# Critical Interface State-Space Pooling

- ID: 2680
- Canonical URL: https://synthcore.org/idea/2680/critical-interface-state-space-pooling
- API JSON: https://synthcore.org/api/idea/2680.json
- API Markdown: https://synthcore.org/api/idea/2680.md
- Verification status: unverified
- Source: [arXiv:2608.21791](https://arxiv.org/abs/2608.21791)
- Category: architecture
- Solves: accuracy, sample-efficiency
- ML areas: ssm, attention, embedding
- Math tags: pde, dynamical-systems, spectral-theory, topology
- Ratings: usefulness 4/10; difficulty 5/10; novelty 8/10

## Idea description

Add a fixed or weakly learned interface-localized branch to a sequence model. Set the critical mass term to zero and make the transport coefficient change sign across a learnable interface, producing a localized mode that pools information near a detected transition rather than averaging uniformly over the sequence.

## Mathematical statement

The paper uses the low-energy operator \(\mathcal{D}=-A\,\partial_x^2-\kappa(x)\,\partial_x+m(x)\), where \(A\) is the second-derivative coefficient, \(\kappa(x)\) is a position-dependent kinetic or transport coefficient, and \(m(x)\) is a mass term. In the critical case \(m(x)=0\), the zero-mode equation is \([-A\partial_x^2-\kappa(x)\partial_x]\psi(x)=0\). For constant asymptotic \(\kappa\), the paper gives \(\psi(x)\sim\exp[-(\kappa/A)x]\), so the sign of \(\kappa/A\) determines decay away from an interface. Transfer this by choosing \(\kappa(x)=\kappa_0\tanh((x-x_0)/w)\), where \(x_0\) is the interface position, \(w>0\) is its width, and \(\kappa_0\) controls the asymptotic transport magnitude. Discretize the operator on sequence positions, then use the normalized smallest-singular-vector of the resulting matrix as a localized pooling profile. The mass remains fixed at zero; localization is caused by the transport sign reversal rather than an ordinary potential well.

## Key formulas

- $$\mathcal{D}=-A\,\partial_x^2-\kappa(x)\,\partial_x+m(x)$$
- $$\bigl[-A\,\partial_x^2-\kappa(x)\,\partial_x\bigr]\psi(x)=0\qquad\text{when }m(x)=0$$
- $$\psi(x)\sim\exp\!\left[-\frac{\kappa}{A}x\right]$$
- $$\kappa(x)=\kappa_0\tanh\!\left(\frac{x-x_0}{w}\right),\qquad z=\sum_{i=1}^{N}\psi_i h_i,\qquad \psi=\frac{v_{\min}(D)}{\|v_{\min}(D)\|_2}$$

## Implementation notes

(1) Integration point: insert this as an auxiliary pooling branch before the classifier or before the final transformer block. For a sequence tensor H in R^{B x N x d}, define normalized positions x_i=i/(N-1), a learnable interface position x_0 constrained to [0,1], and positive parameters A, w, and kappa_0. Compute kappa_i=kappa_0*tanh((x_i-x_0)/w). Assemble an N x N finite-difference matrix D whose interior action is D h at index i equal to -A*(h_{i+1}-2h_i+h_{i-1})/dx^2 - kappa_i*(h_{i+1}-h_{i-1})/(2*dx), with dx=1/(N-1). Use reflecting or zero-flux boundary rows. (2) Pseudocode: compute kappa; assemble D; estimate the unit right singular vector psi associated with the smallest singular value of D using power iteration on D^T D; compute z=sum_i psi_i H[:,i,:]; concatenate z with mean pooling or the CLS representation; apply a learned projection. For stable initialization, first use a differentiable soft profile psi_i proportional to exp(-abs(x_i-x_0)/ell), with ell=A/(abs(kappa_i)+epsilon), and normalize it; later compare against the numerical singular-vector profile. (3) The paper supplies the critical condition m=0, the sign-changing kappa profile, the operator, and the exponential localization criterion. Estimate x_0, kappa_0, w, and the singular vector during training; measure localization with inverse participation ratio IPR=sum_i psi_i^4. (4) First experiment: use a 2-layer Transformer or GRU on synthetic length-128 sequences whose label depends on a short event centered at a random change point, then test ECG anomaly classification or another localized-event dataset. Compare mean pooling, CLS attention, convolutional pooling, and this branch at matched parameter count. Track validation accuracy, accuracy at 10%, 25%, and 100% data, interface localization error, IPR, and runtime. Success means higher accuracy or sample efficiency specifically on interface-localized tasks, while no improvement on uniformly distributed tasks is an informative falsification.

## Disclaimer

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