Classification-Diffusion Copula

• Classification-Diffusion Copula is a generative copula model that uses diffusion-time classification to learn inter-variable dependencies while preserving one-dimensional marginals.
• It employs a forward diffusion process using an Ornstein–Uhlenbeck mechanism to gradually forget dependence, transitioning from full dependence to near-independence.
• A neural classifier trained on discretized diffusion times recovers the copula density and score, enabling both likelihood estimation and score-based sampling.

Classification-Diffusion Copula (CDC) is a copula model in which dependence is learned by classifying diffusion-time indices along a forward process that progressively destroys inter-variable dependence while preserving one-dimensional marginals. In its original formulation, CDC operates on copula-scale data, maps it to Gaussian scale, applies an Ornstein–Uhlenbeck noising process, and recovers the copula density and score from the posterior probabilities of diffusion times under a neural classifier (Huk et al., 24 Sep 2025). In later applied work on multivariate financial forecasting, CDC functions as the dependence module in a two-stage Diffusion-Copula framework, paired with separately estimated heavy-tailed marginals so that marginal calibration and joint tail dependence are modeled explicitly rather than through a single end-to-end diffusion objective (Huk et al., 19 May 2026).

1. Definition and copula-theoretic setting

CDC is defined within the standard copula decomposition furnished by Sklar’s theorem. For a continuous random vector $X=(X^1,\dots,X^d)$ with joint density $p$ and marginal densities $p^i$, the density factorization is

$$p(x^1,\dots,x^d) = p^1(x^1)\cdots p^d(x^d)\, c\!\big(P^1(x^1),\dots,P^d(x^d)\big),$$

where $c:[0,1]^d\to(0,\infty)$ is the copula density. The copula captures all inter-variable dependence, while the marginals account for the one-dimensional behavior (Huk et al., 24 Sep 2025).

CDC works only on the copula component. This is done by moving from data scale $x\in\mathbb{R}^d$ to copula scale

$u^i=P^i(x^i),\qquad u\in[0,1]^d,$

and then to Gaussian scale

$z^i=\Phi^{-1}(u^i),\qquad z\in\mathbb{R}^d.$

Because copula densities are density ratios, they are invariant under diffeomorphisms, so modeling on the Gaussian scale is equivalent for dependence (Huk et al., 24 Sep 2025).

The term “classification” in CDC refers to classifying diffusion time, not to supervised label prediction. In the financial forecasting application, this is stated explicitly: the classification labels are the diffusion time steps $T_s$, and the classifier outputs are used as density ratios to form $c(u)$ and, via gradients, to construct a score for sampling (Huk et al., 19 May 2026). This naming convention is a recurring source of confusion because “classification” in the CDC literature denotes a density-ratio estimation mechanism rather than a conventional task over semantic classes.

2. Dependence-forgetting diffusion

The core construction is a forward diffusion that forgets dependence while keeping marginals unchanged. On Gaussian scale, CDC uses the component-wise Ornstein–Uhlenbeck process

$p$0

with explicit solution

$p$1

If $p$2 has standard normal marginals, then each $p$3 remains $p$4 for all $p$5, so after mapping back via $p$6, every marginal remains $p$7 (Huk et al., 24 Sep 2025).

This yields a continuous path of copulas $p$8. For all $p$9, $p^i$0 is a valid copula sample, and $p^i$1 converges to the independence copula

$p^i$2

in Kullback–Leibler divergence at rate $p^i$3 (Huk et al., 24 Sep 2025). The forward time variable therefore indexes a continuum from full dependence at $p^i$4 to practical independence at large $p^i$5.

The same paper also introduces a correlated OU variant,

$p^i$6

with unit diagonals in $p^i$7. In that case, marginals remain standard normal, but the limiting distribution becomes $p^i$8, so the path terminates at a Gaussian copula rather than the independence copula (Huk et al., 24 Sep 2025). This provides a nontrivial base copula when complete decorrelation is undesirable.

The significance of this design is structural. The forward process is constrained to alter dependence only. Marginal preservation is not learned statistically; it is guaranteed by construction. This separation is central to CDC and distinguishes it from end-to-end multivariate diffusion models in which marginal calibration and dependence learning are entangled.

3. Density-ratio recovery by time classification

CDC discretizes diffusion time as

$p^i$9

with $$p(x^1,\dots,x^d) = p^1(x^1)\cdots p^d(x^d)\, c\!\big(P^1(x^1),\dots,P^d(x^d)\big),$$0 chosen large enough that the forward process is effectively at stationarity. A neural classifier is trained to output

$$p(x^1,\dots,x^d) = p^1(x^1)\cdots p^d(x^d)\, c\!\big(P^1(x^1),\dots,P^d(x^d)\big),$$1

with uniform class priors over the time labels (Huk et al., 24 Sep 2025).

The key identity is that the copula density can be recovered as a probability ratio. For $$p(x^1,\dots,x^d) = p^1(x^1)\cdots p^d(x^d)\, c\!\big(P^1(x^1),\dots,P^d(x^d)\big),$$2 and $$p(x^1,\dots,x^d) = p^1(x^1)\cdots p^d(x^d)\, c\!\big(P^1(x^1),\dots,P^d(x^d)\big),$$3,

$$p(x^1,\dots,x^d) = p^1(x^1)\cdots p^d(x^d)\, c\!\big(P^1(x^1),\dots,P^d(x^d)\big),$$4

Under the correlated OU variant, the formula acquires an explicit Gaussian correction: $$p(x^1,\dots,x^d) = p^1(x^1)\cdots p^d(x^d)\, c\!\big(P^1(x^1),\dots,P^d(x^d)\big),$$5 Thus the learned classifier is simultaneously a time-posterior estimator and a direct copula density estimator (Huk et al., 24 Sep 2025).

CDC also extracts the copula score from classifier gradients. If $$p(x^1,\dots,x^d) = p^1(x^1)\cdots p^d(x^d)\, c\!\big(P^1(x^1),\dots,P^d(x^d)\big),$$6 denotes the copula at time $$p(x^1,\dots,x^d) = p^1(x^1)\cdots p^d(x^d)\, c\!\big(P^1(x^1),\dots,P^d(x^d)\big),$$7, then

$$p(x^1,\dots,x^d) = p^1(x^1)\cdots p^d(x^d)\, c\!\big(P^1(x^1),\dots,P^d(x^d)\big),$$8

where

$$p(x^1,\dots,x^d) = p^1(x^1)\cdots p^d(x^d)\, c\!\big(P^1(x^1),\dots,P^d(x^d)\big),$$9

This identity is what makes CDC a score-based generative model rather than only a likelihood model (Huk et al., 24 Sep 2025).

Training combines a cross-entropy term for time-label prediction with a score-matching term implemented as a noise-prediction mean-squared error. For sampled time $c:[0,1]^d\to(0,\infty)$0, dependent sample $c:[0,1]^d\to(0,\infty)$1, and $c:[0,1]^d\to(0,\infty)$2, the synthetic corruption is

$c:[0,1]^d\to(0,\infty)$3

The loss is

$c:[0,1]^d\to(0,\infty)$4

Under the assumption that the model class can represent the true conditional probabilities $c:[0,1]^d\to(0,\infty)$5, any minimizer recovers the true classifier; at optimality, CDC recovers both the exact copula density $c:[0,1]^d\to(0,\infty)$6 and the exact copula score $c:[0,1]^d\to(0,\infty)$7 (Huk et al., 24 Sep 2025).

4. Sampling, architectures, and relation to the Reflection Copula

Sampling in CDC is performed on Gaussian scale through a reverse diffusion. The initialization is

$c:[0,1]^d\to(0,\infty)$8

followed by backward iteration over $c:[0,1]^d\to(0,\infty)$9. At each step, classifier outputs are converted into a copula score, and a DDPM-style reverse update is applied. The final sample is mapped back to copula scale as

$x\in\mathbb{R}^d$0

CDC is therefore both a direct copula density estimator and a score-based sampler (Huk et al., 24 Sep 2025).

For scientific and tabular data, the classifier is implemented as a ResNet-style MLP with 6 hidden layers of width 512, Swish activations, and a classifier head producing $x\in\mathbb{R}^d$1 logits over times. For images such as digits, MNIST, and CIFAR, the model uses a DDPM-style U-Net with timestep embeddings and an extra classification head over diffusion times. Training uses Adam with learning rates around $x\in\mathbb{R}^d$2–$x\in\mathbb{R}^d$3; the number of time classes $x\in\mathbb{R}^d$4 is typically 8–256 depending on data. CDC uses either KL-space or linear discretization in $x\in\mathbb{R}^d$5, with $x\in\mathbb{R}^d$6 chosen so the forward process is effectively at stationarity (Huk et al., 24 Sep 2025).

The same paper introduces a complementary flow-based construction, the Reflection Copula. It lives directly on $x\in\mathbb{R}^d$7, preserves uniform marginals through coordinate-wise reflections, and learns an expected velocity field $x\in\mathbb{R}^d$8 for deterministic backward ODE sampling. Reflection Copula is designed for fast sampling and does not provide density evaluation, whereas CDC provides both density and score (Huk et al., 24 Sep 2025).

Aspect	CDC	Reflection Copula
State space	Gaussian scale	Copula scale $x\in\mathbb{R}^d$9
Learned object	Time classifier	Velocity field
Output capability	Density and score	Fast sampling
Reverse dynamics	Stochastic reverse diffusion	Deterministic ODE

This division of labor is methodologically important. CDC is the likelihood-oriented member of the framework, while Reflection Copula is the expedient sampler. A plausible implication is that the two models target different operational regimes rather than competing notions of copula learning.

5. Empirical behavior in scientific data, images, and financial forecasting

On scientific and structured tabular datasets, CDC is reported to achieve state-of-the-art likelihoods and Wasserstein-$u^i=P^i(x^i),\qquad u\in[0,1]^d,$0 distances, while Reflection Copula provides competitive or best Wasserstein-$u^i=P^i(x^i),\qquad u\in[0,1]^d,$1 performance (Huk et al., 24 Sep 2025). On Magic ($u^i=P^i(x^i),\qquad u\in[0,1]^d,$2), CDC attains LL $u^i=P^i(x^i),\qquad u\in[0,1]^d,$3 versus the best baseline $u^i=P^i(x^i),\qquad u\in[0,1]^d,$4 (Ratio copula), and W2 $u^i=P^i(x^i),\qquad u\in[0,1]^d,$5, with Reflection at $u^i=P^i(x^i),\qquad u\in[0,1]^d,$6 versus Vine $u^i=P^i(x^i),\qquad u\in[0,1]^d,$7 and Gaussian $u^i=P^i(x^i),\qquad u\in[0,1]^d,$8. On Dry_Bean ($u^i=P^i(x^i),\qquad u\in[0,1]^d,$9), CDC attains LL $z^i=\Phi^{-1}(u^i),\qquad z\in\mathbb{R}^d.$0 versus Ratio $z^i=\Phi^{-1}(u^i),\qquad z\in\mathbb{R}^d.$1 and Vine $z^i=\Phi^{-1}(u^i),\qquad z\in\mathbb{R}^d.$2, with W2 $z^i=\Phi^{-1}(u^i),\qquad z\in\mathbb{R}^d.$3; Reflection yields W2 $z^i=\Phi^{-1}(u^i),\qquad z\in\mathbb{R}^d.$4, equal to Vine and better than Gaussian $z^i=\Phi^{-1}(u^i),\qquad z\in\mathbb{R}^d.$5. On Robocup ($z^i=\Phi^{-1}(u^i),\qquad z\in\mathbb{R}^d.$6), CDC attains LL $z^i=\Phi^{-1}(u^i),\qquad z\in\mathbb{R}^d.$7 versus Ratio $z^i=\Phi^{-1}(u^i),\qquad z\in\mathbb{R}^d.$8 and Gaussian $z^i=\Phi^{-1}(u^i),\qquad z\in\mathbb{R}^d.$9; Reflection gives W2 $T_s$0, CDC $T_s$1, and Vine and Gaussian both $T_s$2 (Huk et al., 24 Sep 2025).

In high-dimensional image experiments, CDC remains competitive in dimensions where classical copulas degrade sharply. On digits, CDC achieves LL $T_s$3 versus Ratio $T_s$4, Vine $T_s$5, and Gaussian $T_s$6, with best W2 $T_s$7, although its FID $T_s$8 is worse than Reflection and Gaussian. On MNIST, CDC attains LL $T_s$9 versus Ratio $c(u)$0, Vine $c(u)$1, and Gaussian $c(u)$2, with best W2 $c(u)$3 and best FID $c(u)$4. On CIFAR, CDC attains LL $c(u)$5 versus Ratio $c(u)$6 and Gaussian $c(u)$7, with best W2 $c(u)$8, while Reflection achieves the best FID $c(u)$9 and CDC the second best $p$00. Qualitatively, only CDC and Reflection produce recognizable digits and cars on the copula scale (Huk et al., 24 Sep 2025).

In multivariate financial forecasting, CDC is embedded in a two-stage Diffusion-Copula architecture. Each asset’s one-step-ahead marginal predictive distribution $p$01 is modeled independently by a deep Mixture Density Network with Normal, Laplace, and Student-$p$02 components, and the resulting PIT values

$p$03

are passed to CDC to model dependence (Huk et al., 19 May 2026). The implementation assumes that dependence on predictors $p$04 is absorbed entirely into the marginals, so the copula density is learned on i.i.d. copula-scale data.

Applied to 9 cryptocurrencies at 10-minute frequency over Q1–Q3 2022, this framework is used to address the “normality bias” of end-to-end multivariate diffusion models. On central predictive metrics, CDC is competitive: RMSE $p$05, MAE $p$06, and CRPS $p$07. Its strongest gains appear in extremes: tail accuracy $p$08 versus $p$09 for CSDI and $p$10 for TMDM, and CRPS on joint extreme days $p$11 versus $p$12 and $p$13 (Huk et al., 19 May 2026). The same study reports that baseline models classify simultaneous market crashes as statistically impossible “Black Swans” with high surprise, whereas the Diffusion-Copula framework identifies them as “Expected Crashes” with low surprise by preserving the correlation structure necessary for contagion modeling (Huk et al., 19 May 2026).

6. Interpretation, misconceptions, and related research

The principal misconception surrounding CDC is terminological. In the CDC literature, “classification” means classification over diffusion time, not over discretized spatial bins and not, in the implemented model, over semantic class labels. The financial application states this directly: they use classification over diffusion time, not over discretized spatial bins in $p$14 (Huk et al., 19 May 2026). The original copula paper similarly explains that the model is learned by classifying diffusion times on Gaussian-scale copula data (Huk et al., 24 Sep 2025).

A second distinction concerns the boundary between implemented method and extension. The CDC paper sketches a class-conditional copula approach in which one would fit a separate CDC per class $p$15, use

$p$16

and then perform Bayes classification. The paper also notes that the current work does not implement such supervised uses (Huk et al., 24 Sep 2025). This suggests that “classification-diffusion copula” has two meanings in the emerging literature: the established one, where time is the class label, and a plausible but currently secondary one, where CDC becomes a building block for generative classification.

CDC also belongs to a broader research line that treats copula estimation as a discriminative or generative-bridging problem. “Your copula is a classifier in disguise” reinterprets copula density estimation as training a classifier to distinguish samples from the joint density from those of the product of independent marginals, recovering the copula density as a density ratio (Huk et al., 2024). A plausible reading is that CDC turns this binary ratio idea into a multi-time, diffusion-indexed density-ratio estimator. In discrete state spaces, “Discrete Copula Diffusion” supplements missing dependency information in discrete diffusion by incorporating another deep generative model termed the copula model, and achieves better (un)conditional text generation using 8 to 32 times fewer denoising steps than the diffusion model alone (Liu et al., 2024).

The diffusion-copula lineage is older than CDC. Earlier work on semiparametric Markov models defined diffusion copulas through an underlying parametric diffusion with nonparametric marginal transformation, yielding parametric dynamic copulas and nonparametric marginal distributions (Bu et al., 2020). Related work on one-dimensional diffusions showed that two diffusion processes are related by a monotone space-time transformation if and only if they share the same serial dependence, expressed through the same copula density up to time transformation (Bibbona et al., 2015). CDC departs from these constructions by replacing closed-form parametric copula dynamics with a learned diffusion-time classifier, but it preserves the same foundational separation between marginals and dependence.

In that sense, CDC is best understood as a modern copula architecture for high-dimensional and multimodal dependence: the forward mechanism forgets dependence without disturbing marginals, and the learned reverse mechanism remembers it through classification. Its technical identity lies in this density-ratio interpretation, its practical value lies in providing both likelihoods and score-based sampling, and its main conceptual contribution is to treat dependence strength itself as a diffusion coordinate.