OctoInformBeta

OpenAI preprint, September 2026 · §3. Uniform compression from partitions · §3

Lemma 3.1 [Logarithmic profile]: Weights γh(j)\gamma_h(j) summing to one build φh\varphi_h from balls of radius j<hj<h; dilating each radius jj to 3j+23j+2 costs at most log⁡3/log⁡(h+1)\log 3/\log(h+1)

The logarithmic profile is an average of ball indicators.

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

Take a graph TT and its distance dTd_T, with values in {0,1,2,…}∪{∞}\{0,1,2,\dots\}\cup\{\infty\} and dT(x,y)=∞d_T(x,y)=\infty when xx and yy lie in different components. For a distance dd and a natural number kk, the indicator 𝟏{d≤k}\mathbf 1\{d\le k\} is 11 when d≤kd\le k and 00 otherwise, ∞\infty being larger than every natural number. Read at d=dT(x,y)d=d_T(x,y) as a function of yy, it is the indicator of the ball of radius kk about xx.

A profile turns distances into weights between 00 and 11. The question is how the cutoff profile of radius hh, for a positive integer hh, is assembled from the indicators of the balls of radii 0,1,…,h−10,1,\dots,h-1, and how much the assembly grows when each ball of radius jj is replaced by the ball of radius 3j+23j+2.

The Profile and Its Weights

For a positive integer hh, the cutoff profile φh\varphi_h of radius hh is the real function on {0,1,2,…}∪{∞}\{0,1,2,\dots\}\cup\{\infty\} with φh(∞)=0\varphi_h(\infty)=0 and, at a finite distance dd, φh(d)=max⁡{0,1−log⁡(d+1)log⁡(h+1)}.\varphi_h(d)=\max\Bigl\{0,\;1-\frac{\log(d+1)}{\log(h+1)}\Bigr\}.(1)✦

For a positive integer hh and a natural number jj, the logarithmic increment γh(j)\gamma_h(j) is the real number γh(j)=log⁡(j+2)−log⁡(j+1)log⁡(h+1).\gamma_h(j)=\frac{\log(j+2)-\log(j+1)}{\log(h+1)}.(2)✦

The Claim

Theorem 1. Let hh be a positive integer. Then γh(j)>0for every natural number j,\gamma_h(j)>0\quad\text{for every natural number } j,(3)✦ ∑j=0h−1γh(j)=1,\sum_{j=0}^{h-1}\gamma_h(j)=1,(4)✦ and for every d∈{0,1,2,…}∪{∞}d\in\{0,1,2,\dots\}\cup\{\infty\}, φh(d)=∑0≤j<hd≤jγh(j),\varphi_h(d)=\sum_{\substack{0\le j<h\\ d\le j}}\gamma_h(j),(5)✦ φh(d)=∑j=0h−1γh(j)𝟏{d≤j},\varphi_h(d)=\sum_{j=0}^{h-1}\gamma_h(j)\,\mathbf 1\{d\le j\},(6)✦ ∑0≤j<hd≤3j+2γh(j)≤φh(d)+log⁡3log⁡(h+1),\sum_{\substack{0\le j<h\\ d\le 3j+2}}\gamma_h(j)\le\varphi_h(d)+\frac{\log 3}{\log(h+1)},(7)✦ ∑j=0h−1γh(j)𝟏{d≤3j+2}≤φh(d)+log⁡3log⁡(h+1).\sum_{j=0}^{h-1}\gamma_h(j)\,\mathbf 1\{d\le 3j+2\}\le\varphi_h(d)+\frac{\log 3}{\log(h+1)}.(8)✦

In the paper, the result reads: for every positive integer hh, the weights γj=(log⁡(j+2)−log⁡(j+1))/log⁡(h+1)\gamma_j=(\log(j+2)-\log(j+1))/\log(h+1), 0≤j<h0\le j<h, the page’s γh(j)\gamma_h(j), are positive and sum to one, and for every d∈{0,1,2,…}∪{∞}d\in\{0,1,2,\dots\}\cup\{\infty\} one has φh(d)=∑j=0h−1γj𝟏{d≤j}\varphi_h(d)=\sum_{j=0}^{h-1}\gamma_j\mathbf 1\{d\le j\} and ∑j=0h−1γj𝟏{d≤3j+2}≤φh(d)+log⁡3/log⁡(h+1)\sum_{j=0}^{h-1}\gamma_j\mathbf 1\{d\le 3j+2\}\le\varphi_h(d)+\log 3/\log(h+1).

What Is Checked

The claim is stated over the logarithmic increments γh(j)\gamma_h(j) of (2) and the cutoff profile φh\varphi_h of (1), for a positive integer hh and every distance dd in {0,1,2,…}∪{∞}\{0,1,2,\dots\}\cup\{\infty\}.