betaFeedback

WorksPhysics and Astronomy

Efficient Block Encodings for Ensembles of Molecular Hamiltonians

Abstract

Accurately characterizing molecular properties and reactivity often requires electronic-structure calculations across an ensemble of molecular geometries. Quantum algorithms that process such ensembles coherently could address this challenge, but they require Hamiltonian block encoding circuits that combine advanced factorization techniques with sublinear scaling in ensemble size. Here, we introduce SVD-THC, which extends tensor hypercontraction to sets of Hamiltonians defined at different nuclear geometries through a singular value decomposition of the ensemble. We use SVD-THC to construct a single multiplexed block encoding circuit for all geometries in an ensemble. We then evaluate the approach on ensembles generated by varying internal rotational coordinates in 1-bromo-2-chloroethane as well as in glycine and alanine dipeptides, corresponding to established model systems for peptide conformational landscapes, with up to 5,184 geometries. Across several resolutions of the internal-coordinate grids, the Toffoli complexity of the resulting circuits grows only weakly with the number of geometries, changing by at most 12% over grids that differ in size by up to 576×\times. In an energy-thresholding circuit, this yields 35–200×\times fewer Toffoli gates than slicing together single-geometry block encodings. This approach could help make coherent processing of molecular geometry ensembles a useful primitive for fault-tolerant quantum computing.

Introduction

Ground state energy estimation is often described as the central use case for quantum computing in chemistry [1, 2, 3, 4, 5, 6]. Indeed, a large body of work has been dedicated to the solution of the electronic structure problem, where the quantum computer would find the ground state of an electronic Hamiltonian for a fixed set of atomic coordinates. However, this static image of quantum chemistry rarely delivers answers to pharmaceutically relevant questions [7, 8]. Many interesting properties can instead be derived from the shape of the potential energy surface (PES) – the energy functional across atomic geometries: In geometry optimization, one searches for the minimal point in the PES which is where the molecule is most stable. During a molecular dynamics (MD) run, one follows the nuclei as their positions evolve over time which naturally generates the low-energy conformations or allows one to follow reagents as they chemically transform into products. The slope of the PES indicates the direction in which the nuclei experience a force during the time evolution [9, 10, 11, 12, 13]. Rotatable bonds determine the relative positions of different active parts of small-molecule drugs. The lowest-energy dihedral allows one to understand how stable the bioactive conformation is. Rotational profiles illuminate where barriers to torsional movement exist and which areas of phase space are accessible by thermal motion. That informs the decision whether a certain conformation of a drug molecule is stable and whether other forms are accessible under thermal motion. These ensemble functions provide important data that helps make decisions in the drug discovery cycle [14]. To learn these properties in quantum computing, one would either have to solve the electronic structure problem for multiple geometries, or encode electronic information of multiple geometries in a state of the form

∣Ψ⟩=∑θcθ∣θ⟩∣ψθ,0⟩,(1)\lvert\Psi\rangle= \sum_{\theta} c_{\theta}\lvert\theta\rangle\lvert\psi_{\theta,0}\rangle, \tag*{(1)}

where ∣ψθ,0⟩\lvert\psi_{\theta,0}\rangle are electronic ground states of the geometry vector θ\boldsymbol{\theta} encoded in some index state ∣θ⟩\lvert\theta\rangle [15, 16, 17]. With such a state, one could aggregate information across multiple geometries. The stretch goal would be to eventually encode a version of the PES itself, where the coefficients cθc_{\theta} become a function of the ground state energies [13, 18]. For the preparation of a PES state in Eq. (1), as well as operations such as free energy calculation and geometry optimization, one needs coherent and efficient access to electronic structure Hamiltonians of all geometries [19, 20]. That is, we need to efficiently encode Hamiltonians that are a functional of the atomic coordinates.

To be competitive with methods relying on repeated energy estimation and classical postprocessing for each geometry, one must aim for a compression sublinear in the number of geometries considered. For a list of MM geometries θ\boldsymbol{\theta}, the total cost of encoding the geometry-dependent Hamiltonian H(θ)H(\boldsymbol{\theta}) must scale as o(M)o(M) times the average cost of encoding a single-geometry Hamiltonian. Since the cost of encoding single-geometry Hamiltonians has gradually been reduced through years of intense research, a tight baseline is set: by replacing time evolution with a quantum walk operator, qubitization has greatly accelerated Hamiltonian simulation through block encodings [21, 22, 23, 24, 25]. Factorization algorithms such as single factorization, sparse factorization [26], double factorization (DF) [27, 28], tensor hypercontraction (THC) [29, 30] as well as double-factorized tensor hypercontraction (DFTHC) [31] exploit patterns in the Hamiltonians to compress block encoding circuits and speed up the quantum algorithm. Recently, the block-invariant symmetry shift [32, 33] has been used to push this compression towards its theoretical limit [34]. To be competitive with repeated queries to single-geometry energy estimation, Hamiltonian simulation for geometric ensembles must draw inspiration from the same tools. In Hamiltonian factorization techniques, the strategy is to reduce the amount of redundant information loaded onto the quantum computer and instead generate the data procedurally from a compressed representation obtained with classical preprocessing. The same strategy could be applied to geometric ensembles. Relevant ensembles of Hamiltonians often share a large similarity and differ only in a few coordinates. It is common that the atomic composition and even the relative positions of atoms in ligands further removed from the active center remain the same. A set of Hamiltonians generated over a grid of coordinates may consequently have a lot of redundant information. This gives us confidence that we can devise a Hamiltonian loading scheme that makes use of this extra degree of redundancy and that is capable of constructing a coherent block encoding of the full ensemble that is cheaper than loading each geometry sequentially.

In this paper, we develop SVD-THC, a compression technique that fits an ensemble of MM electronic structure Hamiltonians into the same block encoding. This technique allows us to time-evolve an ensemble of various geometries, encoded into a quantum state, as a whole. In the pursuit of an algorithmic complexity sublinear in MM, we propose a singular value decomposition (SVD) of the ensemble, that expresses the electronic interactions of all geometries as a superposition of fewer than MM, θ\boldsymbol{\theta}-independent Hamiltonians. All these SVD Hamiltonians are compressed with THC and BLISS techniques, which calls for a multiplexed THC circuit that is efficient for a set of Hamiltonians with individual THC ranks.

We apply our SVD-THC compression technique to the Hamiltonian ensembles of 1-bromo-2-chloroethane (BCE), glycine dipeptide and alanine dipeptide. These model systems increase the hardness of modeling from a one-dimensional PES to a challenging two-dimensional benchmark for machine-learned interatomic potentials (MLIPs) [35]. For each of these three molecular systems we perform the numerical decomposition of the Hamiltonian ensemble and investigate potential energy surfaces. Quantum resource estimates (QREs) are then obtained in the form of block encoding 1-norms, Toffoli gate counts and active volume (AV) for the energy thresholding circuit, which isolates the low-energy subspace of an ensemble. We discuss the compression factors achieved through SVD-THC, by comparing the thresholding circuit to its counterpart in the picture of repeated single-geometry THC circuits.

We find that the advantage of SVD-THC grows with the number of geometries. Since a single SVD-THC circuit serves the whole ensemble, its cost stays nearly constant as we refine the grid, whereas repeating single-geometry circuits grows in proportion to the number of geometries. On the largest grids, our approach reduces the Toffoli count by 35–41× for BCE with 73 geometries, 144× for glycine dipeptide with 2,664 geometries, and 201× for alanine dipeptide with 5,184 geometries, while keeping the CCSD correlation energy within 0.6 mHa for all configurations. Even against the most economical alternative, corresponding to a separate phase estimation for each geometry, we obtain a 10–60× reduction. Ultimately, we find that the compression starts paying off for grids with tens of geometries, with an advantage that continues to grow as the grid is further refined.

The rest of this paper is organized as follows: first we lay out the problem setting and review BLISS-THC in Section II, and develop the SVD-THC compression in Section III, before we draw its quantum circuit in Section IV. In § 1.4, we present numerical results for the SVD-THC factorization applied to our benchmark molecules, and in Section VI, we obtain QREs for the thresholding circuit. We close the paper with the conclusions and outlook in § 1.6.

Problem Setting and Background

In this section, we define the problem setting by describing ensembles of geometry-dependent Hamiltonians mathematically. Since Hamiltonian factorization is going to have an important role in this work, we briefly review the high-level ideas behind BLISS-THC [33].

Parameterized electronic structure Hamiltonians

We consider the parameterized electronic structure Hamiltonian in second quantization,

H(θ)=E01+∑pq∑σhpq(θ)apσ†aqσ+12∑pqrs∑στgpqrs(θ)apσ†aqτ†asτarσ(2)H(\theta) = \mathcal{E}_0 \mathbf{1} + \sum_{pq} \sum_{\sigma} h_{pq}(\theta) a^\dagger_{p\sigma} a_{q\sigma} + \frac{1}{2} \sum_{pqrs} \sum_{\sigma\tau} g_{pqrs}(\theta) a^\dagger_{p\sigma} a^\dagger_{q\tau} a_{s\tau} a_{r\sigma} \tag*{(2)}

where p,q,r,sp,q,r,s label NN spatial orbitals, σ,τ∈{↑,↓}\sigma,\tau\in\{\uparrow,\downarrow\} denote spin, and hpqh_{pq} and gpqrsg_{pqrs} are the one- and two-electron integrals in chemist’s notation, assumed to be real. The scalar E0\mathcal{E}_0 contains the nuclear repulsion energy and the mean-field contribution from excluded occupied orbitals. Although this term is typically omitted in block encoding schemes because its contribution can be added during classical postprocessing, it must be retained here since it contributes to the relative energies between configurations. Note that Eq. (2) supplies Hamiltonian integrals as a function of θ\theta; obtaining the full ground-state energy E(θ)E(\theta) still requires solving the electronic structure problem at each geometry.

The parameter vector, θ=(θ1,…,θd)\boldsymbol{\theta} = (\theta_1,\ldots,\theta_d), contains dd arbitrary coordinates. For example, θi\theta_i could be a torsion angle, which quantifies the rotation about a chemical bond: for four consecutively bonded atoms A–B–C–D, it is the angle between the plane spanned by A, B, and C and the plane spanned by B, C, and D. If the bond lies along the main chain, or backbone, of a molecule such as a peptide, this angle is called a backbone dihedral. We sample these coordinates on a product grid θn\boldsymbol{\theta}_n, where n=(n1,…,nd)\boldsymbol{n} = (n_1,\ldots,n_d), with ni∈{0,…,Mi−1}n_i \in\{0,\ldots,M_i-1\}. The grid therefore contains M=∏i=1dMiM = \prod_{i=1}^{d} M_i total configurations. For quantum circuit constructions in this work, we address each configuration using a single flattened index θ∈{0,…,M−1}\theta\in\{0,\ldots,M-1\}, and we use θn\boldsymbol{\theta}_n and θ\theta interchangeably, with computational states ∣θ⟩|\theta\rangle indexing geometries in quantum registers.

A naive block encoding of Eq. (2) treats every grid point independently. Its cost therefore scales linearly with MM and does not exploit any similarity between neighboring geometries. Our proposed approach instead expands the parameterized Hamiltonian in a data-dependent basis: up to a geometry-dependent energy shift, H(θ)H(\theta) is written as a linear combination of KK fixed, geometry-independent operators HℓH_\ell with geometry-dependent coefficients. Each HℓH_\ell has the form of a molecular electronic-structure Hamiltonian, i.e., a sum of one- and two-body fermionic terms, so that established factorization techniques can be applied to it. We focus on BLISS-THC for this manuscript but other common factorizations should be compatible. Note that the construction we present below is agnostic to how the integrals are obtained and applies generally in the all-electron, frozen-core, and active-space settings. We illustrate this explicitly in § 1.4 using the frozen-core picture, where core electrons are projected out and the remaining orbitals are Löwdin orthogonalized. By contrast, selecting a consistent active space that varies smoothly across the grid is more difficult since active space orbitals can mix or reorder as the geometry changes. We defer this problem to future work.

BLISS-THC

For a specific fermionic Hamiltonian, such as the typical electronic structure Hamiltonian, which corresponds to a single configuration of Eq. (2), our construction uses the block-invariant symmetry-shifted tensor hypercontraction (BLISS-THC) factorization introduced in Ref. [33]. For completeness, we summarize the BLISS-THC approach based on the Majorana representation. Expanding the annihilation operator as apσ=(γ^pσ,0+iγ^pσ,1)/2a_{p\sigma} = (\hat{\gamma}_{p\sigma,0} + i\hat{\gamma}_{p\sigma,1})/2, the block-invariant electronic structure Hamiltonian may be written as,

HBI=i2∑pq,σκpq(BI)γ^pσ,0γ^qσ,1−18∑pqrs,στgpqrs(BI)γ^pσ,0γ^qσ,1γ^rτ,0γ^sτ,1,(3)H_{\mathrm{BI}} = \frac{i}{2} \sum_{pq,\sigma} \kappa^{(\mathrm{BI})}_{pq} \hat{\gamma}_{p\sigma,0}\hat{\gamma}_{q\sigma,1} - \frac{1}{8} \sum_{pqrs,\sigma\tau} g^{(\mathrm{BI})}_{pqrs} \hat{\gamma}_{p\sigma,0}\hat{\gamma}_{q\sigma,1}\hat{\gamma}_{r\tau,0}\hat{\gamma}_{s\tau,1}, \tag*{(3)}

with block-invariant coefficients

κpq(BI)=hpq−12∑rgprrq+∑rgprrq(BI)−α1δpq+12βpqη,(4)\kappa^{(\mathrm{BI})}_{pq} = h_{pq} - \frac{1}{2}\sum_r g_{prrq} + \sum_r g^{(\mathrm{BI})}_{prrq} - \alpha_1\delta_{pq} + \frac{1}{2}\beta_{pq}\eta, \tag*{(4)}
gpqrs(BI)=gpqrs−α2δpqδrs−12(βpqδrs+δpqβrs).(5)g^{(\mathrm{BI})}_{pqrs} = g_{pqrs} - \alpha_2\delta_{pq}\delta_{rs} - \frac{1}{2}\left(\beta_{pq}\delta_{rs}+\delta_{pq}\beta_{rs}\right). \tag*{(5)}

Here, the scalars α1\alpha_1 and α2\alpha_2, together with the one-body matrix βpq\beta_{pq}, parameterize a shift by a function of the particle number operator N^\hat{N} that vanishes identically in the η\eta-electron sector. Thus, within that sector, HBIH_{\mathrm{BI}} has the same eigenvalue spectrum as HH, while its 1-norm can become substantially smaller. The block-invariant two-body tensor admits the THC decomposition,

gpqrs(BI)=∑μ,ν=1Rζμνχpμχqμχrνχsν.(6)g^{(\mathrm{BI})}_{pqrs} = \sum_{\mu,\nu=1}^{R} \zeta_{\mu\nu}\chi^\mu_p\chi^\mu_q\chi^\nu_r\chi^\nu_s. \tag*{(6)}

where RR is the THC rank. The leaf unit vectors χμ\chi^\mu are implemented on the quantum computer using Givens rotations, while the core tensor ζ~μν\tilde{\zeta}_{\mu\nu} is loaded through state preparation. The resulting block encoding has 1-norm,

λ~THC=∑k∣t~k∣+12∑μ,ν∣ζ~μν∣−14∑μ∣ζ~μμ∣(7)\tilde{\lambda}_{\mathrm{THC}} = \sum_k \lvert\tilde{t}_k\rvert+ \frac{1}{2}\sum_{\mu,\nu}\lvert\tilde{\zeta}_{\mu\nu}\rvert- \frac{1}{4}\sum_\mu\lvert\tilde{\zeta}_{\mu\mu}\rvert \tag*{(7)}

where t~k\tilde{t}_k are the eigenvalues of κ(BI)\kappa^{(\mathrm{BI})}. Eq. (7) is the objective that BLISS-THC minimizes along with the residual reconstruction error, and it is the per-mode input to everything that follows.

SVD-parameterized BLISS-THC

The previous section described how BLISS-THC compresses a single Hamiltonian at a fixed geometry. We now extend this construction to an ensemble of Hamiltonians evaluated on a product grid of geometries. The objective here is to represent the full ensemble using a small number of geometry-independent Hamiltonians HℓH_\ell, where each Hamiltonian is multiplied by a scalar function of θ\theta. Finding such a representation allows us to apply a tensor factorization method, such as BLISS-THC, to each mode separately. We name this technique SVD-THC.

A natural starting point requires obtaining the geometry-dependent functions defined from a fixed basis. Fourier, Chebyshev, and Walsh-Hadamard transforms are possible choices [36]. One of their advantages is that their functions are known ahead of time and are well-behaved. Fourier and Chebyshev bases are effective when the Hamiltonian coefficients vary smoothly across the grid. Walsh-Hadamard bases can be useful when the coefficients have a piecewise or otherwise simple structure. However, these assumptions are not always appropriate for constrained potential-energy-surface scans. In this work, we consider potential energy scans of two kinds. In a rigid scan, only the scanned coordinates change and every other internal coordinate is held fixed. In a relaxed scan, at each point of such a scan, the unconstrained coordinates are geometry optimized resulting in molecular configurations that can change abruptly as the scanned coordinates vary. Relaxed scans give physically relevant energy surfaces, but as we show in § 1.4, they are also harder to compress. Conical intersections can also introduce additional nonanalytic behavior, including cusps. A fixed basis may therefore require many modes to represent these types of features accurately.

In this work, we propose constructing the geometry basis directly from the electronic tensors themselves. The basis should therefore adapt to the variation present in the Hamiltonian, including non-smooth features caused by geometry optimization or conical intersections. This basis is obtained from a singular value decomposition applied to the integral tensors sampled across the grid. For large grids, we compute the relevant singular vectors through an M×MM \times M Gram matrix rather than by forming the much larger orbital-space covariance matrices. This section describes the full construction. We first reshape the integral tensors into snapshot matrices. We then use a single joint geometry basis for the one- and two-body terms. Finally, we show how each retained mode defines a geometry-independent chemistry-like Hamiltonian that can be compressed with BLISS-THC.

Snapshot matrices

We begin by reshaping the one- and two-electron integral tensors sampled across the geometry grid. At every grid point, both tensors are computed in the consistent orbital representation defined in Sec. II A. We then reshape each tensor into a snapshot matrix, whose rows correspond to geometries and whose columns correspond to flattened orbital indices. For the one-body channel, we use the Majorana one-body coefficient that appears in the electronic structure Hamiltonian,

κpq(θ)=hpq(θ)+∑rgpqrr(θ)−12∑rgprrq(θ).(8)\kappa_{pq}(\theta) = h_{pq}(\theta) + \sum_{r} g_{pqrr}(\theta) - \frac{1}{2}\sum_{r} g_{prrq}(\theta). \tag*{(8)}

We work with κpq\kappa_{pq}, rather than hpqh_{pq}, because κpq\kappa_{pq} is the one-body matrix that enters the downstream Majorana Hamiltonian representation. The corresponding snapshot matrix is defined as,

Aθ,(pq)⇔κpq(θ),A∈RM×N2.(9)A_{\theta,(pq)} \Leftrightarrow\kappa_{pq}(\theta), \qquad A \in\mathbb{R}^{M \times N^{2}}. \tag*{(9)}

For the two-body channel, we retain only the entries that are unique under the usual eight-fold permutational symmetry. The number of unique entries is therefore given by n8=npair(npair+1)/2n_{8} = n_{\mathrm{pair}}(n_{\mathrm{pair}} + 1)/2, where npair=N(N+1)/2n_{\mathrm{pair}} = N(N + 1)/2. The two-body snapshot matrix is therefore defined as,

Bθ,[pqrs]⇔gpqrs(θ),B∈RM×n8,(10)B_{\theta,[pqrs]} \Leftrightarrow g_{pqrs}(\theta), \qquad B \in\mathbb{R}^{M \times n_{8}}, \tag*{(10)}

where [pqrs][pqrs] labels one unique index. All inner products between two-body snapshots are computed using only these unique entries, so symmetry-related entries are not counted multiple times.

Geometry SVD in a joint basis

At this stage, we have separate snapshot matrices for the one- and two-body integral tensors. While we could perform a singular value decomposition of each matrix independently, this would produce one set of basis vectors for the one-body terms and another set for the two-body terms. Although this approach can compress each tensor separately, it would require the circuit to prepare two different geometry-dependent decompositions. It would also prevent us from using a single set of basis vectors and BLISS coefficients when constructing the Hamiltonian block encoding. We find that this results in a less efficient block encoding of the ensemble. We therefore propose using the single joint basis obtained by defining the normalized Gram matrix,

G=cos⁡2(w)AATtr⁡(AAT)+sin⁡2(w)BBTtr⁡(BBT).(11)\mathcal{G} = \cos^{2}(w)\frac{AA^{\mathsf{T}}}{\operatorname{tr}(AA^{\mathsf{T}})} + \sin^{2}(w)\frac{BB^{\mathsf{T}}}{\operatorname{tr}(BB^{\mathsf{T}})}. \tag*{(11)}

The angle ww sets the relative weight of the two channels, interpolating between a basis fixed entirely by the one-body channel (w=0w = 0), and one fixed entirely by the two-body (w=π/2w = \pi/2) channel and could be optimized per ensemble. We adopt equal weighting, w=π/4w = \pi/4, throughout; all numerical results reported below use this choice. Since each term has unit trace, both the one- and two-body channels contribute equally to the overall basis. This prevents the channel with the larger overall Frobenius norm from determining the basis by itself. The Gram matrix has dimension M×MM \times M, which is much smaller than the covariance matrices ATAA^{\mathsf{T}}A and BTBB^{\mathsf{T}}B when the number of grid points is smaller than the number of tensor entries. Since M≪n8M \ll n_{8} for the systems considered here, we use this Gram-matrix formulation, which is also known as the method of snapshots [37, 38]. The compact basis representation is obtained by diagonalizing G\mathcal{G},

G=∑ℓ=0M−1ωℓVℓ(θ)Vℓ(θ′).(12)\mathcal{G} = \sum_{\ell=0}^{M-1} \omega_{\ell} V_{\ell}(\theta)V_{\ell}(\theta'). \tag*{(12)}

where Vℓ(θ)V_{\ell}(\theta) is an orthonormal eigenvector defined over the geometry grid and ωℓ\omega_{\ell} measures the importance of each joint geometry profile, ordered so that ω0≥ω1≥⋯\omega_{0} \ge\omega_{1} \ge\cdots. The one- and two-body mode tensors are obtained by projecting the snapshot matrices onto the joint basis,

Kpq(ℓ)=∑θVℓ(θ)Aθ,pq,Gpqrs(ℓ)=∑θVℓ(θ)Bθ,pqrs.(13)\mathcal{K}^{(\ell)}_{pq} = \sum_{\theta} V_{\ell}(\theta) A_{\theta,pq}, \qquad G^{(\ell)}_{pqrs} = \sum_{\theta} V_{\ell}(\theta) B_{\theta,pqrs}. \tag*{(13)}

The corresponding reconstruction using the leading KK basis modes is then given by,

κpq(θ)≈∑ℓ=0K−1Vℓ(θ)Kpq(ℓ),gpqrs(θ)≈∑ℓ=0K−1Vℓ(θ)Gpqrs(ℓ).(14)\kappa_{pq}(\theta) \approx\sum_{\ell=0}^{K-1} V_{\ell}(\theta) \mathcal{K}^{(\ell)}_{pq}, \qquad g_{pqrs}(\theta) \approx\sum_{\ell=0}^{K-1} V_{\ell}(\theta) G^{(\ell)}_{pqrs}. \tag*{(14)}

These relations define the proposed compressed basis representation.

Effective mode Hamiltonians

By substituting Eq. (14) into the parameterized Majorana Hamiltonian H(θ)H(\theta), we obtain a geometry-independent chemistry-like Hamiltonian for each pair of mode tensors (K(ℓ),G(ℓ))(\mathcal{K}^{(\ell)},G^{(\ell)}):

H~ℓ=i2∑pq,σKpq(ℓ)γ^pσ,0γ^qσ,1−18∑pqrs,στGpqrs(ℓ)γ^pσ,0γ^qσ,1γ^rτ,0γ^sτ,1.(15)\widetilde{H}_{\ell} = \frac{i}{2} \sum_{pq,\sigma} \mathcal{K}^{(\ell)}_{pq} \hat{\gamma}_{p\sigma,0}\hat{\gamma}_{q\sigma,1} - \frac{1}{8} \sum_{pqrs,\sigma\tau} G^{(\ell)}_{pqrs} \hat{\gamma}_{p\sigma,0}\hat{\gamma}_{q\sigma,1}\hat{\gamma}_{r\tau,0}\hat{\gamma}_{s\tau,1}. \tag*{(15)}

The full parameterized Hamiltonian is therefore,

HSVD(θ)=Eθ+∑ℓ=0K−1Vℓ(θ)H~ℓ,(16)H_{\mathrm{SVD}}(\theta) = \mathcal{E}_{\theta} + \sum_{\ell=0}^{K-1} V_{\ell}(\theta) \widetilde{H}_{\ell}, \tag*{(16)}

where all geometry dependence in the operator component is carried by the mode vectors Vℓ(θ)V_{\ell}(\theta). The scalar Eθ\mathcal{E}_{\theta} is the additive nuclear repulsion and frozen-core energy contribution defined in Sec. II A. Eθ\mathcal{E}_{\theta} remains outside the block encoding and is restored through the phase register when total energies are required. Sec. VI A describes this procedure, and shows that the constant may be used to replace the approximate Hartree-Fock energy of the compressed Hamiltonian with the exact one. Since each H~ℓ\widetilde{H}_{\ell} has the structure of a chemistry Hamiltonian, we may apply tensor factorization techniques to each mode independently. Here, we use BLISS-THC to compress the four-index tensors Gpqrs(ℓ)G^{(\ell)}_{pqrs} into mode-dependent core tensors ζ~μν(ℓ)\widetilde{\zeta}^{(\ell)}_{\mu\nu} and leaf tensors χℓμ\chi^{\ell\mu}. The one-body terms are represented by the corresponding shifted eigenvalues t~k(ℓ)\widetilde{t}^{(\ell)}_{k}. Here, the tilde denotes the unnormalized quantities returned by the BLISS-THC fit. Evaluating Eq. (7) for each mode gives the corresponding block encoding 1-norm,

σℓ=∑k∣t~k(ℓ)∣+12∑μ,ν∣ζ~μν(ℓ)∣−14∑μ∣ζ~μμ(ℓ)∣.(17)\sigma_{\ell} = \sum_{k} \left|\widetilde{t}^{(\ell)}_{k}\right| + \frac{1}{2}\sum_{\mu,\nu}\left|\widetilde{\zeta}^{(\ell)}_{\mu\nu}\right| - \frac{1}{4}\sum_{\mu}\left|\widetilde{\zeta}^{(\ell)}_{\mu\mu}\right|. \tag*{(17)}

Writing H~ℓBI\widetilde{H}^{\mathrm{BI}}_{\ell} for the operator that these fitted quantities reconstruct, that is, Eq. (15) with K(ℓ)\mathcal{K}^{(\ell)} and G(ℓ)G^{(\ell)} replaced by their block-invariant BLISS-THC representations, we define the normalized mode Hamiltonian Hℓ=H~ℓBI/σℓH_{\ell} = \widetilde{H}^{\mathrm{BI}}_{\ell}/\sigma_{\ell}, which has unit 1-norm by construction. We note that H~ℓ\widetilde{H}_{\ell} is the exact mode Hamiltonian of Eq. (16) obtained from the SVD, whereas H~ℓBI\widetilde{H}^{\mathrm{BI}}_{\ell} is its BLISS-THC approximation. Since the BLISS symmetry shift vanishes in the fixed η\eta-electron sector, Eq. (16) becomes

HSVD(θ)=Eθ+∑ℓ=0K−1σℓVℓ(θ)Hℓ.(18)H_{\mathrm{SVD}}(\theta) = \mathcal{E}_{\theta} + \sum_{\ell=0}^{K-1} \sigma_{\ell} V_{\ell}(\theta) H_{\ell}. \tag*{(18)}

This is the representation that we block encode below. The operators HℓH_{\ell} contain the compressed operator structure, σℓ\sigma_{\ell} sets the energy scale of each mode, and Vℓ(θ)V_{\ell}(\theta) contains its geometry dependence.

Quantum circuit

To encode the electronic structure of various geometries in the same wave function, we introduce separate registers: a physical register ss, representing an electronic system on 2N2N qubits, and an angle register aa that stores the flattened geometry index θ\theta as computational state ∣θ⟩\lvert\theta\rangle. A typical state could look like

∣Φ⟩=∑θcθ∣θ⟩a∣ϕθ⟩s,(19)\lvert\Phi\rangle= \sum_{\theta} c_{\theta}\lvert\theta\rangle_{a}\lvert\phi_{\theta}\rangle_{s}, \tag*{(19)}

where cθc_{\theta} are coefficients and ∣ϕθ⟩\lvert\phi_{\theta}\rangle are some states representing an electronic wave function. Following the arguments of Sec. III, we will address the subspace of every geometry θ\theta encoded in the ensemble with the parameterized Hamiltonian H^θ=∑ℓσℓVℓ(θ)Hℓ\widehat{H}_{\theta} = \sum_{\ell} \sigma_{\ell}V_{\ell}(\theta)H_{\ell}. For electronic structure calculations over ensembles of geometries, we require a Hamiltonian multiplexor, a quantum circuit that assigns every ∣θ⟩\lvert\theta\rangle subspace of aa its own Hamiltonian H^θ\widehat{H}_{\theta}:

mux⁡θH^θ=∑θ∣θ⟩⟨θ∣a⊗H^θ.(20)\operatorname{mux}_{\theta}\widehat{H}_{\theta} = \sum_{\theta} \lvert\theta\rangle\langle\theta\rvert_{a} \otimes\widehat{H}_{\theta}. \tag*{(20)}

Throughout this paper, we will use the notation mux⁡iOi=∑i∣i⟩⟨i∣⊗Oi\operatorname{mux}_{i} O_{i} = \sum_{i}\lvert i\rangle\langle i\rvert\otimes O_{i} for various elements OiO_{i} and indices ii in some auxiliary registers.

In this section, we define the SVD-THC circuit, a quantum program for mux⁡θB[H^θ/λ]\operatorname{mux}_{\theta}\mathcal{B}[\widehat{H}_{\theta}/\lambda], the multiplexed block encoding B[H^θ/λ]\mathcal{B}[\widehat{H}_{\theta}/\lambda] of the Hamiltonians H^θ\widehat{H}_{\theta} normalized by a universal 1-norm λ\lambda. Using the special structure of H^θ\widehat{H}_{\theta}, the SVD-THC circuit only multiplexes over θ\theta values for the preparation of an auxiliary state ∝∑ℓσℓ∣Vℓ∣∣ℓ⟩\propto\sum_{\ell}\sqrt{\sigma_{\ell}\lvert V_{\ell}\rvert}\lvert\ell\rangle. In an implementation for mux⁡ℓB[Hℓ]\operatorname{mux}_{\ell}\mathcal{B}[H_{\ell}], we compress the block encodings B[Hℓ]\mathcal{B}[H_{\ell}] of the SVD Hamiltonians HℓH_{\ell} with BLISS-THC. The objective is for the SVD-THC circuit to be cheaper than a circuit mux⁡θB[H^θ′/λmax⁡]\operatorname{mux}_{\theta}\mathcal{B}[\widehat{H}'_{\theta}/\lambda_{\max}], where the Hamiltonians H^θ′=H(θ)−Eθ\widehat{H}'_{\theta} = H(\theta) - \mathcal{E}_{\theta} based on the uncompressed system in Eq. (2) are block-encoded with BLISS-THC using its own universal 1-norm λmax⁡\lambda_{\max}, and where the multiplexor over θ\theta is implemented through slicing, i.e., through a sequence of gates

∣θ⟩⟨θ∣a⊗B[H^θ′/λmax⁡]+∑θ′≠θ∣θ′⟩⟨θ′∣a⊗1(21)\lvert\theta\rangle\langle\theta\rvert_{a} \otimes\mathcal{B}[\widehat{H}'_{\theta}/\lambda_{\max}] + \sum_{\theta'\ne\theta}\lvert\theta'\rangle\langle\theta'\rvert_{a} \otimes1 \tag*{(21)}

for θ=0…(M−1)\theta= 0\ldots(M-1). This sliced-THC algorithm will serve as a baseline for the compression achieved using SVD-THC. Next, we will develop the circuit of SVD-THC, focusing first on the broad strokes including the preparation of the auxiliary state, before turning our attention to the implementation of mux⁡ℓB[Hℓ]\operatorname{mux}_{\ell}\mathcal{B}[H_{\ell}].

SVD-THC circuit

The SVD-THC circuit, our compressed implementation of the ensemble block encoding mux⁡θB[H^θ/λ]\operatorname{mux}_{\theta} B[\hat{H}_{\theta}/\lambda], is depicted in Figure 1(a). Within this algorithm, a state preparation routine prep⁡o,θ\operatorname{prep}_{o,\theta} encodes features of the Hamiltonians H^θ\hat{H}_{\theta} in a log⁡(K+1)\log(K+1)-qubit register b, a single qubit m, as well as a garbage register g. A comprehensive list of all quantum registers in this paper can be found in Table I. In the ∣θ⟩\lvert\theta\rangle subspace of register a, prep⁡o,θ\operatorname{prep}_{o,\theta} prepares the state

RegisterDescription
aAngle register of O(log⁡M)O(\log M) qubits, holding the index ∣θ⟩\lvert\theta\rangle.
bSVD register, holding the index ∣ℓ⟩\lvert\ell\rangle in O(log⁡K)O(\log K) qubits.
fRegister holding a phase gradient state for catalytic rotation synthesis.
gAuxiliary register for the outer state preparation prep⁡o,θ\operatorname{prep}_{o,\theta}.
jAuxiliary register for the inner state preparation prep⁡i,ℓ\operatorname{prep}_{i,\ell}.
mTemporary qubit encoding the signs of Vℓ(θ)V_{\ell}(\theta).
qAuxiliary register for the THC circuit.
q1_1Temporary qubit switching between 1- and 2-body operators in the THC circuit.
q4_4, q5_5Auxiliary registers of log⁡(max⁡ℓRℓ)\log(\max_{\ell} R_{\ell}) qubits each, holding the THC indices ∣μ⟩\lvert\mu\rangle, ∣ν⟩\lvert\nu\rangle respectively.
s2N2N-qubit register representing spin-orbitals of the molecular system.
uAuxiliary qubit in the Generalized Quantum Signal Processing [39] routine.
vFlag qubit padding the block encoding B[H^θ/λ]B[\hat{H}_{\theta}/\lambda] to the 1-norm λ\lambda.
wFlag qubit for the thresholding circuit of Figure 7.
xUnion of all permanent qubits of b, g, j, q as well as v.

Table I. Comprehensive list of all quantum registers encountered in this work.

Figure 1. (a) SVD-THC circuit for mux⁡θB[H^θ/λ]\operatorname{mux}_{\theta} \mathcal{B}[\hat{H}_{\theta}/\lambda], with respect to angle indices θ\theta in register a. The subroutine labeled ‘ℓ=K\ell= K’ flips a qubit v in the ∣K⟩\lvert K\rangle subspace of register b. Diamond-shaped controls are our notation for general gate inputs. The Z gate on m encodes a phase. All quantum registers are listed in Table I. (b) Outer state preparation. A θ\theta-multiplexed version of Alias Sampling [40]. After a uniform state preparation USP on the b register, a joint index x is computed in order to load index aliases alt⁡x\operatorname{alt}_{x} and coefficient data keep⁡x\operatorname{keep}_{x}, along with signs sign⁡x\operatorname{sign}_{x} and their aliases sign⁡x′\operatorname{sign}'_{x} efficiently. To this end, within the ∣ℓ⟩\lvert\ell\rangle subspace of b, a temporary register is initialized in ∣ℓM⟩\lvert\ell M\rangle, and the θ\theta values in a are added to it. In-place adders (subtractors) add (subtract) the values of the registers under the gates labeled ‘+’ (‘−’) to (from) the values in the registers under the gates labeled ‘=’. Hexagon-shaped gates denote data-loader inputs. The gate labeled ‘a<ba < b’ is a comparator, flipping a bit depending on the values in g2g_{2} and g3g_{3}. ‘H’ denotes a tower of Hadamard gates on all qubits in register g3g_{3}. From this circuit we learn that all sub-registers of g except g3g_{3}, as well as qubit m can be considered temporary. (c) THC circuits multiplexed over the ℓ\ell index in b, featuring a modified Alias Sampling routine prep⁡i,ℓ\operatorname{prep}_{i,\ell} that uses a junk register j to prepare a block encoding register q for Select⁡ℓ\operatorname{Select}_{\ell}.

prep⁡o,θ∣0⟩bgm=∑ℓ=0K∣ξθ,ℓ∣∣ℓ⟩b∣Yθ,ℓ⟩g∣sign⁡θ,ℓ⟩m(22)\operatorname{prep}_{o,\theta}\lvert0\rangle_{bgm} = \sum_{\ell=0}^{K} \sqrt{\lvert\xi_{\theta,\ell}\rvert}\lvert\ell\rangle_b\lvert Y_{\theta,\ell}\rangle_g\lvert\operatorname{sign}_{\theta,\ell}\rangle_m \tag*{(22)}

in which ∣ℓ⟩\lvert\ell\rangle are computational basis states, ∣Yθ,ℓ⟩\lvert Y_{\theta,\ell}\rangle are irrelevant garbage states and bits sign⁡θ,ℓ∈{0,1}\operatorname{sign}_{\theta,\ell} \in\{0,1\} encode signs of some coefficients ξθ,ℓ\xi_{\theta,\ell}, such that

(−1)sign⁡θ,ℓ∣ξθ,ℓ∣=ξθ,ℓ.(23)(-1)^{\operatorname{sign}_{\theta,\ell}}\lvert\xi_{\theta,\ell}\rvert= \xi_{\theta,\ell}. \tag*{(23)}

The coefficients themselves are chosen to match the SVD parameters of the Hamiltonian in Eq. (18) via

ξθ,ℓ=σℓVℓ(θ)λ,(24)\xi_{\theta,\ell} = \frac{\sigma_{\ell}V_{\ell}(\theta)}{\lambda}, \tag*{(24)}

for ℓ∈[0,K−1]\ell\in[0,K-1]. Since a Hamiltonian ∑ℓξθ,ℓHℓ\sum_{\ell}\xi_{\theta,\ell}H_{\ell} may have a different 1-norm at every value of θ\theta due to the θ\theta-dependent outcome of the sum ∑ℓσℓ∣Vℓ(θ)∣\sum_{\ell}\sigma_{\ell}\lvert V_{\ell}(\theta)\rvert, we use the subspace ∣K⟩\lvert K\rangle of register b as padding. With what is essentially a large Toffoli gate in the center of Figure 1(a), we flip a qubit v in the subspace of ℓ=K\ell=K, thus removing the amplitude ξθ,K\xi_{\theta,K} from the operator encoded in the ∣0⟩\lvert0\rangle subspace of m, v, g, as well as the auxiliary registers j and q. The Hamiltonian

⟨0∣B[H^θ/λ]∣0⟩gjmqv=∑ℓ=0K−1ξθ,ℓHℓ(25)\langle0\lvert B[\hat{H}_{\theta}/\lambda]\lvert0\rangle_{gjmqv} = \sum_{\ell=0}^{K-1}\xi_{\theta,\ell}H_{\ell} \tag*{(25)}

is therefore subnormalized by the θ\theta-dependent factor, 1−∣ξθ,K∣1 - |\xi_{\theta,K}|. There is some value θ∗\theta^* of θ\theta that realizes the maximum 1-norm for which no padding is needed such that the ∣K⟩|K\rangle subspace of b vanishes: ξθ∗,K=0\xi_{\theta^*,K} = 0. The 1-norm of the SVD-THC is thus,

λ=max⁡θ∑ℓ=0K−1σℓ∣Vℓ(θ)∣=∑ℓ=0K−1σℓ∣Vℓ(θ∗)∣.(26)\lambda= \max_{\theta} \sum_{\ell=0}^{K-1} \sigma_\ell|V_\ell(\theta)| = \sum_{\ell=0}^{K-1} \sigma_\ell|V_\ell(\theta^*)| . \tag*{(26)}

The circuit for mux⁡θprepo⁡θ\operatorname{mux}_{\theta}\operatorname{prepo}_{\theta}, depicted in Figure 1(b), is a multiplexed version of quantum Alias Sampling. In the ∣θ⟩|\theta\rangle-subspace of register a, Alias Sampling subjects the uniform state ∝∑ℓ∣ℓ⟩\propto\sum_{\ell}| \ell\rangle over all the indices ℓ=0,…,K\ell= 0, \ldots, K, to a subroutine that retains only the (keep⁡θ,ℓ)(\operatorname{keep}_{\theta,\ell})-portion of each index ∣ℓ⟩|\ell\rangle and assigns the rest to an alias index alt⁡θ,ℓ\operatorname{alt}_{\theta,\ell}. Here keep⁡θ,ℓ∈[0,1)\operatorname{keep}_{\theta,\ell} \in[0,1) represent fixed-point numbers, while alt⁡θ,ℓ∈[0,K]\operatorname{alt}_{\theta,\ell} \in[0,K] are integers. In the process, it also sets the sign qubit m with the values sign⁡θ,ℓ\operatorname{sign}_{\theta,\ell} and sign⁡θ,ℓ′\operatorname{sign}'_{\theta,\ell} in the kept and re-assigned subspaces, respectively. Each index ∣ℓ⟩|\ell\rangle turns into

∣ℓ⟩b∣0⟩g∣0⟩m↦keep⁡θ,ℓ∣ℓ⟩b∣Zθ,ℓ,0⟩g∣sign⁡θ,ℓ⟩m+1−keep⁡θ,ℓ∣alt⁡θ,ℓ⟩b∣Zθ,ℓ,1⟩g∣sign⁡θ,ℓ′⟩m,(27)|\ell\rangle_b|0\rangle_g|0\rangle_m \mapsto\sqrt{\operatorname{keep}_{\theta,\ell}}|\ell\rangle_b|Z_{\theta,\ell,0}\rangle_g|\operatorname{sign}_{\theta,\ell}\rangle_m + \sqrt{1-\operatorname{keep}_{\theta,\ell}}|\operatorname{alt}_{\theta,\ell}\rangle_b|Z_{\theta,\ell,1}\rangle_g|\operatorname{sign}'_{\theta,\ell}\rangle_m , \tag*{(27)}

where ∣Zθ,ℓ,i⟩|Z_{\theta,\ell,i}\rangle are garbage states. The values of alt, keep, sign and sign′ are chosen appropriately to encode the coefficients ξθ,ℓ\xi_{\theta,\ell} and signs in Eq. (22). Within the mux⁡θprepo⁡θ\operatorname{mux}_{\theta}\operatorname{prepo}_{\theta} circuit, all these values are loaded into registers g and m from the index pairs in registers a and b. This double-indexing can cause a cost overhead in data-loading, which forces us to find a joint index x(ℓ,θ)x(\ell,\theta). Since we are expecting neither K+1K+1 nor MM to be an integer power of two, i.e., a number 2k2^k with integer kk, the data-loader in the circuit of Figure 1(b) would have an O(Klog⁡M)O(K \log M) overhead of Toffoli gates (left elbows).

QROM circuits have a Toffoli cost linear in the number of payload items only when they load from a continuous index [40, 41]. When loading from a and b as if they were one giant register, the QROM would have to iterate over binary numbers θ⊕ℓ\theta\oplus\ell, formed by appending indices θ\theta and ℓ\ell, which are not continuous. Since the integers (M−1)⊕ℓ(M-1) \oplus\ell and 0⊕(ℓ+1)0 \oplus(\ell+1) do not differ by 1, every interface between pairs of adjacent indices (M−1)⊕ℓ(M-1) \oplus\ell and 0⊕(ℓ+1)0 \oplus(\ell+1) incurs O(log⁡M)O(\log M) Toffoli gates in the unitary iteration procedure.

With the data-loader already being the dominant cost of mux⁡θprepo⁡θ\operatorname{mux}_{\theta}\operatorname{prepo}_{\theta}, we want to avoid this scenario by joining ℓ\ell and θ\theta into a continuous index x(ℓ,θ)=Mℓ+θx(\ell,\theta) = M\ell+ \theta. To this end we use a trick from Ref. [27], and first load the basis of the more significant index in xx – meaning here the ℓ\ell-dependent part of xx – into a temporary register. With the basis MℓM\ell being linear in ℓ\ell, this temporary register may be initialized in ∣Mℓ⟩|M\ell\rangle by a multiplication circuit rather than a data-loader. In the second step, the joint index ∣x(ℓ,θ)⟩|x(\ell,\theta)\rangle is finalized by adding the value of the less significant index, here θ\theta, into the temporary register, simply done with an in-place adder originating from register a. Basis-loading and addition have complexity O(K+log⁡M)O(K+\log M), which is preferable to the O(Klog⁡M)O(K\log M) overhead incurred by loading alt, keep, sign and sign′ values from a discontinuous index.

Multiplexed THC circuit

The SVD-THC circuit still contains a block encoding multiplexer mux⁡ℓB[Hℓ]\operatorname{mux}_{\ell}\mathcal{B}[H_\ell], and while we expect this subroutine, depicted in Figure 1(c), to be the most expensive part of the entire circuit, its complexity should be greatly reduced compared to the sliced-THC algorithm.

For one, we are multiplexing THC Hamiltonians over the index ℓ\ell which has a smaller range than θ\theta. One can hope for a compression when K<MK < M assuming comparable THC ranks and parameters.

We also do not want to multiplex B[Hℓ]\mathcal{B}[H_\ell] by slicing. Instead, we modify the THC circuit of Ref. [33], such that we share subroutines such as the Givens rotations between all block encodings B[Hℓ]\mathcal{B}[H_\ell], and simply extend internal data-loaders by another index ℓ\ell.

One challenge here is to account for Hamiltonians HℓH_\ell having individual THC ranks. The THC rank RℓR_\ell of HℓH_\ell is a compression parameter of the block encoding. In fact, a leading factor in the complexity of the THC circuit is O(poly⁡(Rℓ))O(\operatorname{poly}(R_\ell)). In this paper, we construct a multiplexed THC circuit, such that Hamiltonians with smaller rank contribute less to the overall complexity, striving for a scaling of ∑ℓpoly⁡(Rℓ)\sum_\ell\operatorname{poly}(R_\ell) rather than K×poly⁡(max⁡ℓRℓ)K \times\operatorname{poly}(\max_\ell R_\ell).

The THC circuits encode the Hamiltonians

Hℓ=18∑μ=0Rℓ−1∑ν=0ν≠μRℓ−1ζℓμν∑σ,τ=↑,↓Zσ[ϑ^ℓμ]Zτ[ϑ^ℓν](28)H_\ell= \frac{1}{8}\sum_{\mu=0}^{R_\ell-1}\sum_{\substack{\nu=0\\ \nu\ne\mu}}^{R_\ell-1}\zeta_{\ell\mu\nu}\sum_{\sigma,\tau=\uparrow,\downarrow} Z_\sigma[\hat{\vartheta}_{\ell\mu}]Z_\tau[\hat{\vartheta}_{\ell\nu}] \tag*{(28)}
+12∑μ=0Rℓ−1ζℓμμ∑σ,τ=↑,↓σ≠τZσ[ϑ^ℓμ]Zτ[ϑ^ℓμ](29)+\frac{1}{2}\sum_{\mu=0}^{R_\ell-1}\zeta_{\ell\mu\mu}\sum_{\substack{\sigma,\tau=\uparrow,\downarrow\\ \sigma\ne\tau}} Z_\sigma[\hat{\vartheta}_{\ell\mu}]Z_\tau[\hat{\vartheta}_{\ell\mu}] \tag*{(29)}
+12∑μ=0N−1ζℓμRℓ∑σ=↑,↓Zσ[ϑ~ℓμ](30)+\frac{1}{2}\sum_{\mu=0}^{N-1}\zeta_{\ell\mu R_\ell}\sum_{\sigma=\uparrow,\downarrow}Z_\sigma[\widetilde{\vartheta}_{\ell\mu}] \tag*{(30)}

with normalized ‘machine’ coefficients ζℓμν\zeta_{\ell\mu\nu} such that ∑μ≤ν∣ζℓμν∣=1\sum_{\mu\le\nu}|\zeta_{\ell\mu\nu}|=1, and where ζℓμν=ζℓνμ\zeta_{\ell\mu\nu}=\zeta_{\ell\nu\mu} for μ,ν<Rℓ\mu,\nu<R_\ell. The Hamiltonian features rotated occupation operators

Zσ[ϑ]=Gσ†(ϑ)(1−2a0,σ†a0,σ)Gσ(ϑ)(31)Z_\sigma[\vartheta]=G_\sigma^\dagger(\vartheta)\left(1-2a^\dagger_{0,\sigma}a_{0,\sigma}\right)G_\sigma(\vartheta) \tag*{(31)}

parameterized by sets of angles ϑ=ϑ0,ϑ1,⋯ ,ϑN−2\vartheta=\vartheta_0,\vartheta_1,\cdots,\vartheta_{N-2}, and Givens rotations

Gσ(ϑ)=Gσ,0(ϑ0)⋅Gσ,1(ϑ1)⋯Gσ,N−2(ϑN−2)(32)G_\sigma(\vartheta)=G_{\sigma,0}(\vartheta_0)\cdot G_{\sigma,1}(\vartheta_1)\cdots G_{\sigma,N-2}(\vartheta_{N-2}) \tag*{(32)}

with

Gσ,k(ϑk)=exp⁡[ϑk2(ak+1,σ†ak,σ−ak,σ†ak+1,σ)].(33)G_{\sigma,k}(\vartheta_k)=\exp\left[\frac{\vartheta_k}{2}\left(a^\dagger_{k+1,\sigma}a_{k,\sigma}-a^\dagger_{k,\sigma}a_{k+1,\sigma}\right)\right]. \tag*{(33)}

The machine coefficients ζℓμν\zeta_{\ell\mu\nu} and angles ϑ^ℓμ\hat{\vartheta}_{\ell\mu}, ϑ~ℓμ\widetilde{\vartheta}_{\ell\mu} are matched with the parameters t~μ(ℓ)\widetilde{t}^{(\ell)}_\mu and ζ~μν(ℓ)\widetilde{\zeta}^{(\ell)}_{\mu\nu} and χkℓμ\chi^{\ell\mu}_k from Section III as

ζℓμν=1σℓ{−t~μ(ℓ)for ν=Rℓ, μ<N;ζ~μν(ℓ)for ν∈[0,Rℓ), μ∈[0,ν];ζ~μμ(ℓ)/4for ν∈[0,Rℓ), μ=ν;0else,(34)\zeta_{\ell\mu\nu}=\frac{1}{\sigma_\ell} \begin{cases} -\widetilde{t}^{(\ell)}_\mu& \text{for } \nu=R_\ell,\ \mu<N;\\ \widetilde{\zeta}^{(\ell)}_{\mu\nu} & \text{for } \nu\in[0,R_\ell),\ \mu\in[0,\nu];\\ \widetilde{\zeta}^{(\ell)}_{\mu\mu}/4 & \text{for } \nu\in[0,R_\ell),\ \mu=\nu;\\ 0 & \text{else}, \end{cases} \tag*{(34)}

and

χkℓμ=cos⁡(2ϑ^ℓμk)∏m<ksin⁡(2ϑ^ℓμm)(35)\chi^{\ell\mu}_k=\cos(2\hat{\vartheta}_{\ell\mu k})\prod_{m<k}\sin(2\hat{\vartheta}_{\ell\mu m}) \tag*{(35)}

for k∈[0,N−2]k \in[0, N-2] as well as χN−1ℓμ=∏ksin⁡(2ϑ~ℓμk)\chi_{N-1}^{\ell\mu} = \prod_k \sin(2\tilde{\vartheta}_{\ell\mu k}), where ϑ~ℓμk\tilde{\vartheta}_{\ell\mu k} are the components of ϑ~ℓμ\tilde{\vartheta}_{\ell\mu}.

The parameters ϑ\vartheta split into two groups: a group of angles ϑ~ℓμ\tilde{\vartheta}_{\ell\mu} that encode two-body operators, and the groups of angles ϑ~eℓμ\tilde{\vartheta}_{e\ell\mu} encoding one-body operators. In addition to the superindex ℓ\ell, two-body angles depend on THC indices μ\mu and ν\nu in the range from 0 to Rℓ−1R_\ell-1, while one-body angles depend on orbital indices in the range of [0,N−1][0,N-1].

The block encoding circuit is decomposed into an inner prep and Select circuit:

B[Hℓ]=prep⁡i,ℓ†Select⁡ℓprep⁡i,ℓ(36)\mathcal{B}[H_\ell] = \operatorname{prep}_{i,\ell}^{\dagger}\operatorname{Select}_{\ell}\operatorname{prep}_{i,\ell} \tag*{(36)}

A circuit for prep⁡i,ℓ\operatorname{prep}_{i,\ell} is presented in Appendix B. While somewhat similar to the multiplexed Alias Sampling of prep⁡o,ϑ\operatorname{prep}_{o,\vartheta}, the inner prep has some more intricate design choices such as a multiplexed USP. In the ∣ℓ⟩\lvert\ell\rangle subspace of b, prep⁡i,ℓ\operatorname{prep}_{i,\ell} prepares registers j and q in the state

prep⁡i,ℓ∣0⟩jq=∑ν=0Rℓ∑μ≤ν∣ζℓμν∣∣μ⟩q4∣ν⟩q5∣Ωℓμν⟩(37)\operatorname{prep}_{i,\ell}\lvert0\rangle_{jq} = \sum_{\nu=0}^{R_\ell}\sum_{\mu\leq\nu}\sqrt{\lvert\zeta_{\ell\mu\nu}\rvert}\lvert\mu\rangle_{q_4}\lvert\nu\rangle_{q_5}\lvert\Omega_{\ell\mu\nu}\rangle \tag*{(37)}

with the main indices ∣μ⟩\lvert\mu\rangle and ∣ν⟩\lvert\nu\rangle encoded in registers q₄ and q₅, respectively. The state ∣Ωℓμν⟩\lvert\Omega_{\ell\mu\nu}\rangle encodes additional information: a success flag on qubit q₀, a flag for the one-body terms in the case ν=Rℓ\nu=R_\ell on qubit q₁, the sign of ζℓμν\zeta_{\ell\mu\nu} in qubit q₂, as well as a junk state on register j.

The multiplexed Select routine, depicted in Figure 2, encodes two-body operators

Figure 2. Quantum circuit for mux⁡ℓSelect⁡ℓ\operatorname{mux}_{\ell}\operatorname{Select}_{\ell}. This circuit loads a set of different rotation angles ϑp\vartheta_p and ϑq\vartheta_q that serve as the input for sequences of Givens rotations (orange) implementing the rotated operators mux⁡pZ[ϑp]\operatorname{mux}_p Z[\vartheta_p] and mux⁡qZ[ϑq]\operatorname{mux}_q Z[\vartheta_q] respectively, using in-place adders/subtractors into a phase gradient state in register ff. Here p(ℓ,ν)p(\ell,\nu) and q(ℓ,b,μ)q(\ell,b,\mu) are contiguized numbers from indices (ℓ,b,μ,ν)(\ell,b,\mu,\nu) in registers (b,q1,q4,q5)(b,q_1,q_4,q_5). Since the value ℓ\ell in bb determines the range of μ\mu and ν\nu, we load ℓ\ell-dependent bases QℓQ_\ell and PℓP_\ell into respective temporary registers, which then receive values of μ+bˉN\mu+\bar{b}N (blue) and ν\nu (green) via in-place addition, where bˉ\bar{b} is the opposite value of bb. Note that the value b=1b=1 in q1q_1 flags the subspace in which angles for one-body terms must be loaded. The remainder of the circuit follows the construction in [33], with qubit q2q_2 holding information about the sign of the encoded operator, q3q_3 is a sort of coin qubit making the circuit self-inverse, q6q_6 flags the case where μ=ν\mu=\nu after a comparator while q7q_7 and q8q_8 switch the spin-up s↑s_{\uparrow} and spin-down parts s↓s_{\downarrow} of the simulator register.

Zσ[ϑ~ℓμ]Zτ[ϑ~ℓν]+Zσ[ϑ~ℓν]Zτ[ϑ~ℓμ](38)Z_\sigma[\tilde{\vartheta}_{\ell\mu}]Z_\tau[\tilde{\vartheta}_{\ell\nu}] + Z_\sigma[\tilde{\vartheta}_{\ell\nu}]Z_\tau[\tilde{\vartheta}_{\ell\mu}] \tag*{(38)}

in the ∣μ⟩∣ν⟩\lvert\mu\rangle\lvert\nu\rangle subspace of q₄ and q₅ for μ,ν∈[0,Rℓ)\mu,\nu\in[0,R_\ell). The circuit itself has two parts: the first half of the circuit encodes operator mux⁡ℓμZσ[ϑ~ℓμ]\operatorname{mux}_{\ell\mu}Z_\sigma[\tilde{\vartheta}_{\ell\mu}] with μ\mu in q₄ and ℓ\ell in b, while the second half implements mux⁡ℓνZσ[ϑ~ℓν]\operatorname{mux}_{\ell\nu}Z_\sigma[\tilde{\vartheta}_{\ell\nu}] with ν\nu in q₅. Within the ∣ℓ⟩∣ν⟩\lvert\ell\rangle\lvert\nu\rangle subspace of their respective registers, the angles ϑ~ℓν\tilde{\vartheta}_{\ell\nu} are loaded into a temporary register t, before the orange gates in Figure 2 encode the Givens rotations Gσ(i)(ϑ~ℓν)G_\sigma^{(i)}(\tilde{\vartheta}_{\ell\nu}) by continuously adding different cells of t into a phase gradient state in f [42]. This procedure applies to both halves of the circuit, but the data-loading of the second half is more straightforward. Rather than loading data ϑ~ℓν\tilde{\vartheta}_{\ell\nu} from the double index ℓ⊕ν\ell\oplus\nu, we will load the same angles ϑ~p\tilde{\vartheta}_p from a continuous index pp of the form

p(ℓ,ν)=Pℓ+ν(39)p(\ell,\nu)=P_\ell+\nu \tag*{(39)}

where the basis Pℓ=∑m<ℓRmP_\ell=\sum_{m<\ell}R_m is loaded from b into another temporary register and the index ν\nu from q₅ is added to it. The corresponding gates are highlighted green in Figure 2. Note that ν\nu could be out of bounds: although the joint index pp is designed for ν\nu to not exceed the ℓ\ell-dependent values of Rℓ−1R_\ell-1, the state in (37) allows for the indices ν=Rℓ\nu=R_\ell. However, in this case the bit q₁ will be set, and the ZZ operator between the Givens rotations is not applied, making the angles of the Givens rotations irrelevant. In this ∣1⟩∣ν⟩∣Rℓ⟩\lvert1\rangle\lvert\nu\rangle\lvert R_\ell\rangle subspace of the qubits (q1,q4,q5)(q_1,q_4,q_5), the second half of the multiplexed Select circuit is switched off entirely, and its first part encodes the one-body operators mux⁡ℓμZσ[ϑ~eℓμ]\operatorname{mux}_{\ell\mu}Z_\sigma[\tilde{\vartheta}_{e\ell\mu}] for μ∈[0,N)\mu\in[0,N). For the first half of the circuit this means also that, in addition to loading angles ϑ~ℓμ\tilde{\vartheta}_{\ell\mu} in the ∣0⟩\lvert0\rangle subspace of q₁, angles ϑ~eℓμ\tilde{\vartheta}_{e\ell\mu} need to be loaded in the ∣1⟩\lvert1\rangle subspace of q₁. We combine ϑ~\tilde{\vartheta} and ϑ~e\tilde{\vartheta}_e into a triple-indexed set of parameters ϑℓbμ\vartheta_{\ell b\mu}, where ∣b⟩\lvert b\rangle is the index of q₁. The values themselves will be loaded as ϑq\vartheta_q from a joint index qq with

q(ℓ,b,μ)=Qℓ+bˉN+μ(40)q(\ell,b,\mu)=Q_\ell+\bar{b}N+\mu \tag*{(40)}

where bˉ=b+1(mod2)\bar{b}=b+1\pmod{2}, and Qℓ=ℓN+∑m<ℓRmQ_{\ell}=\ell N+\sum_{m<\ell}R_m. The gates involved in the computation of the joint index qq are highlighted light blue in Figure 2, where we XOR the number NN into a temporary register, and then perform an in-place addition from register q4q_4. The resulting values are added to a different temporary register, into which the values of the basis QℓQ_{\ell} have been loaded from bb. The SVD-THC compression and circuit are deployed to molecular systems in the next sections.

Compression results

In this section we present numerical results for three benchmark molecular systems, highlighting both the efficacy of the compressed block encoding and the challenges associated with it. We first introduce the molecules, the procedure used to generate their geometries, and the resulting potential energy surfaces. We then define the accuracy metric used throughout, examining how the compression error converges with the number of retained SVD modes. Next we compare the effective 1-norm of each mode with the bare norm that the factorization penalizes.

Molecular systems and geometries

We evaluate the parameterized compression scheme on three molecules. The first is 1-bromo-2-chloroethane (BCE), scanned over the Cl–C–C–Br torsion φ\varphi on a 73-point grid. The second is the glycine dipeptide, scanned over the backbone dihedrals (φ,ψ)(\varphi,\psi) on a 37×7237 \times72 grid. The third is the alanine dipeptide, scanned over the two backbone dihedrals of a 72×7272 \times72 grid. The BCE torsion and the φ\varphi axis of glycine dipeptide are sampled at Chebyshev-Gauss-Lobatto points on [0∘,180∘][0^\circ,180^\circ], which cluster near the endpoints, and all other axes uniformly in steps of 5∘5^\circ.

All molecular geometries were optimized using the g-xTB tight-binding method based on the xtb program [43, 44], with harmonic restraints holding the scanned dihedrals and all other coordinates relaxed to very tight convergence. Figure 3 shows the resulting potential energy surfaces. All correlated energies use the STO-3G basis. We freeze the core orbitals of a restricted Hartree-Fock (RHF) calculation and build the active space from the remaining valence atomic orbitals, projected against the core and Löwdin orthogonalized [45]. This gives the frozen-core active spaces (26e, 20o) for BCE, (52e, 46o) for glycine dipeptide and (58e, 52o) for alanine dipeptide, in which correlation energies are computed with CCSD. Freezing the core affects relative energies along a torsion negligibly, since core orbitals are spatially compact and barely respond to the rotation. The minimal basis set remains the dominant approximation, and is simply used as a proof-of-concept.

Potential energy surfaces for relaxed and rigid BCE, glycine dipeptide, and alanine dipeptide

Figure 3. Potential energy surfaces (g-xTB) for (a) relaxed and rigid BCE, as well as the (b) glycine and (c) alanine dipeptides vs the backbone (φ\varphi, ψ\psi), on respectively 37 ×\times 72 and 72 ×\times 72 grids. Each panel carries an inset showing the molecule and the coordinate being scanned. Panel (a) shows a filmstrip of ten scanned geometries. Panels (b) and (c) show the two backbone dihedrals as colored dashed lines over the four atoms that define them; all remaining internal coordinates are drawn in gray and relaxed at every grid point.

Both BCE and the glycine dipeptide carry a mirror symmetry that allows us to halve the sampled domain. Rotating φ\varphi to −φ-\varphi produces the mirror image of BCE, resulting in E(φ)=E(−φ)E(\varphi)=E(-\varphi). The physical range therefore requires only φ∈[0∘,180∘]\varphi\in[0^\circ,180^\circ], which is the domain of Figure 3(a) and of the 73-point grid in Table III. Glycine dipeptide has no side chain, so its CαC_{\alpha} atom carries two equivalent hydrogens rather than a stereocenter. Its Ramachandran surface therefore possesses the analogous mirror symmetry E(φ,ψ)=E(−φ,−ψ)E(\varphi,\psi)=E(-\varphi,-\psi). Since ψ\psi is sampled over its full range, the point at (−φ,ψ)(-\varphi,\psi) coincides with the already sampled point at (φ,−ψ)(\varphi,-\psi), so only φ\varphi must be halved. This gives the 37×7237 \times72 grid of Table III rather than 72×7272 \times72. The methyl side chain of alanine dipeptide makes its CαC_{\alpha} a stereocenter and breaks this symmetry, so its 72×7272 \times72 grid samples the full surface in both angles.

systemgridMMKKλ\lambda (Ha)medianmaxmedianmax
1-bromo-2-chloroethane73×173 \times1731453.80.0530.4250.2330.541
glycine dipeptide37×7237 \times722,66490339.90.1020.5170.2120.589
alanine dipeptide72×7272 \times725,184150416.40.0740.8240.1840.569

Table III. Summary of the three block encodings: the torsional grid and number of geometries, the number of retained SVD modes KK, the 1-norm λ\lambda (Hartree), the absolute ℓ2\ell_2 residual of the per-mode THC factorization (Hartree) summarized over the KK modes, and the absolute error in the CCSD correlation energy (mHa) summarized over all geometries of the grid. No geometry exceeds the target error of 0.6 mHa for any of the three systems.

For BCE, we compute both a relaxed scan, in which every other internal coordinate is re-optimized at each value of φ\varphi, and a rigid scan, in which only φ\varphi is rotated while all other coordinates remain fixed at their global-minimum configuration values. Relaxation represents the physical scenario where the molecule can adjust its bond lengths, bond angles, and remaining torsions in response to the changing dihedral. A rigid scan prevents these adjustments and therefore overestimates the energy away from the minimum. Figure 3(a) shows that this effect is substantial. The relaxed barrier is 8.1 kcal/mol, compared with 18.7 kcal/mol for the rigid scan. Evaluating CCSD in the active space on the same geometries gives barriers of 5.0 and 18.7 kcal/mol, respectively. The two levels of theory agree closely for the rigid scan but differ substantially once relaxation is allowed.

The focus for this manuscript is primarily on relaxed potential energy surfaces because they provide an effective low-dimensional description of the energy landscape and they most closely follow the practical workflows in computational drug design. Once the remaining internal coordinates have been minimized at each value of φ\varphi, their energetic response is absorbed into the exposed coordinates rather than being represented explicitly. This reduction becomes especially valuable for higher-dimensional scans, where optimizing secondary coordinates can eliminate dimensions that would otherwise require additional Hamiltonian parameters. Relaxation can therefore simplify the overall representation, but it makes the associated integrals harder to compress, as illustrated by Figure 4(a). Reduced compressibility can also occur due to near-avoided crossings, competing minima, or changes in the identity of the optimized structure, where the minimizing geometry may vary rapidly or develop a cusp. For larger molecules such as the glycine and alanine dipeptides, rigid scans could introduce collisions between different parts of the molecule, and we therefore report relaxed surfaces only for the dipeptides, since they provide the more physically relevant energy landscapes.

Correlation energy error versus retained SVD rank $K$ for BCE, glycine dipeptide, and alanine dipeptide

Figure 4. Correlation energy error versus retained SVD rank KK for (a) BCE including rigid and relaxed convergence, (b) relaxed glycine dipeptide, (c) relaxed alanine dipeptide. Gray dashed line is the 0.6 mHa target. Dashed lines of the same color give the median error over the grid, added as a reference. Every point is evaluated at all geometries of the full grid.

Finally, since the proposed compression scheme is applied to a discretized surface, it is also useful to record how much the energy changes between adjacent grid points. Table II reports this quantity for each full grid and its nested sub-grids, at both levels of theory. Apart from the 3 ×\times 3 alanine dipeptide grid, which is too coarse to resolve the surface, the mean step between neighboring points decreases monotonically as the grid is refined, in every system and at both levels, and the maximum step follows the same trend less regularly. Refinement therefore resolves the existing surface rather than exposing new structure. Since variations below the error tolerance cannot be distinguished from the compression error, we treat the finest grids as resolving the relevant structure of the surfaces.

gridMMminmeanmaxspanminmeanmaxspan
g-xTB (mHa)CCSD (mHa)
1-bromo-2-chloroethane (BCE)
73, 1D φ\varphi100.0212.4936.06212.950.0911.4963.5348.02
190.0091.3663.30012.950.0060.8091.8938.02
370.0010.6961.65212.950.0000.4090.9588.02
550.0000.4641.64812.950.0000.2730.9358.02
730.0000.3480.83612.950.0000.2060.4838.02
Glycine dipeptide
10×181800.0061.94114.19433.800.0051.3448.06715.23
19×183420.0011.3757.57434.120.0001.0365.45316.21
19×366840.0001.0037.58034.240.0000.7155.45316.21
37×361,3320.0000.7063.85534.380.0000.5393.60116.21
37×722,6640.0000.5093.85534.380.0000.3683.92416.28
Alanine dipeptide
3×391.6745.6058.5989.520.1301.8483.9845.49
9×9810.0023.87219.30233.610.0262.8479.75915.33
18×183240.0112.14712.70337.580.0011.6857.28116.39
36×361,2960.0011.1166.42337.580.0020.8844.28517.95
72×725,1840.0000.5653.24338.790.0000.4524.71318.94

Table 2. Energy landscape of the scanned geometries at two levels of theory. For each nested sub-grid, absolute energy differences between neighboring grid points are reported through the minimum, mean and maximum over all neighboring pairs, together with the energy span Emax⁡−Emin⁡E_{\max} - E_{\min}. Total energies are evaluated on the same geometries with g-xTB, and CCSD with a frozen core.

Accuracy metric

Let EcorrCCSD(⋅)E_{\mathrm{corr}}^{\mathrm{CCSD}}(\cdot) denote the CCSD correlation energy of a Hamiltonian, and let HSVDH_{\mathrm{SVD}} denote compressed approximations obtained at rank KK as in Eq. (18) while H(θn)H(\theta_n) are the unfactorized Hamiltonians of Eq. (2). We measure the error as the worst case over the full sampled grid,

ε∞(K)=max⁡n∣EcorrCCSD(HSVD(θn))−EcorrCCSD(H(θn))∣,(41)\varepsilon_{\infty}(K) = \max_{n} \left| E_{\mathrm{corr}}^{\mathrm{CCSD}}(H_{\mathrm{SVD}}(\theta_n)) - E_{\mathrm{corr}}^{\mathrm{CCSD}}(H(\theta_n)) \right|, \tag*{(41)}

evaluated at every grid point. We compare correlation energies since the exact Hartree-Fock energy at each geometry can be restored by modifying the constant term EθE_{\theta} in Eq. (18), ensuring that the energy reported by the circuit differs from the exact one only through its correlation energy. We target ε∞≤0.6\varepsilon_{\infty} \leq0.6 mHa, and refer to the value of ε∞(K)\varepsilon_{\infty}(K) at the chosen KK as the compression floor.

Overall, we target an accuracy of 1.6 mHa, corresponding to 1 kcal/mol, in the energy that SVD-THC reports at every geometry. We divide this total error budget between two sources of error introduced by the algorithm. Data compression and bit truncation are allotted 0.6 mHa, as imposed through Eq. (41) above as a worst case over all MM Hamiltonians of the ensemble. An error of ϵ=1\epsilon= 1 mHa is reserved as the resolution limit for quantum phase estimation tasks that we will introduce in Sec. VI, setting the complexity of those algorithms to N=O(λ/ϵ)\mathcal{N} = O(\lambda/\epsilon). Sliced-THC approaches are evaluated with the same error budget and are validated in the same way as SVD-THC, where each per-geometry BLISS-THC factorization is accepted at the smallest THC rank whose CCSD correlation energy error is below 0.6 mHa. Both sides of every comparison are therefore held to the same CCSD accuracy criterion.

Convergence and rank structure

Figure 4 shows the convergence of ε∞(K)\varepsilon_{\infty}(K) as well as the median energy error with respect to the number of retained SVD modes KK for the three systems. Figure 4(a) highlights the effect of relaxation on compressibility. The rigid BCE scan contains no relaxation-induced structure and reaches ε∞=2.3×10−3\varepsilon_{\infty} = 2.3 \times10^{-3} mHa by K=9K = 9, whereas the relaxed scan does not reach similar accuracy until K=24K = 24. For alanine dipeptide, ϵ∞\epsilon_\infty first falls below the target at K=100K = 100, stays between 0.53 and 0.59 mHa up to K=120K = 120, and drops to 0.22 mHa at K=130K = 130. Over this plateau the maximum is attained at a single geometry, whose error is removed only once modes 121 to 130 are retained. The curve is not strictly monotone, because truncating the SVD is not variational, and adding a mode can increase the error at an individual geometry. The figure also highlights how we select KK for each ensemble. Our approach consists of choosing a value that clears the 0.6 mHa target with some margin rather than at the crossing point itself. For each of these systems, we choose K=14K = 14, 90, and 150 respectively. The margin allows flexibility in the BLISS-THC compression step where more of the error budget can be spent on lowering the ranks RℓR_\ell which has a stronger effect on the total resource cost.

Figure 5 shows the two per-mode quantities that together set the cost of SVD-THC, corresponding to the THC rank RℓR_\ell at which each mode Hamiltonian is fitted, and its effective norm σℓ∣Vℓ(θ∗)∣\sigma_\ell|V_\ell(\theta^*)| at the geometry θ∗\theta^* that attains λ\lambda, as defined in Eq. (26). The modes are ordered by decreasing ωℓ\omega_\ell, so both quantities decay with the mode index. They do not, however, decay at the same rate, and they do not induce the same ordering. We observe a right-skewed rank profile, where the mean rank exceeds the median by up to 55% across the three systems, and the largest mode is fitted at two and a half to four times the average rank. This has consequences for the quantum resource estimates discussed later. Since the Clifford data-loading cost in the SVD-THC circuit scales quadratically with every rank as O(∑ℓRℓ2)O(\sum_\ell R_\ell^2), this short tail of high-rank modes dominates the QROAM cost rather than the number of modes KK itself. It is also why we allocate a different rank per mode, rather than holding it at a single common value. A uniform allocation at max⁡ℓRℓ\max_\ell R_\ell would pay the cost of the most demanding mode KK times over.

Per-mode structure of the three deployed SVD-THC encodings

Figure 5. Per-mode structure of the three deployed SVD-THC encodings of the relaxed molecules in Table III, plotted against the SVD mode index ℓ\ell, with modes ordered by decreasing ωℓ\omega_\ell. Top: THC rank RℓR_\ell, with the mean (dashed) and median (dotted) rank, 31 and 20 for BCE, 146 and 104 for glycine dipeptide, and 153 and 130 for alanine dipeptide. Bottom: effective norm σℓ∣Vℓ(θ∗)∣\sigma_\ell|V_\ell(\theta^*)| on a logarithmic scale, where θ∗\theta^* is the geometry that attains λ\lambda. The effective norms sum exactly to λ\lambda.

The effective norm decays considerably faster, falling by several orders of magnitude across the retained modes, and the three largest modes alone carry 52–92% of λ\lambda. For both dipeptides the first mode is both the most expensive to load and the largest contributor to λ\lambda, but beyond it the orderings generally disagree. In BCE, the highest-rank mode is only the third largest in effective norm. Ultimately, both are front-loaded, where it is observed that half of the QROAM cost is spent on the first 14–16% of the modes, while 90% of λ\lambda is carried by the first 21–24%.

Table III summarizes the decomposition parameters for the three SVD-THC ensembles considered: BCE at K=14K = 14, the glycine dipeptide at K=90K = 90, and the alanine dipeptide at K=150K = 150. The decomposition error for each ensemble is certified against Eq. (41) at every geometry of its full grid, with no failures. For the sliced-THC version of the BCE molecule, block encodings for both rigid and relaxed scans obtain similar accuracy with average THC rank of 51.5 and 51.6 and similar 1-norms averaging 35.8 and 36.4 Hartree, with worst cases of 41.9 and 50.5 Hartree. The SVD-THC technique replaces the sliced-THC encoding scheme with the compressed circuit assembled from either K=9K = 9 or K=14K = 14 tensors. For the relaxed scan, the circuit carries a larger 1-norm than any one single-geometry block encoding, 53.8 vs. 50.5 Hartree, a factor of 1.07, and for the rigid scan 42.7 vs. 41.9 Hartree, a factor of 1.02 is obtained.

Effective norm versus conventional norm

The SVD-THC factorization features two types of norms. The individual terms σℓ∣Vℓ(θ∗)∣\sigma_\ell|V_\ell(\theta^*)| for the norm-attaining geometry θ∗\theta^* are the effective norms plotted in the lower row of Figure 5, and sum to λ=∑ℓσℓ∣Vℓ(θ∗)∣\lambda= \sum_\ell\sigma_\ell|V_\ell(\theta^*)|. The bare THC norms σℓ\sigma_\ell that the factorization regularizes do not. The mean effective norm per mode is 3.84 Ha for BCE, 3.78 Ha for glycine dipeptide and 2.78 Ha for alanine dipeptide, whereas the corresponding mean σℓ\sigma_\ell is 28.2 Ha, 164.8 Ha and 177.7 Ha. Only the former sums to λ\lambda, setting the true rescaling factor for the block encoding.

Applying THC to an individual tensor normally has the target to minimize its 1-norm σℓ\sigma_\ell. In our algorithm, the tensors are always modulated by their coordinate-dependent factor ∣Vℓ(θ∗)∣|V_\ell(\theta^*)|. Figure 6 compares the effective norm with the bare 1-norm σℓ\sigma_\ell that the THC fit regularizes. The two are only loosely correlated, and the orderings they induce diverge most where the spread of ∣Vℓ(θ∗)∣|V_\ell(\theta^*)| is widest: the dipeptides give R2≈0.7R^2 \approx0.7–0.80.8, while BCE remains close to rank-faithful at R2=0.978R^2 = 0.978. The fit therefore places a penalty on a quantity that is not, in general, the actual contribution to the 1-norm of the full algorithm λ\lambda of a certain mode.

Effective norm versus bare THC 1-norm for BCE, glycine dipeptide, and alanine dipeptide

Figure 6. Effective norm σℓ∣Vℓ(θ∗)∣\sigma_{\ell}|V_{\ell}(\boldsymbol{\theta}^{*})| versus the bare THC 1-norm σℓ\sigma_{\ell} that the fit regularizes, one point per mode. The dashed line marks ∣Vℓ(θ∗)∣=1/M|V_{\ell}(\boldsymbol{\theta}^{*})| = 1/\sqrt{M}, the root-mean-square coefficient on an MM-geometry grid. R2R^{2} is the quality of a log-log fit of one quantity against the other.

The normalization of the SVD eigenvectors determines the scale of this effect. Each column VℓV_\ell has unit norm over the MM-point grid, so its root-mean-square value at a single geometry is 1/M1/\sqrt{M}, the dashed line of Figure 6. This places the dashed line for BCE, with M=73M = 73, almost an order of magnitude above those of the dipeptides, showing that it is a property of the grid size rather than of the molecules. More importantly, λ\lambda is the sampled maximum of a function that is smooth in θ\theta. Refining the grid therefore resolves the same bounded surface more finely rather than discovering unboundedly larger values. Without refitting, λ\lambda does not change for relaxed BCE and glycine dipeptide, whose norm-attaining geometry occurs in every nested sub-grid, although their full grids are 7.3× and 14.8× larger than the coarsest ones. For alanine dipeptide, increasing the grid size by 576× changes λ\lambda by 24%, still well below the O(M)O(M) cost of encoding each geometry independently. The resource estimates of Sec. VI B instead re-optimize the operator on each grid, which lowers λ\lambda on the smaller grids (Table IV–Table VII) without changing the query count. Thus, the contributions in Eq. (18) pay for grid resolution only once, yielding sublinear scaling with the number of geometries MM.

MMSliced-THCSliced-THCSliced-THCSVD-THC qubit comparableSVD-THC qubit comparableSVD-THC qubit comparableSVD-THC qubit comparableSVD-THC Toffoli optimalSVD-THC Toffoli optimalSVD-THC Toffoli optimalSVD-THC Toffoli optimal
MMQubitsAV (101110^{11})Toff. (101010^{10})QubitsAV (101110^{11})Toff. (10910^{9})Impr.QubitsAV (101110^{11})Toff. (10910^{9})Impr.
103979.521.144582.281.975.8×6202.251.926.0×
1939818.12.174612.302.0010.9×6232.271.9611.1×
3739935.24.234642.322.0320.9×6262.291.9821.4×
5539952.46.294652.362.0730.4×6272.332.0331.0×
7340069.58.354672.352.0640.5×6292.332.0241.4×

Table IV. Resource estimates of the energy thresholding circuit for the rigid BCE torsion scan of different grid sizes containing qubit, Toffoli and AV counts for versions of the algorithm featuring sliced-THC as well as the qubit-comparable and Toffoli-optimal SVD-THC. ‘Impr.’ is the ratio of the Toffoli counts of sliced-THC and SVD-THC versions of the thresholding circuit. Active volumes and Toffolis are given in the units shown below each column heading. The qubit-comparable setting is the narrowest SVD-THC circuit whose qubit highwater exceeds that of the sliced-THC circuit at its Toffoli minimum, and the Toffoli-optimal setting is the minimum of the Toffoli count, highlighted in Figure 9. In the sliced-THC circuits the QROAM workspace is reused. SVD-THC compression is redone for each grid, with λ=42.7\lambda= 42.7 Ha on the full grid and λ=40.4\lambda= 40.4–41.741.7 Ha on the smaller grids.

GridMMQubitsAV (101310^{13})Toff. (101110^{11})QubitsAV (101410^{14})Toff. (101110^{11})Impr.QubitsAV (101410^{14})Toff. (101110^{11})Impr.
3×391,1821.341.071,3023.215.380.20×8,4703.082.750.39×
9×9811,18512.19.611,3123.205.411.8×8,4803.072.783.5×
18×183241,18748.338.41,3183.275.457.0×8,4863.142.8213.6×
36×361,2961,1891931541,3243.355.5327.8×8,4923.222.9053.0×
72×725,1841,1917726151,3303.635.70108.0×8,4983.503.06200.6×

Table VII. Resource estimates of the thresholding circuit for alanine dipeptide in different grid sizes, including relative improvements between sliced-THC and two different SVD-THC versions of the algorithm. Columns and settings as in Table IV; λ=416.4\lambda= 416.4 Ha on the full grid and 268.7–376.5 Ha on the smaller grids.

Quantum resource estimates

In this section, we introduce an application of our new primitive for electronic structure computation on geometric ensembles, and discuss its utility in comparison to alternative quantum algorithms. To this end, we analyze the quantum resource requirements for a representative application of our SVD-THC circuit: the energy thresholding algorithm divides an ensemble state into two subspaces, one with energy above a delimiter value Π\Pi, and one with energy below. It has potential applications in tasks such as geometry optimization and the preparation of the PES state in Eq. (1). After defining the thresholding circuit, we provide quantum resource estimates for the qubit-comparable and gate-optimal limits of the space-time trade-offs in QROAM data-loaders [46]. These quantum resource estimates are compared to the baseline version of the thresholding algorithm using the sliced-THC, as well as an alternative, where each geometry is targeted with a separate and independent quantum computation.

Energy thresholding

Let ∣ψθ,k⟩|\psi_{\theta,k}\rangle be the kk-th eigenstate of the Hamiltonian H^θ/λ=∑kEθ,k∣ψθ,k⟩⟨ψθ,k∣\hat{H}_{\theta}/\lambda= \sum_k E_{\theta,k}|\psi_{\theta,k}\rangle\langle\psi_{\theta,k}| with respect to energies Eθ,k∈[−1,1]E_{\theta,k} \in[-1,1], and let Π\Pi be the energy threshold for the Hamiltonian H(θ)H(\theta), which includes the constant offset Eθ\mathcal{E}_{\theta}. Using the multiplexed version of quantum phase estimation, mux⁡θQPEθ\operatorname{mux}_{\theta}\mathrm{QPE}_{\theta} featuring the SVD-THC circuit for mux⁡θB[H^θ/λ]\operatorname{mux}_{\theta}\mathcal{B}[\hat{H}_{\theta}/\lambda], the thresholding circuit in Figure 7 approximately sets a flag qubit ww in the low-energy subspace of an ensemble state (19), such that

Figure 7. Thresholding circuit, using the phase estimation routine QPEθ\mathrm{QPE}_{\theta} and its inverse to entangle eigenstates of the Hamiltonians H^θ\hat{H}_{\theta} for every angle parameter θ\theta in register aa with expressions for their eigenenergies in the phase register pp. θ\theta-dependent energy offsets Eθ′/λ\mathcal{E}'_{\theta}/\lambda are loaded or generated in a temporary register. With an in-place adder, these values correct the energy estimates in the pp register. A comparator circuit ‘Π\Pi’ checks the values in the pp register and flips the flag qubit ww if the values are below a certain energy threshold. Here, the ss register encodes the electronic structure of the simulated system and xx includes all permanent auxiliary registers for the block encoding while qubit uu is used for quantum signal processing. Diamond-shaped gates indicate multiplexed inputs but not necessarily data-loaders.

∑θcθ∣θ⟩a∣ϕθ⟩s↦∑θcθ∣θ⟩a∑k⟨ψθ,k∣ϕθ⟩∣ψθ,k⟩s∣πθ,k⟩w(42)\sum_{\theta} c_{\theta}|\theta\rangle_{a}|\phi_{\theta}\rangle_{s} \mapsto\sum_{\theta} c_{\theta}|\theta\rangle_{a}\sum_{k}\langle\psi_{\theta,k}|\phi_{\theta}\rangle|\psi_{\theta,k}\rangle_{s}|\pi_{\theta,k}\rangle_{w} \tag*{(42)}

with

∣πθ,k⟩={∣1⟩for Eθ,k+Eθ′/λ≤Π,∣0⟩else.(43)|\pi_{\theta,k}\rangle= \begin{cases} |1\rangle& \text{for } E_{\theta,k}+\mathcal{E}'_{\theta}/\lambda\le\Pi,\\ |0\rangle& \text{else.} \end{cases} \tag*{(43)}

Here, Eθ′\mathcal{E}'_{\theta} is a correction of Eθ\mathcal{E}_{\theta}, where we have replaced the mean-field energy of the compressed Hamiltonian by that of the exact one. Writing EHF[⋅]E_{\mathrm{HF}}[\cdot] and Ecorr[⋅]E_{\mathrm{corr}}[\cdot] for the Hartree-Fock and correlation energies of the operator part of a Hamiltonian, that is, excluding its scalar Eθ\mathcal{E}_{\theta}, we load the offset register with

Eθ′=Eθ+EHF[H(θ)]−EHF[HSVD(θ)],(44)\mathcal{E}'_{\theta}=\mathcal{E}_{\theta}+E_{\mathrm{HF}}[H(\theta)]-E_{\mathrm{HF}}[H_{\mathrm{SVD}}(\theta)], \tag*{(44)}

for H(θ)H(\theta) and HSVD(θ)H_{\mathrm{SVD}}(\theta) defined in Eq. (2) and Eq. (18). For the ground state (k=0k=0), the value λEθ,0+Eθ′\lambda E_{\theta,0}+\mathcal{E}'_{\theta} thus describes molecular energies,

Eθ+EHF[H(θ)]+Ecorr[HSVD(θ)],(45)\mathcal{E}_{\theta}+E_{\mathrm{HF}}[H(\theta)]+E_{\mathrm{corr}}[H_{\mathrm{SVD}}(\theta)], \tag*{(45)}

such that the approximate mean-field energy is removed, and only the error of the correlation energy remains. It is ultimately this step that makes the correlation energy error of § 1.4 B the right accuracy metric, with CCSD serving as a classical proxy for the correlation energy that phase estimation resolves. The correction of the energies Eθ,kE_{\theta,k} with a scalar Eθ′/λE'_{\theta}/\lambda requires us to compute the eigenvalues of H^θ/λ\hat{H}_{\theta}/\lambda dynamically. To that end, mux⁡θQPE⁡θ\operatorname{mux}_{\theta}\operatorname{QPE}_{\theta}, writes estimates for the energies Eθ,kE_{\theta,k} into a phase register p:

∑θcθ∣θ⟩a∣ϕθ⟩s↦∑θcθ∣θ⟩a∑k⟨ψθ,k∣ϕθ⟩∣ψθ,k⟩s∣E~θ,k⟩p.(46)\sum_{\theta} c_{\theta}|\theta\rangle_a|\phi_{\theta}\rangle_s \mapsto\sum_{\theta} c_{\theta}|\theta\rangle_a\sum_k \langle\psi_{\theta,k}|\phi_{\theta}\rangle|\psi_{\theta,k}\rangle_s|\widetilde{E}_{\theta,k}\rangle_p . \tag*{(46)}

While ∣E~θ,k⟩|\widetilde{E}_{\theta,k}\rangle are not necessarily computational basis states due to the discretization error of QPE⁡θ\operatorname{QPE}_{\theta}, we may assume that ∣E~θ,k⟩|\widetilde{E}_{\theta,k}\rangle are superpositions of finite-bit energy values ∣E⟩|E\rangle close to Eθ,kE_{\theta,k}. Through an in-place adder, the circuit in Figure 7 corrects these energy values EE by the offsets Eθℓ/λ\mathcal{E}_{\theta}^{\ell}/\lambda loaded into a temporary register, before setting the flag qubit ww. Corrections and QPE are then reversed in an attempt to uncompute the phase register p, which will succeed only approximately.

Note that we expect this correction with values Eθℓ/λ\mathcal{E}_{\theta}^{\ell}/\lambda to be necessary for many other applications, but a regular, qubitization-based QPE circuit would (approximately) write expressions such as ∣arccos⁡Eθ,k⟩|\arccos E_{\theta,k}\rangle rather than energies ∣Eθ,k⟩|E_{\theta,k}\rangle into its phase register, making additive corrections impossible. Algorithms such as the energy thresholding circuit therefore need to run on quantum signal processing, rather than qubitization. This means that the phase kickback operators Ujθ±U_{j\theta\pm} within the multiplexed QPE circuit, depicted in detail in Figure 8(a), must encode time-evolution operators exp⁡(±iH^θtj/λ)\exp(\pm i\hat{H}_{\theta}t_j/\lambda) rather than qubitization eigenvalues exp⁡(±itjarccos⁡Eθ,k)\exp(\pm it_j\arccos E_{\theta,k}). To turn the latter into the former, we employ a generalized quantum signal processing (GQSP) [39] circuit, depicted in Figure 8(b), that repeatedly queries the SVD-THC circuit mux⁡θB[H^θ/λ]\operatorname{mux}_{\theta}\mathcal{B}[\hat{H}_{\theta}/\lambda]. For phase kickback and signal modulation, GQSP introduces a qubit u. For every phase-kickback operator Ujθ±U_{j\theta\pm}, we define a GQSP circuit that uses the Jacobi-Anger expansion polynomial of Ref. [21] with degree dj=O(tj+log⁡εj−1)d_j=O(t_j+\log\varepsilon_j^{-1}), to approximate

Figure 8. (a) Example of a quantum phase estimation routine with a 4-qubit phase register p\mathbf{p}, initially prepared in a window state. Phase oracles Ujθ±\mathcal{U}_{j\theta\pm} kick back unitary signals using bidirectional control (gates labeled ‘±\pm’) on phase qubits pjp_j. The routine is concluded with a quantum Fourier transform QFT. (b) Phase oracles defined by a generalized QSP routine in which u\mathbf{u}-controlled quantum walk operators featuring the block encodings B[H^θ/λ]\mathcal{B}[\hat{H}_{\theta}/\lambda] of H^θ/λ\hat{H}_{\theta}/\lambda, as well as arbitrary 1-qubit unitaries S(ϑij)S(\vartheta_{ij}), parameterized by sets of angles ϑij\vartheta_{ij} for i=0…dji=0\ldots d_j with djd_j being the degree of the encoded polynomial. Starting in ∣0⟩\lvert0\rangle the qubit u\mathbf{u} returns to ∣0⟩\lvert0\rangle approximately towards the end of the subroutine. The shaded region is repeated in the range of i=1…dji=1\ldots d_j. The circuit controls the sign of the Hamiltonians H^θ\hat{H}_{\theta} by a controlled-Z gate between qubits u\mathbf{u} and pjp_j, allowing for double phase-kickback by implementing exp⁡(iH^θtj/λ)\exp(i\hat{H}_{\theta}t_j/\lambda) in the ∣0⟩\lvert0\rangle state of pjp_j and exp⁡(−iH^θtj/λ)\exp(-i\hat{H}_{\theta}t_j/\lambda) in its ∣1⟩\lvert1\rangle state.

Ujθ±≈e±iH^θtj/λ⊗∣0⟩⟨0∣u⊗∣0⟩⟨0∣x+…(47)U_{j\theta\pm}\approx e^{\pm i\hat{H}_{\theta}t_j/\lambda}\otimes|0\rangle\langle0|_u\otimes|0\rangle\langle0|_x+\ldots \tag*{(47)}

up to an error of εj\varepsilon_j in the operator distance. Here the register x is the amalgamation of all permanent auxiliary registers used in the block encoding of Figure 1: x=bg3j5vq{0,3,4,5,7,8}\mathrm{x}=b g_{3}j5vq_{\{0,3,4,5,7,8\}} and tjt_j is a time parameter large enough to estimate the jj-th bits in values for H^θ/λ\hat{H}_{\theta}/\lambda. In an npn_p-qubit QPE routine, we implement a series of multiplexed subroutines Ujθ±U_{j\theta\pm} for j=1…npj=1\ldots n_p, in which the phase qubits pjp_j control the direction ±\pm of the time evolution operators (double phase-kickback). For every such phase oracle we make the choice tj=2j−1πt_j=2^{j-1}\pi and εj=ε2−(np+1−j)\varepsilon_j=\varepsilon2^{-(n_p+1-j)} for some error ε\varepsilon. In that way, our QPE queries the SVD-THC circuit a total of

∑j=1npdj=O(2np+np2+nplog⁡ε−1)(48)\sum_{j=1}^{n_p}d_j=O\left(2^{n_p}+n_p^2+n_p\log\varepsilon^{-1}\right) \tag*{(48)}

times. Due to mixing different approximations to the time evolution inside the QPE circuit, we have a somewhat lower confidence in the estimated phase values. Let us say that in a comparable version of mux⁡θQPE⁡θ\operatorname{mux}_{\theta}\operatorname{QPE}_{\theta}, where the phase kickback operators are replaced by the exact time evolution operators exp⁡(±iH^θtj/λ)\exp(\pm i\hat{H}_{\theta}t_j/\lambda), the probability of a phase estimate ϕ′\phi' being at least Δ\Delta-close to the actual phase ϕ\phi is η\eta, i.e., Pr⁡(∣ϕ−ϕ′∣≤Δ)=η\Pr(|\phi-\phi'|\leq\Delta)=\eta. Then, in the QPE of Figure 8, we have an updated confidence window for phase estimates ϕ∘\phi^{\circ}:

Pr⁡(∣ϕ∘−ϕ∣≤Δ)≥η−12∑jεj(49)\Pr(|\phi^{\circ}-\phi|\leq\Delta)\geq\eta-\frac{1}{2}\sum_j\varepsilon_j \tag*{(49)}
≥η−ε/2,(50)\geq\eta-\varepsilon/2, \tag*{(50)}

as is illustrated in Appendix A. By defining the approximation errors εj=ε2−(np+1−j)\varepsilon_j = \varepsilon2^{-(n_p+1-j)} to become smaller with smaller values of jj, we have reduced the uncertainty on the more significant bits. For QPE circuits in this paper, we use the convention ε=0.1\varepsilon= 0.1, which corresponds to a 5% drop in confidence.

Cost analysis

Based on the energy threshold algorithm, we now provide quantum resource estimates for different grid sizes of the three molecular benchmark systems presented in § 1.4. Specifically, we want to compare the SVD-THC version of the algorithm to the sliced-THC version, in which each geometric subspace is assigned a Hamiltonian via slicing of single-geometry THC circuits. There, each single-geometry block encoding is a BLISS-THC factorization fitted without correlation-energy reweighting, at the smallest rank whose CCSD correlation-energy error is below 0.6 mHa, the same CCSD validation applied to SVD-THC (§ 1.4 B). For SVD-THC, the multiplexed QPE has a Jacobi-Anger schedule with binary time steps, so the cost scales with the power-of-two integer above λ/ε\lambda/\varepsilon rather than smoothly with λ/ε\lambda/\varepsilon itself. A non-integer-power-of-two QPE using Jacobi-Anger expansion is beyond the scope of this work. This introduces a rung-dependent overhead between 1× and 2×, and because the sliced-THC and SVD-THC Hamiltonians can land on different rungs, their apparent speed-up can shift. To somewhat mitigate this effect, reported comparisons use the same Jacobi-Anger counting procedure in both SVD-THC and sliced-THC QPEs. For QPE query complexity N\mathcal{N}, we use the convention N∼⌈πλ/ε⌉\mathcal{N} \sim\lceil\pi\lambda/\varepsilon\rceil based on single-shot statistical confidence, whereas the alternative Holevo-based normalization would use N∼⌈πλ/(2ε)⌉\mathcal{N} \sim\lceil\pi\lambda/(2\varepsilon)\rceil. This is an overall factor-of-two deviation from the quantum phase estimation normalization, independent of the Jacobi-Anger scheduling. In the results below, both SVD-THC and sliced-THC estimates are evaluated under the same convention.

Figure 9 summarizes the space-time trade-offs for SVD-THC across all molecules. Each point is obtained by independently varying the QROAM qubit workspace allocation as well as the Givens angle-loading batching strategy. By varying both the number of angle batches and the QROAM space-time trade-off parameter, we obtain a Pareto curve for both active volume and total Toffoli count. Within these two plots, we observe three features which organize the rest of the section. First, the SVD-THC curves lie below the sliced-THC curves across the whole qubit range. Second, the Toffoli count falls steadily and gives way to a genuine minimum (highlighted by the diamond markers), and rises again once the swap network associated with the QROAM is overprovisioned. Third, the active volume flattens much earlier compared to Toffoli count. For the dipeptides, it comes within 3% of its minimum by about 1400 qubits, and it is essentially flat beyond that. The data-loading cost that dominates the inner prep is set by the total size of the table, rather than how it is partitioned; so past this point the additional qubits only reduce Toffolis, with little effect on the overall active volume.

Space-time trade-off for the sliced-THC and SVD-THC versions of the thresholding circuit

Figure 9. Space-time trade-off for the sliced-THC (open markers, dashed) and SVD-THC version (filled markers, solid) of the thresholding circuit, sweeping the qubit highwater against (a) the total active volume and (b) total Toffoli counts. Each curve is one molecule at its full grid. Outlined diamonds mark each method’s Toffoli minimum point. Outlined circles mark the qubit-comparable point of SVD-THC, defined relative to the Toffoli-optimal point of the sliced-THC multiplexed scheme.

We therefore report two operating points on these curves. The qubit comparable setting is defined as the closest approximation of the SVD-THC version of the circuit to the gate-optimal version of the sliced-THC circuit in terms of qubits. The Toffoli optimal setting is the minimum of the Toffoli count, which sometimes coincides with the true minimum of the active volume curve.

Table IV and Table V summarize the quantum resource estimates for both rigid and relaxed BCE. At the coarsest grid of Table V, which contains ten geometries, the SVD-THC requires 4.7× fewer Toffoli gates at the qubit-comparable setting, or 5.0× at the Toffoli optimal one. Over the full 73-point grid, the overall advantage reaches 32.8× at the qubit-comparable setting, and 34.9× at the Toffoli-optimal one. As expected, the rigid scan shows larger compressibility and provides larger improvements overall. The same pattern holds for the dipeptides, where the SVD-THC 1-norm λ\lambda exceeds the maximal single-geometry 1-norm λmax⁡\lambda_{\max} by a factor of 2.5 and 2.7, yet the Toffoli advantage reaches 41.8× and 108.0× at the qubit-comparable setting, and 144.5× and 200.6× at the Toffoli-optimal setting, for glycine dipeptide and alanine dipeptide respectively.

MMSliced-THCSliced-THCSliced-THCSVD-THC qubit comparableSVD-THC qubit comparableSVD-THC qubit comparableSVD-THC qubit comparableSVD-THC Toffoli optimalSVD-THC Toffoli optimalSVD-THC Toffoli optimalSVD-THC Toffoli optimal
MMQubitsAV (101110^{11})Toff. (101010^{10})QubitsAV (101110^{11})Toff. (10910^{9})Impr.QubitsAV (101110^{11})Toff. (10910^{9})Impr.
103979.781.144593.262.444.7×6223.172.295.0×
1939818.62.174623.302.478.8×6253.212.329.4×
3739936.24.244653.332.5116.9×6283.242.3618.0×
5539953.86.304653.362.5524.7×6283.272.4026.3×
7340071.48.364683.382.5532.8×6313.292.4034.9×

Table V. Resource estimates (qubit, Toffoli and AV counts) of the energy thresholding circuit for the relaxed BCE torsion scan in different grid sizes, with relative improvements between sliced-THC and SVD-THC versions of the algorithm in terms of Toffoli gates. Columns and settings as in Table IV; λ=53.8\lambda= 53.8 Ha on the full grid and 53.253.2–53.853.8 Ha on the smaller grids.

As established in § 1.4 D, the number of retained modes KK, the ranks RℓR_{\ell}, and λ\lambda are properties of the SVD-THC technique rather than of the grid. Both the QROAM cost and λ\lambda vary only weakly as a function of grid size, so the SVD-THC Toffoli count remains within 6% at the qubit-comparable setting and 12% at the Toffoli-optimal one of its coarsest-grid value, while the counts for the sliced-THC version of the algorithm add the resources of the block encodings of every geometry and so grow linearly with the grid size. The improvement over the baseline circuit therefore grows with MM, at the qubit-comparable setting from 4.7× to 32.8× for BCE between 10 and 73 geometries, from 2.9× to 41.8× for glycine dipeptide between 180 and 2,664 geometries, and from 0.20× to 108.0× for alanine dipeptide between 9 and 5,184 geometries, and at the Toffoli-optimal setting from 5.0× to 34.9× (rigid BCE: 6.0× to 41.4×), 10.5× to 144.5× and 0.39× to 200.6× respectively.

Furthermore, we observe an improvement for alanine dipeptide below unity at the beginning. At the coarsest 3×33 \times3 grid, the fixed cost of a circuit sized for the full 150-mode ensemble, whose largest rank is 477, is not yet amortized, and SVD-THC is more expensive than sliced-THC. The advantage becomes positive with enough geometries, which occurs between 9 and 81 geometries at both settings, growing thereafter. This is the circuit-level counterpart of the behavior of λ\lambda discussed in § 1.4 D. Compression becomes profitable once the grid is dense enough that redundant per-geometry data-loading, rather than the shared ensemble, dominates the cost.

The improvements observed in the active volume metric are smaller. At the Toffoli-optimal setting the full-grid active-volume improvements are 29.9× and 21.7× for rigid and relaxed BCE, 17.6× for glycine dipeptide and 22.1× for alanine dipeptide, against Toffoli improvements of 41.4×, 34.9×, 144.5× and 200.6×, so for the dipeptides close to 90% of the Toffoli advantage does not carry over into active volume. To understand why, Figure 10 divides one SVD-THC block encoding query, at the full-grid Toffoli-optimal operating point, into its subroutines: the outer prep (muxθ_\theta prepo,θ_{o,\theta}), which loads the geometry-dependent amplitudes σℓVℓ(θ)/λ\sigma_\ell V_\ell(\theta)/\lambda and is the only part of the circuit that addresses the geometry register; the inner prep (muxℓ_\ell prepi,ℓ_{i,\ell}), which loads amplitudes of the central tensors of the modes set by the outer prep; and Select (muxℓ_\ell Selectℓ_\ell), which applies that mode’s basis rotations and fermionic operators. We observe two features. First, the outer prep accounts for only a small share of the cost, carrying 8–11% of the Toffolis and 6–12% of the active volume, so the one subroutine whose data grows with the number of geometries is also a small share of every query, which is the structural reason the SVD-THC costs of Table IV–Table VII are nearly flat in grid size. Second, the two cost measures attribute the remainder differently. In terms of Toffolis, Select dominates for every system at 55–64% of the total cost. In active volume the dipeptides instead place 61–67% of the cost in the inner prep, while for BCE, whose tensors are an order of magnitude smaller, Select still leads at about 50%. The inner prep consists of a QROAM read of the central tensors, and active volume, unlike the Toffoli count, also prices the Clifford operations. These Cliffords are the dominant contribution when writing the loaded data and their number grows with the size of the table. It is this data-loading that negates most of the Toffoli advantage when counting active volume blocks.

Figure 10. Breakdown of (a) aggregated Toffolis and (b) active volume for a single SVD-THC block encoding query at the full-grid Toffoli-optimal operating point, split into subroutines mux⁡θprep⁡o,θ\operatorname{mux}_{\theta}\operatorname{prep}_{o,\theta}, mux⁡ℓprep⁡i,ℓ\operatorname{mux}_{\ell}\operatorname{prep}_{i,\ell} and mux⁡ℓSelect⁡ℓ\operatorname{mux}_{\ell}\operatorname{Select}_{\ell} from Figure 1, labeled ‘Outer PREP’, ‘Inner PREP’ and ‘SELECT’ in this figure. Other contributions are below 1% for all systems and cost types and are not visible at this scale. The total above each bar is the cost of one query.

GridMMQubitsAV (101410^{14})Toff. (101210^{12})QubitsAV (101410^{14})Toff. (101110^{11})Impr.QubitsAV (101410^{14})Toff. (101110^{11})Impr.
10×181808092.201.918282.246.622.9×8,9881.731.8210.5×
19×183428104.173.628312.246.645.5×8,9911.741.8419.7×
19×366848118.357.248342.266.6610.9×8,9941.751.8638.8×
37×361,33281216.314.18372.296.7021.0×8,9971.781.9074.1×
37×722,66481332.528.28402.366.7541.8×9,0001.851.95144.5×

Table VI. Resource estimates of the energy thresholding circuit for the glycine dipeptide in different grid sizes, including relative improvements of the sliced-THC version and SVD-THC versions of the algorithm in terms of Toffoli gates. Columns and settings as in Table IV; λ=339.9\lambda= 339.9 Ha on the full grid and 321.2–337.3 Ha on the smaller grids.

Since the sliced-THC cost is linear in the number of geometries while the SVD-THC cost is nearly constant, every improvement factor is close to proportional to MM, and the break-even grid size follows directly. In terms of Toffolis, the dipeptides break-even at about 17–23 geometries at the Toffoli-optimal setting and 45–65 at the qubit-comparable one; in active volume, the break-even moves out to 140–210 geometries. These results are qualitatively similar to protocols where instead of a thresholding circuit, an energy estimation is done for each geometry using BLISS-THC. We discuss concrete resource counts in Appendix C.

Conclusion

In this work, we have developed a novel framework for electronic structure calculations across multiple geometries of a molecule, encoded in superposition on a quantum computer. In particular, we have shown that the electronic system of every geometry can be addressed simultaneously beyond the break-even point with methods that treat them separately. With SVD-THC we have provided the key primitive to enable a shared time evolution of all geometric subspaces, allowing the parallel computation of molecular energies in quantum phase estimation. Analyzing the quantum resource counts of SVD-THC circuits across several molecular systems and grid sizes, we achieve a positive compression beyond the aforementioned break-even point for grids above tens of geometries.

On the largest grids considered, we obtain Toffoli reductions of 35×35\times to 41×41\times for BCE, 144×144\times for glycine dipeptide, and 200×200\times for alanine dipeptide. These improvements grow almost in proportion to the number of geometries MM, demonstrating that the cost of representing the ensemble can become substantially smaller than repeating an independent computation for every geometry. The advantage persists when measured in active volume, although it is reduced to 18×18\times to 30×30\times. This difference highlights the cost of the Clifford-heavy data-loading operations that are largely invisible to the Toffoli metric. Reducing this data-loading overhead is therefore an important direction for future work. In particular, techniques for loading compressed representations of the geometry-dependent data could substantially narrow the gap between Toffoli count and active volume. Achieving comparable compression in active volume, even with a modest increase in qubit count, would provide a stronger indication of the practical utility of SVD-THC.

We hope this work presents a useful new primitive that will be a part of algorithms widening the range of scientific questions that can be answered with a quantum computer. For instance the Liouvillian framework for performing Born-Oppenheimer molecular dynamics [20] relies on Hamiltonian simulation controlled on a grid over the nuclear positions. Similarly, quantum-search algorithms [47, 48] could make use of SVD-THC when constructing oracles based on molecular energies, enabling searches for minimum-energy geometries, barrier heights, transition probabilities, or just sampled geometries with an energy below a certain threshold. For instance, an unstructured search for the lowest-energy geometry requires only O(M)O(\sqrt{M}) applications of the thresholding circuit, whereas direct estimation requires a separate phase estimation for each of the MM geometries. An SVD-THC-powered search therefore becomes favorable once the compression advantage over sequential QPE (Appendix C) exceeds M\sqrt{M}. This is already the case for BCE on its full grid, while for the dipeptides the advantage is comparable to M\sqrt{M} at the grid sizes considered here, so the outcome depends on the constant factors of the search. Since the advantage grows roughly in proportion to MM, a clear advantage is expected on moderately larger grids.

The framework is not restricted to grids over internal molecular coordinates or to SVD-THC as the underlying tensor factorization. Other low-rank representations of the electron-repulsion integrals, such as sum-of-squares decompositions, could provide alternative routes for compressing the geometry-dependent Hamiltonians and may offer different trade-offs between classical preprocessing, qubit count, and circuit complexity. Furthermore, our method may be applicable to other types of Hamiltonian ensembles. A particularly relevant example is a conformer ensemble, i.e., a set of low-energy geometries, as generated for instance by classical conformer search methods. In this case, the geometries need not lie on a regular grid. Distinct conformers typically differ by larger and more global rearrangements than neighboring grid points. We might therefore expect the shared low-rank structure exploited by our compression to be less prominent. On the other hand, all conformers share the same atoms, bonding topology and much of their local electronic structure, which is the redundancy that SVD-THC exploits. Quantifying how well the compression works with conformer ensembles is therefore an important next step, in particular for computational drug discovery, where the bioactive conformation of a ligand, its strain energy upon binding, and Boltzmann-weighted molecular properties all hinge on an accurate ranking of conformer energies. Beyond conformers, the global changes between neighboring geometries generated by time evolution in molecular dynamics or along a reaction path present other potential targets for compression [20]. Categorical changes between atom types could be a way to enable thermodynamic integration and alchemical interpolations between ligands [19].

Higher-dimensional potential energy surfaces may offer larger opportunities for compression as the number of geometries grows. At the same time, higher-dimensional grids rapidly increase the cost of classical preprocessing, creating a practical limit on the dimensionality that can be treated directly. One possible way to mitigate this cost is to apply compression to block encodings of the generating functions for Hamiltonian matrix elements rather than to their explicitly computed numerical values. On-the-fly evaluation of Gaussian matrix elements currently carries large prefactors, however, and its crossover with explicit matrix-element encoding occurs at system sizes larger than those considered here [49]. An alternative is to formulate the problem in first-quantized electronic-structure algorithms [50, 51, 52], using a space-centered basis in which the geometry-dependent Hamiltonian can be encoded directly from nuclear and electronic coordinates. Taken together, these results establish SVD-THC as a primitive for quantum-parallel simulation of ensembles of related molecular Hamiltonians. The observed sublinear scaling and large Toffoli reductions show that sharing the representation of many geometries can move ensemble simulation beyond the cost of independent calculations. Further reductions in data-loading and classical preprocessing costs will determine how broadly this advantage can translate into practical quantum algorithms, but the framework provides a route toward using quantum parallelism for molecular problems that require information from many related geometries.

Acknowledgments

The authors would like to thank Eric Johnston for his invaluable contributions to speeding up the QRE code. We would also like to thank Carter M. Gustin for showing us that two rounds of amplitude amplification are indeed sufficient for USP circuits in THC, Modjtaba Shokrian Zini and Bjorn K. Berntson for their advice on QSP, as well as Michael J. Bremner and Philipp T. Ernst for helpful suggestions on the manuscript. The authors thank Clemens Utschig-Utschig for his comments on the manuscript and the support during the project. The authors acknowledge the use of AI in the code development for this project.

References

QPE using quantum signal processing

In this section, we define a quantum phase estimation (QPE) algorithm mixing phase oracles that approximate a time evolution operator differently well. Mixing approximations allows us to combine our SVD-THC circuit with quantum signal processing (QSP) without blowing up the complexity of QPE. The main result of our analysis is that the right recombination of different QSP routines yields a version of QPE that differs very little from an idealized version using queries to the time evolution operator. Indeed, with O(nplog⁡ε−1)O(n_p \log\varepsilon^{-1}) additional queries to the SVD-THC circuit, the statistical confidence of an npn_p-qubit QPE with mixed QSP approximations can be made (ε/2)(\varepsilon/2)-close to the ideal QPE. To prove this statement, let us first introduce some notation. Let

e±iarccos⁡Eθ,k,(51)e^{\pm i \arccos E_{\theta,k}}, \tag*{(51)}

be the eigenvalues of the qubitization walk operators featuring the block encoding of H^θ/λ=∑kEθ,k∣ψθ,k⟩⟨ψθ,k∣\hat{H}_{\theta}/\lambda= \sum_k E_{\theta,k}|\psi_{\theta,k}\rangle\langle\psi_{\theta,k}|, where the signs ‘+’ and ‘−’ correspond to different eigenstates ∣Qθ,k,+⟩|Q_{\theta,k,+}\rangle and ∣Qθ,k,−⟩|Q_{\theta,k,-}\rangle. Using a Jacobi-Anger expansion polynomial in quantum signal processing [21], these two eigenstates can be approximately recombined into a unitary with eigenvalues exp⁡(iEθ,kt)\exp(iE_{\theta,k}t) and eigenstates

∣ψθ,k⟩s∣0⟩x∣0⟩u(52)|\psi_{\theta,k}\rangle_s|0\rangle_x|0\rangle_u \tag*{(52)}

where tt is a real parameter, uu is an auxiliary qubit used in generalized quantum signal processing (GQSP) [39], and xx is the amalgamated auxiliary register for the block encoding B[H^θ/λ]\mathcal{B}[\hat{H}_{\theta}/\lambda], containing all permanent (i.e. non-elbowed) registers of bb, gg, jj, qq and vv. A polynomial of finite degree can unfortunately only implement this time-evolution operator up to some error ε\varepsilon. Suppressing this approximation error is relatively easy, as the degree dd of the polynomial, and therefore the number of queries to mux⁡θB[H^θ/λ]\operatorname{mux}_{\theta}\mathcal{B}[\hat{H}_{\theta}/\lambda] in a QSP circuit has with d=O(t+log⁡ε−1)d=O(t+\log\varepsilon^{-1}) a rather forgiving scaling with ε\varepsilon. However, we can only benefit from this scaling if we admit different approximations to time evolution to QPE.

For the quantum phase estimation routine in Figure 8, we define phase-kickback oracles Ujθ±\mathcal{U}_{j\theta\pm} with GQSP routines depicted on the right-hand side of Figure 8(b). Every such routine Ujθ±\mathcal{U}_{j\theta\pm} shall have different degrees djd_j, time parameters tjt_j and target errors εj\varepsilon_j. Approximating the time evolution operator exp⁡(±iH^θtj/λ)\exp(\pm i\hat{H}_{\theta}t_j/\lambda) within the all-zero component A(Ujθ±)\mathcal{A}(\mathcal{U}_{j\theta\pm}) of the GQSP circuit:

Ujθ±=A(Ujθ±)∣0⟩u∣0⟩x⟨0∣u⟨0∣x+…,(53)\mathcal{U}_{j\theta\pm}=\mathcal{A}(\mathcal{U}_{j\theta\pm})|0\rangle_u|0\rangle_x\langle0|_u\langle0|_x+\ldots, \tag*{(53)}

the approximation error is defined such that

εj≥∥A(Ujθ±)−e±iH^θtj/λ∥.(54)\varepsilon_j \ge\left\lVert\mathcal{A}(\mathcal{U}_{j\theta\pm}) - e^{\pm i\hat{H}_{\theta}t_j/\lambda} \right\rVert. \tag*{(54)}

This is done for all phase-kickback routines Ujθ±\mathcal{U}_{j\theta\pm} in the full QPE with npn_p phase qubits pjp_j for j=1…npj = 1 \ldots n_p. In this QPE, we unfortunately cannot transfer the approximation errors onto the Hamiltonian, i.e. we cannot regard the estimated phase as the energy of an effective Hamiltonian shaped by approximation errors. This is due to every phase oracle Ujθ±\mathcal{U}_{j\theta\pm} being an individual approximation of its corresponding time evolution operator exp⁡(±iH^θtj/λ)\exp(\pm i\hat{H}_{\theta}t_j/\lambda) with

Ujθ±≠(U1θ+)±2j−1.(55)\mathcal{U}_{j\theta\pm} \ne\left(\mathcal{U}_{1\theta+}\right)^{\pm2^{j-1}}. \tag*{(55)}

In fact, we do not even guarantee a clean unitarity of forwards and backwards evolutions as Ujθ+Ujθ−=1\mathcal{U}_{j\theta+}\mathcal{U}_{j\theta-} = 1 is fulfilled only approximately. Rather than one being the inverse of the other, we define Ujθ−\mathcal{U}_{j\theta-} as a version of Ujθ+\mathcal{U}_{j\theta+}, in which the sign of the Hamiltonian block encoding B[H^θ/λ]\mathcal{B}[\hat{H}_{\theta}/\lambda] is flipped, which makes for an inexpensive double phase-kickback by CZ gates between qubits pjp_j and u as shown in Figure 8(b).

So while we cannot hope for a systematic error of the phase, we can make statements about the error of estimated phases. To that end, let us first define the window filter by the coefficients {wr}\{w_r\}. The window state preparation routine creates the superposition

window ∣0⟩p=∑rwr∣r⟩p(56)\mathrm{window}\,|0\rangle_p = \sum_r w_r |r\rangle_p \tag*{(56)}

on the phase register p, where integers rr are represented as a bit string r1r2⋯rnpr_1r_2 \cdots r_{n_p}. We now compare the outputs of the QPE in Figure 8 within the ∣0⟩|0\rangle subspaces of register x and qubit u to a QPE where the phase-kickback routines Ujθ±\mathcal{U}_{j\theta\pm} are perfect time evolution routines exp⁡(±iH^θtj/λ)\exp(\pm i\hat{H}_{\theta}t_j/\lambda). We want to evaluate the difference

∥QFT∑rwr∣r⟩pexp⁡[iH^θ∑m(−1)rmtm/λ]∣ψ⟩s−QFT∑rwr∣r⟩p[∏mA(Umθ(−1)rm)]∣ψ⟩s∥.(57)\left\lVert\mathrm{QFT}\sum_r w_r|r\rangle_p \exp\left[i\hat{H}_{\theta}\sum_m(-1)^{r_m}t_m/\lambda\right]|\psi\rangle_s-\mathrm{QFT}\sum_r w_r|r\rangle_p\left[\prod_m\mathcal{A}\left(\mathcal{U}_{m\theta(-1)^{r_m}}\right)\right]|\psi\rangle_s\right\rVert. \tag*{(57)}

Disregarding the quantum Fourier transform as a constant unitary, we can define the two states as the beginning and end of a telescoping series of states

∣QPEk⟩=∑r=02np−1wr∣r⟩p[∏j=1kei(−1)rjH^θtj/λ][∏m=k+1npA(Umθ(−1)rm)]∣ψ⟩s,(58)|\mathrm{QPE}_k\rangle=\sum_{r=0}^{2^{n_p}-1}w_r|r\rangle_p\left[\prod_{j=1}^{k}e^{i(-1)^{r_j}\hat{H}_{\theta}t_j/\lambda}\right]\left[\prod_{m=k+1}^{n_p}\mathcal{A}\left(\mathcal{U}_{m\theta(-1)^{r_m}}\right)\right]|\psi\rangle_s, \tag*{(58)}

allowing us to evaluate the difference of the states ∣QPEnp⟩|\mathrm{QPE}_{n_p}\rangle and ∣QPE0⟩|\mathrm{QPE}_0\rangle using multiple applications of the triangle inequality:

∥∣QPE0⟩−∣QPEnp⟩∥≤∥∣QPE0⟩−∣QPE1⟩∥+∥∣QPE1⟩−∣QPE2⟩∥+⋯+∥∣QPEnp−1⟩−∣QPEnp⟩∥=∑k=1np∥∣QPEk−1⟩−∣QPEk⟩∥.(59)\begin{aligned} \left\lVert|\mathrm{QPE}_0\rangle-|\mathrm{QPE}_{n_p}\rangle\right\rVert\le\left\lVert|\mathrm{QPE}_0\rangle-|\mathrm{QPE}_1\rangle\right\rVert+\left\lVert|\mathrm{QPE}_1\rangle-|\mathrm{QPE}_2\rangle\right\rVert+\cdots+\left\lVert|\mathrm{QPE}_{n_p-1}\rangle-|\mathrm{QPE}_{n_p}\rangle\right\rVert\\ &=\sum_{k=1}^{n_p}\left\lVert|\mathrm{QPE}_{k-1}\rangle-|\mathrm{QPE}_k\rangle\right\rVert. \tag*{(59)} \end{aligned}

Singular elements of the sum can be evaluated as

∥∣QPEk−1⟩−∣QPEk⟩∥=∥∑r=02np−1wr∣r⟩p[∏j=1k−1ei(−1)rjH^θtj/λ][A(Ukθ(−1)rk)−ei(−1)rkH^θtk/λ]×[∏m=k+1npA(Umθ(−1)rm)]∣ψ⟩s∥(60)\begin{aligned} \left\lVert\lvert\mathrm{QPE}_{k-1}\rangle-\lvert\mathrm{QPE}_{k}\rangle\right\rVert ={}&\left\lVert\sum_{r=0}^{2^{n_p}-1}w_r\lvert r\rangle_p \left[\prod_{j=1}^{k-1}e^{i(-1)^{r_j}\hat{H}_{\theta t_j}/\lambda}\right] \left[\mathcal{A}(U_{k\theta(-1)^{r_k}})-e^{i(-1)^{r_k}\hat{H}_{\theta t_k}/\lambda}\right]\right.\\ &\left.\qquad\times \left[\prod_{m=k+1}^{n_p}\mathcal{A}(U_{m\theta(-1)^{r_m}})\right]\lvert\psi\rangle_s\right\rVert \tag*{(60)} \end{aligned}
≤∑r∣wr∣2∥[A(Ukθ(−1)rk)−ei(−1)rkH^θtk/λ][∏m=k+1npA(Umθ(−1)rm)]∣ψ⟩s∥2(61)\leq\sqrt{\sum_r\lvert w_r\rvert^2 \left\lVert \left[\mathcal{A}(U_{k\theta(-1)^{r_k}})-e^{i(-1)^{r_k}\hat{H}_{\theta t_k}/\lambda}\right] \left[\prod_{m=k+1}^{n_p}\mathcal{A}(U_{m\theta(-1)^{r_m}})\right]\lvert\psi\rangle_s \right\rVert^2} \tag*{(61)}
≤∑r∣wr∣2∥A(Ukθ(−1)rk)−ei(−1)rkH^θtk/λ∥2(62)\leq\sqrt{\sum_r\lvert w_r\rvert^2 \left\lVert\mathcal{A}(U_{k\theta(-1)^{r_k}})-e^{i(-1)^{r_k}\hat{H}_{\theta t_k}/\lambda}\right\rVert^2} \tag*{(62)}
≤∑r∣wr∣2⋅εk,∑r∣wr∣2⏟=1(63)\leq\sqrt{\sum_r\lvert w_r\rvert^2}\cdot\varepsilon_k, \qquad \underbrace{\sum_r\lvert w_r\rvert^2}_{=1} \tag*{(63)}

where first we have used ∥∑rwr∣r⟩Ar∣Ψr⟩∥≤∑r∣wr∣2∥Ar∣Ψr⟩∥2\left\lVert\sum_r w_r\lvert r\rangle A_r\lvert\Psi_r\rangle\right\rVert\leq\sqrt{\sum_r\lvert w_r\rvert^2\left\lVert A_r\lvert\Psi_r\rangle\right\rVert^2} for any operators ArA_r and states ∣Ψr⟩\lvert\Psi_r\rangle, then we have used the fact that operator norm is defined as the supremum of all supporting states, ∥A∥=sup⁡∣Ψ⟩∥A∣Ψ⟩∥\lVert A\rVert=\sup_{\lvert\Psi\rangle}\lVert A\lvert\Psi\rangle\rVert, including the choice ∣Ψ⟩=[∏m>kA(Umθ(−1)rm)]∣ψ⟩\lvert\Psi\rangle=\left[\prod_{m>k}\mathcal{A}(U_{m\theta(-1)^{r_m}})\right]\lvert\psi\rangle. Finally, (A12) is recombined with (A4). Plugging (A13) into (A9), we find the difference between states output by ideal and approximate QPE to be ∑kεk≤ε\sum_k\varepsilon_k\leq\varepsilon, following the definition of εk=ε2−(np+1−k)\varepsilon_k=\varepsilon2^{-(n_p+1-k)} in Section VI A. This result might directly influence the distribution of phase values, i.e., if the output states are

∣p⟩=∑j=02np−1pj∣j⟩p∣χj⟩suxand∣p~⟩=∑j=02np−1p~j∣j⟩p∣χj⟩sux,(64)\lvert p\rangle=\sum_{j=0}^{2^{n_p}-1}\sqrt{p_j}\lvert j\rangle_p\lvert\chi_j\rangle_{\mathrm{sux}} \quad\text{and}\quad \lvert\widetilde{p}\rangle=\sum_{j=0}^{2^{n_p}-1}\sqrt{\widetilde{p}_j}\lvert j\rangle_p\lvert\chi_j\rangle_{\mathrm{sux}}, \tag*{(64)}

for some states ∣χj⟩\lvert\chi_j\rangle and with ∑j∣pj−p~j∣≤ε\sum_j\lvert p_j-\widetilde{p}_j\rvert\leq\varepsilon. In the worst case, we lose ε/2\varepsilon/2 of squared amplitude inside the confidence region of QPE, and gain ε/2\varepsilon/2 in probability of measuring the phase in the outside region of our confidence interval. For instance, if the only phase α\alpha measured is in between 2πy2−np2\pi y2^{-n_p} and 2π(y+1)2−np2\pi(y+1)2^{-n_p}, then py+py+1−p~y−p~y+1≤ε/2p_y+p_{y+1}-\widetilde{p}_y-\widetilde{p}_{y+1}\leq\varepsilon/2. That is, we are just lowering the probability of measuring phases with error π2−np+1\pi2^{-n_p+1} by ε/2\varepsilon/2.

The alternative, in which we do not mix approximations to the time evolution, would have been much less favorable. Using QSP to define a constant phase oracle UU that is ε~\widetilde{\varepsilon}-close to exp⁡(iH^θt1/λ)\exp(i\hat{H}_{\theta t_1}/\lambda) would cause an energy error of size O(λε~)O(\lambda\widetilde{\varepsilon}) while increasing the complexity of QPE by a factor O(log⁡ε~−1)O(\log\widetilde{\varepsilon}^{-1}).

Inner state preparation for SVD-THC

The multiplexed version of the inner state preparation, prep⁡i,ℓ\operatorname{prep}_{i,\ell} utilizes a modified version of Alias Sampling to prepare auxiliary state (37).

The circuit presented in Figure 11 features a data-loader with aliases alt and alt′' for indices ∣μ⟩|\mu\rangle and ∣ν⟩|\nu\rangle respectively, while loading also a sign into q2_2 as well as an alternative sign into the junk register j1_1. All these quantities are triply indexed, with μ≤ν\mu\le\nu, and in the ∣ℓ⟩|\ell\rangle subspace of b, the indices (μ,ν)(\mu,\nu) assume all possible values from (0,0)(0,0) to (N−1,Rℓ)(N-1,R_\ell). This means that the size of registers q4_4 and q5_5 must be padded to the largest possible THC rank, bringing their size to ⌈log⁡(max⁡ℓRℓ+1)⌉\lceil\log(\max_\ell R_\ell+1)\rceil qubits. Since ∣μ⟩|\mu\rangle and ∣ν⟩|\nu\rangle are swapped later in Select, q4_4 and q5_5 must have the same size. In the state-of-the-art THC circuit [33], the indices μ\mu and ν\nu are joined arithmetically into s(μ,ν)=(ν+1)ν/2+μs(\mu,\nu)=(\nu+1)\nu/2+\mu. This contiguizer would also work here, where the indices are padded. We obtain the triply joined index r(ℓ,μ,ν)r(\ell,\mu,\nu) by adding the value of ss to SℓS_\ell, the basis with respect to the most significant index ℓ\ell:

Figure 11. Circuit construction for muxℓ_\ell prepi,ℓ_{i,\ell} of SVD-THC. On the basis of a multiplexed uniform state preparation circuit (muxℓ_\ell USPℓ_\ell), this circuit encodes the state ∝∑μζℓμν∣μ⟩∣ν⟩\propto\sum_{\mu} \zeta_{\ell\mu\nu}|\mu\rangle|\nu\rangle with coefficients ζℓμν\zeta_{\ell\mu\nu}, triple-indexed by the values μ\mu and ν\nu in registers q4q_4 and q5q_5, as well as the superindex ℓ\ell in bb. As a first step, the indices μ\mu and ν\nu are turned into a continuous index s(μ,ν)=ν(ν+1)/2+μs(\mu,\nu)=\nu(\nu+1)/2+\mu with quantum arithmetic circuits. Since the indices μ\mu and ν\nu run in the ranges [0,Rℓ)[0,R_\ell) and [0,Rℓ][0,R_\ell], upper-limited by the ℓ\ell-dependent THC ranks RℓR_\ell, we load the bases SℓS_\ell for the indices r(ℓ,μ,ν)r(\ell,\mu,\nu) into a temporary register. An in-place addition with the s(μ,ν)s(\mu,\nu) values then completes the computation of the index rr, from which we can load altr\mathrm{alt}_r and altr′\mathrm{alt}'_r, the alternative values for μ\mu and ν\nu as well as the keep probabilities keepr\mathrm{keep}_r, along with original and alternative values for the signs, signr\mathrm{sign}_r and signr′\mathrm{sign}'_r. A comparator—the gate labeled ‘a<ba<b’—between the keep probabilities and the values in register j5j_5 controls the swaps of original and alternative values. To flag the scenario of ν=Rℓ\nu=R_\ell in the ∣ℓ⟩|\ell\rangle subspace of bb, we load the THC ranks RℓR_\ell with SℓS_\ell, and set the qubit q1q_1 with what is largely a multi-Toffoli gate, labeled ‘a=ba=b’. To uncompute temporary registers, the in-place addition of rr is reverted.

r(ℓ,μ,ν)=Sℓ+s(μ,ν)(65)r(\ell,\mu,\nu)=S_\ell+s(\mu,\nu) \tag*{(65)}

where

Sℓ=ℓN+∑m<ℓ(Rm+1)Rm2.(66)S_\ell=\ell N+\sum_{m<\ell}\frac{(R_m+1)R_m}{2}. \tag*{(66)}

As part of the circuit, we also set qubit q1q_1 in the subspace ∣ℓ⟩∣ν=Rℓ⟩|\ell\rangle|\nu=R_\ell\rangle of qubits (b,q5)(\mathrm{b},q_5), as a flag for single-particle Hamiltonians. To this end, we load the value of RℓR_\ell in the very same routine loading SℓS_\ell. After the controlled swaps of the circuit, the value can then be XORed into q5q_5, and q1q_1 is set by a multi-Toffoli gate.

A resource state for Alias Sampling is prepared approximately by a multiplexed version of the subroutine USPℓ\mathrm{USP}_\ell. We find

USPℓ∣0⟩q0q4q5=∣1⟩q0∑μ,νΛℓμν∣μ⟩q4∣ν⟩q5(67)\mathrm{USP}_\ell|0\rangle_{q_0q_4q_5}=|1\rangle_{q_0}\sum_{\mu,\nu}\Lambda_{\ell\mu\nu}|\mu\rangle_{q_4}|\nu\rangle_{q_5} \tag*{(67)}

with uniform but constrained coefficients

Λℓμν={1(Rℓ+1)Rℓ/2+N,μ≤ν and ν⊕μ≤Rℓ⊕(N−1)0,else,(68)\Lambda_{\ell\mu\nu}= \begin{cases} \sqrt{\dfrac{1}{(R_\ell+1)R_\ell/2+N}}, & \mu\le\nu\ \text{and}\ \nu\oplus\mu\le R_\ell\oplus(N-1)\\ 0, & \text{else,} \end{cases} \tag*{(68)}

and where qubit q0q_0 certifies success. In regular THC, the USP circuit would start from a uniform superposition ∝∑μ,ν∣μ⟩∣ν⟩\propto\sum_{\mu,\nu}|\mu\rangle|\nu\rangle of all possible values of μ\mu and ν\nu, which can then be constrained with at most two rounds of amplitude amplification, depending on the THC rank. In the multiplexed USPℓ\mathrm{USP}_\ell circuit, depicted in Figure 12, registers q4q_4 and q5q_5 are padded up to the largest THC rank max⁡ℓRℓ\max_\ell R_\ell. For subspaces ∣ℓ⟩|\ell\rangle with dramatically lower ranks RℓR_\ell, this could mean more than two rounds of amplitude amplification. In the circuit, we therefore load bit masks IℓI_\ell and JℓJ_\ell, that would control Hadamard gates in registers q4q_4 and q5q_5, respectively. CZ gates running transversely from the temporary registers holding IℓI_\ell (JℓJ_\ell) into a q4q_4 (q5q_5), right in between a T gate in Y basis Ty=exp⁡(iπ/8Y)T_y=\exp(i\pi/8Y) and its inverse Ty†T_y^\dagger (colored blue in Figure 12), affect the target qubit as

Figure 12. The muxℓ_{\ell} USPℓ_{\ell} routine of the multiplexed THC block encoding, preparing the registers q4q_4 and q5q_5 in the uniform state ∣μ⟩∣ν⟩\lvert\mu\rangle\lvert\nu\rangle with μ≤ν\mu\le\nu and indices in the range from (μ,ν)=(0,0)…(N−1,Rℓ)(\mu,\nu)=(0,0)\ldots(N-1,R_{\ell}), where the THC ranks RℓR_{\ell} depend on the superindex ℓ\ell in register bb. In this toy example of USPℓ_{\ell}, q4q_4 and q5q_5 contain 4 qubits each, and the uniform state is prepared within a single round of (fractional) amplitude amplification. Note that a realistic circuit might require two rounds of amplitude amplification. The ℓ\ell-dependent angles ϕℓ\phi_{\ell} for partial reflections, dimensional integers dℓd_{\ell} as well as bit masks IℓI_{\ell} and JℓJ_{\ell} are loaded into temporary registers with comparators that set the bit under the gates ‘a≤ba\le b’ and ‘a≥ba\ge b’. The state of this check bit controls a flip (gate ‘⊕\oplus’) of the entire register ff, holding a phase gradient state, that effects an inversion of the phase resulting from the subsequent in-place addition of the angles ϕℓ\phi_{\ell} into ff. Due to the ℓ\ell-dependence of RℓR_{\ell}, we prepare a different number of qubits in q4q_4 and q5q_5 in ∣+⟩\lvert+\rangle by using the bit masks IℓI_{\ell} and JℓJ_{\ell} as stencils, where we implement controlled Hadamard gates with CZ gates and T gates in the YY basis, TyT_y, as well as their inverses. After tagging the initial state using Toffoli gates, we set the qubit q0q_0 in the good subspace, using the same checks as before.

Ty†ZTy=H,(69)T_y^\dagger ZT_y=H, \tag*{(69)}

when the respective qubit in the bit mask was set to ∣1⟩|1\rangle. This circuit restricts how many qubits in q4q_4 and q5q_5 are prepared in ∣+⟩|+\rangle. The result is a precursor state, from which the target superposition state can be reached with two rounds of amplitude amplification in every subspace ∣ℓ⟩|\ell\rangle. As the circuit in Figure 12 demonstrates, the registers q4q_4 and q5q_5 may have a static part of ⌈log⁡(min⁡ℓRℓ+1)⌉\lceil\log(\min_\ell R_\ell+1)\rceil qubits, where free Hadamard gates suffice as a bit mask would never switch them off. The constrained subspace ∣Λℓ⟩=∑μ,νΛℓμν∣μ⟩∣ν⟩|\Lambda_\ell\rangle=\sum_{\mu,\nu}\Lambda_{\ell\mu\nu}|\mu\rangle|\nu\rangle is then tagged with quantum arithmetic circuits. From (B4) we can read out the two constraints for ∣μ⟩∣ν⟩|\mu\rangle|\nu\rangle:

  1. the index on q4q_4 is always lower than or equal to the index in q5q_5; μ≤ν\mu\le\nu; and

  1. the number ν⊕μ\nu\oplus\mu, derived from appending the binary representations of μ\mu and ν\nu has an upper bound ν⊕μ≤Rℓ⊕(N−1)\nu\oplus\mu\le R_{\ell} \oplus(N - 1).

Both constraints are tested with comparator circuits and the space in which they are mutually fulfilled is flagged. For the second constraint, the ℓ\ell-dependent upper bound ϑℓ\vartheta_{\ell} has been loaded into a temporary register. Some ℓ\ell-dependent angles are loaded in as well: they are parameters for partial reflections that prevent overcooking the state in the amplification process. Bit masks, ϑℓ\vartheta_{\ell} parameters and angles may be loaded in at the beginning of the circuit, and uncomputed as it terminates. While the ideal state ∣Λℓ⟩\lvert\Lambda_{\ell}\rangle can be prepared exponentially well for all ℓ\ell, we nonetheless elect to set the bit q0q_0 in the ∣Λℓ⟩\lvert\Lambda_{\ell}\rangle subspace of the resulting state, by re-using the tagging circuits from the reflections on ∣Λℓ⟩\lvert\Lambda_{\ell}\rangle.

Comparison against sequential QPE

In this section, we compare the proposed SVD-THC methodology against the simplest alternative without compression, which we refer to as sequential QPE. In this approach, an ordinary single-geometry phase estimation is performed for each of the MM geometries. Each QPE has an individual rank and 1-norm, and quantum resource estimates depend on the average 1-norm λˉ\bar{\lambda} and average rank Rˉ\bar{R}, as well as length Lˉ=N+⟨R(R+1)/2⟩\bar{L} = N + \langle R(R + 1)/2 \rangle. This approach is somewhat different from the sliced-THC thresholding circuit of Sec. VIB, which slices BLISS-THC circuits for all MM geometries coherently, where each Hamiltonian is normalized with the worst-case 1-norm λmax⁡\lambda_{\max} so the circuit is valid at every geometry. As highlighted in Table VIII, sequential QPE, which could very well be used as an alternative to a thresholding circuit, beats a sliced-THC version of the circuit by factors of 3.3×\times to 3.8×\times in Toffoli gates. Against sequential QPE, the full-grid Toffoli improvement of the SVD-THC version of the thresholding circuit using the gate-optimal setting is 11.8×\times and 10.1×\times for rigid and relaxed BCE, 38.2×\times for glycine dipeptide, and 60.2×\times for alanine dipeptide, compared to the respective improvement factors of 41.4×\times, 34.9×\times, 144.5×\times and 200.6×\times on the sliced-THC version of the algorithm.

SystemMMRˉ\bar{R}Lˉ\bar{L}λˉ\bar{\lambda} (Ha)AVToff.λmax⁡\lambda_{\max} (Ha)AVToff.Ratio
BCE rigid7351.51,37935.791.82×10121.82 \times10^{12}2.38×10102.38 \times10^{10}41.886.95×10126.95 \times10^{12}8.35×10108.35 \times10^{10}3.5 ×\times
BCE relaxed7351.61,38636.451.85×10121.85 \times10^{12}2.42×10102.42 \times10^{10}50.457.14×10127.14 \times10^{12}8.36×10108.36 \times10^{10}3.5 ×\times
Glycine dipep.2,664134.39,172131.698.35×10148.35 \times10^{14}7.46×10127.46 \times10^{12}134.953.25×10153.25 \times10^{15}2.82×10132.82 \times10^{13}3.8 ×\times
Alanine dipep.5,184148.811,249149.962.17×10152.17 \times10^{15}1.85×10131.85 \times10^{13}153.687.72×10157.72 \times10^{15}6.15×10136.15 \times10^{13}3.3 ×\times

Table VIII. The two references for the thresholding circuit without compression. Sequential QPE sums MM independent single-geometry phase estimations at the average rank Rˉ\bar{R} and 1-norm λˉ\bar{\lambda} and average length Lˉ=N+⟨R(R+1)/2⟩\bar{L} = N + \langle R(R+1)/2 \rangle. The thresholding circuit with sliced-THC contains the same MM geometry-conditioned BLISS-THC circuits and schedules its Jacobi-Anger expansion at the sampled worst-case 1-norm λmax⁡\lambda_{\max} while reusing workspace. ‘Ratio’ is the relative improvement of sequential QPE over the sliced-THC thresholding circuit in terms of Toffoli gates. Both references are validated with CCSD, as SVD-THC is: each per-geometry factorization is fitted without correlation-energy reweighting and accepted at the smallest THC rank whose CCSD correlation-energy error is below 0.6 mHa.

We attribute this difference to how the two references are counted rather than to the algorithms themselves. First, the thresholding circuit of Figure 7 performs a phase estimation and its inverse in every iteration, and both multiplexed approaches are charged for the two, whereas sequential energy estimation requires only one. Second, the Jacobi-Anger schedule rounds λ/ϵ\lambda/\epsilon up to the next power of two, while the sequential reference uses the integer estimate ⌈πλˉ/ϵ⌉\left\lceil\pi\bar{\lambda}/\epsilon\right\rceil. At ϵ=1\epsilon= 1 mHa this rounding costs a factor of 1.83 and 1.80 for rigid and relaxed BCE, 1.99 for glycine dipeptide, whose λˉ/ϵ\bar{\lambda}/\epsilon lies only 0.5% above a power of two, and 1.75 for alanine dipeptide. The remaining 4–5% reflects a small difference in the per-query cost of the two estimates. We note that the worst-case norm, which exceeds λˉ\bar{\lambda} by 2.5–38%, falls on the same power of two as λˉ\bar{\lambda} in every system and therefore does not increase the query count, and that both references use the same phase-estimation normalization. When the two references are counted consistently, the improvement of SVD-THC takes the closed form,

I=McslicedcSVDNslicedNSVD,(70)\mathcal{I} = M \frac{c_{\mathrm{sliced}}}{c_{\mathrm{SVD}}} \frac{N_{\mathrm{sliced}}}{N_{\mathrm{SVD}}}, \tag*{(70)}

where cxc_x denotes the cost of a single query and NxN_x the Jacobi-Anger query count of one quantum circuit. Equation (70) is independent of the number of phase estimations per iteration, applies equally to the sequential and multiplexed references, and reproduces every full-grid improvement of Sec. VI B. For glycine dipeptide, for example, I=2,664×(6,420/59,212)×(1/2)≈144\mathcal{I} = 2{,}664 \times(6{,}420/59{,}212) \times(1/2) \approx144, where the factor of one half arises because the multiplexed 1-norm, 2.5 times the worst-case per-geometry value, places SVD-THC one step higher on the Jacobi-Anger schedule.

These results suggest that the choice of reference algorithm does not change the underlying advantage of SVD-THC, which is set by the number of geometries and by the relative cost of a compressed query: 9.2× and 12.9× a single-geometry query in Toffolis for glycine and alanine dipeptides, and 76× and 118× in active volume, which explains why the active-volume advantage is an order of magnitude smaller. The two references differ genuinely in two respects. Their qubit counts differ slightly and in either direction, 884 for sequential QPE against 813 for the thresholding circuit for glycine dipeptide, and 1139 against 1191 for alanine dipeptide. Sequential QPE also returns MM classical energies rather than coherent access to the ensemble. For plain energy estimation the sequential approach is therefore sufficient, whereas the thresholding and search applications that motivate this work require coherent access. Because the Jacobi-Anger schedule advances in powers of two, we also recommend that any improvement factor be reported together with the schedule and the number of phase estimations on which it is based.

References

  1. [1]M. Reiher, N. Wiebe, K. M. Svore, D. Wecker, and M. Troyer, Proceedings of the National Academy of Sciences 114, 7555 (2017).
  2. [2]Y. Cao, J. Romero, J. P. Olson, M. Degroote, P. D. Johnson, M. Kieferová, I. D. Kivlichan, T. Menke, B. Peropadre, N. P. D. Sawaya, S. Sim, L. Veis, and A. Aspuru-Guzik, Chemical Reviews 119, 10856 (2019).
  3. [3]B. Bauer, S. Bravyi, M. Motta, and G. K.-L. Chan, Chemical Reviews 120, 12685 (2020).
  4. [4]S. McArdle, S. Endo, A. Aspuru-Guzik, S. C. Benjamin, and X. Yuan, Reviews of Modern Physics 92, 15003 (2020).
  5. [5]A. M. M. Dalzell, S. McArdle, M. Berta, P. Bienias, C.-F. Chen, A. Gilyén, C. T. Hann, M. J. Kastoryano, E. T. Khabiboulline, A. Kubica, and et al., Quantum Algorithms: A Survey of Applications and End-to-end Complexities (Cambridge University Press, 2025).
  6. [6]R. Babbush, R. King, S. Boixo, W. Huggins, T. Khattar, G. H. Low, J. R. McClean, T. O’Brien, and N. C. Rubin, PRX Quantum 7, 020101 (2026).
  7. [7]R. Santagati, A. Aspuru-Guzik, R. Babbush, M. Degroote, L. Gonzalez, E. Kyoseva, N. Moll, M. Oppel, R. Parrish, N. Rubin, M. Streif, C. Tautermann, H. Weiss, N. Wiebe, and C. Utschig-Utschig, Nature Physics (2024).
  8. [8]Y. Zhou, J. Chen, J. Cheng, X. Cao, Y. Zhang, G. Karemorе, M. Zitnik, F. T. Chong, J. Liu, T. Fu, and Z. Liang, npj Drug Discovery 3, 1 (2026).
  9. [9]D. Hao, X. He, A. E. Roitberg, S. Zhang, and J. Wang, Journal of Chemical Theory and Computation 18, 978 (2022).
  10. [10]C. S. Tautermann, D. Seeliger, and J. M. Kriegl, Computational and Structural Biotechnology Journal 13, 111 (2015).
  11. [11]H. B. Schlegel, WIREs Computational Molecular Science 1, 790 (2011).
  12. [12]I. Kassal and A. Aspuru-Guzik, The Journal of Chemical Physics 131, 224102 (2009).
  13. [13]I. Kassal, S. P. Jordan, P. J. Love, M. Mohseni, and A. Aspuru-Guzik, Proceedings of the National Academy of Sciences 105, 18681 (2008).
  14. [14]E. Perola and P. S. Charifson, Journal of Medicinal Chemistry 47, 2499 (2004).
  15. [15]A. Gilyén, S. Arunachalam, and N. Wiebe, In Proceedings of the 30th ACM-SIAM Symposium on Discrete Algorithms (SODA 2019), pp. 1425-1444 10.1137/1.9781611975482.87 (2017).
  16. [16]T. E. O’Brien, M. Streif, N. C. Rubin, R. Santagati, Y. Su, W. J. Huggins, J. J. Goings, N. Moll, E. Kyoseva, M. Degroote, C. S. Tautermann, J. Lee, D. W. Berry, N. Wiebe, and R. Babush, Phys. Rev. Res. 4, 043210 (2022).
  17. [17]M. Steudtner, S. Morley-Short, W. Pol, S. Sim, C. L. Cortes, M. Loipersberger, R. M. Parrish, M. Degroote, N. Moll, R. Santagati, and M. Streif, Quantum 7, 1164 (2023).
  18. [18]T. Kosugi, H. Nishi, and Y.-i. Matsushita, npj Quantum Information 9, 112 (2023).
  19. [19]P.-W. Huang, G. Boyd, G.-L. R. Anselmetti, M. Degroote, N. Moll, R. Santagati, M. Streif, B. Ries, D. Marti-Dafcik, H. Jane, S. Simon, N. Wiebe, T. R. Bromley, and B. Koczor, npj Quantum Information 10.1038/s41534-026-01275-2 (2026).
  20. [20]S. Simon, R. Santagati, M. Degroote, N. Moll, M. Streif, and N. Wiebe, PRX Quantum 5, 010343 (2024).
  21. [21]G. H. Low and I. L. Chuang, Quantum 3, 163 (2019).
  22. [22]D. Poulin, A. Kitaev, D. S. Steiger, M. B. Hastings, and M. Troyer, Physical review letters 121, 010501 (2018).
  23. [23]D. W. Berry, M. Kieferová, A. Scherer, Y. R. Sanders, G. H. Low, N. Wiebe, C. Gidney, and R. Babbush, npj Quantum Information 4, 22 (2018).
  24. [24]A. Gilyén, Y. Su, G. H. Low, and N. Wiebe, Proceedings of the 51st Annual ACM SIGACT Symposium on Theory of Computing (STOC 2019), 193 (2019).
  25. [25]G. H. Low and I. L. Chuang, Phys. Rev. Lett. 118, 010501 (2017).
  26. [26]D. W. Berry, C. Gidney, M. Motta, J. R. McClean, and R. Babbush, Quantum 3, 208 (2019).
  27. [27]V. von Burg, G. H. Low, T. Hän­er, D. S. Steiger, M. Reiher, M. Roetteler, and M. Troyer, Physical Review Research 3, 033055 (2021).
  28. [28]M. Motta, E. Ye, J. R. McClean, Z. Li, A. J. Minnich, R. Babbush, and G. K.-L. Chan, npj Quantum Information 7, 83 (2021), arXiv:1808.02625.DOI
  29. [29]E. G. Hohenstein, R. M. Parrish, and T. J. Martínez, The Journal of Chemical Physics 137, 044103 (2012).
  30. [30]J. Lee, D. W. Berry, C. Gidney, W. J. Huggins, J. R. McClean, N. Wiebe, and R. Babbush, PRX quantum 2, 030305 (2021).
  31. [31]G. H. Low, R. King, D. W. Berry, Q. Han, A. E. DePrince III, A. F. White, R. Babbush, R. D. Somma, and N. C. Rubin, Physical Review X 15, 041016 (2025).
  32. [32]I. Loaiza and A. F. Izmaylov, Journal of Chemical Theory and Computation 19, 8201 (2023).
  33. [33]A. Caesura, C. L. Cortes, W. Pol, S. Sim, M. Steudtner, G.-L. R. Anselmetti, M. Degroote, N. Moll, R. Santagati, M. Streif, and C. S. Tautermann, PRX Quantum 6, 030337 (2025).
  34. [34]C. L. Cortes, D. Rocca, J. F. Gonthier, P. J. Ollitrault, R. M. Parrish, G.-L. R. Anselmetti, M. Degroote, N. Moll, R. Santagati, and M. Streif, Phys. Rev. A 110, 022420 (2024).
  35. [35]N. Fedik, W. Li, N. Lubbers, B. Nebgen, S. Tretiak, and Y. W. Li, Digital Discovery 4, 1158 (2025).
  36. [36]J. Welch, D. Greenbaum, S. Mostame, and A. Aspuru-Guzik, New Journal of Physics 16, 033040 (2014).
  37. [37]L. SIROVICH, Quarterly of Applied Mathematics 45, 561 (1987).
  38. [38]Z. Wang, B. McBee, and T. Iliescu, Journal of Computational and Applied Mathematics 307, 374 (2016), 1st Annual Meeting of SIAM Central States Section, April 11-12, 2015.
  39. [39]D. Motlagh and N. Wiebe, PRX Quantum 5, 020368 (2024).
  40. [40]R. Babbush, C. Gidney, D. W. Berry, N. Wiebe, J. McClean, A. Paler, A. Fowler, and H. Neven, Phys. Rev. X 8, 041015 (2018).
  41. [41]A. M. Childs, D. Maslov, Y. Nam, N. J. Ross, and Y. Su, Proceedings of the national academy of sciences 115, 9456 (2018).
  42. [42]I. D. Kivlichan, J. McClean, N. Wiebe, C. Gidney, A. Aspuru-Guzik, G. K.-L. Chan, and R. Babbush, Phys. Rev. Lett. 120, 110501 (2018).
  43. [43]T. Frotzheim, M. Müller, A. Hansen, and S. Grimme, ChemRxiv 10.26434/chemrxiv-2025-bjxvt (2025).
  44. [44]C. Bannwarth, E. Caldeweyher, S. Ehlert, A. Hansen, P. Pracht, J. Seibert, S. Spicher, and S. Grimme, WIREs Computational Molecular Science 11, e1493 (2021).
  45. [45]P.-O. Löwdin, The Journal of Chemical Physics 18, 365 (1950).
  46. [46]G. H. Low, V. Kliuchnikov, and L. Schaeffer, Quantum 8, 1375 (2024).
  47. [47]L. K. Grover, in Proceedings of the Twenty-Eighth Annual ACM Symposium on Theory of Computing, STOC ’96 (Association for Computing Machinery, New York, NY, USA, 1996) p. 212–219.
  48. [48]C. Durr and P. Hoyer, arXiv e-prints , quant-ph/9607014 (1996).
  49. [49]R. Babbush, D. W. Berry, Y. R. Sanders, I. D. Kivlichan, A. Scherer, A. Y. Wei, P. J. Love, and A. Aspuru-Guzik, Quantum Science and Technology 3, 015006 (2017).
  50. [50]Y. Su, D. W. Berry, N. Wiebe, N. Rubin, and R. Babbush, PRX Quantum 2, 040332 (2021).
  51. [51]F. H. da Jornada, M. Lostaglio, S. Pallister, B. Şahinoğlu, and K. I. Seetharam, arXiv e-prints , arXiv:2504.06348 (2025).arxiv.org/abs/2504.06348
  52. [52]A. Dudkiewicz, A. F. White, G. Hao Low, A. F. DePrince III, M. P. Harrigan, M. Kieferova, R. Babbush, D. W. Berry, and N. C. Rubin, arXiv e-prints , arXiv:2607.15358 (2026).

Paper details

vlnc:W_8ttme6te
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

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

  • Physics and Astronomy

AI

View statement
Looking for related papers…
Contents