Introduction
Many devices perform well only within a narrow window of an internal variable that nobody measures directly. A thermally tuned optical filter is a representative case [1 , 2 ] . Its transmission peaks at one temperature of the filter element. This temperature shifts with fabrication tolerances and ageing. The signal that is convenient to measure online is the transmitted optical power itself. The actuator, a resistive heater along one edge of the supporting plate, is not collocated with the element. The heat it releases must diffuse across the plate while part of it leaks to the surroundings. The control task is therefore to drive the heater so that the filter sits at its transmission peak, through a medium governed by a diffusion equation, using one scalar performance signal and no thermometer.
Extremum-seeking control is the natural tool for a problem of this kind. It adjusts a plant input using measured performance when the optimizing operating condition is unknown. Input dithers reveal local sensitivity, while averaging and singular perturbation provide classical analysis tools [3 ] . Lie-bracket approximations address a related class of oscillatory optimization laws [4 ] . Proportional–integral designs combine sensitivity feedback with input-bias adaptation [5 , 6 ] . Dual-mode designs also address the input needed to sustain the operating condition [7 ] , and phasor-based methods exploit structural information about the plant response [8 ] .
Distributed dynamics attenuate and phase-shift the perturbation between actuation and observation, and the sensitivity estimate must account for both. Earlier applications include convection–reaction reactors [9 , 10 ] . For diffusion actuation, probing-signal design and boundary compensation have been combined with infinite-dimensional stability analysis [11 ] . Extensions address multivariable PDE dynamics [12 ] , reaction–advection–diffusion equations [13 ] , spatially integrated outputs [14 ] , unbiased designs [15 ] , and physics-informed probing [16 ] . Broader developments appear in [17 , 18 ] . These designs compensate the actuation dynamics online from a known model of the PDE.
This paper takes a different route, suggested by the physics of the thermal problem. When the heater power is modulated sinusoidally, the temperature at the heated edge responds as a thermal wave. Its amplitude falls with the inverse square root of the frequency and its phase lags the heat flux by π / 4 \pi/4 π /4 . These are classical properties of diffusion at a periodically heated surface [19 ] . They hold whatever the conductivity, the heat loss, and the plate size, which enter only through a single amplitude factor carrying the sign of the input gain. Placing the sensor on the actuated edge therefore hands the extremum-seeking loop a response whose phase is known before the controller is switched on, and the demodulation can be fixed once and for all. These properties permit a controller with two internal states using only the measured cost. The feedback law requires no numerical PDE coefficients, state-field measurement, or online PDE compensator. Signal injection plays a related role in finite-dimensional virtual-measurement constructions [20 ] , but boundary actuation and point observation require additional trace estimates here.
A design that uses no model of the plate must be shown to work on the plate, and the contribution is a constructive link between averaged stability and implemented performance. We first prove exponential stability of the averaged closed loop with its finite-bandwidth sensitivity filter. A longitudinal reaction contribution incorporates existing damping, and a local derivative bound controls the remaining nonlinearities. An exact harmonic lifting, spatial smoothing, and a boundary-trace argument then give trajectory closeness for all time. The resulting local practical exponential regulation has decreasing state and cost bounds as frequency increases, even at fixed dither amplitude. The comparison and its constants are developed directly, within the broader context of infinite-dimensional averaging [21 ] and input-to-state stability [22 ] .
The objective is regulation toward an attainable local minimum of a boundary-output cost. The theorem is specific to the stated geometry and nonlinear margin. The thermal example illustrates controller performance and spatial regulation, while explicitly distinguishing its observed frequency trends from the conservative sufficient certificate.
Sections II–III specify the problem and controller. Averaged stability is established in Section IV. Sections V–VIII develop the lifting, regularity, comparison, and practical stability arguments. Frequency selection and the thermal filter example conclude the analysis. Readers interested mainly in the design and its performance can pass from Sections II–III directly to Sections VIII–X. Table 1 collects the constants that appear in the final estimates and says where each is defined.
Table 1. Main constants of the analysis and where they are defined.
Problem Formulation
Figure 1(a) shows the physical situation we have in mind. A thin square plate carries a small optical element at the midpoint of one edge. A resistive heater runs along that edge. The power it delivers may be shaped along the edge but is set by a single command u u u . Heat spreads through the plate by conduction, leaks to the surroundings through the two broad faces, and the three remaining edges are insulated. The transmission of the element depends on its own temperature and peaks at a value that is not known in advance, Figure 1 (b), and the only measurement is the transmitted optical power, summarized by a score J J J that is smallest at the peak. Neither the temperature field nor the temperature of the element is measured.
Figure 1. (a) Thin plate heated along one edge, with the optical element at the midpoint of that edge. The three other edges are insulated and heat leaks through the two broad faces. (b) Normalized transmission of the element and the score J = 1 / T − 1 J=1/T-1 J = 1/ T − 1 against the detuning of the element temperature from its optimum, in units of the temperature scale Θ \Theta Θ . Only J J J is measured.
Scaling lengths by the side of the plate, time by its diffusion time, and temperature by the width of the transmission peak turns the plate into the unit square and the temperature excess w w w above ambient into the state of a semilinear diffusion equation with boundary actuation. Section X gives the details. In this dimensionless form, and allowing for a reaction term and a fixed heat load that the thermal model does not need but that cost nothing in the analysis, the plant reads as follows. Let Ω = ( 0 , 1 ) 2 \Omega=(0,1)^2 Ω = ( 0 , 1 ) 2 , s = ( s 1 , s 2 ) s=(s_1,s_2) s = ( s 1 , s 2 ) , and consider
w t = κ Δ w − μ 0 w + F ( s , w ) + d ( s ) , w_t=\kappa\Delta w-\mu_0w+F(s,w)+d(s), w t = κ Δ w − μ 0 w + F ( s , w ) + d ( s ) ,
− κ w s 1 ( 0 , s 2 , t ) = b q ( s 2 ) u ( t ) , -\kappa w_{s_1}(0,s_2,t)=bq(s_2)u(t), − κ w s 1 ( 0 , s 2 , t ) = b q ( s 2 ) u ( t ) ,
w s 1 ( 1 , s 2 , t ) = 0 , w_{s_1}(1,s_2,t)=0, w s 1 ( 1 , s 2 , t ) = 0 ,
w s 2 ( s 1 , 0 , t ) = w s 2 ( s 1 , 1 , t ) = 0 , (1) w_{s_2}(s_1,0,t)=w_{s_2}(s_1,1,t)=0, \tag*{(1)} w s 2 ( s 1 , 0 , t ) = w s 2 ( s 1 , 1 , t ) = 0 , (1)
where κ , μ 0 > 0 \kappa,\mu_0>0 κ , μ 0 > 0 , b ≠ 0 b\ne0 b = 0 , u ∈ R u\in\mathbb{R} u ∈ R , and q ( s 2 ) = 1 + ϑ cos ( π s 2 ) q(s_2)=1+\vartheta\cos(\pi s_2) q ( s 2 ) = 1 + ϑ cos ( π s 2 ) with ϑ ∈ R \vartheta\in\mathbb{R} ϑ ∈ R . The fixed source d d d may be unknown. Here κ \kappa κ is the scaled diffusivity, μ 0 \mu_0 μ 0 accounts for the heat lost through the faces, q q q is the distribution of heater power along the actuated edge, and b b b is the input gain. The sign of b b b is unknown such that the design also covers cooling actuators and sensors of either polarity. Only the noise-free cost
J ( t ) = φ ( C b w ( t ) ) , C b w = w ( 0 , 1 2 ) , (2) J(t)=\varphi(C_b w(t)),\qquad C_b w=w\left(0,\frac{1}{2}\right), \tag*{(2)} J ( t ) = φ ( C b w ( t )) , C b w = w ( 0 , 2 1 ) , (2)
is measured. The field and its boundary value are not separately available.
Since the heater acts through the boundary and the initial temperature need not match the boundary condition, the analysis needs a precise notion of solution. Let X = C ( Ω ‾ ; R ) \mathcal{X}=C(\overline{\Omega};\mathbb{R}) X = C ( Ω ; R ) be the continuous real-valued fields, with ∥ f ∥ ∞ = max s ∈ Ω ‾ ∣ f ( s ) ∣ \lVert f\rVert_\infty=\max_{s\in\overline{\Omega}}|f(s)| ∥ f ∥ ∞ = max s ∈ Ω ∣ f ( s ) ∣ . The distinct symbol C b : X → R C_b:\mathcal{X}\to\mathbb{R} C b : X → R denotes point observation, so ∥ C b ∥ = 1 \lVert C_b\rVert=1 ∥ C b ∥ = 1 . For 0 < γ < 1 0<\gamma<1 0 < γ < 1 , use
∥ f ∥ C γ = ∥ f ∥ ∞ + [ f ] γ , [ f ] γ = sup s ≠ s ~ ∣ f ( s ) − f ( s ~ ) ∣ ∣ s − s ~ ∣ γ . \lVert f\rVert_{C^\gamma}=\lVert f\rVert_\infty+[f]_\gamma,\qquad[f]_\gamma=\sup_{s\ne\widetilde{s}}\frac{|f(s)-f(\widetilde{s})|}{|s-\widetilde{s}|^\gamma}. ∥ f ∥ C γ = ∥ f ∥ ∞ + [ f ] γ , [ f ] γ = s = s sup ∣ s − s ∣ γ ∣ f ( s ) − f ( s ) ∣ .
Let A = κ Δ − μ 0 A=\kappa\Delta-\mu_0 A = κ Δ − μ 0 be the homogeneous Neumann generator, S ( t ) S(t) S ( t ) its semigroup, S − 1 , S − 1 ( t ) S_{-1},S_{-1}(t) S − 1 , S − 1 ( t ) their extrapolations. Let B B B be the scalar boundary-command operator in the extrapolation space. Write K B ( t ) = S − 1 ( t ) B \mathcal{K}_B(t)=S_{-1}(t)B K B ( t ) = S − 1 ( t ) B and F ( x ) = F ( ⋅ , x ( ⋅ ) ) F(x)=F(\cdot,x(\cdot)) F ( x ) = F ( ⋅ , x ( ⋅ )) . Solutions have the mild interpretation
w ( t ) = S ( t − t 0 ) w 0 + ∫ t 0 t S ( t − r ) [ F ( w ( r ) ) + d ] d r w(t)=S(t-t_0)w_0+\int_{t_0}^{t}S(t-r)[F(w(r))+d]\,\mathrm{d}r w ( t ) = S ( t − t 0 ) w 0 + ∫ t 0 t S ( t − r ) [ F ( w ( r )) + d ] d r
+ ∫ t 0 t K B ( t − r ) u ( r ) d r . (3) \qquad+\int_{t_0}^{t}\mathcal{K}_B(t-r)u(r)\,\mathrm{d}r. \tag*{(3)} + ∫ t 0 t K B ( t − r ) u ( r ) d r . (3)
The kernel estimates below establish local integrability in X \mathcal{X} X . This formulation permits continuous initial fields without Neumann compatibility [23 ] .
Assumption 1. The source d d d is continuous. The functions F , F w F,F_w F , F w are jointly continuous near the equilibrium field range and its prescribed enlargement. There are H F , L F < ∞ H_F,L_F<\infty H F , L F < ∞ and 0 < γ < 1 0<\gamma<1 0 < γ < 1 such that
∣ F ( s , ζ ) − F ( s ~ , ζ ~ ) ∣ ≤ H F ∣ s − s ~ ∣ γ + L F ∣ ζ − ζ ~ ∣ (4) |F(s,\zeta)-F(\widetilde{s},\widetilde{\zeta})|
\le H_F|s-\widetilde{s}|^\gamma
+L_F|\zeta-\widetilde{\zeta}| \tag*{(4)} ∣ F ( s , ζ ) − F ( s , ζ ) ∣ ≤ H F ∣ s − s ∣ γ + L F ∣ ζ − ζ ∣ (4)
on that region. The cost φ \varphi φ is C 2 C^2 C 2 on an open interval containing the corresponding boundary values.
Assumption 1 asks only for Hölder continuity in space and Lipschitz continuity in temperature of the reaction, and for a twice differentiable cost. Both are mild for the thermal application, where the reaction is absent and the cost is a smooth function of the element temperature.
Assumption 2. There is a stationary pair ( w ⋆ , u ⋆ ) (w^\star,u^\star) ( w ⋆ , u ⋆ ) of (1) whose observed value z ⋆ = C b w ⋆ z^\star=C_b w^\star z ⋆ = C b w ⋆ satisfies φ ′ ( z ⋆ ) = 0 \varphi'(z^\star)=0 φ ′ ( z ⋆ ) = 0 and h : = φ ′ ′ ( z ⋆ ) > 0 h:=\varphi''(z^\star)>0 h := φ ′′ ( z ⋆ ) > 0 .
In the thermal system, Assumption 2 says that some constant heater power holds the element exactly at the temperature where its transmission peaks, and that the peak is not flat. Thus the scalar output minimum is attainable at steady state. The objective is local regulation of the associated field and bias, not optimization of an arbitrary steady input-to-cost map. Neither the equilibrium nor the cost derivatives are known to the controller. For the nominal estimates, fix H 0 < ∞ H_0<\infty H 0 < ∞ and require
∥ w 0 − w ⋆ ∥ C γ ≤ H 0 . (5) \lVert w_0-w^\star\rVert_{C^\gamma}\le H_0. \tag*{(5)} ∥ w 0 − w ⋆ ∥ C γ ≤ H 0 . (5)
The initial error will also be small in X \mathcal{X} X . This fixed family controls the integrable startup layer of the measured trace.
The design relies on one physical observation. A sinusoidal modulation of the heater power launches a thermal wave into the plate, and at the frequencies of interest the temperature at the heated edge responds with an amplitude that decays like ω − 1 / 2 \omega^{-1/2} ω − 1/2 and a phase that lags the heat flux by π / 4 \pi/4 π /4 , the classical signature of diffusion at a periodically heated surface [19 ] . Conductivity, heat loss, and plate size do not alter this leading behavior; they only set the amplitude. In the present notation, for the linear diffusion equation with F = d = 0 F=d=0 F = d = 0 , the midpoint transfer function is
H b ( i ω ) = b κ μ 0 + i ω coth μ 0 + i ω κ . (6) H_b(i\omega)=\frac{b}{\sqrt{\kappa}\sqrt{\mu_0+i\omega}}\coth\sqrt{\frac{\mu_0+i\omega}{\kappa}}. \tag*{(6)} H b ( iω ) = κ μ 0 + iω b coth κ μ 0 + iω . (6)
Here the square roots are taken with positive real part, and the transverse cosine mode of q q q does not appear because it vanishes at the midpoint. § 5 proves
H b ( i ω ) = χ ω − 1 / 2 e − i π / 4 + O ( ω − 3 / 2 ) (7) H_b(i\omega)=\chi\omega^{-1/2}e^{-i\pi/4}+O(\omega^{-3/2}) \tag*{(7)} H b ( iω ) = χ ω − 1/2 e − iπ /4 + O ( ω − 3/2 ) (7)
where χ = b κ \chi=\frac{b}{\sqrt{\kappa}} χ = κ b . The controller uses the known diffusion factor after separating χ \chi χ . Changing the sign of b b b changes the total transfer phase by π \pi π , but does not change the phase of this factor. The demodulation phase is thus fixed in advance, and χ \chi χ is absorbed into a loop gain shown to be proportional to the positive quantity b 2 κ \frac{b^2}{\sqrt{\kappa}} κ b 2 in § 4 .
Controller and Averaged Target
Fix a 0 > 0 a_0>0 a 0 > 0 , a dither amplitude a ∈ ( 0 , a 0 ] a\in(0,a_0] a ∈ ( 0 , a 0 ] , and λ > 0 \lambda>0 λ > 0 . Choose gains satisfying
0 < k I < λ k P . (8) 0<k_I<\lambda k_P. \tag*{(8)} 0 < k I < λ k P . (8)
Set θ = ω ( t − t 0 ) + θ 0 \theta=\omega(t-t_0)+\theta_0 θ = ω ( t − t 0 ) + θ 0 and ε = ω − 1 / 2 \varepsilon=\omega^{-1/2} ε = ω − 1/2 . The two-state controller is
u = v − k P g ^ + a sin θ , v ˙ = − k I g ^ , (9) u=v-k_P\widehat{g}+a\sin\theta,\qquad\dot{v}=-k_I\widehat{g}, \tag*{(9)} u = v − k P g + a sin θ , v ˙ = − k I g , (9)
with the implemented filter
m = 2 λ ε a sin ( θ + π / 4 ) , m ˙ = 2 λ a ε cos ( θ + π / 4 ) , g ^ = p + m J , p ˙ = − λ p − ( λ m + m ˙ ) J . (10) \begin{aligned}
m&=\frac{2\lambda\varepsilon}{a}\sin(\theta+\pi/4),\qquad\dot{m}=\frac{2\lambda}{a\varepsilon}\cos(\theta+\pi/4),\\
\widehat{g}&=p+mJ,\qquad\dot{p}=-\lambda p-(\lambda m+\dot{m})J. \tag*{(10)}
\end{aligned} m g = a 2 λ ε sin ( θ + π /4 ) , m ˙ = a ε 2 λ cos ( θ + π /4 ) , = p + m J , p ˙ = − λ p − ( λm + m ˙ ) J . (10)
Only the cost (2), the states ( v , p ) (v,p) ( v , p ) , and the dither are used. The known derivative m ˙ \dot{m} m ˙ is computed from time. For absolutely continuous cost traces, the equations imply
g ^ ˙ = − λ g ^ + m J ˙ . (11) \dot{\widehat{g}}=-\lambda\widehat{g}+m\dot{J}. \tag*{(11)} g ˙ = − λ g + m J ˙ . (11)
This is an analysis identity, not an implementation requiring cost differentiation. § 6 supplies the needed trace regularity.
The controller is a proportional–integral extremum seeking controller. The bias v v v is the slowly adapted heater power, − k P g ^ -k_P\widehat{g} − k P g is a proportional correction, and a sin θ a\sin\theta a sin θ is the probing signal. The filter (10) estimates the sensitivity of the cost to the heater power. By (11), it multiplies the rate of change of the measured cost by a copy of the dither advanced by π / 4 \pi/4 π /4 , which is exactly in phase with the rate of change of the cost ripple produced by the thermal wave. It passes the product through a first-order low-pass filter of bandwidth λ \lambda λ . The factor ε / a \varepsilon/a ε / a in m m m normalizes the product. The differentiation of a ripple of amplitude a χ φ ′ a\chi\varphi' a χ φ ′ produces a factor a ω 1 / 2 a\omega^{1/2} a ω 1/2 , that is canceled by ε / a \varepsilon/a ε / a .
Define the signed sensitivity
g ( x ) = χ φ ′ ( C b x ) . (12) g(x)=\chi\varphi'(C_bx). \tag*{(12)} g ( x ) = χ φ ′ ( C b x ) . (12)
At a frozen boundary value z z z , the leading cost oscillation is a χ φ ′ ( z ) sin ( θ − π / 4 ) a\chi\varphi'(z)\sin(\theta-\pi/4) a χ φ ′ ( z ) sin ( θ − π /4 ) . Multiplication of its derivative by m m m has mean λ χ φ ′ ( z ) \lambda\chi\varphi'(z) λ χ φ ′ ( z ) . This explains the modulation in (10), and the exact argument follows below.
Using the sensitivity in (12), the averaged target retains the finite filter bandwidth:
x ˉ ˙ = κ Δ x ˉ − μ 0 x ˉ + F ( s , x ˉ ) + d , − κ x ˉ s 1 ( 0 , s 2 , t ) = b q ( s 2 ) ( v ˉ − k P ξ ˉ ) , v ˉ ˙ = − k I ξ ˉ , ξ ˉ ˙ = − λ ξ ˉ + λ g ( x ˉ ) . (13) \begin{aligned}
\dot{\bar{x}}&=\kappa\Delta\bar{x}-\mu_0\bar{x}+F(s,\bar{x})+d,\\
-\kappa\bar{x}_{s_1}(0,s_2,t)&=bq(s_2)(\bar{v}-k_P\bar{\xi}),\\
\dot{\bar{v}}&=-k_I\bar{\xi},\\
\dot{\bar{\xi}}&=-\lambda\bar{\xi}+\lambda g(\bar{x}). \tag*{(13)}
\end{aligned} x ˉ ˙ − κ x ˉ s 1 ( 0 , s 2 , t ) v ˉ ˙ ξ ˉ ˙ = κ Δ x ˉ − μ 0 x ˉ + F ( s , x ˉ ) + d , = b q ( s 2 ) ( v ˉ − k P ξ ˉ ) , = − k I ξ ˉ , = − λ ξ ˉ + λ g ( x ˉ ) . (13)
The other edges have the conditions in (1). The averaged target is the loop that would be obtained if the sensitivity g g g were measured directly and passed through the same low-pass filter. The dither and the demodulation have disappeared, and the task of the following sections is to show how closely the implemented loop follows it. Write Z ˉ = ( x ˉ , v ˉ , ξ ˉ ) \bar{Z}=(\bar{x},\bar{v},\bar{\xi}) Z ˉ = ( x ˉ , v ˉ , ξ ˉ ) , Z ⋆ = ( w ⋆ , u ⋆ , 0 ) Z^\star=(w^\star,u^\star,0) Z ⋆ = ( w ⋆ , u ⋆ , 0 ) , and X ω = ( w , v , g ^ ) X_\omega=(w,v,\widehat{g}) X ω = ( w , v , g ) . The common state space E = X × R 2 \mathcal{E}=\mathcal{X}\times\mathbb{R}^2 E = X × R 2 has norm
∥ ( e , v , ζ ) ∥ E = ∥ e ∥ ∞ + ∣ v ∣ + ∣ ζ ∣ . (14) \lVert(e,v,\zeta)\rVert_{\mathcal{E}}=\lVert e\rVert_\infty+|v|+|\zeta|. \tag*{(14)} ∥( e , v , ζ ) ∥ E = ∥ e ∥ ∞ + ∣ v ∣ + ∣ ζ ∣. (14)
and the regulation error is
E ω ( t ) = ∥ X ω ( t ) − Z ⋆ ∥ E . (15) E_\omega(t)=\lVert X_\omega(t)-Z^\star\rVert_{\mathcal{E}}. \tag*{(15)} E ω ( t ) = ∥ X ω ( t ) − Z ⋆ ∥ E . (15)
We seek local exponential transients with ultimate state and cost bounds of orders ω − 1 / 2 \omega^{-1/2} ω − 1/2 and ω − 1 \omega^{-1} ω − 1 , respectively, at fixed a a a .
A constant cost offset J c = J + c J_c=J+c J c = J + c leaves the implemented estimate and physical trajectories unchanged under p c = p − m c p_c=p-mc p c = p − m c , including p c ( t 0 ) = p ( t 0 ) − m ( t 0 ) c p_c(t_0)=p(t_0)-m(t_0)c p c ( t 0 ) = p ( t 0 ) − m ( t 0 ) c . An offset-invariant initialization is
p ( t 0 ) = − m ( t 0 ) J ( t 0 ) , g ^ ( t 0 ) = 0. (16) p(t_0)=-m(t_0)J(t_0),\qquad\widehat{g}(t_0)=0. \tag*{(16)} p ( t 0 ) = − m ( t 0 ) J ( t 0 ) , g ( t 0 ) = 0. (16)
Accordingly, (15) measures g ^ \widehat{g} g , whereas p = g ^ − m J p=\widehat{g}-mJ p = g − m J can depend on the absolute cost level. For the optical filter, this means that the controller does not need to know the peak transmission of the element.
Stability of the Averaged System
For the averaged state Z = ( x , v , ξ ) Z=(x,v,\xi) Z = ( x , v , ξ ) , set Z e = ( e , ν , ξ ) = ( x − w ⋆ , v − u ⋆ , ξ ) Z_e=(e,\nu,\xi)=(x-w^\star,v-u^\star,\xi) Z e = ( e , ν , ξ ) = ( x − w ⋆ , v − u ⋆ , ξ ) . We retain a longitudinal linear contribution to the reaction, prove stability of the resulting field–controller system, and then bound the remaining nonlinear terms in the norm (14). The linear estimate includes the sensitivity filter and supplies both a decay rate and a gain for additional boundary commands. These two estimates distinguish boundary disturbances from integrable internal forcing in the nonlinear robustness result.
The argument has three parts. First, a linear model is extracted that keeps the diffusion, the part of the reaction that helps (through a reference slope V 0 V_0 V 0 ), and the controller. Its stability is proved with an energy function that pairs the mean temperature across the plate with the two controller states. Second, the response of this linear model to an additional boundary command is bounded. Third, what the linear model leaves out is shown to be a small Lipschitz perturbation, which can be absorbed using a Gronwall-type argument. The unknown sign of b b b enters twice, once in the effect of the heater on the plate and once in the reconstructed sensitivity. As a result, the loop gain seen by the energy estimate is b 2 h / κ > 0 b^2h/\sqrt{\kappa}>0 b 2 h / κ > 0 .
and
Linear system and constants
Choose V 0 ∈ C ( [ 0 , 1 ] ) V_0\in C([0,1]) V 0 ∈ C ([ 0 , 1 ]) to include stabilizing reaction slopes in the linear calculation. The remainder is F ( s , w ⋆ + e ) − F ( s , w ⋆ ) − V 0 ( s 1 ) e F(s,w^\star+e)-F(s,w^\star)-V_0(s_1)e F ( s , w ⋆ + e ) − F ( s , w ⋆ ) − V 0 ( s 1 ) e . The controller does not use V 0 V_0 V 0 . We define
μ r e f ( s 1 ) = μ 0 − V 0 ( s 1 ) , μ = min [ 0 , 1 ] μ r e f , μ a v = ∫ 0 1 μ r e f ( y ) d y . (17) \mu_{\mathrm{ref}}(s_1)=\mu_0-V_0(s_1),\qquad
\mu=\min_{[0,1]}\mu_{\mathrm{ref}},\qquad
\mu_{\mathrm{av}}=\int_0^1\mu_{\mathrm{ref}}(y)\,\mathrm{d}y. \tag*{(17)} μ ref ( s 1 ) = μ 0 − V 0 ( s 1 ) , μ = [ 0 , 1 ] min μ ref , μ av = ∫ 0 1 μ ref ( y ) d y . (17)
Assumption 3. The damping in (17) satisfies μ > 0 \mu>0 μ > 0 .
Linearizing the sensitivity at the optimum gives
e t = κ Δ e − μ r e f ( s 1 ) e e_t=\kappa\Delta e-\mu_{\mathrm{ref}}(s_1)e e t = κ Δ e − μ ref ( s 1 ) e
− κ e s 1 ( 0 , s 2 , t ) = b q ( s 2 ) ( ν − k P ξ + d u ) -\kappa e_{s_1}(0,s_2,t)=bq(s_2)(\nu-k_P\xi+d_u) − κ e s 1 ( 0 , s 2 , t ) = b q ( s 2 ) ( ν − k P ξ + d u )
ν ˙ = − k I ξ \dot{\nu}=-k_I\xi ν ˙ = − k I ξ
ξ ˙ = − λ ξ + λ χ h C b e (18) \dot{\xi}=-\lambda\xi+\lambda\chi_h C_b e \tag*{(18)} ξ ˙ = − λ ξ + λ χ h C b e (18)
with the boundary conditions described in (1). Let T 0 T_0 T 0 be its unforced semigroup on E \mathcal{E} E , and let S r e f S_{\mathrm{ref}} S ref and K B , r e f K_{B,\mathrm{ref}} K B , ref be the homogeneous Neumann semigroup of A + V 0 A+V_0 A + V 0 and its boundary-input kernel.
We set c ⋆ = h / κ c_\star=h/\sqrt{\kappa} c ⋆ = h / κ , c = b 2 c ⋆ c=b^2c_\star c = b 2 c ⋆ , D = k P − k I / λ > 0 D=k_P-k_I/\lambda>0 D = k P − k I / λ > 0 , Q = 1 + ∣ ϑ ∣ Q=1+|\vartheta| Q = 1 + ∣ ϑ ∣ , and J κ ( β ) = β − 1 + ( κ β ) − 1 / 2 \mathcal{J}_\kappa(\beta)=\beta^{-1}+(\kappa\beta)^{-1/2} J κ ( β ) = β − 1 + ( κ β ) − 1/2 for β > 0 \beta>0 β > 0 . The following constant is chosen
0 < δ L < min { 1 2 c k I , κ k I , ( 3 k I μ + 2 μ a v c ) − 1 , ( 2 D ) − 1 } . (19) 0<\delta_L<\min\left\{\frac{1}{2}\sqrt{\frac{c}{k_I}},\frac{\kappa}{k_I},
\left(\frac{3k_I}{\mu}+\frac{2\mu_{\mathrm{av}}}{c}\right)^{-1},(2D)^{-1}\right\}. \tag*{(19)} 0 < δ L < min { 2 1 k I c , k I κ , ( μ 3 k I + c 2 μ av ) − 1 , ( 2 D ) − 1 } . (19)
We then define the constants
σ = 1 3 min { μ , δ L k I , λ } , L ⊥ = c k I + c D λ + δ L k I . (20) \sigma=\frac{1}{3}\min\{\mu,\delta_Lk_I,\lambda\},
\qquad
L_\perp=\sqrt{ck_I}+\sqrt{cD\lambda}+\delta_Lk_I. \tag*{(20)} σ = 3 1 min { μ , δ L k I , λ } , L ⊥ = c k I + cD λ + δ L k I . (20)
L U = 2 c ⋆ λ k P , L ν = 2 c ⋆ k I k P D , L ξ = 2 c ⋆ λ D . (21) L_U=2\sqrt{c_\star\lambda k_P},\qquad
L_\nu=2\sqrt{\frac{c_\star k_Ik_P}{D}},\qquad
L_\xi=2\sqrt{\frac{c_\star\lambda}{D}}. \tag*{(21)} L U = 2 c ⋆ λ k P , L ν = 2 D c ⋆ k I k P , L ξ = 2 D c ⋆ λ . (21)
P c = c ⋆ − 1 ( 1 / k I − 1 / λ − 1 / λ k P / λ ) , K W = max { 1 , λ max ( P c ) } , P_c=c_\star^{-1}
\begin{pmatrix}
1/k_I & -1/\lambda\\
-1/\lambda& k_P/\lambda
\end{pmatrix},
\qquad
K_W=\max\{1,\lambda_{\max}(P_c)\}, P c = c ⋆ − 1 ( 1/ k I − 1/ λ − 1/ λ k P / λ ) , K W = max { 1 , λ m a x ( P c )} ,
C E = 3 K W 2 + 4 L ⊥ μ − σ . (22) C_E=\frac{\sqrt{3K_W}}{2}+\frac{4L_\perp}{\mu-\sigma}. \tag*{(22)} C E = 2 3 K W + μ − σ 4 L ⊥ . (22)
Here P c > 0 P_c>0 P c > 0 follows from det P c = D / ( c ⋆ 2 k I λ ) > 0 \det P_c=D/(c_\star^2k_I\lambda)>0 det P c = D / ( c ⋆ 2 k I λ ) > 0 , and σ ≤ μ / 3 \sigma\leq\mu/3 σ ≤ μ /3 . Finally, we let
M = 1 + C E [ ∣ b ∣ Q L U J κ ( μ − σ ) + L ν + L ξ ] , (23) M=1+C_E\bigl[|b|QL_U\mathcal{J}_\kappa(\mu-\sigma)+L_\nu+L_\xi\bigr], \tag*{(23)} M = 1 + C E [ ∣ b ∣ Q L U J κ ( μ − σ ) + L ν + L ξ ] , (23)
H B = b 2 ( κ − 1 + μ − 1 + δ L / c ) , G V = H B / σ , H_B=b^2(\kappa^{-1}+\mu^{-1}+\delta_L/c),\qquad
G_V=\sqrt{H_B/\sigma}, H B = b 2 ( κ − 1 + μ − 1 + δ L / c ) , G V = H B / σ ,
and
G B = ∣ b ∣ Q J κ ( μ ) ( 1 + L U G V ) + ( L ν + L ξ ) G V . (24) G_B=|b|Q\mathcal{J}_\kappa(\mu)(1+L_UG_V)+(L_\nu+L_\xi)G_V. \tag*{(24)} G B = ∣ b ∣ Q J κ ( μ ) ( 1 + L U G V ) + ( L ν + L ξ ) G V . (24)
These constants are independent of the nonlinear radius chosen below. Their dependence on the plant and the chosen reaction coefficient is explicit. The restriction on V 0 V_0 V 0 separates the linear transverse modes. It does not require the full nonlinear reaction F F F to be independent of s 2 s_2 s 2 .
Proposition 1. Under Assumptions 2, 3, and (8), system (18) defines a strongly continuous semigroup satisfying
∥ T 0 ( t ) ∥ L ( E ) ≤ M e − σ t , t ≥ 0 , (25) \lVert T_0(t)\rVert_{\mathcal{L}(\mathcal{E})}\leq M e^{-\sigma t},\qquad t\geq0, \tag*{(25)} ∥ T 0 ( t ) ∥ L ( E ) ≤ M e − σ t , t ≥ 0 , (25)
with (20) and (23). Its zero-state response to d u ∈ L l o c ∞ d_u\in L^{{\infty}}_{\mathrm{loc}} d u ∈ L loc ∞ satisfies
∥ Z d u ( t ) ∥ E ≤ G B ∥ d u ∥ ∞ , [ t 0 , t ] , t ≥ t 0 , (26) \lVert Z_{d_u}(t)\rVert_{\mathcal{E}}\leq G_B\lVert d_u\rVert_{\infty,[t_0,t]},\qquad t\geq t_0, \tag*{(26)} ∥ Z d u ( t ) ∥ E ≤ G B ∥ d u ∥ ∞ , [ t 0 , t ] , t ≥ t 0 , (26)
with (24) and the essential-supremum input norm. Continuous initial fields require no boundary compatibility.
Proof. The proof proceeds in three steps. 1. Kernel and transverse modes. Let K N , K r e f K_N,K_{\mathrm{ref}} K N , K ref denote the interval Neumann kernels of K ∂ y y \mathcal{K}_{\partial yy} K ∂ y y and K ∂ y y − μ r e f \mathcal{K}_{\partial yy}-\mu_{\mathrm{ref}} K ∂ y y − μ ref . For nonnegative data, the Neumann heat kernel with a multiplied by e − μ t e^{-\mu t} e − μ t is a supersolution. Comparison gives 0 ≤ K r e f ≤ e − μ t K N 0\leq K_{\mathrm{ref}}\leq e^{-\mu t}K_N 0 ≤ K ref ≤ e − μ t K N and ∥ S r e f ( t ) ∥ ≤ e − μ t \lVert S_{\mathrm{ref}}(t)\rVert\leq e^{-\mu t} ∥ S ref ( t )∥ ≤ e − μ t [24 Sec. 2.3] . Its residual is ( μ r e f − μ ) (\mu_{\mathrm{ref}}-\mu) ( μ ref − μ ) times the nonnegative supersolution. Positivity and preservation of constants by the heat semigroup then give the asserted norm bound for arbitrary signed initial fields. Green’s identity and separation give
K N ( t ; y , 0 ) = 1 + 2 ∑ n ≥ 1 e − κ n 2 π 2 t cos ( n π y ) K_N(t;y,0)=1+2\sum_{n\geq1}e^{-\kappa n^2\pi^2t}\cos(n\pi y) K N ( t ; y , 0 ) = 1 + 2 n ≥ 1 ∑ e − κ n 2 π 2 t cos ( nπ y )
K B , r e f ( t ) ( s ) = b K r e f ( t ; s 1 , 0 ) [ 1 + ϑ e − κ π 2 t cos ( π s 2 ) ] (27) K_{B,\mathrm{ref}}(t)(s)=bK_{\mathrm{ref}}(t;s_1,0)[1+\vartheta e^{-\kappa\pi^2t}\cos(\pi s_2)] \tag*{(27)} K B , ref ( t ) ( s ) = b K ref ( t ; s 1 , 0 ) [ 1 + ϑ e − κ π 2 t cos ( π s 2 )] (27)
∥ K B , r e f ( t ) ∥ ∞ ≤ ∣ b ∣ Q e − μ t ( 1 + 1 / κ π t ) . \lVert K_{B,\mathrm{ref}}(t)\rVert_\infty\leq|b|Qe^{-\mu t}(1+1/\sqrt{\kappa\pi t}). ∥ K B , ref ( t ) ∥ ∞ ≤ ∣ b ∣ Q e − μ t ( 1 + 1/ κπ t ) .
Indeed, the Gaussian sum is at most its integral over [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) , namely ( 2 π κ t ) − 1 (2\sqrt{\pi\kappa t})^{-1} ( 2 π κ t ) − 1 . Thus, for ζ < μ \zeta<\mu ζ < μ , we get
∫ 0 ∞ e ζ t ∥ K B , r e f ( t ) ∥ ∞ d t ≤ ∣ b ∣ Q J κ ( μ − ζ ) . (28) \int_0^\infty e^{\zeta t}\lVert K_{B,\mathrm{ref}}(t)\rVert_\infty\,\mathrm{d}t\leq|b|Q\mathcal{J}_\kappa(\mu-\zeta). \tag*{(28)} ∫ 0 ∞ e ζ t ∥ K B , ref ( t ) ∥ ∞ d t ≤ ∣ b ∣ Q J κ ( μ − ζ ) . (28)
For example, the field component of the mild equation is
e ( t ) = S r e f ( t − t 0 ) e ( t 0 ) + ∫ t 0 t K B , r e f ( t − s ) [ ν ( s ) − k P ξ ( s ) + d u ( s ) ] d s . e(t)=S_{\mathrm{ref}}(t-t_0)e(t_0)
+\int_{t_0}^{t}K_{B,\mathrm{ref}}(t-s)[\nu(s)-k_P\xi(s)+d_u(s)]\,\mathrm{d}s. e ( t ) = S ref ( t − t 0 ) e ( t 0 ) + ∫ t 0 t K B , ref ( t − s ) [ ν ( s ) − k P ξ ( s ) + d u ( s )] d s .
The controller components are ordinary integral equations. Their Lipschitz constants tend to zero with the interval length, as does ∥ K B , r e f ∥ L 1 ( 0 , T ; X ) \lVert K_{B,\mathrm{ref}}\rVert_{L^{1}(0,T;\mathcal{X})} ∥ K B , ref ∥ L 1 ( 0 , T ; X ) . The coupled equations therefore define a Volterra contraction for small time. The linear coefficients permit a uniform restart interval, giving global mild solutions. For zero input, uniqueness, autonomy, and continuity at the initial time give T 0 T_{0} T 0 .
The boundary profile contains a mean and one cosine mode. The actuation therefore forces only two transverse components, whereas an arbitrary initial field can contain every mode. We separate these components to identify the part seen by the sensor and retain the contribution of the remaining field. We write τ = t − t 0 \tau=t-t_{0} τ = t − t 0 and define
e 0 ( s 1 ) = ∫ 0 1 e ( s 1 , s 2 ) d s 2 , e_{0}(s_{1})=\int_{0}^{1}e(s_{1},s_{2})\,\mathrm{d}s_{2}, e 0 ( s 1 ) = ∫ 0 1 e ( s 1 , s 2 ) d s 2 ,
e 1 ( s 1 ) = 2 ∫ 0 1 e ( s 1 , s 2 ) cos ( π s 2 ) d s 2 , e_{1}(s_{1})=2\int_{0}^{1}e(s_{1},s_{2})\cos(\pi s_{2})\,\mathrm{d}s_{2}, e 1 ( s 1 ) = 2 ∫ 0 1 e ( s 1 , s 2 ) cos ( π s 2 ) d s 2 ,
e ⊥ = e − e 0 − e 1 cos ( π s 2 ) . e_{\perp}=e-e_{0}-e_{1}\cos(\pi s_{2}). e ⊥ = e − e 0 − e 1 cos ( π s 2 ) .
These projections commute with S r e f S_{\mathrm{ref}} S ref because V 0 V_{0} V 0 is longitudinal. Only e 0 , e 1 e_{0},e_{1} e 0 , e 1 are forced. The first cosine vanishes at the sensor, so e ⊥ ( t ) = S r e f ( τ ) e ⊥ ( t 0 ) e_{\perp}(t)=S_{\mathrm{ref}}(\tau)e_{\perp}(t_{0}) e ⊥ ( t ) = S ref ( τ ) e ⊥ ( t 0 ) , z ⊥ = C b e ⊥ z_{\perp}=C_{b}e_{\perp} z ⊥ = C b e ⊥ , C b e = e 0 ( 0 ) + z ⊥ C_{b}e=e_{0}(0)+z_{\perp} C b e = e 0 ( 0 ) + z ⊥ and therefore
∣ z ⊥ ( t ) ∣ ≤ 4 e − μ τ ∥ e ( t 0 ) ∥ ∞ . (29) |z_{\perp}(t)|\leq4e^{-\mu\tau}\lVert e(t_{0})\rVert_{\infty}. \tag*{(29)} ∣ z ⊥ ( t ) ∣ ≤ 4 e − μτ ∥ e ( t 0 ) ∥ ∞ . (29)
This expression bounds the remainder projection on X \mathcal{X} X , since the identity, mean, and first-cosine projections have norms at most 1 , 1 , 2 1,1,2 1 , 1 , 2 . Although e 1 e_{1} e 1 is forced, it contributes nothing to the midpoint observation. Thus z ⊥ z_{\perp} z ⊥ acts as an exponentially decaying input to the mean-field and controller subsystem.
Energy and full-field decay. Initially take smooth compatible data and d u = 0 d_{u}=0 d u = 0 . Put η c = ξ / ( b c ⋆ ) \eta_{c}=\xi/(bc_{\star}) η c = ξ / ( b c ⋆ ) and ρ c = − ν / ( b c ⋆ ) + ( k I / λ ) η c \rho_{c}=-\nu/(bc_{\star})+(k_{I}/\lambda)\eta_{c} ρ c = − ν / ( b c ⋆ ) + ( k I / λ ) η c . The mean flux is − c ( ρ c + D η c ) -c(\rho_{c}+D\eta_{c}) − c ( ρ c + D η c ) , and
ρ ˙ c = k I [ e 0 ( 0 ) + z ⊥ ] , \dot{\rho}_{c}=k_{I}[e_{0}(0)+z_{\perp}], ρ ˙ c = k I [ e 0 ( 0 ) + z ⊥ ] ,
η ˙ c = λ [ e 0 ( 0 ) + z ⊥ − η c ] . \dot{\eta}_{c}=\lambda[e_{0}(0)+z_{\perp}-\eta_{c}]. η ˙ c = λ [ e 0 ( 0 ) + z ⊥ − η c ] .
These coordinates use χ h = b c ⋆ \chi_{h}=bc_{\star} χ h = b c ⋆ . The same unknown sign enters the plant flux and the reconstructed sensitivity, so their product in the feedback dynamics is c = b 2 c ⋆ > 0 c=b^{2}c_{\star}>0 c = b 2 c ⋆ > 0 . The coordinate change is used for the stability proof. We define j 0 = ∫ 0 1 e 0 j_{0}=\int_{0}^{1}e_{0} j 0 = ∫ 0 1 e 0 , I e = ∫ 0 1 μ r e f e 0 2 I_{e}=\int_{0}^{1}\mu_{\mathrm{ref}}e_{0}^{2} I e = ∫ 0 1 μ ref e 0 2 , and j μ = ∫ 0 1 μ r e f e 0 j_{\mu}=\int_{0}^{1}\mu_{\mathrm{ref}}e_{0} j μ = ∫ 0 1 μ ref e 0 , where ∥ ⋅ ∥ 2 \lVert\cdot\rVert_{2} ∥ ⋅ ∥ 2 the L 2 ( 0 , 1 ) L^{2}(0,1) L 2 ( 0 , 1 ) norm. We pose
W = 1 2 ∥ e 0 ∥ 2 2 + c ρ c 2 2 k I + c D η c 2 2 λ , V = W + δ L ρ c j 0 . (30) W=\frac{1}{2}\lVert e_{0}\rVert_{2}^{2}+\frac{c\rho_{c}^{2}}{2k_{I}}+\frac{cD\eta_{c}^{2}}{2\lambda},
\qquad
V=W+\delta_{L}\rho_{c}j_{0}. \tag*{(30)} W = 2 1 ∥ e 0 ∥ 2 2 + 2 k I c ρ c 2 + 2 λ cD η c 2 , V = W + δ L ρ c j 0 . (30)
Since ∣ ρ c j 0 ∣ ≤ k I / c W |\rho_{c}j_{0}|\leq\sqrt{k_{I}/c}\,W ∣ ρ c j 0 ∣ ≤ k I / c W , (19) gives W / 2 ≤ V ≤ 3 W / 2 W/2\leq V\leq3W/2 W /2 ≤ V ≤ 3 W /2 . The energy W W W pairs the mean field with the controller coordinates so that the terms involving e 0 ( 0 ) e_{0}(0) e 0 ( 0 ) cancel under integration by parts. Its derivative does not directly penalize ρ c \rho_{c} ρ c . The additional term δ L ρ c j 0 \delta_{L}\rho_{c}j_{0} δ L ρ c j 0 uses the mean balance to obtain that dissipation while keeping V V V comparable to W W W . Integration by parts yields
W ˙ = − κ ∥ e 0 ′ ∥ 2 2 − I e − c D η c 2 + ( c ρ c + c D η c ) z ⊥ , \dot{W}=-\kappa\lVert e_{0}'\rVert_{2}^{2}-I_{e}-cD\eta_{c}^{2}+(c\rho_{c}+cD\eta_{c})z_{\perp}, W ˙ = − κ ∥ e 0 ′ ∥ 2 2 − I e − cD η c 2 + ( c ρ c + cD η c ) z ⊥ ,
j ˙ 0 = − c ρ c − c D η c − j μ . \dot{j}_{0}=-c\rho_{c}-cD\eta_{c}-j_{\mu}. j ˙ 0 = − c ρ c − cD η c − j μ .
Here primes denote spatial derivatives. The identity
e 0 ( 0 ) = j 0 − ∫ 0 1 ( 1 − y ) e 0 ′ ( y ) d y e_{0}(0)=j_{0}-\int_{0}^{1}(1-y)e_{0}'(y)\,\mathrm{d}y e 0 ( 0 ) = j 0 − ∫ 0 1 ( 1 − y ) e 0 ′ ( y ) d y
gives
∣ e 0 ( 0 ) j 0 ∣ ≤ 3 ∥ e 0 ∥ 2 2 + ∥ e 0 ′ ∥ 2 2 2 . |e_{0}(0)j_{0}|\leq\frac{3\lVert e_{0}\rVert_{2}^{2}+\lVert e_{0}'\rVert_{2}^{2}}{2}. ∣ e 0 ( 0 ) j 0 ∣ ≤ 2 3 ∥ e 0 ∥ 2 2 + ∥ e 0 ′ ∥ 2 2 .
Applying the Cauchy–Schwarz inequality to μ r e f \sqrt{\mu_{\mathrm{ref}}} μ ref and μ r e f e 0 \sqrt{\mu_{\mathrm{ref}}}e_{0} μ ref e 0 gives j μ 2 ≤ μ a v I e j_{\mu}^{2}\leq\mu_{\mathrm{av}}I_{e} j μ 2 ≤ μ av I e [25 Sec. 1.1] . Differentiating (30) and using I e ≥ μ ∥ e 0 ∥ 2 2 I_{e}\geq\mu\lVert e_{0}\rVert_{2}^{2} I e ≥ μ ∥ e 0 ∥ 2 2 , the mixed terms are shown to satisfy
δ L k I ∣ e 0 ( 0 ) j 0 ∣ ≤ δ L k I 2 ∥ e 0 ′ ∥ 2 2 + 3 δ L k I 2 μ I e , \delta_{L}k_{I}|e_{0}(0)j_{0}|
\leq\frac{\delta_{L}k_{I}}{2}\lVert e_{0}'\rVert_{2}^{2}
+\frac{3\delta_{L}k_{I}}{2\mu}I_{e}, δ L k I ∣ e 0 ( 0 ) j 0 ∣ ≤ 2 δ L k I ∥ e 0 ′ ∥ 2 2 + 2 μ 3 δ L k I I e ,
δ L c D ∣ ρ c η c ∣ ≤ δ L c 4 ρ c 2 + δ L c D 2 η c 2 , \delta_{L}cD|\rho_{c}\eta_{c}|
\leq\frac{\delta_{L}c}{4}\rho_{c}^{2}+\delta_{L}cD^{2}\eta_{c}^{2}, δ L cD ∣ ρ c η c ∣ ≤ 4 δ L c ρ c 2 + δ L c D 2 η c 2 ,
δ L ∣ ρ c j μ ∣ ≤ δ L c 4 ρ c 2 + δ L μ a v c I e . \delta_{L}|\rho_{c}j_{\mu}|
\leq\frac{\delta_{L}c}{4}\rho_{c}^{2}
+\frac{\delta_{L}\mu_{\mathrm{av}}}{c}I_{e}. δ L ∣ ρ c j μ ∣ ≤ 4 δ L c ρ c 2 + c δ L μ av I e .
Each inequality accounts for a specified part of the dissipation. The condition involving μ a v \mu_{\mathrm{av}} μ av controls the weighted reaction integral, δ L < κ / k I \delta_{L}<\kappa/k_{I} δ L < κ / k I controls the field derivative term, and δ L < ( 2 D ) − 1 \delta_{L}<(2D)^{-1} δ L < ( 2 D ) − 1 controls the mixed controller term. The derivative of the cross term contributes − δ L c ρ c 2 -\delta_{L}c\rho_{c}^{2} − δ L c ρ c 2 . These bounds account for half of this term and at least half of each original dissipative term by (19). Hence
V ˙ ≤ − κ 2 ∥ e 0 ′ ∥ 2 2 − 1 2 I e − δ L c 2 ρ c 2 − c D 2 η c 2 + 2 L ⊥ V ∣ z ⊥ ∣ . (31) \dot{V}\leq-\frac{\kappa}{2}\lVert e_{0}'\rVert_{2}^{2}
-\frac{1}{2}I_{e}
-\frac{\delta_{L}c}{2}\rho_{c}^{2}
-\frac{cD}{2}\eta_{c}^{2}
+2L_{\perp}\sqrt{V}|z_{\perp}|. \tag*{(31)} V ˙ ≤ − 2 κ ∥ e 0 ′ ∥ 2 2 − 2 1 I e − 2 δ L c ρ c 2 − 2 cD η c 2 + 2 L ⊥ V ∣ z ⊥ ∣. (31)
Indeed, V ≥ W / 2 V\geq W/2 V ≥ W /2 bounds the forcing coefficient ∣ c ρ c + c D η c + δ L k I j 0 ∣ |c\rho_{c}+cD\eta_{c}+\delta_{L}k_{I}j_{0}| ∣ c ρ c + cD η c + δ L k I j 0 ∣ by 2 L ⊥ V 2L_{\perp}\sqrt{V} 2 L ⊥ V . The negative terms in (31) dominate 3 σ W ≥ 2 σ V 3\sigma W\geq2\sigma V 3 σ W ≥ 2 σ V . The comparison for V \sqrt{V} V , justified at zero by positive regularization, gives
V ( t ) ≤ e − σ t V ( t 0 ) + L ⊥ ∫ t 0 t e − σ ( t − s ) ∣ z ⊥ ( s ) ∣ d s . \sqrt{V(t)}\leq e^{-\sigma t}\sqrt{V(t_{0})}
+L_{\perp}\int_{t_{0}}^{t}e^{-\sigma(t-s)}|z_{\perp}(s)|\,\mathrm{d}s. V ( t ) ≤ e − σ t V ( t 0 ) + L ⊥ ∫ t 0 t e − σ ( t − s ) ∣ z ⊥ ( s ) ∣ d s .
The energy estimate controls the mean field in L 2 L^{2} L 2 and the controller states. To recover the product norm, we first bound the feedback command ν − k P ξ \nu-k_{P}\xi ν − k P ξ . Its action on the full field is then estimated through the boundary kernel, including both the forced first cosine and the remaining modes. For arbitrary real a 1 , a 2 a_{1},a_{2} a 1 , a 2 , the controller terms in V ≥ W / 2 V\geq W/2 V ≥ W /2 give
∣ a 1 ρ c + a 2 η c ∣ 2 ≤ 4 ( k I a 1 2 c + λ a 2 2 c D ) V . |a_{1}\rho_{c}+a_{2}\eta_{c}|^{2}
\leq4\left(\frac{k_{I}a_{1}^{2}}{c}+\frac{\lambda a_{2}^{2}}{cD}\right)V. ∣ a 1 ρ c + a 2 η c ∣ 2 ≤ 4 ( c k I a 1 2 + cD λ a 2 2 ) V .
We then use ν − k P ξ = − b c ⋆ ( ρ c + D η c ) \nu-k_{P}\xi=-bc_{\star}(\rho_{c}+D\eta_{c}) ν − k P ξ = − b c ⋆ ( ρ c + D η c ) , ν = b c ⋆ [ − ρ c + ( k I / λ ) η c ] \nu=bc_{\star}[-\rho_{c}+(k_{I}/\lambda)\eta_{c}] ν = b c ⋆ [ − ρ c + ( k I / λ ) η c ] , and ξ = b c ⋆ η c \xi=bc_{\star}\eta_{c} ξ = b c ⋆ η c to generate the bounds
∣ ν − k P ξ ∣ ≤ L U V , ∣ ν ∣ ≤ L ν V , ∣ ξ ∣ ≤ L ξ V , |\nu-k_{P}\xi|\leq L_{U}\sqrt{V},\qquad
|\nu|\leq L_{\nu}\sqrt{V},\qquad
|\xi|\leq L_{\xi}\sqrt{V}, ∣ ν − k P ξ ∣ ≤ L U V , ∣ ν ∣ ≤ L ν V , ∣ ξ ∣ ≤ L ξ V ,
with (21). Moreover, W = ∥ e 0 ∥ 2 2 / 2 + ( ν , ξ ) P c ( ν , ξ ) T / 2 W=\lVert e_{0}\rVert_{2}^{2}/2+(\nu,\xi)P_{c}(\nu,\xi)^{\mathsf{T}}/2 W = ∥ e 0 ∥ 2 2 /2 + ( ν , ξ ) P c ( ν , ξ ) T /2 . Thus, for N 0 = ∥ z e ( t 0 ) ∥ ε N_{0}=\lVert z_{e}(t_{0})\rVert_{\varepsilon} N 0 = ∥ z e ( t 0 ) ∥ ε , V ( t 0 ) ≤ 3 κ W N 0 / 2 \sqrt{V(t_{0})}\leq\sqrt{3\kappa WN_{0}/2} V ( t 0 ) ≤ 3 κW N 0 /2 . Using (29) in scalar comparison gives V ( t ) ≤ C E e − σ τ N 0 \sqrt{V(t)}\leq C_{E}e^{-\sigma\tau}N_{0} V ( t ) ≤ C E e − σ τ N 0 , with (22), since ∫ 0 τ e − ( μ − σ ) s d s ≤ ( μ − σ ) − 1 \int_{0}^{\tau}e^{-(\mu-\sigma)s}\,\mathrm{d}s\leq(\mu-\sigma)^{-1} ∫ 0 τ e − ( μ − σ ) s d s ≤ ( μ − σ ) − 1 . The two terms in C E C_{E} C E are as follows. The first term converts the initial product norm into the energy norm. The second term accounts for the decaying observation of the unforced transverse modes. Its denominator is positive since these modes are estimated with the faster heat-semigroup decay μ > σ \mu>\sigma μ > σ . The field mild equation and (28) imply
∥ e ( t ) ∥ ∞ ≤ e − σ τ N 0 + ∣ b ∣ Q L U C E J κ ( μ − σ ) e − σ τ N 0 . \lVert e(t)\rVert_{\infty}
\leq e^{-\sigma\tau}N_{0}
+|b|QL_{U}C_{E}\mathcal{J}_{\kappa}(\mu-\sigma)e^{-\sigma\tau}N_{0}. ∥ e ( t ) ∥ ∞ ≤ e − σ τ N 0 + ∣ b ∣ Q L U C E J κ ( μ − σ ) e − σ τ N 0 .
Adding the controller bounds proves (25).
Boundary gain and mild data.
For zero initial state, we have z ⊥ = 0 z_{\perp}=0 z ⊥ = 0 . The additional boundary flux makes V ˙ \dot V V ˙ gain b [ e 0 ( 0 ) + δ L ρ c ] d u b[e_0(0)+\delta L\rho_c]d_u b [ e 0 ( 0 ) + δ L ρ c ] d u . Using ∣ e 0 ( 0 ) ∣ ≤ ∥ e 0 ∥ 2 + ∥ e 0 ′ ∥ 2 |e_0(0)|\le\|e_0\|_2+\|e'_0\|_2 ∣ e 0 ( 0 ) ∣ ≤ ∥ e 0 ∥ 2 + ∥ e 0 ′ ∥ 2 , Young’s inequality gives
∣ b ∣ ∥ e 0 ′ ∥ 2 ∣ d u ∣ ≤ κ 4 ∥ e 0 ′ ∥ 2 2 + b 2 κ ∣ d u ∣ 2 , |b|\|e'_0\|_2|d_u|\le\frac{\kappa}{4}\|e'_0\|_2^2+\frac{b^2}{\kappa}|d_u|^2, ∣ b ∣∥ e 0 ′ ∥ 2 ∣ d u ∣ ≤ 4 κ ∥ e 0 ′ ∥ 2 2 + κ b 2 ∣ d u ∣ 2 ,
∣ b ∣ ∥ e 0 ∥ 2 ∣ d u ∣ ≤ 1 4 I e + b 2 μ ∣ d u ∣ 2 , |b|\|e_0\|_2|d_u|\le\frac{1}{4}I_e+\frac{b^2}{\mu}|d_u|^2, ∣ b ∣∥ e 0 ∥ 2 ∣ d u ∣ ≤ 4 1 I e + μ b 2 ∣ d u ∣ 2 ,
∣ b ∣ δ L ∣ ρ c ∣ ∣ d u ∣ ≤ δ L c 4 ρ c 2 + b 2 δ L c ∣ d u ∣ 2 . |b|\delta L|\rho_c||d_u|\le\frac{\delta L c}{4}\rho_c^2+\frac{b^2\delta L}{c}|d_u|^2. ∣ b ∣ δ L ∣ ρ c ∣∣ d u ∣ ≤ 4 δ L c ρ c 2 + c b 2 δ L ∣ d u ∣ 2 .
The remaining dissipation dominates 3 σ W / 2 ≥ σ V 3\sigma W/2\ge\sigma V 3 σ W /2 ≥ σ V , so V ˙ ≤ − σ V + H B ∣ d u ∣ 2 \dot V\le-\sigma V+H_B|d_u|^2 V ˙ ≤ − σ V + H B ∣ d u ∣ 2 and V ≤ G V ∥ d u ∥ ∞ , [ t 0 , t ] \sqrt V\le G_V\|d_u\|_{\infty,[t_0,t]} V ≤ G V ∥ d u ∥ ∞ , [ t 0 , t ] . Here integration from V ( t 0 ) = 0 V(t_0)=0 V ( t 0 ) = 0 bounds V ( t ) V(t) V ( t ) by H B ∥ d u ∥ ∞ , [ t 0 , t ] 2 / σ H_B\|d_u\|_{\infty,[t_0,t]}^2/\sigma H B ∥ d u ∥ ∞ , [ t 0 , t ] 2 / σ . Applying (28) to the total command ν − k P ξ + d u \nu-k_P\xi+d_u ν − k P ξ + d u yields
∥ e ( t ) ∥ ∞ ≤ ∣ b ∣ Q J κ ( μ ) ( 1 + L U G V ) ∥ d u ∥ ∞ , [ t 0 , t ] . \|e(t)\|_\infty\le|b|QJ_{\kappa}(\mu)(1+L_UG_V)\|d_u\|_{\infty,[t_0,t]}. ∥ e ( t ) ∥ ∞ ≤ ∣ b ∣ Q J κ ( μ ) ( 1 + L U G V ) ∥ d u ∥ ∞ , [ t 0 , t ] .
The controller coordinates contribute at most
( L ν + L ξ ) G V ∥ d u ∥ ∞ , [ t 0 , t ] . (L_\nu+L_\xi)G_V\|d_u\|_{\infty,[t_0,t]}. ( L ν + L ξ ) G V ∥ d u ∥ ∞ , [ t 0 , t ] .
Adding the field and controller bounds gives (26) with exactly (24) .
3. Passage to mild data. The integration by parts above uses smooth compatible data, but the resulting constants do not depend on higher derivatives of those data. We approximate the initial field for the unforced semigroup estimate and the input for the zero-state boundary estimate. First we approximate a continuous initial field uniformly using tensor cosine sums. We add − ( b / κ ) U 0 q ( s 2 ) s 1 η ( n s 1 ) -(b/\kappa)U_0q(s_2)s_1\eta(ns_1) − ( b / κ ) U 0 q ( s 2 ) s 1 η ( n s 1 ) to impose the initial flux, where U 0 = ν ( t 0 ) − k P ξ ( t 0 ) U_0=\nu(t_0)-k_P\xi(t_0) U 0 = ν ( t 0 ) − k P ξ ( t 0 ) and η ∈ C ∞ ( [ 0 , ∞ ) ) \eta\in C^\infty([0,\infty)) η ∈ C ∞ ([ 0 , ∞ )) equals one near zero and vanishes on [ 1 , ∞ ) [1,\infty) [ 1 , ∞ ) . This correction tends uniformly to zero and preserves the other Neumann conditions since q ′ ( 0 ) = q ′ ( 1 ) = 0 q'(0)=q'(1)=0 q ′ ( 0 ) = q ′ ( 1 ) = 0 . Its derivative at the actuated edge is exactly − ( b / κ ) U 0 q ( s 2 ) -(b/\kappa)U_0q(s_2) − ( b / κ ) U 0 q ( s 2 ) , while its supremum norm is O ( n − 1 ) O(n^{-1}) O ( n − 1 ) . On each finite interval, approximate bounded inputs in L p L^p L p , p > 2 p>2 p > 2 , by smooth inputs with no larger supremum norm, vanishing near t 0 t_0 t 0 for the zero-state estimate. Since p ′ = p / ( p − 1 ) < 2 p'=p/(p-1)<2 p ′ = p / ( p − 1 ) < 2 , (27) gives K B , r e f ∈ L p ′ ( 0 , T ; X ) K_{B,\mathrm{ref}}\in L^{p'}(0,T;\mathcal X) K B , ref ∈ L p ′ ( 0 , T ; X ) . The vanishing near t 0 t_0 t 0 makes the zero initial field compatible before the approximating flux begins. Upon preserving the input norm, the right-hand side of the boundary-gain estimate remains bounded by the norm of the original input. Using Hölder’s inequality, the convolution difference is bounded by ∥ K B , r e f ∥ L p ′ ∥ d u , n − d u ∥ L p → 0 \|K_{B,\mathrm{ref}}\|_{L^{p'}}\|d_{u,n}-d_u\|_{L^p}\to0 ∥ K B , ref ∥ L p ′ ∥ d u , n − d u ∥ L p → 0 , uniformly in time. The Volterra continuous dependence leads (25) and (26) to mild solutions. Compatibility is imposed only on the approximating data and does not restrict the initial fields in the proposition.
Nonlinear margin and robustness
Assumption 4. There exist R , η ⋆ > 0 R,\eta_\star>0 R , η ⋆ > 0 such that Assumption 1 holds for fields with ∥ x − w ⋆ ∥ ∞ ≤ R + η ⋆ \|x-w^\star\|_\infty\le R+\eta_\star ∥ x − w ⋆ ∥ ∞ ≤ R + η ⋆ , the cost is C 2 C^2 C 2 on an open interval containing [ z ⋆ − R − η ⋆ , z ⋆ + R + η ⋆ ] [z^\star-R-\eta_\star,z^\star+R+\eta_\star] [ z ⋆ − R − η ⋆ , z ⋆ + R + η ⋆ ] , and
ℓ ( R ) : = sup s ∈ Ω ∣ η ∣ ≤ R ∣ F w ( s , w ⋆ ( s ) + η ) − V 0 ( s 1 ) ∣ + λ ∣ χ ∣ sup ∣ η ∣ ≤ R ∣ φ ′ ′ ( z ⋆ + η ) − h ∣ < σ / M . (32) \ell(R):=\sup_{\substack{s\in\Omega\\|\eta|\le R}}|F_w(s,w^\star(s)+\eta)-V_0(s_1)|
+\lambda|\chi|\sup_{|\eta|\le R}|\varphi''(z^\star+\eta)-h|<\sigma/M. \tag*{(32)} ℓ ( R ) := s ∈ Ω ∣ η ∣ ≤ R sup ∣ F w ( s , w ⋆ ( s ) + η ) − V 0 ( s 1 ) ∣ + λ ∣ χ ∣ ∣ η ∣ ≤ R sup ∣ φ ′′ ( z ⋆ + η ) − h ∣ < σ / M . (32)
Define the nonlinear decay rate and cost-derivative bounds by
α = σ − M ℓ ( R ) > 0 , \alpha=\sigma-M\ell(R)>0, α = σ − M ℓ ( R ) > 0 ,
Φ j = max ∣ z − z ⋆ ∣ ≤ R + η ⋆ ∣ φ ( j ) ( z ) ∣ , j = 1 , 2. \Phi_j=\max_{|z-z^\star|\le R+\eta_\star}|\varphi^{(j)}(z)|,\qquad j=1,2. Φ j = ∣ z − z ⋆ ∣ ≤ R + η ⋆ max ∣ φ ( j ) ( z ) ∣ , j = 1 , 2.
If F w ( s , w ⋆ ( s ) ) = V 0 ( s 1 ) F_w(s,w^\star(s))=V_0(s_1) F w ( s , w ⋆ ( s )) = V 0 ( s 1 ) , continuity gives ℓ ( R ) → 0 \ell(R)\to0 ℓ ( R ) → 0 as R ↓ 0 R\downarrow0 R ↓ 0 . Otherwise, including the discrepancy, including dependence, the radius must satisfy (32). The larger radius accommodates the harmonic lifting later. The margin controls the reaction left after V 0 V_0 V 0 is retained, so large stabilizing equilibrium slopes need not be small. Plant dissipativity alone does not establish this condition for the coupled feedback system. All quantities in the right-hand side of (32) have already been fixed by the linear calculation. Choosing R R R therefore introduces no circular dependence into M M M or σ \sigma σ . The bounds Φ 1 , Φ 2 \Phi_1,\Phi_2 Φ 1 , Φ 2 are finite by continuity on the enlarged compact trace interval.
Lemma 1. Under Assumptions 1–4 and (8), let Z ~ \widetilde Z Z solve (13). Let Z Z Z solve it with additional terms d w , d v , d ξ d_w,d_v,d_\xi d w , d v , d ξ in the field, bias, and filter equations and boundary command ν − k P ξ + d u \nu-k_P\xi+d_u ν − k P ξ + d u . Assume strongly measurable disturbances with q d = ∥ d w ∥ ∞ + ∥ d v ∥ + ∣ d ξ ∣ ∈ L l o c 1 q_d=\|d_w\|_\infty+\|d_v\|+|d_\xi|\in L^1_{\mathrm{loc}} q d = ∥ d w ∥ ∞ + ∥ d v ∥ + ∣ d ξ ∣ ∈ L loc 1 and d u ∈ L l o c ∞ d_u\in L^\infty_{\mathrm{loc}} d u ∈ L loc ∞ . Writing D Z ( t ) = ∥ Z ( t ) − Z ~ ( t ) ∥ ε D_Z(t)=\|Z(t)-\widetilde Z(t)\|_\varepsilon D Z ( t ) = ∥ Z ( t ) − Z ( t ) ∥ ε , while both states remain in B R ( Z ⋆ ) B_R(Z^\star) B R ( Z ⋆ ) ,
D Z ( t ) ≤ M e − α ( t − t 0 ) D Z ( t 0 ) + G B σ α ∥ d u ∥ ∞ , [ t 0 , t ] + M ∫ t 0 t e − α ( t − s ) q d ( s ) d s . (33) D_Z(t)\le M e^{-\alpha(t-t_0)}D_Z(t_0)+G_B\frac{\sigma}{\alpha}\|d_u\|_{\infty,[t_0,t]}
+M\int_{t_0}^{t}e^{-\alpha(t-s)}q_d(s)\,\mathrm{d}s. \tag*{(33)} D Z ( t ) ≤ M e − α ( t − t 0 ) D Z ( t 0 ) + G B α σ ∥ d u ∥ ∞ , [ t 0 , t ] + M ∫ t 0 t e − α ( t − s ) q d ( s ) d s . (33)
Proof. The nonlinear remainder relative to (18) is
N ( Z e ) = ( F ( w ⋆ + e ) − F ( w ⋆ ) − V 0 e 0 λ χ [ φ ′ ( z ⋆ + C b e ) − h C b e ] ) . \mathcal N(Z_e)=
\begin{pmatrix}
F(w^\star+e)-F(w^\star)-V_0e\\
0\\
\lambda\chi[\varphi'(z^\star+C_be)-hC_be]
\end{pmatrix}. N ( Z e ) = F ( w ⋆ + e ) − F ( w ⋆ ) − V 0 e 0 λ χ [ φ ′ ( z ⋆ + C b e ) − h C b e ] .
The mean-value formula and (32) give N ( 0 ) = 0 \mathcal N(0)=0 N ( 0 ) = 0 and Lip ( N ) ≤ ℓ ( R ) \operatorname{Lip}(\mathcal N)\le\ell(R) Lip ( N ) ≤ ℓ ( R ) on the convex ball B R ( 0 ) B_R(0) B R ( 0 ) . Every mean-value segment stays within the stated ranges, and ∥ C b ∥ = 1 \|C_b\|=1 ∥ C b ∥ = 1 bounds the scalar part by the field difference. Variation of constants relative to T 0 T_0 T 0 gives local well-posedness by contraction. The boundary response is continuous and the internal forcing is locally integrable. The same argument permits continuation while states remain bounded and strictly inside the regularity region. Subtracting the mild equations and using (25) and (26) yields
D Z ( t ) ≤ M e − σ ( t − t 0 ) D Z ( t 0 ) + G B ∥ d u ∥ ∞ , [ t 0 , t ] + M ∫ t 0 t e − σ ( t − s ) [ ℓ ( R ) D Z ( s ) + q d ( s ) ] d s . D_Z(t)\le M e^{-\sigma(t-t_0)}D_Z(t_0)+G_B\|d_u\|_{\infty,[t_0,t]}
+M\int_{t_0}^{t}e^{-\sigma(t-s)}[\ell(R)D_Z(s)+q_d(s)]\,\mathrm{d}s. D Z ( t ) ≤ M e − σ ( t − t 0 ) D Z ( t 0 ) + G B ∥ d u ∥ ∞ , [ t 0 , t ] + M ∫ t 0 t e − σ ( t − s ) [ ℓ ( R ) D Z ( s ) + q d ( s )] d s .
The kernel k 0 ( t ) = M ℓ ( R ) e − σ t k_0(t)=M\ell(R)e^{-\sigma t} k 0 ( t ) = M ℓ ( R ) e − σ t has convolution powers and resolvent
k 0 ∗ n ( t ) = [ M ℓ ( R ) ] n ( n − 1 ) ! t n − 1 e − σ t , k_0^{*n}(t)=\frac{[M\ell(R)]^n}{(n-1)!}t^{n-1}e^{-\sigma t}, k 0 ∗ n ( t ) = ( n − 1 )! [ M ℓ ( R ) ] n t n − 1 e − σ t ,
r 0 ( t ) = ∑ n ≥ 1 k 0 ∗ n ( t ) = M ℓ ( R ) e − α t . r_0(t)=\sum_{n\ge1}k_0^{*n}(t)=M\ell(R)e^{-\alpha t}. r 0 ( t ) = n ≥ 1 ∑ k 0 ∗ n ( t ) = M ℓ ( R ) e − α t .
Iteration of the inequality therefore adds r 0 r_0 r 0 convolved with its free terms. The remainder converges to zero on every finite interval because of the factorial denominator. The identity e − σ t + ( r 0 ∗ e − σ ⋅ ) ( t ) = e − α t e^{-\sigma t}+(r_0*e^{-\sigma\cdot})(t)=e^{-\alpha t} e − σ t + ( r 0 ∗ e − σ ⋅ ) ( t ) = e − α t gives the initial and internal terms in (33). The boundary running supremum is nondecreasing, so its factor is at most G B [ 1 + M ℓ ( R ) / α ] = G B σ / α G_B[1+M\ell(R)/\alpha]=G_B\sigma/\alpha G B [ 1 + M ℓ ( R ) / α ] = G B σ / α . Indeed, its convolution with r 0 r_0 r 0 is bounded by its value at the current time multiplied by ∫ 0 t − t 0 r 0 ( s ) d s ≤ M ℓ ( R ) / α \int_{0}^{t-t_0}r_0(s)\,\mathrm{d}s\le M\ell(R)/\alpha ∫ 0 t − t 0 r 0 ( s ) d s ≤ M ℓ ( R ) / α . This also covers ℓ ( R ) = 0 \ell(R)=0 ℓ ( R ) = 0 , when r 0 = 0 r_0=0 r 0 = 0 . □
Corollary 1. Under Lemma 1, every averaged initial state with ∥ Z ˉ ( t 0 ) − Z ⋆ ∥ E < R / M \lVert\bar{Z}(t_0)-Z^\star\rVert_E<R/M ∥ Z ˉ ( t 0 ) − Z ⋆ ∥ E < R / M generates a unique global mild solution satisfying
∥ Z ˉ ( t ) − Z ⋆ ∥ E ≤ M e − α ( t − t 0 ) ∥ Z ˉ ( t 0 ) − Z ⋆ ∥ E , t ≥ t 0 . (34) \lVert\bar{Z}(t)-Z^\star\rVert_E\leq M e^{-\alpha(t-t_0)}\lVert\bar{Z}(t_0)-Z^\star\rVert_E,\qquad t\geq t_0. \tag*{(34)} ∥ Z ˉ ( t ) − Z ⋆ ∥ E ≤ M e − α ( t − t 0 ) ∥ Z ˉ ( t 0 ) − Z ⋆ ∥ E , t ≥ t 0 . (34)
Only continuity of the initial field is required.
Proof. Set Z ~ = Z ⋆ \widetilde{Z}=Z^\star Z = Z ⋆ and all disturbances to zero in (33). The resulting (34) stays strictly below R R R , excluding a first exit by continuity. The strict initial inequality gives the fixed gap R − M ∥ Z ˉ ( t 0 ) − Z ⋆ ∥ E > 0 R-M\lVert\bar{Z}(t_0)-Z^\star\rVert_E>0 R − M ∥ Z ˉ ( t 0 ) − Z ⋆ ∥ E > 0 . This gap persists on every interval of existence, so the trajectory cannot approach the boundary of the certified region while remaining bounded inside it. On a finite maximal interval, the initial-state term in the mild equation has an endpoint limit by strong continuity of T 0 T_0 T 0 . The nonlinear convolution also has a limit because N \mathcal{N} N is bounded along the trajectory and dominated convergence applies to the semigroup factor. Both the field and controller coordinates therefore have limits in the product space. This limit remains strictly inside the regularity region and restarts the local solution, excluding finite-time loss of existence.
Corollary 1 gives the decaying averaged trajectory used below. Lemma 1 controls the effect of the boundary and internal residuals produced by the exact transformation of the implemented system. The next three sections construct that transformation and bound its residuals. Section V removes the dither from the boundary condition, Section VI establishes the trace regularity that the filter identity (11) requires, and Section VII writes the transformed implemented loop as the averaged target plus explicit disturbances.
Harmonic Lifting and Coordinates
Physically, the dither launches a thermal wave into the plate whose shape is known exactly for the linear part of the model. There is a damped oscillation that penetrates a distance of order κ / ω \sqrt{\kappa/\omega} κ / ω from the heated edge and whose amplitude at the edge is of order a ω − 1 / 2 a\omega^{-1/2} a ω − 1/2 . Subtracting this wave from the temperature field leaves only the slowly varying part that the averaged target describes. The fixed-amplitude boundary dither produces a field whose supremum norm decreases with frequency. Then the subtraction of its exact linear response removes the imposed oscillation from the boundary condition. For q 0 = 1 q_0=1 q 0 = 1 , q 1 = ϑ q_1=\vartheta q 1 = ϑ , we define
Ψ ω ( s ) = ( i ω − A − 1 ) − 1 B = b κ ∑ j = 0 1 q j cos ( j π s 2 ) cosh ( γ j ( 1 − s 1 ) ) γ j sinh γ j , γ j = μ 0 + i ω κ + j 2 π 2 . (35) \begin{aligned}
\Psi_\omega(s)&=(i\omega-A_{-1})^{-1}B\\
&=\frac{b}{\kappa}\sum_{j=0}^{1}q_j\cos(j\pi s_2)
\frac{\cosh(\gamma_j(1-s_1))}{\gamma_j\sinh\gamma_j},\\
\gamma_j&=\sqrt{\frac{\mu_0+i\omega}{\kappa}+j^2\pi^2}. \tag*{(35)}
\end{aligned} Ψ ω ( s ) γ j = ( iω − A − 1 ) − 1 B = κ b j = 0 ∑ 1 q j cos ( j π s 2 ) γ j sinh γ j cosh ( γ j ( 1 − s 1 )) , = κ μ 0 + iω + j 2 π 2 . (35)
All roots have positive real part. The field r ω = a Im ( e i θ Ψ ω ) r_\omega=a\operatorname{Im}(e^{i\theta}\Psi_\omega) r ω = a Im ( e i θ Ψ ω ) solves the linear diffusion equation with boundary command a sin θ a\sin\theta a sin θ . Thus its mild equation subtracts exactly from (3). The lifting uses the original A A A .
Set c h = coth 1 c_h=\coth1 c h = coth 1 , ℓ l i f t = max { 1 , 2 κ } \ell_{\mathrm{lift}}=\max\{1,2\kappa\} ℓ lift = max { 1 , 2 κ } , and L ρ = ∣ χ ∣ c h L_\rho=|\chi|c_h L ρ = ∣ χ ∣ c h ,
L r = ∣ b ∣ Q c h κ , D r = ∣ b ∣ c h ( Q / κ + π ∣ ϑ ∣ ) κ , L r , γ = L r + ( 2 L r ) 1 − γ D r γ , L e = ∣ χ ∣ ( μ 0 2 + 4 κ e 2 − 1 ) . (36) \begin{aligned}
L_r&=\frac{|b|Q c_h}{\sqrt{\kappa}},\\
D_r&=\frac{|b|c_h(Q/\kappa+\pi|\vartheta|)}{\sqrt{\kappa}},\\
L_{r,\gamma}&=L_r+(2L_r)^{1-\gamma}D_r^\gamma,\\
L_e&=|\chi|\left(\frac{\mu_0}{2}+\frac{4\kappa}{e^2-1}\right). \tag*{(36)}
\end{aligned} L r D r L r , γ L e = κ ∣ b ∣ Q c h , = κ ∣ b ∣ c h ( Q / κ + π ∣ ϑ ∣ ) , = L r + ( 2 L r ) 1 − γ D r γ , = ∣ χ ∣ ( 2 μ 0 + e 2 − 1 4 κ ) . (36)
Lemma 2. For ω ≥ ω l i f t \omega\geq\omega_{\mathrm{lift}} ω ≥ ω lift , the following holds
∥ r ω ∥ ∞ ≤ L r a ε , ∥ r ω ∥ C γ ≤ L r , γ a ε 1 − γ (37) \lVert r_\omega\rVert_\infty\leq L_r a\varepsilon,\qquad
\lVert r_\omega\rVert_{C^\gamma}\leq L_{r,\gamma}a\varepsilon^{1-\gamma} \tag*{(37)} ∥ r ω ∥ ∞ ≤ L r a ε , ∥ r ω ∥ C γ ≤ L r , γ a ε 1 − γ (37) uniformly in time and phase. The midpoint trace ρ ω = C b r ω \rho_\omega=C_b r_\omega ρ ω = C b r ω satisfies
ρ ω = a χ ε sin ( θ − π / 4 ) + e ω , (38) \rho_\omega=a\chi\varepsilon\sin(\theta-\pi/4)+e_\omega, \tag*{(38)} ρ ω = a χ ε sin ( θ − π /4 ) + e ω , (38) ∣ e ω ∣ ≤ L e a ε 3 , |e_\omega|\leq L_ea\varepsilon^3, ∣ e ω ∣ ≤ L e a ε 3 , and
∣ ρ ω ∣ ≤ L ρ a ε , ∣ ρ ˙ ω ∣ ≤ L ρ a / ε . (39) |\rho_\omega|\leq L_\rho a\varepsilon,\qquad
|\dot{\rho}_\omega|\leq L_\rho a/\varepsilon. \tag*{(39)} ∣ ρ ω ∣ ≤ L ρ a ε , ∣ ρ ˙ ω ∣ ≤ L ρ a / ε . (39)
Proof. Separation of variables gives (35). The root bounds ∣ γ j ∣ ≥ ω / κ |\gamma_j|\geq\sqrt{\omega/\kappa} ∣ γ j ∣ ≥ ω / κ and Re γ j ≥ ω / ( 2 κ ) ≥ 1 \operatorname{Re}\gamma_j\geq\sqrt{\omega/(2\kappa)}\geq1 Re γ j ≥ ω / ( 2 κ ) ≥ 1 imply, for 0 ≤ y ≤ 1 0\leq y\leq1 0 ≤ y ≤ 1 ,
∣ cosh ( γ j ( 1 − y ) ) sinh γ j ∣ ≤ c h , ∣ sinh ( γ j ( 1 − y ) ) sinh γ j ∣ ≤ c h . \left|\frac{\cosh(\gamma_j(1-y))}{\sinh\gamma_j}\right|\leq c_h,\qquad
\left|\frac{\sinh(\gamma_j(1-y))}{\sinh\gamma_j}\right|\leq c_h. sinh γ j cosh ( γ j ( 1 − y )) ≤ c h , sinh γ j sinh ( γ j ( 1 − y )) ≤ c h .
Consequently ∥ r ω ∥ ∞ ≤ L r a ε \lVert r_\omega\rVert_\infty\leq L_ra\varepsilon ∥ r ω ∥ ∞ ≤ L r a ε and ∥ ∇ r ω ∥ ∞ ≤ D r a \lVert\nabla r_\omega\rVert_\infty\leq D_ra ∥ ∇ r ω ∥ ∞ ≤ D r a . The convex-domain inequality ∥ f ∥ C γ ≤ ( 2 ∥ f ∥ ∞ ) 1 − γ ∥ ∇ f ∥ ∞ \lVert f\rVert_{C^\gamma}\leq(2\lVert f\rVert_\infty)^{1-\gamma}\lVert\nabla f\rVert_\infty ∥ f ∥ C γ ≤ ( 2 ∥ f ∥ ∞ ) 1 − γ ∥ ∇ f ∥ ∞ together with ε ≤ 1 \varepsilon\leq1 ε ≤ 1 proves (37).
Midpoint evaluation gives (6), so ∣ H b ( i ω ) ∣ ≤ L ρ ε |H_b(i\omega)|\leq L_\rho\varepsilon ∣ H b ( iω ) ∣ ≤ L ρ ε , which proves (39). Split the harmonic remainder as
H b − χ ( i ω ) − 1 / 2 = χ [ ( μ 0 + i ω ) − 1 / 2 − ( i ω ) − 1 / 2 ] + χ ( μ 0 + i ω ) − 1 / 2 ( coth γ 0 − 1 ) . H_b-\chi(i\omega)^{-1/2}
=\chi\left[(\mu_0+i\omega)^{-1/2}-(i\omega)^{-1/2}\right]
+\chi(\mu_0+i\omega)^{-1/2}(\coth\gamma_0-1). H b − χ ( iω ) − 1/2 = χ [ ( μ 0 + iω ) − 1/2 − ( iω ) − 1/2 ] + χ ( μ 0 + iω ) − 1/2 ( coth γ 0 − 1 ) .
Integration in μ 0 \mu_0 μ 0 bounds the first term by ∣ χ ∣ μ 0 ω − 3 / 2 / 2 |\chi|\mu_0\omega^{-3/2}/2 ∣ χ ∣ μ 0 ω − 3/2 /2 . For x ω = ω / ( 2 κ ) ≥ 1 x_\omega=\sqrt{\omega/(2\kappa)}\geq1 x ω = ω / ( 2 κ ) ≥ 1 ,
∣ coth γ 0 − 1 ∣ ≤ 2 e − 2 x ω 1 − e − 2 , x ω 2 e − 2 x ω ≤ e − 2 . |\coth\gamma_0-1|\leq\frac{2e^{-2x_\omega}}{1-e^{-2}},\qquad
x_\omega^2e^{-2x_\omega}\leq e^{-2}. ∣ coth γ 0 − 1∣ ≤ 1 − e − 2 2 e − 2 x ω , x ω 2 e − 2 x ω ≤ e − 2 .
These bounds give the value of L e L_e L e in (36), proving (38) and (7).
The second coordinate change acts on the filter state. The product m J mJ m J in g ^ \hat{g} g ^ contains a term at twice the dither frequency and the response to the ripple itself. Both are removed exactly, so that the remaining variable ξ \xi ξ obeys the averaged filter equation up to residuals that Section VII bounds.
We introduce the triangular change of coordinates
x = w − r ω , z = C b x , d 2 , ω = 1 2 λ χ ε 2 φ ′ ( z ) cos ( 2 θ ) , ξ = p + m φ ( z ) − d 2 , ω , Z ω = ( x , v , ξ ) . (40) \begin{aligned}
x&=w-r_\omega,\\
z&=C_bx,\\
d_{2,\omega}&=\frac{1}{2}\lambda\chi\varepsilon^2\varphi'(z)\cos(2\theta),\\
\xi&=p+m\varphi(z)-d_{2,\omega},\\
Z_\omega&=(x,v,\xi). \tag*{(40)}
\end{aligned} x z d 2 , ω ξ Z ω = w − r ω , = C b x , = 2 1 λ χ ε 2 φ ′ ( z ) cos ( 2 θ ) , = p + m φ ( z ) − d 2 , ω , = ( x , v , ξ ) . (40)
Its inverse is w = x + r ω w=x+r_\omega w = x + r ω and p = ξ + d 2 , ω − m φ ( C b x ) p=\xi+d_{2,\omega}-m\varphi(C_bx) p = ξ + d 2 , ω − m φ ( C b x ) . We define
Δ ω = d 2 , ω + m [ φ ( z + ρ ω ) − φ ( z ) ] . \Delta_\omega=d_{2,\omega}+m\left[\varphi(z+\rho_\omega)-\varphi(z)\right]. Δ ω = d 2 , ω + m [ φ ( z + ρ ω ) − φ ( z ) ] .
Then g ^ = ξ + Δ ω \hat{g}=\xi+\Delta_\omega g ^ = ξ + Δ ω exactly.
Lemma 3. Suppose ω ≥ ω l i f t \omega\ge\omega_{\mathrm{lift}} ω ≥ ω lift , ∥ x − w ⋆ ∥ ∞ ≤ R \lVert x-w^\star\rVert_\infty\le R ∥ x − w ⋆ ∥ ∞ ≤ R , and ∥ r ω ∥ ∞ ≤ η ⋆ \lVert r_\omega\rVert_\infty\le\eta_\star ∥ r ω ∥ ∞ ≤ η ⋆ . Then
∣ Δ ω ∣ ≤ K Δ ε 2 (41) |\Delta_\omega| \le K_\Delta\varepsilon^2 \tag*{(41)} ∣ Δ ω ∣ ≤ K Δ ε 2 (41)
where K Δ = λ Φ 1 ( ∣ χ ∣ / 2 + 2 L ρ ) K_\Delta= \lambda\Phi_1(|\chi|/2 + 2L_\rho) K Δ = λ Φ 1 ( ∣ χ ∣/2 + 2 L ρ ) . With
δ ω = L r α ε + K Δ ε 2 , (42) \delta_\omega= L_r\alpha\varepsilon+ K_\Delta\varepsilon^2, \tag*{(42)} δ ω = L r α ε + K Δ ε 2 , (42)
the coordinate error satisfies
∥ X ω − Z ω ∥ E ≤ δ ω . (43) \lVert X_\omega-Z_\omega\rVert_E \le\delta_\omega. \tag*{(43)} ∥ X ω − Z ω ∥ E ≤ δ ω . (43)
Proof. The two terms in Δ ω \Delta_\omega Δ ω are bounded by λ ∣ χ ∣ Φ 1 ε 2 / 2 \lambda|\chi|\Phi_1\varepsilon^2/2 λ ∣ χ ∣ Φ 1 ε 2 /2 and ( 2 λ ε / a ) Φ 1 L ρ a ε (2\lambda\varepsilon/a)\Phi_1L_\rho a\varepsilon ( 2 λ ε / a ) Φ 1 L ρ a ε , respectively. Adding their sum to the field bound proves (43). Only cost derivatives enter these bounds. Under a cost offset, p c = p − m c p_c=p-mc p c = p − m c also leaves ξ \xi ξ unchanged. □
Spatial and Boundary-Trace Regularity
This section addresses the filter identity (11). The analysis multiplies the rate of change of the measured cost by the demodulating signal, so the time derivative of the sensor trace z = C b x z=C_bx z = C b x must exist in an integrable sense. Since the heater acts through the boundary, the response to a change of command is singular at startup and a bounded derivative cannot be expected. In what follows, it is shown that the trace is absolutely continuous and that its derivative, weighted by the averaged decay rate, is bounded uniformly in time and frequency (Lemma 6). Readers willing to accept Lemmas 4–6 can proceed to Section VII.
We establish the spatial bounds and integrable trace derivative needed for the filter transformation, without initial boundary compatibility. The boundary kernel is singular at zero, so a bounded pointwise time derivative is not expected at startup. Spatial Hölder regularity provides the integrable estimate used below. Convolutions start at t 0 t_0 t 0 , so ( S ∗ H ) ( t ) = ∫ t 0 t S ( t − s ) H ( s ) d s (S*H)(t)=\int_{t_0}^{t}S(t-s)H(s)\,\mathrm{d}s ( S ∗ H ) ( t ) = ∫ t 0 t S ( t − s ) H ( s ) d s . With e = x − w ⋆ e=x-w^\star e = x − w ⋆ , H = F ( x + r ω ) − F ( w ⋆ ) H=F(x+r_\omega)-F(w^\star) H = F ( x + r ω ) − F ( w ⋆ ) , and U = v − u ⋆ − k P g ^ U=v-u^\star-k_P\widehat{g} U = v − u ⋆ − k P g , the centered mild equation is
e ( t ) = S ( t − t 0 ) e ( t 0 ) + Y H ( t ) + ( K B ∗ U ) ( t ) , (44) e(t)=S(t-t_0)e(t_0)+Y_H(t)+(K_B*U)(t), \tag*{(44)} e ( t ) = S ( t − t 0 ) e ( t 0 ) + Y H ( t ) + ( K B ∗ U ) ( t ) , (44)
where Y H = S ∗ H Y_H=S*H Y H = S ∗ H .
Kernel and spatial estimates
Lemma 4. For 0 < γ < 1 0<\gamma<1 0 < γ < 1 , define
c S = 2 1 − γ κ − γ / 2 , c_S=2^{1-\gamma}\kappa^{-\gamma/2}, c S = 2 1 − γ κ − γ /2 ,
c A = 2 γ ( 2 + γ / 2 ) Γ ( 1 + γ / 2 ) , c_A=2^\gamma(2+\gamma/2)\Gamma(1+\gamma/2), c A = 2 γ ( 2 + γ /2 ) Γ ( 1 + γ /2 ) ,
c B = ∣ b ∣ Q ( 1 + 2 γ / 2 c S ) ( 1 + 2 π κ ) . (45) c_B=|b|Q(1+2^{\gamma/2}c_S)\left(1+\sqrt{\frac{2}{\pi\kappa}}\right). \tag*{(45)} c B = ∣ b ∣ Q ( 1 + 2 γ /2 c S ) ( 1 + π κ 2 ) . (45)
Here Γ \Gamma Γ is Euler’s gamma function. For t > 0 t>0 t > 0 and f ∈ X f\in\mathcal{X} f ∈ X ,
∥ S ( t ) f ∥ C γ ≤ e − μ 0 t ( 1 + c S t − γ / 2 ) ∥ f ∥ ∞ . (46) \lVert S(t)f\rVert_{C^\gamma}\le e^{-\mu_0t}(1+c_St^{-\gamma/2})\lVert f\rVert_\infty. \tag*{(46)} ∥ S ( t ) f ∥ C γ ≤ e − μ 0 t ( 1 + c S t − γ /2 ) ∥ f ∥ ∞ . (46)
For f ∈ C γ ( Ω ‾ ) f\in C^\gamma(\overline{\Omega}) f ∈ C γ ( Ω ) ,
∥ S ( t ) f ∥ C γ ≤ e − μ 0 t ∥ f ∥ C γ . (47) \lVert S(t)f\rVert_{C^\gamma}\le e^{-\mu_0t}\lVert f\rVert_{C^\gamma}. \tag*{(47)} ∥ S ( t ) f ∥ C γ ≤ e − μ 0 t ∥ f ∥ C γ . (47)
The boundary kernel satisfies
∥ K B ( t ) ∥ C γ ≤ c B e − μ 0 t ( 1 + t − ( 1 + γ ) / 2 ) , (48) \lVert K_B(t)\rVert_{C^\gamma}\le c_Be^{-\mu_0t}(1+t^{-(1+\gamma)/2}), \tag*{(48)} ∥ K B ( t ) ∥ C γ ≤ c B e − μ 0 t ( 1 + t − ( 1 + γ ) /2 ) , (48)
and, for f ∈ C γ ( Ω ‾ ) f\in C^\gamma(\overline{\Omega}) f ∈ C γ ( Ω ) ,
∥ A S ( t ) f ∥ ∞ ≤ e − μ 0 t [ μ 0 ∥ f ∥ ∞ + c A κ γ / 2 t γ / 2 − 1 [ f ] γ ] . (49) \lVert AS(t)f\rVert_\infty\le e^{-\mu_0t}\left[\mu_0\lVert f\rVert_\infty+c_A\kappa^{\gamma/2}t^{\gamma/2-1}[f]_\gamma\right]. \tag*{(49)} ∥ A S ( t ) f ∥ ∞ ≤ e − μ 0 t [ μ 0 ∥ f ∥ ∞ + c A κ γ /2 t γ /2 − 1 [ f ] γ ] . (49)
Proof. Extend f f f evenly and 2 2 2 -periodically in each coordinate to R 2 \mathbb{R}^2 R 2 . The folding map is 1 1 1 -Lipschitz, so ∥ f ~ ∥ ∞ = ∥ f ∥ ∞ \lVert\widetilde{f}\rVert_\infty=\lVert f\rVert_\infty ∥ f ∥ ∞ = ∥ f ∥ ∞ and [ f ~ ] γ = [ f ] γ [\widetilde{f}]_\gamma=[f]_\gamma [ f ] γ = [ f ] γ . With G t ( r ) = ( 4 π κ t ) − 1 exp ( − ∣ r ∣ 2 / ( 4 κ t ) ) G_t(r)=(4\pi\kappa t)^{-1}\exp(-|r|^2/(4\kappa t)) G t ( r ) = ( 4 π κ t ) − 1 exp ( − ∣ r ∣ 2 / ( 4 κ t )) ,
S ( t ) f ( s ) = e − μ 0 t ∫ R 2 G t ( r ) f ~ ( s − r ) d r . (50) S(t)f(s)=e^{-\mu_0t}\int_{\mathbb{R}^2}G_t(r)\widetilde{f}(s-r)\,\mathrm{d}r. \tag*{(50)} S ( t ) f ( s ) = e − μ 0 t ∫ R 2 G t ( r ) f ( s − r ) d r . (50)
The Gaussian has unit mass and ∥ ∇ G t ∥ 1 = π / ( 2 κ t ) ≤ ( κ t ) − 1 / 2 \lVert\nabla G_t\rVert_1=\sqrt{\pi}/(2\sqrt{\kappa t})\le(\kappa t)^{-1/2} ∥ ∇ G t ∥ 1 = π / ( 2 κ t ) ≤ ( κ t ) − 1/2 . Convolution therefore bounds the supremum by e − μ 0 t ∥ f ∥ ∞ e^{-\mu_0t}\lVert f\rVert_\infty e − μ 0 t ∥ f ∥ ∞ and the gradient by e − μ 0 t ( κ t ) − 1 / 2 ∥ f ∥ ∞ e^{-\mu_0t}(\kappa t)^{-1/2}\lVert f\rVert_\infty e − μ 0 t ( κ t ) − 1/2 ∥ f ∥ ∞ . The interpolation inequality [ g ] γ ≤ ( 2 ∥ g ∥ ∞ ) 1 − γ ∥ ∇ g ∥ ∞ γ [g]_\gamma\le(2\lVert g\rVert_\infty)^{1-\gamma}\lVert\nabla g\rVert_\infty^\gamma [ g ] γ ≤ ( 2 ∥ g ∥ ∞ ) 1 − γ ∥ ∇ g ∥ ∞ γ on the convex square proves (46). Applying (50) to Hölder differences of translates proves (47). These are estimates on the Hölder norm. They do not require strong continuity of the semigroup on the full Hölder space.
For the interval Neumann kernel K N K_N K N , separation gives
K B ( t ) ( s ) = b e − μ 0 t K N ( t , s 1 , 0 ) [ 1 + γ e − κ π 2 t cos ( π s 2 ) ] . (51) K_B(t)(s)=be^{-\mu_0t}K_N(t,s_1,0)[1+\gamma e^{-\kappa\pi^2t}\cos(\pi s_2)]. \tag*{(51)} K B ( t ) ( s ) = b e − μ 0 t K N ( t , s 1 , 0 ) [ 1 + γ e − κ π 2 t cos ( π s 2 )] . (51)
By expansion of the the cosine, its positivity implies
0 ≤ K N ( t , y , 0 ) ≤ 1 + 2 ∫ 0 ∞ e − π κ t r 2 d r = 1 + ( π κ t ) − 1 / 2 . 0\le K_N(t,y,0)\le1+2\int_0^\infty e^{-\pi\kappa t r^2}\,\mathrm{d}r=1+(\pi\kappa t)^{-1/2}. 0 ≤ K N ( t , y , 0 ) ≤ 1 + 2 ∫ 0 ∞ e − π κ t r 2 d r = 1 + ( π κ t ) − 1/2 .
Using this bound in (51), we then apply (46) to K B ( t ) = S ( t / 2 ) K B ( t / 2 ) K_B(t)=S(t/2)K_B(t/2) K B ( t ) = S ( t /2 ) K B ( t /2 ) . The expansion of the factors 1 + 2 γ / 2 c S t − γ / 2 1+2^{\gamma/2}c_St^{-\gamma/2} 1 + 2 γ /2 c S t − γ /2 and 1 + 2 / ( π κ ) t − 1 / 2 1+\sqrt{2/(\pi\kappa)}t^{-1/2} 1 + 2/ ( π κ ) t − 1/2 proves (48) with (45). Indeed, every power in the product is bounded by 1 + t − ( 1 + γ ) / 2 1+t^{-(1+\gamma)/2} 1 + t − ( 1 + γ ) /2 , and c B c_B c B is the sum of its coefficients.
We then write S N ( t ) = e μ 0 t S ( t ) S_N(t)=e^{\mu_0t}S(t) S N ( t ) = e μ 0 t S ( t ) . The cancellation of ∫ R 2 ∂ t G t = 0 \int_{\mathbb{R}^2}\partial_tG_t=0 ∫ R 2 ∂ t G t = 0 gives
κ Δ S N ( t ) f ( s ) = ∫ R 2 ∂ t G t ( r ) [ f ~ ( s − r ) − f ~ ( s ) ] d r . \kappa\Delta S_N(t)f(s)=\int_{\mathbb{R}^2}\partial_tG_t(r)[\widetilde{f}(s-r)-\widetilde{f}(s)]\,\mathrm{d}r. κ Δ S N ( t ) f ( s ) = ∫ R 2 ∂ t G t ( r ) [ f ( s − r ) − f ( s )] d r .
The Gaussian identities are such that
∣ ∂ t G t ( r ) ∣ ≤ [ t − 1 + ∣ r ∣ 2 / ( 4 κ t 2 ) ] G t ( r ) , |\partial_tG_t(r)|\le[t^{-1}+|r|^2/(4\kappa t^2)]G_t(r), ∣ ∂ t G t ( r ) ∣ ≤ [ t − 1 + ∣ r ∣ 2 / ( 4 κ t 2 )] G t ( r ) ,
∫ R 2 ∣ r ∣ q G t ( r ) d r = ( 4 κ t ) q / 2 Γ ( 1 + q / 2 ) \int_{\mathbb{R}^2}|r|^qG_t(r)\,\mathrm{d}r=(4\kappa t)^{q/2}\Gamma(1+q/2) ∫ R 2 ∣ r ∣ q G t ( r ) d r = ( 4 κ t ) q /2 Γ ( 1 + q /2 )
with q = γ q=\gamma q = γ and q = γ + 2 q=\gamma+2 q = γ + 2 yield e − μ 0 t ∫ ∣ ∂ t G t ( r ) ∣ ∣ r ∣ γ d r ≤ c A κ γ / 2 t γ / 2 − 1 e^{-\mu_0t}\int|\partial_tG_t(r)||r|^\gamma\,\mathrm{d}r\le c_A\kappa^{\gamma/2}t^{\gamma/2-1} e − μ 0 t ∫ ∣ ∂ t G t ( r ) ∣∣ r ∣ γ d r ≤ c A κ γ /2 t γ /2 − 1 . Since A S ( t ) = e − μ 0 t ( κ Δ − μ 0 ) S N ( t ) AS(t)=e^{-\mu_0t}(\kappa\Delta-\mu_0)S_N(t) A S ( t ) = e − μ 0 t ( κ Δ − μ 0 ) S N ( t ) , we prove (49). □
The integration of the bounds in Lemma 4 yields
I S = μ 0 − 1 + c S Γ ( 1 − γ / 2 ) μ 0 γ / 2 − 1 , I_S=\mu_0^{-1}+c_S\Gamma(1-\gamma/2)\mu_0^{\gamma/2-1}, I S = μ 0 − 1 + c S Γ ( 1 − γ /2 ) μ 0 γ /2 − 1 ,
I B = c B [ μ 0 − 1 + Γ ( ( 1 − γ ) / 2 ) μ 0 ( γ − 1 ) / 2 ] , I_B=c_B[\mu_0^{-1}+\Gamma((1-\gamma)/2)\mu_0^{(\gamma-1)/2}], I B = c B [ μ 0 − 1 + Γ (( 1 − γ ) /2 ) μ 0 ( γ − 1 ) /2 ] ,
I A = 1 + c A κ γ / 2 Γ ( γ / 2 ) μ 0 − γ / 2 . (52) I_A=1+c_A\kappa^{\gamma/2}\Gamma(\gamma/2)\mu_0^{-\gamma/2}. \tag*{(52)} I A = 1 + c A κ γ /2 Γ ( γ /2 ) μ 0 − γ /2 . (52)
They bound the integrals of ∥ S ( t ) ∥ X → C γ \lVert S(t)\rVert_{X\to C^\gamma} ∥ S ( t ) ∥ X → C γ , ∥ K B ( t ) ∥ C γ \lVert K_B(t)\rVert_{C^\gamma} ∥ K B ( t ) ∥ C γ , and ∥ A S ( t ) ∥ C γ → X \lVert AS(t)\rVert_{C^\gamma\to X} ∥ A S ( t ) ∥ C γ → X , respectively. All integrals are finite. Since 0 < γ < 1 0<\gamma<1 0 < γ < 1 , the singularities are integrable at zero, and μ 0 > 0 \mu_0>0 μ 0 > 0 controls the tail.
Lemma 5. Suppose Assumptions 1–4 and (5) hold. Fix 0 < a ≤ a 0 0<a\le a_0 0 < a ≤ a 0 and ω ≥ ω l i f t > 1 \omega\ge\omega_{\mathrm{lift}}>1 ω ≥ ω lift > 1 . Consider a nominal mild solution on I T = [ t 0 , t 0 + T ] I_T=[t_0,t_0+T] I T = [ t 0 , t 0 + T ] satisfying
Z ω ( t ) ∈ B R ( Z ⋆ ) , ∥ r ω ( t ) ∥ ∞ ≤ η ⋆ . (53) Z_\omega(t)\in B_R(Z^\star),\qquad\lVert r_\omega(t)\rVert_\infty\le\eta_\star. \tag*{(53)} Z ω ( t ) ∈ B R ( Z ⋆ ) , ∥ r ω ( t ) ∥ ∞ ≤ η ⋆ . (53)
Set G = R + K Δ G=R+K_\Delta G = R + K Δ and U ⋆ = R + k P G U_\star=R+k_PG U ⋆ = R + k P G . Define
W γ = I S ∥ F ( w ⋆ ) + d ∥ ∞ + I B ∣ u ⋆ ∣ , W_\gamma=I_S\lVert F(w^\star)+d\rVert_\infty+I_B|u^\star|, W γ = I S ∥ F ( w ⋆ ) + d ∥ ∞ + I B ∣ u ⋆ ∣ ,
E i n = H 0 + L r , γ a 0 , (54) E_{\mathrm{in}}=H_0+L_{r,\gamma}a_0, \tag*{(54)} E in = H 0 + L r , γ a 0 , (54)
and
H ∞ = L F ( R + η ⋆ ) , E γ = E i n + I S H ∞ + I B U ⋆ , H γ = H ∞ + 2 H F + L F ( 2 W γ + E γ + L r , γ a 0 ) . (55) H_{\infty}=L_F(R+\eta_\star),
\qquad
E_\gamma=E_{\mathrm{in}}+I_SH_{\infty}+I_BU_\star,
\qquad
H_\gamma=H_{\infty}+2H_F+L_F(2W_\gamma+E_\gamma+L_{r,\gamma}a_0). \tag*{(55)} H ∞ = L F ( R + η ⋆ ) , E γ = E in + I S H ∞ + I B U ⋆ , H γ = H ∞ + 2 H F + L F ( 2 W γ + E γ + L r , γ a 0 ) . (55)
Then
∥ w ⋆ ∥ C γ ≤ W γ , \|w^\star\|_{C^\gamma}\le W_\gamma, ∥ w ⋆ ∥ C γ ≤ W γ ,
sup t ∈ I T ∥ e ( t ) ∥ C γ ≤ E γ , \sup_{t\in I_T}\|e(t)\|_{C^\gamma}\le E_\gamma, t ∈ I T sup ∥ e ( t ) ∥ C γ ≤ E γ ,
sup t ∈ I T ∥ H ( t ) ∥ C γ ≤ H γ . (56) \sup_{t\in I_T}\|H(t)\|_{C^\gamma}\le H_\gamma. \tag*{(56)} t ∈ I T sup ∥ H ( t ) ∥ C γ ≤ H γ . (56)
Moreover, Y H ∈ C 1 ( I T , X ) Y_H\in C^1(I_T,\mathcal{X}) Y H ∈ C 1 ( I T , X ) , with a right derivative at t 0 t_0 t 0 , and
Y H ′ ( t ) = H ( t ) + ∫ t 0 t A S ( t − s ) H ( s ) d s . (57) Y_H'(t)=H(t)+\int_{t_0}^{t}AS(t-s)H(s)\,\mathrm{d}s. \tag*{(57)} Y H ′ ( t ) = H ( t ) + ∫ t 0 t A S ( t − s ) H ( s ) d s . (57)
Then
∫ 0 ∞ ∥ A S ( s ) e ( t 0 ) ∥ ∞ d s ≤ H i n , \int_0^\infty\|AS(s)e(t_0)\|_\infty\,\mathrm{d}s\le H_{\mathrm{in}}, ∫ 0 ∞ ∥ A S ( s ) e ( t 0 ) ∥ ∞ d s ≤ H in ,
sup t ∈ I T ∥ Y H ′ ( t ) ∥ ∞ ≤ L H . (58) \sup_{t\in I_T}\|Y_H'(t)\|_\infty\le L_H. \tag*{(58)} t ∈ I T sup ∥ Y H ′ ( t ) ∥ ∞ ≤ L H . (58)
for frequency-independent constants
H i n = I A E i n , L H = H ∞ + I A H γ . (59) H_{\mathrm{in}}=I_AE_{\mathrm{in}},
\qquad
L_H=H_\infty+I_AH_\gamma. \tag*{(59)} H in = I A E in , L H = H ∞ + I A H γ . (59)
The constants are independent of T T T , initial time, and phase, with a 0 a_0 a 0 and the initial-family bound fixed.
Proof. (41) and ε ≤ 1 \varepsilon\le1 ε ≤ 1 give ∣ q ^ ∣ ≤ G |\widehat q|\le G ∣ q ∣ ≤ G and ∣ U ∣ ≤ U ⋆ |U|\le U_\star ∣ U ∣ ≤ U ⋆ . The reaction Lipschitz bound gives ∥ H ∥ ∞ ≤ H ∞ \|H\|_\infty\le H_\infty ∥ H ∥ ∞ ≤ H ∞ . Applying Lemma 4 to the stationary mild equation at time T 1 > 0 T_1>0 T 1 > 0 yields
∥ w ⋆ ∥ C γ ≤ e − μ 0 T 1 ( 1 + c S T 1 − γ / 2 ) ∥ w ⋆ ∥ ∞ + W γ . \|w^\star\|_{C^\gamma}\le e^{-\mu_0T_1}(1+c_ST_1^{-\gamma/2})\|w^\star\|_\infty+W_\gamma. ∥ w ⋆ ∥ C γ ≤ e − μ 0 T 1 ( 1 + c S T 1 − γ /2 ) ∥ w ⋆ ∥ ∞ + W γ .
Letting T 1 → ∞ T_1\to\infty T 1 → ∞ proves the equilibrium bound in (56), with (54). Only boundedness of F ( w ⋆ ) + d F(w^\star)+d F ( w ⋆ ) + d is needed. The initial-family and lift bounds give ∥ e ( t 0 ) ∥ C γ ≤ H 0 + L r , γ a ε 1 − γ ≤ E i n \|e(t_0)\|_{C^\gamma}\le H_0+L_{r,\gamma}a\varepsilon^{1-\gamma}\le E_{\mathrm{in}} ∥ e ( t 0 ) ∥ C γ ≤ H 0 + L r , γ a ε 1 − γ ≤ E in . Equations (44), (47), and (52) then give the bound E γ E_\gamma E γ . Finally, (4) implies
[ H ] γ ≤ 2 H F + L F ( 2 [ w ⋆ ] γ + [ e ] γ + [ r ω ] γ ) , [H]_\gamma\le2H_F+L_F(2[w^\star]_\gamma+[e]_\gamma+[r_\omega]_\gamma), [ H ] γ ≤ 2 H F + L F ( 2 [ w ⋆ ] γ + [ e ] γ + [ r ω ] γ ) ,
which proves the H γ H_\gamma H γ bound in (55). Thus the spatial regularity of H H H follows after that of e e e , without a circular bootstrap. The forcing H H H is continuous into X \mathcal{X} X . By (49),
∥ A S ( r ) H ( s ) ∥ ∞ ≤ e − μ 0 r ( μ 0 + c A κ γ / 2 − 1 ) H γ . \|AS(r)H(s)\|_\infty\le e^{-\mu_0r}\left(\mu_0+c_A\kappa^{\gamma/2-1}\right)H_\gamma. ∥ A S ( r ) H ( s ) ∥ ∞ ≤ e − μ 0 r ( μ 0 + c A κ γ /2 − 1 ) H γ .
This majorant is integrable, with integral I A H γ I_AH_\gamma I A H γ . The integrand is strongly measurable into X \mathcal{X} X for r > 0 r>0 r > 0 . For each fixed positive r r r , continuity of H H H into X \mathcal{X} X suffices because A S ( r ) AS(r) A S ( r ) is bounded on that space. No continuity of H H H in the stronger Hölder norm is assumed. The time derivative must be justified at the upper integration limit, where A S ( t − s ) AS(t-s) A S ( t − s ) is singular. The factor γ γ / 2 − 1 \gamma^{\gamma/2-1} γ γ /2 − 1 is integrable because γ > 0 \gamma>0 γ > 0 . We therefore integrate the semigroup derivative first and interchange the resulting integrals. This produces an integral of a continuous X \mathcal{X} X -valued function, to which the fundamental theorem of calculus applies. The identity ∫ 0 τ A S ( r ) f d r = S ( τ ) f − f \int_0^\tau AS(r)f\,\mathrm{d}r=S(\tau)f-f ∫ 0 τ A S ( r ) f d r = S ( τ ) f − f for f ∈ C γ f\in C^\gamma f ∈ C γ and Fubini’s theorem give
Y H ( t ) = ∫ t 0 t [ H ( s ) + ( A S ∗ H ) ( s ) ] d s . Y_H(t)=\int_{t_0}^{t}[H(s)+(AS*H)(s)]\,\mathrm{d}s. Y H ( t ) = ∫ t 0 t [ H ( s ) + ( A S ∗ H ) ( s )] d s .
The domination of the convergence with the same majorant makes the integrand continuous, including its right limit at t 0 t_0 t 0 . This proves (57) and, by integration of (49), the constants and bounds in (59)–(58).
Absolute continuity and the weighted trace bound
Evaluating (44) gives
z = z f + k ∗ U , z f = z ⋆ + C b S ( t − t 0 ) e ( t 0 ) + C b Y H , (60) z=z_f+k*U,
\qquad
z_f=z^\star+C_bS(t-t_0)e(t_0)+C_bY_H, \tag*{(60)} z = z f + k ∗ U , z f = z ⋆ + C b S ( t − t 0 ) e ( t 0 ) + C b Y H , (60)
where k = C b K B = b e − μ 0 t K N ( t , 0 , 0 ) k=C_bK_B=be^{-\mu_0t}K_N(t,0,0) k = C b K B = b e − μ 0 t K N ( t , 0 , 0 ) has constant sign. Its norm is
K 0 = ∥ k ∥ L 1 ( 0 , ∞ ) = ∣ b ∣ κ μ 0 coth μ 0 / κ . (61) K_0=\|k\|_{L^1(0,\infty)}
=\frac{|b|}{\sqrt{\kappa\mu_0}}\coth\sqrt{\mu_0/\kappa}. \tag*{(61)} K 0 = ∥ k ∥ L 1 ( 0 , ∞ ) = κ μ 0 ∣ b ∣ coth μ 0 / κ . (61)
Indeed, the integral of k k k is the boundary value of the stationary solution of κ y ′ ′ − μ 0 y = 0 \kappa y''-\mu_0y=0 κ y ′′ − μ 0 y = 0 with − κ y ′ ( 0 ) = b -\kappa y'(0)=b − κ y ′ ( 0 ) = b and y ′ ( 1 ) = 0 y'(1)=0 y ′ ( 1 ) = 0 , which gives (61).
Lemma 6. Under Lemma 5, z = C b x z=C_bx z = C b x is absolutely continuous on I T I_T I T . Define
C 0 = ∣ k p λ − k l ∣ G + 2 k p λ Φ 1 L ρ , c ω = 2 k p λ Φ 1 ε / a . (62) C_0=|k_p\lambda-k_l|G+2k_p\lambda\Phi_1L_\rho,
\qquad
c_\omega=2k_p\lambda\Phi_1\varepsilon/a. \tag*{(62)} C 0 = ∣ k p λ − k l ∣ G + 2 k p λ Φ 1 L ρ , c ω = 2 k p λ Φ 1 ε / a . (62)
If
a ω ≥ 4 k p λ Φ 1 K 0 , (63) a\sqrt{\omega}\ge4k_p\lambda\Phi_1K_0, \tag*{(63)} a ω ≥ 4 k p λ Φ 1 K 0 , (63)
then
sup t ∈ I T ∫ t 0 t e − α ( t − s ) ∣ z ˙ ( s ) ∣ d s ≤ L z , \sup_{t\in I_T}\int_{t_0}^{t}e^{-\alpha(t-s)}|\dot z(s)|\,\mathrm{d}s\le L_z, t ∈ I T sup ∫ t 0 t e − α ( t − s ) ∣ z ˙ ( s ) ∣ d s ≤ L z ,
L z = 2 [ H i n + K 0 U ⋆ + L H + K 0 C 0 α ] . (64) L_z=2\left[H_{\mathrm{in}}+K_0U_\star+\frac{L_H+K_0C_0}{\alpha}\right]. \tag*{(64)} L z = 2 [ H in + K 0 U ⋆ + α L H + K 0 C 0 ] . (64)
The bound is uniform in T T T , initial time, dither phase, and frequency in the stated range.
Proof. The physical trace is known only to be continuous. The existing mild solution already determines H H H , hence Y H Y_H Y H and z f z_f z f . We keep these functions fixed and solve the resulting scalar Volterra equation in a class of absolutely continuous trial traces. Uniqueness in the larger class of continuous traces will then identify this constructed solution with the physical trace. This establishes the regularity required for differentiation with respect to time and the chain rule. Lemma 5 gives z f ′ ∈ L l o c 1 z_f'\in L^1_{\mathrm{loc}} z f ′ ∈ L loc 1 and ∥ z ¨ f ∥ L 1 ( t 0 , t 0 + T 1 ) → 0 \|\ddot z_f\|_{L^1(t_0,t_0+T_1)}\to0 ∥ z ¨ f ∥ L 1 ( t 0 , t 0 + T 1 ) → 0 as T 1 ↓ 0 T_1\downarrow0 T 1 ↓ 0 . Fix R d > 0 R_d>0 R d > 0 smaller than one quarter of the distance of z ( t 0 ) + ρ ω ( t 0 ) z(t_0)+\rho_\omega(t_0) z ( t 0 ) + ρ ω ( t 0 ) from the endpoints of the enlarged trace interval. This distance is positive by (53). On a sufficiently short I = [ t 0 , t 0 + T 1 ] I=[t_0,t_0+T_1] I = [ t 0 , t 0 + T 1 ] , the complete derivative ball
D = { ζ ∈ W 1 , 1 ( I ) : ζ ( t 0 ) = z ( t 0 ) , ∥ ζ ˙ ∥ L 1 ( I ) ≤ R d } \mathcal{D}=\{\zeta\in W^{1,1}(I):\zeta(t_0)=z(t_0),\ \|\dot\zeta\|_{L^1(I)}\le R_d\} D = { ζ ∈ W 1 , 1 ( I ) : ζ ( t 0 ) = z ( t 0 ) , ∥ ζ ˙ ∥ L 1 ( I ) ≤ R d }
has ζ + ρ ω \zeta+\rho_\omega ζ + ρ ω in that interval. Its metric is the L 1 L^1 L 1 distance of the derivatives. The common initial value identifies this set with a closed ball in L 1 ( I ) L^1(I) L 1 ( I ) . Scalar norms below are on I I I unless another interval is indicated.
We implement the controller with J ζ = φ ( ζ + ρ ω ) J_\zeta=\varphi(\zeta+\rho_\omega) J ζ = φ ( ζ + ρ ω ) from the actual initial controller states. Let ζ ^ = m φ ( ζ + ρ ω ) ( ζ + ρ ˙ ω ) \hat\zeta=m^\varphi(\zeta+\rho_\omega)(\zeta+\dot\rho_\omega) ζ ^ = m φ ( ζ + ρ ω ) ( ζ + ρ ˙ ω ) . The trial responses are absolutely continuous. Thus, (11) and the bias equation give
g ^ ˙ ζ = − λ g ^ ζ + f ζ , U ˙ ζ = ( k P λ − k I ) g ^ ζ − k P f ζ . (65) \dot{\hat{g}}_{\zeta}=-\lambda\hat{g}_{\zeta}+f_{\zeta},\qquad
\dot{U}_{\zeta}=(k_{P}\lambda-k_{I})\hat{g}_{\zeta}-k_{P}f_{\zeta}. \tag*{(65)} g ^ ˙ ζ = − λ g ^ ζ + f ζ , U ˙ ζ = ( k P λ − k I ) g ^ ζ − k P f ζ . (65)
Only the trial traces have been differentiated here. Absolute continuity of the physical trace will follow by uniqueness. For d ζ = ∥ ζ ˙ 1 − ζ ˙ 2 ∥ 1 d_{\zeta}=\lVert\dot{\zeta}_{1}-\dot{\zeta}_{2}\rVert_{1} d ζ = ∥ ζ ˙ 1 − ζ ˙ 2 ∥ 1 , the common initial value implies ∥ ζ 1 − ζ 2 ∥ ∞ ≤ d ζ \lVert\zeta_{1}-\zeta_{2}\rVert_{\infty}\le d_{\zeta} ∥ ζ 1 − ζ 2 ∥ ∞ ≤ d ζ . Therefore
∥ f ζ 1 − f ζ 2 ∥ 1 ≤ A T 1 d ζ , A T 1 = ∥ m ∥ ∞ [ Φ 1 + Φ 2 ( R d + ∥ ρ ˙ ω ∥ 1 ) ] . \lVert f_{\zeta_{1}}-f_{\zeta_{2}}\rVert_{1}\le A_{T_{1}}d_{\zeta},
\qquad
A_{T_{1}}=\lVert m\rVert_{\infty}\left[\Phi_{1}+\Phi_{2}\left(R_{d}+\lVert\dot{\rho}_{\omega}\rVert_{1}\right)\right]. ∥ f ζ 1 − f ζ 2 ∥ 1 ≤ A T 1 d ζ , A T 1 = ∥ m ∥ ∞ [ Φ 1 + Φ 2 ( R d + ∥ ρ ˙ ω ∥ 1 ) ] .
Variation of constants in (65) gives ∥ g ^ ζ 1 − g ^ ζ 2 ∥ ∞ ≤ A T 1 d ζ \lVert\hat{g}_{\zeta_{1}}-\hat{g}_{\zeta_{2}}\rVert_{\infty}\le A_{T_{1}}d_{\zeta} ∥ g ^ ζ 1 − g ^ ζ 2 ∥ ∞ ≤ A T 1 d ζ and hence
∥ U ˙ ζ 1 − U ˙ ζ 2 ∥ 1 ≤ L T 1 d ζ , L T 1 = ( ∣ k P λ − k I ∣ T 1 + k P ) A T 1 . (66) \lVert\dot{U}_{\zeta_{1}}-\dot{U}_{\zeta_{2}}\rVert_{1}\le L_{T_{1}}d_{\zeta},
\qquad
L_{T_{1}}=\left(\lvert k_{P}\lambda-k_{I}\rvert T_{1}+k_{P}\right)A_{T_{1}}. \tag*{(66)} ∥ U ˙ ζ 1 − U ˙ ζ 2 ∥ 1 ≤ L T 1 d ζ , L T 1 = ( ∣ k P λ − k I ∣ T 1 + k P ) A T 1 . (66)
This constant stays bounded as T 1 ↓ 0 T_{1}\downarrow0 T 1 ↓ 0 for fixed a , ω , R d a,\omega,R_{d} a , ω , R d . Bounded φ ′ ′ \varphi^{\prime\prime} φ ′′ suffices.
Next, we define T ζ = z f + k ∗ U ζ \mathcal{T}\zeta=z_{f}+k*U_{\zeta} T ζ = z f + k ∗ U ζ . For absolutely continuous U U U , inserting U ( s ) = U ( t 0 ) + ∫ t 0 s U ˙ ( r ) d r U(s)=U(t_{0})+\int_{t_{0}}^{s}\dot{U}(r)\,dr U ( s ) = U ( t 0 ) + ∫ t 0 s U ˙ ( r ) d r and using Fubini gives
( k ∗ U ) ′ = k ( t − t 0 ) U ( t 0 ) + k ∗ U ˙ a.e. (67) (k*U)^{\prime}=k(t-t_{0})U(t_{0})+k*\dot{U}\quad\text{a.e.} \tag*{(67)} ( k ∗ U ) ′ = k ( t − t 0 ) U ( t 0 ) + k ∗ U ˙ a.e. (67)
All trial responses share U ( t 0 ) U(t_{0}) U ( t 0 ) . Thus (66) gives
∥ ( T ζ 1 − T ζ 2 ) ′ ∥ 1 ≤ ∥ k ∥ L 1 ( 0 , T 1 ) L T 1 d ζ . (68) \lVert(\mathcal{T}\zeta_{1}-\mathcal{T}\zeta_{2})^{\prime}\rVert_{1}\le\lVert k\rVert_{L^{1}(0,T_{1})}L_{T_{1}}d_{\zeta}. \tag*{(68)} ∥( T ζ 1 − T ζ 2 ) ′ ∥ 1 ≤ ∥ k ∥ L 1 ( 0 , T 1 ) L T 1 d ζ . (68)
The shared initial trace is gives the same initial command. Therefore, k ( t − t 0 ) U ( t 0 ) k(t-t_{0})U(t_{0}) k ( t − t 0 ) U ( t 0 ) cancels when two responses are subtracted. The remaining operator has norm proportional to ∥ k ∥ L 1 ( 0 , T 1 ) \lVert k\rVert_{L^{1}(0,T_{1})} ∥ k ∥ L 1 ( 0 , T 1 ) , which tends to zero despite the kernel’s singularity at zero. Thus (68) becomes contractive without differentiation of k k k . Also, (65) yields
∥ g ^ ζ ∥ ∞ ≤ ∣ g ^ ( t 0 ) ∣ + ∥ m ∥ ∞ Φ 1 ( R d + ∥ ρ ˙ ω ∥ 1 ) , \lVert\hat{g}_{\zeta}\rVert_{\infty}\le\lvert\hat{g}(t_{0})\rvert+\lVert m\rVert_{\infty}\Phi_{1}\left(R_{d}+\lVert\dot{\rho}_{\omega}\rVert_{1}\right), ∥ g ^ ζ ∥ ∞ ≤ ∣ g ^ ( t 0 )∣ + ∥ m ∥ ∞ Φ 1 ( R d + ∥ ρ ˙ ω ∥ 1 ) ,
so ∥ U ˙ ζ ∥ 1 \lVert\dot{U}_{\zeta}\rVert_{1} ∥ U ˙ ζ ∥ 1 stays uniformly bounded over D \mathcal{D} D . Equation (67) implies
∥ ( T ζ ) ′ ∥ 1 ≤ ∥ z f ′ ∥ 1 + ∥ k ∥ L 1 ( 0 , T 1 ) ( ∣ U ( t 0 ) ∣ + ∥ U ˙ ζ ∥ 1 ) . \lVert(\mathcal{T}\zeta)^{\prime}\rVert_{1}\le\lVert z_{f}^{\prime}\rVert_{1}+\lVert k\rVert_{L^{1}(0,T_{1})}\left(\lvert U(t_{0})\rvert+\lVert\dot{U}_{\zeta}\rVert_{1}\right). ∥( T ζ ) ′ ∥ 1 ≤ ∥ z f ′ ∥ 1 + ∥ k ∥ L 1 ( 0 , T 1 ) ( ∣ U ( t 0 )∣ + ∥ U ˙ ζ ∥ 1 ) .
This tends uniformly to zero. Since T \mathcal{T} T preserves the initial value, it maps D \mathcal{D} D into itself and is a contraction for small T 1 T_{1} T 1 . Its fixed point is absolutely continuous.
For continuous trial traces, variation of constants in the original p p p equation and integration of the bias equation give
∥ U ζ 1 − U ζ 2 ∥ ∞ ≤ Λ T 1 ∥ ζ 1 − ζ 2 ∥ ∞ , \lVert U_{\zeta_{1}}-U_{\zeta_{2}}\rVert_{\infty}\le\Lambda_{T_{1}}\lVert\zeta_{1}-\zeta_{2}\rVert_{\infty}, ∥ U ζ 1 − U ζ 2 ∥ ∞ ≤ Λ T 1 ∥ ζ 1 − ζ 2 ∥ ∞ ,
where
Λ T 1 = ( k P + k I T 1 ) Φ 1 [ ∥ m ∥ ∞ + T 1 ( λ ∥ m ∥ ∞ + ∥ m ˙ ∥ ∞ ) ] . \Lambda_{T_{1}}=(k_{P}+k_{I}T_{1})\Phi_{1}\left[\lVert m\rVert_{\infty}+T_{1}\left(\lambda\lVert m\rVert_{\infty}+\lVert\dot{m}\rVert_{\infty}\right)\right]. Λ T 1 = ( k P + k I T 1 ) Φ 1 [ ∥ m ∥ ∞ + T 1 ( λ ∥ m ∥ ∞ + ∥ m ˙ ∥ ∞ ) ] .
The first factor accounts for the bias integration and proportional term. The bracket bounds the change in p + m J p+mJ p + m J using ∣ J ζ 1 − J ζ 2 ∣ ≤ Φ 1 ∣ ζ 1 − ζ 2 ∣ \lvert J_{\zeta_{1}}-J_{\zeta_{2}}\rvert\le\Phi_{1}\lvert\zeta_{1}-\zeta_{2}\rvert ∣ J ζ 1 − J ζ 2 ∣ ≤ Φ 1 ∣ ζ 1 − ζ 2 ∣ . For fixed a , ω a,\omega a , ω , ∥ k ∥ L 1 ( 0 , T 1 ) Λ T 1 < 1 \lVert k\rVert_{L^{1}(0,T_{1})}\Lambda_{T_{1}}<1 ∥ k ∥ L 1 ( 0 , T 1 ) Λ T 1 < 1 when T 1 T_{1} T 1 is small. Hence the Volterra equation has at most one continuous solution, identifying the fixed point with the physical trace.
At any restart time t 1 t_{1} t 1 , the bound ∥ e ( t 1 ) ∥ C γ ≤ E γ \lVert e(t_{1})\rVert_{C^{\gamma}}\le E_{\gamma} ∥ e ( t 1 ) ∥ C γ ≤ E γ gives
∫ t 1 t 1 + τ 1 ∣ z f , t 1 ′ ( s ) ∣ d s ≤ E γ [ μ 0 τ 1 + 2 c A κ γ / 2 γ τ 1 γ / 2 ] + L H τ 1 . \int_{t_{1}}^{t_{1}+\tau_{1}}\lvert z_{f,t_{1}}^{\prime}(s)\rvert\,ds
\le E_{\gamma}\left[\mu_{0}\tau_{1}+\frac{2c_{A}\kappa^{\gamma/2}}{\gamma}\tau_{1}^{\gamma/2}\right]+L_{H}\tau_{1}. ∫ t 1 t 1 + τ 1 ∣ z f , t 1 ′ ( s )∣ d s ≤ E γ [ μ 0 τ 1 + γ 2 c A κ γ /2 τ 1 γ /2 ] + L H τ 1 .
This tends to zero uniformly in t 1 t_{1} t 1 . On a compact interval satisfying (53) , the trace has a positive margin within its regularity interval and the controller coefficients are bounded. The radius and contraction interval can therefore be chosen uniformly over restart times. Finitely many restarts establish absolute continuity on I T I_{T} I T .
We may now differentiate (60) :
z ˙ = C b A s ( t − t 0 ) e ( t 0 ) + C b Y H ′ + k ( t − t 0 ) U ( t 0 ) + k ∗ U ˙ . (69) \dot{z}=C_{b}As(t-t_{0})e(t_{0})+C_{b}Y_{H}^{\prime}+k(t-t_{0})U(t_{0})+k*\dot{U}. \tag*{(69)} z ˙ = C b A s ( t − t 0 ) e ( t 0 ) + C b Y H ′ + k ( t − t 0 ) U ( t 0 ) + k ∗ U ˙ . (69)
Using (65) for the physical trace, ∣ g ^ ∣ ≤ G \lvert\hat{g}\rvert\le G ∣ g ^ ∣ ≤ G , and the harmonic bounds gives
∣ U ˙ ∣ ≤ C 0 + c ω ∣ z ˙ ∣ . (70) \lvert\dot{U}\rvert\le C_{0}+c_{\omega}\lvert\dot{z}\rvert. \tag*{(70)} ∣ U ˙ ∣ ≤ C 0 + c ω ∣ z ˙ ∣ . (70)
with (62) . Denote the supremum in (64) by V T V_{T} V T . It is finite by absolute continuity. For e α ( t ) = e − α t e_{\alpha}(t)=e^{-\alpha t} e α ( t ) = e − α t , commutation of convolutions gives
e α ∗ ( ∣ k ∣ ∗ ∣ z ˙ ∣ ) = ∣ k ∣ ∗ ( e α ∗ ∣ z ˙ ∣ ) ≤ K 0 V T . e_{\alpha}*\left(\lvert k\rvert*\lvert\dot{z}\rvert\right)=\lvert k\rvert*\left(e_{\alpha}*\lvert\dot{z}\rvert\right)\le K_{0}V_{T}. e α ∗ ( ∣ k ∣ ∗ ∣ z ˙ ∣ ) = ∣ k ∣ ∗ ( e α ∗ ∣ z ˙ ∣ ) ≤ K 0 V T .
Equations (58) , (69) , and (70) consequently yield
V T ≤ H i n + K 0 U ∗ + L H + K 0 C 0 α + c ω K 0 V T . V_{T}\le H_{\mathrm{in}}+K_{0}U_{*}+\frac{L_{H}+K_{0}C_{0}}{\alpha}+c_{\omega}K_{0}V_{T}. V T ≤ H in + K 0 U ∗ + α L H + K 0 C 0 + c ω K 0 V T .
Condition (63) gives c ω K 0 ≤ 1 / 2 c_{\omega}K_{0}\le1/2 c ω K 0 ≤ 1/2 , proving (64) . The unweighted norm K 0 K_{0} K 0 suffices even when α ≥ μ 0 \alpha\ge\mu_{0} α ≥ μ 0 , and all constants are uniform as stated. □ \square □
Trajectory Comparison
The regularity estimates allow us to compute the difference between the transformed implementation and the averaged equations exactly. We then apply the disturbance estimate of Lemma 1. The residuals collect everything the averaged target ignores which include the reaction evaluated on the lifted field, the second-harmonic and Taylor remainders of the cost, and the finite-frequency corrections to the demodulation. Each is proportional to a power of ε = ω − 1 / 2 \varepsilon=\omega^{-1/2} ε = ω − 1/2 .
Exact residual equations
On a nominal existence interval with Z ω ∈ B R ( Z ∗ ) Z_{\omega}\in B_{R}(Z^{*}) Z ω ∈ B R ( Z ∗ ) , impose the hypotheses of Lemmas 5 and 6. Then z = C b x z=C_{b}x z = C b x is absolutely continuous, so the scalar derivatives below exist almost everywhere and their integral identities hold throughout the interval.
The lifting removes the dither from the boundary condition while leaving the reaction evaluated at x + r ω x+r_{\omega} x + r ω . The remaining boundary command is v − k P g ^ v-k_{P}\hat{g} v − k P g ^ . Using g ^ = ξ + Δ ω \hat{g}=\xi+\Delta_{\omega} g ^ = ξ + Δ ω therefore writes the transformed field and bias as the averaged target with disturbances
d ω = F ( x + r ω ) − F ( x ) , d u = − k P Δ ω , d v = − k I Δ ω . (71) d_{\omega}=F(x+r_{\omega})-F(x),\qquad
d_{u}=-k_{P}\Delta_{\omega},\qquad
d_{v}=-k_{I}\Delta_{\omega}. \tag*{(71)} d ω = F ( x + r ω ) − F ( x ) , d u = − k P Δ ω , d v = − k I Δ ω . (71)
Here d v d_{v} d v enters the boundary command. To compute the filter residual, we define
E φ = φ ( z + ρ ω ) − φ ( z ) − a ε g ( x ) sin ( θ − π / 4 ) . (72) E_{\varphi}=\varphi(z+\rho_{\omega})-\varphi(z)-a\varepsilon g(x)\sin(\theta-\pi/4). \tag*{(72)} E φ = φ ( z + ρ ω ) − φ ( z ) − a ε g ( x ) sin ( θ − π /4 ) . (72)
By Taylor’s formula and (38) , we obtain
∣ E φ ∣ ≤ Φ 1 L e a ε 3 + 1 2 Φ 2 L ρ 2 a 2 ε 2 . (73) |E_\varphi| \le\Phi_1 L_e a\varepsilon^3+\frac{1}{2}\Phi_2L_\rho^2a^2\varepsilon^2. \tag*{(73)} ∣ E φ ∣ ≤ Φ 1 L e a ε 3 + 2 1 Φ 2 L ρ 2 a 2 ε 2 . (73)
The first term comes from the error in the linear harmonic response. The second is the quadratic Taylor remainder of the cost. Differentiation of (40) almost everywhere yields
ξ ˙ = − λ ξ − λ Δ ω − m ˙ [ φ ( z + ρ ω ) − φ ( z ) ] + m φ ′ ( z ) z ˙ − d ˙ 2 , ω . \dot{\xi}=-\lambda\xi-\lambda\Delta_\omega-\dot{m}[\varphi(z+\rho\omega)-\varphi(z)]+m\varphi'(z)\dot{z}-\dot{d}_{2,\omega}. ξ ˙ = − λ ξ − λ Δ ω − m ˙ [ φ ( z + ρ ω ) − φ ( z )] + m φ ′ ( z ) z ˙ − d ˙ 2 , ω .
Inserting (72) , use the exact identities
− m ˙ a ε g ( x ) sin ( θ − π / 4 ) = λ g ( x ) [ 1 − sin ( 2 θ ) ] , -\dot{m}\,a\varepsilon g(x)\sin(\theta-\pi/4)=\lambda g(x)[1-\sin(2\theta)], − m ˙ a ε g ( x ) sin ( θ − π /4 ) = λ g ( x ) [ 1 − sin ( 2 θ )] ,
d ˙ 2 , ω = 1 2 λ χ ε 2 φ ′ ′ ( z ) z ˙ cos ( 2 θ ) − λ g ( x ) sin ( 2 θ ) . \dot{d}_{2,\omega}=\frac{1}{2}\lambda\chi\varepsilon^2\varphi''(z)\dot{z}\cos(2\theta)-\lambda g(x)\sin(2\theta). d ˙ 2 , ω = 2 1 λ χ ε 2 φ ′′ ( z ) z ˙ cos ( 2 θ ) − λ g ( x ) sin ( 2 θ ) .
The first identity follows from 2 cos ( θ + π / 4 ) sin ( θ − π / 4 ) = sin ( 2 θ ) − 1 2\cos(\theta+\pi/4)\sin(\theta-\pi/4)=\sin(2\theta)-1 2 cos ( θ + π /4 ) sin ( θ − π /4 ) = sin ( 2 θ ) − 1 . The doubled-frequency terms cancel, leaving
ξ ˙ = − λ ξ + λ g ( x ) + d ξ (74) \dot{\xi}=-\lambda\xi+\lambda g(x)+d_\xi \tag*{(74)} ξ ˙ = − λ ξ + λ g ( x ) + d ξ (74)
where
d ξ = − λ Δ ω − m ˙ E φ + [ m φ ′ ( z ) − 1 2 λ χ ε 2 φ ′ ′ ( z ) cos ( 2 θ ) ] z ˙ . d_\xi=-\lambda\Delta_\omega-\dot{m}E_\varphi+\left[m\varphi'(z)-\frac{1}{2}\lambda\chi\varepsilon^2\varphi''(z)\cos(2\theta)\right]\dot{z}. d ξ = − λ Δ ω − m ˙ E φ + [ m φ ′ ( z ) − 2 1 λ χ ε 2 φ ′′ ( z ) cos ( 2 θ ) ] z ˙ .
Thus (41) and (73) imply
∣ d ξ ∣ ≤ K ξ [ a ε + ε 2 + ( ε / a + ε 2 ) ∣ z ˙ ∣ ] , (75) |d_\xi|\le K_\xi[a\varepsilon+\varepsilon^2+(\varepsilon/a+\varepsilon^2)|\dot{z}|], \tag*{(75)} ∣ d ξ ∣ ≤ K ξ [ a ε + ε 2 + ( ε / a + ε 2 ) ∣ z ˙ ∣ ] , (75)
where
K ξ = max { λ Φ 2 L ρ 2 , λ K Δ + 2 λ Φ 1 L e , 2 λ Φ 1 , 1 2 λ ∣ χ ∣ Φ 2 } . (76) K_\xi=\max\left\{\lambda\Phi_2L_\rho^2,\lambda K_\Delta+2\lambda\Phi_1L_e,2\lambda\Phi_1,\frac{1}{2}\lambda|\chi|\Phi_2\right\}. \tag*{(76)} K ξ = max { λ Φ 2 L ρ 2 , λ K Δ + 2 λ Φ 1 L e , 2 λ Φ 1 , 2 1 λ ∣ χ ∣ Φ 2 } . (76)
More explicitly, the four coefficients in (76) arise from
∣ d ξ ∣ ≤ λ Φ 2 L ρ 2 a ε + ( λ K Δ + 2 λ Φ 1 L e ) ε 2 + ( 2 λ Φ 1 ε a + 1 2 λ ∣ χ ∣ Φ 2 ε 2 ) ∣ z ˙ ∣ . |d_\xi|\le\lambda\Phi_2L_\rho^2a\varepsilon+(\lambda K_\Delta+2\lambda\Phi_1L_e)\varepsilon^2+\left(2\lambda\Phi_1\frac{\varepsilon}{a}+\frac{1}{2}\lambda|\chi|\Phi_2\varepsilon^2\right)|\dot{z}|. ∣ d ξ ∣ ≤ λ Φ 2 L ρ 2 a ε + ( λ K Δ + 2 λ Φ 1 L e ) ε 2 + ( 2 λ Φ 1 a ε + 2 1 λ ∣ χ ∣ Φ 2 ε 2 ) ∣ z ˙ ∣.
The factor m ˙ \dot{m} m ˙ , whose magnitude grows with frequency, turns the quadratic cost remainder into the term proportional to a ε a\varepsilon a ε . The coefficient of ∣ z ˙ ∣ |\dot{z}| ∣ z ˙ ∣ decreases with frequency at fixed a a a . Its effect is controlled by the weighted integral in (64) ; no bounded time derivative at startup is required. Only C 2 C^2 C 2 cost regularity is used. The other residuals satisfy
∥ d w ∥ ∞ ≤ L F L r a ε , \|d_w\|_\infty\le L_F L_r a\varepsilon, ∥ d w ∥ ∞ ≤ L F L r a ε ,
∣ d u ∣ ≤ k P K Δ ε 2 , |d_u|\le k_P K_\Delta\varepsilon^2, ∣ d u ∣ ≤ k P K Δ ε 2 ,
∣ d v ∣ ≤ k I K Δ ε 2 . (77) |d_v|\le k_IK_\Delta\varepsilon^2. \tag*{(77)} ∣ d v ∣ ≤ k I K Δ ε 2 . (77)
The full L F L_F L F is required because d w d_w d w includes V 0 r ω V_0r_\omega V 0 r ω , even when this linear reaction contribution supplies strong damping to the averaged system. The smaller derivative variation ℓ ( R ) \ell(R) ℓ ( R ) cannot replace it.
A bound independent of the comparison interval
In this section, we show that the comparison error bound is given by
R ω = G B σ α k P K Δ ε 2 + M α ( L F L r + K ξ ) a ε + M α ( k I K Δ + K ξ ) ε 2 + M K ξ ( ε / a + ε 2 ) L z . (78) \mathcal{R}_\omega=G_B\frac{\sigma}{\alpha}k_PK_\Delta\varepsilon^2+\frac{M}{\alpha}(L_FL_r+K_\xi)a\varepsilon+\frac{M}{\alpha}(k_IK_\Delta+K_\xi)\varepsilon^2+MK_\xi(\varepsilon/a+\varepsilon^2)L_z. \tag*{(78)} R ω = G B α σ k P K Δ ε 2 + α M ( L F L r + K ξ ) a ε + α M ( k I K Δ + K ξ ) ε 2 + M K ξ ( ε / a + ε 2 ) L z . (78)
We note that every constant on the right is fixed independently of the comparison interval. For fixed a > 0 a>0 a > 0 , R ω = O ( ω − 1 / 2 ) \mathcal{R}_\omega=O(\omega^{-1/2}) R ω = O ( ω − 1/2 ) .
Lemma 7. Under Assumptions 1–4 and (8), fix 0 < a ≤ a 0 0<a\le a_0 0 < a ≤ a 0 and impose (5). Suppose ω ≥ ω l i f t \omega\ge\omega_{\mathrm{lift}} ω ≥ ω lift , L r a ε ≤ η ⋆ L_ra\varepsilon\le\eta_\star L r a ε ≤ η ⋆ , and (63) hold. For transformed nominal and averaged solutions remaining in B R ( Z ⋆ ) B_R(Z^\star) B R ( Z ⋆ ) throughout [ t 0 , t ] [t_0,t] [ t 0 , t ] ,
∥ Z ω ( t ) − Z ˉ ( t ) ∥ ε ≤ M e − α ( t − t 0 ) ∥ Z ω ( t 0 ) − Z ˉ ( t 0 ) ∥ ε + R ω (79) \|Z_\omega(t)-\bar{Z}(t)\|_\varepsilon\le Me^{-\alpha(t-t_0)}\|Z_\omega(t_0)-\bar{Z}(t_0)\|_\varepsilon+\mathcal{R}_\omega \tag*{(79)} ∥ Z ω ( t ) − Z ˉ ( t ) ∥ ε ≤ M e − α ( t − t 0 ) ∥ Z ω ( t 0 ) − Z ˉ ( t 0 ) ∥ ε + R ω (79) With matched initial states, the following bounds hold
∥ Z ω ( t ) − Z ˉ ( t ) ∥ ε ≤ R ω , \|Z_\omega(t)-\bar{Z}(t)\|_\varepsilon\le\mathcal{R}_\omega, ∥ Z ω ( t ) − Z ˉ ( t ) ∥ ε ≤ R ω , ∥ X ω ( t ) − Z ˉ ( t ) ∥ ε ≤ R ω + δ ω . (80) \|X_\omega(t)-\bar{Z}(t)\|_\varepsilon\le\mathcal{R}_\omega+\delta_\omega. \tag*{(80)} ∥ X ω ( t ) − Z ˉ ( t ) ∥ ε ≤ R ω + δ ω . (80) The bounds are independent of the interval length.
Proof. The exact residual equations (71) and (74) meet the hypotheses of Lemma 1. The field and boundary residuals are continuous, and the filter residual is locally integrable by Lemma 6. The boundary contribution in (33) is at most G B σ k P K Δ ε 2 / α G_B\sigma k_PK_\Delta\varepsilon^2/\alpha G B σ k P K Δ ε 2 / α . For the internal terms, (75) and (77) give
M ∫ t 0 t e − α ( t − s ) [ ∥ d w ( s ) ∥ ∞ + ∣ d v ( s ) ∣ + ∣ d ξ ( s ) ∣ ] d s ≤ M α [ ( L F L r + K ξ ) a ε + ( k I K Δ + K ξ ) ε 2 ] + M K ξ ( ε / a + ε 2 ) ∫ t 0 t e − α ( t − s ) ∣ z ˙ ( s ) ∣ d s . \begin{aligned}
M\int_{t_0}^{t}e^{-\alpha(t-s)}[\|d_w(s)\|_\infty+|d_v(s)|+|d_\xi(s)|]\,\mathrm{d}s \\
&\le\frac{M}{\alpha}\left[(L_FL_r+K_\xi)a\varepsilon+(k_IK_\Delta+K_\xi)\varepsilon^2\right] \\
&+MK_\xi(\varepsilon/a+\varepsilon^2)\int_{t_0}^{t}e^{-\alpha(t-s)}|\dot{z}(s)|\,\mathrm{d}s.
\end{aligned} M ∫ t 0 t e − α ( t − s ) [ ∥ d w ( s ) ∥ ∞ + ∣ d v ( s ) ∣ + ∣ d ξ ( s ) ∣ ] d s ≤ α M [ ( L F L r + K ξ ) a ε + ( k I K Δ + K ξ ) ε 2 ] + M K ξ ( ε / a + ε 2 ) ∫ t 0 t e − α ( t − s ) ∣ z ˙ ( s ) ∣ d s .
The last integral is at most L z L_z L z by (64) . Adding the boundary contribution yields (78) and proves (79) . Matched initialization removes the initial-error term, and (43) gives (80) .
This is a conditional comparison on the region where the estimates apply. The next section combines it with averaged decay to show that small initial errors and sufficiently high frequencies keep the implemented solution in that region for all time.
Practical Stability
Table 1 collects the constants that enter the final estimates. By the developments above, the averaged decay and the comparison bound now exclude departure from the region where the estimates were obtained. With (42) and (78) , we let
β ω = R ω + ( M + 1 ) δ ω , \beta_\omega=\mathcal{R}_\omega+(M+1)\delta_\omega, β ω = R ω + ( M + 1 ) δ ω ,
r ⋆ = R / ( 8 M ) . (81) r_\star=R/(8M). \tag*{(81)} r ⋆ = R / ( 8 M ) . (81)
The term δ ω \delta_\omega δ ω bounds the coordinate change, while R ω \mathcal{R}_\omega R ω bounds the error dynamics. The additional M δ ω M\delta_\omega M δ ω in β ω \beta_\omega β ω accounts for the change in the initial condition when applying the averaged decay estimate. This initial contribution decays, which explains the smaller ultimate radius below. We impose the sufficient frequency conditions
ω ≥ ω l i f t , a ω ≥ 4 k P λ Φ 1 K 0 , \omega\ge\omega_{\mathrm{lift}}, \qquad a\sqrt{\omega} \ge4k_P\lambda\Phi_1K_0, ω ≥ ω lift , a ω ≥ 4 k P λ Φ 1 K 0 ,
L r a / ω ≤ η ⋆ , β ω ≤ R / 8. (82) L_ra/\sqrt{\omega} \le\eta_\star, \qquad\beta_\omega\le R/8. \tag*{(82)} L r a / ω ≤ η ⋆ , β ω ≤ R /8. (82)
For each fixed a > 0 a>0 a > 0 , they hold for all frequencies above a finite threshold, since δ ω , R ω → 0 \delta_\omega,\mathcal{R}_\omega\to0 δ ω , R ω → 0 .
Theorem 1. Suppose Assumptions 1–4 and (8) hold. Fix a ∈ ( 0 , a 0 ] a\in(0,a_0] a ∈ ( 0 , a 0 ] and the initial-family bound H 0 H_0 H 0 . Choose ω \omega ω satisfying (82). For every t 0 ≥ 0 t_0\ge0 t 0 ≥ 0 , every dither phase, and every initial state satisfying (5) and E ω ( t 0 ) ≤ r ⋆ E_\omega(t_0)\le r_\star E ω ( t 0 ) ≤ r ⋆ , the implemented loop has a unique global mild solution. Initialize the averaged comparison by
Z ˉ ( t 0 ) = Z ω ( t 0 ) (83) \bar{Z}(t_0)=Z_\omega(t_0) \tag*{(83)} Z ˉ ( t 0 ) = Z ω ( t 0 ) (83) using (40). Then, for all t ≥ t 0 t\ge t_0 t ≥ t 0 ,
∥ Z ω ( t ) − Z ˉ ( t ) ∥ E ≤ R ω , ∥ X ω ( t ) − Z ˉ ( t ) ∥ E ≤ R ω + δ ω , E ω ( t ) ≤ M e − α ( t − t 0 ) E ω ( t 0 ) + β ω . (84) \begin{aligned}
\lVert Z_\omega(t)-\bar{Z}(t)\rVert_E &\le\mathcal{R}_\omega,\\
\lVert X_\omega(t)-\bar{Z}(t)\rVert_E &\le\mathcal{R}_\omega+\delta_\omega,\\
E_\omega(t) &\le M e^{-\alpha(t-t_0)}E_\omega(t_0)+\beta_\omega. \tag*{(84)}
\end{aligned} ∥ Z ω ( t ) − Z ˉ ( t ) ∥ E ∥ X ω ( t ) − Z ˉ ( t ) ∥ E E ω ( t ) ≤ R ω , ≤ R ω + δ ω , ≤ M e − α ( t − t 0 ) E ω ( t 0 ) + β ω . (84) Moreover,
lim sup t → ∞ E ω ( t ) ≤ R ω + δ ω , \limsup_{t\to\infty} E_\omega(t)\le\mathcal{R}_\omega+\delta_\omega, t → ∞ lim sup E ω ( t ) ≤ R ω + δ ω , lim sup t → ∞ ∣ J ( t ) − ϕ ( z ⋆ ) ∣ ≤ 1 2 Φ 2 ( R ω + L r a ε ) 2 . (85) \limsup_{t\to\infty}\lvert J(t)-\phi(z^\star)\rvert\le\frac{1}{2}\Phi_2(\mathcal{R}_\omega+L_ra\varepsilon)^2. \tag*{(85)} t → ∞ lim sup ∣ J ( t ) − ϕ ( z ⋆ )∣ ≤ 2 1 Φ 2 ( R ω + L r a ε ) 2 . (85) The state and cost bounds have orders ω − 1 / 2 \omega^{-1/2} ω − 1/2 and ω − 1 \omega^{-1} ω − 1 at fixed a a a . They are uniform in initial time, phase, admitted initial state, and frequency above the selected threshold.
Proof. For fixed a , ω a,\omega a , ω , the original mild field equation and the controller integral equations are locally contractive: K B \mathcal{K}_B K B is locally integrable and all state nonlinearities are locally Lipschitz. This gives local existence without differentiating J J J .
First justify the initial coordinate bound. Condition (82) implies δ ω ≤ R / [ 8 ( M + 1 ) ] \delta_\omega\le R/[8(M+1)] δ ω ≤ R / [ 8 ( M + 1 )] and R ω ≤ R / 8 \mathcal{R}_\omega\le R/8 R ω ≤ R /8 , so
∥ x ( t 0 ) − w ⋆ ∥ ∞ ≤ r ⋆ + L r a ε ≤ r ⋆ + δ ω < R . \lVert x(t_0)-w^\star\rVert_\infty\le r_\star+L_ra\varepsilon\le r_\star+\delta_\omega<R. ∥ x ( t 0 ) − w ⋆ ∥ ∞ ≤ r ⋆ + L r a ε ≤ r ⋆ + δ ω < R . The lift stays within the enlarged regularity region. Applying Lemma 3 therefore gives
M ∥ Z ω ( t 0 ) − Z ⋆ ∥ E ≤ M ( r ⋆ + δ ω ) < R / 4. (86) M\lVert Z_\omega(t_0)-Z^\star\rVert_E\le M(r_\star+\delta_\omega)<R/4. \tag*{(86)} M ∥ Z ω ( t 0 ) − Z ⋆ ∥ E ≤ M ( r ⋆ + δ ω ) < R /4. (86) By (86), Corollary 1 applies to the matched initial state (83). Its decay estimate (34) keeps Z ˉ \bar{Z} Z ˉ in B R / 4 ( Z ⋆ ) B_{R/4}(Z^\star) B R /4 ( Z ⋆ ) . Before any nominal first exit from B R ( Z ⋆ ) B_R(Z^\star) B R ( Z ⋆ ) , (80) gives
∥ Z ω ( t ) − Z ⋆ ∥ E < R / 4 + R ω ≤ 3 R / 8. \lVert Z_\omega(t)-Z^\star\rVert_E<R/4+\mathcal{R}_\omega\le3R/8. ∥ Z ω ( t ) − Z ⋆ ∥ E < R /4 + R ω ≤ 3 R /8. By continuity, any such exit is excluded.
The inverse transformation bounds w w w , v v v , g ^ \hat{g} g ^ and p = g ^ − m φ ( C b w ) p=\hat{g}-m\varphi(C_bw) p = g ^ − m φ ( C b w ) on every finite interval. Since 3 R / 8 + η ⋆ < R + η ⋆ 3R/8+\eta_\star<R+\eta_\star 3 R /8 + η ⋆ < R + η ⋆ , the physical field remains a positive distance inside the regularity region. At a finite maximal time, the bounded internal perturbation and the strong semigroup continuity yield an endpoint limit of the distributed convolution. The boundary convolution has an endpoint limit by local integrability of K B \mathcal{K}_B K B and boundedness of its command. The controller states also have limits. These facts allow one to restart the local mild problem while excluding finite maximal existence. Boundedness of the cost is used here only for continuation, not in the error constants.
The global comparison, averaged decay, and coordinate estimate now imply
E ω ( t ) ≤ M e − α ( t − t 0 ) [ E ω ( t 0 ) + δ ω ] + R ω + δ ω ≤ M e − α ( t − t 0 ) E ω ( t 0 ) + β ω . (87) \begin{aligned}
E_\omega(t)&\le M e^{-\alpha(t-t_0)}[E_\omega(t_0)+\delta_\omega]+\mathcal{R}_\omega+\delta_\omega\\
&\le M e^{-\alpha(t-t_0)}E_\omega(t_0)+\beta_\omega. \tag*{(87)}
\end{aligned} E ω ( t ) ≤ M e − α ( t − t 0 ) [ E ω ( t 0 ) + δ ω ] + R ω + δ ω ≤ M e − α ( t − t 0 ) E ω ( t 0 ) + β ω . (87) Estimate (87) proves (84) and the state part of (85). For the field alone, lim sup ∥ w − w ⋆ ∥ ∞ ≤ R ω + L r a ε \limsup\lVert w-w^\star\rVert_\infty\le\mathcal{R}_\omega+L_ra\varepsilon lim sup ∥ w − w ⋆ ∥ ∞ ≤ R ω + L r a ε . Since φ ′ ( z ⋆ ) = 0 \varphi'(z^\star)=0 φ ′ ( z ⋆ ) = 0 and ∥ C b ∥ = 1 \lVert C_b\rVert=1 ∥ C b ∥ = 1 , Taylor’s formula bounds the cost gap by Φ 2 ∥ w − w ⋆ ∥ ∞ 2 / 2 \Phi_2\lVert w-w^\star\rVert_\infty^2/2 Φ 2 ∥ w − w ⋆ ∥ ∞ 2 /2 , proving the remaining assertion. All estimates use elapsed time and phase-uniform bounds. Their constants are fixed over the prescribed initial family for each chosen a > 0 a>0 a > 0 . □ \square □
Corollary 2. Under Theorem 1, fix η > 0 \eta>0 η > 0 and strengthen (82) to β ω ≤ min { R / 8 , η / 2 } \beta_\omega\le\min\{R/8,\eta/2\} β ω ≤ min { R /8 , η /2 } . Every admitted state with initial error at most min { r ⋆ , η / ( 2 M ) } \min\{r_\star,\eta/(2M)\} min { r ⋆ , η / ( 2 M )} satisfies E ω ( t ) ≤ η E_\omega(t)\le\eta E ω ( t ) ≤ η for all t ≥ t 0 t\ge t_0 t ≥ t 0 . Every admitted state with error at most r ⋆ r_\star r ⋆ satisfies that accuracy after
t − t 0 ≥ T η : = α − 1 max { 0 , log ( 2 M r ⋆ / η ) } . (88) t-t_0\ge T_\eta:=\alpha^{-1}\max\{0,\log(2M r_\star/\eta)\}. \tag*{(88)} t − t 0 ≥ T η := α − 1 max { 0 , log ( 2 M r ⋆ / η )} . (88) Proof. For the smaller initial radius, or after the delay in (88), the decaying term in (84) is at most η / 2 \eta/2 η /2 . Adding β ω ≤ η / 2 \beta_\omega\le\eta/2 β ω ≤ η /2 proves both assertions. □ \square □
The matched averaged initial state depends on frequency and phase. The result gives practical regulation with the dither active. Invariance of the ultimate ball and existence or uniqueness of a periodic orbit are not asserted. In the thermal setting, Theorem 1 says that a plate started near its optimal temperature profile, with the filter initialized by (16), stays near that profile for all time, that the element temperature settles to within a band of width ω − 1 / 2 \omega^{-1/2} ω − 1/2 around its optimum, and that the residual loss of transmission is of order ω − 1 \omega^{-1} ω − 1 .
Frequency Selection and Scope By (81), we obtain
β ω = Γ 1 ( a ) ω + Γ 2 ω , (89) \beta_\omega=\frac{\Gamma_1(a)}{\sqrt{\omega}}+\frac{\Gamma_2}{\omega}, \tag*{(89)} β ω = ω Γ 1 ( a ) + ω Γ 2 , (89) where
Γ 1 ( a ) = a [ M α ( L F L r + K ξ ) + ( M + 1 ) L r ] + M K ξ L Z a , (90) \Gamma_1(a)=a\left[\frac{M}{\alpha}(L_FL_r+K_\xi)+(M+1)L_r\right]+\frac{MK_\xi L_Z}{a}, \tag*{(90)} Γ 1 ( a ) = a [ α M ( L F L r + K ξ ) + ( M + 1 ) L r ] + a M K ξ L Z , (90) Γ 2 = G B σ α k P K Δ + M α ( k P K Δ + K ξ ) + M K ξ L Z + ( M + 1 ) K Δ . \Gamma_2=G_B\frac{\sigma}{\alpha}k_PK_\Delta+\frac{M}{\alpha}(k_PK_\Delta+K_\xi)+MK_\xi L_Z+(M+1)K_\Delta. Γ 2 = G B α σ k P K Δ + α M ( k P K Δ + K ξ ) + M K ξ L Z + ( M + 1 ) K Δ .
For 0 < s ≤ R / 8 0 < s \le R/8 0 < s ≤ R /8 , solving (89) and imposing the remaining conditions of (82) gives
ω ≥ max { ω l i f t , 4 k p λ Φ 1 K 0 a , L r a η ⋆ , Γ 1 ( a ) + Γ 1 ( a ) 2 + 4 Γ 2 s 2 s } . (91) \sqrt{\omega} \ge\max\left\{\sqrt{\omega_{\mathrm{lift}}},\frac{4k_{\mathrm{p}}\lambda\Phi_{1}K_{0}}{a},\frac{L_{r}a}{\eta_{\star}},\frac{\Gamma_{1}(a)+\sqrt{\Gamma_{1}(a)^{2}+4\Gamma_{2}s}}{2s}\right\}. \tag*{(91)} ω ≥ max { ω lift , a 4 k p λ Φ 1 K 0 , η ⋆ L r a , 2 s Γ 1 ( a ) + Γ 1 ( a ) 2 + 4 Γ 2 s } . (91)
Then β ω ≤ s \beta_{\omega} \le s β ω ≤ s . We use s = R / 8 s=R/8 s = R /8 for Theorem 1 or s = min { R / 8 , η / 2 } s=\min\{R/8,\eta/2\} s = min { R /8 , η /2 } for Corollary 2. The constants in (90) use fixed a 0 a_{0} a 0 , leaving terms proportional to a a a and 1 / a 1/a 1/ a . The terms proportional to a a a come from the lifted field and from the reaction evaluated on it. The term proportional to 1 / a 1/a 1/ a comes from the demodulation gain, which grows as the dither amplitude shrinks. Smaller dither amplitudes do not necessarily improve the bound. The controller does not evaluate (91), but certification requires plant and cost-derivative bounds.
For F ( w ) = − K w − β w 3 F(w)=-Kw-\beta w^{3} F ( w ) = − K w − β w 3 , d = 0 d=0 d = 0 , and φ ( z ) = z 2 / 2 \varphi(z)=z^{2}/2 φ ( z ) = z 2 /2 , with K , β > 0 K,\beta>0 K , β > 0 , the optimum is w ⋆ = u ⋆ = 0 w^{\star}=u^{\star}=0 w ⋆ = u ⋆ = 0 . Choosing V 0 = − K V_{0}=-K V 0 = − K gives μ r e f = μ 0 + K \mu_{\mathrm{ref}}=\mu_{0}+K μ ref = μ 0 + K and ℓ ( R ) = 3 β R 2 \ell(R)=3\beta R^{2} ℓ ( R ) = 3 β R 2 . Thus R ≤ σ / ( 6 M β ) R\le\sqrt{\sigma/(6M\beta)} R ≤ σ / ( 6 M β ) ensures α ≥ σ / 2 \alpha\ge\sigma/2 α ≥ σ /2 for every fixed K K K . The full bound L F = K + 3 β ( R + η ⋆ ) 2 L_{F}=K+3\beta(R+\eta_{\star})^{2} L F = K + 3 β ( R + η ⋆ ) 2 still enters the approximation error, so absorbing damping need not reduce the sufficient frequency. Transverse slope variation remains subject to Assumption 4.
The universal π / 4 \pi/4 π /4 phase is a property of the heated edge. Moving the sensor into the plate adds a frequency-dependent delay and attenuation, which is why the design assumes a sensor on the actuated boundary. For the linear diffusion plant, an interior midpoint sensor at ( d s , 1 / 2 ) (d_{s},1/2) ( d s , 1/2 ) , 0 < d s < 1 0<d_{s}<1 0 < d s < 1 , has
H d s ( i ω ) H b ( i ω ) = cosh [ γ 0 ( 1 − d s ) ] cosh γ 0 = exp [ − ( 1 + i ) d s ω 2 κ ] [ 1 + o ( 1 ) ] . \frac{H_{d_{s}}(i\omega)}{H_{b}(i\omega)}
=
\frac{\cosh[\gamma_{0}(1-d_{s})]}{\cosh\gamma_{0}}
=
\exp\left[-(1+i)d_{s}\sqrt{\frac{\omega}{2\kappa}}\right][1+o(1)]. H b ( iω ) H d s ( iω ) = cosh γ 0 cosh [ γ 0 ( 1 − d s )] = exp [ − ( 1 + i ) d s 2 κ ω ] [ 1 + o ( 1 )] .
Hence the boundary phase in (7), after separating the signed gain, cannot be used unchanged for interior sensing. Both its extra attenuation and phase depend on frequency.
For a prescribed physical cost trace, bounded measurement noise n n n and zero initial perturbation of p p p produce
q n ( t ) = m ( t ) n ( t ) − ∫ t 0 t e − λ ( t − s ) [ λ m ( s ) + m ˙ ( s ) ] n ( s ) d s , q_{n}(t)=m(t)n(t)-\int_{t_{0}}^{t}e^{-\lambda(t-s)}[\lambda m(s)+\dot{m}(s)]n(s)\,\mathrm{d}s, q n ( t ) = m ( t ) n ( t ) − ∫ t 0 t e − λ ( t − s ) [ λm ( s ) + m ˙ ( s )] n ( s ) d s ,
and therefore, almost everywhere,
∣ q n ( t ) ∣ ≤ ( 4 λ ε a + 2 a ε ) ∥ n ∥ ∞ , [ t 0 , t ] . (92) |q_{n}(t)|\le\left(\frac{4\lambda\varepsilon}{a}+\frac{2}{a\varepsilon}\right)\|n\|_{\infty,[t_{0},t]}. \tag*{(92)} ∣ q n ( t ) ∣ ≤ ( a 4 λ ε + a ε 2 ) ∥ n ∥ ∞ , [ t 0 , t ] . (92)
The growing term in (92) concerns noise near the dither frequency. For n = N sin ( θ + ψ ) n=N\sin(\theta+\psi) n = N sin ( θ + ψ ) , the steady filter mean is N ω sin ( π / 4 − ψ ) / a N\sqrt{\omega}\sin(\pi/4-\psi)/a N ω sin ( π /4 − ψ ) / a . Constant offsets instead cancel under consistent initialization. These are prescribed-signal filter calculations, not noisy closed-loop guarantees. The first-order cost ripple also vanishes at the optimum because φ ′ ( z ⋆ ) = 0 \varphi'(z^{\star})=0 φ ′ ( z ⋆ ) = 0 .
The theorem concerns local regulation to an attainable output-cost minimum. Global optimization, input saturation, measurement delays, and additional actuator or sensor dynamics are outside its scope.
Thermal Resonance Alignment
We now return to the plate of Fig. 1(a) and give it numbers. Thermally tunable filters motivate the local optical model [1 , 2 ] .
Thermal model and scaling
Let T ( X , Y , t p h y s ) T(X,Y,t_{\mathrm{phys}}) T ( X , Y , t phys ) be the temperature of a plate occupying ( X , Y ) ∈ ( 0 , L ) 2 (X,Y)\in(0,L)^{2} ( X , Y ) ∈ ( 0 , L ) 2 . The thin-plate approximation treats temperature as uniform through its thickness d p d_{p} d p . In-plane conduction and linear cooling through the two broad faces then give
C v d p ∂ T ∂ t p h y s = k t h d p Δ X , Y T − 2 h f ( T − T a ) , C_{v}d_{p}\frac{\partial T}{\partial t_{\mathrm{phys}}}=k_{\mathrm{th}}d_{p}\Delta_{X,Y}T-2h_{f}(T-T_{a}), C v d p ∂ t phys ∂ T = k th d p Δ X , Y T − 2 h f ( T − T a ) ,
− k t h d p T X ( 0 , Y , t p h y s ) = P ( t p h y s ) L q ( Y / L ) , -k_{\mathrm{th}}d_{p}T_{X}(0,Y,t_{\mathrm{phys}})=\frac{P(t_{\mathrm{phys}})}{L}q(Y/L), − k th d p T X ( 0 , Y , t phys ) = L P ( t phys ) q ( Y / L ) ,
T X ( L , Y , t p h y s ) = 0 , T_{X}(L,Y,t_{\mathrm{phys}})=0, T X ( L , Y , t phys ) = 0 ,
T Y ( X , 0 , t p h y s ) = T Y ( X , L , t p h y s ) = 0. (93) T_{Y}(X,0,t_{\mathrm{phys}})=T_{Y}(X,L,t_{\mathrm{phys}})=0. \tag*{(93)} T Y ( X , 0 , t phys ) = T Y ( X , L , t phys ) = 0. (93)
Here C v C_{v} C v is volumetric heat capacity, k t h k_{\mathrm{th}} k th is thermal conductivity, and h f h_{f} h f is the heat-transfer coefficient of each broad face. The factor two accounts for cooling from both faces. The ambient temperature is T a T_{a} T a , and the three unactuated lateral edges are insulated.
The scalar command P P P is distributed along the heated edge by q ( s 2 ) = 1 + ϑ cos ( π s 2 ) q(s_{2})=1+\vartheta\cos(\pi s_{2}) q ( s 2 ) = 1 + ϑ cos ( π s 2 ) . The normalization ∫ 0 1 q ( s 2 ) d s 2 = 1 \int_{0}^{1}q(s_{2})\,\mathrm{d}s_{2}=1 ∫ 0 1 q ( s 2 ) d s 2 = 1 ensures that integrating the imposed edge flux in (93) gives the total heater power P P P . The nonuniform profile produces a spatially varying temperature along the heated edge, although the optical element measures performance at its midpoint.
Table II lists the illustrative physical and optical parameters used in the reported computation. The optical half-width provides the temperature scale Θ \Theta Θ .
Table II. Physical and optical parameters of the thermal example.
Introduce s = ( X / L , Y / L ) s=(X/L,Y/L) s = ( X / L , Y / L ) and normalize temperature, time, and power as follows
w = ( T − T a ) / Θ , t = t p h y s / t D , w=(T-T_{a})/\Theta,\qquad t=t_{\mathrm{phys}}/t_{D}, w = ( T − T a ) /Θ , t = t phys / t D ,
t D = C v L 2 / k t h , u = P / P s , t_{D}=C_{v}L^{2}/k_{\mathrm{th}},\qquad u=P/P_{s}, t D = C v L 2 / k th , u = P / P s ,
P s = k t h Θ d p . (94) P_{s}=k_{\mathrm{th}}\Theta d_{p}. \tag*{(94)} P s = k th Θ d p . (94)
The diffusion time t D t_{D} t D makes the normalized conduction coefficient unity. The power scale P s P_{s} P s similarly makes the boundary input gain unity, so the edge condition becomes − w s 1 = q u -w_{s1}=q\,u − w s 1 = q u . Substituting (94) into (93) gives (1) with
κ = b = 1 , μ 0 = 2 h f L 2 k t h d p = 1 , F = d = 0. \kappa=b=1,\qquad\mu_0=\frac{2h_fL^2}{k_{\mathrm{th}}d_p}=1,\qquad F=d=0. κ = b = 1 , μ 0 = k th d p 2 h f L 2 = 1 , F = d = 0.
Numerically, t D = 4.5714 s t_D=4.5714\,\mathrm{s} t D = 4.5714 s and P s = 0.14 m W P_s=0.14\,\mathrm{mW} P s = 0.14 mW . Thus a dimensionless input specifies total heater power, and the physical midpoint temperature is T b = T a + Θ C b w T_b=T_a+\Theta C_bw T b = T a + Θ C b w . These relations connect the PDE variables and the physical axes in the performance plots.
Optical score and numerical procedure
Let P i n P_{\mathrm{in}} P in be fixed incident optical power and P t r P_{\mathrm{tr}} P tr the transmitted power. The local Lorentzian model is
P t r P i n = A o p t 1 + ( C b w − z ⋆ ) 2 . \frac{P_{\mathrm{tr}}}{P_{\mathrm{in}}}=\frac{A_{\mathrm{opt}}}{1+(C_bw-z^\star)^2}. P in P tr = 1 + ( C b w − z ⋆ ) 2 A opt .
The temperature half-width equals Θ \Theta Θ , so the detuning C b w − z ⋆ C_bw-z^\star C b w − z ⋆ is dimensionless. The reciprocal-power score plotted in Figure 1 (b) is
J = P i n P t r − 1 = 1 A o p t − 1 + ( C b w − z ⋆ ) 2 A o p t , (95) J=\frac{P_{\mathrm{in}}}{P_{\mathrm{tr}}}-1=\frac{1}{A_{\mathrm{opt}}}-1+\frac{(C_bw-z^\star)^2}{A_{\mathrm{opt}}}, \tag*{(95)} J = P tr P in − 1 = A opt 1 − 1 + A opt ( C b w − z ⋆ ) 2 , (95)
with A o p t = 0.98 A_{\mathrm{opt}}=0.98 A opt = 0.98 . The chosen optimum z ⋆ = 8 z^\star=8 z ⋆ = 8 corresponds to T b ⋆ = 308 K T_b^\star=308\,\mathrm{K} T b ⋆ = 308 K . Only J J J enters the feedback. Neither the alignment temperature nor (95) is evaluated by the controller. The stationary midpoint gain is G 0 = coth ( 1 ) G_0=\coth(1) G 0 = coth ( 1 ) , so u ⋆ = z ⋆ / G 0 u^\star=z^\star/G_0 u ⋆ = z ⋆ / G 0 is attainable. The stationary field w ⋆ w^\star w ⋆ is spatially nonuniform because heat is applied at one edge and lost through both faces (Figure 3 ). Regulation therefore concerns this attainable field and its required input bias. Since φ ′ ′ = 2 / A o p t \varphi^{\prime\prime}=2/A_{\mathrm{opt}} φ ′′ = 2/ A opt , V 0 = 0 V_0=0 V 0 = 0 gives ℓ ( R ) = 0 \ell(R)=0 ℓ ( R ) = 0 , and the averaged error dynamics equal the linear system of Proposition 1.
Figure 3. Stationary optimal temperature field T ⋆ T^\star T ⋆ over the plate. The red line marks the heated edge, along which the nonuniform heater profile q q q shapes the temperature, and the circle marks the optical element, held at 308 K.
Use (9)–(10) with λ = 2 \lambda=2 λ = 2 , k P = 0.35 k_P=0.35 k P = 0.35 , k I = 0.10 k_I=0.10 k I = 0.10 , a = 0.30 a=0.30 a = 0.30 , and ω = 80 \omega=80 ω = 80 (2.7852 Hz). Set t 0 = θ 0 = 0 t_0=\theta_0=0 t 0 = θ 0 = 0 . The initial field is the discrete stationary solution with target C b w = 8 C_bw=8 C b w = 8 , so v 0 v_0 v 0 is its corresponding bias, and ρ 0 = − m ( 0 ) J ( 0 ) \rho_0=-m(0)J(0) ρ 0 = − m ( 0 ) J ( 0 ) as in (16), giving g ^ ( 0 ) = 0 \hat{g}(0)=0 g ^ ( 0 ) = 0 . The target uses (83). The horizon 0 ≤ t ≤ 50 0\leq t\leq50 0 ≤ t ≤ 50 is 228.6 s.
The Python/SciPy computation uses continuous P1 elements (2,009 nodes, 3,840 triangles graded toward the heated edge)
Figure 4. Error T − T ∗ T-T^* T − T ∗ between the implemented and the optimal temperature fields at physical times 0, 9.14, and 228.57 s (left to right); the maximum absolute error falls from 221 to 68.5 to 15.1 mK. The red line marks the heated edge and the circle the optical element. The snapshots display the field component, with neither the bias nor the sensitivity estimate.
and implicit midpoint integration with 64 steps per dither period. Time and mesh refinement change the boundary temperature by at most 0.052 and 0.023 mK, respectively, over 0 ≤ t ≤ 5 0 \le t \le5 0 ≤ t ≤ 5 .
Figure 2 shows regulation from 307.8 toward 308 K. The bias moves from 0.8317 to 0.8531 mW, while the applied power remains between 0.7920 and 0.8951 mW without imposed saturation. Over 40 ≤ t ≤ 50 40 \le t \le50 40 ≤ t ≤ 50 , the peak boundary error is 33.8 mK and the mean cost gap is 5.76 × 10 − 4 5.76 \times10^{-4} 5.76 × 1 0 − 4 , a 98.6% reduction from the fixed-bias gap 4.08 × 10 − 2 4.08 \times10^{-2} 4.08 × 1 0 − 2 . The leading ripple Θ a χ / ω = 33.5 \Theta_{a\chi}/\sqrt{\omega}=33.5 Θ a χ / ω = 33.5 mK contributes a 2 χ 2 / ( 2 A o p t ω ) = 5.74 × 10 − 4 a^2\chi^2/(2A_{\mathrm{opt}}\omega)=5.74 \times10^{-4} a 2 χ 2 / ( 2 A opt ω ) = 5.74 × 1 0 − 4 to the mean gap in (95), consistent with the observed residual loss.
Figure 2. Controller performance at ω = 80 \omega= 80 ω = 80 : (a) boundary temperature, (b) optical cost gap, (c) applied power and integral bias, and (d) boundary-temperature error over the final three dither cycles. The nominal cost in (b) is averaged over complete dither cycles. The baseline holds the initial heater power, and the averaged target uses matched coordinates.
Figures 3 and 4 show what the controller does to the plate as a whole. The optimal field T ∗ T^* T ∗ in Fig. 3 is hottest along the heated edge, where the nonuniform heater profile q q q leaves its imprint, and decays across the plate as heat escapes through the faces; holding the element at 308 K means maintaining this entire profile, and the controller does so without ever measuring it. Figure 4 follows the error T − T ∗ T-T^* T − T ∗ . The initial peak field error is 221 mK near the heated edge. It falls to 68.5 mK after two diffusion times (9.14 s) and to 15.1 mK at the end of the run (228.57 s), inside the ± 34 \pm34 ± 34 mK ripple band of Fig. 2(d); what remains is the thin boundary layer of the persistent dither along the heated edge. The snapshots illustrate the field component at selected times, not the full error E ω E_\omega E ω or an all-time numerical bound.
Increasing ω \omega ω from 20 to 320 reduces the late peak boundary error from 67.7 to 17.0 mK and mean cost gap from 2.33 × 10 − 3 2.33 \times10^{-3} 2.33 × 1 0 − 3 to 1.44 × 10 − 4 1.44 \times10^{-4} 1.44 × 1 0 − 4 . The slopes − 0.50 -0.50 − 0.50 and − 1.00 -1.00 − 1.00 agree with the predicted frequency dependence over this range. The plant, gains, amplitude, physical initial field, bias, and rule g ˉ ( 0 ) = 0 \bar{g}(0)=0 g ˉ ( 0 ) = 0 are fixed. The stored p 0 p_0 p 0 and matched averaged initial state vary with frequency.
Condition (19) permits δ L < 0.78125 \delta_L < 0.78125 δ L < 0.78125 . With δ L = 0.78 \delta_L=0.78 δ L = 0.78 , Proposition 1 gives σ = 0.026 \sigma=0.026 σ = 0.026 and M ≈ 134 M \approx134 M ≈ 134 , compared with the averaged rightmost eigenvalue − 0.153 -0.153 − 0.153 . However, E ω ( 0 ) = 0.373 E_\omega(0)=0.373 E ω ( 0 ) = 0.373 requires R ≥ 8 M E ω ( 0 ) ≈ 400 R \ge8ME_\omega(0) \approx400 R ≥ 8 M E ω ( 0 ) ≈ 400 . Since Φ 1 ≥ 2 R / A o p t \Phi_1 \ge2R/A_{\mathrm{opt}} Φ 1 ≥ 2 R / A opt and K 0 = coth ( 1 ) K_0=\coth(1) K 0 = coth ( 1 ) , the condition a ω ≥ 4 k p Φ 1 K 0 a\sqrt{\omega} \ge4k_p\Phi_1K_0 a ω ≥ 4 k p Φ 1 K 0 in (82) alone then requires ω > 10 8 \omega> 10^8 ω > 1 0 8 . These runs lie outside the sufficient certificate of Theorem 1 and provide performance
evidence rather than verification of its hypotheses. The gap between the certificate and the observed behavior reflects the conservatism of the constants M M M and Φ 1 \Phi_1 Φ 1 . The frequency trends predicted by the theorem are the ones observed.
XI Conclusion
The proposed dual-mode controller uses only boundary cost measurements to regulate a diffusion process near an attainable optimal equilibrium without online PDE compensation. It does so by relying on a property of diffusion rather than on a model of the plant: the π / 4 \pi/4 π /4 phase lag of a periodically heated surface fixes the demodulation, and the unknown signed gain disappears into a positive loop gain. Stability of the averaged system and trajectory estimates valid for all time establish local practical exponential regulation, with residual state and cost errors decreasing as the dither frequency increases. The thermal example illustrates the controller’s performance, with temperature and cost errors consistent with the predicted frequency dependence, and a temperature field that settles on the optimal profile across the whole plate although only the element’s transmission is measured. Future work will address less conservative frequency conditions, measurement noise, interior sensing, and additional device dynamics.