces two distinct spatial losses of compactness. Almost invisible states may escape to infinity, where the coefficients and observation geometry can approach limiting configurations not realized by any finite translation. They may also spread out spatially, so that even the largest unit-cell energy tends to zero. These mechanisms are related to the losses of compactness analyzed by P.-L. Lions’s concentration–compactness principle [22 ] . Splitting into separated pieces, corresponding to dichotomy in that principle, is handled by localization. Corollary 2.13 provides compactly supported states whose observation is arbitrarily small relative to their uniformly local norm. Escape to infinity is handled by the translation hull, which retains every limiting configuration. Spreading is handled by cellwise normalization in L u l o c 2 L^2_{\mathrm{uloc}} L uloc 2 , which keeps a maximizing cell at unit mass. The reconstruction assumption (R) then supplies strong local L 2 L^2 L 2 compactness, ensuring that this mass is retained in the limiting invisible profile.
The abstract criterion. Let Γ ω : X → Z \Gamma_\omega:\mathcal{X}\to\mathcal{Z} Γ ω : X → Z , with X = L 2 ( R d ; C N ) \mathcal{X}=L^2(\mathbb{R}^d;\mathbb{C}^N) X = L 2 ( R d ; C N ) and Z = L 2 ( ( 0 , T ) × R d ; C m ) \mathcal{Z}=L^2((0,T)\times\mathbb{R}^d;\mathbb{C}^m) Z = L 2 (( 0 , T ) × R d ; C m ) , be the observation operator associated with a configuration ω \omega ω . We consider the compact full translation hull
W = { θ h ω 0 : h ∈ R d } . \mathcal{W}=\{\theta_h\omega_0:h\in\mathbb{R}^d\}. W = { θ h ω 0 : h ∈ R d } .
of a reference configuration ω 0 \omega_0 ω 0 , endowed with a continuous translation action. The observation operators satisfy the covariance relation
Γ θ h ω τ h = τ h Γ ω . \Gamma_{\theta_h\omega}\tau_h=\tau_h\Gamma_\omega. Γ θ h ω τ h = τ h Γ ω .
where τ h f ( x ) = f ( x + h ) \tau_h f(x)=f(x+h) τ h f ( x ) = f ( x + h ) acts on both states and observations.
Under the assumptions of Section 2, Theorem 2.1 establishes
∃ c > 0 : ∥ Γ ω 0 f ∥ Z ≥ c ∥ f ∥ X ( f ∈ X ) ⟺ ker L u l o c 2 Γ ω = { 0 } ( ω ∈ W ) . \exists c>0:\quad\|\Gamma_{\omega_0}f\|_{\mathcal{Z}}\ge c\|f\|_{\mathcal{X}}\quad(f\in\mathcal{X})
\quad\Longleftrightarrow\quad
\ker_{L^2_{\mathrm{uloc}}}\Gamma_\omega=\{0\}\quad(\omega\in\mathcal{W}). ∃ c > 0 : ∥ Γ ω 0 f ∥ Z ≥ c ∥ f ∥ X ( f ∈ X ) ⟺ ker L uloc 2 Γ ω = { 0 } ( ω ∈ W ) .
Here L u l o c 2 L^2_{\mathrm{uloc}} L uloc 2 denotes the space of functions whose L 2 L^2 L 2 norms on unit cubes are uniformly bounded, without requiring finite global energy. The observation operators admit canonical extensions to this space. We call the condition on the right generalized uniqueness .
We assume that for each bounded-support input the observed energy on each unit cube depends continuously on ω \omega ω . The influence of input cube j j j on output cube k k k is bounded by κ ( k − j ) \kappa(k-j) κ ( k − j ) for a common κ ∈ ℓ 1 ( Z d ) \kappa\in\ell^1(\mathbb{Z}^d) κ ∈ ℓ 1 ( Z d ) . The remaining assumption is a uniform reconstruction property (R). There are uniformly bounded operators B ω B_\omega B ω such that the errors I − B ω Γ ω I-B_\omega\Gamma_\omega I − B ω Γ ω are collectively compact on bounded windows for inputs of bounded uniformly local energy. In the transport applications, a decomposition of the Gramian Γ ω ∗ Γ ω \Gamma_\omega^*\Gamma_\omega Γ ω ∗ Γ ω into a coercive multiplication operator and a locally compact remainder provides such a reconstruction. Conversely, (R) is necessary for observability (Proposition 2.17). Under (H1)–(H3) on the full hull, observability is therefore equivalent to (R) together with generalized uniqueness (Corollary 2.18).
Localization transfers a failure of the global lower bound to states whose observation is small compared with their largest unit-cell norm. We normalize this norm to one and translate a maximizing cell to the origin. Property (R) gives a strongly convergent subsequence on each bounded window, so the limit has norm one on the central cell. Continuity of the cell energies and summability of κ \kappa κ then show that this limit is invisible for a configuration in the hull.
The requirement of uniqueness over the full translation hull is essential. As illustrated by the examples in § 5 , uniqueness for the original configuration, even in the distributional sense, need not imply observability. The obstruction may instead arise from a limiting configuration at infinity.
Transport systems. We now turn to the main application of the abstract criterion: partially observed first-order transport systems
∂ t y + A ∂ x y = B 1 Ω u , ( t , x ) ∈ ( 0 , T ) × R , (1) \partial_t y + A\partial_x y = B\mathbf{1}_{\Omega}u,\qquad(t,x)\in(0,T)\times\mathbb{R}, \tag*{(1)} ∂ t y + A ∂ x y = B 1 Ω u , ( t , x ) ∈ ( 0 , T ) × R , (1)
where A ∈ R n × n A\in\mathbb{R}^{n\times n} A ∈ R n × n has real spectrum, B ∈ R n × m B\in\mathbb{R}^{n\times m} B ∈ R n × m , and Ω ⊂ R \Omega\subset\mathbb{R} Ω ⊂ R is the control region. The adjoint system
∂ t q + A T ∂ x q = 0 , q ( T ) = q T , (2) \partial_t q + A^{\mathsf T}\partial_xq=0,\qquad q(T)=q_T, \tag*{(2)} ∂ t q + A T ∂ x q = 0 , q ( T ) = q T , (2)
is observed through B T q B^{\mathsf T}q B T q on ( 0 , T ) × Ω (0,T)\times\Omega ( 0 , T ) × Ω . We assume the Kalman condition
rank ( B , A B , … , A n − 1 B ) = n , (3) \operatorname{rank}(B,AB,\ldots,A^{n-1}B)=n, \tag*{(3)} rank ( B , A B , … , A n − 1 B ) = n , (3)
and let K K K be the smallest integer such that rank ( B , A B , … , A K B ) = n \operatorname{rank}(B,AB,\ldots,A^K B)=n rank ( B , A B , … , A K B ) = n .
Two features distinguish this problem. First, the Kalman condition does not guarantee L 2 L^2 L 2 observability, even on the whole line. For
A T = ( λ 1 0 λ ) , B T = ( 1 , 0 ) , A^{\mathsf T}=\begin{pmatrix}\lambda&1\\0&\lambda\end{pmatrix},\qquad B^{\mathsf T}=(1,0), A T = ( λ 0 1 λ ) , B T = ( 1 , 0 ) ,
one has, for the terminal datum f = q T f=q_T f = q T and with s = T − t s=T-t s = T − t ,
B T e s A T ∂ x f ( x ) = f 1 ( x + λ s ) + s ∂ x f 2 ( x + λ s ) . B^{\mathsf T}e^{sA^{\mathsf T}\partial_x}f(x)=f_1(x+\lambda s)+s\partial_xf_2(x+\lambda s). B T e s A T ∂ x f ( x ) = f 1 ( x + λ s ) + s ∂ x f 2 ( x + λ s ) .
Consequently, for every T > 0 T>0 T > 0 , the time-integrated observation satisfies
∫ 0 T ∥ B T q ( T − s , ⋅ ) ∥ 2 2 d s ≍ T ∥ f ∥ Y K , J − 2 = ∥ f 1 ∥ 2 2 + ∥ ∂ x f 2 ∥ 2 2 . \int_0^T\lVert B^{\mathsf T}q(T-s,\cdot)\rVert_2^2\,ds\asymp_T\lVert f\rVert_{\mathcal{Y}_{K,J}^-}^2=\lVert f_1\rVert_2^2+\lVert\partial_xf_2\rVert_2^2. ∫ 0 T ∥ B T q ( T − s , ⋅ ) ∥ 2 2 d s ≍ T ∥ f ∥ Y K , J − 2 = ∥ f 1 ∥ 2 2 + ∥ ∂ x f 2 ∥ 2 2 .
For example, take f 1 = 0 f_1=0 f 1 = 0 and f 2 ( x ) = R − 1 / 2 φ ( x / R ) f_2(x)=R^{-1/2}\varphi(x/R) f 2 ( x ) = R − 1/2 φ ( x / R ) with φ ∈ C c ∞ ( R ) \varphi\in C_c^\infty(\mathbb{R}) φ ∈ C c ∞ ( R ) and ∥ φ ∥ 2 = 1 \lVert\varphi\rVert_2=1 ∥ φ ∥ 2 = 1 . These data have unit L 2 L^2 L 2 norm, while ∥ ∂ x f 2 ∥ 2 2 = R − 1 ∥ φ ′ ∥ 2 2 → 0 \lVert\partial_xf_2\rVert_2^2=R^{-1}\lVert\varphi'\rVert_2^2\to0 ∥ ∂ x f 2 ∥ 2 2 = R − 1 ∥ φ ′ ∥ 2 2 → 0 . Whole-line observation determines, up to norm equivalence, the intrinsic Kalman–Jordan space Y K , J − \mathcal{Y}_{K,J}^- Y K , J − recalled in § 6.1 , where J J J records the Jordan structure. Its dual Y K , J + \mathcal{Y}_{K,J}^+ Y K , J + is the state space for control. A Fourier normalization identifies Y K , J − \mathcal{Y}_{K,J}^- Y K , J − with L 2 L^2 L 2 ; the uniformly local uniqueness condition applies to the normalized datum.
Second, the observation region may be an arbitrary, possibly non-periodic, union of intervals
Ω = ⋃ j ∈ Z ( a j , b j ) , a j < b j ≤ a j + 1 , ℓ ∘ : = inf j ( b j − a j ) > 0 , G ∘ : = sup j ( a j + 1 − b j ) < ∞ , (4) \Omega=\bigcup_{j\in\mathbb{Z}}(a_j,b_j),\qquad a_j<b_j\leq a_{j+1},\qquad\ell_\circ:=\inf_j(b_j-a_j)>0,\qquad G^\circ:=\sup_j(a_{j+1}-b_j)<\infty, \tag*{(4)} Ω = j ∈ Z ⋃ ( a j , b j ) , a j < b j ≤ a j + 1 , ℓ ∘ := j inf ( b j − a j ) > 0 , G ∘ := j sup ( a j + 1 − b j ) < ∞ , (4)
with the case Ω = R \Omega=\mathbb{R} Ω = R included by setting G ∘ = 0 G^\circ=0 G ∘ = 0 . When σ ( A ) ⊂ R ∖ { 0 } \sigma(A)\subset\mathbb{R}\setminus\{0\} σ ( A ) ⊂ R ∖ { 0 } , the gap geometry and the smallest characteristic speed determine the time scale
λ ∘ : = min λ ∈ σ ( A ) ∣ λ ∣ , T ∘ : = G ∘ λ ∘ . (5) \lambda_\circ:=\min_{\lambda\in\sigma(A)}|\lambda|,\qquad T_\circ:=\frac{G^\circ}{\lambda_\circ}. \tag*{(5)} λ ∘ := λ ∈ σ ( A ) min ∣ λ ∣ , T ∘ := λ ∘ G ∘ . (5)
This is the supremum of the times that a characteristic with the smallest speed can spend in a gap of Ω \Omega Ω .
Main contributions. Besides the abstract criterion of Theorem 2.1 , our main results are the following.
Transport with arbitrary Jordan structure (§ 6 ). Under (1.3) and σ ( A ) ⊂ R ∖ { 0 } \sigma(A)\subset\mathbb{R}\setminus\{0\} σ ( A ) ⊂ R ∖ { 0 } , for every T > T ∘ T>T_\circ T > T ∘ , Corollary 5.10 gives
c T ∥ q T ∥ Y K , J − 2 ≤ ∫ 0 T ∫ Ω ∣ B T q ∣ 2 d x d t ≤ C T ∥ q T ∥ Y K , J − 2 . (6) c_T\lVert q_T\rVert_{\mathcal{Y}_{K,J}^-}^2\leq\int_0^T\int_\Omega|B^{\mathsf T}q|^2\,dx\,dt\leq C_T\lVert q_T\rVert_{\mathcal{Y}_{K,J}^-}^2. \tag*{(6)} c T ∥ q T ∥ Y K , J − 2 ≤ ∫ 0 T ∫ Ω ∣ B T q ∣ 2 d x d t ≤ C T ∥ q T ∥ Y K , J − 2 . (6)
Consequently, the controlled system is exactly controllable in Y K , J + \mathcal{Y}_{K,J}^+ Y K , J + with L 2 L^2 L 2 controls. If G ∘ > 0 G^\circ>0 G ∘ > 0 , observability fails for every 0 < T ≤ T ∘ 0<T\leq T_\circ 0 < T ≤ T ∘ , even when no gap attains length G ∘ G^\circ G ∘ , as proved in Proposition C.1 . For a bounded Lipschitz weight, Theorem 5.9 characterizes observability by a positive lower bound on the time integral of the weight along every characteristic, including stationary ones. For a single 2 × 2 2\times2 2 × 2 Jordan block, Proposition 5.12 determines the optimal squared observability cost near T ∘ T_\circ T ∘ . With the intrinsic norm fixed in § 6.7 , C o b s ( T , Ω ) ≳ 12 ( 1 + T ∘ 2 ) ( T − T ∘ ) − 3 C_{\mathrm{obs}}(T,\Omega)\gtrsim12(1+T_\circ^2)(T-T_\circ)^{-3} C obs ( T , Ω ) ≳ 12 ( 1 + T ∘ 2 ) ( T − T ∘ ) − 3 as T ↓ T ∘ T\downarrow T_\circ T ↓ T ∘ . (b) Moving observation regions (§ 7 ). For Ω + h ( t ) \Omega+ h(t) Ω + h ( t ) with ∥ h ′ ′ ∥ L ∞ ( 0 , T ) < λ ∙ \lVert h''\rVert_{L^\infty(0,T)} < \lambda_\bullet ∥ h ′′ ∥ L ∞ ( 0 , T ) < λ ∙ , observability holds exactly when
min λ ∈ σ ( A ) ∣ λ T − h ( T ) + h ( 0 ) ∣ > G ∙ . \min_{\lambda\in\sigma(A)} \lvert\lambda T-h(T)+h(0)\rvert> G^\bullet. λ ∈ σ ( A ) min ∣ λ T − h ( T ) + h ( 0 )∣ > G ∙ .
For an affine motion h ( t ) = h 0 + v t h(t)=h_0+vt h ( t ) = h 0 + v t , the change to a moving frame allows arbitrary v v v and also zero characteristic speeds. If G ∙ > 0 G^\bullet>0 G ∙ > 0 , the condition becomes
T min λ ∈ σ ( A ) ∣ λ − v ∣ > G ∙ . T\min_{\lambda\in\sigma(A)}\lvert\lambda-v\rvert>G^\bullet. T λ ∈ σ ( A ) min ∣ λ − v ∣ > G ∙ .
The relative speeds λ − v \lambda-v λ − v allow a moving region to observe stationary components.
(c) Transport with nonlocal dispersion in R d \mathbb{R}^d R d (§ 8 ). We apply the abstract criterion to
∂ t u + V ⋅ ∇ u = J ∗ u − u \partial_tu+V\cdot\nabla u=J*u-u ∂ t u + V ⋅ ∇ u = J ∗ u − u
with an integrable kernel J J J and a Lipschitz observation weight a : R d → [ 0 , 1 ] a:\mathbb{R}^d\to[0,1] a : R d → [ 0 , 1 ] . For each T > 0 T>0 T > 0 , weighted L 2 L^2 L 2 observability holds if and only if
inf z ∈ R d ∫ 0 T a ( z + t V ) d t > 0. \inf_{z\in\mathbb{R}^d}\int_0^T a(z+tV)\,\mathrm{d}t>0. z ∈ R d inf ∫ 0 T a ( z + t V ) d t > 0.
(d) Anisotropic heat flux (§ 9 ). For the heat equation with a constant positive definite diffusion matrix, we combine whole-space heat-flux measurements with temperature measurements on a subset. Observability is equivalent to thickness of this subset. The observation satisfies (R), whereas its Gramian admits no decomposition of the form (H4)–(H5) when the diffusion matrix has two distinct eigenvalues.
The principal transport and nonlocal applications verify (R) through (H4)–(H5). The additional scope of (R) is structural: it is preserved by the localized output transformations of Lemma 2.20, while (H4)–(H5) need not be, as the heat-flux example illustrates.
Idea of the proofs. The transport analysis begins with a rational Fourier multiplier Q = D H Q=DH Q = D H identifying Y K , J − \mathcal{Y}_{K,J}^{-} Y K , J − with L 2 L^2 L 2 . We use this order to control both frequency regimes without assuming that the Kalman and Jordan projectors commute. After this normalization, the observation operator splits into characteristic translations and an exponentially decaying convolution remainder. Characteristic occupation estimates provide the coercive part of the Gramian, while the Kalman condition and uniqueness along characteristics, combined with a lifting in S ′ \mathcal{S}' S ′ , yield generalized uniqueness on the hull. For nonlocal dispersion, generalized uniqueness follows instead from analyticity in time in coordinates moving with the drift.
Related work. Previous transport results. Earlier work [31 ] establishes localized observation for diagonalizable systems with nonzero speeds and whole-line observation for real Jordan blocks. Here we establish localized observability for arbitrary real Jordan blocks on the line and prove failure at the critical time even when the gap supremum is not attained. The higher-dimensional transport application concerns a scalar equation with one drift velocity and a convolution term.
Hulls and limit operators. Favard’s work on almost-periodic differential equations already used hulls and a separation condition for bounded solutions of every associated homogeneous equation [11 ] .
In operator theory, limit operators characterize Fredholm properties through the behavior of an operator at infinity [27 , 20 , 5 ] . A closer connection is provided by generalized collectively compact operator theory [6 , 5 ] . In particular, [5 Theorem 5.9] derives a uniform lower bound from injectivity for a translation-invariant family of operators of the form I − K I-K I − K , under strict-topology compactness assumptions.
Theorem 2.1 follows this localization and compactness mechanism, but is formulated directly for observation maps Γ ω : X → Z \Gamma_\omega:\mathcal{X}\to\mathcal{Z} Γ ω : X → Z . It uses uniqueness for the underlying evolution equation and continuity only of the observed cell energies. The reconstruction operators B ω B_\omega B ω in (R) need only be uniformly bounded; neither their covariance nor their continuity in ω \omega ω is assumed.
The factorization B ω Γ ω = I − S ω B_\omega\Gamma_\omega=I-S_\omega B ω Γ ω = I − S ω does not by itself allow a direct application of [5 Theorem 5.9] to this family: (R) imposes neither translation invariance nor the strict-topology assumptions on the reconstructed operators. Moreover, injectivity of Γ ω \Gamma_\omega Γ ω need not imply injectivity of a chosen reconstruction B ω Γ ω B_\omega\Gamma_\omega B ω Γ ω . Under (H4)–(H5), the choice B ω = D ω − 1 Γ ω ∗ B_\omega=D_\omega^{-1}\Gamma_\omega^* B ω = D ω − 1 Γ ω ∗ gives the normalized Gramian D ω − 1 Γ ω ∗ Γ ω = I + D ω − 1 K ω D_\omega^{-1}\Gamma_\omega^*\Gamma_\omega=I+D_\omega^{-1}K_\omega D ω − 1 Γ ω ∗ Γ ω = I + D ω − 1 K ω . Those additional assumptions fail for the examples in Sections 2 and § 9 .
For uniformly local states of infinite global energy, the identity ⟨ Γ ω ∗ Γ ω w , w ⟩ = ∥ Γ ω w ∥ 2 \langle\Gamma_\omega^*\Gamma_\omega w,w\rangle=\lVert\Gamma_\omega w\rVert^2 ⟨ Γ ω ∗ Γ ω w , w ⟩ = ∥ Γ ω w ∥ 2 is unavailable. Injectivity of the Gramian therefore cannot be inferred from observation injectivity. Under our hypotheses, Proposition 2.3 nevertheless proves that the two injectivity conditions, imposed on the whole hull, are equivalent. This is a consequence of the criterion. The contribution is the direct observation formulation with (R), rather than a strictly weaker injectivity condition under the same hypotheses. Proposition 2.17, a consequence of the lower-bound transfer in Proposition 2.12 , also shows that (R) is necessary for observability. The characterization in Corollary 2.18 clarifies the role of this assumption within the observation framework.
Proposition 2.12 , which transfers lower bounds between exponents, adapts the block argument of [30 Theorem 2.1] , written there for scalar matrices, to operator-valued blocks and to families with a common profile.
Localization and loss of compactness. Our approach is related to concentration–compactness [22 ] and uniformly local methods [4 ] . Limiting media also play a role in the analysis of damped waves in periodic and asymptotically periodic settings [26 , 17 ] .
Hyperbolic and coupled systems. Our results complement geometric control for hyperbolic systems [2 , 3 , 1 , 16 ] , partial dissipation [9 ] , and algebraic recovery through coupling [19 , 23 , 18 , 34 ] . Unlike the bounded-interval setting of [16 ] , we work on the whole line with possibly non-periodic observation regions. For non-diagonalizable hyperbolic evolution, see also [14 ] .
Unbounded domains and moving observations. For the heat equation on R d \mathbb{R}^{d} R d , observability of the final temperature is characterized by thickness of the observation set [10 , 33 ] . Our bounded-gap regions are thick, but finite propagation introduces a sharp positive observation time when the set has a gap of positive length. Moving observation regions have been studied for waves [28 ] and for systems with stationary components [24 , 7 , 8 ] . For spectral approaches to observability, see [29 , 25 ] .
Outline. Sections 2 and 3 develop the abstract criterion and its geometric ingredients. Section 4 illustrates the necessity of the full hull. Sections 5 and 6 treat fixed and moving observation regions; Section 7 applies the abstract criterion to transport with nonlocal dispersion in R d \mathbb{R}^{d} R d . Section 8 illustrates the reconstruction criterion with anisotropic heat-flux measurements. The appendices collect two auxiliary lemmas and prove the necessity of the Kalman condition and the sharpness of the geometric time threshold.
Notation. We write ⊤ \top ⊤ for transposition and ∗ \ast ∗ for the Hermitian or Hilbert-space adjoint, so that M ∗ = M ‾ ⊤ M^{\ast}=\overline{M}^{\top} M ∗ = M ⊤ . We use f ^ ( ξ ) = ( 2 π ) − d / 2 ∫ R d e − i x ξ f ( x ) d x \widehat{f}(\xi)=(2\pi)^{-d/2}\int_{\mathbb{R}^{d}}e^{-ix\xi}f(x)\,dx f ( ξ ) = ( 2 π ) − d /2 ∫ R d e − i x ξ f ( x ) d x and D x = − i ∂ x D_x=-i\partial_x D x = − i ∂ x . For s = T − t s=T-t s = T − t , q ^ ( T − s , ξ ) = e i s ξ A q ^ T ( ξ ) \widehat{q}(T-s,\xi)=e^{is\xi A}\widehat{q}_{T}(\xi) q ( T − s , ξ ) = e i s ξ A q T ( ξ ) .
An observability criterion via translation hulls in arbitrary dimension
The abstract criterion is formulated for every d ≥ 1 d\ge1 d ≥ 1 . The transport applications of Sections 4–6 use its specialization to d = 1 d=1 d = 1 , retaining several characteristic speeds and arbitrary Jordan structure; Section 7 also applies it for d ≥ 2 d\ge2 d ≥ 2 . A translation hull records the environments seen after spatial recentering. The translated observation sets Ω − x n \Omega-x_n Ω − x n provide a concrete example: their local compactness is explained below and illustrates the choice of topology and the operator assumptions.
Fix T > 0 T>0 T > 0 and put
X = L 2 ( R d ; C N ) , Z = L 2 ( ( 0 , T ) × R d ; C m ) . \mathcal{X}=L^{2}(\mathbb{R}^{d};\mathbb{C}^{N}),\qquad\mathcal{Z}=L^{2}((0,T)\times\mathbb{R}^{d};\mathbb{C}^{m}). X = L 2 ( R d ; C N ) , Z = L 2 (( 0 , T ) × R d ; C m ) .
The operators introduced below map inputs in X \mathcal{X} X to observations in Z \mathcal{Z} Z . Scalar functions of x x x act by multiplication on both spaces. For k ∈ Z d k\in\mathbb{Z}^{d} k ∈ Z d , let Q k = k + [ 0 , 1 ) d Q_k=k+[0,1)^{d} Q k = k + [ 0 , 1 ) d and let 1 k \mathbf{1}_k 1 k denote multiplication by 1 Q k \mathbf{1}_{Q_k} 1 Q k ; the corresponding observation cell is ( 0 , T ) × Q k (0,T)\times Q_k ( 0 , T ) × Q k . For 1 ≤ p ≤ ∞ 1\le p\le\infty 1 ≤ p ≤ ∞ , define
∥ f ∥ ( p ) = ∥ ( ∥ 1 k f ∥ L 2 ) k ∈ Z d ∥ ℓ p . (7) \|f\|_{(p)}=\left\|\left(\|\mathbf{1}_k f\|_{L^{2}}\right)_{k\in\mathbb{Z}^{d}}\right\|_{\ell^{p}}. \tag*{(7)} ∥ f ∥ ( p ) = ( ∥ 1 k f ∥ L 2 ) k ∈ Z d ℓ p . (7)
This norm combines the local L 2 L^{2} L 2 sizes across spatial cells by ℓ p \ell^{p} ℓ p summation. The corresponding spaces are denoted by X p \mathcal{X}_{p} X p and Z p \mathcal{Z}_{p} Z p ; they are Wiener amalgams [12 , 13 , 15 ] . In particular, X 2 = X \mathcal{X}_{2}=\mathcal{X} X 2 = X , Z 2 = Z \mathcal{Z}_{2}=\mathcal{Z} Z 2 = Z , and
X ∞ = L u l o c 2 ( R d ; C N ) , ∥ f ∥ L u l o c 2 = sup k ∥ f ∥ L 2 ( Q k ) . \mathcal{X}_{\infty}=L^{2}_{\mathrm{uloc}}(\mathbb{R}^{d};\mathbb{C}^{N}),\qquad\|f\|_{L^{2}_{\mathrm{uloc}}}=\sup_{k}\|f\|_{L^{2}(Q_k)}. X ∞ = L uloc 2 ( R d ; C N ) , ∥ f ∥ L uloc 2 = k sup ∥ f ∥ L 2 ( Q k ) .
Thus the uniformly local norm bounds the L 2 L^{2} L 2 size in every cell without requiring a finite total L 2 L^{2} L 2 norm.
Throughout, spatial distances and lattice indices use the maximum norm ∣ ⋅ ∣ |\cdot| ∣ ⋅ ∣ ; this includes ∣ k − j ∣ |k-j| ∣ k − j ∣ , the localization profiles, and ∣ x − y ∣ |x-y| ∣ x − y ∣ in (21) . Vector norms are Euclidean and matrix norms are the induced operator norms. Lipschitz constants use the Euclidean distance:
Lip ( χ ) = sup x ≠ y ∣ χ ( x ) − χ ( y ) ∣ ∣ x − y ∣ 2 . \operatorname{Lip}(\chi)=\sup_{x\ne y}\frac{|\chi(x)-\chi(y)|}{|x-y|_{2}}. Lip ( χ ) = x = y sup ∣ x − y ∣ 2 ∣ χ ( x ) − χ ( y ) ∣ .
The inequality ∣ x ∣ 2 ≤ d ∣ x ∣ |x|_{2}\le\sqrt{d}|x| ∣ x ∣ 2 ≤ d ∣ x ∣ accounts for the factors d \sqrt{d} d in Proposition 2.12 .
The operator assumptions
Spatial translations. We fix the convention
( τ h f ) ( x ) = f ( x + h ) , h , x ∈ R d , (8) (\tau_h f)(x)=f(x+h),\qquad h,x\in\mathbb{R}^{d}, \tag*{(8)} ( τ h f ) ( x ) = f ( x + h ) , h , x ∈ R d , (8)
with the same translation of the spatial variable on Z \mathcal{Z} Z . These operators are unitary on X \mathcal{X} X and Z \mathcal{Z} Z and satisfy
τ 0 = I , τ h τ k = τ h + k , τ h − 1 = τ − h . \tau_0=I,\qquad\tau_h\tau_k=\tau_{h+k},\qquad\tau_h^{-1}=\tau_{-h}. τ 0 = I , τ h τ k = τ h + k , τ h − 1 = τ − h .
We use superscripts X \mathcal{X} X and Z \mathcal{Z} Z when it is useful to distinguish translations of inputs and outputs. With this convention, τ x n f \tau_{x_n}f τ x n f expresses the profile f f f in coordinates centered at x n x_n x n .
Translations of the environment. An environment ω \omega ω describes the spatial configuration on which the observation operator depends. Let W \mathcal{W} W be a compact metrizable space of such configurations, equipped with maps θ h : W → W \theta_h:\mathcal{W}\to\mathcal{W} θ h : W → W satisfying
θ 0 ω = ω , θ h ( θ k ω ) = θ h + k ω ( h , k ∈ R d , ω ∈ W ) . (9) \theta_0\omega=\omega,\qquad\theta_h(\theta_k\omega)=\theta_{h+k}\omega\qquad(h,k\in\mathbb{R}^{d},\ \omega\in\mathcal{W}). \tag*{(9)} θ 0 ω = ω , θ h ( θ k ω ) = θ h + k ω ( h , k ∈ R d , ω ∈ W ) . (9)
We require ( h , ω ) ↦ θ h ω (h,\omega)\mapsto\theta_h\omega ( h , ω ) ↦ θ h ω to be jointly continuous. The element θ h ω \theta_h\omega θ h ω is the environment seen after the recentering (8) . The action and the topology of W \mathcal{W} W are part of the model.
Example from the translated observation sets. In one dimension, let
Ω = ⋃ j ∈ Z ( a j , b j ) , a j < b j ≤ a j + 1 , ℓ ∙ : = inf j ( b j − a j ) > 0. \Omega= \bigcup_{j\in\mathbb{Z}}(a_j,b_j), \qquad a_j < b_j \le a_{j+1}, \qquad\ell_{\bullet} := \inf_j(b_j-a_j)>0. Ω = j ∈ Z ⋃ ( a j , b j ) , a j < b j ≤ a j + 1 , ℓ ∙ := j inf ( b j − a j ) > 0.
With the translation convention above, the action on observation sets is
θ h D = D − h , 1 D − h ( x ) = 1 D ( x + h ) . (10) \theta_hD = D-h,\qquad\mathbf{1}_{D-h}(x)=\mathbf{1}_D(x+h). \tag*{(10)} θ h D = D − h , 1 D − h ( x ) = 1 D ( x + h ) . (10)
Indeed, x x x belongs to the recentered set exactly when x + h x+h x + h belongs to the original set. Representing environments by their indicators gives
ω 0 = 1 Ω , ω n = θ x n ω 0 = 1 Ω − x n . \omega_0=\mathbf{1}_{\Omega},\qquad\omega_n=\theta_{x_n}\omega_0=\mathbf{1}_{\Omega-x_n}. ω 0 = 1 Ω , ω n = θ x n ω 0 = 1 Ω − x n .
These translated indicators form a relatively compact family in L l o c 1 ( R ) L^1_{\mathrm{loc}}(\mathbb{R}) L loc 1 ( R ) . To see this, fix a bounded open interval I I I . At most two components meeting I I I are not contained in I I I ; all the others have length at least ℓ ∙ \ell_{\bullet} ℓ ∙ . Thus, uniformly in n n n ,
# { components of Ω − x n meeting I } ≤ ∣ I ∣ ℓ ∙ + 2. \#\{\text{components of }\Omega-x_n\text{ meeting }I\}\le\frac{|I|}{\ell_{\bullet}}+2. # { components of Ω − x n meeting I } ≤ ℓ ∙ ∣ I ∣ + 2.
After intersection with I I I , their endpoints lie in the compact interval I ‾ \overline{I} I . Passing to a subsequence fixes the number of intervals and makes all endpoints converge. The resulting indicators converge in L 1 ( I ) L^1(I) L 1 ( I ) , also when some limiting intervals collapse or become adjacent. A diagonal extraction over I = ( − m , m ) I=(-m,m) I = ( − m , m ) therefore gives
ω n ⟶ ω ∞ = 1 Ω ∞ in L l o c 1 ( R ) . \omega_n\longrightarrow\omega_\infty=\mathbf{1}_{\Omega_\infty}\qquad\text{in }L^1_{\mathrm{loc}}(\mathbb{R}). ω n ⟶ ω ∞ = 1 Ω ∞ in L loc 1 ( R ) .
Locally, the limit is an indicator with finitely many jumps; hence Ω ∞ \Omega_\infty Ω ∞ may be chosen open, up to a null set. For every bounded interval I I I , the convergence means
∥ ω n − ω ∞ ∥ L 1 ( I ) = ∣ ( ( Ω − x n ) △ Ω ∞ ) ∩ I ∣ ⟶ 0. \|\omega_n-\omega_\infty\|_{L^1(I)}=|((\Omega-x_n)\mathbin{\triangle}\Omega_\infty)\cap I|\longrightarrow0. ∥ ω n − ω ∞ ∥ L 1 ( I ) = ∣ (( Ω − x n ) △ Ω ∞ ) ∩ I ∣ ⟶ 0.
Thus the configurations differ on a set of vanishing measure in each fixed observation window. Sets are identified almost everywhere. The orbit closure is compact in this topology, and spatial translations act jointly continuously on it. This supplies a concrete parameter space for the abstract criterion. Here θ h \theta_h θ h translates the observation geometry, while τ h \tau_h τ h translates states and observations.
Operator family. For each ω ∈ W \omega\in\mathcal{W} ω ∈ W , let Γ ω ∈ L ( X , Z ) \Gamma_\omega\in\mathcal{L}(X,Z) Γ ω ∈ L ( X , Z ) . In the indicator parametrization, we write Γ D : = Γ 1 D \Gamma_D:=\Gamma_{\mathbf{1}_D} Γ D := Γ 1 D . The following assumptions describe compatibility with translations and limits of environments, together with uniform localization and local compactness.
(H1) Covariance:
Γ θ h ω τ h X = τ h Z Γ ω ( h ∈ R d , ω ∈ W ) . (11) \Gamma_{\theta_h\omega}\tau_h^X=\tau_h^Z\Gamma_\omega\qquad(h\in\mathbb{R}^d,\ \omega\in\mathcal{W}). \tag*{(11)} Γ θ h ω τ h X = τ h Z Γ ω ( h ∈ R d , ω ∈ W ) . (11)
Recenter the input and the environment together: the resulting observation is the original observation expressed in the new coordinates. Equivalently,
Γ θ h ω = τ h Z Γ ω τ − h X . (12) \Gamma_{\theta_h\omega}=\tau_h^Z\Gamma_\omega\tau_{-h}^X. \tag*{(12)} Γ θ h ω = τ h Z Γ ω τ − h X . (12)
For the translated observation sets, this reads
Γ Ω − x n ( τ x n X f ) = τ x n Z ( Γ Ω f ) . \Gamma_{\Omega-x_n}(\tau_{x_n}^Xf)=\tau_{x_n}^Z(\Gamma_\Omega f). Γ Ω − x n ( τ x n X f ) = τ x n Z ( Γ Ω f ) .
(H2) Continuity of the observed energy in each cell. Whenever ω n → ω \omega_n\to\omega ω n → ω , for every f ∈ X f\in X f ∈ X of bounded spatial support and every k ∈ Z d k\in\mathbb{Z}^d k ∈ Z d ,
∥ 1 k Γ ω n f ∥ Z 2 ⟶ ∥ 1 k Γ ω f ∥ Z 2 . (13) \|\mathbf{1}_k\Gamma_{\omega_n}f\|_Z^2\longrightarrow\|\mathbf{1}_k\Gamma_\omega f\|_Z^2. \tag*{(13)} ∥ 1 k Γ ω n f ∥ Z 2 ⟶ ∥ 1 k Γ ω f ∥ Z 2 . (13)
In particular, strong convergence of the observations for each such f f f is sufficient. For translated observation sets, the condition concerns the energy in the fixed window ( 0 , T ) × Q k (0,T)\times Q_k ( 0 , T ) × Q k .
(H3) Summable localization. There is a common profile κ : Z d → [ 0 , ∞ ) \kappa:\mathbb{Z}^d\to[0,\infty) κ : Z d → [ 0 , ∞ ) such that
∥ 1 k Γ ω 1 j ∥ ≤ κ ( k − j ) , ∑ n ∈ Z d κ ( n ) < ∞ ( j , k ∈ Z d , ω ∈ W ) . (14) \|\mathbf{1}_k\Gamma_\omega\mathbf{1}_j\|\le\kappa(k-j),\qquad\sum_{n\in\mathbb{Z}^d}\kappa(n)<\infty\qquad(j,k\in\mathbb{Z}^d,\ \omega\in\mathcal{W}). \tag*{(14)} ∥ 1 k Γ ω 1 j ∥ ≤ κ ( k − j ) , n ∈ Z d ∑ κ ( n ) < ∞ ( j , k ∈ Z d , ω ∈ W ) . (14)
The block 1 k Γ ω 1 j \mathbf{1}_k\Gamma_\omega\mathbf{1}_j 1 k Γ ω 1 j measures how an input in Q j Q_j Q j contributes to the observation in ( 0 , T ) × Q k (0,T)\times Q_k ( 0 , T ) × Q k .
(R) Uniform reconstruction. For each ω ∈ W \omega\in\mathcal{W} ω ∈ W , there is a bounded linear operator B ω : Z ∞ → X ∞ B_\omega:Z_\infty\to X_\infty B ω : Z ∞ → X ∞ such that
sup ω ∈ W ∥ B ω ∥ L ( Z ∞ , X ∞ ) < ∞ . \sup_{\omega\in\mathcal{W}}\|B_\omega\|_{\mathcal{L}(Z_\infty,X_\infty)}<\infty. ω ∈ W sup ∥ B ω ∥ L ( Z ∞ , X ∞ ) < ∞.
Set S ω = I − B ω Γ ω S_\omega=I-B_\omega\Gamma_\omega S ω = I − B ω Γ ω on X ∞ X_\infty X ∞ , using the uniformly local extension of Γ ω \Gamma_\omega Γ ω given by (H3). For every bounded cube Q ⊂ R d Q\subset\mathbb{R}^d Q ⊂ R d , the set
{ 1 Q S ω v : ω ∈ W , ∥ v ∥ X ∞ ≤ 1 } \{\mathbf{1}_Q S_\omega v:\omega\in\mathcal{W},\ \|v\|_{X_\infty}\le1\} { 1 Q S ω v : ω ∈ W , ∥ v ∥ X ∞ ≤ 1 }
is relatively compact in L 2 ( Q ; C N ) . L^2(Q;\mathbb{C}^N). L 2 ( Q ; C N ) . The role of (R) is to prevent concentration or oscillation defects within bounded spatial windows. It ensures that sequences with vanishing observations retain strong local compactness after reconstruction.
A sufficient condition for reconstruction. Theorem 2.1 uses (H1)–(H3) and (R). The following two Gramian assumptions provide a way to verify (R) in the transport applications.
(H4) A coercive multiplication part:
H ω : = Γ ω ∗ Γ ω = D ω + K ω , ( D ω f ) ( x ) = d ω ( x ) f ( x ) , (15) H_\omega:= \Gamma_\omega^*\Gamma_\omega= D_\omega+ K_\omega,\qquad(D_\omega f)(x)=d_\omega(x)f(x), \tag*{(15)} H ω := Γ ω ∗ Γ ω = D ω + K ω , ( D ω f ) ( x ) = d ω ( x ) f ( x ) , (15)
where d ω d_\omega d ω is measurable and Hermitian, with
0 < d ∙ I ≤ d ω ( x ) ≤ d ∙ I a.e., uniformly in ω . (16) 0<d_\bullet I\le d_\omega(x)\le d^\bullet I\quad\text{a.e., uniformly in }\omega. \tag*{(16)} 0 < d ∙ I ≤ d ω ( x ) ≤ d ∙ I a.e., uniformly in ω . (16)
Moreover, d ω n → d ω d_{\omega_n}\to d_\omega d ω n → d ω locally in measure whenever ω n → ω \omega_n\to\omega ω n → ω . The Gramian encodes the observed energy through ⟨ H ω f , f ⟩ X = ∥ Γ ω f ∥ Z 2 \langle H_\omega f,f\rangle_X=\|\Gamma_\omega f\|_Z^2 ⟨ H ω f , f ⟩ X = ∥ Γ ω f ∥ Z 2 . Its multiplication part acts pointwise and satisfies
⟨ D ω f , f ⟩ X ≥ d ∙ ∥ f ∥ X 2 . \langle D_\omega f,f\rangle_X\ge d_\bullet\|f\|_X^2. ⟨ D ω f , f ⟩ X ≥ d ∙ ∥ f ∥ X 2 .
The remainder K ω K_\omega K ω contains the other interactions and is subject to the compactness requirement in (H5).
(H5) Local collective compactness: for all bounded measurable I , J ⊂ R d I,J\subset\mathbb{R}^d I , J ⊂ R d , the set
⋃ ω ∈ W { 1 I K ω 1 J f : ∥ f ∥ L 2 ( J ) ≤ 1 } (17) \bigcup_{\omega\in\mathcal{W}}\left\{\mathbf{1}_I K_\omega\mathbf{1}_J f:\|f\|_{L^2(J)}\le1\right\} \tag*{(17)} ω ∈ W ⋃ { 1 I K ω 1 J f : ∥ f ∥ L 2 ( J ) ≤ 1 } (17)
is relatively compact in L 2 ( I ; C N ) L^2(I;\mathbb{C}^N) L 2 ( I ; C N ) .
Equivalently, for any sequence of environments ω n \omega_n ω n and inputs f n f_n f n in the unit ball of L 2 ( J ; C N ) L^2(J;\mathbb{C}^N) L 2 ( J ; C N ) , the localized outputs 1 I K ω n 1 J f n \mathbf{1}_I K_{\omega_n}\mathbf{1}_J f_n 1 I K ω n 1 J f n admit a strongly convergent subsequence in L 2 ( I ; C N ) L^2(I;\mathbb{C}^N) L 2 ( I ; C N ) . The compactness thus holds simultaneously over the whole family of environments.
Both (R) and (H5) are collective compactness conditions. A family { G ω } ω ∈ W \{G_\omega\}_{\omega\in\mathcal{W}} { G ω } ω ∈ W is collectively compact on bounded windows for a set V \mathcal{V} V of inputs if, for every bounded cube I I I , the set ⋃ ω ∈ W { 1 I G ω v : v ∈ V } \bigcup_{\omega\in\mathcal{W}}\{\mathbf{1}_I G_\omega v:v\in\mathcal{V}\} ⋃ ω ∈ W { 1 I G ω v : v ∈ V } is relatively compact. This is stronger than compactness of each localized operator 1 I G ω 1 J \mathbf{1}_I G_\omega\mathbf{1}_J 1 I G ω 1 J , and the union over ω \omega ω is what Lemma 2.16 uses, since the environment varies along the extracted sequence. Individual compactness does not suffice: in the scalar case, for e n ( x ) = e i n x 1 1 Q 0 ( x ) e_n(x)=e^{inx_1}\mathbf{1}_{Q_0}(x) e n ( x ) = e in x 1 1 Q 0 ( x ) and G n f = ⟨ f , e n ⟩ e n G_n f=\langle f,e_n\rangle e_n G n f = ⟨ f , e n ⟩ e n , each G n G_n G n has rank one, yet the sequence G n e n = e n G_ne_n=e_n G n e n = e n has no convergent subsequence in L 2 ( Q 0 ) L^2(Q_0) L 2 ( Q 0 ) . Lemma 2.14 collects the closure properties used below.
Under (H3)–(H5), Lemma 2.15 gives (R) with B ω = D ω − 1 Γ ω ∗ B_\omega=D_\omega^{-1}\Gamma_\omega^* B ω = D ω − 1 Γ ω ∗ .
A summable profile with infinite first moment. Take a singleton hull in dimension one. Write U f ( x ) = f ( x + 1 ) Uf(x)=f(x+1) U f ( x ) = f ( x + 1 ) and set
A f = f + ∑ n ≥ 1 a n U n f , a n = 1 2 n ( n + 1 ) , ( Γ f ) ( t , x ) = A f ( x ) . Af=f+\sum_{n\ge1}a_nU^nf,\qquad a_n=\frac{1}{2n(n+1)},\qquad(\Gamma f)(t,x)=Af(x). A f = f + n ≥ 1 ∑ a n U n f , a n = 2 n ( n + 1 ) 1 , ( Γ f ) ( t , x ) = A f ( x ) .
Integer translation preserves the uniformly local norm and ∑ n ≥ 1 a n = 1 / 2 \sum_{n\ge1}a_n=1/2 ∑ n ≥ 1 a n = 1/2 , so the Neumann series gives a bounded inverse of A A A on both X X X and X ∞ X_\infty X ∞ . Thus B z = A − 1 T − 1 ∫ 0 T z ( t , ⋅ ) d t Bz=A^{-1}T^{-1}\int_0^T z(t,\cdot)\,dt B z = A − 1 T − 1 ∫ 0 T z ( t , ⋅ ) d t gives (R) with S = 0 S=0 S = 0 . The family satisfies (H1)–(H3). The block of Γ \Gamma Γ from Q n Q_n Q n to ( 0 , T ) × Q 0 (0,T)\times Q_0 ( 0 , T ) × Q 0 has norm T a n \sqrt{T}\,a_n T a n for n ≥ 1 n\ge1 n ≥ 1 . Hence every admissible localization profile has infinite first moment.
This example also separates (R) from (H4)–(H5). With a 0 = 1 a_0=1 a 0 = 1 , the block of Γ ∗ Γ \Gamma^*\Gamma Γ ∗ Γ from Q 1 Q_1 Q 1 to Q 0 Q_0 Q 0 is T ( ∑ n ≥ 0 a n a n + 1 ) U T\left(\sum_{n\ge0}a_na_{n+1}\right)U T ( ∑ n ≥ 0 a n a n + 1 ) U on L 2 ( Q 1 ) L^2(Q_1) L 2 ( Q 1 ) , a positive multiple of an isometry. A multiplication operator has zero block there, so the localized remainder cannot be compact.
A two-speed example for (H4)–(H5). Take a bounded Lipschitz weight a : R → [ a 0 , 1 ] a:\mathbb{R}\to[a_0,1] a : R → [ a 0 , 1 ] , with a 0 > 0 a_0>0 a 0 > 0 , and let b b b range over its translation hull for locally uniform convergence. In dimension one, with N = 2 N=2 N = 2 and m = 1 m=1 m = 1 , consider
( Γ b f ) ( s , x ) = b ( x ) ( f 1 ( x + s ) + f 2 ( x + 2 s ) ) . (\Gamma_bf)(s,x)=\sqrt{b(x)}\bigl(f_1(x+s)+f_2(x+2s)\bigr). ( Γ b f ) ( s , x ) = b ( x ) ( f 1 ( x + s ) + f 2 ( x + 2 s ) ) .
A change of variables gives Γ b ∗ Γ b = D b + K b \Gamma_b^*\Gamma_b=D_b+K_b Γ b ∗ Γ b = D b + K b , where
d b ( z ) = diag ( ∫ 0 T b ( z − s ) d s , ∫ 0 T b ( z − 2 s ) d s ) , T a 0 I ≤ d b ( z ) ≤ T I . d_b(z)=\operatorname{diag}\left(\int_0^T b(z-s)\,ds,\int_0^T b(z-2s)\,ds\right),\qquad Ta_0I\le d_b(z)\le TI. d b ( z ) = diag ( ∫ 0 T b ( z − s ) d s , ∫ 0 T b ( z − 2 s ) d s ) , T a 0 I ≤ d b ( z ) ≤ T I .
This is the coercive multiplication part in (H4); its continuity follows directly from locally uniform convergence of b b b . The remainder has zero diagonal blocks and off-diagonal kernels
k 12 , b ( z , y ) = 1 0 < y − z < T b ( 2 z − y ) , k 21 , b ( z , y ) = k 12 , b ( y , z ) ‾ . k_{12,b}(z,y)=\mathbf{1}_{0<y-z<T}b(2z-y),\qquad k_{21,b}(z,y)=\overline{k_{12,b}(y,z)}. k 12 , b ( z , y ) = 1 0 < y − z < T b ( 2 z − y ) , k 21 , b ( z , y ) = k 12 , b ( y , z ) .
On bounded input and output windows, these kernels are square integrable, so the localized remainders are Hilbert–Schmidt operators. Moreover, they depend continuously on b b b in Hilbert–Schmidt norm. The weight hull is compact by Arzelà–Ascoli, so these localized remainders form a norm-compact family of compact operators, which gives the collective compactness in (H5). The bounded profile w = ( 1 , − 1 ) T w=(1,-1)^{\mathsf T} w = ( 1 , − 1 ) T nevertheless satisfies Γ b w = 0 \Gamma_bw=0 Γ b w = 0 for every b b b : uniqueness for the full observation remains a separate condition.
We use the notation
∥ k ∥ 1 : = ∑ n ∈ Z d ∣ k ( n ) ∣ , τ k ( L ) : = ∑ ∣ n ∣ > L k ( n ) ( L ∈ R ) . \|k\|_1:=\sum_{n\in\mathbb{Z}^d}|k(n)|,\qquad\tau_k(L):=\sum_{|n|>L}k(n)\quad(L\in\mathbb{R}). ∥ k ∥ 1 := n ∈ Z d ∑ ∣ k ( n ) ∣ , τ k ( L ) := ∣ n ∣ > L ∑ k ( n ) ( L ∈ R ) .
Then τ κ ( L ) ≤ ∥ κ ∥ 1 \tau_{\kappa}(L) \le\|\kappa\|_{1} τ κ ( L ) ≤ ∥ κ ∥ 1 and τ κ ( L ) → 0 \tau_{\kappa}(L) \to0 τ κ ( L ) → 0 as L → ∞ L \to\infty L → ∞ . The block estimate in (H3) gives
∥ 1 k Γ ω f ∥ L 2 ≤ ∑ j ∈ Z d κ ( k − j ) ∥ 1 j f ∥ L 2 . \|\mathbf{1}_{k}\Gamma_{\omega}f\|_{L^{2}} \le\sum_{j\in\mathbb{Z}^{d}}\kappa(k-j)\|\mathbf{1}_{j}f\|_{L^{2}}. ∥ 1 k Γ ω f ∥ L 2 ≤ j ∈ Z d ∑ κ ( k − j ) ∥ 1 j f ∥ L 2 .
In particular, ∥ Γ ω ∥ L ( X , Z ) ≤ ∥ κ ∥ 1 \|\Gamma_{\omega}\|_{\mathcal{L}(X,Z)} \le\|\kappa\|_{1} ∥ Γ ω ∥ L ( X , Z ) ≤ ∥ κ ∥ 1 , and Γ ω \Gamma_{\omega} Γ ω has a canonical extension to uniformly local inputs (Definition 2.9 and Lemma 2.10 ); kernels on L u l o c 2 L^{2}_{\mathrm{uloc}} L uloc 2 are understood in this sense. Hypotheses (H3)–(H4) also make distant inputs uniformly negligible in a fixed observation window (Lemma 2.11 ). Summability of κ \kappa κ suffices for these extension and tail estimates and for the full-hull criterion.
The full-hull criterion
Fix ω 0 ∈ W \omega_{0}\in\mathcal{W} ω 0 ∈ W and assume, for the next theorem, that
W = { θ h ω 0 : h ∈ R d } ‾ . (18) \mathcal{W}=\overline{\{\theta_{h}\omega_{0}:h\in\mathbb{R}^{d}\}}. \tag*{(18)} W = { θ h ω 0 : h ∈ R d } . (18)
The closure is taken in the chosen topology of environments. This is the full translation hull: it contains the reference configuration, every translated configuration, and every limit of such configurations.
For the one-dimensional observation sets, it becomes
ω 0 = 1 Ω , W = { 1 Ω − h : h ∈ R } L l o c 1 ( R ) . \omega_{0}=\mathbf{1}_{\Omega},\qquad\mathcal{W}=\{\mathbf{1}_{\Omega-h}:h\in\mathbb{R}\}^{L^{1}_{\mathrm{loc}}(\mathbb{R})}. ω 0 = 1 Ω , W = { 1 Ω − h : h ∈ R } L loc 1 ( R ) .
Thus an environment in this hull is represented by a local L 1 L^{1} L 1 limit of indicators 1 Ω − x n \mathbf{1}_{\Omega-x_{n}} 1 Ω − x n . The parameters x n x_{n} x n may remain bounded or be unbounded. The operator family and its uniform assumptions therefore cover the original observation geometry, its translates, and all the limit geometries arising from spatial recentering.
Theorem 2.1. Assume (H1)–(H3) and (R), and (2.12). The following are equivalent:
There is c > 0 c>0 c > 0 such that ∥ Γ ω 0 f ∥ Z ≥ c ∥ f ∥ X \|\Gamma_{\omega_{0}}f\|_{Z}\ge c\|f\|_{X} ∥ Γ ω 0 f ∥ Z ≥ c ∥ f ∥ X for all f ∈ X f\in X f ∈ X .
For every ω ∈ W \omega\in\mathcal{W} ω ∈ W , ker L u l o c 2 Γ ω = { 0 } \ker_{L^{2}_{\mathrm{uloc}}}\Gamma_{\omega}=\{0\} ker L uloc 2 Γ ω = { 0 } .
If the first assertion fails, there exist ω ∞ ∈ W \omega_{\infty}\in\mathcal{W} ω ∞ ∈ W and w ∞ ∈ L u l o c 2 w_{\infty}\in L^{2}_{\mathrm{uloc}} w ∞ ∈ L uloc 2 such that
Γ ω ∞ w ∞ = 0 , ∥ w ∞ ∥ L 2 ( Q 0 ) = ∥ w ∞ ∥ L u l o c 2 = 1. \Gamma_{\omega_{\infty}}w_{\infty}=0,\qquad\|w_{\infty}\|_{L^{2}(Q_{0})}=\|w_{\infty}\|_{L^{2}_{\mathrm{uloc}}}=1. Γ ω ∞ w ∞ = 0 , ∥ w ∞ ∥ L 2 ( Q 0 ) = ∥ w ∞ ∥ L uloc 2 = 1. If it holds, there is a common c ∗ > 0 c_{*}>0 c ∗ > 0 such that
∥ Γ ω w ∥ ( p ) ≥ c ∗ ∥ w ∥ ( p ) ( 1 ≤ p ≤ ∞ , ω ∈ W , w ∈ X p ) . \|\Gamma_{\omega}w\|_{(p)}\ge c_{*}\|w\|_{(p)}\qquad(1\le p\le\infty,\ \omega\in\mathcal{W},\ w\in X_{p}). ∥ Γ ω w ∥ ( p ) ≥ c ∗ ∥ w ∥ ( p ) ( 1 ≤ p ≤ ∞ , ω ∈ W , w ∈ X p ) . In particular, the observation controls the uniformly local energy.
Corollary 2.2. Assume (H1)–(H3) and (R), without requiring (2.12). Then
inf ω ∈ W inf f ∈ X ∖ { 0 } ∥ Γ ω f ∥ Z ∥ f ∥ X > 0 \inf_{\omega\in\mathcal{W}}\inf_{f\in X\setminus\{0\}}\frac{\|\Gamma_{\omega}f\|_{Z}}{\|f\|_{X}}>0 ω ∈ W inf f ∈ X ∖ { 0 } inf ∥ f ∥ X ∥ Γ ω f ∥ Z > 0 if and only if ker L u l o c 2 Γ ω = { 0 } \ker_{L^{2}_{\mathrm{uloc}}}\Gamma_{\omega}=\{0\} ker L uloc 2 Γ ω = { 0 } for every ω ∈ W \omega\in\mathcal{W} ω ∈ W . In that case a single positive lower bound works for all ω ∈ W \omega\in\mathcal{W} ω ∈ W and all p ∈ [ 1 , ∞ ] p\in[1,\infty] p ∈ [ 1 , ∞ ] .
Proposition 2.3 (Observation versus Gramian uniqueness). Assume (H1)–(H3) and (R). For ω ∈ W \omega\in\mathcal{W} ω ∈ W let H ω = Γ ω ∗ Γ ω H_{\omega}=\Gamma_{\omega}^{*}\Gamma_{\omega} H ω = Γ ω ∗ Γ ω , extended to L u l o c 2 L^{2}_{\mathrm{uloc}} L uloc 2 by Lemma 2.10 (iii). Then ker L u l o c 2 Γ ω ⊂ ker L u l o c 2 H ω \ker_{L^{2}_{\mathrm{uloc}}}\Gamma_{\omega}\subset\ker_{L^{2}_{\mathrm{uloc}}}H_{\omega} ker L uloc 2 Γ ω ⊂ ker L uloc 2 H ω for every ω ∈ W \omega\in\mathcal{W} ω ∈ W , and the following are equivalent:
(a) ker L u l o c 2 Γ ω = { 0 } \ker_{L^{2}_{\mathrm{uloc}}}\Gamma_{\omega}=\{0\} ker L uloc 2 Γ ω = { 0 } for every ω ∈ W \omega\in\mathcal{W} ω ∈ W ;
(b) ker L u l o c 2 H ω = { 0 } \ker_{L^{2}_{\mathrm{uloc}}}H_{\omega}=\{0\} ker L uloc 2 H ω = { 0 } for every ω ∈ W \omega\in\mathcal{W} ω ∈ W .
Remark 2.4. The implication (a)⇒ \Rightarrow ⇒ (b) uses Corollary 2.2, hence the profile extraction of Lemma 2.16 . It is not obtained from the global energy identity, which is unavailable for w ∉ L 2 w\notin L^{2} w ∈ / L 2 . In applications, (a) follows directly from uniqueness for the evolution equation; see Lemmas 4.1 and 5.8, and Step 3 of the proof of Theorem 7.1. The Gramian condition (b) follows afterwards. The equivalence concerns the conditions on the entire family; no strict inclusion of the two kernels for an individual environment is asserted. Proposition 2.17 below shows that observability also implies (R). Corollary 2.18 combines this necessity with Theorem 2.1 to give a characterization under (H1)–(H3) on the full hull.
The localization tools and the proofs are given in § 2.4 and § 2.5 .
Remark 2.5 (Why several exponents). For 1 ≤ p ≤ q ≤ ∞ 1\le p\le q\le\infty 1 ≤ p ≤ q ≤ ∞ and f f f supported in n n n cells,
∥ f ∥ ( q ) ≤ ∥ f ∥ ( p ) ≤ n 1 / p − 1 / q ∥ f ∥ ( q ) . (19) \|f\|_{(q)}\le\|f\|_{(p)}\le n^{1/p-1/q}\|f\|_{(q)}. \tag*{(19)} ∥ f ∥ ( q ) ≤ ∥ f ∥ ( p ) ≤ n 1/ p − 1/ q ∥ f ∥ ( q ) . (19)
A failure of the L 2 L^{2} L 2 lower bound produces inputs with small L 2 L^{2} L 2 observation. Such inputs may be spread over many cells, so their largest cell norm may be much smaller than their L 2 L^{2} L 2 norm, and a direct norm comparison does not reach the uniformly local norm. Proposition 2.12 instead applies the lower bound in one exponent to cutoffs on blocks of R d \mathbb{R}^{d} R d cells. The commutator error is controlled by the summable profile alone, and the two exponents are compared only inside each block, at the fixed cost R d ∣ 1 / p − 1 / q ∣ R^{d|1/p-1/q|} R d ∣1/ p − 1/ q ∣ .
A concrete sufficient operator class
Proposition 2.6. Let W W W be the compact metrizable parameter space with continuous R d \mathbb{R}^{d} R d -action introduced above. Let V 1 , … , V p ∈ R d V_{1},\ldots,V_{p}\in\mathbb{R}^{d} V 1 , … , V p ∈ R d be distinct, and suppose that, for ω ∈ W \omega\in W ω ∈ W ,
( Γ ω f ) ( t , x ) = ∑ ℓ = 1 p c ℓ , ω ( t , x ) f ( x + t V ℓ ) + ∫ R d k ω ( t , x , y ) f ( y ) d y . (20) (\Gamma_{\omega}f)(t,x)=\sum_{\ell=1}^{p}c_{\ell,\omega}(t,x)f(x+tV_{\ell})+\int_{\mathbb{R}^{d}}k_{\omega}(t,x,y)f(y)\,dy. \tag*{(20)} ( Γ ω f ) ( t , x ) = ℓ = 1 ∑ p c ℓ , ω ( t , x ) f ( x + t V ℓ ) + ∫ R d k ω ( t , x , y ) f ( y ) d y . (20) with measurable coefficients c ℓ , ω : ( 0 , T ) × R d → C m × N c_{\ell,\omega}:(0,T)\times\mathbb{R}^{d}\to\mathbb{C}^{m\times N} c ℓ , ω : ( 0 , T ) × R d → C m × N and kernels k ω : ( 0 , T ) × R d × R d → C m × N k_{\omega}:(0,T)\times\mathbb{R}^{d}\times\mathbb{R}^{d}\to\mathbb{C}^{m\times N} k ω : ( 0 , T ) × R d × R d → C m × N . Assume that sup ℓ , ω ∥ c ℓ , ω ∥ L ∞ < ∞ \sup_{\ell,\omega}\|c_{\ell,\omega}\|_{L^{\infty}}<\infty sup ℓ , ω ∥ c ℓ , ω ∥ L ∞ < ∞ and that
∥ k ω ( t , x , y ) ∥ ≤ C e − α ∣ x − y ∣ ( α > 0 ) , (21) \|k_{\omega}(t,x,y)\|\le Ce^{-\alpha|x-y|}\qquad(\alpha>0), \tag*{(21)} ∥ k ω ( t , x , y ) ∥ ≤ C e − α ∣ x − y ∣ ( α > 0 ) , (21) uniformly in ω \omega ω . Assume the covariance relations
c ℓ , θ h ω ( t , x ) = c ℓ , ω ( t , x + h ) , k θ h ω ( t , x , y ) = k ω ( t , x + h , y + h ) , (22) c_{\ell,\theta_{h}\omega}(t,x)=c_{\ell,\omega}(t,x+h),\qquad k_{\theta_{h}\omega}(t,x,y)=k_{\omega}(t,x+h,y+h), \tag*{(22)} c ℓ , θ h ω ( t , x ) = c ℓ , ω ( t , x + h ) , k θ h ω ( t , x , y ) = k ω ( t , x + h , y + h ) , (22) for all h ∈ R d h\in\mathbb{R}^{d} h ∈ R d , ω ∈ W \omega\in W ω ∈ W , 1 ≤ ℓ ≤ p 1\le\ell\le p 1 ≤ ℓ ≤ p and a.e. ( t , x , y ) (t,x,y) ( t , x , y ) . When ω n → ω \omega_{n}\to\omega ω n → ω , assume local convergence in measure of the coefficients on ( 0 , T ) × R d (0,T)\times\mathbb{R}^{d} ( 0 , T ) × R d , and local L 2 L^{2} L 2 convergence of the kernels on ( 0 , T ) × R d × R d (0,T)\times\mathbb{R}^{d}\times\mathbb{R}^{d} ( 0 , T ) × R d × R d . Put
d ω ( z ) = ∑ ℓ = 1 p ∫ 0 T c ℓ , ω ( t , z − t V ℓ ) ∗ c ℓ , ω ( t , z − t V ℓ ) d t , (23) d_{\omega}(z)=\sum_{\ell=1}^{p}\int_{0}^{T}c_{\ell,\omega}(t,z-tV_{\ell})^{*}c_{\ell,\omega}(t,z-tV_{\ell})\,dt, \tag*{(23)} d ω ( z ) = ℓ = 1 ∑ p ∫ 0 T c ℓ , ω ( t , z − t V ℓ ) ∗ c ℓ , ω ( t , z − t V ℓ ) d t , (23) and assume d ω ≥ d ∙ I > 0 d_{\omega}\ge d_{\bullet}I>0 d ω ≥ d ∙ I > 0 uniformly. Write T ℓ , ω f ( t , x ) = c ℓ , ω ( t , x ) f ( x + t V ℓ ) T_{\ell,\omega}f(t,x)=c_{\ell,\omega}(t,x)f(x+tV_{\ell}) T ℓ , ω f ( t , x ) = c ℓ , ω ( t , x ) f ( x + t V ℓ ) , and suppose that the total cross-velocity family
R ω = ∑ ℓ ≠ r T ℓ , ω ∗ T r , ω R_{\omega}=\sum_{\ell\ne r}T_{\ell,\omega}^{*}T_{r,\omega} R ω = ℓ = r ∑ T ℓ , ω ∗ T r , ω is locally collectively compact in the sense of (H5). Then (H1)–(H5) hold. The last assumption is automatic if d = 1 d=1 d = 1 , and it holds in any dimension if c ℓ , ω ∗ c r , ω = 0 c_{\ell,\omega}^{*}c_{r,\omega}=0 c ℓ , ω ∗ c r , ω = 0 a.e. for ℓ ≠ r \ell\ne r ℓ = r , in particular if p = 1 p=1 p = 1 . A dense orbit is needed only to invoke Theorem 2.1 , not Corollary 2.2.
Proof. Let ρ = T max ℓ ∣ V ℓ ∣ \rho=T\max_{\ell}|V_{\ell}| ρ = T max ℓ ∣ V ℓ ∣ and c ∙ = sup ℓ , ω ∥ c ℓ , ω ∥ ∞ c^{\bullet}=\sup_{\ell,\omega}\|c_{\ell,\omega}\|_{\infty} c ∙ = sup ℓ , ω ∥ c ℓ , ω ∥ ∞ . The block from Q j Q_{j} Q j to ( 0 , T ) × Q k (0,T)\times Q_{k} ( 0 , T ) × Q k of a transport term vanishes unless ∣ k − j ∣ ≤ ρ + 1 |k-j|\le\rho+1 ∣ k − j ∣ ≤ ρ + 1 , and its norm is always at most T 1 / 2 c ∙ T^{1/2}c^{\bullet} T 1/2 c ∙ . The integral term has block norm at most T 1 / 2 C e α e − α ∣ k − j ∣ T^{1/2}Ce^{\alpha}e^{-\alpha|k-j|} T 1/2 C e α e − α ∣ k − j ∣ , by Cauchy–Schwarz on the input unit cell. Thus (H3) holds with
κ ( m ) = T 1 / 2 ( p c ∙ 1 { ∣ m ∣ ≤ ρ + 1 } + C e α e − α ∣ m ∣ ) . \kappa(m)=T^{1/2}\left(pc^{\bullet}\mathbf{1}_{\{|m|\le\rho+1\}}+Ce^{\alpha}e^{-\alpha|m|}\right). κ ( m ) = T 1/2 ( p c ∙ 1 { ∣ m ∣ ≤ ρ + 1 } + C e α e − α ∣ m ∣ ) .
On bounded input and output windows, kernel convergence implies operator-norm convergence of the integral terms. Exponential tails control the remaining output. For each transport term, bounded convergence in measure against the integrable weight ∣ f ( x + t V ℓ ) ∣ 2 |f(x+tV_{\ell})|^{2} ∣ f ( x + t V ℓ ) ∣ 2 gives strong convergence of the observations on bounded-support inputs and hence (H2). The same argument in (23) gives local L 1 L^{1} L 1 convergence of d ω d_{\omega} d ω ; boundedness and the assumed lower bound prove (H4).
Let L ω L_{\omega} L ω be the integral term and T ω = ∑ ℓ T ℓ , ω T_{\omega}=\sum_{\ell}T_{\ell,\omega} T ω = ∑ ℓ T ℓ , ω . Then
K ω = R ω + T ω ∗ L ω + L ω ∗ T ω + L ω ∗ L ω . K_{\omega}=R_{\omega}+T_{\omega}^{*}L_{\omega}+L_{\omega}^{*}T_{\omega}+L_{\omega}^{*}L_{\omega}. K ω = R ω + T ω ∗ L ω + L ω ∗ T ω + L ω ∗ L ω .
The three terms involving L ω L_{\omega} L ω have exponentially bounded kernels, uniformly in ω \omega ω . For the last term, use
∫ R d e − α ∣ x − z ∣ e − α ∣ x − y ∣ d x ≤ e − α ∣ z − y ∣ / 2 ∫ R d e − α ∣ x − z ∣ / 2 d x = C d α − d e − α ∣ z − y ∣ / 2 . \int_{\mathbb{R}^{d}}e^{-\alpha|x-z|}e^{-\alpha|x-y|}\,dx\le e^{-\alpha|z-y|/2}\int_{\mathbb{R}^{d}}e^{-\alpha|x-z|/2}\,dx=C_{d}\alpha^{-d}e^{-\alpha|z-y|/2}. ∫ R d e − α ∣ x − z ∣ e − α ∣ x − y ∣ d x ≤ e − α ∣ z − y ∣/2 ∫ R d e − α ∣ x − z ∣/2 d x = C d α − d e − α ∣ z − y ∣/2 .
Their kernels depend continuously on ω \omega ω in local L 2 L^{2} L 2 : truncate intermediate spatial integrations, use the assumed convergences on bounded sets, then discard uniformly exponentially small tails. Thus, for every bounded I , J I,J I , J ,
∥ 1 I [ ( K ω n − R ω n ) − ( K ω − R ω ) ] 1 J ∥ H S ⟶ 0 ( ω n → ω ) . \|1_{I}[(K_{\omega_{n}}-R_{\omega_{n}})-(K_{\omega}-R_{\omega})]1_{J}\|_{\mathrm{HS}}\longrightarrow0\qquad(\omega_{n}\to\omega). ∥ 1 I [( K ω n − R ω n ) − ( K ω − R ω )] 1 J ∥ HS ⟶ 0 ( ω n → ω ) .
Each block is compact, and its continuous dependence on the compact parameter space makes the family compact in operator norm. Lemma 2.14(a) gives collective compactness. Together with the hypothesis on R ω R_{\omega} R ω , Lemma 2.14(b) proves (H5).
If d = 1 d=1 d = 1 and ℓ ≠ r \ell\ne r ℓ = r , the change of variables z = x + t V ℓ z=x+tV_{\ell} z = x + t V ℓ , z ′ = x + t V r z'=x+tV_{r} z ′ = x + t V r gives the cross-term kernel
1 ( 0 , T ) ( t ) ∣ V r − V ℓ ∣ c ℓ , ω ( t , z − t V ℓ ) ∗ c r , ω ( t , z − t V r ) , t = z ′ − z V r − V ℓ . \frac{\mathbf{1}_{(0,T)}(t)}{|V_{r}-V_{\ell}|}c_{\ell,\omega}(t,z-tV_{\ell})^{*}c_{r,\omega}(t,z-tV_{r}),\qquad t=\frac{z'-z}{V_{r}-V_{\ell}}. ∣ V r − V ℓ ∣ 1 ( 0 , T ) ( t ) c ℓ , ω ( t , z − t V ℓ ) ∗ c r , ω ( t , z − t V r ) , t = V r − V ℓ z ′ − z .
It is bounded, has finite range, and depends continuously on ω \omega ω in local L 2 L^{2} L 2 . By Lemma 2.14(a), this proves the asserted automatic compactness. Under channel orthogonality, each cross term vanishes in any dimension. Finally, for f ∈ X f\in\mathcal{X} f ∈ X , (22) and the change of variables y ↦ y + h y\mapsto y+h y ↦ y + h give
( Γ θ h ω τ h f ) ( t , x ) = ∑ ℓ = 1 p c ℓ , ω ( t , x + h ) f ( x + h + t V ℓ ) + ∫ R d k ω ( t , x + h , y + h ) f ( y + h ) d y = ( Γ ω f ) ( t , x + h ) = ( τ h Γ ω f ) ( t , x ) , \begin{aligned}
(\Gamma_{\theta_{h}\omega}\tau_{h}f)(t,x)=\sum_{\ell=1}^{p}c_{\ell,\omega}(t,x+h)f(x+h+tV_{\ell})+\int_{\mathbb{R}^{d}}k_{\omega}(t,x+h,y+h)f(y+h)\,dy \\
&=(\Gamma_{\omega}f)(t,x+h)=(\tau_{h}\Gamma_{\omega}f)(t,x),
\end{aligned} ( Γ θ h ω τ h f ) ( t , x ) = ℓ = 1 ∑ p c ℓ , ω ( t , x + h ) f ( x + h + t V ℓ ) + ∫ R d k ω ( t , x + h , y + h ) f ( y + h ) d y = ( Γ ω f ) ( t , x + h ) = ( τ h Γ ω f ) ( t , x ) ,
which is (H1). □ \square □
Remark 2.7. Injectivity on global L 2 L^2 L 2 alone is insufficient. For example, the scalar Fourier multiplier with symbol
h ( ξ ) = ∣ ξ − ξ ∗ ∣ 2 2 1 + ∣ ξ − ξ ∗ ∣ 2 2 h(\xi)=\frac{|\xi-\xi_*|_2^2}{1+|\xi-\xi_*|_2^2} h ( ξ ) = 1 + ∣ ξ − ξ ∗ ∣ 2 2 ∣ ξ − ξ ∗ ∣ 2 2
is injective on L 2 L^2 L 2 but has lower norm zero, whereas the plane wave e i ξ ∗ ⋅ x e^{i\xi_*\cdot x} e i ξ ∗ ⋅ x is a nonzero uniformly local null profile. Here ∣ ⋅ ∣ 2 |\cdot|_2 ∣ ⋅ ∣ 2 denotes the Euclidean norm. The criterion requires uniqueness for observation maps on every configuration of the hull. This is the form supplied by the transport uniqueness argument below. The relation with injectivity of Γ ω ∗ Γ ω \Gamma_\omega^*\Gamma_\omega Γ ω ∗ Γ ω on uniformly local states is described in Proposition 2.3 and Remark 2.4. The role of (R) can be seen even on a singleton hull. For Γ f ( t ) = e t f \Gamma f(t)=e^t f Γ f ( t ) = e t f , the Gaussian kernel satisfies (H3) and its Fourier multiplier e − ∣ ξ ∣ 2 2 e^{-|\xi|_2^2} e − ∣ ξ ∣ 2 2 is injective on tempered distributions, hence on L u l o c 2 L^2_{\mathrm{uloc}} L uloc 2 . Indeed one can divide by this multiplier on every compact frequency window. Yet the inputs e i n x 1 φ ( x ) e^{inx_1}\varphi(x) e in x 1 φ ( x ) , for a fixed φ ∈ C c ∞ ( Q 0 d ) \varphi\in C^\infty_c(Q_0^d) φ ∈ C c ∞ ( Q 0 d ) of L 2 L^2 L 2 norm one, have observation norm tending to zero by Plancherel’s theorem and dominated convergence. Assumption (R) excludes this loss of compactness within a cell. An explicit observability cost in terms of geometry requires a separate estimate.
Remark 2.8 (The obstruction for several velocities). For d ≥ 2 d\ge2 d ≥ 2 , a cross term has the form
( K f ) ( z ) = ∫ 0 T a ( t , z ) f ( z + t ( V r − V ℓ ) ) d t . (Kf)(z)=\int_0^T a(t,z)f\bigl(z+t(V_r-V_\ell)\bigr)\,\mathrm{d}t. ( K f ) ( z ) = ∫ 0 T a ( t , z ) f ( z + t ( V r − V ℓ ) ) d t .
If this individual operator is nonzero, choose a compactly supported ψ \psi ψ and a bounded output window on which K ψ ≠ 0 K\psi\ne0 K ψ = 0 , and take 0 ≠ η ⊥ ( V r − V ℓ ) 0\ne\eta\perp(V_r-V_\ell) 0 = η ⊥ ( V r − V ℓ ) . Then f n = e i n η ⋅ x ψ ⇀ 0 f_n=e^{in\eta\cdot x}\psi\rightharpoonup0 f n = e in η ⋅ x ψ ⇀ 0 , while K f n = e i n η ⋅ x K ψ Kf_n=e^{in\eta\cdot x}K\psi K f n = e in η ⋅ x K ψ . The localized output norm stays positive, so K K K is not locally compact.
Orthogonality is therefore a useful sufficient condition, but it is not necessary for the total remainder: different cross terms may cancel. For instance, for scalar input and two output components,
( Γ f ) ( t , x ) = ( f ( x ) + f ( x + t v ) f ( x + t v ) − f ( x + 2 t v ) ) , 0 ≠ v ∈ R d , (\Gamma f)(t,x)=
\begin{pmatrix}
f(x)+f(x+tv)\\
f(x+tv)-f(x+2tv)
\end{pmatrix},
\qquad0\ne v\in\mathbb{R}^d, ( Γ f ) ( t , x ) = ( f ( x ) + f ( x + t v ) f ( x + t v ) − f ( x + 2 t v ) ) , 0 = v ∈ R d ,
has Γ ∗ Γ = 4 T I \Gamma^*\Gamma=4TI Γ ∗ Γ = 4 T I , although its three channel vectors are not pairwise orthogonal. This explains why Proposition 2.6 allows collective compactness of the total cross-velocity family. For the general transport systems considered later, the one-dimensional change of variables is what verifies compactness; the theorem does not assert their multidimensional analogue.
Localization in amalgam spaces
The proof of Theorem 2.1 requires almost invisible inputs whose largest cell norm can be normalized to one. We first extend the operators to uniformly local inputs. We then transfer lower bounds between exponents under (H3) and obtain compactly supported inputs for profile extraction.
The uniformly local extension. At the endpoint p = ∞ p=\infty p = ∞ , compactly supported inputs are not dense. For instance, in the scalar case w ≡ 1 w\equiv1 w ≡ 1 belongs to X ∞ \mathcal{X}_\infty X ∞ , but P L = 1 [ − L , L ] d P_L=1_{[-L,L]^d} P L = 1 [ − L , L ] d ,
∥ w − P L w ∥ ∞ = 1 , P L w ⟶ w in L l o c 2 . \|w-P_Lw\|_{\infty}=1,\qquad P_Lw\longrightarrow w\quad\text{in }L^2_{\mathrm{loc}}. ∥ w − P L w ∥ ∞ = 1 , P L w ⟶ w in L loc 2 .
The extension below is therefore defined cellwise; localization will ensure that it agrees with local limits of truncations.
Definition 2.9 (Localized operators and their uniformly local extension). Let E , F ∈ { X , Z } \mathcal{E},\mathcal{F}\in\{\mathcal{X},\mathcal{Z}\} E , F ∈ { X , Z } and let κ : Z d → [ 0 , ∞ ) \kappa:\mathbb{Z}^d\to[0,\infty) κ : Z d → [ 0 , ∞ ) be summable. An operator G ∈ L ( E , F ) G\in\mathcal{L}(\mathcal{E},\mathcal{F}) G ∈ L ( E , F ) is κ \kappa κ -localized if
∥ 1 k G 1 j ∥ ≤ κ ( k − j ) ( j , k ∈ Z d ) . \|1_kG1_j\|\le\kappa(k-j)\qquad(j,k\in\mathbb{Z}^d). ∥ 1 k G 1 j ∥ ≤ κ ( k − j ) ( j , k ∈ Z d ) . For w ∈ E ∞ w\in\mathcal{E}_\infty w ∈ E ∞ , define G w Gw Gw by
1 k G w : = ∑ j ∈ Z d 1 k G ( 1 j w ) ( k ∈ Z d ) . (24) 1_kGw:=\sum_{j\in\mathbb{Z}^d}1_kG(1_jw)\qquad(k\in\mathbb{Z}^d). \tag*{(24)} 1 k Gw := j ∈ Z d ∑ 1 k G ( 1 j w ) ( k ∈ Z d ) . (24) This series converges absolutely in the L 2 L^2 L 2 space of the k k k -th output cell, since its terms are bounded by κ ( k − j ) ∥ w ∥ ( ∞ ) \kappa(k-j)\|w\|_{(\infty)} κ ( k − j ) ∥ w ∥ ( ∞ ) . Under (H3), it defines Γ ω w \Gamma_\omega w Γ ω w for every w ∈ L u l o c 2 w\in L^2_{\mathrm{uloc}} w ∈ L uloc 2 . Throughout, the notation
ker L u l o c 2 Γ ω : = { w ∈ L u l o c 2 : Γ ω w = 0 } \ker_{L^2_{\mathrm{uloc}}}\Gamma_\omega:=\{w\in L^2_{\mathrm{uloc}}:\Gamma_\omega w=0\} ker L uloc 2 Γ ω := { w ∈ L uloc 2 : Γ ω w = 0 } uses this extension. For these observation maps, Lemma 2.11 identifies it with the local limit of bounded-support truncations.
Lemma 2.10. Let G ∈ L ( E , F ) G\in\mathcal{L}(\mathcal{E},\mathcal{F}) G ∈ L ( E , F ) be κ \kappa κ -localized.
For every p ∈ [ 1 , ∞ ] p\in[1,\infty] p ∈ [ 1 , ∞ ] , (2.18) defines a bounded operator E p → F p \mathcal{E}_p\to\mathcal{F}_p E p → F p of norm at most ∥ κ ∥ 1 \|\kappa\|_1 ∥ κ ∥ 1 , agreeing with G G G on E p ∩ E \mathcal{E}_p\cap\mathcal{E} E p ∩ E . For p < ∞ p<\infty p < ∞ , it is the unique bounded extension from compactly supported inputs.
(ii) If w ∈ E ∞ w \in\mathcal{E}_{\infty} w ∈ E ∞ , L ∈ R L \in\mathbb{R} L ∈ R , and 1 j w = 0 \mathbf{1}_{j}w=0 1 j w = 0 whenever ∣ j − k ∣ ≤ L \lvert j-k\rvert\leq L ∣ j − k ∣ ≤ L , then
∥ 1 k G w ∥ ≤ τ κ ( L ) ∥ w ∥ ( ∞ ) . \lVert\mathbf{1}_{k}Gw\rVert\leq\tau_{\kappa}(L)\lVert w\rVert_{(\infty)}. ∥ 1 k Gw ∥ ≤ τ κ ( L ) ∥ w ∥ ( ∞ ) . (iii) If G 1 ∈ L ( E , F ) G_{1} \in\mathcal{L}(\mathcal{E},\mathcal{F}) G 1 ∈ L ( E , F ) is κ 1 \kappa_{1} κ 1 -localized and G 2 ∈ L ( F , E ′ ) G_{2} \in\mathcal{L}(\mathcal{F},\mathcal{E}^{\prime}) G 2 ∈ L ( F , E ′ ) is κ 2 \kappa_{2} κ 2 -localized, with E ′ ∈ { X , Z } \mathcal{E}^{\prime} \in\{\mathcal{X},\mathcal{Z}\} E ′ ∈ { X , Z } , then G 2 G 1 G_{2}G_{1} G 2 G 1 is ( κ 2 ∗ κ 1 ) (\kappa_{2} * \kappa_{1}) ( κ 2 ∗ κ 1 ) -localized and
( G 2 G 1 ) w = G 2 ( G 1 w ) ( w ∈ E ∞ ) . (G_{2}G_{1})w=G_{2}(G_{1}w) \qquad(w \in\mathcal{E}_{\infty}). ( G 2 G 1 ) w = G 2 ( G 1 w ) ( w ∈ E ∞ ) . Moreover, G ∗ G^{*} G ∗ is κ ~ \widetilde{\kappa} κ -localized, where κ ~ ( m ) = κ ( − m ) \widetilde{\kappa}(m)=\kappa(-m) κ ( m ) = κ ( − m ) .
(iv) Under (H3)–(H4), the operators Γ ω ∗ \Gamma_{\omega}^{*} Γ ω ∗ , H ω H_{\omega} H ω , D ω D_{\omega} D ω , and K ω K_{\omega} K ω have the respective localization profiles
κ ~ , κ ~ ∗ κ , d ∗ δ 0 , κ K : = κ ~ ∗ κ + d ∗ δ 0 , \widetilde{\kappa},\qquad\widetilde{\kappa} * \kappa,\qquad d^{*}\delta_{0},\qquad\kappa_{K}:=\widetilde{\kappa} * \kappa+d^{*}\delta_{0}, κ , κ ∗ κ , d ∗ δ 0 , κ K := κ ∗ κ + d ∗ δ 0 , uniformly in ω \omega ω . For every w ∈ L u l o c 2 w \in L_{\mathrm{uloc}}^{2} w ∈ L uloc 2 ,
H ω w = Γ ω ∗ ( Γ ω w ) = d ω w + K ω w . H_{\omega}w=\Gamma_{\omega}^{*}(\Gamma_{\omega}w)=d_{\omega}w+K_{\omega}w. H ω w = Γ ω ∗ ( Γ ω w ) = d ω w + K ω w . Proof. (i) The cellwise estimate
∥ 1 k G w ∥ ≤ ∑ j κ ( k − j ) ∥ 1 j w ∥ \lVert\mathbf{1}_{k}Gw\rVert\leq\sum_{j}\kappa(k-j)\lVert\mathbf{1}_{j}w\rVert ∥ 1 k Gw ∥ ≤ j ∑ κ ( k − j ) ∥ 1 j w ∥ and Young’s inequality on Z d \mathbb{Z}^{d} Z d give the bound for every p p p . If w ∈ E p ∩ E w \in E_{p}\cap\mathcal{E} w ∈ E p ∩ E , its cell decomposition converges in E \mathcal{E} E , so applying G G G and restricting to the k k k -th output cell gives (24) . For p < ∞ p<\infty p < ∞ , uniqueness follows from the density of compactly supported inputs in E p E_{p} E p .
(ii) Only indices with ∣ j − k ∣ > L \lvert j-k\rvert>L ∣ j − k ∣ > L contribute to (24) .
(iii) On L 2 L^{2} L 2 inputs, inserting the intermediate cell decomposition gives
∥ 1 k G 2 G 1 1 j ∥ ≤ ∑ l κ 2 ( k − l ) κ 1 ( l − j ) = ( κ 2 ∗ κ 1 ) ( k − j ) . \lVert\mathbf{1}_{k}G_{2}G_{1}\mathbf{1}_{j}\rVert\leq\sum_{l}\kappa_{2}(k-l)\kappa_{1}(l-j)=(\kappa_{2}*\kappa_{1})(k-j). ∥ 1 k G 2 G 1 1 j ∥ ≤ l ∑ κ 2 ( k − l ) κ 1 ( l − j ) = ( κ 2 ∗ κ 1 ) ( k − j ) . For w ∈ E ∞ w \in\mathcal{E}_{\infty} w ∈ E ∞ , the double series is absolutely convergent:
∑ l , j ∥ 1 k G 2 1 l G 1 1 j w ∥ ≤ ∥ κ 2 ∥ 1 ∥ κ 1 ∥ 1 ∥ w ∥ ( ∞ ) . \sum_{l,j}\lVert\mathbf{1}_{k}G_{2}\mathbf{1}_{l}G_{1}\mathbf{1}_{j}w\rVert\leq\lVert\kappa_{2}\rVert_{1}\lVert\kappa_{1}\rVert_{1}\lVert w\rVert_{(\infty)}. l , j ∑ ∥ 1 k G 2 1 l G 1 1 j w ∥ ≤ ∥ κ 2 ∥ 1 ∥ κ 1 ∥ 1 ∥ w ∥ ( ∞ ) . Summing first in j j j gives 1 k G 2 ( G 1 w ) \mathbf{1}_{k}G_{2}(G_{1}w) 1 k G 2 ( G 1 w ) , while summing first in l l l gives 1 k ( G 2 G 1 ) w \mathbf{1}_{k}(G_{2}G_{1})w 1 k ( G 2 G 1 ) w . The adjoint profile follows from ∥ 1 j G ∗ 1 k ∥ = ∥ 1 k G 1 j ∥ \lVert\mathbf{1}_{j}G^{*}\mathbf{1}_{k}\rVert=\lVert\mathbf{1}_{k}G\mathbf{1}_{j}\rVert ∥ 1 j G ∗ 1 k ∥ = ∥ 1 k G 1 j ∥ .
(iv) Apply (iii) to Γ ω ∗ Γ ω \Gamma_{\omega}^{*}\Gamma_{\omega} Γ ω ∗ Γ ω . Multiplication by d ω d_{\omega} d ω acts within each cell with norm at most d ∗ d^{*} d ∗ , and K ω = H ω − D ω K_{\omega}=H_{\omega}-D_{\omega} K ω = H ω − D ω . These observations give the stated profiles. Both Gramian identities hold cellwise and therefore also on L u l o c 2 L_{\mathrm{uloc}}^{2} L uloc 2 .
The next lemma expresses the spatial meaning of localization: inputs far from a fixed output window have a uniformly small effect there. It will allow local convergence arguments even when the inputs have infinite global energy.
Lemma 2.11. Assume (H3)–(H4) and put P L = 1 [ − L , L ] d P_{L}=\mathbf{1}_{[-L,L]^{d}} P L = 1 [ − L , L ] d . For every bounded measurable I ⊂ R d I\subset\mathbb{R}^{d} I ⊂ R d , there is a function ε I \varepsilon_{I} ε I , depending only on I I I , d d d , κ \kappa κ , d ∗ d^{*} d ∗ , with ε I ( L ) → 0 \varepsilon_{I}(L)\to0 ε I ( L ) → 0 as L → ∞ L\to\infty L → ∞ , such that
∥ 1 I Γ ω ( 1 − P L ) w ∥ Z + ∥ 1 I K ω ( 1 − P L ) w ∥ X ≤ ε I ( L ) ∥ w ∥ ( ∞ ) ( ω ∈ W , L ≥ 1 , w ∈ L u l o c 2 ) . (25) \lVert\mathbf{1}_{I}\Gamma_{\omega}(1-P_{L})w\rVert_{\mathcal{Z}}+\lVert\mathbf{1}_{I}K_{\omega}(1-P_{L})w\rVert_{\mathcal{X}}\leq\varepsilon_{I}(L)\lVert w\rVert_{(\infty)}\qquad(\omega\in\mathcal{W},\ L\geq1,\ w\in L_{\mathrm{uloc}}^{2}). \tag*{(25)} ∥ 1 I Γ ω ( 1 − P L ) w ∥ Z + ∥ 1 I K ω ( 1 − P L ) w ∥ X ≤ ε I ( L ) ∥ w ∥ ( ∞ ) ( ω ∈ W , L ≥ 1 , w ∈ L uloc 2 ) . (25) In particular, 1 I Γ ω ( P L w ) → 1 I Γ ω w \mathbf{1}_{I}\Gamma_{\omega}(P_{L}w)\to\mathbf{1}_{I}\Gamma_{\omega}w 1 I Γ ω ( P L w ) → 1 I Γ ω w and 1 I K ω ( P L w ) → 1 I K ω w \mathbf{1}_{I}K_{\omega}(P_{L}w)\to\mathbf{1}_{I}K_{\omega}w 1 I K ω ( P L w ) → 1 I K ω w in L 2 L^{2} L 2 . The estimate for Γ ω \Gamma_{\omega} Γ ω alone requires only the summable block bound in (14) .
Proof. By Lemma 2.10 (iv), the two operator families are localized with profiles κ \kappa κ and κ K \kappa_{K} κ K . Choose an integer R I ≥ 0 R_{I}\geq0 R I ≥ 0 such that I ⊂ ⋃ ∣ k ∣ ≤ R I Q k I\subset\bigcup_{\lvert k\rvert\leq R_{I}}Q_{k} I ⊂ ⋃ ∣ k ∣ ≤ R I Q k . If Q j Q_{j} Q j meets the complement of [ − L , L ] d [-L,L]^{d} [ − L , L ] d , then ∣ j ∣ > L − 1 \lvert j\rvert>L-1 ∣ j ∣ > L − 1 . Thus, for ∣ k ∣ ≤ R I \lvert k\rvert\leq R_{I} ∣ k ∣ ≤ R I , the input ( 1 − P L ) w (1-P_{L})w ( 1 − P L ) w vanishes in every cell with ∣ j − k ∣ ≤ L − R I − 1 \lvert j-k\rvert\leq L-R_{I}-1 ∣ j − k ∣ ≤ L − R I − 1 . For either operator G ω ∈ { Γ ω , K ω } G_{\omega}\in\{\Gamma_{\omega},K_{\omega}\} G ω ∈ { Γ ω , K ω } , with its corresponding profile ρ ∈ { κ , κ K } \rho\in\{\kappa,\kappa_{K}\} ρ ∈ { κ , κ K } , Lemma 2.10 (ii) gives
∥ 1 k G ω ( 1 − P L ) w ∥ ≤ τ ρ ( L − R I − 1 ) ∥ w ∥ ( ∞ ) . \lVert\mathbf{1}_{k}G_{\omega}(1-P_{L})w\rVert\leq\tau_{\rho}(L-R_{I}-1)\lVert w\rVert_{(\infty)}. ∥ 1 k G ω ( 1 − P L ) w ∥ ≤ τ ρ ( L − R I − 1 ) ∥ w ∥ ( ∞ ) . Summing the squares over the ( 2 R I + 1 ) d (2R_{I}+1)^{d} ( 2 R I + 1 ) d output cells proves (25) with
ε I ( L ) = ( 2 R I + 1 ) d / 2 ( τ κ ( L − R I − 1 ) + τ κ K ( L − R I − 1 ) ) . \varepsilon_{I}(L)=(2R_{I}+1)^{d/2}\bigl(\tau_{\kappa}(L-R_{I}-1)+\tau_{\kappa_{K}}(L-R_{I}-1)\bigr). ε I ( L ) = ( 2 R I + 1 ) d /2 ( τ κ ( L − R I − 1 ) + τ κ K ( L − R I − 1 ) ) . Both tails tend to zero by summability. The convergence of truncations follows by linearity, since P L w ∈ X P_{L}w\in\mathcal{X} P L w ∈ X .
Transfer between exponents. Let E , F ∈ { X , Z } \mathcal{E},\mathcal{F} \in\{\mathcal{X},\mathcal{Z}\} E , F ∈ { X , Z } . For a localized operator G ∈ L ( E , F ) G \in\mathcal{L}(\mathcal{E},\mathcal{F}) G ∈ L ( E , F ) put
ν p ( G ) : = inf f ∈ E p ∖ { 0 } ∥ G f ∥ ( p ) ∥ f ∥ ( p ) . \nu_p(G) := \inf_{f\in\mathcal{E}_p\setminus\{0\}} \frac{\|Gf\|_{(p)}}{\|f\|_{(p)}}. ν p ( G ) := f ∈ E p ∖ { 0 } inf ∥ f ∥ ( p ) ∥ G f ∥ ( p ) .
The following argument groups cells into larger blocks, as in the localization proof of [30 Theorem 2.1] for scalar matrices.
Proposition 2.12. Let G ⊂ L ( E , F ) \mathcal{G} \subset\mathcal{L}(\mathcal{E},\mathcal{F}) G ⊂ L ( E , F ) have a common summable localization profile κ \kappa κ . If ν p ( G ) ≥ a > 0 \nu_p(G) \ge a > 0 ν p ( G ) ≥ a > 0 for all G ∈ G G \in\mathcal{G} G ∈ G at one exponent p ∈ [ 1 , ∞ ] p \in[1,\infty] p ∈ [ 1 , ∞ ] , then
inf G ∈ G , 1 ≤ q ≤ ∞ ν q ( G ) > 0. \inf_{G\in\mathcal{G},\,1\le q\le\infty}\nu_q(G)>0. G ∈ G , 1 ≤ q ≤ ∞ inf ν q ( G ) > 0.
Proof. Fix G ∈ G G \in\mathcal{G} G ∈ G and q ∈ [ 1 , ∞ ] q \in[1,\infty] q ∈ [ 1 , ∞ ] , and write G k j = 1 k G 1 j G_{kj}=1_kG1_j G k j = 1 k G 1 j , so that ∥ G k j ∥ ≤ κ ( k − j ) \|G_{kj}\| \le\kappa(k-j) ∥ G k j ∥ ≤ κ ( k − j ) . Throughout, 1 / ∞ = 0 1/\infty=0 1/∞ = 0 and s + = max { s , 0 } s_+=\max\{s,0\} s + = max { s , 0 } .
Step 1: block norms. For an integer R ≥ 1 R\ge1 R ≥ 1 , the blocks D l = R l + { 0 , … , R − 1 } d D_l=Rl+\{0,\ldots,R-1\}^d D l = R l + { 0 , … , R − 1 } d , l ∈ Z d l\in\mathbb{Z}^d l ∈ Z d , partition Z d \mathbb{Z}^d Z d . For an input or output v v v put
F l ( v ) = ( ∥ 1 j v ∥ L 2 ) j ∈ D l , M ( v ) = ∥ ( F l ( v ) ) l ∥ ℓ q . F_l(v)=\bigl(\|1_jv\|_{L^2}\bigr)_{j\in D_l},\qquad M(v)=\|(F_l(v))_l\|_{\ell^q}. F l ( v ) = ( ∥ 1 j v ∥ L 2 ) j ∈ D l , M ( v ) = ∥ ( F l ( v ) ) l ∥ ℓ q .
Since # D l = R d \#D_l=R^d # D l = R d , Hölder’s inequality on each block gives
R − d ( 1 / q − 1 / p ) + ∥ v ∥ ( q ) ≤ M ( v ) ≤ R d ( 1 / p − 1 / q ) + ∥ v ∥ ( q ) . (26) R^{-d(1/q-1/p)_+}\|v\|_{(q)}\le M(v)\le R^{d(1/p-1/q)_+}\|v\|_{(q)}. \tag*{(26)} R − d ( 1/ q − 1/ p ) + ∥ v ∥ ( q ) ≤ M ( v ) ≤ R d ( 1/ p − 1/ q ) + ∥ v ∥ ( q ) . (26)
In particular M ( v ) < ∞ M(v)<\infty M ( v ) < ∞ and M ( G v ) < ∞ M(Gv)<\infty M ( G v ) < ∞ for v ∈ E q v\in\mathcal{E}_q v ∈ E q .
Step 2: cutoffs. Choose a Lipschitz function χ : R d → [ 0 , 1 ] \chi:\mathbb{R}^d\to[0,1] χ : R d → [ 0 , 1 ] , equal to one on [ 0 , 1 ] d [0,1]^d [ 0 , 1 ] d and supported in [ − 1 , 2 ] d [-1,2]^d [ − 1 , 2 ] d , and let χ l \chi_l χ l multiply the cell j j j , of an input or an output, by χ ( j / R − l ) \chi(j/R-l) χ ( j / R − l ) . Then χ l = 1 \chi_l=1 χ l = 1 on D l D_l D l , and χ l \chi_l χ l vanishes outside ⋃ b ∈ B D l + b \bigcup_{b\in B}D_{l+b} ⋃ b ∈ B D l + b , where B = { − 1 , 0 , 1 , 2 } d B=\{-1,0,1,2\}^d B = { − 1 , 0 , 1 , 2 } d . Hence
M ( v ) ≤ ∥ ( ∥ χ l v ∥ ( p ) ) l ∥ ℓ q ≤ 4 d M ( v ) . (27) M(v)\le\|(\|\chi_lv\|_{(p)})_l\|_{\ell^q}\le4^dM(v). \tag*{(27)} M ( v ) ≤ ∥ ( ∥ χ l v ∥ ( p ) ) l ∥ ℓ q ≤ 4 d M ( v ) . (27)
Each χ l v \chi_lv χ l v has bounded support, so χ l v ∈ E ∩ E p \chi_lv\in\mathcal{E}\cap\mathcal{E}_p χ l v ∈ E ∩ E p , and the extensions of G G G agree on it by Lemma 2.10 (i).
Step 3: commutators. For n ∈ Z d n\in\mathbb{Z}^d n ∈ Z d let G ( n ) G^{(n)} G ( n ) be the n n n th diagonal, 1 k G ( n ) v = G k , k − n 1 k − n v 1_kG^{(n)}v=G_{k,k-n}1_{k-n}v 1 k G ( n ) v = G k , k − n 1 k − n v . By Lemma 2.10 , G v = ∑ n G ( n ) v Gv=\sum_nG^{(n)}v G v = ∑ n G ( n ) v with absolute convergence in every cell, and
1 k [ G ( n ) , χ l ] v = ( χ ( k − n R − l ) − χ ( k R − l ) ) G k , k − n 1 k − n v . 1_k[G^{(n)},\chi_l]v=\left(\chi\left(\frac{k-n}{R}-l\right)-\chi\left(\frac{k}{R}-l\right)\right)G_{k,k-n}1_{k-n}v. 1 k [ G ( n ) , χ l ] v = ( χ ( R k − n − l ) − χ ( R k − l ) ) G k , k − n 1 k − n v .
The scalar factor is at most c χ min { 1 , ∣ n ∣ / R } c_\chi\min\{1,|n|/R\} c χ min { 1 , ∣ n ∣/ R } , with c χ = max { 1 , d Lip ( χ ) } c_\chi=\max\{1,\sqrt d\operatorname{Lip}(\chi)\} c χ = max { 1 , d Lip ( χ )} . It vanishes unless k k k or k − n k-n k − n lies in R ( l + [ − 1 , 2 ] d ) R(l+[-1,2]^d) R ( l + [ − 1 , 2 ] d ) . Equivalently, the input index j = k − n j=k-n j = k − n lies in ⋃ b ∈ B ′ D l + b \bigcup_{b\in B'}D_{l+b} ⋃ b ∈ B ′ D l + b or in ⋃ b ∈ B ′ D l − ⌊ n / R ⌋ + b \bigcup_{b\in B'}D_{l-\lfloor n/R\rfloor+b} ⋃ b ∈ B ′ D l − ⌊ n / R ⌋ + b , where B ′ = { − 2 , … , 2 } d B'=\{-2,\ldots,2\}^d B ′ = { − 2 , … , 2 } d and the floor is taken componentwise. Since k ↦ k − n k\mapsto k-n k ↦ k − n is a bijection,
∥ ( [ G ( n ) , χ l ] v ) l ∥ ( p ) ≤ c χ min { 1 , ∣ n ∣ R } κ ( n ) ∑ b ∈ B ′ ( F l + b ( v ) + F l − ⌊ n / R ⌋ + b ( v ) ) . \|([G^{(n)},\chi_l]v)_l\|_{(p)}
\le c_\chi\min\left\{1,\frac{|n|}{R}\right\}\kappa(n)
\sum_{b\in B'}\left(F_{l+b}(v)+F_{l-\lfloor n/R\rfloor+b}(v)\right). ∥ ([ G ( n ) , χ l ] v ) l ∥ ( p ) ≤ c χ min { 1 , R ∣ n ∣ } κ ( n ) b ∈ B ′ ∑ ( F l + b ( v ) + F l − ⌊ n / R ⌋ + b ( v ) ) .
For every fixed shift c c c , ∥ ( F l + c ( v ) ) l ∥ ℓ q = M ( v ) \|(F_{l+c}(v))_l\|_{\ell^q}=M(v) ∥ ( F l + c ( v ) ) l ∥ ℓ q = M ( v ) . Summing over n n n by Minkowski’s inequality gives
∥ ( ∥ [ G , χ l ] v ∥ ( p ) ) l ∥ ℓ q ≤ C χ η R M ( v ) , η R = ∑ n ∈ Z d κ ( n ) min { 1 , ∣ n ∣ R } , (28) \|(\|[G,\chi_l]v\|_{(p)})_l\|_{\ell^q}\le C_\chi\eta_RM(v),\qquad
\eta_R=\sum_{n\in\mathbb{Z}^d}\kappa(n)\min\left\{1,\frac{|n|}{R}\right\}, \tag*{(28)} ∥ ( ∥ [ G , χ l ] v ∥ ( p ) ) l ∥ ℓ q ≤ C χ η R M ( v ) , η R = n ∈ Z d ∑ κ ( n ) min { 1 , R ∣ n ∣ } , (28)
with C χ = 2 ⋅ 5 d c χ C_\chi=2\cdot5^dc_\chi C χ = 2 ⋅ 5 d c χ . Since κ ∈ ℓ 1 \kappa\in\ell^1 κ ∈ ℓ 1 , η R → 0 \eta_R\to0 η R → 0 by dominated convergence. Only the summability of κ \kappa κ is used here. All sums converge absolutely, also when p p p or q q q is infinite.
Step 4: conclusion. The lower bound at the exponent p p p , applied to χ l v \chi_lv χ l v , gives
a ∥ χ l v ∥ ( p ) ≤ ∥ G ( χ l v ) ∥ ( p ) ≤ ∥ χ l G v ∥ ( p ) + ∥ [ G , χ l ] v ∥ ( p ) . a\|\chi_lv\|_{(p)}\le\|G(\chi_lv)\|_{(p)}\le\|\chi_lGv\|_{(p)}+\|[G,\chi_l]v\|_{(p)}. a ∥ χ l v ∥ ( p ) ≤ ∥ G ( χ l v ) ∥ ( p ) ≤ ∥ χ l G v ∥ ( p ) + ∥ [ G , χ l ] v ∥ ( p ) .
Take the ℓ q \ell^q ℓ q norm in l l l , and use (27) for v v v and for G v Gv G v , together with (28) :
a M ( v ) ≤ 4 d M ( G v ) + C χ η R M ( v ) . aM(v)\le4^dM(Gv)+C_\chi\eta_RM(v). a M ( v ) ≤ 4 d M ( G v ) + C χ η R M ( v ) .
Fix an integer R R R , depending only on a a a , κ \kappa κ , d d d and χ \chi χ , such that C χ η R ≤ a / 2 C_\chi\eta_R\le a/2 C χ η R ≤ a /2 . Since M ( v ) < ∞ M(v)<\infty M ( v ) < ∞ , we get M ( v ) ≤ 2 ⋅ 4 d a − 1 M ( G v ) M(v)\le2\cdot4^da^{-1}M(Gv) M ( v ) ≤ 2 ⋅ 4 d a − 1 M ( G v ) . Applying (26) to v v v and to G v Gv G v then gives
∥ G v ∥ ( q ) ≥ a 2 ⋅ 4 d R − d ∣ 1 / p − 1 / q ∣ ∥ v ∥ ( q ) ( v ∈ E q ) . (29) \|Gv\|_{(q)}\ge\frac{a}{2\cdot4^d}R^{-d|1/p-1/q|}\|v\|_{(q)}
\qquad(v\in\mathcal{E}_q). \tag*{(29)} ∥ G v ∥ ( q ) ≥ 2 ⋅ 4 d a R − d ∣1/ p − 1/ q ∣ ∥ v ∥ ( q ) ( v ∈ E q ) . (29)
Since ∣ 1 / p − 1 / q ∣ ≤ 1 |1/p-1/q|\le1 ∣1/ p − 1/ q ∣ ≤ 1 and R R R depends neither on G G G nor on q q q , we obtain ν q ( G ) ≥ a / ( 2 ⋅ 4 d R d ) \nu_q(G)\ge a/(2\cdot4^dR^d) ν q ( G ) ≥ a / ( 2 ⋅ 4 d R d ) for all G ∈ G G\in\mathcal{G} G ∈ G and q ∈ [ 1 , ∞ ] q\in[1,\infty] q ∈ [ 1 , ∞ ] . □ \square □
Corollary 2.13. Let G ⊂ L ( E , F ) \mathcal{G}\subset\mathcal{L}(\mathcal{E},\mathcal{F}) G ⊂ L ( E , F ) have a common summable localization profile κ \kappa κ , and let p , q ∈ [ 1 , ∞ ] p,q\in[1,\infty] p , q ∈ [ 1 , ∞ ] . If inf G ∈ G ν p ( G ) = 0 \inf_{G\in\mathcal{G}}\nu_p(G)=0 inf G ∈ G ν p ( G ) = 0 , then for every ε > 0 \varepsilon>0 ε > 0 there are G ∈ G G\in\mathcal{G} G ∈ G and a nonzero compactly supported v ∈ E v\in\mathcal{E} v ∈ E such that
∥ G v ∥ ( q ) ≤ ε ∥ v ∥ ( q ) . \|Gv\|_{(q)}\le\varepsilon\|v\|_{(q)}. ∥ G v ∥ ( q ) ≤ ε ∥ v ∥ ( q ) .
Proof. By Proposition 2.12 applied with q q q as the starting exponent, inf G ∈ G ν q ( G ) = 0 \inf_{G\in\mathcal{G}}\nu_q(G)=0 inf G ∈ G ν q ( G ) = 0 . For q < ∞ q<\infty q < ∞ , the claim follows by density of compactly supported inputs in E q \mathcal{E}_q E q and boundedness of the extension. For q = ∞ q=\infty q = ∞ , take ∥ u ∥ ( ∞ ) = 1 \|u\|_{(\infty)}=1 ∥ u ∥ ( ∞ ) = 1 and ∥ G u ∥ ( ∞ ) < ε / 4 \|Gu\|_{(\infty)}<\varepsilon/4 ∥ G u ∥ ( ∞ ) < ε /4 for some G ∈ G G\in\mathcal{G} G ∈ G . Choose j 0 j_0 j 0 with ∥ 1 j 0 u ∥ > 1 / 2 \|1_{j_0}u\|>1/2 ∥ 1 j 0 u ∥ > 1/2 . Let ψ \psi ψ be compactly supported and Lipschitz, with 0 ≤ ψ ≤ 1 0\le\psi\le1 0 ≤ ψ ≤ 1 and ψ ( 0 ) = 1 \psi(0)=1 ψ ( 0 ) = 1 , and let ψ R \psi_R ψ R multiply cell j j j by ψ ( ( j − j 0 ) / R ) \psi((j-j_0)/R) ψ (( j − j 0 ) / R ) . The difference of its multipliers on cells j , k j,k j , k is at most C min { 1 , ∣ j − k ∣ / R } C\min\{1,|j-k|/R\} C min { 1 , ∣ j − k ∣/ R } . Summing the block bounds gives ∥ [ G , ψ R ] u ∥ ( ∞ ) ≤ C η R \|[G,\psi_R]u\|_{(\infty)}\le C\eta_R ∥ [ G , ψ R ] u ∥ ( ∞ ) ≤ C η R , where η R → 0 \eta_R\to0 η R → 0 is as in the proof above. For R R R with C η R < ε / 4 C\eta_R<\varepsilon/4 C η R < ε /4 the compactly supported v = ψ R u v=\psi_Ru v = ψ R u has ∥ v ∥ ( ∞ ) > 1 / 2 \|v\|_{(\infty)}>1/2 ∥ v ∥ ( ∞ ) > 1/2 and ∥ G v ∥ ( ∞ ) < ε / 2 \|Gv\|_{(\infty)}<\varepsilon/2 ∥ G v ∥ ( ∞ ) < ε /2 , as required. □ \square □
Collective compactness. The following closure properties are used in Lemma 2.15 , Proposition 2.6 and Theorem 7.1.
Lemma 2.14. Let W W W be compact, let I , J ⊂ R d I,J\subset\mathbb{R}^{d} I , J ⊂ R d be bounded, and let { G ω } ω ∈ W \{G_{\omega}\}_{\omega\in W} { G ω } ω ∈ W be localized operators.
(a) If ω ↦ 1 I G ω 1 J \omega\mapsto\mathbf{1}_{I}G_{\omega}\mathbf{1}_{J} ω ↦ 1 I G ω 1 J is continuous in operator norm with compact values, then { 1 I G ω 1 J } ω \{\mathbf{1}_{I}G_{\omega}\mathbf{1}_{J}\}_{\omega} { 1 I G ω 1 J } ω is collectively compact on the unit ball of L 2 ( J ) L^{2}(J) L 2 ( J ) .
(b) Finite sums of collectively compact families are collectively compact. If each family { G ω ( m ) } \{G_{\omega}^{(m)}\} { G ω ( m ) } is collectively compact on the unit ball of L 2 ( J ) L^{2}(J) L 2 ( J ) and sup ω ∥ 1 I ( G ω − G ω ( m ) ) 1 J ∥ → 0 \sup_{\omega}\|\mathbf{1}_{I}(G_{\omega}-G_{\omega}^{(m)})\mathbf{1}_{J}\|\to0 sup ω ∥ 1 I ( G ω − G ω ( m ) ) 1 J ∥ → 0 , then { 1 I G ω 1 J } ω \{\mathbf{1}_{I}G_{\omega}\mathbf{1}_{J}\}_{\omega} { 1 I G ω 1 J } ω is collectively compact on the same ball.
(c) Let { G ω } \{G_{\omega}\} { G ω } be collectively compact on a set V \mathcal{V} V , and let m ω m_{\omega} m ω be matrix-valued multipliers with sup ω ∥ m ω ∥ L ∞ < ∞ \sup_{\omega}\|m_{\omega}\|_{L^{\infty}}<\infty sup ω ∥ m ω ∥ L ∞ < ∞ and m ω n → m ω m_{\omega_{n}}\to m_{\omega} m ω n → m ω locally in measure whenever ω n → ω \omega_{n}\to\omega ω n → ω . Then { m ω G ω } \{m_{\omega}G_{\omega}\} { m ω G ω } is collectively compact on V \mathcal{V} V .
(d) If { G ω } \{G_{\omega}\} { G ω } has a common summable profile and is collectively compact on the unit ball of L 2 ( J ) L^{2}(J) L 2 ( J ) for every bounded J J J , then it is collectively compact on the unit ball of X ∞ X_{\infty} X ∞ .
Proof. (a) The set { 1 I G ω 1 J } ω \{\mathbf{1}_{I}G_{\omega}\mathbf{1}_{J}\}_{\omega} { 1 I G ω 1 J } ω is a norm-compact set of compact operators. A finite ϵ \epsilon ϵ -net of it shows that the union of the images of the unit ball is totally bounded. (b) The same ϵ \epsilon ϵ -net argument applies. (c) Given ( ω n , v n ) (\omega_{n},v_{n}) ( ω n , v n ) with v n ∈ V v_{n}\in\mathcal{V} v n ∈ V , extract a subsequence with ω n → ω \omega_{n}\to\omega ω n → ω and 1 I G ω n v n → g \mathbf{1}_{I}G_{\omega_{n}}v_{n}\to g 1 I G ω n v n → g in L 2 ( I ) L^{2}(I) L 2 ( I ) . Then m ω n 1 I G ω n v n → m ω g m_{\omega_{n}}\mathbf{1}_{I}G_{\omega_{n}}v_{n}\to m_{\omega}g m ω n 1 I G ω n v n → m ω g , by the uniform bound and dominated convergence along a further subsequence. (d) Put P L = 1 [ − L , L ] d P_{L}=\mathbf{1}_{[-L,L]^{d}} P L = 1 [ − L , L ] d . By Lemma 2.10 (ii), as in the proof of Lemma 2.11 , the common summable profile gives
sup ω ∈ W sup ∥ v ∥ ( ∞ ) ≤ 1 ∥ 1 I G ω ( I − P L ) v ∥ L 2 ( I ) ≤ η I ( L ) ⟶ 0. \sup_{\omega\in W}\sup_{\|v\|_{(\infty)}\leq1}\|\mathbf{1}_{I}G_{\omega}(I-P_{L})v\|_{L^{2}(I)}\leq\eta_{I}(L)\longrightarrow0. ω ∈ W sup ∥ v ∥ ( ∞ ) ≤ 1 sup ∥ 1 I G ω ( I − P L ) v ∥ L 2 ( I ) ≤ η I ( L ) ⟶ 0.
For fixed L L L , the inputs P L v P_{L}v P L v lie in a common bounded ball of L 2 ( [ − L , L ] d ) L^{2}([-L,L]^{d}) L 2 ([ − L , L ] d ) , whose images are relatively compact by assumption. The uniform tail estimate extends this compactness to the original inputs.
Extraction of profiles and proofs of the criterion
Lemma 2.15. Assume (H3)–(H5). Then (R) holds with B ω = D ω − 1 Γ ω ∗ B_{\omega}=D_{\omega}^{-1}\Gamma_{\omega}^{*} B ω = D ω − 1 Γ ω ∗ and S ω = − D ω − 1 K ω S_{\omega}=-D_{\omega}^{-1}K_{\omega} S ω = − D ω − 1 K ω on X ∞ X_{\infty} X ∞ .
Proof. By Lemma 2.10 (iii), the reflected localization profile bounds Γ ω ∗ \Gamma_{\omega}^{*} Γ ω ∗ uniformly from Z ∞ Z_{\infty} Z ∞ to X ∞ X_{\infty} X ∞ , and (H4) gives a uniform bound for D ω − 1 D_{\omega}^{-1} D ω − 1 . By Lemma 2.10 (iv), Γ ω ∗ ( Γ ω w ) = d ω w + K ω w \Gamma_{\omega}^{*}(\Gamma_{\omega}w)=d_{\omega}w+K_{\omega}w Γ ω ∗ ( Γ ω w ) = d ω w + K ω w for every w ∈ L u l o c 2 w\in L^{2}_{\mathrm{uloc}} w ∈ L uloc 2 ; hence I − B ω Γ ω = − D ω − 1 K ω I-B_{\omega}\Gamma_{\omega}=-D_{\omega}^{-1}K_{\omega} I − B ω Γ ω = − D ω − 1 K ω on X ∞ X_{\infty} X ∞ .
Fix a bounded cube Q Q Q and a sequence ( ω n , v n ) (\omega_{n},v_{n}) ( ω n , v n ) with ∥ v n ∥ ( ∞ ) ≤ 1 \|v_{n}\|_{(\infty)}\leq1 ∥ v n ∥ ( ∞ ) ≤ 1 . Since W W W is compact, we may assume that ω n → ω \omega_{n}\to\omega ω n → ω . Lemma 2.11 allows us to truncate the input to a larger cube with an error tending uniformly to zero. For each fixed truncation, (H5) gives relative compactness of the restricted outputs of K ω n K_{\omega_{n}} K ω n ; thus the same relative compactness holds before truncation (Lemma 2.14 (d)). By (H4), the inverse multipliers d ω n − 1 d_{\omega_{n}}^{-1} d ω n − 1 converge locally in measure; their uniform bound gives strong convergence on L 2 ( Q ; C N ) L^{2}(Q;\mathbb{C}^{N}) L 2 ( Q ; C N ) . Applying them to a convergent subsequence of the restricted outputs preserves convergence (Lemma 2.14 (c)). This proves (R).
Lemma 2.16. Assume (H2), (H3), and (R). Let ω n → ω ∞ \omega_{n}\to\omega_{\infty} ω n → ω ∞ in W W W , and let w n ∈ L u l o c 2 w_{n}\in L^{2}_{\mathrm{uloc}} w n ∈ L uloc 2 satisfy
∥ w n ∥ ( ∞ ) = ∥ w n ∥ L 2 ( Q 0 ) = 1 , ∥ Γ ω n w n ∥ ( ∞ ) → 0. (30) \|w_{n}\|_{(\infty)}=\|w_{n}\|_{L^{2}(Q_{0})}=1,\qquad\|\Gamma_{\omega_{n}}w_{n}\|_{(\infty)}\to0. \tag*{(30)} ∥ w n ∥ ( ∞ ) = ∥ w n ∥ L 2 ( Q 0 ) = 1 , ∥ Γ ω n w n ∥ ( ∞ ) → 0. (30) Then a subsequence of ( w n ) (w_{n}) ( w n ) converges strongly in L l o c 2 L^{2}_{\mathrm{loc}} L loc 2 to some w ∞ ∈ L u l o c 2 w_{\infty}\in L^{2}_{\mathrm{uloc}} w ∞ ∈ L uloc 2 with Γ ω ∞ w ∞ = 0 \Gamma_{\omega_{\infty}}w_{\infty}=0 Γ ω ∞ w ∞ = 0 and ∥ w ∞ ∥ L 2 ( Q 0 ) = ∥ w ∞ ∥ ( ∞ ) = 1 \|w_{\infty}\|_{L^{2}(Q_{0})}=\|w_{\infty}\|_{(\infty)}=1 ∥ w ∞ ∥ L 2 ( Q 0 ) = ∥ w ∞ ∥ ( ∞ ) = 1 .
Proof. By (R),
w n = B ω n Γ ω n w n + S ω n w n . w_{n}=B_{\omega_{n}}\Gamma_{\omega_{n}}w_{n}+S_{\omega_{n}}w_{n}. w n = B ω n Γ ω n w n + S ω n w n .
The first term tends to zero in X ∞ X_{\infty} X ∞ . By (R) the second has a strongly convergent subsequence on every bounded cube. A diagonal extraction gives a single subsequence with w n → w ∞ w_{n}\to w_{\infty} w n → w ∞ strongly in L l o c 2 L^{2}_{\mathrm{loc}} L loc 2 . Cell norms pass to the limit: ∥ 1 k w ∞ ∥ = lim n ∥ 1 k w n ∥ ≤ 1 \|1_{k}w_{\infty}\|=\lim_{n}\|1_{k}w_{n}\|\leq1 ∥ 1 k w ∞ ∥ = lim n ∥ 1 k w n ∥ ≤ 1 for every k k k , with equality for k = 0 k=0 k = 0 . Hence w ∞ ∈ L u l o c 2 w_{\infty}\in L^{2}_{\mathrm{uloc}} w ∞ ∈ L uloc 2 and ∥ w ∞ ∥ ( ∞ ) = ∥ w ∞ ∥ L 2 ( Q 0 ) = 1 \|w_{\infty}\|_{(\infty)}=\|w_{\infty}\|_{L^{2}(Q_{0})}=1 ∥ w ∞ ∥ ( ∞ ) = ∥ w ∞ ∥ L 2 ( Q 0 ) = 1 .
Fix a cell k k k and P L = 1 [ − L , L ] d P_{L}=\mathbf{1}_{[-L,L]^{d}} P L = 1 [ − L , L ] d . By (H3) and Lemma 2.10 (ii),
∥ 1 k Γ ω ( 1 − P L ) u ∥ ≤ e k ( L ) ∥ u ∥ ( ∞ ) , e k ( L ) : = τ k ( L − ∣ k ∣ − 1 ) ⟶ 0. \|1_{k}\Gamma_{\omega}(1-P_{L})u\|\leq e_{k}(L)\|u\|_{(\infty)},\qquad e_{k}(L):=\tau_{k}(L-|k|-1)\longrightarrow0. ∥ 1 k Γ ω ( 1 − P L ) u ∥ ≤ e k ( L ) ∥ u ∥ ( ∞ ) , e k ( L ) := τ k ( L − ∣ k ∣ − 1 ) ⟶ 0.
For fixed L L L , P L w n → P L w ∞ P_{L}w_{n}\to P_{L}w_{\infty} P L w n → P L w ∞ in X X X and the operator norms are uniformly bounded by ∥ κ ∥ \|\kappa\| ∥ κ ∥ . Applying (H2) to the fixed input P L w ∞ P_{L}w_{\infty} P L w ∞ and taking square roots gives
∥ 1 k Γ ω ∞ P L w ∞ ∥ = lim n ∥ 1 k Γ ω n P L w n ∥ ≤ e k ( L ) , \|1_{k}\Gamma_{\omega_{\infty}}P_{L}w_{\infty}\|=\lim_{n}\|1_{k}\Gamma_{\omega_{n}}P_{L}w_{n}\|\leq e_{k}(L), ∥ 1 k Γ ω ∞ P L w ∞ ∥ = n lim ∥ 1 k Γ ω n P L w n ∥ ≤ e k ( L ) ,
where the inequality follows from (30) and the tail bound. Applying the same bound to w ∞ w_{\infty} w ∞ gives ∥ 1 k Γ ω ∞ w ∞ ∥ ≤ 2 e k ( L ) \|1_{k}\Gamma_{\omega_{\infty}}w_{\infty}\|\leq2e_{k}(L) ∥ 1 k Γ ω ∞ w ∞ ∥ ≤ 2 e k ( L ) . Let L → ∞ L\to\infty L → ∞ and then vary k k k to obtain Γ ω ∞ w ∞ = 0 \Gamma_{\omega_{\infty}}w_{\infty}=0 Γ ω ∞ w ∞ = 0 .
Proof of Theorem 2.1 . Observability implies injectivity and uniform lower bounds. This is the only part of the proof that uses (18) . Assume the first assertion, with constant c > 0 c>0 c > 0 . Covariance and the unitarity of translations on X \mathcal{X} X and Z Z Z give the same lower bound on every translate of ω 0 \omega_{0} ω 0 . Given ω ∈ W \omega\in W ω ∈ W , choose h n h_{n} h n such that θ h n ω 0 → ω \theta_{h_{n}}\omega_{0}\to\omega θ h n ω 0 → ω . Fix a bounded-support input f ∈ X f\in\mathcal{X} f ∈ X . By (H3), the output cell norms are dominated by the fixed ℓ 2 \ell^{2} ℓ 2 sequence κ ∗ ( ∥ 1 j f ∥ ) j \kappa*(\|1_{j}f\|)_{j} κ ∗ ( ∥ 1 j f ∥ ) j . By (H2) their squares converge cell by cell. Dominated convergence for the sum of these squares gives convergence of the global observation norms, so
∥ Γ ω f ∥ Z ≥ c ∥ f ∥ X . \|\Gamma_{\omega}f\|_{Z}\geq c\|f\|_{\mathcal{X}}. ∥ Γ ω f ∥ Z ≥ c ∥ f ∥ X .
By density of bounded-support inputs and boundedness of Γ ω \Gamma_\omega Γ ω , this holds for every f ∈ X f \in\mathcal{X} f ∈ X . Consequently,
inf ω ∈ W ν 2 ( Γ ω ) ≥ c > 0. \inf_{\omega\in\mathcal{W}} \nu_2(\Gamma_\omega) \ge c > 0. ω ∈ W inf ν 2 ( Γ ω ) ≥ c > 0.
Proposition 2.12 , applied to the whole family, gives
c ∗ : = inf ω ∈ W , 1 ≤ p ≤ ∞ ν p ( Γ ω ) > 0. c_* := \inf_{\substack{\omega\in\mathcal{W},\,1\le p\le\infty}} \nu_p(\Gamma_\omega) > 0. c ∗ := ω ∈ W , 1 ≤ p ≤ ∞ inf ν p ( Γ ω ) > 0.
The case p = ∞ p=\infty p = ∞ excludes every nonzero uniformly local kernel element and proves the second assertion.
Failure of observability produces a nonzero invisible profile. Suppose the first assertion fails, so that ν 2 ( Γ ω 0 ) = 0 \nu_2(\Gamma_{\omega_0})=0 ν 2 ( Γ ω 0 ) = 0 . Apply Corollary 2.13 to the singleton family { Γ ω 0 } \{\Gamma_{\omega_0}\} { Γ ω 0 } , from p = 2 p=2 p = 2 to q = ∞ q=\infty q = ∞ . For any sequence ε n ↓ 0 \varepsilon_n\downarrow0 ε n ↓ 0 , it gives compactly supported f n ≠ 0 f_n\ne0 f n = 0 satisfying
∥ Γ ω 0 f n ∥ ( ∞ ) ≤ ε n ∥ f n ∥ ( ∞ ) . \|\Gamma_{\omega_0}f_n\|_{(\infty)}\le\varepsilon_n\|f_n\|_{(\infty)}. ∥ Γ ω 0 f n ∥ ( ∞ ) ≤ ε n ∥ f n ∥ ( ∞ ) .
Normalize ∥ f n ∥ ( ∞ ) = 1 \|f_n\|_{(\infty)}=1 ∥ f n ∥ ( ∞ ) = 1 . Since f n f_n f n meets finitely many cells, there is k n ∈ Z d k_n\in\mathbb{Z}^d k n ∈ Z d such that ∥ f n ∥ L 2 ( Q k n ) = 1 \|f_n\|_{L^2(Q_{k_n})}=1 ∥ f n ∥ L 2 ( Q k n ) = 1 . Set
w n = τ k n f n , ω n = θ k n ω 0 . w_n=\tau_{k_n}f_n,\qquad\omega_n=\theta_{k_n}\omega_0. w n = τ k n f n , ω n = θ k n ω 0 .
Integer translations preserve the amalgam norms. Covariance thus gives
∥ w n ∥ ( ∞ ) = ∥ w n ∥ L 2 ( Q 0 ) = 1 , ∥ Γ ω n w n ∥ ( ∞ ) ≤ ε n . \|w_n\|_{(\infty)}=\|w_n\|_{L^2(Q_0)}=1,\qquad\|\Gamma_{\omega_n}w_n\|_{(\infty)}\le\varepsilon_n. ∥ w n ∥ ( ∞ ) = ∥ w n ∥ L 2 ( Q 0 ) = 1 , ∥ Γ ω n w n ∥ ( ∞ ) ≤ ε n .
By compactness, after passing to a subsequence, ω n → ω ∞ ∈ W \omega_n\to\omega_\infty\in\mathcal{W} ω n → ω ∞ ∈ W . Lemma 2.16 yields
Γ ω ∞ w ∞ = 0 , ∥ w ∞ ∥ L 2 ( Q 0 ) = ∥ w ∞ ∥ ( ∞ ) = 1. \Gamma_{\omega_\infty}w_\infty=0,\qquad\|w_\infty\|_{L^2(Q_0)}=\|w_\infty\|_{(\infty)}=1. Γ ω ∞ w ∞ = 0 , ∥ w ∞ ∥ L 2 ( Q 0 ) = ∥ w ∞ ∥ ( ∞ ) = 1.
This proves the asserted profile statement and contradicts the second assertion whenever injectivity holds throughout the hull.
Proof of Corollary 2.2. If the family has a uniform L 2 L^2 L 2 lower bound, Proposition 2.12 gives a common lower bound for all environments and all p ∈ [ 1 , ∞ ] p\in[1,\infty] p ∈ [ 1 , ∞ ] . In particular, its p = ∞ p=\infty p = ∞ lower bound excludes nonzero uniformly local kernel elements. Conversely, suppose inf ω ∈ W ν 2 ( Γ ω ) = 0 \inf_{\omega\in\mathcal{W}}\nu_2(\Gamma_\omega)=0 inf ω ∈ W ν 2 ( Γ ω ) = 0 . Applying Corollary 2.13 to the entire family, from p = 2 p=2 p = 2 to q = ∞ q=\infty q = ∞ , gives parameters ϖ n ∈ W \varpi_n\in\mathcal{W} ϖ n ∈ W and compactly supported f n ≠ 0 f_n\ne0 f n = 0 such that
∥ f n ∥ ( ∞ ) = 1 , ∥ Γ ϖ n f n ∥ ( ∞ ) ≤ ε n , ε n ↓ 0. \|f_n\|_{(\infty)}=1,\qquad\|\Gamma_{\varpi_n}f_n\|_{(\infty)}\le\varepsilon_n,\qquad\varepsilon_n\downarrow0. ∥ f n ∥ ( ∞ ) = 1 , ∥ Γ ϖ n f n ∥ ( ∞ ) ≤ ε n , ε n ↓ 0.
Choose maximizing cells k n k_n k n and set
w n = τ k n f n , ω n = θ k n ϖ n . w_n=\tau_{k_n}f_n,\qquad\omega_n=\theta_{k_n}\varpi_n. w n = τ k n f n , ω n = θ k n ϖ n .
As above, these satisfy (30) . Compactness gives a convergent subsequence of ( ω n ) (\omega_n) ( ω n ) , and Lemma 2.16 produces a nonzero uniformly local kernel element for its limit. This contradicts injectivity on every configuration. The argument uses compactness and invariance of W \mathcal{W} W , without requiring a dense orbit.
Proof of Proposition 2.3 . If Γ ω w = 0 \Gamma_\omega w=0 Γ ω w = 0 , then H ω w = Γ ω ∗ ( Γ ω w ) = 0 H_\omega w=\Gamma_\omega^*(\Gamma_\omega w)=0 H ω w = Γ ω ∗ ( Γ ω w ) = 0 by Lemma 2.10 (iii). This proves the inclusion and (b)⇒ \Rightarrow ⇒ (a). Assume (a). Corollary 2.2 yields c > 0 c>0 c > 0 such that ∥ Γ ω f ∥ Z ≥ c ∥ f ∥ X \|\Gamma_\omega f\|_{\mathcal{Z}}\ge c\|f\|_{\mathcal{X}} ∥ Γ ω f ∥ Z ≥ c ∥ f ∥ X for all ω ∈ W \omega\in\mathcal{W} ω ∈ W and f ∈ X f\in\mathcal{X} f ∈ X . Since
c 2 ∥ f ∥ X 2 ≤ ∥ Γ ω f ∥ Z 2 = ⟨ H ω f , f ⟩ X ≤ ∥ H ω f ∥ X ∥ f ∥ X , c^2\|f\|_{\mathcal{X}}^2\le\|\Gamma_\omega f\|_{\mathcal{Z}}^2=\langle H_\omega f,f\rangle_{\mathcal{X}}\le\|H_\omega f\|_{\mathcal{X}}\|f\|_{\mathcal{X}}, c 2 ∥ f ∥ X 2 ≤ ∥ Γ ω f ∥ Z 2 = ⟨ H ω f , f ⟩ X ≤ ∥ H ω f ∥ X ∥ f ∥ X ,
we get ν 2 ( H ω ) ≥ c 2 \nu_2(H_\omega)\ge c^2 ν 2 ( H ω ) ≥ c 2 for all ω \omega ω . By Lemma 2.10 (iii), the operators H ω H_\omega H ω share the summable profile κ ~ ∗ κ \widetilde{\kappa}*\kappa κ ∗ κ , so Proposition 2.12 , with E = F = X \mathcal{E}=\mathcal{F}=\mathcal{X} E = F = X , gives inf ω ∈ W ν ∞ ( H ω ) > 0 \inf_{\omega\in\mathcal{W}}\nu_\infty(H_\omega)>0 inf ω ∈ W ν ∞ ( H ω ) > 0 . This excludes nonzero elements of ker L u l o c \ker L_{\mathrm{uloc}} ker L uloc .
Necessity of (R)
Proposition 2.17 (Necessity of (R)). Assume (H3), and suppose that
c : = inf ω ∈ W inf f ∈ X ∖ { 0 } ∥ Γ ω f ∥ Z ∥ f ∥ X > 0. (31) c:=\inf_{\omega\in\mathcal{W}}\inf_{f\in\mathcal{X}\setminus\{0\}}\frac{\|\Gamma_\omega f\|_{\mathcal{Z}}}{\|f\|_{\mathcal{X}}}>0. \tag*{(31)} c := ω ∈ W inf f ∈ X ∖ { 0 } inf ∥ f ∥ X ∥ Γ ω f ∥ Z > 0. (31)
Then each Gramian H ω = Γ ω ∗ Γ ω H_\omega=\Gamma_\omega^*\Gamma_\omega H ω = Γ ω ∗ Γ ω , extended to X ∞ \mathcal{X}_\infty X ∞ by Lemma 2.10 , is invertible on X ∞ \mathcal{X}_\infty X ∞ , and sup ω ∥ H ω − 1 ∥ L ( X ∞ ) < ∞ \sup_\omega\|H_\omega^{-1}\|_{\mathcal{L}(\mathcal{X}_\infty)}<\infty sup ω ∥ H ω − 1 ∥ L ( X ∞ ) < ∞ . The operators
B ω : = H ω − 1 Γ ω ∗ : Z ∞ → X ∞ B_\omega:=H_\omega^{-1}\Gamma_\omega^*:\mathcal{Z}_\infty\to\mathcal{X}_\infty B ω := H ω − 1 Γ ω ∗ : Z ∞ → X ∞
are uniformly bounded and satisfy B ω Γ ω = I B_\omega\Gamma_\omega=I B ω Γ ω = I on X ∞ \mathcal{X}_\infty X ∞ . In particular, (R) holds with S ω = 0 S_\omega=0 S ω = 0 . If (H1)–(H2) and (18) hold, the same conclusion follows as soon as the observability inequality holds for ω 0 \omega_0 ω 0 .
Proof. By (H3) and Lemma 2.10 (iii), the adjoints Γ ω ∗ \Gamma_\omega^* Γ ω ∗ are κ ~ \widetilde{\kappa} κ -localized, and the Gramians H ω H_\omega H ω share the summable profile κ ~ ∗ κ \widetilde{\kappa} * \kappa κ ∗ κ . By (31) ,
c 2 ∥ f ∥ X 2 ≤ ⟨ H ω f , f ⟩ ≤ ∥ H ω f ∥ X ∥ f ∥ X , c^2\lVert f\rVert_X^2 \le\langle H_\omega f,f\rangle\le\lVert H_\omega f\rVert_X\lVert f\rVert_X, c 2 ∥ f ∥ X 2 ≤ ⟨ H ω f , f ⟩ ≤ ∥ H ω f ∥ X ∥ f ∥ X ,
so ν 2 ( H ω ) ≥ c 2 \nu_2(H_\omega) \ge c^2 ν 2 ( H ω ) ≥ c 2 for every ω \omega ω . Proposition 2.12 then gives
c 1 : = inf ω ∈ W , 1 ≤ p ≤ ∞ ν p ( H ω ) > 0. c_1 := \inf_{\omega\in\mathcal{W},\,1\le p\le\infty}\nu_p(H_\omega)>0. c 1 := ω ∈ W , 1 ≤ p ≤ ∞ inf ν p ( H ω ) > 0.
Fix ω \omega ω and write H = H ω H=H_\omega H = H ω . The cellwise identification X 1 = ℓ 1 ( L 2 ( Q k ; C N ) k ∈ Z d ) X_1=\ell^1(L^2(Q_k;\mathbb{C}^N)_{k\in\mathbb{Z}^d}) X 1 = ℓ 1 ( L 2 ( Q k ; C N ) k ∈ Z d ) gives X ∞ ∗ ≅ X 1 X_\infty^*\cong X_1 X ∞ ∗ ≅ X 1 under the summed L 2 L^2 L 2 pairing; see also [13 ] for amalgam-space duality. Let f ∈ X 1 f\in X_1 f ∈ X 1 and ψ ∈ X ∞ \psi\in X_\infty ψ ∈ X ∞ . The terms of the double series ∑ j , k ⟨ 1 k H 1 j f , 1 k ψ ⟩ \sum_{j,k}\langle1_kH1_jf,1_k\psi\rangle ∑ j , k ⟨ 1 k H 1 j f , 1 k ψ ⟩ are bounded by ( κ ~ ∗ κ ) ( k − j ) ∥ 1 j f ∥ ∥ 1 k ψ ∥ (\widetilde{\kappa}*\kappa)(k-j)\lVert1_jf\rVert\lVert1_k\psi\rVert ( κ ∗ κ ) ( k − j ) ∥ 1 j f ∥ ∥ 1 k ψ ∥ , so the series converges absolutely. Summing it in both orders and using H ∗ = H H^*=H H ∗ = H gives
⟨ H f , ψ ⟩ = ⟨ f , H ψ ⟩ ( f ∈ X 1 , ψ ∈ X ∞ ) , (32) \langle Hf,\psi\rangle=\langle f,H\psi\rangle\qquad(f\in X_1,\ \psi\in X_\infty), \tag*{(32)} ⟨ H f , ψ ⟩ = ⟨ f , H ψ ⟩ ( f ∈ X 1 , ψ ∈ X ∞ ) , (32)
where H ψ H\psi H ψ denotes the uniformly local extension. On X 1 X_1 X 1 , the operator H H H is bounded below by c 1 c_1 c 1 , so its range is closed. If ψ ∈ X ∞ ≅ X 1 ∗ \psi\in X_\infty\cong X_1^* ψ ∈ X ∞ ≅ X 1 ∗ annihilates H ( X 1 ) H(X_1) H ( X 1 ) , then (32) gives H ψ = 0 H\psi=0 H ψ = 0 , and ν ∞ ( H ) ≥ c 1 \nu_\infty(H)\ge c_1 ν ∞ ( H ) ≥ c 1 forces ψ = 0 \psi=0 ψ = 0 . Hence H H H is invertible on X 1 X_1 X 1 , with ∥ H − 1 ∥ L ( X 1 ) ≤ c 1 − 1 \lVert H^{-1}\rVert_{\mathcal{L}(X_1)}\le c_1^{-1} ∥ H − 1 ∥ L ( X 1 ) ≤ c 1 − 1 . By (32) , the extension of H H H to X ∞ X_\infty X ∞ is the Banach adjoint of H ∣ X 1 H|_{X_1} H ∣ X 1 . It is therefore invertible on X ∞ X_\infty X ∞ , with the same bound.
Consequently ∥ B ω ∥ L ( Z ∞ , X ∞ ) ≤ c 1 − 1 ∥ κ ∥ 1 \lVert B_\omega\rVert_{\mathcal{L}(Z_\infty,X_\infty)}\le c_1^{-1}\lVert\kappa\rVert_1 ∥ B ω ∥ L ( Z ∞ , X ∞ ) ≤ c 1 − 1 ∥ κ ∥ 1 . For w ∈ X ∞ w\in X_\infty w ∈ X ∞ , Lemma 2.10 (iii) gives Γ ω ∗ ( Γ ω w ) = H ω w \Gamma_\omega^*(\Gamma_\omega w)=H_\omega w Γ ω ∗ ( Γ ω w ) = H ω w , and hence B ω Γ ω w = w B_\omega\Gamma_\omega w=w B ω Γ ω w = w . Under (H1)–(H2) and (18) , the first part of the proof of Theorem 2.1 transfers the lower bound from ω 0 \omega_0 ω 0 to every ω ∈ W \omega\in\mathcal{W} ω ∈ W , with the same constant.
Corollary 2.18 (Characterization of observability). Assume (H1)–(H3) and (18) . The observability inequality holds for ω 0 \omega_0 ω 0 if and only if
(R) the uniform reconstruction property holds, and
(U) ker L u l o c 2 Γ ω = { 0 } \ker_{L^2_{\mathrm{uloc}}}\Gamma_\omega=\{0\} ker L uloc 2 Γ ω = { 0 } for every ω ∈ W \omega\in\mathcal{W} ω ∈ W .
Under (H1)–(H3), neither condition implies the other. The Gaussian observer of Remark 2.7 satisfies (U) but not (R). The two-speed example of Section 2.1 satisfies (H4)–(H5), hence (R), but not (U).
Proof. Sufficiency is Theorem 2.1. For necessity, Proposition 2.17 first gives (R); Theorem 2.1 then gives (U). The Gaussian observer is injective on L u l o c 2 L^2_{\mathrm{uloc}} L uloc 2 but not observable, so by Theorem 2.1 it cannot satisfy (R).
Remark 2.19 (Local versus global). Proposition 2.17 uses the observability inequality to construct an exact left inverse. Condition (R) asks for much less: a uniformly bounded reconstruction modulo errors that are collectively compact on bounded windows. Once uniformly bounded reconstruction operators have been constructed, the compactness of the errors can be checked window by window, for instance through (H4)–(H5). Theorem 2.1 converts this local information, together with uniqueness on the hull, into the global estimate.
Lemma 2.20 (Output transformations). Assume that { Γ ω } ω ∈ W \{\Gamma_\omega\}_{\omega\in\mathcal{W}} { Γ ω } ω ∈ W satisfies (H1), (H3) and (R). Let Z ′ = L 2 ( ( 0 , T ) × R d ; C m ) \mathcal{Z}'=L^2((0,T)\times\mathbb{R}^d;\mathbb{C}^m) Z ′ = L 2 (( 0 , T ) × R d ; C m ) , and let M ω ∈ L ( Z , Z ′ ) M_\omega\in\mathcal{L}(\mathcal{Z},\mathcal{Z}') M ω ∈ L ( Z , Z ′ ) , ω ∈ W \omega\in\mathcal{W} ω ∈ W , satisfy:
(a) covariance: M θ h ω τ h Z = τ h Z ′ M ω M_{\theta_h\omega}\tau_h^Z=\tau_h^{Z'}M_\omega M θ h ω τ h Z = τ h Z ′ M ω for h ∈ R d h\in\mathbb{R}^d h ∈ R d ;
(b) localization: ∥ 1 k M ω 1 j ∥ ≤ ν ( k − j ) \lVert1_kM_\omega1_j\rVert\le\nu(k-j) ∥ 1 k M ω 1 j ∥ ≤ ν ( k − j ) for some ν ∈ ℓ 1 ( Z d ) \nu\in\ell^1(\mathbb{Z}^d) ν ∈ ℓ 1 ( Z d ) , independent of ω \omega ω ;
(c) uniform left invertibility: there are operators L ω ∈ L ( Z ′ , Z ) L_\omega\in\mathcal{L}(\mathcal{Z}',\mathcal{Z}) L ω ∈ L ( Z ′ , Z ) with sup ω ∥ L ω ∥ < ∞ \sup_\omega\lVert L_\omega\rVert<\infty sup ω ∥ L ω ∥ < ∞ and L ω M ω z = z L_\omega M_\omega z=z L ω M ω z = z for all z ∈ Z ∞ z\in\mathcal{Z}_\infty z ∈ Z ∞ .
Then Γ ω ′ : = M ω Γ ω \Gamma'_\omega:=M_\omega\Gamma_\omega Γ ω ′ := M ω Γ ω satisfies (H1), (H3) with profile ν ∗ κ \nu*\kappa ν ∗ κ , and (R) with
B ω ′ = B ω L ω , S ω ′ = S ω . B'_\omega=B_\omega L_\omega,\qquad S'_\omega=S_\omega. B ω ′ = B ω L ω , S ω ′ = S ω . Moreover, ker L u l o c 2 Γ ω ′ = ker L u l o c 2 Γ ω \ker_{L^2_{\mathrm{uloc}}}\Gamma'_\omega=\ker_{L^2_{\mathrm{uloc}}}\Gamma_\omega ker L uloc 2 Γ ω ′ = ker L uloc 2 Γ ω for every ω ∈ W \omega\in\mathcal{W} ω ∈ W .
Proof. Covariance follows from (a) and (H1). By Lemma 2.10(iii), with the obvious change of output space, Γ ω ′ \Gamma'_\omega Γ ω ′ is ( ν ∗ κ ) (\nu*\kappa) ( ν ∗ κ ) -localized, and Γ ω ′ w = M ω ( Γ ω w ) \Gamma'_\omega w=M_\omega(\Gamma_\omega w) Γ ω ′ w = M ω ( Γ ω w ) for w ∈ X ∞ w\in X_\infty w ∈ X ∞ . Hence
B ω ′ Γ ω ′ w = B ω L ω M ω Γ ω w = B ω Γ ω w , B'_\omega\Gamma'_\omega w=B_\omega L_\omega M_\omega\Gamma_\omega w=B_\omega\Gamma_\omega w, B ω ′ Γ ω ′ w = B ω L ω M ω Γ ω w = B ω Γ ω w , so I − B ω ′ Γ ω ′ = S ω I-B'_\omega\Gamma'_\omega=S_\omega I − B ω ′ Γ ω ′ = S ω on X ∞ X_\infty X ∞ , and sup ω ∥ B ω ′ ∥ ≤ sup ω ∥ B ω ∥ sup ω ∥ L ω ∥ \sup_\omega\lVert B'_\omega\rVert\le\sup_\omega\lVert B_\omega\rVert\sup_\omega\lVert L_\omega\rVert sup ω ∥ B ω ′ ∥ ≤ sup ω ∥ B ω ∥ sup ω ∥ L ω ∥ . If Γ ω ′ w = 0 \Gamma'_\omega w=0 Γ ω ′ w = 0 , then Γ ω w = L ω M ω Γ ω w = 0 \Gamma_\omega w=L_\omega M_\omega\Gamma_\omega w=0 Γ ω w = L ω M ω Γ ω w = 0 . The converse inclusion is immediate.
Remark 2.21 (Continuity of the cell energies). Assumption (H2) does not pass automatically from Γ ω \Gamma_\omega Γ ω to M ω Γ ω M_\omega\Gamma_\omega M ω Γ ω , because it controls only cell energies. Two sufficient conditions cover the examples of this paper.
(i) M ω M_\omega M ω is multiplication by a matrix field m ω ( t , x ) m_\omega(t,x) m ω ( t , x ) with sup ω ∥ m ω ∥ L ∞ < ∞ \sup_\omega\lVert m_\omega\rVert_{L^\infty}<\infty sup ω ∥ m ω ∥ L ∞ < ∞ and m ω n → m ω m_{\omega_n}\to m_\omega m ω n → m ω locally in measure, and Γ ω n f → Γ ω f \Gamma_{\omega_n}f\to\Gamma_\omega f Γ ω n f → Γ ω f in L l o c 2 L^2_{\mathrm{loc}} L loc 2 for every bounded-support f f f .
(ii) The output splits into channels, Γ ω = ( Γ ω s , Γ ω e ) \Gamma_\omega=(\Gamma_\omega^s,\Gamma_\omega^e) Γ ω = ( Γ ω s , Γ ω e ) , where Γ s \Gamma^s Γ s has the strong convergence in (i), and M ω = diag ( m ω , I ) M_\omega=\operatorname{diag}(m_\omega,I) M ω = diag ( m ω , I ) with m ω m_\omega m ω as in (i). The cell energy of Γ e \Gamma^e Γ e then converges, since it is the difference between the total energy, controlled by (H2), and the energy of Γ s \Gamma^s Γ s .
Bounded-gap observation sets and compact weight families
We record the geometric and compactness ingredients used in the localized observability argument. These depend only on the characteristic speeds and the observation geometry. Let Λ ⊂ R ∖ { 0 } \Lambda\subset\mathbb{R}\setminus\{0\} Λ ⊂ R ∖ { 0 } be a finite nonempty set, and define
λ ∙ = min λ ∈ Λ ∣ λ ∣ , λ ∙ = max λ ∈ Λ ∣ λ ∣ . \lambda_\bullet=\min_{\lambda\in\Lambda}|\lambda|,\qquad\lambda^\bullet=\max_{\lambda\in\Lambda}|\lambda|. λ ∙ = λ ∈ Λ min ∣ λ ∣ , λ ∙ = λ ∈ Λ max ∣ λ ∣.
For ℓ > 0 \ell>0 ℓ > 0 and G ≥ 0 G\ge0 G ≥ 0 , let G ℓ , G \mathcal{G}_{\ell,G} G ℓ , G be the class of open sets
Ω = ⋃ n ∈ Z ( a n , b n ) , a n < b n ≤ a n + 1 , b n − a n ≥ ℓ , a n + 1 − b n ≤ G , (33) \Omega=\bigcup_{n\in\mathbb{Z}}(a_n,b_n),\qquad a_n<b_n\le a_{n+1},\qquad b_n-a_n\ge\ell,\qquad a_{n+1}-b_n\le G, \tag*{(33)} Ω = n ∈ Z ⋃ ( a n , b n ) , a n < b n ≤ a n + 1 , b n − a n ≥ ℓ , a n + 1 − b n ≤ G , (33)
with endpoints tending to p m ∞ pm\infty p m ∞ as appropriate. We also include Ω = R \Omega=\mathbb{R} Ω = R in this class. For T > 0 T>0 T > 0 and a bounded measurable weight a : R → [ 0 , 1 ] a:\mathbb{R}\to[0,1] a : R → [ 0 , 1 ] , set
m λ , a ( z ) : = ∫ 0 T a ( z − λ s ) d s . (34) m_{\lambda,a}(z):=\int_0^T a(z-\lambda s)\,ds. \tag*{(34)} m λ , a ( z ) := ∫ 0 T a ( z − λ s ) d s . (34)
Lemma 3.1 (Geometric intersection bound). Let O ∈ G L 0 , G 0 O\in\mathcal{G}_{L_0,G_0} O ∈ G L 0 , G 0 , where L 0 > 0 L_0>0 L 0 > 0 and G 0 ≥ 0 G_0\ge0 G 0 ≥ 0 . Every interval J J J of length L > G 0 L>G_0 L > G 0 satisfies
∣ J ∩ O ∣ ≥ min { L 0 , L − G 0 } . (35) |J\cap O|\ge\min\{L_0,L-G_0\}. \tag*{(35)} ∣ J ∩ O ∣ ≥ min { L 0 , L − G 0 } . (35) Proof. The assertion is immediate if O = R O=\mathbb{R} O = R . If J J J contains a whole observed interval, then ∣ J ∩ O ∣ ≥ L 0 |J\cap O|\ge L_0 ∣ J ∩ O ∣ ≥ L 0 . Otherwise, it meets at most two observed intervals, and its unobserved part is contained, up to endpoints, in a single gap. Hence ∣ J ∖ O ∣ ≤ G 0 |J\setminus O|\le G_0 ∣ J ∖ O ∣ ≤ G 0 , giving ∣ J ∩ O ∣ ≥ L − G 0 |J\cap O|\ge L-G_0 ∣ J ∩ O ∣ ≥ L − G 0 .
Lemma 3.2. Assume T > G / λ ∙ T>G/\lambda_\bullet T > G / λ ∙ and choose
0 < δ < min { ℓ 4 , λ ∙ T − G 4 } . 0<\delta<\min\left\{\frac{\ell}{4},\frac{\lambda_\bullet T-G}{4}\right\}. 0 < δ < min { 4 ℓ , 4 λ ∙ T − G } . For Ω ∈ G ℓ , G \Omega\in\mathcal{G}_{\ell,G} Ω ∈ G ℓ , G , define
a Ω ( x ) = { min { 1 , δ − 1 dist ( x , R ∖ Ω ) } , Ω ≠ R , 1 , Ω = R . (36) a_\Omega(x)=
\begin{cases}
\min\{1,\delta^{-1}\operatorname{dist}(x,\mathbb{R}\setminus\Omega)\}, & \Omega\ne\mathbb{R},\\
1, & \Omega=\mathbb{R}.
\end{cases} \tag*{(36)} a Ω ( x ) = { min { 1 , δ − 1 dist ( x , R ∖ Ω )} , 1 , Ω = R , Ω = R . (36) Then 0 ≤ a Ω ≤ 1 Ω 0\le a_\Omega\le\mathbf{1}_\Omega 0 ≤ a Ω ≤ 1 Ω , Lip ( a Ω ) ≤ δ − 1 \operatorname{Lip}(a_\Omega)\le\delta^{-1} Lip ( a Ω ) ≤ δ − 1 , and, for every λ ∈ Λ \lambda\in\Lambda λ ∈ Λ and z ∈ R z\in\mathbb{R} z ∈ R ,
m λ , a Ω ( z ) ≥ 1 ∣ λ ∣ min { ℓ − 2 δ , ∣ λ ∣ T − G − 2 δ } ≥ m 0 , where m 0 : = 1 λ ∙ min { ℓ − 2 δ , λ ∙ T − G − 2 δ } > 0. (37) m_{\lambda,a_\Omega}(z)\ge\frac{1}{|\lambda|}\min\{\ell-2\delta,|\lambda|T-G-2\delta\}\ge m_0,\qquad\text{where}\qquad m_0:=\frac{1}{\lambda^\bullet}\min\{\ell-2\delta,\lambda_\bullet T-G-2\delta\}>0. \tag*{(37)} m λ , a Ω ( z ) ≥ ∣ λ ∣ 1 min { ℓ − 2 δ , ∣ λ ∣ T − G − 2 δ } ≥ m 0 , where m 0 := λ ∙ 1 min { ℓ − 2 δ , λ ∙ T − G − 2 δ } > 0. (37) Moreover, the family
A = { a : R → [ 0 , 1 ] : Lip ( a ) ≤ δ − 1 , m λ , a ( z ) ≥ m 0 for all λ ∈ Λ , z ∈ R } (38) \mathfrak{A}=\left\{a:\mathbb{R}\to[0,1]:\ \operatorname{Lip}(a)\le\delta^{-1},\quad m_{\lambda,a}(z)\ge m_0\quad\text{for all }\lambda\in\Lambda,\ z\in\mathbb{R}\right\} \tag*{(38)} A = { a : R → [ 0 , 1 ] : Lip ( a ) ≤ δ − 1 , m λ , a ( z ) ≥ m 0 for all λ ∈ Λ , z ∈ R } (38) is compact for locally uniform convergence, invariant under spatial translations, and contains every a Ω a_\Omega a Ω .
Proof. The range and Lipschitz bounds follow from the definition. For Ω ≠ R \Omega\ne\mathbb{R} Ω = R , the weight a Ω a_\Omega a Ω equals one on
Ω δ : = ⋃ n ∈ Z ( a n + δ , b n − δ ) . \Omega_\delta:=\bigcup_{n\in\mathbb{Z}}(a_n+\delta,b_n-\delta). Ω δ := n ∈ Z ⋃ ( a n + δ , b n − δ ) .
The intervals of O δ O_\delta O δ have lengths at least ℓ − 2 δ \ell- 2\delta ℓ − 2 δ , and their gaps have lengths at most G + 2 δ G + 2\delta G + 2 δ . The characteristic interval J λ , z : = { z − λ s : 0 < s < T } J_{\lambda,z}:=\{z-\lambda s:0<s<T\} J λ , z := { z − λ s : 0 < s < T } has length ∣ λ ∣ T > G + 2 δ |\lambda|T>G+2\delta ∣ λ ∣ T > G + 2 δ . Therefore, by (35) ,
m λ , a Ω ( z ) ≥ ∣ J λ , z ∩ O δ ∣ ∣ λ ∣ ≥ 1 ∣ λ ∣ min { ℓ − 2 δ , ∣ λ ∣ T − G − 2 δ } ≥ m 0 . m_{\lambda,a_\Omega}(z)\ge\frac{|J_{\lambda,z}\cap O_\delta|}{|\lambda|}\ge\frac{1}{|\lambda|}\min\{\ell-2\delta,|\lambda|T-G-2\delta\}\ge m_0. m λ , a Ω ( z ) ≥ ∣ λ ∣ ∣ J λ , z ∩ O δ ∣ ≥ ∣ λ ∣ 1 min { ℓ − 2 δ , ∣ λ ∣ T − G − 2 δ } ≥ m 0 .
For Ω = R \Omega=\mathbb{R} Ω = R , the occupation equals T T T , and the same bounds hold. The functions in A \mathcal{A} A are uniformly bounded and share a common Lipschitz constant. Arzelà–Ascoli and a diagonal argument give relative compactness for locally uniform convergence. The family is closed: the range and Lipschitz bounds pass to the limit, as do the occupation bounds, since a j → a a_j\to a a j → a locally uniformly implies
m λ , a j ⟶ m λ , a m_{\lambda,a_j}\longrightarrow m_{\lambda,a} m λ , a j ⟶ m λ , a
locally uniformly in z z z .
Thus A \mathcal{A} A is compact. Finally, for a h ( x ) : = a ( x + h ) a_h(x):=a(x+h) a h ( x ) := a ( x + h ) , m λ , a h ( z ) = m λ , a ( z + h ) m_{\lambda,a_h}(z)=m_{\lambda,a}(z+h) m λ , a h ( z ) = m λ , a ( z + h ) , which proves translation invariance.
Why the translation hull is necessary
This section shows, on two explicit examples, that uniqueness for the original configuration does not suffice.
The Gramian of a variable observer
Let V = diag ( λ 1 I d 1 , … , λ p I d p ) V=\operatorname{diag}(\lambda_1 I_{d_1},\ldots,\lambda_p I_{d_p}) V = diag ( λ 1 I d 1 , … , λ p I d p ) , with distinct real λ j \lambda_j λ j and n = ∑ j d j n=\sum_j d_j n = ∑ j d j . For f = ( f 1 , … , f p ) f=(f_1,\ldots,f_p) f = ( f 1 , … , f p ) , set ( U ( t ) f ) j ( x ) = f j ( x − λ j t ) (U(t)f)_j(x)=f_j(x-\lambda_jt) ( U ( t ) f ) j ( x ) = f j ( x − λ j t ) . Given B T = ( C 1 , … , C p ) ∈ C b ∞ ( R ; C m × n ) B^{\mathsf{T}}=(C_1,\ldots,C_p)\in C_b^\infty(\mathbb{R};\mathbb{C}^{m\times n}) B T = ( C 1 , … , C p ) ∈ C b ∞ ( R ; C m × n ) and T > 0 T>0 T > 0 , define the bounded observation operator
O B f ( t , x ) = B T ( x ) U ( t ) f ( x ) . O_Bf(t,x)=B^{\mathsf{T}}(x)U(t)f(x). O B f ( t , x ) = B T ( x ) U ( t ) f ( x ) .
Its translation hull H ( B T ) = { B T ( ⋅ + h ) : h ∈ R } \mathcal{H}(B^{\mathsf{T}})=\{B^{\mathsf{T}}(\cdot+h):h\in\mathbb{R}\} H ( B T ) = { B T ( ⋅ + h ) : h ∈ R } is compact and retains the uniform derivative bounds. The diagonal characteristic contribution to G B = O B ∗ O B G_B=O_B^*O_B G B = O B ∗ O B is the multiplication operator D B = diag ( D 1 , … , D p ) D_B=\operatorname{diag}(D_1,\ldots,D_p) D B = diag ( D 1 , … , D p ) , where
D j ( z ) = ∫ 0 T C j ( z + λ j t ) ∗ C j ( z + λ j t ) d t . D_j(z)=\int_0^T C_j(z+\lambda_jt)^*C_j(z+\lambda_jt)\,\mathrm{d}t. D j ( z ) = ∫ 0 T C j ( z + λ j t ) ∗ C j ( z + λ j t ) d t .
Assume uniform characteristic visibility:
D j ( z ) ≥ d I d j ( z ∈ R , 1 ≤ j ≤ p ) , d > 0. (39) D_j(z)\ge dI_{d_j}\qquad(z\in\mathbb{R},\ 1\le j\le p),\qquad d>0. \tag*{(39)} D j ( z ) ≥ d I d j ( z ∈ R , 1 ≤ j ≤ p ) , d > 0. (39)
This condition is preserved throughout H ( B T ) \mathcal{H}(B^{\mathsf{T}}) H ( B T ) .
The same rank and visibility tests, different limits at infinity
From now on, take
V = diag ( I 2 , 2 I 2 ) , J = ( 0 − 1 1 0 ) , R ( x ) = e x J , C r ( x ) = ( I 2 , − r ( x ) R ( x ) ) . V=\operatorname{diag}(I_2,2I_2),\qquad J=\begin{pmatrix}0&-1\\1&0\end{pmatrix},\qquad R(x)=e^{xJ},\qquad C_r(x)=(I_2,-r(x)R(x)). V = diag ( I 2 , 2 I 2 ) , J = ( 0 1 − 1 0 ) , R ( x ) = e x J , C r ( x ) = ( I 2 , − r ( x ) R ( x )) .
Let 0 ≤ χ ≤ 1 0\le\chi\le1 0 ≤ χ ≤ 1 be smooth and 2 π 2\pi 2 π -periodic, with ω χ = { χ > 0 } ≠ ∅ \omega_\chi=\{\chi>0\}\ne\varnothing ω χ = { χ > 0 } = ∅ a finite union of open bands per period, and let G G G be the maximal gap length. The case χ ≡ 1 \chi\equiv1 χ ≡ 1 is allowed, with G = 0 G=0 G = 0 . We observe
B T ( x ) q ( t , x ) = χ ( x ) ( q 1 ( t , x ) − r ( x ) R ( x ) q 2 ( t , x ) ) , 0 < t < T . B^{\mathsf{T}}(x)q(t,x)=\chi(x)(q_1(t,x)-r(x)R(x)q_2(t,x)),\qquad0<t<T. B T ( x ) q ( t , x ) = χ ( x ) ( q 1 ( t , x ) − r ( x ) R ( x ) q 2 ( t , x )) , 0 < t < T .
For r ∈ C b ∞ ( R ) r\in C_b^\infty(\mathbb{R}) r ∈ C b ∞ ( R ) bounded above and below by positive constants, the unlocalized pointwise observability matrix is
( C r ( x ) C r ( x ) V ) = ( I 2 − r ( x ) R ( x ) I 2 − 2 r ( x ) R ( x ) ) . \begin{pmatrix}C_r(x)\\C_r(x)V\end{pmatrix}
=
\begin{pmatrix}
I_2&-r(x)R(x)\\
I_2&-2r(x)R(x)
\end{pmatrix}. ( C r ( x ) C r ( x ) V ) = ( I 2 I 2 − r ( x ) R ( x ) − 2 r ( x ) R ( x ) ) .
It is uniformly invertible: subtracting the first block row from the second leaves the invertible block − r ( x ) R ( x ) -r(x)R(x) − r ( x ) R ( x ) . For the localized observer, the diagonal Gramian blocks are
D 1 ( z ) = ∫ 0 T χ ( z + t ) 2 d t I 2 , D 2 ( z ) = ∫ 0 T χ ( z + 2 t ) 2 r ( z + 2 t ) 2 d t I 2 . D_1(z)=\int_0^T\chi(z+t)^2\,\mathrm{d}t\,I_2,\qquad D_2(z)=\int_0^T\chi(z+2t)^2r(z+2t)^2\,\mathrm{d}t\,I_2. D 1 ( z ) = ∫ 0 T χ ( z + t ) 2 d t I 2 , D 2 ( z ) = ∫ 0 T χ ( z + 2 t ) 2 r ( z + 2 t ) 2 d t I 2 .
When T > G T>G T > G , every characteristic interval meets ω χ \omega_\chi ω χ . The integrals of χ 2 \chi^2 χ 2 are therefore positive continuous periodic functions, with positive minima; the lower bound on r r r gives (39) . If χ ≡ 1 \chi\equiv1 χ ≡ 1 and r ≥ c r\ge c r ≥ c , then D B ≥ T diag ( I 2 , c 2 I 2 ) D_B\ge T\operatorname{diag}(I_2,c^2I_2) D B ≥ T diag ( I 2 , c 2 I 2 ) .
Lemma 4.1. Suppose r > 0 r>0 r > 0 is smooth and h = r ′ / r h=r'/r h = r ′ / r is nonconstant on every nonempty open interval. If a distributional transport solution satisfies χ ( x ) ( q 1 − r Q 2 ) = 0 \chi(x)(q_1-rQ_2)=0 χ ( x ) ( q 1 − r Q 2 ) = 0 on ( 0 , T ) × R (0,T)\times\mathbb{R} ( 0 , T ) × R , with T > G T>G T > G , then its initial datum is zero. No integrability assumption is needed.
Proof. On each component I I I of ω χ \omega_\chi ω χ , q 1 = r R q 2 q_1=rRq_2 q 1 = r R q 2 . Apply ∂ t + ∂ x \partial_t+\partial_x ∂ t + ∂ x and use ( ∂ t + 2 ∂ x ) q 2 = 0 (\partial_t+2\partial_x)q_2=0 ( ∂ t + 2 ∂ x ) q 2 = 0 to obtain ∂ x q 2 = ( h I 2 + J ) q 2 \partial_xq_2=(h_I2+J)q_2 ∂ x q 2 = ( h I 2 + J ) q 2 in distributions. Thus ∂ x ( r − 1 R ( − x ) q 2 ) = 0 \partial_x(r^{-1}R(-x)q_2)=0 ∂ x ( r − 1 R ( − x ) q 2 ) = 0 , so q 2 ( t , x ) = r ( x ) R ( x ) v I ( t ) q_2(t,x)=r(x)R(x)v_I(t) q 2 ( t , x ) = r ( x ) R ( x ) v I ( t ) for a vector distribution v I v_I v I . Its transport equation and its x x x derivative give
v I ′ + 2 J v I + 2 h ( x ) v I = 0 , h ′ ( x ) v I = 0. v_I' + 2Jv_I + 2h(x)v_I = 0,\qquad h'(x)v_I=0. v I ′ + 2 J v I + 2 h ( x ) v I = 0 , h ′ ( x ) v I = 0.
Since h ′ h' h ′ is not identically zero on I I I , testing in x x x gives v I = 0 v_I=0 v I = 0 and hence q 1 = q 2 = 0 q_1=q_2=0 q 1 = q 2 = 0 on ( 0 , T ) × ω χ (0,T)\times\omega_\chi ( 0 , T ) × ω χ . Write q j ( t , x ) = f j ( x − λ j t ) q_j(t,x)=f_j(x-\lambda_jt) q j ( t , x ) = f j ( x − λ j t ) , with λ 1 = 1 \lambda_1=1 λ 1 = 1 , λ 2 = 2 \lambda_2=2 λ 2 = 2 . Each interval z + λ j ( 0 , T ) z+\lambda_j(0,T) z + λ j ( 0 , T ) has length greater than G G G and meets ω χ \omega_\chi ω χ ; Lemma B.1 , applied with speed − λ j -\lambda_j − λ j , therefore implies f j = 0 f_j=0 f j = 0 .
Theorem 4.2. Let r ( x ) = 2 + sin x r(x)=2+\sin x r ( x ) = 2 + sin x . For the observer χ C R \chi_{\mathbb{C}_R} χ C R above and T > 0 T>0 T > 0 , there is a constant C T < ∞ C_T<\infty C T < ∞ such that
∥ q ( 0 ) ∥ 2 2 ≤ C T ∫ 0 T ∫ R χ ( x ) ∣ q 1 ( t , x ) − ( 2 + sin x ) R ( x ) q 2 ( t , x ) ∣ 2 d x d t \lVert q(0)\rVert_2^2 \le C_T \int_0^T\int_{\mathbb{R}} \chi(x)\lvert q_1(t,x)-(2+\sin x)R(x)q_2(t,x)\rvert^2\,dx\,dt ∥ q ( 0 ) ∥ 2 2 ≤ C T ∫ 0 T ∫ R χ ( x ) ∣ q 1 ( t , x ) − ( 2 + sin x ) R ( x ) q 2 ( t , x ) ∣ 2 d x d t if and only if T > G T>G T > G . In particular, with χ ≡ 1 \chi\equiv1 χ ≡ 1 , observability holds for every T > 0 T>0 T > 0 .
Proof. Suppose first that T > G T>G T > G . Visibility was verified above. By periodicity every hull element is χ ( x + θ ) ( I 2 , − ( 2 + sin ( x + θ ) ) R ( x + θ ) ) \chi(x+\theta)(I_2,-(2+\sin(x+\theta))R(x+\theta)) χ ( x + θ ) ( I 2 , − ( 2 + sin ( x + θ )) R ( x + θ )) . The logarithmic derivative cos x / ( 2 + sin x ) \cos x/(2+\sin x) cos x / ( 2 + sin x ) is real analytic and nonconstant, and so are all its translates. Since R ′ = J R R'=JR R ′ = J R is preserved by translation and the gaps retain the bound G G G , Lemma 4.1 gives distributional uniqueness on every hull element. Proposition 2.6 applies to this hull with characteristic speeds − 1 -1 − 1 and − 2 -2 − 2 , zero integral correction, and the coercivity bound (39) . Together with uniqueness on every hull element, Theorem 2.1 yields the observability estimate.
If 0 < T ≤ G 0<T\le G 0 < T ≤ G , let z z z be the left endpoint of a gap of maximal length. For initial data ( f 1 , 0 ) (f_1,0) ( f 1 , 0 ) the observed energy is ∫ m ( y ) ∣ f 1 ( y ) ∣ 2 d y \int m(y)\lvert f_1(y)\rvert^2\,dy ∫ m ( y ) ∣ f 1 ( y ) ∣ 2 d y , where m ( y ) = ∫ 0 T χ ( y + t ) 2 d t m(y)=\int_0^T\chi(y+t)^2\,dt m ( y ) = ∫ 0 T χ ( y + t ) 2 d t . Since m m m is continuous and m ( z ) = 0 m(z)=0 m ( z ) = 0 , unit packets supported in shrinking neighborhoods of z z z have observation tending to zero.
Theorem 4.3. Let r ( x ) = 2 + tanh x r(x)=2+\tanh x r ( x ) = 2 + tanh x . The observer χ C R \chi_{\mathbb{C}_R} χ C R has distributional uniqueness for every T > G T>G T > G , and satisfies the same uniform algebraic and characteristic visibility conditions as above. Nevertheless, for every fixed T > 0 T>0 T > 0 there are smooth compactly supported initial data f n f_n f n with ∥ f n ∥ 2 = 1 \lVert f_n\rVert_2=1 ∥ f n ∥ 2 = 1 and ∥ O χ C R f n ∥ 2 → 0 \lVert O_{\chi_{\mathbb{C}_R}}f_n\rVert_2\to0 ∥ O χ C R f n ∥ 2 → 0 .
Proof. The logarithmic derivative h ( x ) = sech 2 x / ( 2 + tanh x ) h(x)=\operatorname{sech}^2x/(2+\tanh x) h ( x ) = sech 2 x / ( 2 + tanh x ) is real analytic and nonconstant, so Lemma 4.1 gives uniqueness for T > G T>G T > G . The bounds 1 < r < 3 1<r<3 1 < r < 3 give the algebraic and characteristic visibility for T > G T>G T > G already established.
For the failure of observability, choose a real φ ∈ C c ∞ ( − 1 , 1 ) \varphi\in C^\infty_c(-1,1) φ ∈ C c ∞ ( − 1 , 1 ) with ∥ φ ∥ 2 = 1 \lVert\varphi\rVert_2=1 ∥ φ ∥ 2 = 1 , a unit vector v ∈ R 2 v\in\mathbb{R}^2 v ∈ R 2 , and set
a n ( x ) = n − 1 / 2 φ ( x − n 2 n ) , α = 3 10 , f 1 , n ( x ) = α a n ( x ) R ( 2 x ) v , f 2 , n ( x ) = α 3 a n ( x ) R ( x ) v . a_n(x)=n^{-1/2}\varphi\left(\frac{x-n^2}{n}\right),\qquad\alpha=\frac{3}{\sqrt{10}},\qquad f_{1,n}(x)=\alpha a_n(x)R(2x)v,\qquad f_{2,n}(x)=\frac{\alpha}{3}a_n(x)R(x)v. a n ( x ) = n − 1/2 φ ( n x − n 2 ) , α = 10 3 , f 1 , n ( x ) = α a n ( x ) R ( 2 x ) v , f 2 , n ( x ) = 3 α a n ( x ) R ( x ) v .
Orthogonality of R R R gives ∥ f n ∥ 2 2 = α 2 ( 1 + 1 / 9 ) = 1 \lVert f_n\rVert_2^2=\alpha^2(1+1/9)=1 ∥ f n ∥ 2 2 = α 2 ( 1 + 1/9 ) = 1 . The two transports satisfy
q 1 , n − R q 2 , n = α R ( 2 x − 2 t ) v [ a n ( x − t ) − r ( x ) 3 a n ( x − 2 t ) ] . q_{1,n}-Rq_{2,n}=\alpha R(2x-2t)v\left[a_n(x-t)-\frac{r(x)}{3}a_n(x-2t)\right]. q 1 , n − R q 2 , n = α R ( 2 x − 2 t ) v [ a n ( x − t ) − 3 r ( x ) a n ( x − 2 t ) ] .
Now ∥ a n ′ ∥ 2 = n − 1 ∥ φ ′ ∥ 2 \lVert a_n'\rVert_2=n^{-1}\lVert\varphi'\rVert_2 ∥ a n ′ ∥ 2 = n − 1 ∥ φ ′ ∥ 2 , and
∥ a n ( ⋅ − t ) − a n ( ⋅ − 2 t ) ∥ 2 ≤ t ∥ a n ′ ∥ 2 . \lVert a_n(\cdot-t)-a_n(\cdot-2t)\rVert_2\le t\lVert a_n'\rVert_2. ∥ a n ( ⋅ − t ) − a n ( ⋅ − 2 t ) ∥ 2 ≤ t ∥ a n ′ ∥ 2 .
Whenever a n ( x − 2 t ) ≠ 0 a_n(x-2t)\ne0 a n ( x − 2 t ) = 0 and 0 < t < T 0<t<T 0 < t < T , one has x ≥ n 2 − n x\ge n^2-n x ≥ n 2 − n . Therefore, with ε n = sup x ≥ n 2 − n ∣ 1 − r ( x ) / 3 ∣ → 0 \varepsilon_n=\sup_{x\ge n^2-n}\lvert1-r(x)/3\rvert\to0 ε n = sup x ≥ n 2 − n ∣ 1 − r ( x ) /3 ∣ → 0 and 0 ≤ χ ≤ 1 0\le\chi\le1 0 ≤ χ ≤ 1 ,
∥ O χ C R f n ∥ 2 2 ≤ 2 α 2 ( T 3 3 n 2 ∥ φ ′ ∥ 2 2 + T ε n 2 ) → 0. \lVert O_{\chi_{\mathbb{C}_R}}f_n\rVert_2^2\le2\alpha^2\left(\frac{T^3}{3n^2}\lVert\varphi'\rVert_2^2+T\varepsilon_n^2\right)\to0. ∥ O χ C R f n ∥ 2 2 ≤ 2 α 2 ( 3 n 2 T 3 ∥ φ ′ ∥ 2 2 + T ε n 2 ) → 0.
This contradicts any uniform observability inequality.
Remark 4.4 (What the translated limits detect). Along sequences h ν → + ∞ h_\nu\to+\infty h ν → + ∞ with h ν m o d 2 π → θ h_\nu\bmod2\pi\to\theta h ν mod 2 π → θ , the observers in Theorem 4.3 converge to
C + , θ ( x ) = χ ( x + θ ) ( I 2 , − c + R ( x + θ ) ) , c + = 3 , c − = 1. C_{+,\theta}(x)=\chi(x+\theta)(I_2,-c_+R(x+\theta)),\qquad c_+=3,\qquad c_-=1. C + , θ ( x ) = χ ( x + θ ) ( I 2 , − c + R ( x + θ )) , c + = 3 , c − = 1.
These limits admit the nonzero bounded invisible solutions
q 1 ( t , x ) = R ( 2 x − 2 t + θ ) v , q 2 ( t , x ) = c − − 1 R ( x − 2 t ) v . q_1(t,x)=R(2x-2t+\theta)v,\qquad q_2(t,x)=c_-^{-1}R(x-2t)v. q 1 ( t , x ) = R ( 2 x − 2 t + θ ) v , q 2 ( t , x ) = c − − 1 R ( x − 2 t ) v .
These are L l o c 2 L^2_{\mathrm{loc}} L loc 2 profiles, not L 2 L^2 L 2 states. The normalized packets approximate the limiting profile at positive infinity and have frequencies concentrating near { ± 1 , ± 2 } \{\pm1,\pm2\} { ± 1 , ± 2 } : this obstruction is not the zero-frequency degeneracy of a constant Kalman observer. Thus distributional uniqueness for the original medium does not replace uniqueness on the full hull. The two examples share exactly the same localization χ \chi χ and speeds; only the second has an invisible limiting configuration.
Localized observability with arbitrary real Jordan structure
The intrinsic observation space and Jordan coordinates
To apply the abstract criterion to transport systems with Jordan blocks, we first identify the norm measured by whole-line observation and its dual control space. We recall the characterization from [31 ] , with a short proof, and then pass to Jordan coordinates. Zero characteristic speeds are allowed in the whole-line construction.
Assume that A ∈ R n × n A \in\mathbb{R}^{n \times n} A ∈ R n × n has real spectrum and that B ∈ R n × m B \in\mathbb{R}^{n \times m} B ∈ R n × m satisfies
rank ( B , A B , … , A n − 1 B ) = n . \operatorname{rank}(B,AB,\ldots,A^{n-1}B)=n. rank ( B , A B , … , A n − 1 B ) = n .
Set E − 1 = { 0 } E_{-1}=\{0\} E − 1 = { 0 } , E k = Ran ( B , … , A k B ) E_k=\operatorname{Ran}(B,\ldots,A^kB) E k = Ran ( B , … , A k B ) , and K = min { k : E k = R n } K=\min\{k:E_k=\mathbb{R}^n\} K = min { k : E k = R n } . Let Π k A \Pi_k^A Π k A be the orthogonal projector onto E k ∩ E k − 1 ⊥ E_k \cap E_{k-1}^{\perp} E k ∩ E k − 1 ⊥ . For λ ∈ Λ : = σ ( A ) \lambda\in\Lambda:=\sigma(A) λ ∈ Λ := σ ( A ) , let P λ A P_\lambda^A P λ A be the generalized spectral projector of A ⊤ A^\top A ⊤ and m λ m_\lambda m λ the largest Jordan-block size at λ \lambda λ . Define
A λ , j = 1 j ! B ⊤ ( A ⊤ − λ I ) j P λ A , 0 ≤ j < m λ , A_{\lambda,j}=\frac{1}{j!}B^\top(A^\top-\lambda I)^jP_\lambda^A,\qquad0\leq j<m_\lambda, A λ , j = j ! 1 B ⊤ ( A ⊤ − λ I ) j P λ A , 0 ≤ j < m λ ,
and
W K ( ξ ) = ∑ k = 0 K ∣ ξ ∣ 2 k Π k A , W J ( ξ ) = ∑ λ ∈ Λ ∑ j = 0 m λ − 1 ∣ ξ ∣ 2 j A λ , j ∗ A λ , j , W_K(\xi)=\sum_{k=0}^{K}|\xi|^{2k}\Pi_k^A,\qquad W_J(\xi)=\sum_{\lambda\in\Lambda}\sum_{j=0}^{m_\lambda-1}|\xi|^{2j}A_{\lambda,j}^*A_{\lambda,j}, W K ( ξ ) = k = 0 ∑ K ∣ ξ ∣ 2 k Π k A , W J ( ξ ) = λ ∈ Λ ∑ j = 0 ∑ m λ − 1 ∣ ξ ∣ 2 j A λ , j ∗ A λ , j ,
W ( ξ ) = { W K ( ξ ) , 0 < ∣ ξ ∣ ≤ 1 , W J ( ξ ) , ∣ ξ ∣ > 1. (40) W(\xi)=
\begin{cases}
W_K(\xi), & 0<|\xi|\leq1,\\
W_J(\xi), & |\xi|>1.
\end{cases} \tag*{(40)} W ( ξ ) = { W K ( ξ ) , W J ( ξ ) , 0 < ∣ ξ ∣ ≤ 1 , ∣ ξ ∣ > 1. (40)
The Kalman filtration describes the degeneracy at zero, whereas the nilpotent parts determine the growth at high frequencies. Blocks with the same eigenvalue contribute to a single output and are therefore summed before taking its squared norm. The value of W W W at zero is immaterial.
Theorem 5.1 (Whole-line observation). For every T > 0 T>0 T > 0 , the Gramian
G T ( ξ ) = ∫ 0 T e − i s ξ A B B ⊤ e i s ξ A ⊤ d s (41) G_T(\xi)=\int_0^T e^{-is\xi A}BB^\top e^{is\xi A^\top}\,ds \tag*{(41)} G T ( ξ ) = ∫ 0 T e − i s ξ A B B ⊤ e i s ξ A ⊤ d s (41) satisfies
c T W ( ξ ) ≤ G T ( ξ ) ≤ C T W ( ξ ) , ξ ≠ 0 , (42) c_TW(\xi)\leq G_T(\xi)\leq C_TW(\xi),\qquad\xi\ne0, \tag*{(42)} c T W ( ξ ) ≤ G T ( ξ ) ≤ C T W ( ξ ) , ξ = 0 , (42) *for some c T , C T > 0 c_T,C_T>0 c T , C T > 0 . In particular, W ( ξ ) W(\xi) W ( ξ ) is positive definite for ξ ≠ 0 \xi\ne0 ξ = 0 .
Proof. The Kalman condition gives G T ( ξ ) > 0 G_T(\xi)>0 G T ( ξ ) > 0 for ξ ≠ 0 \xi\ne0 ξ = 0 : if B ⊤ e i s ξ A ⊤ z B^\top e^{is\xi A^\top}z B ⊤ e i s ξ A ⊤ z vanishes on [ 0 , T ] [0,T] [ 0 , T ] , its derivatives at s = 0 s=0 s = 0 imply z = 0 z=0 z = 0 . Near zero, set D ξ = ∑ k = 0 K ( i ξ ) k Π k A D_\xi=\sum_{k=0}^{K}(i\xi)^k\Pi_k^A D ξ = ∑ k = 0 K ( i ξ ) k Π k A . The identities B ⊤ ( A ⊤ ) j Π k A = 0 B^\top(A^\top)^j\Pi_k^A=0 B ⊤ ( A ⊤ ) j Π k A = 0 for j < k j<k j < k give
B ⊤ e i s ξ A ⊤ D ξ − 1 = ∑ k = 0 K s k k ! B ⊤ ( A ⊤ ) k Π k A + O T ( ∣ ξ ∣ ) , B^\top e^{is\xi A^\top}D_\xi^{-1}=\sum_{k=0}^{K}\frac{s^k}{k!}B^\top(A^\top)^k\Pi_k^A+O_T(|\xi|), B ⊤ e i s ξ A ⊤ D ξ − 1 = k = 0 ∑ K k ! s k B ⊤ ( A ⊤ ) k Π k A + O T ( ∣ ξ ∣ ) ,
uniformly for s ∈ [ 0 , T ] s\in[0,T] s ∈ [ 0 , T ] . The limiting polynomial map is injective: each coefficient must vanish, and B ⊤ ( A ⊤ ) k B^\top(A^\top)^k B ⊤ ( A ⊤ ) k is injective on E k ∩ E k − 1 ⊥ E_k\cap E_{k-1}^{\perp} E k ∩ E k − 1 ⊥ . Its Gramian is positive definite, so G T ( ξ ) ≃ T D ξ ∗ D ξ = W K ( ξ ) G_T(\xi)\simeq_T D_\xi^*D_\xi=W_K(\xi) G T ( ξ ) ≃ T D ξ ∗ D ξ = W K ( ξ ) near zero.
At high frequencies, the exact expansion is
B ⊤ e i s ξ A ⊤ z = ∑ λ ∈ Λ e i s ξ λ ∑ j = 0 m λ − 1 s j ( i ξ ) j A λ , j z . B^\top e^{is\xi A^\top}z=\sum_{\lambda\in\Lambda}e^{is\xi\lambda}\sum_{j=0}^{m_\lambda-1}s^j(i\xi)^jA_{\lambda,j}z. B ⊤ e i s ξ A ⊤ z = λ ∈ Λ ∑ e i s ξ λ j = 0 ∑ m λ − 1 s j ( i ξ ) j A λ , j z .
The Gram matrix of the functions s j e i s ξ λ s^je^{is\xi\lambda} s j e i s ξ λ has fixed positive definite diagonal blocks ( ∫ 0 T s j + ℓ d s ) j , ℓ \left(\int_0^T s^{j+\ell}\,ds\right)_{j,\ell} ( ∫ 0 T s j + ℓ d s ) j , ℓ . Its off-diagonal blocks are O T ( ∣ ξ ∣ − 1 ) O_T(|\xi|^{-1}) O T ( ∣ ξ ∣ − 1 ) by integration by parts. Applying this matrix estimate to the coefficients ( i ξ ) j A λ , j z (i\xi)^jA_{\lambda,j}z ( i ξ ) j A λ , j z yields G T ( ξ ) ≃ T W J ( ξ ) G_T(\xi)\simeq_T W_J(\xi) G T ( ξ ) ≃ T W J ( ξ ) for large ∣ ξ ∣ |\xi| ∣ ξ ∣ . The expansion and the Kalman condition also show that W J ( ξ ) > 0 W_J(\xi)>0 W J ( ξ ) > 0 for every ξ ≠ 0 \xi\ne0 ξ = 0 . Positivity and compactness on the remaining frequency annuli complete the proof. □
Definition 5.2. The intrinsic observation space Y K , J − \mathcal{Y}_{K,J}^{-} Y K , J − is the completion of S ( R ; C n ) \mathcal{S}(\mathbb{R};\mathbb{C}^n) S ( R ; C n ) for
∥ f ∥ Y K , J − 2 = ∫ R f ^ ( ξ ) ∗ W ( ξ ) f ^ ( ξ ) d ξ . (43) \|f\|_{\mathcal{Y}_{K,J}^{-}}^2=\int_{\mathbb{R}}\widehat{f}(\xi)^*W(\xi)\widehat{f}(\xi)\,d\xi. \tag*{(43)} ∥ f ∥ Y K , J − 2 = ∫ R f ( ξ ) ∗ W ( ξ ) f ( ξ ) d ξ . (43)
It is realized by measurable Fourier fields with W 1 / 2 f ^ ∈ L 2 W^{1/2}\widehat{f}\in L^2 W 1/2 f ∈ L 2 , modulo equality almost everywhere. Its dual Y K , J + \mathcal{Y}_{K,J}^{+} Y K , J + has norm
∥ γ ∥ Y K , J + 2 = ∫ R γ ^ ( ξ ) ∗ W ( ξ ) − 1 γ ^ ( ξ ) d ξ (44) \|\gamma\|_{\mathcal{Y}_{K,J}^{+}}^2=\int_{\mathbb{R}}\widehat{\gamma}(\xi)^*W(\xi)^{-1}\widehat{\gamma}(\xi)\,d\xi \tag*{(44)} ∥ γ ∥ Y K , J + 2 = ∫ R γ ( ξ ) ∗ W ( ξ ) − 1 γ ( ξ ) d ξ (44)
and pairing ⟨ γ , f ⟩ = ∫ R γ ^ ( ξ ) ∗ f ^ ( ξ ) d ξ \langle\gamma,f\rangle=\int_{\mathbb{R}}\widehat{\gamma}(\xi)^*\widehat{f}(\xi)\,d\xi ⟨ γ , f ⟩ = ∫ R γ ( ξ ) ∗ f ( ξ ) d ξ . Here J J J records the Jordan structure; we write Y K , J ± ( A , B ) \mathcal{Y}_{K,J}^{\pm}(A,B) Y K , J ± ( A , B ) when the pair needs to be specified.
Any fixed positive cutoff in (40) gives an equivalent norm. Elements of Y K , J − \mathcal{Y}_{K,J}^{-} Y K , J − need not be componentwise L 2 L^2 L 2 functions, and no canonical distributional realization is assumed. Evolution and observation will be defined by continuity from Schwartz data.
Observation and evolution. For Schwartz data, put U ( s ) f = F − 1 ( e i s ξ A T f ^ ) U(s)f=\mathcal{F}^{-1}(e^{\mathrm{i}s\xi A^{\mathsf{T}}}\widehat{f}) U ( s ) f = F − 1 ( e i s ξ A T f ) . Plancherel and (42) give
c T ∥ f ∥ Y K , J − 2 ≤ ∫ 0 T ∥ B T U ( s ) f ∥ L 2 2 d s ≤ C T ∥ f ∥ Y K , J − 2 . c_T\|f\|_{Y_{K,J}^{-}}^2\le\int_0^T\|B^{\mathsf{T}}U(s)f\|_{L^2}^2\,\mathrm{d}s\le C_T\|f\|_{Y_{K,J}^{-}}^2. c T ∥ f ∥ Y K , J − 2 ≤ ∫ 0 T ∥ B T U ( s ) f ∥ L 2 2 d s ≤ C T ∥ f ∥ Y K , J − 2 .
Applying the lower bound with T = 1 T=1 T = 1 to U ( t ) f U(t)f U ( t ) f uses the observation on [ t , t + 1 ] ⊂ [ − R , R + 1 ] [t,t+1]\subset[-R,R+1] [ t , t + 1 ] ⊂ [ − R , R + 1 ] for ∣ t ∣ ≤ R |t|\le R ∣ t ∣ ≤ R . On the negative half of this interval use the upper bound at − ξ -\xi − ξ and the identity W ( − ξ ) = W ( ξ ) W(-\xi)=W(\xi) W ( − ξ ) = W ( ξ ) . Together the two halves give
sup ∣ t ∣ ≤ R ∥ U ( t ) f ∥ Y K , J − 2 ≤ 2 C R + 1 c 1 ∥ f ∥ Y K , J − 2 . \sup_{|t|\le R}\|U(t)f\|_{Y_{K,J}^{-}}^2\le\frac{2C_{R+1}}{c_1}\|f\|_{Y_{K,J}^{-}}^2. ∣ t ∣ ≤ R sup ∥ U ( t ) f ∥ Y K , J − 2 ≤ c 1 2 C R + 1 ∥ f ∥ Y K , J − 2 .
Density gives a strongly continuous group on Y K , J − Y_{K,J}^{-} Y K , J − . Its transpose S ( t ) = e − t A ∂ x S(t)=e^{-tA\partial_x} S ( t ) = e − t A ∂ x is a strongly continuous group on Y K , J + Y_{K,J}^{+} Y K , J + .
The whole-line upper bound also implies that every norm satisfying ∥ f ∥ ♯ 2 ≲ ∫ 0 T ∥ ( B T U ( s ) f ) ∣ Ω ∥ L 2 ( Ω ) 2 d s \|f\|_{\sharp}^2\lesssim\int_0^T\|(B^{\mathsf{T}}U(s)f)|_{\Omega}\|_{L^2(\Omega)}^2\,\mathrm{d}s ∥ f ∥ ♯ 2 ≲ ∫ 0 T ∥ ( B T U ( s ) f ) ∣ Ω ∥ L 2 ( Ω ) 2 d s on Schwartz data obeys
∥ f ∥ ♯ ≲ ∥ f ∥ Y K , J − . (45) \|f\|_{\sharp}\lesssim\|f\|_{Y_{K,J}^{-}}. \tag*{(45)} ∥ f ∥ ♯ ≲ ∥ f ∥ Y K , J − . (45)
Thus (43) is the strongest norm controlled by whole-line observation, up to equivalence. For a positive definite Hermitian multiplier Ψ \Psi Ψ , the corresponding whole-line lower observation bound holds exactly when Ψ ( ξ ) ≲ W ( ξ ) \Psi(\xi)\lesssim W(\xi) Ψ ( ξ ) ≲ W ( ξ ) almost everywhere.
Localized observation and control. For every measurable Ω \Omega Ω , the map C Ω f = ( B T f ) ∣ Ω C_{\Omega}f=(B^{\mathsf{T}}f)|_{\Omega} C Ω f = ( B T f ) ∣ Ω , initially defined on Schwartz data, extends boundedly from Y K , J − Y_{K,J}^{-} Y K , J − into U : = L 2 ( Ω ; C m ) \mathcal{U}:=L^2(\Omega;\mathbb{C}^m) U := L 2 ( Ω ; C m ) . Indeed, B T = B T Π 0 A = ∑ λ A λ , 0 B^{\mathsf{T}}=B^{\mathsf{T}}\Pi_0^A=\sum_{\lambda}A_{\lambda,0} B T = B T Π 0 A = ∑ λ A λ , 0 implies
∣ B T z ∣ 2 ≤ c B z ∗ W ( ξ ) z , c B = max { ∥ B T ∥ 2 , # Λ } . |B^{\mathsf{T}}z|^2\le c_Bz^*W(\xi)z,\qquad c_B=\max\{\|B^{\mathsf{T}}\|^2,\#\Lambda\}. ∣ B T z ∣ 2 ≤ c B z ∗ W ( ξ ) z , c B = max { ∥ B T ∥ 2 , #Λ } .
Thus spatial restriction is applied to the L 2 L^2 L 2 output B T f B^{\mathsf{T}}f B T f . Let B Ω = C Ω ′ B_{\Omega}=C_{\Omega}' B Ω = C Ω ′ in the preceding duality and set
Φ t u = ∫ 0 t S ( t − s ) B Ω u ( s ) d s , y ( t ) = S ( t ) y 0 + Φ t u . \Phi_tu=\int_0^tS(t-s)B_{\Omega}u(s)\,\mathrm{d}s,\qquad y(t)=S(t)y_0+\Phi_tu. Φ t u = ∫ 0 t S ( t − s ) B Ω u ( s ) d s , y ( t ) = S ( t ) y 0 + Φ t u .
Since Φ T ′ f = C Ω U ( T − ⋅ ) f \Phi_T'f=C_{\Omega}U(T-\cdot)f Φ T ′ f = C Ω U ( T − ⋅ ) f , Hilbert space duality [21 , 32 ] gives Ran Φ T = Y K , J + \operatorname{Ran}\Phi_T=Y_{K,J}^{+} Ran Φ T = Y K , J + if and only if
∫ 0 T ∥ C Ω U ( T − t ) f ∥ U 2 d t ≥ c ∥ f ∥ Y K , J − 2 , f ∈ Y K , J − , (46) \int_0^T\|C_{\Omega}U(T-t)f\|_{\mathcal{U}}^2\,\mathrm{d}t\ge c\|f\|_{Y_{K,J}^{-}}^2,\qquad f\in Y_{K,J}^{-}, \tag*{(46)} ∫ 0 T ∥ C Ω U ( T − t ) f ∥ U 2 d t ≥ c ∥ f ∥ Y K , J − 2 , f ∈ Y K , J − , (46)
for some c > 0 c>0 c > 0 . In that case, for y 0 , y d ∈ Y K , J + y_0,y_d\in Y_{K,J}^{+} y 0 , y d ∈ Y K , J + , the minimum-norm control steering y 0 y_0 y 0 to y d y_d y d satisfies
∥ u min ∥ L 2 ( 0 , T ; U ) 2 ≤ c − 1 ∥ y d − S ( T ) y 0 ∥ Y K , J + 2 . (47) \|u_{\min}\|_{L^2(0,T;\mathcal{U})}^2\le c^{-1}\|y_d-S(T)y_0\|_{Y_{K,J}^{+}}^2. \tag*{(47)} ∥ u m i n ∥ L 2 ( 0 , T ; U ) 2 ≤ c − 1 ∥ y d − S ( T ) y 0 ∥ Y K , J + 2 . (47)
Jordan coordinates. Choose a fixed real change of coordinates P P P such that
F = P − 1 A T P = ⨁ λ ∈ Λ ( λ I + N λ ) , C = B T P , N λ r λ = 0 , r λ = m λ . F=P^{-1}A^{\mathsf{T}}P=\bigoplus_{\lambda\in\Lambda}(\lambda I+N_{\lambda}),\qquad C=B^{\mathsf{T}}P,\qquad N_{\lambda}^{r_{\lambda}}=0,\qquad r_{\lambda}=m_{\lambda}. F = P − 1 A T P = λ ∈ Λ ⨁ ( λ I + N λ ) , C = B T P , N λ r λ = 0 , r λ = m λ .
All Jordan blocks at the same speed are grouped in the coordinate subspace E λ E_{\lambda} E λ . Its orthogonal projector is P λ = P − 1 P λ A P P_{\lambda}=P^{-1}P_{\lambda}^{A}P P λ = P − 1 P λ A P ; put C λ = C ∣ E λ C_{\lambda}=C|_{E_{\lambda}} C λ = C ∣ E λ and extend maps on E λ E_{\lambda} E λ by zero when needed. The Kalman condition and its minimal index are unchanged:
⋂ j = 0 K ker ( C F j ) = { 0 } . (48) \bigcap_{j=0}^{K}\ker(CF^j)=\{0\}. \tag*{(48)} j = 0 ⋂ K ker ( C F j ) = { 0 } . (48)
On each E λ E_{\lambda} E λ , this implies
⋂ j = 0 r λ − 1 ker ( C λ N λ j ) = { 0 } , (49) \bigcap_{j=0}^{r_{\lambda}-1}\ker(C_{\lambda}N_{\lambda}^j)=\{0\}, \tag*{(49)} j = 0 ⋂ r λ − 1 ker ( C λ N λ j ) = { 0 } , (49)
since a vector in this intersection is killed by every C F j CF^j C F j .
The orthogonal Kalman projectors are now computed in these coordinates. Set
N 0 = C n , N k = ⋂ j = 0 k − 1 ker ( C F j ) ( 1 ≤ k ≤ K + 1 ) , \mathcal{N}_0=\mathbb{C}^n,\qquad\mathcal{N}_k=\bigcap_{j=0}^{k-1}\ker(CF^j)\quad(1\le k\le K+1), N 0 = C n , N k = j = 0 ⋂ k − 1 ker ( C F j ) ( 1 ≤ k ≤ K + 1 ) ,
and let Π k \Pi_k Π k project onto N k ∩ N k + 1 ⊥ \mathcal{N}_k\cap\mathcal{N}_{k+1}^{\perp} N k ∩ N k + 1 ⊥ . Then
C F j Π k = 0 ( j < k ) , C F k ∣ Ran Π k is injective . (50) CF^j\Pi_k=0\quad(j<k),\qquad CF^k|_{\operatorname{Ran}\Pi_k}\text{ is injective}. \tag*{(50)} C F j Π k = 0 ( j < k ) , C F k ∣ Ran Π k is injective . (50)
In general, Π k ≠ P − 1 Π k A P \Pi_k\ne P^{-1}\Pi_k^A P Π k = P − 1 Π k A P , because a nonorthogonal change of coordinates does not preserve orthogonal projections. Thus P λ A P_{\lambda}^A P λ A and P λ P_{\lambda} P λ are spectral projectors, whereas Π k A \Pi_k^A Π k A and Π k \Pi_k Π k are orthogonal projectors for the respective Kalman filtrations.
The identity B T e i ξ A T P = C e i ξ F B^{\mathsf{T}}e^{\mathrm{i}\xi A^{\mathsf{T}}}P=Ce^{\mathrm{i}\xi F} B T e i ξ A T P = C e i ξ F and Theorem 5.1 give
∥ P f ∥ Y K , J − ( A , B ) ≍ ∥ f ∥ Y K , J − ( F T , C T ) . (51) \|Pf\|_{\mathcal{Y}_{K,J}^{-}(A,B)}\asymp\|f\|_{\mathcal{Y}_{K,J}^{-}(F^{\mathsf{T}},C^{\mathsf{T}})}. \tag*{(51)} ∥ P f ∥ Y K , J − ( A , B ) ≍ ∥ f ∥ Y K , J − ( F T , C T ) . (51)
The block-coordinate weight is obtained from (40) by replacing Π k A \Pi_k^{A} Π k A with Π k \Pi_k Π k and A λ , j A_{\lambda,j} A λ , j with C λ N N λ j P λ / j ! C_{\lambda}^{N}N_{\lambda}^{j}P_{\lambda}/j! C λ N N λ j P λ / j ! . The fixed factorials may be omitted without changing norm equivalence. We use the notation Y K , J − \mathcal{Y}_{K,J}^{-} Y K , J − in either coordinate system, with (51) understood.
In the construction below assume 0 ∉ Λ 0\notin\Lambda 0 ∈ / Λ and fix T > 0 T>0 T > 0 , with s = T − t s=T-t s = T − t . The case of a zero speed in Theorem 5.9 follows directly from whole-line observation. The remaining task is to place observation in the L 2 L^{2} L 2 framework of the abstract criterion. We construct a rational normalization that compensates for both the Kalman degeneracy at zero and the growth caused by the nilpotent parts at high frequencies. Although its inverse need not have finite propagation, the normalized observation kernel has exponential decay, which supplies the required localization estimates.
High-frequency layers and a rational normalization
The high-frequency decomposition used here is an increasing filtration of each generalized eigenspace. It differs from the low-frequency Kalman filtration, and the two families of projectors need not commute.
Lemma 5.3. For 0 ≤ j ≤ r λ 0\leq j\leq r_{\lambda} 0 ≤ j ≤ r λ , put
F λ , j = ⋂ ℓ = j r λ − 1 ker ( C λ N λ ℓ ) , F λ , r λ = E λ , \mathcal{F}_{\lambda,j}=\bigcap_{\ell=j}^{r_{\lambda}-1}\ker(C_{\lambda}N_{\lambda}^{\ell}),\qquad\mathcal{F}_{\lambda,r_{\lambda}}=E_{\lambda}, F λ , j = ℓ = j ⋂ r λ − 1 ker ( C λ N λ ℓ ) , F λ , r λ = E λ , and let H λ , j H_{\lambda,j} H λ , j be the orthogonal projector onto
E λ , j = F λ , j + 1 ∩ F λ , j ⊥ , 0 ≤ j < r λ . E_{\lambda,j}=\mathcal{F}_{\lambda,j+1}\cap\mathcal{F}_{\lambda,j}^{\perp},\qquad0\leq j<r_{\lambda}. E λ , j = F λ , j + 1 ∩ F λ , j ⊥ , 0 ≤ j < r λ . These projectors give an orthogonal decomposition of C n \mathbb{C}^{n} C n . Moreover,
C λ N λ ℓ H λ , j = 0 ( ℓ > j ) , C λ N λ j ∣ E λ , j is injective . (52) C_{\lambda}N_{\lambda}^{\ell}H_{\lambda,j}=0\quad(\ell>j),\qquad C_{\lambda}N_{\lambda}^{j}\big|_{E_{\lambda,j}}\text{ is injective}. \tag*{(52)} C λ N λ ℓ H λ , j = 0 ( ℓ > j ) , C λ N λ j E λ , j is injective . (52) There are constants c , C > 0 c,C>0 c , C > 0 such that, for ∣ ξ ∣ ≥ 1 |\xi|\geq1 ∣ ξ ∣ ≥ 1 ,
c ∑ λ , j ∣ ξ ∣ 2 j ∣ H λ , j z ∣ 2 ≤ ∑ λ , ℓ ∣ ξ ∣ 2 ℓ ∣ C λ N λ ℓ P λ z ∣ 2 ≤ C ∑ λ , j ∣ ξ ∣ 2 j ∣ H λ , j z ∣ 2 . (53) c\sum_{\lambda,j}|\xi|^{2j}|H_{\lambda,j}z|^{2}\leq\sum_{\lambda,\ell}|\xi|^{2\ell}|C_{\lambda}N_{\lambda}^{\ell}P_{\lambda}z|^{2}\leq C\sum_{\lambda,j}|\xi|^{2j}|H_{\lambda,j}z|^{2}. \tag*{(53)} c λ , j ∑ ∣ ξ ∣ 2 j ∣ H λ , j z ∣ 2 ≤ λ , ℓ ∑ ∣ ξ ∣ 2 ℓ ∣ C λ N λ ℓ P λ z ∣ 2 ≤ C λ , j ∑ ∣ ξ ∣ 2 j ∣ H λ , j z ∣ 2 . (53) Proof. The subspaces F λ , j \mathcal{F}_{\lambda,j} F λ , j increase with j j j , and F λ , 0 = { 0 } \mathcal{F}_{\lambda,0}=\{0\} F λ , 0 = { 0 } by (49) . This proves the orthogonal decomposition. The vanishing in (52) follows from E λ , j ⊂ F λ , j + 1 E_{\lambda,j}\subset\mathcal{F}_{\lambda,j+1} E λ , j ⊂ F λ , j + 1 . If a vector in E λ , j E_{\lambda,j} E λ , j is also killed by C λ N λ j C_{\lambda}N_{\lambda}^{j} C λ N λ j , it belongs to F λ , j \mathcal{F}_{\lambda,j} F λ , j and must be zero. For completeness, set L R = ∑ λ , j R − j H λ , j L_R=\sum_{\lambda,j}R^{-j}H_{\lambda,j} L R = ∑ λ , j R − j H λ , j , R ≥ 1 R\geq1 R ≥ 1 , and consider the stacked map whose ( λ , ℓ ) (\lambda,\ell) ( λ , ℓ ) component is
R ℓ C λ N λ ℓ P λ L R z = ∑ j ≥ ℓ R ℓ − j C λ N λ ℓ H λ , j z . R^{\ell}C_{\lambda}N_{\lambda}^{\ell}P_{\lambda}L_Rz=\sum_{j\geq\ell}R^{\ell-j}C_{\lambda}N_{\lambda}^{\ell}H_{\lambda,j}z. R ℓ C λ N λ ℓ P λ L R z = j ≥ ℓ ∑ R ℓ − j C λ N λ ℓ H λ , j z . As R → ∞ R\to\infty R → ∞ , this map converges in matrix norm to the map with components C λ N λ ℓ H λ , ℓ z C_{\lambda}N_{\lambda}^{\ell}H_{\lambda,\ell}z C λ N λ ℓ H λ , ℓ z . The latter is injective, by (52) and the orthogonal decomposition. Thus the stacked map is bounded above and below uniformly for large R R R . On a remaining compact interval 1 ≤ R ≤ R 0 1\leq R\leq R_0 1 ≤ R ≤ R 0 , the same conclusion follows by continuity and (49) . Substituting z = L R − 1 w z=L_R^{-1}w z = L R − 1 w proves (53) .
Choose an integer M ≥ 1 M\geq1 M ≥ 1 such that 2 M + 1 > K 2M+1>K 2 M + 1 > K , and define
φ ( ξ ) = ξ 2 M + 1 ( 1 + ξ 2 ) M , H ( ξ ) = ∑ λ , j ( 1 + i φ ( ξ ) ) j H λ , j , D ( ξ ) = ∑ k = 0 K ( i ξ 1 + i ξ ) k Π k . (54) \varphi(\xi)=\frac{\xi^{2M+1}}{(1+\xi^{2})^{M}},\qquad H(\xi)=\sum_{\lambda,j}(1+\mathrm{i}\varphi(\xi))^{j}H_{\lambda,j},\qquad D(\xi)=\sum_{k=0}^{K}\left(\frac{\mathrm{i}\xi}{1+\mathrm{i}\xi}\right)^{k}\Pi_k. \tag*{(54)} φ ( ξ ) = ( 1 + ξ 2 ) M ξ 2 M + 1 , H ( ξ ) = λ , j ∑ ( 1 + i φ ( ξ ) ) j H λ , j , D ( ξ ) = k = 0 ∑ K ( 1 + i ξ i ξ ) k Π k . (54)
Our normalization is
Q ( ξ ) = D ( ξ ) H ( ξ ) . (55) Q(\xi)=D(\xi)H(\xi). \tag*{(55)} Q ( ξ ) = D ( ξ ) H ( ξ ) . (55)
The order of the factors in (55) matters: the two systems of projectors need not commute. Each factor is harmless where the other one acts: H = I + O ( ∣ ξ ∣ 2 M + 1 ) H=I+O(|\xi|^{2M+1}) H = I + O ( ∣ ξ ∣ 2 M + 1 ) near zero, whereas D D D and D − 1 D^{-1} D − 1 are bounded for ∣ ξ ∣ ≥ 1 |\xi|\geq1 ∣ ξ ∣ ≥ 1 . In the proof of Lemma 5.4 this gives D H D − 1 = I + O ( ∣ ξ ∣ 2 M + 1 − K ) DHD^{-1}=I+O(|\xi|^{2M+1-K}) D H D − 1 = I + O ( ∣ ξ ∣ 2 M + 1 − K ) near zero and ∣ D H z ∣ ≍ ∣ H z ∣ |DH z|\asymp|H z| ∣ D H z ∣ ≍ ∣ H z ∣ at infinity. With the reverse order, the same argument would require H ( D − D − 1 ) H − 1 → 0 H(D-D^{-1})H^{-1}\to0 H ( D − D − 1 ) H − 1 → 0 at infinity. When the Kalman projectors mix the high-frequency layers, this term is only O ( ∣ ξ ∣ r − 2 ) O(|\xi|^{r-2}) O ( ∣ ξ ∣ r − 2 ) , with r = max λ r λ r=\max_{\lambda}r_{\lambda} r = max λ r λ , and it need not tend to zero. For instance, let λ ≠ μ \lambda\neq\mu λ = μ be nonzero reals and take
F = ( λ 1 0 0 λ 0 0 0 μ ) , C = ( 1 , 0 , 1 ) . F=\begin{pmatrix}\lambda&1&0\\0&\lambda&0\\0&0&\mu\end{pmatrix},\qquad C=(1,0,1). F = λ 0 0 1 λ 0 0 0 μ , C = ( 1 , 0 , 1 ) .
Then K = 2 K=2 K = 2 . Ran Π 2 \operatorname{Ran}\Pi_2 Ran Π 2 is spanned by ( 1 , μ − λ , − 1 ) T (1,\mu-\lambda,-1)^{\mathsf{T}} ( 1 , μ − λ , − 1 ) T , and H ( ξ ) = diag ( 1 , 1 + i φ ( ξ ) , 1 ) H(\xi)=\operatorname{diag}(1,1+\mathrm{i}\varphi(\xi),1) H ( ξ ) = diag ( 1 , 1 + i φ ( ξ ) , 1 ) . Thus H ( ξ ) H(\xi) H ( ξ ) and Π 2 \Pi_2 Π 2 do not commute for ξ ≠ 0 \xi\neq0 ξ = 0 , and the ( 2 , 1 ) (2,1) ( 2 , 1 ) entry of H ( D − D − 1 ) H − 1 H(D-D^{-1})H^{-1} H ( D − D − 1 ) H − 1 tends to − ( μ − λ ) / ( 2 + ( μ − λ ) 2 ) ≠ 0 -(\mu-\lambda)/(2+(\mu-\lambda)^2)\neq0 − ( μ − λ ) / ( 2 + ( μ − λ ) 2 ) = 0 . This example shows why the small-perturbation argument cannot simply be reversed; noncommutation alone does not disprove norm equivalence for H D HD H D .
Lemma 5.4. The matrix Q ( ξ ) Q(\xi) Q ( ξ ) is invertible for real ξ ≠ 0 \xi\ne0 ξ = 0 , is smooth on the real axis, and has rational entries. For Schwartz data,
∥ Q ( D x ) f ∥ L 2 ≃ ∥ f ∥ Y K , J − . (56) \lVert Q(D_x)f\rVert_{L^2} \simeq\lVert f\rVert_{\mathcal{Y}_{K,J}^{-}}. \tag*{(56)} ∥ Q ( D x ) f ∥ L 2 ≃ ∥ f ∥ Y K , J − . (56) Consequently Q ( D x ) Q(D_x) Q ( D x ) extends to an isomorphism from Y K , J − \mathcal{Y}_{K,J}^{-} Y K , J − onto L 2 ( R ; C n ) L^2(\mathbb{R};\mathbb{C}^n) L 2 ( R ; C n ) .
Proof. For real ξ \xi ξ , the number 1 + i φ ( ξ ) 1+i\varphi(\xi) 1 + i φ ( ξ ) does not vanish. All the projectors in H H H are orthogonal, so H H H and H − 1 H^{-1} H − 1 have rational entries without real poles. Also
H ( ξ ) = I + O ( ∣ ξ ∣ 2 M + 1 ) ( ξ → 0 ) , φ ( ξ ) = ξ + O ( ∣ ξ ∣ − 1 ) ( ∣ ξ ∣ → ∞ ) . H(\xi)=I+O(|\xi|^{2M+1})\quad(\xi\to0),\qquad\varphi(\xi)=\xi+O(|\xi|^{-1})\quad(|\xi|\to\infty). H ( ξ ) = I + O ( ∣ ξ ∣ 2 M + 1 ) ( ξ → 0 ) , φ ( ξ ) = ξ + O ( ∣ ξ ∣ − 1 ) ( ∣ ξ ∣ → ∞ ) .
Near zero, ∥ D ( ξ ) ∥ \lVert D(\xi)\rVert ∥ D ( ξ )∥ is bounded and ∥ D ( ξ ) − 1 ∥ ≤ C ∣ ξ ∣ − K \lVert D(\xi)^{-1}\rVert\le C|\xi|^{-K} ∥ D ( ξ ) − 1 ∥ ≤ C ∣ ξ ∣ − K . Hence
D ( ξ ) H ( ξ ) D ( ξ ) − 1 = I + O ( ∣ ξ ∣ 2 M + 1 − K ) . D(\xi)H(\xi)D(\xi)^{-1}=I+O(|\xi|^{2M+1-K}). D ( ξ ) H ( ξ ) D ( ξ ) − 1 = I + O ( ∣ ξ ∣ 2 M + 1 − K ) .
It follows that ∣ Q ( ξ ) z ∣ ≃ ∣ D ( ξ ) z ∣ \lvert Q(\xi)z\rvert\simeq\lvert D(\xi)z\rvert ∣ Q ( ξ ) z ∣ ≃ ∣ D ( ξ ) z ∣ near zero, with constants uniform in z z z , and the latter norm is uniformly equivalent to the low-frequency part of the Y K , J − \mathcal{Y}_{K,J}^{-} Y K , J − norm. On ∣ ξ ∣ ≥ 1 |\xi|\ge1 ∣ ξ ∣ ≥ 1 , both D D D and D − 1 D^{-1} D − 1 are uniformly bounded. Orthogonality of the high-frequency layers gives
∣ H ( ξ ) z ∣ 2 = ∑ λ , j ( 1 + φ ( ξ ) 2 ) j ∣ H λ , j z ∣ 2 ≃ ∑ λ , j ∣ ξ ∣ 2 j ∣ H λ , j z ∣ 2 . \lvert H(\xi)z\rvert^2=\sum_{\lambda,j}(1+\varphi(\xi)^2)^{j}\lvert H_{\lambda,j}z\rvert^2\simeq\sum_{\lambda,j}|\xi|^{2j}\lvert H_{\lambda,j}z\rvert^2. ∣ H ( ξ ) z ∣ 2 = λ , j ∑ ( 1 + φ ( ξ ) 2 ) j ∣ H λ , j z ∣ 2 ≃ λ , j ∑ ∣ ξ ∣ 2 j ∣ H λ , j z ∣ 2 .
Lemma 5.3 gives the required high-frequency equivalence. Continuity and invertibility handle any compact annulus between the two estimates. This proves (56) by Plancherel. Finally, the range of Q ( D x ) Q(D_x) Q ( D x ) on Schwartz functions is dense in L 2 L^2 L 2 : it contains every function whose Fourier transform is smooth, compactly supported, and supported away from zero. Completion proves the last assertion. □ \square □
The normalized observation kernel
For a Schwartz datum f f f , set u = Q ( D x ) f u=Q(D_x)f u = Q ( D x ) f . In the normalized variable u u u , the observation C e s F ∂ x f Ce^{sF\partial_x}f C e s F ∂ x f is represented by the multiplier
M s ( ξ ) = C e i s ξ F Q ( ξ ) − 1 = C e i s ξ F H ( ξ ) − 1 D ( ξ ) − 1 , 0 ≤ s ≤ T , ξ ≠ 0. (57) M_s(\xi)=Ce^{is\xi F}Q(\xi)^{-1}=Ce^{is\xi F}H(\xi)^{-1}D(\xi)^{-1},\qquad0\le s\le T,\quad\xi\ne0. \tag*{(57)} M s ( ξ ) = C e i s ξ F Q ( ξ ) − 1 = C e i s ξ F H ( ξ ) − 1 D ( ξ ) − 1 , 0 ≤ s ≤ T , ξ = 0. (57)
We will show that M s ( D x ) M_s(D_x) M s ( D x ) is a finite sum of characteristic translations and a convolution operator with an exponentially decaying kernel. This provides the representation needed to apply Proposition 2.6 to the weighted observations in the next subsection.
To fix the normalization of convolution kernels, write
( F 0 k ) ( ξ ) : = ∫ R e − i x ξ k ( x ) d x = 2 π k ^ ( ξ ) . (\mathcal{F}_0k)(\xi):=\int_{\mathbb{R}}e^{-ix\xi}k(x)\,dx=\sqrt{2\pi}\,\widehat{k}(\xi). ( F 0 k ) ( ξ ) := ∫ R e − i x ξ k ( x ) d x = 2 π k ( ξ ) .
Thus k ∗ u = ( F 0 k ) ( D x ) u k*u=(\mathcal{F}_0k)(D_x)u k ∗ u = ( F 0 k ) ( D x ) u , and the symbol e i s λ ξ e^{is\lambda\xi} e i s λ ξ acts as the translation u ( x ) ↦ u ( x + λ s ) u(x)\mapsto u(x+\lambda s) u ( x ) ↦ u ( x + λ s ) . Two elementary identities illustrate the kernel contributions below:
F 0 ( 1 ( − s , 0 ) ) ( ξ ) = e i s ξ − 1 i ξ , F 0 ( e − x 1 ( 0 , ∞ ) ( x ) ) ( ξ ) = 1 1 + i ξ . \mathcal{F}_0(\mathbf{1}_{(-s,0)})(\xi)=\frac{e^{is\xi}-1}{i\xi},\qquad\mathcal{F}_0(e^{-x}\mathbf{1}_{(0,\infty)}(x))(\xi)=\frac{1}{1+i\xi}. F 0 ( 1 ( − s , 0 ) ) ( ξ ) = i ξ e i s ξ − 1 , F 0 ( e − x 1 ( 0 , ∞ ) ( x )) ( ξ ) = 1 + i ξ 1 .
In the first identity, the quotient extends to the value s s s at zero: the cancellation produces a kernel supported on a bounded interval. In the second, the nonreal pole produces an exponentially decaying kernel.
Lemma 5.5. The singularity of M s M_s M s at zero is removable, uniformly for s ∈ [ 0 , T ] s\in[0,T] s ∈ [ 0 , T ] , and
M s ( 0 ) = ∑ k = 0 K s k k ! C F k Π k . (58) M_s(0)=\sum_{k=0}^{K}\frac{s^k}{k!}CF^k\Pi_k. \tag*{(58)} M s ( 0 ) = k = 0 ∑ K k ! s k C F k Π k . (58) Define
V λ ( s ) = ∑ j = 0 r λ − 1 s j j ! C Λ λ j H λ , j . (59) V_\lambda(s)=\sum_{j=0}^{r_\lambda-1}\frac{s^j}{j!}C\Lambda_\lambda^jH_{\lambda,j}. \tag*{(59)} V λ ( s ) = j = 0 ∑ r λ − 1 j ! s j C Λ λ j H λ , j . (59) There are measurable matrix kernels k s : R → C m × n k_s:\mathbb{R}\to\mathbb{C}^{m\times n} k s : R → C m × n and constants c , C T > 0 c,C_T>0 c , C T > 0 such that, for every Schwartz input u u u ,
M s ( D x ) u ( x ) = ∑ λ ∈ Λ V λ ( s ) u ( x + λ s ) + ∫ R k s ( x − y ) u ( y ) d y , ∣ k s ( x ) ∣ ≤ C T e − c ∣ x ∣ . (60) M_s(D_x)u(x)=\sum_{\lambda\in\Lambda}V_\lambda(s)u(x+\lambda s)+\int_{\mathbb{R}}k_s(x-y)u(y)\,dy,\qquad|k_s(x)|\le C_Te^{-c|x|}. \tag*{(60)} M s ( D x ) u ( x ) = λ ∈ Λ ∑ V λ ( s ) u ( x + λ s ) + ∫ R k s ( x − y ) u ( y ) d y , ∣ k s ( x ) ∣ ≤ C T e − c ∣ x ∣ . (60) The constants are independent of s ∈ [ 0 , T ] s\in[0,T] s ∈ [ 0 , T ] . This representation defines uniformly bounded operators on both L 2 L^2 L 2 and L u l o c 2 L^2_{\mathrm{uloc}} L uloc 2 .
Proof. Step 1: removal of the apparent singularity at zero. Put p = 2 M + 1 > K p=2M+1>K p = 2 M + 1 > K . By (50) ,
C e i s ξ Π k = ( i ξ ) k ( s k k ! C F k Π k + ξ R s , k ( ξ ) ) . Ce^{\mathrm{i}s\xi}\Pi_k=(\mathrm{i}\xi)^k\left(\frac{s^k}{k!}CF^k\Pi_k+\xi R_{s,k}(\xi)\right). C e i s ξ Π k = ( i ξ ) k ( k ! s k C F k Π k + ξ R s , k ( ξ ) ) .
where R s , k R_{s,k} R s , k is analytic near zero, with uniform bounds for s ∈ [ 0 , T ] s\in[0,T] s ∈ [ 0 , T ] . Since
D ( ξ ) − 1 = ∑ k = 0 K ( 1 + i ξ i ξ ) k Π k D(\xi)^{-1}=\sum_{k=0}^{K}\left(\frac{1+\mathrm{i}\xi}{\mathrm{i}\xi}\right)^k\Pi_k D ( ξ ) − 1 = k = 0 ∑ K ( i ξ 1 + i ξ ) k Π k
these factors ( i ξ ) k (\mathrm{i}\xi)^k ( i ξ ) k cancel the poles of D − 1 D^{-1} D − 1 . Hence C e i s ξ F D ( ξ ) − 1 Ce^{\mathrm{i}s\xi F}D(\xi)^{-1} C e i s ξ F D ( ξ ) − 1 is analytic at zero and its value there is the right-hand side of (58) .
The additional factor H − 1 H^{-1} H − 1 does not change this value. Indeed, H ( ξ ) − 1 − I = ξ p R H ( ξ ) H(\xi)^{-1}-I=\xi^pR_H(\xi) H ( ξ ) − 1 − I = ξ p R H ( ξ ) with R H R_H R H analytic near zero, whereas D − 1 D^{-1} D − 1 has a pole of order at most K K K . Therefore
C e i s ξ F ( H ( ξ ) − 1 − I ) D ( ξ ) − 1 = ξ p − K R ~ s ( ξ ) , Ce^{\mathrm{i}s\xi F}\bigl(H(\xi)^{-1}-I\bigr)D(\xi)^{-1}=\xi^{p-K}\widetilde R_s(\xi), C e i s ξ F ( H ( ξ ) − 1 − I ) D ( ξ ) − 1 = ξ p − K R s ( ξ ) ,
with R ~ s \widetilde R_s R s analytic near zero. Since p > K p>K p > K , this term vanishes at zero. Thus the full symbol M s M_s M s extends analytically across zero, with all local bounds uniform for s ∈ [ 0 , T ] s\in[0,T] s ∈ [ 0 , T ] .
Step 2: extraction of the characteristic translations. Nilpotency and (52) give
M s ( ξ ) = ∑ λ ∈ Λ e i s λ ξ R λ , s ( ξ ) , R λ , s ( ξ ) = [ ∑ j = 0 r λ − 1 ∑ ℓ = 0 j s ℓ ( i ξ ) ℓ ℓ ! ( 1 + i φ ( ξ ) ) − j C λ N λ ℓ H λ , j ] D ( ξ ) − 1 . (61) M_s(\xi)=\sum_{\lambda\in\Lambda}e^{\mathrm{i}s\lambda\xi}R_{\lambda,s}(\xi),\qquad R_{\lambda,s}(\xi)=\left[\sum_{j=0}^{r_\lambda-1}\sum_{\ell=0}^{j}\frac{s^\ell(\mathrm{i}\xi)^\ell}{\ell!}(1+\mathrm{i}\varphi(\xi))^{-j}C_\lambda N_\lambda^\ell H_{\lambda,j}\right]D(\xi)^{-1}. \tag*{(61)} M s ( ξ ) = λ ∈ Λ ∑ e i s λ ξ R λ , s ( ξ ) , R λ , s ( ξ ) = [ j = 0 ∑ r λ − 1 ℓ = 0 ∑ j ℓ ! s ℓ ( i ξ ) ℓ ( 1 + i φ ( ξ ) ) − j C λ N λ ℓ H λ , j ] D ( ξ ) − 1 . (61)
Since φ ( ξ ) = ξ + O ( ∣ ξ ∣ − 1 ) \varphi(\xi)=\xi+O(|\xi|^{-1}) φ ( ξ ) = ξ + O ( ∣ ξ ∣ − 1 ) and D ( ξ ) − 1 = I + O ( ∣ ξ ∣ − 1 ) D(\xi)^{-1}=I+O(|\xi|^{-1}) D ( ξ ) − 1 = I + O ( ∣ ξ ∣ − 1 ) , only the terms with ℓ = j \ell=j ℓ = j contribute to the limit at infinity. Consequently,
R λ , s ( ξ ) = V λ ( s ) + O ( ∣ ξ ∣ − 1 ) ( ∣ ξ ∣ → ∞ ) , R_{\lambda,s}(\xi)=V_\lambda(s)+O(|\xi|^{-1})\qquad(|\xi|\to\infty), R λ , s ( ξ ) = V λ ( s ) + O ( ∣ ξ ∣ − 1 ) ( ∣ ξ ∣ → ∞ ) ,
uniformly in s s s . The terms e i s λ ξ V λ ( s ) e^{\mathrm{i}s\lambda\xi}V_\lambda(s) e i s λ ξ V λ ( s ) give the translations in (60) . The remainders R λ , s − V λ ( s ) R_{\lambda,s}-V_\lambda(s) R λ , s − V λ ( s ) are rational and vanish at infinity. Their poles lie in a fixed finite set consisting of zero and the nonreal poles of H − 1 H^{-1} H − 1 ; the pole at zero has order at most K K K . Their partial-fraction coefficients depend continuously on s s s and are uniformly bounded on [ 0 , T ] [0,T] [ 0 , T ] .
Step 3: cancellation of the poles at zero. For K ≥ 1 K\ge1 K ≥ 1 , group these contributions across all characteristic speeds:
Z s ( ξ ) = ∑ λ ∈ Λ e i s λ ξ ∑ d = 1 K Z λ , d ( s ) ( i ξ ) d . Z_s(\xi)=\sum_{\lambda\in\Lambda}e^{\mathrm{i}s\lambda\xi}\sum_{d=1}^{K}\frac{Z_{\lambda,d}(s)}{(\mathrm{i}\xi)^d}. Z s ( ξ ) = λ ∈ Λ ∑ e i s λ ξ d = 1 ∑ K ( i ξ ) d Z λ , d ( s ) .
All other terms are analytic at zero, so Step 1 implies that Z s Z_s Z s is analytic there as well. This cancellation must be used before identifying its convolution kernel.
Define
z s ( x ) = ∑ λ ∈ Λ ∑ d = 1 K Z λ , d ( s ) ( x + λ s ) + d − 1 ( d − 1 ) ! , z_s(x)=\sum_{\lambda\in\Lambda}\sum_{d=1}^{K}Z_{\lambda,d}(s)\frac{(x+\lambda s)_+^{d-1}}{(d-1)!}, z s ( x ) = λ ∈ Λ ∑ d = 1 ∑ K Z λ , d ( s ) ( d − 1 )! ( x + λ s ) + d − 1 ,
where x + 0 = 1 ( 0 , ∞ ) ( x ) x_+^0=\mathbf{1}_{(0,\infty)}(x) x + 0 = 1 ( 0 , ∞ ) ( x ) . This function vanishes to the left of all points − λ s -\lambda s − λ s . To their right, it is a polynomial whose coefficient of x m x^m x m , 0 ≤ m ≤ K − 1 0\le m\le K-1 0 ≤ m ≤ K − 1 , is
1 m ! ∑ λ ∈ Λ ∑ d = m + 1 K Z λ , d ( s ) ( λ s ) d − 1 − m ( d − 1 − m ) ! . \frac{1}{m!}\sum_{\lambda\in\Lambda}\sum_{d=m+1}^{K}Z_{\lambda,d}(s)\frac{(\lambda s)^{d-1-m}}{(d-1-m)!}. m ! 1 λ ∈ Λ ∑ d = m + 1 ∑ K Z λ , d ( s ) ( d − 1 − m )! ( λ s ) d − 1 − m .
This is 1 / m ! 1/m! 1/ m ! times the coefficient of ( i ξ ) − m − 1 (\mathrm{i}\xi)^{-m-1} ( i ξ ) − m − 1 in the Laurent expansion of Z s Z_s Z s , hence it vanishes. Thus z s z_s z s is supported in the convex hull of { − λ s : λ ∈ Λ } \{-\lambda s:\lambda\in\Lambda\} { − λ s : λ ∈ Λ } , and both its support and its size are uniformly bounded for s ∈ [ 0 , T ] s\in[0,T] s ∈ [ 0 , T ] .
To verify its Fourier symbol, differentiate in distributions:
∂ x K z s = ∑ λ ∈ Λ ∑ d = 1 K Z λ , d ( s ) ∂ x K − d δ − λ s , ( i ξ ) K ( F 0 z s ) ( ξ ) = ( i ξ ) K Z s ( ξ ) . \partial_x^K z_s=\sum_{\lambda\in\Lambda}\sum_{d=1}^{K}Z_{\lambda,d}(s)\partial_x^{K-d}\delta_{-\lambda s},\qquad
(\mathrm{i}\xi)^K(\mathcal{F}_0z_s)(\xi)=(\mathrm{i}\xi)^KZ_s(\xi). ∂ x K z s = λ ∈ Λ ∑ d = 1 ∑ K Z λ , d ( s ) ∂ x K − d δ − λ s , ( i ξ ) K ( F 0 z s ) ( ξ ) = ( i ξ ) K Z s ( ξ ) .
Hence F 0 z s = Z s \mathcal{F}_0z_s=Z_s F 0 z s = Z s for ξ ≠ 0 \xi\ne0 ξ = 0 and, by continuity, also at zero. This identifies the contribution of the poles at zero with a uniformly bounded, compactly supported kernel. For K = 0 K=0 K = 0 , no poles at zero occur and we set Z s = z s = 0 Z_s=z_s=0 Z s = z s = 0 .
Step 4: exponential decay of the remaining kernels. For a nonreal pole ζ \zeta ζ and an integer ν ≥ 1 \nu\ge1 ν ≥ 1 , the kernel with symbol ( ξ − ζ ) − ν (\xi-\zeta)^{-\nu} ( ξ − ζ ) − ν is
κ ζ , ν ( x ) = { i ν 1 ( 0 , ∞ ) ( x ) x ν − 1 ( ν − 1 ) ! e i ζ x , Im ζ > 0 , ( − i ) ν 1 ( − ∞ , 0 ) ( x ) ( − x ) ν − 1 ( ν − 1 ) ! e i ζ x , Im ζ < 0. \kappa_{\zeta,\nu}(x)=
\begin{cases}
\mathrm{i}^{\nu}\mathbf{1}_{(0,\infty)}(x)\dfrac{x^{\nu-1}}{(\nu-1)!}e^{\mathrm{i}\zeta x}, & \operatorname{Im}\zeta>0,\\
(-\mathrm{i})^{\nu}\mathbf{1}_{(-\infty,0)}(x)\dfrac{(-x)^{\nu-1}}{(\nu-1)!}e^{\mathrm{i}\zeta x}, & \operatorname{Im}\zeta<0.
\end{cases} κ ζ , ν ( x ) = ⎩ ⎨ ⎧ i ν 1 ( 0 , ∞ ) ( x ) ( ν − 1 )! x ν − 1 e i ζ x , ( − i ) ν 1 ( − ∞ , 0 ) ( x ) ( ν − 1 )! ( − x ) ν − 1 e i ζ x , Im ζ > 0 , Im ζ < 0.
The identity F 0 κ ζ , ν = ( ξ − ζ ) − ν \mathcal{F}_{0\kappa_{\zeta,\nu}}=(\xi-\zeta)^{-\nu} F 0 κ ζ , ν = ( ξ − ζ ) − ν follows from the Laplace integral. Each kernel is a polynomial times a decaying exponential on one half-line. Multiplication by e i s λ ξ e^{\mathrm{i}s\lambda\xi} e i s λ ξ replaces it by κ ζ , ν ( x + s λ ) \kappa_{\zeta,\nu}(x+s\lambda) κ ζ , ν ( x + s λ ) . There are finitely many poles, their distances from the real axis are positive, and their orders are fixed. The coefficients and the shifts are uniformly bounded for s ∈ [ 0 , T ] s\in[0,T] s ∈ [ 0 , T ] . Decreasing the exponential rate absorbs all polynomial factors. Adding these kernels to z s z_s z s gives k s k_s k s and proves (60) , including the uniform exponential bound.
Finally, sup s ∈ [ 0 , T ] ∥ k s ∥ L 1 < ∞ \sup_{s\in[0,T]}\|k_s\|_{L^1}<\infty sup s ∈ [ 0 , T ] ∥ k s ∥ L 1 < ∞ , so Young’s inequality gives the L 2 L^2 L 2 bound. Translations are uniformly bounded on L u l o c 2 L^2_{\mathrm{uloc}} L uloc 2 , since a translated unit interval meets at most two unit cells. Minkowski’s inequality then bounds convolution by a constant times ∥ k s ∥ L 1 \|k_s\|_{L^1} ∥ k s ∥ L 1 on this space. Together with the uniform bounds on V λ ( s ) V_\lambda(s) V λ ( s ) , this proves the last assertion.
Characteristic occupation and the Gramian
Let
0 ≤ a ≤ 1 , a ∈ W 1 , ∞ ( R ) , inf λ ∈ Λ , y ∈ R ∫ 0 T a ( y − λ s ) d s ≥ μ > 0. (62) 0\leq a\leq1,\qquad a\in W^{1,\infty}(\mathbb{R}),\qquad\inf_{\lambda\in\Lambda,\ y\in\mathbb{R}}\int_0^T a(y-\lambda s)\,\mathrm{d}s\geq\mu>0. \tag*{(62)} 0 ≤ a ≤ 1 , a ∈ W 1 , ∞ ( R ) , λ ∈ Λ , y ∈ R inf ∫ 0 T a ( y − λ s ) d s ≥ μ > 0. (62)
The occupation integrals are continuous, so essential infima give the same condition. The translation hull H ( a ) \mathscr{H}(a) H ( a ) , taken in the locally uniform topology, is compact and translation invariant; all its elements retain the same occupation and Lipschitz bounds. For b ∈ H ( a ) b\in\mathscr{H}(a) b ∈ H ( a ) , set
( Γ b u ) ( s , x ) = b ( x ) M s ( D x ) u ( x ) , G b = Γ b ∗ Γ b . (\Gamma_bu)(s,x)=\sqrt{b(x)}M_s(D_x)u(x),\qquad G_b=\Gamma_b^*\Gamma_b. ( Γ b u ) ( s , x ) = b ( x ) M s ( D x ) u ( x ) , G b = Γ b ∗ Γ b .
By Lemma 5.5 , these observations are uniformly bounded on L 2 L^2 L 2 and also act locally on L u l o c 2 L^2_{\mathrm{uloc}} L uloc 2 .
Lemma 5.6. The family Γ b \Gamma_b Γ b , b ∈ H ( a ) b\in\mathscr{H}(a) b ∈ H ( a ) , satisfies (H1)–(H5). Its Gramian decomposes as
G b = D b + K b , D b ( y ) = ∑ λ ∈ Λ ∫ 0 T b ( y − λ s ) V λ ( s ) ∗ V λ ( s ) d s . (63) G_b=D_b+K_b,\qquad D_b(y)=\sum_{\lambda\in\Lambda}\int_0^T b(y-\lambda s)V_\lambda(s)^*V_\lambda(s)\,\mathrm{d}s. \tag*{(63)} G b = D b + K b , D b ( y ) = λ ∈ Λ ∑ ∫ 0 T b ( y − λ s ) V λ ( s ) ∗ V λ ( s ) d s . (63) The matrices D b D_b D b are uniformly Lipschitz and satisfy
d 0 I ≤ D b ( y ) ≤ d 1 I (64) d_0I\leq D_b(y)\leq d_1I \tag*{(64)} d 0 I ≤ D b ( y ) ≤ d 1 I (64) for constants d 0 , d 1 > 0 d_0,d_1>0 d 0 , d 1 > 0 independent of b b b and y y y . The remainder has a matrix kernel with the uniform bound
∣ K b ( x , y ) ∣ ≤ C T e − c ∣ x − y ∣ . (65) \lvert K_b(x,y)\rvert\leq C_Te^{-c\lvert x-y\rvert}. \tag*{(65)} ∣ K b ( x , y )∣ ≤ C T e − c ∣ x − y ∣ . (65)
Proof. We first establish coercivity of the matrix in (63) . Each V λ V_\lambda V λ acts on the domain block E λ E_\lambda E λ , so D b D_b D b is block diagonal. For z ∈ E λ z\in E_\lambda z ∈ E λ , the polynomial formula
V λ ( s ) z = ∑ j = 0 r λ − 1 s j j ! C λ N λ j H λ , j z V_\lambda(s)z=\sum_{j=0}^{r_\lambda-1}\frac{s^j}{j!}C_\lambda N_\lambda^jH_{\lambda,j}z V λ ( s ) z = j = 0 ∑ r λ − 1 j ! s j C λ N λ j H λ , j z
and Lemma A.1 , applied to b ( y − λ s ) b(y-\lambda s) b ( y − λ s ) , give
∫ 0 T b ( y − λ s ) ∣ V λ ( s ) z ∣ 2 d s ≥ c 0 ∑ j ∣ C λ N λ j H λ , j z ∣ 2 ≥ c 1 ∣ z ∣ 2 , \int_0^T b(y-\lambda s)\lvert V_\lambda(s)z\rvert^2\,\mathrm{d}s\geq c_0\sum_j\lvert C_\lambda N_\lambda^jH_{\lambda,j}z\rvert^2\geq c_1\lvert z\rvert^2, ∫ 0 T b ( y − λ s ) ∣ V λ ( s ) z ∣ 2 d s ≥ c 0 j ∑ ∣ C λ N λ j H λ , j z ∣ 2 ≥ c 1 ∣ z ∣ 2 ,
where the last inequality follows from (52) . The constants are uniform in b b b and y y y , and there are finitely many speeds. This proves the lower bound in (64) ; the upper and Lipschitz bounds follow directly from the formula for D b D_b D b .
Now apply Proposition 2.6 with
c λ , b ( s , x ) = b ( x ) V λ ( s ) , k b ( s , x , y ) = b ( x ) k s ( x − y ) . c_{\lambda,b}(s,x)=\sqrt{b(x)}V_\lambda(s),\qquad k_b(s,x,y)=\sqrt{b(x)}k_s(x-y). c λ , b ( s , x ) = b ( x ) V λ ( s ) , k b ( s , x , y ) = b ( x ) k s ( x − y ) .
Lemma 5.5 supplies the representation and uniform bounds. Locally uniform convergence b n → b b_n\to b b n → b gives coefficient continuity and local L 2 L^2 L 2 kernel continuity. Translation covariance follows from convolution, and the principal matrix is precisely D b D_b D b . Proposition 2.6 therefore gives (H1)–(H5). In particular, its proof yields
∥ 1 I ( K b n − K b ) 1 J ∥ H S ⟶ 0 for all bounded I , J . \|\mathbf{1}_I(K_{b_n}-K_b)\mathbf{1}_J\|_{\mathrm{HS}}\longrightarrow0\qquad\text{for all bounded }I,J. ∥ 1 I ( K b n − K b ) 1 J ∥ HS ⟶ 0 for all bounded I , J .
Compactness of H ( a ) \mathscr{H}(a) H ( a ) then gives collective compactness uniformly over the hull. The bound (65) supplies localization; it is not, by itself, a proof of collective compactness. For (65) , the cross kernel for λ ≠ ν \lambda\neq\nu λ = ν is bounded by ∥ V λ ∥ ∞ ∥ V ν ∥ ∞ / ∣ λ − ν ∣ \|V_\lambda\|_\infty\|V_\nu\|_\infty/|\lambda-\nu| ∥ V λ ∥ ∞ ∥ V ν ∥ ∞ /∣ λ − ν ∣ and vanishes when ∣ x − y ∣ > T ∣ λ − ν ∣ |x-y|>T|\lambda-\nu| ∣ x − y ∣ > T ∣ λ − ν ∣ . The remaining kernels contain the exponentially decaying remainder from Lemma 5.5 .
Generalized uniqueness on the hull
The uniqueness argument is stated for normalized data in L u l o c 2 L^2_{\mathrm{uloc}} L uloc 2 . It therefore excludes both translated concentration and spreading sequences. No assertion about a nonzero global L 2 L^2 L 2 profile is needed.
Let O M ( R ) O_M(\mathbb{R}) O M ( R ) denote the smooth functions whose derivatives grow at most polynomially. For a matrix symbol m m m with entries in O M O_M O M , m ( D x ) m(D_x) m ( D x ) maps S \mathcal{S} S to S \mathcal{S} S and S ′ \mathcal{S}' S ′ to S ′ \mathcal{S}' S ′ , and ( m 1 m 2 ) ( D x ) = m 1 ( D x ) m 2 ( D x ) (m_1m_2)(D_x)=m_1(D_x)m_2(D_x) ( m 1 m 2 ) ( D x ) = m 1 ( D x ) m 2 ( D x ) on S ′ \mathcal{S}' S ′ . The next lemma makes explicit how the normalized datum u ∈ L u l o c 2 u\in L^2_{\mathrm{uloc}} u ∈ L uloc 2 is related to a solution of the original transport system. The lift need not be unique, but its observation is independent of the choice. Since Q − 1 Q^{-1} Q − 1 may have a pole at zero, the lift cannot be defined by multiplying an arbitrary tempered distribution by Q − 1 Q^{-1} Q − 1 .
Lemma 5.7 (Distributional lift and observation identity). (i) The symbols D D D , H H H , H − 1 H^{-1} H − 1 , Q Q Q , C e i s ξ F Ce^{is\xi F} C e i s ξ F , M s M_s M s and F 0 k s \mathcal{F}_0 k_s F 0 k s have entries in O M O_M O M , with bounds uniform in s ∈ [ 0 , T ] s\in[0,T] s ∈ [ 0 , T ] . For K ≥ 1 K\ge1 K ≥ 1 , the symbols D − 1 D^{-1} D − 1 and Q − 1 Q^{-1} Q − 1 have a pole of order K K K at ξ = 0 \xi=0 ξ = 0 and are smooth elsewhere.
(ii) For every u ∈ L u l o c 2 ( R ; C n ) u\in L^2_{\mathrm{uloc}}(\mathbb{R};\mathbb{C}^n) u ∈ L uloc 2 ( R ; C n ) there is f ∈ S ′ ( R ; C n ) f\in\mathcal{S}'(\mathbb{R};\mathbb{C}^n) f ∈ S ′ ( R ; C n ) with Q ( D x ) f = u Q(D_x)f=u Q ( D x ) f = u .
(iii) The kernel of Q ( D x ) Q(D_x) Q ( D x ) on S ′ \mathcal{S}' S ′ is
N Q = { ∑ k = 1 K g k : g k is a polynomial of degree at most k − 1 with values in Ran Π k } , \mathcal{N}_Q=\left\{\sum_{k=1}^{K}g_k:\ g_k\text{ is a polynomial of degree at most }k-1\text{ with values in }\operatorname{Ran}\Pi_k\right\}, N Q = { k = 1 ∑ K g k : g k is a polynomial of degree at most k − 1 with values in Ran Π k } , and C e s F ∂ x g = 0 Ce^{sF\partial_x}g=0 C e s F ∂ x g = 0 for every g ∈ N Q g\in\mathcal{N}_Q g ∈ N Q and every s ∈ R s\in\mathbb{R} s ∈ R . In particular, the lift in (ii) is unique exactly when K = 0 K=0 K = 0 .
(iv) For every lift f f f as in (ii) and every s ∈ [ 0 , T ] s\in[0,T] s ∈ [ 0 , T ] ,
C e s F ∂ x f = M s ( D x ) u in S ′ . Ce^{sF\partial_x}f=M_s(D_x)u\quad\text{in }\mathcal{S}'. C e s F ∂ x f = M s ( D x ) u in S ′ . The right-hand side is given by the absolutely convergent formula (60) , and on each cell it coincides with the uniformly local extension of Definition 2.9. The map s ↦ C e s F ∂ x f s\mapsto Ce^{sF\partial_x}f s ↦ C e s F ∂ x f defines a distribution on ( 0 , T ) × R (0,T)\times\mathbb{R} ( 0 , T ) × R , equal to the locally square integrable function y ( s , x ) = M s ( D x ) u ( x ) y(s,x)=M_s(D_x)u(x) y ( s , x ) = M s ( D x ) u ( x ) . In particular Γ b u = b y \Gamma_bu=\sqrt{b}\,y Γ b u = b y , with Γ b \Gamma_b Γ b understood as in Theorem 2.1 . If u = Q ( D x ) P − 1 q T u=Q(D_x)P^{-1}q_T u = Q ( D x ) P − 1 q T with q T ∈ S q_T\in\mathcal{S} q T ∈ S , then y ( s , ⋅ ) = B ⊤ q ( T − s , ⋅ ) y(s,\cdot)=B^\top q(T-s,\cdot) y ( s , ⋅ ) = B ⊤ q ( T − s , ⋅ ) .
Proof. (i) The entries of D D D and H − 1 H^{-1} H − 1 are rational functions without real poles that are bounded at infinity, and those of H H H have polynomial growth. The entries of C e i s ξ F Ce^{is\xi F} C e i s ξ F are exponentials e i s λ ξ e^{is\lambda\xi} e i s λ ξ times polynomials in s ξ s\xi s ξ . By Lemma 5.5 , M s M_s M s is analytic at zero and equals ∑ λ e i s λ ξ R λ , s ( ξ ) \sum_{\lambda}e^{is\lambda\xi}R_{\lambda,s}(\xi) ∑ λ e i s λ ξ R λ , s ( ξ ) , where each R λ , s R_{\lambda,s} R λ , s is rational and bounded at infinity; all derivatives of M s M_s M s are therefore bounded on { ∣ ξ ∣ ≥ 1 } \{|\xi|\ge1\} { ∣ ξ ∣ ≥ 1 } . Since k s k_s k s decays exponentially, x j k s ∈ L 1 x^j k_s\in L^1 x j k s ∈ L 1 for every j j j , so F 0 k s \mathcal{F}_0k_s F 0 k s is smooth with bounded derivatives. The last assertion follows from the formula for D − 1 D^{-1} D − 1 in the proof of Lemma 5.5 and from H ( 0 ) = I H(0)=I H ( 0 ) = I .
(ii) For k ≥ 1 k\ge1 k ≥ 1 , define
w k ( x ) = 1 ( k − 1 ) ! ∫ 0 x ( x − t ) k − 1 Π k u ( t ) d t , w_k(x)=\frac{1}{(k-1)!}\int_0^x(x-t)^{k-1}\Pi_k u(t)\,dt, w k ( x ) = ( k − 1 )! 1 ∫ 0 x ( x − t ) k − 1 Π k u ( t ) d t ,
with the oriented integral for x < 0 x<0 x < 0 , and set
v k = ( 1 + ∂ x ) k w k , v 0 = Π 0 u , v = ∑ k = 0 K v k . v_k=(1+\partial_x)^k w_k,\qquad v_0=\Pi_0u,\qquad v=\sum_{k=0}^{K}v_k. v k = ( 1 + ∂ x ) k w k , v 0 = Π 0 u , v = k = 0 ∑ K v k .
The uniformly local bound gives ∣ w k ( x ) ∣ ≤ C k ( 1 + ∣ x ∣ k ) ∥ u ∥ L u l o c 2 |w_k(x)|\le C_k(1+|x|^k)\|u\|_{L^2_{\mathrm{uloc}}} ∣ w k ( x ) ∣ ≤ C k ( 1 + ∣ x ∣ k ) ∥ u ∥ L uloc 2 , so all these distributions are tempered. Since ∂ x k w k = Π k u \partial_x^k w_k=\Pi_k u ∂ x k w k = Π k u and ( 1 + ∂ x ) − k (1+\partial_x)^{-k} ( 1 + ∂ x ) − k is a regular multiplier on S ′ \mathcal{S}' S ′ , we have
D ( D x ) v k = ∂ x k ( 1 + ∂ x ) − k v k = Π k u ( k ≥ 1 ) , D ( D x ) v 0 = Π 0 u . D(D_x)v_k=\partial_x^k(1+\partial_x)^{-k}v_k=\Pi_k u\quad(k\ge1),\qquad D(D_x)v_0=\Pi_0u. D ( D x ) v k = ∂ x k ( 1 + ∂ x ) − k v k = Π k u ( k ≥ 1 ) , D ( D x ) v 0 = Π 0 u .
Thus D ( D x ) v = u D(D_x)v=u D ( D x ) v = u . By (i), H − 1 H^{-1} H − 1 acts on S ′ \mathcal{S}' S ′ , and f = H ( D x ) − 1 v f=H(D_x)^{-1}v f = H ( D x ) − 1 v satisfies Q ( D x ) f = u Q(D_x)f=u Q ( D x ) f = u .
(iii) Let Q ( D x ) g = 0 Q(D_x)g=0 Q ( D x ) g = 0 and set h = H ( D x ) g ∈ S ′ h=H(D_x)g\in\mathcal{S}' h = H ( D x ) g ∈ S ′ . Projecting D ( D x ) h = 0 D(D_x)h=0 D ( D x ) h = 0 onto each Kalman layer and applying ( 1 + ∂ x ) k (1+\partial_x)^k ( 1 + ∂ x ) k gives
Π 0 h = 0 , ∂ x k Π k h = 0 ( 1 ≤ k ≤ K ) . \Pi_0h=0,\qquad\partial_x^k\Pi_kh=0\quad(1\le k\le K). Π 0 h = 0 , ∂ x k Π k h = 0 ( 1 ≤ k ≤ K ) .
Hence Π k h \Pi_kh Π k h is a polynomial of degree at most k − 1 k-1 k − 1 . The symbol H − 1 − I H^{-1}-I H − 1 − I vanishes to order 2 M + 1 > K 2M+1>K 2 M + 1 > K at zero, so its multiplier annihilates these polynomials. Thus g = H ( D x ) − 1 h = h g=H(D_x)^{-1}h=h g = H ( D x ) − 1 h = h . Conversely, every such polynomial sum lies in ker Q ( D x ) \ker Q(D_x) ker Q ( D x ) , proving ker Q ( D x ) = N Q \ker Q(D_x)=\mathcal{N}_Q ker Q ( D x ) = N Q . Finally,
C e s F ∂ x g k = ∑ j ≥ 0 s j j ! C F Π k ∂ x j g k . Ce^{sF\partial_x}g_k=\sum_{j\ge0}\frac{s^j}{j!}CF\Pi_k\partial_x^jg_k. C e s F ∂ x g k = j ≥ 0 ∑ j ! s j C F Π k ∂ x j g k .
The terms with j < k j<k j < k vanish by (50) , and those with j ≥ k j\ge k j ≥ k vanish because deg g k ≤ k − 1 \deg g_k\le k-1 deg g k ≤ k − 1 .
(iv) The removable singularity established in Lemma 5.5 is important here. The identity M s ( ξ ) Q ( ξ ) = C e i s ξ F M_s(\xi)Q(\xi)=Ce^{is\xi F} M s ( ξ ) Q ( ξ ) = C e i s ξ F holds for ξ ≠ 0 \xi\ne0 ξ = 0 , and both sides are continuous, so it holds on the whole real axis. By (i), it is an identity of multipliers on S ′ \mathcal{S}' S ′ , including distributions whose Fourier transform is supported at ξ = 0 \xi=0 ξ = 0 . Applying it to f f f gives C e s F ∂ x f = M s ( D x ) u Ce^{sF\partial_x}f=M_s(D_x)u C e s F ∂ x f = M s ( D x ) u . Two lifts differ by an element of N Q \mathcal{N}_Q N Q , which is invisible by (iii), so the observed distribution does not depend on the choice of lift.
For u ∈ L u l o c 2 u\in L^2_{\mathrm{uloc}} u ∈ L uloc 2 the right-hand side of (60) is an absolutely convergent integral. By Fubini’s theorem on each cell it coincides with the uniformly local extension of Definition 2.9. It also coincides with the action of the multiplier M s M_s M s on S ′ \mathcal{S}' S ′ : since k s k_s k s decays exponentially, it is a rapidly decreasing distribution, and k s ∗ u = ( F 0 k s ) ( D x ) u k_s * u = (\mathcal{F}_0 k_s)(D_x)u k s ∗ u = ( F 0 k s ) ( D x ) u in S ′ \mathcal{S}' S ′ . The map s ↦ e s F ∂ x f s \mapsto e^{sF\partial_x}f s ↦ e s F ∂ x f is smooth from R \mathbb{R} R into S ′ \mathcal{S}' S ′ , since ∂ s \partial_s ∂ s acts through the symbol i ξ F i\xi F i ξ F , so it defines a distribution on ( 0 , T ) × R (0,T) \times\mathbb{R} ( 0 , T ) × R . The kernel representation makes y y y jointly measurable and locally square integrable in space-time, and the two distributions coincide. Hence Γ b u = b y \Gamma_b u = \sqrt{b}\,y Γ b u = b y : the weight multiplies the locally square integrable observation, not the possibly distributional lift. The last assertion follows from C = B ⊤ P C = B^\top P C = B ⊤ P and e s F ∂ x P − 1 = P − 1 e s A ⊤ ∂ x e^{sF\partial_x}P^{-1} = P^{-1}e^{sA^\top\partial_x} e s F ∂ x P − 1 = P − 1 e s A ⊤ ∂ x .
The analyticity of M s M_s M s at the origin is used in (iv) on distributions whose Fourier transform is supported at ξ = 0 \xi= 0 ξ = 0 , such as constants, which belong to L u l o c 2 L^2_{\mathrm{uloc}} L uloc 2 . This is where the choice 2 M + 1 > K 2M + 1 > K 2 M + 1 > K enters: the correction H − 1 − I H^{-1} - I H − 1 − I vanishes to order 2 M + 1 2M + 1 2 M + 1 , which exceeds the order K K K of the pole of D − 1 D^{-1} D − 1 . For a constant u 0 u_0 u 0 one has M s ( D x ) u 0 = M s ( 0 ) u 0 M_s(D_x)u_0 = M_s(0)u_0 M s ( D x ) u 0 = M s ( 0 ) u 0 , which is used at the end of the proof of Lemma 5.8.
Lemma 5.8. For every b ∈ S ( a ) b \in\mathscr{S}(a) b ∈ S ( a ) ,
u ∈ L u l o c 2 , Γ b u = 0 ⟹ u = 0. (66) u \in L^2_{\mathrm{uloc}},\quad\Gamma_b u = 0 \quad\Longrightarrow\quad u = 0. \tag*{(66)} u ∈ L uloc 2 , Γ b u = 0 ⟹ u = 0. (66)
Proof. Let f ∈ S ′ ( R ; C n ) f \in\mathcal{S}'(\mathbb{R};\mathbb{C}^n) f ∈ S ′ ( R ; C n ) be a lift of u u u given by Lemma 5.7 (ii), and put
q ( s ) = e s F ∂ x f , y ( s ) = C q ( s ) . q(s) = e^{sF\partial_x}f,\qquad y(s) = Cq(s). q ( s ) = e s F ∂ x f , y ( s ) = C q ( s ) .
By Lemma 5.7 (iv), y ( s ) = M s ( D x ) u y(s) = M_s(D_x)u y ( s ) = M s ( D x ) u , the function y y y is locally square integrable in space-time, and Γ b u = b y \Gamma_b u = \sqrt{b}\,y Γ b u = b y . The choice of the lift does not affect y y y ; we do not claim that f f f itself is uniquely determined by u u u .
Let O b = { x : b ( x ) > 0 } O_b = \{x : b(x) > 0\} O b = { x : b ( x ) > 0 } and U b = ( 0 , T ) × O b U_b = (0,T) \times O_b U b = ( 0 , T ) × O b . Since Γ b u = b y = 0 \Gamma_b u = \sqrt{b}\,y = 0 Γ b u = b y = 0 , restriction to compact subsets where b b b is bounded below by a positive number gives y = 0 y = 0 y = 0 in D ′ ( U b ) \mathcal{D}'(U_b) D ′ ( U b ) .
The equation is ∂ s q = F ∂ x q \partial_s q = F\partial_x q ∂ s q = F ∂ x q . For 0 ≤ j ≤ K 0 \le j \le K 0 ≤ j ≤ K ,
C F j ∂ x K q = ∂ s j ∂ x K − j y = 0 in D ′ ( U b ) . CF^j\partial_x^K q = \partial_s^j\partial_x^{K-j}y = 0 \quad\text{in }\mathcal{D}'(U_b). C F j ∂ x K q = ∂ s j ∂ x K − j y = 0 in D ′ ( U b ) .
The finite-dimensional injectivity (48) therefore gives
∂ x K q = 0 in D ′ ( U b ) . (67) \partial_x^K q = 0 \quad\text{in }\mathcal{D}'(U_b). \tag*{(67)} ∂ x K q = 0 in D ′ ( U b ) . (67)
For the generalized eigenspace component,
q λ ( s , x ) = e s N λ ∂ x f λ ( x + λ s ) , f λ = P λ f . q_\lambda(s,x) = e^{sN_\lambda\partial_x}f_\lambda(x+\lambda s),\qquad f_\lambda= P_\lambda f. q λ ( s , x ) = e s N λ ∂ x f λ ( x + λ s ) , f λ = P λ f .
The inverse nilpotent exponential is a finite differential operator. Hence
∂ x K f λ ( x + λ s ) = ∑ j = 0 r λ − 1 ( − s ) j j ! N λ j ∂ x K + j q λ ( s , x ) = 0 in D ′ ( U b ) , \partial_x^K f_\lambda(x+\lambda s) = \sum_{j=0}^{r_\lambda-1}\frac{(-s)^j}{j!}N_\lambda^j\partial_x^{K+j}q_\lambda(s,x)=0\quad\text{in }\mathcal{D}'(U_b), ∂ x K f λ ( x + λ s ) = j = 0 ∑ r λ − 1 j ! ( − s ) j N λ j ∂ x K + j q λ ( s , x ) = 0 in D ′ ( U b ) ,
where the second equality follows by spatial differentiation of (67) .
The map ( s , x ) ↦ x + λ s (s,x) \mapsto x+\lambda s ( s , x ) ↦ x + λ s sends U b U_b U b onto the whole real line. Indeed, (62) guarantees, for every y y y , a time s ∈ ( 0 , T ) s \in(0,T) s ∈ ( 0 , T ) with b ( y − λ s ) > 0 b(y-\lambda s)>0 b ( y − λ s ) > 0 . Lemma B.1 , applied componentwise, therefore gives ∂ x K f λ = 0 \partial_x^K f_\lambda= 0 ∂ x K f λ = 0 on R \mathbb{R} R . For K = 0 K=0 K = 0 this already says f = 0 f=0 f = 0 . Otherwise each f λ f_\lambda f λ , and therefore f f f , is a polynomial distribution of degree at most K − 1 K-1 K − 1 .
The multiplier Q Q Q is smooth at zero, so u = Q ( D x ) f u = Q(D_x)f u = Q ( D x ) f is also a polynomial. A polynomial in L u l o c 2 ( R ) L^2_{\mathrm{uloc}}(\mathbb{R}) L uloc 2 ( R ) is constant; write u = u 0 u=u_0 u = u 0 . Lemma 5.5 now gives
M s ( D x ) u 0 = M s ( 0 ) u 0 = ∑ k = 0 K s k k ! C F k Π k u 0 . M_s(D_x)u_0 = M_s(0)u_0 = \sum_{k=0}^{K}\frac{s^k}{k!}CF^k\Pi_k u_0. M s ( D x ) u 0 = M s ( 0 ) u 0 = k = 0 ∑ K k ! s k C F k Π k u 0 .
The weight b b b is not identically zero. Since Γ b u 0 = 0 \Gamma_bu_0=0 Γ b u 0 = 0 , this polynomial in s s s is zero on ( 0 , T ) (0,T) ( 0 , T ) . Every coefficient C F k Π k u 0 CF^k\Pi_k u_0 C F k Π k u 0 is zero. Injectivity in (50) gives Π k u 0 = 0 \Pi_k u_0=0 Π k u 0 = 0 for every k k k , and hence u 0 = 0 u_0=0 u 0 = 0 . This proves (66) .
Localized observability and bounded-gap sets
Theorem 5.9. Assume the Kalman condition (48) and σ ( A ) ⊂ R \sigma(A) \subset\mathbb{R} σ ( A ) ⊂ R . Let T > 0 T>0 T > 0 and let a : R → [ 0 , 1 ] a:\mathbb{R}\to[0,1] a : R → [ 0 , 1 ] be Lipschitz. There exist c T , C T > 0 c_T,C_T>0 c T , C T > 0 such that every terminal datum q T ∈ Y K , γ q_T\in\mathcal{Y}_{K,\gamma} q T ∈ Y K , γ and its adjoint solution satisfy
c T ∥ q T ∥ Y K , γ 2 ≤ ∫ 0 T ∫ R a ( x ) ∣ B ⊤ q ( t , x ) ∣ 2 d x d t ≤ C T ∥ q T ∥ Y K , γ 2 . (68) c_T\lVert q_T\rVert_{\mathcal{Y}_{K,\gamma}}^2 \le\int_0^T\int_{\mathbb{R}}a(x)\lvert B^\top q(t,x)\rvert^2\,dx\,dt \le C_T\lVert q_T\rVert_{\mathcal{Y}_{K,\gamma}}^2. \tag*{(68)} c T ∥ q T ∥ Y K , γ 2 ≤ ∫ 0 T ∫ R a ( x ) ∣ B ⊤ q ( t , x ) ∣ 2 d x d t ≤ C T ∥ q T ∥ Y K , γ 2 . (68) if and only if
min λ ∈ σ ( A ) inf y ∈ R m λ , a ( y ) > 0 , m λ , a ( y ) = ∫ 0 T a ( y − λ s ) d s . \min_{\lambda\in\sigma(A)}\inf_{y\in\mathbb{R}}m_{\lambda,a}(y)>0,\qquad m_{\lambda,a}(y)=\int_0^T a(y-\lambda s)\,ds. λ ∈ σ ( A ) min y ∈ R inf m λ , a ( y ) > 0 , m λ , a ( y ) = ∫ 0 T a ( y − λ s ) d s . If 0 ∈ σ ( A ) 0\in\sigma(A) 0 ∈ σ ( A ) , this condition is equivalent to inf R a > 0 \inf_{\mathbb{R}}a>0 inf R a > 0 . The upper bound holds for every such weight. For fixed A , B , T A,B,T A , B , T , the two bounds are uniform on the hull S ( a ) \mathscr{S}(a) S ( a ) whenever the condition holds. More generally they are uniform on any translation-invariant family of [ 0 , 1 ] [0,1] [ 0 , 1 ] -valued weights that is compact for local uniform convergence and has a common Lipschitz bound and a common positive lower bound on the occupation integrals.
Proof. The upper bound follows from a ≤ 1 a \le1 a ≤ 1 , Plancherel’s theorem and Theorem 5.1 . For necessity, fix a nonzero eigenvector A T v = λ v A^{\mathsf{T}}v = \lambda v A T v = λ v . Kalman’s condition gives B T v ≠ 0 B^{\mathsf{T}}v \ne0 B T v = 0 . For q T = v ϕ q_T = v\phi q T = v ϕ , ϕ ∈ C ∞ ( R ) \phi\in C^\infty(\mathbb{R}) ϕ ∈ C ∞ ( R ) , the observed energy is ∣ B T v ∣ 2 ∫ m λ , a ( y ) ∣ ϕ ( y ) ∣ 2 d y \lvert B^{\mathsf{T}}v\rvert^2 \int m_{\lambda,a}(y)\lvert\phi(y)\rvert^2\,dy ∣ B T v ∣ 2 ∫ m λ , a ( y ) ∣ ϕ ( y ) ∣ 2 d y . The whole-line upper bound gives
T ∣ B T v ∣ 2 ∥ ϕ ∥ 2 2 ≤ C T g l o b ∥ v ϕ ∥ Y K , J − 2 . T\lvert B^{\mathsf{T}}v\rvert^2\lVert\phi\rVert_2^2 \le C_T^{\mathrm{glob}}\lVert v\phi\rVert_{\mathcal{Y}_{K,J}^{-}}^2. T ∣ B T v ∣ 2 ∥ ϕ ∥ 2 2 ≤ C T glob ∥ v ϕ ∥ Y K , J − 2 .
Thus the lower bound in (68) implies m λ , a ≥ c T T / C T g l o b m_{\lambda,a} \ge c_TT/C_T^{\mathrm{glob}} m λ , a ≥ c T T / C T glob almost everywhere and, by continuity, everywhere.
For sufficiency, let μ > 0 \mu> 0 μ > 0 be a common lower bound for the occupation integrals. If 0 ∈ σ ( A ) 0 \in\sigma(A) 0 ∈ σ ( A ) , then m 0 , a = T a m_{0,a} = Ta m 0 , a = T a and a ≥ μ / T a \ge\mu/T a ≥ μ / T . Conversely, if inf a > 0 \inf a > 0 inf a > 0 yields m λ , a ≥ T inf a m_{\lambda,a} \ge T\inf a m λ , a ≥ T inf a at every speed. In this case the lower estimate follows from Theorem 5.1 with the factor μ / T \mu/T μ / T .
Now suppose 0 ∉ σ ( A ) 0 \notin\sigma(A) 0 ∈ / σ ( A ) . For Schwartz data put u = Q ( D x ) P − 1 q T u = Q(D_x)P^{-1}q_T u = Q ( D x ) P − 1 q T . Lemma 5.4 and (51) give ∥ u ∥ 2 ≍ ∥ q T ∥ Y K , J − \lVert u\rVert_2 \asymp\lVert q_T\rVert_{\mathcal{Y}_{K,J}^{-}} ∥ u ∥ 2 ≍ ∥ q T ∥ Y K , J − . For b ∈ S ( a ) b \in\mathfrak{S}(a) b ∈ S ( a ) , Lemma 5.7 (iv) shows that the normalized observation satisfies
( Γ b u ) ( s , x ) = b ( x ) B T q ( T − s , x ) . (\Gamma_bu)(s,x) = \sqrt{b(x)}B^{\mathsf{T}}q(T-s,x). ( Γ b u ) ( s , x ) = b ( x ) B T q ( T − s , x ) .
Lemma 5.6 verifies the hypotheses on the observation family, and Lemma 5.8 proves uniqueness in L u l o c 2 L^2_{\mathrm{uloc}} L uloc 2 for every hull element. Theorem 2.1 therefore gives ∥ Γ b u ∥ Z ≥ c ∗ ∥ u ∥ 2 \lVert\Gamma_bu\rVert_Z \ge c_*\lVert u\rVert_2 ∥ Γ b u ∥ Z ≥ c ∗ ∥ u ∥ 2 with c ∗ > 0 c_* > 0 c ∗ > 0 independent of b b b . This proves the lower bound on Schwartz data. The upper bound makes the observation continuous for the intrinsic norm, so both estimates extend by density to Y K , J − \mathcal{Y}_{K,J}^{-} Y K , J − . For a compact family as in the statement the same two lemmas apply throughout the family, and Corollary 2.2 gives a common lower bound.
Corollary 5.10. Assume (48) and σ ( A ) ⊂ R ∖ { 0 } \sigma(A) \subset\mathbb{R} \setminus\{0\} σ ( A ) ⊂ R ∖ { 0 } . Let Ω \Omega Ω be as in (4) and define T ∙ T_\bullet T ∙ by (5) . For T > 0 T > 0 T > 0 , there is a constant such that
∥ q T ∥ Y K , J − 2 ≤ C T , ℓ ∙ , G ∙ , A , B ∫ 0 T ∫ Ω ∣ B T q ( t , x ) ∣ 2 d x d t (69) \lVert q_T\rVert_{\mathcal{Y}_{K,J}^{-}}^2 \le C_{T,\ell_\bullet,G^\bullet,A,B}\int_0^T\int_\Omega\lvert B^{\mathsf{T}}q(t,x)\rvert^2\,dx\,dt \tag*{(69)} ∥ q T ∥ Y K , J − 2 ≤ C T , ℓ ∙ , G ∙ , A , B ∫ 0 T ∫ Ω ∣ B T q ( t , x ) ∣ 2 d x d t (69)
for every q T ∈ Y K , J − q_T \in\mathcal{Y}_{K,J}^{-} q T ∈ Y K , J − if and only if T > T ∙ T > T_\bullet T > T ∙ . When Ω = R \Omega= \mathbb{R} Ω = R the constant depends only on T , A , B T,A,B T , A , B .
Proof. For Ω = R \Omega= \mathbb{R} Ω = R apply Theorem 5.1 . Otherwise suppose T > T ∙ T > T_\bullet T > T ∙ and apply Lemma 3.2 with the characteristic-speed set Λ \Lambda Λ and the parameters ℓ ∙ , G ∙ , T \ell_\bullet,G^\bullet,T ℓ ∙ , G ∙ , T . It provides a common δ > 0 \delta> 0 δ > 0 , a constant m 0 > 0 m_0 > 0 m 0 > 0 , and weights a Ω ≤ 1 Ω a_\Omega\le1_\Omega a Ω ≤ 1 Ω such that
Lip ( a Ω ) ≤ δ − 1 , ∫ 0 T a Ω ( y − λ s ) d s ≥ m 0 . \operatorname{Lip}(a_\Omega) \le\delta^{-1}, \qquad\int_0^T a_\Omega(y-\lambda s)\,ds \ge m_0. Lip ( a Ω ) ≤ δ − 1 , ∫ 0 T a Ω ( y − λ s ) d s ≥ m 0 .
These weights belong to the compact translation-invariant family A \mathfrak{A} A defined in (38) . The uniform version of Theorem 5.9 therefore applies. Since a Ω ≤ 1 Ω a_\Omega\le1_\Omega a Ω ≤ 1 Ω , its weighted observation is bounded above by the observation on Ω \Omega Ω . This proves (69) . For 0 < T ≤ T ∙ 0 < T \le T_\bullet 0 < T ≤ T ∙ , failure follows from Proposition C.1 , including equality when no gap attains G ∙ G^\bullet G ∙ .
Corollary 5.11. Under the hypotheses of Corollary 5.10, system (1) is exactly controllable in Y K , J + \mathcal{Y}_{K,J}^{+} Y K , J + at time T > 0 T > 0 T > 0 if and only if T > T ∙ T > T_\bullet T > T ∙ . In that case, for y 0 , y d ∈ Y K , J + y_0,y_d \in\mathcal{Y}_{K,J}^{+} y 0 , y d ∈ Y K , J + , the minimum-norm control satisfies
∥ u min ∥ L 2 ( ( 0 , T ) × Ω ) 2 ≤ C T , ℓ ∙ , G ∙ , A , B ∥ y d − e − T A ∂ x y 0 ∥ Y K , J + 2 . \lVert u_{\min}\rVert_{L^2((0,T)\times\Omega)}^2 \le C_{T,\ell_\bullet,G^\bullet,A,B}\lVert y_d-e^{-TA\partial_x}y_0\rVert_{\mathcal{Y}_{K,J}^{+}}^2. ∥ u m i n ∥ L 2 (( 0 , T ) × Ω ) 2 ≤ C T , ℓ ∙ , G ∙ , A , B ∥ y d − e − T A ∂ x y 0 ∥ Y K , J + 2 .
The observation space is optimal in the norm-domination sense of (45) .
Proof. By the HUM equivalence (46) , Corollary 5.10 gives exactly the stated range of control times. The minimum-control estimate (47) yields the cost bound, and whole-line observation gives the norm domination (45) .
Sharp observation cost for a 2 by 2 Jordan block
As above, F = A T F=A^{\mathsf{T}} F = A T , C = B T C=B^{\mathsf{T}} C = B T , and s = T − t s=T-t s = T − t . Let
F = ( λ 1 0 λ ) , C = ( 1 , 0 ) , λ ∈ R ∖ { 0 } . F=\begin{pmatrix}\lambda& 1\\0 & \lambda\end{pmatrix},\qquad C=(1,0),\qquad\lambda\in\mathbb{R}\setminus\{0\}. F = ( λ 0 1 λ ) , C = ( 1 , 0 ) , λ ∈ R ∖ { 0 } .
Here K = 1 K=1 K = 1 , both weights are diag ( 1 , ∣ ξ ∣ 2 ) \operatorname{diag}(1,\lvert\xi\rvert^2) diag ( 1 , ∣ ξ ∣ 2 ) , and
∥ f ∥ Y K , J − 2 = ∥ f 1 ∥ 2 2 + ∥ ∂ x f 2 ∥ 2 2 , C e s F ∂ x f ( x ) = f 1 ( x + λ s ) + s ∂ x f 2 ( x + λ s ) . \lVert f\rVert_{\mathcal{Y}_{K,J}^{-}}^2=\lVert f_1\rVert_2^2+\lVert\partial_xf_2\rVert_2^2,\qquad Ce^{sF\partial_x}f(x)=f_1(x+\lambda s)+s\partial_xf_2(x+\lambda s). ∥ f ∥ Y K , J − 2 = ∥ f 1 ∥ 2 2 + ∥ ∂ x f 2 ∥ 2 2 , C e s F ∂ x f ( x ) = f 1 ( x + λ s ) + s ∂ x f 2 ( x + λ s ) .
The completion imposes no additional L 2 L^2 L 2 condition on f 2 f_2 f 2 . Write T T , Ω ( f ) = ∫ 0 T ∫ Ω ∣ C e s F ∂ x f ∣ 2 d x d s \mathcal{T}_{T,\Omega}(f)=\int_0^T\int_\Omega\lvert Ce^{sF\partial_x}f\rvert^2\,dx\,ds T T , Ω ( f ) = ∫ 0 T ∫ Ω ∣ C e s F ∂ x f ∣ 2 d x d s .
Proposition 5.12. Let Ω ∈ G ℓ , G \Omega\in\mathcal{G}_{\ell,G} Ω ∈ G ℓ , G . For T > G / ∣ λ ∣ T > G/\lvert\lambda\rvert T > G / ∣ λ ∣ set μ T = min { ℓ , ∣ λ ∣ T − G } / ∣ λ ∣ \mu_T=\min\{\ell,\lvert\lambda\rvert T-G\}/\lvert\lambda\rvert μ T = min { ℓ , ∣ λ ∣ T − G } / ∣ λ ∣ . Then
μ T 3 12 ( 1 + T 2 ) ∥ f ∥ Y K , J − 2 ≤ T T , Ω ( f ) ≤ ( T + T 3 3 ) ∥ f ∥ Y K , J − 2 (70) \frac{\mu_T^3}{12(1+T^2)}\lVert f\rVert_{\mathcal{Y}_{K,J}^{-}}^2
\le\mathcal{T}_{T,\Omega}(f)
\le\left(T+\frac{T^3}{3}\right)\lVert f\rVert_{\mathcal{Y}_{K,J}^{-}}^2 \tag*{(70)} 12 ( 1 + T 2 ) μ T 3 ∥ f ∥ Y K , J − 2 ≤ T T , Ω ( f ) ≤ ( T + 3 T 3 ) ∥ f ∥ Y K , J − 2 (70)
for every f ∈ Y K , J − f \in\mathcal{Y}_{K,J}^{-} f ∈ Y K , J − .
Suppose now that the supremum G ∙ G^\bullet G ∙ of the actual gap lengths is positive, and let ℓ ∙ \ell_\bullet ℓ ∙ be the infimum of the observed interval lengths. Set T ∙ = G ∙ / ∣ λ ∣ T_\bullet=G^\bullet/|\lambda| T ∙ = G ∙ /∣ λ ∣ and δ = T − T ∙ \delta=T-T_\bullet δ = T − T ∙ . Write C o b s ( T , Ω ) C_{\mathrm{obs}}(T,\Omega) C obs ( T , Ω ) for the least constant in ∥ f ∥ Y K , J 2 2 ≤ C o b s ( T , Ω ) I T , Ω ( f ) \|f\|_{Y_{\mathcal{K},\mathcal{J}}^2}^2\le C_{\mathrm{obs}}(T,\Omega)I_{T,\Omega}(f) ∥ f ∥ Y K , J 2 2 ≤ C obs ( T , Ω ) I T , Ω ( f ) . For 0 < δ < ℓ ∙ / ∣ λ ∣ 0<\delta<\ell_\bullet/|\lambda| 0 < δ < ℓ ∙ /∣ λ ∣ this constant is exactly
C o b s ( T , Ω ) = 1 λ min ( M δ ) , M d = ∫ T − d T ( 1 s s s 2 ) d s ( 0 < d ≤ T ) . C_{\mathrm{obs}}(T,\Omega)=\frac{1}{\lambda_{\min}(M_\delta)},\qquad M_d=\int_{T-d}^{T}\begin{pmatrix}1&s\\s&s^2\end{pmatrix}\,ds\qquad(0<d\le T). C obs ( T , Ω ) = λ m i n ( M δ ) 1 , M d = ∫ T − d T ( 1 s s s 2 ) d s ( 0 < d ≤ T ) .
In particular,
lim T ↓ T ∙ ( T − T ∙ ) 3 C o b s ( T , Ω ) = 12 ( 1 + T ∙ 2 ) . \lim_{T\downarrow T_\bullet}(T-T_\bullet)^3C_{\mathrm{obs}}(T,\Omega)=12(1+T_\bullet^2). T ↓ T ∙ lim ( T − T ∙ ) 3 C obs ( T , Ω ) = 12 ( 1 + T ∙ 2 ) .
Proof. For Schwartz data set v ( y ) = ( f 1 ( y ) , ∂ x f 2 ( y ) ) T v(y)=(f_1(y),\partial_xf_2(y))^\mathsf{T} v ( y ) = ( f 1 ( y ) , ∂ x f 2 ( y ) ) T and w y ( s ) = 1 Ω ( y − λ s ) w_y(s)=\mathbf{1}_{\Omega}(y-\lambda s) w y ( s ) = 1 Ω ( y − λ s ) . Changing variables gives
I T , Ω ( f ) = ∫ R v ( y ) ∗ M ( y ) v ( y ) d y , M ( y ) = ∫ 0 T w y ( s ) ( 1 s s s 2 ) d s . I_{T,\Omega}(f)=\int_{\mathbb{R}}v(y)^\ast M(y)v(y)\,dy,\qquad M(y)=\int_0^T w_y(s)\begin{pmatrix}1&s\\s&s^2\end{pmatrix}\,ds. I T , Ω ( f ) = ∫ R v ( y ) ∗ M ( y ) v ( y ) d y , M ( y ) = ∫ 0 T w y ( s ) ( 1 s s s 2 ) d s .
Lemma 3.1 gives m : = ∫ 0 T w y d s ≥ μ T m:=\int_0^T w_y\,ds\ge\mu_T m := ∫ 0 T w y d s ≥ μ T . For s ˉ = m − 1 ∫ 0 T s w y ( s ) d s \bar{s}=m^{-1}\int_0^Tsw_y(s)\,ds s ˉ = m − 1 ∫ 0 T s w y ( s ) d s , the bound 0 ≤ w y ≤ 1 0\le w_y\le1 0 ≤ w y ≤ 1 implies that { ∣ s − s ˉ ∣ ≤ r } \{|s-\bar{s}|\le r\} { ∣ s − s ˉ ∣ ≤ r } contains at most 2 r 2r 2 r units of weighted mass. The layer-cake identity therefore yields
∫ 0 T ( s − s ˉ ) 2 w y ( s ) d s ≥ ∫ 0 m / 2 2 r ( m − 2 r ) d r = m 3 12 . \int_0^T(s-\bar{s})^2w_y(s)\,ds\ge\int_0^{m/2}2r(m-2r)\,dr=\frac{m^3}{12}. ∫ 0 T ( s − s ˉ ) 2 w y ( s ) d s ≥ ∫ 0 m /2 2 r ( m − 2 r ) d r = 12 m 3 .
Thus det M ( y ) ≥ m 4 / 12 \det M(y)\ge m^4/12 det M ( y ) ≥ m 4 /12 and tr M ( y ) ≤ m ( 1 + T 2 ) \operatorname{tr}M(y)\le m(1+T^2) tr M ( y ) ≤ m ( 1 + T 2 ) , so λ min ( M ( y ) ) ≥ det M ( y ) / tr M ( y ) ≥ μ T 3 / [ 12 ( 1 + T 2 ) ] \lambda_{\min}(M(y))\ge\det M(y)/\operatorname{tr}M(y)\ge\mu_T^3/[12(1+T^2)] λ m i n ( M ( y )) ≥ det M ( y ) / tr M ( y ) ≥ μ T 3 / [ 12 ( 1 + T 2 )] . Integration proves the lower bound. The upper bound follows from ∣ v 1 + s v 2 ∣ 2 ≤ ( 1 + s 2 ) ∣ v ∣ 2 |v_1+sv_2|^2\le(1+s^2)|v|^2 ∣ v 1 + s v 2 ∣ 2 ≤ ( 1 + s 2 ) ∣ v ∣ 2 and w y ≤ 1 w_y\le1 w y ≤ 1 ; density gives both estimates on the completion.
For the exact formula, let E y = { s ∈ ( 0 , T ) : y − λ s ∈ Ω } E_y=\{s\in(0,T):y-\lambda s\in\Omega\} E y = { s ∈ ( 0 , T ) : y − λ s ∈ Ω } . Lemma 3.1 with the actual parameters ℓ ∙ , G ∙ \ell_\bullet,G^\bullet ℓ ∙ , G ∙ gives ∣ E y ∣ ≥ δ |E_y|\ge\delta ∣ E y ∣ ≥ δ . For any z ∈ C 2 z\in\mathbb{C}^2 z ∈ C 2 put p ( s ) = ∣ z 1 + s z 2 ∣ 2 p(s)=|z_1+sz_2|^2 p ( s ) = ∣ z 1 + s z 2 ∣ 2 . Among measurable subsets of [ 0 , T ] [0,T] [ 0 , T ] of measure at least δ \delta δ , its integral is minimized on an interval I I I of length δ \delta δ . Indeed, when z 2 ≠ 0 z_2\ne0 z 2 = 0 , the sublevel sets of p p p in [ 0 , T ] [0,T] [ 0 , T ] are intervals, and their lengths vary continuously from zero to T T T . Choose I = { p ≤ h } I=\{p\le h\} I = { p ≤ h } of length δ \delta δ . On E y ∖ I E_y\setminus I E y ∖ I the integrand is at least h h h , and on I ∖ E y I\setminus E_y I ∖ E y it is at most h h h . For z 2 = 0 z_2=0 z 2 = 0 the assertion is immediate.
For I = [ u , u + δ ] I=[u,u+\delta] I = [ u , u + δ ] the moment matrix on I I I has determinant δ 4 / 12 \delta^4/12 δ 4 /12 and trace δ [ 1 + ( u + δ / 2 ) 2 ] + δ 3 / 12 \delta[1+(u+\delta/2)^2]+\delta^3/12 δ [ 1 + ( u + δ /2 ) 2 ] + δ 3 /12 . At fixed determinant the smaller eigenvalue decreases as the trace increases. The trace is largest at u = T − δ u=T-\delta u = T − δ . Hence ∫ E y p ( s ) d s ≥ λ min ( M δ ) ∣ z ∣ 2 \int_{E_y}p(s)\,ds\ge\lambda_{\min}(M_\delta)|z|^2 ∫ E y p ( s ) d s ≥ λ m i n ( M δ ) ∣ z ∣ 2 . Taking z = v ( y ) z=v(y) z = v ( y ) and integrating yields C o b s ( T , Ω ) ≤ 1 / λ min ( M δ ) C_{\mathrm{obs}}(T,\Omega)\le1/\lambda_{\min}(M_\delta) C obs ( T , Ω ) ≤ 1/ λ m i n ( M δ ) .
For the opposite bound choose gaps ( α n , β n ) (\alpha_n,\beta_n) ( α n , β n ) of lengths g n → G ∙ g_n\to G^\bullet g n → G ∙ and 0 < ε n < g n 0<\varepsilon_n<g_n 0 < ε n < g n tending to zero. If λ > 0 \lambda>0 λ > 0 , take J n = ( β n − ε n , β n ) J_n=(\beta_n-\varepsilon_n,\beta_n) J n = ( β n − ε n , β n ) , and if λ < 0 \lambda<0 λ < 0 , take J n = ( α n , α n + ε n ) J_n=(\alpha_n,\alpha_n+\varepsilon_n) J n = ( α n , α n + ε n ) . For y ∈ J n y\in J_n y ∈ J n , the path y − λ s y-\lambda s y − λ s lies in the gap for 0 < s < ( g n − ε n ) / ∣ λ ∣ 0<s<(g_n-\varepsilon_n)/|\lambda| 0 < s < ( g n − ε n ) /∣ λ ∣ . Thus E y ⊂ [ T − d n , T ] E_y\subset[T-d_n,T] E y ⊂ [ T − d n , T ] , where d n = T − ( g n − ε n ) / ∣ λ ∣ → δ d_n=T-(g_n-\varepsilon_n)/|\lambda|\to\delta d n = T − ( g n − ε n ) /∣ λ ∣ → δ . Let z n z_n z n be a unit eigenvector for λ min ( M d n ) \lambda_{\min}(M_{d_n}) λ m i n ( M d n ) and choose h n ∈ C c ∞ ( J n ) h_n\in C_c^\infty(J_n) h n ∈ C c ∞ ( J n ) with ∥ h n ′ ∥ 2 = 1 \|h_n'\|_2=1 ∥ h n ′ ∥ 2 = 1 . The datum f n = ( z n , 1 h n ′ , z n , 2 h n ) f_n=(z_{n,1}h_n',z_{n,2}h_n) f n = ( z n , 1 h n ′ , z n , 2 h n ) has unit intrinsic norm and observed energy at most λ min ( M d n ) \lambda_{\min}(M_{d_n}) λ m i n ( M d n ) . Passing to the limit proves equality, including when the supremum of the gap lengths is not attained. Finally det M δ = δ 4 / 12 \det M_\delta=\delta^4/12 det M δ = δ 4 /12 and λ max ( M δ ) / δ → 1 + T ∙ 2 \lambda_{\max}(M_\delta)/\delta\to1+T_\bullet^2 λ m a x ( M δ ) / δ → 1 + T ∙ 2 as T ↓ T ∙ T\downarrow T_\bullet T ↓ T ∙ , which gives the asserted limit.
Here C o b s C_{\mathrm{obs}} C obs compares squared norms. By (47) , the corresponding optimal minimal-control norm cost, for a unit final displacement, is C o b s 1 / 2 C_{\mathrm{obs}}^{1/2} C obs 1/2 ; its blow-up rate is therefore ( T − T ∙ ) − 3 / 2 (T-T_\bullet)^{-3/2} ( T − T ∙ ) − 3/2 .
Translating observation regions
The intrinsic spaces are unchanged for the two classes of translating regions considered below.
Slowly translating observation sets
Theorem 6.1. Assume the Kalman condition and σ ( A ) ⊂ R ∖ { 0 } \sigma(A)\subset\mathbb{R}\setminus\{0\} σ ( A ) ⊂ R ∖ { 0 } , and let Ω \Omega Ω be as in (4) . Fix T > 0 T>0 T > 0 and let h ∈ W 1 , ∞ ( 0 , T ) h\in W^{1,\infty}(0,T) h ∈ W 1 , ∞ ( 0 , T ) satisfy ∥ h ′ ∥ L ∞ ( 0 , T ) ≤ v 0 < λ ∙ \|h'\|_{L^\infty(0,T)}\le v_0<\lambda_\bullet ∥ h ′ ∥ L ∞ ( 0 , T ) ≤ v 0 < λ ∙ , with λ ∙ \lambda_\bullet λ ∙ as in (5) . Use the Lipschitz representative of h h h on [ 0 , T ] [0,T] [ 0 , T ] , and define
Ω t : = Ω + h ( t ) , L λ : = ∣ λ ∣ T − h ( T ) + h ( 0 ) , L ∙ : = min λ ∈ σ ( A ) L λ . \Omega_t:=\Omega+h(t),\qquad L_\lambda:=|\lambda|T-h(T)+h(0),\qquad L_\bullet:=\min_{\lambda\in\sigma(A)}L_\lambda. Ω t := Ω + h ( t ) , L λ := ∣ λ ∣ T − h ( T ) + h ( 0 ) , L ∙ := λ ∈ σ ( A ) min L λ . Then the following assertions are equivalent:
L ∙ > G ∙ L_\bullet>G^\bullet L ∙ > G ∙ ;
there exists c > 0 c>0 c > 0 such that the adjoint solution with any terminal datum q T ∈ Y K , J q_T\in\mathcal{Y}_{\mathcal{K},\mathcal{J}} q T ∈ Y K , J satisfies
c ∥ q T ∥ Y K , J 2 2 ≤ ∫ 0 T ∫ Ω t ∣ B T q ( t , x ) ∣ 2 d x d t ; c\|q_T\|_{\mathcal{Y}_{\mathcal{K},\mathcal{J}}^2}^2\le\int_0^T\int_{\Omega_t}|B^\mathsf{T}q(t,x)|^2\,dx\,dt; c ∥ q T ∥ Y K , J 2 2 ≤ ∫ 0 T ∫ Ω t ∣ B T q ( t , x ) ∣ 2 d x d t ; the system ∂ t y + A ∂ x y = B 1 Ω t u \partial_ty+A\partial_xy=B\mathbf{1}_{\Omega_t}u ∂ t y + A ∂ x y = B 1 Ω t u is exactly controllable in Y K , J + \mathcal{Y}_{\mathcal{K},\mathcal{J}}^+ Y K , J + in time T T T , with controls in L 2 ( ( 0 , T ) × R ; C m ) L^2((0,T)\times\mathbb{R};\mathbb{C}^m) L 2 (( 0 , T ) × R ; C m ) .
The observation in the second assertion is interpreted by completion from Schwartz terminal data. Its upper bound by C T ∥ q T ∥ Y K , J 2 C_T\|q_T\|_{\mathcal{Y}_{\mathcal{K},\mathcal{J}}^2} C T ∥ q T ∥ Y K , J 2 always holds.
In particular, T > G ∙ λ ∙ T>\dfrac{G^\bullet}{\lambda_\bullet} T > λ ∙ G ∙ is sufficient. The observation constants are uniform for fixed A , B , T , v 0 A,B,T,v_0 A , B , T , v 0 on families satisfying ℓ ∙ ≥ ℓ 0 > 0 \ell_\bullet\ge\ell_0>0 ℓ ∙ ≥ ℓ 0 > 0 and L ∙ − G ∙ ≥ η > 0 L_\bullet-G^\bullet\ge\eta>0 L ∙ − G ∙ ≥ η > 0 . The same statements include Ω = R \Omega=\mathbb{R} Ω = R , with G ∙ = 0 G^\bullet=0 G ∙ = 0 .
The bound on h ′ h' h ′ makes each relative characteristic t ↦ λ t − h ( t ) t \mapsto\lambda t-h(t) t ↦ λ t − h ( t ) monotone. Its travelled distance is therefore ∣ λ T − h ( T ) + h ( 0 ) ∣ \lvert\lambda T-h(T)+h(0)\rvert ∣ λ T − h ( T ) + h ( 0 )∣ , which explains the endpoint form of the criterion.
Proof. The whole-line case follows from Theorem 5.1 and HUM. Otherwise use the notation of § 6 and put λ ∙ = max λ ∈ σ ( A ) ∣ λ ∣ \lambda^\bullet=\max_{\lambda\in\sigma(A)}\lvert\lambda\rvert λ ∙ = max λ ∈ σ ( A ) ∣ λ ∣ .
Occupation and the hull. The maps ψ λ ( t ) = λ t − h ( t ) \psi_\lambda(t)=\lambda t-h(t) ψ λ ( t ) = λ t − h ( t ) are monotone and bi-Lipschitz, with ∣ λ ∣ − v 0 ≤ ∣ ψ λ ′ ∣ ≤ ∣ λ ∣ + v 0 \lvert\lambda\rvert-v_0\leq\lvert\psi_\lambda'\rvert\leq\lvert\lambda\rvert+v_0 ∣ λ ∣ − v 0 ≤ ∣ ψ λ ′ ∣ ≤ ∣ λ ∣ + v 0 a.e., and image length L λ L_\lambda L λ . If L ∙ > G ∙ L_\bullet>G^\bullet L ∙ > G ∙ , choose 0 < ε < 1 4 min { ℓ ∙ , L ∙ − G ∙ } 0<\varepsilon<\frac{1}{4}\min\{\ell_\bullet,L_\bullet-G^\bullet\} 0 < ε < 4 1 min { ℓ ∙ , L ∙ − G ∙ } and set a Ω ( x ) = min { 1 , ε − 1 dist ( x , R ∖ Ω ) } a_\Omega(x)=\min\{1,\varepsilon^{-1}\operatorname{dist}(x,\mathbb{R}\setminus\Omega)\} a Ω ( x ) = min { 1 , ε − 1 dist ( x , R ∖ Ω )} , b ( s , x ) = a Ω ( x − h ( T − s ) ) b(s,x)=a_\Omega(x-h(T-s)) b ( s , x ) = a Ω ( x − h ( T − s )) . The intervals on which a Ω = 1 a_\Omega=1 a Ω = 1 have lengths at least ℓ ∙ − 2 ε \ell_\bullet-2\varepsilon ℓ ∙ − 2 ε and gaps at most G ∙ + 2 ε G^\bullet+2\varepsilon G ∙ + 2 ε . Lemma 3.1 and the change of variable z = ψ λ ( t ) z=\psi_\lambda(t) z = ψ λ ( t ) give
∫ 0 T b ( s , y − λ s ) d s ≥ μ : = min { ℓ ∙ − 2 ε , L ∙ − G ∙ − 2 ε } λ ∙ + v 0 > 0. \int_0^T b(s,y-\lambda s)\,\mathrm{d}s\geq\mu:=\frac{\min\{\ell_\bullet-2\varepsilon,L_\bullet-G^\bullet-2\varepsilon\}}{\lambda^\bullet+v_0}>0. ∫ 0 T b ( s , y − λ s ) d s ≥ μ := λ ∙ + v 0 min { ℓ ∙ − 2 ε , L ∙ − G ∙ − 2 ε } > 0.
The bounds 0 ≤ b ≤ 1 0\leq b\leq1 0 ≤ b ≤ 1 , Lip x b ≤ ε − 1 \operatorname{Lip}_x b\leq\varepsilon^{-1} Lip x b ≤ ε − 1 and Lip s b ≤ v 0 ε − 1 \operatorname{Lip}_s b\leq v_0\varepsilon^{-1} Lip s b ≤ v 0 ε − 1 make its full spatial hull compact in C l o c ( [ 0 , T ] × R ) C_{\mathrm{loc}}([0,T]\times\mathbb{R}) C loc ([ 0 , T ] × R ) ; every hull weight β \beta β retains this occupation bound.
Define Γ β u ( s , x ) = β ( s , x ) M s ( D x ) u ( x ) \Gamma_\beta u(s,x)=\sqrt{\beta(s,x)}M_s(D_x)u(x) Γ β u ( s , x ) = β ( s , x ) M s ( D x ) u ( x ) . The proofs of Lemmas 5.6 and 5.8 apply with b ( x ) b(x) b ( x ) replaced by β ( s , x ) \beta(s,x) β ( s , x ) . Proposition 2.6 allows the time-dependent coefficients c λ , β = β V λ c_{\lambda,\beta}=\sqrt{\beta}V_\lambda c λ , β = β V λ and kernels k β ( s , x , y ) = β ( s , x ) k s ( x − y ) k_\beta(s,x,y)=\sqrt{\beta(s,x)}k_s(x-y) k β ( s , x , y ) = β ( s , x ) k s ( x − y ) ; Lemma A.1 , applied to s ↦ β ( s , y − λ s ) s\mapsto\beta(s,y-\lambda s) s ↦ β ( s , y − λ s ) , and Lemma 5.3 give the coercivity of ∑ λ ∫ 0 T β ( s , y − λ s ) V λ ( s ) ∗ V λ ( s ) d s \sum_\lambda\int_0^T\beta(s,y-\lambda s)V_\lambda(s)^*V_\lambda(s)\,\mathrm{d}s ∑ λ ∫ 0 T β ( s , y − λ s ) V λ ( s ) ∗ V λ ( s ) d s ; and in the uniqueness argument the set U b U_b U b is replaced by the open set U β = { ( s , x ) : β ( s , x ) > 0 } U_\beta=\{(s,x):\beta(s,x)>0\} U β = {( s , x ) : β ( s , x ) > 0 } , from which every y y y is reached along each characteristic family by the occupation bound. Lemma B.1 applies on U β U_\beta U β . In the last step of Lemma 5.8, the polynomial s ↦ M s ( 0 ) u 0 s\mapsto M_s(0)u_0 s ↦ M s ( 0 ) u 0 vanishes on a nonempty open time interval, since U β U_\beta U β is open and nonempty; hence it vanishes identically, and (50) gives u 0 = 0 u_0=0 u 0 = 0 . Theorem 2.1 , Lemma 5.4 and b ( s , x ) ≤ 1 Ω + h ( T − s ) ( x ) b(s,x)\leq1_{\Omega+h(T-s)}(x) b ( s , x ) ≤ 1 Ω + h ( T − s ) ( x ) now give the lower bound; whole-line admissibility gives the upper bound and extension by completion.
Necessity, including the critical time. For F v = λ v ≠ 0 Fv=\lambda v\neq0 F v = λ v = 0 with v ≠ 0 v\neq0 v = 0 , Kalman gives C v ≠ 0 Cv\neq0 C v = 0 . Terminal data v φ v\varphi v φ have norm equivalent to ∥ φ ∥ 2 \lVert\varphi\rVert_2 ∥ φ ∥ 2 by whole-line observability, while their observed energy is ∣ C v ∣ 2 ∫ m λ ( y ) ∣ φ ( y ) ∣ 2 d y \lvert Cv\rvert^2\int m_\lambda(y)\lvert\varphi(y)\rvert^2\,\mathrm{d}y ∣ C v ∣ 2 ∫ m λ ( y ) ∣ φ ( y ) ∣ 2 d y , where
m λ ( y ) = ∫ 0 T 1 Ω ( y − λ ( T − t ) − h ( t ) ) d t = ∫ I λ 1 Ω ( y − z ) ρ λ ( z ) d z . m_\lambda(y)=\int_0^T1_\Omega(y-\lambda(T-t)-h(t))\,\mathrm{d}t=\int_{I_\lambda}1_\Omega(y-z)\rho_\lambda(z)\,\mathrm{d}z. m λ ( y ) = ∫ 0 T 1 Ω ( y − λ ( T − t ) − h ( t )) d t = ∫ I λ 1 Ω ( y − z ) ρ λ ( z ) d z .
Here I λ I_\lambda I λ is the image of ( 0 , T ) (0,T) ( 0 , T ) under z λ ( t ) = λ ( T − t ) + h ( t ) z_\lambda(t)=\lambda(T-t)+h(t) z λ ( t ) = λ ( T − t ) + h ( t ) , and ρ λ = ∣ ( z λ − 1 ) ′ ∣ \rho_\lambda=\lvert(z_\lambda^{-1})'\rvert ρ λ = ∣( z λ − 1 ) ′ ∣ there, extended by zero elsewhere. The length of I λ I_\lambda I λ is L λ L_\lambda L λ , and ( ∣ λ ∣ + v 0 ) − 1 ≤ ρ λ ≤ ( ∣ λ ∣ − v 0 ) − 1 (\lvert\lambda\rvert+v_0)^{-1}\leq\rho_\lambda\leq(\lvert\lambda\rvert-v_0)^{-1} (∣ λ ∣ + v 0 ) − 1 ≤ ρ λ ≤ (∣ λ ∣ − v 0 ) − 1 . Here the bounds hold on I λ I_\lambda I λ . The zero extension belongs to L 1 ( R ) L^1(\mathbb{R}) L 1 ( R ) , so ∥ m λ ( ⋅ + r ) − m λ ∥ ∞ ≤ ∥ ρ λ ( ⋅ + r ) − ρ λ ∥ 1 → 0 \lVert m_\lambda(\cdot+r)-m_\lambda\rVert_\infty\leq\lVert\rho_\lambda(\cdot+r)-\rho_\lambda\rVert_1\to0 ∥ m λ ( ⋅ + r ) − m λ ∥ ∞ ≤ ∥ ρ λ ( ⋅ + r ) − ρ λ ∥ 1 → 0 . If L λ < G ∙ L_\lambda<G^\bullet L λ < G ∙ , placing y − I λ y-I_\lambda y − I λ inside a larger gap makes m λ ( y ) = 0 m_\lambda(y)=0 m λ ( y ) = 0 . If L λ = G ∙ L_\lambda=G^\bullet L λ = G ∙ , then y − I λ y-I_\lambda y − I λ is an interval of length G ∙ G^\bullet G ∙ with left endpoint y − sup I λ y-\sup I_\lambda y − sup I λ , so m λ ( y ) ≤ ( ∣ λ ∣ − v 0 ) − 1 m G ∙ ( y − sup I λ ) m_\lambda(y)\leq(\lvert\lambda\rvert-v_0)^{-1}m_{G^\bullet}(y-\sup I_\lambda) m λ ( y ) ≤ (∣ λ ∣ − v 0 ) − 1 m G ∙ ( y − sup I λ ) with m G ∙ m_{G^\bullet} m G ∙ as in Lemma C.2 . That lemma gives inf m λ = 0 \inf m_\lambda=0 inf m λ = 0 , whether or not G ∙ G^\bullet G ∙ is attained. In either case its essential infimum is zero, and normalized smooth packets supported where it is small contradict observability.
Uniformity and controllability. Choose ε < 1 4 min { ℓ 0 , η } \varepsilon<\frac{1}{4}\min\{\ell_0,\eta\} ε < 4 1 min { ℓ 0 , η } uniformly and replace ℓ ∙ , L ∙ − G ∙ \ell_\bullet,L_\bullet-G^\bullet ℓ ∙ , L ∙ − G ∙ in μ \mu μ by ℓ 0 , η \ell_0,\eta ℓ 0 , η . All continuous weights with these Lipschitz and occupation bounds form a compact translation-invariant family. The preceding argument and Corollary 2.2 give uniform constants. Also L ∙ ≥ ( λ ∙ − v 0 ) T L_\bullet\geq(\lambda_\bullet-v_0)T L ∙ ≥ ( λ ∙ − v 0 ) T , proving the sufficient time bound. § 6.1 gives uniform boundedness of B 1 Ω : L 2 → Y K , J + B1_\Omega:L^2\to\mathcal{Y}_{K,J}^+ B 1 Ω : L 2 → Y K , J + ; strong continuity of translations gives strong measurability. The bounded endpoint map has adjoint q T ↦ 1 Ω B T q ( T ) q_T\mapsto1_{\Omega}B^Tq(T) q T ↦ 1 Ω B T q ( T ) . The closed range theorem and the reversible group therefore identify the lower observation bound with exact controllability between arbitrary states. □
Translations at constant speed
For affine translations, a change of frame removes the speed restriction and also permits stationary characteristics of the original system.
Corollary 6.2. Assume the Kalman condition and σ ( A ) ⊂ R \sigma(A)\subset\mathbb{R} σ ( A ) ⊂ R , and let Ω \Omega Ω be as in (4) ; Ω = R \Omega=\mathbb{R} Ω = R is allowed, with G ∙ = 0 G^\bullet=0 G ∙ = 0 . For arbitrary h 0 , v ∈ R h_0,v\in\mathbb{R} h 0 , v ∈ R and T > 0 T>0 T > 0 , set
Ω t : = Ω + h 0 + v t , d ( v ) : = min λ ∈ σ ( A ) ∣ λ − v ∣ . \Omega_t:=\Omega+h_0+vt,\qquad d(v):=\min_{\lambda\in\sigma(A)}\lvert\lambda-v\rvert. Ω t := Ω + h 0 + v t , d ( v ) := λ ∈ σ ( A ) min ∣ λ − v ∣ . Observability in Y K , J − \mathcal{Y}_{K,J}^- Y K , J − on the moving set, and equivalently exact controllability in Y K , J + \mathcal{Y}_{K,J}^+ Y K , J + in time T T T , hold if and only if
G ∙ = 0 or T d ( v ) > G ∙ . G^\bullet=0\quad\text{or}\quad T\,d(v)>G^\bullet. G ∙ = 0 or T d ( v ) > G ∙ . More explicitly, when G ∙ > 0 G^\bullet>0 G ∙ > 0 and v ∉ σ ( A ) v\notin\sigma(A) v ∈ / σ ( A ) the necessary and sufficient condition is
T > G ∙ min λ ∈ σ ( A ) ∣ λ − v ∣ . T>\frac{G^\bullet}{\min_{\lambda\in\sigma(A)}\lvert\lambda-v\rvert}. T > min λ ∈ σ ( A ) ∣ λ − v ∣ G ∙ . When G ∙ > 0 G^\bullet>0 G ∙ > 0 and v ∈ σ ( A ) v\in\sigma(A) v ∈ σ ( A ) , observability fails for every positive time. When G ∙ = 0 G^\bullet=0 G ∙ = 0 , it holds for every T > 0 T>0 T > 0 and every v v v , including resonant velocities.
Proof. Set r ( t , z ) = q ( t , z + h 0 + v t ) r(t,z)=q(t,z+h_0+vt) r ( t , z ) = q ( t , z + h 0 + v t ) . Then ∂ t r + ( A T − v I ) ∂ z r = 0 \partial_t r+(A^{\mathsf T}-vI)\partial_z r=0 ∂ t r + ( A T − v I ) ∂ z r = 0 and r ( T , z ) = q T ( z + h 0 + v T ) r(T,z)=q_T(z+h_0+vT) r ( T , z ) = q T ( z + h 0 + v T ) ; changing variables identifies the moving observation of q q q with the stationary observation of r r r on Ω \Omega Ω . For F = A T F=A^{\mathsf T} F = A T , C = B T C=B^{\mathsf T} C = B T and F v = F − v I F_v=F-vI F v = F − v I , the binomial formula gives ⋂ j = 0 k ker ( C F v j ) = ⋂ j = 0 k ker ( C F j ) \bigcap_{j=0}^{k}\operatorname{ker}(CF_v^j)=\bigcap_{j=0}^{k}\operatorname{ker}(CF^j) ⋂ j = 0 k ker ( C F v j ) = ⋂ j = 0 k ker ( C F j ) for every k k k . Thus the Kalman filtration and its projectors are unchanged. The generalized spectral projectors are also unchanged, and F v − ( λ − v ) I = F − λ I F_v-(\lambda-v)I=F-\lambda I F v − ( λ − v ) I = F − λ I on E λ E_\lambda E λ , so the C λ N j C_\lambda N^j C λ N j coincide. Both intrinsic weights, and hence both completed energy spaces, are identical. Spatial translation is an isometry, giving
∥ r ( T ) ∥ Y K , J ( A − v I ) , B = ∥ q T ∥ Y K , J ( A ) , B . \|r(T)\|_{Y_{\mathcal K,J(A-vI),B}}=\|q_T\|_{Y_{\mathcal K,J(A),B}}. ∥ r ( T ) ∥ Y K , J ( A − v I ) , B = ∥ q T ∥ Y K , J ( A ) , B .
If G ∙ = 0 G^\bullet=0 G ∙ = 0 , the observation set equals R \mathbb{R} R up to countably many endpoints, and Theorem 5.1 applies. If d ( v ) > 0 d(v)>0 d ( v ) > 0 , Corollary 5.10 for A − v I A-vI A − v I proves sufficiency, while Proposition C.1 proves failure for T d ( v ) ≤ G ∙ Td(v)\le G^\bullet T d ( v ) ≤ G ∙ , including equality when the supremal gap is not attained. If G ∙ > 0 G^\bullet>0 G ∙ > 0 and d ( v ) = 0 d(v)=0 d ( v ) = 0 , take an eigenvector F e = v e Fe=ve F e = v e and a nonzero smooth packet r T = e φ r_T=e\varphi r T = e φ supported in a gap. This solution is stationary in the moving frame and invisible, while its whole-line observation T ∣ C e ∣ 2 ∥ φ ∥ 2 2 T|Ce|^2\|\varphi\|_2^2 T ∣ C e ∣ 2 ∥ φ ∥ 2 2 is positive. Finally, the change of frame is also an isometry on controls and on both endpoint state spaces, so HUM gives the stated controllability equivalence.
Taking v = 0 v=0 v = 0 recovers the stationary criterion; for ∣ v ∣ < min λ ∈ σ ( A ) ∣ λ ∣ |v|<\min_{\lambda\in\sigma(A)}|\lambda| ∣ v ∣ < min λ ∈ σ ( A ) ∣ λ ∣ , this agrees with Theorem 6.1 .
Transport with nonlocal dispersion in $\mathbb{R}^d`
Let d ≥ 1 d\ge1 d ≥ 1 , V ∈ R d V\in\mathbb{R}^d V ∈ R d and J ∈ L 1 ( R d ; C ) J\in L^1(\mathbb{R}^d;\mathbb{C}) J ∈ L 1 ( R d ; C ) . For u 0 ∈ L 2 ( R d ) u^0\in L^2(\mathbb{R}^d) u 0 ∈ L 2 ( R d ) , let u u u solve
∂ t u + V ⋅ ∇ u = J ∗ u − u in ( 0 , T ) × R d , u ( 0 ) = u 0 . (71) \partial_tu+V\cdot\nabla u=J*u-u\quad\text{in }(0,T)\times\mathbb{R}^d,\qquad u(0)=u^0. \tag*{(71)} ∂ t u + V ⋅ ∇ u = J ∗ u − u in ( 0 , T ) × R d , u ( 0 ) = u 0 . (71)
For J ≥ 0 J\ge0 J ≥ 0 with ∫ J = 1 \int J=1 ∫ J = 1 , the operator J ∗ u − u J*u-u J ∗ u − u is the classical nonlocal dispersal operator; no sign or normalization is needed below. For a weight a : R d → [ 0 , 1 ] a:\mathbb{R}^d\to[0,1] a : R d → [ 0 , 1 ] , put
μ a : = inf z ∈ R d ∫ 0 T a ( z + t V ) d t . (72) \mu_a:=\inf_{z\in\mathbb{R}^d}\int_0^T a(z+tV)\,\mathrm{d}t. \tag*{(72)} μ a := z ∈ R d inf ∫ 0 T a ( z + t V ) d t . (72)
Balls B ε ( z ) B_\varepsilon(z) B ε ( z ) are Euclidean in this subsection.
Theorem 7.1. Let L ≥ 0 L\ge0 L ≥ 0 and let a : R d → [ 0 , 1 ] a:\mathbb{R}^d\to[0,1] a : R d → [ 0 , 1 ] satisfy Lip ( a ) ≤ L \operatorname{Lip}(a)\le L Lip ( a ) ≤ L .
(i) If μ a ≥ μ > 0 \mu_a\ge\mu>0 μ a ≥ μ > 0 , then
∥ u 0 ∥ L 2 ( R d ) 2 ≤ C ∫ 0 T ∫ R d a ( x ) ∣ u ( t , x ) ∣ 2 d x d t ( u 0 ∈ L 2 ( R d ) ) , (73) \|u^0\|_{L^2(\mathbb{R}^d)}^2\le C\int_0^T\int_{\mathbb{R}^d}a(x)|u(t,x)|^2\,\mathrm{d}x\,\mathrm{d}t\qquad(u^0\in L^2(\mathbb{R}^d)), \tag*{(73)} ∥ u 0 ∥ L 2 ( R d ) 2 ≤ C ∫ 0 T ∫ R d a ( x ) ∣ u ( t , x ) ∣ 2 d x d t ( u 0 ∈ L 2 ( R d )) , (73)
with C C C depending only on d d d , T T T , V V V , J J J , L L L and μ \mu μ .
(ii) If μ a = 0 \mu_a=0 μ a = 0 , then (73) fails for every C C C .
Proof. Step 1. Representation. The operator J f : = J ∗ f − f \mathcal{J}f:=J*f-f J f := J ∗ f − f is bounded on L 2 L^2 L 2 and commutes with translations, and
e t J f = e − t f + K t ∗ f , K t : = e − t ∑ k ≥ 1 t k k ! J ∗ k . (74) e^{t\mathcal{J}}f=e^{-t}f+K_t*f,\qquad K_t:=e^{-t}\sum_{k\ge1}\frac{t^k}{k!}J^{*k}. \tag*{(74)} e t J f = e − t f + K t ∗ f , K t := e − t k ≥ 1 ∑ k ! t k J ∗ k . (74)
The series converges in L 1 L^1 L 1 , uniformly for 0 ≤ t ≤ T 0\le t\le T 0 ≤ t ≤ T , and
∣ K t ∣ ≤ F T : = ∑ r ≥ 1 T r r ! ∣ J ∣ ∗ r , ∥ F T ∥ 1 ≤ e T ∥ J ∥ 1 − 1. (75) |K_t|\le F_T:=\sum_{r\ge1}\frac{T^r}{r!}|J|^{*r},\qquad\|F_T\|_1\le e^{T\|J\|_1}-1. \tag*{(75)} ∣ K t ∣ ≤ F T := r ≥ 1 ∑ r ! T r ∣ J ∣ ∗ r , ∥ F T ∥ 1 ≤ e T ∥ J ∥ 1 − 1. (75)
Since w ( t , x ) : = u ( t , x + t V ) w(t,x):=u(t,x+tV) w ( t , x ) := u ( t , x + t V ) solves ∂ t w = J w \partial_tw=\mathcal{J}w ∂ t w = J w , we have u ( t , x ) = ( e t J u 0 ) ( x − t V ) u(t,x)=(e^{t\mathcal{J}}u^0)(x-tV) u ( t , x ) = ( e t J u 0 ) ( x − t V ) . Thus, for b : R d → [ 0 , 1 ] b:\mathbb{R}^d\to[0,1] b : R d → [ 0 , 1 ] ,
Γ b u 0 ( t , x ) : = b ( x ) 1 / 2 u ( t , x ) = e − t b ( x ) 1 / 2 u 0 ( x − t V ) + ∫ R d b ( x ) 1 / 2 K t ( x − t V − y ) u 0 ( y ) d y . (76) \Gamma_bu^0(t,x):=b(x)^{1/2}u(t,x)=e^{-t}b(x)^{1/2}u^0(x-tV)+\int_{\mathbb{R}^d}b(x)^{1/2}K_t(x-tV-y)u^0(y)\,\mathrm{d}y. \tag*{(76)} Γ b u 0 ( t , x ) := b ( x ) 1/2 u ( t , x ) = e − t b ( x ) 1/2 u 0 ( x − t V ) + ∫ R d b ( x ) 1/2 K t ( x − t V − y ) u 0 ( y ) d y . (76)
Step 2. Verification of (H1)–(H5). Let A L , μ \mathcal{A}_{L,\mu} A L , μ consist of the functions b : R d → [ 0 , 1 ] b:\mathbb{R}^d\to[0,1] b : R d → [ 0 , 1 ] satisfying Lip ( b ) ≤ L \operatorname{Lip}(b)\le L Lip ( b ) ≤ L and μ b ≥ μ \mu_b\ge\mu μ b ≥ μ . With locally uniform convergence and the translation action θ h b = b ( ⋅ + h ) \theta_hb=b(\cdot+h) θ h b = b ( ⋅ + h ) , this family is compact and metrizable by Arzelà–Ascoli. The occupation bound passes to the limit on each compact segment { z + t V : 0 ≤ t ≤ T } \{z+tV:0\le t\le T\} { z + t V : 0 ≤ t ≤ T } . Formula (76) gives covariance. Local uniform convergence of b b b implies convergence of the observed cell energies by dominated convergence.
The transport term has finite range between cells. For the convolution term, put D = 1 + T ∣ V ∣ 2 D=1+T|V|^2 D = 1 + T ∣ V ∣ 2 . Young’s inequality bounds the block from Q j Q_j Q j to Q k Q_k Q k , at each time, by ∫ ∣ k − j ∣ − [ D , D ] F T \int_{|k-j|-[D,D]}F_T ∫ ∣ k − j ∣ − [ D , D ] F T . These integrals are summable in k − j k-j k − j by finite overlap, which proves (H3) after integration in time. The principal multiplication operator D b D_b D b has coefficient d b ( z ) = ∫ 0 T e − 2 t b ( z + t V ) d t ∈ [ e − 2 T μ , T ] d_b(z)=\int_0^T e^{-2t}b(z+tV)\,\mathrm{d}t\in[e^{-2T}\mu,T] d b ( z ) = ∫ 0 T e − 2 t b ( z + t V ) d t ∈ [ e − 2 T μ , T ] , which depends continuously on b b b locally uniformly. This gives (H4).
To verify (H5), choose J m ∈ C c ∞ J_m\in C_c^\infty J m ∈ C c ∞ with J m → J J_m\to J J m → J in L 1 L^1 L 1 . For any α > 0 \alpha>0 α > 0 , let M m = ∥ e α ∣ ⋅ ∣ J m ∥ 1 M_m=\|e^{\alpha|\cdot|}J_m\|_1 M m = ∥ e α ∣ ⋅ ∣ J m ∥ 1 and L m = ∥ e α ∣ ⋅ ∣ J m ∥ ∞ L_m=\|e^{\alpha|\cdot|}J_m\|_\infty L m = ∥ e α ∣ ⋅ ∣ J m ∥ ∞ . The triangle inequality and Young’s inequality give ∥ e α ∣ ⋅ ∣ J m ∗ r ∥ ∞ ≤ L m M m r − 1 \|e^{\alpha|\cdot|}J_m^{*r}\|_\infty\le L_mM_m^{r-1} ∥ e α ∣ ⋅ ∣ J m ∗ r ∥ ∞ ≤ L m M m r − 1 . Thus ∣ K t ( m ) ( x ) ∣ ≤ C m , T e − α ∣ x ∣ |K_t^{(m)}(x)|\le C_{m,T}e^{-\alpha|x|} ∣ K t ( m ) ( x ) ∣ ≤ C m , T e − α ∣ x ∣ on [ 0 , T ] [0,T] [ 0 , T ] . Proposition 2.6 applies to these kernels with one velocity − V -V − V and c 1 , b ( t , x ) = e − t b ( x ) c_{1,b}(t,x)=e^{-t}\sqrt{b(x)} c 1 , b ( t , x ) = e − t b ( x ) . In particular, the bounded-window restrictions of ( Γ b ( m ) ) ∗ Γ b ( m ) − D b (\Gamma_b^{(m)})^*\Gamma_b^{(m)}-D_b ( Γ b ( m ) ) ∗ Γ b ( m ) − D b are collectively compact in b b b . The convolution generators converge in operator norm on L 2 L^2 L 2 , hence their exponentials converge in norm uniformly on [ 0 , T ] [0,T] [ 0 , T ] . It follows that Γ b ( m ) → Γ b \Gamma_b^{(m)}\to\Gamma_b Γ b ( m ) → Γ b and their Gramians converge in norm, uniformly in b b b . Since D b D_b D b is independent of m m m , the restricted remainders for J J J are uniform norm limits of collectively compact families and are collectively compact by Lemma 2.14(b). This proves (H5), and Lemma 2.15 gives (R).
Step 3. Uniqueness on A L , μ \mathcal{A}_{L,\mu} A L , μ . Let b ∈ A L , μ b \in\mathcal{A}_{L,\mu} b ∈ A L , μ and w ∈ L u l o c 2 w \in L^2_{\mathrm{uloc}} w ∈ L uloc 2 with Γ b w = 0 \Gamma_b w=0 Γ b w = 0 . By (75) , the series (24) defining the extension of Γ b \Gamma_b Γ b can be summed under the integral, so Γ b w ( t , x ) = b ( x ) 1 / 2 g ( t ) ( x − t V ) \Gamma_b w(t,x)=b(x)^{1/2}g(t)(x-tV) Γ b w ( t , x ) = b ( x ) 1/2 g ( t ) ( x − t V ) with g ( t ) : = e − t w + K t ∗ w g(t):=e^{-t}w+K_t*w g ( t ) := e − t w + K t ∗ w . Young’s inequality on the cells gives ∥ J ∗ f ∥ L u l o c 2 ≤ 2 d ∥ J ∥ L 1 ∥ f ∥ L u l o c 2 \lVert J*f\rVert_{L^2_{\mathrm{uloc}}}\le2^d\lVert J\rVert_{L^1}\lVert f\rVert_{L^2_{\mathrm{uloc}}} ∥ J ∗ f ∥ L uloc 2 ≤ 2 d ∥ J ∥ L 1 ∥ f ∥ L uloc 2 , so g ( t ) = e − t ∑ k ≥ 0 t k k ! J ∗ k ∗ w g(t)=e^{-t}\sum_{k\ge0}\frac{t^k}{k!}J^{*k}*w g ( t ) = e − t ∑ k ≥ 0 k ! t k J ∗ k ∗ w extends to an entire function of t t t with values in L u l o c 2 L^2_{\mathrm{uloc}} L uloc 2 .
Fix y 0 ∈ R d y_0\in\mathbb{R}^d y 0 ∈ R d . Since ∫ 0 T b ( y 0 + t V ) d t ≥ μ > 0 \int_0^T b(y_0+tV)\,dt\ge\mu>0 ∫ 0 T b ( y 0 + t V ) d t ≥ μ > 0 and b b b is continuous, there are t 0 ∈ ( 0 , T ) t_0\in(0,T) t 0 ∈ ( 0 , T ) and ε > 0 \varepsilon>0 ε > 0 such that ( t 0 − ε , t 0 + ε ) ⊂ ( 0 , T ) (t_0-\varepsilon,t_0+\varepsilon)\subset(0,T) ( t 0 − ε , t 0 + ε ) ⊂ ( 0 , T ) and b ( y + t V ) > 0 b(y+tV)>0 b ( y + t V ) > 0 whenever ∣ t − t 0 ∣ < ε |t-t_0|<\varepsilon ∣ t − t 0 ∣ < ε and y ∈ B ε ( y 0 ) y\in B_\varepsilon(y_0) y ∈ B ε ( y 0 ) . The change of variables x = y + t V x=y+tV x = y + t V at fixed t t t shows that g ( t ) ( y ) = 0 g(t)(y)=0 g ( t ) ( y ) = 0 for almost every ( t , y ) ∈ ( t 0 − ε , t 0 + ε ) × B ε ( y 0 ) (t,y)\in(t_0-\varepsilon,t_0+\varepsilon)\times B_\varepsilon(y_0) ( t , y ) ∈ ( t 0 − ε , t 0 + ε ) × B ε ( y 0 ) . By Fubini’s theorem, the entire function t ↦ 1 B ε ( y 0 ) g ( t ) ∈ L 2 ( B ε ( y 0 ) ) t\mapsto1_{B_\varepsilon(y_0)}g(t)\in L^2(B_\varepsilon(y_0)) t ↦ 1 B ε ( y 0 ) g ( t ) ∈ L 2 ( B ε ( y 0 )) vanishes for almost every t ∈ ( t 0 − ε , t 0 + ε ) t\in(t_0-\varepsilon,t_0+\varepsilon) t ∈ ( t 0 − ε , t 0 + ε ) , hence identically. At t = 0 t=0 t = 0 this gives w = 0 w=0 w = 0 on B ε ( y 0 ) B_\varepsilon(y_0) B ε ( y 0 ) . As y 0 y_0 y 0 is arbitrary, w = 0 w=0 w = 0 .
Step 4. Conclusion of (i). Corollary 2.2, applied to { Γ b } b ∈ A L , μ \{\Gamma_b\}_{b\in\mathcal{A}_{L,\mu}} { Γ b } b ∈ A L , μ , gives c ∗ > 0 c_*>0 c ∗ > 0 , depending only on d d d , T T T , V V V , J J J , L L L , μ \mu μ , such that ∥ Γ b u 0 ∥ Z ≥ c ∗ ∥ u 0 ∥ L 2 \lVert\Gamma_bu^0\rVert_Z\ge c_*\lVert u^0\rVert_{L^2} ∥ Γ b u 0 ∥ Z ≥ c ∗ ∥ u 0 ∥ L 2 for all b ∈ A L , μ b\in\mathcal{A}_{L,\mu} b ∈ A L , μ and u 0 ∈ L 2 u^0\in L^2 u 0 ∈ L 2 . Since a ∈ A L , μ a\in\mathcal{A}_{L,\mu} a ∈ A L , μ and ∥ Γ a u 0 ∥ Z \lVert\Gamma_a u^0\rVert_Z ∥ Γ a u 0 ∥ Z is the integral in (73) , this proves (i) with C = c ∗ − 2 C=c_*^{-2} C = c ∗ − 2 .
Step 5. Necessity. Let μ a = 0 \mu_a=0 μ a = 0 . The function m ( z ) : = ∫ 0 T a ( z + t V ) d t m(z):=\int_0^T a(z+tV)\,dt m ( z ) := ∫ 0 T a ( z + t V ) d t is T L TL T L -Lipschitz. Given k ≥ 1 k\ge1 k ≥ 1 , choose z k z_k z k with m ( z k ) ≤ 1 / k m(z_k)\le1/k m ( z k ) ≤ 1/ k and put ε k : = 1 / ( k ( T L + 1 ) ) \varepsilon_k:=1/(k(TL+1)) ε k := 1/ ( k ( T L + 1 )) , so that m ≤ 2 / k m\le2/k m ≤ 2/ k on B ε k ( z k ) B_{\varepsilon_k}(z_k) B ε k ( z k ) . Fix χ ∈ C c ∞ ( B 1 ( 0 ) ) \chi\in C^\infty_c(B_1(0)) χ ∈ C c ∞ ( B 1 ( 0 )) with ∥ χ ∥ L 2 = 1 \lVert\chi\rVert_{L^2}=1 ∥ χ ∥ L 2 = 1 and a unit vector η ∈ R d \eta\in\mathbb{R}^d η ∈ R d , and set
u k , N 0 ( x ) : = ε k − d / 2 χ ( ( x − z k ) / ε k ) e i N η ⋅ x ( N ≥ 1 ) . u^0_{k,N}(x):=\varepsilon_k^{-d/2}\chi((x-z_k)/\varepsilon_k)e^{iN\eta\cdot x}\qquad(N\ge1). u k , N 0 ( x ) := ε k − d /2 χ (( x − z k ) / ε k ) e i N η ⋅ x ( N ≥ 1 ) .
Then ∥ u k , N 0 ∥ L 2 = 1 \lVert u^0_{k,N}\rVert_{L^2}=1 ∥ u k , N 0 ∥ L 2 = 1 , and (74) together with the change of variables x = y + t V x=y+tV x = y + t V gives
∫ 0 T ∫ R d a ∣ u ∣ 2 d x d t ≤ 2 ∫ m ( y ) ∣ u k , N 0 ( y ) ∣ 2 d y + 2 T sup 0 ≤ t ≤ T ∥ K t ∗ u k , N 0 ∥ L 2 2 ≤ 4 k + 2 T sup 0 ≤ t ≤ T ∥ K t ∗ u k , N 0 ∥ L 2 2 . \int_0^T\int_{\mathbb{R}^d}a|u|^2\,dx\,dt\le2\int m(y)|u^0_{k,N}(y)|^2\,dy+2T\sup_{0\le t\le T}\lVert K_t*u^0_{k,N}\rVert_{L^2}^2\le\frac{4}{k}+2T\sup_{0\le t\le T}\lVert K_t*u^0_{k,N}\rVert_{L^2}^2. ∫ 0 T ∫ R d a ∣ u ∣ 2 d x d t ≤ 2 ∫ m ( y ) ∣ u k , N 0 ( y ) ∣ 2 d y + 2 T 0 ≤ t ≤ T sup ∥ K t ∗ u k , N 0 ∥ L 2 2 ≤ k 4 + 2 T 0 ≤ t ≤ T sup ∥ K t ∗ u k , N 0 ∥ L 2 2 .
The Fourier multiplier of f ↦ K t ∗ f f\mapsto K_t*f f ↦ K t ∗ f is e − t ( e t J 0 ( ξ ) − 1 ) e^{-t}(e^{tJ_0(\xi)}-1) e − t ( e t J 0 ( ξ ) − 1 ) , where J 0 ( ξ ) : = ∫ e − i x ⋅ ξ J ( x ) d x J_0(\xi):=\int e^{-ix\cdot\xi}J(x)\,dx J 0 ( ξ ) := ∫ e − i x ⋅ ξ J ( x ) d x tends to zero as ∣ ξ ∣ → ∞ |\xi|\to\infty ∣ ξ ∣ → ∞ by the Riemann–Lebesgue lemma. This multiplier is bounded by e T ∥ J ∥ L 1 + 1 e^{T\lVert J\rVert_{L^1}}+1 e T ∥ J ∥ L 1 + 1 and tends to zero as ∣ ξ ∣ → ∞ |\xi|\to\infty ∣ ξ ∣ → ∞ , uniformly in t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] . Since the Fourier transform of u k , N 0 u^0_{k,N} u k , N 0 is the translate by N η N\eta N η of a fixed L 2 L^2 L 2 function, dominated convergence shows that the last supremum tends to zero as N → ∞ N\to\infty N → ∞ , for fixed k k k . Choosing N = N k N=N_k N = N k large, we obtain unit data whose observation is at most ( 4 + 2 T ) / k (4+2T)/k ( 4 + 2 T ) / k , so (73) fails.
Corollary 7.2. Let V ≠ 0 V\ne0 V = 0 , e : = V / ∣ V ∣ 2 e:=V/|V|^2 e := V /∣ V ∣ 2 and Ω ∈ G ℓ , G \Omega\in\mathcal{G}_{\ell,G} Ω ∈ G ℓ , G , and consider the slab
S Ω : = { x ∈ R d : x ⋅ e ∈ Ω } . S_\Omega:=\{x\in\mathbb{R}^d:x\cdot e\in\Omega\}. S Ω := { x ∈ R d : x ⋅ e ∈ Ω } . For Ω = ⋃ j ∈ Z ( a j , b j ) \Omega=\bigcup_{j\in\mathbb{Z}}(a_j,b_j) Ω = ⋃ j ∈ Z ( a j , b j ) , let G ∘ : = sup j ∈ Z ( a j + 1 − b j ) ≤ G G^\circ:=\sup_{j\in\mathbb{Z}}(a_{j+1}-b_j)\le G G ∘ := sup j ∈ Z ( a j + 1 − b j ) ≤ G denote the supremum of its gap lengths, with G ∘ = 0 G^\circ=0 G ∘ = 0 if Ω = R \Omega=\mathbb{R} Ω = R .
(i) For every T > G ∘ / ∣ V ∣ 2 T>G^\circ/|V|_2 T > G ∘ /∣ V ∣ 2 there is C C C , depending only on d d d , T T T , V V V , J J J , ℓ \ell ℓ , G ∘ G^\circ G ∘ , such that
∥ u 0 ∥ L 2 2 ≤ C ∫ 0 T ∫ S Ω ∣ u ∣ 2 d x d t ( u 0 ∈ L 2 ( R d ) ) . \lVert u^0\rVert_{L^2}^2\le C\int_0^T\int_{S_\Omega}|u|^2\,dx\,dt\qquad(u^0\in L^2(\mathbb{R}^d)). ∥ u 0 ∥ L 2 2 ≤ C ∫ 0 T ∫ S Ω ∣ u ∣ 2 d x d t ( u 0 ∈ L 2 ( R d )) . For T > G / ∣ V ∣ 2 T>G/|V|_2 T > G /∣ V ∣ 2 , the constant can be chosen uniformly for all Ω ∈ G ℓ , G \Omega\in\mathcal{G}_{\ell,G} Ω ∈ G ℓ , G .
(ii) If G ∘ > 0 G^\circ>0 G ∘ > 0 , this inequality fails for every 0 < T ≤ G ∘ / ∣ V ∣ 2 0<T\le G^\circ/|V|_2 0 < T ≤ G ∘ /∣ V ∣ 2 .
(iii) The system ∂ t y + V ⋅ ∇ y = J ∗ y − y + 1 S Ω h \partial_t y+V\cdot\nabla y=J*y-y+1_{S_\Omega}h ∂ t y + V ⋅ ∇ y = J ∗ y − y + 1 S Ω h is exactly controllable between arbitrary states of L 2 ( R d ) L^2(\mathbb{R}^d) L 2 ( R d ) with h ∈ L 2 ( ( 0 , T ) × S Ω ) h\in L^2((0,T)\times S_\Omega) h ∈ L 2 (( 0 , T ) × S Ω ) if and only if T > G ∘ / ∣ V ∣ 2 T>G^\circ/|V|_2 T > G ∘ /∣ V ∣ 2 .
Proof. (i) Since G ∘ G^\circ G ∘ is the supremum of the gap lengths of Ω \Omega Ω , we have Ω ∈ G ℓ , G ∘ \Omega\in\mathcal{G}_{\ell,G^\circ} Ω ∈ G ℓ , G ∘ , and T > G ∘ / ∣ V ∣ 2 T>G^\circ/|V|_2 T > G ∘ /∣ V ∣ 2 . Lemma 3.2, applied with Λ = { − ∣ V ∣ 2 } \Lambda=\{-|V|_2\} Λ = { − ∣ V ∣ 2 } and the gap bound G ∘ G^\circ G ∘ , therefore provides the weight a Ω a_\Omega a Ω of (36) . Put a ( x ) : = a Ω ( x ⋅ e ) a(x):=a_\Omega(x\cdot e) a ( x ) := a Ω ( x ⋅ e ) . Then 0 ≤ a ≤ 1 S Ω 0\le a\le1_{S_\Omega} 0 ≤ a ≤ 1 S Ω , Lip ( a ) ≤ Lip ( a Ω ) \operatorname{Lip}(a)\le\operatorname{Lip}(a_\Omega) Lip ( a ) ≤ Lip ( a Ω ) , and
∫ 0 T a ( z + t V ) d t = m − ∣ V ∣ 2 , a Ω ( z ⋅ e ) ≥ m 0 > 0 ( z ∈ R d ) , \int_0^T a(z+tV)\,dt=m_{-|V|_2,a_\Omega}(z\cdot e)\ge m_0>0\qquad(z\in\mathbb{R}^d), ∫ 0 T a ( z + t V ) d t = m − ∣ V ∣ 2 , a Ω ( z ⋅ e ) ≥ m 0 > 0 ( z ∈ R d ) , where m 0 m_0 m 0 and Lip ( a Ω ) \operatorname{Lip}(a_\Omega) Lip ( a Ω ) depend only on T T T , ∣ V ∣ 2 |V|_2 ∣ V ∣ 2 , ℓ \ell ℓ , G ∘ G^\circ G ∘ . Theorem 7.1(i), applied with L = Lip ( a Ω ) L=\operatorname{Lip}(a_\Omega) L = Lip ( a Ω ) and μ = m 0 \mu=m_0 μ = m 0 , and the bound a ≤ 1 S Ω a\le1_{S_\Omega} a ≤ 1 S Ω give the asserted inequality, with C C C depending only on d d d , T T T , V V V , J J J , ℓ \ell ℓ , G ∘ G^\circ G ∘ . If T > G / ∣ V ∣ 2 T>G/|V|_2 T > G /∣ V ∣ 2 , the same lemma applied with the common bound G G G yields constants that are uniform over G ℓ , G \mathcal{G}_{\ell,G} G ℓ , G .
(ii) Let
m ( z ) : = ∫ 0 T 1 S Ω ( z + t V ) d t = ∣ ( z ⋅ e , z ⋅ e + T ∣ V ∣ 2 ) ∩ Ω ∣ ∣ V ∣ 2 . m(z):=\int_0^T1_{S_\Omega}(z+tV)\,dt=\frac{\left|(z\cdot e,z\cdot e+T|V|_2)\cap\Omega\right|}{|V|_2}. m ( z ) := ∫ 0 T 1 S Ω ( z + t V ) d t = ∣ V ∣ 2 ∣ ( z ⋅ e , z ⋅ e + T ∣ V ∣ 2 ) ∩ Ω ∣ . This function is ( 2 / ∣ V ∣ 2 ) (2/|V|_2) ( 2/∣ V ∣ 2 ) -Lipschitz, and its infimum is zero. Indeed, if T ∣ V ∣ 2 < G ∘ T|V|_2<G^\circ T ∣ V ∣ 2 < G ∘ , an interval of length T ∣ V ∣ 2 T|V|_2 T ∣ V ∣ 2 fits in a gap of Ω \Omega Ω ; if T ∣ V ∣ 2 = G ∘ T|V|_2=G^\circ T ∣ V ∣ 2 = G ∘ , then m ( z ) = ∣ V ∣ 2 − 1 m G ∘ ( z ⋅ e ) m(z)=|V|_2^{-1}m_{G^\circ}(z\cdot e) m ( z ) = ∣ V ∣ 2 − 1 m G ∘ ( z ⋅ e ) , and Lemma C.2 gives inf m = 0 \inf m=0 inf m = 0 , whether or not G ∘ G^\circ G ∘ is attained. Step 5 of the proof of Theorem 7.1, with a a a replaced by 1 S Ω 1_{S_\Omega} 1 S Ω , uses only these two properties of m m m , and yields unit data with vanishing observations.
(iii) The operator − V ⋅ ∇ + J -V\cdot\nabla+\mathcal{J} − V ⋅ ∇ + J generates a group on L 2 ( R d ) L^2(\mathbb{R}^d) L 2 ( R d ) . Its adjoint is V ⋅ ∇ + J ~ V\cdot\nabla+\widetilde{\mathcal{J}} V ⋅ ∇ + J , where J ~ f : = J ~ ∗ f − f \widetilde{\mathcal{J}}f:=\widetilde{J}*f-f J f := J ∗ f − f and J ~ ( x ) : = J ( − x ) \widetilde{J}(x):=J(-x) J ( x ) := J ( − x ) . With control space L 2 ( S Ω ) L^2(S_\Omega) L 2 ( S Ω ) , the control operator is extension by zero to L 2 ( R d ) L^2(\mathbb{R}^d) L 2 ( R d ) , and its adjoint is restriction to S Ω S_\Omega S Ω . After the
time reversal t ↦ T − t t \mapsto T-t t ↦ T − t , the adjoint evolution is therefore (71) with ( V , J ) (V,J) ( V , J ) replaced by ( − V , J ~ ) (-V,\widetilde{J}) ( − V , J ) , and ∥ J ~ ∥ L 1 = ∥ J ∥ L 1 \|\widetilde{J}\|_{L^1}=\|J\|_{L^1} ∥ J ∥ L 1 = ∥ J ∥ L 1 . Since − V / ∣ V ∣ 2 = − e -V/|V|^2=-e − V /∣ V ∣ 2 = − e , the slab is S Ω = { x : x ⋅ ( − e ) ∈ − Ω } S_{\Omega}=\{x:x\cdot(-e)\in-\Omega\} S Ω = { x : x ⋅ ( − e ) ∈ − Ω } , and − Ω ∈ G ℓ , G -\Omega\in\mathcal{G}_{\ell,\mathcal{G}} − Ω ∈ G ℓ , G has the same gap supremum G ∗ \mathcal{G}^* G ∗ . Parts (i) and (ii), applied to this adjoint system, therefore characterize its observability by T > G ∗ / ∣ V ∣ 2 T>\mathcal{G}^*/|V|^2 T > G ∗ /∣ V ∣ 2 . By Hilbert space duality [21 , 32 ] , surjectivity of the endpoint map is equivalent to this observability inequality.
Example 7.3 (No transport drift). If V = 0 V=0 V = 0 , then μ a = T inf a \mu_a=T\inf a μ a = T inf a . Theorem 7.1 therefore gives weighted observability exactly when inf a > 0 \inf a>0 inf a > 0 . For instance, the Lipschitz weight a ( x ) = min { 1 , ∣ x ∣ 2 } a(x)=\min\{1,|x|^2\} a ( x ) = min { 1 , ∣ x ∣ 2 } does not give observability, even when J J J is nonzero: nonlocal dispersion alone does not provide a uniform observation estimate near a zero of the weight.
Anisotropic heat flux beyond the Gramian assumptions
Let A ∈ R d × d A\in\mathbb{R}^{d\times d} A ∈ R d × d be symmetric positive definite, with eigenvalues 0 < α min ≤ ⋯ ≤ α max 0<\alpha_{\min}\leq\cdots\leq\alpha_{\max} 0 < α m i n ≤ ⋯ ≤ α m a x , and let u ( t ) = P t f : = e t ∇ ⋅ ( A ∇ ) f u(t)=P_tf:=e^{t\nabla\cdot(A\nabla)}f u ( t ) = P t f := e t ∇ ⋅ ( A ∇ ) f on R d \mathbb{R}^d R d . In an anisotropic conductor, the energy identity
∥ f ∥ 2 2 = ∥ P T f ∥ 2 2 + 2 ∫ 0 T ⟨ A ∇ u ( t ) , ∇ u ( t ) ⟩ d t (77) \|f\|_2^2=\|P_Tf\|_2^2+2\int_0^T\langle A\nabla u(t),\nabla u(t)\rangle\,dt \tag*{(77)} ∥ f ∥ 2 2 = ∥ P T f ∥ 2 2 + 2 ∫ 0 T ⟨ A ∇ u ( t ) , ∇ u ( t )⟩ d t (77)
involves the quadratic form ⟨ A ∇ u , ∇ u ⟩ \langle A\nabla u,\nabla u\rangle ⟨ A ∇ u , ∇ u ⟩ . Heat-flux sensors, however, measure Fourier’s flux q = − A ∇ u q=-A\nabla u q = − A ∇ u . We observe this flux on R d \mathbb{R}^d R d and the temperature on a measurable set Ω \Omega Ω . The resulting observability statement is elementary. The point of this section is structural: the measured outputs satisfy (R), but when A A A has two distinct eigenvalues their Gramian admits no decomposition of the form (H4)–(H5).
Theorem 8.1. Fix T > 0 T>0 T > 0 . There is a constant C C C such that
∥ f ∥ 2 2 ≤ C ∫ 0 T ( ∥ A ∇ u ( t ) ∥ 2 2 + ∥ 1 Ω u ( t ) ∥ 2 2 ) d t ( f ∈ L 2 ( R d ) ) (78) \|f\|_2^2\leq C\int_0^T\left(\|A\nabla u(t)\|_2^2+\|\mathbf{1}_{\Omega}u(t)\|_2^2\right)\,dt\qquad(f\in L^2(\mathbb{R}^d)) \tag*{(78)} ∥ f ∥ 2 2 ≤ C ∫ 0 T ( ∥ A ∇ u ( t ) ∥ 2 2 + ∥ 1 Ω u ( t ) ∥ 2 2 ) d t ( f ∈ L 2 ( R d )) (78) if and only if Ω \Omega Ω is thick, that is, inf h ∈ R d ∣ Ω ∩ ( h + [ 0 , L ] d ) ∣ > 0 \inf_{h\in\mathbb{R}^d}|\Omega\cap(h+[0,L]^d)|>0 inf h ∈ R d ∣Ω ∩ ( h + [ 0 , L ] d ) ∣ > 0 for some L > 0 L>0 L > 0 . On the weak-star hull W W W of 1 Ω \mathbf{1}_{\Omega} 1 Ω , the observation family
Γ b f = ( − A ∇ P t f , b 1 / 2 P t f ) , b ∈ W , \Gamma_bf=(-A\nabla P_tf,b^{1/2}P_tf),\qquad b\in W, Γ b f = ( − A ∇ P t f , b 1/2 P t f ) , b ∈ W , satisfies (H1)–(H3) and (R). If A A A has two distinct eigenvalues, then for no b ∈ W b\in W b ∈ W does Γ b ∗ Γ b \Gamma_b^*\Gamma_b Γ b ∗ Γ b admit a decomposition satisfying (H4)–(H5).
Proof. Let W W W be the weak-star closure of { 1 Ω ( ⋅ + h ) : h ∈ R d } \{\mathbf{1}_{\Omega}(\cdot+h):h\in\mathbb{R}^d\} { 1 Ω ( ⋅ + h ) : h ∈ R d } in L ∞ L^\infty L ∞ . Its elements satisfy 0 ≤ b ≤ 1 0\leq b\leq1 0 ≤ b ≤ 1 . The hull is compact and metrizable because L 1 ( R d ) L^1(\mathbb{R}^d) L 1 ( R d ) is separable, and the action θ h b = b ( ⋅ + h ) \theta_hb=b(\cdot+h) θ h b = b ( ⋅ + h ) is jointly continuous by translation continuity in L 1 L^1 L 1 . Covariance follows from translation invariance of the heat flow.
Step 1: energy normalization. Put Γ b 0 f = ( A 1 / 2 ∇ P t f , b 1 / 2 P t f ) \Gamma_b^0f=(A^{1/2}\nabla P_tf,b^{1/2}P_tf) Γ b 0 f = ( A 1/2 ∇ P t f , b 1/2 P t f ) . By (77) ,
( Γ b 0 ) ∗ Γ b 0 = 1 2 I + K b 0 , K b 0 = − 1 2 P 2 T + ∫ 0 T P t b P t d t . (\Gamma_b^0)^*\Gamma_b^0=\frac12I+K_b^0,\qquad K_b^0=-\frac12P_{2T}+\int_0^T P_tbP_t\,dt. ( Γ b 0 ) ∗ Γ b 0 = 2 1 I + K b 0 , K b 0 = − 2 1 P 2 T + ∫ 0 T P t b P t d t .
The heat kernel is p t ( x ) = ( 4 π t ) − d / 2 ( det A ) − 1 / 2 exp ( − x ⋅ A − 1 x / 4 t ) p_t(x)=(4\pi t)^{-d/2}(\det A)^{-1/2}\exp(-x\cdot A^{-1}x/4t) p t ( x ) = ( 4 π t ) − d /2 ( det A ) − 1/2 exp ( − x ⋅ A − 1 x /4 t ) , and ∥ ∇ P t ∥ 2 → 2 ≤ ( 2 e α min t ) − 1 / 2 \|\nabla P_t\|_{2\to2}\leq(2e\alpha_{\min}t)^{-1/2} ∥∇ P t ∥ 2 → 2 ≤ ( 2 e α m i n t ) − 1/2 . Hence sup 0 ≤ b ≤ 1 ∥ K b 0 ∥ L 2 → H 1 ≤ C T \sup_{0\leq b\leq1}\|K_b^0\|_{L^2\to H^1}\leq C_T sup 0 ≤ b ≤ 1 ∥ K b 0 ∥ L 2 → H 1 ≤ C T , and Rellich’s theorem gives (H4)–(H5) with d b = 1 2 d_b=\frac12 d b = 2 1 . The Gaussian bounds for p t p_t p t and ∇ p t \nabla p_t ∇ p t bound the blocks between cells at distance r ≥ 2 r\geq2 r ≥ 2 by C T e − c ( r − 1 ) 2 / T C_Te^{-c(r-1)^2/T} C T e − c ( r − 1 ) 2 / T , and the energy identity bounds the remaining blocks. This proves (H3). By Lemma 2.15, Γ b 0 \Gamma_b^0 Γ b 0 satisfies (R).
Step 2: the measured flux. We have Γ b = M Γ b 0 \Gamma_b=M\Gamma_b^0 Γ b = M Γ b 0 with M = diag ( − A 1 / 2 , 1 ) M=\operatorname{diag}(-A^{1/2},1) M = diag ( − A 1/2 , 1 ) . Lemma 2.20, applied with L = diag ( − A − 1 / 2 , 1 ) L=\operatorname{diag}(-A^{-1/2},1) L = diag ( − A − 1/2 , 1 ) , gives (H1), (H3) and (R) for Γ b \Gamma_b Γ b . The flux channel does not depend on b b b . On a cell Q Q Q , the energy of the temperature channel is the pairing of b b b with 1 Q ∫ 0 T ∣ P t f ∣ 2 d t ∈ L 1 \mathbf{1}_Q\int_0^T|P_tf|^2\,dt\in L^1 1 Q ∫ 0 T ∣ P t f ∣ 2 d t ∈ L 1 , which is weak-star continuous. This proves (H2); see Remark 2.21(ii).
Step 3: uniqueness and thickness. Let Γ b w = 0 \Gamma_b w=0 Γ b w = 0 with w ∈ L l o c 2 w\in L^2_{\mathrm{loc}} w ∈ L loc 2 . Since A A A is invertible, ∇ P t w = 0 \nabla P_tw=0 ∇ P t w = 0 , so P t w P_tw P t w is constant in space, and hence also in time by the heat equation. Since P t w → w P_tw\to w P t w → w in L l o c 2 L^2_{\mathrm{loc}} L loc 2 , the datum w w w is a constant c c c . Then b 1 / 2 c = 0 b^{1/2}c=0 b 1/2 c = 0 forces c = 0 c=0 c = 0 unless b = 0 b=0 b = 0 . If Ω \Omega Ω is thick, choose L , η > 0 L,\eta>0 L , η > 0 such that ∣ Ω ∩ ( h + [ 0 , L ] d ) ∣ ≥ η |\Omega\cap(h+[0,L]^d)|\geq\eta ∣Ω ∩ ( h + [ 0 , L ] d ) ∣ ≥ η for every h h h . Passing to weak-star limits preserves ∫ [ 0 , L ] d b ≥ η \int_{[0,L]^d}b\geq\eta ∫ [ 0 , L ] d b ≥ η , so 0 ∉ W 0\notin W 0 ∈ / W . Conversely, if thickness fails, choose h n h_n h n such that ∣ Ω ∩ ( h n + [ − n , n ] d ) ∣ < 1 / n |\Omega\cap(h_n+[-n,n]^d)|<1/n ∣Ω ∩ ( h n + [ − n , n ] d ) ∣ < 1/ n . Then 1 Ω ( ⋅ + h n ) → 0 \mathbf{1}_{\Omega}(\cdot+h_n)\to0 1 Ω ( ⋅ + h n ) → 0 in L l o c 1 L^1_{\mathrm{loc}} L loc 1 ; the uniform bound by one also gives weak-star convergence in L ∞ L^\infty L ∞ . Thus 0 ∈ W 0\in W 0 ∈ W . Theorem 2.1 now gives the equivalence.
Step 4: no Gramian decomposition. The flux part of Γ b ∗ Γ b \Gamma_b^*\Gamma_b Γ b ∗ Γ b is the Fourier multiplier
m T ( ξ ) = ξ ⋅ A 2 ξ 2 ξ ⋅ A ξ ( 1 − e − 2 T ξ ⋅ A ξ ) . m_T(\xi)=\frac{\xi\cdot A^2\xi}{2\,\xi\cdot A\xi}\left(1-e^{-2T\xi\cdot A\xi}\right). m T ( ξ ) = 2 ξ ⋅ A ξ ξ ⋅ A 2 ξ ( 1 − e − 2 T ξ ⋅ A ξ ) .
Let A e i = α i e i Ae_i=\alpha_ie_i A e i = α i e i , i = 1 , 2 i=1,2 i = 1 , 2 , with ∣ e i ∣ 2 = 1 |e_i|_2=1 ∣ e i ∣ 2 = 1 and α 1 ≠ α 2 \alpha_1\ne\alpha_2 α 1 = α 2 . Fix a unit vector φ ∈ C c ∞ ( Q ) \varphi\in C_c^\infty(Q) φ ∈ C c ∞ ( Q ) in L 2 L^2 L 2 , where Q Q Q is a bounded cube, and put f n , i = φ e i n e i ⋅ x f_{n,i}=\varphi e^{in e_i\cdot x} f n , i = φ e in e i ⋅ x . These sequences are weakly null. By Plancherel and dominated convergence, their flux energies tend to α i / 2 \alpha_i/2 α i /2 . The temperature Gramian maps L 2 L^2 L 2 into H 1 H^1 H 1 uniformly. Its compression to Q Q Q is therefore compact, so its contribution tends to zero. Suppose that Γ b ∗ Γ b = D b + K b \Gamma_b^*\Gamma_b=D_b+K_b Γ b ∗ Γ b = D b + K b satisfies (H4)–(H5). Since ∣ f n , i ∣ = ∣ φ ∣ |f_{n,i}|=|\varphi| ∣ f n , i ∣ = ∣ φ ∣ , both limits would equal ∫ Q d b ∣ φ ∣ 2 \int_Qd_b|\varphi|^2 ∫ Q d b ∣ φ ∣ 2 , because 1 Q K b 1 Q \mathbf{1}_QK_b\mathbf{1}_Q 1 Q K b 1 Q is compact. This gives α 1 = α 2 \alpha_1=\alpha_2 α 1 = α 2 , a contradiction.
Remark 8.2. If A = α I A=\alpha I A = α I , then m T ( ξ ) = α 2 ( 1 − e − 2 T α ∣ ξ ∣ 2 ) m_T(\xi)=\frac{\alpha}{2}(1-e^{-2T\alpha|\xi|^2}) m T ( ξ ) = 2 α ( 1 − e − 2 T α ∣ ξ ∣ 2 ) , and (H4)–(H5) hold directly. The obstruction is created by the anisotropy of the medium, not by an ad hoc choice of output weights. Measuring the plain gradient ∇ u \nabla u ∇ u instead of the flux leads to the same conclusion, with limits 1 / ( 2 α i ) 1/(2\alpha_i) 1/ ( 2 α i ) and M = diag ( A − 1 / 2 , 1 ) M=\operatorname{diag}(A^{-1/2},1) M = diag ( A − 1/2 , 1 ) .
Polynomial coercivity on time weights
Lemma A.1. Let r ≥ 1 r\ge1 r ≥ 1 , T > 0 T>0 T > 0 , and let a : ( 0 , T ) → [ 0 , 1 ] a:(0,T)\to[0,1] a : ( 0 , T ) → [ 0 , 1 ] be measurable, with m = ∫ 0 T a ( s ) d s > 0 m=\int_0^T a(s)\,\mathrm{d}s>0 m = ∫ 0 T a ( s ) d s > 0 . For vectors v 0 , … , v r − 1 v_0,\ldots,v_{r-1} v 0 , … , v r − 1 in a finite-dimensional Hilbert space, set P ( s ) = ∑ j = 0 r − 1 s j v j P(s)=\sum_{j=0}^{r-1}s^jv_j P ( s ) = ∑ j = 0 r − 1 s j v j . There is c r > 0 c_r>0 c r > 0 , depending only on r r r , such that
∫ 0 T a ( s ) ∣ P ( s ) ∣ 2 d s ≥ c r m 2 r − 1 ( 1 + T ) 2 r − 2 ∑ j = 0 r − 1 ∣ v j ∣ 2 . (79) \int_0^T a(s)|P(s)|^2\,\mathrm{d}s\ge\frac{c_r m^{2r-1}}{(1+T)^{2r-2}}\sum_{j=0}^{r-1}|v_j|^2. \tag*{(79)} ∫ 0 T a ( s ) ∣ P ( s ) ∣ 2 d s ≥ ( 1 + T ) 2 r − 2 c r m 2 r − 1 j = 0 ∑ r − 1 ∣ v j ∣ 2 . (79)
Factorials in the coefficients may be absorbed in c r c_r c r .
Proof. Write I = ∫ a ∣ P ∣ 2 I=\int a|P|^2 I = ∫ a ∣ P ∣ 2 . If I = 0 I=0 I = 0 , then P P P vanishes on a set of positive Lebesgue measure and hence is identically zero, component by component. Assume I > 0 I>0 I > 0 and let E = { s ∈ ( 0 , T ) : ∣ P ( s ) ∣ 2 ≤ 2 I / m } E=\{s\in(0,T):|P(s)|^2\le2I/m\} E = { s ∈ ( 0 , T ) : ∣ P ( s ) ∣ 2 ≤ 2 I / m } . Since the weighted mass of its complement is at most m / 2 m/2 m /2 , one has ∫ E a ≥ m / 2 \int_Ea\ge m/2 ∫ E a ≥ m /2 , and therefore ∣ E ∣ ≥ m / 2 |E|\ge m/2 ∣ E ∣ ≥ m /2 . Put d = m / ( 4 r ) d=m/(4r) d = m / ( 4 r ) . A greedy selection produces r r r points s 1 , … , s r s_1,\ldots,s_r s 1 , … , s r in E E E at pairwise distances at least d d d : before the last selection, the union of the forbidden intervals has measure at most 2 ( r − 1 ) d < m / 2 2(r-1)d<m/2 2 ( r − 1 ) d < m /2 . The vector-valued Lagrange formula reads
P ( s ) = ∑ i = 1 r P ( s i ) ∏ h = 1 h ≠ i r s − s h s i − s h . P(s)=\sum_{i=1}^{r}P(s_i)\prod_{\substack{h=1\\h\ne i}}^{r}\frac{s-s_h}{s_i-s_h}. P ( s ) = i = 1 ∑ r P ( s i ) h = 1 h = i ∏ r s i − s h s − s h .
Every denominator has modulus at least d r − 1 d^{r-1} d r − 1 , while every coefficient of a numerator is bounded by C r ( 1 + T ) r − 1 C_r(1+T)^{r-1} C r ( 1 + T ) r − 1 . Consequently
∑ j ∣ v j ∣ 2 ≤ C r ( 1 + T ) 2 r − 2 d − 2 r + 2 max i ∣ P ( s i ) ∣ 2 ≤ C r ( 1 + T ) 2 r − 2 m − 2 r + 1 I . \sum_j|v_j|^2\le C_r(1+T)^{2r-2}d^{-2r+2}\max_i|P(s_i)|^2\le C_r(1+T)^{2r-2}m^{-2r+1}I. j ∑ ∣ v j ∣ 2 ≤ C r ( 1 + T ) 2 r − 2 d − 2 r + 2 i max ∣ P ( s i ) ∣ 2 ≤ C r ( 1 + T ) 2 r − 2 m − 2 r + 1 I .
Rearrangement proves the assertion.
Pullback uniqueness along a family of characteristics
Lemma B.1. Let U ⊂ ( 0 , T ) × R U\subset(0,T)\times\mathbb{R} U ⊂ ( 0 , T ) × R be open, λ ∈ R \lambda\in\mathbb{R} λ ∈ R , and p λ ( s , x ) = x + λ s p_\lambda(s,x)=x+\lambda s p λ ( s , x ) = x + λ s . If p λ ( U ) = R p_\lambda(U)=\mathbb{R} p λ ( U ) = R and g ∈ D ′ ( R ) g\in\mathcal{D}'(\mathbb{R}) g ∈ D ′ ( R ) satisfies g ( x + λ s ) = 0 g(x+\lambda s)=0 g ( x + λ s ) = 0 on U U U , then g = 0 g=0 g = 0 .
Proof. Given z 0 ∈ R z_0\in\mathbb{R} z 0 ∈ R , choose ( s 0 , x 0 ) ∈ U (s_0,x_0)\in U ( s 0 , x 0 ) ∈ U with x 0 + λ s 0 = z 0 x_0+\lambda s_0=z_0 x 0 + λ s 0 = z 0 . In the coordinates ( s , z ) = ( s , x + λ s ) (s,z)=(s,x+\lambda s) ( s , z ) = ( s , x + λ s ) , the image of U U U contains a rectangle J × V J\times V J × V around ( s 0 , z 0 ) (s_0,z_0) ( s 0 , z 0 ) . Choose ρ ∈ C ∞ ( J ) \rho\in C^\infty(J) ρ ∈ C ∞ ( J ) with ∫ ρ = 1 \int\rho=1 ∫ ρ = 1 . For ϕ ∈ C ∞ ( V ) \phi\in C^\infty(V) ϕ ∈ C ∞ ( V ) the function η ( s , x ) = ρ ( s ) ϕ ( x + λ s ) \eta(s,x)=\rho(s)\phi(x+\lambda s) η ( s , x ) = ρ ( s ) ϕ ( x + λ s ) is supported in U U U . Testing there and changing variables yields
0 = ⟨ g ( x + λ s ) , η ⟩ = ⟨ g , ϕ ∫ J ρ ( s ) d s ⟩ = ⟨ g , ϕ ⟩ . 0=\langle g(x+\lambda s),\eta\rangle=\left\langle g,\phi\int_J\rho(s)\,\mathrm{d}s\right\rangle=\langle g,\phi\rangle. 0 = ⟨ g ( x + λ s ) , η ⟩ = ⟨ g , ϕ ∫ J ρ ( s ) d s ⟩ = ⟨ g , ϕ ⟩ .
Thus g g g vanishes near each z 0 z_0 z 0 , proving the claim componentwise.
Obstructions and the sharp geometric threshold
Proposition C.1. If the Kalman rank condition fails, no observation inequality for a nondegenerate norm on Schwartz data can hold, even with observation on the whole line. If A A A has a zero eigenvalue and Ω \Omega Ω has a nonempty open gap, the same conclusion holds for localized observation at every T > 0 T>0 T > 0 .
For a real-spectrum Kalman pair with nonzero speeds and a geometry (1.4) with G ∘ > 0 G^\circ>0 G ∘ > 0 , observability in Y K , J \mathcal{Y}_{K,J} Y K , J fails whenever 0 < T ≤ T ∗ 0<T\le T_\ast 0 < T ≤ T ∗ , with T ∗ T_\ast T ∗ as in (1.5).
The critical case T = T ∗ T=T_\ast T = T ∗ rests on the following elementary fact, which is also used in Theorem 6.1 and Corollary 7.2.
Lemma C.2 (Critical length). Let Ω \Omega Ω be as in (1.4) with G ∘ > 0 G^\circ>0 G ∘ > 0 , and let 0 < L ≤ G ∘ 0<L\le G^\circ 0 < L ≤ G ∘ . The function m L ( y ) : = ∣ ( y , y + L ) ∩ Ω ∣ m_L(y):=|(y,y+L)\cap\Omega| m L ( y ) := ∣ ( y , y + L ) ∩ Ω∣ is 2 2 2 -Lipschitz and inf y ∈ R m L ( y ) = 0 \inf_{y\in\mathbb{R}}m_L(y)=0 inf y ∈ R m L ( y ) = 0 , whether or not G ∘ G^\circ G ∘ is attained.
Proof. The symmetric difference of ( y , y + L ) (y,y+L) ( y , y + L ) and ( z , z + L ) (z,z+L) ( z , z + L ) has measure at most 2 ∣ y − z ∣ 2|y-z| 2∣ y − z ∣ . If ( b j , a j + 1 ) (b_j,a_{j+1}) ( b j , a j + 1 ) is a gap of length g j g_j g j , then m L ( b j ) ≤ ( L − g j ) + m_L(b_j)\le(L-g_j)_+ m L ( b j ) ≤ ( L − g j ) + . Since sup j g j = G ∘ ≥ L \sup_jg_j=G^\circ\ge L sup j g j = G ∘ ≥ L , the infimum of these bounds is zero.
Proof of Proposition C.1. Write s = T − t s=T-t s = T − t .
If Kalman fails, the Cayley–Hamilton theorem gives a nonzero vector v v v such that B T ( A T ) j v = 0 B^{\mathsf T}(A^{\mathsf T})^jv=0 B T ( A T ) j v = 0 for every j ≥ 0 j\ge0 j ≥ 0 . Choose a nonzero scalar Schwartz function h h h with compactly supported Fourier transform. Then
B T e i s ξ A T v h ^ ( ξ ) = 0 ( s ∈ [ 0 , T ] , ξ ∈ R ) , B^{\mathsf T}e^{\mathrm{i}s\xi A^{\mathsf T}}v\widehat{h}(\xi)=0\qquad(s\in[0,T],\ \xi\in\mathbb{R}), B T e i s ξ A T v h ( ξ ) = 0 ( s ∈ [ 0 , T ] , ξ ∈ R ) ,
so the nonzero terminal datum v h vh v h is invisible on the whole line. If instead A T v = 0 A^{\mathsf T}v=0 A T v = 0 for some v ≠ 0 v\ne0 v = 0 , choose h ∈ C c ∞ ( I ) ∖ { 0 } h\in C^\infty_{\mathrm{c}}(I)\setminus\{0\} h ∈ C c ∞ ( I ) ∖ { 0 } , where I I I is an open gap of Ω \Omega Ω . The solution with terminal datum v h vh v h is stationary and invisible.
For the last assertion, assume the stated Kalman and spectral hypotheses. Choose a real eigenvector v ≠ 0 v\ne0 v = 0 such that A T v = λ v A^{\mathsf T}v=\lambda v A T v = λ v and ∣ λ ∣ = λ ∙ |\lambda|=\lambda_\bullet ∣ λ ∣ = λ ∙ . Kalman implies B T v ≠ 0 B^{\mathsf T}v\ne0 B T v = 0 , and e s A T ∂ x ( v h ) ( x ) = v h ( x + λ s ) e^{sA^{\mathsf T}\partial_x}(vh)(x)=vh(x+\lambda s) e s A T ∂ x ( v h ) ( x ) = v h ( x + λ s ) . Let C T g l o b > 0 C_T^{\mathrm{glob}}>0 C T glob > 0 be a constant in the whole-line upper observation estimate. It follows that
T ∣ B T v ∣ 2 ∥ h ∥ L 2 2 = ∫ 0 T ∥ B T e s A T ∂ x ( v h ) ∥ L 2 2 d s ≤ C T g l o b ∥ v h ∥ Y κ , J 2 . T|B^{\mathsf T}v|^2\|h\|_{L^2}^2=\int_0^T\|B^{\mathsf T}e^{sA^{\mathsf T}\partial_x}(vh)\|_{L^2}^2\,\mathrm{d}s\le C_T^{\mathrm{glob}}\|vh\|_{\mathcal{Y}_{\kappa,\mathcal{J}}}^2. T ∣ B T v ∣ 2 ∥ h ∥ L 2 2 = ∫ 0 T ∥ B T e s A T ∂ x ( v h ) ∥ L 2 2 d s ≤ C T glob ∥ v h ∥ Y κ , J 2 .
Thus scalar packets with unit L 2 L^2 L 2 norm have intrinsic norms bounded away from zero, without any frequency modulation. For y ∈ R y\in\mathbb{R} y ∈ R , define
J λ ( y ) : = { y − λ s : 0 < s < T } , m λ ( y ) : = ∫ 0 T 1 Ω ( y − λ s ) d s = ∣ J λ ( y ) ∩ Ω ∣ ∣ λ ∣ . J_\lambda(y):=\{y-\lambda s:0<s<T\},\qquad m_\lambda(y):=\int_0^T\mathbf{1}_\Omega(y-\lambda s)\,\mathrm{d}s=\frac{|J_\lambda(y)\cap\Omega|}{|\lambda|}. J λ ( y ) := { y − λ s : 0 < s < T } , m λ ( y ) := ∫ 0 T 1 Ω ( y − λ s ) d s = ∣ λ ∣ ∣ J λ ( y ) ∩ Ω∣ .
Changing variables gives
∫ 0 T ∫ Ω ∣ B T e s A T ∂ x ( v h ) ( x ) ∣ 2 d x d s = ∣ B T v ∣ 2 ∫ R m λ ( y ) ∣ h ( y ) ∣ 2 d y . \int_0^T\int_\Omega|B^{\mathsf T}e^{sA^{\mathsf T}\partial_x}(vh)(x)|^2\,\mathrm{d}x\,\mathrm{d}s=|B^{\mathsf T}v|^2\int_{\mathbb{R}}m_\lambda(y)|h(y)|^2\,\mathrm{d}y. ∫ 0 T ∫ Ω ∣ B T e s A T ∂ x ( v h ) ( x ) ∣ 2 d x d s = ∣ B T v ∣ 2 ∫ R m λ ( y ) ∣ h ( y ) ∣ 2 d y .
If 0 < T < T ∙ 0<T<T_\bullet 0 < T < T ∙ , there is an open gap of length greater than ∣ λ ∣ T |\lambda|T ∣ λ ∣ T . The terminal points y y y for which J λ ( y ) J_\lambda(y) J λ ( y ) lies in that gap contain a nonempty open interval. A nonzero smooth packet supported in this interval has zero observation and positive intrinsic norm.
Now let T = T ∙ T=T_\bullet T = T ∙ , so that ∣ J λ ( y ) ∣ = G ∙ |J_\lambda(y)|=G^\bullet ∣ J λ ( y ) ∣ = G ∙ . The left endpoint of J λ ( y ) J_\lambda(y) J λ ( y ) is y − : = min { y , y − λ T } y_-:=\min\{y,y-\lambda T\} y − := min { y , y − λ T } , so m λ ( y ) = ∣ λ ∣ − 1 m G ∙ ( y − ) m_\lambda(y)=|\lambda|^{-1}m_{G^\bullet}(y_-) m λ ( y ) = ∣ λ ∣ − 1 m G ∙ ( y − ) with m G ∙ m_{G^\bullet} m G ∙ as in Lemma C.2. That lemma gives points y n y_n y n with m λ ( y n ) → 0 m_\lambda(y_n)\to0 m λ ( y n ) → 0 , whether or not the supremal gap length is attained, and
∣ m λ ( y ) − m λ ( z ) ∣ ≤ 2 ∣ y − z ∣ ∣ λ ∣ . |m_\lambda(y)-m_\lambda(z)|\le\frac{2|y-z|}{|\lambda|}. ∣ m λ ( y ) − m λ ( z ) ∣ ≤ ∣ λ ∣ 2∣ y − z ∣ .
Choose ε n ↓ 0 \varepsilon_n\downarrow0 ε n ↓ 0 and real functions
h n ∈ C c ∞ ( y n − ε n , y n + ε n ) , ∥ h n ∥ L 2 = 1. h_n\in C^\infty_{\mathrm{c}}(y_n-\varepsilon_n,y_n+\varepsilon_n),\qquad\|h_n\|_{L^2}=1. h n ∈ C c ∞ ( y n − ε n , y n + ε n ) , ∥ h n ∥ L 2 = 1.
On their supports,
m λ ( y ) ≤ δ n : = m λ ( y n ) + 2 ε n ∣ λ ∣ ⟶ 0. m_\lambda(y)\le\delta_n:=m_\lambda(y_n)+\frac{2\varepsilon_n}{|\lambda|}\longrightarrow0. m λ ( y ) ≤ δ n := m λ ( y n ) + ∣ λ ∣ 2 ε n ⟶ 0.
Consequently, the real Schwartz data f n : = v h n f_n:=vh_n f n := v h n satisfy
∥ f n ∥ Y κ , J 2 ≥ T ∣ B T v ∣ 2 C T g l o b > 0 , ∫ 0 T ∫ Ω ∣ B T e s A T ∂ x f n ∣ 2 d x d s ≤ ∣ B T v ∣ 2 δ n ⟶ 0. \|f_n\|_{\mathcal{Y}_{\kappa,\mathcal{J}}}^2\ge\frac{T|B^{\mathsf T}v|^2}{C_T^{\mathrm{glob}}}>0,\qquad\int_0^T\int_\Omega|B^{\mathsf T}e^{sA^{\mathsf T}\partial_x}f_n|^2\,\mathrm{d}x\,\mathrm{d}s\le|B^{\mathsf T}v|^2\delta_n\longrightarrow0. ∥ f n ∥ Y κ , J 2 ≥ C T glob T ∣ B T v ∣ 2 > 0 , ∫ 0 T ∫ Ω ∣ B T e s A T ∂ x f n ∣ 2 d x d s ≤ ∣ B T v ∣ 2 δ n ⟶ 0.
Dividing by their intrinsic norms gives unit-norm counterexamples to observability at T = T ∙ T=T_\bullet T = T ∙ , whether or not the supremal gap length is attained.
Acknowledgments
The author would like to express his sincere gratitude to Enrique Zuazua for his valuable discussions, insightful suggestions, and continuous encouragement throughout the development of this work.
The author gratefully acknowledges the hospitality of the Chair for Dynamics, Control, Machine Learning and Numerics at Friedrich-Alexander-Universität Erlangen–Nürnberg, where this research was carried out during a research stay. The author also acknowledges the financial support of the Alexander von Humboldt Foundation through an Alexander von Humboldt Research Fellowship.
References