# Quasiconformal distortion barrier for neural warps

- ID: 258
- Canonical URL: https://synthcore.org/idea/258/quasiconformal-distortion-barrier-for-neural-warps
- API JSON: https://synthcore.org/api/idea/258.json
- API Markdown: https://synthcore.org/api/idea/258.md
- Verification status: mechanism_failed
- Source: [arXiv:2607.04051](https://arxiv.org/abs/2607.04051)
- Category: regularization
- Solves: stability, accuracy, generalization
- ML areas: architecture, regularization, training
- Math tags: geometry, differential-geometry, optimization
- Ratings: usefulness 7/10; difficulty 4/10; novelty 6/10

## Idea description

Train a two-dimensional neural deformation map with the paper's Lp conformal-distortion energy instead of using only a determinant or smoothness penalty. The resulting barrier penalizes near-folds and directional collapse while permitting useful nonrigid deformation, making it suitable for spatial transformers, image registration, and learned coordinate warps.

## Mathematical statement

The paper defines the distortion of an orientation-preserving map f by K(z,f) = ||Df(z)||^2 / J(z,f), where Df(z) is the 2 by 2 derivative, ||Df(z)|| is its operator norm, and J(z,f) = det Df(z) > 0 is the Jacobian determinant. It gives the equivalent complex form K(z,f) = (1 + |mu_f|^2) / (1 - |mu_f|^2), where mu_f = f_bar / f_z is the complex dilatation and |mu_f| < 1 corresponds to positive orientation and finite distortion. The optimized functional is E_p(f) = integral of K(z,f)^p with respect to the area measure, for 1 <= p < infinity. The main theorem states that, in the homotopy class of a homeomorphism between analytically finite hyperbolic Riemann surfaces, an Lp minimizer exists, is unique, and is a diffeomorphism. For a neural map g_theta, use L_QC = (1/N) sum_i K_i^p, with K_i = ||Dg_theta(x_i)||_2^2 / (J_i + epsilon), and penalize negative J_i separately. The transferable mechanism is that K becomes large when J approaches zero or when the singular-value ratio becomes extreme.

## Key formulas

- $$\mathbb{K}(z,f)=\frac{\|Df(z)\|^{2}}{J(z,f)}=\frac{1+|\mu_f|^{2}}{1-|\mu_f|^{2}},\qquad\mu_f(z)=\frac{f_{\overline{z}}(z)}{f_z(z)}.$$
- $$\mathsf{E}_{p}(f:R,S)=\int_{R}{\mathbb{K}}^{p}(z,f)\,d\sigma_{R}(z),\qquad 1\leq p<\infty.$$
- $$\mathcal{L}_{\mathrm{QC}}(\theta)=\frac{1}{N}\sum_{i=1}^{N}\left(\frac{\|Dg_\theta(x_i)\|_2^2}{\max(J_i,\varepsilon)}\right)^{p}+\lambda_{\mathrm{task}}\mathcal{L}_{\mathrm{task}},\qquad J_i=\det Dg_\theta(x_i).$$
- $$\mathcal{L}_{\mathrm{QC}}^{\mathrm{cont}}(\theta)=\frac{1}{N}\sum_i\left(\frac{\|Dg_\theta(x_i)\|_2^2}{J_i+\varepsilon}\right)^{p_t},\qquad p_{t+1}=\min(p_t+\Delta p,p_{\max}).$$

## Implementation notes

First integrate the loss into a neural two-dimensional coordinate map g_theta from the unit square to the image plane. Suitable targets are an image-registration warp, a spatial-transformer sampler, a neural texture-coordinate map, or a planar deformation module. Use g_theta(x) = x + u_theta(x), with u_theta fixed to zero on the boundary or with fixed corner values, so that the learned map remains in the intended homotopy class. At sampled grid points x_i, compute Y_i = g_theta(x_i), the 2 by 2 Jacobian D_i by automatic differentiation or centered finite differences, the determinant J_i = det(D_i), and the singular values of D_i. Set K_i = max(singular_values(D_i))^2 / (J_i + epsilon), compute L_QC = mean(K_i^p), and optimize L_task + lambda_QC L_QC. If J_i is negative, add lambda_fold times softplus(-J_i / tau)^2; do not rely only on determinant clamping. Start with p = 1.5 or 2, then increase p by Delta p after the task loss decreases, stopping at p = 4 or 6. Use random grid samples during training and a dense fixed grid for validation. The mathematics directly supplies D_i, J_i, K_i, and the sampled Lp energy; the global fold rate, minimum determinant, and distortion quantiles are empirical diagnostics. Begin with synthetic checkerboard or MNIST registration using a small CNN, comparing task-only training, task plus squared Jacobian deviation, and task plus quasiconformal distortion. Measure registration error, fold percentage, minimum J, median and 95th-percentile K, and equal-FLOP convergence. Success means fewer folds and lower deformation error at similar task loss.

## Verification

- Status: mechanism_failed
- Mechanism evidence: yes
- Mechanism confirmed: no
- Verdict: Built a boundary-preserving 2D neural warp with automatic Jacobians, QC distortion, fold penalty, diagnostics, and comparisons against task-only and determinant regularization. The near-fold check confirmed K grows approximately as 1/J and the fold penalty activates for negative J, but the requested operator-norm formula does not match the paper's Hilbert–Schmidt/complex-dilatation identity for anisotropic maps. In the mini-experiment, QC produced excellent conditioning (min J=0.999, p95 K=1.001) but substantially worse target error (RMSE=0.128 versus 0.0027 task-only), while all methods had 0% folds; therefore no practical win was observed.

### Mechanism check

- Verdict: Built a boundary-preserving 2D neural warp with automatic Jacobians, QC distortion, fold penalty, diagnostics, and comparisons against task-only and determinant regularization. The near-fold check confirmed K grows approximately as 1/J and the fold penalty activates for negative J, but the requested operator-norm formula does not match the paper's Hilbert–Schmidt/complex-dilatation identity for anisotropic maps. In the mini-experiment, QC produced excellent conditioning (min J=0.999, p95 K=1.001) but substantially worse target error (RMSE=0.128 versus 0.0027 task-only), while all methods had 0% folds; therefore no practical win was observed.
- Confidence: 9/10
- Limitations: Only a small synthetic smooth warp was tested for 500 optimization steps; no image registration, spatial-transformer benchmark, hyperparameter sweep, continuation schedule, equal-FLOP analysis, or multi-seed evaluation was performed. The implementation follows the supplied operator-norm plan, not the paper's Hilbert–Schmidt norm, and uses a strong fixed QC weight that may overconstrain this easy task.

## Artifacts

- [experiment.py](https://synthcore.org/code/42/experiment.py)
- [report.md](https://synthcore.org/code/42/report.md)
- [results.json](https://synthcore.org/code/42/results.json)
- [Download all files as ZIP](https://synthcore.org/download/42)

## Disclaimer

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