Quantum Krylov Methods

• Quantum Krylov methods are algorithms that construct reduced subspaces via successive Hamiltonian applications for efficient quantum simulation and eigenvalue estimation.
• They integrate techniques like Lanczos, Arnoldi, and Gaussian-power protocols to optimize measurement efficiency and mitigate sampling errors on both NISQ and fault-tolerant devices.
• Applications include ground- and excited-state energy estimation, Green’s function computation, and diagnostics of operator growth and quantum chaos in many-body systems.

Quantum Krylov methods are a class of quantum algorithms that generalize the classical Krylov subspace techniques—such as Lanczos and Arnoldi—for large-scale eigenvalue problems, operator dynamics, quantum state evolution, and quantum simulation. Their core involves systematically constructing a reduced subspace generated from successive applications of a generator (typically the Hamiltonian or Liouvillian) on a reference state or operator, then performing variational projection or tridiagonalization within this subspace. Quantum Krylov algorithms underpin numerous contemporary methodologies in quantum simulation, quantum chemistry, quantum dynamics, many-body chaos, and quantum algorithmics, with several competing architectures for subspace construction, measurement, and error control.

1. Foundations of Quantum Krylov Subspaces

Given a Hermitian Hamiltonian $H$ and a state $|\psi_0\rangle$, the classical $m$-dimensional Krylov subspace is

$$\mathcal{K}_m(H,|\psi_0\rangle) = \mathrm{span}\{|\psi_0\rangle,\, H|\psi_0\rangle,\, H^2|\psi_0\rangle, \dots, H^{m-1}|\psi_0\rangle\}.$$

In quantum computing, explicit formation of $H^k|\psi_0\rangle$ is infeasible due to non-unitarity, so quantum Krylov protocols use physical evolution (discrete real/imaginary-time steps or polynomial filters) to approximate Krylov vectors. Alternatively, for operator growth and quantum chaos, Krylov subspaces are generated in operator/Liouvillian space via repeated action of $\mathcal{L}_H[O] = [H,O]$ or its generalizations to open-system dynamics (Nandy et al., 2024, 2207.13603, Chen et al., 2024, Adhikari et al., 2022).

Lanczos and Arnoldi Procedures

For Hermitian generators, Lanczos tridiagonalization constructs an orthonormal Krylov basis $\{|\phi_k\rangle\}$ via the three-term recursion: $$H|\phi_k\rangle = \beta_{k+1}|\phi_{k+1}\rangle + \alpha_k|\phi_k\rangle + \beta_k|\phi_{k-1}\rangle,$$ yielding a tridiagonal matrix whose extremal eigenvalues rapidly converge to those of $H$ up to exponentially small error in $m$ (Nandy et al., 2024).

For non-Hermitian or unitary propagators (e.g., Floquet systems), Arnoldi iteration is used, generating an upper Hessenberg matrix and an orthonormal basis via modified Gram–Schmidt (Nizami et al., 2023).

2. Principal Quantum Krylov Algorithms

Quantum Krylov algorithms differ principally in the mechanisms of subspace construction, measurement schemes, and handling of sampling/statistical errors.

Measurement-Efficient Krylov Subspace Diagonalization

The Gaussian-power (GP) protocol constructs a basis of the form

$$f_k(H) = (H-E_0)^{k-1} \exp\left[-\frac{1}{2}(H-E_0)^2 \tau^2\right]$$

which, via a modified Hermite expansion, can be implemented as a linear combination of real-time evolutions $e^{-iHt}$ amenable to direct quantum implementation (Zhang et al., 2023). This method yields subspaces with low condition number and thus minimal amplification of measurement noise. The GP approach achieves measurement overhead as low as $O(1)$ in subspace ill-conditioning factor $\gamma$, versus $\gamma > 10^4$ for power, Chebyshev, or real-time evolved bases.

Real-Time and Stochastic Krylov Protocols

The real-time quantum Krylov framework constructs the subspace from

$$|\phi_n\rangle = e^{-i n \Delta t H} |\psi_0\rangle.$$

Projected Hamiltonian and overlap matrices are measured via Hadamard tests or symmetry-exploiting circuits. Errors from finite sampling in these observables propagate through the generalized eigenvalue problem and can result in severe instability if the overlap matrix $S$ is ill-conditioned. Various regularization and SVD-truncation schemes mitigate this (Kirby, 2024, Shen et al., 2022, Stair et al., 2022, Lee et al., 2024).

In the rQKD stochastic protocol, time-evolution operators are represented as LCU expansions of fast-forwardable fragments, with stochastic (Monte Carlo) sampling over paths to limit circuit depth. This approach is explicitly tailored to double-factorized Hamiltonians prominent in quantum chemistry (Stair et al., 2022).

Time-Reversal Krylov Methods

For systems with explicit time-reversal symmetry ($THT=-H$), the Krylov Time Reversal (KTR) protocol realizes real-valued overlap matrices using only symmetric time-evolution and projective measurements, bypassing ancilla-controlled unitaries. This yields notable reductions in depth and CNOT count, greatly expanding applicability to NISQ hardware (Mariella et al., 30 Jul 2025).

Unitary Decomposition and Variational Approaches

The QKUD protocol expresses the Hamiltonian as a sum of unitaries, constructing Krylov vectors through repeated application of $(X+X^\dag)/(2\epsilon)$ with $X=i e^{-i\epsilon H}$, thereby achieving $O(\epsilon^2)$ subspace error scaling and exact Krylov representation in the $\epsilon \to 0$ limit (Asthana, 12 Dec 2025). Variational Krylov (KVQA) approaches iteratively prepare orthonormal Lanczos vectors and tridiagonalize the Hamiltonian to compute molecular Green's functions and spectra (Jamet et al., 2021).

3. Error Analysis, Sampling Complexity, and Stability

Quantum Krylov algorithms are characterized by sensitivity to sampling error, gate noise, and subspace conditioning.

General Measurement and Statistical Error Bounds

A general measurement cost bound for quantum Krylov subspace diagonalization is

$$M_{\text{tot}} = \frac{\alpha(\kappa)\, \beta(d)}{16\eta^2},$$

where $\alpha(\kappa)$ encodes the probabilistic failure rate, $\beta(d)$ the subspace size/dependence, and $\eta$ the noise threshold. The basis-dependent ill-conditioning factor $\gamma$ can range from $O(1)$ (for GP) to $10^{12}$ (for power/RTE bases). This cost can be further decomposed into ideal projection cost, subspace size, and basis conditioning (Zhang et al., 2023).

Sampling error propagates through the variance of each matrix entry as $O(1/M)$; their impact on the generalized eigenvalue problem is exponentially amplified by the condition number of the overlap matrix (Lee et al., 2024). The use of shifting techniques and coefficient-splitting can reduce effective measurement cost by factors of $20{-}500$ for small molecular testcases (Lee et al., 2024).

Error Scaling Under Device and Sampling Noise

Rigorous analysis demonstrates that, after suitable regularization, the error in ground-state energy estimates from noisy quantum Krylov algorithms is

$\widetilde{E}_0-E_0 = O(\eta) + O(e^{-O(d)})$

with $\eta$ the maximal per-entry matrix error and $d$ the Krylov dimension, matching the observed numerics and resolving discrepancies with earlier $O(\eta^{2/3})$ bounds (Kirby, 2024). The necessity of SVD truncation of small eigenvalue modes in $S$ is essential to prevent amplified anti-variational bias.

4. Krylov Complexity, Chaos, and Operator Growth

Operator-space Krylov construction provides a robust framework for diagnosing quantum chaos and thermalization in many-body systems (Chen et al., 2024, Nandy et al., 2024, Adhikari et al., 2022). The Lanczos coefficients $b_n$ encode operator growth rates; linear $b_n\sim\alpha n$ is the universal signature of chaos per the operator growth hypothesis, and leads to exponential growth in Krylov complexity: $$\mathcal{K}(t) = \sum_n n |\phi_n(t)|^2 \sim e^{2\alpha t}.$$ However, exponential complexity growth alone is insufficient to diagnose chaos: a refined diagnosis requires the additional "Krylov metric" $K_{mn}$, quantifying operator size and off-diagonal suppression. A many-body system is a fast scrambler if and only if: (i) Krylov complexity grows exponentially; (ii) $K_{nn}\sim n^h$ with $h\in(0,1]$; (iii) off-diagonals $K_{mn}$ are negligible ($m\neq n$) (Chen et al., 2024).

In open quantum systems, operator Krylov complexity saturates at finite value due to non-Hermitian “decay” terms in Lindblad/Kossakowski evolutions, reflecting the suppression of complexity growth by dissipation (2207.13603).

5. Applications in Simulation, Chemistry, and Dynamics

Quantum Krylov methods are implemented for:

• Ground and Excited-State Energy Estimation: Quantum Krylov subspace diagonalization (QKSD) efficiently approximates extremal eigenvalues with reduced sampling demands and moderate quantum resources, showing advantages over quantum phase estimation, especially for early fault-tolerant or NISQ devices (Cortes et al., 2021, Stair et al., 2022).
• Molecular Properties: Krylov-based methods can obtain not only ground-state energies but also analytical derivatives (forces, gradients) and relaxed density matrices, enabling direct computation of forces and excited-state properties (Oumarou et al., 9 Jan 2025).
• Green’s Functions and Dynamical Correlations: Continued-fraction expansions in the Krylov/Lanczos basis yield efficient quantum algorithms for retarded Green's functions, facilitating integration with many-body frameworks such as DMFT (Jamet et al., 2021).
• Linear Systems: Krylov subspace linear solvers leveraging quantum simulation offer reduced scaling in condition number $\kappa$ in contrast to HHL/quantum Fourier solvers, with proven resource and error bounds (Xu et al., 2024).
• Quantum Dynamics Simulation and Fast-Forwarding: Real-time Krylov fast-forwarding algorithms trade off quantum circuit depth with classical subspace dimension to extend simulation times beyond device coherence (Cortes et al., 2022).
• Nonequilibrium Dynamics and Operator Spreading: The Krylov machinery naturally expresses operator growth and hydrodynamization in many-body systems, computes return probabilities, studies MBL/delocalization transitions, and evaluates spectral statistics for systems of up to $9\times 10^9$ Hilbert-space dimension on massively parallel classical hardware (Brenes et al., 2017).

6. Extensions: Time Dependence, Open Systems, and Field Theory

Quantum Krylov methods generalize to time-dependent and driven systems via Arnoldi-recursion with Floquet unitaries, providing new paradigms for studying localization/chaos transitions in periodically driven models (Nizami et al., 2023). In open systems, the construction is adapted to the Lindbladian superoperator, with the operator Krylov chain mapping to a non-Hermitian tight-binding model whose edge localization enforces saturation of complexity (2207.13603, Nandy et al., 2024).

In quantum field theory, Krylov complexity provides a direct analog of the “complexity=volume” holographic conjecture. In some free theories, the Krylov basis coincides with the Fock basis, and the complexity equals mean occupation number, showing true volume-law scaling (Adhikari et al., 2022).

7. Future Directions and Limitations

Outstanding challenges include optimizing measurement and circuit protocols for efficient NISQ/hybrid deployment, robust error mitigation and regularization, deeper connection of Krylov complexity to quantum computational advantage, and extension to higher-dimensional, time-dependent, non-Markovian channels, and non-Hermitian or block-encoded operators (Nandy et al., 2024).

Krylov-based approaches have robust theoretical justification, practical empirical benchmarks, and are already enabling advances in both quantum algorithmics and many-body science. However, their resources and stability remain ultimately conditional on the spectrum of the generator, overlap with reference states, and the details of the quantum measurement and error model employed.