OctoInformBeta

OpenAI preprint, September 2026 · §2. Entropy between two scales

Lemma 2.3: For self-similar measures, with logarithmic growth bounds above the coding dimension exponent

Entropy slopes below one force uniform deficits.

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

The question is whether a bound on logarithmic entropy growth controls the entropy increment at every dyadic scale. A weighted affine iterated function system SS on a finite alphabet II consists of real similarities and probability weights satisfying (1); the slopes may have either sign, and the translations are arbitrary.✦ φi(x)=rix+ti,0<|ri|<1,pi>0,∑i∈Ipi=1.\varphi_i(x)=r_i x+t_i,\qquad 0<|r_i|<1,\qquad p_i>0,\qquad \sum_{i\in I}p_i=1.(1)

For a measure η\eta on ℝ\mathbb R, its weighted component ηi\eta_i is the weighted pushforward in (2).✦ A measure μ\mu on ℝ\mathbb R is self-similar for SS when it satisfies the second identity in (2).✦ ηi=pi(φi)*η,μ=∑i∈Iμi.\eta_i=p_i(\varphi_i)_*\eta,\qquad \mu=\sum_{i\in I}\mu_i.(2)

For a measure η\eta on a measurable space XX and a map f:X→Bf:X\to B, the fiber mass mη,f(b)m_{\eta,f}(b) is the real value of the mass of f−1{b}f^{-1}\{b\};✦ the partition entropy H(η,f)H(\eta,f) is the sum in (3).✦ The subscript ℝ\mathbb R denotes the real value of an extended nonnegative quantity, and 0ln⁡00\ln 0 has value zero. mη,f(b)=(η{x:f(x)=b})ℝ,H(η,f)=−∑b∈Bmη,f(b)ln⁡mη,f(b).m_{\eta,f}(b)=\bigl(\eta\{x:f(x)=b\}\bigr)_{\mathbb R},\qquad H(\eta,f)=-\sum_{b\in B}m_{\eta,f}(b)\ln m_{\eta,f}(b).(3)

The shifted grid cell index λs,u\lambda_{s,u} assigns a point to an integer cell at scale ss and shift uu. For a measure η\eta on ℝ\mathbb R, the shifted grid entropy Hs,u(η)H_{s,u}(\eta) is the partition entropy of the shifted grid cell index λs,u\lambda_{s,u}, as specified in (4).✦ λs,u(x)=⌊xs+u⌋,Hs,u(η)=H(η,λs,u).\lambda_{s,u}(x)=\left\lfloor\frac{x}{s}+u\right\rfloor,\qquad H_{s,u}(\eta)=H(\eta,\lambda_{s,u}).(4)

For a measure η\eta on ℝ\mathbb R, the averaged grid entropy H¯(η,s)\overline H(\eta,s) is the Lebesgue average of the shifted grid entropy Hs,u(η)H_{s,u}(\eta) over shifts uu in [0,1][0,1].✦ For a measure η\eta on ℝ\mathbb R, the dyadic entropy deficit Δs(η)\Delta_s(\eta) at scale ss is the difference between ln⁡2\ln 2 and the averaged entropy increment in (5).✦ $$\overline H(\eta,s)=\int_0^1 H_{s,u}(\eta)\,du,\qquad \Delta_s(\eta)=\ln 2-igl(\overline H(\eta,s)-\overline H(\eta,2s)\bigr).$$(5) These entropies and deficits use natural logarithms; division by ln⁡2\ln 2 gives their base-two normalization.

The Coding Exponent

The symbol probability measure σS\sigma_S of SS is the discrete measure on II with masses pip_i, written in (6) using the unit point mass δi\delta_i at ii.✦ The Bernoulli address measure ℙS\mathbb P_S on IℕI^{\mathbb N} is the countable product of the symbol probability measure σS\sigma_S, as in (6).✦ σS=∑i∈Ipiδi,ℙS=⨂n≥0σS.\sigma_S=\sum_{i\in I}p_i\delta_i,\qquad \mathbb P_S=\bigotimes_{n\geq 0}\sigma_S.(6)

For an address ω∈Iℕ\omega\in I^{\mathbb N} and a nonnegative integer nn, the prefix contraction product Rn(ω)R_n(\omega) is the signed product in (7).✦ The scaled translation term Tn(ω)T_n(\omega) for an address ω∈Iℕ\omega\in I^{\mathbb N} and a nonnegative integer nn is the product of the prefix contraction product Rn(ω)R_n(\omega) and the translation tωnt_{\omega_n};✦ the coding map π:Iℕ→ℝ\pi:I^{\mathbb N}\to\mathbb R assigns to an address the sum of its scaled translation terms, as in (7).✦ Rn(ω)=∏j=0n−1rωj,Tn(ω)=Rn(ω)tωn,π(ω)=∑n=0∞Tn(ω).R_n(\omega)=\prod_{j=0}^{n-1}r_{\omega_j},\qquad T_n(\omega)=R_n(\omega)t_{\omega_n},\qquad \pi(\omega)=\sum_{n=0}^{\infty}T_n(\omega).(7)

The self-similar coding measure ν\nu is the pushforward of the Bernoulli address measure under the coding map: ν=π*ℙS.\nu=\pi_*\mathbb P_S.(8)✦ Its weighted components νi\nu_i are those of (2), applied to the measure in (8).

The rational radii ℛ\mathcal R are the positive rational numbers at most one.✦ For measures η,ζ\eta,\zeta on ℝ\mathbb R, the maximal ball mass ratio Mη,ζ(x)M_{\eta,\zeta}(x) is the supremum of their ratios on closed balls with radii in ℛ\mathcal R, as in (9), where B¯(x,q)={y:|y−x|≤q}\overline B(x,q)=\{y:|y-x|\leq q\}.✦ 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))}.(9)

For measures η,ζ\eta,\zeta on ℝ\mathbb R, the positive logarithmic maximal envelope Lη,ζL_{\eta,\zeta} is the function on ℝ\mathbb R whose value at xx is the nonnegative part of the logarithm in (10).✦ Lη,ζ(x)=max⁡{0,ln⁡(Mη,ζ(x)ℝ)}.L_{\eta,\zeta}(x) =\max\bigl\{0,\ln\bigl(M_{\eta,\zeta}(x)_{\mathbb R}\bigr)\bigr\}.(10)

The symbolic logarithmic maximal envelope EE on IℕI^{\mathbb N} is the function whose value at an address ω\omega is the positive logarithmic maximal envelope of ν\nu relative to νω0\nu_{\omega_0} at π(ω)\pi(\omega), as in (11).✦ The logarithmic Radon-Nikodym derivative JJ on IℕI^{\mathbb N} is the real-valued function obtained by taking the logarithm of the real value of the Radon-Nikodym derivative of ν\nu with respect to νω0\nu_{\omega_0} at π(ω)\pi(\omega), as in (11).✦ E(ω)=Lν,νω0(π(ω)),J(ω)=ln⁡((dνdνω0(π(ω)))ℝ).E(\omega)=L_{\nu,\nu_{\omega_0}}(\pi(\omega)),\qquad J(\omega)=\ln\left( \left(\frac{d\nu}{d\nu_{\omega_0}}(\pi(\omega))\right)_{\mathbb R}\right).(11)

The clipped logarithmic density correction J̃\widetilde J on IℕI^{\mathbb N} is the function whose value at an address ω\omega is J(ω)J(\omega) truncated between zero and E(ω)E(\omega) according to (12).✦ The symbol information observable WW on IℕI^{\mathbb N} is the function whose value at an address ω\omega is the negative logarithm of the initial symbol’s weight.✦ The projection information rate ℐ\mathcal I of SS is the expectation of the symbol information observable WW under ℙS\mathbb P_S minus the expectation of the clipped logarithmic density correction J̃\widetilde J under ℙS\mathbb P_S.✦ J̃(ω)=max⁡{0,min⁡{E(ω),J(ω)}},W(ω)=−ln⁡pω0,ℐ=∫WdℙS−∫J̃dℙS.\begin{aligned} \widetilde J(\omega)&=\max\{0,\min\{E(\omega),J(\omega)\}\}, &W(\omega)&=-\ln p_{\omega_0},\\ \mathcal I&=\int W\,d\mathbb P_S-\int\widetilde J\,d\mathbb P_S. \end{aligned}(12)

The positive Lyapunov exponent χ\chi of SS is the weighted average of the negative base-two logarithms of the absolute slopes, as in (13).✦ The coding dimension exponent α\alpha of SS is the ratio in (13).✦ χ=−∑i∈Ipilog⁡2|ri|,α=ℐχln⁡2.\chi=-\sum_{i\in I}p_i\log_2|r_i|,\qquad \alpha=\frac{\mathcal I}{\chi\ln 2}.(13)

The Claim

Theorem 1. Let SS be a finite weighted affine system and let μ\mu be a self-similar probability measure for SS. For every real D>αD>\alpha, there is a real constant CDC_D such that H¯(μ,s)≤Dln⁡(1/s)+CD(0<s<1).\overline H(\mu,s)\leq D\ln(1/s)+C_D\qquad(0<s<1).(14)✦ Suppose real numbers d<1d<1 and CC satisfy H¯(μ,s)≤dln⁡(1/s)+C\overline H(\mu,s)\leq d\ln(1/s)+C for every 0<s<10<s<1. There is a real δ>0\delta>0 such that δ≤Δv(μ)(v>0).\delta\leq\Delta_v(\mu)\qquad(v>0).(15)✦

In the paper, the result reads: a self-similar probability measure for a finite affine system with 0<|ri|<10<|r_i|<1, exact dimensional of dimension d<1d<1, satisfies inf⁡v>0Δv(μ)>0\inf_{v>0}\Delta_v(\mu)>0, without a separation or orientation assumption.

What Is Checked

The logarithmic growth bound (14) is stated for a self-similar probability measure of a finite weighted affine system, and the deficit assertion (15) assumes a logarithmic entropy bound with a coefficient below one.