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 · submitted 2026-04-20 · score 8/10 (HIGH)

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

This paper resolves a long-standing concern about system-bath Gibbs-state preparation algorithms — the simple, ancilla-cheap alternative to full Lindbladian simulation. Earlier system-bath constructions (Lloyd '25, Hahn '26, Ding '25) inherited a non-commuting Lamb shift in the weak-coupling Lindbladian decomposition L = −i[H_Lamb, ·] + L_KMS; ensuring that the channel's fixed point stays close to the Gibbs state required pushing the support width of the envelope f(t) to infinity, inflating the simulation cost. The authors prove that as long as the transition part of L exactly satisfies KMS detailed balance, the discrete channel Φ_α hides an error-cancellation mechanism between the outer unitary U_S(T) and L_KMS that drives the fixed point to within O(α²) of the Gibbs state under constant interaction time T. End-to-end complexity drops from prior O(ε⁻⁴) / O(ε⁻²) bounds to O(ε⁻¹). For Yuan, the headline relevance is to Y5: structured Goemans–Williamson SDP relaxations via quantum Gibbs-state preparation. Faster, hardware-cheaper Gibbs-state preparation directly improves the resource estimates Y5 uses for the quantum (non-dequantised) regime.

Main contribution

Two informal main theorems plus a rigorous instantiation. Theorem III.1 (fixed-point approximation, informal): Whenever the leading-order Lindbladian decomposes as L = −i[H_Lamb, ·] + L_KMS with L_KMS exactly KMS-detailed-balanced, the unique fixed point of Φ_α obeys ‖ρ_fix(Φ_α) − ρ_β‖₁ = O(ε + α²·t_{mix,Φ_α}(ε)) for all ε > 0. Theorem III.2 (mixing time, informal): If L_KMS has spectral gap λ_gap and a quantitative bound on the Lamb-shift commutator with ρ_β^±1/4, then τ_mix = Õ(α⁻²/λ_gap) and ‖ρ_fix − ρ_β‖₁ = Õ(α²/λ_gap). Combined: end-to-end Hamiltonian-evolution complexity for ε-accuracy Gibbs preparation is Õ(ε⁻¹).

Key theorems / lemmas / algorithms

Theorem III.1 (Fixed-Point Approximation, Informal): Under leading-order approximation Φ_α(ρ) = U_S(T) ∘ exp(α²L) ∘ U_S(T) + O(α⁴) with the KMS decomposition (Eq. 3.2), the unique fixed point obeys ‖ρ_fix − ρ_β‖₁ = O(ε + α⁴ τ_mix(ε)).
Theorem III.2 (Mixing Time, Informal): If L_KMS has spectral gap λ_gap > 0 and ‖ρ_β^−1/4H_Lambρ_β^1/4 − ρ_β^1/4H_Lambρ_β^−1/4‖ = O(λ_gap), then for α = O(λ_gap^1/2), τ_mix = Õ(α⁻²λ_gap⁻¹) and ‖ρ_fix − ρ_β‖₁ = Õ(α²λ_gap⁻¹).
Theorem IV.1 + Corollary IV.2 (end-to-end for Ding's algorithm): Under explicit choices σ = Θ(β²/λ_gap), α² = Θ̃(ε λ_gap²/(β⁴‖H‖)), τ_mix = Õ(β⁵‖H‖²/(λ_gap³ε)); total Hamiltonian-evolution time Õ(n⁶ ε⁻¹ λ₀⁻⁴) for local Hamiltonians with ‖H‖ = Θ(n), λ_gap = Θ(λ₀/n).
Lemma (auxiliary fixed-point construction, §C): Build ρ* = ρ_β + α²E with E = ∫ e^−iHt Y e^iHt dν(t) for a carefully chosen spectral measure ν satisfying ν̂(ω) = ω μ̂(ω)/(1 − μ̂(2ω)); randomising T over distribution μ avoids the singularities at ω = kπ/(2T) that fixed T would produce.
Catalogue of compatible Hamiltonians: applies to high-temperature local spin (Rouzé '24 et al., λ_gap = Ω(1/n)), weakly interacting fermionic systems at all temperatures (Smid '25), weakly interacting spin systems (Tong '24), and 1D local Hamiltonians at all temperatures (Bergamaschi '26).

Detailed walkthrough

The paper sits in the active recent thread of quantum Gibbs-state preparation algorithms. Two major design philosophies coexist: (i) explicit KMS-detailed-balance Lindbladian simulation (Chen–Kastoryano–Gilyén '23, Ding '25), which guarantees the Gibbs state is an exact fixed point but requires implementing complicated jump operators, time-reversed Hamiltonian simulation, and large ancilla overhead; and (ii) the system-bath interaction framework (Lloyd '25, Hahn '26, Ding '25, Wang '25, Scandi '25), where a small bath register is weakly coupled to the system and dissipation arises from tracing out and resetting the bath. The system-bath approach is enormously simpler to implement, but its discrete channel Φ_α only approximates a Lindbladian, and the approximate generator has a Lamb-shift Hermitian term that need not commute with ρ_β — so the channel's true fixed point is biased away from the target Gibbs state.

Existing fixes (Hahn, Ding) take the support width σ of the envelope function f(t) to infinity, which forces the Lamb shift to commute with ρ_β asymptotically but inflates the per-iteration simulation cost and degrades the mixing time. The headline result here is that the asymptotic limit is not needed: if one engineers the system-bath interaction so that the transition part of L exactly satisfies KMS detailed balance (a constructive condition met by the cited Lloyd/Hahn/Ding constructions, with logarithmic-in-ε truncation error in T), then the discrete channel Φ_α at constant interaction time T already exhibits a hidden cancellation that suppresses the Lamb-shift bias.

The mechanism is laid out in the proof overview (§V). Writing ρ_fix = ρ_β + α²E + O(α⁴) and applying Φ_α's three stages (forward Hamiltonian evolution, Lindbladian, backward Hamiltonian evolution), the matching condition at order α² becomes a linear equation for E in the eigenbasis of H. After taking the expectation over a random evolution time T ∼ μ, one obtains E_j,k = ν̂(λ_k−λ_j) · iconic Lamb-shift term, where ν̂(ω) = ω μ̂(ω)/(1 − μ̂(2ω)). Crucially, fixing T (rather than randomising) makes ν̂ singular at ω = kπ/(2T) with non-decaying tails — randomising T over μ smooths out these poles and bounds ‖ν‖_TV. The auxiliary state ρ* = ρ_β + α²E is then proved to be an approximate fixed point with controlled distance to ρ_β; uniqueness of the true fixed point + perturbation theory pins it within trace distance O(α²).

The mixing-time analysis (§II.D, §VI) uses a spectral-gap perturbation argument similar to Wang '25 but adapted to the system-bath setting. The Lamb-shift commutator condition (Eq. 3.5) ensures that L's spectral gap is comparable to that of L_KMS. Combining gives the τ_mix = Õ(α⁻²/λ_gap) scaling. With the explicit parameter choice α² ∝ ε λ_gap²/(β⁴‖H‖) and σ ∝ β²/λ_gap, the end-to-end Hamiltonian-evolution time becomes Õ(β⁵‖H‖²/(λ_gap³ε)). For local Hamiltonians on n qubits with ‖H‖ = Θ(n) and λ_gap = Θ(1/n) (a regime supported by recent rapid-mixing results for high-temperature local spin chains, weakly interacting fermions, weakly interacting spins, and 1D local Hamiltonians), this gives N = Õ(n⁵/(λ₀³ε)) iterations and total evolution time Õ(n⁶/(λ₀⁴ε)). The improvement from O(ε⁻⁴) (Slezak '26 analysis of Ding) and O(ε⁻²) (Wang '25) to O(ε⁻¹) matches the first-order Lindbladian time-step expectation and is reportedly the best ε-dependence known for this algorithm family.

A subtle technical caveat: the application to Hahn '26 / Lloyd '25 (§IV.B) requires a small modification of the bath construction (Setup 4.5) — sampling the bath energy h from a Gaussian compensates for the σ-induced concentration of f̂ at ω=0 that would otherwise shrink L_KMS's spectral gap. This is presented as an algorithmic adjustment that does not change the overall framework but is essential for the spectral-gap lower bound. The paper is honest that the gap-perturbation condition (Eq. 3.5) is technical: even if it fails, mixing might still be efficient in practice — clarifying this is left as open.

Figures

No figures extracted from source — the paper is theory-only and uses inline TikZ for circuit diagrams (no separate raster or PDF figures).

Citations to Yuan's papers

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

Overlap with Y1–Y6

Y5 (Goemans–Williamson via Pauli-sparse Gibbs states) — strongest overlap. Y5 demonstrates exponential speed-ups for structured GW relaxations using quantum Gibbs states at the heart of the algorithm. Any improvement in the cost of preparing those Gibbs states (in terms of ε, β, ‖H‖, λ_gap) directly improves Y5's resource estimates in the genuinely-quantum (non-dequantised) regime. The system-bath construction here is also strictly more hardware-friendly than full Lindbladian simulation, which matches Y5's framing of "what's actually realisable on near-term and early-FT machines". The compatible-Hamiltonian list (1D local Hamiltonians at all temperatures via Bergamaschi '26) is suggestive: Y5's Pauli-sparse Hamiltonians might admit similar rapid-mixing guarantees and become eligible.
Y3 (QAOA DGMVP, thermal noise crossover) — methodological adjacency. Y3 uses thermal-noise modelling (the same Gibbs/dissipative formalism) to argue that thermal relaxation precludes quantum advantage in DGMVP. The Lindbladian/Lamb-shift analysis here is in the same mathematical universe, though the application is preparation rather than noise modelling.
Y1, Y2, Y4, Y6 — no meaningful overlap.

Recommended action for Yuan

Read carefully and use as a citation in any Y5 follow-up. The new Õ(ε⁻¹) end-to-end Hamiltonian-evolution complexity is the strongest known precision dependence for system-bath Gibbs preparation; if Y5's quantum-only analysis cites the older O(ε⁻²) or O(ε⁻⁴) bounds, this paper supersedes them and tightens the resource estimate.
Check whether Y5's Pauli-sparse Hamiltonians fit one of the four rapid-mixing families (high-T local spin, weakly interacting fermions, weakly interacting spins, 1D local at all T). If yes, the constants in Theorem IV.1 give a closed-form n-dependent bound directly applicable to Y5's quantum-regime estimates. If no, this is a concrete open problem worth flagging in Y5's "future work" section.
Discuss with Aolita / Franca (Y5 co-authors) whether the system-bath construction can replace Lindbladian simulation in any of Y5's algorithmic stages — given the implementation simplicity, this could materially change Y5's near-term feasibility story.