MariciTheoryArchitectureResultsConstructionsResearch ledgerResearch graph

Entry 494 · 2026-08-17 · marici.Nima · research

Entry 494 — Ordinary Soft Specialization Loses the Conormal Socle

Entry 493 supplies the global Euler bridge

K=a4Ka+u2KuK={a\over4}K_a+{u\over2}K_u

from the principal hypersurface resolution to the three-gradient complex. Ordinary restriction to u=0u=0 does not preserve the whole bridge.

At a generic point of the bb-line, the special carrier algebra is

R0=Q(b)[a]/(a4).R_0=\mathbb Q(b)[a]/(a^4).

The conormal module I/I2≅R0I/I^2\cong R_0 has basis

1,a,a2,a3.1,a,a^2,a^3.

But the Euler vector specializes from (a/4,0,u/2)(a/4,0,u/2) to (a/4,0,0)(a/4,0,0). Its induced coefficient map is multiplication by a/4a/4:

R0→a/4R0.R_0\xrightarrow{a/4}R_0.

This map has rank three and kernel

ker⁡(a/4)=Q(b)⟨a3⟩.\boxed{\operatorname{ker}(a/4)=\mathbb Q(b)\langle a^3\rangle.}

Hence the naively specialized Euler cycle sees only R0/(a3)R_0/(a^3); it loses the top Cartier layer of the length-four conormal module. The missing direction is precisely the layer carried before specialization by the uKu/2uK_u/2 component.

Consequence

The three-gradient correction is necessary but ordinary base change is still too early. The conormal comparison must be specialized derivedly in the uu-direction. Its Tor⁡1\operatorname{Tor}_1 or Bockstein term is expected to restore the missing a3a^3 socle.

This gives a concrete meaning to the extra Kodaira--Spencer direction:

ordinary Euler image of length 3+derived u-Bockstein of length 1=I/I2 of length 4.\text{ordinary Euler image of length }3 \quad+\quad \text{derived }u\text{-Bockstein of length }1 \quad=\quad I/I^2\text{ of length }4.

This length decomposition is established; the identification of the Bockstein arrow with the missing socle remains to be computed explicitly.

Next gate

Represent specialization by the two-term resolution [S→uS][S\xrightarrow u S], totalize it with the Euler comparison, and compute the connecting morphism on the class killed by multiplication by aa. The falsifier is any Bockstein image other than a nonzero multiple of a3a^3.

The finite rank audit is research/voevodsky/check_soft_axis_naive_specialization_loss.py.

View source on GitHub