Overcoming the Lamb Shift in System-Bath Models via KMS Detailed Balance: High-Accuracy Thermalization with Time-Bounded Interactions

Hongrui Chen, Zhiyan Ding, Ruizhe Zhang · arXiv:2604.15616 · new submission 2026-04-20 · score 8/10

Abstract

We investigate quantum thermal state preparation algorithms based on system-bath interactions and uncover a surprising phenomenon in the weak-coupling regime. We rigorously prove that, if the system-bath interaction is engineered so that the transition part of the approximate Lindbladian generator satisfies the KMS detailed balance condition, then the unique fixed point of the dynamics can be made arbitrarily close to the Gibbs state in the weak-coupling limit, regardless of the structure of the Lamb shift term. Importantly, this remains true even when the approximate Lindbladian differs substantially from the ideal Davies generator and the Lamb shift term does not commute with the thermal state. Our result shows that the role of the KMS detailed balance condition extends well beyond standard Lindbladian dynamics, serving as a general principle for a broader class of dissipative systems. Furthermore, by combining this with a general perturbation framework, we bound the mixing time of the dynamics and establish an end-to-end complexity of O(ε⁻¹) for Gibbs state preparation. These guarantees apply to any Hamiltonian for which the corresponding KMS-detailed-balance Lindbladian is known to mix rapidly.

Executive summary

Chen, Ding, and Zhang resolve a long-standing analytical bottleneck in system-bath thermal state preparation: previous algorithms required the bath-correlation support width σ to be pushed to infinity (so that the Lamb-shift term H_Lamb commutes with the Gibbs state), inflating both simulation cost and mixing time. This paper proves that an unexpected cancellation between the outer unitary evolution U_S(T) and the KMS-detailed-balance dissipative part L_KMS suppresses the steady-state bias even at constant σ, provided the transition part of the effective Lindbladian exactly satisfies KMS detailed balance. The end-to-end complexity for ε-accurate Gibbs state preparation drops from 1/ε⁴ (Slezak et al.) or 1/ε² (Wang et al.) to 1/ε. For Yuan, this is the most directly relevant Gibbs-state-preparation paper of the week: Y5's quantum and quantum-inspired SDP speed-ups are powered by Pauli-sparse Gibbs samplers, and a sharper KMS-detailed-balance complexity bound feeds straight into the end-to-end SDP runtime.

Main contribution

The paper studies the quantum channel Φ_α(ρ) := E[Tr_E(U_α(T)(ρ ⊗ ρ_E)U_α†(T))] arising from one round of weak system-bath coupling (Eq. 1). In the weak-coupling limit, this channel factors as Φ_α ≈ U_S(T) ∘ exp(α²L) ∘ U_S(T) + O(α⁴), with the effective Lindbladian L decomposing into a coherent Lamb-shift term −i[H_Lamb, ·] and a KMS-detailed-balance dissipative term L_KMS. The first main result (Theorem III.1, "Fixed-Point Approximation") proves the fixed point of Φ_α is O(α² · t_mix)-close to the Gibbs state ρ_β as long as the transition part of L exactly satisfies KMS detailed balance — even when H_Lamb does not commute with ρ_β. The second main result (Theorem III.2, "Mixing Time") bounds the mixing time as Õ(α⁻² λ_gap⁻¹), with the precise O(α²·λ_gap⁻¹) fixed-point error implied. Specialised to the Ding et al. (2025) construction, these yield (Theorem III.4 + Corollary III.5) a total Hamiltonian-evolution-time end-to-end complexity of Õ(n⁶ ε⁻¹ λ₀⁻⁴), with explicit applicability to (i) high-temperature local spin Hamiltonians, (ii) weakly interacting fermions at all temperatures, (iii) weakly interacting spins, and (iv) 1D local Hamiltonians at all temperatures.

Key theorems / lemmas / algorithms

Theorem III.1 (Fixed-Point Approximation, Informal): Under the leading-order channel approximation Φ_α(ρ) = U_S(T) ∘ exp(α²L) ∘ U_S(T) + O(α⁴) with L = −i[H_Lamb, ·] + L_KMS, the fixed point satisfies ‖ρ_fix(Φ_α) − ρ_β‖₁ = O(ε + α² t_mix(ε)). The KMS detailed balance condition on L_KMS is essential; without it the steady-state bias does not cancel.
Theorem III.2 (Mixing Time, Informal): If L_KMS has spectral gap λ_gap > 0 and the Lamb-shift commutator defect satisfies ‖ρ_β⁻¹⁄⁴ H_Lamb ρ_β¹⁄⁴ − ρ_β¹⁄⁴ H_Lamb ρ_β⁻¹⁄⁴‖ = O(λ_gap), then for α = O(λ_gap¹⁄²), τ_mix(Φ_α) = Õ(α⁻² λ_gap⁻¹) and ‖ρ_fix − ρ_β‖₁ = Õ(α² λ_gap⁻¹).
Theorem III.4 (End-to-end for Ding et al. construction): Under explicit parameter choices σ = Θ(β² / λ_gap), α² = Θ̃(ε λ_gap² / (β⁴ ‖H‖)), T₀ ≥ 2σ √log((α²β log σ)⁻¹), the mixing time is Õ(β⁵ ‖H‖² / (λ_gap³ ε)) with fixed-point error ≤ ε.
Corollary III.5: For local Hamiltonians with ‖H‖ = Θ(n) and λ_gap = Θ(λ₀/n), the algorithm uses N = Õ(n⁵ ε⁻¹ λ₀⁻³) iterations and total evolution time Õ(n⁶ ε⁻¹ λ₀⁻⁴).
Setup III.7 (Modified Lloyd construction): Decouples σ from the single-qubit bath energy h by sampling h ~ N(β⁻¹, 2β⁻² − 2σ⁻¹); the resulting Lindbladian retains the same decomposition as the Ding setup, yielding the same end-to-end complexity bounds.
Asymptotic cancellation lemma (Section IV, "Heuristic idea"): Writing ρ_fix = ρ_β + α²E + O(α⁴) and expanding through one period of Φ_α, the fixed-point equation reduces to U_S(2T) E U_S†(2T) − E + L_KMS(E) − i[U_S(T) H_Lamb U_S†(T) − H_Lamb, ρ_β] = 0. The outer unitary evolutions are essential — without them, the equation reduces to one whose solution is biased away from ρ_β.

Detailed walkthrough

The setting is quantum Gibbs state preparation through repeated weak coupling between a system register (carrying the target Hamiltonian H) and a small ancillary bath that is reset each iteration. Compared with Lindblad-simulation-based approaches (Chen–Kastoryano–Gilyen 2023, Ding 2025), system-bath models are attractive because the jump operators are not implemented explicitly — they emerge from simulating a simple joint Hamiltonian — but theoretical guarantees have been weaker. Specifically, the effective Lindbladian arising from the leading-order approximation Φ_α(ρ) ≈ U_S(T) ∘ exp(α²L) ∘ U_S(T) (Eq. 2) decomposes as L = −i[H_Lamb, ·] + L_KMS, where the Lamb-shift Hamiltonian H_Lamb in general fails to commute with ρ_β, even when L_KMS is engineered to exactly satisfy KMS detailed balance.

Prior work (Hahn et al. 2026, Ding et al. 2025, Lloyd et al. 2025) addressed this by tuning the support width σ of the envelope function f(t) in the coupling operator to infinity, which forces H_Lamb to commute with ρ_β asymptotically. The trade-off is dire: the per-iteration simulation cost balloons, and the mixing time can grow significantly. Slezak et al. and Wang et al. quantified the end-to-end complexity at O(1/ε⁴) and O(1/ε²) respectively under these asymptotic-σ regimes. Section II of the present paper revisits the literature and articulates Q1 (faithful efficient realisation of Lindbladian thermalisation) and Q2 (algorithmic benefits of KMS-detailed balance) as the open questions; the paper's headline contribution is a rigorous positive answer to Q2 — and a partial answer to Q1.

The technical core is in Section IV (informal proof) and the appendix. The heuristic argument writes the fixed point as a formal asymptotic ρ_fix = ρ_β + α²E + O(α⁴), applies the three stages of Φ_α (forward unitary, dissipation, backward unitary) and matches order-α² terms. The resulting equation for the correction E contains both the dissipative action L_KMS(E) and the conjugated Lamb-shift commutator −i[U_S(T) H_Lamb U_S†(T) − H_Lamb, ρ_β]. The fact that the Lamb-shift contribution appears in a "before-minus-after" conjugated form is the cancellation the authors exploit: when H_Lamb approximately commutes with U_S(T) (which holds for the Ding/Hahn/Lloyd constructions in the weak-coupling regime), the conjugation difference is small and the residual bias is suppressed. Without the outer unitary evolutions, this conjugation cancellation does not occur and the resulting fixed point of exp(α²L) alone retains the O(α⁰) bias. This is the conceptual reason the unitary evolution U_S(T) — often treated as a "technical detail" in prior literature — turns out to be essential.

The mixing-time analysis (Theorem III.2) is built on a perturbation argument: as long as the spectral gap of L_KMS is positive and the Lamb-shift commutator defect is O(λ_gap), the perturbed full Lindbladian L retains a comparable contraction rate. This is followed by a worst-case fixed-point bound that picks up the α²/λ_gap scaling. Combining the mixing-time bound with the channel approximation error gives the α = O((ε λ_gap)¹⁄²) condition that the authors flag explicitly.

Section III.B is the applications block. For the Ding et al. (2025) construction, the authors fix σ = Θ(β²/λ_gap), α² = Θ̃(ε λ_gap² / (β⁴ ‖H‖)), and a randomised evolution time T drawn from a specified distribution. The resulting end-to-end complexity is Õ(β⁵ ‖H‖² / (λ_gap³ ε)) iterations, with each iteration consuming Õ(1/λ_gap) Hamiltonian-evolution time. Critically, this linear dependence on 1/ε is the best possible at first order. Specialised to local Hamiltonians with ‖H‖ = Θ(n) and λ_gap = Θ(λ₀/n) (a regime covered by rapid-mixing results of Rouz et al., Smid et al., Tong et al., Bergamaschi et al. for high-temperature spins, weakly interacting fermions/spins, and 1D local Hamiltonians), the total Hamiltonian evolution time scales as Õ(n⁶ ε⁻¹ λ₀⁻⁴).

For the Hahn et al. and Lloyd et al. setups, the spectral-gap perturbation step is more delicate because the Fourier transform f̂ is concentrated near ω = 0 with width 1/σ, suppressing transitions between distant energy levels and possibly making λ_gap(L_KMS) = O(1/σ). The authors propose Setup III.7, a slight modification of Lloyd that decouples σ from the bath energy parameter h by sampling h ~ N(β⁻¹, 2β⁻² − 2σ⁻¹). The transition rates between distant energy levels are restored by the randomised h, while the upward/downward transitions still satisfy KMS detailed balance. The same end-to-end bounds then carry over.

The discussion (Section V) flags three open questions worth Yuan's attention: (i) the trade-off between α²σ → 0 (small coupling at finite support) and α²σ = Θ(1) (the Wang et al. regime) is not fully understood; (ii) the polynomial n-dependence has two sources — the first-order Lindbladian-approximation mixing time scaling as n², and a spectral-gap perturbation requirement T ~ σ = O(n) — both of which the authors believe can be tightened; (iii) the unitary evolution U_S(T) is essential for the fixed-point bias cancellation but plays no role in the mixing-time analysis; it remains open whether it can be used to accelerate mixing as in recent unitary-drift papers (e.g. PRL 134, 140405).

Figures

No figures extracted from source.

Citations to Yuan's papers

No direct citation to any of Y1–Y6 found in bibliography.

Overlap with Y1–Y6

Y5 (Pauli-sparse Gibbs SDP, GW speed-ups): Directly relevant. Y5's quantum and quantum-inspired SDP speed-ups are built on top of Gibbs state samplers for Pauli-sparse Hamiltonians; the per-sample preparation complexity is the dominant cost in the end-to-end SDP runtime. This paper's improvement from 1/ε⁴ (Slezak) or 1/ε² (Wang) to 1/ε in the precision dependence, and its explicit applicability to high-temperature local spin Hamiltonians and 1D local Hamiltonians at all temperatures, is the right kind of result to feed into Y5's complexity headline. Concretely: if the Pauli-sparse Hamiltonian used inside Y5 has a positive KMS spectral gap (which it does for the high-temperature and 1D regimes covered here), the system-bath construction here gives a near-term-friendly alternative to block-encoded Lindblad simulation and improves the ε-dependence of the per-sample complexity.
Y4 (Grover + ADMM cardinality-constrained BO): Tangential. Y4's quantum subroutine is Grover, not Gibbs sampling. A future hybrid algorithm that uses Gibbs samples to seed warm-starting (cf. Y1) would inherit this improved complexity, but the connection is indirect.
Y1 (warm-started iterative QAOA): Tangential. The KMS construction here is for thermal-state preparation, not optimisation; no method overlap.
Y2, Y3 (portfolio QAOA): No method overlap.
Y6 (PBR test on Heron2): No overlap.

Recommended action for Yuan

Cite in any Y5 follow-up / extended version. The end-to-end Õ(n⁶ ε⁻¹) bound for high-temperature local spin Hamiltonians and 1D local Hamiltonians is the right complexity primitive to plug into Y5's end-to-end SDP runtime. The Pauli sparsity structure assumed by Y5 fits cleanly into the "weakly interacting spin systems" category (Theorem III.4 applicability list).
Read Section IV (informal proof) and Corollary III.5. The asymptotic-cancellation mechanism is conceptually clean and worth understanding for any future Gibbs-prep-based SDP work; Corollary III.5 gives the exact scaling that should be quoted.
Consider emailing the authors (Hongrui Chen / Zhiyan Ding / Ruizhe Zhang at Stanford / UCB / Simons). If Y5's next iteration wants to claim end-to-end SDP complexity from raw Hamiltonian to ε-optimal solution, the authors' framework is the right tool and a co-author conversation could compress the ε-dependence improvement directly into Y5's headline claims.