OctoInformBeta

OpenAI preprint, September 2026 · §1. Introduction

§1 (setting): exact dimensionality and the upper bound used in the proof: Lower Hausdorff dimension and the entropy-rate upper bound accompany the closed-ball limit

Self-similar measures have constant local dimension.

Hover over a blue or green term in the text, or a node in the graph, to find it in the other. Select it to read its declaration both in Lean and in words.

The Setting

Local dimension concerns the decay of the mass of a small ball around a typical point. A weighted affine iterated function system SS on a finite alphabet 𝒜\mathcal A consists of similarities φi\varphi_i and probabilities pip_i satisfying (1): φi(x)=rix+ti,ri,ti∈ℝ,0<|ri|<1,pi>0,∑i∈𝒜pi=1.\varphi_i(x)=r_i x+t_i,\qquad r_i,t_i\in\mathbb R,\qquad 0<|r_i|<1,\qquad p_i>0,\qquad \sum_{i\in\mathcal A}p_i=1.(1)✦ The weighted component ηi\eta_i of a measure η\eta on ℝ\mathbb R is the measure on ℝ\mathbb R given by (2). A measure μ\mu on ℝ\mathbb R is self-similar for SS when it satisfies (3). ηi=pi(φi)*η.\eta_i=p_i(\varphi_i)_*\eta.(2)✦ μ=∑i∈𝒜μi.\mu=\sum_{i\in\mathcal A}\mu_i.(3)✦

The lower Hausdorff dimension dim⁡Hμ\dim_H\mu of μ\mu is the extended nonnegative number defined by positive-mass Borel sets in (4), where dim⁡HE\dim_H E is the Hausdorff dimension of EE: dim⁡Hμ=inf⁡{dim⁡HE:E⊆ℝ is Borel,μ(E)>0}.\dim_H\mu =\inf\{\dim_H E:E\subseteq\mathbb R\text{ is Borel},\ \mu(E)>0\}.(4)✦ For a real number aa, write a+=max⁡{a,0}a_+=\max\{a,0\}.

The Map Walk and the Coding Exponent

A word w=i0⋯in−1w=i_0\cdots i_{n-1} of length nn over 𝒜\mathcal A has probability pwp_w and complete affine map gwg_w as defined in (5); composition is in address order, with φ⌀\varphi_{\varnothing} the identity. pw=∏j=0n−1pij,φw=φi0∘⋯∘φin−1,gw=(rw,tw),φw(x)=rwx+tw.p_w=\prod_{j=0}^{n-1}p_{i_j},\qquad \varphi_w=\varphi_{i_0}\circ\cdots\circ\varphi_{i_{n-1}},\qquad g_w=(r_w,t_w),\quad \varphi_w(x)=r_wx+t_w.(5) The set GnG_n of attainable affine maps for words of length nn and its probability mass function qnq_n are given by (6), with equality of maps requiring equality of both the signed slope and the translation. Gn={gw:w∈𝒜n},qn(g)=∑w∈𝒜ngw=gpw.G_n=\{g_w:w\in\mathcal A^n\},\qquad q_n(g)=\sum_{\substack{w\in\mathcal A^n\\g_w=g}}p_w.(6) The walk entropy HnH_n is the base-two entropy in (7), using 0log⁡20=00\log_2 0=0. Hn=−∑g∈Gnqn(g)log⁡2qn(g).H_n=-\sum_{g\in G_n}q_n(g)\log_2 q_n(g).(7)✦ The map-walk entropy rate hh and the positive Lyapunov exponent χ\chi of SS are the real numbers given by (8). h=infn≥1Hnn,χ=−∑i∈𝒜pilog⁡2|ri|.h=\inf_{n\geq1}\frac{H_n}{n},\qquad \chi=-\sum_{i\in\mathcal A}p_i\log_2|r_i|.(8)

On 𝒜\mathcal A with its discrete σ\sigma-algebra, the symbol probability measure β\beta and the Bernoulli address measure ℙ\mathbb P on 𝒜ℕ\mathcal A^{\mathbb N} are defined by (9), where δi\delta_i denotes the unit point mass at ii. β=∑i∈𝒜piδi,ℙ=⨂j=0∞β.\beta=\sum_{i\in\mathcal A}p_i\delta_i,\qquad \mathbb P=\bigotimes_{j=0}^{\infty}\beta.(9) For an address ω=(ω0,ω1,…)\omega=(\omega_0,\omega_1,\ldots), its prefix contraction RnR_n and scaled translation term TnT_n are the functions in (10). Rn(ω)=∏j=0n−1rωj,Tn(ω)=Rn(ω)tωn.R_n(\omega)=\prod_{j=0}^{n-1}r_{\omega_j},\qquad T_n(\omega)=R_n(\omega)t_{\omega_n}.(10) The coding map π:𝒜ℕ→ℝ\pi:\mathcal A^{\mathbb N}\to\mathbb R and the coding probability measure λ\lambda on ℝ\mathbb R are defined by (11). π(ω)=∑n=0∞Tn(ω),λ=π*ℙ.\pi(\omega)=\sum_{n=0}^{\infty}T_n(\omega),\qquad \lambda=\pi_*\mathbb P.(11)

For measures η,ζ\eta,\zeta on ℝ\mathbb R, the maximal ball mass ratio Mη,ζM_{\eta,\zeta} is the extended nonnegative function on ℝ\mathbb R given by (12), using the set ℛ\mathcal R of rational radii specified there; B¯(x,q)={y∈ℝ:|y−x|≤q}\overline B(x,q)=\{y\in\mathbb R:|y-x|\leq q\} denotes a closed ball. ℛ=ℚ∩(0,1],Mη,ζ(x)=supq∈ℛη(B¯(x,q))ζ(B¯(x,q)).\mathcal R=\mathbb Q\cap(0,1],\qquad M_{\eta,\zeta}(x)=\sup_{q\in\mathcal R} \frac{\eta(\overline B(x,q))}{\zeta(\overline B(x,q))}.(12) The real-valued function LL on [0,∞][0,\infty] is given by L(a)=ln⁡aL(a)=\ln a when 0<a<∞0<a<\infty and by L(0)=L(∞)=0L(0)=L(\infty)=0. For measures η,ζ\eta,\zeta on ℝ\mathbb R, the positive logarithmic maximal envelope Eη,ζE_{\eta,\zeta} is the real-valued function on ℝ\mathbb R given by (13). Eη,ζ(x)=max⁡{0,L(Mη,ζ(x))}.E_{\eta,\zeta}(x)=\max\{0,L(M_{\eta,\zeta}(x))\}.(13)✦ The symbolic envelope AA and logarithmic Radon-Nikodym correction CC are the functions on addresses defined in (14), using the weighted components λi\lambda_i of (2). A(ω)=Eλ,λω0(π(ω)),C(ω)=L(dλdλω0(π(ω))).A(\omega)=E_{\lambda,\lambda_{\omega_0}}(\pi(\omega)),\qquad C(\omega)=L\!\left( \frac{d\lambda}{d\lambda_{\omega_0}}(\pi(\omega))\right).(14) KK is the clipped logarithmic density correction and WW the symbol information observable, both real-valued functions on 𝒜ℕ\mathcal A^{\mathbb N} defined by (15). K(ω)=max⁡{0,min⁡{A(ω),C(ω)}},W(ω)=−ln⁡pω0.K(\omega)=\max\{0,\min\{A(\omega),C(\omega)\}\},\qquad W(\omega)=-\ln p_{\omega_0}.(15) With natural logarithms, (16) defines the projection information rate ℐ\mathcal I and coding dimension exponent dSd_S of SS. ℐ=∫Wdℙ−∫Kdℙ,dS=ℐχln⁡2.\mathcal I=\int W\,d\mathbb P-\int K\,d\mathbb P,\qquad d_S=\frac{\mathcal I}{\chi\ln2}.(16)

The Claim

Theorem 1. Let SS be a weighted affine iterated function system satisfying (1), and let μ\mu be a self-similar probability measure as in (3). For μ\mu-almost every xx, the radial limit (17) holds: limr↓0−L(μ(B¯(x,r)))−ln⁡r=dS.\lim_{r\downarrow0} \frac{-L(\mu(\overline B(x,r)))}{-\ln r}=d_S.(17)✦ Its lower Hausdorff dimension satisfies the identity (18) and bound (19): dim⁡Hμ=(dS)+.\dim_H\mu=(d_S)_+.(18)✦ dim⁡Hμ≤(min{1,hχ})+.\dim_H\mu\leq\left(\min\left\{1,\frac h\chi\right\}\right)_+.(19)✦

In the paper, the result reads: there is a constant d=dim⁡Hμd=\dim_H\mu such that log⁡2μ(B(x,s))/log⁡2s→d\log_2\mu(B(x,s))/\log_2s\to d for μ\mu-almost every xx as s↓0s\downarrow0, with d≤min⁡{1,h/χ}d\leq\min\{1,h/\chi\}, where B(x,s)B(x,s) denotes the open ball of radius ss about xx.

What Is Checked

The statements are over a finite weighted affine system SS, its coding exponent dSd_S, and any self-similar probability measure μ\mu on ℝ\mathbb R.