Learning Hamiltonians at Long Times

Constantin Cedillo Vayson de Pradenne; Hsin-Yuan Huang; Jordan Cotler

REVIEW 2 major objections 2 minor 3 cited by

Learning Hamiltonians at Long Times

T0 review · 2 major / 2 minor · reviewed 2026-06-28 · grok-4.3

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read For broad families of local Hamiltonians, the Hamiltonian is the unique approximately conserved local observable at long times.

desk verdict The paper shows that for broad families of local Hamiltonians, at a random long time t, H is typically the only approximately conserved local observable, which lets you recover it from one snapshot via shadows. read the letter →

arxiv 2606.05690 v1 pith:J5ZIZCLP submitted 2026-06-04 quant-ph

Constantin Cedillo Vayson de Pradenne , Jordan Cotler , Hsin-Yuan Huang This is my paper · ORCID

classification quant-ph

keywords Hamiltonianlearninglocalobservablesconservedquantitiesclassicalshadowsquantumequilibrationlong-timedynamicsmany-bodyHamiltonians

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper establishes that for broad families of local Hamiltonians, at a typical long evolution time t, any normalized sum of local observables orthogonal to the Hamiltonian has a commutator with the time-evolution operator whose squared Frobenius norm per dimension is at least inverse polynomial in the number of qubits. This makes the Hamiltonian the only local quantity that is approximately conserved under the dynamics. The uniqueness directly enables an efficient recovery procedure that treats H as the approximate null vector of a matrix assembled from classical shadow measurements on random product states. A corollary is a weak equilibration statement: the infinite-temperature autocorrelation functions of all local observables orthogonal to H must decay by at least an inverse-polynomial factor.

What carries the argument

The large commutator norm condition ||[U(t), A]||_F that forces any local A orthogonal to H to fail approximate conservation, which in turn identifies H as the unique approximate null vector of the classical-shadow data matrix.

What would settle it

Find a Hamiltonian H from one of the families and a time t such that some normalized local observable A orthogonal to H satisfies (1/2^n) ||[U(t), A]||_F^2 much smaller than 1/poly(n).

Watch

Extended reading notes

Core claim

For broad families of local Hamiltonians, with high probability over H and t, any sum of local observables A that is normalized and orthogonal to H satisfies (1/2^n) ||[U(t), A]||_F^2 >= 1/poly(n). The Hamiltonian is therefore the unique approximately conserved local observable, and we can efficiently recover H, up to scale, as the approximate null vector of a data matrix built from random product-state inputs and classical shadows. As a corollary, we obtain a weak equilibration statement: the infinite-temperature autocorrelation of every sum of local observables orthogonal to H decays by at least an inverse-polynomial amount.

Load-bearing premise

The Hamiltonians belong to broad families for which the high-probability statement over random H and t holds.

Editorial extensions

If this is right

H can be recovered efficiently up to scale from a single long-time evolution via the approximate null vector of the shadow data matrix.
The infinite-temperature autocorrelation of every local observable orthogonal to H decays by at least an inverse-polynomial amount.
No other normalized sum of local observables orthogonal to H remains approximately conserved under U(t).

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Hamiltonian learning remains possible even when the evolution time is arbitrarily large and unknown in advance.
The result gives a quantitative local version of ergodicity for typical local Hamiltonians by bounding how many independent conserved quantities can exist.
The same commutator argument could be tested numerically on small systems to check whether the inverse-polynomial bound is tight.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Desk Editor's Note

read the letter

This paper's central claim is that for broad families of local Hamiltonians, with high probability over H and t, any normalized local observable A orthogonal to H satisfies a lower bound on the squared Frobenius norm of its commutator with U(t) of order 1/poly(n). From this they conclude that H is the unique approximately conserved local observable and give an efficient recovery procedure that builds a data matrix from random product-state inputs and classical shadows to extract H up to scale. A weak equilibration corollary follows for the infinite-temperature autocorrelations of other local sums.

What the work does cleanly is address the long-time single-t regime that is common in experiments but usually avoided in Hamiltonian learning because equilibration erases information. Prior approaches typically need short times or multiple evolution times; here the uniqueness of the conserved quantity is turned into a recovery algorithm that uses only one long snapshot. The use of standard shadow techniques keeps the method concrete and polynomial-time.

The main soft spot is the unspecified scope. The abstract invokes "broad families of local Hamiltonians" and a probability measure over H and t without giving the explicit ensemble or restrictions, so it is unclear whether the statement covers generic physical models or only particular random ensembles, and whether integrable or symmetry-protected cases are excluded by assumption. Without the full proof it is also impossible to check the derivation steps, the precise locality assumptions, or the hidden constants in the poly(n) bound. The stress-test note correctly identifies this as the load-bearing uncertainty.

The paper is aimed at researchers in quantum Hamiltonian learning and many-body dynamics who care about device characterization from limited data. A reader working on long-time quantum information or equilibration would get direct value from the uniqueness result and the single-t algorithm. The thinking is clear and engages the relevant literature on shadows and conserved quantities.

It deserves a serious referee so the families, measure, and proof can be examined in detail. I would send it to peer review.

Referee Report

2 major / 2 minor

Summary. The paper studies learning an unknown n-qubit local Hamiltonian H from the time-evolution operator U = e^{-i H t} at a single, possibly large time t. It proves that for broad families of local Hamiltonians, with high probability over H and t, any normalized sum of local observables A orthogonal to H satisfies (1/2^n) ||[U(t), A]||_F^2 >= 1/poly(n). This establishes H as the unique approximately conserved local observable. The authors provide an efficient recovery algorithm for H (up to scale) using a data matrix constructed from random product-state inputs and classical shadows, and derive a corollary on weak equilibration of infinite-temperature autocorrelations for observables orthogonal to H.

Significance. If the central probabilistic statement holds for relevant physical Hamiltonians, this would be a significant contribution to Hamiltonian learning and quantum dynamics. It addresses the challenging regime of long times where perturbative methods fail, and provides both a uniqueness theorem and a practical recovery method. The explicit construction of the recovery procedure and the equilibration result are strengths. The result could impact experimental quantum simulation by enabling Hamiltonian identification without time-scale restrictions.

major comments (2)

[Abstract] Abstract: The precise definition of 'broad families of local Hamiltonians' and the probability measure over H and t is not provided. This is load-bearing for the central claim, as it determines whether the high-probability uniqueness applies to generic local Hamiltonians or only to specially chosen ensembles that may exclude cases with additional conserved quantities by construction.
[Main result] Main result (theorem establishing the commutator lower bound): The derivation of (1/2^n) ||[U(t),A]||_F^2 >= 1/poly(n) for normalized local A orthogonal to H relies on the unspecified family and measure; without an explicit ensemble (e.g., i.i.d. couplings on a fixed graph), it is unclear whether the uniqueness is derived or partly tautological for the chosen family.

minor comments (2)

[Abstract] Abstract: The poly(n) bound on the commutator could be stated with an explicit degree (e.g., n^{-c} for concrete c) to allow assessment of the decay rate.
[Recovery procedure] Recovery section: The sample complexity and dimension of the data matrix built from product states and shadows should be stated explicitly, including any dependence on n.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive comments highlighting the need for greater clarity on the ensemble and probability measure. We address each major comment below and indicate the revisions we will make to the manuscript.

read point-by-point responses

Referee: [Abstract] Abstract: The precise definition of 'broad families of local Hamiltonians' and the probability measure over H and t is not provided. This is load-bearing for the central claim, as it determines whether the high-probability uniqueness applies to generic local Hamiltonians or only to specially chosen ensembles that may exclude cases with additional conserved quantities by construction.

Authors: We agree that the abstract is high-level and omits the explicit definition. The manuscript defines the families in Section 2 as local Hamiltonians with i.i.d. couplings drawn from a continuous distribution (e.g., standard Gaussian) on a fixed bounded-degree interaction graph, with t drawn uniformly from an interval of length poly(n). This measure ensures the result applies to generic instances without built-in extra conservations. We will revise the abstract to include a brief specification of the ensemble and measure. revision: yes
Referee: [Main result] Main result (theorem establishing the commutator lower bound): The derivation of (1/2^n) ||[U(t),A]||_F^2 >= 1/poly(n) for normalized local A orthogonal to H relies on the unspecified family and measure; without an explicit ensemble (e.g., i.i.d. couplings on a fixed graph), it is unclear whether the uniqueness is derived or partly tautological for the chosen family.

Authors: Theorem 1 in Section 3 explicitly states the ensemble (i.i.d. Gaussian couplings on a fixed graph) and the measure over both H and t. The lower bound is derived via concentration and anticoncentration arguments from random matrix theory, establishing that H is the unique approximately conserved local observable for almost all such instances. The result is not tautological, as the family is broad and the proof shows the absence of extra conservations with high probability. We will add a clarifying remark in the introduction and theorem statement to emphasize the explicit ensemble. revision: partial

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity in the claimed derivation.

full rationale

The paper states a probabilistic theorem: for broad families of local Hamiltonians, with high probability over H and t, any normalized local A orthogonal to H has large commutator norm with U(t). This is presented as a derived statement allowing recovery of H as the approximate null vector of a data matrix from shadows. No equations or steps reduce by construction to self-definition, fitted parameters renamed as predictions, or load-bearing self-citations whose content is unverified. The result is a mathematical existence/probability claim under stated assumptions on the family, self-contained as a proof rather than tautological by input redefinition.

Assumptions & free parameters 0 free parameters · 2 assumptions · 0 invented entities

The central claim rests on standard properties of local operators, the Frobenius norm, and a probabilistic measure over Hamiltonians and times; no data-fitted parameters or new postulated entities are introduced in the abstract.

assumptions (2)

domain assumption Local Hamiltonians are sums of few-body terms with standard operator algebra properties.
Invoked to define the space of observables and the notion of locality used in the commutator bound.
domain assumption Existence of a probability measure over H and t such that the stated lower bound holds with high probability.
The proof statement is conditioned on this measure; the abstract does not detail its construction.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Learning Hamiltonians at Long Times." pith.science (2026). https://pith.science/paper/J5ZIZCLP

@misc{pith2026260605690,
  author       = {Pith},
  title        = {Pith review of: Learning Hamiltonians at Long Times},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/J5ZIZCLP}},
  note         = {Machine review of arXiv:2606.05690}
}

abstract

We study the problem of learning an unknown $n$-qubit Hamiltonian $H$ from $U = e^{-iHt}$ for a single time $t$, where $t$ may be arbitrarily large. For broad families of local Hamiltonians, we prove that, with high probability over $H$ and $t$, any sum of local observables $A$ that is normalized and orthogonal to $H$ satisfies $\tfrac{1}{2^n}\|[U(t),A]\|_F^2 \geq 1/\text{poly}(n)$. The Hamiltonian is therefore the unique approximately conserved local observable, and we can efficiently recover $H$, up to scale, as the approximate null vector of a data matrix built from random product-state inputs and classical shadows. As a corollary, we obtain a weak equilibration statement: the infinite-temperature autocorrelation of every sum of local observables orthogonal to $H$ decays by at least an inverse-polynomial amount.

Figures

Figures reproduced from arXiv: 2606.05690 by the authors.

**Figure 1.** Hamiltonian learning from weak equilibration. (a) Schematic of the mechanism. Under long-time evolution, local observables in the search space generically lose their infinite-temperature autocorrelation, while the Hamiltonian direction is exactly conserved. Classical shadow measurements on random product-state probes produce a data matrix whose approximate null vector recovers H. (b) Weak-equilibration diagnostic… view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Near-Optimal Learning of Local Lindbladians
quant-ph 2026-06 unverdicted novelty 8.0 of 10

Near-optimal algorithm learns local Lindbladians via finite-time probes and classical shadows with Õ(Λ²/ε²) channel uses and matching lower bounds showing dissipative terms block Heisenberg-limited scaling.
Efficient Lindbladian Learning from Constant-Time Pauli Responses
quant-ph 2026-07 conditional novelty 7.0 of 10

Local Pauli responses, inverted through a known generator dictionary, separate coherent from dissipative Lindbladian coefficients and recover all of them to accuracy epsilon from O~(M/epsilon^2) short-time measurements.
Provable learning separation for predicting time-evolution of quantum many-body systems
quant-ph 2026-07 accept novelty 6.0 of 10

A provable exponential quantum-classical learning separation is established for predicting expectation values of time-evolved quantum states under unknown low-intersection Hamiltonians, assuming BQP ⊄ P/poly.

Reference graph

Works this paper leans on

50 extracted references · 5 canonical work pages · cited by 3 Pith papers

[1]

ACM Transactions on Quantum Computing , volume=

Learning quantum processes and Hamiltonians via the Pauli transfer matrix , author=. ACM Transactions on Quantum Computing , volume=. 2024 , publisher=

2024
[2]

Quantum , volume=

Hamiltonian learning via shadow tomography of pseudo-choi states , author=. Quantum , volume=. 2025 , publisher=

2025
[3]

Physical review letters , volume=

Hamiltonian learning and certification using quantum resources , author=. Physical review letters , volume=. 2014 , publisher=

2014
[4]

New Journal of Physics , volume=

Robust online Hamiltonian learning , author=. New Journal of Physics , volume=. 2012 , publisher=

2012
[5]

Proceedings of the 57th Annual ACM Symposium on Theory of Computing , pages=

Learning the structure of any Hamiltonian from minimal assumptions , author=. Proceedings of the 57th Annual ACM Symposium on Theory of Computing , pages=
[6]

Nature Communications , volume=

Practical Hamiltonian learning with unitary dynamics and Gibbs states , author=. Nature Communications , volume=. 2024 , publisher=

2024
[7]

Evans, Robin Harper, and Steven T

Scalable bayesian hamiltonian learning , author=. arXiv preprint arXiv:1912.07636 , year=

work page arXiv 1912
[8]

Nature Physics , volume=

Experimental quantum Hamiltonian learning , author=. Nature Physics , volume=. 2017 , publisher=

2017

Show all 50 references

[9]

arXiv preprint arXiv:2502.11900 , year=

Ansatz-free Hamiltonian learning with Heisenberg-limited scaling , author=. arXiv preprint arXiv:2502.11900 , year=
[10]

Quantum , volume=

Robust and efficient Hamiltonian learning , author=. Quantum , volume=. 2023 , publisher=

2023
[11]

Fan R. K. Chung , title =. 1997 , isbn =

1997
[12]

Czechoslovak Mathematical Journal , volume =

Miroslav Fiedler , title =. Czechoslovak Mathematical Journal , volume =. 1973 , doi =

1973
[13]

Mathematical Research Letters , volume =

Anthony Carbery and James Wright , title =. Mathematical Research Letters , volume =. 2001 , doi =

2001
[14]

Nature Physics , volume =

Hsin-Yuan Huang and Richard Kueng and John Preskill , title =. Nature Physics , volume =. 2020 , doi =

2020
[15]

Tropp , title =

Joel A. Tropp , title =. Foundations of Computational Mathematics , volume =. 2012 , doi =

2012
[16]

Reports on Progress in Physics , volume =

Christian Gogolin and Jens Eisert , title =. Reports on Progress in Physics , volume =. 2016 , doi =

2016
[17]

Lindner , title =

Eyal Bairey and Itai Arad and Netanel H. Lindner , title =. Physical Review Letters , volume =. 2019 , doi =

2019
[18]

Physical Review A , volume =

Christoph Dankert and Richard Cleve and Joseph Emerson and Etera Livine , title =. Physical Review A , volume =. 2009 , doi =

2009
[19]

arXiv preprint arXiv:2604.27838 , year =

Heisenberg-limited Hamiltonian learning without short-time control , author =. arXiv preprint arXiv:2604.27838 , year =. 2604.27838 , archivePrefix =

arXiv
[20]

Mathematical Research Letters , volume =

Carbery, Anthony and Wright, James , title =. Mathematical Research Letters , volume =. 2001 , doi =

2001
[21]

Annals of Physics , volume =

Pierre Pfeuty , title =. Annals of Physics , volume =. 1970 , doi =

1970
[22]

Annals of Physics , volume =

Alexei Kitaev , title =. Annals of Physics , volume =. 2006 , doi =

2006
[23]

Baxter , title =

Rodney J. Baxter , title =. 1982 , isbn =

1982
[24]

Wellner , title =

Adrien Saumard and Jon A. Wellner , title =. Statistics Surveys , volume =. 2014 , doi =

2014
[25]

Horn and Charles R

Roger A. Horn and Charles R. Johnson , title =. 2012 , doi =

2012
[26]

Physical Review Letters , volume =

Peter Reimann , title =. Physical Review Letters , volume =. 2008 , doi =

2008
[27]

Short and Andreas Winter , title =

Noah Linden and Sandu Popescu and Anthony J. Short and Andreas Winter , title =. Physical Review E , volume =. 2009 , doi =

2009
[28]

Short , title =

Anthony J. Short , title =. New Journal of Physics , volume =. 2011 , doi =

2011
[29]

Cory , title =

Nathan Wiebe and Christopher Granade and Christopher Ferrie and David G. Cory , title =. Physical Review A , volume =. 2014 , doi =

2014
[30]

Quantum , volume =

Xiao-Liang Qi and Daniel Ranard , title =. Quantum , volume =. 2019 , doi =

2019
[31]

Short and Terence C

Anthony J. Short and Terence C. Farrelly , title =. New Journal of Physics , volume =. 2012 , doi =

2012
[32]

Short and Andreas Winter , title =

Sandu Popescu and Anthony J. Short and Andreas Winter , title =. Nature Physics , volume =. 2006 , doi =

2006
[33]

Nature Physics , volume =

Jeongwan Haah and Robin Kothari and Ewin Tang , title =. Nature Physics , volume =. 2024 , doi =

2024
[34]

Physical Review Letters , volume =

Hsin-Yuan Huang and Yu Tong and Di Fang and Yuan Su , title =. Physical Review Letters , volume =. 2023 , doi =

2023
[35]

2024 IEEE 65th Annual Symposium on Foundations of Computer Science (FOCS) , pages =

Ainesh Bakshi and Allen Liu and Ankur Moitra and Ewin Tang , title =. 2024 IEEE 65th Annual Symposium on Foundations of Computer Science (FOCS) , pages =. 2024 , doi =

2024
[36]

Proceedings of the 56th Annual ACM Symposium on Theory of Computing (STOC) , year =

Ainesh Bakshi and Allen Liu and Ankur Moitra and Ewin Tang , title =. Proceedings of the 56th Annual ACM Symposium on Theory of Computing (STOC) , year =
[37]

O'Brien and Thomas Schuster , title =

Alicja Dutkiewicz and Thomas E. O'Brien and Thomas Schuster , title =. Quantum , volume =. 2024 , doi =

2024
[38]

Nature Physics , volume=

Sample-efficient learning of interacting quantum systems , author=. Nature Physics , volume=. 2021 , publisher=

2021
[39]

Testing and Learning Structured Quantum

Srinivasan Arunachalam and Arkopal Dutt and Francisco Escudero Guti\'. Testing and Learning Structured Quantum. Communications in Mathematical Physics , volume =. 2026 , doi =

2026
[40]

Efficient and robust estimation of many-qubit

Daniel Stilck Fran. Efficient and robust estimation of many-qubit. Nature Communications , volume =. 2024 , doi =

2024
[41]

Communications in the ACM , volume =

Sitan Chen and Jordan Cotler and Hsin-Yuan Huang and Jerry Li , title =. Communications in the ACM , volume =. 2024 , doi =

2024
[42]

Deutsch , title =

Joshua M. Deutsch , title =. Physical Review A , volume =. 1991 , doi =

1991
[43]

Physical Review E , volume =

Mark Srednicki , title =. Physical Review E , volume =. 1994 , doi =

1994
[44]

Lieb and Derek W

Elliott H. Lieb and Derek W. Robinson , title =. Communications in Mathematical Physics , volume =. 1972 , doi =

1972
[45]

Nielsen and Isaac L

Michael A. Nielsen and Isaac L. Chuang , title =. 2010 , doi =

2010
[46]

G. W. Stewart and Ji-Guang Sun , title =. 1990 , isbn =

1990
[47]

Learning Conservation Laws in Unknown Quantum Dynamics

Zhan, Yongtao and Elben, Andreas and Huang, Hsin-Yuan and Tong, Yu. Learning Conservation Laws in Unknown Quantum Dynamics. PRX Quantum. 2024. doi:10.1103/PRXQuantum.5.010350. arXiv:2309.00774

2024 doi
[48]

Integration with respect to the

Beno\^. Integration with respect to the. Communications in Mathematical Physics , volume =. 2006 , doi =

2006
[49]

arXiv:2510.08499 , year=

Quantum probe tomography , author=. arXiv:2510.08499 , year=
[50]

Physical Review Letters , volume =

Quantum Algorithms for Testing Hamiltonian Symmetry , author =. Physical Review Letters , volume =. 2022 , doi =

2022

Pith tools

Reviewed June 28, 2026 · model on record in the stance chip above.

[1] [1]

ACM Transactions on Quantum Computing , volume=

Learning quantum processes and Hamiltonians via the Pauli transfer matrix , author=. ACM Transactions on Quantum Computing , volume=. 2024 , publisher=

2024

[2] [2]

Quantum , volume=

Hamiltonian learning via shadow tomography of pseudo-choi states , author=. Quantum , volume=. 2025 , publisher=

2025

[3] [3]

Physical review letters , volume=

Hamiltonian learning and certification using quantum resources , author=. Physical review letters , volume=. 2014 , publisher=

2014

[4] [4]

New Journal of Physics , volume=

Robust online Hamiltonian learning , author=. New Journal of Physics , volume=. 2012 , publisher=

2012

[5] [5]

Proceedings of the 57th Annual ACM Symposium on Theory of Computing , pages=

Learning the structure of any Hamiltonian from minimal assumptions , author=. Proceedings of the 57th Annual ACM Symposium on Theory of Computing , pages=

[6] [6]

Nature Communications , volume=

Practical Hamiltonian learning with unitary dynamics and Gibbs states , author=. Nature Communications , volume=. 2024 , publisher=

2024

[7] [7]

Evans, Robin Harper, and Steven T

Scalable bayesian hamiltonian learning , author=. arXiv preprint arXiv:1912.07636 , year=

work page arXiv 1912

[8] [8]

Nature Physics , volume=

Experimental quantum Hamiltonian learning , author=. Nature Physics , volume=. 2017 , publisher=

2017