2026-11-20. Research obstruction; no novelty or full-conjecture claim.
After normalizing a positive final diagonal to one, every PSD matrix can be written
A(u) = [[R+u² vv*, uv], [uv*, 1]],
at u=1, with R PSD. This suggests induction on the order via a polynomial in t=u². All A(u) are PSD for real u: if R=VV*, use Gram rows (V_i,u v_i) followed by (0,...,0,1).
A tempting sufficient claim is that every coefficient of
per(A(u))² − per(|A(u)|²)
is nonnegative. Its constant coefficient is precisely the lower-order Chollet gap. The other coefficients might then have supplied an all-order induction. However the coefficientwise claim fails.
Use real data
V = [[−1,−1], [1,−1], [−1,0]], v=(1,−1,−1),
R = [[2,0,1], [0,2,−1], [1,−1,1]].
The exact polynomials in t=u² are
per(A)=8+8t+6t²+24t³,
per(|A|²)=24+40t+168t²+112t³+110t⁴+12t⁵+24t⁶.
Their difference is
40+88t−8t²+368t³+310t⁴+276t⁵+552t⁶.
The coefficient of t² is negative. border_certificate.py independently enumerates the 24 permutation terms using integer polynomial arithmetic and asserts every coefficient.
This is NOT a violation of Chollet: for t≥0 the difference equals
40(1−t²/10)² + 88t + 368t³ + (1548/5)t⁴ + 276t⁵ + 552t⁶,
which is positive. Thus a full induction would need to retain compensation across polynomial degrees, rather than proving coefficientwise positivity in this basis.
This is a counterexample only to the auxiliary proof strategy developed during this research. No claim is made that the auxiliary observation is previously unpublished, or that it resolves a named open problem.