Quantum Davidson Algorithms

• Quantum Davidson algorithms are hybrid quantum–classical methods that iteratively construct subspaces for accurately determining ground and excited eigenstates.
• They employ quantum imaginary-time evolution and LCU-based corrections to efficiently evaluate matrix elements, significantly reducing measurement overhead.
• Applications in quantum chemistry, dynamics, and strongly correlated materials demonstrate scalable and precise eigenstate computation in complex quantum systems.

Quantum Davidson algorithms refer to a class of hybrid quantum–classical eigensolvers that adapt the classical Davidson (and Jacobi-Davidson) methods to quantum architectures, with a primary focus on the iterative determination of ground and excited eigenstates of large Hermitian operators, most notably quantum Hamiltonians. These algorithms operate by adaptively constructing a low-dimensional Krylov or subspace basis on which the projected eigenproblem can be solved classically, exploiting quantum resources to efficiently prepare basis states and evaluate matrix elements, and classically managing subspace updates, residual evaluation, and diagonalization. Recent advances encompass the Quantum Davidson (QDavidson) family, Quantum Jacobi-Davidson (QJD), Sample-Based QJD (SBQJD), and Delta-Davidson (DELDAV) variants. Applications extend to quantum chemistry, strongly correlated materials, quantum dynamics, and the calculation of molecular response properties.

1. Classical Davidson and Its Quantum Extensions

The Davidson method is a subspace iteration technique designed to access extreme eigenstates of large sparse Hermitian matrices. Given a Hermitian $H$ and an initial reference $|\psi_0\rangle$, the classical algorithm iteratively builds a trial subspace $\{v_1, \dots, v_m\}$ (typically from residuals or correction vectors) and solves the projected Hamiltonian $H' = V^\dagger H V$ to estimate eigenvalues and eigenvectors. The convergence is accelerated by preconditioning residuals with the (approximate) diagonal of $H$. The Jacobi-Davidson variant replaces first-order updates with Newton-type corrections, achieving quadratic local convergence.

Quantum extensions retain this subspace structure but utilize quantum circuits for state preparation, matrix-element measurement, and nonunitary transformations (e.g., LCU-based corrections or QITE), while the subspace expansion, orthonormalization, diagonalization, and preconditioning remain classical. Algorithms such as QDavidson (Tkachenko et al., 2022), its multi-reference generalization (Berthusen et al., 2024), QJD/SBQJD (Zhang et al., 2 Feb 2026), and DELDAV (Guan et al., 2020) implement these hybrid strategies with distinct approaches to subspace growth, correction calculation, and quantum circuit primitives.

2. Quantum Davidson Algorithm (QDavidson): Methodology and Implementation

QDavidson operates iteratively as follows:

• Subspace Construction: Start from initial reference(s) $|\chi_1\rangle$; at each iteration, the projected eigenproblem is solved in the subspace spanned by the current set of basis states; the solution yields approximate eigenpairs $(E_k, |\psi_k\rangle)$ for target states.
• Residual and Correction: Calculate the residual vector $r_k = (H - E_k I) |\psi_k\rangle$. The next basis vector is generated from a preconditioned correction:
  • In classical Davidson: $w_k(i) \approx r_k(i)/(H_{ii} - \sigma)$.
  • In QDavidson: approximation of the action of $(H - \sigma I)^{-1}$ by a short quantum imaginary-time evolution, e.g., $|\psi_0\rangle$0, mapped onto a unitary transformation using QITE (Tkachenko et al., 2022, Berthusen et al., 2024).
• Orthonormalization: The new candidate is orthogonalized against the current subspace via Gram–Schmidt or equivalent overlap subroutines.
• Matrix-element Evaluation: Subspace Hamiltonian and overlap matrices, $|\psi_0\rangle$1, $|\psi_0\rangle$2, are assembled from quantum circuit measurements. Matrix elements $|\psi_0\rangle$3 (for Pauli string $|\psi_0\rangle$4) are accessed using Hadamard- or SWAP-test circuits.
• Classical Diagonalization: Solve $|\psi_0\rangle$5 to update eigenpairs.

The workflow continues until convergence criteria—typically $|\psi_0\rangle$6 or stability in $|\psi_0\rangle$7—are met. Circuit depth per iteration scales with basis-state preparation and QITE steps, not with the total number of basis vectors, yielding modest requirements even as Hilbert space size grows (Tkachenko et al., 2022, Berthusen et al., 2024).

3. Quantum Jacobi-Davidson (QJD) and Sample-Based QJD (SBQJD)

QJD generalizes QDavidson by employing a Newton-type correction equation,

$|\psi_0\rangle$8

for the correction vector, where $|\psi_0\rangle$9 is the current Ritz vector and $\{v_1, \dots, v_m\}$0 the residual. The correction is projected onto the orthogonal complement of $\{v_1, \dots, v_m\}$1, offering at least quadratic local convergence for nontrivial overlap with the eigenstate (Zhang et al., 2 Feb 2026).

• LCU Correction Preparation: Nonunitary updates are mapped to quantum-linear-combination-of-unitaries (LCU) circuits, with amplitude amplification used to boost success probability and ancilla overhead $\{v_1, \dots, v_m\}$2 for $\{v_1, \dots, v_m\}$3 unitaries in the decomposition.
• Preconditioning: Approximate inversion typically uses a diagonal preconditioner $\{v_1, \dots, v_m\}$4.
• Sample-Based QJD (SBQJD): Prior to QJD iterations, the reference is optimized via Sample-Based Quantum Diagonalization (SQDiag): measurement in the computational basis identifies the $\{v_1, \dots, v_m\}$5 most probable basis states, $\{v_1, \dots, v_m\}$6 is diagonalized in that subspace, and the leading eigenvectors are used to initialize the algorithm, dramatically reducing iteration count when the wavefunction is dominated by a few determinants (common in quantum chemistry and diagonally dominant models).

QJD and SBQJD demonstrate an order-of-magnitude reduction in iteration counts and Pauli measurement requirements compared to QDavidson. For example, to achieve $\{v_1, \dots, v_m\}$7 energy accuracy on an 8-qubit diagonally dominant example, QD requires $\{v_1, \dots, v_m\}$8–$\{v_1, \dots, v_m\}$9 iterations, QJD $H' = V^\dagger H V$0–$H' = V^\dagger H V$1, and SBQJD just $H' = V^\dagger H V$2–$H' = V^\dagger H V$3; corresponding measurement counts show advantages of $H' = V^\dagger H V$4–$H' = V^\dagger H V$5 and $H' = V^\dagger H V$6–$H' = V^\dagger H V$7 for QJD and SBQJD, respectively (Zhang et al., 2 Feb 2026).

4. Delta-Davidson (DELDAV) for Interior Eigenproblems

The Delta-Davidson (DELDAV) method addresses the failure of standard Davidson and related subspace methods to access interior eigenstates in highly degenerate, dense spectral regions. DELDAV employs a Chebyshev polynomial-based delta filter centered at energy $H' = V^\dagger H V$8 to construct a narrowband filter,

$H' = V^\dagger H V$9

where $H$0 is a rescaled Hamiltonian and $H$1 are Chebyshev polynomials; this operator selectively amplifies components within a narrow energy window. The filtered vector is incorporated into the subspace, which is then used for Ritz-Galerkin projection and diagonalization to extract Ritz pairs near $H$2.

• Algorithmic Steps: Delta filtering, orthonormalization (Daniel–Gragg–Kaufman–Stewart), subspace rotation, and residual convergence tests.
• Cost and Scaling: The dominant cost is $H$3 sparse operator–vector (matvec) products and $H$4 memory. DELDAV avoids $H$5 factorization required by shift-invert methods and converges robustly for high-density central spectra (Guan et al., 2020).

Applications include identification of many-body localization, quantum chaos diagnostics, and the study of volume-law entanglement in thermalized many-body states.

5. Applications in Quantum Chemistry and Dynamics

Davidson-type algorithms enable efficient computation of excitation energies, response properties, and dynamical observables on quantum computers without explicit construction or diagonalization of exponentially large matrices:

• Molecular Response Properties: In quantum self-consistent linear response (q-sc-LR) theory, the electronic Hessian $H$6 is Hermitian and typically large. Quantum Davidson methods apply by evaluating Hessian–vector products on quantum hardware using parameter-shift or finite-difference circuits and constructing a subspace on which the projected eigenproblem is solved. Applications include calculation of excitation energies, static polarizabilities, $H$7 dispersion coefficients, and damped (complex) linear response for x-ray absorption spectra (Reinholdt et al., 2024).
• Quantum Dynamics: Once the subspace is converged, propagating time evolution is reduced to exponentiation in the low-dimensional subspace, $H$8 with subspace-projected $H$9. Numerical studies for Heisenberg chains up to $|\chi_1\rangle$0 qubits show that QDavidson achieves fidelity $|\chi_1\rangle$1 for $|\chi_1\rangle$2, with subspace size scaling sub-exponentially with $|\chi_1\rangle$3 and shallow per-iteration circuit depths (Berthusen et al., 2024).
• Strong Correlation: Unitary coupled-cluster-based subspaces within QDavidson and SBQJD outperform classical projected coupled cluster at strong correlation, retaining high accuracy in stretched molecular geometries (Reinholdt et al., 2024).

6. Resource Costs, Convergence Behavior, and Numerical Benchmarks

Quantum Davidson-type algorithms have been benchmarked against alternative Krylov, VQE, and diagonalization approaches:

Method	Iterations (8q diag dom)	Measurement Ratio	Convergence Rate	Circuit Depth per Iter	Best Application Scenario
QD	77–146	1×	Linear	Shallow	Generic ground- or excited-state problems
QJD	16–76	1/5–1/20×	Quadratic	Deeper (LCU, amplif.)	High-accuracy, fault-tolerant scenarios
SBQJD	1–3	1/10–1/100×	Quadratic	Deeper	Optimizable reference, few-determinant wfns

• Resource Estimates: Qubit count equals system register plus a logarithmic number of ancillas for LCU and 1–2 for Hadamard- or SWAP-tests. Circuit depth per iteration for QD is generally shallower; QJD and SBQJD require deeper correction circuits but converge in far fewer steps (Zhang et al., 2 Feb 2026).
• Benchmark Results: QJD/SBQJD demonstrate total measurement and iteration costs an order of magnitude or more smaller than QD. Example: 8-qubit Hamiltonian—QD needs $|\chi_1\rangle$4 Pauli measurements, QJD $|\chi_1\rangle$5, SBQJD $|\chi_1\rangle$6 (Zhang et al., 2 Feb 2026).
• Scalability: QDavidson–type subspace size grows sub-exponentially with system size for local models, with circuit depth per iteration determined by state-prep and QITE components, never by exponential scaling (Berthusen et al., 2024).

7. Limitations, Open Questions, and Future Directions

Quantum Davidson algorithms are subject to several open challenges and areas for improvement:

• Shot Noise and Error Mitigation: Statistical uncertainty in matrix element measurement (due to finite shots) and hardware noise can affect non-Hermiticity of projected matrices and convergence stability. Mitigation strategies include higher-order finite-difference stencils, noise symmetrization, error-mitigated estimation, and adapted shot allocation (Reinholdt et al., 2024).
• Preconditioning and Correction Step Engineering: Effective diagonal preconditioners or short imaginary-time unitaries are critical. The development of deeper or block preconditioners, analytical Hessian–vector circuit constructions, and direct parameter-shift implementations is an ongoing area (Tkachenko et al., 2022, Reinholdt et al., 2024).
• Extending to Interior Eigenstates: DELDAV is specifically designed to overcome the exponentially small energy gaps in dense interior spectra, which stall classical or naive Krylov methods (Guan et al., 2020).
• Scalability and Hardware Realization: Subspace solvers remain among the most resource-efficient post-NISQ eigenstate algorithms, but the exponential orbital Hilbert space limits both circuit and classical memory resources for very large systems. Active-space truncation, operator-pool adaptation, and subspace recycling (block-Davidson) are promising directions.
• Generalization to Excited-State Properties and Dynamics: The block and multireference generalizations enable efficient access to low-lying and interior states (see multi-reference QDavidson (Berthusen et al., 2024)), crucial for spectroscopy, transport, and non-equilibrium phenomena.

This synthesis is grounded in the published results and algorithmic procedures reported in the following references: (Zhang et al., 2 Feb 2026, Berthusen et al., 2024, Guan et al., 2020, Reinholdt et al., 2024, Tkachenko et al., 2022).