OctoInformBeta

OpenAI preprint, September 2026 · §5. Disjoint windows of entropy gain

§5 (proof of Theorem 1.1): summing targets and completing the contradiction: Retained translation capacities satisfy a common budget across separated dyadic bands

Entropy budgets bound the coding exponent below.

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 deficit in the small-scale entropy of a self-similar measure can coexist with a larger entropy rate of affine compositions. A weighted affine iterated function system SS on a finite alphabet II consists of similarities with real signed slopes and translations, and positive normalized weights, as in (1). φi(x)=rix+ti,0<|ri|<1,ti∈ℝ,pi>0,∑i∈Ipi=1.\varphi_i(x)=r_i x+t_i,\qquad 0<|r_i|<1,\qquad t_i\in\mathbb R,\qquad p_i>0,\qquad \sum_{i\in I}p_i=1.(1)✦ For a measure μ\mu on ℝ\mathbb R, its weighted component μi\mu_i is the weighted pushforward in (2); self-similarity means the equality there between μ\mu and the sum of these components. μi=pi(φi)*μ,μ=∑i∈Iμi.\mu_i=p_i(\varphi_i)_*\mu,\qquad \mu=\sum_{i\in I}\mu_i.(2)

A finite word w=(w0,…,wn−1)w=(w_0,\ldots,w_{n-1}) has word probability pwp_w and complete affine map gw=(rw,tw)g_w=(r_w,t_w) given by (3); the empty word has probability one and the identity map. pw=∏j=0n−1pwj,φw=φw0∘⋯∘φwn−1,φw(x)=rwx+tw.p_w=\prod_{j=0}^{n-1}p_{w_j},\qquad \varphi_w=\varphi_{w_0}\circ\cdots\circ\varphi_{w_{n-1}},\qquad \varphi_w(x)=r_wx+t_w.(3) The set 𝒢n\mathcal G_n of attainable maps at length nn consists of the distinct pairs gwg_w for words of length nn.✦ The random walk mass mn(g)m_n(g) of an attainable map g∈𝒢ng\in\mathcal G_n is the sum of the probabilities of words with the same signed slope and translation as gg.✦ The random walk entropy HnH_n, its entropy rate hh, and the positive Lyapunov exponent χ\chi use base-two logarithms in (4). mn(g)=∑w∈Ingw=gpw,Hn=−∑g∈𝒢nmn(g)log⁡2mn(g),h=infn≥1Hnn,χ=−∑i∈Ipilog⁡2|ri|.\begin{aligned} m_n(g)&=\sum_{\substack{w\in I^n\\g_w=g}}p_w, &H_n&=-\sum_{g\in\mathcal G_n}m_n(g)\log_2m_n(g),\\ h&=\inf_{n\geq1}\frac{H_n}{n}, &\chi&=-\sum_{i\in I}p_i\log_2|r_i|. \end{aligned}(4) Every finite word ww has strictly positive word probability pwp_w.✦

The symbol depth bib_i is the base-two logarithm of the reciprocal contraction magnitude.✦ The contraction depths b(w)b(w) and b(W)b(W) of finite words are the real numbers given in (5). bi=−log⁡2|ri|,b(w)=−log⁡2|rw|,b(W)=∑j=0|W|−1bWj.b_i=-\log_2|r_i|,\qquad b(w)=-\log_2|r_w|,\qquad b(W)=\sum_{j=0}^{|W|-1}b_{W_j}.(5) For a function FF on finite words, its Bernoulli word average at length HH is 𝔼HF\mathbb E_HF in (6). 𝔼HF=∑w∈IHpwF(w).\mathbb E_HF=\sum_{w\in I^H}p_wF(w).(6)✦

Coding Information

The symbol probability measure κ\kappa on the finite alphabet II, with all subsets measurable, is the weighted sum of point masses κ=∑i∈Ipiδi\kappa=\sum_{i\in I}p_i\delta_i.✦ The Bernoulli address measure ℙ=κ⊗ℕ\mathbb P=\kappa^{\otimes\mathbb N} on IℕI^{\mathbb N} is the countable product of the symbol probability measure κ\kappa.✦ For an address ω∈Iℕ\omega\in I^{\mathbb N}, the prefix contraction product Rn(ω)R_n(\omega) scales the translation at position nn to the term un(ω)u_n(\omega); the coding map π\pi and the self-similar coding measure λ\lambda are defined in (7). Rn(ω)=∏j=0n−1rωj,un(ω)=Rn(ω)tωn,π(ω)=∑n=0∞un(ω),λ=π*ℙ.\begin{aligned} R_n(\omega)&=\prod_{j=0}^{n-1}r_{\omega_j}, &u_n(\omega)&=R_n(\omega)t_{\omega_n},\\ \pi(\omega)&=\sum_{n=0}^{\infty}u_n(\omega), &\lambda&=\pi_*\mathbb P. \end{aligned}(7) The scaled translation term un(ω)u_n(\omega) of an address ω\omega is the prefix contraction product Rn(ω)R_n(\omega) multiplied by the translation tωnt_{\omega_n}.✦

For measures ν,η\nu,\eta on ℝ\mathbb R, their maximal ball mass ratio ℛν,η(x)\mathcal R_{\nu,\eta}(x) is the supremum in (8), over positive rational radii at most one; B¯(x,q)\overline B(x,q) denotes the closed ball. ℛν,η(x)=supq∈ℚ∩(0,1]ν(B¯(x,q))η(B¯(x,q)).\mathcal R_{\nu,\eta}(x) =\sup_{q\in\mathbb Q\cap(0,1]} \frac{\nu(\overline B(x,q))}{\eta(\overline B(x,q))}.(8)✦ For measures ν,η\nu,\eta on ℝ\mathbb R, the positive logarithmic maximal envelope ℒν,η(x)\mathcal L_{\nu,\eta}(x) is the maximum of zero and the natural logarithm of the real value of the maximal ball mass ratio ℛν,η(x)\mathcal R_{\nu,\eta}(x).

The symbol information observable i(ω)i(\omega) is the negative natural logarithm of the initial symbol’s weight.✦ The logarithmic density correction ρ(ω)\rho(\omega) is the natural logarithm of the real Radon-Nikodym derivative of λ\lambda with respect to its weighted component λω0\lambda_{\omega_0} at π(ω)\pi(\omega).✦ The symbolic maximal envelope ee, the clipped correction kk, the projection information rate ℐ\mathcal I, and the coding dimension exponent dcd_{\mathrm c} are defined by (9). i(ω)=−ln⁡pω0,ρ(ω)=ln⁡dλdλω0(π(ω)),e(ω)=ℒλ,λω0(π(ω)),k(ω)=max⁡{0,min⁡{e(ω),ρ(ω)}},ℐ=∫idℙ−∫kdℙ,dc=ℐχln⁡2.\begin{aligned} i(\omega)&=-\ln p_{\omega_0}, &\rho(\omega)&=\ln\frac{d\lambda}{d\lambda_{\omega_0}}(\pi(\omega)),\\ e(\omega)&=\mathcal L_{\lambda,\lambda_{\omega_0}}(\pi(\omega)), &k(\omega)&=\max\{0,\min\{e(\omega),\rho(\omega)\}\},\\ \mathcal I&=\int i\,d\mathbb P-\int k\,d\mathbb P, &d_{\mathrm c}&=\frac{\mathcal I}{\chi\ln2}. \end{aligned}(9)

Entropy and Pair Capacity

For a measure ν\nu and an observable ff, the fiber mass mν,f(z)m_{\nu,f}(z) is the real value of ν{x:f(x)=z}\nu\{x:f(x)=z\}.✦ The partition entropy ℋ(ν,f)\mathcal H(\nu,f) of a measure ν\nu and an observable ff is the sum in (10), with the contribution at zero mass defined to be zero. The shifted grid cell index Qs,uQ_{s,u} gives the shifted grid entropy ℋsu\mathcal H^u_s at scale ss and real offset uu, whose Lebesgue average over u∈[0,1]u\in[0,1] is ℋs\mathcal H_s. The dyadic deficit Δs\Delta_s is defined in the same display. ℋ(ν,f)=−∑zmν,f(z)ln⁡mν,f(z),Qs,u(x)=⌊x/s+u⌋,ℋsu(ν)=ℋ(ν,Qs,u),ℋs(ν)=∫01ℋsu(ν)du,Δs(ν)=ln⁡2−(ℋs(ν)−ℋ2s(ν)).\begin{aligned} \mathcal H(\nu,f)&=-\sum_zm_{\nu,f}(z)\ln m_{\nu,f}(z), &Q_{s,u}(x)&=\lfloor x/s+u\rfloor,\\ \mathcal H_s^u(\nu)&=\mathcal H(\nu,Q_{s,u}), &\mathcal H_s(\nu)&=\int_0^1\mathcal H_s^u(\nu)\,du,\\ \Delta_s(\nu)&=\ln2-\bigl(\mathcal H_s(\nu)-\mathcal H_{2s}(\nu)\bigr). \end{aligned}(10) These entropies use natural logarithms; division by ln⁡2\ln2 gives base-two units.

A measure ν\nu is compactly supported when ν(ℝ\[−R,R])=0\nu(\mathbb R\setminus[-R,R])=0 for some R≥0R\geq0.✦

A finite probability law ν\nu on ℝ\mathbb R consists of a finite support EνE_\nu with weights νx>0\nu_x>0 summing to one.✦ For an integer ℓ\ell, its pairs at scale ℓ\ell are the unordered pairs in ℰℓ(ν)\mathcal E_\ell(\nu) in (11).✦ For a finite probability law ν\nu on ℝ\mathbb R and an integer ℓ\ell, a nonnegative pair weighting ww on ℰℓ(ν)\mathcal E_\ell(\nu) is feasible when its incident pair mass Iw(x)I_w(x) is at most νx\nu_x at every point of EνE_\nu.✦ The fractional matching capacity cℓ(ν)c_\ell(\nu) of a finite probability law ν\nu on ℝ\mathbb R is the supremum of the total feasible pair weight at scale ℓ\ell. ℰℓ(ν)={{x,y}:x,y∈Eν,2−ℓ≤|x−y|<21−ℓ},Iw(x)=12∑e∈ℰℓ(ν)x∈ewe,cℓ(ν)=sup⁡{∑e∈ℰℓ(ν)we:we≥0,Iw(x)≤νxfor every x∈Eν}.\begin{aligned} \mathcal E_\ell(\nu) &=\bigl\{\{x,y\}:x,y\in E_\nu, \ 2^{-\ell}\leq|x-y|<2^{1-\ell}\bigr\},\\ I_w(x)&=\frac12\sum_{\substack{e\in\mathcal E_\ell(\nu)\\x\in e}}w_e,\\ c_\ell(\nu)&=\sup\left\{ \sum_{e\in\mathcal E_\ell(\nu)}w_e: w_e\geq0,\ I_w(x)\leq\nu_x\ \text{for every }x\in E_\nu \right\}. \end{aligned}(11)

For a finite source set EE with weights qvq_v and an observable f:E→Yf:E\to Y, the fiber mass at yy is the sum of qvq_v over f(v)=yf(v)=y.✦ For a real observable ff on a finite source set EE with positive normalized weights qvq_v, the distribution of ff is the finite probability law assigning to each distinct value yy of ff the sum of qvq_v over the source points satisfying f(v)=yf(v)=y.✦ For a real observable ff, positive source weights, and a tag map θ:E→Y\theta:E\to Y, the tag fiber Fθ(v)F_\theta(v), its mass Mθ(v)M_\theta(v), and the conditional observable distribution νθ,vf\nu^f_{\theta,v} are defined by (12). Fθ(v)={u∈E:θ(u)=θ(v)},Mθ(v)=∑u∈Fθ(v)qu,νθ,vf=∑u∈Fθ(v)quMθ(v)δf(u).\begin{aligned} F_\theta(v)&=\{u\in E:\theta(u)=\theta(v)\}, &M_\theta(v)&=\sum_{u\in F_\theta(v)}q_u,\\ \nu^f_{\theta,v} &=\sum_{u\in F_\theta(v)}\frac{q_u}{M_\theta(v)}\delta_{f(u)}. \end{aligned}(12) For strictly positive source weights, the conditional weights qu/Mθ(v)q_u/M_\theta(v) sum to one over the tag fiber Fθ(v)F_\theta(v).✦

Fix a map τ:I→β\tau:I\to\beta to a finite type alphabet. For a map τ:I→β\tau:I\to\beta and a word w∈Inw\in I^n, the word tag sequence τn(w)\tau^n(w) is the sequence in βn\beta^n with entries τ(wj)\tau(w_j).✦ The type-conditioned translation law νw\nu_w is the conditional observable distribution of v↦tvv\mapsto t_v on InI^n, with weights pvp_v and tag map τn\tau^n, at the word ww.✦ For a finite list WW of symbols in II, the type-conditioned translation law νW\nu_W is the conditional observable distribution of v↦tvv\mapsto t_v on I|W|I^{|W|}, with weights pvp_v and tag map τ|W|\tau^{|W|}, at the word formed from WW in address order.✦

Retained Targets

For a real depth bound A′A', an integer band interval [a,B][a,B], and a nonnegative block length nn, the averaged band capacity budget ℬA′(a,B;n)\mathcal B_{A'}(a,B;n) is the weighted sum in (13). ℬA′(a,B;n)=∑w∈Inb(w)≤A′npw∑ℓ=aBcℓ(νw).\mathcal B_{A'}(a,B;n) =\sum_{\substack{w\in I^n\\b(w)\leq A'n}} p_w\sum_{\ell=a}^{B}c_\ell(\nu_w).(13)✦ All integer sums range over inclusive integer intervals and are zero when the interval is empty.

For a depth profile σ:ℕ→ℝ\sigma:\mathbb N\to\mathbb R, an index jj is retained at target L∈ℤL\in\mathbb Z when the two conditions in (14) hold. σ(j+n)−σ(j)≤A′n,a≤L−⌈σ(j)⌉≤B.\sigma(j+n)-\sigma(j)\leq A'n,\qquad a\leq L-\lceil\sigma(j)\rceil\leq B.(14)✦ For a finite word WW, write W<jW_{<j} for its first jj letters and W[j,j+n)W_{[j,j+n)} for the next nn letters, truncating either segment at the end of WW. The prefix depth profile σW\sigma_W of a finite word WW is the function σW(j)=b(W<j)\sigma_W(j)=b(W_{<j}). For a finite word WW, the retained matching capacity CL,jA′,n,a,B(W)C^{A',n,a,B}_{L,j}(W) is the nonnegative real number in (15), determined by the prefix depth profile σW\sigma_W. CL,jA′,n,a,B(W)={cL−⌈σW(j)⌉(νW[j,j+n)),if j is retained at L,0,otherwise.C^{A',n,a,B}_{L,j}(W)= \begin{cases} c_{L-\lceil\sigma_W(j)\rceil}(\nu_{W_{[j,j+n)}}), &\text{if }j\text{ is retained at }L,\\ 0,&\text{otherwise}. \end{cases}(15)✦

The Claim

Theorem 1. Let f:ℤ→ℝf:\mathbb Z\to\mathbb R and let a,B,c,Ta,B,c,T be integers with a≥1a\geq1, c≥0c\geq0, and B+c≤TB+c\leq T. Then ∑1≤L≤Ta≤L−c≤Bf(L−c)=∑ℓ=aBf(ℓ).\sum_{\substack{1\leq L\leq T\\a\leq L-c\leq B}}f(L-c) =\sum_{\ell=a}^{B}f(\ell).(16)✦

Fix nonnegative integers H,j,nH,j,n with j+n≤Hj+n\leq H, a real A′A', and integers a,B,Ta,B,T with a≥1a\geq1. Suppose that every finite word WW satisfies 0≤⌈σW(j)⌉0\leq\lceil\sigma_W(j)\rceil and B+⌈σW(j)⌉≤TB+\lceil\sigma_W(j)\rceil\leq T. Then $$\sum_{L=1}^{T}\mathbb E_H\!left[C^{A',n,a,B}_{L,j}\right] =\mathcal B_{A'}(a,B;n).$$(17)✦

Suppose that symbols with the same type have the same signed slope, and let μ\mu be a compactly supported self-similar probability measure for SS. Let δ∈ℝ\delta\in\mathbb R satisfy δ≤Δs(μ)\delta\leq\Delta_s(\mu) for every s>0s>0. Let M,D,H,qM,D,H,q be nonnegative integers with M>4M>4, D≥1D\geq1, 5ln⁡2/M≤δ/25\ln2/M\leq\delta/2, and bv≤Db_v\leq D for every symbol vv. Choose nonnegative integers ni,Kin_i,K_i and integer endpoints ai,Bia_i,B_i for 1≤i≤q1\leq i\leq q, and a real A′A'. Assume that Kini=HK_in_i=H, ai≥1a_i\geq1, and Bi≤HB_i\leq H for every ii, and that Bk<ai−A′ni−1B_k<a_i-A'n_i-1 whenever k<ik<i. Then δ2∑i=1qKiℬA′(ai,Bi;ni)≤2DHln⁡(M2).\frac{\delta}{2}\sum_{i=1}^{q} K_i\mathcal B_{A'}(a_i,B_i;n_i) \leq 2DH\ln(M^2).(18)✦

Suppose that symbols with the same type have the same signed slope, and let μ\mu be a self-similar probability measure for SS. Let d,C,A,A′d,C,A,A' be real numbers with 0≤d<10\leq d<1, A>0A>0, and χ<A′<A\chi<A'<A. Suppose that ℋs(μ)≤dln⁡(1/s)+C\mathcal H_s(\mu)\leq d\ln(1/s)+C for every 0<s<10<s<1. Then hln⁡2≤dAln⁡2+ln⁡|β|.h\ln2\leq dA\ln2+\ln|\beta|.(19)✦

For every self-similar probability measure μ\mu for SS, the coding dimension exponent satisfies min⁡{1,hχ}≤dc.\min\left\{1,\frac{h}{\chi}\right\}\leq d_{\mathrm c}.(20)✦

In the paper, the result reads: summing targets L=1,…,TL=1,\ldots,T gives a fixed positive contribution from each band, whereas the disjoint-window budget is 2Tlog⁡2M2T\log_2M, yielding 2Tlog⁡2M≥δcqT/(4D)2T\log_2M\geq\delta c qT/(4D) with positive constants δ,c\delta,c and block-depth bound DD, which is impossible for arbitrarily large qq, so d<1d<1 and dχ<hd\chi<h cannot hold together.

What Is Checked

The identities (16) and (17) concern integer target intervals and finite Bernoulli words; the bounds (18) through (20) concern self-similar probability measures and their finite type-conditioned translation laws, with compact support and a uniform deficit required in (18). Every finite word has strictly positive probability✦, and the conditional weights on each tag fiber sum to one.✦