OctoInformBeta

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

Lemma 2.2: An almost-everywhere radial limit controls averaged entropy, with a uniform grid estimate from ball masses

Radial dimension bounds entropy growth.

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 comparison concerns grid entropy at small scales and the radial exponent of ball masses. Write B(x,r)B(x,r) and B¯(x,r)\overline B(x,r) for the open and closed balls of radius rr about x∈ℝx\in\mathbb R.

For a measure ν\nu on a measurable space XX, a map f:X→Af:X\to A and a label a∈Aa\in A, the fiber mass mν,f(a)m_{\nu,f}(a) is the real-valued mass assigned to the set where f=af=a.✦ For a measure ν\nu on a measurable space XX and a map f:X→Af:X\to A, the partition entropy Hν(f)H_\nu(f) of ff is the sum in (1), with a zero summand at a fiber of zero mass. mν,f(a)=ν{x∈X:f(x)=a},Hν(f)=−∑a∈Amν,f(a)ln⁡mν,f(a).m_{\nu,f}(a)=\nu\{x\in X:f(x)=a\}, \qquad H_\nu(f)=-\sum_{a\in A}m_{\nu,f}(a)\ln m_{\nu,f}(a).(1)

A compactly carried measure ν\nu on ℝ\mathbb R is a measure for which some B≥0B\geq0 satisfies (2): ν(ℝ\[−B,B])=0.\nu\bigl(\mathbb R\setminus[-B,B]\bigr)=0.(2)✦

For real numbers s,us,u, the shifted grid cell index ks,u:ℝ→ℤk_{s,u}:\mathbb R\to\mathbb Z is the integer-valued map in (3); for a measure ν\nu on ℝ\mathbb R, the shifted grid entropy Hsu(ν)H_s^u(\nu) of ν\nu is the partition entropy of ks,uk_{s,u}: ks,u(x)=⌊xs+u⌋,Hsu(ν)=Hν(ks,u).k_{s,u}(x)=\left\lfloor\frac{x}{s}+u\right\rfloor, \qquad H_s^u(\nu)=H_\nu(k_{s,u}).(3) For a measure ν\nu on ℝ\mathbb R and a real scale ss, the averaged grid entropy H¯s(ν)\overline H_s(\nu) is the Lebesgue integral of its shifted grid entropies over offsets in [0,1][0,1], as in (4). H¯s(ν)=∫01Hsu(ν)du.\overline H_s(\nu)=\int_0^1 H_s^u(\nu)\,du.(4)✦ These entropies use natural logarithms; division by ln⁡2\ln2 gives entropy in bits.

The Claim

Theorem 1. Let ν\nu be a probability measure on ℝ\mathbb R, let B≥0B\geq0 satisfy (2), and take 0<s≤10<s\leq1, u∈[0,1]u\in[0,1], and d,η≥0d,\eta\geq0. Suppose E⊆ℝE\subseteq\mathbb R satisfies ν(Ec)≤η\nu(E^c)\leq\eta and ν(B(x,s/2))≥(s/4)d\nu(B(x,s/2))\geq(s/4)^d for every x∈Ex\in E. Then the grid entropy satisfies (5): Hsu(ν)≤ln⁡6+dln⁡4+ηln⁡(2B+3)+(d+η)ln⁡(1/s).H_s^u(\nu)\leq \ln6+d\ln4+\eta\ln(2B+3)+(d+\eta)\ln(1/s).(5)✦

Theorem 2. Let ν\nu be a compactly carried probability measure on ℝ\mathbb R. For every s>0s>0 and u∈ℝu\in\mathbb R, the grid entropy satisfies (6): Hsu(ν)≤H¯s(ν)+ln⁡2.H_s^u(\nu)\leq\overline H_s(\nu)+\ln2.(6)✦

Theorem 3. Let ν\nu be a compactly carried probability measure on ℝ\mathbb R, and let d≥0d\geq0. Suppose that, for ν\nu-almost every xx, −ln⁡ν(B¯(x,r))/(−ln⁡r)-\ln\nu(\overline B(x,r))/(-\ln r) tends to dd as r↓0r\downarrow0. For every D>dD>d there is a real constant CC such that (7) holds for every 0<s<10<s<1: H¯s(ν)≤Dln⁡(1/s)+C.\overline H_s(\nu)\leq D\ln(1/s)+C.(7)✦

In the paper, the result reads: for every compactly supported exact-dimensional probability measure ν\nu on ℝ\mathbb R of dimension dd, limsup⁡s↓0max⁡{H¯s(ν)/ln⁡2,Hs0(ν)/ln⁡2}log⁡2(1/s)≤d\displaystyle\limsup_{s\downarrow0} \frac{\max\{\overline H_s(\nu)/\ln2,H_s^0(\nu)/\ln2\}} {\log_2(1/s)}\leq d.

What Is Checked

The bounds are stated for probability measures on ℝ\mathbb R, with compact carriage or the specified bounded interval of full mass, using the fiber-mass entropy of the shifted grid cell map and its Lebesgue average over [0,1][0,1].