# Transport-Signature Density Decoder

- ID: 341
- Canonical URL: https://synthcore.org/idea/341/transport-signature-density-decoder
- API JSON: https://synthcore.org/api/idea/341.json
- API Markdown: https://synthcore.org/api/idea/341.md
- Verification status: mechanism_works
- Source: [arXiv:2607.01696](https://arxiv.org/abs/2607.01696)
- Category: architecture
- Solves: accuracy, stability, generalization
- ML areas: vae, diffusion, world-model, embedding
- Math tags: optimal-transport, geometry, spectral-theory, differential-geometry
- Ratings: usefulness 8/10; difficulty 6/10; novelty 6/10

## Idea description

Replace a pixelwise density decoder with a decoder that predicts coefficients of a transport potential relative to a fixed reference density. The reconstructed density is the pushforward of the reference measure through a differentiable transport map, so positivity and total mass are structural properties rather than learned penalties.

## Mathematical statement

Let \(\rho_0(x)>0\) be a fixed reference probability density on \(\Omega\subset\mathbb R^d\), and let \(\rho_y\) be the target density conditioned on input \(y\). For quadratic transport, the optimal map has the form \(T_y=\nabla\phi_y\), where \((T_y)_\#\rho_0=\rho_y\). Represent the potential as \(\hat\phi_y(x)=\bar\phi(x)+\sum_{j=1}^r a_j(y)\psi_j(x)\), with neural coefficients \(a_j(y)\). The spatial basis can be obtained from the weighted Laplacian \(\mathcal L_{\rho_0}f=-\rho_0^{-1}\nabla\cdot(\rho_0\nabla f)\), using eigenfunctions satisfying \(\mathcal L_{\rho_0}\psi_j=\lambda_j\psi_j\). The decoded density is \(\hat\rho_y=(\nabla\hat\phi_y)_\#\rho_0\), equivalently \(\int g(z)\hat\rho_y(z)\,dz=\int g(\nabla\hat\phi_y(x))\rho_0(x)\,dx\) for every test function \(g\). If the map is invertible, its Jacobian satisfies \(\hat\rho_y(\nabla\hat\phi_y(x))\det(\nabla^2\hat\phi_y(x))=\rho_0(x)\). For the exact map \(T_y\), any approximate map \(\hat T_y\) obeys \(W_2^2(\rho_y,\hat\rho_y)\leq\int_\Omega\|T_y(x)-\hat T_y(x)\|^2\rho_0(x)\,dx\).

## Key formulas

- $$\mathcal L_{\rho_0}f=-\rho_0^{-1}\nabla\cdot(\rho_0\nabla f),\qquad \mathcal L_{\rho_0}\psi_j=\lambda_j\psi_j.$$
- $$\hat\phi_y(x)=\bar\phi(x)+\sum_{j=1}^{r}a_j(y)\psi_j(x),\qquad \hat T_y(x)=\nabla\hat\phi_y(x).$$
- $$\hat\rho_y=(\hat T_y)_\#\rho_0,\qquad \int g(z)\hat\rho_y(z)\,dz=\int g(\hat T_y(x))\rho_0(x)\,dx.$$
- $$W_2^2(\rho_y,\hat\rho_y)\leq\int_\Omega\|T_y(x)-\hat T_y(x)\|^2\rho_0(x)\,dx.$$

## Implementation notes

Integrate this as the output head of a conditional density model, VAE decoder, diffusion auxiliary model, or neural operator. Choose a fixed reference point cloud \(x_i\) with normalized weights \(w_i\) approximating \(\rho_0\), and precompute \(r\) basis functions \(\psi_j\) using the lowest eigenvectors of a finite-element or graph discretization of \(\mathcal L_{\rho_0}\). The network receives condition \(y\) and outputs \(a\in\mathbb R^r\). Compute \(\phi=\bar\phi+B a\) on the reference points, differentiate it spatially to obtain \(z_i=\nabla\hat\phi(x_i)\), and represent the output as weighted particles \((z_i,w_i)\). For a grid output, use a differentiable Gaussian splat or kernel histogram. Pseudocode: `a = net(y); phi = phi_bar + B @ a; T = gradient(phi, x_ref); loss = sinkhorn(T, w, target_samples) + alpha * convexity_penalty; backpropagate(loss)`. The paper's mathematics supplies the pushforward decoder and Wasserstein surrogate; automatic differentiation, Sinkhorn approximation, and convexity penalties are engineering choices. Ensure approximate convexity by using a strongly convex \(\bar\phi\), limiting coefficient magnitudes, or penalizing negative eigenvalues of the Hessian. First test on a two-dimensional translated and deformed Gaussian or continuity-equation dataset. Compare against a pixelwise MLP decoder with equal parameter count and evaluate held-out Wasserstein error, negative predicted mass, total-mass error, and wall-clock training. Success means lower Wasserstein error with fewer latent coefficients and near-zero mass violation.

## Verification

- Status: mechanism_works
- Mechanism evidence: yes
- Mechanism confirmed: no
- Verdict: Built a conditional transport-signature decoder and a pixelwise histogram baseline on a synthetic 2D pushforward family. The transport decoder achieved lower held-out sliced Wasserstein error (0.0224 vs 0.0291) with only 117 parameters versus 8,512, while preserving mass structurally and maintaining positive Jacobian curvature (minimum eigenvalue 0.864). The accuracy and compactness effect is real in this small test, although transport training was slower (16.16 s vs 2.96 s).

### Mechanism check

- Verdict: Built a conditional transport-signature decoder and a pixelwise histogram baseline on a synthetic 2D pushforward family. The transport decoder achieved lower held-out sliced Wasserstein error (0.0224 vs 0.0291) with only 117 parameters versus 8,512, while preserving mass structurally and maintaining positive Jacobian curvature (minimum eigenvalue 0.864). The accuracy and compactness effect is real in this small test, although transport training was slower (16.16 s vs 2.96 s).
- Confidence: 8/10
- Limitations: The experiment uses a synthetic known map, a low-dimensional affine/quadratic potential rather than computed weighted-Laplacian eigenfunctions, sliced Wasserstein rather than full Wasserstein or Sinkhorn, and a softmax histogram baseline that already enforces positivity and unit mass. No image-scale, VAE, diffusion, or real-data evaluation was performed.

## Artifacts

- [report.md](https://synthcore.org/code/53/report.md)
- [results.json](https://synthcore.org/code/53/results.json)
- [transport_decoder_experiment.py](https://synthcore.org/code/53/transport_decoder_experiment.py)
- [Download all files as ZIP](https://synthcore.org/download/53)

## Disclaimer

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