Unknown-structure Hamiltonian learning from Gibbs states at all temperatures

Unsolved ID op_845158213592821f Last edited 16 September 2026
Edit

Problem

Can every bounded-degree local Hamiltonian with unknown interaction support be learned efficiently from copies of its Gibbs state at any fixed inverse temperature?

Fix constants \(k\) and \(g\). Let \(\mathcal P_{n,k}\) be the nonidentity \(n\)-qubit Pauli strings of weight at most \(k\), and consider

\begin{equation} H=\sum_{P\in\mathcal P_{n,k}}a_PP, \quad |a_P|\leq1, \quad \max_j|\{P:a_P\neq0,\ j\in\operatorname{supp}(P)\}|\leq g, \qquad \rho_\beta=\frac{e^{-\beta H}}{\operatorname{Tr}(e^{-\beta H})}. \tag{1} \end{equation}

The nonzero Pauli terms are unknown, while \(\beta>0\) is known and fixed independently of \(n\). Can independent copies of \(\rho_\beta\) in Eq. (1) be used to output coefficients satisfying \(\max_P|\widehat a_P-a_P|\leq\varepsilon\) with probability at least \(2/3\), using \(\operatorname{poly}(n,1/\varepsilon)\) copies and classical time for every fixed \(\beta\)?

Source

This precise formulation is editor wording based on the unresolved direction and limitations documented in the cited primary literature [Narayanan25][Lewis26]; it is not presented as a verbatim conjecture of those authors.

Progress

  • With a supplied bounded-degree list of \(r\) Pauli terms, learning at every fixed temperature is possible. Narayanan’s Theorem 1.6 gives

    \begin{equation} N=O\!\left(r^6(1/\varepsilon)^{O(\beta^2)} +\frac{\log r}{\beta^2\varepsilon^2}\right) \tag{2} \end{equation}

    copies and polynomial computational time for fixed locality, interaction degree, and \(\beta\). The supplied list is essential to this theorem; it is parameter learning rather than unknown-structure learning. [Narayanan25]

    The displayed definitions, constraints, and target bounds are recorded in Eqs. (2).

  • For geometrically local Hamiltonians with a known interaction dictionary, Chen, Anshu, and Nguyen obtain the sharper lattice sample bound

    \begin{equation} N=\widetilde O\!\left(\frac{e^{\operatorname{poly}(\beta)}}{\beta^2\varepsilon^2}\right)\log(n/\delta), \tag{3} \end{equation}

    where \(\delta\) is the failure probability and the tilde suppresses logarithmic factors. Their all-temperature results do not remove the supplied-structure assumption. [Chen25]

    The displayed definitions, constraints, and target bounds are recorded in Eqs. (3).

  • Unknown-structure learning is now solved at sufficiently high temperature. Lewis, Tang, and Wright’s Theorem 5.18 proves, under the normalization above,

    \begin{equation} \beta\leq\frac{1}{1000e^6(2kg+1)^8} \quad\Longrightarrow\quad N=O\!\left(\frac{\log(n/\delta)}{\beta^2\varepsilon^2}\right), \tag{4} \end{equation}

    with classical runtime \(O(n^k\operatorname{poly}(g)\log(n/\delta)/(\beta^2\varepsilon^2))\). Thus neither high-temperature structure learning nor all-temperature known-structure learning should be listed as open. [Lewis26]

    The displayed definitions, constraints, and target bounds are recorded in Eqs. (4).

  • Section 1.4 of the 29 June 2026 preprint explicitly leaves all-temperature structure learning open. Supplying all \(O(n^k)\) candidate Pauli terms to [Narayanan25] does not immediately solve the problem: their candidate interaction graph no longer has bounded degree. [Narayanan25][Lewis26]

Comment

The unresolved conjunction is unknown interaction support, arbitrary fixed positive temperature, and polynomial resources. Constants and polynomial exponents may depend on \(k\), \(g\), and \(\beta\); this question does not demand efficient scaling as the temperature approaches zero with system size. Lewis, Tang, and Wright explicitly leave all-temperature Gibbs-state structure learning open in Section 1.4; no impossibility result is asserted here.

References

[Narayanan25]
S. Narayanan, "Improved algorithms for learning quantum Hamiltonians, via flat polynomials," in Proceedings of the Thirty Eighth Conference on Learning Theory, Proceedings of Machine Learning Research 291, 4360–4385 (2025). Proceedings;linkarXiv
[Chen25]
C.-F. Chen, A. Anshu, and Q. T. Nguyen, "Learning quantum Gibbs states locally and efficiently," arXiv preprint (2025), version 1, 3 April 2025.arXiv
[Lewis26]
L. Lewis, E. Tang, and J. Wright, "Learning the structure of open quantum systems," arXiv preprint (2026), version 1, 29 June 2026.arXiv

Page edit log

  • Record created
  • Last edited
  • Revisions2

View the full history on GitHub

Your contribution is welcome!

Found progress, a correction, or a resolution? Edit this record on GitHub and open a pull request, or report an update with the primary sources. The proposal page explains the available submission route; see the contribution guide for details.

Cite this page

“Unknown-structure Hamiltonian learning from Gibbs states at all temperatures,” Quantum Information and Quantum Computation Open Problem Zoo (QIQCOP Zoo), ID op_845158213592821f, accessed 2026-09-16.

Use the Cite button above for BibTeX and the permanent link.

Cite this problem

Please also cite the primary sources listed under References. Cite this page for the statement, status, and stable identifier.

BibTeX

@incollection{qiqcop_op_845158213592821f,
  title = {Unknown-structure Hamiltonian learning from Gibbs states at all temperatures},
  booktitle = {Quantum Information and Quantum Computation Open Problem Zoo (QIQCOP Zoo)},
  year = {2026},
  howpublished = {\url{https://qiqc-op.com/problem/op_845158213592821f/}},
  note = {Stable ID op_845158213592821f; status: Unsolved; accessed 2026-09-16}
}

Plain text

“Unknown-structure Hamiltonian learning from Gibbs states at all temperatures,” Quantum Information and Quantum Computation Open Problem Zoo (QIQCOP Zoo), https://qiqc-op.com/problem/op_845158213592821f/, ID op_845158213592821f, accessed 2026-09-16.

Share this problem

Permanent link

Identifiers

op_845158213592821f
01M2M9FC22Z4C4YCDYT9XZCB59