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, 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

This paper proves that system-bath-interaction Gibbs-state preparation algorithms — which are far cheaper to run than full Lindbladian simulators because they don't need block-encoded jump operators or clock registers — can reach ε-accuracy in linear time in 1/ε, not the previous 1/ε⁴ or 1/ε². The trick is showing that an exact KMS detailed-balance condition on the transition part of the effective Lindbladian induces an "error cancellation" between the outer Hamiltonian evolution U_S(T) and the noncommuting Lamb-shift term, even at constant interaction time T. For Yuan, this is the most direct method-overlap result of the week with Y5: quantum SDP via Pauli-sparse Gibbs states relies entirely on the cost of preparing a Gibbs state, and a 1/ε bound on that cost translates directly into a tighter end-to-end SDP cost.

Main contribution

Earlier system-bath constructions (Lloyd et al., Hahn et al., Ding et al.) decomposed the leading-order effective Lindbladian as L = -i[H_Lamb, ·] + L_KMS. KMS detailed balance was known to make L_KMS fix ρ_β, but H_Lamb does not commute with ρ_β, so the full L doesn't fix ρ_β and a steady-state bias persists. Prior work removed the bias by sending the envelope-function support width σ to +∞, but this enlarges per-iteration cost and inflates mixing time, yielding only 1/ε²–1/ε⁴ end-to-end scaling. The present authors show that even at constant σ, an error-cancellation mechanism between U_S(T) and L_KMS drives the bias to O(α²), and a perturbation argument bounds the mixing time of the perturbed channel as Õ(α⁻²/λ_gap). Choosing α² = Õ(ελ_gap²/(β⁴‖H‖)) then yields O(ε⁻¹) total Hamiltonian evolution time.

Key theorems / lemmas / algorithms

Theorem III.1 (Fixed-Point Approximation, Informal): If the effective Lindbladian decomposes as L = -i[H_Lamb, ·] + L_KMS with L_KMS KMS-detailed-balanced and Φ_α has a unique fixed point with finite mixing time, then ‖ρ_fix(Φ_α) − ρ_β‖₁ = O(ε + α² t_mix(ε)) for any ε > 0.
Theorem III.2 (Mixing Time, Informal): Under the same setup, if L_KMS has spectral gap λ_gap > 0 and the Lamb shift satisfies a Gibbs-commutation-defect bound of order λ_gap, then for α = O(λ_gap^1/2), τ_mix(Φ_α) = Õ(α⁻²λ_gap⁻¹) and ‖ρ_fix − ρ_β‖₁ = Õ(α²λ_gap⁻¹).
Theorem IV.1 (Rigorous Ding-et-al. end-to-end bound): Under the Ding-et al. setup with explicit Gaussian envelope f(t), bath-energy function g(ω), and random evolution time drawn from μ₀(t/T₀)/T₀ with T₀ = 2σ √log((α²β log σ)⁻¹), τ_mix(ε) = Õ(β⁵‖H‖² / (λ_gap³ ε)).
Corollary IV.2: For a local Hamiltonian with ‖H‖ = Θ(n) and λ_gap = Θ(λ₀/n), the required iteration count is Õ(n⁵ε⁻¹λ₀⁻³) and total evolution time is Õ(n⁶ε⁻¹λ₀⁻⁴).
Setup IV.3 (Modified Lloyd): A "drift the bath-Hamiltonian field h over a Gaussian distribution" tweak that keeps the spectral gap from collapsing as σ → ∞, so the same end-to-end bounds apply.

Detailed walkthrough

Section II reviews the background: KMS detailed balance for channels and Lindbladians (Sec. II.B), spectral gap and mixing time (Sec. II.C), and the system-bath approximation framework (Sec. II.D). The system-bath channel acts as Φ_α(ρ) = E[Tr_E(U_α(T)(ρ ⊗ ρ_E) U_α(T)^†)] under joint Hamiltonian evolution H_α(t) = H + H_E + α(f(t) A_S ⊗ B_E + h.c.). In the weak-coupling limit, Φ_α(ρ) = U_S(T) ° exp(α²L) ° U_S(T)(ρ) + O(α⁴).

Section III states the two general theorems. The crucial conceptual move is that the decomposition L = -i[H_Lamb,·] + L_KMS with exact KMS detailed balance for L_KMS requires careful construction of the system-bath interaction; it does not happen automatically. Once it does, the asymptotic expansion ρ_fix = ρ_β + α² E + O(α⁴) works.

The heuristic argument in Sec. V (eqs. (V.1)–(V.4)) is illuminating: apply forward evolution, Lindbladian, and backward evolution to ρ_fix, use [H, ρ_β] = 0 and L_KMS(ρ_β) = 0, and match the α² terms. The result is a recursion relation for the off-diagonal matrix elements of E in the energy eigenbasis, controlled by the Fourier transform μ̂ of the random-evolution-time distribution. Solving for E gives E = ∫ e^−iHt Y e^iHt dν(t) where ν is a spectral measure satisfying ν̂(ω) = ω μ̂(ω) / (1 − μ̂(2ω)), and the Lamb-shift data Y encodes the bias.

A subtle point hidden here: if the evolution time T were fixed, then ν would have singularities at ω = kπ/(2T) and unbounded total variation, defeating the cancellation. Randomizing T via the gamma-like density μ₀(t) ∝ (t−1)³e^−(t−1) Ι_{t≥1} smooths these singularities. This is one of the most interesting technical insights in the paper — the random evolution time is not a convenience but a necessity.

Section IV applies the framework to Ding et al.'s setup with explicit parameter choices. The role of the Gaussian envelope width σ is reframed: prior work needed σ → ∞, the present work fixes σ = Θ(β²/λ_gap) and trades that for α → 0, which is far cheaper per iteration. The Hahn-et al. and Lloyd-et al. setups have a technical obstruction (the spectral gap of L_KMS can collapse with σ), which the authors address in Setup IV.3 (modified Lloyd) by sampling the bath field h from a Gaussian, restoring uniform spectral-gap bounds and preserving the ε⁻¹ end-to-end scaling.

Section V (proof overview) explains the rigorous argument that bypasses the formal asymptotic expansion. Rather than justifying the expansion term-by-term, the authors construct an explicit candidate ρ^* = ρ_β + α² E using the integral form, verify that ‖ρ^* − ρ_β‖₁ = O_β(α²), show Φ_α(ρ^*) ≈ ρ^* to leading order, and then invoke a stability lemma to conclude that the true fixed point is also close to ρ_β. The mixing-time argument is a perturbation result for primitive channels, in the spirit of Wang et al.'s 2025 analysis.

The paper closes with a comparison table (Tab. I) showing that all three existing system-bath algorithms (Hahn, Lloyd, Ding) — after the modifications proposed here — gain end-to-end complexity O(1/ε), dropping the envelope-width-to-infinity requirement to Θ(1). The open questions left in Sec. VI are noteworthy: improving the n-dependence (currently n⁶) via sharper spectral-gap perturbation, and exploring whether unitary-evolution-driven mixing-time speedups (PhysRevLett.134.140405; Li 2025) might compose with the present cancellation mechanism.

Figures

No figures extracted from source — this is a pure theory paper with only one (table) caption in the LaTeX source.

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): Direct method overlap. Y5's quantum SDP solver consumes quantum Gibbs states as its fundamental subroutine; the end-to-end cost of preparing those Gibbs states is exactly the bottleneck this paper addresses. A move from 1/ε⁴ to 1/ε in Gibbs-state preparation propagates linearly into Y5's SDP runtime. Crucially, Pauli-sparse Hamiltonians (Y5's class of interest) are local Hamiltonians with bounded operator norm, and the rapid-mixing examples this paper enumerates (high-temperature local spin Hamiltonians, weakly interacting spin/fermion systems, 1D local Hamiltonians) are structurally similar — Y5's Pauli-sparse Hamiltonians plausibly fit one of these regimes.
Y4 (Grover + ADMM for cardinality-constrained binary optimization): Indirect overlap. Y4 is search-based, not Gibbs-based, so the algorithmic family differs. However, ADMM-style classical-quantum hybrids that use quantum-prepared Gibbs sampling (e.g., for QUBO Hamiltonians via simulated annealing) would also benefit from cheaper Gibbs sampling. Tangential, not direct.
Y2, Y3 (QAOA for portfolio): No direct overlap. QAOA is a variational gate-based optimizer, not a Lindbladian-based thermal sampler.
Y1 (warm-started QAOA): No direct overlap.
Y6 (PBR test on Heron2): No direct overlap.

Recommended action for Yuan

Read Sec. III and Sec. IV deeply — the theorem statements and the explicit parameter choices in Theorem IV.1 (Ding et al. setup) directly tighten the cost analysis you can quote in any Y5 follow-up. Specifically, the O(1/ε) end-to-end bound replaces the looser bounds in the older Ding/Slezak references; if you cite Gibbs-preparation cost in a quantum SDP context, this is the up-to-date reference.
Check whether the Pauli-sparse Hamiltonians used in Y5 satisfy the rapid-mixing assumptions listed in the "Local spin Hamiltonians in the high-temperature regime" / "Weakly interacting spin systems" itemization (Sec. IV.A). If they do, you immediately inherit λ_gap = Ω(1/n) and Corollary IV.2's Õ(n⁶ε⁻¹) bound for the inner Gibbs-prep subroutine — worth a paragraph in any follow-up draft, or a back-of-envelope check.
Consider emailing the authors (Hongrui Chen at Stanford, Zhiyan Ding at Michigan, Ruizhe Zhang at Purdue) once you've worked through the rapid-mixing matching. A short note pointing out that Y5's Pauli-sparse Hamiltonians offer a concrete application class for their bound would be high-value for them and might prompt a reciprocal citation in their next iteration.