# Parabolic Riesz Feature Preconditioner

- ID: 425
- Canonical URL: https://synthcore.org/idea/425/parabolic-riesz-feature-preconditioner
- API JSON: https://synthcore.org/api/idea/425.json
- API Markdown: https://synthcore.org/api/idea/425.md
- Verification status: mech_ok_no_baseline
- Source: [arXiv:2607.05181](https://arxiv.org/abs/2607.05181)
- Category: geometry
- Solves: stability, accuracy, sample-efficiency
- ML areas: ssm, graph-nn, attention, training
- Math tags: harmonic-analysis, functional-analysis, pde, numerical-analysis
- Ratings: usefulness 6/10; difficulty 5/10; novelty 7/10

## Idea description

Add a learned Riesz-transform branch that extracts normalized spatial gradients after diffusion by a positive parabolic operator. The diffusion branch carries smooth semantic content, while the Riesz branch represents boundaries, motion changes, and graph discontinuities. Resolvent smoothing makes the derivative branch less sensitive to feature noise than directly applying a finite difference.

## Mathematical statement

The paper defines the Riesz transform associated with a second-order operator \(\mathcal{L}\) and first-order differential structure \(\mathbb{D}\) by \(\mathcal{R}_{\mathcal{L}}=\mathbb{D}\mathcal{L}^{-1/2}\). For \(\mathcal{H}=\partial_t-\operatorname{Div}_x(A\nabla_x)\), \(\mathbb{D}\) can be a spatial gradient or a combined spatial and temporal difference. The paper defines \(q_-(\mathcal{H})=\inf\{p\geq1:(\lambda\mathbb{D}\mathcal{E}_\lambda)_{\lambda>0}\text{ is uniformly }L^p\text{-bounded}\}\), where \(\mathcal{E}_\lambda=(1+\lambda^2\mathcal{H})^{-1}\) is the resolvent family. The implementation uses the resolvent approximation \(R_\lambda X=\lambda\mathbb{D}(I+\lambda^2H)^{-1}X\) for \(\mathbb{D}\mathcal{H}^{-1/2}\). Here \(H\) is a discrete parabolic operator, \(\lambda\) is the derivative scale, \(\mathbb{D}\) is a discrete incidence or finite-difference operator, and \((I+\lambda^2H)^{-1}\) suppresses frequencies above the scale before differentiation.

## Key formulas

- $$\mathcal{R}_{\mathcal{L}}\coloneqq\mathbb{D}\mathcal{L}^{-1/2}$$
- $$q_{-}(\mathcal{H})\coloneqq\inf\left\{p\geq1:\ \left(\lambda\mathbb{D}\mathcal{E}_{\lambda}\right)_{\lambda>0}\ \text{is uniformly }\mathrm{L}^{p}\text{ bounded}\right\}$$
- $$\mathcal{E}_{\lambda}=(1+\lambda^{2}\mathcal{H})^{-1},\qquad R_{\lambda}X=\lambda\mathbb{D}(I+\lambda^{2}H)^{-1}X$$
- $$X_{\mathrm{out}}=X+W_sX+gW_rR_{\lambda}X$$

## Implementation notes

(1) Integration point: add the branch after patch embedding or after each early spatiotemporal encoder block, before the nonlinear MLP. Use it for video, event-camera, optical-flow, traffic, or graph signals where information is concentrated at spatial and temporal discontinuities. Let \(X\in\mathbb{R}^{B\times T\times N\times d}\). Construct a sparse graph Laplacian \(L_A\), a causal temporal difference matrix \(D_t\), and \(H=D_t+L_A+\epsilon I\), with \(\epsilon>0\). Let \(\mathbb{D}\) be the spatial incidence matrix, optionally concatenated with a scaled temporal difference. (2) Pseudocode: for each selected scale \(\lambda\), solve \((I+\lambda^2H)Z=X\) using 5--20 preconditioned conjugate-gradient iterations; compute \(R=\lambda\mathbb{D}Z\); apply RMSNorm to \(R\); return \(X_{out}=X+W_sX+gW_rR\), where \(g\) is initialized to zero or 0.1 and learned. Use multiple scales and concatenate the resulting \(R\)'s if memory permits. (3) The mathematical component is the dimensionless factor \(\lambda\mathbb{D}\) combined with the resolvent inverse, which prevents derivative magnitude from increasing uncontrollably with resolution. Estimate the operator norm of \(\lambda\mathbb{D}(I+\lambda^2H)^{-1}\) using five-step power iteration and reduce \(g\) if the estimate exceeds a target such as 2. Select scales from the empirical feature correlation length; the paper does not provide a neural-network-specific optimal scale. (4) First experiment: compare a six-layer graph transformer or small VideoMamba with and without this branch on Moving-MNIST, traffic forecasting, or UCF101 using 64 nodes and 16 time steps. Measure validation accuracy, sensitivity to 10% feature perturbations, gradient-norm variance, and training loss. Compare against direct \(\mathbb{D}X\) and an ordinary convolutional high-pass branch. Success is improved boundary or motion accuracy and lower corruption sensitivity for no more than 20% additional FLOPs.

## Verification

- Status: mech_ok_no_baseline
- Mechanism evidence: yes
- Mechanism confirmed: yes
- Practical verdict: inconclusive
- Verdict: Built a compact Parabolic Riesz Feature Preconditioner MVP using a ring-graph Laplacian and the resolvent operator lambda*D*(I+lambda^2 H)^-1, with RMS normalization and a learned residual gate. The numerical check reproduced the claimed smoothing effect: at n=32/64/128, the Riesz operator norms were about 0.50 versus 2.00 for direct differences, and measured noise amplification was roughly 0.41 versus 1.22–1.54. The toy classifier did not show an accuracy win: all variants reached 1.0 accuracy and the Riesz branch had the highest validation loss, so the result supports stability/noise suppression but not improved task performance.

### Mechanism check

- Verdict: Built a compact Parabolic Riesz Feature Preconditioner MVP using a ring-graph Laplacian and the resolvent operator lambda*D*(I+lambda^2 H)^-1, with RMS normalization and a learned residual gate. The numerical check reproduced the claimed smoothing effect: at n=32/64/128, the Riesz operator norms were about 0.50 versus 2.00 for direct differences, and measured noise amplification was roughly 0.41 versus 1.22–1.54. The toy classifier did not show an accuracy win: all variants reached 1.0 accuracy and the Riesz branch had the highest validation loss, so the result supports stability/noise suppression but not improved task performance.
- Confidence: 8/10
- Limitations: Only a 1-D periodic graph and a small synthetic frequency-classification task were tested. No real video, traffic, graph benchmark, FLOP accounting, multi-scale branch, causal temporal operator, conjugate-gradient approximation, or statistically repeated training runs were evaluated.

### Practical benchmark

- Paired seeds: 8
- Baseline mean: 0
- Idea mean: 0
- p-value: 0.42645
- Paired wins: 4/8
- Benchmark verdict: no significant win

## Artifacts

- [bench_report.json](https://synthcore.org/code/1295/bench_report.json)
- [experiment.py](https://synthcore.org/code/1295/experiment.py)
- [report.md](https://synthcore.org/code/1295/report.md)
- [report_bench_2026-09-04T123408.md](https://synthcore.org/code/1295/report_bench_2026-09-04T123408.md)
- [results.json](https://synthcore.org/code/1295/results.json)
- [stage2_bench.py](https://synthcore.org/code/1295/stage2_bench.py)
- [Download all files as ZIP](https://synthcore.org/download/1295)

## Disclaimer

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