Introduction
Extremum seeking control (ESC) steers an unknown dynamical system toward an extremum of a measured objective function whose analytical form is unavailable [1 ] , [2 ] . In its classical form, the plant is treated as a stable steady-state input-output map that is optimized on a slow time scale, and closed-loop stability follows from averaging and singular perturbation arguments [3 ] , [4 ] . The time-scale separation underlying this analysis limits the achievable transient performance and excludes plants that are not stable under constant inputs. Dual-mode ESC was introduced to remove this restriction [5 ] , [6 ] . It is a proportional–integral (PI) formulation: a proportional action, generated from the measured cost, stabilizes the plant at a feasible optimum on the time scale of the plant, while an integral action adapts a bias toward the unknown constant input that sustains the optimum at equilibrium. In [7 ] , both actions are realized by periodic bounded feedback and the closed loop is analyzed through Lie-bracket averaging, which yields practical asymptotic stability of the unknown optimum for deterministic nonlinear plants. Finite-time and prescribed-time dual-mode designs are developed in [8 ] –[10 ] .
Lie-bracket approximation characterizes the effective vector fields generated by high-frequency, large-amplitude periodic inputs [11 ] and has become a standard framework for the analysis and design of extremum seeking systems [12 ] –[16 ] . A bounded trigonometric dependence on the measured cost [17 ] , [18 ] keeps the update rate bounded irrespective of the cost magnitude, and the same mechanism provides model-free stabilizing feedback under unknown control directions through minimum seeking of control Lyapunov functions [19 ] , [20 ] . The sign independence of Lie-bracket feedback has since been exploited to solve robust output regulation problems without Nussbaum-type gains [21 ] –[24 ] and to optimize mechanical systems that are open-loop unstable [25 ] , [26 ] . Robustness of large-amplitude, high-frequency schemes to bounded input and output disturbances is established in [27 ] , [28 ] , and further extensions address time-varying costs [29 ] , vanishing input oscillations [30 ] , unbiased convergence [31 ] , and initialization-free designs [32 ] . In all of these results the plant is deterministic: disturbances, when present, are bounded functions of time, and the averaged system is an ordinary differential equation.
Stochastic extremum seeking has developed along two lines. In the first, the probing signal is random. Manzie and Krstić [33 ] replace sinusoidal dithers by stochastic perturbations, and Liu and Krstić [34 ] –[37 ] develop stochastic averaging theorems for fast ergodic excitation and use them to analyze extremum seeking for static maps and nonlinear dynamical systems, in continuous and discrete time. In the second, the cost measurement is corrupted by noise, and convergence is obtained with vanishing gains and stochastic approximation arguments [38 ] , [39 ] , through quasi-stochastic approximation [40 ] , or through time-delay approaches to averaging that yield mean-square bounds [41 ] , [42 ] . In both lines the randomness enters through the algorithm or the sensor, the plant itself is deterministic, and the averaged system is again an ordinary differential equation. Persistent process noise acting on the plant state is a different situation. The noise cannot be averaged out, the closed loop is a diffusion whose stability must be assessed with stochastic Lyapunov tools [43 ] –[45 ] , and any averaging of the probing signal must retain the Brownian term [46 ] , [47 ] .
Two recent contributions bring Brownian motion into the Liebracket framework. Scheinker [48 ] extends bounded extremum seeking to Itô processes. A cosine-driven stochastic differential equation is compared with a target diffusion driven by the same Brownian motion on compact time intervals, with application to gradient-free sampling from generative models. Abdelgalil, Krstic, and Poveda [49 , 50 ] introduce stochastic Lie-bracket approximations in which smooth interpolations of a Wiener process replace periodic dithers. The resulting flows approximate Riemannian Langevin dynamics on compact manifolds in mean square, and the Gibbs invariant measure yields infinite-horizon probabilistic guarantees for global optimization. In both formulations the Brownian term is an ingredient of the optimization algorithm rather than a disturbance of the plant. In [48 ] probing and Brownian forcing share the same coefficient functions, and in [50 ] the interpolated Wiener process is itself the dither. The objective is zeroth-order optimization or sampling over the whole state space, and neither stabilization of an open-loop unstable plant nor compensation of an unknown constant input arises.
The present paper considers unknown input-affine plants subject to persistent Brownian disturbances whose diffusion coefficient is independent of the control input and need not share the actuation directions. The controller uses a single scalar cost measurement. The principal contribution is a dual-mode, PI formulation of extremum seeking for this class of controlled diffusions. A bounded cosine feedback generates, through Lie-bracket averaging, a proportional action that stabilizes a feasible minimum of the cost and an integral action that regulates the input to an unknown bias. The averaged closed loop is a PI-regulated diffusion that retains the process noise, and the analysis is carried out directly on the stochastic closed loop at a fixed, sufficiently large excitation frequency. Specifically:
The averaged dynamics of the cosine-controlled diffusion are identified as a PI target diffusion. A local matrix inequality condition for exponential stability of its equilibrium is derived in which the input map enters only through a positive semidefinite product, so that the proportional gain can be selected without knowledge of the control direction.
For the noise-free closed loop, an exact periodic response is constructed near the optimum and a Lyapunov function centered on this response is obtained with bounds that are uniform in the excitation frequency. This construction separates noise-induced deviations from the oscillations caused by probing. The mean-square regulation error is bounded by an exponentially decaying transient plus a term of order σ 2 + ω − 1 \sigma^2+\omega^{-1} σ 2 + ω − 1 , where σ \sigma σ scales the noise and ω \omega ω is the excitation frequency.
Capture and residence estimates follow from stopped Lyapunov arguments [43 , 44 ] . For sufficiently small noise, trajectories starting in a certified compact set enter a prescribed neighbourhood of the optimum and remain there over horizons that grow exponentially with the inverse noise variance, with a failure probability that decreases on the same scale, in the spirit of the exit problem of [45 ] .
Under an additional regional incremental stability condition of stochastic contraction type [51 ] , the closed loop and the averaged diffusion driven by the same Brownian motion are compared uniformly in time, with a mean-square error of order ω − 1 \omega^{-1} ω − 1 that applies until either trajectory leaves the comparison region; see [52 ] for uniform-in-time averaging of diffusions under strong exponential stability.
A planar hydrodynamic saddle trap illustrates the regulation result. The controlled extensional direction is unstable, the passive compressional direction carries Brownian noise, and an unknown offset requires a nonzero holding input. A regional target certificate covers uncertain strain rate, optical amplitude, drift offset, and either actuator sign. The scalar measured objective is a Gaussian optical-intensity score.
Sections II and III introduce the preliminaries, plant model, and stability assumptions. Section IV presents the cosine controller, and Section V derives the averaged proportional–integral dynamics and examines their stability. Section VI constructs the periodic change of coordinates, establishes the finite-horizon approximation, and proves attraction to the noise-free periodic response. Section VII develops the stochastic regulation, capture, and residence estimates, followed by the time-uniform approximation result under incremental stability. Section VIII presents the hydrodynamic-trap model and stochastic simulations, and Section IX concludes the paper. Appendices A and B provide the regional certificate and an analytical bound on radial residence.
Preliminaries
The symbols ∣ ⋅ ∣ |\cdot| ∣ ⋅ ∣ , ∥ ⋅ ∥ \|\cdot\| ∥ ⋅ ∥ , and ∥ ⋅ ∥ F \|\cdot\|_{F} ∥ ⋅ ∥ F denote Euclidean, induced, and Frobenius norms. Write sym A = ( A + A T ) / 2 \operatorname{sym} A=(A+A^{\mathsf T})/2 sym A = ( A + A T ) /2 , let S + + n \mathbb{S}_{++}^{n} S ++ n denote the symmetric positive-definite matrices, and set T = R / ( 2 π Z ) \mathbb{T}=\mathbb{R}/(2\pi\mathbb{Z}) T = R / ( 2 π Z ) . For open sets, U ⊆ D U\subseteq D U ⊆ D means that U ‾ \overline{U} U is compact and contained in D D D . A C j C^{j} C j norm includes spatial derivatives through order j j j . Variables entering trigonometric functions are dimensionless. Constants denoted by C \mathcal{C} C may be enlarged between estimates. Their dependence on a time horizon or a noise bound is indicated when relevant.
All stochastic equations are interpreted in the Itô sense on a filtered probability space satisfying the usual conditions. The Brownian dimension is denoted by d B d_{B} d B , reserving r r r for a Lyapunov level. Solutions are considered up to their maximal lifetime. For an open set U U U , define τ U ( Z ) = inf { t ≥ 0 : Z t ∉ U } \tau_{U}(Z)=\inf\{t\geq0:Z_{t}\notin U\} τ U ( Z ) = inf { t ≥ 0 : Z t ∈ / U } , with inf ∅ = ∞ \inf\varnothing=\infty inf ∅ = ∞ . An integrand multiplied by 1 { t < τ } \mathbf{1}_{\{t<\tau\}} 1 { t < τ } is defined to be zero after exit. For vector fields A A A , B B B and scalar function H H H , the Lie derivative and the Lie bracket are denoted by
L A H = ∇ H T A , [ A , B ] = D B A − D A B (1) L_{A}H=\nabla H^{\mathsf T}A,\qquad[A,B]=DB\,A-DA\,B \tag*{(1)} L A H = ∇ H T A , [ A , B ] = D B A − D A B (1)
respectively.
For a vector- or matrix-valued function F ( θ , z ) F(\theta,z) F ( θ , z ) that is periodic in θ \theta θ , we use the spatial norm
∥ F ∥ C ∗ k : = max 0 ≤ j ≤ k θ ∈ T z ∈ R n ∥ D z j F ( θ , z ) ∥ . \|F\|_{C^{k}_{*}}:=\max_{\substack{0\leq j\leq k\\ \theta\in\mathbb{T}\\ z\in\mathbb{R}^{n}}}\|D_{z}^{j}F(\theta,z)\|. ∥ F ∥ C ∗ k := 0 ≤ j ≤ k θ ∈ T z ∈ R n max ∥ D z j F ( θ , z ) ∥.
Here D z 0 F = F D_{z}^{0}F=F D z 0 F = F . For j = 0 j=0 j = 0 , the norm is the Euclidean norm for vectors and the Frobenius norm for matrices. For $j\geq 1, it is the corresponding induced multilinear norm. Thus the C k C^k C k norm bounds the function and its spatial derivatives through order k k k , uniformly in the phase. Derivatives with respect to θ \theta θ are not included.
A stopped Lyapunov estimate
Consider the Itô diffusion d Z t = a ( t , Z t ) d t + σ S ( t , Z t ) d B t dZ_t=a(t,Z_t)dt+\sigma S(t,Z_t)dB_t d Z t = a ( t , Z t ) d t + σ S ( t , Z t ) d B t on an open U ⊂ R d U\subset\mathbb{R}^d U ⊂ R d , from a deterministic state z 0 z_0 z 0 . The coefficients are continuous and locally Lipschitz in the state, uniformly on finite time intervals. For a nonnegative Φ ∈ C 1 , 2 ( [ 0 , ∞ ) × U ) \Phi\in C^{1,2}([0,\infty)\times U) Φ ∈ C 1 , 2 ([ 0 , ∞ ) × U ) , the time-augmented generator L σ \mathcal{L}_\sigma L σ acts as
L σ Φ = ∂ t Φ + ∇ z Φ T a + 1 2 σ 2 tr ( S S T D z 2 Φ ) . \mathcal{L}_\sigma\Phi=\partial_t\Phi+\nabla_z\Phi^\mathsf{T}a+\frac{1}{2}\sigma^2\operatorname{tr}(SS^\mathsf{T}D_z^2\Phi). L σ Φ = ∂ t Φ + ∇ z Φ T a + 2 1 σ 2 tr ( S S T D z 2 Φ ) .
Setting Q t : = Φ ( t , Z t ) Q_t:=\Phi(t,Z_t) Q t := Φ ( t , Z t ) , Itô’s formula gives
d Q t = ( L σ Φ ) ( t , Z t ) d t + σ ∇ z Φ ( t , Z t ) T S ( t , Z t ) d B t . dQ_t=(\mathcal{L}_\sigma\Phi)(t,Z_t)dt+\sigma\nabla_z\Phi(t,Z_t)^\mathsf{T}S(t,Z_t)dB_t. d Q t = ( L σ Φ ) ( t , Z t ) d t + σ ∇ z Φ ( t , Z t ) T S ( t , Z t ) d B t .
Here, q σ ( Φ ) : = σ 2 ∣ S T ∇ z Φ ∣ 2 q_\sigma(\Phi):=\sigma^2\lvert S^\mathsf{T}\nabla_z\Phi\rvert^2 q σ ( Φ ) := σ 2 ∣ S T ∇ z Φ ∣ 2 denotes the quadratic-variation rate of the local-martingale part of Q t Q_t Q t . Suppose that the closed sublevels { z ∈ U : Φ ( t , z ) ≤ R } \{z\in U:\Phi(t,z)\le R\} { z ∈ U : Φ ( t , z ) ≤ R } lie in one compact subset of U U U , uniformly in t t t , and let τ R = inf { t ≥ 0 : Q t ≥ R } \tau_R=\inf\{t\ge0:Q_t\ge R\} τ R = inf { t ≥ 0 : Q t ≥ R } . Then for Q 0 < R Q_0<R Q 0 < R , compact continuation and continuity give Q τ R = R Q_{\tau_R}=R Q τ R = R at every finite exit.
The following lemma provides a moment bound on Q t Q_t Q t before exit from a Lyapunov sublevel and a bound on the probability of exit.
Lemma 1. Suppose, for Φ < R \Phi<R Φ < R ,
L σ Φ ≤ − λ Φ + c 0 σ 2 , q σ ( Φ ) ≤ d 0 σ 2 Φ , \mathcal{L}_\sigma\Phi\le-\lambda\Phi+c_0\sigma^2,\qquad q_\sigma(\Phi)\le d_0\sigma^2\Phi, L σ Φ ≤ − λ Φ + c 0 σ 2 , q σ ( Φ ) ≤ d 0 σ 2 Φ , where λ > 0 \lambda>0 λ > 0 and c 0 , d 0 ≥ 0 c_0,d_0\ge0 c 0 , d 0 ≥ 0 . For Q 0 < R Q_0<R Q 0 < R ,
E [ Q t 1 { t < τ R } ] ≤ e − λ t Q 0 + c 0 σ 2 λ ( 1 − e − λ t ) . (2) \mathbb{E}[Q_t\mathbf{1}_{\{t<\tau_R\}}]\le e^{-\lambda t}Q_0+\frac{c_0\sigma^2}{\lambda}(1-e^{-\lambda t}). \tag*{(2)} E [ Q t 1 { t < τ R } ] ≤ e − λ t Q 0 + λ c 0 σ 2 ( 1 − e − λ t ) . (2) There are η , C 0 > 0 \eta,C_0>0 η , C 0 > 0 , depending only on λ , c 0 , d 0 \lambda,c_0,d_0 λ , c 0 , d 0 , such that for σ > 0 \sigma>0 σ > 0 and finite T ≥ 0 T\ge0 T ≥ 0 ,
P { τ R ≤ T } ≤ e − η ( R − Q 0 ) / σ 2 + C 0 T e − η R / σ 2 . (3) \mathbb{P}\{\tau_R\le T\}\le e^{-\eta(R-Q_0)/\sigma^2}+C_0Te^{-\eta R/\sigma^2}. \tag*{(3)} P { τ R ≤ T } ≤ e − η ( R − Q 0 ) / σ 2 + C 0 T e − η R / σ 2 . (3) For σ = 0 \sigma=0 σ = 0 , τ R = ∞ \tau_R=\infty τ R = ∞ and Q t ≤ e − λ t Q 0 Q_t\le e^{-\lambda t}Q_0 Q t ≤ e − λ t Q 0 .
Proof. Itô’s formula applied up to τ R \tau_R τ R for e λ t Q t e^{\lambda t}Q_t e λ t Q t yields
e λ t E [ Q t 1 { t < τ R } ] + R E [ e λ τ R 1 { τ R ≤ t } ] ≤ Q 0 + c 0 σ 2 λ ( e λ t − 1 ) . e^{\lambda t}\mathbb{E}[Q_t\mathbf{1}_{\{t<\tau_R\}}]+R\mathbb{E}[e^{\lambda\tau_R}\mathbf{1}_{\{\tau_R\le t\}}]\le Q_0+\frac{c_0\sigma^2}{\lambda}(e^{\lambda t}-1). e λ t E [ Q t 1 { t < τ R } ] + R E [ e λ τ R 1 { τ R ≤ t } ] ≤ Q 0 + λ c 0 σ 2 ( e λ t − 1 ) .
Dropping the exit contribution proves (2). For (3), choose η > 0 \eta>0 η > 0 with η d 0 ≤ λ \eta d_0\le\lambda η d 0 ≤ λ and set F = exp ( η Φ / σ 2 ) F=\exp(\eta\Phi/\sigma^2) F = exp ( η Φ/ σ 2 ) . The generator chain rule gives
L σ F ≤ e η y ( η c 0 − η λ y / 2 ) , y = Φ / σ 2 . \mathcal{L}_\sigma F\le e^{\eta y}(\eta c_0-\eta\lambda y/2),\qquad y=\Phi/\sigma^2. L σ F ≤ e η y ( η c 0 − η λ y /2 ) , y = Φ/ σ 2 .
The right side has a finite upper bound C 0 > 0 C_0>0 C 0 > 0 on y ≥ 0 y\ge0 y ≥ 0 , independent of R , σ R,\sigma R , σ . Thus E F ( T ∧ τ R , Z T ∧ τ R ) ≤ e η Q 0 / σ 2 + C 0 T \mathbb{E}F(T\wedge\tau_R,Z_{T\wedge\tau_R})\le e^{\eta Q_0/\sigma^2}+C_0T E F ( T ∧ τ R , Z T ∧ τ R ) ≤ e η Q 0 / σ 2 + C 0 T . On exit F = e η R / σ 2 F=e^{\eta R/\sigma^2} F = e η R / σ 2 , which proves (3). All stopped stochastic integrals are integrable on finite intervals by compactness. For σ = 0 \sigma=0 σ = 0 , the differential inequality prevents Q Q Q from reaching R R R , and compact continuation gives the claim.
Consider the system
d X t = [ f ( X t ) + G ( X t ) u t ] d t + σ Σ 0 ( X t ) d B t , dX_t=[f(X_t)+G(X_t)u_t]dt+\sigma\Sigma_0(X_t)dB_t, d X t = [ f ( X t ) + G ( X t ) u t ] d t + σ Σ 0 ( X t ) d B t ,
Y t = h ( X t ) , Y_t=h(X_t), Y t = h ( X t ) ,
where X t ∈ D ⊂ R n X_t\in D\subset\mathbb{R}^n X t ∈ D ⊂ R n , u t ∈ R m u_t\in\mathbb{R}^m u t ∈ R m , and Y t ∈ R Y_t\in\mathbb{R} Y t ∈ R . The unknown maps are f : D → R n f:D\to\mathbb{R}^n f : D → R n , G : D → R n × m G:D\to\mathbb{R}^{n\times m} G : D → R n × m , Σ 0 : D → R n × d B \Sigma_0:D\to\mathbb{R}^{n\times d_B} Σ 0 : D → R n × d B , and h : D → R h:D\to\mathbb{R} h : D → R . The scalar σ ≥ 0 \sigma\ge0 σ ≥ 0 scales the diffusion amplitude and is fixed for a given plant. The controller uses only Y Y Y and its own states.
Assumption 1. The domain D D D is open, 1 ≤ m ≤ n 1\le m\le n 1 ≤ m ≤ n , and f , G , Σ 0 , h f,G,\Sigma_0,h f , G , Σ 0 , h are smooth on D D D .
Assumption 2. There exists ( x ⋆ , u ⋆ ) ∈ D × R m (x^\star,u^\star)\in D\times\mathbb{R}^m ( x ⋆ , u ⋆ ) ∈ D × R m such that
f ( x ⋆ ) + G ( x ⋆ ) u ⋆ = 0. f(x^\star)+G(x^\star)u^\star=0. f ( x ⋆ ) + G ( x ⋆ ) u ⋆ = 0.
The cost satisfies ∇ h ( x ⋆ ) = 0 \nabla h(x^\star)=0 ∇ h ( x ⋆ ) = 0 and ∇ 2 h ( x ⋆ ) > 0 \nabla^2h(x^\star)>0 ∇ 2 h ( x ⋆ ) > 0 , and G ( x ⋆ ) G(x^\star) G ( x ⋆ ) has full column rank.
The objective is the regulation of the optimal cost h ⋆ = h ( x ⋆ ) h^\star=h(x^\star) h ⋆ = h ( x ⋆ ) at the corresponding equilibrium pair ( x ⋆ , u ⋆ ) (x^\star,u^\star) ( x ⋆ , u ⋆ ) , despite persistent process noise. These disturbances continually perturb the state, requiring ongoing feedback correction to maintain operation near the desired equilibrium. We therefore seek to derive quantitative bounds on the regulation error and on the probability of reaching and remaining in a prescribed neighborhood of the desired operating point.
We define the input sensitivity
ℓ ( x ) : = G ( x ) T ∇ h ( x ) ∈ R m . \ell(x):=G(x)^\mathsf{T}\nabla h(x)\in\mathbb{R}^m. ℓ ( x ) := G ( x ) T ∇ h ( x ) ∈ R m .
Its components are the directional derivatives of the cost along the columns of G G G . For gains K P , K I ∈ S + + m K_P,K_I\in\mathbb{S}_{++}^m K P , K I ∈ S ++ m , we introduce the bias u b ∈ R m u_b\in\mathbb{R}^m u b ∈ R m , the augmented state z = col ( x , u b ) z=\operatorname{col}(x,u_b) z = col ( x , u b ) , the desired operating point z ⋆ = col ( x ⋆ , u ⋆ ) z^\star=\operatorname{col}(x^\star,u^\star) z ⋆ = col ( x ⋆ , u ⋆ ) , and the domain D = D × R m D=D\times\mathbb{R}^m D = D × R m . The deterministic PI target field is z ˙ = F T ( z ) \dot z=F_T(z) z ˙ = F T ( z ) , where
F T ( z ) = [ f ( x ) + G ( x ) ( u b − K P ℓ ( x ) ) − K I ℓ ( x ) ] . (4) F_T(z)=
\begin{bmatrix}
f(x)+G(x)(u_b-K_P\ell(x))\\
-K_I\ell(x)
\end{bmatrix}. \tag*{(4)} F T ( z ) = [ f ( x ) + G ( x ) ( u b − K P ℓ ( x )) − K I ℓ ( x ) ] . (4)
Here, F T F_T F T specifies the desired averaged dynamics. The controller in Section IV uses the measured cost to produce PI actions through Lie-bracket averaging, without evaluating ℓ ( x ) \ell(x) ℓ ( x ) .
Definition 1. A regional target certificate for a compact initial set K ⊂ D \mathcal{K}\subset D K ⊂ D consists of a bounded open O ⊂ D \mathcal{O}\subset D O ⊂ D containing z ⋆ z^\star z ⋆ , a function V ∈ C 4 V\in C^4 V ∈ C 4 on a neighborhood of O ‾ \overline{\mathcal{O}} O , constants c 1 , c 2 , c 3 > 0 c_1,c_2,c_3>0 c 1 , c 2 , c 3 > 0 , and levels 0 < r 0 < R 0 0<r_0<R_0 0 < r 0 < R 0 . For z ∈ O z\in\mathcal{O} z ∈ O ,
c 1 ∣ z − z ⋆ ∣ 2 ≤ V ( z ) ≤ c 2 ∣ z − z ⋆ ∣ 2 , (5) c_1\lvert z-z^\star\rvert^2\le V(z)\le c_2\lvert z-z^\star\rvert^2, \tag*{(5)} c 1 ∣ z − z ⋆ ∣ 2 ≤ V ( z ) ≤ c 2 ∣ z − z ⋆ ∣ 2 , (5)
∇ V ( z ) T F T ( z ) ≤ − c 3 ∣ z − z ⋆ ∣ 2 . (6) \nabla V(z)^\mathsf{T}F_T(z)\le-c_3\lvert z-z^\star\rvert^2. \tag*{(6)} ∇ V ( z ) T F T ( z ) ≤ − c 3 ∣ z − z ⋆ ∣ 2 . (6)
The set K \mathcal{K} K lies in { z ∈ O : V ( z ) ≤ r 0 } \{z\in\mathcal{O}:V(z)\le r_0\} { z ∈ O : V ( z ) ≤ r 0 } , and S R 0 : = { z ∈ O : V ( z ) ≤ R 0 } S_{R_0}:=\{z\in\mathcal{O}:V(z)\le R_0\} S R 0 := { z ∈ O : V ( z ) ≤ R 0 } is compact and contained in O \mathcal{O} O .
Definition 1 concerns the deterministic averaged system z ˙ = F T ( z ) \dot z=F_T(z) z ˙ = F T ( z ) . This certificate is used below to construct the periodic Lyapunov function Φ ω \Phi_\omega Φ ω for the cosine-controlled system (9), where ω > 0 \omega>0 ω > 0 denotes the controller’s base angular excitation frequency. The stochastic drift and quadratic-variation estimates required by Lemma 1 are then derived from this construction.
Assumption 3. Let B ⊂ D \mathcal{B} \subset\mathcal{D} B ⊂ D be an open neighborhood of z ⋆ z^\star z ⋆ . For every compact K ⊂ B \mathcal{K} \subset\mathcal{B} K ⊂ B , gains K P , K I ∈ S + + m K_P,K_I \in\mathbb{S}^{m}_{++} K P , K I ∈ S ++ m can be chosen so that F T F_T F T admits a regional target certificate for K \mathcal{K} K . The gains may depend on K \mathcal{K} K but are independent of ω \omega ω .
Remark 1. Assumption 3 is motivated by earlier work on deterministic stability analysis of PI and dual-mode extremum seeking [5 , 7 ] . For certain classes of nonlinear systems with convex cost functions, these works develop stabilizing gain choices and Lyapunov constructions under suitable conditions on the dynamics and input sensitivity. Their analyses address both convergence of the plant state to the optimum and compensation of the unknown sustaining input.
For the considered plant class, the present formulation expresses the required deterministic stability property through the regional certificate in Definition 1. Proposition 2 provides a local sufficient condition for the existence of this certificate. The subsequent analysis uses this certificate to quantify the effects of periodic excitation and process noise on the corresponding closed loop.
Dual-Mode Controller
We choose ω > 0 \omega> 0 ω > 0 , distinct positive integers κ 1 , … , κ m \kappa_1,\ldots,\kappa_m κ 1 , … , κ m , and set C P = K P 1 / 2 C_P=K_P^{1/2} C P = K P 1/2 and C I = K I K P − 1 / 2 C_I=K_I K_P^{-1/2} C I = K I K P − 1/2 . For v t ω = col ( v 1 , t ω , … , v m , t ω ) v_t^\omega=\operatorname{col}(v_{1,t}^\omega,\ldots,v_{m,t}^\omega) v t ω = col ( v 1 , t ω , … , v m , t ω ) , the feedback is
d θ t = ω d t , θ 0 ∈ T , v i , t ω = 2 κ i ω cos ( Y t ω + κ i θ t ) , u t ω = u b , t ω + C P v t ω , d u b , t ω = C I v t ω d t . (7) \begin{aligned}
\mathrm{d}\theta_t &= \omega\mathrm{d}t,\qquad\theta_0\in\mathbb{T},\\
v_{i,t}^\omega&= \sqrt{2\kappa_i\omega}\cos(Y_t^\omega+\kappa_i\theta_t),\\
u_t^\omega&= u_{b,t}^\omega+C_Pv_t^\omega,\\
\mathrm{d}u_{b,t}^\omega&= C_Iv_t^\omega\mathrm{d}t. \tag*{(7)}
\end{aligned} d θ t v i , t ω u t ω d u b , t ω = ω d t , θ 0 ∈ T , = 2 κ i ω cos ( Y t ω + κ i θ t ) , = u b , t ω + C P v t ω , = C I v t ω d t . (7)
Here Y t ω = h ( X t ω ) Y_t^\omega=h(X_t^\omega) Y t ω = h ( X t ω ) and u b , 0 ω u_{b,0}^\omega u b , 0 ω is prescribed. The phase and bias have finite-variation paths. The bias is random through Y t ω Y_t^\omega Y t ω , but its update requires neither an output derivative nor integration against d Y t ω \mathrm{d}Y_t^\omega d Y t ω .
For z = col ( x , u b ) z=\operatorname{col}(x,u_b) z = col ( x , u b ) , we define the augmented drift
b ( z ) = [ f ( x ) + G ( x ) u b 0 m ] b(z)=\begin{bmatrix} f(x)+G(x)u_b\\ 0_m \end{bmatrix} b ( z ) = [ f ( x ) + G ( x ) u b 0 m ]
and the diffusion matrix
S 0 ( z ) = [ Σ 0 ( x ) 0 m × d B ] . S_0(z)=\begin{bmatrix}\Sigma_0(x)\\0_{m\times d_B}\end{bmatrix}. S 0 ( z ) = [ Σ 0 ( x ) 0 m × d B ] .
With e i e_i e i the i i i th standard basis vector of R m \mathbb{R}^m R m , we introduce the following augmented-state fields
A i ( z ) = cos ( h ( x ) ) H i ( z ) B i ( z ) = − sin ( h ( x ) ) H i ( z ) , (8) \begin{aligned}
A_i(z)&=\cos(h(x))H_i(z)\\
B_i(z)&=-\sin(h(x))H_i(z), \tag*{(8)}
\end{aligned} A i ( z ) B i ( z ) = cos ( h ( x )) H i ( z ) = − sin ( h ( x )) H i ( z ) , (8)
where
H i ( z ) = [ G ( x ) C P e i C I e i ] . H_i(z)=\begin{bmatrix}G(x)C_Pe_i\\C_Ie_i\end{bmatrix}. H i ( z ) = [ G ( x ) C P e i C I e i ] .
Let p i ( θ ) = 2 κ i cos ( κ i θ ) p_i(\theta)=\sqrt{2\kappa_i}\cos(\kappa_i\theta) p i ( θ ) = 2 κ i cos ( κ i θ ) and q i ( θ ) = 2 κ i sin ( κ i θ ) q_i(\theta)=\sqrt{2\kappa_i}\sin(\kappa_i\theta) q i ( θ ) = 2 κ i sin ( κ i θ ) , and put a ( θ , z ) = ∑ i [ A i ( z ) p i ( θ ) + B i ( z ) q i ( θ ) ] a(\theta,z)=\sum_i[A_i(z)p_i(\theta)+B_i(z)q_i(\theta)] a ( θ , z ) = ∑ i [ A i ( z ) p i ( θ ) + B i ( z ) q i ( θ )] . The augmented closed loop is
d Z t ω = [ b ( Z t ω ) + ω a ( θ t , Z t ω ) ] d t + σ S 0 ( Z t ω ) d B t . (9) \mathrm{d}Z_t^\omega=[b(Z_t^\omega)+\sqrt{\omega}\,a(\theta_t,Z_t^\omega)]\mathrm{d}t+\sigma S_0(Z_t^\omega)\mathrm{d}B_t. \tag*{(9)} d Z t ω = [ b ( Z t ω ) + ω a ( θ t , Z t ω )] d t + σ S 0 ( Z t ω ) d B t . (9)
Its phase-augmented generator A ω , σ \mathcal{A}_{\omega,\sigma} A ω , σ is defined such that
A ω , σ Φ = ω ∂ θ Φ + ∇ z Φ T ( b + ω a ) + 1 2 σ 2 tr ( S 0 S 0 T D z 2 Φ ) . (10) \mathcal{A}_{\omega,\sigma}\Phi=\omega\partial_\theta\Phi+\nabla_z\Phi^\mathsf{T}(b+\sqrt{\omega}a)+\frac{1}{2}\sigma^2\operatorname{tr}(S_0S_0^\mathsf{T}D_z^2\Phi). \tag*{(10)} A ω , σ Φ = ω ∂ θ Φ + ∇ z Φ T ( b + ω a ) + 2 1 σ 2 tr ( S 0 S 0 T D z 2 Φ ) . (10)
With a ω = ( 2 ω ∑ i κ i ) 1 / 2 a_\omega=(2\omega\sum_i\kappa_i)^{1/2} a ω = ( 2 ω ∑ i κ i ) 1/2 , the bounded cosine yields the following bounds
∣ u t ω − u b , t ω ∣ ≤ ∥ C P ∥ a ω , ∣ C I v t ω ∣ ≤ ∥ C I ∥ a ω . \lvert u_t^\omega-u_{b,t}^\omega\rvert\le\lVert C_P\rVert a_\omega,\qquad\lvert C_Iv_t^\omega\rvert\le\lVert C_I\rVert a_\omega. ∣ u t ω − u b , t ω ∣ ≤ ∥ C P ∥ a ω , ∣ C I v t ω ∣ ≤ ∥ C I ∥ a ω .
For each fixed frequency, the bounds are independent of the cost magnitude and grow with ω \sqrt{\omega} ω . Replacing the amplitude in (7) by a frequency-independent constant α i \alpha_i α i results in the bracket coefficient α i 2 / ( 2 κ i ω ) \alpha_i^2/(2\kappa_i\omega) α i 2 / ( 2 κ i ω ) . Hence, the chosen amplitude scaling ω \sqrt{\omega} ω maintains a non-vanishing Lie-bracket coefficient as the frequency increases.
Remark 2. A constant cost offset leaves ℓ ( x ) \ell(x) ℓ ( x ) and F T F_T F T unchanged but generally changes the feedback (7). To make the applied input invariant to such offsets, replace Y t ω Y_t^\omega Y t ω in (7) by Y t ω − w t Y_t^\omega-w_t Y t ω − w t with d w t = γ ( Y t ω − w t ) d t \mathrm{d}w_t=\gamma(Y_t^\omega-w_t)\mathrm{d}t d w t = γ ( Y t ω − w t ) d t , initialized w 0 = Y 0 ω w_0=Y_0^\omega w 0 = Y 0 ω , and fixed γ > 0 \gamma>0 γ > 0 . This modification adds filter dynamics while preserving the averaged dynamics F T ( x , u b ) F_T(x,u_b) F T ( x , u b ) . Applying the regional stability results to this modified controller requires a Lyapunov certificate that also includes the filter state w w w .
Averaged Dynamics and Target Stability
We first identify the averaged PI dynamics and give a local sufficient condition for the target certificate. For a 2 π 2\pi 2 π -periodic function, write ⟨ v ⟩ = ( 2 π ) − 1 ∫ 0 2 π v ( θ ) d θ \langle v\rangle=(2\pi)^{-1}\int_0^{2\pi}v(\theta)\mathrm{d}\theta ⟨ v ⟩ = ( 2 π ) − 1 ∫ 0 2 π v ( θ ) d θ . The normalized input dithers satisfy
⟨ ( ∫ 0 θ p i ( s ) d s ) q i ( θ ) ⟩ = 1. \left\langle\left(\int_0^\theta p_i(s)\mathrm{d}s\right)q_i(\theta)\right\rangle=1. ⟨ ( ∫ 0 θ p i ( s ) d s ) q i ( θ ) ⟩ = 1.
Reversing p i , q i p_i,q_i p i , q i gives − 1 -1 − 1 , while all corresponding averages between distinct channels vanish.
Proposition 1. The fields in (8) satisfy
∑ i = 1 m [ A i , B i ] = − [ G K P ℓ K I ℓ ] . (11) \sum_{i=1}^{m}[A_i,B_i]=-\begin{bmatrix}GK_P\ell\\K_I\ell\end{bmatrix}. \tag*{(11)} i = 1 ∑ m [ A i , B i ] = − [ G K P ℓ K I ℓ ] . (11)
Proof. For h ~ ( z ) = h ( x ) \tilde h(z)=h(x) h ~ ( z ) = h ( x ) , the product rule with (1) gives [ A i , B i ] = − ( L H i h ~ ) H i [A_i,B_i]=-(L_{H_i}\tilde h)H_i [ A i , B i ] = − ( L H i h ~ ) H i . Since L H i h ~ = e i T C P T ℓ L_{H_i}\tilde h=e_i^\mathsf{T}C_P^\mathsf{T}\ell L H i h ~ = e i T C P T ℓ , summation gives
∑ i [ A i , B i ] = − [ G C P C P T ℓ C I C P T ℓ ] . \sum_i[A_i,B_i]=-\begin{bmatrix}GC_PC_P^\mathsf{T}\ell\\C_IC_P^\mathsf{T}\ell\end{bmatrix}. i ∑ [ A i , B i ] = − [ G C P C P T ℓ C I C P T ℓ ] .
The identities C P C P T = K P C_PC_P^\mathsf{T}=K_P C P C P T = K P and C I C P T = K I C_IC_P^\mathsf{T}=K_I C I C P T = K I prove (11). □ \square □
We note that adding the bracket correction in (11) to the drift b b b gives F T F_T F T . The associated averaged diffusion is
d Z ˉ t = F T ( Z ˉ t ) d t + σ S 0 ( Z ˉ t ) d B t . (12) \mathrm{d}\bar Z_t=F_T(\bar Z_t)\mathrm{d}t+\sigma S_0(\bar Z_t)\mathrm{d}B_t. \tag*{(12)} d Z ˉ t = F T ( Z ˉ t ) d t + σ S 0 ( Z ˉ t ) d B t . (12)
Setting σ = 0 \sigma=0 σ = 0 in (12) recovers the deterministic PI target z ˙ = F T ( z ) \dot z=F_T(z) z ˙ = F T ( z ) . By Definition 1, if a regional target certificate is available, then (5), (6) imply that S R 0 S_{R_0} S R 0 is forward invariant, and every solution initialized in S R 0 S_{R_0} S R 0 satisfies
∣ z ( t ) − z ⋆ ∣ ≤ c 2 c 1 e − c 3 t / ( 2 c 2 ) ∣ z ( 0 ) − z ⋆ ∣ \lvert z(t)-z^\star\rvert\le\sqrt{\frac{c_2}{c_1}}e^{-c_3t/(2c_2)}\lvert z(0)-z^\star\rvert ∣ z ( t ) − z ⋆ ∣ ≤ c 1 c 2 e − c 3 t / ( 2 c 2 ) ∣ z ( 0 ) − z ⋆ ∣
Compact continuation gives existence for all t ≥ 0 t \ge0 t ≥ 0 . The following proposition gives a sufficient condition for such a certificate to exist locally.
Proposition 2. Under Assumptions 1–2, define H ⋆ = ∇ 2 h ( x ⋆ ) H_\star= \nabla^2 h(x^\star) H ⋆ = ∇ 2 h ( x ⋆ ) , G ⋆ = G ( x ⋆ ) G_\star= G(x^\star) G ⋆ = G ( x ⋆ ) , and A ⋆ = D x ( f + G u ⋆ ) ( x ⋆ ) A_\star= D_x(f+Gu^\star)(x^\star) A ⋆ = D x ( f + G u ⋆ ) ( x ⋆ ) . If K P ∈ S + + m K_P \in\mathbb{S}_{++}^m K P ∈ S ++ m satisfies
sym ( H ⋆ A ⋆ ) − H ⋆ G ⋆ K P G ⋆ T H ⋆ ≺ 0 , (13) \operatorname{sym}(H_\star A_\star)-H_\star G_\star K_P G_\star^{\mathsf T}H_\star\prec0, \tag*{(13)} sym ( H ⋆ A ⋆ ) − H ⋆ G ⋆ K P G ⋆ T H ⋆ ≺ 0 , (13) then z ⋆ z^\star z ⋆ is locally exponentially stable for every fixed K I ∈ S + + m K_I \in\mathbb{S}_{++}^m K I ∈ S ++ m . Every compact initial set in a sufficiently small neighborhood admits a regional target certificate.
Proof. Set A P = A ⋆ − G ⋆ K P G ⋆ T H ⋆ A_P=A_\star-G_\star K_PG_\star^{\mathsf T}H_\star A P = A ⋆ − G ⋆ K P G ⋆ T H ⋆ . Since D ℓ ( x ⋆ ) = G ⋆ T H ⋆ D\ell(x^\star)=G_\star^{\mathsf T}H_\star D ℓ ( x ⋆ ) = G ⋆ T H ⋆ , the target Jacobian A T : = D F T ( z ⋆ ) A_T:=DF_T(z^\star) A T := D F T ( z ⋆ ) is then
A T = [ A P G ⋆ − K I G ⋆ T H ⋆ 0 ] . A_T=
\begin{bmatrix}
A_P & G_\star\\
-K_I G_\star^{\mathsf T}H_\star& 0
\end{bmatrix}. A T = [ A P − K I G ⋆ T H ⋆ G ⋆ 0 ] .
By (13), Q = − sym ( H ⋆ A P ) ≻ 0 Q=-\operatorname{sym}(H_\star A_P)\succ0 Q = − sym ( H ⋆ A P ) ≻ 0 . Then defining P 0 = diag ( H ⋆ , K I − 1 ) P_0=\operatorname{diag}(H_\star,K_I^{-1}) P 0 = diag ( H ⋆ , K I − 1 ) , we have A T T P 0 + P 0 A T = diag ( − 2 Q , 0 ) A_T^{\mathsf T}P_0+P_0A_T=\operatorname{diag}(-2Q,0) A T T P 0 + P 0 A T = diag ( − 2 Q , 0 ) . An eigenpair A T v = μ v A_Tv=\mu v A T v = μv , v = col ( v x , v b ) ≠ 0 v=\operatorname{col}(v_x,v_b)\ne0 v = col ( v x , v b ) = 0 , satisfies
2 Re ( μ ) v ⋆ P 0 v = − 2 v x ⋆ Q v x . 2\operatorname{Re}(\mu)v^\star P_0v=-2v_x^\star Qv_x. 2 Re ( μ ) v ⋆ P 0 v = − 2 v x ⋆ Q v x .
If Re ( μ ) ≥ 0 \operatorname{Re}(\mu)\ge0 Re ( μ ) ≥ 0 , then v x = 0 v_x=0 v x = 0 and the first block gives G ⋆ v b = 0 G_\star v_b=0 G ⋆ v b = 0 . Full column rank contradicts v ≠ 0 v\ne0 v = 0 , so A T A_T A T is Hurwitz. Consequently, there exists a unique P ∈ S + + m + n P\in\mathbb{S}_{++}^{m+n} P ∈ S ++ m + n such that A T T P + P A T = − I A_T^{\mathsf T}P+PA_T=-I A T T P + P A T = − I . Denoting e : = z − z ⋆ e:=z-z^\star e := z − z ⋆ , smoothness gives
d d t ( e T P e ) ≤ − 1 2 ∣ e ∣ 2 \frac{\mathrm{d}}{\mathrm{d}t}(e^{\mathsf T}Pe)\le-\frac{1}{2}\lvert e\rvert^2 d t d ( e T P e ) ≤ − 2 1 ∣ e ∣ 2
on a sufficiently small closed ball. We choose an outer quadratic sublevel strictly inside that ball. Its interior contains a neighborhood of z ⋆ z^\star z ⋆ , and each compact subset lies below a smaller positive level. These choices give Definition 1, with c 1 = λ min ( P ) c_1=\lambda_{\min}(P) c 1 = λ m i n ( P ) , c 2 = λ max ( P ) c_2=\lambda_{\max}(P) c 2 = λ m a x ( P ) , and c 3 = 1 / 2 c_3=1/2 c 3 = 1/2 . □
For fixed ( H ⋆ , G ⋆ , A ⋆ ) (H_\star,G_\star,A_\star) ( H ⋆ , G ⋆ , A ⋆ ) , conditions (13) and K P ≻ 0 K_P\succ0 K P ≻ 0 form a strict LMI feasibility problem in K P K_P K P , providing a sufficient condition for local exponential stability of the deterministic PI target.
Fix a regional target certificate from Definition 1, with outer sublevel S R 0 = { z ∈ O : V ( z ) ≤ R 0 } S_{R_0}=\{z\in\mathcal{O}:V(z)\le R_0\} S R 0 = { z ∈ O : V ( z ) ≤ R 0 } . For each r 0 < R ≤ R 0 r_0<R\le R_0 r 0 < R ≤ R 0 , define
S R : = { z ∈ O : V ( z ) ≤ R } . S_R:=\{z\in\mathcal{O}:V(z)\le R\}. S R := { z ∈ O : V ( z ) ≤ R } .
This is a closed subset of the compact set S R 0 S_{R_0} S R 0 and is therefore compactly contained in O \mathcal{O} O . Moreover, K ⊂ { z ∈ O : V ( z ) < R } \mathcal{K}\subset\{z\in\mathcal{O}:V(z)<R\} K ⊂ { z ∈ O : V ( z ) < R } .
For the averaged diffusion (12), the time-augmented generator applied to V V V is
( L ‾ σ V ) ( z ) = ∇ V ( z ) T F T ( z ) + σ 2 2 tr ( S 0 ( z ) S 0 ( z ) T D 2 V ( z ) ) . (\overline{\mathcal{L}}_\sigma V)(z)=\nabla V(z)^{\mathsf T}F_T(z)+\frac{\sigma^2}{2}\operatorname{tr}\bigl(S_0(z)S_0(z)^{\mathsf T}D^2V(z)\bigr). ( L σ V ) ( z ) = ∇ V ( z ) T F T ( z ) + 2 σ 2 tr ( S 0 ( z ) S 0 ( z ) T D 2 V ( z ) ) .
The quadratic bounds on V V V from (5) imply ∇ V ( z ⋆ ) = 0 \nabla V(z^\star)=0 ∇ V ( z ⋆ ) = 0 . A local Hessian bound near z ⋆ z^\star z ⋆ , together with compactness away from z ⋆ z^\star z ⋆ , gives a constant L V L_V L V such that
∣ ∇ V ( z ) ∣ ≤ L V ∣ z − z ⋆ ∣ \lvert\nabla V(z)\rvert\le L_V\lvert z-z^\star\rvert ∣ ∇ V ( z )∣ ≤ L V ∣ z − z ⋆ ∣
for z ∈ S R 0 z\in S_{R_0} z ∈ S R 0 . The target certificate and the definition of q σ ( V ) q_\sigma(V) q σ ( V ) consequently give
L ‾ σ V ≤ − c 3 c 2 V + C V σ 2 , \overline{\mathcal{L}}_\sigma V\le-\frac{c_3}{c_2}V+C_V\sigma^2, L σ V ≤ − c 2 c 3 V + C V σ 2 ,
q σ ( V ) ≤ D V σ 2 V , (14) q_\sigma(V)\le D_V\sigma^2V, \tag*{(14)} q σ ( V ) ≤ D V σ 2 V , (14)
throughout S R 0 S_{R_0} S R 0 . One may take
C V : = 1 2 max z ∈ S R 0 ( ∥ S 0 ( z ) ∥ F 2 ∥ D 2 V ( z ) ∥ ) C_V:=\frac{1}{2}\max_{z\in S_{R_0}}\bigl(\lVert S_0(z)\rVert_F^2\lVert D^2V(z)\rVert\bigr) C V := 2 1 z ∈ S R 0 max ( ∥ S 0 ( z ) ∥ F 2 ∥ D 2 V ( z )∥ )
and
D V : = L V 2 c 1 max z ∈ S R 0 ∥ S 0 ( z ) ∥ F 2 . D_V:=\frac{L_V^2}{c_1}\max_{z\in S_{R_0}}\lVert S_0(z)\rVert_F^2. D V := c 1 L V 2 z ∈ S R 0 max ∥ S 0 ( z ) ∥ F 2 .
Both constants are finite and independent of σ \sigma σ and the chosen level R R R .
For Z ‾ 0 = z 0 ∈ K \overline{Z}_0=z_0\in\mathcal{K} Z 0 = z 0 ∈ K , let
τ ‾ R : = inf { t ≥ 0 : V ( Z ‾ t ) ≥ R } \overline{\tau}_R:=\inf\{t\ge0:V(\overline{Z}_t)\ge R\} τ R := inf { t ≥ 0 : V ( Z t ) ≥ R }
be the first exit from { z ∈ O : V ( z ) < R } \{z\in\mathcal{O}:V(z)<R\} { z ∈ O : V ( z ) < R } . Lemma 1 applies to Q t = V ( Z ‾ t ) Q_t=V(\overline{Z}_t) Q t = V ( Z t ) up to τ ‾ R \overline{\tau}_R τ R , using (14). To obtain corresponding estimates for the augmented closed loop (9), we must also account for the oscillations generated by the probing signal.
A periodic change of coordinates removes the leading oscillatory drift. We use the resulting perturbation of F T F_T F T to construct an exact periodic response, then center the target Lyapunov function on that response.
A periodic change of coordinates
Fix a certificate and choose bounded open sets with
S R 0 ⊂ O 1 ⋐ O 2 ⊂ O . S_{R_0}\subset\mathcal{O}_1\Subset\mathcal{O}_2\subset\mathcal{O}. S R 0 ⊂ O 1 ⋐ O 2 ⊂ O .
Choose a smooth cutoff function χ : R n + m → [ 0 , 1 ] \chi:\mathbb{R}^{n+m}\to[0,1] χ : R n + m → [ 0 , 1 ] that equals one on a neighborhood of O ‾ 2 \overline{\mathcal{O}}_2 O 2 and has compact support in O \mathcal{O} O . For each F ∈ { b , S 0 , A i , B i } F\in\{b,S_0,A_i,B_i\} F ∈ { b , S 0 , A i , B i } , define
F ~ ( z ) : = { χ ( z ) F ( z ) , z ∈ O , 0 , z ∉ O . \widetilde{F}(z):=
\begin{cases}
\chi(z)F(z), & z\in\mathcal{O},\\
0, & z\notin\mathcal{O}.
\end{cases} F ( z ) := { χ ( z ) F ( z ) , 0 , z ∈ O , z ∈ / O .
Define the averaged drift of the extended system by
F ~ T ( z ) : = b ~ ( z ) + ∑ i = 1 m [ A ~ i , B ~ i ] ( z ) . \widetilde{F}_T(z):=\widetilde{b}(z)+\sum_{i=1}^{m}[\widetilde{A}_i,\widetilde{B}_i](z). F T ( z ) := b ( z ) + i = 1 ∑ m [ A i , B i ] ( z ) .
Since the cutoff equals one near O ‾ 2 \overline{\mathcal{O}}_2 O 2 , the extended fields and their derivatives agree with the original ones there. In particular,
F ~ T ( z ) = F T ( z ) , z ∈ O 2 , \widetilde{F}_T(z)=F_T(z),\qquad z\in\mathcal{O}_2, F T ( z ) = F T ( z ) , z ∈ O 2 ,
where F T F_T F T is the deterministic target drift defined in (4).
We construct the coordinate transformation using the extended coefficients, whose spatial derivatives are globally bounded. The resulting identities apply to the original system whenever the physical and transformed states remain in O 2 \mathcal{O}_2 O 2 .
For the normal-form construction in this subsection and the auxiliary comparison in the proof of Proposition 3, we suppress the tildes and write b , S 0 , A i , B i b,S_0,A_i,B_i b , S 0 , A i , B i for the compactly supported extensions. Within these calculations, a = ∑ i ( A i p i + B i q i ) a=\sum_i(A_i p_i+B_i q_i) a = ∑ i ( A i p i + B i q i ) and F T = b + ∑ i [ A i , B i ] F_T=b+\sum_i[A_i,B_i] F T = b + ∑ i [ A i , B i ] are formed from the extended fields. Outside these calculations, the symbols retain their original definitions. The two versions agree on O 2 O_2 O 2 , where the cutoff equals one. Thus the identities obtained for the extended equations apply to the original equations whenever the relevant states remain in O 2 O_2 O 2 .
Throughout, we set ε = ω − 1 / 2 \varepsilon=\omega^{-1/2} ε = ω − 1/2 . Then (9) has drift b + ε − 1 a b+\varepsilon^{-1}a b + ε − 1 a and d θ t = ε − 2 d t d\theta_t=\varepsilon^{-2}dt d θ t = ε − 2 d t . We denote the zero-mean periodic corrector as
U 1 ( θ , z ) = ∑ i 2 / κ i [ A i ( z ) sin ( κ i θ ) − B i ( z ) cos ( κ i θ ) ] . U_1(\theta,z)=\sum_i\sqrt{2/\kappa_i}\left[A_i(z)\sin(\kappa_i\theta)-B_i(z)\cos(\kappa_i\theta)\right]. U 1 ( θ , z ) = i ∑ 2/ κ i [ A i ( z ) sin ( κ i θ ) − B i ( z ) cos ( κ i θ ) ] .
It satisfies ∂ θ U 1 = a \partial_\theta U_1=a ∂ θ U 1 = a . Let U 2 U_2 U 2 be the unique zero-mean periodic solution of
∂ θ U 2 = D z a U 1 − ⟨ D z a U 1 ⟩ (15) \partial_\theta U_2=D_za\,U_1-\langle D_za\,U_1\rangle \tag*{(15)} ∂ θ U 2 = D z a U 1 − ⟨ D z a U 1 ⟩ (15)
and define
Γ ε ( θ , ζ ) = ζ + ε U 1 ( θ , ζ ) + ε 2 U 2 ( θ , ζ ) . \Gamma_\varepsilon(\theta,\zeta)=\zeta+\varepsilon U_1(\theta,\zeta)+\varepsilon^2U_2(\theta,\zeta). Γ ε ( θ , ζ ) = ζ + ε U 1 ( θ , ζ ) + ε 2 U 2 ( θ , ζ ) .
All norms in the next lemma are global spatial norms for the extensions, uniform in phase.
Lemma 2. There exist ε 0 , C > 0 \varepsilon_0,C>0 ε 0 , C > 0 such that Γ ε ( θ , ⋅ ) \Gamma_\varepsilon(\theta,\cdot) Γ ε ( θ , ⋅ ) is a diffeomorphism for 0 < ε ≤ ε 0 0<\varepsilon\leq\varepsilon_0 0 < ε ≤ ε 0 and every phase, with
∥ Γ ε − id ∥ C 3 + ∥ Γ ε − 1 − id ∥ C 3 ≤ C ε . (16) \|\Gamma_\varepsilon-\operatorname{id}\|_{C^3}+\|\Gamma_\varepsilon^{-1}-\operatorname{id}\|_{C^3}\leq C\varepsilon. \tag*{(16)} ∥ Γ ε − id ∥ C 3 + ∥ Γ ε − 1 − id ∥ C 3 ≤ C ε . (16) For the extended closed loop, ζ t = Γ ε − 1 ( θ t , Z t ω ) \zeta_t=\Gamma_\varepsilon^{-1}(\theta_t,Z_t^\omega) ζ t = Γ ε − 1 ( θ t , Z t ω ) satisfies
d ζ t = [ F T ( ζ t ) + ε R ε ( θ t , ζ t ) + σ 2 B ε ( θ t , ζ t ) ] d t + σ S ^ ε ( θ t , ζ t ) d B t . (17) d\zeta_t=\left[F_T(\zeta_t)+\varepsilon R_\varepsilon(\theta_t,\zeta_t)+\sigma^2B_\varepsilon(\theta_t,\zeta_t)\right]dt+\sigma\widehat S_\varepsilon(\theta_t,\zeta_t)dB_t. \tag*{(17)} d ζ t = [ F T ( ζ t ) + ε R ε ( θ t , ζ t ) + σ 2 B ε ( θ t , ζ t ) ] d t + σ S ε ( θ t , ζ t ) d B t . (17) The remainder fields are periodic in phase, independent of σ \sigma σ , and satisfy ∥ R ε ∥ C 2 ≤ C \|R_\varepsilon\|_{C^2}\leq C ∥ R ε ∥ C 2 ≤ C and
∥ B ε ∥ C 2 + ∥ S ^ ε − S 0 ∥ C 2 ≤ C ε . (18) \|B_\varepsilon\|_{C^2}+\|\widehat S_\varepsilon-S_0\|_{C^2}\leq C\varepsilon. \tag*{(18)} ∥ B ε ∥ C 2 + ∥ S ε − S 0 ∥ C 2 ≤ C ε . (18) The constants are independent of ε , θ , σ \varepsilon,\theta,\sigma ε , θ , σ . For the original coefficients, (17) holds until Z t ω Z_t^\omega Z t ω exits O 1 O_1 O 1 .
Proof. The probing-signal identities give ⟨ D z a U 1 ⟩ = ∑ i [ A i , B i ] = F T − b \langle D_za\,U_1\rangle=\sum_i[A_i,B_i]=F_T-b ⟨ D z a U 1 ⟩ = ∑ i [ A i , B i ] = F T − b . The smooth extensions give bounded spatial derivatives of U 1 U_1 U 1 through order six and U 2 U_2 U 2 through order five. Put P ε = U 1 + ε U 2 P_\varepsilon=U_1+\varepsilon U_2 P ε = U 1 + ε U 2 and J ε = D ζ Γ ε = I + ε D ζ P ε J_\varepsilon=D_\zeta\Gamma_\varepsilon=I+\varepsilon D_\zeta P_\varepsilon J ε = D ζ Γ ε = I + ε D ζ P ε . Choose ε 0 \varepsilon_0 ε 0 so that ε ∥ D ζ P ε ∥ ∞ ≤ 1 / 2 \varepsilon\|D_\zeta P_\varepsilon\|_\infty\leq1/2 ε ∥ D ζ P ε ∥ ∞ ≤ 1/2 . The equation ζ = z − ε P ε ( θ , ζ ) \zeta=z-\varepsilon P_\varepsilon(\theta,\zeta) ζ = z − ε P ε ( θ , ζ ) is a contraction on R n + m \mathbb{R}^{n+m} R n + m , defining the inverse. Differentiating this identity proves (16) and gives
1 2 ∣ v − w ∣ ≤ ∣ Γ ε ( θ , v ) − Γ ε ( θ , w ) ∣ ≤ 3 2 ∣ v − w ∣ . (19) \frac{1}{2}|v-w|\leq|\Gamma_\varepsilon(\theta,v)-\Gamma_\varepsilon(\theta,w)|\leq\frac{3}{2}|v-w|. \tag*{(19)} 2 1 ∣ v − w ∣ ≤ ∣ Γ ε ( θ , v ) − Γ ε ( θ , w ) ∣ ≤ 2 3 ∣ v − w ∣. (19)
With unmarked arguments evaluated at ( θ , ζ ) (\theta,\zeta) ( θ , ζ ) , Taylor’s formula and (15) give
b ( Γ ε ) + ε − 1 a ( θ , Γ ε ) − ε − 2 ∂ θ Γ ε = F T + ε Q ε , b(\Gamma_\varepsilon)+\varepsilon^{-1}a(\theta,\Gamma_\varepsilon)-\varepsilon^{-2}\partial_\theta\Gamma_\varepsilon=F_T+\varepsilon Q_\varepsilon, b ( Γ ε ) + ε − 1 a ( θ , Γ ε ) − ε − 2 ∂ θ Γ ε = F T + ε Q ε ,
where
Q ε = ∫ 0 1 D b ( ζ + s ε P ε ) P ε d s + D ζ a U 2 + ∫ 0 1 ( 1 − s ) D ζ 2 a ( θ , ζ + s ε P ε ) [ P ε , P ε ] d s . Q_\varepsilon=\int_0^1Db(\zeta+s\varepsilon P_\varepsilon)P_\varepsilon\,ds+D_\zeta a\,U_2+\int_0^1(1-s)D_\zeta^2a(\theta,\zeta+s\varepsilon P_\varepsilon)[P_\varepsilon,P_\varepsilon]\,ds. Q ε = ∫ 0 1 D b ( ζ + s ε P ε ) P ε d s + D ζ a U 2 + ∫ 0 1 ( 1 − s ) D ζ 2 a ( θ , ζ + s ε P ε ) [ P ε , P ε ] d s .
The two corrector equations cancel the ε − 1 \varepsilon^{-1} ε − 1 and order-one oscillatory terms. The integral formula gives a uniform C 2 C^2 C 2 bound for Q ε Q_\varepsilon Q ε . Setting
R ε = J ε − 1 [ Q ε − D ζ P ε F T ] R_\varepsilon=J_\varepsilon^{-1}[Q_\varepsilon-D_\zeta P_\varepsilon F_T] R ε = J ε − 1 [ Q ε − D ζ P ε F T ]
therefore gives J ε − 1 ( F T + ε Q ε ) = F T + ε R ε J_\varepsilon^{-1}(F_T+\varepsilon Q_\varepsilon)=F_T+\varepsilon R_\varepsilon J ε − 1 ( F T + ε Q ε ) = F T + ε R ε and ∥ R ε ∥ C 2 ≤ C \|R_\varepsilon\|_{C^2}\leq C ∥ R ε ∥ C 2 ≤ C .
Itô’s formula for Z = Γ ε ( θ , ζ ) Z=\Gamma_\varepsilon(\theta,\zeta) Z = Γ ε ( θ , ζ ) gives
d Z = ε − 2 ∂ θ Γ ε d t + J ε d ζ + σ 2 2 ∑ j = 1 d B D ζ 2 Γ ε [ S ^ ε , j , S ^ ε , j ] d t . dZ=\varepsilon^{-2}\partial_\theta\Gamma_\varepsilon\,dt+J_\varepsilon d\zeta+\frac{\sigma^2}{2}\sum_{j=1}^{d_B}D_\zeta^2\Gamma_\varepsilon[\widehat S_{\varepsilon,j},\widehat S_{\varepsilon,j}]dt. d Z = ε − 2 ∂ θ Γ ε d t + J ε d ζ + 2 σ 2 j = 1 ∑ d B D ζ 2 Γ ε [ S ε , j , S ε , j ] d t .
Here S ^ ε , j \widehat S_{\varepsilon,j} S ε , j is the j j j th column and D ζ 2 Γ ε [ v , w ] D_\zeta^2\Gamma_\varepsilon[v,w] D ζ 2 Γ ε [ v , w ] is the vector of component-Hessian bilinear forms. Matching diffusion and drift determines
S ^ ε = J ε − 1 S 0 ( Γ ε ) \widehat S_\varepsilon=J_\varepsilon^{-1}S_0(\Gamma_\varepsilon) S ε = J ε − 1 S 0 ( Γ ε )
and
B ε = − 1 2 J ε − 1 ∑ j = 1 d B D ζ 2 Γ ε [ S ^ ε , j , S ^ ε , j ] . B_\varepsilon=-\frac{1}{2}J_\varepsilon^{-1}\sum_{j=1}^{d_B}D_\zeta^2\Gamma_\varepsilon[\widehat S_{\varepsilon,j},\widehat S_{\varepsilon,j}]. B ε = − 2 1 J ε − 1 j = 1 ∑ d B D ζ 2 Γ ε [ S ε , j , S ε , j ] .
Thus σ 2 B ε \sigma^2B_\varepsilon σ 2 B ε is the coordinate-induced Itô correction. The corrector bounds give J ε − 1 − I = O ( ε ) J_\varepsilon^{-1}-I=O(\varepsilon) J ε − 1 − I = O ( ε ) and D ζ 2 Γ ε = O ( ε ) D_\zeta^2\Gamma_\varepsilon=O(\varepsilon) D ζ 2 Γ ε = O ( ε ) in C 2 C^2 C 2 , proving (18). These formulas contain no σ \sigma σ .
Finally, the distance from O ‾ 1 \overline O_1 O 1 to R n + m ∖ O 2 \mathbb{R}^{n+m}\setminus O_2 R n + m ∖ O 2 is positive. We decrease ε 0 \varepsilon_0 ε 0 until the displacement in (16) is smaller than this distance. Both Z t ω Z_t^\omega Z t ω and Γ ε − 1 ( θ t , Z t ω ) \Gamma_\varepsilon^{-1}(\theta_t,Z_t^\omega) Γ ε − 1 ( θ t , Z t ω ) then lie in O 2 O_2 O 2 before exiting O 1 O_1 O 1 , where the extensions and original fields agree. □ \square □
The averaged diffusion (12) retains the plant’s process-noise term while replacing the oscillatory feedback by its effective PI drift. The following proposition compares (12) with the original closed loop (9), up to a common exit time.
Proposition 3. Under Lemma 2, fix K ⊂ U ⊂ O 1 \mathcal K\subset U\subset O_1 K ⊂ U ⊂ O 1 with K \mathcal K K compact and U U U open. Systems (9) and (12) are subject to the same Brownian motion and common deterministic initial state z 0 z_0 z 0 . Set τ ω = τ U ( Z ω ) ∧ τ U ( Z ‾ ) \tau^\omega=\tau_U(Z^\omega)\wedge\tau_U(\overline Z) τ ω = τ U ( Z ω ) ∧ τ U ( Z ) . For every finite T > 0 T>0 T > 0 and σ ‾ ≥ 0 \overline\sigma\geq0 σ ≥ 0 , there exist ω 0 , C K , T , σ ‾ > 0 \omega_0,C_{\mathcal K,T,\overline\sigma}>0 ω 0 , C K , T , σ > 0 such that
sup z 0 ∈ K , θ 0 ∈ T E sup t ≤ T ∧ τ ω ∣ Z t ω − Z ‾ t ∣ 2 ≤ C K , T , σ ‾ / ω (20) \sup_{z_0\in\mathcal K,\theta_0\in\mathbb T}\mathbb E\sup_{t\leq T\wedge\tau^\omega}|Z_t^\omega-\overline Z_t|^2\leq C_{\mathcal K,T,\overline\sigma}/\omega \tag*{(20)} z 0 ∈ K , θ 0 ∈ T sup E t ≤ T ∧ τ ω sup ∣ Z t ω − Z t ∣ 2 ≤ C K , T , σ / ω (20) for ω ≥ ω 0 \omega\geq\omega_0 ω ≥ ω 0 and 0 ≤ σ ≤ σ ‾ 0\leq\sigma\leq\overline\sigma 0 ≤ σ ≤ σ .
Proof. First use the extended versions of (9) and (12), with solutions Z ω , ε Z^{\omega,\varepsilon} Z ω , ε , Z ‾ \overline Z Z . Put ζ t = Γ ε − 1 ( θ t , Z t ω , ε ) \zeta_t=\Gamma_\varepsilon^{-1}(\theta_t,Z_t^{\omega,\varepsilon}) ζ t = Γ ε − 1 ( θ t , Z t ω , ε ) and e t = ζ t − Z ‾ t e_t=\zeta_t-\overline Z_t e t = ζ t − Z t . Subtracting (12) from (17) gives d e t = r t d t + Ξ t d B t de_t=r_tdt+\Xi_tdB_t d e t = r t d t + Ξ t d B t , where
r t = F T ( ζ t ) − F T ( Z ‾ t ) + ε R ε ( θ t , ζ t ) + σ 2 B ε ( θ t , ζ t ) r_t=F_T(\zeta_t)-F_T(\overline Z_t)+\varepsilon R_\varepsilon(\theta_t,\zeta_t)+\sigma^2B_\varepsilon(\theta_t,\zeta_t) r t = F T ( ζ t ) − F T ( Z t ) + ε R ε ( θ t , ζ t ) + σ 2 B ε ( θ t , ζ t )
and
Ξ t = σ [ S ^ ε ( θ t , ζ t ) − S 0 ( Z ‾ t ) ] . \Xi_t=\sigma[\widehat S_\varepsilon(\theta_t,\zeta_t)-S_0(\overline Z_t)]. Ξ t = σ [ S ε ( θ t , ζ t ) − S 0 ( Z t )] .
The extended F T , S 0 F_T,S_0 F T , S 0 are globally Lipschitz, so Lemma 2 gives
∣ e 0 ∣ 2 ≤ C ε 2 , ∣ r t ∣ 2 + ∥ Ξ t ∥ F 2 ≤ C σ ‾ ( ∣ e t ∣ 2 + ε 2 ) . |e_0|^2\leq C\varepsilon^2,\qquad|r_t|^2+\|\Xi_t\|_F^2\leq C_{\overline\sigma}(|e_t|^2+\varepsilon^2). ∣ e 0 ∣ 2 ≤ C ε 2 , ∣ r t ∣ 2 + ∥ Ξ t ∥ F 2 ≤ C σ ( ∣ e t ∣ 2 + ε 2 ) .
The integral error equation, Cauchy–Schwarz, and Doob’s inequality with Itô’s isometry imply, for E ( t ) = E sup s ≤ t ∣ e s ∣ 2 \mathcal E(t)=\mathbb E\sup_{s\leq t}|e_s|^2 E ( t ) = E sup s ≤ t ∣ e s ∣ 2 and t ≤ T t\leq T t ≤ T ,
E ( t ) ≤ 3 ∣ e 0 ∣ 2 + 3 T ∫ 0 t E ∣ r s ∣ 2 d s + 12 ∫ 0 t E ∥ Ξ s ∥ F 2 d s ≤ C T , σ ‾ ε 2 + C T , σ ‾ ∫ 0 t E ( s ) d s . (21) \begin{aligned}
\mathcal E(t)\leq3|e_0|^2+3T\int_0^t\mathbb E|r_s|^2\,ds+12\int_0^t\mathbb E\|\Xi_s\|_F^2\,ds \\
&\leq C_{T,\overline\sigma}\varepsilon^2+C_{T,\overline\sigma}\int_0^t\mathcal E(s)\,ds. \tag*{(21)}
\end{aligned} E ( t ) ≤ 3∣ e 0 ∣ 2 + 3 T ∫ 0 t E ∣ r s ∣ 2 d s + 12 ∫ 0 t E ∥ Ξ s ∥ F 2 d s ≤ C T , σ ε 2 + C T , σ ∫ 0 t E ( s ) d s . (21)
Grönwall gives E ( T ) ≤ C T , δ ε 2 \mathcal{E}(T) \le C T,\delta\varepsilon^2 E ( T ) ≤ C T , δ ε 2 . The identity Z ω , e − Z ˉ e = Γ ε ( θ , ζ ) − ζ + e Z^{\omega,e}-\bar{Z}^{e}=\Gamma_{\varepsilon}(\theta,\zeta)-\zeta+e Z ω , e − Z ˉ e = Γ ε ( θ , ζ ) − ζ + e and (16) give the same order in physical coordinates. The extended and original solutions agree through τ ω \tau^\omega τ ω by pathwise uniqueness and continuity. Restriction to this interval and ε 2 = ω − 1 \varepsilon^2=\omega^{-1} ε 2 = ω − 1 prove the result.
Periodic response of the noise-free closed loop
We construct a periodic trajectory near z ⋆ z^\star z ⋆ for fixed frequency. It is the reference for the stochastic estimates, rather than an equilibrium of the time-periodic physical system.
Lemma 3. For a certified target and sufficiently small ε \varepsilon ε , there is a unique 2 π 2\pi 2 π -periodic C 1 C^1 C 1 function χ ε \chi_\varepsilon χ ε in a fixed sufficiently small supremum-norm neighborhood of zero satisfying
ω d χ ε d θ = F T ( z ⋆ + χ ε ) + ε R ε ( θ , z ⋆ + χ ε ) . (22) \omega\frac{d\chi_\varepsilon}{d\theta}=F_T(z^\star+\chi_\varepsilon)+\varepsilon R_\varepsilon(\theta,z^\star+\chi_\varepsilon). \tag*{(22)} ω d θ d χ ε = F T ( z ⋆ + χ ε ) + ε R ε ( θ , z ⋆ + χ ε ) . (22)
It obeys $|\chi_\varepsilon|_\infty\le C\varepsilon). The physical response
z ω ( θ ) = Γ ε ( θ , z ⋆ + χ ε ( θ ) ) (23) z^\omega(\theta)=\Gamma_\varepsilon(\theta,z^\star+\chi_\varepsilon(\theta)) \tag*{(23)} z ω ( θ ) = Γ ε ( θ , z ⋆ + χ ε ( θ )) (23)
satisfies sup θ ∣ z ω ( θ ) − z ⋆ ∣ ≤ C ε \sup_\theta|z^\omega(\theta)-z^\star|\le C\varepsilon sup θ ∣ z ω ( θ ) − z ⋆ ∣ ≤ C ε . For every initial phase, t ↦ z ω ( θ 0 + ω t ) t\mapsto z^\omega(\theta_0+\omega t) t ↦ z ω ( θ 0 + ω t ) is a 2 π / ω 2\pi/\omega 2 π / ω -periodic solution of (9) with σ = 0 \sigma=0 σ = 0 .
Proof. Expanding (5)–(6) at z ⋆ z^\star z ⋆ gives P = 1 2 D 2 V ( z ⋆ ) ≻ 0 P=\frac{1}{2}D^2V(z^\star)\succ0 P = 2 1 D 2 V ( z ⋆ ) ≻ 0 and A T T P + P A T ⪯ − c 3 I A_T^{\mathsf T}P+PA_T\preceq-c_3I A T T P + P A T ⪯ − c 3 I , where A T = D F T ( z ⋆ ) A_T=DF_T(z^\star) A T = D F T ( z ⋆ ) . Thus ∥ e A T t ∥ ≤ M A e − α A t \|e^{A_Tt}\|\le M_Ae^{-\alpha_At} ∥ e A T t ∥ ≤ M A e − α A t . We define N ( e ) = F T ( z ⋆ + e ) − A T e N(e)=F_T(z^\star+e)-A_Te N ( e ) = F T ( z ⋆ + e ) − A T e . Locally, ∣ N ( e ) ∣ ≤ L N ∣ e ∣ 2 |N(e)|\le L_N|e|^2 ∣ N ( e ) ∣ ≤ L N ∣ e ∣ 2 and ∥ D N ( e ) ∥ ≤ 2 L N ∣ e ∣ \|DN(e)\|\le2L_N|e| ∥ D N ( e ) ∥ ≤ 2 L N ∣ e ∣ . Let L R L_R L R bound R ε R_\varepsilon R ε and its first spatial derivative, uniformly in ε \varepsilon ε .
On continuous periodic functions, we define
( T ε χ ) ( θ ) = ∫ 0 ∞ e A T s [ N ( χ ( θ − ω s ) ) + ε R ε ( θ − ω s , z ⋆ + χ ( θ − ω s ) ) ] d s . (\mathcal{T}_\varepsilon\chi)(\theta)=\int_0^\infty e^{A_Ts}\left[N(\chi(\theta-\omega s))+\varepsilon R_\varepsilon(\theta-\omega s,z^\star+\chi(\theta-\omega s))\right]\,ds. ( T ε χ ) ( θ ) = ∫ 0 ∞ e A T s [ N ( χ ( θ − ω s )) + ε R ε ( θ − ω s , z ⋆ + χ ( θ − ω s )) ] d s .
Writing c A = M A / α A c_A=M_A/\alpha_A c A = M A / α A , we choose ρ ⋆ > 0 \rho_\star>0 ρ ⋆ > 0 with B 2 ρ ⋆ ( z ⋆ ) ⊂ O 1 B_{2\rho_\star}(z^\star)\subset\mathcal{O}_1 B 2 ρ ⋆ ( z ⋆ ) ⊂ O 1 and 2 c A L N ρ ⋆ ≤ 1 / 4 2c_AL_N\rho_\star\le1/4 2 c A L N ρ ⋆ ≤ 1/4 . We can then take c A ε L R ≤ min { 1 / 4 , ρ ⋆ / 2 } c_A\varepsilon L_R\le\min\{1/4,\rho_\star/2\} c A ε L R ≤ min { 1/4 , ρ ⋆ /2 } . On the closed ball ∥ χ ∥ ∞ ≤ ρ ⋆ \|\chi\|_\infty\le\rho_\star ∥ χ ∥ ∞ ≤ ρ ⋆ , the map has a contraction constant of at most c A ( 2 L N ρ ⋆ + ε L R ) ≤ 1 / 2 c_A(2L_N\rho_\star+\varepsilon L_R)\le1/2 c A ( 2 L N ρ ⋆ + ε L R ) ≤ 1/2 and ∥ T ε 0 ∥ ∞ ≤ c A ε L R ≤ ρ ⋆ / 2 \|\mathcal{T}_\varepsilon0\|_\infty\le c_A\varepsilon L_R\le\rho_\star/2 ∥ T ε 0 ∥ ∞ ≤ c A ε L R ≤ ρ ⋆ /2 . It maps the ball into itself, giving a unique fixed point and ∥ χ ε ∥ ∞ ≤ 2 c A L R ε \|\chi_\varepsilon\|_\infty\le2c_AL_R\varepsilon ∥ χ ε ∥ ∞ ≤ 2 c A L R ε .
For q ( t ) = χ ε ( θ 0 + ω t ) q(t)=\chi_\varepsilon(\theta_0+\omega t) q ( t ) = χ ε ( θ 0 + ω t ) , the fixed-point identity is
q ( t ) = ∫ − ∞ t e A T ( t − s ) [ N ( q ( s ) ) + ε R ε ( θ 0 + ω s , z ⋆ + q ( s ) ) ] d s . q(t)=\int_{-\infty}^{t}e^{A_T(t-s)}[N(q(s))+\varepsilon R_\varepsilon(\theta_0+\omega s,z^\star+q(s))]\,ds. q ( t ) = ∫ − ∞ t e A T ( t − s ) [ N ( q ( s )) + ε R ε ( θ 0 + ω s , z ⋆ + q ( s ))] d s .
The integrand is continuous and bounded after its exponential factor is removed. Differentiation gives (21). Conversely, variation of constants for any bounded complete periodic solution, followed by resetting the starting time to − ∞ -\infty − ∞ , gives this same identity. This observation proves uniqueness in the fixed ball. Phase translations preserve the supremum norm, so the estimates are uniform in frequency. The displacement bounds keep the response in O 1 \mathcal{O}_1 O 1 , and the exact transformation with σ = 0 \sigma=0 σ = 0 proves the assertion for (9). Periodicity follows from that of Γ ε , χ ε \Gamma_\varepsilon,\chi_\varepsilon Γ ε , χ ε .
Centering on the periodic response
For σ = 0 \sigma=0 σ = 0 , write Z t ω , 0 Z_t^{\omega,0} Z t ω , 0 for the solution of (9) from state z 0 z_0 z 0 and phase θ 0 \theta_0 θ 0 . Define
Ψ ω ( θ , z ) = Γ ε − 1 ( θ , z ) − χ ε ( θ ) . \Psi_\omega(\theta,z)=\Gamma_\varepsilon^{-1}(\theta,z)-\chi_\varepsilon(\theta). Ψ ω ( θ , z ) = Γ ε − 1 ( θ , z ) − χ ε ( θ ) .
It maps z ω ( θ ) z^\omega(\theta) z ω ( θ ) to z ⋆ z^\star z ⋆ . Along (9) with σ = 0 \sigma=0 σ = 0 , y t = Ψ ω ( θ t , Z t ω , 0 ) y_t=\Psi_\omega(\theta_t,Z_t^{\omega,0}) y t = Ψ ω ( θ t , Z t ω , 0 ) satisfies
d y t = [ F T ( y t ) + E ε ( θ t , y t ) ] d t , (24) dy_t=[F_T(y_t)+E_\varepsilon(\theta_t,y_t)]dt, \tag*{(24)} d y t = [ F T ( y t ) + E ε ( θ t , y t )] d t , (24)
where
E ε ( θ , y ) = F T ( y + χ ε ( θ ) ) − F T ( y ) − F T ( z ⋆ + χ ε ( θ ) ) + F T ( z ⋆ ) + ε R ε ( θ , y + χ ε ( θ ) ) − ε R ε ( θ , z ⋆ + χ ε ( θ ) ) . E_\varepsilon(\theta,y)=F_T(y+\chi_\varepsilon(\theta))-F_T(y)
-F_T(z^\star+\chi_\varepsilon(\theta))+F_T(z^\star)
+\varepsilon R_\varepsilon(\theta,y+\chi_\varepsilon(\theta))
-\varepsilon R_\varepsilon(\theta,z^\star+\chi_\varepsilon(\theta)). E ε ( θ , y ) = F T ( y + χ ε ( θ )) − F T ( y ) − F T ( z ⋆ + χ ε ( θ )) + F T ( z ⋆ ) + ε R ε ( θ , y + χ ε ( θ )) − ε R ε ( θ , z ⋆ + χ ε ( θ )) .
The extended F T F_T F T has bounded second derivatives and R ε R_\varepsilon R ε has uniformly bounded first derivatives. Therefore ∥ D y E ε ∥ ≤ C ε \|D_yE_\varepsilon\|\le C\varepsilon ∥ D y E ε ∥ ≤ C ε . Since E ε ( θ , z ⋆ ) = 0 E_\varepsilon(\theta,z^\star)=0 E ε ( θ , z ⋆ ) = 0 , integration along the segment from z ⋆ z^\star z ⋆ to y y y gives
∣ E ε ( θ , y ) ∣ ≤ C ε ∣ y − z ⋆ ∣ . (25) |E_\varepsilon(\theta,y)|\le C\varepsilon|y-z^\star|. \tag*{(25)} ∣ E ε ( θ , y ) ∣ ≤ C ε ∣ y − z ⋆ ∣. (25)
Set
Φ ω ( θ , z ) = V ( Ψ ω ( θ , z ) ) . (26) \Phi_\omega(\theta,z)=V(\Psi_\omega(\theta,z)). \tag*{(26)} Φ ω ( θ , z ) = V ( Ψ ω ( θ , z )) . (26)
The uniform displacement bounds ensure Ψ ω ( θ , O 1 ) ⊂ O 2 \Psi_\omega(\theta,\mathcal{O}_1)\subset\mathcal{O}_2 Ψ ω ( θ , O 1 ) ⊂ O 2 for small ε \varepsilon ε , so this function is defined on a neighborhood of T × O ‾ 1 \mathbb{T}\times\overline{\mathcal{O}}_1 T × O 1 . For 0 < s ≤ R 0 0<s\le R_0 0 < s ≤ R 0 , put S s ⋆ = { y ∈ O : V ( y ) < s } \mathcal{S}_s^\star=\{y\in\mathcal{O}:V(y)<s\} S s ⋆ = { y ∈ O : V ( y ) < s } and define
D s ω ( θ ) = { Γ ε ( θ , y + χ ε ( θ ) ) : y ∈ S s ⋆ } . (27) \mathcal{D}_s^\omega(\theta)=\{\Gamma_\varepsilon(\theta,y+\chi_\varepsilon(\theta)):y\in\mathcal{S}_s^\star\}. \tag*{(27)} D s ω ( θ ) = { Γ ε ( θ , y + χ ε ( θ )) : y ∈ S s ⋆ } . (27)
Their closures lie in a fixed compact subset of O 1 \mathcal{O}_1 O 1 for small ε \varepsilon ε . The defining diffeomorphism maps sublevel boundaries to boundaries, where Φ ω = s \Phi_\omega=s Φ ω = s .
Proposition 4. Fix r 0 < r < R 0 r_0<r<R_0 r 0 < r < R 0 . For sufficiently large ω \omega ω , with λ = c 3 / ( 2 c 2 ) \lambda=c_3/(2c_2) λ = c 3 / ( 2 c 2 ) ,
A ω , 0 Φ ω ≤ − λ Φ ω (28) \mathcal{A}_{\omega,0}\Phi_\omega\le-\lambda\Phi_\omega \tag*{(28)} A ω , 0 Φ ω ≤ − λ Φ ω (28)
on D R 0 ω ( θ ) \mathcal{D}_{R_0}^\omega(\theta) D R 0 ω ( θ ) , and
a 1 ∣ z − z ω ( θ ) ∣ 2 ≤ Φ ω ( θ , z ) ≤ a 2 ∣ z − z ω ( θ ) ∣ 2 , (29) a_1|z-z^\omega(\theta)|^2\le\Phi_\omega(\theta,z)\le a_2|z-z^\omega(\theta)|^2, \tag*{(29)} a 1 ∣ z − z ω ( θ ) ∣ 2 ≤ Φ ω ( θ , z ) ≤ a 2 ∣ z − z ω ( θ ) ∣ 2 , (29)
*with a 1 = 4 c 1 / 9 a_1=4c_1/9 a 1 = 4 c 1 /9 and a 2 = 4 c 2 a_2=4c_2 a 2 = 4 c 2 . Every z 0 ∈ D r ω ( θ 0 ) z_0\in\mathcal{D}_r^\omega(\theta_0) z 0 ∈ D r ω ( θ 0 ) , 0 < s ≤ R 0 0<s\le R_0 0 < s ≤ R 0 , generates a complete noise-free solution that remains in D s ω ( θ t ) \mathcal{D}_s^\omega(\theta_t) D s ω ( θ t ) and satisfies
∣ Z t ω , 0 − z ω ( θ t ) ∣ 2 ≤ a 2 a 1 e − λ t ∣ z 0 − z ω ( θ 0 ) ∣ 2 . (30) |Z_t^{\omega,0}-z^\omega(\theta_t)|^2\le\frac{a_2}{a_1}e^{-\lambda t}|z_0-z^\omega(\theta_0)|^2. \tag*{(30)} ∣ Z t ω , 0 − z ω ( θ t ) ∣ 2 ≤ a 1 a 2 e − λ t ∣ z 0 − z ω ( θ 0 ) ∣ 2 . (30)
Moreover, K ⊂ D r ω ( θ ) \mathcal{K}\subset\mathcal{D}_r^\omega(\theta) K ⊂ D r ω ( θ ) for every phase. The constants and threshold are uniform over the phase and initial state.
Proof. On S R 0 \mathcal{S}_{R_0} S R 0 , ∣ ∇ V ( y ) ∣ ≤ L V ∣ y − z ⋆ ∣ |\nabla V(y)|\le L_V|y-z^\star| ∣∇ V ( y ) ∣ ≤ L V ∣ y − z ⋆ ∣ . Equations (6), (23), and (24) give
A ω , 0 Φ ω ≤ − ( c 3 − C L V ε ) ∣ y − z ⋆ ∣ 2 ≤ − λ V ( y ) \mathcal{A}_{\omega,0}\Phi_\omega\le-(c_3-C L_V\varepsilon)|y-z^\star|^2\le-\lambda V(y) A ω , 0 Φ ω ≤ − ( c 3 − C L V ε ) ∣ y − z ⋆ ∣ 2 ≤ − λV ( y )
for C L V ε ≤ c 3 / 2 CL_V\varepsilon\le c_3/2 C L V ε ≤ c 3 /2 . Applying (19) to y + χ ε y+\chi_\varepsilon y + χ ε and z ⋆ + χ ε z^\star+\chi_\varepsilon z ⋆ + χ ε , then using (5), proves (28). Until first exit from D R 0 ω ( θ t ) \mathcal{D}_{R_0}^\omega(\theta_t) D R 0 ω ( θ t ) , (27) gives Φ ω ( θ t , Z t ω , 0 ) ≤ e − λ t Φ ω ( θ 0 , z 0 ) < s \Phi_\omega(\theta_t,Z_t^{\omega,0})\le e^{-\lambda t}\Phi_\omega(\theta_0,z_0)<s Φ ω ( θ t , Z t ω , 0 ) ≤ e − λ t Φ ω ( θ 0 , z 0 ) < s . At a finite exit, continuity would give Φ ω = s \Phi_\omega=s Φ ω = s , which is a contradiction. The fixed compact containment permits continuation for all time. Applying (28) at the initial and current states yields (29).
Finally, (16) and the bound on χ ε \chi_{\varepsilon} χ ε give
∣ Ψ ω ( θ , z ) − z ∣ ≤ C Ψ ε \lvert\Psi_{\omega}(\theta,z)-z\rvert\le C_{\Psi}\varepsilon ∣ Ψ ω ( θ , z ) − z ∣ ≤ C Ψ ε
for ( θ , z ) ∈ T × K (\theta,z)\in\mathbb{T}\times\mathcal{K} ( θ , z ) ∈ T × K with C Ψ C_{\Psi} C Ψ independent of ε \varepsilon ε and θ \theta θ . We choose δ > 0 \delta>0 δ > 0 such that the closed neighborhood K δ : = { y : dist ( y , K ) ≤ δ } \mathcal{K}_{\delta}:=\{y:\operatorname{dist}(y,\mathcal{K})\le\delta\} K δ := { y : dist ( y , K ) ≤ δ } is contained in O \mathcal{O} O , and let L K L_{\mathcal{K}} L K bound ∣ ∇ V ∣ \lvert\nabla V\rvert ∣ ∇ V ∣ there. Choose ε 1 > 0 \varepsilon_{1}>0 ε 1 > 0 sufficiently small such that C Ψ ε 1 < δ C_{\Psi}\varepsilon_{1}<\delta C Ψ ε 1 < δ and L K C Ψ ε 1 < r − r 0 L_{\mathcal{K}}C_{\Psi}\varepsilon_{1}<r-r_{0} L K C Ψ ε 1 < r − r 0 .
For z ∈ K z\in\mathcal{K} z ∈ K , θ ∈ T \theta\in\mathbb{T} θ ∈ T , and 0 < ε ≤ ε 1 0<\varepsilon\le\varepsilon_{1} 0 < ε ≤ ε 1 , the point y = Ψ ω ( θ , z ) y=\Psi_{\omega}(\theta,z) y = Ψ ω ( θ , z ) and the segment joining it to z z z lie in K δ ⊂ O \mathcal{K}_{\delta}\subset\mathcal{O} K δ ⊂ O . The gradient bound therefore gives
Φ ω ( θ , z ) = V ( y ) ≤ V ( z ) + L K ∣ y − z ∣ ≤ r 0 + L K C Ψ ε < r . \begin{aligned}
\Phi_{\omega}(\theta,z)&=V(y)\le V(z)+L_{\mathcal{K}}\lvert y-z\rvert\\
&\le r_{0}+L_{\mathcal{K}}C_{\Psi}\varepsilon<r.
\end{aligned} Φ ω ( θ , z ) = V ( y ) ≤ V ( z ) + L K ∣ y − z ∣ ≤ r 0 + L K C Ψ ε < r .
Thus y ∈ S r < y\in S_{r}^{<} y ∈ S r < . Since z = Γ ε ( θ , y + χ ε ( θ ) ) z=\Gamma_{\varepsilon}(\theta,y+\chi_{\varepsilon}(\theta)) z = Γ ε ( θ , y + χ ε ( θ )) , the definition (26) yields K ⊂ D R ω ( θ ) \mathcal{K}\subset D_{R}^{\omega}(\theta) K ⊂ D R ω ( θ ) for every θ ∈ T \theta\in\mathbb{T} θ ∈ T and 0 < ε ≤ ε 1 0<\varepsilon\le\varepsilon_{1} 0 < ε ≤ ε 1 . □
Stochastic Regulation and Long-Horizon Approximation
The periodic function Φ ω \Phi_{\omega} Φ ω transfers the deterministic decay estimate to the physical diffusion. We first obtain regulation and capture estimates under the target certificate. A separate incremental condition then provides a time-uniform comparison with the bracket diffusion.
Regulation about the periodic response
For (9), define the exit time
τ R ω = inf { t ≥ 0 : Z t ω ∉ D R ω ( θ t ) } , \tau_{R}^{\omega}=\inf\{t\ge0:Z_{t}^{\omega}\notin D_{R}^{\omega}(\theta_{t})\}, τ R ω = inf { t ≥ 0 : Z t ω ∈ / D R ω ( θ t )} ,
where the sublevels are defined in (26).
Theorem 1. Let Assumptions 1–3 hold. Fix a certificate for K \mathcal{K} K , levels r 0 < r < R < R 0 r_{0}<r<R<R_{0} r 0 < r < R < R 0 , and a finite noise bound σ ˉ > 0 \bar{\sigma}>0 σ ˉ > 0 . There exists ω K > 0 \omega_{\mathcal{K}}>0 ω K > 0 such that, for every ω ≥ ω K \omega\ge\omega_{\mathcal{K}} ω ≥ ω K , the closed loop (9) satisfies
E [ ∣ Z t ω − z ⋆ ∣ 2 1 { t < τ R ω } ] ≤ C e − λ t ∣ z 0 − z ⋆ ∣ 2 + C ( σ 2 + ω − 1 ) (31) \mathbb{E}\left[\lvert Z_{t}^{\omega}-z^{\star}\rvert^{2}\mathbf{1}_{\{t<\tau_{R}^{\omega}\}}\right]\le Ce^{-\lambda t}\lvert z_{0}-z^{\star}\rvert^{2}+C(\sigma^{2}+\omega^{-1}) \tag*{(31)} E [ ∣ Z t ω − z ⋆ ∣ 2 1 { t < τ R ω } ] ≤ C e − λ t ∣ z 0 − z ⋆ ∣ 2 + C ( σ 2 + ω − 1 ) (31) where λ = c 3 / ( 2 c 2 ) \lambda=c_{3}/(2c_{2}) λ = c 3 / ( 2 c 2 ) . Relative to the response (22),
E [ ∣ Z t ω − z ω ⋆ ( θ t ) ∣ 2 1 { t < τ R ω } ] ≤ M e − λ t ∣ z 0 − z ω ⋆ ( θ 0 ) ∣ 2 + C σ 2 ( 1 − e − λ t ) . (32) \mathbb{E}\left[\lvert Z_{t}^{\omega}-z_{\omega}^{\star}(\theta_{t})\rvert^{2}\mathbf{1}_{\{t<\tau_{R}^{\omega}\}}\right]\le Me^{-\lambda t}\lvert z_{0}-z_{\omega}^{\star}(\theta_{0})\rvert^{2}+C\sigma^{2}(1-e^{-\lambda t}). \tag*{(32)} E [ ∣ Z t ω − z ω ⋆ ( θ t ) ∣ 2 1 { t < τ R ω } ] ≤ M e − λ t ∣ z 0 − z ω ⋆ ( θ 0 ) ∣ 2 + C σ 2 ( 1 − e − λ t ) . (32) For σ > 0 \sigma>0 σ > 0 and finite T ≥ 0 T\ge0 T ≥ 0 ,
P { τ R ω ≤ T } ≤ e − η ( R − r ) / σ 2 + C 0 T e − η R / σ 2 . (33) \mathbb{P}\{\tau_{R}^{\omega}\le T\}\le e^{-\eta(R-r)/\sigma^{2}}+C_{0}Te^{-\eta R/\sigma^{2}}. \tag*{(33)} P { τ R ω ≤ T } ≤ e − η ( R − r ) / σ 2 + C 0 T e − η R / σ 2 . (33) The positive constants and frequency threshold are uniform for z 0 ∈ K z_{0}\in\mathcal{K} z 0 ∈ K , θ 0 ∈ T \theta_{0}\in\mathbb{T} θ 0 ∈ T , 0 ≤ σ ≤ σ ˉ 0\le\sigma\le\bar{\sigma} 0 ≤ σ ≤ σ ˉ , and t ≥ 0 t\ge0 t ≥ 0 .
Proof. Proposition 4 gives Φ ω ( θ 0 , z 0 ) < r \Phi_{\omega}(\theta_{0},z_{0})<r Φ ω ( θ 0 , z 0 ) < r and (27). On D R ω ( θ ) D_{R}^{\omega}(\theta) D R ω ( θ ) , write y = Ψ ω ( θ , z ) y=\Psi_{\omega}(\theta,z) y = Ψ ω ( θ , z ) . Section VI-C places z z z in a fixed compact subset of O 1 \mathcal{O}_{1} O 1 and y y y in S R 0 S_{R_{0}} S R 0 .
Using (16), we choose uniform bounds ∥ D z Ψ ω ∥ ≤ J Ψ \lVert D_{z}\Psi_{\omega}\rVert\le J_{\Psi} ∥ D z Ψ ω ∥ ≤ J Ψ and ∥ D z 2 Ψ ω ∥ ≤ H Ψ \lVert D_{z}^{2}\Psi_{\omega}\rVert\le H_{\Psi} ∥ D z 2 Ψ ω ∥ ≤ H Ψ . Let G V , H V G_{V},H_{V} G V , H V bound ∣ ∇ V ∣ , ∥ D 2 V ∥ \lvert\nabla V\rvert,\lVert D^{2}V\rVert ∣ ∇ V ∣ , ∥ D 2 V ∥ on S R 0 S_{R_{0}} S R 0 , and let S ∗ = max O 1 ∥ S 0 ∥ F S_{\ast}=\max_{\mathcal{O}_{1}}\lVert S_{0}\rVert_{F} S ∗ = max O 1 ∥ S 0 ∥ F . The chain rule applied to (25) gives
∇ z Φ ω = ( D z Ψ ω ) T ∇ V ( y ) , D z 2 Φ ω [ v , w ] = D 2 V ( y ) [ D z Ψ ω v , D z Ψ ω w ] + ∇ V ( y ) D z 2 Ψ ω [ v , w ] . \begin{aligned}
\nabla_{z}\Phi_{\omega}&=(D_{z}\Psi_{\omega})^{\mathsf{T}}\nabla V(y),\\
D_{z}^{2}\Phi_{\omega}[v,w]&=D^{2}V(y)[D_{z}\Psi_{\omega}v,D_{z}\Psi_{\omega}w]
+\nabla V(y)D_{z}^{2}\Psi_{\omega}[v,w].
\end{aligned} ∇ z Φ ω D z 2 Φ ω [ v , w ] = ( D z Ψ ω ) T ∇ V ( y ) , = D 2 V ( y ) [ D z Ψ ω v , D z Ψ ω w ] + ∇ V ( y ) D z 2 Ψ ω [ v , w ] .
Thus ∥ D z 2 Φ ω ∥ ≤ H Φ : = J Ψ 2 H V + G V H Ψ \lVert D_{z}^{2}\Phi_{\omega}\rVert\le H_{\Phi}:=J_{\Psi}^{2}H_{V}+G_{V}H_{\Psi} ∥ D z 2 Φ ω ∥ ≤ H Φ := J Ψ 2 H V + G V H Ψ . (10) yields
A ω , σ Φ ω = A ω , 0 Φ ω + 1 2 σ 2 tr ( S 0 S 0 T D z 2 Φ ω ) ≤ − λ Φ ω + c 0 σ 2 , (34) \begin{aligned}
\mathcal{A}_{\omega,\sigma}\Phi_{\omega}
&=\mathcal{A}_{\omega,0}\Phi_{\omega}
+\frac{1}{2}\sigma^{2}\operatorname{tr}(S_{0}S_{0}^{\mathsf{T}}D_{z}^{2}\Phi_{\omega})\\
&\le-\lambda\Phi_{\omega}+c_{0}\sigma^{2}, \tag*{(34)}
\end{aligned} A ω , σ Φ ω = A ω , 0 Φ ω + 2 1 σ 2 tr ( S 0 S 0 T D z 2 Φ ω ) ≤ − λ Φ ω + c 0 σ 2 , (34)
where c 0 = S ∗ 2 H Φ / 2 c_{0}=S_{\ast}^{2}H_{\Phi}/2 c 0 = S ∗ 2 H Φ /2 . The estimate ∣ ∇ V ( y ) ∣ ≤ L V ∣ y − z ⋆ ∣ \lvert\nabla V(y)\rvert\le L_{V}\lvert y-z^{\star}\rvert ∣ ∇ V ( y )∣ ≤ L V ∣ y − z ⋆ ∣ and (5) similarly give
σ 2 ∣ S 0 T ∇ z Φ ω ∣ 2 ≤ d 0 σ 2 Φ ω (35) \sigma^{2}\lvert S_{0}^{\mathsf{T}}\nabla_{z}\Phi_{\omega}\rvert^{2}\le d_{0}\sigma^{2}\Phi_{\omega} \tag*{(35)} σ 2 ∣ S 0 T ∇ z Φ ω ∣ 2 ≤ d 0 σ 2 Φ ω (35)
where d 0 = S ∗ 2 J Ψ 2 L V 2 / c 1 d_{0}=S_{\ast}^{2}J_{\Psi}^{2}L_{V}^{2}/c_{1} d 0 = S ∗ 2 J Ψ 2 L V 2 / c 1 . All bounds are independent of phase, σ \sigma σ , and sufficiently large ω \omega ω .
The compact containment and boundary value Φ ω = R \Phi_{\omega}=R Φ ω = R allow us to apply Lemma 1 to ( t , z ) ↦ Φ ω ( θ 0 + ω t , z ) (t,z)\mapsto\Phi_{\omega}(\theta_{0}+\omega t,z) ( t , z ) ↦ Φ ω ( θ 0 + ω t , z ) . This gives
E [ Φ ω ( θ t , Z t ω ) 1 { t < τ R ω } ] ≤ e − λ t Φ ω ( θ 0 , z 0 ) + c 0 σ 2 λ ( 1 − e − λ t ) . (36) \mathbb{E}\left[\Phi_{\omega}(\theta_{t},Z_{t}^{\omega})\mathbf{1}_{\{t<\tau_{R}^{\omega}\}}\right]
\le e^{-\lambda t}\Phi_{\omega}(\theta_{0},z_{0})+\frac{c_{0}\sigma^{2}}{\lambda}(1-e^{-\lambda t}). \tag*{(36)} E [ Φ ω ( θ t , Z t ω ) 1 { t < τ R ω } ] ≤ e − λ t Φ ω ( θ 0 , z 0 ) + λ c 0 σ 2 ( 1 − e − λ t ) . (36)
The lower bound in (28) and (35) give
E [ ∣ Z t ω − z ω ⋆ ( θ t ) ∣ 2 1 { t < τ R ω } ] ≤ e − λ t a 1 Φ ω ( θ 0 , z 0 ) + c 0 σ 2 λ a 1 ( 1 − e − λ t ) . \mathbb{E}\left[\lvert Z_{t}^{\omega}-z_{\omega}^{\star}(\theta_{t})\rvert^{2}\mathbf{1}_{\{t<\tau_{R}^{\omega}\}}\right]
\le\frac{e^{-\lambda t}}{a_{1}}\Phi_{\omega}(\theta_{0},z_{0})+\frac{c_{0}\sigma^{2}}{\lambda a_{1}}(1-e^{-\lambda t}). E [ ∣ Z t ω − z ω ⋆ ( θ t ) ∣ 2 1 { t < τ R ω } ] ≤ a 1 e − λ t Φ ω ( θ 0 , z 0 ) + λ a 1 c 0 σ 2 ( 1 − e − λ t ) .
At the initial state, the upper bound gives Φ ω ( θ 0 , z 0 ) ≤ a 2 ∣ z 0 − z ω ⋆ ( θ 0 ) ∣ 2 \Phi_{\omega}(\theta_{0},z_{0})\le a_{2}\lvert z_{0}-z_{\omega}^{\star}(\theta_{0})\rvert^{2} Φ ω ( θ 0 , z 0 ) ≤ a 2 ∣ z 0 − z ω ⋆ ( θ 0 ) ∣ 2 . Substitution proves (31) with M = a 2 / a 1 M=a_{2}/a_{1} M = a 2 / a 1 and C ≥ c 0 / ( λ a 1 ) C\ge c_{0}/(\lambda a_{1}) C ≥ c 0 / ( λ a 1 ) .
For σ > 0 \sigma>0 σ > 0 , apply (3) with Q 0 = Φ ω ( θ 0 , z 0 ) ≤ r Q_{0}=\Phi_{\omega}(\theta_{0},z_{0})\le r Q 0 = Φ ω ( θ 0 , z 0 ) ≤ r . Since R − Q 0 ≥ R − r R-Q_{0}\ge R-r R − Q 0 ≥ R − r , replacing Q 0 Q_{0} Q 0 by r r r in that upper bound proves (32).
Finally, Lemma 3 gives sup θ ∣ z ω ⋆ ( θ ) − z ⋆ ∣ ≤ C ω ω − 1 / 2 \sup_{\theta}\lvert z_{\omega}^{\star}(\theta)-z^{\star}\rvert\le C_{\omega}\omega^{-1/2} sup θ ∣ z ω ⋆ ( θ ) − z ⋆ ∣ ≤ C ω ω − 1/2 . The inequality ∣ x + y ∣ 2 ≤ 2 ∣ x ∣ 2 + 2 ∣ y ∣ 2 \lvert x+y\rvert^{2}\le2\lvert x\rvert^{2}+2\lvert y\rvert^{2} ∣ x + y ∣ 2 ≤ 2 ∣ x ∣ 2 + 2 ∣ y ∣ 2 therefore yields
∣ Z t ω − z ⋆ ∣ 2 ≤ 2 ∣ Z t ω − z ω ⋆ ( θ t ) ∣ 2 + 2 C 0 2 ω \lvert Z_{t}^{\omega}-z^{\star}\rvert^{2}\le2\lvert Z_{t}^{\omega}-z_{\omega}^{\star}(\theta_{t})\rvert^{2}+\frac{2C_{0}^{2}}{\omega} ∣ Z t ω − z ⋆ ∣ 2 ≤ 2 ∣ Z t ω − z ω ⋆ ( θ t ) ∣ 2 + ω 2 C 0 2
before exit, and
∣ z 0 − z ω ⋆ ( θ 0 ) ∣ 2 ≤ 2 ∣ z 0 − z ⋆ ∣ 2 + 2 C 0 2 ω . \lvert z_{0}-z_{\omega}^{\star}(\theta_{0})\rvert^{2}\le2\lvert z_{0}-z^{\star}\rvert^{2}+\frac{2C_{0}^{2}}{\omega}. ∣ z 0 − z ω ⋆ ( θ 0 ) ∣ 2 ≤ 2 ∣ z 0 − z ⋆ ∣ 2 + ω 2 C 0 2 .
Taking the pre-exit expectation in the first inequality and substituting the orbital estimate and the second inequality gives
E [ ∣ Z t ω − z ⋆ ∣ 2 1 { t < τ R ω } ] ≤ 4 M e − λ t ∣ z 0 − z ⋆ ∣ 2 + 2 c 0 λ a 1 σ 2 + ( 4 M + 2 ) C 0 2 ω . \mathbb{E}\left[\lvert Z_{t}^{\omega}-z^{\star}\rvert^{2}\mathbf{1}_{\{t<\tau_{R}^{\omega}\}}\right]
\le4Me^{-\lambda t}\lvert z_{0}-z^{\star}\rvert^{2}+\frac{2c_{0}}{\lambda a_{1}}\sigma^{2}+\frac{(4M+2)C_{0}^{2}}{\omega}. E [ ∣ Z t ω − z ⋆ ∣ 2 1 { t < τ R ω } ] ≤ 4 M e − λ t ∣ z 0 − z ⋆ ∣ 2 + λ a 1 2 c 0 σ 2 + ω ( 4 M + 2 ) C 0 2 .
We then choose
C ≥ max { 4 M , 2 c 0 λ a 1 , ( 4 M + 2 ) C 0 2 } . C\ge\max\left\{4M,\frac{2c_{0}}{\lambda a_{1}},(4M+2)C_{0}^{2}\right\}. C ≥ max { 4 M , λ a 1 2 c 0 , ( 4 M + 2 ) C 0 2 } .
The preceding right-hand side is then bounded by C e − λ t ∣ z 0 − z ⋆ ∣ 2 + C ( σ 2 + ω − 1 ) Ce^{-\lambda t}\lvert z_{0}-z^{\star}\rvert^{2}+C(\sigma^{2}+\omega^{-1}) C e − λ t ∣ z 0 − z ⋆ ∣ 2 + C ( σ 2 + ω − 1 ) , which establishes (30). This choice of C C C is independent of time, the initial state and phase, and the frequency and noise intensity within the stated ranges. □
Since ∇ h ( x ⋆ ) = 0 \nabla h(x^\star)=0 ∇ h ( x ⋆ ) = 0 , a Taylor expansion near x ⋆ x^\star x ⋆ and the compactness away from it give ∣ h ( x ) − h ⋆ ∣ ≤ C h ∣ x − x ⋆ ∣ 2 |h(x)-h^\star|\le C_h|x-x^\star|^2 ∣ h ( x ) − h ⋆ ∣ ≤ C h ∣ x − x ⋆ ∣ 2 on the physical analysis region. The expected absolute cost error before exit therefore satisfies (31) with a changed constant.
The next corollary strengthens the regulation estimate to a trajectory-level statement. With high probability, the closed-loop state enters a prescribed neighborhood of z ⋆ z^\star z ⋆ by a fixed capture time and remains there throughout a subsequent interval. The proof first bounds the noise fluctuations that could prevent capture, then applies the exit estimate of Lemma 1 from the state reached at the capture time. Both failure probabilities decrease exponentially with the inverse noise variance, allowing exponentially long residence intervals.
Corollary 1. Under Theorem 1, for each ρ > 0 \rho>0 ρ > 0 there exist T a > 0 T_a>0 T a > 0 , ω ρ ≥ ω K \omega_\rho\ge\omega_\mathcal{K} ω ρ ≥ ω K , and 0 < σ ρ ≤ σ ˉ 0<\sigma_\rho\le\bar{\sigma} 0 < σ ρ ≤ σ ˉ such that
P { τ R ω > T a + T , sup T a ≤ t ≤ T a + T ∣ Z t ω − z ⋆ ∣ < ρ } ≥ 1 − e − b c a p / σ 2 − e − b 1 / σ 2 − C 0 T e − b 2 / σ 2 . (37) \mathbb{P}\left\{\tau_R^\omega>T_a+T,\ \sup_{T_a\le t\le T_a+T}|Z_t^\omega-z^\star|<\rho\right\}\ge1-e^{-b_{\mathrm{cap}}/\sigma^2}-e^{-b_1/\sigma^2}-C_0Te^{-b_2/\sigma^2}. \tag*{(37)} P { τ R ω > T a + T , T a ≤ t ≤ T a + T sup ∣ Z t ω − z ⋆ ∣ < ρ } ≥ 1 − e − b cap / σ 2 − e − b 1 / σ 2 − C 0 T e − b 2 / σ 2 . (37) The constants b c a p , b 1 , b 2 , C 0 > 0 b_{\mathrm{cap}},b_1,b_2,C_0>0 b cap , b 1 , b 2 , C 0 > 0 are uniform for z 0 ∈ K z_0\in\mathcal{K} z 0 ∈ K , θ 0 ∈ T \theta_0\in\mathbb{T} θ 0 ∈ T , ω ≥ ω ρ \omega\ge\omega_\rho ω ≥ ω ρ , 0 < σ ≤ σ ρ 0<\sigma\le\sigma_\rho 0 < σ ≤ σ ρ , and finite T ≥ 0 T\ge0 T ≥ 0 . For T = T ⋆ e a / σ 2 T=T_\star e^{a/\sigma^2} T = T ⋆ e a / σ 2 with fixed T ⋆ > 0 T_\star>0 T ⋆ > 0 and 0 < a < b 2 0<a<b_2 0 < a < b 2 , the failure bound is at most ( 2 + C 0 T ⋆ ) e − b a / σ 2 (2+C_0T_\star)e^{-b_a/\sigma^2} ( 2 + C 0 T ⋆ ) e − b a / σ 2 , where b a = min { b c a p , b 1 , b 2 − a } > 0 b_a=\min\{b_{\mathrm{cap}},b_1,b_2-a\}>0 b a = min { b cap , b 1 , b 2 − a } > 0 .
Proof. We start by choosing 0 < r a < R a < R 0<r_a<R_a<R 0 < r a < R a < R with R a / α 1 < ρ / 2 \sqrt{R_a}/\alpha_1<\rho/2 R a / α 1 < ρ /2 . By Lemma 3, we can increase ω ρ \omega_\rho ω ρ such that sup θ ∣ z ω ⋆ ( θ ) − z ⋆ ∣ < ρ / 2 \sup_\theta|z_\omega^\star(\theta)-z^\star|<\rho/2 sup θ ∣ z ω ⋆ ( θ ) − z ⋆ ∣ < ρ /2 . The lower bound in (29) then ensures that D R a ω ( θ ) \mathcal{D}_{R_a}^\omega(\theta) D R a ω ( θ ) lies inside the ρ \rho ρ -ball about z ⋆ z^\star z ⋆ for every phase. We then choose T a > 0 T_a>0 T a > 0 such that e − λ T a r < r a / 4 e^{-\lambda T_a}r<r_a/4 e − λ T a r < r a /4 .
Let Q t = Φ ω ( θ t , Z t ω ) Q_t=\Phi_\omega(\theta_t,Z_t^\omega) Q t = Φ ω ( θ t , Z t ω ) up to the outer exit time τ R ω \tau_R^\omega τ R ω . We first estimate the probability of the capture event G : = { τ R ω > T a , Q T a < r a } \mathcal{G}:=\{\tau_R^\omega>T_a,\ Q_{T_a}<r_a\} G := { τ R ω > T a , Q T a < r a } . To control the accumulated noise before T a T_a T a , we define
M t : = σ ∫ 0 t ∧ τ R ω e λ s ∇ z Φ ω ( θ s , Z s ω ) T S 0 ( Z s ω ) d B s . M_t:=\sigma\int_0^{t\wedge\tau_R^\omega}e^{\lambda s}\nabla_z\Phi_\omega(\theta_s,Z_s^\omega)^\mathsf{T}S_0(Z_s^\omega)\,dB_s. M t := σ ∫ 0 t ∧ τ R ω e λ s ∇ z Φ ω ( θ s , Z s ω ) T S 0 ( Z s ω ) d B s .
Applying Itô’s formula to e λ t Q t e^{\lambda t}Q_t e λ t Q t up to τ R ω \tau_R^\omega τ R ω and using (34) gives
Q t R ≤ e − λ t R Q 0 + c 0 σ 2 λ + e − λ t R M t (38) Q_{t_R}\le e^{-\lambda t_R}Q_0+\frac{c_0\sigma^2}{\lambda}+e^{-\lambda t_R}M_t \tag*{(38)} Q t R ≤ e − λ t R Q 0 + λ c 0 σ 2 + e − λ t R M t (38)
for t R : = t ∧ τ R ω t_R:=t\wedge\tau_R^\omega t R := t ∧ τ R ω . Since Q s < R Q_s<R Q s < R before exit, the expression (35) gives
⟨ M ⟩ T a ≤ d 0 R σ 2 ∫ 0 T a e 2 λ s d s ≤ H a σ 2 , \langle M\rangle_{T_a}\le d_0R\sigma^2\int_0^{T_a}e^{2\lambda s}\,ds\le H_a\sigma^2, ⟨ M ⟩ T a ≤ d 0 R σ 2 ∫ 0 T a e 2 λ s d s ≤ H a σ 2 ,
where H a : = max { 1 , d 0 R ( e 2 λ T a − 1 ) / ( 2 λ ) } H_a:=\max\{1,d_0R(e^{2\lambda T_a}-1)/(2\lambda)\} H a := max { 1 , d 0 R ( e 2 λ T a − 1 ) / ( 2 λ )} .
We then set Δ a = 1 4 min { R − r , r a } \Delta_a=\frac14\min\{R-r,r_a\} Δ a = 4 1 min { R − r , r a } and choose 0 < σ ρ ≤ σ ˉ 0<\sigma_\rho\le\bar{\sigma} 0 < σ ρ ≤ σ ˉ such that c 0 σ ρ 2 / λ ≤ Δ a c_0\sigma_\rho^2/\lambda\le\Delta_a c 0 σ ρ 2 / λ ≤ Δ a . On the event { sup t ≤ T a M t < Δ a } \{\sup_{t\le T_a}M_t<\Delta_a\} { sup t ≤ T a M t < Δ a } , inequality (38) implies Q t R ≤ r + 2 Δ a < R Q_{t_R}\le r+2\Delta_a<R Q t R ≤ r + 2 Δ a < R for every t ≤ T a t\le T_a t ≤ T a . An outer exit would require Q τ R ω = R Q_{\tau_R^\omega}=R Q τ R ω = R , so τ R ω > T a \tau_R^\omega>T_a τ R ω > T a . At the capture time, the same estimate gives
Q T a ≤ e − λ T a r + 2 Δ a < 3 4 r a < r a . Q_{T_a}\le e^{-\lambda T_a}r+2\Delta_a<\frac34r_a<r_a. Q T a ≤ e − λ T a r + 2 Δ a < 4 3 r a < r a .
Thus a failure of G \mathcal{G} G requires the martingale M t M_t M t to reach Δ a \Delta_a Δ a before T a T_a T a .
For γ > 0 \gamma>0 γ > 0 , the process exp ( γ M t − γ 2 ⟨ M ⟩ t / 2 ) \exp(\gamma M_t-\gamma^2\langle M\rangle_t/2) exp ( γ M t − γ 2 ⟨ M ⟩ t /2 ) is a nonnegative supermartingale [53 ] . It is stopped when M t M_t M t first reaches Δ a \Delta_a Δ a , or at T a T_a T a . We then choose γ = Δ a / ( H a σ 2 ) \gamma=\Delta_a/(H_a\sigma^2) γ = Δ a / ( H a σ 2 ) . This gives
P ( G c ) ≤ P { sup t ≤ T a M t ≥ Δ a } ≤ exp ( − Δ a 2 2 H a σ 2 ) . \mathbb{P}(\mathcal{G}^{c})\le\mathbb{P}\left\{\sup_{t\le T_a}M_t\ge\Delta_a\right\}\le\exp\left(-\frac{\Delta_a^2}{2H_a\sigma^2}\right). P ( G c ) ≤ P { t ≤ T a sup M t ≥ Δ a } ≤ exp ( − 2 H a σ 2 Δ a 2 ) .
Hence the capture failure probability is at most e − b c a p / σ 2 e^{-b_{\mathrm{cap}}/\sigma^2} e − b cap / σ 2 , with b c a p : = Δ a 2 / ( 2 H a ) b_{\mathrm{cap}}:=\Delta_a^2/(2H_a) b cap := Δ a 2 / ( 2 H a ) .
It remains to monitor the departures after capture. On G \mathcal{G} G , let τ R a ω \tau_{R_a}^\omega τ R a ω be the first exit from D R a ω ( θ t ) \mathcal{D}_{R_a}^\omega(\theta_t) D R a ω ( θ t ) after T a T_a T a . The bounds (34) –(35) hold on this smaller sublevel, uniformly in phase. Applying Lemma 1 conditionally on F T a \mathcal{F}_{T_a} F T a , with Q T a < r a Q_{T_a}<r_a Q T a < r a , yields
P { τ R a ω ≤ T a + T ∣ F T a } ≤ e − η ( R a − r a ) / σ 2 + C 0 T e − η R a / σ 2 on G . \mathbb{P}\{\tau_{R_a}^\omega\le T_a+T\mid\mathcal{F}_{T_a}\}\le e^{-\eta(R_a-r_a)/\sigma^2}+C_0Te^{-\eta R_a/\sigma^2}\quad\text{on }\mathcal{G}. P { τ R a ω ≤ T a + T ∣ F T a } ≤ e − η ( R a − r a ) / σ 2 + C 0 T e − η R a / σ 2 on G .
On G ∩ { τ R a ω > T a + T } \mathcal{G}\cap\{\tau_{R_a}^\omega>T_a+T\} G ∩ { τ R a ω > T a + T } , the trajectory remains in the ρ \rho ρ -ball during [ T a , T a + T ] [T_a,T_a+T] [ T a , T a + T ] and does not leave the outer sublevel before the interval ends. Hence the capture failure bound to the conditional exit bound proves (37) , with b 1 : = η ( R a − r a ) b_1:=\eta(R_a-r_a) b 1 := η ( R a − r a ) and b 2 : = η R a b_2:=\eta R_a b 2 := η R a .
Finally, for T = T ⋆ e a / σ 2 T=T_\star e^{a/\sigma^2} T = T ⋆ e a / σ 2 and 0 < a < b 2 0<a<b_2 0 < a < b 2 , the failure bound becomes
e − b c a p / σ 2 + e − b 1 / σ 2 + C 0 T ⋆ e − ( b 2 − a ) / σ 2 ≤ ( 2 + C 0 T ⋆ ) e − b a / σ 2 , e^{-b_{\mathrm{cap}}/\sigma^2}+e^{-b_1/\sigma^2}+C_0T_\star e^{-(b_2-a)/\sigma^2}
\le(2+C_0T_\star)e^{-b_a/\sigma^2}, e − b cap / σ 2 + e − b 1 / σ 2 + C 0 T ⋆ e − ( b 2 − a ) / σ 2 ≤ ( 2 + C 0 T ⋆ ) e − b a / σ 2 ,
where b a = min { b c a p , b 1 , b 2 − a } > 0 b_a=\min\{b_{\mathrm{cap}},b_1,b_2-a\}>0 b a = min { b cap , b 1 , b 2 − a } > 0 . All choices use the uniform bounds of Theorem 1, so the capture time and constants are independent of the initial state, initial phase, and frequency above ω ρ \omega_\rho ω ρ . □
The estimate (20) bounds an expected trajectory supremum on each fixed interval. We now bound the mean-square discrepancy at every time with one constant. This requires decay of the separation between averaged trajectories driven by the same Brownian motion.
Assumption 4. For the fixed gains, region O 2 \mathcal{O}_2 O 2 , and noise bound σ ˉ \bar{\sigma} σ ˉ , there exist P ∈ S + n + m P\in\mathbb{S}_+^{n+m} P ∈ S + n + m and α > 0 \alpha>0 α > 0 such that, for z , z ′ ∈ O 2 z,z'\in\mathcal{O}_2 z , z ′ ∈ O 2 and 0 ≤ σ ≤ σ ˉ 0\le\sigma\le\bar{\sigma} 0 ≤ σ ≤ σ ˉ ,
2 ( z − z ′ ) T P [ F T ( z ) − F T ( z ′ ) ] + σ 2 ∥ P 1 / 2 [ S 0 ( z ) − S 0 ( z ′ ) ] ∥ F 2 ≤ − α ∣ z − z ′ ∣ P 2 . (39) 2(z-z')^\mathsf{T}P[F_T(z)-F_T(z')]+\sigma^2\|P^{1/2}[S_0(z)-S_0(z')]\|_F^2\le-\alpha|z-z'|_P^2. \tag*{(39)} 2 ( z − z ′ ) T P [ F T ( z ) − F T ( z ′ )] + σ 2 ∥ P 1/2 [ S 0 ( z ) − S 0 ( z ′ )] ∥ F 2 ≤ − α ∣ z − z ′ ∣ P 2 . (39)
where ∣ e ∣ P 2 = e T P e |e|_P^2=e^\mathsf{T}Pe ∣ e ∣ P 2 = e T P e .
Remark 3. This condition is stronger than the equilibrium certificate on the same region. Its verification requires structural information or uncertainty bounds that are not supplied by the scalar cost alone. It is used only for the approximation result. On convex O 2 \mathcal{O}_2 O 2 , a sufficient test is
D F T T P + P D F T + σ ˉ 2 ∑ j = 1 d B D S 0 , j T P D S 0 , j ⪯ − α P . DF_T^\mathsf{T}P+PDF_T+\bar{\sigma}^2\sum_{j=1}^{d_B}DS_{0,j}^\mathsf{T}PDS_{0,j}\preceq-\alpha P. D F T T P + P D F T + σ ˉ 2 j = 1 ∑ d B D S 0 , j T P D S 0 , j ⪯ − α P .
Here S 0 , j S_{0,j} S 0 , j is the j j j th diffusion column. Integration along a segment and Jensen’s inequality imply (39) . A Hurwitz target linearization gives this test on a sufficiently small neighborhood for small enough σ ˉ \bar{\sigma} σ ˉ , not automatically on the prescribed region.
Theorem 2. Let Assumptions 1–3 and 4 hold. Fix an open U U U with K ⊂ U ⋐ O 1 K \subset U \Subset O_1 K ⊂ U ⋐ O 1 , and couple (9) and (12) with the same initial state and Brownian motion. Put τ c ω = τ U ( Z ω ) ∧ τ U ( Z ˉ ) \tau_c^\omega=\tau_U(Z^\omega)\wedge\tau_U(\bar Z) τ c ω = τ U ( Z ω ) ∧ τ U ( Z ˉ ) . There exist ω 0 , C > 0 \omega_0,C>0 ω 0 , C > 0 such that
sup z 0 ∈ K , θ 0 ∈ T t ≥ 0 E [ ∣ Z t ω − Z ˉ t ∣ 2 1 { t < τ c ω } ] ≤ C ω (40) \sup_{\substack{z_0\in K,\theta_0\in\mathbb{T}\\t\geq0}}\mathbb{E}\left[|Z_t^\omega-\bar Z_t|^2\mathbf{1}_{\{t<\tau_c^\omega\}}\right]\leq\frac{C}{\omega} \tag*{(40)} z 0 ∈ K , θ 0 ∈ T t ≥ 0 sup E [ ∣ Z t ω − Z ˉ t ∣ 2 1 { t < τ c ω } ] ≤ ω C (40)
for ω ≥ ω 0 \omega\geq\omega_0 ω ≥ ω 0 and 0 ≤ σ ≤ σ ˉ 0\leq\sigma\leq\bar\sigma 0 ≤ σ ≤ σ ˉ .
Proof. Before τ c ω \tau_c^\omega τ c ω , define ζ t = Γ − 1 ( θ t , Z t ω ) \zeta_t=\Gamma^{-1}(\theta_t,Z_t^\omega) ζ t = Γ − 1 ( θ t , Z t ω ) and e t = ζ t − Z ˉ t e_t=\zeta_t-\bar Z_t e t = ζ t − Z ˉ t . The buffer and (16) keep both arguments in O 2 O_2 O 2 for small ε \varepsilon ε . Their error equation is
d e t = [ F T ( ζ t ) − F T ( Z ˉ t ) + r t ] d t + σ ( D t + J t ) d B t , de_t=[F_T(\zeta_t)-F_T(\bar Z_t)+r_t]dt+\sigma(D_t+J_t)dB_t, d e t = [ F T ( ζ t ) − F T ( Z ˉ t ) + r t ] d t + σ ( D t + J t ) d B t ,
where r t = ε R ε ( θ t , ζ t ) + σ 2 B ε ( θ t , ζ t ) r_t=\varepsilon R_\varepsilon(\theta_t,\zeta_t)+\sigma^2B_\varepsilon(\theta_t,\zeta_t) r t = ε R ε ( θ t , ζ t ) + σ 2 B ε ( θ t , ζ t ) , D t = S 0 ( ζ t ) − S 0 ( Z ˉ t ) D_t=S_0(\zeta_t)-S_0(\bar Z_t) D t = S 0 ( ζ t ) − S 0 ( Z ˉ t ) , and J t = S ^ ε ( θ t , ζ t ) − S 0 ( ζ t ) J_t=\widehat S_\varepsilon(\theta_t,\zeta_t)-S_0(\zeta_t) J t = S ε ( θ t , ζ t ) − S 0 ( ζ t ) . Lemma 2 gives ∣ r t ∣ + ∥ J t ∥ F ≤ C ε |r_t|+\|J_t\|_F\leq C\varepsilon ∣ r t ∣ + ∥ J t ∥ F ≤ C ε for σ ≤ σ ˉ \sigma\leq\bar\sigma σ ≤ σ ˉ . A Lipschitz bound for the extension of S 0 S_0 S 0 gives ∥ D t ∥ F ≤ L S ∣ e t ∣ \|D_t\|_F\leq L_S|e_t| ∥ D t ∥ F ≤ L S ∣ e t ∣ .
For W t = e t T P e t W_t=e_t^{\mathsf T}Pe_t W t = e t T P e t , the generator of the coupled process satisfies
L p a i r W = 2 e t T P [ F T ( ζ t ) − F T ( Z ˉ t ) + r t ] + σ 2 ∥ P 1 / 2 ( D t + J t ) ∥ F 2 ≤ − α W t + 2 e t T P r t + 2 σ 2 ⟨ P 1 / 2 D t , P 1 / 2 J t ⟩ F + σ 2 ∥ P 1 / 2 J t ∥ F 2 , \begin{aligned}
\mathcal{L}_{\mathrm{pair}}W={}&2e_t^{\mathsf T}P[F_T(\zeta_t)-F_T(\bar Z_t)+r_t]\\
&+\sigma^2\|P^{1/2}(D_t+J_t)\|_F^2\\
\leq{}&-\alpha W_t+2e_t^{\mathsf T}Pr_t\\
&+2\sigma^2\langle P^{1/2}D_t,P^{1/2}J_t\rangle_F+\sigma^2\|P^{1/2}J_t\|_F^2,
\end{aligned} L pair W = ≤ 2 e t T P [ F T ( ζ t ) − F T ( Z ˉ t ) + r t ] + σ 2 ∥ P 1/2 ( D t + J t ) ∥ F 2 − α W t + 2 e t T P r t + 2 σ 2 ⟨ P 1/2 D t , P 1/2 J t ⟩ F + σ 2 ∥ P 1/2 J t ∥ F 2 ,
where (38) bounds the terms containing only F T , D t F_T,D_t F T , D t , and ⟨ A , B ⟩ F = tr ( A T B ) \langle A,B\rangle_F=\operatorname{tr}(A^{\mathsf T}B) ⟨ A , B ⟩ F = tr ( A T B ) . The remainder is at most C ε W t + C ε 2 C\varepsilon\sqrt{W_t}+C\varepsilon^2 C ε W t + C ε 2 . Young’s inequality yields
L p a i r W ≤ − 1 2 α W t + C 1 ε 2 , t < τ c ω . (41) \mathcal{L}_{\mathrm{pair}}W\leq-\frac{1}{2}\alpha W_t+C_1\varepsilon^2,\qquad t<\tau_c^\omega. \tag*{(41)} L pair W ≤ − 2 1 α W t + C 1 ε 2 , t < τ c ω . (41)
The stopped stochastic integral is square integrable on each finite interval. Applying (40) to e α ( t ∧ τ c ω ) / 2 W t ∧ τ c ω e^{\alpha(t\wedge\tau_c^\omega)/2}W_{t\wedge\tau_c^\omega} e α ( t ∧ τ c ω ) /2 W t ∧ τ c ω and dropping its nonnegative exit contribution gives
E [ W t 1 { t < τ c ω } ] ≤ e − α t / 2 W 0 + 2 C 1 α ε 2 ( 1 − e − α t / 2 ) . \mathbb{E}[W_t\mathbf{1}_{\{t<\tau_c^\omega\}}]\leq e^{-\alpha t/2}W_0+\frac{2C_1}{\alpha}\varepsilon^2(1-e^{-\alpha t/2}). E [ W t 1 { t < τ c ω } ] ≤ e − α t /2 W 0 + α 2 C 1 ε 2 ( 1 − e − α t /2 ) .
Let C Γ C_\Gamma C Γ denote a constant in (16). The common initialization implies W 0 ≤ ∥ P ∥ C Γ 2 ε 2 W_0\leq\|P\|C_\Gamma^2\varepsilon^2 W 0 ≤ ∥ P ∥ C Γ 2 ε 2 by (16). The same bound and Z t ω = Γ ε ( θ t , ζ t ) Z_t^\omega=\Gamma_\varepsilon(\theta_t,\zeta_t) Z t ω = Γ ε ( θ t , ζ t ) give
∣ Z t ω − Z ˉ t ∣ 2 ≤ 2 C Γ 2 ε 2 + 2 W t λ min ( P ) . |Z_t^\omega-\bar Z_t|^2\leq2C_\Gamma^2\varepsilon^2+\frac{2W_t}{\lambda_{\min}(P)}. ∣ Z t ω − Z ˉ t ∣ 2 ≤ 2 C Γ 2 ε 2 + λ m i n ( P ) 2 W t .
Combining these inequalities proves (39), uniformly in time, phase, initial state, and σ ≤ σ ˉ \sigma\leq\bar\sigma σ ≤ σ ˉ . □
To relate this pre-exit estimate to a growing interval, we choose S R 0 ⊂ U ⋐ O 1 S_{R_0}\subset U\Subset O_1 S R 0 ⊂ U ⋐ O 1 . For a sufficiently large frequency, both D R ω ( θ ) D_R^\omega(\theta) D R ω ( θ ) and S R = { V ≤ R } S_R=\{V\leq R\} S R = { V ≤ R } lie in U U U . With τ ˉ R = inf { t : V ( Z ˉ t ) ≥ R } \bar\tau_R=\inf\{t:V(\bar Z_t)\geq R\} τ ˉ R = inf { t : V ( Z ˉ t ) ≥ R } , the union bound gives
P { τ c ω ≤ T } ≤ P { τ R ω ≤ T } + P { τ ˉ R ≤ T } . \mathbb{P}\{\tau_c^\omega\leq T\}\leq\mathbb{P}\{\tau_R^\omega\leq T\}+\mathbb{P}\{\bar\tau_R\leq T\}. P { τ c ω ≤ T } ≤ P { τ R ω ≤ T } + P { τ ˉ R ≤ T } .
Apply (32) to the first term and Lemma 1 with (14) to the second. Taking the smaller exponential coefficient gives
P { τ c ω ≤ T } ≤ δ σ ( T ) : = 2 e − η ∗ ( R − r ) / σ 2 + C ∗ T e − η ∗ R / σ 2 \mathbb{P}\{\tau_c^\omega\leq T\}\leq\delta_\sigma(T):=2e^{-\eta_*(R-r)/\sigma^2}+C_*Te^{-\eta_*R/\sigma^2} P { τ c ω ≤ T } ≤ δ σ ( T ) := 2 e − η ∗ ( R − r ) / σ 2 + C ∗ T e − η ∗ R / σ 2
for constants η ∗ , C ∗ > 0 \eta_*,C_*>0 η ∗ , C ∗ > 0 . Since { τ c ω > T } ⊂ { t < τ c ω } \{\tau_c^\omega>T\}\subset\{t<\tau_c^\omega\} { τ c ω > T } ⊂ { t < τ c ω } for t ≤ T t\leq T t ≤ T , (39) implies
sup 0 ≤ t ≤ T E [ ∣ Z t ω − Z ˉ t ∣ 2 ∣ τ c ω > T ] ≤ C ω [ 1 − δ σ ( T ) ] \sup_{0\leq t\leq T}\mathbb{E}[|Z_t^\omega-\bar Z_t|^2\mid\tau_c^\omega>T]\leq\frac{C}{\omega[1-\delta_\sigma(T)]} 0 ≤ t ≤ T sup E [ ∣ Z t ω − Z ˉ t ∣ 2 ∣ τ c ω > T ] ≤ ω [ 1 − δ σ ( T )] C
whenever δ σ ( T ) < 1 \delta_\sigma(T)<1 δ σ ( T ) < 1 . For T = T ∗ e a / σ 2 T=T_*e^{a/\sigma^2} T = T ∗ e a / σ 2 with 0 < a < η ∗ R 0<a<\eta_*R 0 < a < η ∗ R , δ σ ( T ) → 0 \delta_\sigma(T)\to0 δ σ ( T ) → 0 as σ ↓ 0 \sigma\downarrow0 σ ↓ 0 . The supremum here is outside the expectation, as in Theorem 2. The estimate bounds the conditional mean-square error at every time in [ 0 , T ] [0,T] [ 0 , T ] , given that both trajectories remain in U U U throughout that interval. The supremum is taken over these expectations, rather than over the error along each individual trajectory.
Brownian-Particle Confinement in a Hydrodynamic Saddle Trap
Physical model and measured objective
Consider a planar hydrodynamic trap with controlled outlet-axis stagnation-point position x s = x 0 + β c u x_s=x_0+\beta_cu x s = x 0 + β c u [54 , 55 , 56 ] . The physical particle position ( x p , y p ) (x_p,y_p) ( x p , y p ) satisfies the local overdamped model
d x p = ϵ ˙ f ( x p − x s ) d τ + 2 D B d W 1 , τ , dx_p=\dot\epsilon_f(x_p-x_s)d\tau+\sqrt{2D_B}\,dW_{1,\tau}, d x p = ϵ ˙ f ( x p − x s ) d τ + 2 D B d W 1 , τ ,
d y p = − ϵ ˙ f y p d τ + 2 D B d W 2 , τ . dy_p=-\dot\epsilon_fy_p\,d\tau+\sqrt{2D_B}\,dW_{2,\tau}. d y p = − ϵ ˙ f y p d τ + 2 D B d W 2 , τ .
where ϵ ˙ f > 0 \dot\epsilon_f>0 ϵ ˙ f > 0 is the strain rate, D B > 0 D_B>0 D B > 0 is the translational diffusivity, and W 1 , W 2 W_1,W_2 W 1 , W 2 are independent Wiener processes. The offset x s x_s x s is unknown and β c ≠ 0 \beta_c\neq0 β c = 0 is the signed command sensitivity. All coefficients are fixed during each realization.
Normalize position by the detection-spot width w > 0 w>0 w > 0 and time by a reference strain rate ϵ ˙ 0 > 0 \dot\epsilon_0>0 ϵ ˙ 0 > 0 . Let x = x p / w x=x_p/w x = x p / w , y = y p / w y=y_p/w y = y p / w , and t = ϵ ˙ 0 τ t=\dot\epsilon_0\tau t = ϵ ˙ 0 τ , the dynamics become
d x t = ( a x t + d + b u t ) d t + σ d B 1 , t , dx_t=(ax_t+d+bu_t)dt+\sigma\,dB_{1,t}, d x t = ( a x t + d + b u t ) d t + σ d B 1 , t ,
d y t = − a y t d t + σ d B 2 , t , (42) dy_t=-ay_t\,dt+\sigma\,dB_{2,t}, \tag*{(42)} d y t = − a y t d t + σ d B 2 , t , (42)
where a = ϵ ˙ f / ϵ ˙ 0 a=\dot\epsilon_f/\dot\epsilon_0 a = ϵ ˙ f / ϵ ˙ 0 , d = − a x s / w d=-ax_s/w d = − a x s / w , b = − a β c / w b=-a\beta_c/w b = − a β c / w , and σ 2 = 2 D B / ( ϵ ˙ 0 w 2 ) \sigma^2=2D_B/(\dot\epsilon_0w^2) σ 2 = 2 D B / ( ϵ ˙ 0 w 2 ) . The controller receives only the optical score
Y t = h ( x t , y t ) = h 0 − I 0 exp [ − x t 2 + y t 2 2 ] , (43) Y_t=h(x_t,y_t)=h_0-I_0\exp\left[-\frac{x_t^2+y_t^2}{2}\right], \tag*{(43)} Y t = h ( x t , y t ) = h 0 − I 0 exp [ − 2 x t 2 + y t 2 ] , (43)
with constant background h 0 h_0 h 0 and peak intensity I 0 > 0 I_0>0 I 0 > 0 . Minimizing this score maximizes detected intensity, and the controller forms no position estimate. This idealized instantaneous measurement omits the photon-arrival sensing and spatial scanning used for position estimation in single-molecule ABEL traps [57 ] .
The plant has state X = ( x , y ) T ∈ R 2 X=(x,y)^{\mathsf T}\in\mathbb{R}^2 X = ( x , y ) T ∈ R 2 , input vector G = ( b , 0 ) T G=(b,0)^{\mathsf T} G = ( b , 0 ) T , and normalized diffusion matrix Σ 0 = I 2 \Sigma_0=I_2 Σ 0 = I 2 . Thus, noise acts along the non-actuated, passively stable y y y direction. The feasible nominal minimum is X ∗ = 0 X^*=0 X ∗ = 0 , with sustaining input u ∗ = − d / b u^*=-d/b u ∗ = − d / b , value h ∗ = h 0 − I 0 h^*=h_0-I_0 h ∗ = h 0 − I 0 , and Hessian ∇ 2 h ( 0 ) = I 0 I 2 ≻ 0 \nabla^2h(0)=I_0I_2\succ0 ∇ 2 h ( 0 ) = I 0 I 2 ≻ 0 . Writing φ ( x , y ) = exp [ − ( x 2 + y 2 ) / 2 ] \varphi(x,y)=\exp[-(x^2+y^2)/2] φ ( x , y ) = exp [ − ( x 2 + y 2 ) /2 ] , the input sensitivity is ℓ ( x , y ) = b I 0 φ ( x , y ) \ell(x,y)=bI_0\varphi(x,y) ℓ ( x , y ) = b I 0 φ ( x , y ) . The scripts denote the strain-rate ratio a a a by ϵ \epsilon ϵ , whereas the averaging parameter is ω − 1 / 2 \omega^{-1/2} ω − 1/2 .
Controller and integral compensation
We use (7) with κ 1 = 1 \kappa_1=1 κ 1 = 1 , k P = 4 k_P=4 k P = 4 , and k I = 1 k_I=1 k I = 1 . With κ P = k P b 2 I 0 \kappa_P=k_Pb^2I_0 κ P = k P b 2 I 0 , the averaged diffusion is
d x ˉ t = [ a x ˉ t + d + b u ˉ b , t − κ P x ˉ t φ ( x ˉ t , y ˉ t ) ] d t + σ d B 1 , t , d y ˉ t = − a y ˉ t d t + σ d B 2 , t , d u ˉ b , t = − k I b I 0 x ˉ t φ ( x ˉ t , y ˉ t ) d t . (44) \begin{aligned}
d\bar x_t&=[a\bar x_t+d+b\bar u_{b,t}-\kappa_P\bar x_t\varphi(\bar x_t,\bar y_t)]dt+\sigma\,dB_{1,t},\\
d\bar y_t&=-a\bar y_t\,dt+\sigma\,dB_{2,t},\\
d\bar u_{b,t}&=-k_IbI_0\bar x_t\varphi(\bar x_t,\bar y_t)\,dt. \tag*{(44)}
\end{aligned} d x ˉ t d y ˉ t d u ˉ b , t = [ a x ˉ t + d + b u ˉ b , t − κ P x ˉ t φ ( x ˉ t , y ˉ t )] d t + σ d B 1 , t , = − a y ˉ t d t + σ d B 2 , t , = − k I b I 0 x ˉ t φ ( x ˉ t , y ˉ t ) d t . (44)
To isolate integral compensation, we compare against the matched proportional-only controller obtained by fixing u b = 0 u_b=0 u b = 0 while retaining the same proportional gain and probing signal. At y = 0 y=0 y = 0 , its deterministic average is
x ˙ = a x + d − κ P x e − x 2 / 2 . (45) \dot{x}=ax+d-\kappa_Pxe^{-x^2/2}. \tag*{(45)} x ˙ = a x + d − κ P x e − x 2 /2 . (45)
The drift at the optical minimum equals d d d , so this point is not an equilibrium when d ≠ 0 d\ne0 d = 0 . For κ P > a \kappa_P>a κ P > a and sufficiently small d d d , the locally stable equilibrium satisfies x d = d / ( κ P − a ) + O ( d 3 ) x_d=d/(\kappa_P-a)+O(d^3) x d = d / ( κ P − a ) + O ( d 3 ) . The dual-mode deterministic average admits the optical minimum as an equilibrium through the sustaining bias u b ∗ = − d / b u_b^*=-d/b u b ∗ = − d / b .
Regional target stability
Consider the uncertainty family a ∈ [ 0.8 , 1.2 ] a\in[0.8,1.2] a ∈ [ 0.8 , 1.2 ] , ∣ b ∣ ∈ [ 0.9 , 1.1 ] \lvert b\rvert\in[0.9,1.1] ∣ b ∣ ∈ [ 0.9 , 1.1 ] , I 0 ∈ [ 0.9 , 1.1 ] I_0\in[0.9,1.1] I 0 ∈ [ 0.9 , 1.1 ] , and ∣ d ∣ ≤ 0.25 \lvert d\rvert\le0.25 ∣ d ∣ ≤ 0.25 . Define ϱ = b ( u b − u ∗ ) = b u b + d \varrho=b(u_b-u^*)=bu_b+d ϱ = b ( u b − u ∗ ) = b u b + d , ξ = col ( x , y , ϱ ) \xi=\operatorname{col}(x,y,\varrho) ξ = col ( x , y , ϱ ) , and k = b 2 I 0 = κ I k=b^2I_0=\kappa_I k = b 2 I 0 = κ I . Appendix A shows that
V ( ξ ) = 1 2 ( x 2 + y 2 ) + ϱ 2 2 k − 1 4 x ϱ (46) V(\xi)=\frac{1}{2}(x^2+y^2)+\frac{\varrho^2}{2k}-\frac{1}{4}x\varrho \tag*{(46)} V ( ξ ) = 2 1 ( x 2 + y 2 ) + 2 k ϱ 2 − 4 1 x ϱ (46)
satisfies the uniform bounds
0.298 ∣ ξ ∣ 2 ≤ V ( ξ ) ≤ 0.749 ∣ ξ ∣ 2 (47) 0.298\lvert\xi\rvert^2\le V(\xi)\le0.749\lvert\xi\rvert^2 \tag*{(47)} 0.298 ∣ ξ ∣ 2 ≤ V ( ξ ) ≤ 0.749 ∣ ξ ∣ 2 (47)
and V ˙ ≤ − 0.1549 ∣ ξ ∣ 2 \dot{V}\le-0.1549\lvert\xi\rvert^2 V ˙ ≤ − 0.1549 ∣ ξ ∣ 2 along the noise-free target for ∣ ξ ∣ < 0.7 \lvert\xi\rvert<0.7 ∣ ξ ∣ < 0.7 . In the original error coordinates e = col ( x , y , u b − u ∗ ) e=\operatorname{col}(x,y,u_b-u^*) e = col ( x , y , u b − u ∗ ) , the corresponding constants in Definition 1 are c 1 = 0.24138 c_1=0.24138 c 1 = 0.24138 , c 2 = 0.90629 c_2=0.90629 c 2 = 0.90629 , and c 3 = 0.125469 c_3=0.125469 c 3 = 0.125469 .
The outer sublevel { V ≤ R 0 } \{V\le R_0\} { V ≤ R 0 } , with R 0 = 0.14 R_0=0.14 R 0 = 0.14 , lies strictly inside ∣ ξ ∣ < 0.7 \lvert\xi\rvert<0.7 ∣ ξ ∣ < 0.7 , since R 0 < 0.298 ( 0.7 ) 2 R_0<0.298(0.7)^2 R 0 < 0.298 ( 0.7 ) 2 . At the common initialization ( x 0 , y 0 , u b , 0 ) = ( 0.2 , 0.1 , 0 ) (x_0,y_0,u_{b,0})=(0.2,0.1,0) ( x 0 , y 0 , u b , 0 ) = ( 0.2 , 0.1 , 0 ) , the nominal choice k = 1 k=1 k = 1 , d = 0.25 d=0.25 d = 0.25 gives V ( ξ 0 ) = 0.04375 V(\xi_0)=0.04375 V ( ξ 0 ) = 0.04375 . Over the entire uncertainty family, V ( ξ 0 ) ≤ 0.080367 V(\xi_0)\le0.080367 V ( ξ 0 ) ≤ 0.080367 , so r 0 = 0.09 < R 0 r_0=0.09<R_0 r 0 = 0.09 < R 0 covers this initialization uniformly.
Since the noise acts additively on ( x , y ) (x,y) ( x , y ) and V x x + V y y = 2 V_{xx}+V_{yy}=2 V xx + V y y = 2 , the averaged generator and quadratic-variation rate satisfy
L σ V ≤ − 0.1549 0.749 V + σ 2 , q σ ( V ) = σ 2 [ ( x − ϱ / 4 ) 2 + y 2 ] ≤ 2 σ 2 V (48) \mathcal{L}_{\sigma}V\le-\frac{0.1549}{0.749}V+\sigma^2,\qquad
q_{\sigma}(V)=\sigma^2\left[(x-\varrho/4)^2+y^2\right]\le2\sigma^2V \tag*{(48)} L σ V ≤ − 0.749 0.1549 V + σ 2 , q σ ( V ) = σ 2 [ ( x − ϱ /4 ) 2 + y 2 ] ≤ 2 σ 2 V (48)
on the certified sublevel. These estimates verify the regional target certificate for Theorem 1 and the averaged bounds used in Lemma 1.
Simulation protocol
Simulations use Euler–Maruyama integration with nominal parameters a = b = I 0 = 1 a=b=I_0=1 a = b = I 0 = 1 , d = 0.25 d=0.25 d = 0.25 , and h 0 = θ 0 = 0 h_0=\theta_0=0 h 0 = θ 0 = 0 , and the initialization specified above. Table I summarizes the ensemble sizes and simulation horizons. Paired dual-mode and proportional-only runs use identical Wiener increments.
Table I. Simulation settings. N N N is the number of paths and T T T is the normalized time horizon.
Curves labelled “period averaged” are moving averages of the simulated paths, rather than trajectories of the averaged diffusion (43). Shaded bands span the pointwise empirical fifth to ninety-fifth percentiles. Proportional-only equilibrium references are computed from (44).
Regulation, offset compensation, and frequency dependence
For σ = 0.05 \sigma=0.05 σ = 0.05 and ω = 200 \omega=200 ω = 200 , the filtered mean positions settle near the spot center and the mean bias approaches u ∗ = − 0.25 u^*=-0.25 u ∗ = − 0.25 (Figure 1 ). The cost remains near its minimum, with residual fluctuations from probing and Brownian motion.
Figure 1. Nominal trap response over 128 paths, showing positions, bias, and cost. Thin curves show one realization and thick curves show period-filtered ensemble means. Shading spans pointwise 5–95% empirical quantiles of the unfiltered paths. Dotted lines mark nominal targets.
In Figure 2 , the filtered proportional-only means agree with the equilibria predicted by (44), approximately 0.0837 for d = 0.25 d=0.25 d = 0.25 and 0.3646 for d = 1 d=1 d = 1 . The dual-mode means remain near zero. The d = 1 d=1 d = 1 case is a stress test outside the certified uncertainty family.
Figure 2. Dual-mode and matched frozen-bias simulations for d = 0.25 d=0.25 d = 0.25 and d = 1 d=1 d = 1 . Upper panels: filtered ensemble mean and empirical 5–95% bands, with the stable proportional-only equilibrium indicated. Lower panels: filtered particle positions over t ∈ [ 10 , 20 ] t \in[10,20] t ∈ [ 10 , 20 ] . Circles mark normalized radii 0.5 and 1 for visualization, not a stopping criterion used in these runs. The d = 1 d=1 d = 1 case is outside the certified offset family.
Figure 3 shows regulation for either actuator sign, with the bias approaching + 0.25 +0.25 + 0.25 when b = − 1 b=-1 b = − 1 . For nominal b = 1 b=1 b = 1 , the frequency sweep uses ω ∈ { 25 , 50 , 100 , 200 , 400 , 800 } \omega\in\{25,50,100,200,400,800\} ω ∈ { 25 , 50 , 100 , 200 , 400 , 800 } and σ ∈ { 0 , 0.05 , 0.1 } \sigma\in\{0,0.05,0.1\} σ ∈ { 0 , 0.05 , 0.1 } . The reported error averages ∣ Z − z ∗ ∣ 2 \lvert Z-z^*\rvert^2 ∣ Z − z ∗ ∣ 2 over paths and t ∈ [ 10 , 20 ] t\in[10,20] t ∈ [ 10 , 20 ] , where Z = col ( x , y , u b ) Z=\operatorname{col}(x,y,u_b) Z = col ( x , y , u b ) and z ∗ = col ( 0 , 0 , − 0.25 ) z^*=\operatorname{col}(0,0,-0.25) z ∗ = col ( 0 , 0 , − 0.25 ) .
Figure 3. Left and center: one-period-filtered position and bias for both actuator signs. Right: reported sample/time mean squared augmented error over t ∈ [ 10 , 20 ] t \in[10,20] t ∈ [ 10 , 20 ] . Solid curves use the stored raw state and dotted curves use its moving average. The dashed line is the leading ripple energy (49) . The frequency sweep uses one noise-free path or 32 noisy paths per point, not 128.
The leading-order ripple energy is
E r i p p l e ( ω ) = κ P b 2 + κ I 2 / κ P ω = 4.25 ω . (49) E_{\mathrm{ripple}}(\omega)=\frac{\kappa_Pb^2+\kappa_I^2/\kappa_P}{\omega}=\frac{4.25}{\omega}. \tag*{(49)} E ripple ( ω ) = ω κ P b 2 + κ I 2 / κ P = ω 4.25 . (49)
The nominal linearized averaged diffusion has stationary covariance P l i n = σ 2 diag ( 1 / 6 , 1 / 2 , 1 / 6 ) P_{\mathrm{lin}}=\sigma^2\operatorname{diag}(1/6,1/2,1/6) P lin = σ 2 diag ( 1/6 , 1/2 , 1/6 ) . Adding its trace gives the local reference 4.25 / ω + 5 σ 2 / 6 4.25/\omega+5\sigma^2/6 4.25/ ω + 5 σ 2 /6 , which agrees with the simulated errors at ω = 200 \omega=200 ω = 200 (Table II) and supports the ripple–noise interpretation.
Table 2. Raw mean squared error at ω = 200 \omega=200 ω = 200 and the local analytical reference.
Large-offset capture
The offsets d = 1.8 d=1.8 d = 1.8 and 2.2 2.2 2.2 lie outside the certified uncertainty family. In 64-path runs to T = 12 T=12 T = 12 , all proportional-only paths exit the strip ∣ x ∣ < 1.8 |x|<1.8 ∣ x ∣ < 1.8 at both offsets. The dual-mode exit fractions are 0% and approximately 19%, respectively (Figure 4 ), illustrating improved transient capture under large offsets.
Figure 4. Large-offset stress tests, 64 paths per case. Left: medians of the one-period-filtered, grid-stopped positions. “Lost” denotes first exit from ∣ x ∣ < 1.8 |x| < 1.8 ∣ x ∣ < 1.8 by T = 12 T = 12 T = 12 and is rounded to a whole percentage. Stopped paths remain in the plotted medians. Right: positive-axis intersections for the proportional-only average, illustrating disappearance of its stable branch. This panel does not display every equilibrium on R \mathbb{R} R .
Paths are frozen at their first recorded exit. The plotted medians include these paths and moving-average post-processing, so they describe the full stopped ensemble rather than surviving paths alone. The right panel shows only x ≥ 0 x \ge0 x ≥ 0 and omits the unstable equilibria at x < 0 x<0 x < 0 .
The simulations assume ideal stagnation-point actuation and instantaneous intensity feedback. At the nominal settings, the probing signal requires a stagnation-point displacement amplitude of 40 w 40w 40 w at frequency ω ϵ 0 / ( 2 π ) \omega\epsilon_0/(2\pi) ω ϵ 0 / ( 2 π ) . A physical implementation must accommodate this excursion and probing frequency.
Conclusion
A dual-mode extremum-seeking controller with cosine modulation was analyzed for unknown input-affine controlled diffusions. An exact deterministic periodic response separates probing ripple from process noise in the Lyapunov analysis. Regional moment bounds and exponential capture and residence estimates establish practical regulation from certified initial sets over horizons that grow as the noise decreases. Under an additional incremental condition, the same-noise approximation is uniform in mean square before common exit. The hydrodynamic trap example verifies a regional target certificate and illustrates regulation and compensation of unknown offsets using scalar optical feedback. Future work will examine the certified frequency requirement under sensing and actuation constraints.
Verification of the Regional Trap Certificate
Let k = b 2 I 0 k=b^2I_0 k = b 2 I 0 and φ = e − ( x 2 + y 2 ) / 2 \varphi=e^{-(x^2+y^2)/2} φ = e − ( x 2 + y 2 ) /2 . Differentiating (45) along the noise-free dynamics in (43) gives
V ˙ = − a y 2 − [ x φ ] M ( a , k , φ ) [ x φ ] (50) \dot{V}=-ay^2-\begin{bmatrix}x & \varphi\end{bmatrix}M(a,k,\varphi)\begin{bmatrix}x\\\varphi\end{bmatrix} \tag*{(50)} V ˙ = − a y 2 − [ x φ ] M ( a , k , φ ) [ x φ ] (50)
where c x = 1 − a / 4 + ( k − 1 ) φ c_{\mathrm{x}} = 1 - a/4 + (k - 1)\varphi c x = 1 − a /4 + ( k − 1 ) φ and
M ( a , k , φ ) = [ 15 4 k φ − a − c x / 2 − c x / 2 1 4 ] . (51) M(a,k,\varphi) =
\begin{bmatrix}
\frac{15}{4}k\varphi-a & -c_{\mathrm{x}}/2 \\
-c_{\mathrm{x}}/2 & \frac{1}{4}
\end{bmatrix}. \tag*{(51)} M ( a , k , φ ) = [ 4 15 k φ − a − c x /2 − c x /2 4 1 ] . (51)
On ∣ ξ ∣ < 0.7 |\xi| < 0.7 ∣ ξ ∣ < 0.7 , φ ≥ e − 0.245 > 0.7827 \varphi\ge e^{-0.245} > 0.7827 φ ≥ e − 0.245 > 0.7827 . The matrix M M M is affine in each variable over the box a ∈ [ 0.8 , 1.2 ] a \in[0.8,1.2] a ∈ [ 0.8 , 1.2 ] , k ∈ [ 0.729 , 1.331 ] k \in[0.729,1.331] k ∈ [ 0.729 , 1.331 ] , and φ ∈ [ 0.7827 , 1 ] \varphi\in[0.7827,1] φ ∈ [ 0.7827 , 1 ] , so every value is a convex combination of its eight vertex matrices. At each vertex, exact rational evaluation gives M 11 − 0.1549 > 0.7848 M_{11} - 0.1549 > 0.7848 M 11 − 0.1549 > 0.7848 and det ( M − 0.1549 I 2 ) > 3.02 × 10 − 4 \det(M - 0.1549I_2) > 3.02 \times10^{-4} det ( M − 0.1549 I 2 ) > 3.02 × 1 0 − 4 . Hence M ⪰ 0.1549 I 2 M \succeq0.1549I_2 M ⪰ 0.1549 I 2 throughout the box. Since a ≥ 0.8 a \ge0.8 a ≥ 0.8 , (50) yields
V ˙ ≤ − 0.1549 ∣ ξ ∣ 2 \dot{V} \le-0.1549|\xi|^2 V ˙ ≤ − 0.1549∣ ξ ∣ 2
throughout ∣ ξ ∣ < 0.7 |\xi| < 0.7 ∣ ξ ∣ < 0.7 , proving the required decay bound.
Writing V ( ξ ) = ξ T P ( k ) ξ V(\xi) = \xi^{\mathsf{T}}P(k)\xi V ( ξ ) = ξ T P ( k ) ξ gives
P ( k ) = 1 2 [ 1 0 − 1 4 0 1 0 − 1 4 0 1 / k ] . P(k) = \frac{1}{2}
\begin{bmatrix}
1 & 0 & -\frac{1}{4} \\
0 & 1 & 0 \\
-\frac{1}{4} & 0 & 1/k
\end{bmatrix}. P ( k ) = 2 1 1 0 − 4 1 0 1 0 − 4 1 0 1/ k .
Since P P P is affine in 1 / k 1/k 1/ k , endpoint checks using Sylvester’s criterion give 0.298 I 3 ≺ P ( k ) ≺ 0.749 I 3 0.298I_3 \prec P(k) \prec0.749I_3 0.298 I 3 ≺ P ( k ) ≺ 0.749 I 3 , proving (47) . Consequently, { V ≤ 0.14 } ⊂ { ∣ ξ ∣ < 0.7 } \{V \le0.14\} \subset\{|\xi| < 0.7\} { V ≤ 0.14 } ⊂ { ∣ ξ ∣ < 0.7 } .
At the common initialization, V 0 = 0.025 + d 2 / ( 2 k ) − 0.05 d V_0 = 0.025 + d^2/(2k) - 0.05d V 0 = 0.025 + d 2 / ( 2 k ) − 0.05 d is maximized at d = − 0.25 d = -0.25 d = − 0.25 and k = 0.729 k = 0.729 k = 0.729 , giving V 0 < 0.080367 < r 0 = 0.09 V_0 < 0.080367 < r_0 = 0.09 V 0 < 0.080367 < r 0 = 0.09 . For e = col ( x , y , u b − u ∗ ) e = \operatorname{col}(x,y,u_b-u^*) e = col ( x , y , u b − u ∗ ) , the transformation ξ = diag ( 1 , 1 , b ) e \xi= \operatorname{diag}(1,1,b)e ξ = diag ( 1 , 1 , b ) e gives 0.81 ∣ e ∣ 2 ≤ ∣ ξ ∣ 2 ≤ 1.21 ∣ e ∣ 2 0.81|e|^2 \le|\xi|^2 \le1.21|e|^2 0.81∣ e ∣ 2 ≤ ∣ ξ ∣ 2 ≤ 1.21∣ e ∣ 2 and hence the certificate constants stated in the main text.
The additive-noise Itô correction is σ 2 \sigma^2 σ 2 . Moreover,
2 V − [ ( x − ϱ / 4 ) 2 + y 2 ] = ( 1 k − 1 16 ) ϱ 2 ≥ 0 , 2V-\left[(x-\varrho/4)^2+y^2\right]=\left(\frac{1}{k}-\frac{1}{16}\right)\varrho^2\ge0, 2 V − [ ( x − ϱ /4 ) 2 + y 2 ] = ( k 1 − 16 1 ) ϱ 2 ≥ 0 ,
so q σ ( V ) ≤ 2 σ 2 V q_{\sigma}(V)\le2\sigma^2V q σ ( V ) ≤ 2 σ 2 V . Together with the decay estimate, these bounds establish (47) and complete the regional certificate for Theorem 1. □
Passive-Axis Bound on Radial Residence
For L > 0 L>0 L > 0 , define τ s p o t ( L ) = inf { t ≥ 0 : x t 2 + y t 2 ≥ L 2 } \tau_{\mathrm{spot}}(L)=\inf\{t\ge0:x_t^2+y_t^2\ge L^2\} τ spot ( L ) = inf { t ≥ 0 : x t 2 + y t 2 ≥ L 2 } and τ y ( L ) = inf { t ≥ 0 : ∣ y t ∣ ≥ L } \tau_y(L)=\inf\{t\ge0:|y_t|\ge L\} τ y ( L ) = inf { t ≥ 0 : ∣ y t ∣ ≥ L } . These satisfy τ s p o t ( L ) ≤ τ y ( L ) \tau_{\mathrm{spot}}(L)\le\tau_y(L) τ spot ( L ) ≤ τ y ( L ) pathwise.
For σ > 0 \sigma>0 σ > 0 , the mean exit time m ( y ) = E y τ y ( L ) m(y)=\mathbb{E}_y\tau_y(L) m ( y ) = E y τ y ( L ) of the passive coordinate solves σ 2 2 m ′ ′ − a y m ′ = − 1 \frac{\sigma^2}{2}m''-aym'=-1 2 σ 2 m ′′ − a y m ′ = − 1 with m ( ± L ) = 0 m(\pm L)=0 m ( ± L ) = 0 . Integration gives
E τ s p o t ( L ) ≤ E τ y ( L ) = π a ∫ a ∣ y 0 ∣ / σ a L / σ e s 2 erf ( s ) d s \mathbb{E}\tau_{\mathrm{spot}}(L)\le\mathbb{E}\tau_y(L)=\frac{\sqrt{\pi}}{a}\int_{\sqrt{a}|y_0|/\sigma}^{\sqrt{a}L/\sigma}e^{s^2}\operatorname{erf}(s)\,ds E τ spot ( L ) ≤ E τ y ( L ) = a π ∫ a ∣ y 0 ∣/ σ a L / σ e s 2 erf ( s ) d s
for an initial state inside the disk. This bound is independent of the controller. An optical intensity threshold I min ∈ ( 0 , I 0 ) I_{\min}\in(0,I_0) I m i n ∈ ( 0 , I 0 ) corresponds to L = 2 log ( I 0 / I min ) L=\sqrt{2\log(I_0/I_{\min})} L = 2 log ( I 0 / I m i n ) .