WorksPhysics and Astronomy
Spin-adapted sample-based quantum diagonalization with randomized time evolution
Abstract
Quantum algorithms based on the classical post-processing of measurement samples have emerged as effective and robust methods to approximate low-energy wavefunctions of many-body quantum systems on present-day and early-fault-tolerant quantum devices. This paper introduces a generic framework for enforcing spin symmetry within these approaches. We tailor this framework to SqDRIFT, a sample-based quantum diagonalization variant that uses the qDRIFT randomized compilation protocol to generate the quantum samples. We discuss how three key components of SqDRIFT (the initial-state preparation, the construction of the time-evolution circuits, and the classical post-processing) can be adapted to incorporate spin symmetry. We show that spin-pure circuits alone are not sufficient, since the sampled determinants generally span a subspace that is not closed under , and we restore this closure by completing the sampled electronic configurations. In classical simulations of 4-aminobenzonitrile and of an model complex, and in experiments on an IBM quantum processor for the oxo-Mn(Salen) complex, this spin-adapted post-processing yields roots that are exact spin eigenfunctions at every subspace size. On hardware, it does so with diagonalized subspaces two to three orders of magnitude smaller than those of the conventional constructions for the same circuit depth.
Introduction
Solving the ground-state problem for quantum-chemical Hamiltonians is one of the most promising applications of quantum computers. Despite that, realizing this promise in practice, on present-day quantum computers, has proven challenging. While approaches based on the variational quantum eigensolver can hardly be scaled to systems with more than 20 electrons in 20 orbitals [1], methods based on quantum simulation primitives, such as quantum phase estimation and quantum subspace approaches, yield very deep circuits due to the complexity of implementing the time-evolution operator [2, 3]. Sample-based approaches, such as quantum selected configuration interaction [4, 5, 6] and sample-based quantum diagonalization (SQD) [7, 8, 9, 10, 11], have recently emerged as an intermediate solution for executing large-scale chemical simulations on current quantum hardware. The key idea underlying these approaches is to solve the ground-state problem classically within a subspace that is constructed by sampling from a (collection of) quantum circuit(s). Sample-based methods in which circuits are constructed from time-evolution primitives, such as sample-based Krylov quantum diagonalization (SKQD) [12], are especially appealing: under the assumption that the ground state is concentrated on a small (polynomially scaling) number of determinants, they inherit convergence guarantees similar to those of quantum phase estimation. Since the exact time-evolution operator is too costly to implement on present-day quantum computers, it must be approximated when applying these approaches to chemical Hamiltonians. In SqDRIFT [13], the time-evolution operator is approximated through the qDRIFT randomized protocol [14]. Early applications of SqDRIFT to large-scale quantum-chemical calculations [15, 16] have demonstrated the potential of the method.
In this work, we further extend SqDRIFT to include spin symmetry. From a practical perspective, accounting for spin symmetry is a critical requirement of any method aimed at simulating strongly correlated systems. In fact, many strongly correlated systems feature non-trivial spin energetics [17, 18, 19], which can be resolved only using electronic-structure methods properly accounting for this symmetry. While spin-adapted variants of many state-of-the-art classical approaches for simulating strongly correlated systems are nowadays available, including tensor networks [20, 21, 22] and selected configuration interaction (SCI) [23, 24, 25, 26], this is not yet the case for most sample-based quantum approaches, and in particular for those based on time evolution. This work fills this gap and discusses how the three key components underlying SqDRIFT, i.e., the initial-state preparation [27, 28, 29], the qDRIFT randomization protocol [14], and the classical post-processing can be modified to account for the Hamiltonian spin symmetry. Specifically, we discuss how to express the second-quantized form of the electronic-structure Hamiltonian in a spin-invariant way, such that each individual term that is sampled in qDRIFT commutes with . Moreover, we discuss how to modify the classical post-processing originally proposed in Ref. [7] to ensure that the diagonalization subspace is closed under , so that the wavefunctions obtained through subspace diagonalization are eigenfunctions of . Our proposed procedure follows the approaches presented in Refs. [5, 25], which we combine with configuration recovery [7]. We validate the approach through classical simulations of the excited states of 4-aminobenzonitrile (ABN) and of the quintet–triplet splitting of the [Fe(NH)] model complex (which we will refer to as FeN). Our extension enables applying the pipeline to data obtained from quantum hardware, as we show with experiments on an IBM Heron processor targeting the spin energetics of the oxo-Mn(Salen) complex, a transition-metal complex that has been extensively studied with tensor-network approaches [30, 31, 32]. The FeN simulations show that spin-adapted circuits alone do not yield spin-pure eigenstates of the subspace Hamiltonian, since the sampled determinants span a subspace that is not closed under , whereas spin completion does at every subspace size. On hardware, where only spin-adapted post-processing is applied, spin completion yields exact spin eigenfunctions at comparable accuracy, while requiring diagonalization of subspaces two to three orders of magnitude smaller than those required by previous construction strategies at the same excitation level. Our work lays the groundwork for constructing an SqDRIFT-based pipeline that can target strongly correlated systems with non-trivial spin energetics [17, 33].
Theory
We focus on the non-relativistic electronic-structure Hamiltonian formulated in second quantization within a finite orbital basis, given by:
where and are one- and two-electron integrals, respectively, and () is the creation (annihilation) operator of the fermionic mode with spin . In the following, we first describe how to modify the SqDRIFT pipeline to ensure that the circuits prepare spin-pure wave functions. Then, we describe how to generalize the classical post-processing of quantum samples to enforce spin completeness of the sampled subspaces.
We first define determinants and electronic configurations, since these terms are often used interchangeably in the literature, whereas here they have distinct meanings. Given an ordered orbital basis set, we refer to electronic configuration (or, simply, configuration) as a vector of integers whose -th entry is 0, 1, or 2 if orbital is empty, singly occupied, or doubly occupied, respectively. Note that, in an electronic configuration, the information about singly occupied orbitals is not spin-resolved. Conversely, determinants here refers to Slater determinants
where is the vacuum state and . The determinant is thus identified by a bitstring
of length , where are the occupation strings of the and spin orbitals, respectively, and we write . Slater determinants form an orthonormal basis of the antisymmetric (fermionic) sector of the many-particle Hilbert space. Hence, any molecular wavefunction can be expressed as
In a Slater determinant, every electron has a well-defined spin. Hence, Slater determinants are eigenfunctions of , with eigenvalue , where () is the number of () electrons, but in general not of the total spin . However, the non-relativistic electronic structure Hamiltonian in Eq. (1) fulfills the following commutation relations:
Hence, a common eigenbasis of , , and exists. In particular, the eigenstates of can be chosen to be eigenfunctions of the total spin. While preparing a linear combination of Slater determinants with the same guarantees that the final wavefunction is an eigenfunction, this is not true for the total spin since Slater determinants are, in general, not eigenfunctions of . Fortunately, we can define linear combinations of Slater determinants that possess the same spatial, spin, and permutational symmetry as the physical eigenstates. These basis states are called configuration state functions (CSFs). Any molecular wavefunction can be expressed as a linear combination of CSFs
Several strategies to construct CSFs exist, such as the graphical unitary group approach (GUGA) [34, 35], which is widely adopted in (classical) quantum-chemistry software [36, 37]. The GUGA approach has been combined with different variants of modern CI approaches, including selected CI [26] and full configuration interaction quantum Monte Carlo (FCIQMC) [23, 38]. In the following, we illustrate how CSFs can be used as the initial state for spin-adapted SqDRIFT calculations and how the relationship between CSFs and Slater determinants can be used to efficiently post-process bitstrings sampled from quantum devices.
SKQD and SqDRIFT
Sample-based approaches diagonalize the Hamiltonian given in Eq. (1) in a subspace defined by the samples collected from a (collection of) quantum circuit(s). Among the different sample-based quantum diagonalization (SQD) variants, we focus in this work on SqDRIFT, which can be interpreted as a randomized variant of
SKQD [12]. SKQD constructs the subspace by sampling from quantum Krylov states, defined as:
where the time-evolution operator is compiled using Trotter product formulas. Conversely, in SqDRIFT, the operator is approximated through the qDRIFT randomized compilation protocol [14]. To this end, we decompose the Hamiltonian as
where we intentionally do not specify, for the moment, the form of the operators , as it differs between the spin-adapted SqDRIFT variants discussed below. However, we require, without loss of generality, that and that each has unit spectral norm, [14]. The exact time-evolution channel is then approximated by
where each randomized evolution sequence
is weighted by the probability
In other words, each of the factors in Eq. (9) is obtained by independently drawing a term of Eq. (7) with probability and evolving under it for the fixed time . Terms with large coefficients are therefore sampled more often, while the evolution time of each factor is the same regardless of the sampled term. The channel in Eq. (8) approximates with an error, in diamond norm, of [14]. Crucially, this bound depends on the Hamiltonian only through and not on the number of terms in Eq. (7). This makes qDRIFT particularly appealing for electronic-structure Hamiltonians, whose number of terms grows as , whereas the depth of Trotter circuits grows with the number of terms.
SqDRIFT uses this randomized construction to approximate the Krylov states of Eq. (6). For each Krylov index , we set , draw random sequences according to Eq. (10), and measure each circuit in the computational basis times. Since the probability of measuring the bitstring from the state is
each sample collected in this way is distributed according to the qDRIFT approximation of the time-evolved state, without the mixed state ever being prepared explicitly. The bitstrings collected over all Krylov indices and random realizations are then post-processed and used to build the subspace in which the Hamiltonian is diagonalized classically (see Appendix § 1.1.4 for the batching and configuration-recovery settings). Note that, as argued in Ref. [13], it suffices that the sampling distribution, accumulated over Krylov indices and random realizations, be approximately uniform over the support of the ground state. Deviations from the exact time-evolution channel only redistribute probability among determinants, which is harmless as long as every determinant in the support retains a non-negligible sampling probability.
Spin-adapted initial states for SqDRIFT
The choice of the initial state and of the corresponding state-preparation primitive is a key step of SqDRIFT: as discussed in Ref. [13], the ability of SqDRIFT to sample the support of the exact ground-state wavefunction efficiently depends on (and increases with) its overlap with the initial state. When targeting a state with total spin , the initial state should be an eigenstate of with eigenvalue , e.g., a CSF. In fact, it has been shown that, for strongly correlated systems, a CSF or a short linear combination of CSFs can have a substantially larger overlap with the exact ground state than a single determinant expressed in the same orbital basis [28, 29], although, for iron–sulfur clusters, optimizing the orbitals of a single determinant has been shown to yield an even larger overlap [29]. Since the convergence guarantees of SqDRIFT become tighter as this overlap increases [13], such a choice is expected to improve the efficiency of SqDRIFT as well. Moreover, choosing a CSF as the initial state in an SqDRIFT calculation ensures that, if the time-evolution circuit that is applied on this initial state preserves the spin (see Section II C), the wavefunction generated by each SqDRIFT circuit is an eigenstate of as well. Quantum circuits that prepare a CSF with a given number of electrons and spin quantum numbers and have been proposed [27, 28, 39], and can be used as the state-preparation step of SqDRIFT without modifying the rest of the workflow. For simplicity, for given values of and , we choose—without loss of generality—as initial state the CSF associated with the largest value . Since also commutes with the spin ladder operators, , the components of a spin multiplet, which are connected by , are exactly degenerate, so that the value of can be chosen freely. If the configuration has exactly singly occupied orbitals, the CSF with corresponds to a single determinant in which all singly occupied orbitals host electrons, so that . Our choice is motivated by the fact that, on current noise-limited quantum processors, reducing as much as possible the depth associated with the state-preparation part enables executing deeper, qDRIFT-randomized time-evolution circuits that yield more accurate SqDRIFT calculations [13]. This choice has two further advantages. First, the sector contains no states with total spin lower than , so that its lowest root cannot acquire lower-spin character. Second, since each factor of the spin-adapted circuits of Section II C commutes with (up to the splitting error discussed there), which do not change orbital occupations, the probability of sampling a given electronic configuration from the time-evolved state does not depend on . Within the sector, this probability is distributed over up to determinants, with the number of singly occupied orbitals of the configuration. Since this binomial coefficient decreases with , for a given number of samples, the sampled distribution is most concentrated and the spin-completed subspace (see Section II D) is smallest for . However, our choice restricts the initial state to configurations with exactly singly occupied orbitals. For target states dominated by configurations with more than singly occupied orbitals, such as open-shell singlets, every CSF of the dominant configuration is a linear combination of several determinants, and even the preparation of a single such CSF requires circuits whose additional depth may be prohibitive on current noise-limited hardware [27, 28, 39, 40, 41]. As hardware noise decreases towards the fault-tolerant era, such CSFs, as well as linear combinations thereof, may become viable initial states.
Spin-adapted circuits for SqDRIFT
Section II B discussed how to initialize quantum circuits using states with the correct spin. Starting from these, we build the SqDRIFT circuits shown in Eqs. (6) and (9). Given the commutation relationships in Eq. (4), the exact time-evolution operator preserves the spin of the initial state . However, product formulas, such as Trotter–Suzuki decompositions, preserve spin only if each individual term of the Hamiltonian decomposition commutes with . This is not the case for the spin-orbital terms of Eq. (1). The same holds for qDRIFT, since each randomized realization is a product of exponentials of individual terms we refer to as excitations. To avoid breaking the spin symmetry in our circuit, we must rewrite the Hamiltonian in such a way that each sampled term preserves spin by itself. To do so, we take advantage of the global SU(2) spin invariance of the non-relativistic Hamiltonian in Eq. (1). We introduce the spin-adapted excitation operators
which are the spin-free one-body operators and generate the subgroup U() of spin-independent orbital rotations. These operators are scalars with respect to spin rotations (singlets) and therefore commute with all total-spin operators (, , ).
Consequently, any polynomial in the , and the exponential of any Hermitian combination of them, also commutes with . Expressed in terms of the , the Hamiltonian reads
where collects the one-body contribution arising from the reordering of the two-body term.
The individual terms in Eq. (13) are not Hermitian, since , and their exponentials are therefore not unitary. Since is Hermitian and, for real orbitals, all integrals are real, we can write , which yields
with
Each and is Hermitian and commutes with . Conjugate pairs that coincide because of the permutational symmetry of the integrals, namely and (with ), are merged into a single term.
Finally, we take these Hermitian terms as the decomposition in Eq. (7). For a term with coefficient and operator , we set and , which satisfies the normalization requirements of Section II A. The component of proportional to the identity, which only contributes a global phase, is dropped before normalizing. Since every commutes with , each randomized realization in Eq. (9) would preserve the total spin of if each factor were implemented exactly. Under the Jordan–Wigner mapping, however, the Pauli strings of a two-body term do not, in general, all commute. We therefore partition them into mutually commuting groups, , and implement each factor as , in which every group exponential is exact. Each group conserves and , so particle number and are preserved exactly, whereas the individual groups do not commute with , so the total spin is preserved only up to a splitting error of order per factor. One-body terms and two-body terms whose Pauli strings all commute () are implemented exactly. We refer to the circuits built with spin-adapted excitation operators as spin-adapted.
Spin-adapted post-processing of bitstrings
When the initial state is prepared in the target spin sector and the qDRIFT terms are sampled from the Hermitized spin-adapted Hamiltonian of Eq. (14), the circuits prepare states that are spin pure up to the splitting error of Section II C, in the absence of hardware noise. However, the diagonalization subspace is built from measurement samples of these circuits. In the SqDRIFT setting, samples from the quantum device correspond to Slater determinants, which are eigenstates of but, in general, not of . More importantly, the sampled subspace is not guaranteed to be closed under . We say that a subspace is closed under if for all .
In previous sample-based quantum algorithms, a strategy to mitigate the spin contamination of the final wavefunction has been proposed [7]. Let us denote by the set of unique bitstrings sampled from the quantum device, and by the subspace spanned by the corresponding Slater determinants. Moreover, we denote by and the sets of unique - and -strings, respectively. From these sets we define the subspace
namely the subspace spanned by all determinants obtained by combining any sampled -string with any sampled -string. We will refer to this as the symmetrized cartesian subspace produced by the collected samples. When the numbers of and electrons coincide, we can symmetrize the bitstrings found in the two sectors by defining
and perform the diagonalization in the subspace determined by every ordered pair that can be obtained from , namely
extends the original subspace and mitigates spin contamination, while additionally enforces invariance under the global exchange of and strings, further reducing it. is the symmetrized cartesian subspace spanned by the collected samples. However, neither subspace is closed under . This can be shown by looking at the action of the spin operator on the Slater determinants. In atomic units () and in terms of , the total-spin operator reads
where is the only term that does not have a diagonal action on the Slater determinants. In second quantization,
which moves an electron from orbital to orbital while moving a electron from to . Given a Slater determinant , at fixed we have
where and are the -th and -th canonical unit vectors. This relation holds up to a fermionic sign and only if and ; otherwise, the result vanishes. Having does not imply , showing that is not closed under . Of course, as the size of the sampled subspace grows it becomes increasingly likely that the strings connected by these spin-exchange moves are also included in , eventually yielding a spin-closed subspace. However, in order to reduce the (classical) computational resources required by sample-based quantum algorithms, this property is desirable at every subspace size, and for both open- and closed-shell systems.
For this reason, we adopt a different procedure to complete the set of sampled determinants such that the resulting subspace is closed under . First, we note that CSFs have exactly the property that we need: every CSF is an eigenstate, and therefore any subspace spanned by CSFs is closed under . Moreover, CSFs are simultaneous eigenstates of and . This implies that the Slater determinants forming a given CSF all share the same eigenvalue of and, since spin operators do not change orbital occupations, belong to the same electronic configuration. In general, this does not identify a specific CSF, since the same set of Slater determinants enters several CSFs with different coefficients [42, 43]. However, we are only interested in the determinant subspace spanned by the CSFs, not in the individual CSFs, as the diagonalization is performed in the determinant basis. This means that we only have to identify all the other Slater determinants sharing the same configuration and , i.e., all ways of distributing the open-shell electrons among the singly occupied orbitals at fixed , while doubly occupied and empty orbitals remain fixed. The singly occupied orbitals host electrons, of which are and are . Therefore, the number of such determinants equals the number of ways of distributing the open-shell electrons among the singly occupied orbitals, i.e.
which remains small for moderate numbers of open shells (e.g., 924 for and ). With this, we end up with a set of spin-completed bitstrings spanning a subspace that is closed under , in which the classical diagonalization is performed. The resulting procedure, which we refer to as spin completion, is summarized in Algorithm 1.
More information about how we performed the diagonalizations in the spin-adapted subspace and how we post-processed the samples and the wavefunctions can be found in Appendices A and B. In practice, spin completion is applied to each sub-sampled batch, after configuration recovery, and the subspace is truncated by discarding whole configurations, which preserves closure under . In the next section, we present results on different chemical phenomena that we investigated using the spin-adapted workflow both in numerical simulation and on quantum hardware.
Algorithm 1: Spin completion of the sampled bitstrings
Input: Set $\mathcal{X}$ of unique bitstrings $\mathbf{x}=(\mathbf{a},\mathbf{b})$ sampled from the quantum device; $(N_{\alpha},N_{\beta})$, $M_{S}=(N_{\alpha}-N_{\beta})/2$
Output: Spin-completed set $\mathcal{X}_{\mathrm{spin}}$
1 $\mathcal{X}\leftarrow\{\mathbf{x}\in\mathcal{X}:|\mathbf{a}|=N_{\alpha},|\mathbf{b}|=N_{\beta}\}$; ▷ post-select
2 $\mathcal{X}_{\mathrm{spin}}\leftarrow\varnothing,\ C\leftarrow\varnothing$; ▷ done: $C$
3 foreach $\mathbf{x}=(\mathbf{a},\mathbf{b})\in\mathcal{X}$ do
4 $\mathbf{n}\leftarrow\mathbf{a}+\mathbf{b}$; ▷ configuration
5 if $\mathbf{n}\notin C$ then
6 $C\leftarrow C\cup\{\mathbf{n}\}$;
7 $\mathcal{O}\leftarrow\{p:n_{p}=1\},\ n_{\mathrm{single}}\leftarrow|\mathcal{O}|$;
8 $n_{\alpha}^{\mathrm{open}}\leftarrow n_{\mathrm{single}}/2+M_{S}$;
9 foreach $A\subseteq\mathcal{O}$ with $|A|=n_{\alpha}^{\mathrm{open}}$ do
10 $a'_{p}\leftarrow 1$ if $n_{p}=2$ or $p\in A$, else $0$;
11 $b'_{p}\leftarrow 1$ if $n_{p}=2$ or $p\in\mathcal{O}\setminus A$, else $0$;
12 $\mathcal{X}_{\mathrm{spin}}\leftarrow\mathcal{X}_{\mathrm{spin}}\cup\{(\mathbf{a}',\mathbf{b}')\}$;
13 return $\mathcal{X}_{\mathrm{spin}}$;
Results
Accounting for spin symmetry is important in the description of many quantum chemical phenomena, from ground-state to excited-state properties. In the following, we present different use cases, including experiments on real hardware.
Excited-state characterization of 4-aminobenzonitrile (ABN)
The study of the low-lying energy spectrum of molecular compounds is essential to characterize their photochemical properties. In many systems, the interplay between locally excited (LE) and charge-transfer (CT) states is of particular relevance, for instance in organic light-emitting diodes based on thermally activated delayed fluorescence [44, 45] and in organic photovoltaics [46]. An LE state occurs when an electron is promoted from an occupied to an unoccupied molecular orbital within the same localized region, or fragment, of a molecule. These states are usually characterized by a large oscillator strength, leading to bright, fast radiative decay. In CT states, instead, an electron is transferred from an electron-donor fragment to an electron-acceptor fragment, which can be part of the same molecule (intramolecular CT) or belong to two distinct molecules (intermolecular CT). CT states exhibit large changes in the dipole moment, large Stokes shifts (i.e., large energy differences between the absorption and emission maxima), and lower radiative rates than pure LE states. Aminobenzonitriles are a widely studied class of molecules, as they are prototypical systems for the interplay between LE and intramolecular CT states [47, 48, 49]. Their emission properties can be controlled by structural modifications or by external parameters such as solvent polarity and temperature, with important applications in materials design. Here, we focus on 4-aminobenzonitrile (ABN, HN-CH-CN) to characterize its lowest LE and CT states. The molecule consists of a benzene ring substituted with an electron-donating amino group (-NH) and, in the para position, an electron-withdrawing cyano group (-CN), as shown in panel (a) of Figure 1.
Figure 1. Low-lying spectrum of ABN (simulation). (a) Structure of ABN (HN–CH–CN); calculations use the 6-31G basis and an (18e, 14o) active space. (b,c) of the five lowest roots against the number of random circuits, for the raw (b) and spin-completed (c) samples; dashed lines mark the exact singlet and triplet values. (d) Energy error of the LE state against the number of determinants in the subspace, for . Raw SqDRIFT samples, their spin completion, and the Cartesian product are compared against variational HCI. The dashed line marks chemical accuracy (1.6 mHa).
From a Hartree–Fock calculation in the 6-31G basis set, we select an active space of 18 electrons in 14 orbitals, (18e, 14o). We then emulate SqDRIFT classically, sampling the circuits with a tensor-network simulator. For each of the three Krylov times, we draw random realizations with excitations each, and collect samples per circuit. The Hamiltonian is then diagonalized in the subspace spanned by the unique sampled determinants. Since we target singlet excited states, we compute for the five lowest roots in order to identify and exclude triplet (and higher-multiplicity) states. As shown in panel (b) of Figure 1, diagonalizing in the subspace spanned by the raw samples yields spin-contaminated roots when up to 100 random circuits are used, which makes it difficult to classify the excited states by their spin. After the spin completion described in § 2, every root is a spin eigenfunction, here a singlet or a triplet, at every number of circuits (panel (c)).
Although this is not a general criterion for distinguishing LE from CT states, we use the total spin and the change in dipole moment along the main molecular axis to identify the first singlet excited state as the LE state. Panel (d) shows the convergence of its energy towards the FCI reference as a function of the dimension of the diagonalized matrix for four subspace constructions: the raw SqDRIFT samples, their spin completion, the Cartesian product of Eq. (18), and, as a classical selected-CI reference, variational heat-bath configuration interaction (HCI) [50, 51, 52], as implemented in DICE [53]. The raw samples converge slowly and remain above chemical accuracy over the whole range, with an error of about mHa at the largest subspace of about determinants. Spin completion reaches chemical accuracy at about determinants, and an error of about mHa below determinants. The Cartesian product eventually attains the smallest errors, down to about Ha, but only at about determinants, roughly of the FCI space ( determinants), and it reaches chemical accuracy only beyond about determinants. HCI reaches chemical accuracy at about determinants and an error of about 0.5 mHa at its largest subspace of determinants. The large subspaces produced by HCI are mainly due to the fact that we are considering the union of all the subspaces relevant to the excited states up to the target LE one [52]. Among the constructions considered, spin completion therefore provides the most compact subspace at chemical accuracy, about three and seven times smaller than those required by the Cartesian product and HCI, respectively. More details on the calculations on ABN are given in Appendix C.
Spin-state splitting in an Fe(II)N4 model complex
Iron–nitrogen sites are the catalytic motif of Fe-N-C materials for the oxygen reduction reaction [54, 55], whose activity depends on the spin state of the iron centre [56, 57]. For the parent iron(II) porphyrin, the ordering of the triplet and quintet states is a notoriously difficult multi-reference problem, which requires large active spaces and self-consistent orbital optimization to be resolved [19, 31, 58]. The quantity of interest is therefore not a single total energy but a difference between states of different multiplicity, which places a correspondingly sharper demand on the subspace: the two legs of the splitting must be converged consistently, and each must be an eigenstate of for the difference to correspond to a physical spin-state gap at all.
We examine the quintet–triplet splitting of a (in short FeN4) complex using an (8e, 11o) active space, where the FCI splitting is 0.3048 Ha (191 kcal/mol) and the exact spin difference is .
The SqDRIFT circuits start from a single determinant in each sector: the Hartree–Fock determinant for ( orbitals , orbitals ) and, for , the determinant minimizing ( orbitals , orbitals ), which lies 0.109 Ha below the Hartree–Fock determinant in this orbital basis. In both determinants all open-shell electrons have spin, so each is an exact eigenfunction of with , and the circuits start in the target spin sector. The and sectors are sampled independently and paired at equal numbers of sampled bitstrings—the sampler’s own control variable—since the same number of samples expands to different determinant counts in each sector.
Figure 2 shows that only the spin-complete construction yields a gap between genuine spin eigenstates: its spin error is zero to solver precision at every subspace size, and it reaches a gap energy error of Ha at a total dimension of , an order of magnitude below the full CI dimension of the pair. The raw and spin-adapted subspaces carry residual spin errors of –, comparable to their energy errors, so their apparent accuracy on is not accompanied by spin purity. The Cartesian product achieves the lowest absolute error ( Ha) only by diagonalizing determinants, within 10% of the full space.
Figure 2. Spin-state splitting in an FeN4 centre (simulation). Error on the quintet–triplet gap , with the two states computed as the lowest roots of the and sectors and in an active space, against the total number of determinants in the two subspaces. (a) Gap energy error and (b) gap spin error (with ), for the raw sampled determinants (SqDRIFT), their Cartesian product , the spin-completed and target-spin-projected subspace (spin-complete), and spin-adapted sampling. The latter is the raw-determinant construction applied to counts from the spin-adapted circuits, without completion or projection: although the circuits prepare states that are spin pure up to the splitting error of Section II C, the sampled set of determinants is not closed under , which causes the residual error. Marker shape denotes the construction and color the number of excitations (); the dashed line marks the summed full CI dimension of the two sectors (101 640). Spin errors of the spin-complete construction are below , i.e., at solver precision, and are drawn clamped onto the shaded band. spin pure.
Because the error cancellation between the two legs is accidental, a gap can appear more accurate than either sector individually, which explains the non-monotonic curves. Moreover, the cancellation is strongest for the spin measure, where two equally contaminated legs can produce a small difference in while neither state is
Spin energetics of the oxo-Mn(Salen) complex
High-valent manganese–oxo species are the proposed active intermediates of the Jacobsen–Katsuki epoxidation [59, 60, 61], and their strongly correlated Mn 3d/O 2p manifold makes them a standard benchmark for multireference electronic-structure methods. The oxo-Mn(Salen) complex studied here is small enough to admit a converged DMRG reference in a (32e, 24o) active space (see Appendix C), yet retains the dense low-lying spin spectrum that makes spin symmetry difficult to control.
Previous studies indicate that the ground state of oxo-Mn(Salen) is a singlet, closely followed by a triplet, whereas the lowest quintet lies significantly higher in energy [30, 62, 63]. Our DMRG references in the (32e, 24o) active space reproduce this ordering: the triplet lies only 0.48 mHa (0.30 kcal/mol) above the singlet, while the quintet is higher by 149 mHa (94 kcal/mol). Each state is computed as the lowest root of its own sector (, 1, and 2 for the singlet, triplet, and quintet, respectively). The triplet and quintet legs are thus protected from lower-spin contamination by construction, whereas the sector contains both the singlet and the component of the triplet, only 0.48 mHa apart. This near-degeneracy makes the singlet leg a stringent test of spin purity: the residual energy errors of the sampled subspaces discussed below are of the same order as the gap, so in a subspace that is not closed under
nothing prevents the lowest root from acquiring triplet character, and its multiplicity can only be checked a posteriori through .
Figure 3 compares the two SqDRIFT constructions against HCI for each multiplicity. The spin errors (panels d–f) separate the methods unambiguously. The spin-completed subspaces are spin pure to solver precision at every subspace size and for all three states, so their curves sit on the clamped band throughout. The Cartesian product, by contrast, retains a contamination of 4.6– in even at its largest subspaces, decreasing only slowly with ; the Cartesian product is not closed under , and spin purity is recovered only incidentally as the subspace approaches the full space. HCI was performed in two different ways: we first ran DICE with different variational thresholds (HCI line in the figure), then we applied the spin-completing post-processing described in § 2 D and performed a final subspace diagonalization (HCI spin-adapt curve).
Figure 3. Spin energetics of the oxo-Mn(Salen) molecular compound (hardware). Energy error (a–c) and spin error (d–f) against the subspace dimension, for the singlet, triplet, and quintet states, computed in the , , and sectors, respectively, in the (32e, 24o) active space. Marker shape denotes the subspace construction and color the number of excitations (); HCI and its spin-adapted variant (see main text and Appendix C) are shown as references in red. The dashed line marks chemical accuracy (1.6 mHa). Classical and quantum spin-complete errors are omitted from the plot, as their values are substantially lower than those of the other methods (machine precision).
The energy errors (panels a–c) show why this matters. At the spin-complete subspace reaches 0.84 mHa for the singlet—within chemical accuracy—at a diagonalized dimension of , whereas the Cartesian product attains a lower 0.35 mHa only at determinants, a subspace larger by a factor of . The same pattern holds for the triplet (0.39 mHa at versus 0.19 mHa at ) and for the quintet. Therefore, even on hardware and in the presence of noise, spin adaptation buys exact spin symmetry and a reduction in the matrix to be diagonalized, at comparable accuracy (both within chemical accuracy) and at fixed .
HCI remains more accurate in absolute terms, reaching errors of the order of Ha, but at subspace dimensions of – and without exact spin purity. For this system, the residual contamination of the Cartesian-product and HCI roots is small in practice. As a rough estimate, a deviation in corresponds to a weight of () of the triplet (quintet) in the singlet root, and of of the quintet in the triplet root; for , this shifts the energies by at most about 0.03 mHa, well below the singlet–triplet gap. The advantage of spin adaptation is therefore not the repair of a contamination large enough to swap the two states, but a guarantee: in a spin-complete subspace every root is an exact spin eigenfunction, so the multiplicity of the lowest root is determined exactly rather than inferred from . This guarantee comes with a diagonalized dimension two to three orders of magnitude smaller than that of the Cartesian product at the same number of excitations . For more details on the chemical calculations and the quantum experiments, see Appendices C and § 1.1.4.
Conclusions
In this work, we have introduced a general spin-adapted formulation of SqDRIFT. We have discussed how its three key components—the initial-state preparation, the time-evolution circuits, and the classical post-processing—can be generalized to account for spin symmetry. Based on both classical simulations and experiments on quantum hardware, we have demonstrated that the resulting spin-adapted SqDRIFT variant significantly reduces the classical overhead, i.e. the dimension of the diagonalized subspace, while guaranteeing spin-pure roots.
Our hardware experiments employed only the spin-adapted classical post-processing, i.e., spin completion (Section IIID); future work will also execute spin-adapted circuits on quantum hardware. This could require future generations of quantum devices, since preparing generic CSFs and implementing spin-adapted SqDRIFT circuits (for a fixed number of excitations ) substantially increase the circuit depth. As an example, for the FeN model, the two-qubit depth before routing is 8.8 and 4.3 times that of the standard circuits for and , respectively.
Although our framework relies on the Jordan–Wigner fermion-to-qubit mapping, the space spanned by the CSFs could also be encoded directly in the qubit register, using, e.g., the mappings of Refs. [41, 64]. On the one hand, such mappings would make the adaptation of the post-processing unnecessary, since each sample would directly correspond to a CSF. On the other hand, the time-evolution operator would become considerably deeper in this basis, so that such approaches would be practical only on fault-tolerant hardware. Finally, although we have focused on SqDRIFT, the framework readily extends to SKQD [12]. Spin completion also applies directly to SQD variants based on variational ansätze, whereas enforcing spin symmetry at the level of the variational circuit is less straightforward: SQD has mostly been combined with the LUCJ ansatz [65], which does not conserve the total spin. However, spin-adapted ansätze tailored to molecular wave functions [66], including coupled-cluster-based ones [67], have recently been proposed and could be combined with the framework presented here.
AI use statement
We used large language models to assist with writing and debugging parts of the code and with language editing of the manuscript. The authors designed the method and the experiments, checked the code, the results and every cited work, and take full responsibility for the content of the paper.
Author contributions
S.B. conceived the idea, performed classical calculations and experiments on quantum hardware. S.B. and S.P. wrote the code for the spin-adapted workflow. S.P. and A.B. performed classical quantum chemistry calculations, classical simulations and experiments on quantum hardware. A.B., F.T. and I.T. supervised the project. All authors wrote and revised the manuscript.
Acknowledgments
We thank Davide Cugini, Davide Zanin, Laurin Fischer, Javier Robledo-Moreno and Max Rossmannek for fruitful discussions. This research was supported by the project RESQUE (Rethinking Quantum Simulations in the Quantum Utility Era, grant number 20QU-1_225229) and by the National Center of Competence in Research (NCCR) SPIN (grant number 225153), both funded by the Swiss National Science Foundation.
Diagonalizing in the spin-adapted subspace
Section Section II D showed how the sampled determinants can be post-processed so that the diagonalization subspace is closed under . The orthogonal projector onto then commutes with , and so does the projected Hamiltonian, . Its eigenfunctions can therefore be chosen with a definite total spin (and are, barring accidental degeneracies between different spin sectors), so they can be filtered directly by their spin.
When only a specific spin sector is of interest, computational effort spent on other sectors can be avoided. Here, we illustrate an additional step that can be performed by taking advantage of the structure of the spin-adapted subspace and that can be used to guarantee that Davidson–Liu [68, 69] diagonalization is performed in a fixed-total-spin subspace. Diagonalizing yields the orthonormal eigenvectors with the target spin, which form the columns of the isometry , where is the dimension of the spin-completed subspace [42]. Each configuration with open shells contributes columns, so that over the configurations in . Crucially, the spin operator is block diagonal per configuration, as it only rearranges the spins of the open shells and it does not change which orbitals are doubly or singly occupied. At fixed each block depends only on ; its eigenvectors are therefore computed once per distinct open-shell count and reused. In this work, we constructed the matrix of explicitly; for larger systems, schemes that never form this matrix, such as GUGA [34], may become necessary. Once the matrix is obtained, we define the transformed Hamiltonian
and use this to perform the Davidson–Liu diagonalization in a single spin sector. The iterations are started from the determinant used as the initial state of the SqDRIFT circuits, projected onto the target spin sector. We refer to this restriction of the diagonalization to a single total-spin sector as spin projection. The Davidson method belongs to a class of algorithms that approximate the eigenvalues and eigenvectors of a large matrix by projecting it onto a smaller subspace and minimizing the Rayleigh quotient [70]. The Rayleigh–Ritz procedure also provides an estimate of the distance from the exact eigenpair, which is used to expand the search subspace iteratively. In particular, given as the pair of approximated eigenvector and eigenvalue at iteration of the Davidson algorithm, the residual vector is:
While the exact correction that would make vanish cannot be computed, Davidson proposed to approximate the optimal correction vector at every iteration as
where acts as a diagonal preconditioner (the division is element-wise). In the projected case, the same iterations are performed for in . Given the approximate eigenpair , with , the residual reads
so that never needs to be formed explicitly: is mapped back to the determinant basis, acts there, and the result is projected again onto the target spin sector. The correction vector is built from the diagonal of ,
where and run only over the determinants of the configuration to which column belongs, since is block diagonal. This projection is not required to obtain spin-pure wave functions, since the eigenfunctions in already have a definite total spin; it reduces the dimension of the diagonalized matrix and the number of roots required when a given multiplicity is targeted. All spin-completed results in § 2 use it, with the target multiplicity of each state.
Spin-preserving post-processing of the wave function
Noise in the quantum device broadens the distribution of the bitstrings sampled from the prepared circuits. This affects the quantum-selected subspace by introducing determinants with negligible weight in the low-energy states (the so-called deadwood). For both constructions, noise is addressed via the self-consistent configuration recovery of Ref. [7]; for the spin-completed subspaces, bitstrings that still have the wrong after recovery are discarded before spin completion. Deadwood, on the other hand, is removed a posteriori by discarding determinants that contribute negligibly to the low-energy states. This procedure, often called wave-function trimming, has been used in both classical and quantum subspace methods [11, 71] to reduce the computational effort of subsequent processing of the wave function, without compromising the accuracy of the method.
Trimming is performed by ordering the determinants by their CI contribution and removing those whose amplitudes sit below a given threshold. The squared amplitudes are commonly used as a proxy for the contribution of each determinant to the energy. However, this might compromise the spin purity of our wave function. For this reason we perform a spin-preserving trimming, following Ref. [72]. In this approach, the determinants are again grouped by electronic configurations and the contribution of a single configuration is computed as
where is the number of determinants belonging to the same electronic configuration and are the CI coefficients of such determinants, as defined in Eq. (3). Discarding whole configurations according to this weight yields truncated CI expansions that remain eigenfunctions of —provided the untrimmed wavefunction is itself an eigenfunction of —since is block diagonal in the configurations. In Ref. [72], another criterion is discussed by introducing the mean electronic configuration contribution
which avoids favoring configurations whose large total weight stems from many low-weight determinants. In this work, we truncated by configuration weight, which resulted in the most compact wavefunctions at a given accuracy. All hardware experiments on the oxo-Mn(Salen) complex use configuration-weight truncation.
Details on the classical chemical calculations
This Appendix collects the settings of the classical calculations underlying § 2; Figure 4 and Figure 5 provide additional information on the subspace sizes and spin errors discussed there.
Figure 4. Subspace sizes for ABN (simulation). Dimension of the ABN subspace in the sector as a function of the number of random circuits sampled, for : raw SqDRIFT samples, their spin completion, and the Cartesian product. The dashed line marks the FCI dimension (4 008 004).
Figure 5. Spin purity for the FeN quintet (simulation). Spin error of the lowest root of the FeN model () against the number of determinants, for the four constructions of Figure 2. Marker shape denotes the construction and color the number of excitations (; spin-adapted sampling at only). The dashed line marks the FCI dimension of the sector (25 410). Spin errors of the spin-complete construction are below , i.e., at solver precision, and are drawn clamped onto the shaded band.
ABN. The active space was selected from a Hartree–Fock calculation in the 6-31G basis set using the AVAS active space selection method [73] (see Table I for the Cartesian coordinates of the molecule). Let be the number of spin strings with 9 electrons in 14 orbitals. The FCI reference was computed in the full sector of the active space—which contains determinants—and the HCI references with DICE [53] for variational thresholds .
FeN model. The [Fe(N)] complex is a square-planar model with Fe–N distance fixed at the iron porphyrin value (see Table II for the Cartesian coordinates of the molecule). The active space is prepared with AVAS [73] after a Hartree–Fock calculation in the def2-SVP basis set [74]. The (, ) and (, ) sectors contain 76 230 and 25 410 determinants, respectively.
oxo-Mn(Salen). The geometry of the oxo-Mn(Salen) complex was optimized for the singlet state at the CASSCF(10e,10o) level of theory, using the 6-31G basis set [31, 62]. The corresponding Cartesian coordinates are provided in Table III. Following the procedure described in Ref. [31], we selected an active space comprising 32 electrons in 24 orbitals (32e, 24o) and performed DMRG calculations to obtain reference wavefunctions and energies. HCI calculations for the complex were performed with DICE [53] without PT2 correction and with variational thresholds . For the HCI spin-adapt calculations, we collected the samples obtained with DICE at a fixed variational threshold , applied the spin-completion procedure described in Section II D, and performed a final diagonalization in DICE keeping this spin-completed subspace fixed. CSF-based variants of HCI have been proposed [26], but their usage would require a completely new code, which is beyond the scope of this work.
Details on the quantum experiments
ABN. To showcase the capabilities of the spin-adapted workflow in the study of excited states, we performed a classical simulation of SqDRIFT on the ABN model. For each of the three Krylov times, we draw random realizations with excitations each, and collect samples per circuit. Figure 4 shows how the size of the subspaces of Figure 1(d) grows with the number of circuits. As explained in the main text, the dimension of the symmetrized Cartesian product is given by the product of the numbers of distinct and strings (the square of given the symmetrization), each bounded by . The Cartesian product reaches about determinants at 200 circuits and grows slowly afterwards, covering about one third of the FCI space at 500 circuits. Spin completion enlarges the raw subspace by a factor that decreases from about twelve at 10 circuits to about four at 500, since an increasing fraction of the determinants added by completion has already been sampled. At 500 circuits, the spin-completed subspace is still more than an order of magnitude smaller than the Cartesian product.
FeN. SqDRIFT was performed using a classical noiseless simulator on the (8e,11o) active space of the FeN model in order to study the quintet–triplet gap. For each sector and each of the Krylov times , we sampled random circuits with excitations, collecting samples per circuit. Since , the Cartesian product is of Eq. (16), without symmetrization. is evaluated on the eigenvector embedded in the full sector. For the spin-complete construction, the -axis of Figure 2 and Figure 5 counts the determinants of ; the matrix diagonalized after spin projection (Appendix A) is smaller.
Fig. 5 resolves the quintet leg of the gap in Figure 2. The sector contains no states with (), and a deviation corresponds to a mixture of and components with total weight between and . The spin-complete construction is spin pure to solver precision at every size by definition. The raw and spin-adapted-sampling subspaces show spin errors that decrease from about at 50 determinants to below at 3000, with no appreciable difference between the two samplers. This is consistent with Section II D: the residual error stems from the sampled set of determinants not being closed under , not from the circuits. The Cartesian product reaches errors of about only as its dimension approaches that of the full sector. Finally, in this sector spin completion enlarges the sampled subspace by less than a factor of three, compared with the factor of four to twelve found for ABN at . This agrees with the argument of Section II B that completion adds the fewest determinants for .
oxo-Mn(Salen). Experiments on the spin-state energetics of the oxo-Mn(Salen) complex were performed on the IBM Heron r3 quantum processing unit (QPU), ibm_boston. For each spin state, we applied SqDRIFT using Krylov times and randomly generated circuit realizations for each of , and 15 excitations. The quantum circuits were initialized with different states for singlet, triplet and quintet study. As discussed in the main text, for simplicity we considered initial states with with singly occupied orbitals, represented by a single determinant. Given the (32e, 24o) active space we prepared a (16, 16) initial state for the singlet, (17, 15) for the triplet and (18, 14) for the quintet. When preparing the Hamiltonian for the qDRIFT randomization protocol, terms diagonal in the occupation-number basis are excluded from the decomposition, so the circuits approximate the evolution under the off-diagonal part of . Given that only the sampled support enters the subspace, this does not affect the method. We additionally tracked the sampled excitations to identify those that had no effect on the quantum state. Such trivial excitations, which include, for example, excitations of electrons between already occupied orbitals, were discarded and resampled. This filtering affects only the final circuit depth, as the qDRIFT evolution time is adjusted to account for the discarded excitations. In this way, we avoid introducing additional noise associated with unnecessary circuit operations. Finally, each circuit was sampled times, yielding a total of 1.5 million shots per spin state and value of , i.e., 4.5 million per spin state and 13.5 million in total.
| Atom | Atom | ||||||
| N | 3.22465175 | −0.00016435 | −0.00016248 | C | −2.37901020 | 0.00002585 | 0.00007608 |
| N | −3.52808419 | 0.00003935 | 0.00012624 | H | 1.67124777 | −2.13635152 | −0.00008321 |
| C | 1.85428028 | 0.00004661 | −0.00011584 | H | 1.67125967 | 2.13648964 | −0.00011838 |
| C | 1.13987449 | −1.20419643 | −0.00007314 | H | −0.77189917 | −2.13075023 | 0.00002373 |
| C | 1.13990772 | 1.20432135 | −0.00009317 | H | −0.77189568 | 2.13073759 | −0.00001173 |
| C | −0.94910189 | 0.00001219 | 0.00001215 | H | 3.73679291 | 0.84648597 | −0.00032793 |
| C | −0.23864408 | −1.20022502 | −0.00000950 | H | 3.73662970 | −0.84690143 | −0.00031335 |
| C | −0.23860907 | 1.20023043 | −0.00002947 |
Table 1. Optimized Cartesian coordinates (Å) of ABN.
After collecting the hardware samples, we applied both Cartesian and spin-adapted post-processing, including configuration recovery, to the same dataset. For the Cartesian subspace, we generated three batches of samples each by sub-sampling the original dataset and performed five rounds of configuration recovery. Since holds only for the singlet state, we did not construct the symmetrized Cartesian subspace, thereby ensuring consistency across spin states.
For the spin-completed post-processing, we generated three sub-sampled batches of samples each and
| Atom | Atom | ||||||
| Fe | 0.0000 | 0.0000 | 0.0000 | N | 0.0000 | 2.0100 | 0.0000 |
| N | 2.0100 | 0.0000 | 0.0000 | H | 0.0000 | 2.6100 | 0.8000 |
| H | 2.6100 | 0.0000 | 0.8000 | H | 0.0000 | 2.6100 | −0.8000 |
| H | 2.6100 | 0.0000 | −0.8000 | H | −0.8000 | 2.6100 | 0.0000 |
| H | 2.6100 | 0.8000 | 0.0000 | N | 0.0000 | −2.0100 | 0.0000 |
| N | −2.0100 | 0.0000 | 0.0000 | H | 0.0000 | −2.6100 | 0.8000 |
| H | −2.6100 | 0.0000 | 0.8000 | H | 0.0000 | −2.6100 | −0.8000 |
| H | −2.6100 | 0.0000 | −0.8000 | H | 0.8000 | −2.6100 | 0.0000 |
| H | −2.6100 | −0.8000 | 0.0000 | ||||
| Atom | Atom | ||||||
| Mn | −0.0128264866 | −0.0002779912 | −0.1091894544 | H | 0.8908035853 | 3.6241374839 | 1.0039698993 |
| O | −1.1037647823 | 1.0246639195 | −1.2821459116 | H | 2.5024254412 | −2.4854165868 | 1.5399007635 |
| O | −0.4319309790 | −1.5910632792 | −1.0423314985 | H | 2.7858100041 | 1.4098930461 | 0.8661311665 |
| N | 0.7243918073 | 1.6205280457 | 0.7709992473 | H | 1.1490336627 | −0.0652673733 | 2.9502674019 |
| N | 1.3008767173 | −0.9287768925 | 1.0635839133 | H | 2.0638511115 | 2.0769335299 | 2.3353274192 |
| C | −1.2886186814 | 2.2834208702 | −1.2181338272 | H | 2.7993768711 | −0.4085469833 | 2.4534782102 |
| C | −0.7050865946 | 3.1821062059 | −0.3604271375 | H | −1.9892347032 | 2.6686747336 | −1.9454961675 |
| C | 0.3534920334 | 2.8195308082 | 0.5207059889 | H | −0.0986239324 | −3.4441353556 | −1.6764363233 |
| C | 0.2177419816 | −2.6945497951 | −0.9662726987 | Cl | 1.8821942342 | 0.3098852687 | −1.6599082362 |
| C | 1.2181378357 | −3.0250490115 | −0.0961995911 | O | −1.1395010178 | −0.1703680057 | 0.9263551082 |
| C | 1.7180430400 | −2.1229929449 | 0.8899899444 | H | −0.9312248148 | 4.2251864392 | −0.4696082573 |
| C | 1.9393113446 | 1.3525171542 | 1.5367699106 | H | 1.6695378989 | −3.9953914470 | −0.1633815245 |
| C | 1.8333722479 | −0.0604962603 | 2.1055467704 |
Table 2. Optimized Cartesian coordinates (Å) of [Fe(NH)].
performed 50 rounds of configuration recovery. The difference in batch sizes reflects the substantially smaller subspaces generated by the spin-adapted approach for a given number of samples. This allowed us to increase the batch size beyond samples without exceeding memory constraints, whereas larger Cartesian batches were computationally infeasible due to memory limitations. The number of configuration recovery rounds also differed between the two approaches: Cartesian post-processing typically converged within a few iterations, whereas the spin-completed energies continued to improve over several tens of iterations. Despite the larger number of iterations, the smaller subspaces kept the total computational time lower than or comparable to that of the Cartesian workflow. Selected subspace diagonalization is performed through the SBD diagonalizer [9]. In the case of the spin-completed subspaces we applied the spin projection of Appendix A.
References
- [1]D. Wecker, M. B. Hastings, and M. Troyer, Phys. Rev. A 92, 042303 (2015).
- [2]J. Lee, D. W. Berry, C. Gidney, W. J. Huggins, J. R. McClean, N. Wiebe, and R. Babbush, PRX Quantum 2, 030305 (2021).
- [3]J. J. Goings, A. White, J. Lee, C. S. Tautermann, M. Degroote, C. Gidney, T. Shiozaki, R. Babbush, and N. C. Rubin, Proc. Natl. Acad. Sci. U.S.A. 119, e2203533119 (2022).
- [4]K. Kanno, M. Kohda, R. Imai, S. Koh, K. Mitarai, W. Mizukami, and Y. O. Nakagawa, Phys. Rev. Research 8, 023268 (2026).
- [5]K. Sugisaki, S. Kanno, T. Itoko, R. Sakuma, and N. Yamamoto, Physical Chemistry Chemical Physics 27, 20869 (2025).
- [6]M. Mikkelsen and Y. O. Nakagawa, Physical Review Research 7, 043043 (2025).
- [7]J. Robledo-Moreno, M. Motta, H. Haas, et al., Sci. Adv. 11, eadu9991 (2025).
- [8]S. Barison, J. Robledo-Moreno, and M. Motta, Quantum Science and Technology 10, 025034 (2025).
- [9]T. Shirakawa, J. Robledo-Moreno, T. Itoko, et al., Closed-loop calculations of electronic structure on a quantum processor and a classical supercomputer at full scale (2025), arXiv:2511.00224 [quant-ph].arxiv.org/abs/2511.00224
- [10]D. Kaliakin, A. Shajan, F. Liang, J. Robledo-Moreno, Z. Li, A. Mitra, M. Motta, C. Johnson, A. A. Saki, S. Das, I. Sitdikov, A. Mezzacapo, and K. M. Merz, Jr., Communications Physics 8, 396 (2025).
- [11]K. M. Merz, Jr., A. Shajan, D. Kaliakin, et al., Crossing the 12,000-atom barrier with heterogeneous quantum-classical supercomputing: quantum chemistry of protein-ligand complexes (2026), arXiv:2605.01138.arxiv.org/abs/2605.01138
- [12]J. Yu, J. Robledo-Moreno, J. T. Josue, et al., Quantum-centric algorithm for sample-based Krylov diagonalization (2025), arXiv:2501.09702.arxiv.org/abs/2501.09702
- [13]S. Piccinelli, A. Baiardi, S. Barison, M. Rossmannek, A. Carrera Vázquez, F. Tacchino, S. Mensa, E. Altamura, A. Alavi, M. Motta, J. Robledo-Moreno, W. Kirby, K. Sharma, A. Mezzacapo, and I. Tavernelli, Quantum chemistry with provable convergence via randomized sample-based Krylov quantum diagonalization (2025), arXiv:2508.02578.arxiv.org/abs/2508.02578
- [14]E. Campbell, Phys. Rev. Lett. 123, 070503 (2019).
- [15]I. Rønnević, F. Paschke, Y. Gao, L.-A. Lieske, L. A. Gödde, S. Barison, S. Piccinelli, A. Baiardi, I. Tavernelli, J. Repp, F. Albrecht, H. L. Anderson, and L. Gross, Science 392, eadae3321 (2026).
- [16]S. Piccinelli, S. Barison, A. Baiardi, F. Tacchino, J. Repp, I. Rønnević, F. Albrecht, H. L. Anderson, L. Gross, A. Curioni, and I. Tavernelli, Exploring pathways towards quantum advantage in quantum chemistry: the case of a molecule with half-Möbius topology (2026), arXiv:2603.08696.arxiv.org/abs/2603.08696
- [17]S. Sharma, K. Sivalingam, F. Neese, and G. K.-L. Chan, Nat. Chem. 6, 927 (2014).
- [18]Z. Li, S. Guo, Q. Sun, and G. K.-L. Chan, Nat. Chem. 11, 1026 (2019).
- [19]G. Li Namini and A. Alavi, J. Phys. Chem. A 122, 4935 (2018).
- [20]S. Sharma and G. K.-L. Chan, J. Chem. Phys. 136, 124121 (2012).
- [21]S. Keller and M. Reiher, J. Chem. Phys. 144, 134101 (2016).
- [22]P. Schmoll, S. Singh, M. Rizzi, and R. Orús, Ann. Phys. 419, 168232 (2020).
- [23]W. Dobrautz, S. D. Smart, and A. Alavi, J. Chem. Phys. 151, 094104 (2019).
- [24]G. Li Manni, W. Dobrautz, and A. Alavi, J. Chem. Theory Comput. 16, 2202 (2020).
- [25]V. G. Chilkuri, T. Applencourt, K. Gasperich, P.-F. Loos, and A. Scemama, New Electron Correlation Methods and their Applications, and Use of Atomic Orbitals with Exponential Asymptotes, Advances in Quantum Chemistry, Vol. 83 (Elsevier, 2021) pp. 65–81.DOI
- [26]M. Ugandi and M. Roemelt, J. Comput. Chem. 44, 2374 (2023).
- [27]A. Carbone, D. E. Galli, M. Motta, and B. Jones, Symmetry 14, 624 (2022).
- [28]D. Marti-Dafcik, H. G. A. Burton, and D. P. Tew, Phys. Rev. Res. 7, 013191 (2025).
- [29]P. J. Ollitrault, C. L. Cortes, J. F. Gonthier, R. M. Parrish, D. Rocca, G.-L. Anselmetti, M. Degroote, N. Moll, R. Santagati, and M. Streif, Phys. Rev. Lett. 133, 250601 (2024).
- [30]S. Wouters, T. Bogaerts, P. Van Der Voort, V. Van Speybroeck, and D. Van Neck, J. Chem. Phys. 140, 241103 (2014).
- [31]R. Olivares-Amaya, W. Hu, N. Nakatani, S. Sharma, J. Yang, and G. K.-L. Chan, J. Chem. Phys. 142, 034102 (2015).
- [32]C. J. Stein and M. Reiher, J. Chem. Theory Comput. 12, 1760 (2016).
- [33]N. A. Bogdanov, G. Li Manni, S. Sharma, O. Gunnarsson, and A. Alavi, Nature Physics 18, 190 (2022).
- [34]I. Shavitt, Int. J. Quantum Chem. 12, 131 (1977).
- [35]I. Shavitt, Int. J. Quantum Chem. 14, 5 (1978).
- [36]H. Lischka, T. Müller, P. G. Szalay, I. Shavitt, R. M. Pitzer, and R. Shepard, WIREs Comput. Mol. Sci. 1, 191 (2011).
- [37]G. Li Manni, I. Fdez. Galván, A. Alavi, et al., J. Chem. Theory Comput. 19, 6933 (2023).
- [38]W. Dobrautz, O. Weser, N. A. Bogdanov, A. Alavi, and G. Li Manni, J. Chem. Theory Comput. 17, 5684 (2021).
- [39]K. Sugisaki, S. Yamamoto, S. Nakazawa, K. Toyota, K. Sato, D. Shiomi, and T. Takui, Chem. Phys. Lett. 737, 100002 (2019).
- [40]M. Liu and Z. Li, Minimal building blocks for molecular quantum circuits with exact spin symmetry (2026), arXiv:2609.37101 [quant-ph].arxiv.org/abs/2609.37101
- [41]J. Burkat and N. Fitzpatrick, The quantum paldus transform: Efficient circuits with applications (2025), arXiv:2506.09151 [quant-ph].arxiv.org/abs/2506.09151
- [42]R. Pauncz, Spin Eigenfunctions: Construction and Use (Plenum Press, New York, 1979).
- [43]T. Helgaker, P. Jørgensen, and J. Olsen, Molecular Electronic-Structure Theory (Wiley, Chichester, 2000).DOI
- [44]H. Uoyama, K. Goushi, K. Shizu, H. Nomura, and C. Adachi, Nature 492, 234 (2012).
- [45]M. K. Etherington, J. Gibson, H. F. Higginbotham, T. J. Penfold, and A. P. Monkman, Nat. Commun. 7, 13680 (2016).
- [46]K. Vandewal, S. Albrecht, E. T. Hoke, et al., Nat. Mater. 13, 63 (2014).
- [47]Z. R. Grabowski, K. Rotkiewicz, and W. Rettig, Chem. Rev. 103, 3899 (2003).
- [48]M. Segado, I. Gómez, and M. Reguero, Physical Chemistry Chemical Physics 18, 6861 (2016).
- [49]P.-F. Loos, M. Comin, X. Blase, and D. Jacquemin, Journal of Chemical Theory and Computation 17, 3666 (2021).
- [50]A. A. Holmes, N. M. Tubman, and C. J. Umrigar, Journal of Chemical Theory and Computation 12, 3674 (2016).
- [51]S. Sharma, A. A. Holmes, G. Jeanmairet, A. Alavi, and C. J. Umrigar, Journal of Chemical Theory and Computation 13, 1595 (2017).
- [52]A. A. Holmes, C. J. Umrigar, and S. Sharma, The Journal of Chemical Physics 147, 164111 (2017).
- [53]S. Sharma et al., Dice, https://github.com/sanshar/Dice (2026).
- [54]M. Lefèvre, E. Proietti, F. Jaouen, and J.-P. Dodelet, Science 324, 71 (2009).
- [55]A. Zitolo, V. Goellner, V. Armel, M.-T. Sougrati, T. Mineva, L. Stievano, E. Fonda, and F. Jaouen, Nature Materials 14, 937 (2015).
- [56]K. Liu, J. Fu, Y. Lin, T. Luo, G. Ni, H. Li, Z. Lin, and M. Liu, Nature Communications 13, 2075 (2022).
- [57]T. Zheng, J. Wang, Z. Xia, G. Wang, and Z. Duan, Journal of Materials Chemistry A 11, 19360 (2023).
- [58]G. Li Manni, S. D. Smart, and A. Alavi, J. Chem. Theory Comput. 12, 1245 (2016).
- [59]W. Zhang, J. L. Loebach, S. R. Wilson, and E. N. Jacobsen, J. Am. Chem. Soc. 112, 2801 (1990).
- [60]R. Irie, K. Noda, Y. Ito, M. Matsumoto, and T. Katsuki, Tetrahedron Letters 31, 7345 (1990).
- [61]E. M. McGarrigle and D. G. Gilheany, Chemical Reviews 105, 1563 (2005).
- [62]J. Ivanic, J. R. Collins, and S. K. Burt, The Journal of Physical Chemistry A 108, 2314 (2004).
- [63]J. S. Sears and C. D. Sherrill, The Journal of Chemical Physics 124, 144314 (2006).
- [64]A. Gandon, A. Baiardi, M. Rossmannek, W. Dobrautz, and I. Tavernelli, PRX Quantum 6, 030306 (2025).
- [65]M. Motta, K. J. Sung, K. B. Whaley, M. Head-Gordon, and J. Shee, Chem. Sci. 14, 11213 (2023).
- [66]E. R. Kjellgren, K. M. Ziems, P. Reinholdt, S. P. A. Sauer, S. Coriani, and J. Kongsted, Spin-adapted Multistate Contracted VQE for Excited States and Properties on Quantum Hardware (2026), arXiv:2606.15489.arxiv.org/abs/2606.15489
- [67]S. G. Mehendale, B. Peng, N. Govind, and Y. Alexeev, J. Phys. Chem. A 127, 4526 (2023).
- [68]E. R. Davidson, J. Comput. Phys. 17, 87 (1975).
- [69]B. Liu, in Numerical Algorithms in Chemistry: Algebraic Methods, edited by C. Moler and I. Shavitt (Lawrence Berkeley Laboratory, Berkeley, CA, 1978) pp. 49–53, report LBL-8158.
- [70]Y. Saad, Numerical Methods for Large Eigenvalue Problems, revised ed., Classics in Applied Mathematics, Vol. 66 (SIAM, Philadelphia, 2011).
- [71]H. Zhang and M. Otten, From random determinants to the ground state (2025), arXiv:2511.14734.arxiv.org/abs/2511.14734
- [72]L. Bytautas and K. Ruedenberg, in Electron Correlation Methodology, edited by A. K. Wilson and K. A. Peterson (American Chemical Society, 2007).
- [73]E. R. Sayfutyarova, Q. Sun, G. K.-L. Chan, and G. Knizia, Journal of Chemical Theory and Computation 13, 4063 (2017).
- [74]F. Weigend and R. Ahlrichs, Phys. Chem. Chem. Phys. 7, 3297 (2005).




