△ Mechanism confirmed, baseline not beaten 2026

Dimension-Free Brenier Transport Layer

Implementation & benchmark of arXiv:2608.15906 — Dimension-Free Lipschitz Bounds for Brenier Maps to Compactly Supported Log-Concave Targets

Usefulness7/10

Difficulty6/10

Novelty7/10

Architecture Stability Accuracy Sample efficiency Vae Normalization Diffusion Optimal transport Convex analysis Geometry

Source paper: Dimension-Free Lipschitz Bounds for Brenier Maps to Compactly Supported Log-Concave Targets arXiv:2608.15906 ⓘ · analyzed Sep 1, 2026

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 Not peer-reviewed.

Idea description

Build a neural transport layer by parameterizing a convex potential whose gradient maps a semi-log-concave latent distribution into a compact convex data domain. Use the paper's dimension-free Lipschitz certificate to set the layer's Jacobian scale, initialize the potential, and reject or regularize parameter updates that create excessive curvature. The goal is a bounded-output transport module that is less sensitive to latent dimension than diameter-based spectral heuristics.

Formulas

$$x\longmapsto\frac{1}{2}\langle Qx,x\rangle-V(x)\ \text{is convex}$$

$$w_K(v):=\sup_{y\in K}\langle y,v\rangle-\inf_{y\in K}\langle y,v\rangle$$

$$\partial_{vv}\Phi\leq 0.587\sqrt{\langle Qv,v\rangle}\,w_K(v)$$

$$\operatorname{Lip}(\nabla\Phi)\leq 0.587\sqrt{\|Q\|_{\mathrm{op}}}\,\operatorname{diam}(K)$$

Mathematical statement

Let the source probability measure be d\mu(x)=Z_\mu^{-1}e^{-V(x)}dx on \mathbb{R}^d, where V is finite and Z_\mu=\int e^{-V(x)}dx. Let Q\succ0 be symmetric positive definite and assume x\mapsto \frac12\langle Qx,x\rangle-V(x) is convex; in the smooth case this means
abla^2V(x)\preceq Q. Let
u be a compactly supported log-concave probability measure with support K, and let \Phi be the convex Brenier potential transporting \mu to
u via T(x)=
abla\Phi(x). For a direction v, define the support width w_K(v)=\sup_{y\in K}\langle y,v\rangle-\inf_{y\in K}\langle y,v\rangle. The theorem gives the distributional directional Hessian bound \partial_{vv}\Phi\le 0.587\sqrt{\langle Qv,v\rangle}\,w_K(v). Consequently, T has a globally Lipschitz representative with \operatorname{Lip}(T)\le0.587\sqrt{\|Q\|_{op}}\operatorname{diam}(K), where \|Q\|_{op} is the largest eigenvalue of Q and \operatorname{diam}(K)=\sup_{y,z\in K}|y-z|. The neural adaptation uses this as a target upper bound for the Jacobian operator norm.

Implementation notes

(1) Integration point: use this as the final latent-to-data map in a convex-potential normalizing flow or VAE decoder. Parameterize \Phi_ heta(x)=\frac12|x|^2+f_ heta(x) with an input-convex neural network, so T_ heta(x)=
abla\Phi_ heta(x) is monotone. Assume a latent Gaussian with covariance \Sigma; set Q=\Sigma^{-1}, or estimate a conservative Q from the latent density. Choose a known compact convex output domain K, such as [-1,1]^d or a data bounding box. (2) Pseudocode: sample latent x; compute T=grad_x(Phi_theta(x)); estimate J=jacobian(T,x) on a small probe batch; set L*=0.587*sqrt(lambda_max(Q))*diam(K); optimize the task or transport loss plus lambda*relu(||J||_2-L*)^2. For tighter anisotropic control, sample unit directions v and penalize relu(v^T J v-0.587*sqrt(v^T Q v)*width_K(v))^2. Optionally rescale the final map by L*/max(estimated Jacobian norm,L*) during initialization. (3) The theorem supplies the constant 0.587 and the target bound; Jacobian norms, empirical widths for non-box K, and approximation errors are estimated numerically. (4) First experiment: train a 2D/8D convex-potential OT map from Gaussian latents to a uniform distribution on a box or convex polygon, comparing an unconstrained ICNN, spectral-normalized ICNN, and this curvature-capped version. Measure transport cost, maximum Jacobian norm, exploding-gradient frequency, and held-out likelihood or reconstruction error. Success is lower worst-case Jacobian and fewer unstable runs at equal transport cost, without dimension-dependent degradation as d increases.

Verification

Mechanism confirmed, baseline not beaten

Stage 1 · Toy mechanism gate: Passed ✓

Stage 2 · Mechanism transferred to benchmark: Confirmed ✓

Stage 2 · Practical benchmark result: No measurable effect

Methodology: 8 paired seeds, permutation test (p<0.05) against a learning-rate-tuned baseline. How verification works

Stage 1 — Mechanism check agent confidence 7/10

Built a reproducible certificate check and a small PyTorch projected-update transport experiment. The analytic product Brenier map had supremum Jacobian 0.797885, below the directional certificate 1.174 in d=1,2,8; projected optimization maintained zero post-update cap violations while the uncapped baseline reached Jacobian 3.92. Final MSE was effectively unchanged between methods (about 0.00123 in d=1, 0.00125 in d=2, and 0.001245 in d=8), so the observed effect is real stability control without toy-task accuracy loss, but not an accuracy improvement.

Agent confidence: 7/10

Baseline: Peak post-update Jacobian 3.92; cap violations 45, 33, and 8 for d=1,2,8; final MSE 0.001231, 0.001251, 0.001245.
Idea: Projected certificate cap peak Jacobian 1.174, 1.660, and 3.321; zero cap violations; final MSE 0.001230, 0.001250, 0.001245 for d=1,2,8.

Limitations: The neural map is only a one-parameter diagonal tanh surrogate, not a trained input-convex neural network or a general Brenier approximation. The global certificate used diam([-1,1]^d), which grows as sqrt(d), so dimension-free directional behavior was checked analytically but not demonstrated as a learned high-dimensional advantage. No spectral-normalized ICNN, likelihood, reconstruction metric, FLOP comparison, or multi-seed robustness study was run.

How to run: python3 experiment.py

Stage 2 — Benchmark vs. tuned baseline (arXiv:2608.15906) Sep 2, 2026

No significant difference vs. the tuned baseline (0 metric points, p=1, wins 0 / 8 paired seeds).

Benchmark: Dynamics forecasting
Model: rnn_small
Paired seeds: 8
Baseline mean: 0
Idea mean: 0
Effect (Δ): 0 (n/a; negative = idea better)
Wins: 0 / 8 paired seeds
p-value: 1 (permutation test, 20 000 shuffles)
Smallest detectable effect: ±0.0%
Mechanism: Confirmed ✓
Practical effect: No effect
Baseline tuning: swept over 0 configs

Limitations:

The registered dynamics track does not directly test compact-domain Brenier transport or an ICNN architecture. The certificate cap was loose for this task, so it never activated; no dimension sweep, spectral-normalized ICNN comparator, likelihood evaluation, or deliberately high-curvature transport task was tested.

How to run: python3 bench_experiment.py

Verdict computed by deterministic test code from paired-seed statistics — not by the language model.

Stage 2 — Benchmark Sep 2, 2026 latest Failed ✗

Agent confidence: 9/10

Implemented the certificate-based Jacobian projection on the structurally matched registered dynamics track using identical rnn_small systems and equal-budget training. The mechanism signature was confirmed from trained-model behavior, but the cap never activated and the paired test MSE difference was exactly zero (p=1.0), so there is no observed benchmark win.

Baseline: Best lr=0.01; 8-seed mean test MSE 1.0635505304890103e-05; observed maximum Jacobian 0.5639464855194092.
Comparison: delta_mean=0.0, idea_wins=0/8, p_value=1.0, verdict=no measurable effect
Idea: Best lr=0.01 and certificate cap=2.0334276480858615; 8-seed mean test MSE 1.0635505304890103e-05; observed maximum Jacobian 0.5639464855194092; zero cap events.

Benchmark result