# Max-Volume Transport Skeleton

- ID: 342
- Canonical URL: https://synthcore.org/idea/342/max-volume-transport-skeleton
- API JSON: https://synthcore.org/api/idea/342.json
- API Markdown: https://synthcore.org/api/idea/342.md
- Verification status: mech_ok_no_baseline
- Source: [arXiv:2607.01696](https://arxiv.org/abs/2607.01696)
- Category: training
- Solves: sample-efficiency, scalability, accuracy
- ML areas: data-augmentation, world-model, embedding, curriculum
- Math tags: linear-algebra, approximation-theory, optimal-transport, numerical-analysis
- Ratings: usefulness 7/10; difficulty 5/10; novelty 7/10

## Idea description

Use a maximal-volume cross approximation of the parameter-by-space transport-signature matrix to select informative training conditions and compact spatial features. This provides an active-learning alternative to random snapshot selection or ordinary PCA, targeting parameters that are difficult to interpolate from the current reduced representation.

## Mathematical statement

Let \(S(\mu,x)\) be the transport signature for parameter \(\mu\) and spatial index \(x\), such as a discretized weighted-Laplacian potential or its coefficients in a fine basis. After discretization, \(S\in\mathbb R^{N_\mu\times N_x}\). Choose row indices \(I\) and column indices \(J\), each of size \(r\), and form the skeleton approximation \(\hat S=S[:,J]S[I,J]^{-1}S[I,:]\). The maximal-volume criterion seeks indices maximizing \(|\det S[I,J]|\), which favors linearly independent rows and columns and generally improves interpolation conditioning. Exact optimization is combinatorial, so pivoted QR or greedy determinant-increasing swaps are practical substitutes. For an unselected parameter \(\mu\), define the residual \(e(\mu)=\|S(\mu,:)-S(\mu,J)S[I,J]^{-1}S[I,:]\|_2\). Adding the parameter with largest residual is an active-learning rule for generating the next expensive PDE or transport snapshot.

## Key formulas

- $$\hat S=S[:,J]S[I,J]^{-1}S[I,:],\qquad |I|=|J|=r.$$
- $$(I,J)\approx\operatorname*{arg\,max}_{|I|=|J|=r}\left|\det S[I,J]\right|.$$
- $$e(\mu)=\left\|S(\mu,:)-S(\mu,J)S[I,J]^{-1}S[I,:]\right\|_2.$$
- $$\mu_{\mathrm{new}}=\operatorname*{arg\,max}_{\mu\in\mathcal P_{\mathrm{pool}}}e(\mu).$$

## Implementation notes

Use this as an offline active-learning and basis-construction procedure for a conditional density model or learned reduced-order simulator. Start with a pool of \(N_\mu\) candidate conditions and compute an approximate transport signature for each condition by solving discrete optimal transport from the fixed reference distribution. Stack the signatures into \(S\). Obtain initial row and column selections using pivoted QR on \(S\) and \(S^\mathsf{T}\), or initialize randomly and perform greedy max-volume swaps. For each candidate swap, update or recompute \(C=S[I,J]^{-1}\) and accept the swap if the determinant-volume proxy increases. Train the neural coefficient predictor only on the selected parameter rows \(I\), using either the selected signature columns \(S[I,J]\) or reduced coefficients as targets. During refinement, evaluate the residual `e(mu)` for every unselected pool point, add the largest-residual condition to the expensive training set, recompute the skeleton, and stop when the maximum residual is below a chosen tolerance. The mathematical quantities are the skeleton approximation, determinant criterion, and residual; QR, approximate OT, and numerical matrix inversion are implementation approximations. Begin with a two-dimensional advection or continuity-equation problem, 1000 candidate parameters, and budgets of 8, 16, and 32 snapshots. Compare random selection, PCA plus random selection, and max-volume selection. Report held-out Wasserstein error, worst-case error, condition number of \(S[I,J]\), and number of expensive snapshots. Success is reaching a fixed target error with fewer PDE solves and reducing worst-case error rather than only average error.

## Verification

- Status: mech_ok_no_baseline
- Mechanism evidence: yes
- Mechanism confirmed: yes
- Practical verdict: inconclusive
- Verdict: Built a readable max-volume transport-skeleton MVP using pivoted QR cross selection and stable linear solves. The rank-4 exactness check achieved 2.38e-16 relative error; across ranks 4, 8, 12, and 16, QR residuals were 0.184, 0.0129, 0.00112, and 7.01e-5, while random median residuals were 7.99, 6.09, 4.74, and 0.973. Cross determinants were dramatically larger and conditioning substantially better than random crosses, so the proposed mechanism manifested clearly in this toy setting.

### Mechanism check

- Verdict: Built a readable max-volume transport-skeleton MVP using pivoted QR cross selection and stable linear solves. The rank-4 exactness check achieved 2.38e-16 relative error; across ranks 4, 8, 12, and 16, QR residuals were 0.184, 0.0129, 0.00112, and 7.01e-5, while random median residuals were 7.99, 6.09, 4.74, and 0.973. Cross determinants were dramatically larger and conditioning substantially better than random crosses, so the proposed mechanism manifested clearly in this toy setting.
- Confidence: 8/10
- Limitations: Used translated Gaussian proxy signatures rather than solving a PDE or optimal-transport problem. Did not train a neural coefficient predictor, compare PCA, compute Wasserstein error, measure wall-clock expensive-snapshot savings, or perform exact max-volume swaps.

### Practical benchmark

- Paired seeds: 8
- Baseline mean: 0
- Idea mean: 0
- p-value: 0.64895
- Paired wins: 4/8
- Benchmark verdict: no significant win

## Artifacts

- [REPORT.md](https://synthcore.org/code/116/REPORT.md)
- [bench_report.json](https://synthcore.org/code/116/bench_report.json)
- [experiment.py](https://synthcore.org/code/116/experiment.py)
- [report.md](https://synthcore.org/code/116/report.md)
- [report_bench_2026-08-30T145812.md](https://synthcore.org/code/116/report_bench_2026-08-30T145812.md)
- [results.json](https://synthcore.org/code/116/results.json)
- [stage2_bench.py](https://synthcore.org/code/116/stage2_bench.py)
- [transport_track.py](https://synthcore.org/code/116/transport_track.py)
- [Download all files as ZIP](https://synthcore.org/download/116)

## Disclaimer

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