Double-Factorized Quantum Phase Estimation

• DF QPE is a fault-tolerant quantum phase estimation method that compresses electronic Hamiltonians via double factorization for ground-state energy estimation.
• It reduces the O(N⁴) complexity to an L-fragment form, enabling systematic qubitized block-encoding tailored for resource-estimated, error-corrected simulations.
• Benchmark studies indicate that while DF QPE achieves high accuracy, its current deployment is constrained by significant hardware overhead and limited solvability regions.

Double-Factorized Quantum Phase Estimation (DF QPE) is a fault-tolerant quantum phase estimation workflow for electronic-structure Hamiltonians in which the two-electron integral tensor is first compressed by double factorization and the resulting Hamiltonian is then simulated through qubitized block-encoding. In the QB-GSEE benchmark, DF QPE is defined specifically as a resource-estimated, error-corrected algorithm with surface-code overheads explicitly costed, rather than as a noise-adaptive NISQ procedure. Its purpose is ground-state energy estimation at a prescribed accuracy and confidence level for Hamiltonians supplied in FCIDUMP form, with logical and physical costs derived from the structure of the double-factorized Hamiltonian and from assumptions about state preparation, overlap, and hardware performance (Bellonzi et al., 14 Aug 2025).

1. Formal definition and Hamiltonian representation

DF QPE begins from the standard second-quantized electronic Hamiltonian in a finite orbital basis,

$$H = \sum_{i,j} h_{ij}\, a_i^\dagger a_j + \frac{1}{2}\sum_{i,j,k,l} h_{ijkl}\, a_i^\dagger a_k^\dagger a_l a_j,$$

where $a_i^\dagger$ and $a_j$ are fermionic creation and annihilation operators and $h_{ij}$, $h_{ijkl}$ are one- and two-electron integrals provided via FCIDUMP. The defining structural step is the re-expression of the four-index tensor $h_{ijkl}$ in the double-factorized form

$$H = \sum_{i,j} h_{ij}\, a_i^\dagger a_j + \frac{1}{2} \sum_{\ell = 0}^{L-1} \lambda_\ell \Big(\sum_{i,j} [g^{(\ell)}]_{ij}\, a_i^\dagger a_j \Big)^2, \tag{DF-H}$$

where each fragment $\ell$ is specified by a scalar DF eigenvalue $\lambda_\ell$, a matrix $g^{(\ell)} \in \mathbb{R}^{N \times N}$, and the integer $a_i^\dagger$0, the DF rank (Bellonzi et al., 14 Aug 2025).

The benchmark description emphasizes the compression role of this representation. The original Hamiltonian contains $a_i^\dagger$1 two-body terms, whereas double factorization groups them into $a_i^\dagger$2 fragments, each expressed as the square of a quadratic fermionic form. For realistic chemistry, $a_i^\dagger$3 empirically and is upper bounded by $a_i^\dagger$4. The expression “double-factorized” therefore refers to the Hamiltonian compression technique internal to the QPE pipeline, not to a modification of phase estimation itself (Bellonzi et al., 14 Aug 2025).

After factorization, the Hamiltonian is mapped to qubits through Jordan–Wigner encoding in the QB-GSEE benchmark, and the quantum algorithm proceeds through qubitized block-encoding. In that setting, the target of phase estimation is the unitary encoding of $a_i^\dagger$5, from which the ground-state energy is inferred. This places DF QPE within the class of systematic, non-variational eigenvalue estimation methods: the benchmark text states that QPE with qubitization is systematic and non-variational, and can in principle obtain arbitrarily accurate eigenvalues (Bellonzi et al., 14 Aug 2025).

2. End-to-end pipeline in the QB-GSEE benchmark

The benchmarked DF QPE workflow begins with classical preprocessing from an FCIDUMP instance. The first step is to parse the one- and two-electron integrals and construct the fermionic Hamiltonian with $a_i^\dagger$6 spin-orbitals and $a_i^\dagger$7 electrons. The second step is the computation of the double-factorized decomposition, following the procedures attributed in the benchmark description to von Burg et al. and Motta et al. The resulting DF features are not used only for simulation; they are also treated as benchmark hardness descriptors. The explicitly listed quantities are the DF rank $a_i^\dagger$8, the ordered DF eigenvalues $a_i^\dagger$9, and the DF eigenvalue gap $a_j$0 (Bellonzi et al., 14 Aug 2025).

The benchmark adopts a fixed truncation threshold of $a_j$1 on the DF eigenvalues. Components below that threshold are discarded, and the associated truncation error is assumed negligible. The stated motivation is prior evidence of approximately chemical-accuracy errors at that threshold for large systems together with weak dependence of logical resources on tighter truncation. The paper identifies this as a simplification relative to earlier work in which DF truncation was tuned against CCSD(T) truncation error (Bellonzi et al., 14 Aug 2025).

The truncated DF Hamiltonian is then mapped to a Pauli decomposition,

$a_j$2

with $a_j$3 qubits, typically $a_j$4 under Jordan–Wigner. The benchmark defines the Hamiltonian one-norm

$a_j$5

and represents the interaction structure by a hypergraph $a_j$6, in which vertices correspond to qubits and hyperedges correspond to nonzero Pauli strings. Feature extraction includes the Pauli string count $a_j$7, edge-order statistics $a_j$8, vertex-degree statistics $a_j$9, and edge-weight statistics $h_{ij}$0. The benchmark text states that LCU/qubitization-based simulation costs scale with $h_{ij}$1 and $h_{ij}$2, whereas Trotter-based methods would depend more directly on commutators and structure; DF QPE in QB-GSEE uses qubitization rather than Trotterization (Bellonzi et al., 14 Aug 2025).

At the simulation layer, the benchmark assumes qubitized QPE with DF block-encoding following von Burg et al. A block-encoding of $h_{ij}$3 is implemented for some normalization $h_{ij}$4, and qubitization is used to simulate $h_{ij}$5 within QPE without Trotterization error. The benchmark description attributes the practical implementation to pyLIQTR for double-factorized qubitized simulation and to Qualtran for physical costing. The role of double factorization is central because each fragment

$h_{ij}$6

can be implemented efficiently using orbital rotations and quadratic fermionic operations, and the total cost scales roughly linearly in $h_{ij}$7 times the cost per fragment (Bellonzi et al., 14 Aug 2025).

3. Ground-state preparation, accuracy targets, and costing model

The QPE stage is formulated under a promise model in which there exists an easy-to-prepare state $h_{ij}$8, or pure state $h_{ij}$9, having overlap at least $h_{ijkl}$0 with the true ground state $h_{ijkl}$1,

$h_{ijkl}$2

In the benchmark implementation of DF QPE, the initial state is taken to be the dominant Configuration State Function (CSF). The ground-state overlap is estimated by DMRG using the block2 code: DMRG yields an approximate ground state, and its overlap with the dominant CSF is used in the QPE failure analysis. The benchmark further states that this overlap parameter affects both the shot count and the overall runtime, since lower overlap requires more repetitions of QPE to recover the ground-state eigenvalue with the target confidence (Bellonzi et al., 14 Aug 2025).

The accuracy target is defined rigorously in terms of the spectrum $h_{ijkl}$3 and an output $h_{ijkl}$4 satisfying

$h_{ijkl}$5

For QB-GSEE, the targets are $h_{ijkl}$6, approximately chemical accuracy, and confidence $h_{ijkl}$7. The reported runtime and qubit counts are those needed to satisfy this accuracy/confidence pair, with all algorithmic parameters selected so that the combined failure probability from imperfect overlap, block-encoding errors, DF truncation error as treated in the benchmark, and hardware errors is at most $h_{ijkl}$8 (Bellonzi et al., 14 Aug 2025).

The complexity discussion in the benchmark is empirical rather than asymptotic in closed form, but several parameters are identified as dominant: the number of spin-orbitals $h_{ijkl}$9, the number of electrons $h_{ijkl}$0, the DF rank $h_{ijkl}$1, the Hamiltonian one-norm $h_{ijkl}$2, the overlap parameter $h_{ijkl}$3, and the target precision and failure rate. At the conceptual logical level, the benchmark describes the simulation cost per fragment as $h_{ijkl}$4, the total simulation-step cost as roughly linear in $h_{ijkl}$5,

$h_{ijkl}$6

and the qubitized simulation cost to precision $h_{ijkl}$7 as scaling like $h_{ijkl}$8, where $h_{ijkl}$9 is an effective Hamiltonian normalization closely related to $$H = \sum_{i,j} h_{ij}\, a_i^\dagger a_j + \frac{1}{2} \sum_{\ell = 0}^{L-1} \lambda_\ell \Big(\sum_{i,j} [g^{(\ell)}]_{ij}\, a_i^\dagger a_j \Big)^2, \tag{DF-H}$$0 and DF parameters. A rough structural form is therefore

$$H = \sum_{i,j} h_{ij}\, a_i^\dagger a_j + \frac{1}{2} \sum_{\ell = 0}^{L-1} \lambda_\ell \Big(\sum_{i,j} [g^{(\ell)}]_{ij}\, a_i^\dagger a_j \Big)^2, \tag{DF-H}$$1

with additional multiplicative overhead from amplitude amplification when the overlap is small and from QPE repetition to achieve the required confidence (Bellonzi et al., 14 Aug 2025).

The physical resource model uses surface-code error correction with catalyzed AutoCCZ factories. Five DF QPE solver configurations are described; all employ qubitized DF QPE, four catalyzed AutoCCZ factories for magic-state distillation, and the same algorithmic core, differing only in physical assumptions (Bellonzi et al., 14 Aug 2025).

Configuration	UUID	Assumptions
Baseline solver (serial)	5dad4064-cd11-412f-85cb-d722afe3b3de	$$H = \sum_{i,j} h_{ij}\, a_i^\dagger a_j + \frac{1}{2} \sum_{\ell = 0}^{L-1} \lambda_\ell \Big(\sum_{i,j} [g^{(\ell)}]_{ij}\, a_i^\dagger a_j \Big)^2, \tag{DF-H}$$2, $$H = \sum_{i,j} h_{ij}\, a_i^\dagger a_j + \frac{1}{2} \sum_{\ell = 0}^{L-1} \lambda_\ell \Big(\sum_{i,j} [g^{(\ell)}]_{ij}\, a_i^\dagger a_j \Big)^2, \tag{DF-H}$$3, single QPU, shots serial
Parallelized solver	4b07b89f-c66f-4e72-8c24-df3e4222cb41	Same hardware as baseline, but all shots parallelized across enough QPUs
Faster cycle, same error rate (serial)	2610d8de-bd3a-469e-9a80-473e8988755f	$$H = \sum_{i,j} h_{ij}\, a_i^\dagger a_j + \frac{1}{2} \sum_{\ell = 0}^{L-1} \lambda_\ell \Big(\sum_{i,j} [g^{(\ell)}]_{ij}\, a_i^\dagger a_j \Big)^2, \tag{DF-H}$$4, $$H = \sum_{i,j} h_{ij}\, a_i^\dagger a_j + \frac{1}{2} \sum_{\ell = 0}^{L-1} \lambda_\ell \Big(\sum_{i,j} [g^{(\ell)}]_{ij}\, a_i^\dagger a_j \Big)^2, \tag{DF-H}$$5, serial
Lower physical error rate, same cycle (serial)	5d768520-b3d0-4292-bbb4-9776fa128107	$$H = \sum_{i,j} h_{ij}\, a_i^\dagger a_j + \frac{1}{2} \sum_{\ell = 0}^{L-1} \lambda_\ell \Big(\sum_{i,j} [g^{(\ell)}]_{ij}\, a_i^\dagger a_j \Big)^2, \tag{DF-H}$$6, $$H = \sum_{i,j} h_{ij}\, a_i^\dagger a_j + \frac{1}{2} \sum_{\ell = 0}^{L-1} \lambda_\ell \Big(\sum_{i,j} [g^{(\ell)}]_{ij}\, a_i^\dagger a_j \Big)^2, \tag{DF-H}$$7, serial
Both faster and lower error (serial)	f6b36bde-be4a-4eee-975b-2c5f7e553f5f	$$H = \sum_{i,j} h_{ij}\, a_i^\dagger a_j + \frac{1}{2} \sum_{\ell = 0}^{L-1} \lambda_\ell \Big(\sum_{i,j} [g^{(\ell)}]_{ij}\, a_i^\dagger a_j \Big)^2, \tag{DF-H}$$8, $$H = \sum_{i,j} h_{ij}\, a_i^\dagger a_j + \frac{1}{2} \sum_{\ell = 0}^{L-1} \lambda_\ell \Big(\sum_{i,j} [g^{(\ell)}]_{ij}\, a_i^\dagger a_j \Big)^2, \tag{DF-H}$$9, serial

For each Hamiltonian and solver configuration, Qualtran determines the required code distance and logical qubit count, T-count and total logical cycles, runtime under the cycle-time and parallelization assumptions, and whether the resulting hardware success rate satisfies the required overall success probability. If that success rate is not achieved, no runtime is reported and the instance is counted as unsolved (Bellonzi et al., 14 Aug 2025).

4. Benchmark performance and solvability interpretation

Within QB-GSEE, a DF QPE instance is considered solved when the estimated ground-state energy satisfies the $\ell$0 accuracy target and the total runtime, including error-correction overhead, remains below the runtime requirement for the specified benchmark configuration. The benchmark does not compare a measured QPE output directly to a reference energy; instead, the resource model is constructed so that the stated accuracy and confidence follow from the assumed Hamiltonian representation, overlap estimate, and hardware model (Bellonzi et al., 14 Aug 2025).

The quantitative result reported for DF QPE, using one of the serial improved-hardware models, is the following: 131 tasks attempted, 4 tasks solved, and a solvability ratio of $\ell$1. The benchmark stresses that this solvability ratio is not simply the fraction $\ell$2. Rather, an SVM classifier is trained on problem-feature vectors and solver success or failure labels; then 10,000 synthetic points are sampled in a two-dimensional PCA latent space, and the fraction of latent points with predicted success probability greater than $\ell$3 defines the reported solvable-space fraction (Bellonzi et al., 14 Aug 2025).

The comparison with classical solvers is correspondingly stark. Optimized SHCI, defined in the benchmark as SHCI+PT plus orbital optimization and extrapolation, attains a solvability ratio of $\ell$4 and solves 148 of 226 attempted tasks. SHCI variants with fixed thresholds have solvability ratios between $\ell$5 and $\ell$6. DMRG, using the first run with lowest variational energy, attains a solvability ratio of $\ell$7 and solves 107 of 228 attempted tasks. By the benchmark’s ML-based latent-space metric as well as by task counts, the solvable region for DF QPE is therefore very small relative to SHCI and smaller than that of DMRG (Bellonzi et al., 14 Aug 2025).

The PCA solvability maps are interpreted carefully in the benchmark text. DMRG displays a structured boundary that correlates with lower entanglement or correlation indicators, while DF QPE exhibits only a very small high-solvability region in the latent space. The benchmark explicitly warns, however, that “unsolved” for DF QPE usually means the physical runtime estimate exceeds the allowed runtime or that no feasible hardware success rate exists under the stated assumptions. It is therefore not evidence that QPE, as an abstract algorithm, fails on those Hamiltonians; it is evidence that, under the benchmark’s current fault-tolerant resource estimates and hardware models, DF QPE is not yet practically deployable for most of the benchmark set (Bellonzi et al., 14 Aug 2025).

5. Relation to split-evolution QPE and alternative DF-QPE realizations

A distinct but directly relevant development is split-evolution quantum phase estimation (SE-QPE), which modifies canonical QPE for particle-conserving Hamiltonians by replacing controlled time evolution with a CSWAP-based interference construction between a target register and a reference register. In that framework, the connection to DF QPE arises when the time-evolution operator is implemented through a Trotterized double-factorized chemistry Hamiltonian. The SE-QPE analysis therefore intersects not with the qubitized DF QPE of QB-GSEE, but with a different realization of “double-factorized QPE”: QPE whose time-evolution blocks are built from Motta-style double factorization and Trotterization (Rowe et al., 16 Apr 2026).

In the SE-QPE paper, the double-factorized Hamiltonian after fermion-to-qubit mapping is written as

$\ell$8

with basis-rotation unitaries $\ell$9 implemented as products of Givens rotations. Time evolution is then approximated by first- or second-order Trotter formulas built from these DF primitives, and canonical QPE applies controlled powers of the resulting $\lambda_\ell$0. SE-QPE instead replaces each controlled-$\lambda_\ell$1 block by a split-evolution gadget $\lambda_\ell$2, using a factorization $\lambda_\ell$3 for the $\lambda_\ell$4-th phase bit. Because the two factors are identical functions of the same Hamiltonian, they share an eigenbasis, satisfying the condition under which SE-QPE preserves the phase-register statistics of canonical QPE up to a known offset (Rowe et al., 16 Apr 2026).

The resource consequences in that Trotterized DF setting are explicit. Over a range of FeMoco active spaces, SE-QPE is reported to reduce time-evolution resources, with asymptotic reductions of about $\lambda_\ell$5 in CX count, $\lambda_\ell$6 in $\lambda_\ell$7 count, and an asymptotic depth ratio of $\lambda_\ell$8 for CX layers. The advantage is strongest at higher phase powers, because the elimination of controlled-simulation overhead and the ability to run two uncontrolled evolutions in parallel outweigh the additive CSWAP overhead. The same work therefore recommends mixed strategies in which canonical controlled DF-Trotter blocks are used for lower phase bits and SE-QPE blocks for higher phase bits (Rowe et al., 16 Apr 2026).

This comparison highlights an important terminological point. In the QB-GSEE benchmark, DF QPE denotes fault-tolerant qubitized phase estimation on a double-factorized Hamiltonian. In the SE-QPE study, “DF QPE” refers to QPE whose evolution blocks are implemented by Trotterized double-factorized chemistry circuits. The common element is the same double-factorized representation of the chemistry Hamiltonian, but the simulation primitive differs: qubitization in the benchmarked resource study, Trotterized evolution in the split-evolution redesign (Bellonzi et al., 14 Aug 2025).

6. Strengths, bottlenecks, dataset bias, and projected directions

The benchmark attributes several strengths to DF QPE. First, QPE with qubitization is systematic and non-variational, and in principle supports arbitrarily accurate eigenvalue estimation. Second, the DF representation compresses realistic chemistry Hamiltonians in a way that directly affects simulation cost. Third, DF-derived descriptors are themselves informative hardness features: the appendix-level SHAP analysis reported in the benchmark identifies the double-factorized spectral gap as the most important feature for classifying solvability in some contexts, suggesting that DF-based parameters are effective predictors of algorithmic difficulty (Bellonzi et al., 14 Aug 2025).

The same benchmark also isolates the principal bottlenecks. The most prominent is fault-tolerant overhead. Even under what the text calls very optimistic hardware parameters, such as $\lambda_\ell$9 cycle time and $g^{(\ell)} \in \mathbb{R}^{N \times N}$0 physical error rate, DF QPE is feasible only for a small subset of instances under the imposed runtime limits. Physical resource requirements in logical qubits, magic-state factories, and T-count remain extremely large. The benchmark further states that most DF QPE failures arise because runtime exceeds limits, not because of an intrinsic algorithmic inability. Improvements in cycle time or physical error rate help but do not radically alter the overall solvability ratio within the current benchmark (Bellonzi et al., 14 Aug 2025).

Additional bottlenecks arise from preprocessing and state preparation. Overlap estimates rely on DMRG through block2, so for strongly correlated systems where DMRG itself is more difficult, the overlap estimation pipeline may become less reliable or more expensive. The fixed $g^{(\ell)} \in \mathbb{R}^{N \times N}$1 DF truncation is also treated as negligible by assumption; the benchmark notes that in extremely strongly correlated systems this may fail, implying either non-negligible model error or the need for tighter truncation, which would increase DF rank and cost (Bellonzi et al., 14 Aug 2025).

A major interpretive issue is benchmark composition. The paper states that many benchmark Hamiltonians are drawn from datasets tailored to SHCI and related classical approaches, including small and medium molecular systems with near-FCI energies already accessible to SHCI+PT and DMRG. The stated consequence is a bias toward classical solvers: optimized SHCI achieves near-universal solvability, DMRG performs substantially on low-entanglement systems, and DF QPE is evaluated precisely where classical approaches are already strong while fault-tolerant quantum hardware remains far from practical scale. The benchmark therefore argues for adding more challenging, strongly correlated systems, including larger multireference problems, long-range correlated systems, challenging transition-metal complexes beyond current SHCI reach, and strongly correlated model Hamiltonians such as Hubbard and Heisenberg instances in regimes known to be difficult for classical methods (Bellonzi et al., 14 Aug 2025).

The future directions named in the benchmark are correspondingly algorithmic, architectural, and methodological. They include more efficient tensor hypercontraction or better double-factorization strategies; better state preparation through MPS-based state preparation or more advanced initial ansätze for QPE; more efficient block encodings and symmetry exploitation, including BLISS; hardware co-design; and refined resource models with improved error budgets and adaptive truncation. A plausible implication is that DF QPE is being framed less as a presently competitive solver on SHCI-friendly chemistry instances than as a structured target for coordinated progress in Hamiltonian compression, initialization, fault-tolerant architecture, and benchmark design (Bellonzi et al., 14 Aug 2025).

In that sense, the benchmark’s central conclusion is narrowly scoped rather than universal. It states that “DF QPE is currently constrained by hardware and algorithmic limitations,” while simultaneously treating it as a method whose competitiveness may change as hardware improves and as benchmark instances shift toward strongly correlated, classically difficult regimes. The current evidence therefore supports a differentiated view: DF QPE is not shown to be a fundamentally inadequate method, but it is shown to be a method whose practical deployment is presently dominated by resource overhead and by the choice of problem instances used to evaluate it (Bellonzi et al., 14 Aug 2025).