The Comon conjecture asserts that the tensor decomposition rank and symmetric tensor rank
of a symmetric tensor are equal (Comon et
al. 2008). Writing these ranks as and
, respectively, the inequality
holds because every symmetric
tensor decomposition is also an unrestricted
tensor decomposition.
An initial proposed counterexample over the complex numbers (Shitov 2018) contained an error in the lower bound on its symmetric tensor rank (Draisma 2024). However, the conjecture was subsequently refuted over the real numbers by a fourth-order symmetric tensor of dimension 208 with tensor decomposition rank 761 and symmetric tensor rank 762 (Shitov 2020). Further development of this construction yielded a sixth-order counterexample over the real numbers (Wang and Seigal 2023).
Third-order counterexamples extend the refutation to every infinite field of characteristic different from 2 and 3, including both the real and complex numbers (Shitov 2024). A later explicit fourth-order symmetric tensor over the complex numbers has dimension 2430, tensor decomposition rank 9603, and symmetric tensor rank at least 9604 (Shitov 2026).
A smaller third-order symmetric tensor has rational entries,
size
, and ranks
and
over the rational,
real, and complex numbers
(Lovitz 2026). Its tensor direct sums also yield
larger gaps between the two ranks. Over both the real
and complex numbers, Li (2026a) proved that two
copies have
|
(1)
|
while three copies have
|
(2)
|
These tensors have sizes and
, respectively.
Consequently, the difference
can be 2 or 3, not only 1.
The fourth-order complex counterexample and the smaller third-order example both involved mathematical AI assistance. In the former, GPT-5.6 SOL helped simplify the construction (Shitov 2026). For the latter, GPT-6 Astra was used throughout the project, including to obtain an initial proof that Lovitz (2026) subsequently verified and simplified.
Li (2026b) reports a smaller fourth-order counterexample over the real numbers. For , the tensor
defined by
|
(3)
|
is a Hankel tensor, where is the Kronecker delta.
Its reported tensor decomposition rank
is 10 and its symmetric tensor rank is 11.
The tensor direct sum of
copies has the corresponding ranks
and
for every positive integer
, giving an unbounded rank gap in order four. These
assertions are over the real numbers and do not establish
a counterexample for complex Hankel
tensors.