WorksPhysics and Astronomy
Virtual Autopurification in Translationally Invariant Quantum States
Abstract
We introduce virtual autopurification (VAP), a protocol for virtual purification that operates entirely within a single translationally invariant system. Virtual purification suppresses thermal fluctuations and decoherence noise by amplifying the dominant pure component of a mixed state, but its standard implementation requires multiple copies of the state and global entangled measurements. VAP avoids the multiple-copy overhead by constructing virtual copies from spatially separated subregions and performing entangled measurements only between them. Assuming clustering of correlations, we show that VAP reproduces the expectation values of the standard virtual purification protocol, with discrepancies that decay exponentially in both the buffer length and the inter-subregion distance. Numerical simulations support the theoretical predictions and demonstrate accurate implementation of virtual cooling with reduced resource requirements. Our results establish VAP as a resource-efficient and scalable approach to quantum-state purification in near-term devices.
Introduction
Introduction.— Among digital quantum simulation tasks, thermal state preparation is a central application of quantum computing with potential impact across condensed-matter physics, quantum chemistry, and high-energy physics. The goal is to prepare a Gibbs state for a given Hamiltonian and temperature on quantum devices, which is essential for faithfully reproducing realistic experimental conditions.
Despite substantial theoretical progress [1, 2, 3, 4] and experimental demonstrations [5, 6, 7, 8], reaching lower temperatures is still difficult, because thermal fluctuations and decoherence must be suppressed strongly.
A promising route to alleviating these difficulties is virtual purification, a post-processing technique that extracts low-temperature observables from noisy or higher-temperature states [9, 10, 11, 12, 13, 14]. By amplifying the dominant pure component of a mixed state using powers of as via swap techniques, virtual purification suppresses thermal noise and decoherence. For a Gibbs state , this amounts to a virtual cooling , increasing the inverse temperature by a factor of . However, the
standard formulation, or fully virtual purification (FVP), requires globally entangled measurements over multiple physical copies, and moreover incurs exponential sampling overhead with system size [15].
A first step toward reducing such an overhead was provided by localized virtual purification (LVP) [16], which exploits the exponential clustering property in geometrically local systems to replace the globally entangled measurement with the one supported on the vicinity of the target observable of interest. While such a technique lifts the need for exponential sampling overhead, it still leaves the common crucial property in condensed matter problems: translational symmetry. Under clustered correlations, each local window of support can be expected to exhibit nearly identical reduced density matrices, while it has remained theoretically unclear whether it can be used as an operational resource for virtual purification.
In this work, we formulate this idea as virtual autopurification (VAP), a copy-free virtual purification protocol for translationally invariant quantum states, and prove its validity by providing a performance guarantee. As illustrated in Figure 1, spatially separated subregions in a translationally invariant system act as approximately independent copies. Assuming exponential clustering, we show that VAP reproduces expectation values of the standard purification protocol up to a bias that decays exponentially in both the buffer length and the inter-subregion separation, while circumventing the need for exponential overhead with respect to the system size (see Table 1). Numerical simulations corroborate these bounds and demonstrate accurate virtual cooling with substantially reduced resources, establishing VAP as a scalable route to virtual purification.
Figure 1. Schematic illustration of the VAP protocol. Identical subregions are chosen in a translationally invariant Gibbs state , with supporting the observable and providing a correlation buffer. Entangled measurements between distant subregions realize virtual cooling to temperature .
| FVP | LVP | VAP (ours) | |
| # copies per shot | 1 | ||
| Sampling overhead | |||
| Clustering assumption | none | Observable clustering | Trace-norm clustering |
Table I. Comparison of fully virtual purification (FVP), localized virtual purification (LVP), and virtual autopurification (VAP). Here, is an index for a virtually purified state and is the total number of lattice sites. Observable clustering and trace-norm clustering are given by Eqs. (1) and (6), respectively.
Virtual purification.— We begin by briefly reviewing the existing purification protocols for simulating the Gibbs state of a geometrically-local Hamiltonian . Given a quantum state and an integer , FVP estimates an observable supported on a local region via as introduced in Refs. [9, 10, 11], which is designed to amplify the dominant component of . To implement FVP on a quantum device, one uses the cyclic shift operator acting on copies, which gives . FVP therefore requires globally entangled operation for measuring and , which can be performed, for example, via Hadamard tests [17].
LVP reduces these measurement resources by restricting the cyclic shift to a local subsystem , where is a buffer surrounding and is the remainder of the system [16]. The resulting estimator is , where . Its accuracy follows from observable clustering: for normalized operators and supported on disjoint regions,
where is the distance between the regions. This property holds, in particular, for Gibbs states of local Hamiltonians at sufficiently high temperatures [18, 19, 20, 21, 22]. Under this condition, the FVP–LVP error satisfies
where is the distance between and , as shown in Figure 1. Thus, LVP reproduces FVP up to exponentially small corrections while avoiding global measurements and substantially reducing the sampling overhead.
Virtual autopurification.— The essential idea of VAP is to eliminate the multiple-copy requirement of LVP by constructing virtual copies within a single translationally invariant quantum state. When correlations decay exponentially, two spatially separated regions behave approximately independently; thus the LVP estimator can be reproduced by performing an entangled measurement between distant, isomorphic subregions of a single copy, as illustrated in Figure 1.
For clarity, we present the analytic calculation for the case . Let be the subsystem used in LVP and be a translated copy of separated by an inter-copy distance . The remainder of the system is denoted by . The VAP estimator is defined as
where swaps the two subsystems and . Assuming that the state is translationally invariant as , the LVP estimator can be rewritten as
The difference between the two estimators is given by
where .
If the quantum state on and were completely separated, so that , the LVP and VAP estimators would coincide. In general, however, this factorization does not hold exactly. We therefore assume exponential clustering at the level of reduced density operators, which we refer to as trace-norm clustering:
This is a stronger assumption than the clustering of observables in Eq. (1). This assumption exactly holds for one-dimensional systems at arbitrary temperatures [23]. In higher-dimensional systems, it is also known to hold under the assumption that the effective Hamiltonians associated with local reduced density operators of the Gibbs state remain local [24].
Building on this clustering property, we analyze the VAP–LVP discrepancy through second order in , a parameter that scales the boundary couplings between the virtual-copy regions and the remainder of the system. Concretely, together with the high-temperature condition specified in Lemma 1 of the Supplemental Material (SM) [25], we obtain the following bound.
Theorem 1. Assuming trace-norm clustering and sufficiently high temperature indicating quasi-locality of the imaginary-time evolution, the difference between the FVP and VAP estimators is suppressed with both the buffer length and the inter-copy separation as
This theorem follows by combining the previously established bound between FVP and LVP in Eq. (2) with the proposition below.
Proposition 1. Under the assumptions stated in Theorem 1, the difference between the LVP and VAP estimators decays exponentially with the inter-copy distance:
Proof sketch. We compare the two density operators and in Eq. (5) perturbatively around a decoupled Hamiltonian. Let , where contains the interactions within , , and , while contains the boundary couplings to . By introducing a coupling parameter for , we consider the perturbed thermal states . The VAP and LVP estimators coincide up to first order in , while the second-order deviation of the density operators is given by
Here, denotes the thermal state at , and encodes the two-copy correlations,
with and denoting the corresponding effective operator associated with . We further define the reduced operator , with obtained analogously by tracing out . The corresponding scalar function is defined as . Substituting Eq. (9) into the expression for the difference between the VAP and LVP estimators in Eq. (5), we obtain four classes of terms. The scalar term vanishes because it multiplies the zeroth-order expectation value of the centered observable. For the terms involving and , Lemma 1 of SM [25] approximates the imaginary-time-evolved boundary interactions by local operators with exponentially small errors. The local-observable clustering condition in Eq. (1) then gives an bound. The term retains correlations between reduced density operators even after the imaginary-time-evolved boundary interactions are replaced by their local approximations, and therefore requires the stronger condition in Eq. (6), which gives the same exponential bound. Combining these estimates gives the exponential suppression stated in Proposition 1. The detailed derivation for is presented in SM [25]. Since connected contributions at second order in involve at most two virtual copies, the same analysis applies to any fixed .
Theorem 1 bounds the systematic bias of VAP relative to FVP. We quantify the required resources by the space–time overhead, defined as the measurement cost multiplied by the number of copies required per shot. For estimating the expectation value of a Pauli operator , the FVP variance is bounded as
This motivates defining the measurement cost of FVP as , following the Supplemental Material of Ref. [16]. By the same reasoning, the measurement costs of LVP and VAP are given by and , respectively. Combining these measurement costs with the number of physical copies required per shot, we define the space–time overhead as , and where denotes the cyclic shift among the spatially separated subregions used as virtual copies within a single system. Their leading scaling is
Here, and are the numbers of sites in the full system and region , respectively, and is the thermal entropy density. Thus, LVP and VAP replace the dependence on the total system size with that on the local subsystem size , while VAP further removes the prefactor by requiring only one physical copy per shot.
Numerical simulations
We numerically test the theoretical predictions above and benchmark the performance of VAP. As a benchmark, we consider the one-dimensional transverse-field Ising model with open boundary conditions,
where () are Pauli operators at site . In the following calculations, we focus on the critical point . To access large system sizes and directly manipulate thermal mixed states, we employ matrix-product density operators (MPDOs) [26, 27, 28] implemented using the ITensor library [29, 30]. Following the subsystem arrangement shown in Figure 2 (inset), we study how the distances and control the differences among the three estimators for . For each method, we define the effective reduced density operator on region as , and where is a complement of the subsystem . We refer to these as effective reduced density operators because is generally non-Hermitian, while still reproducing the correct VAP expectation values.
Figure 2. (a), (b) Trace distances between the effective reduced density operators of FVP, LVP, and VAP as functions of and , respectively. (c), (d) Corresponding space–time overheads . The subsystem configuration is shown in the inset. We set , , and , with five additional boundary sites at each end of the chain, giving .
Figure 2 shows the trace distances between the effective reduced density operators and the corresponding space–time overheads for the three protocols, as functions of and . The exponential decay of the FVP–LVP discrepancy with is consistent with the theoretical result established in Ref. [16]. Likewise, the exponential decay of the LVP–VAP discrepancy with supports the validity of Proposition 1. At the same time, the numerical results also suggest a stronger scaling than that stated in Proposition 1, since the LVP–VAP discrepancy appears to decay exponentially in both and :
Consequently, when is sufficiently large for the LVP–VAP discrepancy to become negligible, the VAP approximation error relative to FVP is governed primarily by . The space–time overhead further highlights the practical advantage of LVP and VAP. For fixed , the suppression of the LVP overhead relative to FVP scales as , where . Thus, the LVP overhead is exponentially smaller than the FVP overhead in the size of . For fixed , remains constant as increases, explaining the nearly constant LVP and VAP overheads. VAP further reduces the overhead relative to LVP by approximately a factor of , since it requires only one physical copy per shot.
Figure 3 plots the space–time overhead for the number of virtual copies varied from to . As shown in
Figure 3. (a) Space–time overhead of the VAP protocol as a function of for several numbers of virtual copies . (b) Ratio of the LVP overhead to the VAP overhead as a function of for , , , and . The calculations use , , and . For each , the virtual copies are arranged to maximize their inter-copy separations; for , this gives .
Fig. 3(a), the overhead of VAP decreases toward lower temperatures. This behavior is consistent with the entropy scaling in Eq. (12): lowering the temperature reduces and hence the space–time overhead of VAP. Figure 3(b) shows , i.e., the advantage of VAP over LVP. In the high-temperature regime, the expected factor-of- advantage arising from the smaller number of copies required by VAP is clearly observed. For the fixed subsystem geometries considered here, the LVP-to-VAP overhead ratio exceeds the copy-number factor at lower temperatures, indicating an additional reduction in sampling overhead. This comparison isolates the variance-related cost and does not account for possible increases in and required to maintain a fixed approximation bias as the correlation length grows.
Finally, we numerically examine the robustness of VAP to violations of translational invariance in the Gibbs state due to miscalibration or algorithmic errors. To this end, we consider the one-dimensional Heisenberg model with a random magnetic field,
where () denotes the spin operator at site . Each is independently drawn from a uniform distribution on , where sets the strength of the random field. The random fields represent spatially varying calibration errors that cause the prepared state to be the Gibbs state of a Hamiltonian different from the target one. This setting allows us to test whether VAP suppresses the resulting deviation from the target Gibbs state. Figure 4 shows the numerical results for and . In both cases, VAP substantially reduces the disorder-induced deviation of the energy, yielding values closer to the clean reference than to the disordered reference. This behavior suggests that VAP has an effective disorder-averaging effect and can extract the approximately translationally invariant component of the thermal state. A possible interpretation is that combining the density operators associated with multiple virtual copies acts more like averaging the underlying local Hamiltonians than directly averaging the density operators themselves. Unlike FVP and LVP, which purify copies of the same spatially inhomogeneous state, VAP combines distinct spatial regions and can partially average out their local deviations. These results therefore suggest that VAP may mitigate spatially inhomogeneous errors that are not directly addressed by conventional virtual-purification protocols.
Figure 4. VAP energy for the random-field Heisenberg model for (a) and (b) . The data are averaged over 10 disorder realizations. The gray dashed line indicates the energy of the original disordered thermal state, while the black dotted line indicates the corresponding clean reference without disorder ().
Summary and discussion.— We have presented virtual autopurification (VAP), a protocol that performs virtual purification entirely within a single translationally invariant quantum state. By using spatially separated subregions as approximate virtual replicas, VAP removes the need for multiple physical copies while retaining the purification mechanism of fully virtual purification (FVP) [9, 10, 11]. Under clustering assumptions, we showed that the VAP estimator reproduces the results of FVP and localized virtual purification (LVP) [16], with errors that decay exponentially with the buffer length and the inter-copy separation. Numerical simulations corroborated these predictions and demonstrated accurate virtual cooling with reduced space–time overhead. Numerical results in the End Matter further show that VAP can simultaneously realize virtual cooling and mitigate local depolarizing noise, suggesting robustness against this form of decoherence noise.
To implement virtual purification protocols on real quantum devices [12], it is crucial to estimate nonlinear quantities such as and efficiently. Beyond the naive generalized swap trick, several approaches have recently been developed for this purpose, including methods based on query access to a unitary that prepares a purification of [31], randomized measurements [32, 33], and classical shadows [34, 35]. However, such estimator-level approaches do not by themselves remove the exponential sampling overhead intrinsic to virtual purification. In contrast, LVP and VAP exploit physical properties of , such as clustering and translational invariance, to reduce this overhead at the level of the purification protocol itself. This protocol-level reduction is complementary to nonlinear-estimation techniques and to other error-mitigation or error-correction layers, including quantum subspace expansion [36, 37] and approaches that combine virtual purification with quantum error correction [38]. Related single-state approaches to virtual purification have also been proposed, such as applying the inverse of the state-preparation circuit [39] or approximating the SWAP operator in the eigenbasis of the target observable [40]. Compared with these approaches, VAP provides a rigorous error analysis controlled by physical properties of the state.
The idea that a single thermal state can contain substantial information about the underlying Hamiltonian has appeared in distinct contexts. In quantum information, Hamiltonian learning from Gibbs states has been actively studied [41, 42, 43]. In statistical mechanics, related questions have been discussed for reconstructing a Hamiltonian from a single thermal eigenstate [44]. These perspectives suggest that a single thermal state can encode more information than is directly accessible through conventional expectation values. From this viewpoint, VAP can be regarded as a concrete protocol for extracting lower-temperature information encoded in a single Gibbs state. Exploring what other physical information can be accessed from thermal states is an interesting direction for future work.
Acknowledgments.— A. I. was supported by the Center of Innovation for Sustainable Quantum AI (JST Grant No. JPMJPF2221) and JSPS KAKENHI (Grant No. 25K17308). H. H. was supported by the MEXT Quantum Leap Flagship Program (MEXT Q-LEAP) (Grant No. JPMXS0112019794), JST PRESTO (Grant No. JPMJPR25F7), and JSPS KAKENHI (Grant No. JP24K16979). N. Y. was supported by JST (Grant No. JPMJPF2221), JST CREST (Grant No. JPMJCR23I4), JST ASPIRE (Grant No. JPMJAP2316), JST ERATO (Grant No. JPMJER2302), JST Moonshot R&D (Grant No. JPMJMS256J), IBM Quantum, Google Quantum AI, and the Institute for AI and Beyond at the University of Tokyo.
Data availability.— The code used to generate the numerical results reported in this work is publicly available on GitHub [45].
End matter
Analysis under stochastic noise
We consider the effect of local depolarizing noise. The depolarizing noise acting on the th qubit is defined by
where denotes the error rate. The local depolarizing noise channel is given by applying this noise independently to all sites:
In the numerical calculations below, we consider a thermal state subjected to local depolarizing noise, and study a setting in which VAP simultaneously performs virtual cooling and error mitigation.
We present numerical results for and , with the number of virtual copies varied from to . Figure 5(a) shows the VAP energy plotted against for a fixed error rate . The dashed line denotes the exact result, and the VAP data collapse well onto it. Figure 5(b) shows the deviation of the VAP energy from its noiseless value at . For all , the noise-induced deviation grows polynomially with . Figure 5(c) shows the deviation of the VAP energy from the exact result. For all error rates considered, this deviation is suppressed exponentially with increasing . Although we have not developed a theory of error mitigation in VAP, these results numerically demonstrate that, under a realistic error model, VAP can simultaneously realize virtual cooling and error mitigation.
Figure 5. Thermal-state results in the presence of local depolarizing noise. (a) VAP energy rescaled by , plotted as a function of for . The dashed line indicates the exact result. (b) Absolute deviation of the energy from its noiseless value at , shown as a function of for . (c) Absolute deviation of the energy from the exact result, shown as a function of for .
To qualitatively understand Figure 5(b), consider a simple noninteracting toy model,
For this model, the thermal state factorizes into single-site mixed states with polarization , and local depolarizing noise reduces the polarization as with . Virtual purification with copies then gives an effective polarization
so that, for small , the energy deviation behaves as
where
At finite , the thermal state contains excited-state components whose weights are modified by stochastic noise. When these weights are raised to the th power, a first-order noise correction in one factor combines with the noiseless thermal weight in the remaining factors. As a result, a term linear in remains after virtual purification. For a pure ground state, incoherent errors can introduce weight in orthogonal excited states. Virtual purification raises these noise-induced weights to the th power, so their contribution to the error typically scales as . In the toy model, the energy deviation in the pure ground-state limit is
For , we have . Thus, at fixed , the coefficient decays exponentially with up to the prefactor , consistent with Figure 5(c).
Setup and Notation
Setup
We consider a quantum many-body system on a lattice equipped with the graph distance , with a local Hamiltonian
where each term is supported on a subset of the lattice. We assume that every nonzero interaction term satisfies and . We assume a uniform bound on the local interaction strength,
Here, bounds the number of sites involved in each interaction, is the maximum spatial range of the interactions, and sets the typical local energy scale.
As illustrated schematically in Figure S1, we partition the system into disjoint regions , , , , . Region supports the local observable of interest . We assume that its support contains at most sites:
Figure S1. Schematic picture of the spatial partition used in the analysis. Regions and form buffers around and , respectively, while denotes the remainder of the system. The buffer length and the inter-copy separation are also indicated.
Region is a buffer surrounding , and we set
We further choose to be a translated copy of , spatially separated from it. The remainder of the system is denoted by .
Throughout the analysis, we use two characteristic distances: the buffer length between and , and the inter-copy separation between and . These two length scales, shown in Figure S1, govern the error of the LVP approximation and the additional error introduced in VAP, respectively. We assume to exclude direct interactions between and .
We denote the mixed state of the full system by . For any region , the reduced state on is written as
where is the complement of . When the global state is translationally invariant and is a translated copy of , we have
Swap operator
Let denote the swap operator exchanging the two isomorphic subsystems and :
with trivial action on the rest of the system. We will repeatedly use the identity
where and act on isomorphic subsystems.
LVP and VAP estimators (for n=2)
For the remainder of the Supplemental Material, we focus on the case . Given a local observable supported on , we define the localized virtual purification (LVP) expectation value as
The virtual autopurification (VAP) estimator is defined by
Using the identity in the previous subsection, one obtains the exact relation
Perturbative setup
To control correlations between and through , we decompose the Hamiltonian into intra-region and boundary contributions:
Here, and . For example, , and collects interactions straddling the boundary :
and similarly for . We introduce a coupling parameter as
and define the corresponding Gibbs state
The original Hamiltonian and Gibbs state are recovered as and . In § 2, we expand the relevant quantities around and retain terms through second order in . When no confusion can arise, we leave the dependence of reduced states implicit.
Proof of Proposition 1
In this section, we prove Proposition 1 by a perturbative expansion in the coupling up to second order. The main goal is to bound the difference between the VAP and LVP estimators while keeping track of their normalization factors. Throughout the proof, we use the perturbative setup introduced in Sec. S0.
Clustering assumption
We begin by stating the clustering assumption used in the proof of Proposition 1. Throughout this section, we assume that the reduced density operators of distant regions satisfy exponential clustering in trace norm:
for disjoint regions and , with some correlation length .
This assumption immediately implies the standard exponential clustering of connected correlations. Indeed, for operators and supported on and , respectively, we have
Exact identity for the VAP-LVP difference
We now rewrite the VAP-LVP difference in a form suitable for perturbative analysis. For this purpose, we first establish explicitly that the first-order contribution to the relevant density-operator difference vanishes. Let , and denote its reduced states on , , and by , , and , respectively. Since contains no couplings between these regions, the unperturbed Gibbs state factorizes as .
The first-order correction to the normalized Gibbs state is linear in the perturbation . Here and below, and denote the coefficients of in the expansions of and , respectively. If only is retained, remains decoupled from for all , so this perturbation contributes only to the first-order correction of . Similarly, contributes only . Therefore,
On the other hand, expanding the product of the reduced states gives
Thus, the zeroth- and first-order terms coincide, and hence
Consequently, the difference starts at second order in . Therefore, up to , the normalization factors can be replaced by their zeroth-order values, which yields
Here
From this point on, we redefine by shifting it as
With this convention,
Second-order expansion
Having established the cancellation of the first-order contribution, we now evaluate the leading second-order term explicitly. We work in the interaction picture with respect to and define
For any operator , we denote the thermal expectation value in the unperturbed state by
It is convenient to introduce the fluctuation operators
and
With these definitions, the normalized Gibbs state admits the Duhamel expansion
We next trace out the region in order to isolate the effective coupling between and . In the subscripts of the effective operators, and are omitted. The labels and indicate the boundaries from which these operators arise, although the operators may act on the full regions and , respectively. The effective boundary operators are
where
We then define
which captures the connected second-order contribution between the two boundary couplings through . From , we further define
Collecting all second-order terms, we obtain
Equation (S36) reduces the problem to bounding the four second-order contributions separately, which we do in the next subsection.
Bounds on the second-order terms
Reduction of the numerator
Substituting Eq. (58) into Eq. (43), we find that the estimator difference is determined by four second-order contributions associated with , , , and . To evaluate these terms, it is convenient to introduce the imaginary-time evolved operator
In particular, we will use .
It is also convenient to rewrite the relevant traces in terms of the auxiliary state
which naturally appears after applying the swap identity and collecting the thermal weights. We denote expectation values with respect to by .
Vanishing of the scalar term (s,t)
We first note that the contribution proportional to vanishes. Indeed, since is a scalar, it can be pulled out of the trace, and we obtain
By construction, the observable has been centered so that
Therefore, the scalar contribution does not contribute to the VAP–LVP difference at order .
Bounds on the single-boundary terms B(s,t) and B'(s,t)
We first consider the contribution containing . Applying the swap identity in Eq. (30) to the normalized trace in Eq. (43), and identifying the unperturbed Gibbs states of the translated copies and under the swap, we obtain
In the second equality, we used and the cyclicity of the trace. The last equality follows from the definition of as a trace over and the definition of . Expanding , we obtain
Here, the products of expectation values arise from the factorization of between , , and . In the first term, we used , since these operators act on disjoint regions. The contribution containing is treated in the same way.
To bound the connected correlators in Eq. (63), we use a quasi-locality property of the imaginary-time evolved boundary interactions. Roughly speaking, although and are not strictly local, they can be approximated by operators supported near the corresponding boundaries, with errors decaying exponentially in the truncation distance. This allows us to combine quasi-locality with exponential clustering to obtain an exponential bound in the inter-copy distance .
The following lemma is closely related to standard quasi-locality estimates for imaginary-time evolution. For completeness, we provide a self-contained proof below. The argument follows the usual nested-commutator expansion together with standard bounds on multicommutators; see, e.g., Appendix E.1 and Lemma 13 of Ref. [46].
Lemma 1. Let be a local Hamiltonian term supported on , and let be a region containing . Define
where denotes the restriction of the Hamiltonian to the region . Assume , and let
where is the complement of . Then
where
Proof. We compare the full imaginary-time evolution generated by with the truncated one generated by . Expanding both in nested commutators, we obtain
where and .
Each -fold commutator in Eq. (68) can be written as a sum over sequences of local Hamiltonian terms whose supports overlap successively, starting from the original support of . Since each Hamiltonian term has spatial diameter at most , the support of the operator can expand by at most distance at each commutator step. Therefore, if
no sequence of overlapping local terms can reach the complement . In that case, every term contributing to is already contained entirely inside , and hence coincides with the corresponding term in . It follows that
Thus the difference in Eq. (68) receives contributions only from orders .
For the remaining terms with , we bound the norm of the difference naively by estimating each nested commutator separately. Using
together with the assumption
and the fact that each interaction term acts on at most sites, one obtains the standard estimate
and similarly for . Hence
Substituting this into Eq. (68), we obtain
Since , the sum is geometric and converges:
This proves
The same estimate holds for the opposite convention , since replacing by changes only the signs of the nested-commutator coefficients and not their norm bounds.
Using this opposite-sign counterpart of Lemma 1, we truncate the imaginary-time evolved boundary operators and within . Let denote the regions retained when truncating at distances and from the and boundaries, respectively. The corresponding truncated operators are denoted by and . Since
Lemma 1 implies
Therefore,
where
Similarly,
Moreover, using the support condition in Eq. (25) and Eq. (28) of Ref. [47], for we have
Combining these bounds yields
Although need not be supported on , it remains supported within , because its imaginary-time evolution is generated by alone. At this stage, the weaker clustering assumption is sufficient. Using the exponential clustering of connected correlations introduced in Eq. (39), we obtain
Substituting this into the previous bound, we obtain
Choosing
and defining
we arrive at
Applying the same reasoning to the other connected term in Eq. (63), we conclude that the contributions containing and are exponentially suppressed in .
Bound on the double-boundary term
We finally consider the contribution containing . In contrast to the single-boundary terms, the estimate of this term requires the stronger trace-norm clustering assumption introduced in Eq. (38).
Including the normalization factor in Eq. (43), we define
and such that it satisfies
The swap identity in Eq. (30) places the factor of first in both cross terms. Moving a Gibbs factor through the swapped interaction and using gives, for the first ordering,
For the opposite ordering, the factors must also be reordered inside their thermal trace. Thermal cyclicity shifts their time argument, so that
The factorizable subtraction terms transform in the same way. Combining the two orderings, we obtain
Figure S2 schematically shows the supports of and . In particular, note that the two operators have overlapping support in region .
For the first term in Eq. (91), we bound
Applying the same quasi-local truncation as above, we obtain
For the last term, the stronger trace-norm clustering assumption in Eq. (6) is needed. Tracing over the truncation regions and shown in Figure S2, we have

Figure S2. Schematic illustration of the supports of and . The operator acts on the boundary region between and , shown in light red. The operator acts on the interior of and on the part of adjacent to , shown in light blue. The regions and indicate the parts of retained by the corresponding truncations.
Using Eq. (38), this is bounded by
Substituting this back into the previous estimate, we obtain
Choosing
and defining
we arrive at
Therefore, the first term in Eq. (91) is bounded by
The second term in Eq. (91) is treated in the same way. Hence the contribution containing is also exponentially suppressed in .
Collecting the bounds on the , , , and contributions, we conclude that each second-order term in Eq. (43) decays exponentially with . Therefore,
for some effective correlation length . This proves Proposition 1.
References
- [1]F. G. S. L. Brandão and M. J. Kastoryano, Finite correlation length implies efficient preparation of quantum thermal states, Communications in Mathematical Physics 365, 1–16 (2018).arxiv.org/abs/1609.07877
- [2]C.-F. Chen, M. Kastoryano, F. G. S. L. Brandão, and A. Gilyén, Efficient quantum thermal simulation, Nature 646, 561–566 (2025).DOI
- [3]C. Rouzé, D. Stilck França, and A. M. Alhambra, Optimal quantum algorithm for gibbs state preparation, Phys. Rev. Lett. 136, 060601 (2026).DOI
- [4]C. Rouzé, D. Stilck França, and A. M. Alhambra, Efficient thermalization and universal quantum computing with quantum gibbs samplers, Nature Physics (2026).
- [5]J. Wu and T. H. Hsieh, Variational thermal quantum simulation via thermofield double states, Phys. Rev. Lett. 123, 220502 (2019).DOI
- [6]M. Motta, C. Sun, A. T. K. Tan, M. J. O’Rourke, E. Ye, A. J. Minnich, F. G. S. L. Brandão, and G. K.-L. Chan, Determining eigenstates and thermal states on a quantum computer using quantum imaginary time evolution, Nature Physics 16, 205–210 (2019).DOI
- [7]D. Zhu, S. Johri, N. M. Linke, K. A. Landsman, C. Huerta Alderete, N. H. Nguyen, A. Y. Matsuura, T. H. Hsieh, and C. Monroe, Generation of thermofield double states and critical ground states with a quantum computer, Proceedings of the National Academy of Sciences 117, 25402–25406 (2020).
- [8]R. Sagastizabal, S. P. Premaratne, B. A. Klaver, M. A. Rol, V. Negîrneac, M. S. Moreira, X. Zou, S. Johri, N. Muthusubramanian, M. Beekman, C. Zachariadis, V. P. Ostroukh, N. Haider, A. Bruno, A. Y. Matsuura, and L. DiCarlo, Variational preparation of finite-temperature states on a quantum computer, npj Quantum Information 7, 130 (2021).DOI
- [9]J. Cotler, S. Choi, A. Lukin, H. Gharibyan, T. Grover, M. E. Tai, M. Rispoli, R. Schittko, P. M. Preiss, A. M. Kaufman, M. Greiner, H. Pichler, and P. Hayden, Quantum virtual cooling, Phys. Rev. X 9, 031013 (2019).
- [10]B. Koczor, Exponential error suppression for near-term quantum devices, Phys. Rev. X 11, 031057 (2021).
- [11]W. J. Huggins, S. McArdle, T. E. O’Brien, J. Lee, N. C. Rubin, S. Boixo, K. B. Whaley, R. Babbush, and J. R. McClean, Variational distillation for quantum error mitigation, Phys. Rev. X 11, 041036 (2021).DOI
- [12]T. E. O’Brien et al., Purification-based quantum error mitigation of pair-correlated electron simulations, Nature Physics 19, 1787–1792 (2023).DOI
- [13]Z. Cai, Resource-efficient purification-based quantum error mitigation, arXiv:2107.02779 (2021).
- [14]P. Czarnik, A. Arrasmith, L. Cincio, and P. J. Coles, Qubit-efficient exponential suppression of errors, arXiv:2102.06056 (2021).arxiv.org/abs/2102.06056
- [15]P. Vikstald, G. Ferrini, and S. Puri, Study of noise in virtual distillation circuits for quantum error mitigation, Quantum 8, 1441 (2024).
- [16]H. Hakoshima, S. Endo, K. Yamamoto, Y. Matsuzaki, and N. Yoshioka, Localized virtual purification, Phys. Rev. Lett. 133, 080601 (2024).DOI
- [17]A. K. Eckert, C. M. Alves, D. K. L. Oi, M. Horodecki, P. Horodecki, and L. C. Kwek, Direct estimations of linear and nonlinear functionals of a quantum state, Phys. Rev. Lett. 88, 217901 (2002).
- [18]H. Araki, Gibbs states of a one dimensional quantum lattice, Communications in Mathematical Physics 14, 120–157 (1969).DOI
- [19]Y. M. Park and H. J. Yoo, Uniqueness and clustering properties of gibbs states for classical and quantum unbounded spin systems, Journal of Statistical Physics 80, 223–271 (1995).DOI
- [20]D. Ueltschi, Cluster expansions and correlation functions, Moscow Mathematical Journal 4, 511–522 (2004).
- [21]M. Kliesch, C. Gogolin, M. J. Kastoryano, A. Riera, and J. Eisert, Locality of temperature, Phys. Rev. X 4, 031019 (2014).
- [22]J. Fröhlich and D. Ueltschi, Some properties of correlations of quantum lattice systems in thermal equilibrium, Journal of Mathematical Physics 56, 053302 (2015).DOI
- [23]A. Bluhm, A. Capel, and A. Pérez-Hernández, Exponential decay of mutual information for gibbs states of local hamiltonians, Quantum 6, 650 (2022).DOI
- [24]A. Bluhm, A. Capel, and A. Pérez-Hernández, Strong decay of correlations for gibbs states in any dimension, Journal of Statistical Physics 192, 134 (2025).
- [25]See Supplemental Material for details (URL to be added).
- [26]F. Verstraete, J. J. García-Ripoll, and J. I. Cirac, Matrix product density operators: Simulation of finite-temperature and dissipative systems, Phys. Rev. Lett. 93, 207204 (2004).pubmed.ncbi.nlm.nih.gov/15600964
- [27]M. Zwolak and G. Vidal, Mixed-state dynamics in one-dimensional quantum lattice systems: A time-dependent superoperator renormalization algorithm, Phys. Rev. Lett. 93, 207205 (2004).DOI
- [28]A. E. Feiguin and S. R. White, Finite-temperature density matrix renormalization using an enlarged hilbert space, Phys. Rev. B 72, 220401 (2005).DOI
- [29]M. Fishman, S. R. White, and E. M. Stoudenmire, The ITensor Software Library for Tensor Network Calculations, SciPost Phys. Codebases , 4 (2022).DOI
- [30]M. Fishman, S. R. White, and E. M. Stoudenmire, Codebase release 0.3 for ITensor, SciPost Phys. Codebases , 4 (2022).DOI
- [31]Y. Zhang, Y. Wu, Y. Zhou, and X. Yuan, Measuring less to learn more: Quadratic speedup in learning nonlinear properties of quantum states, Phys. Rev. Lett. 136, 130602 (2026).DOI
- [32]Z. Du, Y. Tang, A. Elben, I. Roth, J. Eisert, and Z. Liu, Optimal randomized measurements for a family of nonlinear quantum properties, PRX Quantum 7, 010360 (2026).arxiv.org/abs/2505.09206
- [33]Z. Li, D. Chen, D. Qin, Y. Yang, and Y. Zhou, Quantum nonlinear properties from a single measurement setting, arXiv:2605.09958 (2026).arxiv.org/abs/2605.09958
- [34]A. Seif, Z.-P. Cian, S. Zhou, S. Chen, and L. Jiang, Shadow distillation: Quantum error mitigation with classical shadows for near-term quantum processors, PRX Quantum 4, 010303 (2023).DOI
- [35]Q. Liu, Z. Li, X. Yuan, H. Zhu, and Y. Zhou, Auxiliary-free replica shadows: Efficient estimation of multiple nonlinear quantum properties, Phys. Rev. Lett. 136, 100602 (2026).DOI
- [36]N. Yoshioka, H. Hakoshima, Y. Matsuzaki, Y. Tokunaga, Y. Suzuki, and S. Endo, Generalized quantum subspace expansion, Phys. Rev. Lett. 129, 020502 (2022).DOI
- [37]B. Yang, N. Yoshioka, H. Harada, S. Hakkaku, Y. Tokunaga, H. Hakoshima, K. Yamamoto, and S. Endo, Resource-efficient generalized quantum subspace expansion, Phys. Rev. Appl. 23, 054021 (2025).DOI
- [38]H. Kwon, C. Oh, Y. Lim, H. Jeong, S-W. Lee, and L. Jiang, Virtual purification complements quantum error correction in quantum metrology, npj Quantum Information 12, 88 (2026).DOI
- [39]M. Huo and Y. Li, Dual-state purification for practical quantum error mitigation, Phys. Rev. A 105, 022427 (2022).DOI
- [40]A. Karim, H. J. Vallury, and M. Usman, Fictitious copy quantum error mitigation, arXiv:2603.09302 (2026).arxiv.org/abs/2603.09302
- [41]A. Anshu, S. Arunachalam, T. Kuwahara, and M. Soleimanifar, Sample-efficient learning of interacting quantum systems, Nature Physics 17, 931–935 (2021).DOI
- [42]J. Haah, R. Kothari, and E. Tang, Learning quantum Hamiltonians from high-temperature Gibbs states and real-time evolutions, Nature Physics 20, 1027–1031 (2024).DOI
- [43]A. Bakshi, A. Liu, A. Moitra, and E. Tang, Learning quantum Hamiltonians at any temperature in polynomial time, in Proceedings of the 56th Annual ACM Symposium on Theory of Computing, STOC ’24 (ACM, 2024), p. 1470–1477.
- [44]J. R. Garrison and T. Grover, Does a single eigenstate encode the full hamiltonian?, Phys. Rev. X 8, 021026 (2018).DOI
- [45]A. Iwaki, VAP Simulations, GitHub repository, https://github.com/AtshushiIwaki/vap-simulations (2026).
- [46]T. Kuwahara, A. M. Alhambra, and A. Anshu, Improved thermal area law and quasilinear time algorithm for quantum Gibbs states, Phys. Rev. X 11, 011047 (2021).DOI
- [47]A. M. Alhambra, Quantum many-body systems in thermal equilibrium, PRX Quantum 4, 040201 (2023).DOI





