ARITHMETIC / LEARNING · EXPERIMENT RECORD 20261105
Every correlation.
Every saved update.
Replay a finite Fourier learner on actual coupled sextic-character targets. Inspect its full210-coordinate loss, retained raw zeros, signed readouts and exact rational certificates.
This page replays saved rational states. It performs no browser training. Plots and numerical values are double-precision display projections; the linked source files contain the complete exact arithmetic.
FINITE INSTANCE · D = 6000 · U = 60 · H = 625 · PRINCIPAL ν
Open the saved record.
Requested amplitude error
τ = 10−12
Loading and verifying the compact display record…
Loss and readout certificates
- Full weighted loss L
- CC certificate²
- CM certificate²
- Requested τ²
Signed CC and CM readouts
μ(q), complex profile conjugation, all h deletions and the exact diagonal remain included. Residuals are projected from exact differences, before floating-point subtraction can lose small errors.
ALL210 ORDERED OFF-DIAGONAL CORRELATIONS
A whole vector, not one scalar.
Selected complex residual
Exact basis coordinates
A pair (a,b) means a+bζ₆, with |a+bζ₆|² =a²+ab+b².
ONE RELEVANT FINITE CHARACTER-FEATURE BLOCK
A saved phase, an inspectable F.
Display-only evaluation of Fg(e) =Σ gᵣ conjugate(θᵣ(e)) from the selected saved coefficients. Phase choices and exact targets come from the saved block samples. This computes a finite feature value; it does not execute an optimizer or enumerate the dense residue group.
SCOPE, SOURCE AND EXACTNESS
A finite certificate has a precise boundary.
All 15 good squarefree ideal columns at the four activated norms 7507,7519,8281,10507 are included:2+4+1+8 ideals. One declared common smooth four-bump profile vanishes at every other integer norm. The real and complex profiles use exact rational nodal values; no approximate exponential labels are used.
The2262 literal rows retain all six units and S-powers. The separate 2148-row set is a guaranteed nonprincipal subset, not the entire nonprincipal family. Every row case has4530 raw column zeros, and every run has64 nonzero negative-μ(q) CC orientations.
The full finite projection loss is evaluated exactly through Parseval. Metric GD uses g←(g+Γ)/2 in the declared λᵢM metric, so L decreases by exactly a factor 1/4. Ordinary real-coordinate GD is a separate 50-update control. These methods are labelled separately throughout.
All four metric runs reach both signed readout tolerances and the full-coordinate tolerance:53/55/53/55 executed updates. Fixed-finite-instance termination permits unrestricted exact resources; it gives no conductor-, range-, profile- or runtime-uniform guarantee.
This is supervised Fourier coefficient learning from exact labels, with no claimed asymptotic advantage over evaluating those labels. It establishes no new analytic saving, covariance positivity, PSD or positive arithmetic determinant, generalization, zero location, Lean theorem, quantum advantage or GRH proof.