Elliptic Schedule
A family of variance-preserving interpolation schedules derived from conformal mapping, explicitly incorporating the condition number κ of the target distribution and reducing to the trigonometric schedule as κ approaches 1.
Timeline (last 90 days)
The research proposes a new framework based on the analytic properties of ODE trajectories, deriving a family of elliptic schedules via conformal mapping and, paired with a customized discretization, obtaining the Zoloflow algorithm.
The elliptic schedule is a family of variance-preserving schedules whose parameterized form involves Jacobi elliptic functions and explicitly includes the condition number κ of the target distribution.
As κ approaches 1, the elliptic schedule degenerates into the classical trigonometric schedule, so it can be regarded as a generalization of the trigonometric schedule.
The elliptic schedule adaptively adjusts the path shape according to the geometric properties of the target distribution, making numerical integration more stable.
The ODE trajectory is extended to the complex time plane, and the larger the analytic region, the smaller the local error of numerical integration; the elliptic schedule is given by a conformal mapping that maps a strip region to an elliptic arc.
For log-concave distributions, the elliptic schedule requires only Õ(√d polylog(κ, 1/ε)) score evaluations to achieve a 2-Wasserstein error of ε.
The authors conduct experiments on synthetic benchmarks and existing image generation datasets, verifying that the elliptic schedule yields higher sampling quality when the condition number is large and is more computationally efficient when paired with Zoloflow.
All Facts (7)
The research proposes a new framework based on the analytic properties of ODE trajectories, deriving a family of elliptic schedules via conformal mapping and, paired with a customized discretization, obtaining the Zoloflow algorithm.
50%UnverifiedThe elliptic schedule is a family of variance-preserving schedules whose parameterized form involves Jacobi elliptic functions and explicitly includes the condition number κ of the target distribution.
50%UnverifiedAs κ approaches 1, the elliptic schedule degenerates into the classical trigonometric schedule, so it can be regarded as a generalization of the trigonometric schedule.
50%UnverifiedThe elliptic schedule adaptively adjusts the path shape according to the geometric properties of the target distribution, making numerical integration more stable.
50%UnverifiedThe ODE trajectory is extended to the complex time plane, and the larger the analytic region, the smaller the local error of numerical integration; the elliptic schedule is given by a conformal mapping that maps a strip region to an elliptic arc.
50%UnverifiedFor log-concave distributions, the elliptic schedule requires only Õ(√d polylog(κ, 1/ε)) score evaluations to achieve a 2-Wasserstein error of ε.
50%UnverifiedThe authors conduct experiments on synthetic benchmarks and existing image generation datasets, verifying that the elliptic schedule yields higher sampling quality when the condition number is large and is more computationally efficient when paired with Zoloflow.
50%