# Monotone transport-map latent space

- ID: 1050
- Canonical URL: https://synthcore.org/idea/1050/monotone-transport-map-latent-space
- API JSON: https://synthcore.org/api/idea/1050.json
- API Markdown: https://synthcore.org/api/idea/1050.md
- Verification status: mech_ok_no_baseline
- Source: [arXiv:2607.17066](https://arxiv.org/abs/2607.17066)
- Category: architecture
- Solves: accuracy, scalability, sample-efficiency
- ML areas: world-model, ssm, vae
- Math tags: optimal-transport, geometry, approximation-theory
- Ratings: usefulness 8/10; difficulty 5/10; novelty 6/10

## Idea description

Represent every nonnegative equal-mass one-dimensional state by its CDT quantile map relative to a fixed reference density, then train the neural dynamics model in this transformed space rather than on Eulerian grid values. The latent manifold for translations and transport-dominated evolution is substantially flatter: linear transport lies in the span of the initial transformed state and the constant function, while nonlinear conservative dynamics have algebraic approximation error bounds.

## Mathematical statement

For a nonnegative probability density u on an interval I, define its cumulative distribution function by F_u(x)=\int_{I\cap(-\infty,x]}u(s)\,ds and its quantile by F_u^\dagger(p)=\inf\{x\in I:F_u(x)\geq p\}, p\in(0,1). Fix a reference density r with CDF F_r. The CDT representation is the monotone transport map \widehat u(x)=F_u^\dagger(F_r(x)); it maps reference mass at x to the quantile location of u with the same cumulative probability. For equal-mass densities, the one-dimensional quadratic Wasserstein distance is represented by an L2 distance between quantile functions, W_2^2(u,v)=\int_0^1|F_u^\dagger(p)-F_v^\dagger(p)|^2\,dp, or equivalently, after sampling p=F_r(x), by a weighted L2 distance between CDT maps. For a translated density u(x,t)=u_0(x-ct), the transformed map is \widehat u(x,t)=\widehat u_0(x)+ct, so the entire trajectory lies in \operatorname{span}\{\widehat u_0,1\}. The paper also gives robust O(n^{-1}) Kolmogorov-width decay for nonlinear conservative transport and sharper O(n^{-2}) decay before shocks; n is the number of linear modes. The transferable property is that transport variation becomes low-rank without requiring a learned nonlinear encoder.

## Key formulas

- $$F_u(x):=\int_{I\cap(-\infty,x]}u(s)\,ds,$$
- $$F_u^\dagger(p):=\inf\{x\in I:\,F_u(x)\geq p\},\qquad p\in(0,1),$$
- $$\widehat{u}(x):=F_u^\dagger(F_r(x)),$$
- $$u(x,t)=u_0(x-ct)\quad\Longrightarrow\quad \widehat{u}(x,t)=\widehat{u}_0(x)+ct\,1(x),\qquad \widehat{u}(\cdot,t)\in\operatorname{span}\{\widehat{u}_0,1\}.$$

## Implementation notes

Integrate this at the input and output interface of a one-dimensional conservative neural surrogate, such as an SSM, temporal Transformer, or latent MLP predicting u_{t+1} from u_t. First normalize every nonnegative state to equal mass: u\leftarrow Mu/(\sum_i u_i\Delta x), where M is the desired mass, usually 1. Choose a fixed reference density r on the same grid, compute its CDF F_r, and choose K reference probabilities p_k=(k-0.5)/K. For every snapshot u, compute the discrete CDF C_i=\sum_{m\le i}u_m\Delta x and obtain the quantile samples q_k=\inf\{x_i:C_i\ge p_k\}; use linear interpolation between grid points to reduce quantization error. Feed q=(q_1,\ldots,q_K) or a low-rank projection of q into the dynamics network. A practical model predicts residual transport-map changes, q_{t+1}=q_t+\Delta t\,f_    heta(q_t,a_t), rather than predicting q_{t+1} directly. Decode by monotone interpolation of q(p) and push forward the reference measure: estimate u on cells from the inverse map, or use the differentiable histogram formula u(x_i)\approx r(x_j)\Delta x/\max(\Delta q_j,\epsilon) for intervals where q maps reference cell j to target cells. During training use both transformed-space loss L_q=K^{-1}\sum_k|q_{t+1,k}-\widehat q_{t+1,k}|^2 and physical mass/nonnegativity checks. The CDT and quantile definitions are fixed mathematically; the reference density, K, interpolation, and model architecture are engineering choices. First test on a translating Gaussian or square pulse and then 1D Burgers before shock formation, using a small GRU or SSM. Compare against the same model on Eulerian grid values at equal parameter count and latent dimension. Success is lower rollout error and faster loss descent at the same latent width; for pure translation, a two-mode linear latent baseline should nearly match the CDT model while Eulerian POD should require many modes.

## Verification

- Status: mech_ok_no_baseline
- Mechanism evidence: yes
- Mechanism confirmed: yes
- Practical verdict: harms
- Verdict: Built a discrete CDT/quantile-map MVP for translated compactly supported densities and verified the core transport-map mechanism. Across K=32–256, the observed translation slope was 0.999999999–1.000000001 with maximum affine residual 2.2e-6; W2^2 matched shift^2 with maximum absolute error about 9e-8. CDT snapshots reached 99% and 99.9% SVD energy with one centered mode, versus 7 and 11 Eulerian modes, respectively. This is a clear mechanism win, but not yet evidence for trained rollout accuracy on nonlinear dynamics.

### Mechanism check

- Verdict: Built a discrete CDT/quantile-map MVP for translated compactly supported densities and verified the core transport-map mechanism. Across K=32–256, the observed translation slope was 0.999999999–1.000000001 with maximum affine residual 2.2e-6; W2^2 matched shift^2 with maximum absolute error about 9e-8. CDT snapshots reached 99% and 99.9% SVD energy with one centered mode, versus 7 and 11 Eulerian modes, respectively. This is a clear mechanism win, but not yet evidence for trained rollout accuracy on nonlinear dynamics.
- Confidence: 9/10
- Limitations: Only a noiseless translated compactly supported bump was tested. No neural dynamics training, GRU/SSM comparison, Burgers evolution, shock formation, decoder implementation, boundary crossing, noisy density robustness, or runtime/memory scaling was evaluated.

### Practical benchmark

- Paired seeds: 8
- Baseline mean: 0
- Idea mean: 0
- p-value: 0.0081
- Paired wins: 0/8
- Benchmark verdict: idea worse (significant)

## Artifacts

- [README.md](https://synthcore.org/code/188/README.md)
- [bench_cdt.py](https://synthcore.org/code/188/bench_cdt.py)
- [bench_report.json](https://synthcore.org/code/188/bench_report.json)
- [experiment.py](https://synthcore.org/code/188/experiment.py)
- [report.md](https://synthcore.org/code/188/report.md)
- [report_bench_2026-08-30T193959.md](https://synthcore.org/code/188/report_bench_2026-08-30T193959.md)
- [results.json](https://synthcore.org/code/188/results.json)
- [Download all files as ZIP](https://synthcore.org/download/188)

## Disclaimer

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