Variational quantum thermalizers based on weakly-symmetric nonunitary multi-qubit operations

Elias Zapusek; Florentin Reiter; Kristina Kirova; Michael Marthaler; Walter Hahn

REVIEW 3 major objections 6 minor 112 references

Variational quantum thermalizers based on weakly-symmetric nonunitary multi-qubit operations

T0 review · 3 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A variational quantum thermalizer whose unitary layers are interleaved with weakly-symmetric multi-qubit dissipative channels prepares Gibbs states of paradigmatic spin models at all temperatures, with fidelity above 0.99 in the Ising…

desk verdict The weak/strong symmetry analysis and dissipative gate designs are a keeper, but the entropy estimator at the core of the cost function has a backwards rescaling and a gaming-prone regularizer, so the numerical claims need major revision before I would trust them. read the letter →

arxiv 2502.09698 v2 pith:ICN53Q4S submitted 2025-02-13 quant-ph

Elias Zapusek , Kristina Kirova , Walter Hahn , Michael Marthaler , Florentin Reiter This is my paper

classification quant-ph

keywords variationalquantumthermalizerGibbsstatepreparationdissipationengineeringweaksymmetrymulti-qubitnonunitaryoperationsentropyestimationtransverse-fieldIsingmodelsimulation

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 argues that interleaving a variational quantum thermalizer's unitary layers with weakly-symmetric multi-qubit nonunitary operations, realized by engineering dissipation rather than by measurements or ancillas, lets the algorithm prepare Gibbs states of spin models at every temperature, including the intermediate regime where prior VQTs lose fidelity. The failure mode it fixes is specific: at intermediate temperatures the target state is mixed but entangled, and single-qubit noise channels cannot reach it, whereas two-qubit dissipative jumps can. To train the ansatz the paper develops a scaled subsystem entropy estimate and a symmetry analysis showing that weakly symmetric channels can move population between symmetry sectors while strongly symmetric ones cannot, which explains why some noise models cap fidelity at 0.5. If the claim holds, near-term quantum computers without ancilla qubits could prepare thermal states of many-body models, with the Ising model demonstrated at fidelity above 0.99 across temperatures.

What carries the argument

The load-bearing object is the weakly symmetric multi-qubit nonunitary layer: a quantum channel that commutes with the symmetry representation as a whole, so that each individual jump operator may fail to commute and can transfer population between symmetry sectors. These layers are implemented by dissipation engineering, in which a weak drive and an auxiliary decaying oscillator produce effective multi-qubit jump operators such as $L_{\kappa,0}^{\mathrm{eff}} = \sqrt{\kappa_f}\,|11\rangle\langle 10| + \sqrt{\kappa_{af}}\,|01\rangle\langle 00|$, with rates tuned by detunings. Training minimizes the free energy $F = \beta\langle H\rangle - S$, with the von Neumann entropy $S$ estimated by the scaled subsystem entropy $S_{n_a} = \frac{n_a}{n} S(\cdot|_{n_a})$, justified by a white-noise model of random circuits interleaved with Pauli channels, and regularized by $(1-|\Delta S|)$ to prevent the optimizer from inflating subsystem entropy without raising the full entropy.

What would settle it

Run the converged six-qubit transverse-field Ising ansatz at an intermediate temperature, perform full tomography on the output, and compare the exact von Neumann entropy with the scaled three-qubit entropy used during training; if the difference is not small for the trained circuit, the cost function is not evaluating free energy and the near-unit fidelities would not be reproduced by an exact-entropy optimization.

Watch

Extended reading notes

Core claim

The central claim is that a variational ansatz alternating parameterized unitaries with weakly-symmetric multi-qubit nonunitary channels prepares the Gibbs state $\rho_G = e^{-\beta H}/Z$ of a Hamiltonian at all temperatures. The multi-qubit channels are engineered dissipative jumps: a weak drive excites selected multi-qubit states into a strongly shifted manifold, and decay of an auxiliary oscillator completes a nonunitary step, so no ancilla qubits or measurements are required. Numerically, the paper reports that for the Ising model the multi-qubit operations achieve near-perfect state preparation with fidelity above 0.99 across temperatures, and for the transverse-field Ising model at criticality and the Heisenberg model with a transverse field they significantly outperform single-qubit noise, especially at the intermediate temperatures where previous thermalizers struggle. The paper also shows that a strongly symmetric channel is confined to one symmetry sector, bounding its high-temperature fidelity by 0.5, while a weakly symmetric channel can mix sectors and reach the Gibbs state, and it proves that its approximate dissipative evolution approaches an efficient noncommutative Gibbs sampler in diamond norm as $\mathcal{O}(\beta)$ at high temperatures.

Load-bearing premise

The scaled subsystem entropy estimate assumes that the full $n$-qubit output state's von Neumann entropy is faithfully captured by the entropy of a small ($n_a = 3$ or $4$) translation-invariant subsystem; if the optimizer drives the circuit to a state whose subsystem entropy grows while the full entropy stays put, the free-energy cost function misranks states and the reported high fidelities would not survive.

Editorial extensions

If this is right

Multi-qubit nonunitary layers give near-perfect thermal state preparation for the Ising model, with fidelity above 0.99 at all tested temperatures, whereas single-qubit channels fall short.
For the transverse-field Ising model at criticality and the Heisenberg model with a transverse field, weakly-symmetric two-qubit dissipators outperform single-qubit noise and cannot be emulated by single-qubit channels combined with unitaries.
Weak symmetry, rather than strong symmetry, is the right design target for Gibbs state preparation because it allows population to move between symmetry sectors; strong symmetry caps fidelity at 0.5 for high temperatures.
The approximate dissipative dynamics reproduces the evolution of an efficient noncommutative Gibbs sampler at high temperatures, with diamond-norm deviation scaling as $\mathcal{O}(\beta)$ as $\beta \to 0$.
Restricting the ansatz to symmetric unitaries and weakly symmetric channels roughly halves the number of optimization iterations needed for convergence.

Reading between the lines

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

A direct test the paper does not run: optimize the same ansatz with an exact entropy estimator (full tomography for small systems) and compare converged fidelities; if they match, the scaled subsystem entropy is not the limiting factor, and if they do not, the reported 0.99 fidelities depend on the approximation.
The weak-symmetry design principle should transfer to other settings with blockade symmetries, such as lattice gauge theories and generalized Gibbs ensembles, where strong symmetry also freezes the wrong sectors.
Because the white-noise justification of the entropy approximation is derived for random circuits, the method's reliability for non-random trained circuits is an open question; the regularization term is a partial patch, not a proof.
The same entropy-splitting idea, applied to non-translation-invariant Hamiltonians by summing entropies of tractable subsystems, is mentioned but untested in the paper and could extend the method to quantum Boltzmann machine training.

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.

Desk Editor's Note

read the letter

Colleague,

This one is worth a referee's time, but not as it stands. The genuinely new piece is the weak/strong symmetry analysis for nonunitary layers in a variational thermalizer, and the dissipative two-qubit gate designs. The explanation for why a strongly symmetric channel caps fidelity at 0.5 at high temperature is clean and correct, and the Ising/TFIM/Heisenberg numerics are plausible if you trust the cost function. The authors also ship a data repo, which is more than most papers in this area do.

The soft spots are real. Eq (5) as printed has the subsystem rescaling backwards: for a translation-invariant state with entropy density s, an n_a-qubit subsystem has entropy n_a s, so the full entropy is (n/n_a) S(n_a), not (n_a/n) S(n_a). On six qubits with n_a=3 that is a factor-four error. I assume it is a typo and the simulations used the corrected scaling, but the text cannot be implemented as written. The Appendix's random-circuit error analysis is internally consistent but does not justify the cost function for the non-random circuits the optimizer actually finds. The authors see this and try to fix it with Eq (10), but multiplying the whole cost by 1-|Delta S| is not a regularizer. When the two subsystem estimates disagree strongly, the cost is forced toward zero regardless of the physical state, so the optimizer can exploit exactly the failure mode the term is meant to suppress. That needs a real fix: an additive penalty, or a cross-check against exact entropy training on the main models.

The main fidelity plots are best-of-initialization, so the high numbers in Figs 5, 6 and 8 are not evidence of robustness. Appendix A does give a ten-seed average comparison showing the scaled subsystem entropy tracks exact entropy for the Heisenberg model, which is the strongest evidence in the paper and should be promoted to the main text. But the 'all temperatures' claim rests on six-qubit rings and best-of runs, so it is overstated.

My take: the symmetry contribution is a keeper, and the dissipative gate engineering is well grounded in prior trapped-ion work. The entropy estimator and its regularization are the load-bearing part and are not presently sound as written. This is a major-revision paper, not a reject. I would send it to referees, and I'd tell the authors to fix the scaling, replace the multiplicative 'regularizer' with something that cannot zero out the cost, and report seed statistics or run-to-run spreads.

Referee Report

3 major / 6 minor

Summary. The paper proposes a variational quantum thermalizer (VQT) that alternates unitary layers with nonunitary channels realized through dissipation engineering, and uses weak and strong symmetries to design both the unitary and the nonunitary elements. The central claim is that this approach prepares Gibbs states of paradigmatic spin models (Ising, transverse-field Ising, and Heisenberg with a transverse field) at all temperatures. Since the von Neumann entropy of the full state is not directly observable, the authors introduce a scaled subsystem entropy as a surrogate in a free-energy cost function, together with a multiplicative consistency factor based on two subsystem sizes. Numerical simulations on six qubits report high fidelities with the exact Gibbs state, and a symmetry-sector argument is used to explain why strongly symmetric channels are incapable of preparing high-temperature states. Appendices contain derivations of the entropy approximation, the weak/strong symmetry formalism, a connection to an efficient quantum Gibbs sampler, and a physical implementation of the multi-qubit nonunitary operations.

Significance. If the central claim is established, the paper would provide a practical, ancilla-free route to thermal-state preparation on near-term devices, with a clear design principle (weak versus strong symmetry) for choosing nonunitary operations. The dissipation-engineering implementations in Appendix D are concrete and connect to existing trapped-ion and superconducting platforms. The paper also contributes a benchmark comparison of entropy-estimation methods and a data-availability link, which support reproducibility. However, the main numerical results depend crucially on the entropy surrogate, and as written the surrogate contains a scaling error and a regularization whose behavior is not controlled, so the present version does not yet substantiate the 'all temperatures' claim.

major comments (3)

[Section IV, Eq. (5)] The rescaling in Eq. (5) is inverted. For a translation-invariant state with entropy density s, the output of the n_a-qubit circuit has entropy n_a s, so the full n-qubit entropy is (n/n_a) S_{n_a}(A|_{n_a}), not (n_a/n) S_{n_a}(A|_{n_a}). As written, Eq. (5) underestimates the full entropy by a factor (n/n_a)^2, which is a factor 4 for the n=6, n_a=3 case used later. This also contradicts the text's statement that the approximation is exact for the fully mixed state, and it is inconsistent with the error analysis in Eqs. (8)-(9), where S_n and S_{n_a} are compared without any rescaling. Because Eq. (5) defines the entropy term in the cost function (10), the algorithm as written does not minimize a free energy; this needs to be corrected and the numerical experiments re-evaluated with the correct scaling.
[Section IV, Eq. (10)] The multiplicative factor (1-|ΔS|) is not a regularizer in any standard sense. When |ΔS| approaches 1, the cost is forced toward zero independent of both the energy and the entropy, so states with maximally inconsistent subsystem entropy estimates become global minima of the cost landscape; when |ΔS| exceeds 1, the cost becomes negative, which the optimizer can exploit. For the chosen parameters n=6, n_a=3, n_b=4, entropy differences of order log 2 are reachable, and under the published Eq. (5) even product states give |ΔS|=7/6 log2 ≈ 0.808, with larger differences possible. The paper provides no argument that the optimizer cannot drive ΔS into the problematic regime, and no alternative additive consistency penalty is analyzed.
[Section IV and Section VI, Figs. 5, 6, 8] The entropy surrogate is justified analytically only for random circuit instances, as the authors themselves note in Section IV when they state that the optimizer can increase the subsystem entropy without changing the full-system entropy. The main numerical claims for the multi-qubit ansätze are reported only as best-of-initialization runs, and no cross-check against training with the exact entropy is provided for those ansätze. Since the cost function is an approximation, selecting the run with the lowest approximate cost may preferentially select runs that exploit the surrogate error. To support the 'all temperatures' claim, the paper should report, for the multi-qubit ansätze, a comparison between approximate-entropy and exact-entropy training (analogous to Fig. 9 for the single-qubit case), together with averages or standard deviations rather than only the best run.

minor comments (6)

[Section IV, Eq. (4)] The free energy is defined as F = β⟨H⟩ - S, which differs from the standard Helmholtz free energy E - T S by a factor T; this is equivalent as a cost function but should be flagged to avoid confusion.
[Section IV, after Eq. (5)] The statement that the entropy approximation is exact for the fully mixed state is only true with the corrected rescaling; as written, Eq. (5) gives n_a^2/n log2 instead of n log2 for an n-qubit completely mixed state.
[Appendix A 2, after Eq. (A4)] The formula for the single-qubit entropy is missing a minus sign; it should read S = -(1-Λ/2) log(1-Λ/2) - (Λ/2) log(Λ/2).
[Section VI, first paragraph] The description 'we neglect the fluctuations due to the inefficient optimization procedure' indicates that all main figures are best-of-initialization results; reporting the average and standard deviation over initializations would substantially strengthen the empirical claims.
[Appendix B 4, Eq. (B22)] The notation 'θ({111, XXX, X11, 1XX}) = {0, π, π, 0}' should specify the ordering of group elements so that the phases can be matched to the character table that follows.
[Section IV, Eq. (10)] The factor (1-|ΔS|) is described as a regularization term, but it is a multiplicative factor on the entire cost; this terminology is misleading and should be changed.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the Gibbs-state fidelity is benchmarked externally and the free-energy cost is minimized rather than fit to the target; the paper's own limitation passages are soundness caveats, not input-output identifications.

full rationale

The paper's central derivation chain is not circular. The variational parameters are trained by minimizing a free-energy cost, and the reported success metric is the Uhlmann-Jozsa fidelity against the exact Gibbs state (Sec. VI), an external benchmark that is not built into the cost function. The entropy surrogate in Eq. (5) is a heuristic justified by a random-circuit white-noise model (Eqs. 6-9) and by numerical benchmarking; its failure mode is explicitly acknowledged in Sec. IV: 'the optimizer that minimizes the free energy can reduce its cost by increasing the entropy of the subsystem, while the entropy of the entire system is not changed.' The regularization in Eq. (10) is a heuristic, and the statement that 'the global minimum of the cost landscape does not produce the target Gibbs state' is a soundness and robustness limitation, not a circular reduction. Similarly, Eq. (5) appears to contain a scaling typo because the infinite-temperature limit is not exact as written, but this is an algebraic or reproducibility issue, not a case of the result being equivalent to its inputs by construction. The high-temperature connection to the Gibbs sampler of Ref. [4] in App. C is an independent bound, and the physical implementation of the nonunitary layers relies on externally demonstrated dissipation engineering (Ref. [57]) and standard adiabatic elimination (Ref. [96]). The many self-citations in the paper support experimental feasibility and standard methods rather than carrying the load of the central claim that the optimized states match the exact Gibbs states. Therefore, no circular step is established.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

The central claim rests on the standard open-system framework (Lindblad dynamics, adiabatic elimination), a translation-invariant entropy estimator whose rigor is only established for random circuits, and several hand-chosen ansatz hyperparameters. No new physical entities are introduced; the nonunitary operations are engineered dissipators derived from existing techniques.

free parameters (4)

Subsystem sizes n_a=3, n_b=4 = 3 and 4 qubits
Used in the scaled subsystem entropy and regularization (Sec IV, Eq 10). The choice is not derived from an optimality condition; it sets the bias-variance trade-off of the cost function and affects which states minimize the approximate free energy.
Regularization weight (implicit coefficient 1 on 1-|ΔS|) = 1
The regularized cost (Eq 10) multiplies the approximate free energy by 1-|ΔS| with no tunable prefactor or schedule; this is an ad hoc fix for the optimizer exploiting entropy approximation error.
TFIM channel parameter q in L=Z+qY = not stated (variational or hand-set)
The nonunitary layer for the transverse-field Ising model is generated by L=Z+qY (Sec VI B); the value or training rule for q is not given in the main text, and performance depends on it.
Heisenberg jump rates κ_f and κ_af = tuned through detunings, values not listed
The two-qubit dissipators in Eqs D17-D19 depend on κ_f and κ_af controlled by detunings δ, Δ and coupling g; the paper describes the mechanism but does not report the values used in Fig 8.

assumptions (6)

standard math Lindblad master equation describes the open-system dynamics
Used throughout Sec III and App D to model dissipation engineering; the engineered channel is e^{tL}.
domain assumption Adiabatic elimination in the κ >> Ω limit yields the stated effective jump operators
App D Eqs D9-D19; assumes the oscillator cooling dominates the drive so that excited states can be eliminated. This is a physical approximation, not a theorem.
domain assumption Translation invariance of the target Hamiltonian and ansatz permits rescaling subsystem entropy
Sec IV Eq 5; exact only for translation-invariant or product states. The authors acknowledge this in Sec IV and discuss how the optimizer can exploit the approximation error.
domain assumption Random-circuit white-noise approximation (Eq 6) describes output states of the noisy ansatz
Taken from Ref [63] and used to derive Eqs 8-9; assumes a high-fidelity pure-state component, logarithmic depth, and local Pauli noise, conditions not verified for the trained circuits.
standard math Gibbs state commutes with symmetries of the Hamiltonian
Sec V Eq 12; follows from [R_g,H]=0 and spectral calculus.
standard math Subadditivity of von Neumann entropy
Used in App A 3 to build variational entropy upper bounds and the regularization term.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Variational quantum thermalizers based on weakly-symmetric nonunitary multi-qubit operations." pith.science (2026). https://pith.science/paper/ICN53Q4S

@misc{pith2026250209698,
  author       = {Pith},
  title        = {Pith review of: Variational quantum thermalizers based on weakly-symmetric nonunitary multi-qubit operations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ICN53Q4S}},
  note         = {Machine review of arXiv:2502.09698}
}

read the original abstract

We propose incorporating multi-qubit nonunitary operations in Variational Quantum Thermalizers (VQTs). VQTs are hybrid quantum-classical algorithms that generate the thermal (Gibbs) state of a given Hamiltonian, with applications in quantum algorithms and simulations. However, current algorithms struggle at intermediate temperatures, where the target state is nonpure but exhibits entanglement. We devise multi-qubit nonunitary operations that harness weak symmetries and thereby improve the performance of the algorithm. Utilizing dissipation engineering, we create these nonunitary multi-qubit operations without the need for measurements or additional qubits. To train the ansatz, we develop and benchmark novel methods for entropy estimation of quantum states, expanding the toolbox for quantum state characterization. We demonstrate that our approach can prepare thermal states of paradigmatic spin models at all temperatures. Our work thus creates new opportunities for simulating open quantum many-body systems.

Figures

Figures reproduced from arXiv: 2502.09698 by the authors.

**Figure 2.** FIG. 2. Relative error of the entropy approximation. The an [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** FIG. 3. The effect of symmetries on gradient variances. We [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Figures from the paper (5 more)

**Figure 4.** Figure 4: FIG. 4. The effect of unitary symmetries on model training. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

**Figure 5.** Figure 5: FIG. 5. Ising model. We plot the fidelity with the ideal tar [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

**Figure 7.** Figure 7: FIG. 7. Nonunitary multi-qubit gate operation. Two-body [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]

**Figure 9.** Figure 9: FIG. 9. Comparison of entropy estimation methods. (a) Fidelity with the target state for the lowest final cost. The circuit [PITH_FULL_IMAGE:figures/full_fig_p012_9.png]

**Figure 10.** Figure 10: FIG. 10. Engineering of nonunitary multi-qubit operation. [PITH_FULL_IMAGE:figures/full_fig_p018_10.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

112 extracted references · 52 canonical work pages

[1]

Estimate by sampling from
[2]

Estimate entropy by simulating subsystem
[3]

internal symmetry internal & spatial no symmetry (c) FIG

Return updated parameters . internal symmetry internal & spatial no symmetry (c) FIG. 1. Overview. (a) The hybrid algorithm. The quantum circuit begins with an initial unitary layer U0, followed by m layers, which are composed of unitary operations Uθj interleaved with nonunitary layers Eλj . The parameters of all layers are trained using the free energy ...
[4]

The entropy of a subsystem of size na is computed and rescaled to ap- proximate the entropy of the full system

Scaled Subsystem Entropy In the main text, we introduce a scaled entropy approx- imation for translation-invariant systems. The entropy of a subsystem of size na is computed and rescaled to ap- proximate the entropy of the full system. For further details, refer to Sec. IV of the main text
[5]

[19] derive an analytic approxima- tion of the entropy of an ansatz that interleavesm unitary layers with m layers of depolarizing channels

Analytic Entropy approximation The authors of Ref. [19] derive an analytic approxima- tion of the entropy of an ansatz that interleavesm unitary layers with m layers of depolarizing channels. The au- thors calculate the entropy of the state produced by an ansatz that first applies all m layers of the depolarizing channel followed bym unitary layers. The a...
[6]

(A6) Here ρj signifies the partial trace over all subsystems but the j-th one

V ariational Entropy Estimation Evaluating the entropy of a product state is as simple as S(ρ1 ⊗ρ2 ⊗ · · · ⊗ρn) = S(ρ1) +S(ρ2) +· · ·+S(ρn). (A6) Here ρj signifies the partial trace over all subsystems but the j-th one. For a general state, the entropy is upper bounded by the sum of entropies of subsystems S(ρ) ≤ S(ρ1) + S(ρ2) + · · ·+ S(ρn). (A7) Further...
[7]

This allows us to benchmark the scaled subsystem entropy against the variational entropy estimation

Comparison for translation invariant system We consider a system with translation invariance. This allows us to benchmark the scaled subsystem entropy against the variational entropy estimation. We consider the Heisenberg model with a transverse field H = X k XkXk+1 + YjYk+1 + ZkZk+1 + X k Xk. (A10) We initialize the ansatz in the symmetry sector of the g...
[8]

These definitions can be found in any group theory textbook [88]

Preliminaries We first review relevant concepts from group and rep- resentation theory. These definitions can be found in any group theory textbook [88]. For a review in the context of quantum machine learning consider [89]. Symmetries are organized in groups. Recall the defini- tion of a (finite) group. Definition B.1 (Group) A group consists of a set G ...

Show all 112 references

[9]

For all g, h, c∈ G, (g · h) · c = g · (h · c)

Composition is associative. For all g, h, c∈ G, (g · h) · c = g · (h · c)
[10]

There exists an identity element e in G such that for any g in G e,·g = g · e = g
[11]

For every element g in G there exists an element g−1 such that g · g−1 = g−1 · g = e. When we consider how a symmetry group acts on a phys- ical system, we require a mechanism to translate the ab- stract group into the physical Hilbert space that describes the system. This mec...
[12]

[74], we introduce symmetry conditions for quantum channels

W eak and strong symmetry conditions Following Ref. [74], we introduce symmetry conditions for quantum channels. Consider a model described by Hamiltonian H with symmetry group G with unitary representation Rg for g in G. Definition B.6 (W eak symmetry) A channel E sat- isfies...
[13]

Symmetries of Lindbladian We consider quantum systems described by Lindblad master equations for the implementation of quantum channels. The dynamics of a quantum system coupled to a Markovian reservoir can be captured by a Master equation in Lindblad form [91]: L(ρ) = − i ℏ [...
[14]

Rg X i KiρK † i # = 1 |G| X g χα(g) Tr

Symmetry sectors One can use symmetries to partition a quantum sys- tem’s Hilbert space into distinct subspaces. The main advantage is that they allow us to reduce the dimension of the problem we have to consider greatly simplifying it [92, 93] The projector onto the symmetry ...
[15]

We can simplify the im- plementation of the jump operator by applying a unitary 17 transformation

Engineering of the single-qubit TFIM jump The jump operator used for the TFIM has a form L = √p (Z + dY ) , (D2) with p and d real parameters. We can simplify the im- plementation of the jump operator by applying a unitary 17 transformation. Application of a unitary operator U...
[16]

(D12) We consider a physical system of two qubits and four levels {|0⟩ , |1⟩ , |e⟩ , |f ⟩} coupled to an oscillator mode with ladder operator a and b

Engineering of the two-qubit Heisenberg jumps The jump operators we use to prepare thermal states of the Heisenberg model are: Leff κ,0 = √κf |11⟩ ⟨10| + √κaf |01⟩ ⟨00| , (D11) Leff κ,1 = √κf |00⟩ ⟨01| + √κaf |10⟩ ⟨11| . (D12) We consider a physical system of two qubits and fo...
[17]

R. P. Feynman, Int. J. Theor. Phys. 21, 467 (1982)

1982
[18]

Peruzzo, J

A. Peruzzo, J. McClean, P. Shadbolt, M.-H. Yung, X.-Q. Zhou, P. J. Love, A. Aspuru-Guzik, and J. L. O’Brien, Nat. Commun. 5, 4213 (2014)

2014
[19]

Beguˇ si´ c, J

T. Beguˇ si´ c, J. Gray, and G. K.-L. Chan, Sci. Adv. 10, eadk4321 (2024)

2024
[20]

C.-F. Chen, M. J. Kastoryano, and A. Gily´ en, An effi- cient and exact noncommutative quantum Gibbs sampler (2023), arXiv:2311.09207

2023
[21]

Y. Cao, J. Romero, J. P. Olson, M. Degroote, P. D. John- son, M. Kieferov´ a, I. D. Kivlichan, T. Menke, B. Per- opadre, N. P. D. Sawaya, S. Sim, L. Veis, and A. Aspuru- Guzik, Chem. Rev. 119, 10856 (2019)

2019
[22]

Mazzola, R

G. Mazzola, R. Helled, and S. Sorella, Phys. Rev. Lett. 120, 025701 (2018)

2018
[23]

E. Gull, O. Parcollet, and A. J. Millis, Phys. Rev. Lett. 110, 216405 (2013)

2013
[24]

Bauer, S

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

2020
[25]

F. G. Brandao and K. M. Svore, in 2017 IEEE 58th An- nual Symposium on Foundations of Computer Science (FOCS) (IEEE, Berkeley, CA, 2017) pp. 415–426

2017
[26]

Watts, Y

O. Watts, Y. Kikuchi, and L. Coopmans, Quantum Semidefinite Programming with Thermal Pure Quantum States (2023), arXiv:2310.07774

2023 arXiv
[27]

Verdon, M

G. Verdon, M. Broughton, and J. Biamonte, (2019), arXiv:1712.05304

2019 arXiv
[28]

Zoufal, A

C. Zoufal, A. Lucchi, and S. Woerner, Quantum Machine Intelligence 3, 7 (2021)

2021
[29]

M. H. Amin, E. Andriyash, J. Rolfe, B. Kulchytskyy, and R. Melko, Phys. Rev. X 8, 021050 (2018)

2018
[30]

Bravyi, A

S. Bravyi, A. Chowdhury, D. Gosset, and P. Wocjan, Nat. Phys. 18, 1367 (2022)

2022
[31]

Martyn and B

J. Martyn and B. Swingle, Phys. Rev. A 100, 032107 (2019)

2019
[32]

A. N. Chowdhury, G. H. Low, and N. Wiebe, (2020), arXiv:2002.00055

2020 arXiv
[33]

Y. Wang, G. Li, and X. Wang, Phys. Rev. Appl 16, 054035 (2021)

2021
[34]

Warren, L

A. Warren, L. Zhu, N. J. Mayhall, E. Barnes, and S. E. Economou, Adaptive variational algorithms for quantum Gibbs state preparation (2022), arXiv:2203.12757

2022 arXiv
[35]

Foldager, A

J. Foldager, A. Pesah, and L. K. Hansen, Sci. Rep. 12, 3862 (2022)

2022
[36]

Consiglio, in Numerical Computations: Theory and Algorithms, Vol

M. Consiglio, in Numerical Computations: Theory and Algorithms, Vol. 14478 (Springer Nature Switzerland, Cham, 2025) pp. 56–70, series Title: Lecture Notes in Computer Science

2025
[37]

Ilin and I

Y. Ilin and I. Arad, IEEE Trans. Quantum Eng. 6, 1 (2025)

2025
[38]

Fromm, O

M. Fromm, O. Philipsen, M. Spannowsky, and C. Winterowd, Simulating Z2 Lattice Gauge Theory with the Variational Quantum Thermalizer (2023), arXiv:2306.06057

2023 arXiv
[39]

X. Wang, Y. Chai, M. Demidik, X. Feng, K. Jansen, and C. T¨ uys¨ uz, Symmetry enhanced variational quantum imaginary time evolution (2023), arXiv:2307.13598

2023 arXiv
[40]

Selisko, M

J. Selisko, M. Amsler, T. Hammerschmidt, R. Drautz, and T. Eckl, Quantum Sci. Technol. 9, 015026 (2024)

2024
[41]

Polla, Y

S. Polla, Y. Herasymenko, and T. E. O’Brien, Phys. Rev. A 104, 012414 (2021)

2021
[42]

Holmes, G

Z. Holmes, G. Muraleedharan, R. D. Somma, Y. Subasi, and B. S ¸ahino˘ glu, Quantum6, 825 (2022)

2022
[43]

Temme, T

K. Temme, T. J. Osborne, K. G. Vollbrecht, D. Poulin, and F. Verstraete, Nature 471, 87 (2011)

2011
[44]

Yung and A

M.-H. Yung and A. Aspuru-Guzik, Proceedings of the National Academy of Sciences 109, 754 (2012)

2012
[45]

Wocjan and K

P. Wocjan and K. Temme, Communications in Mathe- matical Physics 402, 3201 (2023)

2023
[46]

P. Rall, C. Wang, and P. Wocjan, Quantum 7, 1132 (2023)

2023
[47]

C.-F. Chen, M. J. Kastoryano, F. G. S. L. Brand˜ ao, and A. Gily´ en, Quantum Thermal State Preparation (2023), arXiv:2303.18224. 19

2023 arXiv
[48]

Rouz´ e, D

C. Rouz´ e, D. S. Fran¸ ca, and A. M. Alhambra, (2024), arXiv:2403.12691

2024 arXiv
[49]

Preskill, Quantum 2, 79 (2018)

J. Preskill, Quantum 2, 79 (2018)

2018
[50]

Scarani, M

V. Scarani, M. Ziman, P. ˇStelmachoviˇ c, N. Gisin, and V. Buˇ zek, Phys. Rev. Lett.88, 097905 (2002)

2002
[51]

Wu and T

J. Wu and T. H. Hsieh, Phys. Rev. Lett. 123, 220502 (2019)

2019
[52]

Kraus, H

B. Kraus, H. P. B¨ uchler, S. Diehl, A. Kantian, A. Micheli, and P. Zoller, Phys. Rev. A 78, 042307 (2008)

2008
[53]

Reiter, D

F. Reiter, D. Reeb, and A. S. Sørensen, Phys. Rev. Lett. 117, 040501 (2016)

2016
[54]

Farhi, J

E. Farhi, J. Goldstone, and S. Gutmann, A Quan- tum Approximate Optimization Algorithm (2014), arXiv:1411.4028

2014 arXiv
[55]

Jamio lkowski, Rep

A. Jamio lkowski, Rep. Math. Phys. 3, 275 (1972)

1972
[56]

Choi, Linear Algebra and its Applications 10, 285 (1975)

M.-D. Choi, Linear Algebra and its Applications 10, 285 (1975)

1975
[57]

M. J. Kastoryano, F. Reiter, and A. S. Sørensen, Phys. Rev. Lett. 106, 090502 (2011)

2011
[58]

Y. Lin, J. P. Gaebler, F. Reiter, T. R. Tan, R. Bowler, A. S. Sørensen, D. Leibfried, and D. J. Wineland, Nature 504, 415 (2013)

2013
[59]

Morigi, J

G. Morigi, J. Eschner, C. Cormick, Y. Lin, D. Leibfried, and D. J. Wineland, Phys. Rev. Lett.115, 200502 (2015)

2015
[60]

D. C. Cole, J. J. Wu, S. D. Erickson, P.-Y. Hou, A. C. Wilson, D. Leibfried, and F. Reiter, New J. Phys. 23, 073001 (2021)

2021
[61]

D. C. Cole, S. D. Erickson, G. Zarantonello, K. P. Horn, P.-Y. Hou, J. J. Wu, D. H. Slichter, F. Reiter, C. P. Koch, and D. Leibfried, Phys. Rev. Lett. 128, 080502 (2022)

2022
[62]

Malinowski, C

M. Malinowski, C. Zhang, V. Negnevitsky, I. Rojkov, F. Reiter, T.-L. Nguyen, M. Stadler, D. Kienzler, K. K. Mehta, and J. P. Home, Phys. Rev. Lett. 128, 080503 (2021)

2021
[63]

Doucet, F

E. Doucet, F. Reiter, L. Ranzani, and A. Kamal, Phys. Rev. Res. 2, 023370 (2020)

2020
[64]

Singkanipa and D

P. Singkanipa and D. A. Lidar, Beyond unital noise in variational quantum algorithms: noise-induced barren plateaus and fixed points (2024)

2024
[65]

P. M. Harrington, E. J. Mueller, and K. W. Murch, Na- ture Reviews Physics 4, 660 (2022)

2022
[66]

Kronwald, F

A. Kronwald, F. Marquardt, and A. A. Clerk, New J. Phys. 16, 063058 (2014)

2014
[67]

G. S. Agarwal and S. Huang, Phys. Rev. A 93, 043844 (2016)

2016
[68]

Rojkov, D

I. Rojkov, D. Layden, P. Cappellaro, J. Home, and F. Re- iter, Phys. Rev. Lett. 128, 140503 (2022)

2022
[69]

J. T. Barreiro, M. M¨ uller, P. Schindler, D. Nigg, T. Monz, M. Chwalla, M. Hennrich, C. F. Roos, P. Zoller, and R. Blatt, Nature 470, 486 (2011)

2011
[70]

Reiter, A

F. Reiter, A. S. Sørensen, P. Zoller, and C. A. Muschik, Nat. Commun. 8, 1822 (2017)

2017
[71]

Wang and J

C. Wang and J. M. Gertler, Phys. Rev. Res. 1, 033198 (2019)

2019
[72]

Rojkov, M

I. Rojkov, M. Simoni, E. Zapusek, F. Reiter, and J. Home, Stabilization of cat-state manifolds using non- linear reservoir engineering (2024)

2024
[73]

Van Mourik, E

M. Van Mourik, E. Zapusek, P. Hrmo, L. Gerster, R. Blatt, T. Monz, P. Schindler, and F. Reiter, Phys. Rev. Lett. 132, 040602 (2024)

2024
[74]

Reiter, L

F. Reiter, L. Tornberg, G. Johansson, and A. S. Sørensen, Phys. Rev. A 88, 032317 (2013)

2013
[75]

Zapusek, A

E. Zapusek, A. Javadi, and F. Reiter, Quantum Sci. Tech- nol. 8, 015001 (2023)

2023
[76]

Bakshi, A

A. Bakshi, A. Liu, A. Moitra, and E. Tang, (2024), arXiv:2403.16850

2024 arXiv
[77]

S. Wang, E. Fontana, M. Cerezo, K. Sharma, A. Sone, L. Cincio, and P. J. Coles, Nat. Commun. 12, 6961 (2021)

2021
[78]

Boixo, S

S. Boixo, S. V. Isakov, V. N. Smelyanskiy, R. Babbush, N. Ding, Z. Jiang, M. J. Bremner, J. M. Martinis, and H. Neven, Nat. Phys. 14, 595 (2018)

2018
[79]

A. M. Dalzell, N. Hunter-Jones, and F. G. S. L. Brand˜ ao, Random quantum circuits transform local noise into global white noise (2021)

2021
[80]

J.-F. Chen, K. S. Rai, P. Emonts, D. Farina, M. P lodzie´ n, P. Grzybowski, M. Lewenstein, and J. Tura, Boosting thermalization of classical and quantum many-body sys- tems (2024), arXiv:2411.03420

2024
[81]

Zheng, C

H. Zheng, C. Kang, G. S. Ravi, H. Wang, K. Setia, F. T. Chong, and J. Liu, in 2023 IEEE International Confer- ence on Quantum Computing and Engineering (QCE) (IEEE, Bellevue, W A, USA, 2023) pp. 236–245

2023
[82]

Sauvage, M

F. Sauvage, M. Larocca, P. J. Coles, and M. Cerezo, Quantum Sci. Technol. 9, 015029 (2024)

2024
[83]

B. T. Gard, L. Zhu, G. S. Barron, N. J. Mayhall, S. E. Economou, and E. Barnes, npj Quantum Inf. 6, 10 (2020)

2020
[84]

Park, Efficient ground state preparation in varia- tional quantum eigensolver with symmetry breaking lay- ers (2021)

C.-Y. Park, Efficient ground state preparation in varia- tional quantum eigensolver with symmetry breaking lay- ers (2021)

2021
[85]

J. J. Meyer, M. Mularski, E. Gil-Fuster, A. A. Mele, F. Arzani, A. Wilms, and J. Eisert, PRX Quantum 4, 010328 (2023)

2023
[86]

Larocca, F

M. Larocca, F. Sauvage, F. M. Sbahi, G. Verdon, P. J. Coles, and M. Cerezo, PRX Quantum 3, 030341 (2022)

2022
[87]

Schatzki, M

L. Schatzki, M. Larocca, Q. T. Nguyen, F. Sauvage, and M. Cerezo, npj Quantum Inf. 10, 12 (2024)

2024
[88]

Muser, E

T. Muser, E. Zapusek, V. Belis, and F. Reiter, Phys. Rev. A 110, 032434 (2024)

2024
[89]

Buˇ ca and T

B. Buˇ ca and T. Prosen, New J. Phys.14, 073007 (2012)

2012
[90]

de Groot, A

C. de Groot, A. Turzillo, and N. Schuch, Quantum 6, 856 (2022)

2022
[91]

Misguich and P

G. Misguich and P. Sindzingre, J. Phys.: Condens. Mat- ter 19, 145202 (2007)

2007
[92]

Y. Y. Lifshitz, E. Bairey, E. Arbel, G. Aleksandrowicz, H. Landa, and I. Arad, Practical Quantum State Tomog- raphy for Gibbs states (2023)

2023
[93]

A. Gu, L. Cincio, and P. J. Coles, Nat. Commun. 15, 312 (2024)

2024
[94]

Scholl, M

P. Scholl, M. Schuler, H. J. Williams, A. A. Eberharter, D. Barredo, K.-N. Schymik, V. Lienhard, L.-P. Henry, T. C. Lang, T. Lahaye, A. M. L¨ auchli, and A. Browaeys, Nature 595, 233 (2021)

2021
[95]

Y. Kim, A. Eddins, S. Anand, K. X. Wei, E. Van Den Berg, S. Rosenblatt, H. Nayfeh, Y. Wu, M. Zale- tel, K. Temme, and A. Kandala, Nature 618, 500 (2023)

2023
[96]

Sachdev, Quantum Phase Transitions, 2nd ed

S. Sachdev, Quantum Phase Transitions, 2nd ed. (Cam- bridge University Press, 2011)

2011
[97]

N. Moll, P. Barkoutsos, L. S. Bishop, J. M. Chow, A. Cross, D. J. Egger, S. Filipp, A. Fuhrer, J. M. Gam- betta, M. Ganzhorn, et al., Quantum Sci. Technol. 3, 030503 (2018)

2018
[98]

Consiglio, J

M. Consiglio, J. Settino, A. Giordano, C. Mastroianni, F. Plastina, S. Lorenzo, S. Maniscalco, J. Goold, and T. J. G. Apollaro, Phys. Rev. A 110, 012445 (2024)

2024
[99]

Reiter, F

F. Reiter, F. Lange, S. Jain, M. Grau, J. P. Home, and Z. Lenarˇ ciˇ c, Phys. Rev. Res.3, 033142 (2021). 20

2021
[100]

Schmale and H

T. Schmale and H. Weimer, Stabilizing quantum sim- ulations of lattice gauge theories by dissipation (2024), arXiv:2404.16454

2024 arXiv
[101]

M. M. Wauters, E. Ballini, A. Biella, and P. Hauke, Symmetry-protection Zeno phase transition in monitored lattice gauge theories (2024), arXiv:2405.18504

2024 arXiv
[102]

https://github.com/Elias-Hubb/VQT
[103]

Virtanen, R

P. Virtanen, R. Gommers, T. E. Oliphant, M. Haber- land, T. Reddy, D. Cournapeau, E. Burovski, P. Peter- son, W. Weckesser, J. Bright, et al., Nat. Methods 17, 261 (2020)

2020
[104]

Fulton and J

W. Fulton and J. Harris, Representation Theory, Gradu- ate Texts in Mathematics, Vol. 129 (Springer New York, New York, NY, 2004)

2004
[105]

Ragone, P

M. Ragone, P. Braccia, Q. T. Nguyen, L. Schatzki, P. J. Coles, F. Sauvage, M. Larocca, and M. Cerezo arxiv.2210.07980 (2022)

2022 arXiv
[106]

M. A. Nielsen and I. L. Chuang, Quantum computation and quantum information, 10th ed. (Cambridge Univer- sity Press, Cambridge ; New York, 2010)

2010
[107]

Lindblad, Communications in Mathematical Physics 48, 119 (1976)

G. Lindblad, Communications in Mathematical Physics 48, 119 (1976)

1976
[108]

Marvian, Phys

I. Marvian, Phys. Rev. B 95, 045111 (2017)

2017
[109]

de Groot, D

C. de Groot, D. T. Stephen, A. Molnar, and N. Schuch, J. Phys. A: Math. Theor. 53, 335302 (2020)

2020
[110]

Lewandowski and J

Z. Lewandowski and J. Szynal, J. Comput. Appl. Math. 99, 529 (1998)

1998
[111]

Terashima and M

H. Terashima and M. Ueda, International Journal of Quantum Information 03, 633 (2005)

2005
[112]

Reiter and A

F. Reiter and A. S. Sørensen, Phys. Rev. A 85, 032111 (2012)

2012

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Estimate by sampling from

[2] [2]

Estimate entropy by simulating subsystem

[3] [3]

internal symmetry internal & spatial no symmetry (c) FIG

Return updated parameters . internal symmetry internal & spatial no symmetry (c) FIG. 1. Overview. (a) The hybrid algorithm. The quantum circuit begins with an initial unitary layer U0, followed by m layers, which are composed of unitary operations Uθj interleaved with nonunitary layers Eλj . The parameters of all layers are trained using the free energy ...

[4] [4]

The entropy of a subsystem of size na is computed and rescaled to ap- proximate the entropy of the full system

Scaled Subsystem Entropy In the main text, we introduce a scaled entropy approx- imation for translation-invariant systems. The entropy of a subsystem of size na is computed and rescaled to ap- proximate the entropy of the full system. For further details, refer to Sec. IV of the main text

[5] [5]

[19] derive an analytic approxima- tion of the entropy of an ansatz that interleavesm unitary layers with m layers of depolarizing channels

Analytic Entropy approximation The authors of Ref. [19] derive an analytic approxima- tion of the entropy of an ansatz that interleavesm unitary layers with m layers of depolarizing channels. The au- thors calculate the entropy of the state produced by an ansatz that first applies all m layers of the depolarizing channel followed bym unitary layers. The a...

[6] [6]

(A6) Here ρj signifies the partial trace over all subsystems but the j-th one

V ariational Entropy Estimation Evaluating the entropy of a product state is as simple as S(ρ1 ⊗ρ2 ⊗ · · · ⊗ρn) = S(ρ1) +S(ρ2) +· · ·+S(ρn). (A6) Here ρj signifies the partial trace over all subsystems but the j-th one. For a general state, the entropy is upper bounded by the sum of entropies of subsystems S(ρ) ≤ S(ρ1) + S(ρ2) + · · ·+ S(ρn). (A7) Further...

[7] [7]

This allows us to benchmark the scaled subsystem entropy against the variational entropy estimation

Comparison for translation invariant system We consider a system with translation invariance. This allows us to benchmark the scaled subsystem entropy against the variational entropy estimation. We consider the Heisenberg model with a transverse field H = X k XkXk+1 + YjYk+1 + ZkZk+1 + X k Xk. (A10) We initialize the ansatz in the symmetry sector of the g...

[8] [8]

These definitions can be found in any group theory textbook [88]

Preliminaries We first review relevant concepts from group and rep- resentation theory. These definitions can be found in any group theory textbook [88]. For a review in the context of quantum machine learning consider [89]. Symmetries are organized in groups. Recall the defini- tion of a (finite) group. Definition B.1 (Group) A group consists of a set G ...