# PLMS Copula Augmentation Layer

- ID: 2577
- Canonical URL: https://synthcore.org/idea/2577/plms-copula-augmentation-layer
- API JSON: https://synthcore.org/api/idea/2577.json
- API Markdown: https://synthcore.org/api/idea/2577.md
- Verification status: unverified
- Source: [arXiv:2608.19608](https://arxiv.org/abs/2608.19608)
- Category: regularization
- Solves: generalization, sample-efficiency
- ML areas: data-augmentation, embedding, vae
- Math tags: probability, statistics, geometry
- Ratings: usefulness 5/10; difficulty 4/10; novelty 8/10

## Idea description

Generate pairs of latent variables with exactly uniform marginals but non-Gaussian, asymmetric dependence by applying a randomly chosen PLMS map to one uniform latent coordinate. The coupling can expose a model to controlled concordant, discordant, or piecewise-dependent examples without changing either marginal distribution.

## Mathematical statement

A PLMS map is specified by a partition 0=a_0<\cdots<a_N=1. On interval (a_{i-1},a_i), define \alpha_i(x)=(x-a_{i-1})/(a_i-a_{i-1}) if the branch is increasing and \alpha_i(x)=(a_i-x)/(a_i-a_{i-1}) if it is decreasing. Every branch maps its interval onto (0,1), so if U\sim\mathrm{Unif}(0,1) and V=\alpha(U), then V\sim\mathrm{Unif}(0,1) even though (U,V) is dependent; this is the deterministic copula C_{e,\alpha}. For two such factorizable couplings, the paper proves     au_{C_{e,\alpha}*C_{\beta,e}}=\rho_{C_{e,\alpha}*C_{\beta,e}}=    au_{C_{e,\alpha}}    au_{C_{\beta,e}}, where     au and \rho are Kendall's tau and Spearman's rho and * is the Markov product. The augmentation exploits exact marginal preservation and the multiplicative dependence composition law, rather than assuming a Gaussian correlation model.

## Key formulas

- $$\alpha_i(x)=\begin{cases}\dfrac{x-a_{i-1}}{a_i-a_{i-1}},&\text{increasing branch},\\[4pt]\dfrac{a_i-x}{a_i-a_{i-1}},&\text{decreasing branch},\end{cases}\qquad x\in(a_{i-1},a_i).$$
- $$U\sim\mathrm{Unif}(0,1),\qquad V=\alpha(U),\qquad V\sim\mathrm{Unif}(0,1).$$
- $$\rho_{C_{e,\alpha}*C_{e,\beta}}=\tau_{C_{e,\alpha}*C_{e,\beta}}=\tau_{C_{e,\alpha}}\,\tau_{C_{e,\beta}}.$$
- $$\widetilde z_1=F_1^{-1}(U),\qquad \widetilde z_2=F_2^{-1}(V).$$

## Implementation notes

Integrate this at the latent augmentation point of a VAE, contrastive encoder, or multimodal embedding model. Given a minibatch of latent vectors z in R^d, choose two coordinates or two examples, transform each selected coordinate to an empirical uniform value using U=rank(z)/(B+1), and sample a PLMS map alpha. Store a partition 0=a_0<...<a_N=1 and a sign s_i in {+1,-1} for each branch. For every scalar u, locate interval i and compute v=(u-a_{i-1})/(a_i-a_{i-1}) when s_i=+1, otherwise v=(a_i-u)/(a_i-a_{i-1}). Map back through the desired marginal inverse CDF, or simply use v as a latent coordinate. Pseudocode: u = rank(z_coord)/(B+1); i = bucketize(u,a); v = where(sign[i]>0,(u-a[i-1])/(a[i]-a[i-1]),(a[i]-u)/(a[i]-a[i-1])); z_aug = replace(z_coord,v). Estimate achieved dependence empirically with pairwise concordance, while the uniform-marginal guarantee and branch formulas come directly from the PLMS construction. Start with CIFAR-10 SimCLR using a 2-layer projection head and compare independent noise, Gaussian-copula noise, and PLMS coupling at equal augmentation count and FLOPs. Test partitions with two to eight branches and random sign patterns. Success is higher linear-probe accuracy or contrastive retrieval accuracy at equal training compute, with marginal histograms remaining uniform and measured tau matching the intended coupling.

## Disclaimer

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