Sign conjecture for Steiner distance hyperdeterminants

About 2 years old · traced to

Let TT be a tree on nn vertices, let MM be its order-kk Steiner distance hypermatrix, and write sgn⁡(x)\operatorname{sgn}(x) for 11 if x>0x>0, −1-1 if x<0x<0, and 00 if x=0x=0.

Sign conjecture. For any tree on nn vertices,

sgn⁡(det⁡(M))=(−1)n−1\operatorname{sgn}(\det(M))=(-1)^{n-1}

when this quantity is nonzero.

The conjecture appears in earlier work and has been checked numerically for (k,n)=(4,4)(k,n)=(4,4), (4,5)(4,5), and (6,4)(6,4). It is trivially true for n≤3n\leq 3 and odd kk, and follows from Graham–Pollak when k=2k=2; its general status is not established in the supplied text.

References

Primary source

Joshua Cooper and Zhibin Du, “Determinants of Steiner Distance Hypermatrices”, arXiv:2505.10501 (2025).

Additional references

2 papers in this index state this conjecture (2024–2025). The statement above is taken from the most recent of them; the others are arXiv:2403.02287.

Progress summary

Refreshed
Claimed solved

An unverified posted proof claims to settle the sign conjecture for every tree and every order, while the latest published paper still leaves it open.

The conjecture asserts that the sign of the Steiner-distance hyperdeterminant of a tree is (−1)n−1(-1)^{n-1} whenever it is nonzero. Cooper and Du’s 2025 paper proves related tree-independence of the determinant’s value, but explicitly retains this sign statement as an open conjecture.

Known results

  • Trivial for n≤3n\leq 3 and odd kk.
  • Graham–Pollak proves the case k=2k=2.
  • Cooper and Du prove the case n=2n=2 for every kk.
  • Numerical checks cover (k,n)=(4,4),(4,5),(6,4)(k,n)=(4,4),(4,5),(6,4); even kk and n≥2n\geq2 also give nonzero determinants.

Posted attempt (posted August 21, 2026)

A reader-written argument claims a complete proof for all k≥2k\geq2 and all trees when the determinant is nonzero, using edge-cut coordinates, separated gradient equations, and a resultant sign computation. The attempt has not been independently verified.

Current status (as of August 2026): The conjecture is proved only in the recorded special cases, while a complete posted proof claim is unverified.

Sources

Solutions 1

ProofThis solution needs a summarySee full solutionHide full solution

Let TT be a tree on n≥1n\geq1 vertices, and let MM be its order-kk Steiner distance tensor, where k≥2k\geq2. We prove that whenever det⁡(M)≠0\det(M)\ne0,

sgn⁡det⁡(M)=(−1)n−1.\operatorname{sgn}\det(M)=(-1)^{n-1}.

This settles the sign conjecture of Cooper–Du, Conjecture 1. The key is to use edge-cut coordinates: all but one of the gradient equations then separate into one-variable equations. Their nonreal solutions occur in conjugate pairs, while the unique real solution contributes a positive factor.

We use the cut identity from Cooper–Du, Theorem 2.1 and the normalized resultant and Poisson formula recalled in Zheng, Section 2, Theorems 2, 4 and 5. The known vanishing for odd order when n≥3n\geq3, and the case n=2n=2 for every order, are already recorded in Cooper–Du; the latter is their Proposition 4.1. Thus it suffices to prove the assertion for even k≥2k\geq2 and n≥2n\geq2. For n=1n=1, every entry is zero.

Cut coordinates and normalization

Put m=n−1m=n-1 and d=k−1d=k-1. Its entries count the edges of the smallest subtree containing the indexed vertices, with repetitions allowed. Write its associated form as

F(x)=∑i1,…,ik=1nMi1…ikxi1⋯xik.F(x)=\sum_{i_1,\ldots,i_k=1}^{n} M_{i_1\ldots i_k}x_{i_1}\cdots x_{i_k}.

The symmetric tensor determinant used here is

det⁡(M)=Res⁡(1k∂F∂x1,…,1k∂F∂xn),\det(M)=\operatorname{Res} \left(\frac1k\frac{\partial F}{\partial x_1}, \ldots,\frac1k\frac{\partial F}{\partial x_n}\right),

with the normalization Res⁡(x1d,…,xnd)=1\operatorname{Res}(x_1^d,\ldots,x_n^d)=1.

Root the tree at vertex nn, and order the other vertices so that descendants precede ancestors. For each nonroot vertex ii, let aia_i be the sum of xvx_v over its rooted subtree, and set s=∑vxvs=\sum_vx_v. The linear map

x⟼(a1,…,am,s)x\longmapsto(a_1,\ldots,a_m,s)

has a lower triangular matrix with diagonal entries all equal to 11. In particular, its determinant is 11.

An edge belongs to the subtree spanned by a tuple precisely when that tuple meets both sides of its cut. Summing over tuples therefore gives the known cut identity

Φ(a,s)=msk−∑i=1m(aik+(s−ai)k),\begin{aligned} \Phi(a,s)&=m s^k\\ &\quad-\sum_{i=1}^{m}\bigl(a_i^k+(s-a_i)^k\bigr), \end{aligned}

where Φ\Phi is FF expressed in the new coordinates. Its normalized gradient, in the variable order (a1,…,am,s)(a_1,\ldots,a_m,s), consists of

Pi=(s−ai)d−aid(1≤i≤m)P_i=(s-a_i)^d-a_i^d\qquad(1\leq i\leq m)

and

P0=msd−∑i=1m(s−ai)d.P_0=m s^d-\sum_{i=1}^{m}(s-a_i)^d.

Both the variable change and the corresponding inverse-transpose change of gradient equations have determinant 11. The composition rule for normalized homogeneous resultants consequently gives the exact equality

det⁡(M)=Res⁡(P1,…,Pm,P0).\det(M)=\operatorname{Res}(P_1,\ldots,P_m,P_0).

The order of the equations matches the order of the variables, including P0P_0 for the last variable ss.

The product and its sign

Because kk is even, dd is odd. At s=0s=0 we have Pi=−2aidP_i=-2a_i^d for 1≤i≤m1\leq i\leq m, so these leading forms have no common nonzero zero. Their normalized resultant is

R∞=Res⁡(−2a1d,…,−2amd)=(−2)mdm−1.\begin{aligned} R_\infty &=\operatorname{Res}(-2a_1^d,\ldots,-2a_m^d)\\ &=(-2)^{m d^{m-1}}. \end{aligned}

After setting s=1s=1, the first mm equations separate as p(ai)=0p(a_i)=0, where

p(z)=(1−z)d−zd.p(z)=(1-z)^d-z^d.

This polynomial has degree dd and dd distinct roots. Indeed, z=0z=0 is not a root, and its roots are exactly

z=11+ζ,ζd=1.z=\frac{1}{1+\zeta},\qquad \zeta^d=1.

The denominators are nonzero because dd is odd, and distinct ζ\zeta give distinct roots. Moreover, the only real root is z=1/2z=1/2: on the real line, the odd power map is injective, so (1−z)d=zd(1-z)^d=z^d forces 1−z=z1-z=z.

Let R\mathcal R be this root set. The affine common zeros of P1,…,PmP_1,\ldots,P_m are exactly Rm\mathcal R^m, each with multiplicity one: their Jacobian is diagonal with nonzero entries p′(ai)p'(a_i). Applying the Poisson formula in the last variable ss gives

det⁡(M)=(−2)mdm∏a∈Rmh(a),h(a)=m−∑i=1m(1−ai)d.\begin{aligned} \det(M) &=(-2)^{m d^m} \prod_{a\in\mathcal R^m}h(a),\\ h(a)&=m-\sum_{i=1}^{m}(1-a_i)^d. \end{aligned}

Assume det⁡(M)≠0\det(M)\ne0. Every factor in this product is then nonzero. Complex conjugation preserves Rm\mathcal R^m and pairs each nonreal tuple aa with a distinct tuple a‾\overline a. Since hh has real coefficients, each such pair contributes

h(a)h(a‾)=∣h(a)∣2>0.h(a)h(\overline a)=|h(a)|^2>0.

There is just one real tuple, namely (1/2,…,1/2)(1/2,\ldots,1/2), and its factor is

h(1/2,…,1/2)=m(1−2−d)>0.h(1/2,\ldots,1/2)=m(1-2^{-d})>0.

Thus the entire product is positive. Since dmd^m is odd, the prefactor has sign (−1)m(-1)^m. Therefore

sgn⁡det⁡(M)=(−1)m=(−1)n−1,\operatorname{sgn}\det(M)=(-1)^m=(-1)^{n-1},

as required. Together with the previously established odd-order and two-vertex cases, this proves Conjecture 1 for every order k≥2k\geq2 and every tree whenever its Steiner distance hyperdeterminant is nonzero.