We prove optimal trace-norm relaxation for the one-dimensional Lindbladian associated with the Witten differential a=h∂x+V′ which annihilates the classical Gibbs density: if λ1(h) is the first positive eigenvalue of H=a∗a, then the relaxation rate is γh=λ1(h)/(2h). It applies to initial operators whose Schwartz kernels, after a Gibbs conjugation, satisfy L2 estimates for the restriction to the diagonal and for the normal derivative. An appendix by Chat GPT 6 presents a stronger result specialised to positive initial data.
Statement of the result and motivation
With motivation and references provided below, we study the evolution governed by the following superoperator:
LT=h1(aTa∗−21(a∗aT+Ta∗a)),a:=h∂x+V′(x).(1)
Since the jump operator, a, is given by a Witten differential [25] (and H:=a∗a is the scalar Witten Laplacian), we refer to L as the Witten Lindbladian. The assumptions on V are given in (1.8). By adding an irrelevant h-dependent constant we assume that
∫Re−2V(x)/hdx=1.(2)
For initial conditions for the Lindblad evolution, we consider T∈L1(L2(R)), the space of trace-class operators on L2(R),
Tu(x)=∫Re−(V(x)+V(y))/hr(x,y)u(y)dy,r∈C1,
where, with dμ(y):=e−2V(y)/hdy, r satisfies
r∣{(x,x)}∈L2(dμ),∣∂yr(x,y)∣≤G0(y)∈L2(dμ).(3)
Theorem 1. *There is a constant Ch,V>0 such that for T above with trT=1, we have for t≥0 and Π:=∣e−V/h⟩⟨e−V/h∣,
Remarks. 1. The regularity assumptions on r are chosen to keep the proof elementary. The pointwise differentiability in y is likely to be replaceable by an appropriate absolute-continuity condition. We do not pursue this endpoint question here, but some restrictions on T are always needed, as L does not have a spectral gap on L1 (easily seen in the case of V(x)=x2. The more substantial problem is the extension to higher dimensions: our methods are very much one dimensional.
Tracing the constants in the proof, one may take
Ch,V≤max(2eγh,1,2Ah,V(1+γh)1/2)V(V),
where (with S defined in (2.3))
Ah,V=⟨x⟩(1−S)−1/2L2(μ)→L2(μ),
and V(V) was defined in (3.2). For 0<h≤h0, the assumptions on V give Ah,V=OV(1) and γh=OV(1). Thus the h-dependence provided by the proof is contained in V(V). It is bounded as h→0 in the one-well case, while for a nondegenerate double well of barrier height S0 the proof gives
Ch,V=OV(heS0/h)=OV(γhh).
We do not claim that this is the optimal common prefactor in the estimate.
When, in the theorem, T≥0, we can replace the hypotheses by r∣{(x,x)}∈L∞ and the assumption that trHT=∥aT1/2∥22<∞. The result then follows from the following inequality proved at the of this introduction:
At our request, Chat GPT 6 upgraded this by weakening the hypothesis to r∣{(x,x)}∈Lq(dμ), q>1. That result, Theorem 2, is presented in the appendix.
When V(x)=x2/2, (1.1) is, after the standard normalization of the annihilation operator, the one-mode pure-loss channel, with transmissivity e−2t; see [15, 24, 6, 7]. The general potential replaces the harmonic-oscillator annihilation operator by the Witten differential h∂x+V′. Related first-order factorizations for anharmonic oscillators appear in [9, 21, 22].
Lindblad generators [13, 19] provide the standard Markovian framework for dissipative quantum dynamics. Recently, they have also appeared in quantum Gibbs sampling; see [4, 8] for finite-dimensional constructions and [11, 2] for related infinite-dimensional work. A model based on the simpler equation considered here appeared in Leng–Ding–Chen–Lin [17], where Witten differentials were used as Lindblad jump operators. The stationary state Π is obtained from the square root e−V/h of the Gibbs density. The trace norm is natural here since
∣tr((T−Π)O)∣≤∥T−Π∥L1∥O∥
for every bounded observable O.
The diagonal evolution is governed by the Witten Laplacian H. The theorem says that the full Lindblad evolution, including the off-diagonal part of the kernel, has the same optimal relaxation rate γh. For a self-contained one-dimensional account of the semiclassical asymptotics of λ1(h), and references, see [20]. If V has exactly one critical point, a unique nondegenerate minimum x0, harmonic approximation gives
γh=V′′(x0)+O(h).(7)
For a multiwell potential, γh is exponentially small and is governed by the Eyring–Kramers law. Thus the result includes the metastable regime. Related semiclassical Lindblad–Fokker–Planck constructions are discussed in [16, 10, 18, 23].
We now impose, in addition to the normalization (1.2),
These assumptions are chosen for a short proof rather than for maximal generality. Since −H/(2h) generates a contraction semigroup on L2(R), the Davies construction gives the trace-class contraction semigroup used below; see [10], §1, Remark. The remaining bounds are used in the maximum-principle, weighted, and tail estimates. We write
H:=a∗a=−h2∂x2+∣V′∣2−hV′′,ν:=e−V/h,Π:=∣ν⟩⟨ν∣.(9)
Then, aν=0, kerH=Cν, and, from (1.2), ∥ν∥L2=1. The assumptions (1.8) imply that H has compact resolvent and hence that λ1(h)>0.
Proof of (1.6). By the spectral theorem, T=∑j≥0λj∣φj⟩⟨φj∣, λj≥0, ∑j≥0λj=1, where the φj are orthonormal in L2(R). Writing φj=νfj we introduce the ℓ2-valued function F(x):={λjfj(x)}j∈N∈L2(R;ℓ2(N),dμ). It follows that
Acknowledgements. This note grew out of discussions between the authors and ChatGPT 5.6 about improving an earlier probabilistic argument with a non-optimal decay rate γh/2[23]. In particular, the key Volterra factorisation argument was suggested by ChatGPT and developed to a substantial extent through these discussions. MZ gratefully acknowledges partial support from the Simons Foundation through Targeted Grant Award No. 896630, “Moiré Materials Magic”.
The conjugated kernel and the maximum principle
In the language of the theorem in §1, let T0=T and Tt=etLT0. Using (1.9), write
Multiplication by ν gives the unitary equivalences of (unbounded) operators on L2(dμ) (on the left) with operators on L2(dx) (on the right):
Q≃−2ha∗a,S≃−2haa∗.(18)
The operator aa∗ has no zero mode, and its spectrum agrees with the nonzero spectrum of H=a∗a. Hence
∥etQf−μ(f)∥L2(μ)≤e−γht∥f−μ(f)∥L2(μ),(19)
∥etSg∥L2(μ)≤e−γht∥g∥L2(μ).(20)
Since ∫r0(x,x)dμ(x)=trT0=1, this gives
∥rt(x,x)−1∥L2(μ)≤e−γht∥r0(x,x)−1∥L2(μ).(21)
For the whole-space comparison we will use the second condition in (1.3) and the maximum principle for degenerate heat equations. Thus we start with
Lemma 2. Let w∈C2([0,T]×R2;R) have at most polynomial growth, and satisfy
(∂t−M+V′′(y))w≥0,w(0,x,y)≥0.(22)
Thenw(t,x,y)≥0 for (t,x,y)∈[0,T]×R2.
Proof. For C0 in (1.8) put
z=e−C0tw,q(y)=−V′′(y)−C0≤0,P=∂t−M−q(y).
Then (2.10) gives Pz≥0. Let ρ(x,y)=1+x2+y2. For every integer k≥1, (1.8) and direct differentiation give Mρk≤Ckρk (see (2.2) for the definition of M).
Choose m so that ∣z(t,x,y)∣≤CTρ(x,y)m on [0,T]×R2 and put
φ(t,x,y)=eAtρ(x,y)m+1.
Since q≤0, choosing A>Cm+1 gives Pφ>0.
For ε>0, set v=z+εφ. It is positive at t=0, and it tends to +∞ as x2+y2→∞, uniformly for 0≤t≤T. If it first vanished at (t0,x0,y0), then
∂tv≤0,∂xv=∂yv=0,∇x,y2v≥0.
In particular, (∂x+∂y)2v≥0. At the contact point the drift and zeroth-order terms vanish, and therefore
Pv=∂tv−21h(∂x+∂y)2v≤0.
But Pv=Pz+εPφ>0, a contradiction. Thus z+εφ>0 and letting ε→0+ proves the claim. □
The next proposition propagates the estimate (1.3) on the initial condition:
Proposition 3.For allt≥0, and forSdefined in (2.3),
∣∂yrt(x,y)∣≤(etSG0)(y),x,y∈R.(23)
Proof. Put ut=∂yrt. By (2.4),
∂tut=(M−V′′(y))ut.
Let gt=etSG0 and regard gt(y) as a function of (x,y). Since M reduces to Q on functions of y alone, it satisfies the same equation. For bounded smooth data, Lemma 2, applied for every θ∈R to
gt(y)−Re(e−iθut(x,y)),
gives (2.11). The same lemma shows that the closed semigroup generated by M−V′′(y) is positivity preserving. Since u0 and G0(y) belong to L2(R2;dμ(x)dμ(y)), the order inequalities
−G0(y)≤Re(e−iθu0(x,y))≤G0(y)
pass to the L2 closure. This proves the result for the stated G0∈L2(μ). □
For quadratic confinement the Volterra estimate below requires one spatial weight. It is obtained using smoothing for positive times:
Lemma 4. There is Ch,V>0 such that, for g∈L2(μ) and t>0,
has conjugated kernel r0(x,y)=1+f1(y), so that r0(x,x)−1=f1(x)∈L2(μ) and
∣∂yr0(x,y)∣=∣f1′(y)∣.
Hence it satisfies the hypotheses of the theorem, and
L(∣ν⟩⟨φ1∣)=−γh∣ν⟩⟨φ1∣.
Thus no larger uniform exponent is possible. □
Appendix by ChatGPT 6
The following strengthens Remark 3 in §1, under the same assumptions (1.2) and (1.8). Function norms ∥⋅∥p are in Lp(μ), and inner products are linear in the first factor.
Theorem 2. Let T≥0, trT=1, and let d(x)=r(x,x) be its conjugated diagonal, as in (1.10). If
For d∈L∞(μ), replace ∥d∥q by ∥d∥∞; the estimate then holds with an absolute constant.
Proof. Let F(x)={λjfj(x)}j be the ℓ2-valued function introduced in the proof of (1.6). Equations (1.10)–(1.11) give
r(x,y)=⟨F(x),F(y)⟩,d(x)=∥F(x)∥2,h2∥F′∥22=E(T).
We first suppose that F∈C∞(R;CN) and ∥F∥2=1; the approximation is given below.
For any positive trace-one operator A, the operator A−Π has at most one negative eigenvalue, since its quadratic form is nonnegative on ν⊥. Its trace is zero, so
∥A−Π∥L1=2∥A−Π∥≤2∥A−Π∥L2;(36)
see also [5]. It therefore suffices to estimate the Hilbert–Schmidt norm. To this end, put
Indeed, for complex Hilbert-space vectors u,v, with α=∥u∥, β=∥v∥,
⟨u,v⟩−21(α2+β2)≤21(α+β)∥u−v∥≤∥αu−βv∥.
The first inequality follows by Cauchy–Schwarz, the second by expanding squared norms. Apply them to u=F(x), v=F(y) and use ∥(∥F∥F′)′∥≤2∥F∥∥F′∥=b′; also ∣d′∣≤b′.
For p=2q/(q+1)∈(1,2), Hölder also gives
∥b′∥p≤2∥F∥2q∥F′∥2=h2∥d∥qE(T).(39)
To propagate (A.4), use M, Q, S from (2.2)–(2.3) and put Pt=etQ, dt=Ptd, bt=Ptb, kt=rt−21(dt(x)+dt(y)); dt(x)=rt(x,x) by (2.5). On x<y, both kt and bt(y)−bt(x) solve ∂tu=Mu and vanish on x=y. The barrier proof of Lemma 2, with the zeroth-order term omitted, applied to bt(y)−bt(x)−Re(eiθkt) for every θ, propagates the first bound. Apply the same argument to dt(y)−dt(x) for the second bound, and reflect for x>y. Smooth majorants converging uniformly to b′ justify comparison when b′ is not smooth.
Set σt=∥bt−μ(bt)∥2. Using 2∥f−μ(f)∥22=∬∣f(x)−f(y)∣2dμ(x)dμ(y) gives
∥dt−1∥2≤σt,∥kt∥L2(μ⊗μ)≤2σt.
The L2(μ⊗μ) norm of 21(dt(x)+dt(y))−1 is ∥dt−1∥2/2. Since ∂xQ=S∂x, we have bt′=etSb′. Thus (A.2) and the Poincaré inequality associated with (2.6)–(2.7) give
∥etLT−Π∥L1≤32σt≤λ1(h)32h∥etSb′∥2.(40)
The confinement assumptions (1.8) give c∣x∣≤∣V′(x)∣≤C∣x∣, xV′(x)>0, near infinity. Writing ρ=e−2V/h and defining m to the median of μ (∫−∞mρ(s)ds=21), these bounds give, for large x>0,
Hence F+(x)log(1/F+(x))∫mxρ(s)−1ds is bounded with the similar argument for F−. Thus μ satisfies a logarithmic Sobolev inequality [1] [BR03, Theorem 1], and [Gr75, Theorem 6] gives ∥Ptp∥Lp→L2≤1 for some tp<∞; take t2=0. Since V′′≥−C0, comparison gives 0≤etSG≤eC0tPtG for G≥0. The spectral bound for S in (2.8) therefore yields
∥etSb′∥2≤e−γh(t−tp)eC0tp∥b′∥p,t≥tp.(41)
Equations (A.5)–(A.7) prove (A.1) for t≥tp. For t≤tp, contractivity, the pure-state trace-distance formula and Cauchy–Schwarz give
∥etLT−Π∥L1≤21−tr(ΠT)≤2λ1(h)E(T),
using H≥λ1(h)(I−Π). Since ∥d∥q≥1, this covers short times. For q=∞, take p=2, tp=0; the constant is 62.
Finally, (1.11) and d∈Lq(μ) show that F∈H1(μ;ℓ2)∩L2q(μ;ℓ2). For finite q, component and spatial cutoffs, then mollification and normalization, give smooth finite-dimensional Fn converging in both spaces, with ∥Fn∥2=1. Thus ∥Tn−T∥L1≤2∥Fn−F∥2→0, and the energies and diagonal norms converge. Contractivity passes (A.1) to T. For q=∞, the same approximation gives ∥Fn∥∞≤(1+o(1))∥F∥∞. □
References
[1][BR03] F. Barthe and C. Roberto, Sobolev inequalities for probability measures on the real line, Studia Math. 159 (2003), no. 3, 481–497.DOI
[2][BRS26] S. Becker, C. Rouzé and R. Salzmann, Quantum Gibbs sampling in infinite dimensions: generation, mixing times and circuit implementation, arXiv:2604.01192.arxiv.org/abs/2604.01192
[4]C-F. Chen, M. J. Kastoryano, and A. Gilyén, An efficient and exact noncommutative quantum Gibbs sampler, arXiv:2311.09207.arxiv.org/abs/2311.09207
[5]P. J. Coles, M. Cerezo, and L. Cincio, Strong bound between trace distance and Hilbert–Schmidt distance for low-rank states, Phys. Rev. A 100 (2019), 022103.DOI
[6]G. De Palma, D. Trevisan, and V. Giovannetti, Gaussian states minimize the output entropy of the one-mode quantum attenuator, IEEE Trans. Inform. Theory 63 (2017), no. 1, 728–737.DOI
[7]G. De Palma, D. Trevisan, and V. Giovannetti, The one-mode quantum-limited Gaussian attenuator and amplifier have Gaussian maximizers, Ann. Henri Poincaré 19 (2018), no. 10, 2919–2953.DOI
[8]Z. Ding, B. Li, and L. Lin, Efficient quantum Gibbs samplers with Kubo–Martin–Schwinger detailed balance condition, Comm. Math. Phys. 406 (2025), Paper No. 67.
[9]S. Gao, Dissipative quantum dynamics at surfaces: A nonlinear-coupling model, Phys. Rev. B 60 (1999), 15609.DOI
[10]J. Galkowski and M. Zworski, with an appendix by Z. Huang and M. Zworski, Classical–quantum correspondence in Lindblad evolution, J. Math. Phys. 66 (2025), Paper No. 091503.
[11]J. Galkowski and M. Zworski, Localised Davies generators for (pseudo)differential operators, arXiv:2604.00306.arxiv.org/abs/2604.00306
[12]I. C. Gohberg and M. G. Krein, Theory and Applications of Volterra Operators in Hilbert Space, Translations of Mathematical Monographs 24, American Mathematical Society, 1970.
[13]V. Gorini, A. Kossakowski, and E. C. G. Sudarshan, Completely positive dynamical semigroups of N-level systems, J. Math. Phys. 17 (1976), 821–825.DOI
[15]A. S. Holevo and R. F. Werner, Evaluating capacities of bosonic Gaussian channels, Phys. Rev. A 63 (2001), 032312.
[16]F. Hernández, D. Ranard, and C. J. Riedel, Classical correspondence beyond the Ehrenfest time for open quantum systems with general Lindbladians, Comm. Math. Phys. 406 (2025), Paper No. 4.
[17]J. Leng, Z. Ding, Z. Chen, and L. Lin, Operator-level quantum acceleration of non-logconcave sampling, Proc. Natl. Acad. Sci. USA 123 (2026), e2512789123.arxiv.org/abs/2505.05301
[24]C. Weedbrook, S. Pirandola, R. García-Patrón, N. J. Cerf, T. C. Ralph, J. H. Shapiro, and S. Lloyd, Gaussian quantum information, Rev. Mod. Phys. 84 (2012), 621–669.
[25]E. Witten, Supersymmetry and Morse theory, J. Differential Geom. 17 (1982), no. 4, 661–692.DOI