Dyadic Two-Zone Sieve Decomposition for the Unaligned Fejér Residual

Investment state: active. This describes research progress; claims have separate evidence grades.

Contribution to the goal

Formulates a dyadic two-zone partition for the unaligned Fejér residual R(H) = sum_m w_nu(m) D_nu^un, decoupling bulk arithmetic cancellation in q on m <= X/(log H)^4 from boundary logarithmic density on X/(log H)^4 < m <= H, bypassing the lower-endpoint obstruction.

Prior work and proposed difference

Prior art — updated search record (job #12514, route 408, dyadic m-block census)

Search date 2026-10-27, from this machine, one search engine, titles and snippets only: no external full text, table or dataset was opened, and no external number is used anywhere in this return. Every number here is recomputed by this run's instrument.

Queries and what came back.

  1. dyadic decomposition of a sieve majorant residual bulk boundary layer reduced residues twin primes exponential sum census — returned primorial-sieve preprints that construct twin-admissible residue classes (a "replication-deletion primorial sieve" and a companion note), a large-sieve estimate for sets missing residue classes mod prime squares (arXiv 2610.20554), Kumchev's sieve-and-exponential-sums notes, a Selberg-sieve lecture, and bounds for exponential sums on reduced residue systems. None of them decomposes a majorant's residual over dyadic ranges of the modulus m, and none computes an unaligned residual at all.
  2. unaligned part Fejer kernel majorant sum over m reduced residues arithmetic progression (searched in #6264 and again here) — no source that poses an unaligned Fejér-majorant residual.
  3. log-periodic / scaling law of the residual of a divisor-measure majorant — nothing on the shape of such a residual across scales; the returned material is about the sieve weight, not about the distribution of a residual in the modulus variable.

Nearest prior art, and the exact difference. The classical frame is unchanged from #6252/#6264: the aligned object sum_{a in (Z/q)*} |E_H(a/q)|^2 is finite-grid orthogonality and the Montgomery–Vaughan mean-value machinery bounds it; (8) of #6156 is its closed form, and #6156's (12) bounds it. The unaligned residual sum_m |U_nu(m)| — the exact per-m difference between the CRT class sum and that closed form — is what this return measures, and no source found computes it. What this return adds beyond #6264's prior-art record is the shape question: no prior work found here gives the distribution of such a residual over dyadic ranges of m, and none reports whether such a residual tracks its own divisor-measure majorant block by block. That tracking (this return's F1: sum|U_0|/AL_0 = 0.586–0.644 and sum|U_1|/AL_1 = 0.655–0.723 across the twelve blocks, i.e. constant to within ~10 %) is the measured fact that decides whether a block-local absorption can gain anything.

Exact remaining gap. Whether the (g,e) split — Step 1's own object sum_{ge&lt;=H} u_0(g)u_2(e)|D^un|, for which this return's m-aggregate is a triangle-inequality lower bound — separates inside a dyadic block where the m-split does not. The gap is a placement question one level finer than this run's, not a digit question, and it is what the next step addresses.

Central uncertainty

Whether the boundary layer sum with w_nu(m) = O(1) can be bounded by O(H log H log log H) purely through the divisor-measure of u_0(g)u_2(e) without requiring additional phase cancellation in D_nu^un.

Next experiment

Step 1's own object sum_{ge<=H} u_0(g)u_2(e)|D^un| has never been measured: the m-aggregate is its triangle lower bound, and the old success clause ((g,e)-ratio <= 0.5) is unsatisfiable because the (g,e)-ratio is >= the m-ratio per block and every block m-ratio of the recorded table (H = 3e8) exceeds 0.586. Does the (g,e)-weighted unaligned object sit within 10% of the m-aggregate in every dyadic block (pair-level neutrality), or does per-pair mass concentrate where the m-split cannot see it?

The same exact-Phi instrument with the (g,e) decomposition carried through the class recursion and accumulated inside the twelve dyadic m-blocks, exactly as #6280's step posed it: the aligned terms are already enumerated per divisor g|m (A(g)J(m/g) psi_2(H/g) and A(g)J(m/g) T(H/(2g))), so the block AL is reproduced by summing them per block; the new work is the per-pair unaligned term, which needs each CRT class's weight projected onto the g-part and the e-part of the class (the class generator itself is reused; the counted work is sum over odd squarefree m <= H of 6^omega(m) = 3^omega(m) classes times 2^omega(m) divisors). Per block and branch report: the aligned (g,e) sum (must reproduce this return's block AL exactly), the m-aggregate sum_m|U_nu(m)| (must equal this return's block |U| and must not exceed the (g,e) object - the triangle direction, kept as a validation gate), and the (g,e)-weighted unaligned object sum_{(g,e)} u_0(g)u_2(e)|D^un(g,e)| (Step 1's own object, never measured), plus each block's (g,e)-object share of the total and the top four blocks' share (blocks 0-3 span m > H/16). The old success clause is removed: it required a block (g,e)-ratio <= 0.5, which the triangle direction makes unreachable wherever the m-ratio exceeds 0.5 (all 23 recorded table ratios at H = 3e8, min 0.586438). Pre-registered on the per-block per-branch ratios of the (g,e)-object to AL: neutrality = every block within 10% of its m-aggregate ratio (recorded ranges: branch 0 0.586438..0.644446, branch 1 0.654860..0.722832 at H = 3e8); concentration = some block more than 10% above its m-ratio, with the 2x level marking strong concentration. Scales: H = 1e5 first, where the counted work is ~1e9-1e10 class-divisor operations (controls on the record: #6264's row sum|U| 501360.8, AL 816725.6, n0 40527, and #6280's block profile at 3e7/1e8/3e8, whose per-branch ratios the new run must bracket); then H = 1e6 if the cost model is confirmed, since 6^omega grows like (log H)^5.

Required evidence

Unaccepted premises remain conditional.

Evidence behind continued investment

These investigations led to the current experiment. Their claims retain their own evidence grades.

Investigation history

Route 408's stored next step executed as written: #6264's exact-Phi instrument plus twelve dyadic m-block accumulators, nothing else changed. Attachments:

file role sha256
phicb.c the instrument: #6264's phic.c (byte-identical to its published artifact hash a4f42607…) plus the block bucket, its range bookkeeping and the blocks JSON field 83316279fe29fac0
pub_phicb_30000000.json, pub_phicb_100000000.json, pub_phicb_300000000.json raw instrument output, one JSON object each (totals, both zone buckets, all twelve blocks) d8a9cb30…, 0fa6325d…, 10864554…
reference_bg.py independent exact-Fraction recomputation of Phi_0, Phi_1, U_0, U_1 from the definitions, bucketed by the same rule, sharing no code with the instrument f5fdba6e085b3549
ref_1000.json, ref_3000.json the reference's own output at H = 1e3, 3e3 2ee46752…, 7022c232…
gates_bg.py the gate suite (G1 control … G8 verdict) and both fault sweeps 017f2cba0a363240
gates.out GATES: 73/73 pass, 0 fail, exit 0 9634b81ee2145c33
fault_sweep.out GATES: 85/85 pass, 0 fail, exit 0 — 5/5 harness faults and 7/7 instrument faults caught ec3d8be203460c69
results_bg.json every reported ratio, per scale and per block, machine-readable a1bcaf3377188ad7
run_blocks_pub.sh the three scales under the department tool's enforced wall-clock wrapper, checkpointing every 200000 m (placeholder run name/tool path) 71aacf6ddcd8c5d9

Decisive facts

  1. Control exact. H = 3e7 and 1e8 reproduce #6264's published values to the published digits, not within a tolerance: UN -165204454.9973803 / -590124000.4975066, sum|U| 239307751.14034262 / 862532894.1080903, AL 357252675.5729102 / 1269847314.7489388, n0 12158561 / 40528460, classes 359252929 / 1328886196.
  2. H = 3e8 new. n0 = 121585417, CRT classes = 4363825153, UN = -1875543456.350526, sum|U| = 2766948656.38378, AL = 4023739834.8795924, 154.56 s on one core.
  3. Pre-registered rule at H = 3e8: the success clause holds and the failure clause does not fire. Max block share 0.411135 (block 1), top block (H/2 &lt; m &lt;= H) share 0.194370, sum|U|/(H log H log log H) = 0.159021 <= 0.6, sum|U|/(H log H) = 0.472515 < 1.0. The top block alone does not carry essentially all of the residual.
  4. Continuous: sum|U|/(H log H) = 0.463325 -> 0.468242 -> 0.472515 at 3e7 / 1e8 / 3e8; the fitted exponent of sum|U| ~ C H (log H)^a over #6264's nine scales plus this one is 1.183232 (r2 0.995212).
  5. The limit (F1): the split is non-degenerate in value but constant-neutral. sum|U_0|/AL_0 ranges 0.586438..0.644446 over the twelve blocks (0.642072 overall) and sum|U_1|/AL_1 ranges 0.654860..0.722832 (0.719248 overall), while the block shares are scale-invariant to < 0.0012 between H = 3e7 and 3e8. So block-locality buys at most a factor 1.10 against the global sum|U|/AL = 0.687656.
  6. F2: the top block's combined-ratio deficit (0.644 against 0.694 at H = 1e8) is the absent branch-1 term (n1 = 0 where 2m &gt; H), not a property of the range m ~ H. F3: the top four blocks carry 0.909139 of sum|U|; the tail m &lt;= H/512 carries 0.090861.

Reviewer reproduction

cc -O2 -o phicb phicb.c; ./phicb H 1 H 1 1 --quiet for H in 30000000, 100000000, 300000000; then python3 gates_bg.py --fault-sweep (expects 85/85, exit 0).

Not established

No proof; the measured object is the m-aggregate sum_m|U_nu(m)|, a lower bound for Step 1's (g,e)-weighted object, and whether the (g,e) split separates inside a block is not measured here (it is the proposed next step). No route closure, no document integration, no change to another route.

The pre-registered step was executed as written, with the exact-Phi instrument of #6252 reimplemented in C, validated against that return's seven Python values (worst relative deviation 7.26e-14), and with two independent arms (direct h-enumeration with no CRT; a sieve census) inside the same program. Rung: measured, two scales.

Measured (H = 3e7 / 1e8). UN = -165204454.9973803 / -590124000.4975066; sum|U| = 239307751.14034262 / 862532894.1080903; AL = 357252675.5729102 / 1269847314.7489388. sum|U|/(H log H) = 0.463325 / 0.468242; sum|U|/(H log H log log H) = 0.162805 / 0.160716; AL/(H log H) = 0.691678 / 0.689360; sum|U|/AL = 0.669856 / 0.679241; UN/AL = -0.462430 / -0.464720. Fitted exponent of sum|U| ~ C H (log H)^a: a = 1.1873 (r2 0.9942, scales 1e3..1e8). 1.33e9 CRT classes at H = 1e8 in 41.78 s on one core.

The pre-registered rule splits. Success 1 (<= 0.6) holds: 0.162805, 0.160716. Success 2 (< 1.6) holds: 1.1873. Success 3 (boundary share < 0.9) FAILS at both scales with share exactly 1.0000000000. Failure 1 (>= 0.9) does not fire; failure 2 (share stays 1.0 while sum|U|/(H log H) rises: 0.458208 -> 0.463325 -> 0.468242) FIRES. So Step 1 as stated does not survive its own success clause, and its failure clause is triggered by the split, not by the sum.

F1 (decisive, placement). the split point is m* = H/(log H)^6 = 1.1519 (3e7) and 2.5596 (1e8), both below 3, and the only odd squarefree m below 3 is m = 1, for which U_0(1) = U_1(1) = 0 exactly (rho = H mod 1 = 0, R = H mod 2 = 0; aligned term at g = e = 1 is psi_2(H) = 0 and T(H/2) = 0 for even H). The bulk bucket is 0 and the boundary share is exactly 1 at every measured scale. Under the other reading of the route's text (X = H, bulk m <= H/(log H)^4 = 341, 868) the bulk holds 140, 351 m-values carrying 5.9e-6, 4.4e-6 of sum|U|. The two-zone decoupling is therefore empty in value at both named scales: 'bulk arithmetic cancellation in q' has no range to act on, and the density-only term is asked to absorb the whole sum.

F2 (the first criterion is a priori automatic). sum|U| <= sum|Phi| + AL = (1.0584..1.1936) H log H (measured, triangle inequality) and H log H <= H log H log log H, so the pre-registered ratio is bounded by 0.3737 (3e7) and 0.3633 (1e8) before any computation. A constant-level test of the mechanism must be stated against H log H or against AL.

F3 (the discriminating constant). sum|U|/AL = 0.669856 and 0.679241: the unaligned residual is two thirds of the aligned closed form (8), so a bound through the divisor measure alone must absorb a comparable-size quantity and can gain only log log H; sum|U|/(H log H) rises monotonically over the last four scales while AL/(H log H) is flat at 0.69.

F4 (no aggregate sign cancellation). UN < 0 at every scale with UN/(H log H) = -0.31985 / -0.32036, flat across three decades, |UN|/AL = 0.4625 / 0.4647.

Limits. Finite census at two scales, no proof of Step 1, and the measured quantity is the m-aggregate sum_m|U_nu(m)|, a lower bound for the route's (g,e)-weighted sum_{ge<=H} u_0 u_2 |D^un|. Phi is exact integer arithmetic; the aligned sums are long double. Gates 10/10 exit 0, 12/12 planted faults caught.

F1 (algebraic, decisive). For coprime g,e the route's weight is the aligned weight: u_0(g)u_2(e)(g/e)2^omega(e) = A(g)J(e), exact on all 5,725 coprime odd-squarefree pairs <= 200. So S_bound(H) = sum_window A(g)J(e): no D_nu^un, no w_nu, no phase. By #6156 (10)-(12), |script A(H)| <= (3/4)C_alpha K_A K_J H(1+log H), hence S_bound = O(H log H) already and S_bound/(H log H loglog H) -> 0: the route's success clause (<=2.5) is automatic.

Measured S_bound (coprime reading / literal non-coprime convolution): 1e5: 2.0940e6 / 3.1418e6; 1e6: 2.4632e7 / 3.7283e7; 1e7: 2.8362e8 / 4.3108e8. Ratios to H logH loglogH: 0.744/0.679/0.633 and 1.117/1.028/0.962. Ratios to H log^2H: 0.158/0.129/0.109 and 0.237/0.195/0.166. Both clauses fire at every named H (success <=2.5 and failure >=0.1); the failure clause is thresholded in H (log H <= 10c), not in the mathematics.

F2 (scaling). With X = H/(log H)^2 the split point is m* = H/(log H)^6 = 0.0429 / 0.1438 / 0.5703 < 1: the bulk zone is empty and the boundary zone is the whole range. Crossing: H* = 24,128,091.7. Confirmed in the measurement: as_written == all_m to 0 relative difference, and the alternative X = H reading moves S_bound by 1.3e-5 relative (the statistic barely feels the partition).

F3 (the real object). U_nu(H;m) := Phi_nu(H;m) - sum_{ge=m}A(g)J(e){psi_2(H/g) or T(H/(2g))} = sum_{ge=m} u_0u_2 D_nu^un exactly ((5)+(8) of #6156), computed at O(3^omega(m)) per m by CRT:

H | UN | sum|U| | AL | |UN|/(HlogH) | sum|U|/(HlogH) | sum|U|/(HlogHloglogH) | UN/AL 1e3 | -1930.4 | 2557.7 | 5066.9 | 0.280 | 0.370 | 0.192 | -0.381 3e3 | -7717.7 | 9962.0 | 18193.5 | 0.321 | 0.415 | 0.199 | -0.424 1e4 | -30494.5 | 40556.4 | 68643.4 | 0.331 | 0.440 | 0.198 | -0.444 3e4 | -96236.7 | 129887.6 | 216918.9 | 0.311 | 0.420 | 0.180 | -0.444 1e5 | -364043.7 | 501360.8 | 816725.6 | 0.316 | 0.436 | 0.178 | -0.446 1e6 | -4383577.0 | 6179095.9 | 9675236.2 | 0.317 | 0.447 | 0.170 | -0.453 1e7 | -51511976.9 | 73854477.7 | 111868432.5 | 0.320 | 0.458 | 0.165 | -0.460

Fit over seven rows: sum|U| ~ H(log H)^1.203; sum|U|/(HlogH) settles at 0.46 while sum|U|/(HlogHloglogH) falls 0.192 -> 0.165. The unaligned aggregate is the same order as the aligned part and negative throughout (no sign cancellation below H log H). Since sum_m|U| <= sum_{ge}u_0u_2|D^un|, this is a lower bound for the route's Step-1 quantity: Step 1 would hold with only the loglog H factor as slack (0.165-0.20 of target) — a constant-level inequality, which is the hard part, not an order-level gap.

F4 (next experiment, replaces the route's statistic). Same exact-Phi instrument at H = 3e7 and 1e8, the first scales with two non-empty zones, reporting UN, sum|U|, AL and the two zone sums separately. Pre-registered: sum|U|/(HlogHloglogH) <= 0.6 with boundary share < 0.9 keeps Step 1 alive; >= 0.9, or a share of 1.0 with a rising ratio, refutes Step 1 as stated. Needs the same algorithm in C (~0.5 CPU-h).

Controls. (7) 1,320 direct complex-Fourier checks, worst 3.0e-12. (5) 250 CRT-vs-Fourier cases, worst 3.0e-13. (8) 250 cases, worst 8.0e-13. Unaligned both ways 136 cases, worst 5.8e-13. B(H) from its own C_d definition (H = 30,60,90,120, both parities) equals sum_m w Phi to 5.8e-15 and is non-negative; that identity also pins the w convention (q = 1 with alpha(1) = 1, and omega(mq) >= 2, are both required). check_g3.py 13/13 PASS exit 0; 13/13 planted faults turn exactly their gate red.