arXiv:2610.03669
Optimal kernel functions for linear combination of Hamiltonian simulation
Harriet Apel, Bjorn K. Berntson
Abstract
Linear combination of Hamiltonian simulation (LCHS) provides an asymptotically optimal framework for the quantum simulation of linear non-unitary dynamics. The LCHS approach approximately implements the Peano--Baker propagator through an integral representation determined by a kernel function. We establish an LCHS approximation theorem for a broadened class of admissible kernel functions, defined by natural complex-analytic conditions. In the standard time-independent block-encoding access model, we explicitly construct, for each target error $\varepsilon$, the unique kernel function in this class minimizing the query-cost functional and show that the minimum of this functional has the asymptotic expansion $\tfrac{2\mathrm{e}}π\big(\log\frac{1}{\varepsilon}-\log\log\frac{1}{\varepsilon}+o(1)\big)$ as ${\varepsilon\downarrow0}$. The provably optimal leading coefficient $2\mathrm{e}/π$ improves by a factor of approximately 2.1 on the coefficient obtained empirically by Low and Somma through numerical minimization over their family of optimally scaling kernel functions. We construct a spectrally convergent quadrature with explicit bounds on operator error and linear combination of unitaries (LCU) subnormalization, proving that this asymptotic improvement persists after discretization into a finite linear combination of unitaries suitable for implementation. Our results are based on Hardy space theory and convex analysis.
AI Summary
Linear combination of Hamiltonian simulation (LCHS) represents dissipative linear evolution as an integral of unitary propagators. This paper asks which integral kernel minimizes the trade-off between the Hamiltonian-simulation cutoff $R$ and the LCU subnormalization $\alpha_R$, subject to a prescribed approximation-error bound. It broadens admissible kernels to a Hardy-space class on lower-half-plane strips and proves an LCHS approximation theorem under these weaker boundary assumptions. Within this class, it solves the infinite-dimensional optimization exactly: the unique optimum uses the full lower half-plane and an explicit kernel with branch points at $\pm R$. Its parameters follow from a single scalar transcendental equation. The minimum continuous cost is $Q_\star(\varepsilon)=\frac{2e}{\pi}[\log(1/\varepsilon)-\log\log(1/\varepsilon)+o(1)]$. The leading coefficient is about $2.14$ times smaller than an empirical fit for the optimized Low–Somma kernel family. The proof uses inner–outer factorization, harmonic measure, and equality conditions in Jensen’s inequality. A double-exponential change of variables yields quadrature with exponential discretization convergence and double-exponential tail convergence, together with explicit operator-error and subnormalization bounds. These preserve the cost expansion after discretization, with total error at most $2\varepsilon$. Optimality concerns the proxy $R\alpha_R$ in a time-independent block-encoding model and a sufficient error bound, not unrestricted algorithmic complexity or actual operator error. State-dependent preparation factors and non-query gate costs are excluded; nonuniform node loading adds overhead. Concurrent independent work obtains the same kernel form and leading coefficient, whereas this analysis establishes global optimality and uniqueness within the stated class.
OpenAI GPT-5.6, GPT-6.0, and GPT-6.1 were used to assist with verifying and revising proofs, writing code for the numerical results, and editing the manuscript.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.
- Comments
- 66 pages, 1 figure