# Dyadic Resolvent Filter Layer

- ID: 1226
- Canonical URL: https://synthcore.org/idea/1226/dyadic-resolvent-filter-layer
- API JSON: https://synthcore.org/api/idea/1226.json
- API Markdown: https://synthcore.org/api/idea/1226.md
- Verification status: mechanism_failed
- Source: [arXiv:2607.21278](https://arxiv.org/abs/2607.21278)
- Category: architecture
- Solves: accuracy, stability, scalability
- ML areas: graph-nn, ssm, regularization
- Math tags: spectral-theory, functional-analysis, harmonic-analysis, numerical-analysis
- Ratings: usefulness 7/10; difficulty 6/10; novelty 6/10

## Idea description

Replace a single polynomial graph filter or dense inverse with a multiscale layer that applies unitary propagations at dyadic times and combines them according to the dyadic resolvent identity. For a symmetric graph operator, this implements a stable rational spectral filter that can selectively retain or suppress frequency bands while exposing logarithmic multiscale structure.

## Mathematical statement

Proposition 1 states that for any self-adjoint operator A on a Hilbert space, with unitary evolution U_t = exp(-itA), and lambda greater than zero, the resolvent satisfies (A - i lambda)^(-1) = i(1 - exp(-lambda) U_1)^(-1) - i sum over k of 2^(-k)(1 + exp(-lambda/2^k) U_(2^(-k)))^(-1). Here A is a real symmetric graph operator such as a normalized Laplacian, U_t is its unitary propagator, lambda is a positive damping or spectral scale, and each inverse is an operator inverse. The identity converges in the strong operator topology. Each U_t preserves the feature norm because U_t is unitary, while the dyadic factors 2^(-k) explicitly weight progressively finer time scales. For implementation, use a finite truncation K and approximate every inverse by a finite Neumann expansion: (I + c U_t)^(-1) X approximately equals sum from j=0 to J of (-c)^j U_t^j X, where c = exp(-lambda/2^k). The resulting feature map approximates a complex resolvent; concatenate real and imaginary parts and apply a learned real projection.

## Key formulas

- $$(A-i\lambda)^{-1}=i\,(1-e^{-\lambda}U_{1})^{-1}-i\sum_{k=1}^{\infty}\frac{1}{2^{k}}\bigl(1+e^{-\lambda/2^{k}}U_{2^{-k}}\bigr)^{-1},\qquad U_t=e^{-itA},\quad \lambda>0.$$
- $$(A-i\lambda)^{-1}=i\sum_{j=0}^{\infty}e^{-j\lambda}U_j-i\lim_{\ell\to\infty}\sum_{k=1}^{\ell}\sum_{j=0}^{\infty}(-1)^j2^{-k}e^{-j\lambda/2^k}U_{j2^{-k}}.$$
- $$\widehat{R}_{K,J}(X)=i\sum_{j=0}^{J}e^{-j\lambda}U_1^jX-i\sum_{k=1}^{K}\frac{1}{2^k}\sum_{j=0}^{J}(-1)^j e^{-j\lambda/2^k}U_{2^{-k}}^jX.$$
- $$U_t^*U_t=I,\qquad \|U_tX\|_2=\|X\|_2\quad\text{when }A=A^*.$$

## Implementation notes

(1) Integration point: insert the layer between graph message-passing blocks. Let L = D^(-1/2)(D - A_g)D^(-1/2) be a symmetric normalized graph Laplacian, let X in R^(n x d) be node features, and set A = L after rescaling its spectrum to a bounded interval. Maintain real and imaginary feature tensors X_R and X_I, or concatenate the two channels and use a learned real projection at the end. (2) Pseudocode: set Y = 0; for j from 0 to J, compute Z_j = U_1^j X recursively and add i exp(-j lambda) Z_j to Y. For k from 1 to K, set c_k = exp(-lambda / 2^k), initialize Z = X, add -i 2^(-k) Z to Y, then repeat J times: Z = -c_k U_(2^(-k)) Z; add -i 2^(-k) Z to Y. Return concat(real(Y), imag(Y)) W_out plus a residual X W_res. (3) Compute U_t X using Chebyshev approximation to exp(-itL), Lanczos expmv, or a fixed-degree polynomial approximation. The paper determines the dyadic coefficients and resolvent structure; estimate truncation and exponential-approximation errors empirically by comparing K,J with K+1,J+2 and measuring the relative feature difference. Keep lambda fixed initially, then optionally learn lambda through lambda = softplus(theta) plus a small positive floor. (4) First experiment: use Cora and ogbn-arxiv, comparing a 2-layer GCN, ChebNet with the same number of sparse graph multiplications, and this layer with K in {2,4,6} and J in {2,4,8}. Report validation accuracy, accuracy under random edge deletion, feature norm growth across 8 stacked layers, and wall-clock time. A positive result is higher accuracy or perturbation robustness at matched sparse-operation count, together with bounded feature norms and a smoother training loss than the polynomial baseline.

## Verification

- Status: mechanism_failed
- Mechanism evidence: yes
- Mechanism confirmed: no
- Verdict: Built a readable MVP implementing exact finite-K dyadic resolvents, finite Neumann dyadic features, unitary propagation, and a synthetic graph classification comparison. The mechanism manifested: increasing K by 2 reduced exact resolvent error by approximately 4x, consistent with the predicted O(2^-K) tail; unitary propagation preserved norms to double precision; and Neumann convergence became substantially slower as lambda decreased because exp(-lambda/2^k) approaches 1 at fine scales. However, the dyadic feature layer did not win the mini-experiment: it achieved 0.850 validation accuracy versus 0.925 for the degree-6 polynomial baseline and was 4.3-5.9x slower, with larger feature norm ratios.

### Mechanism check

- Verdict: Built a readable MVP implementing exact finite-K dyadic resolvents, finite Neumann dyadic features, unitary propagation, and a synthetic graph classification comparison. The mechanism manifested: increasing K by 2 reduced exact resolvent error by approximately 4x, consistent with the predicted O(2^-K) tail; unitary propagation preserved norms to double precision; and Neumann convergence became substantially slower as lambda decreased because exp(-lambda/2^k) approaches 1 at fine scales. However, the dyadic feature layer did not win the mini-experiment: it achieved 0.850 validation accuracy versus 0.925 for the degree-6 polynomial baseline and was 4.3-5.9x slower, with larger feature norm ratios.
- Confidence: 9/10
- Limitations: Only a small synthetic 80-node graph and closed-form ridge classifier were tested; no Cora or ogbn-arxiv, learned lambda, Chebyshev/Lanczos propagation, edge-deletion robustness, deep stacking, GPU timing, or broad hyperparameter search was evaluated.

## Artifacts

- [experiment.py](https://synthcore.org/code/495/experiment.py)
- [report.md](https://synthcore.org/code/495/report.md)
- [results.json](https://synthcore.org/code/495/results.json)
- [Download all files as ZIP](https://synthcore.org/download/495)

## Disclaimer

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