Elliptic Sampling: Optimal Interpolant Schedules for Gaussians and Beyond

Cole Becker, Taylan Kargin, Kimia Hassibi, Asuman Ozdaglar, Pablo A. ParriloarXiv 2026 · arXiv

In the rapidly evolving world of generative AI, "diffusion models" and "stochastic interpolants" have become the gold standard for turning random noise into high-fidelity images. But beneath the hood of these models lies a tricky numerical problem: how do we efficiently solve the differential equations that transport noise to data?

A new paper from researchers at MIT, "Elliptic Sampling: Optimal interpolant schedules for Gaussians and Beyond," introduces a principled framework for designing these sampling paths. By looking at the problem through the lens of complex analysis, they’ve developed a new family of "elliptic schedules" that significantly outperform standard methods, especially when you're on a tight computational budget.

TL;DR

The authors introduce elliptic schedules and a new sampling algorithm called Zoloflow, which are derived from the analyticity properties of ODE trajectories. This approach achieves theoretical optimality for Gaussian targets and provides a massive boost in efficiency for complex tasks like image generation, producing high-quality results in as few as four steps.

What Problem Does This Paper Address?

Generative modeling typically involves two major design choices:

  1. The Interpolant Schedule: The path that defines how noise is gradually removed to reveal the data.
  2. The Discretization Scheme: The numerical method (like Euler or Heun) used to step along that path.

Historically, these have been designed separately and often heuristically. Standard schedules, like the "trig" schedule, work well for simple, uniform data but struggle when the data distribution is "stretched" or poorly conditioned (i.e., has a high condition number $\kappa$). The central question the authors tackle is: Given a fixed budget of neural network evaluations, how should we choose both the schedule and the numerical solver jointly?

Methods

The researchers propose a design criterion based on complex analyticity. They discovered that the numerical behavior of a solver depends on how far the solution of the probability-flow ODE can be extended into the "complex time" plane without hitting a singularity.

Using conformal mapping, they solved for the schedule that maximizes this "holomorphic width." This led to a family of variance-preserving schedules parameterized by Jacobi elliptic functions ($cn, sn, am$).

They then paired this elliptic schedule with a tailored discretization to create Zoloflow. This algorithm is deeply connected to Zolotarev’s optimal rational function approximants—a classical mathematical tool for finding the best possible way to approximate functions like $x^{-1/2}$ (which is essentially what a model does when sampling a Gaussian).

Elliptic schedules for different condition numbers $\kappa$. As $\kappa$ increases, the schedule becomes more "curved" to handle the wider spectrum of the data.

Datasets

The authors evaluated their methods across a range of targets:

  • Synthetic Benchmarks: Gaussian distributions with varying condition numbers, strongly log-concave mixtures, and 2D toy targets (like a "fern frond" and a fractal tree).
  • Image Generation: They tested on pretrained models including SiT-XL/2 (trained on ImageNet) and Stable Diffusion 3.5 Medium (a 2.5B parameter text-to-image model).

Key Insights and Ideas

  • Joint Design is Key: The schedule and the solver are intrinsically coupled. Zoloflow succeeds because the discretization is mathematically "aware" of the elliptic schedule's properties.
  • Conditioning Matters: Most standard schedules assume the data is isotropic (uniform in all directions). Elliptic schedules explicitly incorporate the condition number $\kappa$, allowing the model to navigate "stretched" data distributions much more effectively.
  • Rational vs. Polynomial Convergence: Standard solvers act like polynomial approximants, which converge slowly for wide data spectra. Zoloflow acts like a rational approximant, which converges exponentially faster.
  • Smoothing as Early Stopping: For complex, non-Gaussian data, the authors show that a tiny bit of Gaussian smoothing (effectively stopping the ODE just before it hits the "data" endpoint) allows the elliptic schedule to maintain its efficiency guarantees.

Key Results

The quantitative gains are striking:

  • Gaussian Sampling: On a Gaussian target with $\kappa=100$, Zoloflow reaches machine-level precision in just 12 evaluations. Standard solvers like Euler or DPM-Solver++ require significantly more steps to reach even $10^{-2}$ error.
  • Image Quality at Low Steps: In text-to-image generation with Stable Diffusion 3.5, the elliptic schedule produced noticeably better images than linear or trigonometric schedules when limited to only 4 oracle calls.
Figure 1: Side-by-side comparison of schedules. Images are generated by Stable Diffusion 3.5 Medium (CFG = 4.5) using DPM-Solver++ with only 4 neural network evaluations (4 oracle calls). From left to right, the schedules are the target-rescaled linear, trigonometric, and elliptic (ours). The schedules are rescaled by the hyperparameters shown in Table 2. The text prompt is shown below the images.
  • Efficiency: For log-concave targets, the elliptic schedule achieves a query complexity that is polylogarithmic in both the condition number and the desired precision, a significant theoretical improvement.
Error vs. score evaluations. Zoloflow (black dashed line) drops to near-zero error much faster than any standard solver-schedule combination.

Limitations and Future Work

While the theory is robust, the authors note a few areas for further exploration:

  • Approximate Scores: The current theory assumes the model provides "exact" scores. In reality, neural networks have approximation errors, and extending these guarantees to "noisy" scores is an open challenge.
  • Parameter Selection: For non-Gaussian data (like natural images), the optimal "width" $\kappa$ isn't always obvious. While the authors provide a covariance-based heuristic, more principled ways to select these parameters for complex manifolds could be developed.
  • General Discretization: Zoloflow's specific update rule is tailored for Gaussians; finding a similarly "optimal" rule for general log-concave distributions is still on the horizon.

Summary updated

Abstract

Generative modeling based on stochastic interpolants requires specifying two key choices: an interpolant schedule bridging the initial and target distributions, and an efficient discretization of the resulting probability-flow ODE. In this work, we introduce a new framework for joint schedule/discretization design based on the analyticity properties of the (complexified) ODE trajectories, which in turn ensure good numerical behavior. Our formulation can be solved exactly via conformal mapping, yielding a natural family of variance-preserving elliptic schedules, parametrized by Jacobi elliptic functions. Pairing an elliptic schedule with a tailored discretization, we obtain Zoloflow, a new sampling algorithm closely connected to Zolotarev's optimal rational function approximants. Crucially, both the elliptic schedule and Zoloflow explicitly incorporate the condition number $κ$ of the target distribution and reduce to the classical trig schedule in the limit $κ\rightarrow 1$. For Gaussian targets, Zoloflow achieves minimax optimality among all oblivious score-span algorithms, attaining total-variation error $\varepsilon$ using $Θ(\logκ\,\log(\sqrt d/\varepsilon))$ (exact) score queries. For the log-concave case, the elliptic schedule allows sampling with at most $\varepsilon$ 2-Wasserstein-error by requiring merely $\widetilde{O}(\sqrt{d} \, \mathrm{polylog}(κ, 1/\varepsilon))$ score evaluations. Complementing these theoretical results, we empirically evaluate the elliptic schedule on general (non-log-concave) target distributions, including detailed comparisons with alternative schedule proposals on both synthetic benchmarks and existing image generation datasets.
BibTeX
@misc{becker2026elliptic,
  author = {Cole Becker and Taylan Kargin and Kimia Hassibi and Asuman Ozdaglar and Pablo A. Parrilo},
  title = {{Elliptic Sampling: Optimal Interpolant Schedules for Gaussians and Beyond}},
  year = {2026},
  eprint = {2611.18542},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.18542},
}
Figure