Pith. sign in

REVIEW 3 major objections 5 minor 123 references

Learning to stabilize nonequilibrium phases of matter with active feedback using partial information

T0 review · 3 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Reinforcement-learning agents trained on partial observations learn stochastic feedback policies that drive stabilizer circuits into area-law entangled steady states for any nonzero disentangling bias p, and above a critical information thr

desk verdict RL-trained feedback with partial information in Clifford circuits is a genuinely new mechanism; the finite-size evidence is solid, the thermodynamic-limit claim is a reasonable but unproven extrapolation. read the letter →

arxiv 2508.06612 v1 pith:F455LNAR submitted 2025-08-08 quant-ph cond-mat.dis-nncond-mat.quant-gascond-mat.stat-mechcond-mat.str-el

Giovanni Cemin , Markus Schmitt , Marin Bukov This is my paper
classification quant-phcond-mat.dis-nncond-mat.quant-gascond-mat.stat-mechcond-mat.str-el PACS 03.67.-a
keywords reinforcementlearningentanglementphasesstabilizercircuitsactivefeedbackpartialinformationarea-lawvolume-lawnonequilibriumsteadystates
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

This paper asks how much information a controller needs to stop entanglement from spreading in a many-body quantum system, and reports that a learned feedback strategy needs only partial state information. The authors train reinforcement-learning agents to choose where to apply disentangling gates in (1+1)-dimensional stabilizer circuits of up to 128 qubits, using observations of the stabilizer tableau in which each stabilizer is kept with probability q. Above a p-dependent information threshold, the learned policies produce steady states with area-law entanglement, whereas random and greedy alternatives retain a volume-law phase; with full information, the volume-law phase disappears entirely, and any small positive bias p suffices. The agent's mechanism is not a simple heuristic: it keeps entanglement near zero at approximately equidistant bonds, creating pyramidal profiles that split the system into weakly entangled subsystems. If this extrapolates to larger systems, it means information-driven active feedback can engineer nonequilibrium phases that passive dynamics cannot reach, and it points toward control using only single-shot measurements.

What carries the argument

Two ingredients carry the argument. First, a scalable active-feedback loop: at each turn the agent observes a (possibly partial) stabilizer tableau, outputs a bond location, and an optimally disentangling Clifford gate is applied there; the reward is negative normalized total entanglement, and the policy is a transformer encoder trained with Proximal Policy Optimization. Second, the mechanism the policy discovers: entanglement bottlenecks. By holding certain roughly equidistant bonds at near-zero entanglement, the agent splits the chain into weakly entangled subsystems, which appears in the steady state as self-similar pyramids in the bond-entanglement profile. A simplified rate model with n

What would settle it

Run the same learned-feedback training at N=256 for p=0.05 and inspect the steady state. The central claim predicts the pyramid count doubles to 8 and $\langle S_{\mathrm{tot}}\rangle/\mathcal{N}$ keeps decreasing roughly as $(1-p)/(pN)$. If the pyramid count saturates, the entanglement density plateaus, or a single-cut entropy at mid-chain grows linearly with N, the area-law claim fails.

Watch

Extended reading notes

Core claim

In the stabilizer-circuit model considered here, a bias p selects whether an optimally disentangling two-qubit Clifford gate or a random Clifford gate acts, while the RL agent chooses the location of the disentangling gate from a tableau observation. The central discovery is that learned policies solve the entanglement-suppression task in a way neither random nor greedy strategies do. With full information, the learned strategies eliminate the Clifford phase transition: for every fixed p>0, the normalized total entanglement $\langle S_{\mathrm{tot}}\rangle/\mathcal{N}$ decreases with system size and appears to vanish in the thermodynamic limit, so the volume-law phase survives only at p=0. W

Load-bearing premise

The area-law conclusion rests on the assumption that the self-similar pyramid pattern seen from N=64 to N=128 (pyramid count doubling from 2 to 4 at p=0.05) keeps doubling as N grows, so total entanglement grows only linearly and the normalized density continues to fall roughly as 1/(pN); only four system sizes, with visible deviations at p=0.05 and p=0.10, currently support that extrapolation.

Editorial extensions

If this is right

  • If the area-law claim holds, the volume-law phase of this circuit model collapses to the single point p=0 for learned strategies: any nonzero disentangling bias keeps the steady state area-law.
  • Partial information suffices only above q_c(p); the same RL pipeline reverts to volume-law entanglement below the threshold, and the required information decreases as p grows.
  • Active feedback is essential: the learned strategy adapts in real time, whereas fixed human-designed bottleneck placement becomes unstable and reaches maximally entangled states at low p.
  • The asymptotic law $\langle S_{\mathrm{tot}}\rangle/\mathcal{N}\sim(1-p)/(pN)$ gives a quantitative target for larger simulations and a discriminator between area-law scaling and finite-size volume-law behavior.
  • The steady states are genuine nonequilibrium states: information-based control breaks detailed balance, and maintaining them carries a thermodynamic cost for erasing the controller's record.

Reading between the lines

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

  • If the bottleneck mechanism is generic rather than Clifford-specific, the same learned feedback should suppress volume-law entanglement in Haar-random circuits, where the paper notes passive dynamics have no transition; moderate system sizes could already show pyramidal suppression.
  • The conjecture that q=0 works for p at or above the greedy threshold implies that an experiment needs only to detect and target nearest-neighbor Bell-like pairs from single-shot measurements, not reconstruct the full quantum state, removing a major postselection bottleneck.
  • The information threshold q_c(p) resembles a phase transition in the space of control policies; near it, RL training time or sample complexity may diverge with system size, making the threshold observable as a learning transition rather than only a steady-state property.
  • A concrete prediction beyond the paper's data: at fixed small p, doubling N should double the number of pyramids and roughly halve the normalized entanglement, so N=256 at p=0.05 should show eight pyramids if the self-similar scaling continues.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper studies active feedback control of entanglement in (1+1)-dimensional stabilizer circuits using reinforcement learning. At each time step a disentangling two-qubit gate is placed according to a learned strategy π(a|o) based on partial observations of the stabilizer tableau, while random Clifford gates are applied with complementary probability. The authors compare random, greedy, and learned strategies and report that, for full information, learned strategies above a critical information threshold produce area-law steady-state entanglement for any bias p>0, eliminating the volume-law phase of the passive unitary-circuit game. They identify pyramid-shaped entanglement bottlenecks as the control mechanism and propose a simplified scaling model. For partial information, they report a critical information fraction qc(p) separating area-law from volume-law steady states. Simulations cover N=16,32,64,128, and code, data, and trained models are deposited.

Significance. If the thermodynamic-limit claim were established, the paper would demonstrate a new capability: information-driven reinforcement learning can engineer nonequilibrium steady states with qualitatively different entanglement scaling than passive dynamics, using only partial state information. The work is clearly presented, compares learned strategies against meaningful baselines, and identifies an interpretable mechanism (bottlenecks/pyramids) rather than a black-box policy. Strengths include the reproducible deposition of code, data, and trained models, the direct simulation of a well-defined stochastic process with no fitted parameters in the central results, and the detailed RL methodology. The main limitation is that the headline claim of area-law scaling for any p>0 in the thermodynamic limit rests on a four-point finite-size extrapolation with single training runs, so the significance of the result in its strongest form is not yet fully supported.

major comments (3)
  1. [Sec. IV B, Fig. 4(b)] The central claim that learned strategies eliminate the volume-law phase for any fixed p>0 is an extrapolation from N=16,32,64,128, with each point apparently coming from a single trained policy and no training-seed error bars. The self-similar pyramid argument rests on one doubling step (2 to 4 pyramids from N=64 to N=128 at p=0.05). The paper itself notes deviations from the proposed (pN)^{-1} law and concedes that training can converge to local minima that eliminate pyramids. Please provide larger system sizes (e.g., N=256) and multiple independent training seeds, or revise the claim to a finite-size evidence statement.
  2. [Eqs. (6) and (7)] The simplified scaling model is algebraically inconsistent. Equation (6), read as p = 1 - (p/N)n_B, gives n_B = N(1-p)/p and hence Stot/N ~ p/[(1-p)N], not the printed 1/(p/(1-p)N + 1) ~ (1-p)/(pN). If Eq. (6) was instead intended as p = (1-p/N)^{n_B}, then n_B ~ N ln(1/p)/p and Stot/N ~ p/[N ln(1/p)], also not Eq. (7). This error must be corrected before the (pN)^{-1} scaling can be used to support the thermodynamic-limit inference.
  3. [Sec. IV C, Fig. 5(a)] The partial-information critical threshold qc(p) is also inferred from single trained models per (p,N,q), and the phase boundary in Fig. 5(b) is a guide-to-eye line without an operational definition or uncertainty estimate. Repeated training seeds and a finite-size scaling analysis are needed to establish the bifurcation and the claimed trade-off between q and p.
minor comments (5)
  1. [Acknowledgments] Typo: 'AKNOWLEDGMENTS' should be 'ACKNOWLEDGMENTS'.
  2. [Appendix D 6] The text refers to the 'Zotero repository' but the availability statement and reference [85] point to Zenodo; please correct the name.
  3. [Appendix D 1] The environment description contains a duplicated paragraph describing the step dynamics; please remove the repetition.
  4. [Eq. (2)] The notation N for both the system size and the normalization factor in Eq. (2) is confusing. Consider using a different symbol for the normalization, e.g., N_norm or S_max.
  5. [Appendix C 4 / Sec. IV B] The statement that learned strategies 'cannot be replaced by simple human-designed control rules' is supported by single realizations of the pyramid strategy in Fig. 11. Reporting statistics over multiple stochastic realizations would make the claim more robust.

Circularity Check

0 steps flagged · score 1.0 of 10

No load-bearing circularity: central results are direct simulations; the scaling model is an explicit explanatory toy model, and self-citations are background only.

full rationale

I walked the paper's derivation chain and found no step where a claimed prediction is equivalent to its input by construction. The central entanglement results (Sec. IV B, Fig. 4) are direct numerical measurements of a well-defined stabilizer-circuit stochastic process. The RL reward is indeed $r_t=-\langle S_{\rm tot}\rangle/N$ (Eq. 5), i.e., the same order parameter that is later reported; however, the paper's physical claim is not merely that the agent lowers the reward, but that the optimized policy changes the scaling with system size, producing area-law behavior for fixed $p>0$ and above an information threshold $q_c(p)$. That scaling is not encoded in the reward and is measured after training on independent circuits, so the optimization objective does not make the scaling claim circular. The simplified scaling model in Eqs. (6)--(7) is explicitly introduced as a model that 'captures essential features and provides useful intuition' while 'disregarding some underlying physics'; it is a stated-ansatz explanatory construction, not a fit to the numerical data, and the authors themselves note deviations from its $(pN)^{-1}$ prediction due to finite-size effects and local minima in the RL loss landscape. Thus it does not constitute a fitted parameter renamed as a prediction. Self-citations (Refs. [40,45,46]) appear, but they are used for background or for contrast with earlier RL disentangling work, not as the load-bearing justification for the central claims; the random-strategy critical point is checked against the external Ref. [37]. The main caveat is that the thermodynamic-limit statement rests on finite-size extrapolation from $N=64$ to $N=128$ with limited pyramid-doubling evidence and acknowledged deviations; that is an inference-robustness concern, not circularity.

Assumptions & free parameters 0 free parameters · 7 assumptions · 0 invented entities

The core computation is a well-defined stochastic simulation, so no fitted parameters are needed for the central entanglement measurements. The main burden falls on modeling assumptions: the row-drop model of partial information, the finite-size extrapolation, and the simplified scaling model used to interpret the learned mechanism. The paper labels the q to 0 statement as a conjecture, which is honest, but it remains an unproved load-bearing assumption for part of the phase diagram.

assumptions (7)
  • standard math Gottesman-Knill theorem: Clifford circuits with classical control are efficiently simulable.
    Invoked in Sec. II and App. E.2 to justify simulating up to N=128 qubits.
  • standard math Clipped gauge relation: entanglement entropy across a cut equals half the number of stabilizer endpoints crossing the cut.
    Used in App. F to compute S(a) and to construct the disentangling gate lookup; taken from prior literature.
  • domain assumption Partial observability is modeled by dropping tableau rows independently with probability 1-q.
    Sec. IV C defines the entire partial-information framework through this O(o|s) map; the qc(p) threshold is defined relative to this model.
  • domain assumption PPO training converges to a near-optimal policy for each (p,N,q).
    No optimality certificate is provided; hyperparameters required tuning and multiple checkpoint restarts, especially at N=128.
  • ad hoc to paper Self-similar pyramid scaling continues beyond N=128.
    Sec. IV B infers Stot~N from observing 2 pyramids at N=64 and 4 pyramids at N=128; the thermodynamic-limit claim assumes this doubling persists.
  • ad hoc to paper Simplified model assumptions: independent bond entanglement probabilities p/N and (1-p)/N and steady-state condition Eq. (6).
    Sec. IV B uses this model to derive (1-p)/(pN) scaling; the authors acknowledge it neglects correlations and simultaneous adjacent entangling events.
  • domain assumption In the q to 0 limit, targeting Bell-like pairs is sufficient to maintain area-law scaling for p>=pgreedy_c.
    Conjecture in Sec. IV C and App. C.5, supported only by 71-76% targeting probabilities from a single trained model.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning to stabilize nonequilibrium phases of matter with active feedback using partial information." pith.science (2026). https://pith.science/paper/F455LNAR

@misc{pith2026250806612,
  author       = {Pith},
  title        = {Pith review of: Learning to stabilize nonequilibrium phases of matter with active feedback using partial information},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/F455LNAR}},
  note         = {Machine review of arXiv:2508.06612}
}
read the original abstract

We investigate the role of information in active feedback control of quantum many-body systems using reinforcement learning. Active feedback breaks detailed balance, enabling the engineering of steady states and dynamical phases of matter otherwise inaccessible in equilibrium. We train reinforcement learning agents using partial state information to prevent entanglement spreading in (1+1)-dimensional stabilizer circuits with up to 128 qubits. We find that, above a critical information threshold, learned near-optimal strategies are non-greedy, stochastic, and reduce volume-law entangled steady states to area-law scaling. The agents achieve this by placing a series of bottlenecks that induce pyramidal structures in the long-time spatial entanglement distribution, which effectively split the system and reduce the maximum accessible entanglement. Crucially, learned strategies are inherently out of equilibrium and require real-time active feedback; we find that the learned behavior cannot be replaced by simple human-designed control rules. This work establishes the foundations for classically implemented, information-driven individual control of many interacting quantum degrees of freedom, demonstrating the capabilities of reinforcement learning to stabilize and uncover novel critical properties of many-body nonequilibrium steady states.

Figures

Figures reproduced from arXiv: 2508.06612 by the authors.

Figure 1
Figure 1. FIG. 1 [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. FIG. 2 [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. FIG. 3 [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (10 more)
Figure 4
Figure 4. Figure 4: FIG. 4 [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: (b). The data points with error bars correspond to those from [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: FIG. 6 [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]
Figure 7
Figure 7. Figure 7: FIG. 7 [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]
Figure 8
Figure 8. Figure 8: FIG. 8 [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]
Figure 9
Figure 9. Figure 9: FIG. 9 [PITH_FULL_IMAGE:figures/full_fig_p015_9.png]
Figure 10
Figure 10. Figure 10: FIG. 10 [PITH_FULL_IMAGE:figures/full_fig_p015_10.png]
Figure 12
Figure 12. Figure 12: FIG. 12 [PITH_FULL_IMAGE:figures/full_fig_p016_12.png]
Figure 13
Figure 13. Figure 13: After processing through all encoder layers, the data is flattened and passed through a final dense layer to produce the desired output dimension. Each encoder block consists of four main compo￾nents. First, a multi-head self-attention mechanism em￾ploys several paral…
Figure 14
Figure 14. Figure 14: FIG. 14 [PITH_FULL_IMAGE:figures/full_fig_p024_14.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

123 extracted references · 50 canonical work pages

  1. [1]

    Measuring transient times to steady state 13

  2. [2]

    Enhanced data visualization 14

  3. [3]

    Critical point determination via Binder cumulant 14

  4. [4]

    Stability of human-designed pyramid strategies 14

  5. [5]

    Reinforcement learning methodology 16

    Learned strategies with vanishing information 14 D. Reinforcement learning methodology 16

  6. [6]

    Environment setup and dynamics 16

  7. [7]

    Training Algorithm 17

  8. [8]

    Policy Architecture 18

Show all 123 references
  1. [9]

    Computational Resources 19

  2. [10]

    Stabilizer circuits 20

    Hyperparameters 20 E. Stabilizer circuits 20

  3. [11]

    Disentangling Clifford gates 23

    Simulating Stabilizer circuits 22 F. Disentangling Clifford gates 23

  4. [12]

    Efficient clipping algorithm 23 a. Performance Considerations 25 Appendix A: Description of supplementary videos The paper is accompanied by twelve supplementary videos illustrating the dynamics of the entanglement pro- file [see Fig.1(b)] for both greedy and learned strategie...

  5. [13]

    To this end, we simulate 2 11 independent trajec- 14 tories, each initialized in the product state |0⟩⊗N , for various values of the bias parameter p and system size N

    Measuring transient times to steady state In this section, we outline the measurement of the tran- sient time required for the circuit to reach the steady state. To this end, we simulate 2 11 independent trajec- 14 tories, each initialized in the product state |0⟩⊗N , for vari...

  6. [14]

    2 but with improved visual clarity through sep- arate panels

    Enhanced data visualization Figure 9 presents the same dataset shown in the main text Fig. 2 but with improved visual clarity through sep- arate panels. This avoids overlapping curves of different strategies and better highlights the distinct scaling be- haviors and critical t...

  7. [15]

    The fourth- order cumulant is defined as: U = 1 − ⟨x4⟩ 3⟨x2⟩2 (C2) where x = Stot/N is the normalized total entanglement

    Critical point determination via Binder cumulant To determine the critical bias for the greedy strategy, we employ Binder cumulant analysis [84]. The fourth- order cumulant is defined as: U = 1 − ⟨x4⟩ 3⟨x2⟩2 (C2) where x = Stot/N is the normalized total entanglement. Figure 10...

  8. [16]

    Stability of human-designed pyramid strategies In Section IV B, we noted that learned strategies, while appearing simple, are surprisingly difficult to replicate using human-designed approaches. As evidence, we con- sider the failure of a human-designed ”pyramid strategy”: for...

  9. [17]

    Here, we present numerical evidence supporting this conjecture

    Learned strategies with vanishing information In Section IV C, we conjecture that for p > pgreedy c , area-law scaling persists as q → 0. Here, we present numerical evidence supporting this conjecture. The underlying intuition is that, as the system size grows, the gate densit...

  10. [18]

    II), we implement a custom RL environment

    Environment setup and dynamics To learn and investigate entanglement-reducing strate- gies in Clifford circuits (Sec. II), we implement a custom RL environment. The RL pipeline, outlined in Sec. IV A and Fig. 3, trains an agent to learn a policy π(a|s) that selects positions a...

  11. [19]

    Policy gradient methods directly learn a parameterized policy πθ by updating the parameters θ to maximize expected rewards

    T raining Algorithm To train the RL agent, we use the Proximal Policy Op- timization (PPO) algorithm [63], a state-of-the-art policy gradient method. Policy gradient methods directly learn a parameterized policy πθ by updating the parameters θ to maximize expected rewards. The...

  12. [20]

    Specifically, we employ the en- coder component of the transformer architecture illus- trated in Fig

    Policy Architecture Both the actor and critic are modeled with a transformer-based architecture implemented using the Flax Python library. Specifically, we employ the en- coder component of the transformer architecture illus- trated in Fig. 13. The encoder architecture stacks ...

  13. [21]

    T raining Process We employ a warmup cosine decay learning rate sched- ule which improves training stability, facilitates conver- gence, and enables soft restarts from checkpoints – partic- ularly at large system sizes ( N = 64, 128) where training required multiple checkpoint...

  14. [22]

    However, the dominant bottleneck is the transient time required for the physical dynamics to reach the steady state, which scales as Ttr ∝ N 3

    Computational Resources All training are conducted on a high-performance com- puting system with the following specifications: • Intel Xeon Platinum 8360Y CPUs (36 cores at 2.40GHz, dual CPU per node) • 72 execution nodes with 1 TB RAM and 4 Nvidia A100-80GB GPUs each Resource...

  15. [23]

    While common defaults were used when possible, some fine- tuning was required for each model

    Hyperparameters Table I summarizes the hyperparameters used for training across system sizes N ∈ 16, 32, 64, 128. While common defaults were used when possible, some fine- tuning was required for each model. The complete list of hyperparameter configurations is available in th...

  16. [24]

    The cen- tral component of the algorithm is the so-called tableau that represents the n stabilizer and the n destabilizer gen- erators

    Simulating Stabilizer circuits First introduced by Aaronson and Gottesman [54], the Tableau algorithm represents an efficient method to sim- ulate stabilizer circuits on classical computers. The cen- tral component of the algorithm is the so-called tableau that represents the ...

  17. [25]

    This is done at each time step of the dynamics, and an efficient clipping algorithm is essential to keep computa- tion time manageable

    Efficient clipping algorithm Given a stabilizer state |ψ⟩, mapping it to disentan- gling gates requires the tableau to be in clipped gauge. This is done at each time step of the dynamics, and an efficient clipping algorithm is essential to keep computa- tion time manageable. A...

  18. [26]

    The row must not have been processed

  19. [27]

    It must have an endpoint on the current column (left or right, depending on direction)

  20. [28]

    Once the pivot is found, it is used to eliminate all other rows with an endpoint on the same column

    Among these, the row with the smallest support is chosen. Once the pivot is found, it is used to eliminate all other rows with an endpoint on the same column. This pro- ceeds as follows:

  21. [29]

    Identify the pivot row

  22. [30]

    Create a boolean mask to select rows that need updating

  23. [31]

    This parallelized approach significantly reduces compu- tation time

    Apply a single bitwise XOR operation to update all selected rows in parallel. This parallelized approach significantly reduces compu- tation time. Finally, the clipped gauge also enables direct computa- tion of entanglement. The number of stabilizers crossing a given cut deter...

  24. [32]

    Landauer, IBM J

    R. Landauer, IBM J. Res. Dev. 5, 183 (1961)

    1961

  25. [33]

    Salath´ e, P

    Y. Salath´ e, P. Kurpiers, T. Karg, C. Lang, C. K. An- dersen, A. Akin, S. Krinner, C. Eichler, and A. Wallraff, Phys. Rev. Appl. 9, 034011 (2018)

    2018

  26. [34]

    Sakaguchi, S

    A. Sakaguchi, S. Konno, F. Hanamura, W. Asavanant, H. Yamazaki, A. Youssefi, K. Takase, J.-i. Yoshikawa, N. C. Menicucci, K. Ping, K. Miyata, H. Ogawa, P. Marek, R. Filip, and A. Furusawa, Nat. Commun. 14, 3817 (2023)

    2023

  27. [35]

    Iqbal, N

    M. Iqbal, N. Tantivasadakarn, T. M. Gatterman, J. A. Gerber, K. Gilmore, D. Gresh, A. Hankin, N. Hewitt, C. V. Horst, M. Matheny, T. Mengle, B. Neyenhuis, M. Foss-Feig, A. Vishwanath, R. Verresen, and H. Dreyer, Commun. Phys. 7, 205 (2024)

    2024

  28. [36]

    Buhrman, M

    H. Buhrman, M. Folkertsma, B. Loff, and N. M. P. Neu- mann, Quantum 8, 1552 (2024)

    2024

  29. [37]

    Wendin, Rep

    G. Wendin, Rep. Prog. Phys. 80, 106001 (2017)

    2017

  30. [38]

    Preskill, arXiv:2208.08064 (2022)

    J. Preskill, arXiv:2208.08064 (2022)

    2022 arXiv

  31. [39]

    D. J. Gross, P. Zoller, and A. Sevrin, eds., The Physics of Quantum Information: Proceedings of the 28th Solvay Conference on Physics, World Scientific Series on Quan- tum Science and Technology (World Scientific, Singa- pore, 2023) brussels, Belgium, 19–21 May 2022

    2023

  32. [40]

    J. D. Biamonte, M. E. S. Morales, and D. E. Koh, Phys. Rev. A 101, 012349 (2020)

    2020

  33. [41]

    A. J. Daley, I. Bloch, C. Kokail, S. Flannigan, N. Pearson, M. Troyer, and P. Zoller, Nature 607, 667 (2022)

    2022

  34. [42]

    G. E. Fux, E. Tirrito, M. Dalmonte, and R. Fazio, Phys. Rev. Res. 6, L042030 (2024)

    2024

  35. [43]

    Q. Zhao, Y. Zhou, and A. M. Childs, Nat. Phys. 10.1038/s41567-025-02945-2 (2025)

    2025 doi

  36. [44]

    Aolita, F

    L. Aolita, F. de Melo, and L. Davidovich, Rep. Prog. Phys. 78, 042001 (2015)

    2015

  37. [45]

    Bauer, S

    B. Bauer, S. Bravyi, M. Motta, and G. K. Chan, Chem. Rev. 120, 12685 (2020)

    2020

  38. [46]

    D ´ ıez-Valle, D

    P. D ´ ıez-Valle, D. Porras, and J. J. Garc ´ ıa-Ripoll, Phys. Rev. A 104, 062426 (2021)

    2021

  39. [47]

    Niedermeier, J

    M. Niedermeier, J. L. Lado, and C. Flindt, Phys. Rev. Res. 6, 033325 (2024)

    2024

  40. [48]

    A. C. Nakhl, T. Quella, and M. Usman, Phys. Rev. A 109, 032413 (2024)

    2024

  41. [49]

    G. J. Mooney, G. A. L. White, C. D. Hill, and L. C. L. Hollenberg, npj Quantum Inf. 7, 27 (2021)

    2021

  42. [50]

    Zhang, T

    X.-M. Zhang, T. Li, and X. Yuan, Phys. Rev. Lett. 129, 230504 (2022)

    2022

  43. [51]

    G.-Y. Zhu, N. Tantivasadakarn, A. Vishwanath, S. Trebst, and R. Verresen, Phys. Rev. Lett. 131, 200201 (2023)

    2023

  44. [52]

    E. H. Chen, G.-Y. Zhu, R. Verresen, A. Seif, E. Ba¨ umer, D. Layden, N. Tantivasadakarn, G. Zhu, S. Shel- don, A. Vishwanath, S. Trebst, and A. Kandala, arXiv:2309.02863 (2023)

    2023 arXiv

  45. [53]

    Islam, R

    R. Islam, R. Ma, P. M. Preiss, M. E. Tai, A. Lukin, M. Rispoli, and M. Greiner, Nature 528, 77 (2015)

    2015

  46. [54]

    Vermersch, A

    B. Vermersch, A. Elben, M. Dalmonte, J. I. Cirac, and P. Zoller, Phys. Rev. A 97, 023604 (2018)

    2018

  47. [55]

    Brydges, M

    M. Brydges, M. Martin, D. Eldredge, J. H. Wesenberg, D. H. J. K. Dahlberg, C. A. Muschik, M. B. Plenio, A. M. Rey, and C. Monroe, Science 364, 260 (2019)

    2019

  48. [56]

    Vasseur and J

    R. Vasseur and J. E. Moore, J. Stat. Mech. 2016, 064010 (2016)

    2016

  49. [57]

    Nahum, S

    A. Nahum, S. Vijay, and J. Haah, Phys. Rev. X8, 021014 (2018)

    2018

  50. [58]

    Swingle, Nat

    B. Swingle, Nat. Phys. 14, 988 (2018)

    2018

  51. [59]

    Y. Li, X. Chen, and M. P. A. Fisher, Phys. Rev. B 100, 134306 (2019)

    2019

  52. [60]

    M. P. Fisher, V. Khemani, A. Nahum, and S. Vijay, Annu. Rev. Condens. Matter Phys. 14, 335 (2023)

    2023

  53. [61]

    Y. Li, X. Chen, and M. P. A. Fisher, Phys. Rev. B 98, 205136 (2018)

    2018

  54. [62]

    Skinner, J

    B. Skinner, J. Ruhman, and A. Nahum, Phys. Rev. X 9, 031009 (2019)

    2019

  55. [63]

    A. Chan, R. M. Nandkishore, M. Pretko, and G. Smith, Phys. Rev. B 99, 224307 (2019)

    2019

  56. [64]

    C. Noel, P. Niroula, D. Zhu, A. Risinger, L. Egan, D. Biswas, M. Cetina, A. V. Gorshkov, M. J. Gullans, D. A. Huse, and C. Monroe, Nat. Phys. 18, 760 (2022)

    2022

  57. [65]

    J. C. Hoke, M. Ippoliti, E. Rosenberg, D. Abanin, R. Acharya, T. I. Andersen, M. Ansmann, F. Arute, K. Arya, A. Asfaw, ..., and P. Roushan, Nature 622, 481 (2023)

    2023

  58. [66]

    J. Koh, S. Sun, M. Motta, A. J. Minnich, and ..., Nat. Phys. 19, 1314 (2023)

    2023

  59. [67]

    Nahum, J

    A. Nahum, J. Ruhman, S. Vijay, and J. Haah, Phys. Rev. X 7, 031016 (2017)

    2017

  60. [68]

    Morral-Yepes, A

    R. Morral-Yepes, A. Smith, S. Sondhi, and F. Pollmann, PRX Quantum 5, 010309 (2024)

    2024

  61. [69]

    Morral-Yepes, M

    R. Morral-Yepes, M. Langer, A. Gammon-Smith, B. Kraus, and F. Pollmann, arXiv:2507.05055 (2025)

    2025 arXiv

  62. [70]

    R. S. Sutton and A. G. Barto, Reinforcement Learning: An Introduction, 2nd ed. (MIT Press, 2018)

    2018

  63. [71]

    Bukov, A

    M. Bukov, A. G. R. Day, D. Sels, P. Weinberg, A. Polkovnikov, and P. W. Claeys, Phys. Rev. X 8, 031086 (2018)

    2018

  64. [72]

    F¨ osel, P

    T. F¨ osel, P. Tighineanu, T. Weiss, and F. Marquardt, Phys. Rev. X 8, 031084 (2018)

    2018

  65. [73]

    M. Y. Niu, S. Boixo, V. N. Smelyanskiy, and H. Neven, npj Quantum Inf. 5, 33 (2019). 26

    2019

  66. [74]

    M. M. Wauters, E. Panizon, G. B. Mbeng, and G. E. Santoro, Phys. Rev. Res. 2, 033446 (2020)

    2020

  67. [75]

    S.-F. Guo, F. Chen, Q. Liu, M. Xue, J.-J. Chen, J.-H. Cao, T.-W. Mao, M. K. Tey, and L. You, Phys. Rev. Lett. 126, 060401 (2021)

    2021

  68. [76]

    Metz and M

    F. Metz and M. Bukov, Nat. Mach. Intell. 5, 780 (2023)

    2023

  69. [77]

    Tashev, S

    P. Tashev, S. Petrov, F. Metz, and M. Bukov, arXiv preprint arXiv:2406.07884 (2024), arXiv:2406.07884 [quant-ph]

    2024 arXiv

  70. [78]

    J. Olle, R. Zen, M. Puviani, and F. Marquardt, npj Quan- tum Inf. 10, 126 (2024)

    2024

  71. [79]

    V. B. Sørdal and J. Bergli, Phys. Rev. A 100, 042314 (2019)

    2019

  72. [80]

    S. R. P, arXiv:2305.06177 (2023)

    2023 arXiv

  73. [81]

    P. A. Erdman, R. Czupryniak, B. Bhandari, A. N. Jor- dan, F. No´ e, J. Eisert, and G. Guarnieri, Quantum Sci. Technol. 10, 025047 (2025)

    2025

  74. [82]

    Gottesman, Phys

    D. Gottesman, Phys. Rev. A 54, 1862 (1996)

    1996

  75. [83]

    Gottesman, Phys

    D. Gottesman, Phys. Rev. A 57, 127–137 (1998)

    1998

  76. [84]

    Gottesman, arXiv:quant-ph/9807006 (1998)

    D. Gottesman, arXiv:quant-ph/9807006 (1998)

    1998 arXiv

  77. [85]

    Aaronson and D

    S. Aaronson and D. Gottesman, Phys. Rev. A 70, 052328 (2004)

    2004

  78. [86]

    S. Sang, Z. Li, T. H. Hsieh, and B. Yoshida, PRX Quan- tum 4, 040332 (2023)

    2023

  79. [87]

    Richter, O

    J. Richter, O. Lunt, and A. Pal, Phys. Rev. Res. 5, L012031 (2023)

    2023

  80. [88]

    O. Lunt, M. Szyniszewski, and A. Pal, Phys. Rev. B 104, 155111 (2021)

    2021

  81. [89]

    Makki, N

    N. Makki, N. Lang, and H. P. B¨ uchler, Phys. Rev. Res. 6, 013278 (2024)

    2024

  82. [90]

    J. Z. Zhuang and colleagues, Phys. Rev. Research 5, L042043 (2023)

    2023

  83. [91]

    The critical value pc is smaller than 0 .5 due to an asym- metry in the gates. The entanglement-reducing strategy exclusively places gates that reduce entanglement, while the random strategy samples gates uniformly from the Clifford group C2 – most of which increase entangleme...

  84. [92]

    Nandkishore and D

    R. Nandkishore and D. A. Huse, Annu. Rev. Condens. Matter Phys. 6, 15 (2015)

    2015

  85. [93]

    The RL environment is conceptually distinct from the notion of environment in open systems in physics

  86. [94]

    Schulman, F

    J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov, arXiv:1707.06347 (2017)

    2017 arXiv

  87. [95]

    Freeman, E

    D. Freeman, E. Frey, A. Tsividis, S. Yazdani, J. Knight, J. S. Park, S. Levine, and H. Michalewski, Brax - a dif- ferentiable physics engine for large scale rigid body sim- ulation (2021)

    2021

  88. [96]

    Konda and J

    V. Konda and J. Tsitsiklis, in Adv. Neural Inf. Process. Syst., Vol. 12 (MIT Press, 1999)

    1999

  89. [97]

    Bradbury, R

    J. Bradbury, R. Frostig, P. Hawkins, M. J. Johnson, C. Leary, D. Maclaurin, G. Necula, A. Paszke, J. Van- derPlas, S. Wanderman-Milne, and Q. Zhang, JAX: com- posable transformations of Python+NumPy programs (2018)

    2018

  90. [98]

    Cramer, M

    M. Cramer, M. B. Plenio, S. T. Flammia, R. Somma, D. Gross, S. D. Bartlett, O. Landon-Cardinal, D. Poulin, and Y. Liu, Nat. Commun. 1, 149 (2010)

    2010

  91. [99]

    B. P. Lanyon, C. Maier, M. Holz¨ apfel, T. Baumgratz, C. Hempel, P. Jurˇ cevi´ c, I. Dhand, A. S. Buyskikh, A. J. Daley, M. Cramer, M. B. Plenio, R. Blatt, and C. F. Roos, Nat. Phys. 13, 1158 (2017)

    2017

  92. [100]

    Rocchetto, S

    A. Rocchetto, S. Aaronson, S. Severini, G. Carvacho, D. Poderini, I. Agresti, M. Bentivegna, and F. Sciarrino, Sci. Adv. 5, eaau1946 (2019)

    2019

  93. [101]

    A Bell-like pair is a two qubits state that is maximally entangled and locally equivalent (up to single-qubit Clif- ford operation) to a Bell pair

  94. [102]

    P¨ utz, S

    M. P¨ utz, S. J. Garratt, H. Nishimori, S. Trebst, and G. Zhu, arXiv:2504.12385 (2025)

    2025 arXiv

  95. [103]

    Reuer, J

    K. Reuer, J. Landgraf, T. F¨ osel, J. O’Sullivan, L. Beltr´ an, A. Akin, G. Norris, A. Remm, M. Kerschbaum, J. Besse, F. Marquardt, A. Wallraff, and C. Eichler, Nat. Com- mun. 14, 7138 (2023)

    2023

  96. [104]

    D. T. R. Nagy, C. Czab´ an, B. Bak´ o, P. H´ aga, Z. Kallus, and Z. Zimbor´ as, arXiv:2410.18284 (2024)

    2024 arXiv

  97. [105]

    Wen, Rev

    X.-G. Wen, Rev. Mod. Phys. 89, 041004 (2017)

    2017

  98. [106]

    Senthil, Annu

    T. Senthil, Annu. Rev. Condens. Matter Phys. 6, 299 (2015)

    2015

  99. [107]

    Y. Zhou, Q. Zhao, X. Yuan, and X. Ma, NPJ Quantum Inf. 5, 83 (2019)

    2019

  100. [108]

    B. M. Terhal, Rev. Mod. Phys. 87, 307 (2015)

    2015

  101. [109]

    Wang and H

    Z. Wang and H. Tang, arXiv:2412.20380 (2024)

    2024 arXiv

  102. [110]

    F. J. Cao and M. Feito, Phys. Rev. E 79, 041118 (2009)

    2009

  103. [111]

    J. V. Koski, V. F. Maisi, T. Sagawa, and J. P. Pekola, Phys. Rev. Lett. 113, 030601 (2014)

    2014

  104. [112]

    Sagawa and M

    T. Sagawa and M. Ueda, Phys. Rev. E 85, 021104 (2012)

    2012

  105. [113]

    Prech and P

    K. Prech and P. P. Potts, Phys. Rev. Lett. 133, 140401 (2024)

    2024

  106. [114]

    J. M. Horowitz and M. Esposito, Phys. Rev. X 4, 031015 (2014)

    2014

  107. [115]

    Binder, Phys

    K. Binder, Phys. Rev. Let. 47, 693 (1981)

    1981

  108. [116]

    Cemin, M

    G. Cemin, M. Schmitt, and M. Bukov, Data, code and models., https://doi.org/10.5281/zenodo. 16672380 (2025), available upon reasonable request

    2025 doi

  109. [117]

    Gottesman, Surviving as a quantum computer in a classical world (2024), textbook manuscript preprint

    D. Gottesman, Surviving as a quantum computer in a classical world (2024), textbook manuscript preprint

    2024

  110. [118]

    Gottesman, arXiv:quant-ph/9705052 (1997)

    D. Gottesman, arXiv:quant-ph/9705052 (1997)

    1997 arXiv

  111. [119]

    M. A. Nielsen and I. L. Chuang, Quantum Computa- tion and Quantum Information: 10th Anniversary Edi- tion (Cambridge University Press, Cambridge, 2010)

    2010

  112. [120]

    The normalizer of a subset S in a group G is the set of el- ements of G that leave the set S fixed under conjugation, i.e., NG(S) := {g ∈ G|gS = Sg} = {g ∈ G|gSg−1 = S}

  113. [121]

    Koenig and J

    R. Koenig and J. A. Smolin, J. Math. Phys. 55, 122202 (2014)

    2014

  114. [122]

    Ozols, Clifford Group, Essay, University of Waterloo (2008)

    M. Ozols, Clifford Group, Essay, University of Waterloo (2008)

    2008

  115. [123]

    Iwould be in S, but −I does not stabilize anything

    If S ∈ S, then S can only have a phase of ±1, not ±i; for in the latter case S2 = −I . . . Iwould be in S, but −I does not stabilize anything

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.