<!-- All digits below are copied verbatim from the certified production run of this
     work (2026-06-26); none were recomputed for this file. -->

# Threshold banana (1,1,1,9) — non-unipotent connection coefficient, closed form

Three-loop banana K3 at masses-squared (1,1,1,9), i.e. on the additive
normal-threshold coalescence sqrt(M4) = sqrt(M1)+sqrt(M2)+sqrt(M3). The order-4
Picard–Fuchs operator has threshold indicial polynomial

    18432 (rho-1)^3 (2 rho - 3)   =>   exponents {1, 1, 1, 3/2}

so the local monodromy at t = p^2 -> 0 is NON-unipotent (eigenvalue
e^{3 pi i} = -1 on the t^{3/2} branch; order 2). The single new function-level
datum is the connection coefficient c_{3/2} of the holomorphic period onto that
branch,

    varpi_0(t) = A0 Phi^(0)(t) + A1 Phi^(1)(t) + A2 Phi^(2)(t) + c_{3/2} Phi_{3/2}(t).

## Closed form (function-level, closed)

    c_{3/2} = - sqrt(3) / (36 pi)            equivalently  432 pi^2 c_{3/2}^2 = 1

PSLQ/LLL relations (primary relation height 36; squared relation coefficient 432):
  - 36 c_{3/2} pi + sqrt(3) = 0                      residual 194.2 d
  - -2 log|c_{3/2}| - 2 log pi - 4 log 2 - 3 log 3 = 0   residual 196.1 d
  - 432 pi^2 c_{3/2}^2 = 1                           residual 192.8 d

Where the constants come from: in the Hankel-tail derivation (FeynmanGeometry Sec. 2.4; this bundle's threshold_hankel_tail.py) the
non-oscillatory tail of J0(y)^3 J0(3y) is -sqrt(3)/(18 pi^2) y^-3 + ..., where 1/sqrt(3) = (m1 m2 m3 m4)^(-1/2) and 1/pi^2
comes from the four Bessel asymptotics; the Mellin transform of H0^(1) at lambda = -1 supplies 2^-3 Gamma(-1/2)^2 = pi/2, giving
-sqrt(3)/(36 pi). One may also read 36 = (sum_i sqrt(M_i))^2, the normal threshold; at this mass point the two readings coincide. The Chowla–Selberg Gamma(a/15)
CM-period ring of the level-15 K3 motive is PSLQ-NEGATIVE (zero coefficients in
every relation).

## Certified value (Route B, Julia/Arb, 768-bit balls, 195 digits)

c_{3/2} =
-0.01531469153949422359753684717536026226103851334351779591109350056363973065192416096778833413397696914465827970348884947342805812313051692629590881161944422277729854364955194835822349292058384942464

Im part rigorously |Im| < 3.3e-198 (Arb ball). Ball-certified accuracy 649-650 bits.

Sign convention: the sign is branch-dependent — the order-2 monodromy negates the
t^{3/2} branch, so |c_{3/2}| = sqrt(3)/(36 pi) is the convention-robust statement.
(Route A's mpmath e^{-i pi} lands on arg = +pi via roundoff, so Route A reports +|c_{3/2}|.)

## How it was extracted (the projector trick)

Exact spectral projector onto the -1 eigenspace of the threshold monodromy M_0:

    P_{-1} = -(M_0 - 1)^3 / 8 ,    P_{-1} . varpi_0_state = c_{3/2} . Phi_{3/2}_state  (exactly)

(M_0 - 1)^3 annihilates the rank-3 unipotent log tower (nilpotent index 3) and acts
as -8 on Phi_{3/2}, killing algebraically the ~1e4 unipotent contamination that had
walled the earlier 4x4 lu_solve extraction at 3.3-3.7 digits.

Two independent transport routes, sharing ONLY the exact-rational PF operator:
  - Route A: Python/mpmath, s-chart, holomorphic-period multinomial seed
    (Boenisch-Fischbach-Klemm-Nega-Safari, arXiv:2008.10574), small CCW circle
    |s|=s_dec around s=0. s_dec-independence 102.5 d, two-base-point agreement 95.9 d (110 working digits);
    re-run at 210 and 225 working digits the two runs agree to 204 d (internal floor 204 d).
  - Route B: Julia/Arb ball-arithmetic theta-companion transport, z=-1/s chart,
    operator-recursion seed (0/155 mismatches vs the multinomial seed), large CW
    circle around all six finite singularities. s_dec in {1/4, 7/20} agree to all
    195 ball-certified digits; row consistency 231.2 d.
Two-method agreement A vs B: 195 d, every ball-certified digit of Route B (the mpmath route re-run at 210/225 working digits;
the earlier 80.7 d figure was the record's 80-digit printout cap, not the route's accuracy).

The non-unipotency itself was certified four independent ways (indicial polynomial,
Frobenius monodromy, parity, and series-contrast checks), all from the same
exact-rational Picard–Fuchs operator.

## Honest caveats

  - No external Feynman oracle: a distinct-mass AMFlow evaluation was not
    available for this configuration. Cross-validation = two-independent-method
    gate + s_dec-independence + 4-row consistency (>= 100 d) + height-36 PSLQ
    closure at 195 d.
  - Both routes share the exact-rational PF operator (built once via CRT +
    rational reconstruction, annihilation residual exactly 0).
