betaFeedback

WorksMathematics

Small-Gain Analysis of Robust Event-Triggered Extremum Seeking for Scalar Maps

Abstract

We study a sample-and-hold implementation of gradient extremum seeking for a scalar quadratic map in which the control signal is refreshed only when an event occurs. The event generator acts on the demodulated gradient signal as it is actually measured, rather than on its averaged version, and contains a strictly positive offset, following the small-gain philosophy for robust event-triggered control. Rather than invoking averaging theorems, we work directly with the exact dynamics of the parameter estimation error. In the scalar setting, the demodulated gradient has an explicit sinusoidal structure, which is exploited through a change of variables that absorbs the zero-mean oscillatory perturbation, together with a quadratic Lyapunov function and the persistence of excitation of the probing signal. For a sufficiently fast dither, the estimation error is shown to decay exponentially to a residual neighborhood of the origin whose size is quantified by the dither amplitude and frequency, and by the relative threshold and the offset of the event generator. All constants, including the input-to-state stability gain of the extremum seeking loop, the admissible dither frequency and the positive lower bound on the time between two consecutive events, are given in closed form, so that Zeno behavior is excluded. A simulation example illustrates the results.

Introduction

Extremum seeking (ES) is a non-model-based adaptive technique that drives the input of an uncertain map toward the point where the output is optimal, using nothing more than output measurements and a periodic probing signal [1, 2, 3]. After the stability proof in [2], the method has been extended in several directions, including multi-parameter and Newton-type schemes [4], maps with delays and distributed-parameter actuation [5, 6], and applications such as source seeking with autonomous vehicles [7].

When the optimizer runs on an embedded or networked platform, refreshing the control input continuously or at a fixed fast rate squanders bandwidth and processor time. Event-triggered control replaces time-driven updates by state-driven ones: the actuator value is recomputed only when a prescribed triggering condition fails [8, 9]. The idea has been developed, among many other settings, for nonlinear systems with dynamic triggers [10]. Bringing event-triggered updates into ES is more recent: the gradient scheme of [11] was followed by a Newton-type version [12], by a discrete-time multivariable formulation [13], and by game-theoretic settings in which several players seek a Nash equilibrium [14]. All of these works share one methodological choice: the trigger is designed on the average model, and the behavior of the true oscillating loop is inferred afterwards from averaging theorems.

Here we take a different path, motivated by the small-gain approach to robust event-triggered control of nonlinear systems [15, 16], which rests on input-to-state stability (ISS) and on the small-gain theory for interconnections [17]. We regard the event-triggered ES loop as the feedback interconnection of two blocks: the dynamics of the parameter estimation error, and the event generator. The sampling error and the oscillations produced by the dither enter this interconnection as disturbances. We focus on the scalar (single-input) case, for which the demodulated gradient admits a fully explicit sinusoidal decomposition and all the constants of the analysis can be written in closed form. The contributions are:

  • a static triggering rule with a positive offset, evaluated on the measured gradient estimate itself (not on its average), which tolerates the persistent oscillatory perturbations that are intrinsic to ES;

  • an ISS estimate of the ES dynamics with respect to the sampling error, derived on the exact error equation, without averaging, through the sign structure and the persistence of excitation of the demodulated gradient;

  • a small-gain condition coupling the ISS gain of the ES dynamics with the gain of the event generator, which yields local exponential practical stability and an explicit residual set;

  • a closed-form, uniform lower bound on the inter-event times, which rules out Zeno behavior.

The results are the scalar specialization of a multivariable theory developed by the authors; in the scalar case the extra assumptions on probing frequencies and gain structure disappear and every constant becomes transparent.

Notation: ∣⋅∣|\cdot| is the absolute value, sgn⁡(⋅)\operatorname{sgn}(\cdot) the sign function and N={0,1,2,…}\mathbb{N} = \{0,1,2,\ldots\}. For a function ff and an interval II, sup⁡I∣f∣:=sup⁡τ∈I∣f(τ)∣\sup_I |f| := \sup_{\tau\in I}|f(\tau)|.

Problem Statement

We consider the scalar static map

y(t)=Q(θ(t))=Q∗+H∗2(θ(t)−θ∗)2,(1)y(t) = Q(\theta(t)) = Q^* + \frac{H^*}{2}\left(\theta(t)-\theta^*\right)^2, \tag*{(1)}

where θ(t)∈R\theta(t) \in\mathbb{R} is the input that can be manipulated in real time, y(t)∈Ry(t) \in\mathbb{R} is the measured output, θ∗∈R\theta^{*} \in\mathbb{R} is the optimizer, Q∗=Q(θ∗)Q^{*} = Q(\theta^{*}) is the optimal value and H∗∈RH^{*} \in\mathbb{R} is the curvature (second derivative) of the map.

Assumption 1: The constants θ∗\theta^{*}, Q∗Q^{*} and H∗H^{*} are unknown, but H∗≠0H^{*} \ne0 and sgn⁡(H∗)\operatorname{sgn}(H^{*}) is known.

Since any smooth map with a nondegenerate extremum agrees with (1) up to second order near θ∗\theta^{*}, the quadratic model captures the local behavior of general scalar ES problems while keeping the demodulated signal explicitly computable. When H∗>0H^{*} > 0 the task is to minimize QQ, and when H∗<0H^{*} < 0 the task is to maximize it; in both cases θ(t)\theta(t) must be steered to a small vicinity of θ∗\theta^{*}. We write

h:=∣H∗∣>0.(2)h := \lvert H^{*}\rvert> 0. \tag*{(2)}

Event-Triggered Extremum Seeking

Probing, Demodulation and Error Variable

The scheme is depicted in Figure 1. A single sinusoid is used both to perturb the input and to demodulate the output [3]:

Figure 1. Robust event-triggered extremum seeking for a scalar map. The event generator monitors the measured gradient estimate G^(t)\hat{G}(t) and releases a new sample G^(tk)\hat{G}(t_k) to the zero-order hold (ZOH) only when the triggering condition is violated.

S(t)=M(t)=asin⁡(ωt),a>0, ω>0,(3)S(t) = M(t) = a\sin(\omega t), \qquad a > 0,\ \omega> 0, \tag*{(3)}

with period

T:=2πω.(4)T := \frac{2\pi}{\omega}. \tag*{(4)}

The dither is superimposed on the current estimate of the optimizer, θ(t)=θ^(t)+S(t)\theta(t) = \hat{\theta}(t) + S(t), where θ^(t)\hat{\theta}(t) is the state of an integrator. With the estimation error

θ~(t):=θ^(t)−θ∗,(5)\tilde{\theta}(t) := \hat{\theta}(t) - \theta^{*}, \tag*{(5)}

one has θ(t)−θ∗=θ~(t)+S(t)\theta(t) - \theta^{*} = \tilde{\theta}(t) + S(t) and, from (1),

y(t)=Q∗+H∗2θ~2(t)+H∗S(t)θ~(t)+H∗2S2(t).(6)y(t) = Q^{*} + \frac{H^{*}}{2}\tilde{\theta}^{2}(t) + H^{*}S(t)\tilde{\theta}(t) + \frac{H^{*}}{2}S^{2}(t). \tag*{(6)}

Structure of the Gradient Estimate

The gradient estimate is obtained by demodulation,

G^(t):=M(t)y(t).(7)\hat{G}(t) := M(t)y(t). \tag*{(7)}

Multiplying (6) by M(t)=S(t)=asin⁡(ωt)M(t) = S(t) = a\sin(\omega t) and using the identities 2sin⁡2(ωt)=1−cos⁡(2ωt)2\sin^{2}(\omega t) = 1 - \cos(2\omega t) and 4sin⁡3(ωt)=3sin⁡(ωt)−sin⁡(3ωt)4\sin^{3}(\omega t) = 3\sin(\omega t) - \sin(3\omega t), we arrive at the exact decomposition

G^(t)=a22H∗β(t)θ~(t)+dq(t,θ~(t))+Δ(t),(8)\hat{G}(t) = \frac{a^{2}}{2}H^{*}\beta(t)\tilde{\theta}(t) + d_q(t,\tilde{\theta}(t)) + \Delta(t), \tag*{(8)}

where

β(t):=1−cos⁡(2ωt)=2sin⁡2(ωt)∈[0,2],(9)\beta(t) := 1 - \cos(2\omega t) = 2\sin^{2}(\omega t) \in[0,2], \tag*{(9)}
dq(t,θ~):=a2H∗sin⁡(ωt)θ~2,(10)d_q(t,\tilde{\theta}) := \frac{a}{2}H^{*}\sin(\omega t)\tilde{\theta}^{2}, \tag*{(10)}
Δ(t):=S(t)(Q∗+H∗2S2(t))=a(Q∗+38H∗a2)sin⁡(ωt)−H∗a38sin⁡(3ωt).(11)\Delta(t) := S(t)\left(Q^{*} + \frac{H^{*}}{2}S^{2}(t)\right) = a\left(Q^{*} + \frac{3}{8}H^{*}a^{2}\right)\sin(\omega t) - \frac{H^{*}a^{3}}{8}\sin(3\omega t). \tag*{(11)}

The first term in (8) is the useful part of the estimate: it is linear in θ~\tilde{\theta} and its coefficient has the sign of H∗H^{*} and average value a22H∗\frac{a^{2}}{2}H^{*} over one period. The term dqd_q is quadratic in the error and vanishes at θ~=0\tilde{\theta} = 0. The term Δ(t)\Delta(t) does not depend on θ~\tilde{\theta} and does not vanish: it is a persistent oscillation generated by the dither and by the offset Q∗Q^{*} of the map. From (11),

∣Δ(t)∣≤dˉ:=a(∣Q∗∣+a2h2),∀t≥0.(12)\lvert\Delta(t)\rvert\le\bar{d} := a\left(\lvert Q^{*}\rvert+ \frac{a^{2}h}{2}\right), \qquad\forall t \ge0. \tag*{(12)}

Moreover, since ∣β(t)∣≤2\lvert\beta(t)\rvert\le2 and ∣sin⁡(ωt)∣≤1\lvert\sin(\omega t)\rvert\le1,

∣G^(t)∣≤ℓ∣θ~(t)∣+ah2θ~2(t)+dˉ,ℓ:=a2h,(13)\lvert\hat{G}(t)\rvert\le\ell\lvert\tilde{\theta}(t)\rvert+ \frac{ah}{2}\tilde{\theta}^{2}(t) + \bar{d}, \qquad\ell:= a^{2}h, \tag*{(13)}
∣dq(t,θ~)∣≤ah2θ~2.(14)\lvert d_q(t,\tilde{\theta})\rvert\le\frac{ah}{2}\tilde{\theta}^{2}. \tag*{(14)}

The next lemma gathers the three properties of the probing signal on which the whole analysis rests.

Lemma 1: For all t≥0t \ge0, the following statements hold.

(i) (Sign structure) β(t)≥0\beta(t) \ge0 and β(t)≤2\beta(t) \le2.

(ii) (Persistence of excitation) ∫tt+Tβ(τ) dτ=T\int_{t}^{t+T}\beta(\tau)\,\mathrm{d}\tau= T.

(iii) (Bounded primitive) The TT-periodic zero-mean primitive of Δ\Delta, namely

ΨΔ(t):=−aω(Q∗+38H∗a2)cos⁡(ωt)+H∗a324ωcos⁡(3ωt),(15)\Psi_{\Delta}(t) := -\frac{a}{\omega}\left(Q^{*} + \frac{3}{8}H^{*}a^{2}\right)\cos(\omega t) + \frac{H^{*}a^{3}}{24\omega}\cos(3\omega t), \tag*{(15)}

satisfies Ψ˙Δ(t)=Δ(t)\dot{\Psi}_{\Delta}(t) = \Delta(t) and

∣ΨΔ(t)∣≤κΔω,κΔ:=a∣Q∗∣+512ha3.(16)\lvert\Psi_{\Delta}(t)\rvert\le\frac{\kappa_{\Delta}}{\omega}, \qquad\kappa_{\Delta} := a\lvert Q^{*}\rvert+ \frac{5}{12}ha^{3}. \tag*{(16)}

Proof: Items (i) and (ii) follow from β(t)=2sin⁡2(ωt)\beta(t) = 2\sin^{2}(\omega t) and from ∫tt+Tcos⁡(2ωτ) dτ=0\int_{t}^{t+T}\cos(2\omega\tau)\,\mathrm{d}\tau= 0, since 2ωT=4π2\omega T = 4\pi. Item (iii) is obtained by integrating each sinusoid in (11) termwise and bounding ∣Q∗+38H∗a2∣≤∣Q∗∣+38ha2\lvert Q^{*} + \frac{3}{8}H^{*}a^{2}\rvert\le\lvert Q^{*}\rvert+ \frac{3}{8}ha^{2}, so that ∣ΨΔ∣≤aω(∣Q∗∣+38ha2+124ha2)\lvert\Psi_{\Delta}\rvert\le\frac{a}{\omega}(\lvert Q^{*}\rvert+ \frac{3}{8}ha^{2} + \frac{1}{24}ha^{2}).

Items (i) and (ii) state that the instantaneous adaptation gain never has the wrong sign and has a strictly positive mean; together they replace the averaging step of the classical analysis.

Sample-and-Hold Emulation

The adaptation gain is a scalar k∈Rk \in\mathbb{R} with

sgn⁡(k)=−sgn⁡(H∗),kˉ:=∣k∣.(17)\operatorname{sgn}(k) = -\operatorname{sgn}(H^{*}), \qquad\bar{k} := \lvert k\rvert. \tag*{(17)}

so that kH∗=−kˉhkH^*=-\bar{k}h: a negative gain is used for minimization and a positive one for maximization. The estimator is the integrator θ^˙(t)=u(t)\dot{\hat{\theta}}(t)=u(t) and, since θ∗\theta^* is constant, θ~˙(t)=u(t)\dot{\tilde{\theta}}(t)=u(t). The continuous-time law u=kG^u=k\hat{G} is emulated with a zero-order hold: the control value is recomputed only at the instants (tk)k∈N(t_k)_{k\in\mathbb{N}}, t0=0t_0=0, delivered by the event generator,

u(t)=kG^(tk),t∈[tk,tk+1),k∈N.(18)u(t)=k\hat{G}(t_k),\qquad t\in[t_k,t_{k+1}),\quad k\in\mathbb{N}. \tag*{(18)}

The mismatch between the held and the current estimates is the sampling error

e(t):=G^(tk)−G^(t),t∈[tk,tk+1),k∈N.(19)e(t):=\hat{G}(t_k)-\hat{G}(t),\qquad t\in[t_k,t_{k+1}),\quad k\in\mathbb{N}. \tag*{(19)}

Collecting (8), (18) and (19), the estimation error obeys, on each interval [tk,tk+1)[t_k,t_{k+1}),

θ~˙(t)=−kˉha22β(t)θ~(t)+kΔ(t)+kχ(t),(20)\dot{\tilde{\theta}}(t)=-\frac{\bar{k}ha^2}{2}\beta(t)\tilde{\theta}(t)+k\Delta(t)+k\chi(t), \tag*{(20)}

where the lumped input

χ(t):=e(t)+dq(t,θ~(t))(21)\chi(t):=e(t)+d_q(t,\tilde{\theta}(t)) \tag*{(21)}

collects the sampling error, which is shaped by the event generator, and the quadratic term, which vanishes at the origin. (20) is exact. Observe that dqd_q appears twice: inside χ\chi, where it drives the error, and inside G^\hat{G}, which is the signal the event generator watches.

Triggering Rule with Offset

Because Δ(t)\Delta(t) does not vanish at θ~=0\tilde{\theta}=0, neither does G^(t)\hat{G}(t); in fact G^(t)\hat{G}(t) crosses zero at every half period. A purely relative rule of the form ∣e∣≤σ∣G^∣|e|\leq\sigma|\hat{G}| would therefore force events to accumulate near those zero crossings. This is avoided by adding a positive offset, in the spirit of [16 Sec. 3.4.1].

Definition 1 (Static event generator with offset): Given σ>0\sigma>0 and ε>0\varepsilon>0, the sequence of event times {tk}k∈N\{t_k\}_{k\in\mathbb{N}} is defined by t0=0t_0=0 and

tk+1:=inf⁡{t>tk:∣e(t)∣>σ∣G^(t)∣+ε},(22)t_{k+1}:=\inf\left\{t>t_k:|e(t)|>\sigma|\hat{G}(t)|+\varepsilon\right\}, \tag*{(22)}

with the convention tk+1=+∞t_{k+1}=+\infty when the set is empty. At each tkt_k the control value (18) is refreshed.

Implementing (22) requires only G^(t)\hat{G}(t) and its last transmitted value G^(tk)\hat{G}(t_k); θ∗\theta^*, Q∗Q^* and H∗H^* play no role in it. By construction, along the solutions,

∣e(t)∣≤σ∣G^(t)∣+ε,∀t≥0.(23)|e(t)|\leq\sigma|\hat{G}(t)|+\varepsilon,\qquad\forall t\geq0. \tag*{(23)}

The event generator as a gain. Combining (23) with (13) and (14), the input (21) satisfies

∣χ(t)∣≤σℓ∣θ~(t)∣+(1+σ)ah2θ~2(t)+δ,δ:=σˉdˉ+ε,|\chi(t)|\leq\sigma\ell|\tilde{\theta}(t)|+\frac{(1+\sigma)ah}{2}\tilde{\theta}^2(t)+\delta,\qquad\delta:=\bar{\sigma}\bar{d}+\varepsilon,

the factor 1+σ1+\sigma being the price of the twofold appearance of dqd_q noted above. Consequently, for any R>0R>0, as long as ∣θ~(t)∣≤R|\tilde{\theta}(t)|\leq R,

∣χ(t)∣≤ρR∣θ~(t)∣+δ,ρR:=σa2h+(1+σ)ah2R.(24)|\chi(t)|\leq\rho_R|\tilde{\theta}(t)|+\delta,\qquad\rho_R:=\sigma a^2h+\frac{(1+\sigma)ah}{2}R. \tag*{(24)}

Following [15, 16], the event generator together with the quadratic term is viewed as an operator from θ~\tilde{\theta} to χ\chi with linear gain ρR\rho_R and offset δ\delta on the ball of radius RR. Feeding χ\chi into (20), which returns θ~\tilde{\theta}, closes the loop of Fig. 2. Section IV quantifies the gain of each block and shows that, when their product is small enough, the interconnection is stable.

Main Results

Throughout, we use the excitation level

λ:=a2h>0,(25)\lambda:=a^2h>0, \tag*{(25)}

which is the mean value over one period of a2hβ(t)a^2h\beta(t); equivalently, kˉλ/2\bar{k}\lambda/2 is the average damping coefficient of (20).

ISS of the Extremum Seeking Dynamics

We first treat χ\chi in (20) as an arbitrary exogenous signal, disregarding how it is produced. In the framework of [15, 16], the resulting estimate is the property that the plant is assumed to possess; here it is derived rather than assumed. Item (a) below is an ISS-Lyapunov inequality for the discrete-time system obtained by sampling the Lyapunov function once per dither period, in the sense of [18].

Lemma 2 (ISS with respect to χ\chi): Let Assumption 1 hold and let the gain satisfy (17). Consider (20) driven by an arbitrary piecewise continuous input χ\chi, and define

z(t):=θ~(t)−kΨΔ(t),W(t):=z2(t),(26)z(t):=\tilde{\theta}(t)-k\Psi_{\Delta}(t),\qquad W(t):=z^2(t), \tag*{(26)}
m0:=kˉλ8,γe:=48λ,δω:=kˉκΔω,νˉ:=kˉ2κΔλ.(27)m_0:=\frac{\bar{k}\lambda}{8},\qquad\gamma_e:=\frac{48}{\lambda},\qquad\delta_\omega:=\frac{\bar{k}\kappa_{\Delta}}{\omega},\qquad\bar{\nu}:=\bar{k}^2\frac{\kappa_{\Delta}}{\lambda}. \tag*{(27)}

If

ω≥ω0:=16kˉkˉλ,(28)\omega\geq\omega_0:=16\bar{k}\bar{k}\lambda, \tag*{(28)}

then, for all t≥0t\geq0:

(a) (One-period ISS-Lyapunov inequality)

W(t+T)≤(1−m0T)W(t)+T64kˉληt2,(29)W(t+T)\leq(1-m_0T)W(t)+T\frac{64}{\bar{k}\lambda}\eta_t^2, \tag*{(29)}

with ηt:=kˉsup⁡[t,t+T]∣χ∣+νˉ\eta_t:=\bar{k}\sup_{[t,t+T]}|\chi|+\bar{\nu};

(b) (ISS estimate)

∣θ~(t)∣≤4exp⁡(−m02t)∣θ~(0)∣+γesup⁡[0,t]∣χ∣+53δω.(30)|\tilde{\theta}(t)|\leq4\exp\left(-\frac{m_0}{2}t\right)|\tilde{\theta}(0)|+\gamma_e\sup_{[0,t]}|\chi|+53\delta_\omega. \tag*{(30)}

Moreover, the right-hand side of (30) involves χ\chi only through its restriction to [0,t][0,t].

Proof: Step 1 (Removing the oscillatory forcing). Although kΔ(t)k\Delta(t) never dies out, it integrates to zero over each period. Since Ψ˙Δ=Δ\dot{\Psi}_{\Delta}=\Delta, the variable zz in (26) satisfies

z˙(t)=−kˉ2β(t)z(t)+kχ(t)+ν(t),(31)\dot{z}(t)=-\frac{\bar{k}}{2}\beta(t)z(t)+k\chi(t)+\nu(t), \tag*{(31)}
ν(t):=−kkˉλ2β(t)ΨΔ(t),\nu(t):=-\frac{k\bar{k}\lambda}{2}\beta(t)\Psi_{\Delta}(t),

where we used −kkˉa22βθ~=−kˉλ2β(z+kΨΔ)-\frac{k\bar{k}a^{2}}{2}\beta\tilde{\theta}=-\frac{\bar{k}\lambda}{2}\beta(z+k\Psi_{\Delta}). From Lemma 1(i) and (16),

∣ν(t)∣≤νˉ,∣θ~(t)−z(t)∣≤δω.(32)|\nu(t)|\le\bar{\nu},\qquad|\tilde{\theta}(t)-z(t)|\le\delta_{\omega}. \tag*{(32)}

Step 2 (Derivative of WW). Along (31), for τ∈[t,t+T]\tau\in[t,t+T],

W˙(τ)=−kˉq(τ)+2z(τ)(kχ(τ)+ν(τ))≤−kˉq(τ)+2ηt∣z(τ)∣,(33)\dot{W}(\tau)=-\bar{k}q(\tau)+2z(\tau)(k\chi(\tau)+\nu(\tau)) \le-\bar{k}q(\tau)+2\eta_t|z(\tau)|, \tag*{(33)}

with q(τ):=λβ(τ)z2(τ)≥0q(\tau):=\lambda\beta(\tau)z^{2}(\tau)\ge0. Since β\beta touches zero twice per period, the first term cannot provide a pointwise decrease of WW; the decrease emerges only after integration over a period, where Lemma 1(ii) can be exploited.

Step 3 (Decrease over one period). Let L0:=kˉλL_0:=\bar{k}\lambda. By (4) and (28), L0T=2πkˉλ/ω≤1/8L_0T=2\pi\bar{k}\lambda/\omega\le1/8. From (31) and Lemma 1(i), ∣z˙(τ)∣≤L0∣z(τ)∣+ηt|\dot{z}(\tau)|\le L_0|z(\tau)|+\eta_t on [t,t+T][t,t+T], so Grönwall’s inequality [19] and e1/8<2e^{1/8}<2 give

∣z(τ)∣≤2(∣z(t)∣+ηtT),(34)|z(\tau)|\le2\left(|z(t)|+\eta_tT\right), \tag*{(34)}
∣z(τ)−z(t)∣≤2L0T∣z(t)∣+2ηtT.(35)|z(\tau)-z(t)|\le2L_0T|z(t)|+2\eta_tT. \tag*{(35)}

For fixed τ\tau, x↦λβ(τ)∣x∣x\mapsto\lambda\beta(\tau)|x| is a seminorm, hence q(τ)≥12λβ(τ)z2(t)−λβ(τ)(z(τ)−z(t))2q(\tau)\ge\frac{1}{2}\lambda\beta(\tau)z^{2}(t)-\lambda\beta(\tau)(z(\tau)-z(t))^{2}. Integrating over [t,t+T][t,t+T], using β≤2\beta\le2, Lemma 1(ii), (35) and L0T≤1/8L_0T\le1/8,

∫tt+Tq(τ) dτ≥Tλ2z2(t)−16λT3(L02z2(t)+ηt2)≥Tλ4z2(t)−16λT3ηt2.(36)\begin{aligned} \int_t^{t+T}q(\tau)\,\mathrm{d}\tau &\ge\frac{T\lambda}{2}z^{2}(t)-16\lambda T^{3}\left(L_0^{2}z^{2}(t)+\eta_t^{2}\right)\\ &\ge\frac{T\lambda}{4}z^{2}(t)-16\lambda T^{3}\eta_t^{2}. \tag*{(36)} \end{aligned}

Also, from (34), ∫tt+T∣z∣dτ≤2T(∣z(t)∣+ηtT)\int_t^{t+T}|z|\mathrm{d}\tau\le2T(|z(t)|+\eta_tT). Integrating (33) and using 16kˉλT=16L0T≤216\bar{k}\lambda T=16L_0T\le2,

W(t+T)−W(t)≤−kˉλT4z2(t)+4ηtT∣z(t)∣+6ηt2T2.(37)W(t+T)-W(t)\le-\frac{\bar{k}\lambda T}{4}z^{2}(t)+4\eta_tT|z(t)|+6\eta_t^{2}T^{2}. \tag*{(37)}

By Young’s inequality xy≤x22ϖ+ϖy22xy\le\frac{x^{2}}{2\varpi}+\frac{\varpi y^{2}}{2} with x=4ηtx=4\eta_t, y=∣z(t)∣y=|z(t)| and ϖ=kˉλ/4\varpi=\bar{k}\lambda/4,

W(t+T)≤(1−kˉλT8)W(t)+T32kˉληt2+6T2ηt2,W(t+T)\le\left(1-\frac{\bar{k}\lambda T}{8}\right)W(t)+T\frac{32}{\bar{k}\lambda}\eta_t^{2}+6T^{2}\eta_t^{2},

and since 6T≤32/(kˉλ)6T\le32/(\bar{k}\lambda) (because L0T≤1/8<16/3L_0T\le1/8<16/3), inequality (29) follows.

Step 4 (ISS estimate). Fix t>0t>0 and set χˉ:=sup⁡[0,t]∣χ∣\bar{\chi}:=\sup_{[0,t]}|\chi|, ηˉ:=kˉχˉ+νˉ\bar{\eta}:=\bar{k}\bar{\chi}+\bar{\nu} and μ0:=64ηˉ2/(kˉλ)\mu_0:=64\bar{\eta}^{2}/(\bar{k}\lambda). Whenever [iT,(i+1)T]⊂[0,t][iT,(i+1)T]\subset[0,t], the window constant obeys ηiT≤ηˉ\eta_{iT}\le\bar{\eta}, and (29) becomes the linear recursion W((i+1)T)≤cW(iT)+Tμ0W((i+1)T)\le cW(iT)+T\mu_0, whose coefficient c:=1−m0Tc:=1-m_0T lies in (0,1)(0,1) because m0T=L0T/8≤1/64m_0T=L_0T/8\le1/64. Iterating, for all j∈Nj\in\mathbb{N} with jT≤tjT\le t,

W(jT)≤cjW(0)+μ0m0(1−cj)≤e−m0jTW(0)+μ0m0,(38)W(jT)\le c^{j}W(0)+\frac{\mu_0}{m_0}(1-c^{j})\le e^{-m_0jT}W(0)+\frac{\mu_0}{m_0}, \tag*{(38)}

where 1−x≤e−x1-x\le e^{-x} was used. Now take τ∈[jT,min⁡{(j+1)T,t}]\tau\in[jT,\min\{(j+1)T,t\}]. By (34) on [jT,τ][jT,\tau], (38), x+y≤x+y\sqrt{x+y}\le\sqrt{x}+\sqrt{y} and em0T/2≤2e^{m_0T/2}\le2,

∣z(τ)∣≤4e−m02τ∣z(0)∣+2μ0m0+2ηˉT.(39)|z(\tau)|\le4e^{-\frac{m_0}{2}\tau}|z(0)|+2\sqrt{\frac{\mu_0}{m_0}}+2\bar{\eta}T. \tag*{(39)}

Here 2μ0/m0=2512 ηˉ/(kˉλ)≤46ηˉ/(kˉλ)2\sqrt{\mu_0/m_0}=2\sqrt{512}\,\bar{\eta}/(\bar{k}\lambda)\le46\bar{\eta}/(\bar{k}\lambda) and, from L0T≤1L_0T\le1, 2kˉT≤2/λ2\bar{k}T\le2/\lambda. Substituting ηˉ=kˉχˉ+νˉ\bar{\eta}=\bar{k}\bar{\chi}+\bar{\nu}, the coefficient of χˉ\bar{\chi} in (39) is at most 48/λ=γe48/\lambda=\gamma_e, while the terms in νˉ\bar{\nu} add up to at most 46νˉ/(kˉλ)+2νˉT≤48δω46\bar{\nu}/(\bar{k}\lambda)+2\bar{\nu}T\le48\delta_{\omega}, by the definition of νˉ\bar{\nu} in (27) and kˉλT≤1\bar{k}\lambda T\le1. Finally, (32) gives ∣θ~(τ)∣≤∣z(τ)∣+δω|\tilde{\theta}(\tau)|\le|z(\tau)|+\delta_{\omega} and ∣z(0)∣≤∣θ~(0)∣+δω|z(0)|\le|\tilde{\theta}(0)|+\delta_{\omega}, which contributes 5δω5\delta_{\omega} and yields (30). Only the values of χ\chi on [0,t][0,t] were used. ■

Small-Gain Theorem

Theorem 1. Let Assumption 1 hold and consider the map (1) in feedback with the held gradient law (18), (17) and the event generator of Definition 1. Suppose that the parameters R>0R>0, σ>0\sigma>0 and ϵ>0\epsilon>0 fulfill the small-gain condition

κ:=γeρR<1,i.e.,σa2h+(1+σ)ah2R<a2h48,(40)\kappa:=\gamma_e\rho_R<1,\qquad\text{i.e.,}\qquad\sigma a^{2}h+\frac{(1+\sigma)ah}{2}R<\frac{a^{2}h}{48}, \tag*{(40)}

and the offset condition

γeδ≤(1−κ)R4,i.e.,σdˉ+ϵ≤(1−κ)a2hR192.(41)\gamma_e\delta\le\frac{(1-\kappa)R}{4},\qquad\text{i.e.,}\qquad\sigma\bar{d}+\epsilon\le\frac{(1-\kappa)a^{2}hR}{192}. \tag*{(41)}

with γe\gamma_e, ρR\rho_R, δ\delta and dˉ\bar{d} in (27), (24) and (12). Let the dither frequency satisfy

ω>ω∗:=max⁡{16kˉλ(λ+ρR),212kˉλΔ(1−κ)Rˉ}.(42)\omega>\omega^{*}:=\max\left\{16\bar{k}\lambda\left(\lambda+\rho_R\right),\frac{212\bar{k}\lambda\Delta}{(1-\kappa)\bar{R}}\right\}. \tag*{(42)}

Then, for every initial condition with ∣θ(0)−θ∗∣≤(1−κ)R/8|\theta(0)-\theta^{*}|\le(1-\kappa)R/8, the solution exists for all t≥0t\ge0, satisfies ∣θ(t)∣<R|\theta(t)|<R, and

∣θ(t)−θ∗∣≤Mθe−m2t+a+rˉ,(43)|\theta(t)-\theta^{*}|\le M_{\theta}e^{-\frac{m}{2}t}+a+\bar{r}, \tag*{(43)}
∣y(t)−Q∗∣≤Mye−mt+2h(a2+rˉ2),(44)|y(t)-Q^{*}|\le M_y e^{-mt}+2h\left(a^{2}+\bar{r}^{2}\right), \tag*{(44)}

where m:=(1−κ2)kˉλ/8m:=(1-\kappa^{2})\bar{k}\lambda/8, Mθ:=4∣θ(0)−θ∗∣M_{\theta}:=4|\theta(0)-\theta^{*}|, My:=hMθ2M_y:=hM_{\theta}^{2} and

rˉ≤3γeδ1−κ2+O(1ω).(45)\bar{r}\le\frac{3\gamma_e\delta}{\sqrt{1-\kappa^{2}}}+O\left(\frac{1}{\omega}\right). \tag*{(45)}

In addition, the inter-event times satisfy tk+1−tk≥τ∗>0t_{k+1}-t_k\ge\tau^{*}>0 for all k∈Nk\in\mathbb{N}, with τ∗\tau^{*} given in (51) below, so that Zeno behavior does not occur.

Proof. Let t∞:=lim⁡k→∞tk≤+∞t_{\infty}:=\lim_{k\to\infty}t_k\le+\infty. Between events the control is constant, so θ~\tilde{\theta} is continuous and piecewise affine on [0,t∞)[0,t_{\infty}) and cannot escape in finite time. Part A shows that ∣θ~(t)∣<R|\tilde{\theta}(t)|<R on [0,t∞)[0,t_{\infty}) and establishes the convergence estimates; Part B uses only the bound ∣θ~∣<R|\tilde{\theta}|<R to prove that t∞=+∞t_{\infty}=+\infty.

A. Stability analysis.

Step 1 (The ball ∣θ~∣<R|\tilde{\theta}|<R is invariant). Since S(0)=0S(0)=0, θ~(0)=θ(0)−θ∗\tilde{\theta}(0)=\theta(0)-\theta^{*} and thus 4∣θ~(0)∣≤(1−κ)R/24|\tilde{\theta}(0)|\le(1-\kappa)R/2. By (42), ω≥ω0\omega\ge\omega_0 and 53δω=53kˉκΔ/ω<(1−κ)R/453\delta_{\omega}=53\bar{k}\kappa\Delta/\omega<(1-\kappa)R/4. If the claim were false, continuity of θ~\tilde{\theta} would provide a first time

te∈(0,t∞)t_e \in(0,t_\infty) at which ∣θ~(te)∣=R|\tilde{\theta}(t_e)|=R. Up to that time the bound (24) is in force, giving ∣χ∣≤ρRR+δ|\chi|\leq\rho_R R+\delta on [0,te][0,t_e]. Because (30) only sees χ\chi on [0,te][0,t_e] and χ\chi is piecewise continuous, we may evaluate Lemma 2(b) at t=tet=t_e (continuing χ\chi by zero beyond tet_e, for definiteness). With (40) and (41),

∣θ~(te)∣≤4∣θ~(0)∣+κR+γeδ+5δω<(1−κ)R2+κR+(1−κ)R4+(1−κ)R4=R,|\tilde{\theta}(t_e)|\leq4|\tilde{\theta}(0)|+\kappa R+\gamma_e\delta+5\delta_\omega <\frac{(1-\kappa)R}{2}+\kappa R+\frac{(1-\kappa)R}{4}+\frac{(1-\kappa)R}{4}=R,

a contradiction. Thus (24) can be used freely on [0,t∞)[0,t_\infty).

Step 2 (Closing the loop). From (24) and (32), ∣χ(τ)∣≤ρR∣z(τ)∣+ρRδω+δ|\chi(\tau)|\leq\rho_R|z(\tau)|+\rho_R\delta_\omega+\delta. Then (31) gives ∣z˙(τ)∣≤L∣z(τ)∣+η^|\dot{z}(\tau)|\leq L|z(\tau)|+\hat{\eta} with

L:=kˉ(λ+ρR),η^:=kˉ(δ+ρRδω)+νˉ.L:=\bar{k}(\lambda+\rho_R),\qquad\hat{\eta}:=\bar{k}(\delta+\rho_R\delta_\omega)+\bar{\nu}.

By (42), LT≤1/8LT\leq1/8, so Grönwall’s inequality yields ∣z(τ)∣≤2(∣z(t)∣+η^T)|z(\tau)|\leq2(|z(t)|+\hat{\eta}T) on [t,t+T][t,t+T] and the input of (29) obeys

ηt≤2kˉρR∣z(t)∣+η′,η′:=kˉδ+kˉρR(2η^T+δω)+νˉ.(46)\eta_t\leq2\bar{k}\rho_R|z(t)|+\eta', \qquad \eta':=\bar{k}\delta+\bar{k}\rho_R(2\hat{\eta}T+\delta_\omega)+\bar{\nu}. \tag*{(46)}

Step 3 (Small gain). Inserting (46) into (29) and using the weighted Young inequality (x+y)2≤98x2+9y2(x+y)^2\leq\frac{9}{8}x^2+9y^2,

W(t+T)≤(1−m0T+T288kˉρR2λ)W(t)+T576η′2kˉλ.(47)W(t+T)\leq\left(1-m_0T+T\frac{288\bar{k}\rho_R^2}{\lambda}\right)W(t)+T\frac{576\eta'^2}{\bar{k}\lambda}. \tag*{(47)}

The quantity 288kˉρR2/λ288\bar{k}\rho_R^2/\lambda is the loop gain of the interconnection in Figure 2 as seen by the Lyapunov function. Since ρR=κλ/48\rho_R=\kappa\lambda/48 by (40), this loop gain equals κ2kˉλ/8=κ2m0\kappa^2\bar{k}\lambda/8=\kappa^2m_0, and κ<1\kappa<1 is exactly what is needed to retain a positive decrease rate. Therefore, with

Interconnection view of the closed loop

Figure 2. Interconnection view of the closed loop. Stability holds under the small-gain condition γeρR<1\gamma_e\rho_R<1.

m:=(1−κ2)m0=(1−κ2)kˉλ8,μ:=576η′2kˉλ,(48)m:=(1-\kappa^2)m_0=\frac{(1-\kappa^2)\bar{k}\lambda}{8}, \qquad \mu:=\frac{576\eta'^2}{\bar{k}\lambda}, \tag*{(48)}

we obtain W(t+T)≤(1−mT)W(t)+TμW(t+T)\leq(1-mT)W(t)+T\mu, with mT≤m0T≤1/64mT\leq m_0T\leq1/64.

Step 4 (Convergence). The recursion W(t+T)≤(1−mT)W(t)+TμW(t+T)\leq(1-mT)W(t)+T\mu has the same form as in Step 4 of Lemma 2; running that argument again with (m,μ,η^)(m,\mu,\hat{\eta}) instead of (m0,μ0,ηˉ)(m_0,\mu_0,\bar{\eta}) gives

∣θ~(t)∣≤4e−m2t∣θ~(0)∣+rˉ,(49)|\tilde{\theta}(t)|\leq4e^{-\frac{m}{2}t}|\tilde{\theta}(0)|+\bar{r}, \tag*{(49)}
rˉ:=136η′kˉλ1−κ2+2η^T+5δω,\bar{r}:=\frac{136\eta'}{\bar{k}\lambda\sqrt{1-\kappa^2}}+2\hat{\eta}T+5\delta_\omega,

where

2μm=24608η′kˉλ1−κ2≤136η′kˉλ1−κ22\sqrt{\frac{\mu}{m}}=\frac{2\sqrt{4608}\eta'}{\bar{k}\lambda\sqrt{1-\kappa^2}} \leq\frac{136\eta'}{\bar{k}\lambda\sqrt{1-\kappa^2}}

was used. All three ingredients of rˉ\bar{r} are explicit: η′=kˉδ+O(1/ω)\eta'=\bar{k}\delta+O(1/\omega), η^T=O(1/ω)\hat{\eta}T=O(1/\omega) and δω=O(1/ω)\delta_\omega=O(1/\omega); since 136/λ≤3γe136/\lambda\leq3\gamma_e, this is (45).

Step 5 (Bounds on θ\theta and yy). From θ−θ∗=θ~+S\theta-\theta^*=\tilde{\theta}+S, ∣S∣≤a|S|\leq a and (49), estimate (43) holds with Mθ=4∣θ(0)−θ∗∣M_\theta=4|\theta(0)-\theta^*|.

Moreover, ∣y−Q∗∣=h2(θ−θ∗)2≤h2(Mθe−m2t+rˉ+a)2|y-Q^*|=\frac{h}{2}(\theta-\theta^*)^2\leq\frac{h}{2}(M_\theta e^{-\frac{m}{2}t}+\bar{r}+a)^2, and two applications of (x+y)2≤2x2+2y2(x+y)^2\leq2x^2+2y^2 give (44) with My=hMθ2M_y=hM_\theta^2.

Exclusion of Zeno behavior

We now use only the invariance established in Step 1. With ϕ(t):=θ~(t)+S(t)=θ(t)−θ∗\phi(t):=\tilde{\theta}(t)+S(t)=\theta(t)-\theta^* one has ∣ϕ∣≤R+a|\phi|\leq R+a and y=Q∗+H∗2ϕ2y=Q^*+\frac{H^*}{2}\phi^2, whence

∣y(t)∣≤yˉ:=∣Q∗∣+h2(R+a)2,∣G^(t)∣≤ayˉ.|y(t)|\leq\bar{y}:=|Q^*|+\frac{h}{2}(R+a)^2,\qquad|\hat{G}(t)|\leq a\bar{y}.

Writing Sk:=S(tk)S_k:=S(t_k), yk:=y(tk)y_k:=y(t_k), ϕk:=ϕ(tk)\phi_k:=\phi(t_k), the sampling error (19) splits as e(t)=(Sk−S(t))yk+S(t)(yk−y(t))e(t)=(S_k-S(t))y_k+S(t)(y_k-y(t)), t∈[tk,tk+1)t\in[t_k,t_{k+1}).

Since ∣Sk−S(t)∣≤aω(t−tk)|S_k-S(t)|\leq a\omega(t-t_k), ∣yk−y(t)∣=h2∣ϕk−ϕ(t)∣∣ϕk+ϕ(t)∣≤h(R+a)∣ϕk−ϕ(t)∣|y_k-y(t)|=\frac{h}{2}|\phi_k-\phi(t)||\phi_k+\phi(t)|\leq h(R+a)|\phi_k-\phi(t)| and, because θ~˙=kG^(tk)\dot{\tilde{\theta}}=k\hat{G}(t_k) is constant between events, ∣ϕk−ϕ(t)∣≤(kˉayˉ+aω)(t−tk)|\phi_k-\phi(t)|\leq(\bar{k}a\bar{y}+a\omega)(t-t_k), we conclude that

∣e(t)∣≤Le(t−tk),t∈[tk,tk+1),(50)|e(t)|\leq L_e(t-t_k),\qquad t\in[t_k,t_{k+1}), \tag*{(50)}
Le:=aωyˉ+ah(R+a)(kˉayˉ+aω).L_e:=a\omega\bar{y}+ah(R+a)(\bar{k}a\bar{y}+a\omega).

At an event, by (22) and continuity of ee on [tk,tk+1)[t_k,t_{k+1}), ∣e(tk+1)∣≥σ∣G^(tk+1)∣+ε≥ε|e(t_{k+1})|\geq\sigma|\hat{G}(t_{k+1})|+\varepsilon\geq\varepsilon, while e(tk)=0e(t_k)=0. Hence (50) gives

tk+1−tk≥τ∗:=εLe>0,∀k∈N.(51)t_{k+1}-t_k\geq\tau^*:=\frac{\varepsilon}{L_e}>0,\qquad\forall k\in\mathbb{N}. \tag*{(51)}

As τ∗\tau^* does not depend on kk, either finitely many events occur or tk→∞t_k\to\infty; in both cases t∞=+∞t_\infty=+\infty, the solution is defined for all t≥0t\geq0 and Steps 2–5 of Part A hold globally in time. ▪

Discussion

Remark 1 (Small-gain reading): In (40) two gains are multiplied: γe\gamma_e, which measures how strongly the ES dynamics amplify χ\chi into θ~\tilde{\theta} (Lemma 2), and ρR\rho_R, which measures how strongly the event generator and the quadratic term send θ~\tilde{\theta} back into χ\chi (see (24)). This is the one-dimensional, linear-gain version of the composition condition α−1∘γ∘ρ<Id⁡\alpha^{-1}\circ\gamma\circ\rho<\operatorname{Id} of [15], [16]. Two parameters feed ρR\rho_R: the relative tolerance σ\sigma, and the size RR of the region where the result is claimed, which enters because the quadratic part of the map is treated as an input. It is worth observing that kˉ\bar{k} drops out of γe=48/(a2h)\gamma_e=48/(a^2h), whereas a larger excitation level λ\lambda lowers it. Condition (40) is in the standard form κ<1\kappa<1, and the strict inequality is used only to keep the decrease rate m=(1−κ2)m0m=(1-\kappa^2)m_0 positive; accordingly, the decay rate, the admissible initial error, the admissible radius in (45) all degrade as κ→1\kappa\to1, the factor (1−κ2)−1/2(1-\kappa^2)^{-1/2} in (45) being the counterpart of the familiar (1−κ)−1(1-\kappa)^{-1} of the classical small-gain estimate γeσ/(1−κ)\gamma_e\sigma/(1-\kappa). Had the cross term in Step 3 of the proof been split with the symmetric inequality (x+y)2≤2x2+2y2(x+y)^2\leq2x^2+2y^2, the loop gain would have read 512kˉρR2/λ512\bar{k}\rho_R^2/\lambda and only κ<3/4\kappa<3/4 could have been certified; the weighted splitting removes this artificial restriction at the cost of a larger constant multiplying η′2\eta'^2.

Remark 2 (Design procedure): Dividing (40) by ahah, the small-gain condition becomes σa+1+σ2R<a48\sigma a+\frac{1+\sigma}{2}R<\frac{a}{48}, which suggests the following order of choices:

  1. the gain kk and the amplitude aa, which fix λ\lambda, ℓ\ell, dˉ\bar{d} and γe\gamma_e;

  2. a provisional threshold σ0<1/48\sigma_0<1/48;

  1. a radius

0<R<2a(1−48σ0)48(1+σ0),(52)0 < R < \frac{2a(1-48\sigma_0)}{48(1+\sigma_0)}, \tag*{(52)}

which reserves an explicit share of the small-gain margin for the quadratic term, and the corresponding κ0:=γeρR∣σ=σ0<1\kappa_0 := \gamma_e\rho_R|_{\sigma=\sigma_0}<1;

  1. the threshold σ≤min⁡{σ0,(1−κ0)a2hR/(384dˉ)}\sigma\le\min\{\sigma_0,(1-\kappa_0)a^2hR/(384\bar d)\} and the offset ε≤(1−κ0)a2hR/384\varepsilon\le(1-\kappa_0)a^2hR/384, so that (41) holds;

  1. the frequency ω>ω∗\omega>\omega^* in (42).

Since ρR\rho_R, and hence κ\kappa, grows with σ\sigma, lowering σ\sigma in step 4) preserves (40) and gives κ≤κ0\kappa\le\kappa_0. Inequality (52) forces R<a/24R<a/24, i.e., the region in which convergence is certified is comparable to the probing amplitude. This is inherent to a statement proved on the exact, non-averaged loop. Being sufficient conditions obtained through worst-case bounds, (40)–(42) are conservative, and Section V shows satisfactory behavior for parameters well beyond them.

Remark 3 (Trade-offs): Three mechanisms determine the size of the final neighborhood in (43): the probing amplitude aa, which is unavoidable in ES; the oscillation Δ(t)\Delta(t), whose effect scales as 1/ω1/\omega; and the event generator, which contributes a term proportional to γeδ\gamma_e\delta. Without a strictly positive ε\varepsilon no uniform bound of the type (51) could hold, because G^\hat G keeps oscillating through zero even when θ~=0\tilde\theta=0; what is obtained is practical stability in the presence of a non-vanishing matched disturbance, a situation also addressed in [20]. Since LeL_e in (50) grows linearly with ω\omega, one has τ∗=O(ε/ω)\tau^*=O(\varepsilon/\omega): a faster dither shrinks the residual set but demands faster updates, in agreement with the intuition that the held control must track the probing oscillation.

Numerical Example

The scalar map (1) is simulated with

H∗=4,Q∗=1,θ∗=1,H^*=4,\qquad Q^*=1,\qquad\theta^*=1,

so that QQ has a minimum and, by (17), a negative gain must be used. The probing signal (3) has amplitude a=0.2a=0.2 and frequency ω=3\omega=3 rad/s, the adaptation gain is k=−1k=-1, and the integrator starts at θ^(0)=2\hat\theta(0)=2, i.e., one unit away from the optimizer. The event generator of Definition 1 is run with σ=0.1\sigma=0.1 and ε=0.01\varepsilon=0.01, and the triggering condition (22) is evaluated at every step of a fixed-step integration with step size 10−310^{-3} s over a horizon of 200 s. For these values the excitation level is λ=a2h=0.16\lambda=a^2h=0.16, so the mean time constant of (20) is 2/(kˉλ)=12.52/(\bar k\lambda)=12.5 s; moreover, dˉ=0.216\bar d=0.216 and δ=σdˉ+ε≈0.032\delta=\sigma\bar d+\varepsilon\approx0.032. The relative threshold σ=0.1\sigma=0.1 lies well above the value 1/481/48 allowed by Remark 2, so the example also serves to show how conservative the sufficient conditions are.

Figure 3 and Figure 4 show the input and the output of the map. Starting from y(0)=3y(0)=3, the input θ(t)\theta(t) descends toward θ∗\theta^* within roughly 25 s, a duration compatible with the time constant dictated by λ\lambda, and the output is brought from its initial excursions, which reach almost 4 because of the dither, to the vicinity of Q∗=1Q^*=1. After the transient, θ(t)\theta(t) carries the

Input $\theta(t)$ of the map and optimizer $\theta^*$.

Figure 3. Input θ(t)\theta(t) of the map and optimizer θ∗\theta^*.

Output $y(t)$ of the map and minimum value $Q^*$.

Figure 4. Output y(t)y(t) of the map and minimum value Q∗Q^*.

fast oscillation of amplitude a=0.2a=0.2 imposed by the probing signal, which is the irreducible term α\alpha in (43), superimposed on a slow, low-amplitude excursion about θ∗\theta^*: the mean value of θ\theta undershoots to about 0.75 near t=45t=45 s, then drifts to about 1.2 near t=90t=90 s and oscillates with decreasing amplitude afterwards. This slow component is the signature of the sample-and-hold implementation: between two events the integrator is driven by a constant u=kG^(tk)u=k\hat G(t_k) which, owing to the persistent term Δ(tk)\Delta(t_k) in (8), is not exactly zero even when θ^˙(tk)\dot{\hat\theta}(t_k) is small, so that θ^\hat\theta moves linearly until the triggering rule detects the discrepancy and refreshes the control. The excursion stays within approximately 0.25 of θ∗\theta^*, and the corresponding output remains below 1.3 for all t>25t>25 s, in agreement with the practical-stability nature of Theorem 1, whose residual set is shaped not only by aa but also by the event-generator term in (45).

The control signal u(t)=kG^(tk)u(t)=k\hat G(t_k) is displayed in Figure 5. As imposed by (18), it is piecewise constant, with only 13 distinct levels over the whole run. Its magnitude peaks at about 0.053 around t=15t=15 s, when the estimation error and hence the useful component a22H∗θ~\frac{a^2}{2}H^*\tilde\theta of (8) are largest, and then decays: after t≈45t\approx45 s the held values do not exceed 0.008 in absolute value. The distribution of the events, shown in Figure 6, follows the same pattern. Ten of the thirteen updates occur during the transient, within the first 45 s, while ∣G^∣|\hat G| is still appreciable and the relative term σ∣G^∣\sigma|\hat{G}| in (22) is easily exceeded; the remaining three occur near t=92t=92, 155 and 190 s, separated by intervals of about one minute, during which the offset ε\varepsilon dominates the threshold and keeps the hold active even though G^(t)\hat{G}(t) keeps oscillating through zero. No accumulation of events is observed anywhere, and the shortest interval between consecutive events is of the order of one second, orders of magnitude larger than the guaranteed bound (51), which for R=1.2R=1.2 (a ball that contains the whole trajectory, since ∣θ~(0)∣=1|\tilde{\theta}(0)|=1) gives yˉ=4.92\bar{y}=4.92, Le≈4.7L_e\approx4.7 and τ∗=ε/Le≈2\tau^*=\varepsilon/L_e\approx2 ms.

Control signal $u(t)=k\hat{G}(t_k)$

Figure 5. Control signal u(t)=kG^(tk)u(t)=k\hat{G}(t_k) delivered by the zero-order hold.

Instants of control update

Figure 6. Instants of control update: fixed-step implementation (circles, 200 001 updates) and event-triggered implementation (stars, 13 updates).

Against the 200 001 updates of the time-triggered implementation, the event-triggered law converges to a neighborhood of the optimizer with 13 updates, a reduction above 99.99% in control transmissions. The price is the slow excursion of θ\theta discussed above, which decreasing ε\varepsilon and σ\sigma, or increasing ω\omega, would reduce at the expense of more updates (Remark 3).

Conclusions

We have analyzed an event-triggered implementation of gradient extremum seeking for scalar quadratic maps in which the triggering rule acts on the measured gradient estimate and includes a positive offset. The scalar structure of the demodulated signal, namely a nonnegative and persistently exciting adaptation coefficient plus an explicit zero-mean sinusoidal perturbation, allowed us to bypass averaging theory altogether: the ISS gain of the extremum seeking dynamics with respect to the sampling error was computed on the exact error equation, and closed-loop stability was obtained from a small-gain condition with the gain of the event generator, in which the quadratic term of the map is handled as part of the input. All design constants, the minimum dither frequency and the lower bound on the inter-event times are explicit. Dynamic triggers and Newton-based schemes are natural next steps.

References

  1. [1]K. B. Ariyur and M. Krstić, Real-Time Optimization by Extremum-Seeking Control. NJ, USA: Wiley, 2003.
  2. [2]M. Krstić and H.-H. Wang, “Stability of extremum seeking feedback for general nonlinear dynamic systems,” Automatica, vol. 36, pp. 595–601, 2000.DOI
  3. [3]M. Krstić, “Extremum seeking control,” in Encyclopedia of Systems and Control (J. Baillieul and T. Samad, eds.), pp. 413–416, London: Springer, 2014.
  4. [4]A. Ghaffari, M. Krstić, and D. Nešić, “Multivariable Newton-based extremum seeking,” Automatica, vol. 48, no. 8, pp. 1759–1767, 2012.
  5. [5]T. R. Oliveira, M. Krstić, and D. Tsakibano, “Extremum seeking for static maps with delays,” IEEE Transactions on Automatic Control, vol. 62, no. 4, pp. 1911–1926, 2017.DOI
  6. [6]T. R. Oliveira and M. Krstić, Extremum Seeking through Delays and PDEs. Society for Industrial and Applied Mathematics, 2022.DOI
  7. [7]C. Zhang, D. Arnold, N. Ghods, A. Siranosian, and M. Krstić, “Source seeking with non-holonomic unicycle without position measurement and with tuning of forward velocity,” Systems & Control Letters, vol. 56, no. 3, pp. 245–252, 2007.DOI
  8. [8]P. Tabuada, “Event-triggered real-time scheduling of stabilizing control tasks,” IEEE Transactions on Automatic Control, vol. 52, no. 9, pp. 1680–1685, 2007.DOI
  9. [9]W. P. M. H. Heemels, K. H. Johansson, and P. Tabuada, “An introduction to event-triggered and self-triggered control,” in 51st IEEE Conference on Decision and Control (CDC), pp. 3270–3285, 2012.DOI
  10. [10]P. Zhang, T. Liu, and Z.-P. Jiang, “Input-to-state stabilization of nonlinear discrete-time systems with event-triggered controllers,” Systems & Control Letters, vol. 103, pp. 16–22, 2017.DOI
  11. [11]V. H. P. Rodrigues, T. R. Oliveira, L. Hsu, M. Diagne, and M. Krstić, “Event-triggered and periodic event-triggered extremum seeking control,” Automatica, vol. 174, p. 112161, 2025.
  12. [12]V. H. P. Rodrigues, T. R. Oliveira, M. Krstić, and P. Tabuada, “Event-triggered Newton extremum seeking for multivariable optimization,” arXiv preprint arXiv:2601.14416, 2026.
  13. [13]V. H. P. Rodrigues, T. R. Oliveira, M. Krstić, and F. Allgöwer, “Event-triggered discrete-time multivariable extremum seeking systems,” arXiv preprint arXiv:2607.25223, 2026.
  14. [14]V. H. P. Rodrigues, T. R. Oliveira, M. Krstić, and T. Başar, “Distributed event-triggered Nash equilibrium seeking,” Automatica, vol. 192, p. 113164, 2026.
  15. [15]T. Liu and Z.-P. Jiang, “A small-gain approach to robust event-triggered control of nonlinear systems,” IEEE Transactions on Automatic Control, vol. 60, no. 8, pp. 2072–2085, 2015.
  16. [16]T. Liu, P. Zhang, and Z.-P. Jiang, Robust Event-Triggered Control of Nonlinear Systems. Singapore: Springer, 2020.DOI
  17. [17]P. Pepe and Z.-P. Jiang, “A Lyapunov–Krasovskii methodology for ISS and iISS of time-delay systems,” Systems & Control Letters, vol. 55, no. 12, pp. 1006–1014, 2006.DOI
  18. [18]Z.-P. Jiang and Y. Wang, “A converse Lyapunov theorem for discrete-time systems with disturbances,” Systems & Control Letters, vol. 45, no. 1, pp. 49–58, 2002.DOI
  19. [19]H. K. Khalil, Nonlinear Systems. Upper Saddle River, New Jersey: Prentice Hall, 3rd ed., 2002.
  20. [20]I. Karafyllis and M. Krstić, “Deadzone-adapted disturbance suppression control for global practical IOS and zero asymptotic gain to matched uncertainties,” Systems & Control Letters, vol. 185, p. 105746, 2024.

Paper details

vlnc:W_35v989hh
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 11

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents