Spin Hamiltonian as Matrix-Free Linear Map

Abstract: We present an algorithm that computes the action of a generic spin Hamiltonian on a state vector on the fly, entirely avoiding explicit matrix assembly. This is achieved through mixed-radix indexing of the full tensor-product basis, which translates local spin operations into simple integer offsets. The result is an explicit framework for evaluating single- and two-site terms across arbitrary spin lattices, including mixed-spin systems. Our construction bridges the basis-indexing logic familiar from exact diagonalization with the matrix-free state-update philosophy of address-based frameworks. By writing the indexing logic in closed form, a single uniform loop applies to every site regardless of its local Hilbert-space dimension. The method is parallelizable and memory-conserving, and can be extended to restricted basis or truncated bosonic levels.

• The paper presents a matrix-free framework that applies generic spin Hamiltonians on-the-fly without assembling the full matrix.
• The method uses mixed-radix indexing to deterministically calculate local operator actions, reducing operator memory by factors of 10^6–10^8 while maintaining competitive matvec performance.
• The framework naturally extends to restricted bases and heterogeneous systems, enabling exact propagation in complex many-body simulations and facilitating efficient operator updates.

Matrix-Free Construction of Generic Spin Hamiltonians

Overview

The paper "Spin Hamiltonian as Matrix-Free Linear Map" (2606.02169) introduces an explicit, uniform framework for the on-the-fly application of generic spin Hamiltonians to state vectors without assembly of the global Hamiltonian matrix. The core innovation lies in mixed-radix indexing of the tensor-product basis, enabling deterministic logic for local operator actions regardless of site-dependent Hilbert space dimension or physical content. This approach is directly compatible with Krylov eigensolvers, is parallelizable, achieves dramatic reductions in operator memory usage, and extends naturally to restricted or truncated bases.

Mixed-Radix Integer Encoding and Operator Application

The exponential scaling of the Hilbert space dimension in many-body quantum systems motivates matrix-free techniques. The paper formalizes mixed-radix indexing, whereby a basis state $\{n_1, n_2, ..., n_N\}$ is mapped to a global integer index $I$ via $$I = \sum_{k=1}^N n_k \left(\prod_{\ell=k+1}^N d_\ell\right)$$ where $d_k$ are site-specific Hilbert space dimensions. This encoding generalizes bitwise representations used in spin-$1/2$ systems to arbitrary local dimensions (higher spins, bosons).

Given this index, the action of one-site or two-site Hamiltonian terms reduces to deterministic integer offset calculations—local operator actions correspond to simple shifts of site-specific indices. Diagonal terms are direct, while ladder operators (raising/lowering) and pairwise interactions require algebraic reconstruction of source indices. All necessary updates are performed per target state, enabling cache-friendly parallelization and minimal synchronization overhead.

Algorithmic Structure and Benchmarks

The algorithm is formalized for both single-site and two-site terms, where each target vector component is updated by local operators via index shift logic. For two-site interactions, both sites are decoded and candidate source indices generated. The algorithm only accesses the required source vector components and updates the target amplitude using local matrix elements, entirely bypassing explicit matrix construction or storage.

Benchmarking against sparse matrix assembly on canonical test cases (uniform spin-1 chain, mixed-spin chain, 2D spin-$1/2$ lattice) demonstrates several key findings:

• Operator memory usage is reduced by factors of $10^6$--$10^8$, with metadata consuming kilobytes versus gigabytes for explicit storage.
• Steady-state matrix-free matvec rates are competitive with sparse matvec, typically performing similar work per target index.
• Wall times favor matrix-free methods when operator assembly is a bottleneck, e.g., for repeated Hamiltonian rebuilds, parameter sweeps, or large systems.

Where full sparse assembly can yield slightly faster repeated matvecs, the matrix-free approach prevails for workflows requiring frequent operator modification or where explicit storage is prohibitive.

Extension to Restricted Bases

Beyond the full product basis, the method generalizes to restricted bases. In direct-product truncations (local restrictions), mixed-radix logic persists. For globally constrained sectors (e.g., fixed magnetization), basis states are stored non-contiguously and require a label-to-index lookup, implemented via hash tables, combinatorial ranking, or custom lexicographic indexing [DanceQ_2025, vcufar2026rimu]. The matrix-free update is then conditioned on admissibility via the lookup mechanism.

Numerical Complexity and Practical Scaling

The computational complexity per matrix-vector multiplication is $O(D(N+|P|))$, similar to sparse matvec for $D$-dimensional space and $I$0 interacting pairs. The operator-side memory scales as $I$1, and no expensive $I$2 matrix assembly step is incurred. Dense vector storage remains the primary constraint, but basis restriction reduces this bottleneck when physically justified.

Implications and Comparison with Conventional Methods

The matrix-free mixed-radix framework aligns with and extends prior address-based exact diagonalization approaches [QuSpinPartI2017, KahanamokuMeyer2023dynamite], but uniquely provides a closed-form, uniform, site-agnostic implementation for arbitrary local dimensions, mixed species, and spin-boson systems. Compared to symmetry-aware packages, tensor-network algorithms, or state-of-the-art sparse matvecs, this approach is optimal for systems lacking strong global symmetries, with heterogeneous local content, and frequent operator updates where assembly costs dominate.

Practically, this enables exact propagation of large systems at machine precision, provided dense vector storage fits in memory. The construction is robust for workflows involving time-dependent drives, disorder averaging, and mixed-spin geometries. Theoretically, the algebraic encoding of local state transitions provides a unified language for many-body operator application, facilitating extensions to higher-order terms and complex spin-boson models.

Contradictory and Strong Claims

• Operator memory for matrix-free updates is independent of Hilbert space dimension $I$3, contradicting conventional wisdom that storage must scale with $I$4 for all many-body Hamiltonians.
• No explicit sparse matrix needs to be formed, even for site-dependent or restricted bases, eliminating the assembly bottleneck commonly faced in large-scale exact diagonalization.
• Benchmark results demonstrate that ground state energies agree up to numerical precision ($I$5) across all methods, confirming the exactness of matrix-free propagation.

Future Directions

The deterministic mixed-radix algebra may inform scalable simulations of multi-species systems, facilitate hybrid classical-quantum algorithms, and provide a backend for matrix-free quantum Monte Carlo or tensor-network wrappers. Further optimizations—especially label-index lookup for constrained bases, and adaptation for time-dependent Hamiltonians—are promising. The approach may also enhance massively parallel implementations [VanBeeumen2020matrixfree] and serve as a bridge for state reduction techniques used in NMR and electron spin resonance.

Conclusion

The paper presents an explicit matrix-free construction for generic spin Hamiltonians via mixed-radix product-basis indexing. This method obviates operator storage scaling, matches sparse matvec performance, and is strongly advantageous for workflows characterized by frequent Hamiltonian modification or heterogeneous site content. Extensions to restricted bases, higher-order terms, and bosonic modes are seamless, positioning the framework as a robust tool in many-body numerical simulation.

---