2026-11-20. Research notes only. No resolution or novelty claim.
Literature gate
Chollet's 1982 conjecture is per(A∘B) ≤ per(A)per(B) for complex Hermitian PSD matrices. Original source: J. Chollet, American Mathematical Monthly 89 (1982), 57–58, DOI 10.1080/00029890.1982.11995380. Yair Lavi's updated status page https://yair-lavi.com/mincs-list/conjecture-30/index.html explicitly marks the full problem open, while listing orders through six and rank at most three as covered by recent work. Pant–Singh https://arxiv.org/abs/2604.24192 is a primary 2026 source proving structural classes, not the unrestricted assertion. Searches for full proofs/counterexamples found restricted results and disproofs of other permanent conjectures, not a full resolution of Chollet.
The standard Cauchy–Schwarz reduction, attributed already to Chollet, is to B=conjugate(A). For permutation products a_sigma, b_sigma, |sum a_sigma b_sigma| ≤ sqrt(sum |a_sigma|² sum |b_sigma|²). Consequently it suffices to prove per(|A|²) ≤ per(A)². This reduction is not new.
Counterexample reconnaissance
search.py uses Gram matrices V V* to enforce PSD and the Glynn sign formula with exact analytic gradients (floating evaluation) to seek violations of the full universal conjecture. The gradient was checked at one random order-four/rank-at-most-three point by central finite differences, maximum error 2.3e-10; this is a spot check, not universal validation. results.json records 72 starts at selected orders 7–12 with Gram column counts 4–6 (rank upper bounds, not certified final ranks). fullrank.json records another 84 starts at orders 7, 8, 9, 10, and 12, using near-full/full Gram column counts. Five reported ratios exceeded one only at floating-point noise scale; none passed the candidate threshold log(R)>1e-7, and none is a certified violation. These are not proofs, not exhaustive searches, and not claimed dimension-specific mathematical results. Floating candidates would require exact Gram reconstruction and exact permanent certificates before being considered evidence of disproof.
Exact obstruction to a proposed proof route
A tempting stronger claim is that for every correlation matrix A, the ratio
R(t) = per(|(1−t)I+tA|²) / per((1−t)I+tA)²
is nonincreasing on [0,1]. Since R(0)=1, this would prove Chollet. It is false, as follows. This is an obstruction to the strategy, NOT a counterexample to Chollet and not claimed as a previously unknown result.
Let ζ be a primitive sixth root of unity and A_ij=(1+ζ^(i−j))/2 for 0≤i,j<6. This is the Gram matrix of unit vectors (1,ζ^i)/sqrt(2), up to the harmless choice of inner-product convention, and has rank two. A(t)=(1−t)I+tA is PSD for 0≤t≤1.
For a matrix M with unit diagonal, multilinearity gives
(d/dt) per((1−t)I+tM) at t=1
= n per(M) − sum_i per(M with row/column i deleted).
For H(t)=|A(t)|², diagonal entries are one and off-diagonal entries are t²|a_ij|², so H'(1)=2(H−I), where H=|A|².
Here H is the rational circulant matrix with first row
(1, 3/4, 1/4, 0, 1/4, 3/4).
Exact permutation enumeration gives per(H)=4005/256 and each principal five-minor has permanent 51/8. These statements are verified with rational arithmetic by radial_certificate.py.
For completeness, the Gram permanent formula follows by expanding ∏_j(z+ζ^j w)/sqrt(2): in the squared Gaussian norm, monomials z^k w^(n−k) have norm k!(n−k)!. The product is (z6−w6)/8 up to a unit phase, giving per(A)=2·6!/2^6=45/2. Deleting any one factor leaves five-degree coefficients all of modulus one, giving each principal minor permanent
sum_(k=0)^5 k!(5−k)! / 2^5 = 39/4.
Therefore
R'(1)/R(1)
= 2·6[(39/4)/(45/2) − (51/8)/(4005/256)]
= 138/445 > 0.
Continuity gives an interval immediately left of 1 on which R increases. The endpoint ratio itself is only 89/2880<1, so there is no contradiction to Chollet.
What remains
The monotonicity shortcut is ruled out exactly. Numerical search did not find a counterexample. No general positivity argument controlling permutation-phase cancellation has been established. Any future proof cannot simply assert monotonicity of this ratio under identity contraction.