betaFeedback

Reading v1View latest version

WorksEngineering

Proportional Neuro-Impulsive Control of SISO LTI Systems

Abstract

This paper explores a controller that is at the crossroads of neuromorphic, i.e., spike-based, and impulsive control systems. The controller consists of a single integrate-and-fire-like neuron that generates an input proportional to the system output when the neuron variable reaches a threshold. We prove that any static output-feedback stabilizable single-input single-output (SISO) linear time-invariant (LTI) plant can be made semi-globally asymptotically stable (sGAS) in closed loop with the proposed neuro-impulsive controller. This improves on existing fixed-amplitude neuromorphic controllers, which only achieve practical stability, but at the price of using graded spiking inputs. Moreover, we provide an algorithm for choosing appropriate controller parameters given a set of admissible initial conditions. A numerical example of an inverted pendulum linearized around its upright equilibrium illustrates the method.

Introduction

Understanding and reproducing biological control laws and feedback loops is an old but ever-reappearing goal in control and in science at large. Norbert Wiener’s cybernetics [1] is one well-known attempt, trying to unify the study of feedback loops, whether biological, physical, or even social. In a similar spirit, control engineers have recently taken an interest in neuromorphic engineering, a field that studies how to reproduce neuronal dynamics in physical substrates and how to use them to compute, sense, and actuate [2]. The field started as an electronics subfield and has achieved its most impressive results there, but its link with control has been clear since its inception [3] and has recently attracted new interest [4]–[7].

Neuron models come in many shapes and forms [8], and so do neuromorphic controllers. Most rely on a common idea: a dynamic element, the neuron, evolves according to its input until a threshold is reached; a discrete event, the spike, is then triggered and the element is reset. The spike may be a true discrete event produced by a hard threshold, as in the integrate-and-fire models [9], or embedded in the dynamics as a fast but continuous trace, as in conductance-based and mixed-feedback models [4], [10], [11]. The models differ greatly, but they all agree on one point: the neuron produces a control pulse of fixed amplitude (almost fixed for continuous spikes) at each spike.

Controlling with discrete events is not unique to neuromorphic engineering. As early as 1963, integral pulse frequency modulation (IPFM) [12] described controllers that, also inspired by neural communication, generate fixed-amplitude pulses with a frequency proportional to the input signal, and that line of work went as far as characterizing the resulting practical stability zone [13]. This is close, in spirit, to controllers relying on an integrate-and-fire neuron as a spike generator [6], [7], [14], as both trigger an event when the integral of an error reaches a threshold. In a similar way, impulsive control was created to analyze systems that follow continuous dynamics most of the time but undergo discrete transitions at time instants that can either be fixed or depend on the state [15].

The way neuromorphic controllers have largely been conceptualized leaves at least two open problems. Since every spike has the same amplitude, the correction sent to the plant does not scale with the plant’s distance from its target. Close to the desired unstable fixed point, fixed-amplitude spikes can therefore only lead to practical stability, with oscillation being unavoidable [6], [7]. Far from it, spikes must be delivered many times to have a large effect, so the spike frequency grows with the distance to the target and the inter-spike times can only be bounded semi-globally. The two situations pull against each other: minimizing the oscillation requires spikes with a small effect, but keeping the spike frequency low far from the target requires spikes with a large effect. High spike frequencies are not a problem in theory, but depending on the actuator they may be undesirable in practice. In power electronics, for example, the total switching loss grows with the switching frequency [16 Ch. 4]. This prompts an apparently obvious fix: if spike strength were proportional to the state, as in continuous control, would this not give asymptotic stability while softly bounding the spiking rate? This kind of proportional, state-dependent transition is not unheard of in impulsive control [17], but, to the best of the authors’ knowledge, the tools developed there do not fit the neuromorphic setting: they either ask the jump matrix to be strictly contracting [18], or they require satisfying discrete-time Lyapunov linear matrix inequalities on the flow–jump cycle for all possible dwell times [17]. Sections IV and V detail why neither applies, in general, in the neuromorphic control setting.

The controller we study here keeps the integrate-and-fire behavior driven by the plant output, but makes its spike amplitude proportional to that output at firing time. Because the spike now shrinks with the output, the limitation to practical stability disappears. This only removes the first of the two problems: the spike frequency still grows far from the target, but each spike delivers a larger correction, so fewer are needed. Although a graded-spike actuator might not be readily available, this work is a first step toward a design using an array of fixed-amplitude neurons working at different output scales.

Our first contribution is the formulation of what we call a single-input single-output (SISO) linear time-invariant (LTI) neuro-impulsive control system (NICS), and the proof that any such system built for a static output feedback (SOF) stabilizable SISO LTI plant is semi-globally asymptotically stable (sGAS), using the hybrid system formalism [19] and an emulation approach [6, 20], in which the spiking controller is designed to emulate the closed-loop behavior of a stabilizing continuous-time SOF. The second contribution is an algorithm returning control parameters that make the NICS sGAS from a given set of initial conditions. It is tested by stabilizing a linearized inverted pendulum at its upright equilibrium.

The paper is structured as follows. Section II fixes notation, Section III states the control problem, Section IV develops the intermediate results, Section V presents the main result and its proof and derives an algorithm from it, Section VI illustrates the approach on a numerical example, and Section VII concludes the paper. Appendix A contains the proofs of the intermediate results.

All code and data are available on GitHub: https://github.com/ChristianFernandezLorden/ECC27-FernandezLorden-SISONeuroImpulsive.

Preliminaries

We denote by R\mathbb{R}, C\mathbb{C}, and N\mathbb{N} the sets of real, complex, and natural numbers, the latter including 0. Given a positive-definite matrix PP, the energy norm of xx is ∥x∥P=xTPx\lVert x\rVert_P = \sqrt{x^{\mathsf{T}}Px}, and ∥M∥P\lVert M\rVert_P and μP(M)\mu_P(M) denote the matrix norm and the logarithmic norm [21] of a square matrix MM induced by it. Lemmas holding for any vector norm are written without a subscript. The logarithmic norm property we use is eμ(A)t≤e∥A∥te^{\mu(A)t} \leq e^{\lVert A\rVert t}. For a symmetric MM, λmax⁡(M)\lambda_{\max}(M) and λmin⁡(M)\lambda_{\min}(M) are its largest and smallest eigenvalues, and a square matrix is Hurwitz if all its eigenvalues have negative real part. By W0(x)W_0(x) we denote the principal branch of the Lambert WW function [22], defined for x≥−1/ex \geq-1/e.

We consider hybrid systems in the formalism of [19], meaning a system H\mathcal{H} is defined as

H:{z˙=f(z),z∈C,z+=g(z),z∈D,(1)\mathcal{H}:\left\{ \begin{aligned} \dot{z} &= f(z), && z \in C,\\ z^+ &= g(z), && z \in D, \end{aligned} \right. \tag*{(1)}

with z∈Rnz \in\mathbb{R}^n the state, ff and CC the flow map and the flow set, and gg and DD the jump map and the jump set. Hybrid time domains and solutions of (1) are defined as in [19 Definitions 2.3 and 2.6], so each solution zz has a domain dom⁡z=⋃i[ti,ti+1]×{i}\operatorname{dom} z = \bigcup_i [t_i,t_{i+1}] \times\{i\} with 0=t0≤t1≤⋯0 = t_0 \leq t_1 \leq\cdots, and we write zi:=z(ti,i)z_i := z(t_i,i) for the state after the iith jump (z0z_0 being the initial condition). We only consider maximal solutions in the sense of [19 Definition 2.7]. A controlled hybrid system is a system (1) whose data [19 Definition 2.2] are parametrized by θ∈Θ\theta\in\Theta, the set of admissible control parameters.

Definition 1 (Semi-global asymptotic stability). A set A⊂U\mathcal{A} \subset\mathcal{U} of a controlled hybrid system with state space U\mathcal{U} and control parameters θ∈Θ\theta\in\Theta is said to be semi-globally asymptotically stable (sGAS) if, for every compact set Z⊂U\mathcal{Z} \subset\mathcal{U} of initial conditions, there exists θ∈Θ\theta\in\Theta such that A\mathcal{A} is pre-asymptotically stable from every z0∈Zz_0 \in\mathcal{Z} in the sense of [19 Definition 3.6].

Control system setup and problem statement

We consider SISO LTI systems

x˙=Ax+Bu,y=Cx,(2)\dot{x} = Ax + Bu,\qquad y = Cx, \tag*{(2)}

with scalar input u(t)∈Ru(t) \in\mathbb{R}, state x(t)∈Rnx(t) \in\mathbb{R}^n, and scalar output y(t)∈Ry(t) \in\mathbb{R} at time t∈R≥0t \in\mathbb{R}_{\geq0}. We only consider systems (2) that are unstable, i.e., AA is not Hurwitz, and satisfy the following assumption.

Assumption 2 (Static output-feedback (SOF) stabilizability). The set K={K∈R>0:A−BKC is Hurwitz}\mathcal{K} = \{K \in\mathbb{R}_{>0}: A - BKC \text{ is Hurwitz}\} is non-empty.1

A necessary and sufficient condition for Assumption 2 is given in [23 Theorem 5].

This paper considers the stabilization of the origin of an LTI system (2) using a feedback hybrid controller inspired by neural dynamics, and aims to show that any SISO LTI system stabilizable by SOF can also be stabilized by it. Consider the hybrid control system (1) with data

f((xξ)):=(Ax∣Cx∣+Ib),C:=Rn×[0,Δ],(3)f\left(\begin{pmatrix}x\\ \xi\end{pmatrix}\right) := \begin{pmatrix}Ax\\ |Cx|+I_b\end{pmatrix},\qquad C := \mathbb{R}^n \times[0,\Delta], \tag*{(3)}
g((xξ)):=((I−αBC)x0),D:=Rn×[Δ,∞),(4)g\left(\begin{pmatrix}x\\ \xi\end{pmatrix}\right) := \begin{pmatrix}(I-\alpha BC)x\\ 0\end{pmatrix},\qquad D := \mathbb{R}^n \times[\Delta,\infty), \tag*{(4)}

where ξ∈R≥0\xi\in\mathbb{R}_{\geq0} is the neuron state, Δ∈R>0\Delta\in\mathbb{R}_{>0} the firing threshold, Ib∈R>0I_b \in\mathbb{R}_{>0} a constant bias current, and α∈R>0\alpha\in\mathbb{R}_{>0} the spike gain. Δ\Delta, IbI_b, and α\alpha are the design parameters of (3). Observe that (3) is the feedback interconnection of (2) and a SISO neuro-impulsive controller with state ξ\xi that generates a control impulse proportional to y=Cxy = Cx each time the neural threshold is reached. We call this closed loop a SISO LTI neuro-impulsive control system (NICS).

The flow of ξ\xi in (3a) is inspired by the well-known integrate-and-fire dynamics [9], already used in a hybrid closed-loop model in [6, 14], but the jump (3b) replaces the usual fixed-amplitude jump of the plant state with a jump proportional to the plant output at firing time, like a Dirac impulse sent as input to (2). With fixed-amplitude spikes [6 Section III], a control input taking both signs requires at least two neurons, one per sign, whereas (3b) already carries the sign of yy, so one neuron driven by the nonnegative input ∣Cx∣+Ib|Cx|+I_b suffices.

Such a system has a unique maximal solution from every initial condition (x0,ξ0)∈C∪D=Rn×R≥0(x_0,\xi_0) \in C \cup D = \mathbb{R}^n \times\mathbb{R}_{\geq0} (see Appendix A). Since ξ≥Ib\xi\geq I_b during flows and ξ\xi is reset to 0 at each jump, this solution jumps infinitely many times. The problem we address is to design, for a system (2) satisfying Assumption 2 and a compact set Z⊂Rn×R≥0\mathcal{Z} \subset\mathbb{R}^n \times\mathbb{R}_{\geq0} of initial conditions, the parameters Δ\Delta, IbI_b, and α\alpha such that the set {0}×[0,Δ]\{0\} \times[0,\Delta] is asymptotically stable for (3) with a basin of attraction containing Z\mathcal{Z}. In the rest of the paper, NICS is a shorthand for SISO LTI NICS and z:=(x,ξ)z := (x,\xi) denotes the state of (3).

Remark 3. The bias current IbI_b ensures an upper bound Δ/Ib\Delta/I_b on the time between two consecutive spikes, also called maximal dwell time, as shown in Lemma 7. Asymptotic stability comes from the spike being proportional to the plant output, but our analysis also requires Ib>0I_b > 0. Since ∣Cx∣|Cx| disappears from the trigger near the origin, without IbI_b, the time between spikes diverges to infinity as xi→0x_i \to0 and the effective gain α/T(xi)\alpha/T(x_i) of § 3 goes to zero.

Impact map analysis

Note that BCBC is a rank-one matrix, so I−αBCI-\alpha BC must have 11 as an eigenvalue as soon as n>1n>1, which means the jump map can never be a contraction. Between jumps, u=0u=0 and the plant follows its open-loop flow, which is not a contraction either since AA is not Hurwitz. Stability can therefore only be proven from the combined effect of the flow and the jump, which rules out methods relying on contracting jump matrices, e.g., [18], and leads us to define the so-called impact map.

Definition 4 (Impact map). The impact map of a solution of a hybrid system H\mathcal{H} that satisfies the uniqueness conditions of [19 Proposition 2.11] is the function F:Rn→Rn\mathcal{F}:\mathbb{R}^n\to\mathbb{R}^n that maps the hybrid system state immediately after a jump to its state immediately after the next jump:

zi+1=F(zi)whenever (ti+1,i+1)∈dom⁡z.(5)z_{i+1}=\mathcal{F}(z_i)\quad\text{whenever }(t_{i+1},i+1)\in\operatorname{dom}z. \tag*{(5)}

In NICSs (3), the impact map (Definition 4) of a solution is a linear flow followed by a linear jump, hence linear and, for any jump after the first, it is determined by a transition matrix

Φ(TΔ,Ib(x),α):=[(I−αBC)eATΔ,Ib(x)0n×101×n0],(6)\Phi(T_{\Delta,I_b}(x),\alpha):= \begin{bmatrix} (I-\alpha BC)e^{AT_{\Delta,I_b}(x)} & 0_{n\times1}\\ 0_{1\times n} & 0 \end{bmatrix}, \tag*{(6)}

with TΔ,Ib(x):Rn→R>0T_{\Delta,I_b}(x):\mathbb{R}^n\to\mathbb{R}_{>0} the time it takes ξ\xi to reach Δ\Delta from a post-jump state z=(x,0)z=(x,0), i.e., the dwell time [19 Section 2.4] from zz. The zeros on the bottom row represent the reset of ξ\xi, and for brevity we denote the non-zero block by Φ(x)\Phi(x) and the dwell time TΔ,Ib(x)T_{\Delta,I_b}(x) by T(x)T(x).

To first order in T(x)T(x), the transition matrix (I−αBC)eAT(x)(I-\alpha BC)e^{AT(x)} agrees with the flow matrix e(A−(α/T(x))BC)T(x)e^{(A-(\alpha/T(x))BC)T(x)} of a continuous SOF loop with gain α/T(x)\alpha/T(x), which Lemma 5 makes quantitative.

In Lemmas 5–9, zz denotes any maximal solution of (3) and i∈N≥1i\in\mathbb{N}_{\geq1}, so that ξi=0\xi_i=0. The first flow, which starts from an arbitrary ξ0\xi_0, is treated separately in the proof of Theorem 11.

Lemma 5 (Continuous-feedback emulation). Let i∈N≥1i\in\mathbb{N}_{\geq1} and K∈RK\in\mathbb{R}, and define κ(xi):=K−α/T(xi)\kappa(x_i):=K-\alpha/T(x_i). If T(xi)<min⁡(∥A∥−1,∥A−BKC∥−1)T(x_i)<\min\left(\lVert A\rVert^{-1},\lVert A-BKC\rVert^{-1}\right), then

∥Φ(xi)∥≤∥e(A−BKC)T(xi)∥+C1(κ(xi))T(xi)+C2(κ(xi))T(xi)2+C3(κ(xi))T(xi)3,(7)\lVert\Phi(x_i)\rVert\leq\left\lVert e^{(A-BKC)T(x_i)}\right\rVert+C_1(\kappa(x_i))T(x_i)+C_2(\kappa(x_i))T(x_i)^2+C_3(\kappa(x_i))T(x_i)^3, \tag*{(7)}

where

C1(κ):=∥κ∥∥BC∥,(8)C_1(\kappa):=\lVert\kappa\rVert\lVert BC\rVert, \tag*{(8)}
C2(κ):=∥BC∥((∥K∥+∥κ∥)∥A∥+K22∥CB∥),(9)C_2(\kappa):=\lVert BC\rVert\left((\lVert K\rVert+\lVert\kappa\rVert)\lVert A\rVert+\frac{K^2}{2}\lVert CB\rVert\right), \tag*{(9)}
C3(κ):=(∥A∥3+∥A−BKC∥3)2e−52+(∥K∥+∥κ∥)∥BC∥∥A∥2(e−2).(10)C_3(\kappa):=\left(\lVert A\rVert^3+\lVert A-BKC\rVert^3\right)\frac{2e-5}{2}+(\lVert K\rVert+\lVert\kappa\rVert)\lVert BC\rVert\lVert A\rVert^2(e-2). \tag*{(10)}

Therefore, for any solution of (3), the ratio α/T(x)\alpha/T(x) is the effective static output-feedback gain that the spiking controller emulates when flowing from a state z=(x,0)z=(x,0) and jumping, and κ(x)\kappa(x) is the deviation of that effective gain from the reference gain KK. Note that this ratio is not a design parameter: α\alpha is, but T(x)T(x) is determined by the plant dynamics through the firing condition. Since κ(x)\kappa(x) is the coefficient of the first-order term of the difference between the transition matrix and the continuous flow, the two are close for small T(x)T(x) when ∣κ(x)∣|\kappa(x)| is small.

For Hurwitz A−BKCA-BKC and a positive-definite Q∗∈Rn×nQ^*\in\mathbb{R}^{n\times n}, the Lyapunov equation

(A−BKC)TP∗+P∗(A−BKC)=−Q∗(11)(A-BKC)^\mathsf{T}P^*+P^*(A-BKC)=-Q^* \tag*{(11)}

has a unique positive-definite solution P∗P^*, which defines the Lyapunov function V(x)=xTP∗xV(x)=x^\mathsf{T}P^*x of the continuous loop x˙=(A−BKC)x\dot{x}=(A-BKC)x. Standard Lyapunov decay estimate then guarantees that ∥e(A−BKC)T∥P∗≤e−λ0T/2\lVert e^{(A-BKC)T}\rVert_{P^*}\leq e^{-\lambda_0T/2}, with λ0:=λmin⁡(Q∗)/λmax⁡(P∗)>0\lambda_0:=\lambda_{\min}(Q^*)/\lambda_{\max}(P^*)>0 [24 Section 8.5]. Lemma 5 expressed in the P∗P^*-norm then bounds ∥Φ(x)∥P∗\lVert\Phi(x)\rVert_{P^*} by e−λ0T(x)/2+∑k=13C‾k(κ(x))T(x)ke^{-\lambda_0T(x)/2}+\sum_{k=1}^{3}\overline{C}_k(\kappa(x))T(x)^k. As Lemma 6 shows, under the right conditions this bound is smaller than 11, which means that the transition matrix (5) contracts xx.

Lemma 6 (Existence of a contracting (α,T(x))(\alpha,T(x)) range). Suppose Assumption 2 is satisfied, let K∈KK\in\mathcal{K}, let Q∗∈Rn×nQ^*\in\mathbb{R}^{n\times n} be positive definite, and let P∗P^* be the solution of the Lyapunov equation (8) and λ0:=λmin⁡(Q∗)/λmax⁡(P∗)>0\lambda_0:=\lambda_{\min}(Q^*)/\lambda_{\max}(P^*)>0. Let C‾1,C‾2,C‾3\overline{C}_1,\overline{C}_2,\overline{C}_3 be the constants (7) of Lemma 5 expressed in the P∗P^*-norm. Suppose the following holds for at least one κ‾∈(0,λ0/(2∥BC∥P∗))\overline{\kappa}\in(0,\lambda_0/(2\lVert BC\rVert_{P^*})):

T(xi)<−(C‾2+λ028)+(C‾2+λ028)2+4C‾3(λ02−C‾1)2C‾3,(12)T(x_i)<\frac{-\left(\overline{C}_2+\frac{\lambda_0^2}{8}\right)+\sqrt{\left(\overline{C}_2+\frac{\lambda_0^2}{8}\right)^2+4\overline{C}_3\left(\frac{\lambda_0}{2}-\overline{C}_1\right)}}{2\overline{C}_3}, \tag*{(12)}
∣κ(xi)∣<κ‾.(13)|\kappa(x_i)|<\overline{\kappa}. \tag*{(13)}

Then ∥Φ(xi)∥P∗≤g(T(xi))<1\lVert\Phi(x_i)\rVert_{P^*}\leq g(T(x_i))<1, where g(T):=e−λ0T/2+C‾1T+C‾2T2+C‾3T3<1g(T):=e^{-\lambda_0T/2}+\overline{C}_1T+\overline{C}_2T^2+\overline{C}_3T^3<1 for every T>0T>0 satisfying (9). Here C‾1:=C1(κ‾)\overline{C}_1:=C_1(\overline{\kappa}), C‾2:=C2(κ‾)\overline{C}_2:=C_2(\overline{\kappa}), and C‾3:=C3(κ‾)\overline{C}_3:=C_3(\overline{\kappa}) are the constants (7) of Lemma 5 expressed in the P∗P^*-norm.

κ(x)\kappa(x) and T(x)T(x) are not controller parameters. But κ(x)\kappa(x) depends on the dynamics only through T(x)T(x), and since Lemma 5 and (9) impose upper bounds on T(x)T(x) while (10) bounds α/T(x)\alpha/T(x) to a range, it suffices to bound T(x)T(x) in terms of the controller parameters, which the next lemma does.

Lemma 7 (Per-jump dwell time bounds). For any i∈N≥1i\in\mathbb{N}_{\geq1},

0<T(xi)≤T‾(xi)≤T‾,(14)0<T(x_i)\leq\overline{T}(x_i)\leq\overline{T}, \tag*{(14)}

with

T‾:=ΔIb(15)\overline{T} := \frac{\Delta}{I_b} \tag*{(15)}

and

T‾(xi):=1μ(A)ln⁡(W0(∥C∥∥xi∥Ibe∥C∥∥xi∥Ib+Δμ(A)Ib)∥C∥∥xi∥Ib).(16)\underline{T}(x_i) := \frac{1}{\mu(A)} \ln\left(\frac{W_0\left(\frac{\lVert C\rVert\lVert x_i\rVert}{I_b}e^{\frac{\lVert C\rVert\lVert x_i\rVert}{I_b}}+\frac{\Delta\mu(A)}{I_b}\right)}{\frac{\lVert C\rVert\lVert x_i\rVert}{I_b}}\right). \tag*{(16)}

The upper bound (15) alone is enough for Lemma 5 and condition (9), as both only require T(x) to be small enough. Condition (10), however, requires α/T(x)\alpha/T(x) to lie in (K−κ‾,K+κ‾)(K-\overline{\kappa}, K+\overline{\kappa}) for every possible T(x), which an upper bound cannot give. Multiplying the range by T(x), the condition becomes finding a single α\alpha between two linear functions of T(x) over the interval of Lemma 7. Its existence is not trivial. The next lemma gives a sufficient condition for existence using both bounds of Lemma 7.

Lemma 8 (Sufficient condition on IbI_b and Δ\Delta for an admissible α/T(x)\alpha/T(x) range). Let 0<a<b0 < a < b and i∈N≥1i \in\mathbb{N}_{\geq1}, and suppose the following holds

Ib>Δ(a/b)μ(A)ln⁡((a/b)Δμ(A)∥C∥∥xi∥W0((a/b)Δμ(A)∥C∥∥xi∥e(a/b)Δμ(A)∥C∥∥xi∥+(1−a/b)Δμ(A))).(17)I_b > \Delta\frac{(a/b)\mu(A)}{\ln\left(\frac{(a/b)\Delta\mu(A)}{\lVert C\rVert\lVert x_i\rVert}W_0\left(\frac{(a/b)\Delta\mu(A)}{\lVert C\rVert\lVert x_i\rVert}e^{\frac{(a/b)\Delta\mu(A)}{\lVert C\rVert\lVert x_i\rVert}+(1-a/b)\Delta\mu(A)}\right)\right)}. \tag*{(17)}

Then there exists α∈R>0\alpha\in\mathbb{R}_{>0} such that a<α/T<ba < \alpha/T < b for all T∈[T‾(xi),T‾]T \in[\underline{T}(x_i),\overline{T}].

Suppose the parameters of (3) are chosen so that Φ(xi)\Phi(x_i) is contractive, i.e., ∥xi∥P∗>∥xi+1∥P∗\lVert x_i\rVert_{P^*} > \lVert x_{i+1}\rVert_{P^*}. The lower bound (13) depends on ∥x∥\lVert x\rVert, and this dependency propagates to Lemma 8, which becomes harder to satisfy as that bound shrinks. So parameters making Φ(xi)\Phi(x_i) contractive also make the next transition contractive if the bound is monotonically decreasing in ∥x∥\lVert x\rVert.

Lemma 9 (Dwell time lower bound monotonicity). Let Δ,Ib∈R>0\Delta, I_b \in\mathbb{R}_{>0}. The lower bound T‾(x)\underline{T}(x) defined by (13) depends on x∈Rn∖{0}x \in\mathbb{R}^n \setminus\{0\} only through ∥x∥\lVert x\rVert, and is a continuous, strictly decreasing function of ∥x∥∈R>0\lVert x\rVert\in\mathbb{R}_{>0}.

In Lemma 8, IbI_b is easy to isolate but Δ\Delta is not, and the dwell time conditions of Lemmas 5 and 6 are met once they hold at T‾=Δ/Ib\overline{T} = \Delta/I_b, which motivates a common IbI_b lower bound.

Definition 10 (IbI_b total lower bound). Suppose Assumption 2 is satisfied. Given xx, a gain K∈KK \in\mathcal{K}, a positive-definite Q∗∈Rn×nQ^* \in\mathbb{R}^{n\times n}, a threshold Δ∈R>0\Delta\in\mathbb{R}_{>0}, and κ‾∈R>0\overline{\kappa} \in\mathbb{R}_{>0}, the value Ib‾(x,K,Q∗,Δ,κ‾)\underline{I_b}(x,K,Q^*,\Delta,\overline{\kappa}), denoted Ib‾(x)\underline{I_b}(x) for brevity, is the infimum of the Ib∈R>0I_b \in\mathbb{R}_{>0} simultaneously satisfying the condition of Lemma 8 with a=K−κ‾a = K-\overline{\kappa} and b=K+κ‾b = K+\overline{\kappa} (which satisfy2 0<a<b0 < a < b) and the T‾(x)\underline{T}(x) conditions of Lemmas 5 and 6 at the worst case T‾(x)=T‾=Δ/Ib\underline{T}(x) = \overline{T} = \Delta/I_b, all in the energy norm of Lemma 6.

Each of these conditions is a strict lower bound on IbI_b, so all of them hold if and only if Ib>Ib‾(x)I_b > \underline{I_b}(x).

Stability of NICSS

In the previous section we showed that the impact map (5) contracts the LTI state xx when the dwell time and the effective gain satisfy Lemma 6. Definition 10, built on Lemmas 7 and 8, turns these into conditions on the controller parameters. This is enough to show that suitable IbI_b and α\alpha exist for every Δ>0\Delta> 0 and every compact set of initial conditions such that the set {0}×[0,Δ]\{0\}\times[0,\Delta] is asymptotically stable and attracts all solutions starting in it.

Theorem 11 (Existence of a stabilizing spike controller for static-output-feedback-stabilizable SISO linear systems). Consider a NICS (3) for which Assumption 2 holds. Then, for every Δ>0\Delta> 0, the set {0}×[0,Δ]\{0\}\times[0,\Delta] is sGAS in the sense of Definition 1 with θ=(Ib,α)\theta=(I_b,\alpha) and Θ=R>02\Theta=\mathbb{R}^2_{>0}.

Proof. Let K∈KK \in\mathcal{K} and let Q∗∈Rn×nQ^* \in\mathbb{R}^{n\times n} be positive definite. Then the Lyapunov equation (8) has a unique positive-definite solution P∗∈Rn×nP^* \in\mathbb{R}^{n\times n}, and Q∗Q^* and P∗P^* define λ0:=λmin⁡(Q∗)/λmax⁡(P∗)>0\lambda_0 := \lambda_{\min}(Q^*)/\lambda_{\max}(P^*) > 0 (see Section IV). Let Δ>0\Delta> 0 and κ‾∈(0,λ0/(2∥BC∥P∗))\overline{\kappa} \in(0,\lambda_0/(2\lVert BC\rVert_{P^*})). For any compact set Z⊂Rn×R≥0\mathcal{Z}\subset\mathbb{R}^n\times\mathbb{R}_{\geq0} of initial conditions, let x^∈arg⁡max⁡x:(x,ξ)∈Z∥x∥P∗\widehat{x}\in\arg\max_{x:(x,\xi)\in\mathcal{Z}}\lVert x\rVert_{P^*} be the worst-case initial condition, where we assume without loss of generality that x^≠0\widehat{x}\ne0. Every norm below is in the P∗P^*-norm, and μ:=μP∗(A)≥0\mu:=\mu_{P^*}(A)\geq0 since AA is not Hurwitz.

Since the system may start with ξ0=0\xi_0=0, the first flow can be arbitrarily short and the first jump must be treated separately. Let Ib>0I_b>0 and α∈(0,(K+κ‾)T‾)\alpha\in(0,(K+\overline{\kappa})\overline{T}), with T‾=Δ/Ib\overline{T}=\Delta/I_b as in (12), and let z=(x,ξ)z=(x,\xi) be a maximal solution of the NICS (3) from z0=(x0,ξ0)z_0=(x_0,\xi_0). Its first jump occurs at a time t1∈[0,T‾]t_1\in[0,\overline{T}], since t1=0t_1=0 if ξ0≥Δ\xi_0\geq\Delta, as the system cannot flow from such a state, and t1≤(Δ−ξ0)/Ib≤T‾t_1\leq(\Delta-\xi_0)/I_b\leq\overline{T} if ξ0<Δ\xi_0<\Delta. After this jump, the neuron state is reset to 0 and the LTI state is x1=(I−αBC)eAt1x0x_1=(I-\alpha BC)e^{At_1}x_0. Using ∥I−αBC∥≤1+α∥BC∥\lVert I-\alpha BC\rVert\leq1+\alpha\lVert BC\rVert and the logarithmic norm property ∥eAt1∥≤eμt1≤eμT‾\lVert e^{At_1}\rVert\leq e^{\mu t_1}\leq e^{\mu\overline{T}}, we have ∥x1∥≤c(Ib)∥x0∥\lVert x_1\rVert\leq c(I_b)\lVert x_0\rVert with

c(Ib):=(1+(K+κ‾)∥BC∥ΔIb)eμΔIb.(18)c(I_b):=\left(1+(K+\overline{\kappa})\lVert BC\rVert\frac{\Delta}{I_b}\right)e^{\frac{\mu\Delta}{I_b}}. \tag*{(18)}

Both factors are at least 1 and non-increasing in IbI_b, so c≥1c\geq1 and cc is non-increasing.

The first jump can leave the set of initial conditions, so the parameters must be chosen for an enlarged set. Let Ib,0:=Ib‾(x^)I_{b,0}:=\underline{I_b}(\widehat{x}) be the bound of Definition 10 at the worst-case initial condition, let c‾:=c(Ib,0)\overline{c}:=c(I_{b,0}), and let x~∈Rn\widetilde{x}\in\mathbb{R}^n be any state with ∥x~∥=c‾∥x^∥\lVert\widetilde{x}\rVert=\overline{c}\lVert\widehat{x}\rVert. The parameters are then

Ib>max⁡(Ib,0‾,Ib‾(x~)),α∈((K−κ‾)T‾,(K+κ‾)T‾(x~)).(19)I_b>\max(\underline{I_{b,0}},\underline{I_b}(\widetilde{x})),\qquad\alpha\in((K-\overline{\kappa})\overline{T},(K+\overline{\kappa})\underline{T}(\widetilde{x})). \tag*{(19)}

The range of α\alpha is non-empty since the condition of Lemma 8 holds at x~\widetilde{x}, and every α\alpha in it satisfies 0<α<(K+κ‾)T‾0<\alpha<(K+\overline{\kappa})\overline{T} since T‾(x~)≤T‾\underline{T}(\widetilde{x})\leq\overline{T}, so the bound on the first jump applies. Since Ib‾\underline{I_b} and cc is non-increasing, c(Ib)≤c‾c(I_b)\leq\overline{c}, so ∥x1∥≤c‾∥x0∥≤∥x~∥\lVert x_1\rVert\leq\overline{c}\lVert x_0\rVert\leq\lVert\widetilde{x}\rVert for every initial condition with ∥x0∥≤∥x^∥\lVert x_0\rVert\leq\lVert\widehat{x}\rVert and ξ0≥0\xi_0\geq0, which includes every z0∈Zz_0\in\mathcal{Z}.

Now consider the solutions zz of the NICS (3) with the parameters (16). After the first jump, every post-jump state

  1. Need equation 14 uncertain malformed. Image readable: denominator ln( (a/b) Δ μ(A)/||C|| ||xi|| / W0( (a/b) Δ μ(A)/||C|| ||xi|| e^{ ||C|| ||xi|| + (1-a/b) Δ μ(A) /?}) ). Let's inspect image. It shows: Ib > Δ (a/b) μ(A) / ln( [ (a/b) Δ μ(A) / ||C|| ||xi|| ] / W0( (a/b) Δ μ(A)/||C|| ||xi|| e^{||C|| ||xi|| +(1-a/b) Δ μ(A)} ) ) Actually numerator fraction atop, denominator is ln( fraction (a/b)Δμ / ||C||||xi|| divided by W0(...)). transcribe best. Our generated had placement wrong. use visual.

Also theorem labels likely theorem numbering, statement fences okay. Definition footnote marker is superscript 2, should use 2 perhaps yes.

Heading uppercase normalized title case? headings transcribe title likely "Stability of NICSS", drop number. Fine. Require: A system model as in (2), a compact set of initial conditions Z\mathcal{Z}, and a gain KK such that Assumption 2 holds.

  1. Choose a positive-definite Q∗Q^* and compute P∗P^* using the Lyapunov continuous-time equation.

  2. Find x^∈arg⁡max⁡x:(x,ξ)∈Z∥x∥P∗\hat{x} \in\arg\max_{x:(x,\xi)\in\mathcal{Z}}\|x\|_{P^*}.

  3. Choose Δ>0\Delta> 0.

  4. Choose a κˉ∈(0,λ02(∥BC∥P∗))\bar{\kappa} \in\left(0,\frac{\lambda_0}{2(\|BC\|_{P^*})}\right).

  5. Choose Ib>Ib(xˉ)I_b>I_b(\bar{x}), with xˉ\bar{x} being any state that satisfies ∥xˉ∥P∗=c(Ib(xˉ))∥x^∥P∗\|\bar{x}\|_{P^*}=c(I_b(\bar{x}))\|\hat{x}\|_{P^*}, with cc as defined in (15).

  6. Compute the minimum and maximum dwell times T‾(xˉ)\underline{T}(\bar{x}) and T‾\overline{T} using Lemma 7.

  7. Choose α∈((K−κˉ)T‾,(K+κˉ)T‾(xˉ))\alpha\in\left((K-\bar{\kappa})\overline{T},(K+\bar{\kappa})\underline{T}(\bar{x})\right).

is of the form (xi,0)(x_i,0) with i∈N≥1i\in\mathbb{N}_{\geq1}, so the transition matrix (6) describes the evolution of xix_i. Let gg be the function of Lemma 6. Since gg is convex on R≥0\mathbb{R}_{\geq0}, its maximum over [T‾(xˉ),T‾][\underline{T}(\bar{x}),\overline{T}] is reached at an endpoint, and both endpoints satisfy (12) since T‾\overline{T} does, so

γˉ:=max⁡(g(T‾(xˉ)),g(T‾))<1.(20)\bar{\gamma}:=\max\left(g(\underline{T}(\bar{x})),g(\overline{T})\right)<1. \tag*{(20)}

Suppose ∥xi∥≤∥xˉ∥\|x_i\|\leq\|\bar{x}\| for some i∈N≥1i\in\mathbb{N}_{\geq1}. If xi=0x_i=0, the LTI state stays at 00. Otherwise, Lemmas 7 and 9 give T(xi)∈[T‾(xˉ),T‾]T(x_i)\in[\underline{T}(\bar{x}),\overline{T}], so ∣κ(xi)∣<κˉ|\kappa(x_i)|<\bar{\kappa} by the choice of α\alpha, and the dwell time conditions of Lemmas 5 and 6 hold since they hold at T‾\overline{T} by Definition 10. Lemma 6 then bounds the norm of the transition matrix by ∥Φ(xi)∥≤g(T(xi))≤γˉ\|\Phi(x_i)\|\leq g(T(x_i))\leq\bar{\gamma}, so ∥xi+1∥≤γˉ∥xi∥<∥xˉ∥\|x_{i+1}\|\leq\bar{\gamma}\|x_i\|<\|\bar{x}\|. Since ∥x1∥≤∥xˉ∥\|x_1\|\leq\|\bar{x}\|, repeating this argument gives, for every i∈N≥1i\in\mathbb{N}_{\geq1},

∥xi∥≤γˉi−1∥x1∥≤γˉi−1c∥x0∥.(21)\|x_i\|\leq\bar{\gamma}^{i-1}\|x_1\|\leq\bar{\gamma}^{i-1}c\|x_0\|. \tag*{(21)}

Since the system can only flow for a maximum of T‾\overline{T}, the maximum norm it can attain from a state xix_i is eμT‾∥xi∥e^{\mu\overline{T}}\|x_i\|. Since c≥1c\geq1, it follows that, for every (t,i)∈dom⁡z(t,i)\in\operatorname{dom}z,

∥x(t,i)∥≤ceμT‾γˉmax⁡(i−1,0)∥x0∥.(22)\|x(t,i)\|\leq ce^{\mu\overline{T}}\bar{\gamma}^{\max(i-1,0)}\|x_0\|. \tag*{(22)}

Note also that every maximal solution is complete and non-Zeno, since 0<T‾(xˉ)≤T(xi)≤T‾0<\underline{T}(\bar{x})\leq T(x_i)\leq\overline{T} for every i∈N≥1i\in\mathbb{N}_{\geq1}, and that t≤(i+1)T‾t\leq(i+1)\overline{T} for every (t,i)∈dom⁡z(t,i)\in\operatorname{dom}z.

Attractivity follows since t+i→∞t+i\to\infty implies i→∞i\to\infty, so (22) with γˉ<1\bar{\gamma}<1 yields lim⁡t+i→∞∥x(t,i)∥=0\lim_{t+i\to\infty}\|x(t,i)\|=0 and thus lim⁡t+i→∞x(t,i)=0\lim_{t+i\to\infty}x(t,i)=0. Stability follows since, for every ε>0\varepsilon>0, δ:=min⁡{∥x^∥,εe−μT‾/c}\delta:=\min\{\|\hat{x}\|,\varepsilon e^{-\mu\overline{T}}/c\} is such that ∥x0∥≤δ\|x_0\|\leq\delta, with any ξ0≥0\xi_0\geq0, implies ∥x(t,i)∥≤ε\|x(t,i)\|\leq\varepsilon for every (t,i)∈dom⁡z(t,i)\in\operatorname{dom}z by (22).

Finally, the neuron state ξ\xi is always contained in [0,Δ][0,\Delta], except possibly at t=0t=0, so asymptotic stability of 00 for xx gives asymptotic stability of the set {0}×[0,Δ]\{0\}\times[0,\Delta] for z.□z. \square

The theorem shows that parameters always exist for which the LTI state contracts at each flow–jump cycle by emulating a SOF controller, and thus converges to 00. Since ξ\xi stays in [0,Δ][0,\Delta] and does not count for stability, the attractor is the set {0}×[0,Δ]\{0\}\times[0,\Delta] rather than a point.

The proof is constructive: Figure 1 presents an algorithm based on it, designing the controller from the plant parameters and a set of initial conditions. The algorithm chooses six parameters: the controller parameters (Δ,Ib,α)(\Delta,I_b,\alpha) and three intermediate ones (Q∗,P∗,κˉ)(Q^*,P^*,\bar{\kappa}) that follow from the emulation approach, Q∗Q^* and P∗P^* coming from the Lyapunov equation of the emulated continuous system and κˉ\bar{\kappa} being the maximum deviation of the effective gain. Interestingly, the choice of Δ\Delta is completely free: it may be optimized against some objective, but a stabilizing controller exists for every positive Δ\Delta.

Figure 1. Algorithm to design a stable NICS from a stabilizing gain and a set of initial conditions. In Section VI we take Q∗,P∗Q^*,P^* maximizing λ0\lambda_0 by convex optimization (Julia, CONVEX and SCS), Δ=∥C∥P∗∥x^∥P∗eμP∗(A)\Delta=\|C\|_{P^*}\|\hat{x}\|_{P^*}e^{\mu_{P^*}(A)}, κˉ\bar{\kappa} minimizing Ib(xˉ)I_b(\bar{x}), Ib=1.5Ib(xˉ)I_b=1.5I_b(\bar{x}), and α\alpha in the middle of its range.

One may ask whether stability could be established with methods relying on parameter-dependent LMIs such as [17]. With the jump map constrained to the form I−αBCI-\alpha BC, such results can certify a given set of parameters, but they cannot directly prove that stabilizing (α,Δ,Ib)(\alpha,\Delta,I_b) exist for the whole class of SOF stabilizable SISO systems.

One may also wonder why the term ∣Cx∣|Cx| is included in (3). If it is removed, firing becomes periodic, and global stability can be proven with Lemmas 5 and 6 (periodic case already covered by [17 Theorem 2.1]). The output-driven firing, i.e., the integrate-and-fire mechanism, may have two benefits: faster convergence than periodic firing and, when K=(Kmin⁡,∞)K=(K_{\min},\infty) and ∣1−αCB∣≤1|1-\alpha CB|\leq1, bounded trajectories for all Δ\Delta and IbI_b.

Numerical Example

We consider an inverted pendulum linearized around its upright equilibrium, with a simplified motor model. It is a SISO LTI system with state x=(θ,θ˙,I)∈R3x=(\theta,\dot{\theta},I)\in\mathbb{R}^3, input u∈Ru\in\mathbb{R}, and matrices

A=[010p1−p2p30−p4−p5],B=[00p6],C=[100],(23)A=\begin{bmatrix}0&1&0\\p_1&-p_2&p_3\\0&-p_4&-p_5\end{bmatrix},\qquad B=\begin{bmatrix}0\\0\\p_6\end{bmatrix},\qquad C=\begin{bmatrix}1&0&0\end{bmatrix}, \tag*{(23)}

with p1,p2,p3,p4,p5,p6∈R>0p_1,p_2,p_3,p_4,p_5,p_6\in\mathbb{R}_{>0}.

We consider p1=1p_1=1, p2=2p_2=2, p3=1p_3=1, p4=1p_4=1, p5=1p_5=1, and p6=2p_6=2. Assumption 2 holds with K=(2,18)K=(2,18). We choose K‾=3.2\overline{K}=3.2 for our design. Designing for a compact set Z\mathcal{Z} whose largest element in the P∗P^*-norm is (0.5,0.5,0)(0.5,0.5,0) and applying the algorithm outlined in Figure 1 with the choices in gray yields Ib=3767.007I_b=3767.007, Δ=11.611\Delta=11.611, and α=0.009854\alpha=0.009854.

Figure 2 compares the closed loop under the parameters returned by the algorithm with a “relaxed” choice that keeps αIb/Δ\alpha I_b/\Delta constant but divides IbI_b by 3000 and Δ\Delta by 4, and with the continuous loop being emulated. The relaxed controller emulates the same gain at the equilibrium but no longer fits the bounds established by (20). Since IbI_b is reduced much more than Δ\Delta, it achieves a far larger maximum dwell time, lowering the frequency of actuation at the origin by a factor of 750.

Figure 2. Trajectories of the controlled linearized inverted pendulum from (θ,θ˙,I)=(2.5,2.5,0.0)(\theta,\dot{\theta},I)=(2.5,2.5,0.0), for the parameters returned by the algorithm (blue), the relaxed parameters of Section VI (red), and the emulated continuous loop (green). From top to bottom: θ\theta, θ˙\dot{\theta}, II, and the P∗P^*-energy norm of the state, for which the theoretical minimum decay is also plotted.

Two observations stand out. First, the algorithm’s parameters yield trajectories closely following the continuous system, as expected from an emulation approach, whereas the relaxed controller diverges from it quickly even though it emulates the same gain at the origin. Second, the controller can sit outside the bounds fixed by the proof and still stabilize the plant: the original one starts five times outside its design set, and the relaxed one also uses parameters beyond the bounds of Theorem 11. The returned parameters are therefore only sufficient, and emulating a single gain KK ignores the whole range of stabilizing gains the plant admits.

Conclusion

We have designed a controller at the crossroads of neuromorphic and traditional impulsive control, using a single integrate-and-fire-like neural variable with a spike amplitude proportional to the output, and shown that one such neuron semi-globally asymptotically stabilizes any SOF stabilizable SISO LTI system. We also proposed an algorithm to design the controller from the plant parameters, and tested it in simulations on a linearized inverted pendulum.

At this stage the contribution is mostly conceptual: an actuator that can deliver pulses of any amplitude can usually deliver a continuous signal as well, and the proportional spike then gives up the main practical advantage of neuromorphic control. The result is still useful, as it is the shrinking of the spikes with the state, and not only the firing time, that decides between asymptotic and practical stability. This can still be used with fixed-amplitude spikes by building arrays of neurons working at different scales so that the total correction is quantized but keeps decreasing with the state, much like the motor units recruited in biological muscle depending on the required torque [25]; such an array should converge with fewer spikes and a smaller ultimate bound than a single neuron pair. The work could also be extended to the multiple-input multiple-output case, which would probably require average dwell times, and it would be interesting to see how much of the conservatism of the proof and of the algorithm can be pruned.

Appendix

Proofs

Proof of uniqueness of solutions of NICSs (3). Points (a) and (c) of [19 Proposition 2.11] hold for NICSs since the flow map (3a) is globally Lipschitz, so solutions of the flow are unique [26 Theorem 3.2], and the jump map (3b) is single-valued.

flow map (3a) is globally Lipschitz, so solutions of the flow are unique [26 Theorem 3.2], and the jump map (3b) is single-valued. Point (b) of [19 Proposition 2.11] only concerns the states in C∩D\mathcal{C}\cap\mathcal{D}, i.e., z=(x,Δ)z=(x,\Delta) with x∈Rnx\in\mathbb{R}^n. But since the flow map (3a) always returns a strictly positive value for the neuron variable ξ\xi, there can exist no trajectory z∗:[0,ε]→Rn+1z^*:[0,\varepsilon]\to\mathbb{R}^{n+1} with z∗(0)=(x,Δ)z^*(0)=(x,\Delta) such that z˙∗(t)=f(z∗(t))\dot{z}^*(t)=f(z^*(t)) for almost all t∈[0,ε]t\in[0,\varepsilon] and z∗(t)∈Cz^*(t)\in\mathcal{C} for all t∈(0,ε]t\in(0,\varepsilon]. Indeed, if z˙∗(t)=f(z∗(t))\dot{z}^*(t)=f(z^*(t)), then ξ∗(t)>Δ\xi^*(t)>\Delta for any t∈(0,ε]t\in(0,\varepsilon] since ξ˙∗>0\dot{\xi}^*>0. So point (b) of [19 Proposition 2.11] is also satisfied and the solution to NICSs from any starting point in C∪D=Rn×R≥0\mathcal{C}\cup\mathcal{D}=\mathbb{R}^n\times\mathbb{R}_{\geq0} is unique. □

Proof of Lemma 5. Consider any maximal solution zz of the NICS and any i∈N≥1i\in\mathbb{N}_{\geq1}. For brevity, we write TT, κ\kappa, and Φ\Phi for T(xi)T(x_i), κ(xi)\kappa(x_i), and Φ(xi)\Phi(x_i) in this proof. The transition matrix of the LTI state can be expressed as

Φ=Φ−e(A−BKC)T+e(A−BKC)T(24)\Phi=\Phi-e^{(A-BKC)T}+e^{(A-BKC)T} \tag*{(24)}

which, in norms, gives

∥Φ∥≤∥e(A−BKC)T∥+∥Φ−e(A−BKC)T∥.(25)\lVert\Phi\rVert\leq\lVert e^{(A-BKC)T}\rVert+\lVert\Phi-e^{(A-BKC)T}\rVert. \tag*{(25)}

Using the Taylor expansion of both exponentials, the difference between the transition matrix and the exponential can be written as

Φ−e(A−BKC)T=T(K−αT)BC+∑m=2∞Tmm!(A−mαTBC)Am−1−∑m=2∞Tmm!(A−BKC)m.(26)\begin{aligned} \Phi-e^{(A-BKC)T} &=\mathrm{T}\left(K-\frac{\alpha}{\mathrm{T}}\right)BC\\ &\quad+\sum_{m=2}^{\infty}\frac{\mathrm{T}^m}{m!}\left(A-m\frac{\alpha}{\mathrm{T}}BC\right)A^{m-1}\\ &\quad-\sum_{m=2}^{\infty}\frac{\mathrm{T}^m}{m!}(A-BKC)^m. \tag*{(26)} \end{aligned}

Then, by setting κ=K−α/T\kappa=K-\alpha/\mathrm{T}, developing the first term of each sum together, using the bound ∑m=k∞1m!am<ak∑m=k∞1m!\sum_{m=k}^{\infty}\frac{1}{m!}a^m<a^k\sum_{m=k}^{\infty}\frac{1}{m!} valid if 0<a<10<a<1 (with a=T∥A∥a=\mathrm{T}\lVert A\rVert and a=T∥A−BKC∥a=\mathrm{T}\lVert A-BKC\rVert), and using the identity ∑m=0∞1m!=e\sum_{m=0}^{\infty}\frac{1}{m!}=e, the constants C1(κ)C_1(\kappa), C2(κ)C_2(\kappa), and C3(κ)C_3(\kappa) of the lemma can be computed. □

Proof of Lemma 6. Consider any maximal solution zz of the NICS and any i∈N≥1i\in\mathbb{N}_{\geq1}, and write T for T(xi)T(x_i) in this proof. From [24 Section 8.5], the standard Lyapunov decay estimate gives

∥e(A−BKC)T∥P∗≤e−λ0T/2,λ0:=λmin⁡(Q∗)λmax⁡(P∗).\left\lVert e^{(A-BKC)T}\right\rVert_{P^*}\leq e^{-\lambda_0T/2},\qquad\lambda_0:=\frac{\lambda_{\min}(Q^*)}{\lambda_{\max}(P^*)}.

Since the constants (7) are non-decreasing in ∣κ∣\lvert\kappa\rvert and ∣κ(xi)∣<κˉ\lvert\kappa(x_i)\rvert<\bar{\kappa} by (10), Ck(κ(xi))≤CˉkC_k(\kappa(x_i))\leq\bar{C}_k for k∈{1,2,3}k\in\{1,2,3\}. Plugging this in the results of Lemma 5 expressed in the P∗P^*-norm yields

∥Φ(xi)∥P∗≤e−λ0T/2+Cˉ1T+Cˉ2T2+Cˉ3T3,(27)\left\lVert\Phi(x_i)\right\rVert_{P^*}\leq e^{-\lambda_0T/2}+\bar{C}_1T+\bar{C}_2T^2+\bar{C}_3T^3, \tag*{(27)}

since T≤min⁡(∥A∥P∗−1,∥A−BKC∥P∗−1)T\leq\min\left(\lVert A\rVert_{P^*}^{-1},\lVert A-BKC\rVert_{P^*}^{-1}\right).

At T=0T=0 the right-hand side of (24) equals one and, since its second derivative with respect to TT is strictly positive for any TT, there exists at most one interval (0,T∗)(0,T^*) where (24) is smaller than 1, with T∗>0T^*>0 the point at which the right-hand side of (24) equals one. Since the first derivative of the right-hand side of (27) at T=0T=0 equals Cˉ1−λ0/2\bar{C}_1-\lambda_0/2, the interval only exists when Cˉ1=κˉ∥BC∥P∗<λ0/2\bar{C}_1=\bar{\kappa}\lVert BC\rVert_{P^*}<\lambda_0/2, which is the condition on κˉ\bar{\kappa} of the lemma.

Using the inequality e−λ0T/2≤1−λ02T+λ028T2e^{-\lambda_0T/2}\leq1-\frac{\lambda_0}{2}T+\frac{\lambda_0^2}{8}T^2, valid for T≥0T\geq0, a lower bound on T∗T^* can be found by solving the resulting cubic equation. Since T=0T=0 is a root, the cubic factors as T(Cˉ2T2+(Cˉ2+λ028)T+(Cˉ1−λ02))=0T\left(\bar{C}_2T^2+\left(\bar{C}_2+\frac{\lambda_0^2}{8}\right)T+\left(\bar{C}_1-\frac{\lambda_0}{2}\right)\right)=0, and the positive root of the remaining quadratic equation is the bound (9) stated in the lemma.

Proof of Lemma 7. Consider any maximal solution zz of the NICS and any i∈N≥1i\in\mathbb{N}_{\geq1}, and write TT for T(xi)T(x_i) in this proof. Starting from the state (xi,0)(x_i,0), the LTI state flows as eAτxie^{A\tau}x_i, so bounding the dwell time TT amounts to bounding the T>0T>0 that makes the neuron state

∫0Tξ˙ dτ=∫0T(∣CeAτxi∣+Ib)dτ=IbT+∫0T∣CeAτxi∣ dτ(28)\int_0^T \dot{\xi}\,\mathrm{d}\tau=\int_0^T\left(\lvert Ce^{A\tau}x_i\rvert+I_b\right)\mathrm{d}\tau =I_bT+\int_0^T\lvert Ce^{A\tau}x_i\rvert\,\mathrm{d}\tau \tag*{(28)}

equal to Δ\Delta.

To compute the upper bound, we can lower bound (25), which is easily done by ignoring the effect of the input ∣Cx∣\lvert Cx\rvert. To compute the lower bound, we can upper bound (25) by using the properties of compatible vector and matrix norms and logarithmic norms, ∣CeAτxi∣≤∥C∥eμ(A)τ∥xi∥\lvert Ce^{A\tau}x_i\rvert\leq\lVert C\rVert e^{\mu(A)\tau}\lVert x_i\rVert, then solving the integral. This yields the equation

IbT(xi)+∥C∥∥xi∥μ(A)(eμ(A)T(xi)−1)=Δ,(29)I_bT(x_i)+\frac{\lVert C\rVert\lVert x_i\rVert}{\mu(A)}\left(e^{\mu(A)T(x_i)}-1\right)=\Delta, \tag*{(29)}

assuming μ(A)≠0\mu(A)\neq0; the case μ(A)=0\mu(A)=0 reduces to a linear equation. This transcendental equation admits a unique real solution, which can be expressed using the principal branch of the Lambert WW function.

Proof of Lemma 8. Consider any maximal solution zz of the NICS and any i∈N≥1i\in\mathbb{N}_{\geq1}. A single α\alpha with α/T(xi)∈(a,b)\alpha/T(x_i)\in(a,b) for every T(xi)∈[T‾(xi),T‾]T(x_i)\in[\underline{T}(x_i),\overline{T}] exists if and only if aT‾<bT(xi)a\overline{T}<bT(x_i). Using the expressions from Lemma 7 and rearranging, this condition can be expressed as

aΔIb<b(ΔIb+1μ(A)(∥C∥∥xi∥Ib−W0(∥C∥∥xi∥IbeΔμ(A)IbeΔμ(A)Ib))).(30)a\frac{\Delta}{I_b}<b\left(\frac{\Delta}{I_b}+\frac{1}{\mu(A)}\left(\frac{\lVert C\rVert\lVert x_i\rVert}{I_b}-W_0\left(\frac{\lVert C\rVert\lVert x_i\rVert}{I_b}e^{\frac{\Delta\mu(A)}{I_b}}e^{\frac{\Delta\mu(A)}{I_b}}\right)\right)\right). \tag*{(30)}

By rearranging and using the upper bound W0(x)<ln⁡(x+1)W_0(x)<\ln(x+1), valid for x>0x>0, the inequality is implied by

∥C∥∥xi∥Ib<e−(a/b)Δμ(A)Ib(1−e(a/b−1)Δμ(A)Ibe−∥C∥∥xi∥Ib).(31)\frac{\lVert C\rVert\lVert x_i\rVert}{I_b}<e^{-(a/b)\frac{\Delta\mu(A)}{I_b}}\left(1-e^{(a/b-1)\frac{\Delta\mu(A)}{I_b}}e^{-\frac{\lVert C\rVert\lVert x_i\rVert}{I_b}}\right). \tag*{(31)}

Then, using the lower bound 1−e−x>x/(x+1)1-e^{-x}>x/(x+1), valid for x>0x>0, and rearranging, the existence is implied by

∥C∥∥xi∥Ib(1−a/b)Δμ(A)+∥C∥∥xi∥(1−a/b)Δμ(A)+∥C∥∥xi∥e−(a/b)Δμ(A)Ib((1−a/b)Δμ(A)+∥C∥∥xi∥)<−∥C∥∥xi∥.(32)\frac{\lVert C\rVert\lVert x_i\rVert}{I_b}\frac{(1-a/b)\Delta\mu(A)+\lVert C\rVert\lVert x_i\rVert}{(1-a/b)\Delta\mu(A)+\lVert C\rVert\lVert x_i\rVert} e^{-(a/b)\frac{\Delta\mu(A)}{I_b}} \left((1-a/b)\Delta\mu(A)+\lVert C\rVert\lVert x_i\rVert\right) <-\lVert C\rVert\lVert x_i\rVert. \tag*{(32)}

assuming μ(A)>0\mu(A)>0.

Finally, since the inequality is monotone in IbI_b and is satisfied as Ib→∞I_b\to\infty, the lower bound on IbI_b satisfying the inequality can be found by solving the equality. This transcendental equation admits a unique real solution, expressible using the principal branch of the Lambert WW function.

Proof of Lemma 9. Writing (26) for a generic state xx and using implicit differentiation, the derivative of the lower bound from Lemma 7 with respect to ∥x∥\lVert x\rVert can easily be computed as

dT(x)d∥x∥=−∂(26)∂∥x∥∂(26)∂T=−∥C∥μ(A)(eμ(A)T(x)−1)Ib+∥C∥∥x∥eμ(A)T(x)(33)\frac{\mathrm{d}T(x)}{\mathrm{d}\lVert x\rVert} =-\frac{\frac{\partial(26)}{\partial\lVert x\rVert}}{\frac{\partial(26)}{\partial T}} =-\frac{\frac{\lVert C\rVert}{\mu(A)}\left(e^{\mu(A)T(x)}-1\right)}{I_b+\lVert C\rVert\lVert x\rVert e^{\mu(A)T(x)}} \tag*{(33)}

and since the numerator and the denominator are strictly positive, the derivative is always strictly negative, assuming μ(A)≠0\mu(A)\neq0.

Acknowledgment

Claude Opus 5 and 5.5 (Anthropic) were used to edit the phrasing throughout and to draft initial versions of some passages in Section I. All technical content, results, and proofs are the authors’ own, and the authors reviewed all resulting text and are responsible for it.

References

  1. [1]N. Wiener, Cybernetics or Control and Communication in the Animal and the Machine. The MIT Press, [1948] 2019.
  2. [2]C. Mead, “Neuromorphic electronic systems,” Proceedings of the IEEE, vol. 78, no. 10, pp. 1629–1636, 1990.DOI
  3. [3]S. P. Deweerth, L. Nielsen, C. A. Mead, and K. J. Astrom, “A simple neuron servo,” IEEE Transactions on Neural Networks, vol. 2, no. 2, pp. 248–251, 1991.pubmed.ncbi.nlm.nih.gov/18276378
  4. [4]L. Ribar and R. Sepulchre, “Neuromorphic Control: Designing Multiscale Mixed-Feedback Systems,” IEEE Control Systems Magazine, vol. 41, no. 6, pp. 34–63, 2021.DOI
  5. [6]E. Petri, K. J. A. Scheres, E. Steur, and W. P. M. H. Heemels, “Emulation-based Neuromorphic Control for the Stabilization of LTI Systems,” 2025.arxiv.org/abs/2511.11875
  6. [7]L. Eilers, J. Stapmanns, C. Dias, and J.-P. Pfister, “On the stability of event-based control with neuronal dynamics,” in 2026 European Control Conference (ECC), Jul. 2026, pp. 2361–2368.arxiv.org/abs/2511.18015
  7. [8]E. Izhikevich, “Which model to use for cortical spiking neurons?” IEEE Transactions on Neural Networks, vol. 15, no. 5, pp. 1063–1070, 2004.pubmed.ncbi.nlm.nih.gov/15484883
  8. [9]A. N. Burkitt, “A Review of the Integrate-and-fire Neuron Model: I. Homogeneous Synaptic Input,” Biological Cybernetics, vol. 95, no. 1, pp. 1–19, 2006.pubmed.ncbi.nlm.nih.gov/16622699
  9. [10]A. Franci, G. Drion, V. Seutin, and R. Sepulchre, “A Balance Equation Determines a Switch in Neuronal Excitability,” PLOS Computational Biology, vol. 9, no. 5, p. e1003040, 2013.DOI
  10. [11]R. Sepulchre, G. Drion, and A. Franci, “Control Across Scales by Positive and Negative Feedback,” Annual Review of Control, Robotics, and Autonomous Systems, vol. 2, pp. 89–113, 2019.DOI
  11. [12]C. C. Li and R. W. Jones, “Integral pulse frequency modulated control systems,” IFAC Proceedings Volumes, vol. 1, no. 2, pp. 186–195, 1963.DOI
  12. [13]E. A. King-Smith and J. R. Cumpston, “Periodic cycles in integral pulse frequency modulation (IPFM) systems,” International Journal of Control, vol. 9, no. 1, pp. 1–16, 1969.
  13. [14]E. Petri, K. J. A. Scheres, E. Steur, and W. P. M. H. Heemels, “Analysis of a simple neuromorphic controller for linear systems: A hybrid systems perspective,” in 2024 IEEE 63rd Conference on Decision and Control (CDC), 2024, pp. 8578–8583.arxiv.org/abs/2409.06353
  14. [15]W. M. Haddad, V. Chellaboina, and S. G. Nersesov, Impulsive and Hybrid Dynamical Systems: Stability, Dissipativity, and Control. Princeton University Press, 2006.
  15. [16]R. W. Erickson and D. Maksimovic, Fundamentals of Power Electronics. Cham: Springer International Publishing, 2020.
  16. [17]C. Briat, “Convex conditions for robust stability, analysis and stabilization of linear aperiodic impulsive and sampled-data systems under dwell-time constraints,” Automatica, vol. 49, no. 11, pp. 3449–3457, 2013.arxiv.org/abs/1304.1998
  17. [18]J. P. Hespanha, D. Liberzon, and A. R. Teel, “Lyapunov conditions for input-to-state stability of impulsive systems,” Automatica, vol. 44, no. 11, pp. 2735–2744, 2008.DOI
  18. [19]R. Goebel, R. G. Sanfelice, and A. R. Teel, Hybrid Dynamical Systems: Modeling, Stability, and Robustness. Princeton University Press, 2012.
  19. [20]D. Nesic, A. R. Teel, and D. Carnevale, “Explicit Computation of the Sampling Period in Emulation of Controllers for Nonlinear Sampled-Data Systems,” IEEE Transactions on Automatic Control, vol. 54, no. 3, pp. 619–624, 2009.DOI
  20. [21]T. Ström, “On logarithmic norms,” SIAM Journal on Numerical Analysis, vol. 12, no. 5, pp. 741–753, 1975.DOI
  21. [22]R. M. Corless, G. H. Gonnet, D. E. G. Hare, D. J. Jeffrey, and D. E. Knuth, “On the Lambert W function,” Advances in Computational Mathematics, vol. 5, no. 1, pp. 329–359, 1996.
  22. [23]A. Astolfi and P. Colaneri, “Static output feedback stabilization: from linear to nonlinear and back,” in Nonlinear control in the Year 2000, A. Isidori, F. Lamnabhi-Lagarrigue, and W. Respondek, Eds. London: Springer, 2001, pp. 49–71.
  23. [24]J. P. Hespanha, Linear Systems Theory: Second Edition. Princeton University Press, 2018.DOI
  24. [25]A. J. Fuglevand, D. A. Winter, and A. E. Patla, “Models of recruitment and rate coding organization in motor-unit pools,” Journal of Neurophysiology, vol. 70, no. 6, pp. 2470–2488, 1993.pubmed.ncbi.nlm.nih.gov/8120594
  25. [26]T. C. Sideris, Ordinary Differential Equations and Dynamical Systems, ser. Atlantis Studies in Differential Equations. Paris: Atlantis Press, 2013, vol. 2.DOI

Paper details

vlnc:W_67wapf92
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 1

downloads · 0

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

  • Engineering

Claude Opus 5 (bare)

View statement
Looking for related papers…
Contents