Hybrid Quantum-Classical Algorithm for Hamiltonian Simulation

Abstract: We introduce a hybrid classical-quantum algorithm for simulating a Hamiltonian of the form $H= \sum_{i=1}^K H_i = \sum_{i=1}^K H_{i_1} \otimes H_{i_2} \otimes \cdots \otimes H_{i_M}$. Given that the entries of all ${ H_{i_1}, H_{i_2} , \cdots , H_{i_M}}$ (for all $i$) are classically known, we present a procedure (with three variants) in which these operators are classically diagonalized, and then this information is fed into three possible quantum procedures to obtain the block-encoding of $H$. The evolution operator $\exp(-iHt)$ is then obtained using the standard block-encoding/quantum singular value transformation framework. In the case where ${H_i}_{i=1}^K$ commute pairwise, our method can be trivially extended to the case with time-dependent coefficients. We provide a detailed discussion of the efficient regime of our hybrid framework and compare it with existing quantum simulation algorithms. Our algorithm can serve as a useful complement to existing quantum simulation algorithms, thereby expanding the reach of quantum computers for practically simulating physical systems. As a side contribution, we will show how the recent technique called \textit{randomized truncation to a quantum state} developed by Harrow, Lowe, and Witteveen [arXiv preprint arXiv:2510.08518, 2025] can be applied to the context of quantum simulation and particularly quantum state preparation, for which the latter can be of independent interest.

• The paper introduces a hybrid simulation algorithm that combines classical diagonalization with quantum block-encoding for efficient Hamiltonian simulation.
• It details three quantum methodologies—spectral projector encoding, statistical sampling, and sparse randomized approximations—to achieve accurate block-encodings.
• The approach offers exponential savings in precision scaling for local Hamiltonian models and extends classical-to-quantum state preparation techniques.

Hybrid Quantum-Classical Algorithm for Hamiltonian Simulation

Introduction and Context

This work introduces a hybrid quantum-classical scheme for Hamiltonian simulation tailored to models composed as sums of tensor products of small, classically known Hermitian matrices. The central objective is to circumvent typical quantum input models—such as sparse-oracle access and LCU—by directly leveraging explicit classical matrix data and classical diagonalization, then proceeding with resource-efficient quantum steps to produce block-encodings of the global Hamiltonian. The protocol also supports a restricted class of time-dependent Hamiltonians and enables application of recent classical-quantum state-sparsification results to quantum-state preparation.

Problem Definition and Assumptions

The targeted Hamiltonians are of the structure:

$$H = \sum_{i=1}^K H_i = \sum_{i=1}^K H_{i_1} \otimes H_{i_2} \otimes \cdots \otimes H_{i_M},$$

where each component $H_{i_j}$ is a small $d \times d$ Hermitian matrix with fully specified classical entries. The approach exploits this explicit knowledge, allowing a hybrid treatment hinging on conventional classical linear algebra for diagonalization (which is tractable for the intended parameter regime) and subsequent block-encoding for quantum simulation.

The dominant regime of applicability is when, for each $i$, most tensor factors are identity matrices (or equally trivial to encode), that is, when each $H_i$ is supported nontrivially only on a small, fixed number $|R_i|$ subsystems with $|R_i| = \mathcal{O}(1)$. This captures common instances from quantum many-body physics, e.g., lattice systems with few-body couplings, and covers prominent models like TFIM, Heisenberg, and stabilizer Hamiltonians.

Algorithmic Structure

Classical Component

• Each $H_{i_j}$ is diagonalized efficiently, rendering spectra $\{ \lambda_{i_j}(k) \}$ and local eigenbases.
• For $H_i$, the spectrum is constructed as the products of the spectra of its nontrivial factors.
• The overall cost is $H_{i_j}$0, but this is reduced when identities predominate.

Quantum Component

The quantum processing adopts the block-encoding and quantum singular value transformation (QSVT) paradigm.

There are three primary quantum approaches:

1. Direct block-encoding via spectral projectors: With classically computed eigenbases, state preparation and projective block-encoding are carried out for all nontrivial tensor factors, composing the encoding of $H_{i_j}$1 by summing over spectral projectors weighted by eigenvalue products.
2. Statistical sampling approach: Instead of exhaustive enumeration, this method samples tensor product eigenstates according to absolute eigenvalue-weighted distributions, accumulates the block-encodings of these samples, and uses concentration bounds (central limit, Chebyshev) to ensure operator-norm approximation within prescribed error.
3. Sparse randomized state approximation (for nonnegative terms): Utilizing the classical randomized truncation technique from (Harrow et al., 9 Oct 2025), dense eigenstates are replaced with convex combinations of sparse approximate states, prepared with efficient quantum circuits, and used in the construction of block-encodings.

All approaches culminate in a block-encoding of $H_{i_j}$2, which is then used in a standard QSVT/Jacobi-Anger-based simulation procedure to approximate $H_{i_j}$3 up to additive error $H_{i_j}$4.

Complexity Highlights:

In the case that for all $H_{i_j}$5, $H_{i_j}$6 and $H_{i_j}$7 is small (qubits/qudits), quantum circuit depth scales as $H_{i_j}$8. When Hamiltonian terms are highly local and the system sparse in operator support, this yields practical complexity in both quantum time and ancillary width.

Specialization to Commuting, Time-dependent Hamiltonians

If the constituent $H_{i_j}$9 terms commute pairwise and appear with efficiently integrable time-dependent coefficients $d \times d$0, the algorithm is naturally generalized. The time-evolution operator can be realized without time-ordering complications using

$d \times d$1

with only minor overhead relative to the time-independent case.

Numerical and Structural Claims

• No quantum oracle access to $d \times d$2 (sparse/LCU) is required—all input is classical.
• The approach naturally supports structured Hamiltonians (tensor product sums with local terms) and achieves circuit complexity:

$d \times d$3

in the dominant regime.

• For lattice-type Hamiltonians, the method achieves exponential savings in precision scaling, e.g., logarithmic dependence on $d \times d$4, compared to some Trotter or Taylor-series lattice schemes, which scale polynomially or worse.
• By exploiting the randomized truncation method for quantum state preparation, the paper extends the applicability of classical-to-quantum state preparation techniques to previously intractable dense states at polynomial resource cost, provided error is tolerated and classical resources are available.

Relation to Prior Work

Most quantum simulation protocols (e.g., sparse-access, LCU, Trotterization, product formulas) assume quantum oracles for matrix entries/unitaries and do not capitalize on explicit classical input for tensor factors. The hybrid approach supplements existing methodologies by occupying the scenario where the whole classical tensor structure is available, and quantum oracles are unavailable or too costly to realize.

It particularly offers advantages when:

• The system size is large, but each Hamiltonian term acts locally (i.e., $d \times d$5 is independent of $d \times d$6).
• The matrix sparsity $d \times d$7 is high, making sparse-access-based methods inapplicable or highly inefficient.
• Time-dependent Hamiltonian simulation is required only for commuting terms—commuting, time-dependent simulation is handled efficiently, whereas the general time-ordered case remains outside the scope.

Implications and Prospects

Practical Impact

This hybrid scheme is highly relevant for near-term quantum devices and simulation applications where models are designed (or compiled) such that explicit, sparse, block-structured representations are known and prediagonalizable. It significantly lowers the overhead for embedding complex Hamiltonians into quantum circuits compared to black-box input-model-based methods.

The integration of classical randomized state truncation into quantum workflows further promises practical improvements for quantum state preparation beyond simulation. The approximate, resource-efficient preparation of dense states via classical pre-processing and quantum block-encoding offers a valuable tool for variational algorithms, initialization in quantum chemistry, and neural-network quantum state representations.

Theoretical Significance and Future Directions

The hybrid paradigm elucidated here presents an explicit algorithmic bridge between classical and quantum simulation, useful for both algorithm analysis and experimental implementations. Future directions include:

• Extension to highly nonlocal Hamiltonians with moderate $d \times d$8, potentially by exploiting symmetries or additional classical pre-processing.
• Investigation of more general time-dependent Hamiltonians, ideally extending beyond commuting families, possibly with controlled Trotter error analysis.
• Further development of quantum algorithms that use classical pre-computation to minimize quantum circuit complexity for other primitives (e.g., measurement, state tomography, and Lindbladian evolution).

Conclusion

This paper provides a rigorous and efficient hybrid framework for Hamiltonian simulation, leveraging explicit classical knowledge of Hamiltonian constituents and modern block-encoding techniques. It establishes a complementary path to quantum simulation, enhancing both the practical simulation capabilities of quantum computers for physical models and the broader quantum algorithmic toolkit, including quantum state preparation backed by advanced classical algorithms.