betaFeedback

WorksMathematics

Optimal Dynamic Quantizers for Periodically Time-Varying Systems via Cyclic Reformulation

Abstract

This paper addresses the design of optimal dynamic quantizers for discrete-time linear periodically time-varying (LPTV) plants with discrete-valued inputs. Optimal dynamic quantizers, which minimize the worst-case output error caused by quantization, are well established for linear time-invariant (LTI) plants, whereas the corresponding design for plants with periodically time-varying dynamics has not been addressed. For square LPTV plants having relative degree one at every time phase, this paper derives an optimal periodic dynamic quantizer in closed form. The quantizer is constructed by applying the LTI theory to the cycled plant obtained by cyclic reformulation and reading off its periodic coefficient matrices block by block; its optimality is then verified directly, showing that it attains a performance lower bound determined by the largest instantaneous input-to-output gain over the time phases. The worst-case output error of general periodic dynamic quantizers is also characterized. The optimal quantizer is shown to be stable if and only if all invariant zeros of the cycled plant lie in the open unit disk, which is the condition that governs stable causal inversion of LPTV plants. A numerical example with comparisons to conventional quantizers illustrates the results.

Introduction

Control systems with discrete-valued input signals arise in many practical situations, including mechanical systems with ON/OFF actuators, networked control systems with finite communication bandwidth, and power electronics with switching converters. The design of quantizers that map continuous-valued control signals to discrete-valued ones is a central issue in such systems.

For linear time-invariant (LTI) plants, Azuma and Sugie [2] established a systematic framework for the design of optimal dynamic quantizers, which minimize the worst-case output error caused by quantization. In particular, Azuma and Sugie [2] derived a closed-form optimal dynamic quantizer that cancels the influence of past quantization errors on the output, leaving only the instantaneous error determined by the input-to-output gain CBCB. The stability of this quantizer is governed by the zeros of the plant; for nonminimum phase plants, Azuma and Sugie [3] developed a linear programming-based synthesis, and Sawada and Shin [27] proposed an LMI-based invariant set approach for stable suboptimal quantizer design.

The framework has also been extended to nonlinear plants (Azuma and Sugie, [5]) and to decentralized quantizer structures (Minami et al., [18]), and Fujimoto and Minami [11] recently proposed a data-driven design that avoids the need for an explicit plant model. Dynamic quantizers have been applied to feedback control (Minami et al., [17]; Azuma and Sugie, [4]), to networked control systems with communication rate constraints (Okajima et al., [24]; Almakhles et al., [1]) or with time-varying delays and packet losses (Li et al., 2015), and to encrypted control (Teranishi and Kogiso, [29]).

Quantization and discrete-valued control have also been studied in other contexts. From a signal processing perspective, Quevedo and Goodwin [26] proposed a multistep optimal converter that generalizes ΔΣ\Delta\Sigma modulation, and Ohno and Tariq [20] addressed noise shaping filter optimization for quantizers. In quantized feedback stabilization, quantizers whose sensitivity is adjusted online were introduced by Brockett and Liberzon [9], and quantization with dynamic or encoding–decoding mechanisms continues to be studied in iterative learning control [15] and in state estimation over networks [6, 16]. Finally, discrete-valued control has been approached through sum-of-absolute-values optimization [13].

Despite these advances, optimal dynamic quantizer design for plants with periodically time-varying dynamics has not been addressed. Linear periodically time-varying (LPTV) systems arise in multirate sampled-data systems [10], in systems with periodic scheduling of resources, and in plants subject to periodic operating conditions [30]. When the actuator of such a system accepts only discrete-valued inputs, the quantizer must be designed for a periodic plant. In a preliminary study, Okajima et al. [23] proposed a lifting-based periodically time-varying dynamic quantizer for networked control systems with periodically varying communication rate constraints, enabling effective quantization at non-integer bit rates such as 1.5 bits per sampling period. However, that work addressed a time-varying quantizer structure under fixed LTI plant dynamics. Dynamic quantizers for multirate control systems were also studied by Morita et al. [19], but for an LTI plant combined with hold devices of different periods.

This paper derives a closed-form optimal periodic dynamic quantizer for square LPTV plants having relative degree one at every time phase, i.e., plants whose instantaneous input-to-output gain is nonsingular at every phase; this is the periodic counterpart of the LTI result of Azuma and Sugie [2]. The quantizer is obtained through cyclic reformulation [7, 8], which transforms an LPTV system into an equivalent LTI system at the original sampling rate while preserving the step-by-step input–output correspondence. This technique has been applied to robust estimator synthesis [12], system identification [22], and stable inversion of LPTV plants [21]. Applying the LTI theory to the cycled plant yields an LTI quantizer whose coefficient matrices retain the cyclic block structure, so that a periodically time-varying quantizer with the original state dimension can be read off block by block; its optimality is then verified directly in the original time domain.

The contributions of this paper are as follows. First, the optimal dynamic quantizer design problem for LPTV plants is formulated, and a closed-form optimal periodic quantizer is derived together with the minimum achievable worst-case output error, generalizing the LTI result of Azuma and Sugie [2]. Second, the worst-case output error of a general periodic dynamic quantizer is characterized in terms of phase-wise impulse response matrices. Third, the optimal quantizer is shown to be stable if and only if the plant is periodically minimum phase, i.e., all invariant zeros of the cycled plant lie in the open unit disk, which is exactly the condition for stable causal inversion of the plant [21].

This paper is organized as follows. Section 2 presents the preliminaries on cyclic reformulation, reviews the existing optimal dynamic quantizer theory for LTI systems, and states the problem formulation for LPTV systems. Section 3 presents the main results: the performance analysis of general periodic quantizers, the optimal quantizer, its stability analysis, and a numerical example. Section 4 discusses the role of cyclic reformulation and the limitations of the results, and Section 5 concludes the paper.

Notation: Let R\mathbb{R}, R+\mathbb{R}_{+}, Z\mathbb{Z}, Z+\mathbb{Z}_{+}, and N\mathbb{N} denote the real numbers, positive reals, integers, nonnegative integers, and positive integers, respectively. For a matrix M={Mij}M = \{M_{ij}\}, abs⁡(M):={∣Mij∣}\operatorname{abs}(M) := \{|M_{ij}|\} denotes the elementwise absolute value, and ∥M∥\lVert M \rVert denotes the ℓ∞\ell_{\infty}-induced norm (maximum absolute row sum). For a square matrix MM, ρ(M)\rho(M) denotes the spectral radius, and MM is called Schur stable if ρ(M)<1\rho(M) < 1. For a vector xx, ∥x∥\lVert x \rVert is the ℓ∞\ell_{\infty}-norm. For a signal sequence X={x0,x1,…}X = \{x_0, x_1, \ldots\}, ∥X∥:=sup⁡k∥xk∥\lVert X \rVert:= \sup_k \lVert x_k \rVert; the space ℓ∞m\ell^{m}_{\infty} denotes the set of bounded sequences in Rm\mathbb{R}^{m}. The symbols InI_n and 0n×m0_{n \times m} denote the n×nn \times n identity matrix and the n×mn \times m zero matrix (the subscripts are omitted when the dimensions are clear from the context), 1n\boldsymbol{1}_n is the nn-vector of ones, and ⊗\otimes denotes the Kronecker product. For a square system (A,B,C)(A, B, C) with zero feedthrough, the invariant zeros are the values z0∈Cz_0 \in\mathbb{C} at which the rank of the system pencil [z0I−A−BC0]\begin{bmatrix} z_0 I-A & -B \\ C & 0 \end{bmatrix} drops below its normal rank. The static quantizer q:Rm→Vmq : \mathbb{R}^{m} \to\mathbb{V}^{m} with V:=dZ\mathbb{V} := d\mathbb{Z} and quantization interval d∈R+d \in\mathbb{R}_{+} is the nearest-neighbor quantizer satisfying abs⁡(q(μ)−μ)≤d21m\operatorname{abs}(q(\mu)-\mu) \le\frac{d}{2}\boldsymbol{1}_m for all μ∈Rm\mu\in\mathbb{R}^{m}. Throughout this paper, all subscripts of NN-periodic matrices (such as AkA_k, BkB_k, CkC_k, and their derived quantities) are interpreted modulo NN without further notice; that is, Ak=Ak mod NA_k = A_{k \bmod N} (e.g., A−1=AN−1A_{-1} = A_{N-1} and CN+1=C1C_{N+1} = C_1).

Preliminaries and Problem Formulation

Cyclic Reformulation of Periodically Time-Varying Systems

Consider a discrete-time LPTV system with period NN (N∈NN \in\mathbb{N}):

x(k+1)=Akx(k)+Bku(k),y(k)=Ckx(k),(1)\begin{aligned} x(k+1) &= A_kx(k)+B_ku(k),\\ y(k) &= C_kx(k), \tag*{(1)} \end{aligned}

where x∈Rnx \in\mathbb{R}^{n}, u∈Rmu \in\mathbb{R}^{m}, y∈Rpy \in\mathbb{R}^{p}, and Ak+N=AkA_{k+N}=A_k, Bk+N=BkB_{k+N}=B_k, Ck+N=CkC_{k+N}=C_k for all k∈Z+k \in\mathbb{Z}_{+}. Here, Dk=0D_k=0 is assumed for simplicity, which is natural for many sampled-data systems.

The cyclic reformulation [7, 8] transforms this LPTV system into a time-invariant system. Letting i:=k mod N∈{0,1,…,N−1}i := k \bmod N \in\{0,1,\ldots,N-1\} denote the time phase, the cycled input signal u~(k)∈RNm\tilde{u}(k) \in\mathbb{R}^{Nm} is constructed by placing u(k)u(k) in the ii-th block position and setting all other blocks to zero:

u~(k)=ei⊗u(k),i:=k mod N,(2)\tilde{u}(k)=e_i\otimes u(k),\qquad i:=k\bmod N, \tag*{(2)}

where ei∈RNe_i \in\mathbb{R}^{N} is the (i+1)(i+1)-th standard unit vector. For example, u~(0)=[u(0)T,0,…,0]T\tilde{u}(0)=[u(0)^{\mathsf{T}},0,\ldots,0]^{\mathsf{T}}, u~(1)=[0,u(1)T,0,…,0]T\tilde{u}(1)=[0,u(1)^{\mathsf{T}},0,\ldots,0]^{\mathsf{T}}, and so on. The cycled state x~(k)∈RNn\tilde{x}(k) \in\mathbb{R}^{Nn} and cycled output y~(k)∈RNp\tilde{y}(k) \in\mathbb{R}^{Np} are defined analogously.

The cycled system is the following time-invariant system:

x~(k+1)=A~x~(k)+B~u~(k),y~(k)=C~x~(k),(3)\begin{aligned} \tilde{x}(k+1)&=\tilde{A}\tilde{x}(k)+\tilde{B}\tilde{u}(k),\\ \tilde{y}(k)&=\tilde{C}\tilde{x}(k), \tag*{(3)} \end{aligned}

where A~\tilde{A} and B~\tilde{B} have a block-circulant structure (the nonzero block of the jj-th block-column appears in block-row j+1j+1) and C~\tilde{C} is block-diagonal:

A~=[00⋯0AN−1A00⋯000A1⋱⋮⋮⋮⋱⋱000⋯0AN−20]∈RNn×Nn,(4)\tilde{A}= \begin{bmatrix} 0 & 0 & \cdots& 0 & A_{N-1}\\ A_0 & 0 & \cdots& 0 & 0\\ 0 & A_1 & \ddots& \vdots& \vdots\\ \vdots& \ddots& \ddots& 0 & 0\\ 0 & \cdots& 0 & A_{N-2} & 0 \end{bmatrix} \in\mathbb{R}^{Nn\times Nn}, \tag*{(4)}
B~=[00⋯0BN−1B00⋯000B1⋱⋮⋮⋮⋱⋱000⋯0BN−20]∈RNn×Nm,(5)\tilde{B}= \begin{bmatrix} 0 & 0 & \cdots& 0 & B_{N-1}\\ B_0 & 0 & \cdots& 0 & 0\\ 0 & B_1 & \ddots& \vdots& \vdots\\ \vdots& \ddots& \ddots& 0 & 0\\ 0 & \cdots& 0 & B_{N-2} & 0 \end{bmatrix} \in\mathbb{R}^{Nn\times Nm}, \tag*{(5)}
C~=diag⁡(C0,C1,…,CN−1)∈RNp×Nn.(6)\tilde{C}=\operatorname{diag}(C_0,C_1,\ldots,C_{N-1})\in\mathbb{R}^{Np\times Nn}. \tag*{(6)}

With x~(0)=e0⊗x(0)\tilde{x}(0)=e_0\otimes x(0) and the cycled input (3), the trajectories of (4)–(5) satisfy x~(k)=ei⊗x(k)\tilde{x}(k)=e_i\otimes x(k) and y~(k)=ei⊗y(k)\tilde{y}(k)=e_i\otimes y(k) with i=k mod Ni=k\bmod N [8 Chap. 6]. The cycled system thus reproduces the LPTV system step by step at the original sampling rate: at each time kk, only the ii-th block of each cycled signal is nonzero, and it equals the original signal.

Optimal Dynamic Quantizers for LTI Systems

This subsection reviews the key results of Azuma and Sugie (2008a,b) for LTI plants. Consider the system Σ\Sigma composed of an LTI plant PP and a dynamic quantizer QQ. The plant is

P:x(k+1)=Ax(k)+Bv(k),y(k)=Cx(k),(7)P:\qquad x(k+1)=Ax(k)+Bv(k),\qquad y(k)=Cx(k), \tag*{(7)}

where x∈Rnx \in\mathbb{R}^{n}, v∈Rmv \in\mathbb{R}^{m}, y∈Rpy \in\mathbb{R}^{p}. The dynamic quantizer is of the form

Q:{ξ(k+1)=Aξ(k)+B1u(k)+B2v(k),v(k)=q(Cξ(k)+u(k)),(8)Q:\left\{\begin{aligned} \xi(k+1)&=A\xi(k)+\mathcal{B}_{1}u(k)+\mathcal{B}_{2}v(k),\\ v(k)&=q(C\xi(k)+u(k)), \end{aligned}\right. \tag*{(8)}

where ξ∈Rnξ\xi\in\mathbb{R}^{n_{\xi}} with ξ(0)=0\xi(0)=0 and uu is the continuous-valued input to Σ\Sigma. The maximum output difference is defined as

E(Q):=sup⁡U∈ℓ∞m∥Y(x0,U)−Y(x0,Q(U))∥,(9)E(Q):=\sup_{U\in\ell_{\infty}^{m}}\left\|Y(x_{0},U)-Y(x_{0},Q(U))\right\|, \tag*{(9)}

where Y(x0,U)Y(x_{0},U) and Y(x0,Q(U))Y(x_{0},Q(U)) denote the output sequences of PP with initial state x0x_{0} driven by UU directly and through QQ, respectively. The following lemma summarizes Theorems 1(i) and 2 of Azuma and Sugie [2].

Lemma 1 For the system Σ\Sigma, if m=pm=p and CBCB is nonsingular, then

E(Q)≥∥abs⁡(CB)∥⋅d2(10)E(Q) \geq\|\operatorname{abs}(CB)\|\cdot\frac{d}{2} \tag*{(10)}

for any dynamic quantizer QQ of the form (10). Moreover, an optimal dynamic quantizer is given by

Qc:{ξ(k+1)=Aξ(k)+Bv(k)−Bu(k),v(k)=q(−(CB)−1CAξ(k)+u(k)),(11)Q^{c}:\left\{\begin{aligned} \xi(k+1)&=A\xi(k)+Bv(k)-Bu(k),\\ v(k)&=q\left(-(CB)^{-1}CA\xi(k)+u(k)\right), \end{aligned}\right. \tag*{(11)}

and the minimum value of E(Q)E(Q) is E(Qc)=∥abs⁡(CB)∥⋅d2E(Q^{c})=\|\operatorname{abs}(CB)\|\cdot\frac{d}{2}.

Since B1=−B\mathcal{B}_{1}=-B and B2=B\mathcal{B}_{2}=B in QcQ^{c} (i.e., B1=−B2\mathcal{B}_{1}=-\mathcal{B}_{2}), the state equation reads ξ(k+1)=Aξ(k)+B(v(k)−u(k))\xi(k+1)=A\xi(k)+B(v(k)-u(k)), so that ξ\xi tracks the accumulated quantization effect in the plant state coordinates, and the term −(CB)−1CAξ(k)-(CB)^{-1}CA\xi(k) compensates for the influence of past quantization errors on the next output y(k+1)y(k+1).

The following result (Azuma and Sugie [2]) characterizes the performance of a general dynamic quantizer.

Lemma 2 For the system Σ\Sigma composed of PP in (9) and QQ in (10) with B1=−B2\mathcal{B}_{1}=-\mathcal{B}_{2}, let w(k):=v(k)−(Cξ(k)+u(k))w(k):=v(k)-(C\xi(k)+u(k)) denote the static quantization error satisfying ∥w(k)∥≤d/2\|w(k)\|\leq d/2. Define the augmented system Σa\Sigma_{a} driven by ww:

A:=[ABC0A+B2C],B2:=[BB2],C:=[C0].(12)\mathcal{A}:=\begin{bmatrix} A & BC \\ 0 & A+\mathcal{B}_{2}C \end{bmatrix},\qquad \mathcal{B}_{2}:=\begin{bmatrix} B \\ \mathcal{B}_{2} \end{bmatrix},\qquad \mathcal{C}:=\begin{bmatrix} C & 0 \end{bmatrix}. \tag*{(12)}

If A\mathcal{A} is Schur stable, then

E(Q)=∥∑k=0∞abs⁡(CAkB2)∥⋅d2.(13)E(Q)=\left\|\sum_{k=0}^{\infty}\operatorname{abs}\left(\mathcal{C}\mathcal{A}^{k}\mathcal{B}_{2}\right)\right\|\cdot\frac{d}{2}. \tag*{(13)}

In (14), AA, BB, CC are the plant matrices in (9), and the calligraphic A\mathcal{A}, B2\mathcal{B}_{2}, C\mathcal{C} are the quantizer matrices in (10). The boldface matrices describe the augmented state z:=[(x−xr)⊤,ξ⊤]⊤z:=[(x-x_{r})^{\top},\xi^{\top}]^{\top} via z(k+1)=Az(k)+B2w(k)z(k+1)=\mathcal{A}z(k)+\mathcal{B}_{2}w(k), where xx and xrx_{r} are the plant states driven through QQ and directly by uu, respectively. The (1, 2) block BCBC of A\mathcal{A} is the coupling from the quantizer state ξ\xi to the plant state error.

For the optimal quantizer QcQ^{c} in (13), A=A\mathcal{A}=A, B2=B\mathcal{B}_{2}=B, and C=−(CB)−1CA\mathcal{C}=-(CB)^{-1}CA. Substituting these into (14), the (2, 2) block A+B2CA+\mathcal{B}_{2}\mathcal{C} of A\mathcal{A} becomes F:=A−B(CB)−1CAF:=A-B(CB)^{-1}CA, and CAkB2=CFkB\mathcal{C}\mathcal{A}^{k}\mathcal{B}_{2}=CF^{k}B. Since CF=0CF=0, this vanishes for all k≥1k\geq1, so that only the term CB2=CBC\mathcal{B}_{2}=CB remains in (15), in agreement with Lemma 1. The matrix FF is the state matrix of the one-step-delayed inverse of PP when CBCB is nonsingular (Silverman [28]). Moreover, CF=0CF=0 yields the identity

\begin{bmatrix} I & 0 \\ -C & zI \end{bmatrix} \begin{bmatrix} zI-A & -B \\ C & 0 \end{bmatrix} \begin{bmatrix} I & 0 \\ -(CB)^{-1}CA & I \end{bmatrix} = \begin{bmatrix} zI-F & -B \\ 0 & CB \end{bmatrix}, \tag*{(14)}\end{bmatrix}

whose determinant gives zmdet⁡[zI−A−BC0]=det⁡(zI−F)det⁡(CB)z^{m}\det\begin{bmatrix}zI-A&-B\\ C&0\end{bmatrix}=\det(zI-F)\det(CB). The system pencil of PP thus has full normal rank, and the eigenvalues of FF are the n−mn-m invariant zeros of PP together with mm eigenvalues at the origin. Hence, QcQ^{c} is stable if and only if all invariant zeros of PP lie in the open unit disk, i.e., PP is minimum phase. When PP is nonminimum phase, the linear programming-based approach of Azuma and Sugie [3] and the invariant set analysis approach of Sawada and Shin [27] provide stable suboptimal quantizers with guaranteed performance bounds.

Problem Formulation

The system considered in this paper is illustrated in Figure 1. The system Σ\Sigma consists of a dynamic quantizer QQ cascaded with an LPTV plant PP. The quantizer QQ receives a continuous-valued input u(k)u(k) and produces a discrete-valued output v(k)∈Vmv(k) \in\mathbb{V}^{m}, which drives the plant PP to generate the output y(k)y(k). As a reference, the output yr(k)y_r(k) denotes the response of the same plant PP driven directly by u(k)u(k) without quantization. The design objective is to find QQ that minimizes the worst-case output difference between y(k)y(k) and yr(k)y_r(k) over all admissible inputs.

Figure 1. Block diagram of the system Σ\Sigma (top) and the reference system (bottom). The dynamic quantizer QQ maps the continuous-valued input uu to the discrete-valued signal v∈Vmv \in\mathbb{V}^{m}, which drives the LPTV plant PP. The reference output yry_r is obtained by applying uu directly to PP.

The LPTV plant PP in (1)–(2) is now driven by the discrete-valued control input v∈Vmv \in\mathbb{V}^{m} rather than uu:

P:x(k+1)=Akx(k)+Bkv(k),y(k)=Ckx(k).(15)P:\quad x(k+1)=A_kx(k)+B_kv(k),\quad y(k)=C_kx(k). \tag*{(15)}

The reference system is the same plant driven directly by uu, i.e., xr(k+1)=Akxr(k)+Bku(k)x_r(k+1)=A_kx_r(k)+B_ku(k) and yr(k)=Ckxr(k)y_r(k)=C_kx_r(k), with the common initial state xr(0)=x(0)=x0x_r(0)=x(0)=x_0. The dynamic quantizer for this periodic plant is allowed to be periodically time-varying with the same period NN:

Q:{ξ(k+1)=Akξξ(k)+B1,ku(k)+B2,kv(k),v(k)=q(Ckξξ(k)+u(k)),(16)Q:\quad \begin{cases} \xi(k+1)=A_k^\xi\xi(k)+B_{1,k}u(k)+B_{2,k}v(k),\\ v(k)=q(C_k^\xi\xi(k)+u(k)), \end{cases} \tag*{(16)}

where ξ∈Rnξ\xi\in\mathbb{R}^{n_\xi} is the quantizer state with ξ(0)=0\xi(0)=0, u∈Rmu \in\mathbb{R}^{m} is the continuous-valued input to the system Σ\Sigma, and all coefficient matrices AkξA_k^\xi, B1,kB_{1,k}, B2,kB_{2,k}, CkC_k are NN-periodic. The quantizer state dimension nξn_\xi is a design parameter that may differ from the plant state dimension nn; as shown in Theorem 2, the optimal quantizer has nξ=nn_\xi=n.

Analogously to the LTI case (11), the maximum output difference between the quantized output y(k)y(k) and the reference output yr(k)y_r(k) is defined as

E(Q):=sup⁡U∈ℓ∞m∥Yr(x0,U)−Y(x0,Q(U))∥,(17)E(Q) := \sup_{U\in\ell_\infty^m} \left\|Y_r(x_0,U)-Y(x_0,Q(U))\right\|, \tag*{(17)}

where Yr(x0,U)Y_r(x_0,U) denotes the output sequence of PP driven directly by UU (i.e., yr(k)y_r(k) in the lower part of Figure 1), and Y(x0,Q(U))Y(x_0,Q(U)) denotes the output sequence of PP driven through QQ (i.e., y(k)y(k) in the upper part of Figure 1). Since both outputs start from the same initial state x0x_0, their difference, and hence E(Q)E(Q), does not depend on x0x_0.

The following assumptions are imposed.

Assumption 1 The dimensions of vv and yy are the same (m=pm=p).

Assumption 2 For given quantization interval d∈R+d\in\mathbb{R}_+, V:=dZ\mathbb{V}:=d\mathbb{Z}, and the static quantizer qq is of the nearest-neighbor type satisfying ∣q(μ)−μ∣≤d21m\lvert q(\mu)-\mu\rvert\leq\frac{d}{2}\mathbf{1}_m for all μ∈Rm\mu\in\mathbb{R}^m.

Assumption 3 The matrix Ck+1BkC_{k+1}B_k is nonsingular for every k∈{0,1,…,N−1}k\in\{0,1,\ldots,N-1\}.

Remark 1 Assumption 3 is the periodic counterpart of the condition that CBCB is nonsingular in the LTI case. It requires that the system has relative degree one at every time instant: the input v(k)v(k) directly affects the next output through y(k+1)=Ck+1(Akx(k)+Bkv(k))y(k+1)=C_{k+1}(A_kx(k)+B_kv(k)) via the nonsingular gain Ck+1BkC_{k+1}B_k. When N=1N=1, it reduces to CBCB being nonsingular. The case where this assumption fails is discussed in § 4.

Problem 1 For the system Σ\Sigma composed of the periodic plant PP in (15) and the periodic quantizer QQ in (16), under Assumptions 1–3, find QQ minimizing E(Q)E(Q) in (17) and determine the minimum value.

In § 3, the worst-case output error of a general periodic quantizer is first characterized. An optimal quantizer is then constructed by applying the LTI results of § 2.2 to the cycled plant, and its optimality is verified directly in the original time domain.

Main Results

Performance Analysis of Periodic Dynamic Quantizers

The LTI performance characterization (Lemma 2) is first extended to the LPTV setting. Consider a quantizer QQ of the form (16) with B1,k=−B2,k\mathcal{B}_{1,k}=-\mathcal{B}_{2,k} for all kk, and let w(k):=v(k)−(Ckξ(k)+u(k))w(k):=v(k)-(C_k\xi(k)+u(k)) denote the static quantization error, which satisfies ∥w(k)∥≤d/2\lVert w(k)\rVert\leq d/2. Analogously to (12), define the NN-periodic matrices

Ak:=[AkBkCk0Ak+B2,kCk],B2,k:=[BkB2,k],\mathbf{A}_k:=\begin{bmatrix} A_k & B_kC_k \\ 0 & A_k+B_{2,k}C_k \end{bmatrix},\qquad \mathbf{B}_{2,k}:=\begin{bmatrix} B_k \\ B_{2,k} \end{bmatrix},
Ck:=[Ck0],(18)\mathbf{C}_k:=\begin{bmatrix} C_k & 0 \end{bmatrix}, \tag*{(18)}

with the same typographic convention as in the LTI case. The augmented state z:=[(x−xr)T,ξT]Tz:=[(x-x_r)^\mathsf{T},\xi^\mathsf{T}]^\mathsf{T} obeys z(k+1)=Akz(k)+B2,kw(k)z(k+1)=\mathbf{A}_kz(k)+\mathbf{B}_{2,k}w(k), where xx and xrx_r are the plant states driven through QQ and directly by uu, respectively, and the output error e(k):=yr(k)−y(k)e(k):=y_r(k)-y(k) is given by e(k)=−Ckz(k)e(k)=-\mathbf{C}_kz(k).

Since ξ(0)=0\xi(0)=0 and x(0)=xr(0)=x0x(0)=x_r(0)=x_0, it follows that z(0)=0z(0)=0, and the output error at time τ\tau is expressed as

e(τ)=−∑l=1τCτΨ(τ,τ−l+1)B2,τ−lw(τ−l),(19)e(\tau)=-\sum_{l=1}^{\tau}\mathbf{C}_{\tau}\boldsymbol{\Psi}(\tau,\tau-l+1)\mathbf{B}_{2,\tau-l}w(\tau-l), \tag*{(19)}

where the state transition matrix is defined by

Ψ(j,i):={Aj−1Aj−2⋯Ai,j>i,I,j=i.(20)\boldsymbol{\Psi}(j,i):= \begin{cases} \mathbf{A}_{j-1}\mathbf{A}_{j-2}\cdots\mathbf{A}_i, & j>i,\\ I, & j=i. \end{cases} \tag*{(20)}

The impulse response matrix characterizing the contribution of w(τ−l)w(\tau-l) to e(τ)e(\tau) at output time phase i:=τ mod Ni:=\tau\bmod N is defined as

Hl(i):=CiΨ(i,i−l+1)B2,i−l,l=1,2,…,(21)H_l(i):=\mathbf{C}_i\boldsymbol{\Psi}(i,i-l+1)\mathbf{B}_{2,i-l},\quad l=1,2,\ldots, \tag*{(21)}

where all subscripts of A\mathbf{A} in Ψ(i,i−l+1)\boldsymbol{\Psi}(i,i-l+1) are reduced modulo NN. Note that the input matrix B2,i−l\mathbf{B}_{2,i-l} varies with ll, reflecting the periodically changing disturbance input channel. In the LTI case, (21) reduces to Hl=CAl−1B2H_l=\mathbf{C}\mathbf{A}^{l-1}\mathbf{B}_2.

Theorem 1. For the system Σ\Sigma with a quantizer QQ of the form (16) satisfying B1,k=−B2,k\mathbf{B}_{1,k}=-\mathbf{B}_{2,k} for all kk, suppose the monodromy matrix AN−1AN−2⋯A0\mathbf{A}_{N-1}\mathbf{A}_{N-2}\cdots\mathbf{A}_0 of the augmented periodic system (18) is Schur stable. Then

E(Q)=max⁡i∈{0,1,…,N−1}∥∑l=1∞abs⁡(Hl(i))∥⋅d2.(22)E(Q)=\max_{i\in\{0,1,\ldots,N-1\}}\left\|\sum_{l=1}^{\infty}\operatorname{abs}(H_l(i))\right\|\cdot\frac{d}{2}. \tag*{(22)}

Proof. By periodicity, (19) reads e(τ)=−∑l=1τHl(i)w(τ−l)e(\tau)=-\sum_{l=1}^{\tau}H_l(i)w(\tau-l) with i=τ mod Ni=\tau\bmod N. The Schur stability of the monodromy matrix implies that ∥Hl(i)∥\|H_l(i)\| decays geometrically in ll, so the series in (22) converges. Since ∥w(k)∥≤d/2\|w(k)\|\le d/2, one has ∥e(τ)∥≤∥∑l=1τabs⁡(Hl(i))∥⋅d/2\|e(\tau)\|\le\left\|\sum_{l=1}^{\tau}\operatorname{abs}(H_l(i))\right\|\cdot d/2 for every τ\tau, and hence E(Q)E(Q) does not exceed the right-hand side of (22). Conversely, for any given ξ(k)\xi(k), the input u(k)u(k) can be chosen such that w(k)w(k) is arbitrarily close to any prescribed point of the box [−d/2,d/2]m[-d/2,d/2]^m, because the argument of qq in (16) ranges over Rm\mathbb{R}^m as u(k)u(k) does. Fix a phase ii and a row index jj attaining the maximum absolute row sum of ∑l=1∞abs⁡(Hl(i))\sum_{l=1}^{\infty}\operatorname{abs}(H_l(i)). For τ\tau with τ mod N=i\tau\bmod N=i, choose u(0),…,u(τ−1)u(0),\ldots,u(\tau-1) successively so that each component of w(τ−l)w(\tau-l), l=1,…,τl=1,\ldots,\tau, approaches d/2d/2 in magnitude with the sign opposite to that of the corresponding entry of the jj-th row of Hl(i)H_l(i). The jj-th component of e(τ)e(\tau) then approaches d/2d/2 times the jj-th row sum of ∑l=1τabs⁡(Hl(i))\sum_{l=1}^{\tau}\operatorname{abs}(H_l(i)), which converges to that of the infinite sum as τ→∞\tau\to\infty. Hence E(Q)E(Q) is not smaller than the right-hand side of (22). □\square

Since Ak\mathbf{A}_k is block upper triangular, the hypothesis of Theorem 1 requires that both the plant monodromy matrix AN−1⋯A0\mathbf{A}_{N-1}\cdots\mathbf{A}_0 and the monodromy matrix of Ak+B2,kCk\mathbf{A}_k+\mathbf{B}_{2,k}\mathbf{C}_k be Schur stable. The right-hand side of (22) coincides with the LTI expression (13) evaluated for the cycled augmented system, because the ii-th block-row of CˇAˇl−1Bˇ2\check{\mathbf{C}}\check{\mathbf{A}}^{l-1}\check{\mathbf{B}}_2 contains Hl(i)H_l(i) as its only nonzero block, where Aˇ\check{\mathbf{A}}, Bˇ2\check{\mathbf{B}}_2, Cˇ\check{\mathbf{C}} are constructed from Ak\mathbf{A}_k, B2,k\mathbf{B}_{2,k}, Ck\mathbf{C}_k via (4)–(6).

Optimal Dynamic Quantizer for Periodic Systems

The optimal quantizer is derived by applying Lemma 1 to the cycled plant Pˇ\check{P}, i.e., the LTI system (3)–(5) with the input uˇ\check{u} replaced by vˇ\check{v}, and by extracting the periodic coefficient matrices from the resulting LTI quantizer.

Theorem 2. For the LPTV system Σ\Sigma under Assumptions 1–3, every NN-periodic dynamic quantizer QQ of the form (16) satisfies

E(Q)≥max⁡i∈{0,1,…,N−1}∥abs⁡(Ci+1Bi)∥⋅d2.(23)E(Q)\ge\max_{i\in\{0,1,\ldots,N-1\}}\|\operatorname{abs}(\mathbf{C}_{i+1}\mathbf{B}_i)\|\cdot\frac{d}{2}. \tag*{(23)}

Moreover, an optimal periodically time-varying dynamic quantizer is given by

Q∗:{ξ(k+1)=Akξ(k)+Bkv(k)−Bku(k),v(k)=q(−(Ck+1Bk)−1Ck+1Akξ(k)+u(k)),(24)Q^*: \begin{cases} \xi(k+1)=\mathbf{A}_k\xi(k)+\mathbf{B}_k v(k)-\mathbf{B}_k u(k),\\ v(k)=q\left(-(\mathbf{C}_{k+1}\mathbf{B}_k)^{-1}\mathbf{C}_{k+1}\mathbf{A}_k\xi(k)+u(k)\right), \end{cases} \tag*{(24)}

where ξ∈Rn\xi\in\mathbb{R}^n and ξ(0)=0\xi(0)=0, and the minimum value of E(Q)E(Q) is

E(Q∗)=max⁡i∈{0,1,…,N−1}∥abs⁡(Ci+1Bi)∥⋅d2.(25)E(Q^*)=\max_{i\in\{0,1,\ldots,N-1\}}\|\operatorname{abs}(\mathbf{C}_{i+1}\mathbf{B}_i)\|\cdot\frac{d}{2}. \tag*{(25)}

Proof. Steps 1 and 2 derive Q∗Q^* from the LTI theory through cyclic reformulation; Steps 3 and 4 verify its optimality directly in the original time domain.

Step 1: Optimal LTI quantizer for the cycled plant. Direct computation from (4)–(6) gives

C~B~=Π⋅diag⁡(C1B0,C2B1,…,C0BN−1),(26)\widetilde{C}\widetilde{B}=\Pi\cdot\operatorname{diag}(C_1B_0,C_2B_1,\ldots,C_0B_{N-1}), \tag*{(26)}
C~A~=Π⋅diag⁡(C1A0,C2A1,…,C0AN−1),(27)\widetilde{C}\widetilde{A}=\Pi\cdot\operatorname{diag}(C_1A_0,C_2A_1,\ldots,C_0A_{N-1}), \tag*{(27)}

where Π\Pi is the block permutation matrix mapping block-column jj to block-row j+1j+1. Since Π\Pi is nonsingular and each Ci+1BiC_{i+1}B_i is nonsingular by Assumption 3, C~B~\widetilde{C}\widetilde{B} is nonsingular, and Lemma 1 applied to P~\widetilde{P} gives the optimal LTI quantizer

Q~∗:ξ~(k+1)=A~ξ~(k)+B~v~(k)−B~u~(k),v~(k)=q~(−(C~B~)−1C~A~ξ~(k)+u~(k)),(28)\widetilde{Q}^*:\quad \begin{aligned} \widetilde{\xi}(k+1)&=\widetilde{A}\widetilde{\xi}(k)+\widetilde{B}\widetilde{v}(k)-\widetilde{B}\widetilde{u}(k),\\ \widetilde{v}(k)&=\widetilde{q}\left(-(\widetilde{C}\widetilde{B})^{-1}\widetilde{C}\widetilde{A}\widetilde{\xi}(k)+\widetilde{u}(k)\right), \end{aligned} \tag*{(28)}

where q~:RNm→VNm\widetilde{q}:\mathbb{R}^{Nm}\to\mathcal{V}^{Nm} is the nearest-neighbor quantizer acting componentwise. From (26) and (27),

(C~B~)−1C~A~=diag⁡((C1B0)−1C1A0, (C2B1)−1C2A1, …, (C0BN−1)−1C0AN−1),(29)(\widetilde{C}\widetilde{B})^{-1}\widetilde{C}\widetilde{A} =\operatorname{diag}\left((C_1B_0)^{-1}C_1A_0,\ (C_2B_1)^{-1}C_2A_1,\ \ldots,\ (C_0B_{N-1})^{-1}C_0A_{N-1}\right), \tag*{(29)}

which is block-diagonal with NN blocks of size m×nm\times n.

Step 2: Parameter extraction. The coefficient matrices of Q~∗\widetilde{Q}^* retain the cyclic structure (A~\widetilde{A} and B~\widetilde{B} are block-circulant, and (29) is block-diagonal), so that extracting their blocks gives (24). Indeed, since q~\widetilde{q} acts componentwise and q(0)=0q(0)=0, the output equation of (28) preserves the zero-block pattern of ξ~(k)\widetilde{\xi}(k) and u~(k)\widetilde{u}(k). Hence, as in § 2.1, the cycled input u~(k)=ei⊗u(k)\widetilde{u}(k)=e_i\otimes u(k) with ξ~(0)=0\widetilde{\xi}(0)=0 yields ξ~(k)=ei⊗ξ(k)\widetilde{\xi}(k)=e_i\otimes\xi(k) and v~(k)=ei⊗v(k)\widetilde{v}(k)=e_i\otimes v(k), i=kmod  Ni=k\mod N, where ξ\xi and vv are the state and output of (24). Therefore, Q~∗\widetilde{Q}^* is the cycled representation of the NN-periodic quantizer (24), which has state dimension nn; the quantizer Q∗Q^* of the theorem is thus obtained.

Step 3: Performance of Q∗Q^*. Let xx and xrx_r be the plant states driven through Q∗Q^* and directly by uu, respectively. By the state equation of (24), x−xrx-x_r and ξ\xi obey the same recursion with the same initial value; hence ξ(k)=x(k)−xr(k)\xi(k)=x(k)-x_r(k) and the output error is e(k):=yr(k)−y(k)=−Ckξ(k)e(k):=y_r(k)-y(k)=-C_k\xi(k). With w(k)w(k) denoting the static quantization error, the output equation of (24) gives v(k)−u(k)=w(k)−(Ck+1Bk)−1Ck+1Akξ(k)v(k)-u(k)=w(k)-(C_{k+1}B_k)^{-1}C_{k+1}A_k\xi(k). Then

e(k+1)=−Ck+1(Akξ(k)+Bk(v(k)−u(k)))=−Ck+1Bkw(k).(30)\begin{aligned} e(k+1)&=-C_{k+1}\left(A_k\xi(k)+B_k(v(k)-u(k))\right)\\ &=-C_{k+1}B_kw(k). \tag*{(30)} \end{aligned}

Since e(0)=0e(0)=0 and ∥w(k)∥≤d/2\lVert w(k)\rVert\le d/2, it follows that ∥e(k+1)∥≤∥abs⁡(Ck+1Bk)∥⋅d/2\lVert e(k+1)\rVert\le\lVert\operatorname{abs}(C_{k+1}B_k)\rVert\cdot d/2 for all kk, so that E(Q∗)E(Q^*) does not exceed the right-hand side of (25).

Step 4: Lower bound. The argument of Azuma and Sugie [2] for the first time step is applied at each time phase. Consider an arbitrary QQ of the form (16) and fix i∈{0,1,…,N−1}i\in\{0,1,\ldots,N-1\}. Let u(k)=0u(k)=0 for k<ik<i. Since ξ(0)=0\xi(0)=0 and q(0)=0q(0)=0, it follows that v(k)=0v(k)=0 and ξ(k)=0\xi(k)=0 for k<ik<i, and hence ξ(i)=0\xi(i)=0 and x(i)=xr(i)x(i)=x_r(i). Therefore, v(i)=q(u(i))v(i)=q(u(i)) and

e(i+1)=−Ci+1Bi(q(u(i))−u(i)).(31)e(i+1)=-C_{i+1}B_i\left(q(u(i))-u(i)\right). \tag*{(31)}

As u(i)u(i) ranges over Rm\mathbb{R}^m, the quantization error q(u(i))−u(i)q(u(i))-u(i) comes arbitrarily close to every point of the box [−d/2,d/2]m[-d/2,d/2]^m, so that sup⁡u(i)∥e(i+1)∥≥∥abs⁡(Ci+1Bi)∥⋅d/2\sup_{u(i)}\lVert e(i+1)\rVert\ge\lVert\operatorname{abs}(C_{i+1}B_i)\rVert\cdot d/2. Since E(Q)E(Q) is not smaller than this supremum for every ii, (23) follows. Applying (23) to Q∗Q^* and combining it with Step 3 proves (25); hence Q∗Q^* is optimal. □

Remark 2 (Interpretation) The quantizer Q∗Q^* in (24) has the following structure at each time kk: (i) the state satisfies ξ(k)=x(k)−xr(k)\xi(k)=x(k)-x_r(k), i.e., it tracks the accumulated quantization effect in the plant state coordinates; (ii) the compensation term −(Ck+1Bk)−1Ck+1Akξ(k)-(C_{k+1}B_k)^{-1}C_{k+1}A_k\xi(k) cancels the effect of past errors on y(k+1)y(k+1), so that, by (30), only the current quantization error reaches the next output, through the instantaneous gain Ck+1BkC_{k+1}B_k. In terms of Theorem 1, H1(i)=CiBi−1H_1(i)=C_iB_{i-1} holds irrespective of the quantizer coefficients, whereas Q∗Q^* achieves Hl(i)=0H_l(i)=0 for all l≥2l\ge2, in parallel with CAkB2=0CA^kB_2=0 for k≥1k\ge1 in the LTI case of § 2.2. Only the term H1(i)H_1(i) thus remains in (22), in agreement with (25), and the performance is dictated by the time phase with the largest gain.

Remark 3 (Relation to the LTI result). Theorem 2 is the periodic counterpart of Lemma 1, with the same three statements: a lower bound valid for every quantizer of the considered form, an optimal quantizer, and the minimum value. When N=1N = 1, it reduces to Lemma 1, and (24) reduces to (11). The minimum (25) equals the optimum ∥abs⁡(C˘B˘)∥⋅d/2\lVert\operatorname{abs}(\breve{C}\breve{B})\rVert\cdot d/2 that Lemma 1 gives for the cycled plant P˘\breve{P}, since each block-row of C˘B˘\breve{C}\breve{B} contains exactly one nonzero block. The lower bound, however, does not follow from Lemma 1 applied to P˘\breve{P}, whose worst case is taken over all inputs in ℓ∞m\ell_\infty^m rather than over cycled inputs only; it is therefore established directly in Step 4.

Stability of the Optimal Quantizer

The optimal quantizer Q∗Q^* is not always stable. Substituting the output equation of (24) into the state equation gives the error dynamics

ξ(k+1)=Fkξ(k)+Bkw(k),(32)\xi(k+1)=F_k\xi(k)+B_kw(k), \tag*{(32)}

where

Fk:=Ak−Bk(Ck+1Bk)−1Ck+1Ak,k=0,1,…,N−1,(33)F_k:=A_k-B_k(C_{k+1}B_k)^{-1}C_{k+1}A_k,\qquad k=0,1,\ldots,N-1, \tag*{(33)}

and w(k)w(k) is the static quantization error, which satisfies ∥w(k)∥≤d/2\lVert w(k)\rVert\le d/2. The unforced part of (32) is an NN-periodic linear system with state matrices FkF_k, which is asymptotically stable if and only if the monodromy matrix

Φ:=FN−1FN−2⋯F0(34)\Phi:=F_{N-1}F_{N-2}\cdots F_0 \tag*{(34)}

is Schur stable [8]. Accordingly, Q∗Q^* is said to be stable if Φ\Phi is Schur stable, in which case the quantizer state ξ(k)\xi(k) remains bounded because ww is bounded.

Theorem 3. Under Assumptions 1–3, the optimal quantizer Q∗Q^* in (24) is stable if and only if all invariant zeros of the cycled plant P˘\breve{P} lie in the open unit disk.

Proof. Let F˘:=A˘−B˘(C˘B˘)−1C˘A˘\breve{F}:=\breve{A}-\breve{B}(\breve{C}\breve{B})^{-1}\breve{C}\breve{A}. Since (C˘B˘)−1C˘A˘(\breve{C}\breve{B})^{-1}\breve{C}\breve{A} is block-diagonal with blocks (Ck+1Bk)−1Ck+1Ak(C_{k+1}B_k)^{-1}C_{k+1}A_k by (29), F˘\breve{F} has the block-circulant structure of A˘\breve{A} in (4) with each AkA_k replaced by FkF_k. Consequently, F˘N=diag⁡(Φ0,Φ1,…,ΦN−1)\breve{F}^{N}=\operatorname{diag}(\Phi_0,\Phi_1,\ldots,\Phi_{N-1}) with Φj:=Fj+N−1⋯Fj+1Fj\Phi_j:=F_{j+N-1}\cdots F_{j+1}F_j and Φ0=Φ\Phi_0=\Phi. The matrices Φj\Phi_j are cyclic permutations of the same product and therefore share the same nonzero eigenvalues, so that ρ(F˘)N=ρ(F˘N)=ρ(Φ)\rho(\breve{F})^N=\rho(\breve{F}^N)=\rho(\Phi); that is, Φ\Phi is Schur stable if and only if F˘\breve{F} is. On the other hand, C˘B˘\breve{C}\breve{B} is nonsingular by (26), and applying (14) to the cycled plant P˘\breve{P} shows that the eigenvalues of F˘\breve{F} are the invariant zeros of P˘\breve{P} together with NmNm eigenvalues at the origin. Hence F˘\breve{F} is Schur stable if and only if all invariant zeros of P˘\breve{P} lie in the open unit disk.

:::

Remark 4 (Periodically minimum phase plants). A plant satisfying the zero condition of Theorem 3 is called periodically minimum phase [21]. The invariant zeros of P˘\breve{P} correspond to the periodic invariant zeros of PP [8] and coincide with the transmission zeros of P˘\breve{P} when P˘\breve{P} is reachable and observable.

Remark 5 (Connection to stable inversion). Inversion of LPTV plants was studied by Perdon et al. [25], and van Zundert and Oomen [30] constructed bounded noncausal inverses for nonminimum phase LPTV plants; an explicit periodic realization of the one-step-delayed inverse of PP was obtained via cyclic reformulation by Okajima [21]. The matrix FkF_k and the gain −(Ck+1Bk)−1Ck+1Ak-(C_{k+1}B_k)^{-1}C_{k+1}A_k of the compensation term in (24) coincide with the state and output matrices of this realization, whose stability is characterized by the same zero condition as in Theorem 3. The optimal quantizer thus performs a partial inversion of the plant at each time step: the compensation term applies the one-step-ahead inverse to the accumulated error state ξ\xi to cancel the output contribution of past errors.

Numerical Example

Consider a 3-periodic (N=3N = 3) SISO system with

A0=[01−0.50.8],A1=[0.30.5−0.40.6],(35)A_0 = \begin{bmatrix} 0 & 1 \\ -0.5 & 0.8 \end{bmatrix}, \quad A_1 = \begin{bmatrix} 0.3 & 0.5 \\ -0.4 & 0.6 \end{bmatrix}, \tag*{(35)}
A2=[0.20.4−0.30.5],(36)A_2 = \begin{bmatrix} 0.2 & 0.4 \\ -0.3 & 0.5 \end{bmatrix}, \tag*{(36)}
B0=[01],B1=[0.51],B2=[0.51],(37)B_0 = \begin{bmatrix} 0 \\ 1 \end{bmatrix}, \quad B_1 = \begin{bmatrix} 0.5 \\ 1 \end{bmatrix}, \quad B_2 = \begin{bmatrix} 0.5 \\ 1 \end{bmatrix}, \tag*{(37)}
C0=[10],C1=[0.51],C2=[10.5],(38)C_0 = \begin{bmatrix} 1 & 0 \end{bmatrix}, \quad C_1 = \begin{bmatrix} 0.5 & 1 \end{bmatrix}, \quad C_2 = \begin{bmatrix} 1 & 0.5 \end{bmatrix}, \tag*{(38)}

d=2d = 2.

The matrices AkA_k and BkB_k are those of the examples in Okajima [21]. The plant monodromy matrix A2A1A0A_2A_1A_0 has spectral radius ρ≈0.20\rho\approx0.20, so the plant is stable.

Verification of Assumption 3: C1B0=1C_1B_0 = 1, C2B1=1C_2B_1 = 1, and C0B2=0.5C_0B_2 = 0.5, all nonzero.

Optimal quantizer performance: From (25), E(Q∗)=max⁡{∣C1B0∣,∣C2B1∣,∣C0B2∣}⋅d/2=max⁡{1,1,0.5}⋅1=1E(Q^*) = \max\{|C_1B_0|, |C_2B_1|, |C_0B_2|\} \cdot d/2 = \max\{1, 1, 0.5\} \cdot1 = 1.

Stability analysis: The state matrices (33) are

F0=[010−0.5],F1=[0.250.1−0.5−0.2],F2=[00−0.7−0.3],(39)F_0 = \begin{bmatrix} 0 & 1 \\ 0 & -0.5 \end{bmatrix}, \quad F_1 = \begin{bmatrix} 0.25 & 0.1 \\ -0.5 & -0.2 \end{bmatrix}, \quad F_2 = \begin{bmatrix} 0 & 0 \\ -0.7 & -0.3 \end{bmatrix}, \tag*{(39)}

and the monodromy matrix is

Φ=F2F1F0=[000−0.02].(40)\Phi= F_2F_1F_0 = \begin{bmatrix} 0 & 0 \\ 0 & -0.02 \end{bmatrix}. \tag*{(40)}

The eigenvalues of Φ\Phi are {0,−0.02}\{0, -0.02\}, so Q∗Q^* is stable and, by Theorem 3, the plant is periodically minimum phase. The error cancellation property C1F0=0C_1F_0 = 0, C2F1=0C_2F_1 = 0, and C0F2=0C_0F_2 = 0 can also be verified directly.

Simulation results: The proposed Q∗Q^* is compared with two conventional quantizers: the first-order ΔΣ\Delta\Sigma modulator (see, e.g., Quevedo and Goodwin [26]) v(k)=q(ξ(k)+u(k))v(k) = q(\xi(k)+u(k)), ξ(k+1)=ξ(k)+u(k)−v(k)\xi(k+1) = \xi(k)+u(k)-v(k), a widely used dynamic quantizer that feeds back the accumulated quantization error without using any plant information, and the static quantizer v(k)=q(u(k))v(k) = q(u(k)). The input signal is u(k)=3cos⁡(0.06(k+1))+2sin⁡(0.13(k+1))+sin⁡(0.31(k+1))u(k) = 3\cos(0.06(k+1)) + 2\sin(0.13(k+1)) + \sin(0.31(k+1)) for k=0,1,…,499k = 0, 1, \ldots, 499. Figure 2 shows u(k)u(k) and the quantized input v(k)v(k) produced by Q∗Q^* and by the ΔΣ\Delta\Sigma modulator.

Input $u(k)$ and quantized input $v(k)$ for the numerical example ($d = 2$)

Figure 2. Input u(k)u(k) and quantized input v(k)v(k) for the numerical example (d=2d = 2). Top: optimal periodic Q∗Q^*. Bottom: ΔΣ\Delta\Sigma modulator.

Figure 3 shows the output error e(k)=yr(k)−y(k)e(k) = y_r(k)-y(k) for the three quantizers. The simulation is consistent with Theorem 2: max⁡k∣e(k)∣=0.997\max_k |e(k)| = 0.997 for Q∗Q^*, which is below E(Q∗)=1E(Q^*) = 1 (dashed lines). In contrast, the ΔΣ\Delta\Sigma modulator yields max⁡k∣e(k)∣=1.889\max_k |e(k)| = 1.889, and the static quantizer yields max⁡k∣e(k)∣=2.319\max_k |e(k)| = 2.319.

Output errors for the numerical example

Figure 3. Output errors for the numerical example (d=2d = 2). Top: optimal periodic Q∗Q^* (max∣e∣|e| = 0.997, bound = 1.0). Middle: ΔΣ\Delta\Sigma modulator (max∣e∣|e| = 1.889). Bottom: static quantizer (max∣e∣|e| = 2.319). Dashed lines: ±E(Q∗)\pm E(Q^*).

Figure 4 compares the plant output under the reference input u(k)u(k) (no quantization, yielding yr(k)y_r(k)), the optimal periodic quantizer Q∗Q^*, and the static quantizer. The output under Q∗Q^* closely tracks the reference yry_r, while the static quantizer exhibits visible deviation.

Output comparison for the numerical example

Figure 4. Output comparison for the numerical example (d=2d = 2). The output under Q∗Q^{*} (red dashed) closely follows the reference yry_r (blue solid), while the static quantizer (green dashed) shows noticeable deviation.

The maximum errors observed for a single input only bound E(⋅)E(\cdot) from below. The worst-case errors of the two conventional quantizers are therefore evaluated by Theorem 1, which applies to both: they are of the form (16) with B1,k=−B2,kB_{1,k} = -B_{2,k} (the ΔΣ\Delta\Sigma modulator with Ak=1A_k = 1, B1,k=1B_{1,k} = 1, B2,k=−1B_{2,k} = -1, Ck=1C_k = 1, and the static quantizer with all coefficients zero), and the corresponding augmented monodromy matrices are Schur stable because the plant is stable. Table 1 lists the resulting values of E(⋅)E(\cdot) together with the observed maximum errors, for d=2d = 2 and d=0.5d = 0.5; all values of E(⋅)E(\cdot) scale linearly with dd. In terms of the worst-case error, Q∗Q^* achieves a reduction of about 65% relative to both conventional quantizers, whereas the reduction observed in the simulation relative to the ΔΣ\Delta\Sigma modulator (47–50%) is specific to the input used. For this plant, the ΔΣ\Delta\Sigma modulator, which uses no plant information, offers almost no worst-case improvement over the static quantizer; the improvement achieved by Q∗Q^* stems from incorporating the plant dynamics into the quantizer.

Q∗Q^*QΔΣQ_{\Delta\Sigma}Static
d=2.0d = 2.0E(⋅)E(\cdot)1.0002.8702.885
Observed0.9971.8892.319
d=0.5d = 0.5E(⋅)E(\cdot)0.2500.7170.721
Observed0.2480.4960.556

Table 1. Worst-case output error E(⋅)E(\cdot) computed from (22) and maximum error observed in the simulation for the numerical example in § 3.4.

Discussion

The role of cyclic reformulation in the above results is twofold. First, it is constructive: the optimal LTI quantizer of the cycled plant retains the cyclic block structure, so that the periodic quantizer Q∗Q^* with the original state dimension nn is read off block by block (Steps 1 and 2 of the proof of Theorem 2); no new optimization machinery is needed, and the optimality of Q∗Q^* is then confirmed by an elementary per-step argument. Second, it links the stability of Q∗Q^* to a time-invariant object, namely the invariant zeros of Pˇ\check{P} (Theorem 3). Lifting (Chen and Francis [10]), by contrast, stacks the inputs and outputs over one period and evolves on the NN-step block time scale, so that the per-step causal structure required of a quantizer operating at the original sampling rate must be imposed separately.

The performance (25) shows that the achievable accuracy is dictated by the largest instantaneous gain max⁡i∥abs⁡(Ci+1Bi)∥\max_i \lVert\operatorname{abs}(C_{i+1}B_i)\rVert: a single time phase with large input-to-output coupling dominates the performance even if the gains at the other phases are small. Where the periodic structure is a design choice, it should therefore be chosen to reduce the largest of these gains. In addition, stability couples all NN time phases through the product Φ=FN−1⋯F0\Phi= F_{N-1}\cdots F_0 (§ 3.3). Implementing Q∗Q^* requires only a few matrix–vector products of dimension nn and one static quantization per time step, together with the storage of the $N sets of coefficient matrices.

The present results require Ck+1BkC_{k+1}B_k to be nonsingular at every time phase (Assumption 3) and, for stability, a periodically minimum phase plant (Theorem 3). If the plant instead has uniform periodic relative degree r≥2r \ge2, i.e., its first nonzero periodic Markov parameter is of order rr and nonsingular at every time phase [21], the first nonzero Markov parameter of the cycled plant is CˇAˇr−1Bˇ\check{C}\check{A}^{r-1}\check{B}, which is nonsingular and has the periodic Markov parameters as its only nonzero blocks; the closed-form design is then expected to carry over with Ck+1BkC_{k+1}B_k replaced by these parameters, which is left for future work. When the relative degree is not uniform, as in multirate input–output patterns where Ck+1Bk=0C_{k+1}B_k = 0 at some phases but not at others, CˇBˇ\check{C}\check{B} is singular without being zero, and no closed-form optimal quantizer is currently available. For nonminimum phase plants, the linear programming approach [3] and the LMI-based invariant set approach [27] can be applied to the LTI cycled plant Pˇ\check{P} to obtain stable suboptimal quantizers. However, the resulting quantizer has state dimension NnNn in general and is not guaranteed to retain the cyclic structure from which an NN-periodic quantizer can be recovered, so that this structure has to be imposed in the synthesis.

Conclusion

This paper has presented an optimal dynamic quantizer for square LPTV plants having relative degree one at every time phase. The quantizer Q∗Q^* is obtained in closed form by applying the LTI theory to the cycled plant and reading off the periodic coefficient matrices block by block, and it is shown directly to attain the minimum worst-case output error, which is determined by the largest instantaneous gain max⁡i∥abs⁡(Ci+1Bi)∥\max_i \lVert\operatorname{abs}(C_{i+1}B_i)\rVert. The worst-case error of general periodic dynamic quantizers has also been characterized and used to quantify the advantage of Q∗Q^{*} over conventional quantizers in a numerical example. The quantizer Q∗Q^{*} embeds the one-step-delayed inverse of the plant, and it is stable if and only if the plant is periodically minimum phase, which is the condition for stable causal inversion of LPTV plants [21].

Future work includes the extension to feedback configurations, as done for LTI plants by Minami et al. [17], to plants with uniform periodic relative degree greater than one, and to plants with non-uniform relative degree, such as multirate systems in which Assumption 3 fails at some time phases.

Declaration of generative AI and AI-assisted technologies in the manuscript preparation process

During the preparation of this work, the author used Claude (Anthropic) in order to refine English expressions and improve readability throughout the manuscript. After using this tool, the author reviewed and edited the content as needed and takes full responsibility for the content of the published article.

Funding

This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.

References

  1. [1]Almakhles, D., Swain, A. K., Nasiri, A., & Patel, N. (2017). An adaptive two-level quantizer for networked control systems. IEEE Transactions on Control Systems Technology, 25(3), 1084–1091.DOI
  2. [2]Azuma, S., & Sugie, T. (2008a). Optimal dynamic quantizers for discrete-valued input control. Automatica, 44(2), 396–406.DOI
  3. [3]Azuma, S., & Sugie, T. (2008b). Synthesis of optimal dynamic quantizers for discrete-valued input control. IEEE Transactions on Automatic Control, 53(9), 2064–2075.DOI
  4. [4]Azuma, S., & Sugie, T. (2008c). Stability analysis of quantized feedback systems including optimal dynamic quantizers. In Proc. 47th IEEE Conference on Decision and Control (pp. 3392–3397).DOI
  5. [5]Azuma, S., & Sugie, T. (2012). Dynamic quantization of nonlinear control systems. IEEE Transactions on Automatic Control, 57(4), 875–888.DOI
  6. [6]Bai, X., Li, G., Wang, Z., Zhao, Z., & Dong, H. (2025). Recursive unscented Kalman filtering for power distribution networks under hybrid attacks: Tackling dynamic quantization effects. IEEE Internet of Things Journal, 12(22), 48993–49003.DOI
  7. [7]Bittanti, S., & Colaneri, P. (2000). Invariant representations of discrete-time periodic systems. Automatica, 36(12), 1777–1793.DOI
  8. [8]Bittanti, S., & Colaneri, P. (2009). Periodic Systems: Filtering and Control. London: Springer-Verlag.DOI
  9. [9]Brockett, R. W., & Liberzon, D. (2000). Quantized feedback stabilization of linear systems. IEEE Transactions on Automatic Control, 45(7), 1279–1289.DOI
  10. [10]Chen, T., & Francis, B. A. (1995). Optimal Sampled-Data Control Systems. London: Springer-Verlag.DOI
  11. [11]Fujimoto, Y., & Minami, Y. (2026). Data-driven design of dynamic quantizers applicable to nonminimum phase systems. IFAC Journal of Systems and Control, 35, 100369.DOI
  12. [12]Hosoe, Y., Miyamoto, M., & Hagiwara, T. (2019). Cycling-based synthesis of robust output estimators for uncertain LPTV systems. SICE Journal of Control, Measurement, and System Integration, 12(2), 39–46.DOI
  13. [13]Ikeda, T., Nagahara, M., & Ono, S. (2017). Discrete-valued control of linear time-invariant systems by sum-of-absolute-values optimization. IEEE Transactions on Automatic Control, 62(6), 2750–2763.DOI
  14. [15]Liu, T., Shen, D., & Wang, J. (2025). Adaptive quantized iterative learning control using encoding-decoding strategy. IEEE Transactions on Cybernetics, 55(3), 1359–1371.DOI
  15. [16]Meng, X., Wang, Z., Wang, F., & Chen, Y. (2023). Distributed fusion filtering for nonlinear time-varying systems over amplify-and-forward relay networks: An H∞ quantized framework. IEEE Transactions on Network Science and Engineering, 10(6), 3597–3608.DOI
  16. [17]Minami, Y., Azuma, S., & Sugie, T. (2007). An optimal dynamic quantizer for feedback control with discrete-valued signal constraints. In Proc. 46th IEEE Conference on Decision and Control (pp. 2259–2264).DOI
  17. [18]Minami, Y., Azuma, S., & Sugie, T. (2009). Optimal decentralized dynamic quantizers for discrete-valued input control: A closed form solution and experimental evaluation. In Proc. 2009 American Control Conference (pp. 4367–4372).DOI
  18. [19]Morita, R., Azuma, S., & Sugie, T. (2011). Dynamic quantization of multi-rate control systems. Transactions of the Institute of Systems, Control and Information Engineers, 24(5), 109–118 (in Japanese).DOI
  19. [20]Ohno, S., & Tariq, M. R. (2017). Optimization of noise shaping filter for quantizer with error feedback. IEEE Transactions on Circuits and Systems I: Regular Papers, 64(4), 918–930.DOI
  20. [21]Okajima, H. (2026). Stable inversion of discrete-time linear periodically time-varying systems via cyclic reformulation. Automatica, in press. arXiv:2603.23147.arxiv.org/abs/2603.23147
  21. [22]Okajima, H., Fujimoto, Y., Oku, H., & Kondo, H. (2025). Cyclic reformulation based system identification for periodically time-varying systems. IEEE Access, 13, 26483–26493.DOI
  22. [23]Okajima, H., Hosoe, Y., Hagiwara, T., & Minami, Y. (2019). Basic idea of periodically time-varying dynamic quantizer in networked control systems. In Proc. 2019 SICE Annual Conference (pp. 883–889).DOI
  23. [24]Okajima, H., Sawada, K., & Matsunaga, N. (2016). Dynamic quantizer design under communication rate constraints. IEEE Transactions on Automatic Control, 61(10), 3190–3196.DOI
  24. [25]Perdon, A. M., Conte, G., & Longhi, S. (1992). Invertibility and inversion of linear periodic systems. Automatica, 28(3), 645–648.
  25. [26]Quevedo, D. E., & Goodwin, G. C. (2005). Multistep optimal analog-to-digital conversion. IEEE Transactions on Circuits and Systems I: Regular Papers, 52(3), 503–515.DOI
  26. [27]Sawada, K., & Shin, S. (2011). Synthesis of dynamic quantizers for quantized feedback systems within invariant set analysis framework. In Proc. 2011 American Control Conference (pp. 1662–1667).
  27. [28]Silverman, L. M. (1969). Inversion of multivariable linear systems. IEEE Transactions on Automatic Control, 14(3), 270–276.
  28. [29]Teranishi, K., & Kogiso, K. (2021). ElGamal-type encryption for optimal dynamic quantizer in encrypted control systems. SICE Journal of Control, Measurement, and System Integration, 14(1), 59–66.DOI
  29. [30]van Zundert, J., & Oomen, T. (2019). Stable inversion of LPTV systems with application in position-dependent and non-equidistantly sampled systems. International Journal of Control, 92(5), 1022–1032.

Paper details

vlnc:W_uw2w86pv
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

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

  • Mathematics

Claude (Anthropic)

View statement
Looking for related papers…
Contents