Optimization and uniform error
Logarithmic vertical scale. Cyan: certified uniform bound. Gold: exact objective, displayed approximately. Dashed: requested uniform error. These are different quantities.
VALIDATED DIRICHLET TARGETS · EVERY UPDATE REPLAYABLE
Three completed primitive Dirichlet functions, a compact complex disk, and an executed rational readout optimizer. Inspect every coefficient update and the five sources of the proved uniform error.
This readout trains J(b) = ½ Σ |bₙ − tₙ|² on rational Fourier midpoint targets. Its difference from least squares at the true circle nodes is certified by outward coefficient rectangles. The coefficients use the normalized variable z = (s − ½)/(6/5).
Loading preserved optimizer records…
Truncation parameters change with the request. The interval-radius test controls precision retries; all five saved sample bounds are rounded upward at a request-dependent scale. Labels at true circle nodes are recomputed, rather than reusing the old fixed-data majorant.
Logarithmic vertical scale. Cyan: certified uniform bound. Gold: exact objective, displayed approximately. Dashed: requested uniform error. These are different quantities.
The optimization residual decreases. The fixed sample, aliasing, Taylor and arithmetic terms remain. A small training loss alone does not certify the function.
Cyan: selected iterate; muted: final saved iterate. Color points follow the input circle angle. Decimal rendering is illustrative; the exact rational certificate supplies the uniform bound.
Download complete exact certificate ↗
Fresh adaptive producer · Preserved fixed producer · Definitions and proof
The fresh adaptive run certifies requested 10⁻²⁴ error for all three conductors, using degree 48 and 88 newly evaluated true circle labels per family. The sampler changes its omitted-term cutoff, time cutoff, quadrature order, primitive series and interval precision. The requested-error construction has a termination proof in an unbounded-resource mathematical model; no practical runtime guarantee is asserted for extreme requests.
The validated theta/Mellin labels are frozen at 64 nodes per conductor. The program scans degrees 0–63 and accepts a requested rational tolerance only if its exact budget has strict slack. Otherwise it refuses. The smallest computed base budgets are about 7.29 × 10⁻¹⁷, 1.74 × 10⁻¹⁶ and 3.37 × 10⁻¹⁶ for conductors 3, 4 and 5; these are limits of this certificate, not impossibility results for approximation.
The preserved fixed-label API retains its floor; the fresh adaptive driver supplies the additional sampling and node-count choices. Both are finite linear readouts with a documented objective, not a trained quantum architecture or an arithmetic spectral determinant. They supply neither zero-location control nor a proof of GRH. No new Lean theorem is claimed.