Quantum Many-body Scars through the Lens of Correlation Matrix

Introduction. With rapid advances in preparing and manipulating quantum states across various artificial quantum platforms, there has been growing interest in understanding how a generic isolated quantum system thermalizes after a quench [1, 2, 3, 4, 5, 6]. The universal thermalization imposes stringent constraints on the eigenstates of generic interacting quantum systems, and a widely accepted framework is the eigenstate thermalization hypothesis (ETH), which posits that locally an eigenstate encodes the information of a thermal ensemble [7, 8, 9, 10, 11]. Except for systems that are integral [12, 13, 14] or with emergent integrability such as the many-body localized systems [15, 16, 5], the validity of ETH has been confirmed in numerous numerical studies [17, 3]. Moreover, it was believed that the ETH applies in the strong sense that it holds for every eigenstate with finite energy density [18].

However, the ETH has recently been challenged by persistent oscillations of local observables after a quench in a Rydberg atom chain [19]. Shortly after, it is shown that the ETH in the effective PXP model is valid only in the weak sense: it breaks down for a measure-zero subset of eigenstates [20, 21]. These special outliers, drawing analogy to single-particle phenomena [22], were dubbed quantum many-body scars (QMBS), or simply scar states. This discovery sparked intense ongoing research, significantly expanding our understanding of QMBS [23, 24, 25]. For example, Shiraishi and Mori have constructed a framework to embed a target space spanned by scar states in a nonintegrable model [26]. Families of Hamiltonians hosting QMBS with closed analytic expressions and equidistant energy separation have also been constructed from symmetry and quasisymmetry perspectives [27, 28, 29, 30, 31, 32]. In parallel, in the spirit of $\eta$-paring states in the Hubbard model [33, 34], the Mark–Lin–Motrunich (MLM) framework [35] and the equivalent spectrum generating algebra (SGA) formalism [36] have provided unified explanations of scar states discovered across different models [37, 38, 39, 40, 41, 42, 43].

However, due to the varying principles underlying those constructions, a unified framework for QMBS is still lacking [24]. As a result, the origin of QMBS in specific models is often addressed on a case-by-case basis. Moreover, theoretical analysis can often become quite involved, making it difficult to identify potential connections between different frameworks. For example, it is unclear whether the second family of QMBS in the spin-1 XY model can be captured by existing formalisms [41, 44]. Additionally, controlling the scar subspace is often challenging. On the numerical side, the entanglement entropy is commonly used to identify QMBS due to their characteristic low-entanglement nature [45]. However, because of the limited system sizes accessible via exact diagonalization, finite-size effects can make it difficult to reliably distinguish QMBS using this approach.

In this Letter, we show that these problems can be well addressed using the correlation matrix spectrum as a tool for characterizing and diagnosing QMBS. For exact QMBS, including those not captured by previous formalisms, we find the dimensionality of the kernel of the correlation matrix—being an integer and immune to finite-size effects— differs from that of thermal eigenstates. This number thus provides a sharp qualitative probe of QMBS, superior to the quantitative probe of entanglement entropy. For approximate QMBS, such as those in the PXP model, the distinct feature of numerous approximate zero eigenvalues in the correlation matrix spectrum also allow us to single out QMBS from thermal eigenstates.

Correlation Matrix. Our work is primarily motivated by the information perspective of the ETH. According to this hypothesis, the diagonal matrix element of a local observable, $\hat{O}$, with respect to a thermal eigenstate $\ket{n}$ approaches the thermal ensemble average as the system size increases [8, 9, 3]

where $Z=\mathrm{Tr}\,e^{-\beta\hat{H}}$ is the partition function and $\beta$ represents the inverse temperature (with $k_{B}=1$). This implies that a thermal eigenstate encodes sufficient information to infer the underlying Hamiltonian, consistent with the fact that a wavefunction has exponentially many components, while a local Hamiltonian has only extensively many parameters. In other words, one might not able to reconstruct the Hamiltonian from a low-entangled QMBS. These ideas are solidified by using the correlation matrix, originally introduced in [46], as an operational tool for reconstructing the underlying local Hamiltonian from a single eigenstate. Let $\ket{\psi}$ be a wave function, and let $\mathcal{V}$ denote a real vector space of Hermitian operators spanned a set of operator basis $\{\hat{L}_{i}\}$ that acts on $\ket{\psi}$. For an arbitrary Hermitian operator $\hat{O}=\sum_{i}w_{i}\hat{L}_{i}\in\mathcal{V}$, we have the the following inequality

where $\hat{O}\ket{\psi}=\xi\ket{\psi}+\ket{\psi_{\perp}}$, and $\ket{\psi_{\perp}}$ is orthogonal to $\ket{\psi}$, i.e., $\langle\psi|\psi_{\perp}\rangle=0$. This inequality can be casted into a quadratic form

where $\mathbf{w}\equiv(w_{1},w_{2},\cdots)^{T}$ is a vector of coefficients, and $\mathbf{M}$ is the real symmetric correlation matrix defined as

The correlation matrix $\mathbf{M}$ is positive semidefinite, and equality in the quadratic form is attained only when $\mathbf{w}$ belongs to its null space, $\mathbf{w}\in\mathrm{Ker}(\mathbf{M})$. In this case, the corresponding $\hat{O}$ is referred to as an eigenoperator of $\ket{\psi}$, meaning $\hat{O}\ket{\psi}=\xi\ket{\psi}$. Let $N_{0}\equiv\mathrm{dim}(\mathrm{Ker}(\mathbf{M}))$ denote the dimensionality of the null space of $\mathbf{M}$. We can then construct $N_{0}$ linear-independent eigenoperators of $\ket{\psi}$. If the original Hamiltonian belongs to the chosen operator space $\mathcal{V}$, then the uniqueness of Hamiltonian reconstruction from thermal eigenstates is equivalent to $N_{0}=1$ [46]. We shall see below exact QMBS are characterized by a often much larger $N_{0}$.

Exact QMBS. A paradigmatic model that hosts exact QMBS is the one-dimensional Affleck–Kennedy–Lieb–Tasaki (AKLT) Hamiltonian

where $\hat{\mathbf{S}}_{i}$ is the spin-1 operator at site $i$, and $\hat{\mathbb{I}}$ is the identity operator [47, 37, 38]. The known series of QMBS $\ket{\mathcal{S}_{2n}}$ have eigenenergies of $E_{n}=2n$ and are captured by the SGA formalism. Their explicit form is given by

where $\ket{G}$ is the ground state of the AKLT model and $\hat{Q}^{+}=\sum_{i=1}^{L}(-1)^{i}(\hat{S}_{i}^{+})^{2}$ with $\hat{S}_{i}^{+}$ being the spin-1 raising operator at site $i$ [36, 35]. Interestingly, this scar tower $\{\ket{\mathcal{S}_{2n}}\}$ is stable against certain disorder perturbations: they persist as eigenstates of the following disordered Hamiltonian $\hat{H}=\hat{H}_{0}+\hat{H}_{\text{dis}}$,

where $n$ indexes lattice sites, $m$ and $m^{\prime}$ are restricted to take $-2,-1,0$ in the summation, $\xi_{m,m^{\prime}}^{n}$ are arbitrary complex numbers, and $\ket{J_{j,m}}$ inside the parenthesis is the total spin $j$, total magnetization $m$ state of two spins at sites $n$ and $n+1$ [35].

To show that the number of correlation matrix zeros $N_{0}$ can distinguish scar states from thermal eigenstates, we start with simple range-2 local basis $\hat{L}_{i}$. Just like Pauli matrices for spin-$1/2$ systems, we can use the standard generators of the SU(3) group, the eight Gell-Mann matrices $\lambda_{a}(a=1,2,\cdots,8)$, together with the identity matrix $\lambda_{0}$ to define the operator bases $\hat{L}_{i}$,

where $n=1,2,\cdots,L$ labels the lattice sites ($L+1$ is identified as $1$) and $\hat{L}_{i}$ acts nontrivially only on sites $n$ and $n+1$. Here $a=0,1,\cdots,8$ and $b=1,2,\cdots,8$ so that we include all single-site and two-site operators but exclude the trivial identity operator, which makes the correlation matrix $M$ of dimension $72L\times 72L$. Since we can linearly combine $\hat{L}_{i}$ to construct the Hamiltonian (7), we have $N_{0}\geq 1$ for every eigenstate. As illustrated in Fig. 1, eigenstates with energies $E_{n}=0,2,\cdots,L$ indeed exist and are numerically check to correspond to SGA scar states $\ket{\mathcal{S}_{2n}}$. Additionally, other scar states with anomalously low bipartite entanglement entropies are observed. On the side of correlation matrix spectrum, all thermal eigenstates unanimously have $N_{0}=1$, meaning the original Hamiltonian can be reconstructed from a single thermal eigenstate [46]. In contrast, all the scar states exhibit larger values of $N_{0}$, making the correlation matrix zeros a reliable qualitative probe for scar states. This stands in contrast to the quantitative probe of entanglement entropy, which is prone to finite-size effects.

Furthermore, analytical results otherwise difficult to derive can be obtained. Further analysis shows that, at least for $L=6,8,\cdots,16$, the minimal number of $N_{0}$ of SGA scar states satisfies $N_{0}^{(\text{min})}=9L+3$. This number has to be compared with the known $9L+2$ Hermitian eigenoperators of the SGA scar state: the Hamiltonian $\hat{H}_{0}$; the total magnetization $\hat{S}^{z}=\sum_{i}\hat{S}_{i}^{z}$; $9L$ Hermitian disorder terms in $\hat{H}_{\text{dis}}$. This means there is an additional linearly independent range-2 operator, denoted as $\hat{A}$, that is the eigenoperator of all SGA scar states. Thus, the most general range-2 Hamiltonian hosting the SGA scar states is $\hat{H}+\xi\hat{A}$ where $\xi$ is an arbitrary real number. The explicit form of $\hat{A}$ is given by

where $\hat{A}_{n}$ in matrix representation reads

with $c_{n}=(-11/3,-11/3,-5,1,1,53/3,53/3,7)$ accurate to machine precision. We have also tested that for even system sizes ranging from $L=6$ to $L=16$, the constructed operator $\hat{A}$ annihilates all the SGA scars in the AKLT model. Although the prime numbers appearing in $c_{n}$ suggest the difficulty of deriving this result analytically, it can be easily obtained by analyzing the null space of the correlation matrix numerically. For the additional scar states, we have successfully obtained their analytical expressions as well ¹¹1The theoretical analysis is quite involved, and the detailed study in left in Appendix A.. In short, these states are highly excited states of the AKLT Hamiltonian $\hat{H}_{0}$ that are preserved under inclusion of the disorder term $\hat{H}_{\text{dis}}$.

Scar Embedding and Compatibility. In addition to providing a qualitative diagnosis of scar states, the correlation matrix framework enables flexible manipulation of the scar state subspace and offers insight into its internal structure. Let $\mathcal{A}_{\ket{\psi}}$ denote the eigenoperator space of $\ket{\psi}$, which is the operator space spanned by $\{\hat{A}=\sum_{i}w_{i}\hat{L}_{i}\}$, where $\mathbf{w}$ is the null eigenvector of the correlation matrix $\mathbf{M}$ constructed from $\ket{\psi}$. Given two eigenstates $\ket{\psi}$ and $\ket{\phi}$ of a Hamiltonian $\hat{H}$, if $\mathcal{A}_{\ket{\psi}}\subset\mathcal{A}_{\ket{\phi}}$, then we can always select an operator $\hat{A}\notin\mathcal{A}_{\ket{\phi}}$ from $\mathcal{A}_{\ket{\phi}}$ and add it to the original Hamiltonian to construct a new Hamiltonian,

of which $\ket{\phi}$ but not $\ket{\psi}$ remains an eigenstate. To illustrate this, we use the disordered AKLT model (7) as an example. Scar states in this model include the ferromagnetic state $\ket{F}=\ket{11\cdots 11}$ where $m_{i}\equiv 1$, the single-magnon state,

where $n=0,1,\cdots,L-1$ in the summation and $k=2\pi m/L$ ($m=0,1,\cdots,L-1$) is the quasimomentum, and several other families such as the SGA tower ²²2Refer to Appendix A for details.. A detailed analysis of the eigenoperator spaces reveals that,

and that $\mathcal{A}_{\ket{1_{k}}}$ is not a subset of the eigenoperator space of any other scar. Moreover, although the dimension of $\mathcal{A}_{\ket{1_{k}}}$ is the same for all $k$, their individual eigenoperator spaces are different. This means that we can construct a Hamiltonian to retain any given $\ket{1_{k}}$ and $\ket{F}$ as the only scar states. We can also add terms to make $\ket{F}$ the unique scar state. However, for Hamiltonian with with on-site and nearest-neighbor interactions, if $\ket{1_{k}}$ is a scar state, $\ket{F}$ is necessarily an eigenstate. As shown in Fig. 2, we can construct a Hamiltonian $\hat{H}^{\prime}=\hat{H}+\sum_{i}\xi_{i}\hat{A}_{i}$ where $\hat{A}_{i}\in\mathcal{A}_{\ket{1_{\pi}}}$ and $\xi_{i}$ are random numbers to embed scar space $\mathcal{S}=\operatorname{span}\{\ket{1_{\pi}},\ket{F}\}$ spanned by $\ket{1_{\pi}}$ and $\ket{F}$. Similarly, since the eigenoperator space $\mathcal{A}_{\ket{G}}$ is not a subset of any other scar, we can construct a Hamiltonian with $\ket{G}$ as the only scar state ³³3Refer to Appendix B for an in-depth analysis of the scar eigenoperator spaces and scar embedding examples..

Unification and Generality. Our method of using correlation matrix zeros $N_{0}$ to probe exact QMBS unifies many mathematical frameworks. To illustrate the generality of our approach, we consider three prominent examples: the Shiraishi–Mori framework, the non-Abelian Lie algebra construction, and the SGA formalism. In the Shiraishi–Mori formalism [26], the scarring Hamiltonian takes the following form

where $\hat{P}_{i}$ are a set of local projectors that annihilate the scar subspace $\mathcal{S}$, $\hat{h}_{i}$ are some local Hermitian operators, and $\hat{H}^{\prime}$ is some Hamiltonian that satisfies $[\hat{P}_{i},\hat{H}^{\prime}]\mathcal{S}=0$. It follows that for any scar state $\ket{\psi}\in\mathcal{S}$, the projector $\hat{P}_{i}$ is an eigenoperator of $\ket{\psi}$ and each $\hat{P}_{i}$ contributes a zero eigenvector to the correlation matrix $\mathbf{M}$ for appropriate large operator space $\mathcal{V}$. This means that for any scar state $N_{0}\geq N_{p}$ where $N_{p}$ is the number of projectors $\hat{P}_{i}$, which distinguish it from thermal eigenstates. Similarly, in the “tunnels to towers” construction [32], the scarring Hamiltonian take the following form

where $\hat{H}_{\text{sys}}$ is a Hamiltonian with non-Abelian symmetry to endow the spectrum with a degenerate subspace, $\hat{H}_{\text{SG}}$ consists of a linear combination of generators in the Cartan subalgebra to lift the degeneracy, and a final Hamiltonian $\hat{H}_{\text{A}}$ that annihilates a targeted subspace $\mathcal{S}$ is needed to break symmetries and promote $\mathcal{S}$ to a scar subspace. In this case, $\hat{H}_{\text{A}}$ is also an eigenoperator for the scar states, making it detectable via the correlation matrix approach.

To demonstrate that our method also captures QMBS in the SGA formalism [36], it is convenient to use the equivalent Mark–Lin–Motrunich (MLM) framework [35]

where $\mathcal{S}$ is the scar subspace, $\hat{H}$ is the Hamiltonian, and $\hat{Q}^{+}$ is the ladder operator that repeatedly acts on a reference eigenstate $\ket{\psi_{0}}$ to generates the whole scar tower $\{\ket{\psi_{n}}\equiv(\hat{Q}^{+})^{n}\ket{\psi_{0}}\}$. Therefore, as long as $\hat{H}$ and $\hat{Q}^{+}$ are sum of local operators with maximum range $k$, the left hand, denoted as $\hat{A}$, of (16) is of sum of local operators. Although $\hat{A}$ is in general not Hermitian, this can be easily remedied by choose the operator basis $\{\hat{L}_{i}\}$ to be non-Hermitian and subsequently modify the form of correlation matrix $M$ as

The correlation matrix $M$ is then a positive semidefinite Hermitian matrix and annihilator $\hat{A}$ is also one-to-one correspondence with the null eigenvector of $\mathbf{M}$. In fact, as demonstrated in the disordered AKLT model, even using a Hermitian operator basis with a real symmetric $\mathbf{M}$ is often sufficient in detecting scar states.

Approximate QMBS. For several paradigmatic exact QMBS we have tested ⁴⁴4Refer to Appendix C for extra examples of QMBS in the spin-1 XY model which also includes references [54]. , all scar states distinguish themselves from thermal eigenstates by having a much larger $N_{0}$. However, for approximate scars with neither known analytic expressions nor can be captured by previous mathematical formalisms, the single number $N_{0}$ may be inadequate. A prime example of this is the well-known PXP model, described by the following Hamiltonian [20, 21]

where $\hat{P}_{i}=(1-\hat{Z}_{i})/2$ is the projection operator at site $i$, and $\hat{X}_{i}$ and $\hat{Z}_{i}$ are standard Pauli operators. Except for special eigenstates in the middle of the spectrum, the analytic expressions of QMBS are, in general, unavailable [52]. In our correlation matrix study, we choose the matrix representation of range-$k$ operator basis $\hat{L}_{i}$ acting on lattice sites $n+1,n+2,\cdots,n+k$ as

where $\sigma_{a_{j}}$ denotes identity and three Pauli matrices $(X,Y,Z)$ for $a_{j}=0$ and $a_{j}=1,2,3$, respectively. As before, $a_{1}$ is restricted to $1,2,3$ while other $a_{i}$ take $0,1,2,3$. This ensures that our operator basis excludes trivial identity operators but includes all nontrivial local operators up to range $k$. We employ periodic boundary conditions in our study. As shown in Fig. 3(b), the scar state (colored black) has the same $N_{0}$ as those of thermal eigenstates, which is associated with the Rydberg blockade condition. However, the correlation matrix spectrum reveals marked differences between the scar states and the thermal states. To quantify these differences, we introduce the following measure analogous to the role of entanglement entropy in quantifying quantum entanglement

to characterize the correlation matrix spectrum, where the summation over $\lambda_{i}$ is restricted to nonzero values and $N$ is the total number of nonzero $\lambda_{i}$. As can be seen from Fig. 3(c), the introduced quantity $n_{c}$ effectively distinguishes scar states from thermal eigenstates. Moreover, $n_{c}$ is less sensitive to finite-size effects compared to entanglement entropy, making it a more reliable probe.

Summary. We have introduced a new method to detect quantum many-body scars (QMBS) using the correlation matrix spectrum. For exact QMBS, we demonstrate that the integer $N_{0}$, the number of zero eigenvalues, can qualitatively distinguish scar states from thermal eigenstates. In this regard, $N_{0}$ functions similarly to how symmetry and topological numbers are used to classify phases of matter. Our method not only unifies many previous framework of QMBS, similar to the recent commutant algebra formalism [53] which we learned after the main idea is developed, but also is quite accessible and flexible. It enables the straightforward derivation of otherwise obscure analytic results, manipulation of the scar subspace, and insights into its internal constraints. For approximate QMBS where $N_{0}$ may fall short, one can introduce quantities such as $n_{c}$ to characterize scar states, leveraging the presence of multiple approximate zero eigenvalues in the spectrum. Given the versatility of the correlation matrix approach, we anticipate that its applications will extend beyond the study of QMBS.

Acknowledgments. We thank Hui Zhai, Lei Pan, and Shang Liu for numerous useful discussions. We also thank Shang Liu for carefully reading our manuscript and valuable feedback. We acknowledge support by NSFC Grant No. 12304288 (Z. Y.), No. 12247101 (Z. Y.), and No. 12374477 (P. Z.).