Estimator and numerical conventions
Single-pass and diffusion models
The released vdc-denoiser-m64-v1 is a single-pass
denoiser. It maps an empirical pair histogram, with its configured
conditioning channels, to a density grid. A marginal projection then
adjusts the grid toward uniform marginals.
The package also supports iterative diffusion_unet
checkpoints. Their reverse diffusion settings do not apply to the
released denoiser. Load architectures from the configuration stored in
the checkpoint, using the public loader.
Density grids
A uniform m × m grid uses cell centers
(i + 0.5) / m. The array contains density values, so a
cell's probability mass is its density divided by m**2. The
first array axis corresponds to the first variable.
Iterative proportional fitting alternates row and column normalizations. A finite iteration count leaves a numerical residual; it does not guarantee exactly uniform continuous marginals. The implementation supports both density- and mass-space calculations and log-space projection utilities.
Conditional transforms and vines
Each edge stores a density grid and cached conditional CDF tables. D/C-vine recursion uses those transforms to construct the next tree's inputs. Edges within a tree may be processed in batches, while the trees remain sequential.
HFuncLookup.pdf interpolates density values; conditional
tables use their own normalization and interpolation. These are
approximate counterparts. Therefore projection of every grid does not by
itself imply exact continuous normalization of the assembled vine or
exact agreement between its sampling and likelihood. The sampling
implementation has boundary and inverse-round-trip regression tests.
Information estimates
Bivariate MI is computed by grid quadrature of
c * log(c) in nats. For the fitted vine,
logpdf_decomposed sums log-density contributions over its
edges. The mean on held-out observations estimates total correlation
under the model's assumptions. Fitted edge contributions can be
negative; their interpretation differs from nonnegative population
conditional mutual information.
Limits and reproducibility
- The released model assumes continuous marginals and simplified pair copulas.
- Fixed grid resolution limits near-deterministic dependence and fine tail detail.
- Conditional calibration and deep-tree recursion need task-specific evaluation.
- Changing grids, batch sizes, precision or interpolation is a numerical change.
- Fast fitting and compiled sampling have qualified settings, described in their fitting and sampling guides.
- Paper experiments use explicit model hashes and software pins in the paper reproduction guide, not unspecified current defaults.
Probit-coordinate and transformed-density training options are experimental. They should not be interpreted as proof that a checkpoint supports nonuniform cells throughout vine likelihood and sampling.