Stabilisability and beta representations

Carl P. Dettmann, University of Bristol, UK · 2025

Read Stabilisability and beta representations by Carl P. Dettmann, University of Bristol, UK (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2509-08532.txt.

Introduction

Writing numbers with the smallest possible average digit is subtler than the greedy recipe suggests, shows Carl P. Dettmann in a solo 2025 dynamical systems paper. Beta representations expand numbers in a non-integer base with unrestricted digits, and minimising the average digit is the goal. The problem is equivalent to a one-dimensional switched system obtained from a stabilisable two-dimensional linear system.

For a countable set of bases, the usual greedy expansion is optimal. On an interval of bases, strictly smaller averages are achievable, beating greed. A numerical upper bound and a conditional lower bound cover all bases. Number representations underpin coding and dynamics alike, and this result maps exactly where the obvious algorithm stops being best.

Catalog record

CREATORCarl P. Dettmann, University of Bristol, UK
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE28 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2509-08532

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Stabilisability and beta representations Carl P. Dettmann, University of Bristol, UK Abstract We consider a one dimensional affine switched system obtained from a formal limit of a two dimensional linear system. We show this is equivalent to minimising the average digit in beta representations with unrestricted digits. We give a countable set of β \beta for which the result is given by the usual (greedy) beta expansion, an interval of values for which it is strictly less, and a conditional lower bound for all β \beta . 1 Introduction Joint spectral properties of sets of matrices are notoriously challenging to characterise. For example, for the joint spectral radius \rho_{\infty} [ Rot60 ] it is known that the question \rho_{\infty}\leq 1 is undecidable [ Blo00 ] . Here, we consider the stabilisability radius \tilde{\rho} [ Jun17 ] . Given a discrete time switched linear system x(k+1)=A_{\sigma_{k}}x(k),\qquad x(0)=x_{0} (1) defined by a set of matrices \mathcal{M}=\{A_{\sigma}|\sigma\in\Sigma\}\subset\mathbb{R}^{n\times n} , \tilde{\rho} characterises the ability, using a knowledge of the state of the system, to stabilise it (that is, send x(k) to zero as quickly as possible) by controlling the switching signal \{\sigma_{k}\} : \displaystyle\tilde{\rho}(\mathcal{M}) \displaystyle=\sup_{x\in\mathbb{R}^{n}\setminus\{0\}}\tilde{\rho}_{x}(\mathcal{M}) (2) \displaystyle\tilde{\rho}_{x}(\mathcal{M}) \displaystyle=\inf\{\lambda\geq 0\;|\;\exists M>0,\{\sigma_{k}\}\;s.t.\;\|x(k)\|\leq M\lambda^{k}\|x\|\;\forall k\in\mathbb{N}\} where x(k) is determined by Eq. ( 1 ) with x(0)=x . Many engineering control problems (mechanical, power, traffic, etc.) involve control of multiple variables with a selection of actions [ Zha16 ] . This is also the case for treatment regimes of viruses and cancer, eliminating the pathogens in the presence of mutation, drug resistance and drug toxicity [ And21 ] . Switched systems have also been used to understand strategies of microbes that switch between dormant and active states [ Bla21 ] . Even a switched linear system defined by two general 2\times 2 matrices is poorly understood (see Section 2 below), so here we simplify still further, and consider the one dimensional affine system \displaystyle u(k+1) \displaystyle=a_{\sigma_{k}}(u(k)) (3) \displaystyle a_{\sigma}(u) \displaystyle=\left\{\begin{array}[]{cc}u-1&\sigma=1\\ \beta u&\sigma=2\end{array}\right. where u\in\mathbb{R} , \beta\in\mathbb{R}_{>1} and we choose \sigma_{k} to minimise the proportion of a_{1} transformations while keeping u bounded. Clearly this is possible only if u(0)\geq 0 and \sigma_{k}=2 if u(k)<1 . In Section 2 we obtain Eq. ( 3 ) by taking an appropriate limit of a previously studied system of two 2\times 2 matrices. The derivation is non-rigorous, however it motivates the above system as relevant to known problems and containing at least some of their essential features. The switched affine system is also related to beta representations, where u(0) is written as a sum of inverse powers of \beta with integer coefficients (“digits”) d_{j}\in\mathbb{Z}_{\geq 0} . If \beta=10 we have the usual decimal representation of real numbers, and Rényi introduced similar systems for non-integer base \beta [ Rén57 ] , which have been much studied more recently; see for example Ref. [ Gho25 , Tak24 ] and references therein. Usually there is a requirement that d_{j}<\beta , but here we do not place a bound on the digits. Section 3 shows how the affine switched system corresponds to minimising average digits in beta representations. The supremum over u(0) of the minimum average digit is denoted by \displaystyle\bar{d}(\beta) \displaystyle=\sup_{u\in\mathbb{R}_{+}}\bar{d}(\beta,u) (6) \displaystyle\bar{d}(\beta,u) \displaystyle=\inf_{\{d_{j}\}}\{\limsup_{k\to\infty}\frac{1}{k}\sum_{j=0}^{k-1}d_{j}:u=\sum_{j=0}^{\infty}\frac{d_{j}}{\beta^{j}}\} If we always choose \sigma_{k}=1 where possible (that is, when u(k)\geq 1 ), this corresponds to always choosing the largest digit, and gives what is known as the greedy beta representation, or beta expansion. We denote the average digit in this case by \displaystyle\bar{d}^{(\beta E)}(\beta) \displaystyle=\sup_{u\in\mathbb{R}_{+}}\bar{d}^{(\beta E)}(\beta,u) (7) \displaystyle\bar{d}^{(\beta E)}(\beta,u) \displaystyle=\limsup_{k\to\infty}\frac{1}{k}\sum_{j=0}^{k-1}d_{j}^{(\beta E)}(u) where d_{j}^{(\beta E)}(\beta,u) is j th digit of the beta expansion of u and evidently \bar{d}^{(\beta E)}(\beta)\geq\bar{d}(\beta) for all \beta . Section 4 presents three theorems concerning this problem. Theorem 1 gives a countable set of \beta for which \bar{d}^{(\beta E)}(\beta)=\bar{d}(\beta) . Theorem 2 gives a non-trivial upper bound on \bar{d}(\beta) for 2<\beta<2.28879 showing that \bar{d}(\beta)<\bar{d}^{(\beta E)}(\beta) here. Theorem 3 gives a conditional lower bound on \bar{d}(\beta) for all \beta\in(1,\infty) . These theorems together with a numerical upper bound are illustrated in Figure 1 . The proofs of these theorems are given in the remaining sections. Figure 1: Upper curve: \bar{d}^{(\beta E)}(\beta) . Black dots: Elements of \{\rho\}\cup MB where \bar{d}^{(\beta E)}(\beta)=\bar{d}(\beta) from Theorem 1. The small section of curve for 2<\beta<2.289 gives the upper bound for \bar{d}(\beta) from Theorem 2. The nonmonotonic yellow curve gives the numerical upper bound, to k=12 . The lower curve is the conditional lower bound for \bar{d}(\beta) from Theorem 3. 2 Simplifying switched linear systems Consider the switched linear system, Eq. ( 1 ) with two 2\times 2 matrices as follows \mathcal{M}(\theta,c,\beta)=\left\{A_{1}=\left(\begin{matrix}\cos\theta&\sin\theta\\ -\sin\theta&\cos\theta\end{matrix}\right),\qquad A_{2}=\left(\begin{matrix}c&0\\ 0&\beta c\end{matrix}\right)\right\} (8) where \beta\geq c^{-1}>1 . Without knowledge of the initial condition x(0) , it is not possible to stabilise the system, since any product of these matrices has determinant at least equal to 1. However, given the initial condition, and for almost all values of \theta , the matrix A_{1} can be used to rotate arbitrarily close to the x_{1} -axis, after which the matrix A_{2} reduces \|x\| . The case \mathcal{M}(\pi/6,1/2,4) was proposed in Ref. [ Sta79 ] and also discussed in Ref. [ Sta94 ] . In Ref. [ Det20 ] , \mathcal{M}(\theta,c,c^{-2}) was considered, especially the special case \mathcal{M}(\pi/4,1/2,4) . For the latter example, the best bounds given were 0.707\approx\frac{1}{\sqrt{2}}\leq\tilde{\rho}(\mathcal{M}(\pi/4,1/2,4))<0.9^{1/4}\approx 0.974 (9) This is a large gap as the system is still poorly understood. In this paper, we seek to simplify the problem further, whilst retaining its essential properties. Assume that \theta=O(\epsilon) and u=\arctan(x_{2}/x_{1})/\theta=O(1) for some small \epsilon>0 . Then to linear order in \epsilon , the actions of the matrices A_{1} and A_{2} become Eq. ( 3 ). Also, the optimal stabilisation (choice of switching signal to make x(k)\to 0 as quickly as possible) now corresponds to minimising the proportion of a_{1} transformations. This simplification neglects two effects of the original system, firstly the nonlinear terms, of which the leading term is O(\epsilon^{3}) with a negative coefficient in the a_{2} equation of Eq. ( 3 ). Secondly, whilst it can be assumed that small rotations can eventually direct the system near the stable x_{1} -axis so that \arctan(x_{2}/x_{1}) is small as desired, it is also possible that a large number of small rotations could, if necessary, rotate around to a more favourable configuration, for example, exactly onto the x_{1} -axis. However for very small \epsilon this would be a huge penalty, so the set of initial conditions where it improves the stabilisation should be vanishingly small. In this paper, we consider Eq. ( 3 ) as an interesting system in its own right. 3 Beta representations A beta representation expresses a non-negative real number to base \beta\in\mathbb{R}_{>1} , u=\sum_{j=0}^{\infty}\frac{d_{j}}{\beta^{j}} (10) where d_{j}\in\mathbb{Z}_{\geq 0} . Note that conventionally, d_{j}<\beta , and either j\geq 1 for u\in[0,1] or j>-\infty for u\in\mathbb{R}_{\geq 0} . There is a 1:1 correspondence between the \{d_{j}\} of the beta representation and the \{\sigma_{k}\} of the system ( 3 ). From the beta representation, apply a_{1} if d_{0}>0 ; this reduces d_{0} by 1. Otherwise apply a_{2} ; this performs a shift on the d_{j} . From the switched system, define the digit d_{j} to be the number of sequential a_{1} transformations following the j th a_{2} transformation. Assuming the simplifications in

Read the complete text