# Conditional-Transport Discrete Reverse Diffusion

- ID: 1342
- Canonical URL: https://synthcore.org/idea/1342/conditional-transport-discrete-reverse-diffusion
- API JSON: https://synthcore.org/api/idea/1342.json
- API Markdown: https://synthcore.org/api/idea/1342.md
- Verification status: beats_baseline
- Source: [arXiv:2607.23947](https://arxiv.org/abs/2607.23947)
- Category: sampling
- Solves: accuracy, stability, sample-efficiency
- ML areas: diffusion, diffusion-sampling, training
- Math tags: probability, stochastic-processes, statistics, information-theory
- Ratings: usefulness 8/10; difficulty 6/10; novelty 5/10

## Idea description

Replace the standard Gaussian affine reverse step with a conditional transport kernel learned from the forward transition. Given a noisy state x_{k+1}, the model predicts a full conditional distribution for x_k using a monotone conditional CDF or an autoregressive normalizing flow. This represents multimodal and state-dependent reverse transitions that cannot be captured by a single Gaussian mean and variance.

## Mathematical statement

Let the forward process be x_{k+1}=h_k(x_k,u_k), where u_k is independent process noise and p_k is the marginal density of x_k. The exact reverse kernel follows from Bayes' rule: p(x_k given x_{k+1}) equals p(x_{k+1} given x_k) times p_k(x_k), divided by p_{k+1}(x_{k+1}). A reverse sampler should implement x_k=T_k(x_{k+1},z_k), where z_k is standard normal and T_k transports the base noise into the conditional reverse distribution. In one dimension, T_k is the inverse conditional CDF evaluated at the Gaussian CDF of z_k. For vector states, use a Rosenblatt transform based on a sequence of conditional CDFs. The paper's independence mechanism can be monitored through the conditional characteristic function Psi(omega,x)=E[exp(j omega transpose h(x,u)) given x]. If reverse residual noise is independent of the conditioning state, its characteristic function should not vary with that state.

## Key formulas

- $$x_{k+1}=h_k(x_k,u_k),\qquad u_k\perp\!\!\!\perp(x_0,\ldots,x_k).$$
- $$p(x_k\mid x_{k+1})=\frac{p(x_{k+1}\mid x_k)p_k(x_k)}{p_{k+1}(x_{k+1})}.$$
- $$x_k=T_k(x_{k+1},z_k),\qquad z_k\sim\mathcal N(0,I),\qquad T_k=F_{k,x_{k+1}}^{-1}\circ\Phi\quad\text{in one dimension}.$$
- $$\Psi(\omega,x)=E\left[\exp\left(j\omega^{\top}h(x,u)\right)\mid x\right].$$

## Implementation notes

1. Integration point: use this reverse kernel inside a discrete DDPM or latent diffusion sampler. Retain the ordinary forward schedule x_{k+1}=sqrt(1-epsilon_k)x_k+sqrt(epsilon_k)u_k, but replace the Gaussian reverse mean-and-variance head with a conditional autoregressive flow T_theta(z; x_{k+1}, k). Use rational-quadratic splines or masked affine coupling layers, conditioned on x_{k+1} and a timestep embedding. 2. Pseudocode: sample a clean training example x_k; generate x_{k+1} with the known forward kernel; compute flow parameters theta=g_theta(x_{k+1},k); draw z from a standard normal; evaluate x_hat=T_theta(z; x_{k+1}); optimize the exact flow negative log likelihood, equal to one half the squared norm of z minus the log absolute Jacobian determinant plus a constant. For generation, start from the terminal noise state and repeatedly sample z and apply the learned inverse-time flow. 3. Taken from the paper: Bayes' reverse-kernel identity, conditional-CDF transport, and the requirement that reverse noise be independent of the reversed state. Estimated empirically: flow parameters, conditional likelihoods, and density ratios. Measure independence by training a classifier to predict x_{k+1} from standardized reverse residuals. 4. First cheap experiment: use a two-dimensional mixture of Gaussians with a non-Gaussian initial distribution and compare a standard affine Gaussian reverse model against a conditional spline flow. Evaluate held-out reverse conditional NLL, estimated conditional KL divergence, final Wasserstein distance, and residual-prediction AUC. The prediction is that the affine model's conditional KL will become large in regions where the exact reverse conditional is skewed or multimodal, while the flow will have lower KL and residual AUC near 0.5. As a control, repeat with a Gaussian initial state; the performance gap should substantially shrink because the reverse conditional is then closer to the affine Gaussian assumption.

## Verification

- Status: beats_baseline
- Mechanism evidence: yes
- Mechanism confirmed: yes
- Practical verdict: helps
- Verdict: Built an exact 1D conditional-CDF transport MVP for a Gaussian-mixture reverse kernel, alongside a moment-matched affine Gaussian baseline. Bayes identity was verified to median relative error 2.0e-16, PIT sup deviation was 0.030, and the predicted bimodality threshold noise variance 0.00387 matched the sweep exactly. The affine conditional KL increased with mixture separation (0 at d=0 to 0.063 at d=5) and noise variance (0.018 at 0.03 to 0.871 at 4), while exact transport has KL 0 by construction. This demonstrates the mechanism, but not a practical trained-flow improvement.

### Mechanism check

- Verdict: Built an exact 1D conditional-CDF transport MVP for a Gaussian-mixture reverse kernel, alongside a moment-matched affine Gaussian baseline. Bayes identity was verified to median relative error 2.0e-16, PIT sup deviation was 0.030, and the predicted bimodality threshold noise variance 0.00387 matched the sweep exactly. The affine conditional KL increased with mixture separation (0 at d=0 to 0.063 at d=5) and noise variance (0.018 at 0.03 to 0.871 at 4), while exact transport has KL 0 by construction. This demonstrates the mechanism, but not a practical trained-flow improvement.
- Confidence: 8/10
- Limitations: The MVP uses an analytically known 1D posterior and numerical inverse CDF, not a learned autoregressive flow, rational-quadratic spline, multidimensional Rosenblatt transform, or DDPM sampler. The residual diagnostic uses residual magnitude with a median conditioning-state split rather than a trained classifier. No neural training, wall-clock, FLOP, image-quality, or high-dimensional comparison was tested.

### Practical benchmark

- Paired seeds: 8
- Baseline mean: 0
- Idea mean: 0
- p-value: 0.0468
- Paired wins: 6/8
- Benchmark verdict: idea better (significant)

## Artifacts

- [bench_report.json](https://synthcore.org/code/358/bench_report.json)
- [conditional_transport_bench.py](https://synthcore.org/code/358/conditional_transport_bench.py)
- [conditional_transport_experiment.py](https://synthcore.org/code/358/conditional_transport_experiment.py)
- [report.md](https://synthcore.org/code/358/report.md)
- [report_bench_2026-08-31T092046.md](https://synthcore.org/code/358/report_bench_2026-08-31T092046.md)
- [results.json](https://synthcore.org/code/358/results.json)
- [Download all files as ZIP](https://synthcore.org/download/358)

## Disclaimer

AI-generated research hypothesis, automatically tested. Not peer-reviewed.