Existence of positive tensor rank decompositions for unital symmetric real tensors

About 4 years old · traced to

Let PabcP_{abc} be a unital symmetric real tensor, meaning that its indices are symmetric and its unit conditions are P1ab=Pa1b=Pab1=δabP_{1ab}=P_{a1b}=P_{ab1}=\delta_{ab}. A positive tensor rank decomposition is a representation

Pabc=∑i=1RβipaipbipciP_{abc}=\sum_{i=1}^R \beta_i p_a^i p_b^i p_c^i

with positive coefficients and real vectors pip^i. Existence conjecture. Every unital symmetric real tensor has a positive tensor rank decomposition. Numerical experiments for N=3,4,5N=3,4,5 found such decompositions for 1,000 random tensors, but no proof is given, so the general assertion remains open.

References

Primary source

Dennis Obster, “Tensors and Algebras: An Algebraic Spacetime Interpretation for Tensor Models”, arXiv:2203.03633 (2023).

Progress summary

Refreshed
Claimed progress

A reader-submitted argument claims to prove the conjecture, but no independent verification has been found.

The conjecture asks whether every unital symmetric real tensor admits a decomposition using positive coefficients. The catalogued literature had not reported a resolution as of March 2022.

Community submission (unverified; posted 2026-09-05)

A submitted proof reduces the problem to matching prescribed mass, mean, covariance, and third moments. It uses paired atoms to realize each cubic term while preserving the first two moments, then adds symmetric atoms to complete the covariance and a final atom to restore total mass. The argument claims this yields the required positive decomposition in every finite dimension, but it has not been independently checked.

Current status (as of September 2026): The conjecture remains without an independently verified proof; a community submission claims a complete proof, so the asserted resolution is unconfirmed.

Sources

Solutions 1

ProofAn AI-assisted constructive proof of Obster’s positive-decomposition conjecture using finite moment matching, with strictly positive unit contractions.See full solutionHide full solution

Proof of Conjecture 4.7 in finite dimension.

Let P∈Sym⁡3(RN)P\in\operatorname{Sym}^3(\mathbb R^N), N≥1N\ge1, with unit α\alpha, meaning P(α,u,v)=⟨u,v⟩P(\alpha,u,v)=\langle u,v\rangle. We construct P=∑iϕi⊗3P=\sum_i\phi_i^{\otimes3} with ⟨α,ϕi⟩>0\langle\alpha,\phi_i\rangle>0, the positivity condition in Obster, equations (4.9)–(4.10). Minimal length is unnecessary (footnote 11).

Choose orthonormal coordinates with α=re0\alpha=re_0, r>0r>0. Then Q=rPQ=rP has unit e0e_0. If Q=∑iηi⊗3Q=\sum_i\eta_i^{\otimes3} with (ηi)0>0(\eta_i)_0>0, then ϕi=r−1/3ηi\phi_i=r^{-1/3}\eta_i decomposes PP, with α⋅ϕi=r2/3(ηi)0>0\alpha\cdot\phi_i=r^{2/3}(\eta_i)_0>0. Hence assume α=e0\alpha=e_0.

Put d=N−1d=N-1. The unit fixes P000=1P_{000}=1, P00a=0P_{00a}=0, and P0ab=δabP_{0ab}=\delta_{ab} for a,b>0a,b>0; write the remaining block as TT. It suffices to construct positive weighted atoms satisfying

∑iwi=1,∑iwixi=0,∑iwixixiT=Id,∑iwixi⊗3=T,\sum_iw_i=1,\qquad \sum_iw_ix_i=0,\qquad \sum_iw_ix_ix_i^T=I_d,\qquad \sum_iw_ix_i^{\otimes3}=T,

because ϕi=wi1/3(1,xi)\phi_i=w_i^{1/3}(1,x_i) then works. For d=0d=0, use one atom.

By polarization, write T=∑j=1mcjvj⊗3T=\sum_{j=1}^m c_jv_j^{\otimes3}, allowing m=0m=0. Choose

ε=12(1+m+∑j∥vj∥2),C=Id−ε∑jvjvjT.\varepsilon=\frac1{2(1+m+\sum_j\|v_j\|^2)},\qquad C=I_d-\varepsilon\sum_jv_jv_j^T.

Thus C≻0C\succ0 and mε<1/2m\varepsilon<1/2. For each jj, set

sj=cj/ε,aj=1+max⁡(sj,0),bj=1+max⁡(−sj,0).s_j=c_j/\varepsilon,\quad a_j=1+\max(s_j,0),\quad b_j=1+\max(-s_j,0).

Place weights

εaj(aj+bj) at ajvj,εbj(aj+bj) at −bjvj.\frac{\varepsilon}{a_j(a_j+b_j)}\ \text{at }a_jv_j,\qquad \frac{\varepsilon}{b_j(a_j+b_j)}\ \text{at }-b_jv_j.

These are positive. This pair has mass ε/(ajbj)\varepsilon/(a_jb_j), first moment zero, second moment εvjvjT\varepsilon v_jv_j^T, and third moment

ε(aj−bj)vj⊗3=cjvj⊗3.\varepsilon(a_j-b_j)v_j^{\otimes3}=c_jv_j^{\otimes3}.

Together the pairs therefore have mass M≤mε<1/2M\le m\varepsilon<1/2, mean zero, second moment Id−CI_d-C, and third moment TT.

Diagonalize C=∑k=1dλkukukTC=\sum_{k=1}^d\lambda_ku_ku_k^T, with λk>0\lambda_k>0 and orthonormal uku_k. Choose A>0A>0 with ∑kλk/A2<1−M\sum_k\lambda_k/A^2<1-M. Add atoms ±Auk\pm Au_k, each with weight λk/(2A2)\lambda_k/(2A^2). They supply second moment CC, with zero first and third moments. Put the remaining positive mass at the origin. All required moments now hold. ∎

The existence statement also follows from established moment theory; see Blekherman–Fialkow (2020), Corollary 2.2. The construction above is self-contained.

Developed with OpenAI Codex and reviewed by Claude Opus 5.