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

Authors: Hongrui Chen, Zhiyan Ding, Ruizhe Zhang · arXiv: 2604.15616 · new submission, quant-ph · 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

Chen, Ding, and Zhang prove that the system-bath interaction framework for quantum Gibbs state preparation can achieve high-accuracy thermal state output even when the residual Lamb shift does not commute with ρ_β, provided the dissipative (transition) part of the effective Lindbladian satisfies KMS detailed balance exactly. They identify a hidden cancellation between the outer Hamiltonian-evolution slabs U_S(T) and the dissipative Lindbladian inside the discrete repeated-interaction channel Φ_α, and turn it into a rigorous end-to-end Gibbs preparation complexity of Õ(β⁵ ‖H‖² λ_gap^-3 ε^-1) — linear in 1/ε, improving the prior 1/ε² and 1/ε⁴ bounds. For Yuan, this is the most directly relevant primitive in today’s digest: the same KMS-detailed-balance Gibbs state preparation that powers Yuan’s Goemans-Williamson speed-ups in Y5.

Main contribution

The paper resolves an open question (their Q2): whether KMS detailed balance can confer algorithmic benefits in the system-bath interaction framework, where the effective Lindbladian decomposes as L(ρ) = -i[H_Lamb, ρ] + L_KMS(ρ) and H_Lamb generally fails to commute with ρ_β. Prior analyses (Hahn et al., Lloyd et al., Ding et al., Wang et al.) all required the envelope-function support width to diverge so that H_Lamb commuted asymptotically with ρ_β, paying a polynomial overhead in 1/ε. The authors instead show that for finite (constant) interaction time T and KMS-detailed-balance L_KMS, the discrete channel Φ_α = U_S(T) · e^α²L · U_S(T) + O(α⁴) exhibits a cancellation that suppresses the steady-state bias to O(α²). Combined with a perturbation argument that bounds the spectral gap of the full L by that of L_KMS, this yields the linear-in-1/ε end-to-end runtime guarantee.

Key theorems / lemmas / algorithms

Theorem III.1 (Fixed-Point Approximation, Informal): If Φ_α(ρ) = U_S(T) ∘ exp(α²L) ∘ U_S(T)(ρ) + O(α⁴) with L = -i[H_Lamb, ·] + L_KMS and L_KMS KMS-detailed-balanced, then for any ε>0 the unique fixed point obeys ‖ρ_fix(Φ_α) - ρ_β‖₁ = O(ε + α⁴τ_mix(ε)) = O(ε + α² t_mix(ε)).
Theorem III.2 (Mixing Time, Informal): If additionally 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) the mixing time satisfies τ_mix = Õ(α^-2λ_gap^-1) and ‖ρ_fix - ρ_β‖₁ = Õ(α²λ_gap^-1).
Theorem III.3 (End-to-end, rigorous, applied to Ding et al. 2025): Under the explicit Gaussian-bath setup with σ = Θ(β²/λ_gap) and α² = ÕΘ(ελ_gap²/(β⁴‖H‖)), the mixing time is Õ(β⁵‖H‖²λ_gap^-3ε^-1) with fixed-point error ≤ ε.
Corollary III.4: For local n-qubit Hamiltonians with ‖H‖ = Θ(n), λ_gap = Θ(λ₀/n), the iteration count is N = Õ(n⁵ε^-1λ₀^-3) and total Hamiltonian-evolution time is Õ(n⁶ε^-1λ₀^-4).
Setup III.5 (modified Lloyd-style construction): A new bath construction with H_E = -hZ/2, h ~ N(β^-1, 2β^-2 - 2σ^-1), that retains a uniform spectral-gap lower bound while keeping the bath always reset to |0⟩.
Lemma B.1 (nu existence, §C): Construction of a probability distribution μ whose Fourier transform ν̂(ω) = ωμ̂(ω)/(1 - μ̂(2ω)) has bounded total variation. Randomising the per-iteration evolution time T ∼ μ is essential — for fixed T the corresponding spectral measure has unbounded TV norm. The chosen μ₀(t) = (t-1)³e^-(t-1)/6 · 1_t≥1 realises the bound.
Application table (Table I): Compares fixed-point error, mixing time, interaction strength, support width, and end-to-end cost across Langbehn et al., Shtanko-Chen, Hagan et al., Hahn et al., Lloyd et al., Scandi et al., and Ding et al.; all prior linear-in-ε rows showed support-width +∞ with cost 1/ε² or 1/ε⁴; the present work upgrades all three modified algorithms to support-width Θ(1) with end-to-end cost O(1/ε).

Detailed walkthrough

The paper’s technical core lives in Sections III, IV, and the heuristic of Section IV is the cleanest entry point. The system-bath channel takes the form Φ_α(ρ) = E[Tr_E(U_α(T)(ρ ⊗ ρ_E)U_α(T)^†)] with engineered total Hamiltonian H_α(t) = H + H_E + α(f(t)A_S ⊗ B_E + h.c.). In the weak-coupling limit, the channel admits the asymptotic decomposition Φ_α ≈ U_S(T) ∘ e^α²L ∘ U_S(T), with L decomposing into the noncommuting Lamb shift and a KMS-detailed-balance dissipator. The standard continuous Lindbladian intuition is that a noncommuting Lamb shift biases the steady state, since [H_Lamb, ρ_β] ≠ 0 means ρ_β is not a fixed point of L. Existing work has worked around this by sending the envelope width σ to infinity so that [H_Lamb, ρ_β] = O(1/σ) vanishes, but this inflates simulation cost.

The authors’ key insight is that the discrete channel Φ_α — with finite-time conjugation by U_S(T) on each side — behaves differently. Section IV (informal proof) writes the fixed-point ansatz ρ_fix = ρ_β + α²E + O(α⁴), applies the three stages (forward U_S(T), Lindbladian e^α²L, backward U_S(T)), uses [H, ρ_β] = 0 and L_KMS(ρ_β) = 0, and matches α² terms. After taking the expectation over the random T ∼ μ, one obtains an explicit equation for E in the energy eigenbasis: E_j,k = μ̂(2(λ_k-λ_j))E_j,k + μ̂(λ_k-λ_j) · (-i(H_Lamb)_j,k(e^-βλ_k-e^-βλ_j))/Z_β. Solving and introducing a spectral measure ν with ν̂(ω) = ωμ̂(ω)/(1-μ̂(2ω)) rewrites E = ∫ e^-iHt Y e^iHt dν(t) for a Y that packages the Lamb shift and thermal data. The final bound becomes ‖E‖₁ = O(‖Y‖₁‖ν‖_TV).

This is where the randomisation of T enters as a non-trivial design choice: if T is fixed, ν̂(ω) = ωe^iTω/(1-e^2iTω) has poles at ω = kπ/(2T), the corresponding measure ν(t) has a non-decaying tail, and the TV norm diverges. Lemma B.1 picks μ₀(t) = (t-1)³e^-(t-1)/6 · 1_t≥1, and shows that this Gamma-tail randomisation makes ν have bounded TV norm. Section IV then upgrades the heuristic to a rigorous proof: rather than justifying the α-expansion directly, the authors construct a candidate trace-one Hermitian operator ρ* = ρ_β + α²E using the integral formula, verify it is an approximate fixed point of Φ_α, and use a perturbation-of-fixed-points argument to conclude ‖ρ_fix - ρ*‖₁ ≤ O(α⁴ · τ_mix).

Section III.B applies the general Theorems III.1-III.2 to three concrete prior algorithms: Ding et al. (Setup III.6, Gaussian bath with width σ), Hahn et al., and Lloyd et al. (a single-qubit bath). For Ding et al., Theorem III.3 plugs in σ = Θ(β²/λ_gap) and α² = ÕΘ(ελ_gap²/(β⁴‖H‖)) and reads off the mixing time and corollary scaling. For Hahn et al. and Lloyd et al., a complication is that their Fourier-localised f makes &hat;f concentrate near ω=0 with width σ^-1; for large σ, transitions between distant energy levels are suppressed and the spectral gap of L_KMS can decay. The authors therefore propose a modification (Setup III.5): keep the small-σ envelope but sample the bath frequency h from a Gaussian distribution. This decouples the spectral-gap bound from σ and recovers the same end-to-end O(1/ε) guarantee, while keeping the bath always initialised in |0⟩.

The mixing-time analysis (Section IV.B-D) uses the framework of Wang et al. 2025, treating the Lamb shift as a perturbation to L_KMS and bounding the contraction rate. Crucially, the perturbation bound only needs the technical condition ‖ρ_β^-1/4H_Lambρ_β^1/4 - ρ_β^1/4H_Lambρ_β^-1/4‖ = O(λ_gap), not the much stronger [H_Lamb, ρ_β] = 0. This is what unlocks the constant-T regime. Section III.C lists concrete Hamiltonian classes whose L_KMS spectral gap λ_gap = Ω(1/n) is already known: high-temperature local spin Hamiltonians, weakly interacting fermionic systems at all temperatures, weakly interacting spin systems at all temperatures, and 1D local Hamiltonians at all temperatures — for each of which the present framework now yields N = Õ(n⁵/ε) iterations and Õ(n⁶/ε) total evolution time.

No figures extracted from source — the paper is theory-only with one comparison table and no \begin{figure} blocks.

Citations to Yuan's papers

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

Overlap with Y1–Y6

Y5 — GW speed-ups via Gibbs states + Pauli sparsity (arXiv:2510.08292) [HIGH overlap, method axis]. Y5’s exponential speed-ups for structured Goemans-Williamson SDPs depend on preparing quantum Gibbs states e^-βH/Z for Pauli-sparse Hamiltonians. The present paper provides exactly such a preparation primitive with the best known O(1/ε) precision dependence and an end-to-end Hamiltonian-evolution-time guarantee tied to spectral-gap assumptions that hold for the local spin / fermionic / 1D classes most relevant to MaxCut-style SDP reductions. If Y5’s end-to-end runtime currently uses an older Gibbs primitive (e.g., Chen-Kastoryano-Gilyen, or Hahn et al.’s system-bath construction with σ → ∞), substituting this paper’s system-bath result with constant T tightens the ε-dependence and removes the support-width penalty.
Y3 — QAOA DGMVP portfolio (arXiv:2410.16265) [LOW direct overlap]. No method overlap, but Y3’s ‘thermal relaxation precludes quantum advantage’ conclusion is in the same regime as the dissipative-thermalisation analyses here. The dissipative dynamics that break Y3’s QAOA are the same kind that this paper controls rigorously. Conceptually adjacent only.
Y1, Y2, Y4, Y6 — no method/scope/conclusion overlap.

Recommended action for Yuan

Read the proof overview (Section IV) and Theorem III.3 carefully. The randomised-T trick + the explicit μ₀(t) = (t-1)³e^-(t-1)/6 distribution is a non-obvious algorithmic choice that delivers the linear-in-1/ε scaling. Y5’s next iteration could potentially adopt the same randomisation if Gibbs preparation is on the runtime critical path.
Check whether Y5’s analysis already uses the O(1/ε²) Wang et al. 2025 bound — if so, plug in this paper’s O(1/ε) bound and recompute the end-to-end SDP runtime. Even a constant-factor improvement in the ε-dependence shifts the regime where Y5’s claimed exponential speed-ups become numerically demonstrable on near-term hardware.
Email Chen / Ding / Zhang. Their framework explicitly targets early fault-tolerant Gibbs preparation; Y5’s Pauli-sparse Hamiltonians fall squarely within the ‘weakly interacting spin systems at arbitrary temperature’ regime where they prove λ_gap = Ω(1/n). A natural collaboration is whether Pauli-sparse structure can sharpen the n⁶/ε runtime further.