# Adjoint spectral opening for stable pooling

- ID: 2619
- Canonical URL: https://synthcore.org/idea/2619/adjoint-spectral-opening-for-stable-pooling
- API JSON: https://synthcore.org/api/idea/2619.json
- API Markdown: https://synthcore.org/api/idea/2619.md
- Verification status: unverified
- Source: [arXiv:2608.20399](https://arxiv.org/abs/2608.20399)
- Category: architecture
- Solves: stability, accuracy, memory
- ML areas: cnn, memory, regularization
- Math tags: convex-analysis, harmonic-analysis, functional-analysis, numerical-analysis
- Ratings: usefulness 5/10; difficulty 6/10; novelty 6/10

## Idea description

Construct a learnable spectral pooling block as an erosion followed by its adjoint dilation, making the resulting opening idempotent, increasing, and anti-extensive. The block can suppress frequencies outside a learned passband while guaranteeing that applying it twice does not continue changing the representation, which is useful in multi-stage CNN pyramids and U-Net skip paths.

## Mathematical statement

For an adjunction (epsilon, delta), the paper gives epsilon(f) <=_2 g if and only if f <=_1 delta(g). Proposition 2.4 states that epsilon commutes with infima, delta commutes with suprema, epsilon is anti-extensive (epsilon(f) <= f), delta is extensive (g <= delta(g)), and gamma = delta composed with epsilon is an opening: gamma^2 = gamma, gamma(f) <= f, and gamma is increasing. In the adapted finite spectral model, f is a nonnegative frequency magnitude, epsilon_psi is a quotient erosion, and delta_psi is implemented as the order-adjoint supremum map delta_psi(g)(xi) = g(xi)(psi(xi)+epsilon). The discrete pair satisfies epsilon_psi(f) <= g if and only if f <= delta_psi(g) pointwise. Therefore gamma_psi(f) = delta_psi(epsilon_psi(f)) is an idempotent spectral filter. Because arbitrary FFT magnitudes will not necessarily satisfy every continuous-lattice regularity assumption, idempotence and the adjunction should be unit-tested directly on the discretized tensors rather than assumed from the theorem.

## Key formulas

- $$\varepsilon(f)\leq_2 g\iff f\leq_1\delta(g)$$
- $$\gamma:=\delta\circ\varepsilon,\qquad \gamma^2=\gamma,\qquad \gamma(f)\leq f$$
- $$\varepsilon_{\psi}(f)(\xi)=\frac{f(\xi)}{\psi(\xi)+\epsilon},\qquad \delta_{\psi}(g)(\xi)=g(\xi)\bigl(\psi(\xi)+\epsilon\bigr)$$
- $$\gamma_{\psi}(f)(\xi)=\delta_{\psi}(\varepsilon_{\psi}(f))(\xi)$$

## Implementation notes

Use this as a pooling or skip-connection operator in a frequency-aware U-Net or CNN. At a chosen resolution, compute `X=fft2(x)`, magnitude `f=abs(X)`, and phase `p=X/(f+eps)`. Parameterize a positive spectral structuring element `psi=softplus(theta)+psi_min`; initialize theta so psi is large outside the desired passband and near one inside it. Compute the erosion `e=f/(psi+eps)`, then the adjoint dilation `d=e*(psi+eps)`. In exact arithmetic d=f, so a literal same-grid quotient/multiplication pair is only a validity test, not a useful pooling layer. To make the opening nontrivial, insert a low-resolution projection P between them: downsample e to a coarse Fourier grid, optionally apply a nonnegative max/average aggregation, upsample it with U, and set `o=U(P(e))*(psi+eps)`, clipping `o` to f to retain anti-extensivity. Alternatively define delta as a learned supremum over neighboring frequency bins, then verify the adjunction numerically. Reconstruct `y=real(ifft2(o*p))`, and use y as the pooled branch or as a filtered skip feature. The code should test `||gamma(gamma(f))-gamma(f)||`, `gamma(f)<=f`, and monotonicity on random nonnegative f. First cheap experiment: compare a standard stride-2 convolution, average-pool-plus-convolution, and this opening block in a small U-Net on CIFAR-10 or 128x128 BSD denoising. Track aliasing under frequencies above the post-stride Nyquist limit, validation accuracy or PSNR, and repeated-application drift. Success means lower high-frequency alias energy and materially smaller representation drift when the opening is applied twice, with no worse accuracy at equal memory.

## Disclaimer

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