betaFeedback

WorksPhysics and Astronomy

Markov-Shielded Typicality for Thermal Quantum States

Abstract

Pure states of quantum many-body systems at thermal equilibrium are known to be typical, in a sense that a randomly chosen pure state from an appropriate energy shell reproduces the Gibbs-state expectation values of most observables acting on a small subsystem. However, such typical states are intrinsically stochastic, conventionally generated from the random high-temperature state, suffer from statistical fluctuations, and do not provide a continuous connection to the ground state. Here, we introduce a deterministic variational framework for constructing a Markov-Shielded Typical (MST) state, a unique thermal pure-state representative whose local reduced density matrices closely approximate those of the Gibbs state, from the ground state up to infinite-temperature limit. Our approach reformulates free-energy minimization by replacing the nonlocal von Neumann entropy with a local, computationally tractable Markov entropy, where we establish rigorous upper and lower bounds of MST free energy to lie in the very vicinity of the thermal free energy. This framework extends the concept of typicality to deterministic pure-state representations of thermal equilibrium, offering a scalable wave-function approach with potential applications to both classical and quantum computing.

main_202611.pdf

Markov-Shielded Typicality for Thermal Quantum States

Introduction

The traditional framework of statistical mechanics states that the thermal equilibrium is described by an ensemble of mixed states called the Gibbs state. Gibbs state is often introduced as a minimizer of the generalized free energy for a Hamiltonian HH, given as a functional of the density operator ρ\rho:

F~(ρ;T)=Tr⁡[ρH]−TSvN(ρ),(1)\widetilde{F}(\rho; T) = \operatorname{Tr}[\rho H] - T S_{\mathrm{vN}}(\rho), \tag*{(1)}

where SvN(ρ)=−Tr⁡[ρlog⁡ρ]S_{\mathrm{vN}}(\rho) = - \operatorname{Tr}[\rho\log\rho] is the von Neumann entropy. At temperature T=1/βT = 1/\beta, the variational principle naturally yields the well-known form of the Gibbs density operator, ρG(β)=e−βH/Z(β)\rho_{\mathrm{G}}(\beta) = e^{-\beta H}/Z(\beta) with Z(β)=Tr⁡[e−βH]Z(\beta) = \operatorname{Tr}[e^{-\beta H}], which gives the thermal entropy Sth=SvN(ρG)S_{\mathrm{th}} = S_{\mathrm{vN}}(\rho_{\mathrm{G}}) and the thermal free energy Fth(T)=F~(ρG,T)F_{\mathrm{th}}(T) = \widetilde{F}(\rho_{\mathrm{G}}, T), accordingly.

Modern theories have updated this paradigm to claim that this ensemble operator, which costs numbers of averaging that grows exponentially with system size NN, and is often unreachable numerically, is unnecessary, and instead one may refer to an individual pure state [1, 2], which is later named the thermal pure quantum (TPQ) state [3, 4, 5]. To clarify their relationships, one needs to compare the open and isolated systems. Suppose that the open system AA is connected to a bath and equilibrates after a long enough timescale to a Gibbs state based on a canonical mixture in connection to a bath. On the other hand, if we consider a mixed state in the system AA and purify it by attaching a sufficient auxiliary state BB, which is proven to be always possible, A+BA+B form an isolated system. The eigenstate thermalization hypothesis [6, 7, 8] claims that the majority of states realized on isolated systems, say A+BA+B, in a far-from-equilibrium environment, can evolve in time to a true equilibrium. The philosophy of canonical typicality [1, 2, 9, 10] asserts that a system AA, whether embedded in an isolated TPQ state or coupled to a thermal bath, exhibits statistically almost indistinguishable reduced properties. This equivalence is illustrated schematically in Figure 1(a). The density operator of AA obtained from the pure state of A+BA+B by tracing out the degrees of freedom of BB, ρA=Tr⁡B∣ψTPQ(β)⟩⟨ψTPQ(β)∣\rho_A = \operatorname{Tr}_B |\psi_{\mathrm{TPQ}}(\beta)\rangle\langle\psi_{\mathrm{TPQ}}(\beta)|, reproduces the mixed-state ensemble with the Boltzmann weight naturally emerging in the process. An “overwhelming majority” of TPQ states is thus thermodynamically equivalent to the Gibbs state when one considers physical observables within a sufficiently large “subsystem” AA.

Figure 1. Schematic illustration of (a) TPQ states (A+BA+B) and Gibbs state, and (b) the free energy landscape of how the TPQ state evolves from the high temperature limit with imaginary time. Inversely, our MST state starts from T=0T=0 and traces the most typical state among the TPQ shells toward the highest temperature.

The realization of TPQ state in computers [11, 12, 13] is demonstrated by utilizing the randomness of unitary matrices, which embodies the concept of typical sets. Choosing a pure state uniformly at random, ∣ψ0⟩|\psi_0\rangle, which represents the highest TT state, the imaginary-time evolution cools the state down to a temperature β−1\beta^{-1} as ∣ψTPQ(β)⟩∝e−βH/2∣ψ0⟩|\psi_{\mathrm{TPQ}}(\beta)\rangle\propto e^{-\beta H/2}|\psi_0\rangle. Its density matrix, ρTPQ=Tr⁡B∣ψTPQ⟩⟨ψTPQ∣\rho_{\mathrm{TPQ}} = \operatorname{Tr}_B |\psi_{\mathrm{TPQ}}\rangle\langle\psi_{\mathrm{TPQ}}|, agrees with the density matrix of a Gibbs state stochastically at a very high rate.

However, ρTPQ\rho_{\mathrm{TPQ}} suffers from two issues by definition. First, ρTPQ\rho_{\mathrm{TPQ}} is a stochastic pure-state representative that has substantial fluctuations in finite-size systems that depend on the choice of ∣ψ0⟩|\psi_0\rangle. This fluctuation is proven to be bounded by e−Sthe^{-S_{\mathrm{th}}} [4], which increases rapidly at lower temperatures because, intuitively, the law of large numbers loses its power, which becomes serious at the limited system sizes available for quantum many-body calculations. This is the implication of the “overwhelming majority”, and within that random fluctuation, the states obtained are mostly “typical.” Second, it is constructured by cooling the highest temperature state ∣ψ0⟩|\psi_0\rangle via the imaginary time evolution, which never reaches the ground state. Here, accessing low-temperature Gibbs states can be even more challenging, as illustrated by recent studies of quasi-local quantum Gibbs samplers [14, 15, 16]. These samplers prepare the Gibbs state ρG\rho_G through Lindblad dissipation, and rigorous proofs establish convergence to ρG\rho_G in polynomial time. However, their results are restricted to a high-temperature regime, well above the characteristic energy scale of the model. This fact highlights both the importance and the challenge of constructing deterministic pure-state representations of thermal equilibrium over a broad temperature range.

Then, the question is whether we can prepare an ideal thermal pure state in a computer that overcomes the above two issues inherent to TPQ state. To qualify how each TPQ state closely resembles the Gibbs state, the framework called microscopic thermal equilibrium (MITE) [17] was introduced. The pure state is said to be MITE if its local density matrices of all subsystems up to a specified size called kk-local (with k<N/2k < N/2) are equal to those of the Gibbs state within an error ϵ\epsilon. In this work, we clarify that a thermal state can be uniquely determined variationally, and mimics MITE. To generate such a state, we propose a variational framework using a generalized Markov free energy F~M\widetilde{F}_{M} [18, 19] as a functional of a pure state, which is practically accessible in a quantum many-body calculation for arbitrary temperature. Denoting by Spure\mathcal{S}_{\mathrm{pure}} the set of pure-state density matrices, we define

ψMST(T)=arg⁡min⁡ψ∈Spure F~M(ψ;T).(2)\psi_{\mathrm{MST}}(T)=\underset{\psi\in\mathcal{S}_{\mathrm{pure}}}{\arg\min}\ \widetilde{F}_{M}(\psi;T). \tag*{(2)}

We call ψMST(T)=∣ψMST(T)⟩⟨ψMST(T)∣\psi_{\mathrm{MST}}(T)=|\psi_{\mathrm{MST}}(T)\rangle\langle\psi_{\mathrm{MST}}(T)| a Markov-shielded typical (MST) state. Unlike the conventional free energy (1) which includes the nonlocal von Neumann entropy being strictly zero for the pure state, F~M\widetilde{F}_{M} is defined based on local quantities and gives for any subsystem an optimal solution. Notably, the ground state, which is unique, continues to Eq. (2) at finite temperature, which makes this MST state the utmost typical and different in philosophy from a TPQ state as a stochastic pure-state representative, as illustrated in Figure 1(b).

We show analytically that, under physically motivated assumptions, the Markov free energy of the MST state closely approximates the thermodynamic free energy, and that approximate MITE states are nearly optimal candidates for MST optimization. Using a tensor network ansatz for a one-dimensional (1D) spin system, we also demonstrate that the optimized MST states closely reproduce the local reduced density matrices of the Gibbs state, providing numerical evidence for the existence of MITE.

Preliminaries

We consider a quantum many-body system described by a density matrix ρ\rho. Using its reduced density matrix, ρA=Tr⁡Ac[ρ]\rho^{A}=\operatorname{Tr}_{A^{c}}[\rho], for a subsystem AA, the entanglement entropy S(A)ρ=SvN(ρA)S(A)_{\rho}=S_{\mathrm{vN}}(\rho^{A}) is obtained, where we omit the subscript ρ\rho unless otherwise needed. If ρ\rho is a thermal equilibrium state, then S(A)ρS(A)_{\rho} is equivalent to the thermal entropy SthS_{\mathrm{th}} of AA, and we derive a Markov entropy that replaces it.

Quantum Markov property

Let us consider a lattice system and assign a site index by drawing a 1D path. We now focus on site kk, and denote a set of sites labeled by numbers smaller than kk as {<k}\{< k\}, the neighborhood of site kk as Nk\mathcal{N}_{k}, and Mk={<k}∖Nk\mathcal{M}_{k}=\{< k\}\setminus\mathcal{N}_{k} is termed Markov shield. We consider the set of three regions AA, BB, CC as kk, Mk\mathcal{M}_{k} and {<k}∖Mk\{< k\}\setminus\mathcal{M}_{k}, respectively, to introduce the notion of shielding. We say that BB shields AA from CC if all paths along the lattice bonds which start in AA and end in CC go through BB. Figure 2 shows two examples that (a) BB shields AA from CC, and (b) which does not.

Figure 2. In the MED method, lattice sites are labeled according to a specific ordering: (a) in 1D systems, numbering proceeds from left to right, and (b) in 2D systems, from the top left. The Markov shield Mk\mathcal{M}_k (blue) is defined as the intersection between the set of sites labeled with indices smaller than kk, {<k}\{ < k \} (yellow), and the neighborhood of site kk, Nk\mathcal{N}_k (green). The parameter mm denotes the width of the Markov shield. If we take A={k}A = \{k\}, B=MkB = \mathcal{M}_k, and C={<k}\MkC = \{ < k \} \backslash\mathcal{M}_k, then in case (a), BB shields AA from CC, whereas in case (b), it does not.

Remind that in the classical probabilistic theory, if a distribution of XX is determined solely by YY independent of ZZ, it is equivalent to the vanishing of the mutual information between XX and ZZ conditioned by YY, which is called Markov property, as is common in the Monte Carlo simulation. The mutual information between XX and ZZ conditioned by YY is defined as

I(X:Z∣Y)=H(XY)+H(YZ)−H(XYZ)−H(Y),(3)I(X : Z \mid Y) = H(XY) + H(YZ) - H(XYZ) - H(Y), \tag*{(3)}

where H(X)H(X) is the Shannon entropy of a random variable XX. Similarly, in quantum systems, by reading off the Shannon entropy as von Neumann entropy, the zero of the mutual information for the subsystems AA, BB, and CC, given as

I(A:C∣B)ρ=S(AB)+S(BC)−S(ABC)−S(B),(4)I(A : C \mid B)_{\rho} = S(AB) + S(BC) - S(ABC) - S(B), \tag*{(4)}

implies a quantum Markov property. Again in Figure 2, if the quantum state ρ\rho on a given lattice satisfies I(A:C∣B)ρ=0I(A : C \mid B)_{\rho} = 0 for arbitrary choice of BB that shields AA from CC, it is called a quantum Markov network. Indeed, I(A:C∣B)ρ=0I(A : C \mid B)_{\rho} = 0 is both a necessary and a sufficient condition to recover the information about AA using the information solely from BB, known as the Petz recovery map [20, 21]. The quantum Markov networks are equivalent to the Gibbs states of a commuting Hamiltonian [22], which is the quantum version of the Hammersley-Clifford theorem [23]. The general Gibbs states have an approximate Markov property [24, 25, 26, 27, 28]. In 1D systems, the mutual information is bounded by

I(A:C∣B)ρG≤c1e−m/ξ,(5)I(A : C \mid B)_{\rho_G} \le c_1 e^{-m/\xi}, \tag*{(5)}

where c1c_1 and ξ\xi are temperature-dependent constants, and mm is the size of the Markov shield BB [25]. Here, the polynomial prefactor in the original bound is absorbed into the constants by slightly weakening the exponential decay rate.

Markov entropy decomposition

We now introduce the Markov entropy decomposition (MED), a method developed and refined to provide a lower bound on the thermodynamic free energy [18, 19]. Since a pure state has vanishing von Neumann entropy by definition, SvN(∣ψ⟩⟨ψ∣)=0S_{\mathrm{vN}}(|\psi\rangle\langle\psi|) = 0, the direct application of Eq. (1) does not make sense. The MED instead allows us to evaluate an upper bound of the von Neumann entropy SvNS_{\mathrm{vN}} using only local quantities. We first rewrite it using the conditional entropy S(A∣B)=S(AB)−S(B)S(A \mid B) = S(AB) - S(B):

SvN(ρ)=S(1)+∑k=2N[S({≤k})−S({<k})]=∑k=1NS(k∣{<k}).(6)\begin{aligned} S_{\mathrm{vN}}(\rho) &= S(1) + \sum_{k=2}^{N} [S(\{ \le k \}) - S(\{ < k \})] \\ &= \sum_{k=1}^{N} S(k \mid\{ < k \}). \tag*{(6)} \end{aligned}

Employing the inequality of strong subadditivity S(A∣BC)≤S(A∣B)S(A \mid BC) \le S(A \mid B) [29], the conditional entropy is bounded by the local quantity using a Markov shield:

S(k∣{<k})≤S(k∣{<k}∩Nk)=S(k∣Mk).(7)S(k \mid\{ < k \}) \le S(k \mid\{ < k \} \cap\mathcal{N}_k) = S(k \mid\mathcal{M}_k). \tag*{(7)}

The von Neumann entropy is bounded by their sum as

SvN(ρ)=∑k=1NS(k∣{<k})≤SM(ρ),(8)S_{\mathrm{vN}}(\rho) = \sum_{k=1}^{N} S(k \mid\{ < k \}) \le S_M(\rho), \tag*{(8)}

where we use the Markov entropy defined as

SM(ρ)=∑k=1NS(k∣Mk).(9)S_M(\rho) = \sum_{k=1}^{N} S(k \mid\mathcal{M}_k). \tag*{(9)}

The generalized free energy Eq. (1) is then bounded by the generalized Markov free energy F~M\widetilde{F}_M as

F~M(ρ;T)=Tr⁡[ρH]−TSM(ρ)≤F~(ρ;T).(10)\widetilde{F}_M(\rho; T) = \operatorname{Tr}[\rho H] - T S_M(\rho) \le\widetilde{F}(\rho; T). \tag*{(10)}

Since Eq. (10) holds for any given ρ\rho, the inequality is preserved upon minimizing both sides with respect to ρ\rho. The minimization of the generalized Markov free energy F~M(ρ;T)\widetilde{F}_M(\rho; T) gives the Markov free energy FM(T)F_M(T), and we find,

FM(T)=min⁡ρF~M(ρ;T)≤min⁡ρF~(ρ;T)=Fth(T).(11)F_M(T) = \min_{\rho} \widetilde{F}_M(\rho; T) \le\min_{\rho} \widetilde{F}(\rho; T) = F_{\mathrm{th}}(T). \tag*{(11)}

This means that FM(T)F_M(T) gives a lower bound for the thermodynamic free energy Fth(T)F_{\mathrm{th}}(T). In numerical calculations, the MED algorithm optimizes a set of local density matrices subject to consistency constraints on their overlapping regions [18, 19].

Although these constraints ensure consistency among overlapping marginals, they do not necessarily guarantee the existence of a global physical state whose reduced density matrices reproduce all the optimized marginals. This reflects the intrinsic difficulty of the quantum marginal problem [30, 31]. Nevertheless, the MED algorithm is rigorously guaranteed to provide a lower bound on the thermal free energy.

Theoretical framework

We consider a 1D quantum lattice system consisting of NN sites with open boundary conditions (OBC) and a local degrees of freedom dd at each site, described by a finite-range Hamiltonian

H=∑ihi,(12)H=\sum_i h_i, \tag*{(12)}

where hih_i denotes a local interaction term belonging to site ii. We assume that the number of hih_i is formally smaller than or equal to NN and that their operator norms are bounded by

J=max⁡i∥hi∥op,(13)J=\max_i\lVert h_i\rVert_{\mathrm{op}}, \tag*{(13)}

for some finite JJ where ∥⋅∥op\lVert\cdot\rVert_{\mathrm{op}} denotes the operator norm.

Markov-shielded typical state

We now pin down the relationships between the thermal free energy FthF_{\mathrm{th}} minimized via Eq. (2) and the other free energies about the mixed and pure states, which are schematically illustrated in Figure 3. The Markov free energy of the MST state lies in between the lower and upper bounds of FthF_{\mathrm{th}}, mediated by the inequalities between the free energies of the MED and MITE, which we formulate in the following.

Figure 3. Schematic illustration of our proofs in Sec. III A. From Theorem 1, we find F~M(ψMST)/N≤F~M(ψMITE)/N≤F~M(ρG)/N+δ(ϵ,T)≤Fth/N+δ(ϵ,T)\widetilde{F}_M(\psi_{\mathrm{MST}})/N \leq\widetilde{F}_M(\psi_{\mathrm{MITE}})/N \leq\widetilde{F}_M(\rho_G)/N+\delta(\epsilon,T) \leq F_{\mathrm{th}}/N+\delta(\epsilon,T). Similarly, from Theorem 2, Fth/N−c2Te−m/ξ′≤F~M(ρMED)/N≤F~M(ψMST)/NF_{\mathrm{th}}/N-c_2Te^{-m/\xi'} \leq\widetilde{F}_M(\rho_{\mathrm{MED}})/N \leq\widetilde{F}_M(\psi_{\mathrm{MST}})/N. These inequalities guarantee that the MST free energy is bounded from both directions by FthF_{\mathrm{th}}.

We consider the setup illustrated in Figure 2(a), where BB shields AA from CC. Based on Eq. (10), the difference between the two free energies is given by the conditional mutual information as

F~(ρ;T)−F~M(ρ;T)=T∑k=1N[S(k∣Mk)−S(k∣{<k})]=T∑k=1NI(k:{<k}\Mk∣Mk).(14)\begin{aligned} \widetilde{F}(\rho;T)-\widetilde{F}_M(\rho;T)=T\sum_{k=1}^{N}\left[S(k\mid\mathcal{M}_k)-S\left(k\mid\{<k\}\right)\right] \\ &=T\sum_{k=1}^{N}I\left(k:\{<k\}\backslash\mathcal{M}_k\mid\mathcal{M}_k\right). \tag*{(14)} \end{aligned}

Plugging ρG\rho_G into Eq. (14) and combining it with Eq. (5), we obtain

0≤Fth(T)−F~M(ρG;T)N≤c1Te−m/ξ,(15)0\leq\frac{F_{\mathrm{th}}(T)-\widetilde{F}_M(\rho_G;T)}{N}\leq c_1Te^{-m/\xi}, \tag*{(15)}

where we used F~(ρG;T)=Fth(T)\widetilde{F}(\rho_G;T)=F_{\mathrm{th}}(T). This shows that the Markov free energy density of the Gibbs state approaches the thermal free energy density exponentially with increasing Markov shield size mm.

Next, we introduce the pure state that approximates MITE [17].

Definition 1. A pure state ψMITE∈Spure\psi_{\mathrm{MITE}}\in\mathcal{S}_{\mathrm{pure}} is said to be an ϵ\epsilon-approximate kk-local MITE state if

D(ψMITEA,ρGA)≤ϵ(16)D\left(\psi_{\mathrm{MITE}}^A,\rho_G^A\right)\leq\epsilon \tag*{(16)}

for every contiguous subsystem AA with ∣A∣≤k|A|\leq k, where D(ρ,σ)=Tr⁡[∣ρ−σ∣]/2D(\rho,\sigma)=\operatorname{Tr}[|\rho-\sigma|]/2 denotes the trace distance.

Hereafter, we refer to such a state as an (ϵ,k)(\epsilon,k)-MITE state for brevity. The following lemma bounds the difference between the Markov free energies of an (ϵ,m+1)(\epsilon,m+1)-MITE state and the Gibbs state. In what follows, we assume that each local Hamiltonian term is supported within at most m+1m+1 consecutive sites.

Lemma 1. For any (ϵ,m+1)(\epsilon,m+1)-MITE state ψMITE\psi_{\mathrm{MITE}},

∣F~M(ψMITE;T)−F~M(ρG;T)∣N≤δ(ϵ,T),(17)\frac{\left|\widetilde{F}_M(\psi_{\mathrm{MITE}};T)-\widetilde{F}_M(\rho_G;T)\right|}{N}\leq\delta(\epsilon,T), \tag*{(17)}

where

δ(ϵ,T)=2Jϵ+T[2ϵlog⁡d+(1+ϵ)h2(ϵ1+ϵ)]\delta(\epsilon,T)=2J\epsilon+T\left[2\epsilon\log d+(1+\epsilon)h_2\left(\frac{\epsilon}{1+\epsilon}\right)\right]
=O(ϵTlog⁡(1/ϵ))(T>0, ϵ→0),(18)=\mathcal{O}(\epsilon T\log(1/\epsilon))\quad(T>0,\ \epsilon\to0), \tag*{(18)}

and h2(x)=−xlog⁡x−(1−x)log⁡(1−x)h_2(x)=-x\log x-(1-x)\log(1-x) is the binary entropy.

Proof. By the definition of (ϵ,m+1)(\epsilon,m+1)-MITE, the energy difference satisfies

∣Tr⁡[(ψMITE−ρG)H]∣≤2∑i∥hi∥opD(ψMITEAi,ρGAi)\left|\operatorname{Tr}\left[(\psi_{\mathrm{MITE}}-\rho_G)H\right]\right|\leq2\sum_i\lVert h_i\rVert_{\mathrm{op}}D\left(\psi_{\mathrm{MITE}}^{A_i},\rho_G^{A_i}\right)
≤2NJϵ.(19)\leq2N J\epsilon. \tag*{(19)}

For the entropy term, the continuity bound for quantum conditional entropy [32] gives

∣S(k∣Mk)ψMITE−S(k∣Mk)ρG∣≤2ϵlog⁡d+(1+ϵ)h2(ϵ1+ϵ).(20)\left|S(k \mid\mathcal{M}_k)_{\psi_{\mathrm{MITE}}} - S(k \mid\mathcal{M}_k)_{\rho_G}\right| \le2\epsilon\log d + (1+\epsilon)h_2\left(\frac{\epsilon}{1+\epsilon}\right). \tag*{(20)}

Summing over all sites and combining this with the energy bound yields the stated result.

If an (ϵ,m+1)(\epsilon,m+1)-MITE state exists and δ(ϵ,T)\delta(\epsilon,T) is small, its Markov free energy and that of the MST state provide approximate lower bounds on the thermodynamic free energy. Combining the previous lemma with Eq. (15) and the MST variational principle, we obtain the following theorem.

Theorem 1. Assuming the existence of an (ϵ,m+1)(\epsilon,m+1)-MITE state ψMITE\psi_{\mathrm{MITE}}, we have

F~M(ψMITE;T)−Fth(T)N≤δ(ϵ,T).(21)\frac{\widetilde{F}_M(\psi_{\mathrm{MITE}};T)-F_{\mathrm{th}}(T)}{N}\leq\delta(\epsilon,T). \tag*{(21)}

By the variational principle in Eq. (2), the same bound holds for ψMST\psi_{\mathrm{MST}}.

Proof. From Eq. (15), we have

F~M(ρG;T)≤Fth(T).(22)\widetilde{F}_M(\rho_G;T)\leq F_{\mathrm{th}}(T). \tag*{(22)}

Combining this with Lemma 1, we obtain

F~M(ψMITE;T)≤F~M(ρG;T)+Nδ(ϵ,T)≤Fth(T)+Nδ(ϵ,T),(23)\begin{aligned} \widetilde{F}_M(\psi_{\mathrm{MITE}};T)&\leq\widetilde{F}_M(\rho_G;T)+N\delta(\epsilon,T)\\ &\leq F_{\mathrm{th}}(T)+N\delta(\epsilon,T), \tag*{(23)} \end{aligned}

which yields the stated bound. By the variational definition of the MST state in Eq. (2),

F~M(ψMST;T)≤F~M(ψMITE;T),(24)\widetilde{F}_M(\psi_{\mathrm{MST}};T)\leq\widetilde{F}_M(\psi_{\mathrm{MITE}};T), \tag*{(24)}

so the same bound holds for ψMST\psi_{\mathrm{MST}}.

This theorem now proves that FthF_{\mathrm{th}} gives the upper bound of F~M(ψMITE)\widetilde{F}_M(\psi_{\mathrm{MITE}}), which is the free energy of the pure state mimicking the Gibbs state and satisfying at most the canonical typicality condition.

The next step is to show that FthF_{\mathrm{th}} bounds the pure state free energy from the opposite side. For that purpose, we introduce the MED state which is defined as a minimizer of the Markov free energy over all physical density matrices:

ρMED=arg⁡min⁡ρF~M(ρ;T).(25)\rho_{\mathrm{MED}}=\mathop{\arg\min}_{\rho}\widetilde{F}_M(\rho;T). \tag*{(25)}

We assume that ρMED\rho_{\mathrm{MED}} satisfies the approximate Markov property

I(A:C∣B)ρMED≤c2e−m/ξ′,(26)I(A:C\mid B)_{\rho_{\mathrm{MED}}}\leq c_2e^{-m/\xi'}, \tag*{(26)}

for every tripartition associated with a Markov shield as shown in Figure 2(a), where c2c_2 and ξ′\xi' are temperature-dependent constants. Under this assumption, the Markov free energy of the MST state provides an approximate upper bound on the thermodynamic free energy when mm is sufficiently large.

Theorem 2. Assuming the approximate Markov property of the MED state in Eq. (26), the MST state satisfies

Fth(T)−F~M(ψMST;T)N≤c2Te−m/ξ′.(27)\frac{F_{\mathrm{th}}(T)-\widetilde{F}_M(\psi_{\mathrm{MST}};T)}{N}\leq c_2Te^{-m/\xi'}. \tag*{(27)}

Proof. By the approximate Markov property of ρMED\rho_{\mathrm{MED}} and Eq. (14), we have

0≤F~(ρMED;T)−F~M(ρMED;T)≤c2NTe−m/ξ′.(28)0\leq\widetilde{F}(\rho_{\mathrm{MED}};T)-\widetilde{F}_M(\rho_{\mathrm{MED}};T) \leq c_2NTe^{-m/\xi'}. \tag*{(28)}

Since Fth(T)≤F~(ρMED;T)F_{\mathrm{th}}(T)\leq\widetilde{F}(\rho_{\mathrm{MED}};T), it follows that

F~M(ρMED;T)≥Fth(T)−c2NTe−m/ξ′.(29)\widetilde{F}_M(\rho_{\mathrm{MED}};T)\geq F_{\mathrm{th}}(T)-c_2NTe^{-m/\xi'}. \tag*{(29)}

By the variational definitions of the MED and MST states,

F~M(ρMED;T)≤F~M(ψMST;T).(30)\widetilde{F}_M(\rho_{\mathrm{MED}};T)\leq\widetilde{F}_M(\psi_{\mathrm{MST}};T). \tag*{(30)}

Combining these inequalities yields the bound stated.

The above two theorems give the following corollary:

Corollary 1. Assuming the existence of an (ϵ,m+1)(\epsilon,m+1)-MITE state ψMITE\psi_{\mathrm{MITE}} and the approximate Markov property of the MED state in Eq. (26), the MST state satisfies

−c2Te−m/ξ′≤F~M(ψMST;T)−Fth(T)N≤δ(ϵ,T).(31)-c_2Te^{-m/\xi'}\leq\frac{\widetilde{F}_M(\psi_{\mathrm{MST}};T)-F_{\mathrm{th}}(T)}{N}\leq\delta(\epsilon,T). \tag*{(31)}

For any such MITE state,

0≤F~M(ψMITE;T)−F~M(ψMST;T)N≤δ(ϵ,T)+c2Te−m/ξ′.(32)0\leq\frac{\widetilde{F}_M(\psi_{\mathrm{MITE}};T)-\widetilde{F}_M(\psi_{\mathrm{MST}};T)}{N} \leq\delta(\epsilon,T)+c_2Te^{-m/\xi'}. \tag*{(32)}

When both δ(ϵ,T)\delta(\epsilon,T) and c2Te−m/ξ′c_2Te^{-m/\xi'} are small or approaches zero, the MST Markov free energy approximates the thermodynamic free energy, and approximate MITE state becomes the optimal candidate for MST optimization. These results provide a theoretical basis for the MST variational principle in Eq. (2), where the MST serves as the most typical thermal equilibrium state, as far as the minimum F~M\widetilde{F}_M is safely identified. We notice that the exact MITE states with ϵ=0\epsilon=0 have been explicitly constructed at infinite temperature [33, 34, 35, 36, 37], but not at T<∞T<\infty, meaning that our construction and numerical demonstration together provide evidence for their existence.

MST algorithm

Supported by the proof in the previous subsection, we propose a numerical approach to obtain the MST state variationally. Here, any types of description of the wave

Algorithm 1 Variational optimization of MST states

Require: Hamiltonian HH, variational wave function ∣ψ(θ)⟩|\psi(\boldsymbol{\theta})\rangle, Markov free energy F~M\widetilde{F}_M, temperature grid 0=T0<T1<⋯<TL=Tmax⁡0 = T_0 < T_1 < \cdots< T_L = T_{\max}

Ensure: Optimized variational states {∣ψ(θℓ)⟩}ℓ=0L\{|\psi(\boldsymbol{\theta}_{\ell})\rangle\}_{\ell=0}^{L}

  1. Initialize θ0\boldsymbol{\theta}_0 by approximately minimizing the ground-state energy

  2. for ℓ=1,…,L\ell= 1,\ldots,L do

  3. Set θ←θℓ−1\boldsymbol{\theta} \leftarrow\boldsymbol{\theta}_{\ell-1}

  4. Minimize F~M(∣ψ(θ)⟩⟨ψ(θ)∣;Tℓ)\widetilde{F}_M(|\psi(\boldsymbol{\theta})\rangle\langle\psi(\boldsymbol{\theta})|;T_{\ell}) with respect to θ\boldsymbol{\theta}

  5. Store the optimized parameters as θℓ\boldsymbol{\theta}_{\ell}

  6. end for

  7. return {∣ψ(θℓ)⟩}ℓ=0L\{|\psi(\boldsymbol{\theta}_{\ell})\rangle\}_{\ell=0}^{L}

function ∣ψ(θ)⟩|\psi(\boldsymbol{\theta})\rangle using a finite number of variational parameters θ\boldsymbol{\theta} can be used, such as variational Monte Carlo methods [38], Boltzmann machine [39], and density matrix renormalization group [40, 41].

The optimization process involves minimizing the Markov free energy with respect to the variational parameters θ\boldsymbol{\theta}. One needs to straightforwardly optimize the wave function by using the Markov free energy F~M(∣ψ⟩⟨ψ∣,T)\widetilde{F}_M(|\psi\rangle\langle\psi|,T) as a complex cost function for each given TT. However, because F~M\widetilde{F}_M is a functional of θ\boldsymbol{\theta} with many local optima, we start from the ground state that minimizes the Hamiltonian, and then using the solution ∣ψ(θT)⟩|\psi(\boldsymbol{\theta}_T)\rangle at TT as an initial state, obtain ∣ψ(θT+ΔT)⟩|\psi(\boldsymbol{\theta}_{T+\Delta T})\rangle, by gradually raising the temperature up to Tmax⁡T_{\max}. The temperature-continuation procedure is summarized in Algorithm 1.

Although computing the ground state is classified as QMA-complete [42, 43, 44], implying that it is generally hard, most of the physical models with short-range interactions, do not apply to these worst-case scenarios. For instance, by focusing on frustration-free systems, where the ground state can be obtained easily, we can circumvent this problem.

We also notice the crucial difference between our MST and MED algorithm. The MED algorithm optimizes the locally consistent density matrices, whereas our approach considers the Markov free energy F~M\widetilde{F}_M as a functional of a global pure state ψ\psi. The phase space to be optimized also differ; the MED optimization problem is always convex because it is formulated in terms of local density operators. The MST state is defined over the pure state manifold, ψ∈Spure\psi\in\mathcal{S}_{\mathrm{pure}} in Eq. (2), which makes the optimization problem nonconvex. We demonstrate in our simulation that we are indeed able to reach the minimum by continuously tracking the variational solution from the ground state as the temperature increases.

Numerical demonstrations

MST-MPS implementation

For the numerical demonstration, we employ a matrix product state (MPS) ansatz [45] to optimize the Markov free energy in Spure\mathcal{S}_{\mathrm{pure}}. Since the standard MPS-form suffers a severe area-law bound of the entanglement entropy in practice, we adopt the representation called TPQ-MPS with auxiliary degrees of freedom attached to both ends of the chain, which we have previously succeeded in recovering the volume-law entanglement entropy for the description of the thermal state [46]:

∣ψ(θ)⟩=∑{α}∑{i}Aα0α1[1]i1Aα1α2[2]i2⋯AαN−1αN[N]iN∣α0,i1,i2,…,iN,αN⟩=tensor-network representation(33)\begin{aligned} |\psi(\boldsymbol{\theta})\rangle= \sum_{\{\alpha\}}\sum_{\{i\}} A^{[1]i_1}_{\alpha_0\alpha_1} A^{[2]i_2}_{\alpha_1\alpha_2} \cdots A^{[N]i_N}_{\alpha_{N-1}\alpha_N} |\alpha_0,i_1,i_2,\ldots,i_N,\alpha_N\rangle\\ &= \text{tensor-network representation} \tag*{(33)} \end{aligned}

which we have previously developed for calculating the so-called TPQ-MPS [46]. The advantage of this form is that it could maximize the amount of entanglement stored in the pure state, allowing us to have a volume-law entanglement entropy when measured from the center sites. This time, we use it as a variational ansatz and call it MST-MPS. The bond dimension α0\alpha_0 and αN\alpha_N are set to 1 for the ground state, while it increases naturally when we increase TT, due to the increasing entanglement. When starting at finite TT one needs to set α0=αN=χ\alpha_0 = \alpha_N = \chi, the same as other bond dimensions (see Supplementary information [47]).

The variational process consists of updating the nn-th tensor given below by setting the orthogonal center to that tensor to keep its canonical form:

tensor=∑αβ∑inAαβ[n]in∣α,in,β⟩,(34)\text{tensor} = \sum_{\alpha\beta}\sum_{i_n} A^{[n]i_n}_{\alpha\beta}|\alpha,i_n,\beta\rangle, \tag*{(34)}

and a set of processes of shifting nn from the left edge to the right edge and then back to the left, constitutes a single sweep of the update. At each step, the nn-th tensor is optimized to maximize the Markov entropy, where its contribution is mediated through the reduced density matrices entering the k=1:Nk = 1:N-th (m+1)(m+1)-local region of S(k∣Mk)S(k|M_k) in Eq. (9), which is connected to the nn-th tensor via the virtual bonds α\alpha or β\beta. Although the variational parameters are the local nn-th tensor, and this update does not change the remaining N−1N-1 tensors themselves, all the other marginals (local density matrices) change simultaneously, making this process nonlocal. At the same time, the whole set of MPS retains its pure state property by construction, namely all marginals automatically preserve the consistency condition as part of a single pure state.

This is in sharp contrast to the numerical implementation of the MED; the standard formulation is such that a single local density operator (marginal) is updated in a way to satisfy consistency constraints with its neighboring marginals that have overlaps. However, this does not imply that all these marginals are part of the same global quantum pure state in a compatible manner. Such global compatibility is computationally too demanding.

We use ITensors.jl for the MPS calculations [48, 49] and Optim.jl for the BFGS optimization of each tensor [50]. The system size and bond dimension were fixed at N=48N=48 and χ=8\chi=8, respectively. We set ΔT=0.1\Delta T=0.1 and Tmax⁡=2.0T_{\max}=2.0, and at each temperature, we perform up to maximally 10 sweeps, while setting the convergence of FM/NF_M/N between sweeps to be less than 10−410^{-4}. The computational bottleneck of this procedure is the diagonalization of a series of (m+1)(m+1)-site reduced density operators to evaluate the Markov entropy. The cost of optimizing a single matrix scales as O(Nd3m)O(Nd^{3m}), and a full sweep through the chain costs O(N2d3m)O(N^2d^{3m}), where dd is the local degree of freedom. For m=4m=4, the computation time per sweep took about 3 hours for a standard workstation.

In the following demonstrations, we use the 1D transverse-field Ising model with OBC,

H=−∑i=1N−1σizσi+1z−g∑i=1Nσix(35)H=-\sum_{i=1}^{N-1}\sigma_i^z\sigma_{i+1}^z-g\sum_{i=1}^{N}\sigma_i^x \tag*{(35)}

where σiα\sigma_i^\alpha (α=x,y,z\alpha=x,y,z) denotes the Pauli operator at site ii.

Escape from random fluctuations

We first compare the MST state obtained using the MST-MPS ansatz with TPQ states generated in the full Hilbert space, focusing on a small system of size N=16N=16.

We plot in Figure 4(a-c) the energy E=⟨H⟩E=\langle H\rangle, and magnetization Mα=∑i=1NσiαM_\alpha=\sum_{i=1}^{N}\sigma_i^\alpha in α=x\alpha=x and zz directions. For the TPQ state, we plot 10 different samples for the initial choice of random state, ∣ψ0⟩|\psi_0\rangle, and for the MST, only one sample as data points, which we tested for the shield of m=1,2,3,4m=1,2,3,4. One finds that at m=3m=3, the curve almost falls onto the exact diagonalization curve. Whereas, the standard TPQ state cannot get rid of the large random fluctuation. Notice that this random fluctuation is not the numerical error but the outcome of typicality, and is reduced when the system size becomes larger due to the entropic effect, but even at relatively large NN, it largely affects the data at low temperature. Picking up a single TPQ state thus provides an increasingly inaccurate result at low temperatures, where statistical fluctuations are more pronounced. This proves the physical picture shown in Figure 1 that the MST can track the most desirable curve by a one-shot calculation, which, for the TPQ state, is the curve obtained after taking a number of random sampling averages.

Figure 4. Comparison of the one-shot calculations of the TPQ state obtained by the imaginary time evolution using the full Hilbert space (blue lines), and the MST states with m=1,2,3,4m=1,2,3,4 using the MPS ansatz. (a) Energy, (b) xx and (c) zz magnetization of the transverse-field Ising chain with g=0.5g=0.5 for N=16N=16. Solid blue lines are the 10 different samples of the TPQ state, blue region is their variance, and the broken line is the exact result.

Numerical accuracy

We compare the MST-MPS results with high-precision reference values obtained via the purification method [51, 52, 53], which we refer to as numerically “exact.” In Figure 5(a-d) we plot the Markov entropy SMS_M as a function of E/NE/N, and TT dependent Markov free energy FMF_M, energy EE, and magnetization MxM_x, MzM_z, which are all local quantities. The accuracy of the MST-MPS systematically improves with mm and for m=4m=4, the data over the whole temperature range converge to the exact one. As we find in SMS_M, even in the vicinity of the quantum critical point at g=1g=1 and T→0T\to0, we find a remarkable agreement (for other data at g=1g=1 see Supplementary information [47]).

Figure 5. Physical quantities of the transverse-field Ising chain for N=48N = 48. We use the MST-MPS method for various sizes of Markov shields m=1,2,3,4m = 1, 2, 3, 4 and the purification method [51] as the exact results (dashed lines). (a) SM/NS_M/N as a function of E/NE/N for g=0.5g = 0.5 and 1. Temperature dependence of (b) Markov free energy FM/NF_M/N and energy E/NE/N, (c) transverse magnetization ⟨Mx⟩/N\langle M_x\rangle/N, and longitudinal magnetization ⟨Mz⟩/N\langle M_z\rangle/N, (d) specific heat per site calculated in two different ways: C1C_1 from Hamiltonian fluctuations, and C2C_2 from temperature derivatives of energy, calculated for g=0.5g = 0.5. For the critical case at g=1.0g = 1.0, see Supplementary Information [47].

Figure 5(d) shows the specific heat computed using the two definitions:

C1=⟨H2⟩−⟨H⟩2T2,(36)C_1=\frac{\langle H^2\rangle-\langle H\rangle^2}{T^2}, \tag*{(36)}
C2=d⟨H⟩dT≃⟨H⟩T+ΔT−⟨H⟩TΔT.(37)C_2=\frac{d\langle H\rangle}{dT}\simeq\frac{\langle H\rangle_{T+\Delta T}-\langle H\rangle_T}{\Delta T}. \tag*{(37)}

Notice that since the MST state can accurately reproduce only the local properties of the Gibbs state, C1C_1 becomes less reliable, whereas C2C_2 gives a much better agreement with the exact solutions. The latter precision can be further improved by controlling ΔT\Delta T.

How typical?

To measure the degree of canonical typicality of the MST, we now compute the trace distance D(ρ,σ)D(\rho,\sigma) of the reduced density operators ρ,σ\rho,\sigma of the MST-MPS and the Gibbs state obtained via the purification method, respectively. D(ρ,σ)D(\rho,\sigma) gives an upper bound on the difference in expectation values of arbitrary observables. In Figure 6(a), we show the trace distance at T=0.2T = 0.2 and g=0.5g = 0.5 for the nn contiguous subsystem sites taken from the center of the system. The value of DD is less than 0.06 for most cases, and with increasing the size of the shield mm up to 4, DD is almost fully suppressed up to n≲6n \lesssim6, demonstrating that the MST state reproduces the local reduced density matrices of the Gibbs state. Even at high enough temperature T=2.0T = 2.0, this tendency works (see Figure 6(b)), while for such high TT, the coherence of the quantum state itself is almost lost, so that the falling off of DD at n≳m+1n \gtrsim m + 1 from zero does not indicate the degradation of the quality of the thermal state, as the physical quality in that range is determined by possibly up to two site correlations. Another observation is that D∼0D \sim0 at n≤m+1n \leq m + 1 for all temperature range, meaning that if we want to analyze the quantity supported by a certain size scale, we need to prepare mm of the same order. The canonical typicality of the MST is thus attained regardless of the temperature as far as we properly confine ourselves to the kk-local support of the observables that rely on DD.

Figure 6. Transverse-field Ising chain of size N=48N = 48 and g=0.5g = 0.5. (a,b) Trace distance between local reduced density operators of the MST-MPS and those of the Gibbs state for various Markov shield sizes m=1,2,3,4m = 1, 2, 3, 4. The local density operators are computed for nn-site subsystems extracted from the center of the chain at T=0.2T = 0.2 and 2.02.0. Entanglement properties of the MST-MPS for various temperatures T=0.0T = 0.0–1.01.0. (c) Entanglement entropy of nn-site subsystems centered in the middle of the chain, and (d) the one across each cut position ii along the chain.

Entanglement property

The entanglement entropy (EE) is the basic property that characterizes the quantum state. In Figure 6(c), we show the EE for the nn-site subsystem from the center; For the first ∼10\sim10 sites, the entropy grows linearly, showing the volume law entanglement, which saturates for larger nn. This profile is almost the same as we found

for the TPQ-MPS state after a random average over a few samples [46]. Many of the meaningful physical properties at finite temperature can be measured within the lengthscale of this volume-law range.

Next, in Figure 6(d) we show the EE across the cut that divides the system into ii and (N−i)(N - i) sites on the left and right-hand sides, respectively. The EE profile is mostly flat and increases with temperature. At intermediate temperatures, there arises an asymmetry, which may reflect the directional nature of the Markov shielding used for MST-MPS. In the standard MPS without auxiliaries, EE is known to start from nearly zero at the edges and increase toward the center, known as the Page curve [54, 55], which seriously restricted the amount of entanglement. Here, with the aid of auxiliaries at the edges, the almost site-independent, nearly translationally symmetric nature is preserved, which we have previously found to be important for the volume law in terms of nn [46].

Discussion

We have established a pathway to access the most typical state within all the typical thermal pure quantum (TPQ) states that belong to the microcanonical shell, which is the Markov-shielded typical (MST) state. The TPQ states suffer inevitable random fluctuations, as they are generated from an infinite-temperature random state by cooling them down to the target temperature through the imaginary time evolution. These fluctuations are not numerical errors but are intrinsic and are significantly enhanced at lower temperatures and for smaller system sizes [56, 57]. Our MST state coincides with the random averages over all these TPQ states. We enabled the construction of the MST state as a minimizer of the fictitious free energy that uses the Markov entropy, which

is a local quantity, instead of the nonlocal von Neumann entropy. So far, there had been no numerical methods starting from the high-temperature side that smoothly connect to the ground state. The present scheme enables the variational approach for arbitrary temperatures, from the ground state to the infinite temperature, with a substantial accuracy, and more importantly, there is no restriction on the choice of the wave functional form as far as it is physical in the sense that it is globally defined consistently.

As the previous theories of typicality are all built on randomness [1, 2, 9, 10, 58], a typical pure state based on variational principles may appear incompatible with it at first glance, while the following interpretation is possible: Similarly to Landau’s generalized free energies designed as functions of order parameters, the Markov free energy has some “randomness” as a phase space parameter, and shows a landscape with many local minima that are all regarded as the typical states. A set of MST states are those that continuously connect the ground state, except for the system that undergo a first-order transition, they continue to form the unique lowest free-energy manifold of states. In that respect, the MST provides a deterministic pure-state representation of thermal equilibrium that satisfies the canonical typicality condition.

We finally make some remarks on the related works. Scalet et al. [59] have recently proposed a looser condition than the quantum Markov property to justify the MED. They showed that under a clustering condition, one can efficiently compute a state that is locally equivalent to the Gibbs state using the MED method. This idea can be straightforwardly applied to extend our MST approach to higher-dimensional systems.

There have also been numerical advances very recently. One is the imaginary-time-evolved entangled antipodal pair (EAP) state [35], that approaches again from the high temperature, while avoiding explicit initial randomness. While, this EAP state shall be an analogue of a purification method rather than the thermal pure state, since it is locally equivalent to a purified mixed state, particularly from the perspective of numerical algorithms [60]. Another route starting from the ground state to a finite temperature was proposed by Corboz et al. [61] using tree tensor networks. The method succeeded in reaching the lowest temperature part, yielding an approximate full Gibbs state, while the description seemingly falls off from the thermal state at T≳0.1T \gtrsim0.1, possibly because they rely on the assumption that the ground state is well described by a conformal field theory, or the shape of the tensor and the bond dimension may restrict the inclusion of the major excited states at higher temperatures. Additionally, the number of excited states that can be represented is restricted by the bond dimension of the tensor network, which might somehow be avoided by combining with our MST framework.

The problem of finding an MST can be viewed from a broader perspective. Given a Gibbs state at some arbitrary temperature, one may rephrase our MST as whether there exists a pure state whose local reduced density matrices reproduce all (m+1)(m+1)-local marginals of the Gibbs state. This question is closely related to the quantum marginal problem, which is known to be QMA-complete in general [62]. There, one asks whether a given collection of local density matrices is compatible with a global quantum state. Although several analytical approaches have been proposed [63], the practical construction of large-scale many-body quantum states satisfying prescribed local constraints remains highly challenging. In this context, MST provides a concrete pure-state variational framework as part of the quantum marginal problem. We also note that certain classes of quantum marginal problems obeying a volume-law structure can be solved analytically by mixed states that maximize the entropy [64], which is complementary to our approach.

Our results also suggest a promising route to representing thermal equilibrium on quantum devices using MITE states, which can serve as a complement to the Gibbs ensemble. As demonstrated in Sec. IV, MST-MPS can accurately reproduce local thermal properties with relatively small bond dimensions. Recent advances have shown that MPS can be prepared using logarithmic-depth quantum circuits [65, 66, 67, 68], or even constant-depth circuits with measurements and feedback [69, 70, 71, 72, 73, 74]. These developments highlight the potential of MST-MPS for efficient quantum simulation of thermal properties.

Acknowledgments

A. I. thanks Akihiro Hokkyo for helpful discussions on the definition of MITE. A. I. was supported by the Center of Innovations for Sustainable Quantum AI (JST Grant No. JPMJPF2221) and other JSPS KAKENHI (No. 25K17308). C. H. was supported by a Grant-in-Aid for Transformative Research Areas “The Natural Laws of Extreme Universe—A New Paradigm for Space-time and Matter from Quantum Information” (Grant No. 21H05191) and other JSPS KAKENHI (No. 21K03440).

Declarations

Data availability

The code used to generate the numerical results reported in this work is publicly available on GitHub [75].

Supplementary Information of ``Markov-Shielded Typicality for Thermal Quantum States''

Atsushi Iwaki; Chisa Hotta

supple20260320_iwaki.pdf

Supplementary Information of ``Markov-Shielded Typicality for Thermal Quantum States''

Optimization of MST-MPS

In this section, we describe the implementation details of the MST-MPS optimization. In our implementation based on ITensors.jl [1, 2], the MPS with auxiliary degrees of freedom is stored as a standard MPS defined on an enlarged physical Hilbert space. More specifically, we introduce Naux=log⁡2χN_{\mathrm{aux}} = \log_{2}\chi auxiliary sites on both the left and right sides. Accordingly, the total number of sites in the MPS is given by Ntot=N+2NauxN_{\mathrm{tot}} = N + 2N_{\mathrm{aux}}. For this MPS with auxiliary sites, we define a function that evaluates the Markov free-energy functional. We then optimize each of the NtotN_{\mathrm{tot}} tensors in sequence using the BFGS algorithm implemented in Optim.jl [3], sweeping from the left edge to the right edge and then back again. One such round trip is defined as a single sweep. The optimization is repeated over multiple sweeps and terminated when the change in the Markov free energy per sites becomes smaller than 10−410^{-4}.

Figure S1 shows the behavior of the cost function, namely the Markov free-energy functional, during each local tensor update for the case of N=16N = 16, χ=8\chi= 8, and m=3m = 3, when the temperature is increased from T=0.9T = 0.9 to T=1.0T = 1.0. For clarity, in this example we do not terminate the optimization according to the free-energy tolerance, and instead explicitly show the results for 4 sweeps. The cost function decreases rapidly during the 1st sweep, although the accuracy is not yet sufficient at that stage. After 2 sweeps, the result reaches an accuracy of approximately 10−410^{-4} compared with the exact value. After the 3rd and 4th sweeps, the value changes only marginally and appears to be saturated. This residual deviation should be interpreted not as a failure of the optimization, but rather as the intrinsic limitation of the free-energy approximation achievable within the MST-MPS ansatz. For practical purposes, the numerical accuracy is already sufficient at this stage.

Figure S1. Behavior of the cost function, namely the Markov free-energy functional, during the tensor updates in the MST-MPS optimization. Results are shown for the one-dimensional transverse-field Ising model with N=16N = 16 and g=0.5g = 0.5, when the temperature is raised from T=0.9T = 0.9 to T=1.0T = 1.0. A single sweep consists of updating the tensors from the left edge to the right edge and back again. The plot shows four sweeps in total, with the boundaries between sweeps marked by gray dotted lines. The exact value obtained by the purification method is indicated by the orange dashed line, while the shaded region represents values within 10−410^{-4} of the exact result.

MST-MPS calculations at the critical point

To support the reliability of MST-MPS shown in Fig. 5 in the main text, we perform the same set of calculations for different parameters of the transverse-field Ising chain. The value g=0.5g = 0.5, mainly used in the main-text figures, corresponds to a gapped ground state in one dimension. Therefore, we show the results of the gapless

critical point g=1g = 1, for which the relation between SM/NS_M/N and E/NE/N is already shown in Fig. 5(a) of the main text.

Figure S2 shows the temperature dependence of the Markov free energy FM/NF_M/N, the energy per site E/NE/N, and the magnetizations ⟨Mx⟩/N\langle M_x\rangle/N and ⟨Mz⟩/N\langle M_z\rangle/N for N=48N = 48 and m=1,2,3,4m = 1, 2, 3, 4, allowing for a direct comparison with Fig. 5(b-d) of the main text. In contrast to what might be expected near criticality, the MST-MPS results do not show a substantial deterioration in accuracy or an increase in fluctuations compared with the g=0.5g = 0.5 case, although the correlation length ξ\xi, which strongly affects the quantum Markov property, is expected to diverge near the critical point.

Figure S2. Physical quantities of the transverse-field Ising chain for N=48N = 48 at the critical point g=1.0g = 1.0. We use the MST-MPS method for various sizes of Markov shields m=1,2,3,4m = 1, 2, 3, 4 and the purification method [4] as the exact results (dashed lines). Temperature dependence of (a) the Markov free energy FM/NF_M/N and the energy per site E/NE/N, (b) the transverse and longitudinal magnetizations, ⟨MX⟩/N\langle M_X\rangle/N and ⟨MZ⟩/N\langle M_Z\rangle/N, and (c) the specific heat per site, evaluated in two different ways: C1C_1 from Hamiltonian fluctuations and C2C_2 from the temperature derivative of the energy.

Entanglement with auxiliary sites

The MST-MPS method proposed in this work is initialized from the ground state. Accordingly, we adopt an MPS ansatz with auxiliary sites as the variational manifold. At zero temperature, the physical and auxiliary degrees of freedom are completely disentangled, so that the effective bond dimension connected to the auxiliary1: [email protected] sector is equal to one. In practice, the corresponding unused part of the variational space is represented numerically by zero-filled components. As the temperature is raised, entanglement between the physical system and the auxiliary sites is progressively generated. This auxiliary entanglement provides the variational freedom necessary to construct an MST state that is locally equivalent to the Gibbs state. At the same time, the effective bond dimension in the auxiliary sector becomes larger as the temperature increases.

Figure S3 shows the entanglement entropy across bipartitions at the left and right edges of the system. At T=0T = 0, both quantities are zero, which indicates that the variational state is simply the ground state and does not rely on the auxiliary degrees of freedom. As the temperature increases, the bond entanglement also increases, demonstrating that the auxiliary sector is gradually activated to represent the finite-temperature MST state. The asymmetry between the left and right entanglement entropies reflects the intrinsic asymmetry of the Markov entropy.

References

Figure S3. Temperature dependence of the entanglement entropy across the bipartition at the left edge (blue circles) and the right edge (red squares), for N=48N = 48, g=0.5g = 0.5, χ=8\chi= 8, and m=4m = 4.

References

  1. [1]S. Popescu, A. J. Short, and A. Winter, Entanglement of the foundations of statistical mechanics, Nature Phys. 2, 754 (2006).DOI
  2. [2]S. Goldstein, J. L. Lebowitz, R. Tumulka, and N. Zanghì, Canonical typicality, Phys. Rev. Lett. 96, 050403 (2006).
  3. [3]S. Sugiura and A. Shimizu, Thermal pure quantum states at finite temperature, Phys. Rev. Lett. 108, 240401 (2012).pubmed.ncbi.nlm.nih.gov/23004239
  4. [4]S. Sugiura and A. Shimizu, Canonical thermal pure quantum state, Phys. Rev. Lett. 111, 010401 (2013).
  5. [5]M. Hyuga, S. Sugiura, K. Sakai, and A. Shimizu, Thermal pure quantum states of many-particle systems, Phys. Rev. B 90, 121110(R) (2014).DOI
  6. [6]J. M. Deutsch, Quantum statistical mechanics in a closed system, Phys. Rev. A 43, 2046 (1991).pubmed.ncbi.nlm.nih.gov/9905246
  7. [7]M. Srednicki, Chaos and quantum thermalization, Phys. Rev. E 50, 888 (1994).DOI
  8. [8]M. Rigol, V. Dunjko, and M. Olshanii, Thermalization and its mechanism for generic isolated quantum systems, Nature 452, 854–858 (2008).
  9. [9]P. Reimann, Typicality for generalized microcanonical ensembles, Phys. Rev. Lett. 99, 160404 (2007).DOI
  10. [10]A. Sugita, Nonlinear Phenom. Complex Syst 10, 192 (2007).
  11. [11]M. Imada and M. Takahashi, Quantum transfer monte carlo method for finite temperature properties and quantum molecular dynamics method for dynamical correlation functions, Journal of the Physical Society of Japan 55, 3354 (1986).
  12. [12]J. Jaklič and P. Prelovšek, Lanczos method for the calculation of finite-temperature quantities in correlated systems, Phys. Rev. B 49, 5065 (1994).
  13. [13]A. Hams and H. De Raedt, Fast algorithm for finding the eigenvalue distribution of very large matrices, Phys. Rev. E 62, 4365 (2000).DOI
  14. [14]C.-F. Chen, M. Kastoryano, F. G. S. L. Brandão, and A. Gilyén, Efficient quantum thermal simulation, Nature 646, 561–566 (2025).DOI
  15. [15]C. Rouzé, D. Stilck França, and A. M. Alhambra, Optimal quantum algorithm for Gibbs state preparation, Phys. Rev. Lett. 136, 060601 (2026).DOI
  16. [16]C. Rouzé, D. Stilck França, and A. M. Alhambra, Efficient thermalization and universal quantum computing with quantum Gibbs samplers, Nature Physics 22, 1123 (2026).
  17. [17]S. Goldstein, D. A. Huse, J. L. Lebowitz, and R. Tumulka, Thermal equilibrium of a macroscopic quantum system in a pure state, Phys. Rev. Lett. 115, 100402 (2015).DOI
  18. [18]D. Poulin and M. B. Hastings, Markov entropy decomposition: A variational dual for quantum belief propagation, Phys. Rev. Lett. 106, 080403 (2011).pubmed.ncbi.nlm.nih.gov/21405554
  19. [19]A. J. Ferris and D. Poulin, Algorithms for the markov entropy decomposition, Phys. Rev. B 87, 205126 (2013).DOI
  20. [20]D. Petz, Sufficient subalgebras and the relative entropy of states of a von neumann algebra, Communications in Mathematical Physics 105, 123–131 (1986).DOI
  21. [21]D. Petz, Sufficiency of channels over von neumann algebras, The Quarterly Journal of Mathematics 39, 97–108 (1988).DOI
  22. [22]W. Brown and D. Poulin, Quantum markov networks and commuting hamiltonians, arXiv:1206.0755 (2012).arxiv.org/abs/1206.0755
  23. [23]J. M. Hammersley and P. Clifford, Markov fields on finite graphs and lattices (1971), unpublished manuscript.
  24. [24]K. Kato and F. G. S. L. Brandão, Quantum approximate markov chains are thermal, Communications in Mathe- matical Physics 370, 117 (2019).DOI
  25. [25]T. Kuwahara, Clustering of conditional mutual informa- tion and quantum markov structure at arbitrary temper- atures, Physical Review X 15, 041010 (2025).
  26. [26]K. Kato and T. Kuwahara, On the clustering of con- ditional mutual information via dissipative dynamics, arXiv:2504.02235 (2025).
  27. [27]C.-F. Chen and C. Rouzé, Quantum gibbs states are lo- cally markovian, arXiv:2504.02208 (2025).arxiv.org/abs/2504.02208
  28. [28]T. H. Yang, Improved estimate of local markovianity for quantum gibbs states, arXiv:2609.38007 (2025).arxiv.org/abs/2609.38007
  29. [29]E. H. Lieb and M. B. Ruskai, Proof of the strong subaddi- tivity of quantum-mechanical entropy, Journal of Math- ematical Physics 14, 1938–1941 (1973).
  30. [30]Y.-K. Liu, Proc. RANDOM (Springer-Verlag, Berlin, 2006) p. 438.
  31. [31]C. Schilling, The quantum marginal problem, in Math- ematical Results in Quantum Mechanics (WORLD SCI- ENTIFIC, 2014) p. 165–176.arxiv.org/abs/1404.1085
  32. [32]A. Winter, Tight uniform continuity bounds for quantum entropies: Conditional entropy, relative entropy distance and energy constraints, Communications in Mathemati- cal Physics 347, 291–313 (2016).DOI
  33. [33]M. Doroudiani and V. Karimipour, Planar maximally en- tangled states, Phys. Rev. A 102, 012427 (2020).DOI
  34. [34]Y.-L. Wang, Planar k-uniform states: a generalization of planar maximally entangled states, Quantum Informa- tion Processing 20, 271 (2021).arxiv.org/abs/2106.12209
  35. [35]Y. Chiba and Y. Yoneta, Exact thermal eigenstates of nonintegrable spin chains at infinite temperature, Phys. Rev. Lett. 133, 170404 (2024).DOI
  36. [36]S. Mohapatra, S. Moudgalya, and A. C. Balram, Exact volume-law entangled zero-energy eigenstates in a large class of spin models, Phys. Rev. Lett. 134, 210403 (2025).DOI
  37. [37]A. Hokkyo, Exact thermal stabilizer eigenstates at infi- nite temperature, arXiv:2601.16177 (2026).arxiv.org/abs/2601.16177
  38. [38]W. L. McMillan, Ground state of liquid he⁴, Phys. Rev. 138, A442 (1965).
  39. [39]G. Carleo and M. Troyer, Solving the quantum many- body problem with artificial neural networks, Science 355, 602–606 (2017).DOI
  40. [40]S. R. White, Density matrix formulation for quantum renormalization groups, Phys. Rev. Lett. 69, 2863 (1992).pubmed.ncbi.nlm.nih.gov/10046608
  41. [41]S. R. White, Density-matrix algorithms for quantum renormalization groups, Phys. Rev. B 48, 10345 (1993).pubmed.ncbi.nlm.nih.gov/10007313
  42. [42]A. Kitaev, A. Shen, and M. Vyalyi, Classical and Quan- tum Computation, Graduate studies in mathematics (American Mathematical Society, 2002).
  43. [43]D. Aharonov, D. Gottesman, S. Irani, and J. Kempe, The power of quantum systems on a line, Communications in Mathematical Physics 287, 41 (2009).
  44. [44]S. Hallgren, D. Nagaj, and S. Narayanaswami, The local hamiltonian problem on a line with eight states is qma- complete, Quantum Info. Comput. 13, 721–750 (2013).arxiv.org/abs/1312.1469
  45. [45]M. Fannes, B. Nachtergaele, and R. F. Werner, Finitely correlated states on quantum spin chains, Commun. Math. Phys. 144, 443 (1992).DOI
  46. [46]A. Iwaki, A. Shimizu, and C. Hotta, Thermal pure quan- tum matrix product states recovering a volume law en- tanglement, Phys. Rev. Res. 3, L020215 (2021).DOI
  47. [47]See Supplementary Information for details.
  48. [48]M. Fishman, S. R. White, and E. M. Stoudenmire, The ITensor Software Library for Tensor Network Calcula- tions, SciPost Phys. Codebases , 4 (2022).DOI
  49. [49]M. Fishman, S. R. White, and E. M. Stoudenmire, Code- base release 0.3 for ITensor, SciPost Phys. Codebases , 4 (2022).
  50. [50]P. K. Mogensen and A. N. Riseth, Optim: A mathe- matical optimization package for julia, Journal of Open Source Software 3, 615 (2018).DOI
  51. [51]A. E. Feiguin and S. R. White, Finite-temperature den- sity matrix renormalization using an enlarged hilbert space, Phys. Rev. B 72, 220401 (2005).DOI
  52. [52]M. Zwolak and G. Vidal, Mixed-state dynamics in one- dimensional quantum lattice systems: A time-dependent superoperator renormalization algorithm, Physical Re- view Letters 93, 207205 (2004).
  53. [53]F. Verstraete, J. J. García-Ripoll, and J. I. Cirac, Ma- trix product density operators: Simulation of finite- temperature and dissipative systems, Phys. Rev. Lett. 93, 207204 (2004).pubmed.ncbi.nlm.nih.gov/15600964
  54. [54]D. N. Page, Average entropy of a subsystem, Phys. Rev. Lett. 71, 1291 (1993).DOI
  55. [55]Y. O. Nakagawa, M. Watanabe, H. Fujita, and S. Sug- iura, Universality in volume-law entanglement of scram- bled pure quantum states, Nature Communications 9, 1635 (2018).DOI
  56. [56]A. Iwaki and C. Hotta, Purity of thermal mixed quantum states, Phys. Rev. B 106, 094409 (2022).DOI
  57. [57]A. Iwaki and C. Hotta, Sample complexity of matrix product states at finite temperature, Phys. Rev. B 109, 224410 (2024).DOI
  58. [58]K. Landsman, Typical = random, Axioms 12, 727 (2023).DOI
  59. [59]S. O. Scalet, A. Capel, A. N. Chowdhury, H. Fawzi, O. Fawzi, I. H. Kim, and A. N. Tikku, Classical estima- tion of the free energy and quantum gibbs sampling from the markov entropy decomposition, arXiv:2504.17405 (2025).
  60. [60]Y. Yoneta, Thermal pure states for systems with antiu- nitary symmetries and their tensor network representa- tions, Phys. Rev. Res. 6, L042062 (2024).DOI
  61. [61]D. Cocchiarella and M. C. Bańuls, Low-temperature gibbs states with tensor networks, Physical Review Re- search 7, 033184 (2025).
  62. [62]Y.-K. Liu, M. Christandl, and F. Verstraete, Quantum computational complexity of the n-representability prob- lem: Qma complete, Phys. Rev. Lett. 98, 110503 (2007).pubmed.ncbi.nlm.nih.gov/17501036
  63. [63]X.-D. Yu, T. Simnacher, N. Wyderka, H. C. Nguyen, and O. Gühne, A complete hierarchy for the pure state marginal problem in quantum mechanics, Nature Com- munications 12, 1012 (2021).DOI
  64. [64]I. H. Kim, Entropy scaling law and the quantum marginal problem, Physical Review X 11, 021039 (2021).
  65. [65]Z.-Y. Wei, D. Malz, and J. I. Cirac, Efficient adiabatic preparation of tensor network states, Physical Review Research 5, L022037 (2023).arxiv.org/abs/2209.01230
  66. [66]D. Malz, G. Stal, Z.-Y. Wei, and J. I. Cirac, Prepa- ration of matrix product states with log-depth quantum circuits, Physical Review Letters 132, 040404 (2024).DOI
  67. [67]M. Scheer, A. Baiardi, E. B. Marty, Z.-Y. Wei, and D. Malz, Renormalization-group-based prepara- tion of matrix product states on up to 80 qubits, arXiv:2510.24681 (2025).arxiv.org/abs/2510.24681
  68. [68]K. Murota, F. Sauvage, M. Ballarin, G. Matos, and E. Ri- naldi, Exact log-depth preparation of highly entangled matrix product states, arXiv:2606.24475 (2026).arxiv.org/abs/2606.24475
  69. [69]K. C. Smith, A. Khan, B. K. Clark, S. Girvin, and T.-C. Wei, Constant-depth preparation of matrix product states with adaptive quantum circuits, PRX Quantum 5, 030344 (2024).DOI
  70. [70]Y. Zhang, S. Gopalakrishnan, and G. Styliaris, Characterizing matrix-product states and projected entangled-pair states preparable via measurement and feedback, PRX Quantum 5, 040304 (2024).DOI
  71. [71]R. Sahay and R. Verresen, Classifying one-dimensional quantum states prepared by a single round of measurements, PRX Quantum 6, 010329 (2025).DOI
  72. [72]D. T. Stephen and O. Hart, Preparing matrix product states via fusion: Constraints and extensions, Physical Review B 112, 235127 (2025).DOI
  73. [73]R. Sahay and R. Verresen, Finite-depth preparation of tensor network states from measurement, arXiv:2404.17087 (2024).arxiv.org/abs/2404.17087
  74. [74]Y. Li, A. Mitra, and T.-C. Lu, State preparation via measurement and feedback: pushing relations, state structures, and non-invertible symmetries, arXiv:2608.08821 (2026).arxiv.org/abs/2608.08821
  75. [75]A. Iwaki, Markov-shielded typicality for 1D spin chains, GitHub repository, https://github.com/AtsushiIwaki/markov-shielded-typicality (2026).

Paper details

vlnc:W_bpmsy2s5
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 4

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents