# Spectrally identifiable phaseless recurrent layer

- ID: 1461
- Canonical URL: https://synthcore.org/idea/1461/spectrally-identifiable-phaseless-recurrent-layer
- API JSON: https://synthcore.org/api/idea/1461.json
- API Markdown: https://synthcore.org/api/idea/1461.md
- Verification status: mech_ok_no_baseline
- Source: [arXiv:2607.26705](https://arxiv.org/abs/2607.26705)
- Category: architecture
- Solves: accuracy, sample-efficiency, stability
- ML areas: rnn, ssm, graph-nn, embedding
- Math tags: spectral-theory, linear-algebra, graph-theory, dynamical-systems
- Ratings: usefulness 7/10; difficulty 6/10; novelty 8/10

## Idea description

Build a complex-valued recurrent or graph-neural layer whose hidden state evolves under a fixed graph Schrödinger operator and is exposed to the downstream network only through coordinate magnitudes at several times. Choose the diagonal potential so that the spectrum has unique unordered pair sums, the squared-eigenvector matrix is invertible, and every eigenvector pair overlaps in at least one observed coordinate; the resulting temporal intensity code is theoretically injective up to one unavoidable global phase.

## Mathematical statement

The paper studies the real symmetric operator $H_Q=\Delta_G+Q$, where $\Delta_G$ is the graph Laplacian and $Q=\operatorname{diag}(q_1,\ldots,q_n)$ is a real diagonal potential. For an initial complex state $u_0\in\mathbb{C}^n$, the observable is $y_j(t)=|e^{-itH_Q}u_0(j)|$ for vertices $j$ and times $t\in[0,T]$. Writing $H_Q\phi_k=\lambda_k\phi_k$ with orthonormal real eigenvectors $\phi_k$, the sufficient criterion uses Condition $B_2$: all values $\lambda_j+\lambda_k$ for $1\leq j\leq k\leq n$ are pairwise distinct; Condition $(I)$: the squared-eigenvector matrix $S\in\mathbb{R}^{n\times n}$ with $S_{jk}=\phi_k(j)^2$ is invertible; and Condition $(S)$: every pair of eigenvectors has a vertex $j$ with $\phi_k(j)\phi_\ell(j)\neq0$. Under these conditions, the magnitude trajectory determines $u_0$ up to multiplication by a scalar $e^{i\theta}$.

## Key formulas

- $$H_Q=\Delta_G+Q,$$
- $$y_j(t)=\left|e^{-itH_Q}u_0(j)\right|,$$
- $$S_{jk}=\phi_k(j)^2,\qquad \det S\neq 0,$$
- $$\left\{\lambda_j+\lambda_k:1\leq j\leq k\leq n\right\}\ \text{pairwise distinct},\qquad \forall k\neq\ell\ \exists j:\ \phi_k(j)\phi_\ell(j)\neq0.$$

## Implementation notes

1. Integration point: replace the hidden-state transition in a small complex-valued RNN or graph neural network with a fixed graph Schrödinger encoder. Let $h\in\mathbb{C}^n$ be the incoming hidden state, choose a connected graph Laplacian $\Delta_G$, and compute $K$ intensity observations $z_r=|U_rh|$ with $U_r=\exp(-i\tau_rH_Q)$ and $\tau_r=r\Delta t$. Concatenate $z_0,\ldots,z_{K-1}$ and feed them to a real MLP or classifier. 2. Pseudocode: initialize $H=\Delta_G+\operatorname{diag}(q)$; precompute $U=\operatorname{matrix\_exponential}(-i\Delta tH)$; set $x=h$; for $r=0,\ldots,K-1$, append $|x|$ to a list and set $x=Ux$; concatenate the list and apply the readout. Use $K=4$--$16$ and normalize each magnitude vector by its RMS. 3. During initialization, eigendecompose $H=\Phi\Lambda\Phi^T$, calculate $S=\Phi\odot\Phi$, reject potentials with small $\sigma_{\min}(S)$, reject near-collisions among all pair sums $\lambda_j+\lambda_k$, and reject eigenvector pairs with no overlapping support. In floating point, use thresholds such as $\sigma_{\min}(S)>10^{-2}$ and minimum pair-sum separation above $10^{-3}$ after spectral normalization. The uniqueness theorem is imported from the paper; the time grid and conditioning thresholds are empirical. 4. First experiment: compare a 2-layer complex graph-RNN using this encoder against a same-parameter real GRU and a generic magnitude-only encoder on synthetic graph-sequence classification. Generate random complex initial states and labels invariant to global phase, train at equal parameter count and FLOPs, and test robustness to random global phase and additive sensor noise. Success is higher accuracy and lower variance across global phases, ideally matching the full complex-state baseline while exposing only magnitude observations.

## Verification

- Status: mech_ok_no_baseline
- Mechanism evidence: yes
- Mechanism confirmed: yes
- Practical verdict: harms
- Verdict: Built a fixed complex graph Schrödinger encoder that exposes only coordinate magnitudes over multiple times, with spectral diagnostics and a reproducible toy classification benchmark. For n=6, the Jacobian rank rose from 6 at K=1 to the predicted 11=2n−1 at K=2 and remained 11 thereafter; global-phase perturbation changed observations by only 6.7e-16. Increasing random potential scale from 0.02 to 1.0 increased median pair-sum separation from 9.6e-5 to 3.1e-2 and sigma_min(S) from 0.0026 to 0.106, while multi-time classification reached 0.878 versus 0.603 for a single-time magnitude baseline. The mechanism manifested, but this does not establish superiority over a trained GRU or full recurrent architecture.

### Mechanism check

- Verdict: Built a fixed complex graph Schrödinger encoder that exposes only coordinate magnitudes over multiple times, with spectral diagnostics and a reproducible toy classification benchmark. For n=6, the Jacobian rank rose from 6 at K=1 to the predicted 11=2n−1 at K=2 and remained 11 thereafter; global-phase perturbation changed observations by only 6.7e-16. Increasing random potential scale from 0.02 to 1.0 increased median pair-sum separation from 9.6e-5 to 3.1e-2 and sigma_min(S) from 0.0026 to 0.106, while multi-time classification reached 0.878 versus 0.603 for a single-time magnitude baseline. The mechanism manifested, but this does not establish superiority over a trained GRU or full recurrent architecture.
- Confidence: 8/10
- Limitations: The experiment used a path graph and a linear logistic readout, not a trained complex RNN, GRU, or graph-neural sequence model. It checked local differential identifiability rather than proving or numerically exhaustively testing global injectivity, and did not perform FLOP-matched recurrent training or broad graph/noise/time-grid sweeps.

### Practical benchmark

- Paired seeds: 8
- Baseline mean: 0
- Idea mean: 0
- p-value: 0.0081
- Paired wins: 0/8
- Benchmark verdict: idea worse (significant)

## Artifacts

- [bench_report.json](https://synthcore.org/code/572/bench_report.json)
- [prediction_report.json](https://synthcore.org/code/572/prediction_report.json)
- [report.md](https://synthcore.org/code/572/report.md)
- [report_bench_2026-08-31T182012.md](https://synthcore.org/code/572/report_bench_2026-08-31T182012.md)
- [results.json](https://synthcore.org/code/572/results.json)
- [spectral_phaseless_experiment.py](https://synthcore.org/code/572/spectral_phaseless_experiment.py)
- [stage2_bench.py](https://synthcore.org/code/572/stage2_bench.py)
- [Download all files as ZIP](https://synthcore.org/download/572)

## Disclaimer

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