# Kriging-guided coordinate sampling

- ID: 1354
- Canonical URL: https://synthcore.org/idea/1354/kriging-guided-coordinate-sampling
- API JSON: https://synthcore.org/api/idea/1354.json
- API Markdown: https://synthcore.org/api/idea/1354.md
- Verification status: unverified
- Source: [arXiv:2607.24283](https://arxiv.org/abs/2607.24283)
- Category: training
- Solves: sample-efficiency, accuracy
- ML areas: data-augmentation, world-model, embedding
- Math tags: probability, statistics, optimization, approximation-theory
- Ratings: usefulness 6/10; difficulty 5/10; novelty 6/10

## Idea description

Train a coordinate MLP or neural operator using locations selected by an ordinary-kriging estimate of the unresolved field rather than by uniform random sampling. At each acquisition round, estimate the local reconstruction variance from the current labeled set and query points with the largest variance, optionally weighted by their application importance.

## Mathematical statement

The paper models the spatial field as \(\Gamma(\bm{x})=\mu(\bm{x})+S(\bm{x})>0\), where \(\bm{x}\in\mathcal{D}\subset\mathbb{R}^{3}\), \(\mu\) is a deterministic mean, and \(S\) is a zero-mean Gaussian random field. For a target point \(\bm{x}\), let \(\mathcal{N}_{m}(\bm{x})=\{\bm{x}_{1},\ldots,\bm{x}_{m}\}\) be its selected neighboring measurement locations, let \(\gamma_{ij}\) denote the paper's spatial variogram or covariance quantity between \(\bm{x}_{i}\) and \(\bm{x}_{j}\), and let \(\gamma_{i0}\) denote the same quantity between neighbor \(\bm{x}_{i}\) and target \(\bm{x}\). Ordinary-kriging weights \(\lambda_{1},\ldots,\lambda_{m}\) and multiplier \(\upsilon\) solve \(\bm{R}^{\mathcal{S}}\bm{\Lambda}^{\mathcal{S}}=\bm{r}^{\mathcal{S}}\), with \(\bm{R}^{\mathcal{S}}=\begin{bmatrix}\gamma_{11}&\cdots&\gamma_{1m}&1\\\vdots&&\vdots&\vdots\\\gamma_{m1}&\cdots&\gamma_{mm}&1\\1&\cdots&1&0\end{bmatrix}\), \(\bm{\Lambda}^{\mathcal{S}}=[\lambda_{1},\ldots,\lambda_{m},-\upsilon]^{\mathsf T}\), and \(\bm{r}^{\mathcal{S}}=[\gamma_{10},\ldots,\gamma_{m0},1]^{\mathsf T}\). The constraint \(\sum_i\lambda_i=1\) makes the estimator reproduce an unknown constant mean. Its local mean-squared interpolation error is \(\sigma_{\epsilon_{\bm{x}}}^{2}=2\sum_i\lambda_i\gamma_{i0}-\sum_i\lambda_i\sum_j\lambda_j\gamma_{ij}\). This scalar becomes the acquisition score for selecting new neural-network labels.

## Key formulas

- $$\Gamma(\bm{x})=\mu(\bm{x})+S(\bm{x})>0,\qquad S(\bm{x})\sim\mathcal{N}(0,\sigma^2(\bm{x})).$$
- $$\bm{R}^{\mathcal{S}}\bm{\Lambda}^{\mathcal{S}}=\bm{r}^{\mathcal{S}},\quad \bm{\Lambda}^{\mathcal{S}}=\begin{bmatrix}\lambda_1&\cdots&\lambda_m&-\upsilon\end{bmatrix}^{\mathsf T},\quad \bm{r}^{\mathcal{S}}=\begin{bmatrix}\gamma_{10}&\cdots&\gamma_{1m}&1\end{bmatrix}^{\mathsf T},$$
$$\bm{R}^{\mathcal{S}}=\begin{bmatrix}\gamma_{11}&\cdots&\gamma_{1m}&1\\\vdots&&\vdots&\vdots\\\gamma_{m1}&\cdots&\gamma_{mm}&1\\1&\cdots&1&0\end{bmatrix}.$$
- $$\sigma_{\epsilon_{\bm{x}}}^{2}=2\sum_{\bm{x}_{n}\in\mathcal{N}_{m}(\bm{x})}\lambda_n\gamma_{n0}-\sum_{\bm{x}_{n}\in\mathcal{N}_{m}(\bm{x})}\lambda_n\left(\sum_{\bm{x}_{j}\in\mathcal{N}_{m}(\bm{x})}\lambda_j\gamma_{nj}\right).$$
- $$a(\bm{x})=w(\bm{x})\,\sigma_{\epsilon_{\bm{x}}}^{2},\qquad \bm{x}_{\mathrm{new}}=\arg\max_{\bm{x}\in\mathcal{C}\setminus\mathcal{S}}a(\bm{x}),$$

## Implementation notes

(1) Integration point: use this as the data-acquisition sampler for a coordinate MLP \(f_\theta(\bm{x})\), neural field, or neural operator whose labels \(y(\bm{x})\) are expensive. It changes which coordinates enter the supervised minibatches; the network architecture and optimizer remain unchanged. Maintain a labeled set \(\mathcal{S}\) and a finite candidate pool \(\mathcal{C}\). (2) Pseudocode: initialize \(\mathcal{S}\) with a small uniform or Latin-hypercube design; train \(f_\theta\) for \(T\) steps; compute residuals \(e_i=y(\bm{x}_i)-f_\theta(\bm{x}_i)\); estimate a variogram from residual pairs, for example \(\hat\gamma(\bm{x},\bm{x}')=\tfrac12(e(\bm{x})-e(\bm{x}'))^2\), with distance-bin averaging and a nugget for numerical stability; for each candidate \(\bm{x}\), select its \(m\) nearest labeled points, construct \(\bm{R}^{\mathcal{S}}\) and \(\bm{r}^{\mathcal{S}}\), solve the augmented linear system, evaluate the paper's \(\sigma_{\epsilon_{\bm{x}}}^{2}\), and append the top \(B\) candidates after nonmaximum suppression. Repeat until the query budget is exhausted. (3) The kriging system and error formula are fixed by the paper; variogram parameters, residuals, neighbor lists, and the optional importance weight \(w(\bm{x})\) are estimated empirically. Use Cholesky with diagonal jitter rather than explicit matrix inversion. Refit the variogram every few acquisition rounds, not every gradient step. (4) First experiment: fit a 4-layer SIREN or Fourier-feature MLP to a 2D or 3D analytic field and to a stored PDE-simulation field using 1%, 5%, and 10% label budgets. Compare uniform random, Latin hypercube, farthest-point, and kriging-guided acquisition with identical network updates and queried-label counts. Evaluate dense-grid MSE, worst-region MSE, and wall-clock overhead. Success is lower dense-grid MSE at the same number of labels, or reaching a target MSE with at least 25% fewer expensive field evaluations; reject the method if kriging overhead dominates or if residual variograms are too noisy to improve over farthest-point sampling.

## Disclaimer

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