We characterize observability of strongly coupled parabolic systems from arbitrary space-time measurable sets of positive measure. For constant coercive diffusion matrices and arbitrary constant zeroth-order coupling and observation matrices, observability holds if and only if the generalized Kalman rank condition is satisfied at every Dirichlet eigenvalue. The diffusion matrix need not be diagonalizable, and the number of components is arbitrary. The observation is measured in L1, yielding null controllability with bounded controls by duality. The proof combines a Remez-type inequality for exponential polynomials with a quantitative analysis of the mode-dependent Kalman condition. The resulting finite-dimensional observability estimate is uniform over the spatial modes up to a polynomial frequency loss. A spectral inequality and the Lebeau–Robbiano strategy then give observability of the full system.
Introduction and main results
Background and motivation
Let Ω⊂Rd be a bounded Lipschitz, locally star-shaped domain, and let T>0. We consider the following coupled parabolic system
for some α>0. This assumption ensures that the system is parabolic in the sense of Petrovskii and, in particular, that (1.1) is well-posed. More precisely, for every z0∈L2(Ω;Rm), there exists a unique solution z∈C([0,T];L2(Ω;Rm)). The system is referred to as strongly coupled since the coupling occurs through the principal part of the differential operator.
Let B∈Rm×r be an observation matrix, and let D⊂(0,T)×Ω be a measurable set of positive measure. Our main objective is to characterize the triples (A,M,B) for which there exists a constant C>0 such that
∥z(T)∥L2(Ω;Rm)2≤C∫D∣BTz(t,x)∣2dxdt,(3)
for every initial datum z0∈L2(Ω;Rm). This inequality expresses that the energy of the full system at time T can be estimated from partial observations on an arbitrary space-time measurable set of positive measure.
A natural starting point is the finite-dimensional counterpart
Key words and phrases. Strongly coupled parabolic systems, observability from measurable sets, Kalman rank condition, null-controllability. Given any measurable set E⊂(0,T) of positive measure, the corresponding observability inequality is
∣z(T)∣2≤C∫E∣B⊤z(t)∣2dt,∀z0∈Rm.(5)
It is a classical consequence of the Kalman observability criterion (see [7 Chapter 1]) that this inequality holds if and only if
rank[BMB⋯Mm−1B]=m.(6)
For the parabolic system (1.1), this finite-dimensional picture appears after projecting onto the Dirichlet eigenmodes of −Δ. More precisely, if (λk,ek) is a Dirichlet eigenpair and
zk(t):=∫Ωz(t,x)ek(x)dx,
then
zk′(t)yk(t)=(−λkA⊤+M⊤)zk(t),=B⊤zk(t),(7)
is the corresponding finite-dimensional system together with its observed output. The Kalman criterion applied to this system leads directly to the following frequency-dependent rank condition:
rank[B(−λkA+M)B⋯(−λkA+M)m−1B]=m,∀k≥1.(8)
The purpose of this paper is to show that this condition characterizes observability of the full parabolic system, even when observations are restricted to an arbitrary measurable subset of space-time.
The observability of coupled parabolic systems has been extensively studied over the past decades. We briefly review some key results, with particular emphasis on observability from measurable sets, to highlight the main motivation and novelty of the present work.
Scalar case. For the scalar heat equation, corresponding to m=1, the coupling issue disappears. We may assume that A=a>0 and B=b=0. The observability inequality (1.3) was first established for cylindrical observation sets D=(0,T)×ω, with ω⊂Ω a nonempty open subset, in the seminal works [13, 16]. Several intermediate results retained a product structure in space and time. Wang [24] obtained bounded null controls supported on E×ω, where E⊂(0,T) has positive measure and ω⊂Ω is nonempty and open. Phung and Wang [20] established related observability estimates for parabolic equations from such sets, using quantitative unique continuation at a single time. Under analyticity assumptions, Apraiz and Escauriaza [4] treated measurable spatial observation sets, and Zhang [25] proved observability for the heat equation on products E×ω with both factors measurable and of positive measure, for domains with real-analytic boundary. The product structure was subsequently removed by Apraiz, Escauriaza, Wang, and Zhang [3], who established observability from arbitrary space-time measurable sets of positive measure.
A further direction concerns spatial sets of zero Lebesgue measure. Burq and Moyano [6] established spectral inequalities and heat observability from spatial sets of positive (d−δ)-dimensional Hausdorff measure, for sufficiently small δ>0, allowing elliptic operators with Lipschitz coefficients. Green, Le Balc’h, Martin, and Orsoni [14] proved observability of the heat equation on bounded C1 domains from spatial sets of Hausdorff dimension strictly greater than d−1, with the observation measured by a spatial supremum integrated in time. Huang, G. Wang, and M. Wang [15] subsequently refined this geometric criterion using logarithmic-type Hausdorff contents. Their observability estimates also cover certain observation sets of Hausdorff dimension exactly d−1.
Weakly coupled case. A prototypical example of a weakly coupled parabolic system is when A=Im and M∈Rm×m. Here, the coupling only occurs through the zeroth-order term and does not affect the principal part of the differential operator. For cylindrical observation sets D=(0,T)×ω, with ω nonempty and open, Ammar Khodja et al. [1] established observability under the classical Kalman rank condition
rank[BMB⋯Mm−1B]=m.
We refer to [2] for a survey of controllability results for coupled parabolic systems and of the role played by Kalman-type conditions. Escauriaza, Montaner, and Zhang [10] subsequently obtained observability from arbitrary space-time measurable sets of positive measure for a class of weakly coupled systems with two components.
First-order couplings have also been studied for controls supported on (0,T)×ω, with ω nonempty and open. Duprez and Lissy [8] characterized null and approximate controllability for a class of constant-coefficient systems with m equations and m−1 controls, allowing both zeroth- and first-order couplings. Their approach combines the fictitious control method, algebraic solvability, and Carleman estimates. For two equations with variable coefficients, they obtained further sufficient conditions and counterexamples in [9]. Extending our measurable-set result to first-order couplings would require additional arguments: spatial derivatives do not, in general, preserve the Dirichlet eigenspaces, so the independent modal systems used here are no longer available.
Strongly coupled case. We now return to the system (1.1), where the coupling occurs through the principal part. For cylindrical observation sets D=(0,T)×ω, with ω nonempty and open, Lissy and Zuazua [18] established observability under the generalized Kalman rank condition (1.8). Their argument can also be adapted to measurable spatial sets ω⊂Ω of positive measure, by using the spectral inequality from measurable sets established in [3].
More recently, Fu, Wang, Yu, and Zhu [12] addressed the significantly more delicate question of observability from arbitrary space-time measurable sets. They considered the particular two-component system (1.1) with
AT=(ab−ba),B=(10),MT=0,a>0,b=0.
A major difficulty in this setting is that high-frequency oscillations induced by the coupling may lead to cancellations in the observed component. Consequently, the pointwise-in-time interpolation inequalities used to derive observability from measurable sets in the scalar and weakly coupled settings [3, 10] may fail, as demonstrated in [12 Proposition 3.3].
To overcome this obstruction, the authors of [12] established an integral interpolation inequality based on a Remez-type estimate, thereby avoiding the need for pointwise-in-time interpolation inequalities. Their proof relies on an explicit representation of the solutions, made possible by the particular structure of the coupling matrix.
A related quantitative problem was recently studied by Zhu [26] for diffusion matrices aI+εK, where a>0 and K is skew-symmetric, with observation on arbitrary measurable space-time sets. Under a Kalman rank condition, that work determines sharp blow-up rates of the observability cost as ε→0, as well as the corresponding minimum-norm control costs. At the level of individual modes, [26 Proposition 2.6] already uses the Turán–Nazarov inequality to pass to measurable time sets in arbitrary dimension.
Building on the integral interpolation approach of [12], we prove a complete characterization under the generalized Kalman rank condition (1.8) for general coercive diffusion matrices and arbitrary zeroth-order coupling. This includes non-diagonalizable diffusion matrices and rank conditions that depend on the spatial mode. The main quantitative ingredient is a uniform finite-dimensional observability estimate with polynomial dependence on the spatial frequency. It combines a consequence of Nazarov’s inequality for exponential polynomials with a quantitative analysis of the mode-dependent Kalman matrix. Together with the Lebeau–Robbiano strategy, this yields observability from arbitrary space-time measurable sets.
Main results
We are now in a position to state our main result.
Theorem 1.1.Assume the parabolicity condition(2). Then the following assertions are equivalent.
For everyT>0and every measurable setD⊂(0,T)×Ωwith∣D∣>0, there exists a constantC=C(A,M,B,T,D,Ω)>0such that every solution of(1)satisfies
∥z(T)∥L2(Ω;Rm)≤C∫D∣BTz(t,x)∣dxdt.(9)
Remark 1.2. We make a few comments on Theorem 1.1.
Theorem 1.1 extends the parabolic observability result of Lissy and Zuazua [18] from cylindrical sets (0,T)×ω, with ω nonempty and open, to arbitrary space-time measurable sets of positive measure, under the same generalized Kalman rank condition. It also extends [12 Theorem 1.1] from their two-component model to arbitrary coupling and observation matrices satisfying the parabolicity and Kalman rank conditions.
The observability inequality (9) is stronger than the standard L2-L2 observability inequality (3), since the observation is measured in L1(D) rather than L2(D). We state the result in this form because the L1 estimate naturally arises in our proof.
The necessity of the generalized Kalman rank condition (8) for the system is straightforward. Indeed, assume that (8) fails for some index j0≥1. Then there exists v∈Rm∖{0} such that
BT(−λj0AT+MT)kv=0,k=0,…,m−1.
It follows from the Cayley–Hamilton theorem that BTet(−λj0AT+MT)v=0 for every t≥0. Let ej0 be a nonzero Dirichlet eigenfunction of −Δ with eigenvalue λj0>0. Then
z(t,x)=et(−λj0AT+MT)vej0(x)
solves (1) and satisfies BTz=0 in (0,T)×Ω, while z(T)=0. This contradicts (9). Thus, the main contribution of the paper is the proof of sufficiency.
By duality, Theorem 1.1 yields null controllability with bounded controls supported on arbitrary measurable sets.
Corollary 1.3.Assume(2). The Kalman rank condition(8)holds if and only if, for everyT>0, every measurable setD⊂(0,T)×Ωof positive measure, and everyy0∈L2(Ω;Rm), there existsu∈L∞(D;Rr)such that the solution of
Proof. Let φ solve the backward adjoint equation −∂tφ−ATΔφ=MTφ, with zero Dirichlet data and terminal datum φT∈L2(Ω;Rm). Applying Theorem 1.1 to z(t)=φ(T−t) on the reflected set {(T−t,x):(t,x)∈D} gives
∥φ(0)∥L2(Ω;Rm)≤C∥BTφ∥L1(D;Rr).
Thus the functional BTφ↦−⟨y0,φ(0)⟩ is well defined and bounded on a subspace of L1(D;Rr), with norm at most C∥y0∥L2. The Hahn–Banach theorem and L1–L∞ duality give a representing control u satisfying (11) and
−⟨y0,φ(0)⟩=∫Du⋅BTφdxdt.
The duality identity
⟨y(T),φT⟩=⟨y0,φ(0)⟩+∫Du⋅BTφdxdt=0
then yields y(T)=0. Conversely, if (8) fails at an eigenvalue λk, there is a nonzero v such that BTet(−λkAT+MT)v=0 for all t≥0. The corresponding backward mode φ(t,x)=e(T−t)(−λkAT+MT)vek(x) has BTφ=0 and φ(0)=0. Taking y0=φ(0) in the duality identity precludes y(T)=0. □
Example 1.4. Consider a system with three components and scalar observation, with
A=110011001,M=000010002,B=100.
The matrix A is non-diagonalizable, with a single Jordan block of size three, and satisfies ⟨AX,X⟩≥(1−1/2)∣X∣2 for all X∈R3. For Dλ=−λA+M, direct computation gives
Hence (8) holds at every Dirichlet eigenvalue. Theorem 1.1 therefore yields observability from any measurable space-time set of positive measure by observing only the first component BTz=z1. This example lies outside both the two-component model of [12] and the diffusion family aI+εK with K skew-symmetric considered in [26].
Strategy of the proofs
The main difficulty is to combine partial observation of the components with observation on an arbitrary measurable subset of space-time. As shown in [12], the oscillations generated by the principal coupling can prevent pointwise-in-time interpolation estimates. We therefore work with time-integrated observations. The new ingredient is a quantitative estimate for general matrix evolutions, which treats non-diagonalizable couplings and controls the dependence on the spatial frequency.
Uniform propagation from measurable time sets. After projection onto a Laplace eigenmode and rescaling time, the observed quantity is BTe−tGjη, where Gj=AT−MT/λj. Its scalar components are exponential polynomials, with polynomial factors accounting for the Jordan blocks. Proposition 3.2 extends Nazarov’s inequality to these functions by approximating the polynomial factors with finite differences of exponentials. The number of frequencies remains at most m, and the constants are uniform in j. In particular, the estimate does not require uniform bounds on Jordan bases or separation of the eigenvalues. This replaces the special trigonometric representation used in [12] and permits general coupling matrices. For comparison, the Carleman approach of Fernández-Cara, González-Burgos, and de Teresa [11 Theorem 1.1] for non-diagonalizable diffusion matrices and space-time dependent zeroth-order coefficients requires Jordan blocks of size at most four. In our constant-matrix setting, the exponential-polynomial argument imposes no restriction on the size of the Jordan blocks.
Direct quantitative observability of each mode. Quantitative estimates for fast controls in finite dimension provide an important background for this step. Seidman [21] identified the small-time growth of the minimum L2 control cost in terms of the Kalman structure; Seidman and Yong [22] treated more general Lq control norms. Related constructive estimates based on rank conditions can be found in [5, 23]. In the coupled parabolic setting, Lissy and Zuazua [18 Lemma 3.1] already track both the time horizon and the mode parameter in finite-dimensional observability estimates used in their Lebeau–Robbiano argument. Here too, a small-time estimate for a fixed matrix is insufficient: we must control simultaneously the interval length ℓ, the measurable observation set E, and the spatial frequency λj. In particular, the constant hidden in a fixed-matrix estimate cannot be allowed to depend arbitrarily on j. The estimate below makes these dependencies explicit for general coercive diffusion matrices and arbitrary zeroth-order coupling.
The central estimate is Proposition 3.4. Its proof separates recovery of the observation derivatives from recovery of the state. Cayley–Hamilton gives a scalar differential equation, with uniformly bounded coefficients, for each observed component. A Taylor expansion, together with polynomial norm equivalence, then bounds the first m observation derivatives by the time integral of the observation. Using the weighted sum of derivatives ∑k=0m−1ℓk∣f(k)(0)∣ makes the Taylor remainder absorbable on an interval whose length is independent of the conditioning of the Kalman matrix. The state is subsequently recovered from these derivatives using the rank condition. Since the entries of the observability matrix are polynomial in 1/λj, a nonzero minor yields a loss bounded by (1+λj)m(m−1). Combining the two estimates with the Remez inequality gives
Thus all modes are covered by the same observability argument, with no additional negative power of ℓ. The polynomial dependence on frequency is essential for the subsequent spectral argument.
From individual modes to the full system. Minkowski’s inequality combines the mode estimates without a loss depending on the number of retained modes. The spatial spectral inequality from measurable sets then gives the low-frequency observation estimate. Together with high-frequency parabolic dissipation, it yields the integral interpolation inequality of Theorem 3.1 by the Lebeau–Robbiano strategy. Finally, a density-point and telescoping argument, following [3, 12, 20], gives Theorem 1.1. These final steps use established methods. The contribution lies in the uniform matrix estimates that allow them to apply under the general Kalman rank condition.
Preliminaries
In this section, we first introduce the basic notation used throughout the paper. We then recall several lemmas that are crucial for the proof of Theorem 1.1 and can be found in the literature.
Let N+:={1,2,…} and R+:=(0,+∞). For d∈N+, r∈R+, and x0∈Rd, we denote by Brd(x0) the closed ball in Rd centered at x0 with radius r. In particular, we write Brd:=Brd(0). For any Lebesgue measurable set E⊂Rd, ∣E∣ denotes its Lebesgue measure. Given a matrix Q∈Rm×n with m,n∈N+, we denote by QT its transpose. For a Banach space X, ∥⋅∥X denotes its norm. In the finite-dimensional Euclidean space Rd, the norm ∥⋅∥Rd is the standard Euclidean norm; when no confusion can arise, we abbreviate it as ∣⋅∣. If X is a Hilbert space, ⟨⋅,⋅⟩X denotes its inner product.
Throughout the paper, the dependence of constants on the fixed domain Ω is left implicit unless otherwise stated.
We next recall the following technical lemma, which is standard in the derivation of observability inequalities for the heat equation from space-time measurable sets; see [3, 12].
Lemma 2.1. Suppose that T>0 and BRd(x0)⊂Ω for some R>0 and x0∈Ω. Let D⊂(0,T)×BRd(x0) be a measurable set with ∣D∣>0. Set
Dt:={x∈Ω:(t,x)∈D}for t∈(0,T),(12)
E:={t∈(0,T):∣Dt∣≥2T∣D∣}.(13)
Then, the following statements hold:
:::`
Dt is a measurable subset of Ω for a.e. t∈(0,T).
For any f∈L1((0,T)×Ω), the function t↦∫ΩχDt(x)∣f(t,x)∣dx is integrable over (0,T).
E is a measurable subset of (0,T) and ∣E∣≥2∣BRd∣∣D∣.
χE(t)χDt(x)≤χD(t,x) in (0,T)×Ω.
Let {λj}j∈N+, with 0<λ1<λ2≤⋯ and λj→+∞ as j→+∞, be the eigenvalues of −Δ under the homogeneous Dirichlet boundary condition. Let {ej}j≥1 be the corresponding Dirichlet eigenfunctions of −Δ on Ω, constituting an orthonormal basis of L2(Ω).
The following lemma will also be used to prove Theorem 1.1.
Lemma 2.2. Suppose that T>0 and B4Rd(x0)⊂Ω for some R>0 and x0∈Ω. Let D⊂(0,T)×BRd(x0) be measurable with ∣D∣>0. Let E and Dt be defined as in Lemma 2.1. For every fixed Λ1>0, there exists a constant C(R,∣D∣/T)>0 such that, for almost every t∈E, every family {aj}⊂Rr, and every Λ>Λ1,
Squaring this estimate and inserting it into (2.4) yields (2.3). □
Integral interpolation inequality
This section is devoted to proving the following integral-type interpolation observability inequality for (1.1), which plays a crucial role in the proof of Theorem 1.1.
We write S(t)=et(A⊤Δ+M⊤) for the solution semigroup with homogeneous Dirichlet boundary conditions. In this section, for an initial datum z∈L2(Ω;Rm), we use the convention z(t)=S(t)z.
We fix Λ1≥max{1,2∥M∥/α} and use the observation sets constructed in Section 2. More precisely, let D⊂(0,T)×BRd(x0) have positive measure, with B4Rd(x0)⊂Ω, and set
Dt={x∈Ω:(t,x)∈D},E={t∈(0,T):∣Dt∣≥2T∣D∣}.
By Lemma 2.1, these slices are measurable for almost every t, the corresponding observation integrals are measurable, and ∣E∣>0. Lemma 2.2 provides a constant D=D(R,∣D∣/T)>0 such that, for almost every t∈E, every Λ>Λ1, and every family {aj}⊂Rr,
These properties are preserved when E is replaced by a measurable subset of positive measure, or when the time interval and the observation pair (E,Dt) are translated together. In both cases, the spectral constant D is inherited from the original pair; it is not recomputed from the restricted set. The result below applies to these pairs as well. All constants may depend on the fixed domain Ω.
Theorem 3.1. Assume (1.2) and (1.8), and let (E,Dt) be an observation pair described above. There exists a constant C8:=C8(A,M,B,Λ1,m,D,Ω)>0 such that, for every θ∈(0,1), every 0<S1<S2≤T satisfying ∣E∩[S1,S2]∣>0, and every z∈L2(Ω;Rm),
We split z(t)=PΛz(t)+(I−PΛ)z(t) into low and high frequencies, where PΛ is the spectral projection
PΛu=λj≤Λ∑⟨u,ej⟩ej,u∈L2(Ω).(18)
The projection PΛ acts componentwise on vector-valued functions and commutes with constant matrices.
In Section 3.1, we prove the low-frequency estimate (3.26) for ∥PΛz(t)∥L2(Ω;Rm). To achieve this, we first establish in Proposition 3.2 a Remez–type inequality for functions of the form f(t)=∑ν=1se−μνtPν(t), where Pν is a polynomial and μ1,…,μs are the distinct eigenvalues of A−(M∈Rm×m. Based on Proposition 3.2 and the generalized Kalman rank condition (1.8), we prove the unified observability estimate in Proposition 3.4, which is the key ingredient in the proof of (3.26).
In Section 3.2, we combine the low-frequency estimate (3.26) with the Lebeau–Robbiano strategy to prove Theorem 3.1.
For each z∈L2(Ω;R2), let {zj}j∈N+:={(z1,j,z2,j)T}j∈N+⊂R2 denote the Fourier coefficients of z with respect to the orthonormal basis {ej}j∈N+. Then
and used it to prove the low-frequency estimate. However, for general matrices A,M∈Rm×m and B∈Rm×r, simple solution formulas such as (19) are no longer available. For each z∈L2(Ω;Rm), the solution of (1) has the expansion
denote the Fourier coefficients of z with respect to the orthonormal basis {ej}j∈N+.
We first analyze the form of each component of BTe−(AT−λjMT)tη for η∈Rm, and then establish the corresponding Remez-type inequality.
Set
Gj:=AT−λjMT.
For all j≥1, we have the uniform bound
μ∈σ(Gj)max∣Reμ∣≤∥Gj∥≤∥A∥+λ1∥M∥=:KG.(22)
Proposition 3.2. Let μ1,…,μs be the distinct eigenvalues of Gj∈Rm×m, with algebraic multiplicities d1,…,ds, respectively, so that ∑ν=1sdν=m. For η∈Rm, the following statements hold.
(i) Any scalar component of BTe−tGjη can be written as
f(t)=ν=1∑se−μνtPν(t),ν=1∑s(degPν+1)≤m,(23)
where each Pν is a polynomial with complex coefficients; zero polynomials are omitted from the sum of degrees.
(ii) Let j≥1, let I⊂R be an interval with 0<∣I∣≤1, and let E⊂I be a measurable subset with ∣E∣>0. Then there exists a constant C=C(A,M,m,λ1,r)>0, independent of j, such that
Proof. (i) By the Jordan canonical form, there exists an invertible matrix P∈Cm×m such that Gj=PJP−1, where J=diag(J1,…,Js), where Jν groups all Jordan blocks associated with μν. Thus Jν has size dν and satisfies
Jν=μνIdν+Nν,Nνdν=0,1≤ν≤s,(26)
Here Nν is the direct sum of the corresponding nilpotent Jordan blocks; it need not be a single shift matrix.
By the Jordan decomposition,
e−tGj=Pe−JtP−1=Pdiag(e−J1t,…,e−Jst)P−1,
where
e−Jνt=e−μνtk=0∑dν−1k!(−t)kNνk.(27)
The expansion of the exponential and the identity Nνdν=0 show that each component of e−Jνt is of the form e−μνtPν(t), with degPν≤dν−1. Hence, any scalar component of BTe−tGjη can be written as
where aν,k∈C. Omitting zero polynomials from the sum of degrees, we obtain
ν=1∑s(degPν+1)≤m.
(ii) Our strategy is as follows: we construct an exponential polynomial fε of order N such that
fε→funiformly on I.
We apply Nazarov’s inequality from [19] to fε and then pass to f by an approximation argument.
Lemma 3.3 (Nazarov’s inequality). Let
p(t)=j=1∑Ncjeαjt
be an exponential polynomial of order N, where αj∈C and cj∈C. Let I⊂R be an interval, and let E⊂I be a measurable subset with positive Lebesgue measure. Then there exists a constant C>0, depending only on N, such that
which is an exponential polynomial with at most m distinct frequencies. Indeed, for each fixed ν, the index r ranges from 0 to k, and since k≤dν−1, we have 0≤r≤dν−1. Hence, after collecting terms with identical frequencies, the total number Nε of distinct frequencies in fε satisfies
Nε≤ν=1∑sdν=m.
Moreover, the uniform convergence pν,k,ε→tke−μνt on I as ε→0+ (cf. (3.15)) implies
fε⟶funiformly on I.
For 0<ε≤1, the real parts of the frequencies are uniformly bounded by maxν∣Re(μν)∣+m. Applying (3.13) to fε yields
For every j≥1, (3.7) and ∣I∣≤1 show that the exponential factor in the preceding inequality is bounded uniformly with respect to j. Thus, taking the limit ε→0+ yields the generalized estimate
Here, f≍g means that there exists a constant C>0 such that f≤Cg and g≤Cf. Then (3.10) follows from (3.19) and (3.20). □
We next establish a direct observability estimate valid for every mode. Throughout the remainder of this section, constants may depend on the fixed matrices, their dimensions, and Ω, but not on the mode index. Set p=m(m−1) and ω=∥M∥.
Proposition 3.4.Assume (1.2) and (1.8). The following estimates hold for everyj≥1, with a constantC=C(A,M,B,m,Ω)>0. For0<ℓ≤1, every measurableE⊂(0,ℓ)of positive measure, and everyη∈Rm,
∣η∣≤∣E∣mC(1+λj)p∫E∣BTe−tGjη∣dt.(36)
For everyL>0, every measurableF⊂(0,L)of positive measure, and everyξ∈Rm,
On the scalar polynomial space of degree at most m−1, the map q↦∫01∣q(s)∣ds is a norm. Each derivative evaluation at zero is a continuous linear functional on this finite-dimensional space, so
∣q(k)(0)∣≤Cm∫01∣q(s)∣ds,0≤k≤m−1.
Apply this to the vector polynomial Q(s)=P(ℓs) as follows. For each k with f(k)(0)=0, choose the unit vector vk=f(k)(0)/∣f(k)(0)∣ and use the scalar polynomial qk(s)=vk⋅Q(s). Since qk(k)(0)=ℓk∣f(k)(0)∣ and ∣qk(s)∣≤∣Q(s)∣, we get
ℓk∣f(k)(0)∣≤Cm∫01∣Q(s)∣ds.
The same inequality is immediate for a zero derivative. Summing over k and changing variables gives Jℓ≤mCmℓ−1∫0ℓ∣P(t)∣dt, as claimed.
We next estimate the remainder. Using its integral representation and the bound on f(m) obtained above, we have, for 0≤t≤ℓ,
This proves (38) on the whole range 0<ℓ≤1. In particular, ℓ0 depends only on m and K, and not on B or the conditioning of the Kalman matrix.
Step 2: Quantitative recovery of the state. Consider the polynomial matrix
O(s)=BTBT(AT−sMT)⋮BT(AT−sMT)m−1.
By the Kalman rank condition (8), O(1/λj) has rank m for every j.
Fix once and for all an m×m minor S(s) of O(s) whose determinant d(s) is not identically zero. Such a minor exists because O(1/λj) has rank m. This choice is fixed throughout the argument, independently of the mode index j; hence the coefficients of S and d depend only on the fixed matrices A,M,B. Each row has degree at most m−1, and the determinant is a sum of products of m entries; hence degd≤m(m−1)=p. Factor out its zero at the origin:
d(s)=sqh(s),0≤q≤p,h(0)=0.
The polynomial h is also fixed, rather than chosen separately for each j. By continuity, we may choose s0∈(0,1], depending only on h, such that ∣h(s)∣≥∣h(0)∣/2 for 0≤s≤s0. Set
C0=∣h(0)∣20≤s≤s0max∥adjS(s)∥<∞.
The cofactor formula gives
∥S(s)−1∥=sq∣h(s)∣∥adjS(s)∥≤C0s−q,0<s≤s0.
Thus s0 and C0 depend on the fixed matrices A,M,B through the chosen minor, but neither depends on j nor on the evaluation point s=1/λj. It follows that, for λj≥s0−1,
For the finitely many indices with λj<s0−1, define
κj=∣η∣=1min∣O(1/λj)η∣>0.
Positivity follows from injectivity and compactness of the unit sphere. If this finite set is nonempty, its minimum κ∗>0 satisfies ∣η∣≤κ∗−1∣O(1/λj)η∣ for each of these modes. Enlarging C thus gives
∣η∣≤C(1+λj)pk=0∑m−1∣BTGjkη∣,j≥1.(40)
The dependence on Ω includes the lower bound λ1 and these finitely many singular-value bounds. This is the only step that distinguishes small and large modes.
Step 3: Observation on measurable sets. Since f(k)(0)=(−1)kBTGjkη, (3.24) and (38) imply
The factors ℓ−m and ℓm cancel, yielding (36). Both estimates are uniform over all modes.
Step 4: Arbitrary time intervals. Partition (0,L) into N=⌈L⌉ intervals of equal length at most one. One of them, I=(a,b), satisfies ∣F∩I∣≥∣F∣/N. Applying (36) to e−aGjξ and the translated set (F∩I)−a gives
∣e−aGjξ∣≤C(1+λj)p(∣F∣N)m∫F∩I∣B⊤e−tGjξ∣dt.
To propagate this estimate to time L, let v(t)=e−tGjv0. The coercivity assumption (2), which also holds for A⊤, gives
Combining this bound with the preceding estimate, N≤1+L, and F∩I⊂F proves (37). If λjα≥ω, the first exponential factor is at most one, so the propagation is contractive and no exponential loss is needed. □
We now use Proposition 3.4 to estimate the low-frequency part.
Theorem 3.5.Assume(2)and(8). There existsC=C(A,M,B,m,Ω)>0such that, for0<S1<S2≤T, every measurableE⊂(0,T)withδ=∣E∩(S1,S2)∣>0, everyΛ>0, and every initial datumz∈L2(Ω;Rm),
For λj≤Λ, (43) implies ∣aj∣≤HΛ∫E∣bj(t)∣dt, where
HΛ=∣E∣mCeωΔ(1+Λ)p+1(1+Δ)m.
Minkowski’s integral inequality then gives
λj≤Λ∑∣aj∣21/2≤HΛ∫Eλj≤Λ∑∣bj(t)∣21/2dt.
Finally, Parseval’s identity yields
BTPΛz(t)L2(Ω;Rr)2=λj≤Λ∑∣bj(t)∣2,(45)
and the corresponding identity for PΛz(S2) proves (42). □
Proof of Theorem 3.1
Proof. We use the low-frequency estimate (42) in three steps. Throughout this proof, ∥⋅∥q denotes the Lq(Ω) norm with values in the appropriate Euclidean space. Set
Here C5>0 depends only on the fixed parameters. For the coefficients bj(t) defined in (44), the spectral inequality (16) gives, for almost every t∈E,
λj≤Λ∑∣bj(t)∣2≤DeDΛχDtBTPΛz(t)12.(48)
Here we used ∑λj≤Λbj(t)ej=BTPΛz(t). Insert (48) into (42) and enlarge C5 to obtain (46).
Step 2: Frequency splitting. For θ∈(0,1), we seek an estimate of the form
∥z(S2)∥2≤δm−1C7(ε(Λ)−γδI(z)+ε(Λ)∥z∥2),(49)
where γ=θ/(1−θ) and
ε(Λ)=e−c(Λ−Λ1)S1/2,Λ>Λ1.(50)
By (2) and the choice of Λ1, the high-frequency part satisfies
∥(I−PΛ)z(t)∥2≤Che−cΛt∥z∥2,t≥0,Λ>Λ1,(51)
with Ch,c>0 depending only on A and M. The triangle inequality, (46), and (51) yield
Step 3: Optimization. Write X=∥z(S2)∥2, Y=I(z)/δ, Z=∥z∥2, and Π=C7/δm−1. As Λ ranges over (Λ1,∞), ε(Λ) ranges over (0,1), so (3.33) reads
X≤Π(ε−γY+εZ),0<ε<1.
Since δ≤S2 and C6≥1, we have Π≥eωS2. The energy estimate X≤eωS2Z therefore extends this bound to ε≥1. For Y,Z>0, choose ε=(Y/Z)1−θ to get
∥z(S2)∥2≤δm−12C7(δI(z))1−θ∥z∥2θ.
The same conclusion holds if Y or Z vanishes, by a limiting argument or uniqueness. Finally, (3.37) gives 2C7≤C8exp(C8[S2/(1−θ)+1/(θS1)]) for a constant C8 depending only on the fixed parameters. This proves (3.2). □
Proof of Theorem 1.1
We follow the telescoping argument of [3, 12, 20], using Theorem 3.1. In this proof, ∥⋅∥q denotes the Lq(Ω) norm with values in the appropriate Euclidean space.
Step 1: Localization. By [12 Lemma 2.3], there are (t0,x0)∈D and R>0 such that
Since BRd(x0)⊂Ω and DR⊂(0,T)×BRd(x0), this is an observation pair of Section 3. Lemmas 2.1 and 2.2 give ∣ER∣>0 and (3.1) with the fixed constant D=D(R,∣DR∣/T). We will also use
χER(t)χ(DR)t(x)≤χDR(t,x).(55)
Step 2: Interpolation. We recall the following density-point lemma from [3, 17, 20].
Lemma 4.1. Let E⊂(0,T) have positive measure and let ℓ be a density point of E. For every μ>1, there exists ℓ1∈(ℓ,T) such that
Restrict ER to (ℓj+1,ℓj) and translate time by ℓj+2. The resulting observation pair has the same spectral constant D, independently of j. Apply Theorem 3.1 to z(ℓj+2+⋅) with S1=ℓj+1−ℓj+2 and S2=ℓj−ℓj+2. Changing back to the original time variable yields
Sum over even j≥2. The remainder tends to zero, because z∈C([0,T];L2) and the exponential weights tend to zero. The integration intervals are disjoint, hence
X. Chen is grateful to Professor Gengsheng Wang and Professor Shanlin Huang for helpful discussions, particularly regarding the proof and construction of Proposition 3.2.
Disclosure of AI assistance
The authors used AI tools for proofreading and reformulating parts of the manuscript, preparing the literature discussion and locating references, and for assistance in simplifying their initial proof of Proposition 3.4. That initial proof was inspired by the approach of Lissy and Zuazua [18] and treated low and high frequencies separately. The AI-assisted simplification led to the unified observability argument presented here. The authors take full responsibility for the mathematical content and the final text of the article.
References
School of Mathematics (Zhuhai), Sun Yat-Sen University, Zhuhai 519082, Guangdong, China
[1]F. Ammar-Khodja, A. Benabdallah, C. Dupaix, and M. González-Burgos, A Kalman rank condition for the localized distributed controllability of a class of linear parabolic systems, J. Evol. Equ. 9 (2009), no. 2, 267–291.
[2]F. Ammar-Khodja, A. Benabdallah, M. González-Burgos, and L. de Teresa, Recent results on the controllability of linear coupled parabolic problems: A survey, Math. Control Relat. Fields 1 (2011), no. 3, 267–306.DOI
[3]J. Apraiz, L. Escauriaza, G. Wang, and C. Zhang, Observability inequalities and measurable sets, J. Eur. Math. Soc. 16 (2014), no. 11, 2433–2475.arxiv.org/abs/1202.4876
[4]J. Apraiz and L. Escauriaza, Null-control and measurable sets, ESAIM Control Optim. Calc. Var. 19 (2013), no. 1, 239–254.arxiv.org/abs/1107.4132
[5]D. Azé and O. Cârjă, Fast controls and minimum time, Control Cybernet. 29 (2000), 887–894.
[6]N. Burq and I. Moyano, Propagation of smallness and control for heat equations, J. Eur. Math. Soc. 25 (2023), no. 4, 1349–1377.arxiv.org/abs/1912.07402
[7]J.-M. Coron, Control and Nonlinearity, Mathematical Surveys and Monographs, vol. 136, American Mathematical Society, Providence, RI, 2007.
[8]M. Duprez and P. Lissy, Indirect controllability of some linear parabolic systems of m equations with m − 1 controls involving coupling terms of zero or first order, J. Math. Pures Appl. 106 (2016), no. 5, 905–934.DOI
[9]M. Duprez and P. Lissy, Positive and negative results on the internal controllability of parabolic equations coupled by zero- and first-order terms, J. Evol. Equ. 18 (2018), no. 2, 659–680.
[10]L. Escauriaza, S. Montaner, and C. Zhang, Observation from measurable sets for parabolic analytic evolutions and applications, J. Math. Pures Appl. 104 (2015), no. 5, 837–867.
[11]E. Fernández-Cara, M. González-Burgos, and L. de Teresa, Controllability of linear and semilinear non-diagonalizable parabolic systems, ESAIM Control Optim. Calc. Var. 21 (2015), no. 4, 1178–1204.
[12]X. Fu, G. Wang, H. Yu, and X. Zhu, Observability from measurable sets for strongly coupled parabolic systems via single-component observation, Preprint, arXiv:2604.13599 (2026).arxiv.org/abs/2604.13599
[13]A. V. Fursikov and O. Yu. Imanuvilov, Controllability of Evolution Equations, Lecture Notes Series, vol. 34, Seoul National University, Seoul, 1996.
[14]A. W. Green, K. Le Balc’h, J. Martin, and M.-A. Orsoni, On the dimension of observable sets for the heat equation, Pure Appl. Anal. 7 (2025), no. 3, 639–660.
[15]S. Huang, G. Wang, and M. Wang, Observability inequality, log-type Hausdorff content and heat equations, Comm. Math. Phys. 407 (2026), no. 2, article 33, doi:10.1007/s00220-025-05529-6.DOI
[16]G. Lebeau and L. Robbiano, Contrôle exact de l’équation de la chaleur, Comm. Partial Differential Equations 20 (1995), no. 1–2, 335–356.
[17]J.-L. Lions, Optimal Control of Systems Governed by Partial Differential Equations, Grundlehren der mathematischen Wissenschaften, vol. 170, Springer-Verlag, Berlin, 1971.
[18]P. Lissy and E. Zuazua, Internal observability for coupled systems of linear partial differential equations, SIAM J. Control Optim. 57 (2019), no. 2, 832–853.DOI
[19]F. L. Nazarov, Local estimates for exponential polynomials and their applications to inequalities of the uncertainty principle type, Algebra i Analiz 5 (1993), no. 4, 3–66; English transl., St. Petersburg Math. J. 5 (1994), 663–717.
[20]K. D. Phung and G. Wang, An observability estimate for parabolic equations from a measurable set in time and its applications, J. Eur. Math. Soc. 15 (2013), no. 2, 681–703.DOI
[21]T. I. Seidman, How violent are fast controls?, Math. Control Signals Systems 1 (1988), no. 1, 89–95.
[22]T. I. Seidman and J. Yong, How violent are fast controls?, II, Math. Control Signals Systems 9 (1996), no. 4, 327–340.
[23]R. Triggiani, Constructive steering control functions for linear systems and abstract rank conditions, J. Optim. Theory Appl. 74 (1992), 347–367.DOI
[24]G. Wang, L∞-null controllability for the heat equation and its consequences for the time optimal control problem, SIAM J. Control Optim. 47 (2008), no. 4, 1701–1720.DOI
[25]C. Zhang, An observability estimate for the heat equation from a product of two measurable sets, J. Math. Anal. Appl. 396 (2012), no. 1, 7–12.DOI
[26]X. Zhu, Sharp observability costs for strongly coupled parabolic systems with vanishing principal coupling, Preprint, arXiv:2609.35087 (2026).arxiv.org/abs/2609.35087