A preprint reports that a computational framework called SAKE closely reproduced an exact calculation for transporting nonlinear spectroscopic responses between nearby quantum models. In the main benchmark, its third-order expansion also reproduced the dominant pathway amplitudes and the way those pathways mixed in a four-level excitonic dimer.
Building a local map
SAKE is meant to provide a local map of how a response changes when a model’s controls shift. It uses nested forward-mode automatic differentiation—a way for software to calculate derivatives as inputs vary—to generate first-, second- and third-order derivative tensors for the system’s Liouvillian, the mathematical operator used to represent its dynamics. Those tensors feed a Duhamel expansion, a local series for transporting the response.
The validation used a four-level excitonic dimer with su(2) × su(2) symmetry. The authors compared second- and third-order transported operators with exact projected transport matrices from direct calculations. The working basis contained six third-order pathway diagrams covering rephasing and nonrephasing sectors, with GSB, SE and ESA pathways. The reference controls were λ = (J, κ) = (0, 0), while the target was a representative random point without optimization or tuning.
A close match, with visible mixing
The third-order result closely agreed with the exact projected transport matrix and reproduced the dominant pathway amplitudes and their mixing. The matrix was not simply diagonal: diagonal or near-diagonal weights were approximately 0.87–0.91, while lower-rephasing off-diagonal coefficients included 1.13, 0.74, 0.38 and 0.36. In practical terms, a pathway could retain much of its original weight while also feeding into other pathways; RP pathways mixed within RP, and NRP pathways within NRP.
Synthetic recovery
The authors then used SAKE as a local surrogate—a stand-in for a direct calculation during optimization—to recover model controls from three synthetic targets. Starting from the uncoupled reference, the procedure made six to eight direct trial calculations. The final relative pathway-state residuals, a measure of remaining mismatch, were 2.25 × 10−3, 2.52 × 10−3 and 7.92 × 10−6.
The corresponding Euclidean errors in the recovered controls were 7.26 × 10−2 meV, 6.72 × 10−2 meV and 2.96 × 10−4 meV. Because the targets were synthetic, these are benchmark reconstruction figures rather than measurements from an experiment.
Where the time was spent
The reported computational advantage came from reusing local charts during repeated optimization calls. A direct RP+NRP response took 2.17 seconds in the stated benchmark, while building a third-order chart took 9.1 seconds. Surrogate optimization took 0.044 seconds per step; including direct validation, an accepted step cost 11.5 seconds, with 2.2 seconds assigned to validation. Each recovery required six to eight steps, and the timings were hardware dependent.
The pathways did not move uniformly
Pathway mixing also varied by sector. Across the three recovered cases, the off-diagonal Frobenius fractions—the study’s measure of pathway mixing—were 0.708, 0.707 and 0.672. In RP, mixing mainly involved GSB and SE while ESA remained comparatively isolated. In NRP, GSB and ESA mixed while SE remained unmixed and was only renormalized.
A benchmark with a narrow reach
The benchmark has a built-in boundary. The finite reference pathway space was not exactly closed under transport, so projections discarded residuals outside the chosen basis. As a result, multiplying local projected matrices need not produce a transitive route between endpoints; the paper used a global projection for its final mechanistic interpretation.
The expansion is local and was evaluated through third order in a four-level dimer with two controls, so the reported agreement does not establish accuracy for larger systems, more controls or points farther from the reference. The inverse exercise also did not show that a single spectrum can uniquely separate the coherent and dissipative controls J and κ. The paper notes that a spectrum may constrain combinations of them more strongly and suggests additional waiting times, polarization sequences or physical priors for future experimental applications.
The study reports no statistical confidence intervals, and the synthetic recovery was not evaluated with measurement noise.
The implementation is available
The code and numerical validation data are stated to be openly available in the Duhamel Transport repository on the codex/sake-autodiff-dimer branch. An archived Zenodo release contains the implementation, tutorials, benchmark scripts and validation workflows. The manuscript is identified as arXiv:2608.20132v1, dated 20 August 2026, with no journal listed in the supplied metadata.
Paper data and sources
Original title: SAKE: Spectral Autodiff Kernel Expansion for Geometric Liouvillian Transport. A Differential-Geometric Framework for Response Transport in Quantum Dynamical Systems
Authors: Eric R. Bittner, Carlos Silva-Acuna, Hao Li, Simon Paiva-Ortega
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text